From 86d8a2d184dcd902a5bf0b25099e67471f8e7b08 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 15 Feb 2021 12:11:01 +0100
Subject: [PATCH 001/342] introduced namespaces & common superclass for BEs

---
 CMakeLists.txt                       |   3 +
 src/thorin/CMakeLists.txt            |  12 +-
 src/thorin/be/backends.cpp           | 201 ++++++++++++++++++++++++++
 src/thorin/be/backends.h             |  62 ++++++++
 src/thorin/be/{ => c}/c.cpp          |   4 +-
 src/thorin/be/{ => c}/c.h            |   7 +-
 src/thorin/be/{llvm => c}/opencl.cpp |  11 +-
 src/thorin/be/{llvm => c}/opencl.h   |   9 +-
 src/thorin/be/llvm/amdgpu.cpp        |   2 +-
 src/thorin/be/llvm/amdgpu.h          |   4 +
 src/thorin/be/llvm/cpu.cpp           |   2 +-
 src/thorin/be/llvm/cpu.h             |   2 +-
 src/thorin/be/llvm/cuda.cpp          |   2 +-
 src/thorin/be/llvm/cuda.h            |   2 +-
 src/thorin/be/llvm/hls.cpp           |   2 +-
 src/thorin/be/llvm/hls.h             |   2 +-
 src/thorin/be/llvm/llvm.cpp          | 204 ++-------------------------
 src/thorin/be/llvm/llvm.h            |  33 +----
 src/thorin/be/llvm/nvvm.cpp          |   2 +-
 src/thorin/be/llvm/nvvm.h            |   4 +
 src/thorin/be/llvm/parallel.cpp      |   8 +-
 src/thorin/be/llvm/runtime.cpp       |   2 +-
 src/thorin/be/llvm/runtime.h         |  14 +-
 src/thorin/be/llvm/vectorize.cpp     |   2 +-
 src/thorin/config.h.in               |   1 +
 25 files changed, 328 insertions(+), 269 deletions(-)
 create mode 100644 src/thorin/be/backends.cpp
 create mode 100644 src/thorin/be/backends.h
 rename src/thorin/be/{ => c}/c.cpp (99%)
 rename src/thorin/be/{ => c}/c.h (85%)
 rename src/thorin/be/{llvm => c}/opencl.cpp (53%)
 rename src/thorin/be/{llvm => c}/opencl.h (55%)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index d60f032e2..4c7d8b0f7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -54,6 +54,9 @@ endif()
 if(THORIN_PROFILE)
     set(THORIN_ENABLE_PROFILING TRUE)
 endif()
+if(LLVM_FOUND)
+    set(THORIN_ENABLE_LLVM TRUE)
+endif()
 if(RV_FOUND)
     set(THORIN_ENABLE_RV TRUE)
 endif()
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 6d15a8cd2..15e11073e 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -30,8 +30,8 @@ set(THORIN_SOURCES
     analyses/scope.h
     analyses/verify.cpp
     analyses/verify.h
-    be/c.cpp
-    be/c.h
+    be/c/c.cpp
+    be/c/c.h
     be/kernel_config.h
     tables/allnodes.h
     tables/arithoptable.h
@@ -82,7 +82,9 @@ set(THORIN_SOURCES
     util/symbol.h
     util/types.h
     util/utility.h
-)
+    be/backends.cpp
+    be/backends.h
+    )
 
 if(LLVM_FOUND)
     list(APPEND THORIN_SOURCES
@@ -98,8 +100,8 @@ if(LLVM_FOUND)
         be/llvm/amdgpu.h
         be/llvm/nvvm.cpp
         be/llvm/nvvm.h
-        #be/llvm/opencl.cpp
-        #be/llvm/opencl.h
+            be/c/opencl.cpp
+            be/c/opencl.h
         be/llvm/parallel.cpp
         be/llvm/runtime.inc
         be/llvm/runtime.cpp
diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
new file mode 100644
index 000000000..6349c4e90
--- /dev/null
+++ b/src/thorin/be/backends.cpp
@@ -0,0 +1,201 @@
+#include "backends.h"
+
+#include "thorin/analyses/scope.h"
+
+#ifdef THORIN_ENABLE_LLVM
+#include "thorin/be/llvm/cpu.h"
+#include "thorin/be/llvm/nvvm.h"
+#include "thorin/be/llvm/amdgpu.h"
+#include "thorin/be/llvm/cuda.h"
+#include "thorin/be/llvm/hls.h"
+#include "thorin/be/c/opencl.h"
+#include "thorin/transform/codegen_prepare.h"
+#endif
+
+namespace thorin {
+
+static void get_kernel_configs(Importer& importer,
+                               const std::vector<Continuation*>& kernels,
+                               Cont2Config& kernel_config,
+                               std::function<std::unique_ptr<KernelConfig> (Continuation*, Continuation*)> use_callback)
+{
+    importer.world().opt();
+
+    auto exported_continuations = importer.world().exported_continuations();
+    for (auto continuation : kernels) {
+        // recover the imported continuation (lost after the call to opt)
+        Continuation* imported = nullptr;
+        for (auto exported : exported_continuations) {
+            if (exported->name() == continuation->name())
+                imported = exported;
+        }
+        if (!imported) continue;
+
+        visit_uses(continuation, [&] (Continuation* use) {
+            auto config = use_callback(use, imported);
+            if (config) {
+                auto p = kernel_config.emplace(imported, std::move(config));
+                assert_unused(p.second && "single kernel config entry expected");
+            }
+            return false;
+        }, true);
+
+        continuation->destroy_body();
+    }
+}
+
+static const Continuation* get_alloc_call(const Def* def) {
+    // look through casts
+    while (auto conv_op = def->isa<ConvOp>())
+        def = conv_op->op(0);
+
+    auto param = def->isa<Param>();
+    if (!param) return nullptr;
+
+    auto ret = param->continuation();
+    if (ret->num_uses() != 1) return nullptr;
+
+    auto use = *(ret->uses().begin());
+    auto call = use.def()->isa_continuation();
+    if (!call || use.index() == 0) return nullptr;
+
+    auto callee = call->callee();
+    if (callee->name() != "anydsl_alloc") return nullptr;
+
+    return call;
+}
+
+static uint64_t get_alloc_size(const Def* def) {
+    auto call = get_alloc_call(def);
+    if (!call) return 0;
+
+    // signature: anydsl_alloc(mem, i32, i64, fn(mem, &[i8]))
+    auto size = call->arg(2)->isa<PrimLit>();
+    return size ? static_cast<uint64_t>(size->value().get_qu64()) : 0_u64;
+}
+
+Backends::Backends(World& world, int opt, bool debug)
+: cuda(world)
+, nvvm(world)
+, opencl(world)
+, amdgpu(world)
+, hls(world)
+{
+    // determine different parts of the world which need to be compiled differently
+    Scope::for_each(world, [&] (const Scope& scope) {
+        auto continuation = scope.entry();
+        Continuation* imported = nullptr;
+        if (is_passed_to_intrinsic(continuation, Intrinsic::CUDA))
+            imported = cuda.import(continuation)->as_continuation();
+        else if (is_passed_to_intrinsic(continuation, Intrinsic::NVVM))
+            imported = nvvm.import(continuation)->as_continuation();
+        else if (is_passed_to_intrinsic(continuation, Intrinsic::OpenCL))
+            imported = opencl.import(continuation)->as_continuation();
+        else if (is_passed_to_intrinsic(continuation, Intrinsic::AMDGPU))
+            imported = amdgpu.import(continuation)->as_continuation();
+        else if (is_passed_to_intrinsic(continuation, Intrinsic::HLS))
+            imported = hls.import(continuation)->as_continuation();
+        else
+            return;
+
+        imported->set_name(continuation->unique_name());
+        imported->make_exported();
+        continuation->set_name(continuation->unique_name());
+
+        for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
+            imported->param(i)->set_name(continuation->param(i)->unique_name());
+
+        kernels.emplace_back(continuation);
+    });
+
+    // get the GPU kernel configurations
+    if (!cuda.world().empty()   ||
+        !nvvm.world().empty()   ||
+        !opencl.world().empty() ||
+        !amdgpu.world().empty()) {
+        auto get_gpu_config = [&] (Continuation* use, Continuation* /* imported */) {
+            // determine whether or not this kernel uses restrict pointers
+            bool has_restrict = true;
+            DefSet allocs;
+            for (size_t i = LaunchArgs::Num, e = use->num_args(); has_restrict && i != e; ++i) {
+                auto arg = use->arg(i);
+                if (!arg->type()->isa<PtrType>()) continue;
+                auto alloc = get_alloc_call(arg);
+                if (!alloc) has_restrict = false;
+                auto p = allocs.insert(alloc);
+                has_restrict &= p.second;
+            }
+
+            auto it_config = use->arg(LaunchArgs::Config)->as<Tuple>();
+            if (it_config->op(0)->isa<PrimLit>() &&
+                it_config->op(1)->isa<PrimLit>() &&
+                it_config->op(2)->isa<PrimLit>()) {
+                return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> {
+                it_config->op(0)->as<PrimLit>()->qu32_value().data(),
+                it_config->op(1)->as<PrimLit>()->qu32_value().data(),
+                it_config->op(2)->as<PrimLit>()->qu32_value().data()
+                }, has_restrict);
+            }
+            return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> { -1, -1, -1 }, has_restrict);
+        };
+        get_kernel_configs(cuda,   kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(nvvm,   kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(opencl, kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(amdgpu, kernels, kernel_config, get_gpu_config);
+    }
+
+    // get the HLS kernel configurations
+    if (!hls.world().empty()) {
+        auto get_hls_config = [&] (Continuation* use, Continuation* imported) {
+            HLSKernelConfig::Param2Size param_sizes;
+            for (size_t i = 3, e = use->num_args(); i != e; ++i) {
+                auto arg = use->arg(i);
+                auto ptr_type = arg->type()->isa<PtrType>();
+                if (!ptr_type) continue;
+                auto size = get_alloc_size(arg);
+                if (size == 0)
+                    world.edef(arg, "array size is not known at compile time");
+                auto elem_type = ptr_type->pointee();
+                size_t multiplier = 1;
+                if (!elem_type->isa<PrimType>()) {
+                    if (auto array_type = elem_type->isa<ArrayType>())
+                        elem_type = array_type->elem_type();
+                }
+                if (!elem_type->isa<PrimType>()) {
+                    if (auto def_array_type = elem_type->isa<DefiniteArrayType>()) {
+                        elem_type = def_array_type->elem_type();
+                        multiplier = def_array_type->dim();
+                    }
+                }
+                auto prim_type = elem_type->isa<PrimType>();
+                if (!prim_type)
+                    world.edef(arg, "only pointers to arrays of primitive types are supported");
+                auto num_elems = size / (multiplier * num_bits(prim_type->primtype_tag()) / 8);
+                // imported has type: fn (mem, fn (mem), ...)
+                param_sizes.emplace(imported->param(i - 3 + 2), num_elems);
+            }
+            return std::make_unique<HLSKernelConfig>(param_sizes);
+        };
+        get_kernel_configs(hls, kernels, kernel_config, get_hls_config);
+    }
+
+#ifdef THORIN_ENABLE_LLVM
+    cpu_cg = std::make_unique<llvm_be::CPUCodeGen>(world, opt, debug);
+
+    if (!nvvm.  world().empty()) nvvm_cg   = std::make_unique<llvm_be::NVVMCodeGen  >(nvvm  .world(), kernel_config,      debug);
+    if (!amdgpu.world().empty()) amdgpu_cg = std::make_unique<llvm_be::AMDGPUCodeGen>(amdgpu.world(), kernel_config, opt, debug);
+#else
+    // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
+#endif
+    // TODO: Fix the C-based backends
+    //if (!cuda.  world().empty()) cuda_cg   = std::make_unique<CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
+    //if (!opencl.world().empty()) opencl_cg = std::make_unique<OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
+    //if (!hls.   world().empty()) hls_cg    = std::make_unique<HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
+}
+
+CodeGen::CodeGen(World& world, bool debug)
+: world_(world)
+, debug_(debug)
+{}
+
+}
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
new file mode 100644
index 000000000..b9c309889
--- /dev/null
+++ b/src/thorin/be/backends.h
@@ -0,0 +1,62 @@
+#ifndef THORIN_BACKENDS_H
+#define THORIN_BACKENDS_H
+
+#include "thorin/transform/importer.h"
+#include "thorin/be/kernel_config.h"
+
+namespace thorin {
+
+class CodeGen {
+protected:
+    CodeGen(World& world, bool debug);
+public:
+    virtual void emit(std::ostream& stream) = 0;
+
+    /// @name getters
+    //@{
+    World& world() const { return world_; }
+    bool debug() const { return debug_; }
+    //@}
+
+private:
+    World& world_;
+    bool debug_;
+};
+
+struct LaunchArgs {
+    enum {
+        Mem = 0,
+        Device,
+        Space,
+        Config,
+        Body,
+        Return,
+        Num
+    };
+};
+
+struct Backends {
+    Backends(World& world, int opt, bool debug);
+
+    Cont2Config kernel_config;
+    std::vector<Continuation*> kernels;
+
+    // TODO use arrays + loops for this
+    Importer cuda;
+    Importer nvvm;
+    Importer opencl;
+    Importer amdgpu;
+    Importer hls;
+
+    // TODO use arrays + loops for this
+    std::unique_ptr<CodeGen> cpu_cg;
+    std::unique_ptr<CodeGen> cuda_cg;
+    std::unique_ptr<CodeGen> nvvm_cg;
+    std::unique_ptr<CodeGen> opencl_cg;
+    std::unique_ptr<CodeGen> amdgpu_cg;
+    std::unique_ptr<CodeGen> hls_cg;
+};
+
+}
+
+#endif
diff --git a/src/thorin/be/c.cpp b/src/thorin/be/c/c.cpp
similarity index 99%
rename from src/thorin/be/c.cpp
rename to src/thorin/be/c/c.cpp
index 2c6c580af..0a822ed83 100644
--- a/src/thorin/be/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -7,7 +7,7 @@
 #include "thorin/analyses/schedule.h"
 #include "thorin/analyses/scope.h"
 #include "thorin/util/stream.h"
-#include "thorin/be/c.h"
+#include "c.h"
 
 #include <cmath>
 #include <sstream>
@@ -15,7 +15,7 @@
 #include <cctype>
 #include <variant>
 
-namespace thorin {
+namespace thorin::c_be {
 
 class CCodeGen;
 
diff --git a/src/thorin/be/c.h b/src/thorin/be/c/c.h
similarity index 85%
rename from src/thorin/be/c.h
rename to src/thorin/be/c/c.h
index 54bfcc944..01fe38354 100644
--- a/src/thorin/be/c.h
+++ b/src/thorin/be/c/c.h
@@ -4,13 +4,14 @@
 #include <cstdint>
 #include <iostream>
 
-#include "thorin/world.h"
 #include "thorin/be/kernel_config.h"
 
 namespace thorin {
 
 class World;
 
+namespace c_be {
+
 enum class Lang : uint8_t {
     C99,        ///< Flag for C99
     HLS,        ///< Flag for HLS
@@ -19,7 +20,9 @@ enum class Lang : uint8_t {
 };
 
 void emit_c(World&, const Cont2Config& kernel_config, std::ostream& stream, Lang lang, bool debug);
-void emit_c_int(World&, std::ostream& stream);
+void emit_c_int(World&, Stream& stream);
+
+}
 
 }
 
diff --git a/src/thorin/be/llvm/opencl.cpp b/src/thorin/be/c/opencl.cpp
similarity index 53%
rename from src/thorin/be/llvm/opencl.cpp
rename to src/thorin/be/c/opencl.cpp
index 225123993..c93910f55 100644
--- a/src/thorin/be/llvm/opencl.cpp
+++ b/src/thorin/be/c/opencl.cpp
@@ -1,22 +1,21 @@
-#include "thorin/be/llvm/opencl.h"
+#include "opencl.h"
 
 #include <fstream>
 #include <stdexcept>
 
 #include "thorin/primop.h"
 #include "thorin/world.h"
-#include "thorin/be/c.h"
+#include "thorin/be/c/c.h"
 
-
-namespace thorin {
+namespace thorin::c_be {
 
 OpenCLCodeGen::OpenCLCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
+    : CodeGen(world, debug)
     , kernel_config_(kernel_config)
 {}
 
 void OpenCLCodeGen::emit(std::ostream& stream) {
-    thorin::emit_c(world(), kernel_config_, stream, Lang::OPENCL, debug());
+    emit_c(world(), kernel_config_, stream, Lang::OPENCL, debug());
 }
 
 }
diff --git a/src/thorin/be/llvm/opencl.h b/src/thorin/be/c/opencl.h
similarity index 55%
rename from src/thorin/be/llvm/opencl.h
rename to src/thorin/be/c/opencl.h
index 87815bb23..3b472777e 100644
--- a/src/thorin/be/llvm/opencl.h
+++ b/src/thorin/be/c/opencl.h
@@ -1,19 +1,18 @@
 #ifndef THORIN_BE_LLVM_OPENCL_H
 #define THORIN_BE_LLVM_OPENCL_H
 
-#include "thorin/be/llvm/llvm.h"
+#include "../backends.h"
+#include "../llvm/llvm.h"
 
-namespace thorin {
+namespace thorin::c_be {
 
-class OpenCLCodeGen : public CodeGen {
+class OpenCLCodeGen : public thorin::CodeGen {
 public:
     OpenCLCodeGen(World& world, const Cont2Config&, int opt, bool debug);
 
     void emit(std::ostream& stream) override;
 
 protected:
-    virtual std::string get_alloc_name() const override { THORIN_UNREACHABLE; /*alloc not supported in OpenCL*/; }
-
     const Cont2Config& kernel_config_;
 };
 
diff --git a/src/thorin/be/llvm/amdgpu.cpp b/src/thorin/be/llvm/amdgpu.cpp
index a4dd61954..859948be4 100644
--- a/src/thorin/be/llvm/amdgpu.cpp
+++ b/src/thorin/be/llvm/amdgpu.cpp
@@ -3,7 +3,7 @@
 #include "thorin/primop.h"
 #include "thorin/world.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 AMDGPUCodeGen::AMDGPUCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
diff --git a/src/thorin/be/llvm/amdgpu.h b/src/thorin/be/llvm/amdgpu.h
index ce1f97d9a..33866fc63 100644
--- a/src/thorin/be/llvm/amdgpu.h
+++ b/src/thorin/be/llvm/amdgpu.h
@@ -7,6 +7,8 @@ namespace thorin {
 
 class Load;
 
+namespace llvm_be {
+
 class AMDGPUCodeGen : public CodeGen {
 public:
     AMDGPUCodeGen(World& world, const Cont2Config&, int opt, bool debug);
@@ -23,4 +25,6 @@ class AMDGPUCodeGen : public CodeGen {
 
 }
 
+}
+
 #endif
diff --git a/src/thorin/be/llvm/cpu.cpp b/src/thorin/be/llvm/cpu.cpp
index d98e394a2..ed6e92772 100644
--- a/src/thorin/be/llvm/cpu.cpp
+++ b/src/thorin/be/llvm/cpu.cpp
@@ -8,7 +8,7 @@
 #include <llvm/Target/TargetMachine.h>
 #include <llvm/Target/TargetOptions.h>
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 CPUCodeGen::CPUCodeGen(World& world, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
diff --git a/src/thorin/be/llvm/cpu.h b/src/thorin/be/llvm/cpu.h
index 0a348a3e4..7607cdcb6 100644
--- a/src/thorin/be/llvm/cpu.h
+++ b/src/thorin/be/llvm/cpu.h
@@ -3,7 +3,7 @@
 
 #include "thorin/be/llvm/llvm.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 class CPUCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/cuda.cpp b/src/thorin/be/llvm/cuda.cpp
index 1711f43a1..28b9824e7 100644
--- a/src/thorin/be/llvm/cuda.cpp
+++ b/src/thorin/be/llvm/cuda.cpp
@@ -7,7 +7,7 @@
 #include "thorin/world.h"
 #include "thorin/be/c.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 CUDACodeGen::CUDACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
diff --git a/src/thorin/be/llvm/cuda.h b/src/thorin/be/llvm/cuda.h
index bc2d84321..3fe4472eb 100644
--- a/src/thorin/be/llvm/cuda.h
+++ b/src/thorin/be/llvm/cuda.h
@@ -3,7 +3,7 @@
 
 #include "thorin/be/llvm/llvm.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 class CUDACodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/hls.cpp b/src/thorin/be/llvm/hls.cpp
index 3427b464b..125722774 100644
--- a/src/thorin/be/llvm/hls.cpp
+++ b/src/thorin/be/llvm/hls.cpp
@@ -7,7 +7,7 @@
 #include "thorin/world.h"
 #include "thorin/be/c.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 HLSCodeGen::HLSCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
diff --git a/src/thorin/be/llvm/hls.h b/src/thorin/be/llvm/hls.h
index 5b033fef0..7143a675e 100644
--- a/src/thorin/be/llvm/hls.h
+++ b/src/thorin/be/llvm/hls.h
@@ -3,7 +3,7 @@
 
 #include "thorin/be/llvm/llvm.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 class HLSCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 43101097c..5f267f947 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -36,27 +36,19 @@
 #include "thorin/type.h"
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
-#include "thorin/be/llvm/cpu.h"
-#include "thorin/be/llvm/nvvm.h"
-#include "thorin/be/llvm/amdgpu.h"
-#include "thorin/be/llvm/cuda.h"
-#include "thorin/be/llvm/hls.h"
-#include "thorin/be/llvm/opencl.h"
-#include "thorin/transform/codegen_prepare.h"
 #include "thorin/util/array.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 CodeGen::CodeGen(World& world,
                  llvm::CallingConv::ID function_calling_convention,
                  llvm::CallingConv::ID device_calling_convention,
                  llvm::CallingConv::ID kernel_calling_convention,
                  int opt, bool debug)
-    : world_(world)
+    : thorin::CodeGen(world, debug)
     , context_(new llvm::LLVMContext())
     , module_(new llvm::Module(world.name(), *context_))
     , opt_(opt)
-    , debug_(debug)
     , dibuilder_(module())
     , function_calling_convention_(function_calling_convention)
     , device_calling_convention_(device_calling_convention)
@@ -183,7 +175,7 @@ llvm::Type* CodeGen::convert(const Type* type) {
                 return types_[type] = llvm_type;
             }
 
-            auto env_type = convert(Closure::environment_type(world_));
+            auto env_type = convert(Closure::environment_type(world()));
             ops.push_back(env_type);
             auto fn_type = llvm::FunctionType::get(ret, ops, false);
             auto ptr_type = llvm::PointerType::get(fn_type, 0);
@@ -288,10 +280,10 @@ std::unique_ptr<llvm::Module>& CodeGen::emit() {
         // Darwin only supports dwarf2
         if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin())
             module().addModuleFlag(llvm::Module::Warning, "Dwarf Version", 2);
-        dicompile_unit_ = dibuilder_.createCompileUnit(llvm::dwarf::DW_LANG_C, dibuilder_.createFile(world_.name(), llvm::StringRef()), "Impala", opt() > 0, llvm::StringRef(), 0);
+        dicompile_unit_ = dibuilder_.createCompileUnit(llvm::dwarf::DW_LANG_C, dibuilder_.createFile(world().name(), llvm::StringRef()), "Impala", opt() > 0, llvm::StringRef(), 0);
     }
 
-    Scope::for_each(world_, [&] (const Scope& scope) { emit(scope); });
+    Scope::for_each(world(), [&] (const Scope& scope) { emit(scope); });
 
     if (debug()) dibuilder_.finalize();
 
@@ -785,15 +777,15 @@ llvm::Value* CodeGen::emit_(const Def* def) {
             llvm::Value* env = nullptr;
             if (is_thin(closure->op(1)->type())) {
                 if (is_type_unit(val->type())) {
-                    env = emit(world_.bottom(Closure::environment_type(world_)));
+                    env = emit(world().bottom(Closure::environment_type(world())));
                 } else {
-                    env = emit(world_.cast(Closure::environment_type(world_), val));
+                    env = emit(world().cast(Closure::environment_type(world()), val));
                 }
             } else {
                 world().wdef(def, "closure '{}' is leaking memory, type '{}' is too large", def, agg->op(1)->type());
                 auto alloc = emit_alloc(irbuilder, val->type(), nullptr);
                 irbuilder.CreateStore(emit(val), alloc);
-                env = irbuilder.CreatePtrToInt(alloc, convert(Closure::environment_type(world_)));
+                env = irbuilder.CreatePtrToInt(alloc, convert(Closure::environment_type(world())));
             }
             llvm_agg = irbuilder.CreateInsertValue(llvm_agg, closure_fn, 0);
             llvm_agg = irbuilder.CreateInsertValue(llvm_agg, env, 1);
@@ -946,7 +938,7 @@ llvm::Value* CodeGen::emit_(const Def* def) {
     if (auto vector = def->isa<Vector>()) {
         llvm::Value* vec = llvm::UndefValue::get(convert(vector->type()));
         for (size_t i = 0, e = vector->num_ops(); i != e; ++i)
-            vec = irbuilder.CreateInsertElement(vec, emit(vector->op(i)), emit(world_.literal_pu32(i, vector->loc())));
+            vec = irbuilder.CreateInsertElement(vec, emit(vector->op(i)), emit(world().literal_pu32(i, vector->loc())));
 
         return vec;
     }
@@ -1295,182 +1287,4 @@ llvm::Value* CodeGen::create_tmp_alloca(llvm::IRBuilder<>& irbuilder, llvm::Type
 
 //------------------------------------------------------------------------------
 
-static void get_kernel_configs(Importer& importer,
-    const std::vector<Continuation*>& kernels,
-    Cont2Config& kernel_config,
-    std::function<std::unique_ptr<KernelConfig> (Continuation*, Continuation*)> use_callback)
-{
-    importer.world().opt();
-
-    auto exported_continuations = importer.world().exported_continuations();
-    for (auto continuation : kernels) {
-        // recover the imported continuation (lost after the call to opt)
-        Continuation* imported = nullptr;
-        for (auto exported : exported_continuations) {
-            if (exported->name() == continuation->name())
-                imported = exported;
-        }
-        if (!imported) continue;
-
-        visit_uses(continuation, [&] (Continuation* use) {
-            auto config = use_callback(use, imported);
-            if (config) {
-                auto p = kernel_config.emplace(imported, std::move(config));
-                assert_unused(p.second && "single kernel config entry expected");
-            }
-            return false;
-        }, true);
-
-        continuation->destroy_body();
-    }
-}
-
-static const Continuation* get_alloc_call(const Def* def) {
-    // look through casts
-    while (auto conv_op = def->isa<ConvOp>())
-        def = conv_op->op(0);
-
-    auto param = def->isa<Param>();
-    if (!param) return nullptr;
-
-    auto ret = param->continuation();
-    if (ret->num_uses() != 1) return nullptr;
-
-    auto use = *(ret->uses().begin());
-    auto call = use.def()->isa_continuation();
-    if (!call || use.index() == 0) return nullptr;
-
-    auto callee = call->callee();
-    if (callee->name() != "anydsl_alloc") return nullptr;
-
-    return call;
-}
-
-static uint64_t get_alloc_size(const Def* def) {
-    auto call = get_alloc_call(def);
-    if (!call) return 0;
-
-    // signature: anydsl_alloc(mem, i32, i64, fn(mem, &[i8]))
-    auto size = call->arg(2)->isa<PrimLit>();
-    return size ? static_cast<uint64_t>(size->value().get_qu64()) : 0_u64;
-}
-
-Backends::Backends(World& world, int opt, bool debug)
-    : cuda(world)
-    , nvvm(world)
-    , opencl(world)
-    , amdgpu(world)
-    , hls(world)
-{
-    // determine different parts of the world which need to be compiled differently
-    Scope::for_each(world, [&] (const Scope& scope) {
-        auto continuation = scope.entry();
-        Continuation* imported = nullptr;
-        if (is_passed_to_intrinsic(continuation, Intrinsic::CUDA))
-            imported = cuda.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::NVVM))
-            imported = nvvm.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::OpenCL))
-            imported = opencl.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::AMDGPU))
-            imported = amdgpu.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::HLS))
-            imported = hls.import(continuation)->as_continuation();
-        else
-            return;
-
-        imported->set_name(continuation->unique_name());
-        imported->make_exported();
-        continuation->set_name(continuation->unique_name());
-
-        for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
-            imported->param(i)->set_name(continuation->param(i)->unique_name());
-
-        kernels.emplace_back(continuation);
-    });
-
-    // get the GPU kernel configurations
-    if (!cuda.world().empty()   ||
-        !nvvm.world().empty()   ||
-        !opencl.world().empty() ||
-        !amdgpu.world().empty()) {
-        auto get_gpu_config = [&] (Continuation* use, Continuation* /* imported */) {
-            // determine whether or not this kernel uses restrict pointers
-            bool has_restrict = true;
-            DefSet allocs;
-            for (size_t i = LaunchArgs::Num, e = use->num_args(); has_restrict && i != e; ++i) {
-                auto arg = use->arg(i);
-                if (!arg->type()->isa<PtrType>()) continue;
-                auto alloc = get_alloc_call(arg);
-                if (!alloc) has_restrict = false;
-                auto p = allocs.insert(alloc);
-                has_restrict &= p.second;
-            }
-
-            auto it_config = use->arg(LaunchArgs::Config)->as<Tuple>();
-            if (it_config->op(0)->isa<PrimLit>() &&
-                it_config->op(1)->isa<PrimLit>() &&
-                it_config->op(2)->isa<PrimLit>()) {
-                return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> {
-                    it_config->op(0)->as<PrimLit>()->qu32_value().data(),
-                    it_config->op(1)->as<PrimLit>()->qu32_value().data(),
-                    it_config->op(2)->as<PrimLit>()->qu32_value().data()
-                }, has_restrict);
-            }
-            return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> { -1, -1, -1 }, has_restrict);
-        };
-        get_kernel_configs(cuda,   kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(nvvm,   kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(opencl, kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(amdgpu, kernels, kernel_config, get_gpu_config);
-    }
-
-    // get the HLS kernel configurations
-    if (!hls.world().empty()) {
-        auto get_hls_config = [&] (Continuation* use, Continuation* imported) {
-            HLSKernelConfig::Param2Size param_sizes;
-            for (size_t i = 3, e = use->num_args(); i != e; ++i) {
-                auto arg = use->arg(i);
-                auto ptr_type = arg->type()->isa<PtrType>();
-                if (!ptr_type) continue;
-                auto size = get_alloc_size(arg);
-                if (size == 0)
-                    world.edef(arg, "array size is not known at compile time");
-                auto elem_type = ptr_type->pointee();
-                size_t multiplier = 1;
-                if (!elem_type->isa<PrimType>()) {
-                    if (auto array_type = elem_type->isa<ArrayType>())
-                        elem_type = array_type->elem_type();
-                }
-                if (!elem_type->isa<PrimType>()) {
-                    if (auto def_array_type = elem_type->isa<DefiniteArrayType>()) {
-                        elem_type = def_array_type->elem_type();
-                        multiplier = def_array_type->dim();
-                    }
-                }
-                auto prim_type = elem_type->isa<PrimType>();
-                if (!prim_type)
-                    world.edef(arg, "only pointers to arrays of primitive types are supported");
-                auto num_elems = size / (multiplier * num_bits(prim_type->primtype_tag()) / 8);
-                // imported has type: fn (mem, fn (mem), ...)
-                param_sizes.emplace(imported->param(i - 3 + 2), num_elems);
-            }
-            return std::make_unique<HLSKernelConfig>(param_sizes);
-        };
-        get_kernel_configs(hls, kernels, kernel_config, get_hls_config);
-    }
-
-    cpu_cg = std::make_unique<CPUCodeGen>(world, opt, debug);
-
-    if (!nvvm.  world().empty()) nvvm_cg   = std::make_unique<NVVMCodeGen  >(nvvm  .world(), kernel_config,      debug);
-    if (!amdgpu.world().empty()) amdgpu_cg = std::make_unique<AMDGPUCodeGen>(amdgpu.world(), kernel_config, opt, debug);
-
-    // TODO
-    //if (!cuda.  world().empty()) cuda_cg   = std::make_unique<CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
-    //if (!opencl.world().empty()) opencl_cg = std::make_unique<OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
-    //if (!hls.   world().empty()) hls_cg    = std::make_unique<HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
-}
-
-//------------------------------------------------------------------------------
-
 }
diff --git a/src/thorin/be/llvm/llvm.h b/src/thorin/be/llvm/llvm.h
index 855249b65..7f41d51e8 100644
--- a/src/thorin/be/llvm/llvm.h
+++ b/src/thorin/be/llvm/llvm.h
@@ -9,6 +9,7 @@
 #include "thorin/config.h"
 #include "thorin/continuation.h"
 #include "thorin/analyses/schedule.h"
+#include "thorin/be/backends.h"
 #include "thorin/be/llvm/runtime.h"
 #include "thorin/be/kernel_config.h"
 #include "thorin/transform/importer.h"
@@ -17,7 +18,9 @@ namespace thorin {
 
 class World;
 
-class CodeGen {
+namespace llvm_be {
+
+class CodeGen : public thorin::CodeGen {
 protected:
     CodeGen(World& world,
             llvm::CallingConv::ID function_calling_convention,
@@ -29,15 +32,13 @@ class CodeGen {
 
     /// @name getters
     //@{
-    World& world() const { return world_; }
     llvm::LLVMContext& context() { return *context_; }
     llvm::Module& module() { return *module_; }
     const llvm::Module& module() const { return *module_; }
-    virtual void emit(std::ostream& stream);
     int opt() const { return opt_; }
-    bool debug() const { return debug_; }
     //@}
 
+    void emit(std::ostream& stream) override;
     std::unique_ptr<llvm::Module>& emit();
 
 protected:
@@ -96,11 +97,9 @@ class CodeGen {
     void emit_vectorize(u32, llvm::Function*, llvm::CallInst*);
     void emit_phi_arg(llvm::IRBuilder<>&, const Param*, llvm::Value*);
 
-    World& world_;
     std::unique_ptr<llvm::LLVMContext> context_;
     std::unique_ptr<llvm::Module> module_;
     int opt_;
-    bool debug_;
 
 protected:
     std::unique_ptr<llvm::TargetMachine> machine_;
@@ -128,27 +127,7 @@ class CodeGen {
 template<class T>
 llvm::ArrayRef<T> llvm_ref(const Array<T>& array) { return llvm::ArrayRef<T>(array.begin(), array.end()); }
 
-struct Backends {
-    Backends(World& world, int opt, bool debug);
-
-    Cont2Config kernel_config;
-    std::vector<Continuation*> kernels;
-
-    // TODO use arrays + loops for this
-    Importer cuda;
-    Importer nvvm;
-    Importer opencl;
-    Importer amdgpu;
-    Importer hls;
-
-    // TODO use arrays + loops for this
-    std::unique_ptr<CodeGen> cpu_cg;
-    std::unique_ptr<CodeGen> cuda_cg;
-    std::unique_ptr<CodeGen> nvvm_cg;
-    std::unique_ptr<CodeGen> opencl_cg;
-    std::unique_ptr<CodeGen> amdgpu_cg;
-    std::unique_ptr<CodeGen> hls_cg;
-};
+} // namespace llvm_be
 
 } // namespace thorin
 
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index ab3923f16..1fd8c7818 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -16,7 +16,7 @@
 #include "thorin/primop.h"
 #include "thorin/world.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 NVVMCodeGen::NVVMCodeGen(World& world, const Cont2Config& kernel_config, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::PTX_Device, llvm::CallingConv::PTX_Kernel, 0, debug)
diff --git a/src/thorin/be/llvm/nvvm.h b/src/thorin/be/llvm/nvvm.h
index ee1f84485..f637304c4 100644
--- a/src/thorin/be/llvm/nvvm.h
+++ b/src/thorin/be/llvm/nvvm.h
@@ -7,6 +7,8 @@ namespace thorin {
 
 class Load;
 
+namespace llvm_be {
+
 class NVVMCodeGen : public CodeGen {
 public:
     NVVMCodeGen(World& world, const Cont2Config&, bool debug); // NVVM-specific optimizations are run in the runtime
@@ -37,4 +39,6 @@ class NVVMCodeGen : public CodeGen {
 
 }
 
+}
+
 #endif
diff --git a/src/thorin/be/llvm/parallel.cpp b/src/thorin/be/llvm/parallel.cpp
index 764a7869f..128728fad 100644
--- a/src/thorin/be/llvm/parallel.cpp
+++ b/src/thorin/be/llvm/parallel.cpp
@@ -1,6 +1,6 @@
 #include "thorin/be/llvm/llvm.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 enum {
     PAR_ARG_MEM,
@@ -31,7 +31,7 @@ Continuation* CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation*
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world_.tuple_type(continuation->arg_fn_type()->ops().skip_front(PAR_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->ops().skip_front(PAR_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -119,7 +119,7 @@ Continuation* CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* c
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world_.tuple_type(continuation->arg_fn_type()->ops().skip_front(FIB_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->ops().skip_front(FIB_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -195,7 +195,7 @@ Continuation* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* co
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world_.tuple_type(continuation->arg_fn_type()->ops().skip_front(SPAWN_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->ops().skip_front(SPAWN_NUM_ARGS)));
     llvm::Value* closure = nullptr;
     if (closure_type->isStructTy()) {
         closure = llvm::UndefValue::get(closure_type);
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index 7344605a0..b7b38070c 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -13,7 +13,7 @@
 #include "thorin/be/llvm/llvm.h"
 #include "thorin/be/llvm/runtime.inc"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 Runtime::Runtime(llvm::LLVMContext& context,
                  llvm::Module& target)
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index 84a7cb78d..c125deed5 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -8,22 +8,10 @@
 
 #include "thorin/world.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 class CodeGen;
 
-struct LaunchArgs {
-    enum {
-        Mem = 0,
-        Device,
-        Space,
-        Config,
-        Body,
-        Return,
-        Num
-    };
-};
-
 class Runtime {
 public:
     Runtime(llvm::LLVMContext&, llvm::Module& target);
diff --git a/src/thorin/be/llvm/vectorize.cpp b/src/thorin/be/llvm/vectorize.cpp
index 0a9e5e578..529ac844b 100644
--- a/src/thorin/be/llvm/vectorize.cpp
+++ b/src/thorin/be/llvm/vectorize.cpp
@@ -36,7 +36,7 @@
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
 
-namespace thorin {
+namespace thorin::llvm_be {
 
 struct VectorizeArgs {
     enum {
diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index bb6f494ec..5e061fe28 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -3,6 +3,7 @@
 
 #cmakedefine01 THORIN_ENABLE_CHECKS
 #cmakedefine01 THORIN_ENABLE_PROFILING
+#cmakedefine01 THORIN_ENABLE_LLVM
 #cmakedefine01 THORIN_ENABLE_RV
 
 #endif

From 07127e553e418b4da1a6b7cb74b28e18010e7678 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 15:40:14 +0100
Subject: [PATCH 002/342] make the C-based backends not based on
 llvm_be::CodeGen

---
 src/thorin/CMakeLists.txt   | 12 ++++++------
 src/thorin/be/backends.cpp  | 14 ++++++--------
 src/thorin/be/c/c.cpp       |  5 +++--
 src/thorin/be/c/c.h         | 19 +++++++++++++++++--
 src/thorin/be/c/cuda.cpp    |  9 +++++++++
 src/thorin/be/c/cuda.h      | 15 +++++++++++++++
 src/thorin/be/c/hls.cpp     |  9 +++++++++
 src/thorin/be/c/hls.h       | 15 +++++++++++++++
 src/thorin/be/c/opencl.cpp  | 14 +-------------
 src/thorin/be/c/opencl.h    | 10 ++--------
 src/thorin/be/llvm/cuda.cpp | 21 ---------------------
 src/thorin/be/llvm/cuda.h   | 22 ----------------------
 src/thorin/be/llvm/hls.cpp  | 21 ---------------------
 src/thorin/be/llvm/hls.h    | 22 ----------------------
 14 files changed, 83 insertions(+), 125 deletions(-)
 create mode 100644 src/thorin/be/c/cuda.cpp
 create mode 100644 src/thorin/be/c/cuda.h
 create mode 100644 src/thorin/be/c/hls.cpp
 create mode 100644 src/thorin/be/c/hls.h
 delete mode 100644 src/thorin/be/llvm/cuda.cpp
 delete mode 100644 src/thorin/be/llvm/cuda.h
 delete mode 100644 src/thorin/be/llvm/hls.cpp
 delete mode 100644 src/thorin/be/llvm/hls.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 15e11073e..b0c18ea7c 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -32,6 +32,12 @@ set(THORIN_SOURCES
     analyses/verify.h
     be/c/c.cpp
     be/c/c.h
+    be/c/cuda.cpp
+    be/c/cuda.h
+    be/c/hls.cpp
+    be/c/hls.h
+    be/c/opencl.cpp
+    be/c/opencl.h
     be/kernel_config.h
     tables/allnodes.h
     tables/arithoptable.h
@@ -90,18 +96,12 @@ if(LLVM_FOUND)
     list(APPEND THORIN_SOURCES
         be/llvm/cpu.cpp
         be/llvm/cpu.h
-        #be/llvm/cuda.cpp
-        #be/llvm/cuda.h
-        #be/llvm/hls.cpp
-        #be/llvm/hls.h
         be/llvm/llvm.cpp
         be/llvm/llvm.h
         be/llvm/amdgpu.cpp
         be/llvm/amdgpu.h
         be/llvm/nvvm.cpp
         be/llvm/nvvm.h
-            be/c/opencl.cpp
-            be/c/opencl.h
         be/llvm/parallel.cpp
         be/llvm/runtime.inc
         be/llvm/runtime.cpp
diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 6349c4e90..6458bfca1 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -6,11 +6,10 @@
 #include "thorin/be/llvm/cpu.h"
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu.h"
-#include "thorin/be/llvm/cuda.h"
-#include "thorin/be/llvm/hls.h"
-#include "thorin/be/c/opencl.h"
-#include "thorin/transform/codegen_prepare.h"
 #endif
+#include "thorin/be/c/cuda.h"
+#include "thorin/be/c/hls.h"
+#include "thorin/be/c/opencl.h"
 
 namespace thorin {
 
@@ -187,10 +186,9 @@ Backends::Backends(World& world, int opt, bool debug)
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
-    // TODO: Fix the C-based backends
-    //if (!cuda.  world().empty()) cuda_cg   = std::make_unique<CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
-    //if (!opencl.world().empty()) opencl_cg = std::make_unique<OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
-    //if (!hls.   world().empty()) hls_cg    = std::make_unique<HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
+    if (!cuda.  world().empty()) cuda_cg   = std::make_unique<c_be::CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
+    if (!opencl.world().empty()) opencl_cg = std::make_unique<c_be::OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
+    if (!hls.   world().empty()) hls_cg    = std::make_unique<c_be::HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
 }
 
 CodeGen::CodeGen(World& world, bool debug)
diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 0a822ed83..daf1b4f18 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -1554,8 +1554,9 @@ std::string CCodeGen::tuple_name(const TupleType* tuple_type) {
 
 //------------------------------------------------------------------------------
 
-void emit_c(World& world, const Cont2Config& kernel_config, Stream& stream, Lang lang, bool debug) {
-    CCodeGen(world, kernel_config, stream, lang, debug).emit();
+void CodeGen::emit(std::ostream &stream) {
+    Stream s(stream);
+    CCodeGen(world(), kernel_config_, s, lang_, debug_).emit();
 }
 
 void emit_c_int(World& world, Stream& stream) {
diff --git a/src/thorin/be/c/c.h b/src/thorin/be/c/c.h
index 01fe38354..818eaed81 100644
--- a/src/thorin/be/c/c.h
+++ b/src/thorin/be/c/c.h
@@ -4,7 +4,7 @@
 #include <cstdint>
 #include <iostream>
 
-#include "thorin/be/kernel_config.h"
+#include "thorin/be/backends.h"
 
 namespace thorin {
 
@@ -19,7 +19,22 @@ enum class Lang : uint8_t {
     OPENCL      ///< Flag for OpenCL
 };
 
-void emit_c(World&, const Cont2Config& kernel_config, std::ostream& stream, Lang lang, bool debug);
+class CodeGen : public thorin::CodeGen {
+public:
+    CodeGen(World& world, const Cont2Config& kernel_config, Lang lang, bool debug)
+    : thorin::CodeGen(world, debug)
+    , kernel_config_(kernel_config)
+    , lang_(lang)
+    , debug_(debug) {}
+
+    void emit(std::ostream& stream) override;
+
+private:
+    const Cont2Config& kernel_config_;
+    Lang lang_;
+    bool debug_;
+};
+
 void emit_c_int(World&, Stream& stream);
 
 }
diff --git a/src/thorin/be/c/cuda.cpp b/src/thorin/be/c/cuda.cpp
new file mode 100644
index 000000000..815173a1b
--- /dev/null
+++ b/src/thorin/be/c/cuda.cpp
@@ -0,0 +1,9 @@
+#include "cuda.h"
+
+namespace thorin::c_be {
+
+CUDACodeGen::CUDACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
+    : CodeGen(world, kernel_config, Lang::CUDA, debug)
+{}
+
+}
diff --git a/src/thorin/be/c/cuda.h b/src/thorin/be/c/cuda.h
new file mode 100644
index 000000000..a6200e294
--- /dev/null
+++ b/src/thorin/be/c/cuda.h
@@ -0,0 +1,15 @@
+#ifndef THORIN_BE_LLVM_CUDA_H
+#define THORIN_BE_LLVM_CUDA_H
+
+#include "c.h"
+
+namespace thorin::c_be {
+
+class CUDACodeGen : public CodeGen {
+public:
+    CUDACodeGen(World &world, const Cont2Config &, int opt, bool debug);
+};
+
+}
+
+#endif
diff --git a/src/thorin/be/c/hls.cpp b/src/thorin/be/c/hls.cpp
new file mode 100644
index 000000000..66660f431
--- /dev/null
+++ b/src/thorin/be/c/hls.cpp
@@ -0,0 +1,9 @@
+#include "hls.h"
+
+namespace thorin::c_be {
+
+HLSCodeGen::HLSCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
+    : CodeGen(world, kernel_config, Lang::HLS, debug)
+{}
+
+}
diff --git a/src/thorin/be/c/hls.h b/src/thorin/be/c/hls.h
new file mode 100644
index 000000000..53ea1ad6a
--- /dev/null
+++ b/src/thorin/be/c/hls.h
@@ -0,0 +1,15 @@
+#ifndef THORIN_BE_LLVM_HLS_H
+#define THORIN_BE_LLVM_HLS_H
+
+#include "c.h"
+
+namespace thorin::c_be {
+
+class HLSCodeGen : public CodeGen {
+public:
+    HLSCodeGen(World& world, const Cont2Config&, int opt, bool debug);
+};
+
+}
+
+#endif
diff --git a/src/thorin/be/c/opencl.cpp b/src/thorin/be/c/opencl.cpp
index c93910f55..5d54ad4d5 100644
--- a/src/thorin/be/c/opencl.cpp
+++ b/src/thorin/be/c/opencl.cpp
@@ -1,21 +1,9 @@
 #include "opencl.h"
 
-#include <fstream>
-#include <stdexcept>
-
-#include "thorin/primop.h"
-#include "thorin/world.h"
-#include "thorin/be/c/c.h"
-
 namespace thorin::c_be {
 
 OpenCLCodeGen::OpenCLCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, debug)
-    , kernel_config_(kernel_config)
+    : CodeGen(world, kernel_config, Lang::OPENCL, debug)
 {}
 
-void OpenCLCodeGen::emit(std::ostream& stream) {
-    emit_c(world(), kernel_config_, stream, Lang::OPENCL, debug());
-}
-
 }
diff --git a/src/thorin/be/c/opencl.h b/src/thorin/be/c/opencl.h
index 3b472777e..3d082077f 100644
--- a/src/thorin/be/c/opencl.h
+++ b/src/thorin/be/c/opencl.h
@@ -1,19 +1,13 @@
 #ifndef THORIN_BE_LLVM_OPENCL_H
 #define THORIN_BE_LLVM_OPENCL_H
 
-#include "../backends.h"
-#include "../llvm/llvm.h"
+#include "c.h"
 
 namespace thorin::c_be {
 
-class OpenCLCodeGen : public thorin::CodeGen {
+class OpenCLCodeGen : public CodeGen {
 public:
     OpenCLCodeGen(World& world, const Cont2Config&, int opt, bool debug);
-
-    void emit(std::ostream& stream) override;
-
-protected:
-    const Cont2Config& kernel_config_;
 };
 
 }
diff --git a/src/thorin/be/llvm/cuda.cpp b/src/thorin/be/llvm/cuda.cpp
deleted file mode 100644
index 28b9824e7..000000000
--- a/src/thorin/be/llvm/cuda.cpp
+++ /dev/null
@@ -1,21 +0,0 @@
-#include "thorin/be/llvm/cuda.h"
-
-#include <fstream>
-#include <stdexcept>
-
-#include "thorin/primop.h"
-#include "thorin/world.h"
-#include "thorin/be/c.h"
-
-namespace thorin::llvm_be {
-
-CUDACodeGen::CUDACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
-    , kernel_config_(kernel_config)
-{}
-
-void CUDACodeGen::emit(std::ostream& stream) {
-    thorin::emit_c(world(), kernel_config_, stream, Lang::CUDA, debug());
-}
-
-}
diff --git a/src/thorin/be/llvm/cuda.h b/src/thorin/be/llvm/cuda.h
deleted file mode 100644
index 3fe4472eb..000000000
--- a/src/thorin/be/llvm/cuda.h
+++ /dev/null
@@ -1,22 +0,0 @@
-#ifndef THORIN_BE_LLVM_CUDA_H
-#define THORIN_BE_LLVM_CUDA_H
-
-#include "thorin/be/llvm/llvm.h"
-
-namespace thorin::llvm_be {
-
-class CUDACodeGen : public CodeGen {
-public:
-    CUDACodeGen(World& world, const Cont2Config&, int opt, bool debug);
-
-    void emit(std::ostream& stream) override;
-
-protected:
-    virtual std::string get_alloc_name() const override { return "malloc"; }
-
-    const Cont2Config& kernel_config_;
-};
-
-}
-
-#endif
diff --git a/src/thorin/be/llvm/hls.cpp b/src/thorin/be/llvm/hls.cpp
deleted file mode 100644
index 125722774..000000000
--- a/src/thorin/be/llvm/hls.cpp
+++ /dev/null
@@ -1,21 +0,0 @@
-#include "thorin/be/llvm/hls.h"
-
-#include <fstream>
-#include <stdexcept>
-
-#include "thorin/primop.h"
-#include "thorin/world.h"
-#include "thorin/be/c.h"
-
-namespace thorin::llvm_be {
-
-HLSCodeGen::HLSCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
-    , kernel_config_(kernel_config)
-{}
-
-void HLSCodeGen::emit(std::ostream& stream) {
-    thorin::emit_c(world(), kernel_config_, stream, Lang::HLS, debug());
-}
-
-}
diff --git a/src/thorin/be/llvm/hls.h b/src/thorin/be/llvm/hls.h
deleted file mode 100644
index 7143a675e..000000000
--- a/src/thorin/be/llvm/hls.h
+++ /dev/null
@@ -1,22 +0,0 @@
-#ifndef THORIN_BE_LLVM_HLS_H
-#define THORIN_BE_LLVM_HLS_H
-
-#include "thorin/be/llvm/llvm.h"
-
-namespace thorin::llvm_be {
-
-class HLSCodeGen : public CodeGen {
-public:
-    HLSCodeGen(World& world, const Cont2Config&, int opt, bool debug);
-
-    void emit(std::ostream& stream) override;
-
-protected:
-    virtual std::string get_alloc_name() const override { THORIN_UNREACHABLE; /*alloc not supported in HLS*/; }
-
-    const Cont2Config& kernel_config_;
-};
-
-}
-
-#endif

From ed29b0e84f033068b6b026f403212442b6b01fba Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 16:22:48 +0100
Subject: [PATCH 003/342] remove _be suffix from backend namespaces

---
 src/thorin/be/backends.cpp       | 12 ++++++------
 src/thorin/be/c/c.cpp            |  2 +-
 src/thorin/be/c/c.h              |  2 +-
 src/thorin/be/c/cuda.cpp         |  2 +-
 src/thorin/be/c/cuda.h           |  2 +-
 src/thorin/be/c/hls.cpp          |  2 +-
 src/thorin/be/c/hls.h            |  2 +-
 src/thorin/be/c/opencl.cpp       |  2 +-
 src/thorin/be/c/opencl.h         |  2 +-
 src/thorin/be/llvm/amdgpu.cpp    |  2 +-
 src/thorin/be/llvm/amdgpu.h      |  4 +++-
 src/thorin/be/llvm/cpu.cpp       |  2 +-
 src/thorin/be/llvm/cpu.h         |  4 +++-
 src/thorin/be/llvm/llvm.cpp      |  2 +-
 src/thorin/be/llvm/llvm.h        |  4 +++-
 src/thorin/be/llvm/nvvm.cpp      |  2 +-
 src/thorin/be/llvm/nvvm.h        |  4 +++-
 src/thorin/be/llvm/parallel.cpp  |  2 +-
 src/thorin/be/llvm/runtime.cpp   |  2 +-
 src/thorin/be/llvm/runtime.h     |  4 +++-
 src/thorin/be/llvm/vectorize.cpp |  2 +-
 21 files changed, 36 insertions(+), 26 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 6458bfca1..e3cbf14a7 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -179,16 +179,16 @@ Backends::Backends(World& world, int opt, bool debug)
     }
 
 #ifdef THORIN_ENABLE_LLVM
-    cpu_cg = std::make_unique<llvm_be::CPUCodeGen>(world, opt, debug);
+    cpu_cg = std::make_unique<llvm::CPUCodeGen>(world, opt, debug);
 
-    if (!nvvm.  world().empty()) nvvm_cg   = std::make_unique<llvm_be::NVVMCodeGen  >(nvvm  .world(), kernel_config,      debug);
-    if (!amdgpu.world().empty()) amdgpu_cg = std::make_unique<llvm_be::AMDGPUCodeGen>(amdgpu.world(), kernel_config, opt, debug);
+    if (!nvvm.  world().empty()) nvvm_cg   = std::make_unique<llvm::NVVMCodeGen  >(nvvm  .world(), kernel_config,      debug);
+    if (!amdgpu.world().empty()) amdgpu_cg = std::make_unique<llvm::AMDGPUCodeGen>(amdgpu.world(), kernel_config, opt, debug);
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
-    if (!cuda.  world().empty()) cuda_cg   = std::make_unique<c_be::CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
-    if (!opencl.world().empty()) opencl_cg = std::make_unique<c_be::OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
-    if (!hls.   world().empty()) hls_cg    = std::make_unique<c_be::HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
+    if (!cuda.  world().empty()) cuda_cg   = std::make_unique<c::CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
+    if (!opencl.world().empty()) opencl_cg = std::make_unique<c::OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
+    if (!hls.   world().empty()) hls_cg    = std::make_unique<c::HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
 }
 
 CodeGen::CodeGen(World& world, bool debug)
diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index daf1b4f18..ce93ef399 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -15,7 +15,7 @@
 #include <cctype>
 #include <variant>
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 class CCodeGen;
 
diff --git a/src/thorin/be/c/c.h b/src/thorin/be/c/c.h
index 818eaed81..368ceb94e 100644
--- a/src/thorin/be/c/c.h
+++ b/src/thorin/be/c/c.h
@@ -10,7 +10,7 @@ namespace thorin {
 
 class World;
 
-namespace c_be {
+namespace c {
 
 enum class Lang : uint8_t {
     C99,        ///< Flag for C99
diff --git a/src/thorin/be/c/cuda.cpp b/src/thorin/be/c/cuda.cpp
index 815173a1b..ec7c92ab1 100644
--- a/src/thorin/be/c/cuda.cpp
+++ b/src/thorin/be/c/cuda.cpp
@@ -1,6 +1,6 @@
 #include "cuda.h"
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 CUDACodeGen::CUDACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, kernel_config, Lang::CUDA, debug)
diff --git a/src/thorin/be/c/cuda.h b/src/thorin/be/c/cuda.h
index a6200e294..d55b8e9de 100644
--- a/src/thorin/be/c/cuda.h
+++ b/src/thorin/be/c/cuda.h
@@ -3,7 +3,7 @@
 
 #include "c.h"
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 class CUDACodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/c/hls.cpp b/src/thorin/be/c/hls.cpp
index 66660f431..ec2bbb43e 100644
--- a/src/thorin/be/c/hls.cpp
+++ b/src/thorin/be/c/hls.cpp
@@ -1,6 +1,6 @@
 #include "hls.h"
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 HLSCodeGen::HLSCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, kernel_config, Lang::HLS, debug)
diff --git a/src/thorin/be/c/hls.h b/src/thorin/be/c/hls.h
index 53ea1ad6a..3f05ebf79 100644
--- a/src/thorin/be/c/hls.h
+++ b/src/thorin/be/c/hls.h
@@ -3,7 +3,7 @@
 
 #include "c.h"
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 class HLSCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/c/opencl.cpp b/src/thorin/be/c/opencl.cpp
index 5d54ad4d5..295dd9b47 100644
--- a/src/thorin/be/c/opencl.cpp
+++ b/src/thorin/be/c/opencl.cpp
@@ -1,6 +1,6 @@
 #include "opencl.h"
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 OpenCLCodeGen::OpenCLCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, kernel_config, Lang::OPENCL, debug)
diff --git a/src/thorin/be/c/opencl.h b/src/thorin/be/c/opencl.h
index 3d082077f..0e4cab1da 100644
--- a/src/thorin/be/c/opencl.h
+++ b/src/thorin/be/c/opencl.h
@@ -3,7 +3,7 @@
 
 #include "c.h"
 
-namespace thorin::c_be {
+namespace thorin::c {
 
 class OpenCLCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/amdgpu.cpp b/src/thorin/be/llvm/amdgpu.cpp
index 859948be4..bd80736d0 100644
--- a/src/thorin/be/llvm/amdgpu.cpp
+++ b/src/thorin/be/llvm/amdgpu.cpp
@@ -3,7 +3,7 @@
 #include "thorin/primop.h"
 #include "thorin/world.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 AMDGPUCodeGen::AMDGPUCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
diff --git a/src/thorin/be/llvm/amdgpu.h b/src/thorin/be/llvm/amdgpu.h
index 33866fc63..80e325023 100644
--- a/src/thorin/be/llvm/amdgpu.h
+++ b/src/thorin/be/llvm/amdgpu.h
@@ -7,7 +7,9 @@ namespace thorin {
 
 class Load;
 
-namespace llvm_be {
+namespace llvm {
+
+namespace llvm = ::llvm;
 
 class AMDGPUCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/cpu.cpp b/src/thorin/be/llvm/cpu.cpp
index ed6e92772..ffd06af06 100644
--- a/src/thorin/be/llvm/cpu.cpp
+++ b/src/thorin/be/llvm/cpu.cpp
@@ -8,7 +8,7 @@
 #include <llvm/Target/TargetMachine.h>
 #include <llvm/Target/TargetOptions.h>
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 CPUCodeGen::CPUCodeGen(World& world, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
diff --git a/src/thorin/be/llvm/cpu.h b/src/thorin/be/llvm/cpu.h
index 7607cdcb6..462ca3e21 100644
--- a/src/thorin/be/llvm/cpu.h
+++ b/src/thorin/be/llvm/cpu.h
@@ -3,7 +3,9 @@
 
 #include "thorin/be/llvm/llvm.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
+
+namespace llvm = ::llvm;
 
 class CPUCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 5f267f947..5a2c867ce 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -38,7 +38,7 @@
 #include "thorin/analyses/scope.h"
 #include "thorin/util/array.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 CodeGen::CodeGen(World& world,
                  llvm::CallingConv::ID function_calling_convention,
diff --git a/src/thorin/be/llvm/llvm.h b/src/thorin/be/llvm/llvm.h
index 7f41d51e8..518db9ef6 100644
--- a/src/thorin/be/llvm/llvm.h
+++ b/src/thorin/be/llvm/llvm.h
@@ -18,7 +18,9 @@ namespace thorin {
 
 class World;
 
-namespace llvm_be {
+namespace llvm {
+
+namespace llvm = ::llvm;
 
 class CodeGen : public thorin::CodeGen {
 protected:
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index 1fd8c7818..3a109ab84 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -16,7 +16,7 @@
 #include "thorin/primop.h"
 #include "thorin/world.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 NVVMCodeGen::NVVMCodeGen(World& world, const Cont2Config& kernel_config, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::PTX_Device, llvm::CallingConv::PTX_Kernel, 0, debug)
diff --git a/src/thorin/be/llvm/nvvm.h b/src/thorin/be/llvm/nvvm.h
index f637304c4..541530868 100644
--- a/src/thorin/be/llvm/nvvm.h
+++ b/src/thorin/be/llvm/nvvm.h
@@ -7,7 +7,9 @@ namespace thorin {
 
 class Load;
 
-namespace llvm_be {
+namespace llvm {
+
+namespace llvm = ::llvm;
 
 class NVVMCodeGen : public CodeGen {
 public:
diff --git a/src/thorin/be/llvm/parallel.cpp b/src/thorin/be/llvm/parallel.cpp
index 128728fad..16247bf47 100644
--- a/src/thorin/be/llvm/parallel.cpp
+++ b/src/thorin/be/llvm/parallel.cpp
@@ -1,6 +1,6 @@
 #include "thorin/be/llvm/llvm.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 enum {
     PAR_ARG_MEM,
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index b7b38070c..ee653bccb 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -13,7 +13,7 @@
 #include "thorin/be/llvm/llvm.h"
 #include "thorin/be/llvm/runtime.inc"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 Runtime::Runtime(llvm::LLVMContext& context,
                  llvm::Module& target)
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index c125deed5..3035704ef 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -8,7 +8,9 @@
 
 #include "thorin/world.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
+
+namespace llvm = ::llvm;
 
 class CodeGen;
 
diff --git a/src/thorin/be/llvm/vectorize.cpp b/src/thorin/be/llvm/vectorize.cpp
index 529ac844b..12ba8b840 100644
--- a/src/thorin/be/llvm/vectorize.cpp
+++ b/src/thorin/be/llvm/vectorize.cpp
@@ -36,7 +36,7 @@
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
 
-namespace thorin::llvm_be {
+namespace thorin::llvm {
 
 struct VectorizeArgs {
     enum {

From c0bdbed96d423cdcbea43316c4cbdb8f6c4b5e68 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 16:26:32 +0100
Subject: [PATCH 004/342] nuke redundant wrapper classes

---
 src/thorin/CMakeLists.txt  |  6 ------
 src/thorin/be/backends.cpp | 10 ++++------
 src/thorin/be/c/cuda.cpp   |  9 ---------
 src/thorin/be/c/cuda.h     | 15 ---------------
 src/thorin/be/c/hls.cpp    |  9 ---------
 src/thorin/be/c/hls.h      | 15 ---------------
 src/thorin/be/c/opencl.cpp |  9 ---------
 src/thorin/be/c/opencl.h   | 15 ---------------
 8 files changed, 4 insertions(+), 84 deletions(-)
 delete mode 100644 src/thorin/be/c/cuda.cpp
 delete mode 100644 src/thorin/be/c/cuda.h
 delete mode 100644 src/thorin/be/c/hls.cpp
 delete mode 100644 src/thorin/be/c/hls.h
 delete mode 100644 src/thorin/be/c/opencl.cpp
 delete mode 100644 src/thorin/be/c/opencl.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index b0c18ea7c..81bf603a8 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -32,12 +32,6 @@ set(THORIN_SOURCES
     analyses/verify.h
     be/c/c.cpp
     be/c/c.h
-    be/c/cuda.cpp
-    be/c/cuda.h
-    be/c/hls.cpp
-    be/c/hls.h
-    be/c/opencl.cpp
-    be/c/opencl.h
     be/kernel_config.h
     tables/allnodes.h
     tables/arithoptable.h
diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index e3cbf14a7..3ad7a55b8 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -7,9 +7,7 @@
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu.h"
 #endif
-#include "thorin/be/c/cuda.h"
-#include "thorin/be/c/hls.h"
-#include "thorin/be/c/opencl.h"
+#include "thorin/be/c/c.h"
 
 namespace thorin {
 
@@ -186,9 +184,9 @@ Backends::Backends(World& world, int opt, bool debug)
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
-    if (!cuda.  world().empty()) cuda_cg   = std::make_unique<c::CUDACodeGen  >(cuda  .world(), kernel_config, opt, debug);
-    if (!opencl.world().empty()) opencl_cg = std::make_unique<c::OpenCLCodeGen>(opencl.world(), kernel_config, opt, debug);
-    if (!hls.   world().empty()) hls_cg    = std::make_unique<c::HLSCodeGen   >(hls   .world(), kernel_config, opt, debug);
+    if (!cuda.  world().empty()) cuda_cg   = std::make_unique<c::CodeGen>(cuda  .world(), kernel_config, c::Lang::CUDA  , debug);
+    if (!opencl.world().empty()) opencl_cg = std::make_unique<c::CodeGen>(opencl.world(), kernel_config, c::Lang::OPENCL, debug);
+    if (!hls.   world().empty()) hls_cg    = std::make_unique<c::CodeGen>(hls   .world(), kernel_config, c::Lang::HLS   , debug);
 }
 
 CodeGen::CodeGen(World& world, bool debug)
diff --git a/src/thorin/be/c/cuda.cpp b/src/thorin/be/c/cuda.cpp
deleted file mode 100644
index ec7c92ab1..000000000
--- a/src/thorin/be/c/cuda.cpp
+++ /dev/null
@@ -1,9 +0,0 @@
-#include "cuda.h"
-
-namespace thorin::c {
-
-CUDACodeGen::CUDACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, kernel_config, Lang::CUDA, debug)
-{}
-
-}
diff --git a/src/thorin/be/c/cuda.h b/src/thorin/be/c/cuda.h
deleted file mode 100644
index d55b8e9de..000000000
--- a/src/thorin/be/c/cuda.h
+++ /dev/null
@@ -1,15 +0,0 @@
-#ifndef THORIN_BE_LLVM_CUDA_H
-#define THORIN_BE_LLVM_CUDA_H
-
-#include "c.h"
-
-namespace thorin::c {
-
-class CUDACodeGen : public CodeGen {
-public:
-    CUDACodeGen(World &world, const Cont2Config &, int opt, bool debug);
-};
-
-}
-
-#endif
diff --git a/src/thorin/be/c/hls.cpp b/src/thorin/be/c/hls.cpp
deleted file mode 100644
index ec2bbb43e..000000000
--- a/src/thorin/be/c/hls.cpp
+++ /dev/null
@@ -1,9 +0,0 @@
-#include "hls.h"
-
-namespace thorin::c {
-
-HLSCodeGen::HLSCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, kernel_config, Lang::HLS, debug)
-{}
-
-}
diff --git a/src/thorin/be/c/hls.h b/src/thorin/be/c/hls.h
deleted file mode 100644
index 3f05ebf79..000000000
--- a/src/thorin/be/c/hls.h
+++ /dev/null
@@ -1,15 +0,0 @@
-#ifndef THORIN_BE_LLVM_HLS_H
-#define THORIN_BE_LLVM_HLS_H
-
-#include "c.h"
-
-namespace thorin::c {
-
-class HLSCodeGen : public CodeGen {
-public:
-    HLSCodeGen(World& world, const Cont2Config&, int opt, bool debug);
-};
-
-}
-
-#endif
diff --git a/src/thorin/be/c/opencl.cpp b/src/thorin/be/c/opencl.cpp
deleted file mode 100644
index 295dd9b47..000000000
--- a/src/thorin/be/c/opencl.cpp
+++ /dev/null
@@ -1,9 +0,0 @@
-#include "opencl.h"
-
-namespace thorin::c {
-
-OpenCLCodeGen::OpenCLCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, kernel_config, Lang::OPENCL, debug)
-{}
-
-}
diff --git a/src/thorin/be/c/opencl.h b/src/thorin/be/c/opencl.h
deleted file mode 100644
index 0e4cab1da..000000000
--- a/src/thorin/be/c/opencl.h
+++ /dev/null
@@ -1,15 +0,0 @@
-#ifndef THORIN_BE_LLVM_OPENCL_H
-#define THORIN_BE_LLVM_OPENCL_H
-
-#include "c.h"
-
-namespace thorin::c {
-
-class OpenCLCodeGen : public CodeGen {
-public:
-    OpenCLCodeGen(World& world, const Cont2Config&, int opt, bool debug);
-};
-
-}
-
-#endif

From e9b6930ddd17db414f6d639022537c1748a53315 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 17:42:50 +0100
Subject: [PATCH 005/342] fix formatting

---
 src/thorin/be/backends.cpp | 14 +++++++-------
 src/thorin/be/c/c.h        |  9 +++++----
 2 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 3ad7a55b8..c8bf65f98 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -72,11 +72,11 @@ static uint64_t get_alloc_size(const Def* def) {
 }
 
 Backends::Backends(World& world, int opt, bool debug)
-: cuda(world)
-, nvvm(world)
-, opencl(world)
-, amdgpu(world)
-, hls(world)
+    : cuda(world)
+    , nvvm(world)
+    , opencl(world)
+    , amdgpu(world)
+    , hls(world)
 {
     // determine different parts of the world which need to be compiled differently
     Scope::for_each(world, [&] (const Scope& scope) {
@@ -190,8 +190,8 @@ Backends::Backends(World& world, int opt, bool debug)
 }
 
 CodeGen::CodeGen(World& world, bool debug)
-: world_(world)
-, debug_(debug)
+    : world_(world)
+    , debug_(debug)
 {}
 
 }
diff --git a/src/thorin/be/c/c.h b/src/thorin/be/c/c.h
index 368ceb94e..bfb98e307 100644
--- a/src/thorin/be/c/c.h
+++ b/src/thorin/be/c/c.h
@@ -22,10 +22,11 @@ enum class Lang : uint8_t {
 class CodeGen : public thorin::CodeGen {
 public:
     CodeGen(World& world, const Cont2Config& kernel_config, Lang lang, bool debug)
-    : thorin::CodeGen(world, debug)
-    , kernel_config_(kernel_config)
-    , lang_(lang)
-    , debug_(debug) {}
+        : thorin::CodeGen(world, debug)
+        , kernel_config_(kernel_config)
+        , lang_(lang)
+        , debug_(debug)
+    {}
 
     void emit(std::ostream& stream) override;
 

From 6bbed3a9c8c921921e63fa43c595bfc5f1ad734d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 18:29:07 +0100
Subject: [PATCH 006/342] generalize backends logic

---
 src/thorin/be/backends.cpp | 58 +++++++++++++++++---------------------
 src/thorin/be/backends.h   | 18 ++++--------
 2 files changed, 31 insertions(+), 45 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index c8bf65f98..dc4fd6953 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -72,28 +72,22 @@ static uint64_t get_alloc_size(const Def* def) {
 }
 
 Backends::Backends(World& world, int opt, bool debug)
-    : cuda(world)
-    , nvvm(world)
-    , opencl(world)
-    , amdgpu(world)
-    , hls(world)
+    : device_cgs({})
 {
+    for (int backend = 0; backend < BackendCount; backend++)
+        importers_.emplace_back(world);
+
     // determine different parts of the world which need to be compiled differently
     Scope::for_each(world, [&] (const Scope& scope) {
         auto continuation = scope.entry();
         Continuation* imported = nullptr;
-        if (is_passed_to_intrinsic(continuation, Intrinsic::CUDA))
-            imported = cuda.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::NVVM))
-            imported = nvvm.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::OpenCL))
-            imported = opencl.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::AMDGPU))
-            imported = amdgpu.import(continuation)->as_continuation();
-        else if (is_passed_to_intrinsic(continuation, Intrinsic::HLS))
-            imported = hls.import(continuation)->as_continuation();
-        else
-            return;
+        for (int backend = 0; backend <= BackendCount; backend++) {
+            if (backend == BackendCount) return;
+            if (is_passed_to_intrinsic(continuation, Intrinsic(int(Intrinsic::AcceleratorBegin) + backend))) {
+                imported = importers_[backend].import(continuation)->as_continuation();
+                break;
+            }
+        }
 
         imported->set_name(continuation->unique_name());
         imported->make_exported();
@@ -106,10 +100,10 @@ Backends::Backends(World& world, int opt, bool debug)
     });
 
     // get the GPU kernel configurations
-    if (!cuda.world().empty()   ||
-        !nvvm.world().empty()   ||
-        !opencl.world().empty() ||
-        !amdgpu.world().empty()) {
+    if (!importers_[Cuda  ].world().empty() ||
+        !importers_[NVVM  ].world().empty() ||
+        !importers_[OpenCL].world().empty() ||
+        !importers_[AMDGPU].world().empty()) {
         auto get_gpu_config = [&] (Continuation* use, Continuation* /* imported */) {
             // determine whether or not this kernel uses restrict pointers
             bool has_restrict = true;
@@ -135,14 +129,14 @@ Backends::Backends(World& world, int opt, bool debug)
             }
             return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> { -1, -1, -1 }, has_restrict);
         };
-        get_kernel_configs(cuda,   kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(nvvm,   kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(opencl, kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(amdgpu, kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(importers_[Cuda  ], kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(importers_[NVVM  ], kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(importers_[OpenCL], kernels, kernel_config, get_gpu_config);
+        get_kernel_configs(importers_[AMDGPU], kernels, kernel_config, get_gpu_config);
     }
 
     // get the HLS kernel configurations
-    if (!hls.world().empty()) {
+    if (!importers_[HLS].world().empty()) {
         auto get_hls_config = [&] (Continuation* use, Continuation* imported) {
             HLSKernelConfig::Param2Size param_sizes;
             for (size_t i = 3, e = use->num_args(); i != e; ++i) {
@@ -173,20 +167,20 @@ Backends::Backends(World& world, int opt, bool debug)
             }
             return std::make_unique<HLSKernelConfig>(param_sizes);
         };
-        get_kernel_configs(hls, kernels, kernel_config, get_hls_config);
+        get_kernel_configs(importers_[HLS], kernels, kernel_config, get_hls_config);
     }
 
 #ifdef THORIN_ENABLE_LLVM
     cpu_cg = std::make_unique<llvm::CPUCodeGen>(world, opt, debug);
 
-    if (!nvvm.  world().empty()) nvvm_cg   = std::make_unique<llvm::NVVMCodeGen  >(nvvm  .world(), kernel_config,      debug);
-    if (!amdgpu.world().empty()) amdgpu_cg = std::make_unique<llvm::AMDGPUCodeGen>(amdgpu.world(), kernel_config, opt, debug);
+    if (!importers_[NVVM  ].world().empty()) device_cgs[NVVM  ] = std::make_unique<llvm::NVVMCodeGen  >(importers_[NVVM  ].world(), kernel_config,      debug);
+    if (!importers_[AMDGPU].world().empty()) device_cgs[AMDGPU] = std::make_unique<llvm::AMDGPUCodeGen>(importers_[AMDGPU].world(), kernel_config, opt, debug);
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
-    if (!cuda.  world().empty()) cuda_cg   = std::make_unique<c::CodeGen>(cuda  .world(), kernel_config, c::Lang::CUDA  , debug);
-    if (!opencl.world().empty()) opencl_cg = std::make_unique<c::CodeGen>(opencl.world(), kernel_config, c::Lang::OPENCL, debug);
-    if (!hls.   world().empty()) hls_cg    = std::make_unique<c::CodeGen>(hls   .world(), kernel_config, c::Lang::HLS   , debug);
+    if (!importers_[Cuda  ].world().empty()) device_cgs[Cuda  ] = std::make_unique<c::CodeGen>(importers_[Cuda  ].world(), kernel_config, c::Lang::CUDA  , debug);
+    if (!importers_[OpenCL].world().empty()) device_cgs[OpenCL] = std::make_unique<c::CodeGen>(importers_[OpenCL].world(), kernel_config, c::Lang::OPENCL, debug);
+    if (!importers_[HLS   ].world().empty()) device_cgs[HLS   ] = std::make_unique<c::CodeGen>(importers_[HLS   ].world(), kernel_config, c::Lang::HLS   , debug);
 }
 
 CodeGen::CodeGen(World& world, bool debug)
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index b9c309889..0879b4426 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -41,20 +41,12 @@ struct Backends {
     Cont2Config kernel_config;
     std::vector<Continuation*> kernels;
 
-    // TODO use arrays + loops for this
-    Importer cuda;
-    Importer nvvm;
-    Importer opencl;
-    Importer amdgpu;
-    Importer hls;
-
-    // TODO use arrays + loops for this
     std::unique_ptr<CodeGen> cpu_cg;
-    std::unique_ptr<CodeGen> cuda_cg;
-    std::unique_ptr<CodeGen> nvvm_cg;
-    std::unique_ptr<CodeGen> opencl_cg;
-    std::unique_ptr<CodeGen> amdgpu_cg;
-    std::unique_ptr<CodeGen> hls_cg;
+
+    enum { Cuda, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
+    std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
+private:
+    std::vector<Importer> importers_;
 };
 
 }

From 0f279243f27be9ab5d3c5e3e3f126ee943c16387 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 18:43:22 +0100
Subject: [PATCH 007/342] generalize gpu kernel setup

---
 src/thorin/be/backends.cpp | 61 +++++++++++++++++---------------------
 src/thorin/be/backends.h   |  3 ++
 2 files changed, 31 insertions(+), 33 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index dc4fd6953..a7b5d0eab 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -99,40 +99,35 @@ Backends::Backends(World& world, int opt, bool debug)
         kernels.emplace_back(continuation);
     });
 
-    // get the GPU kernel configurations
-    if (!importers_[Cuda  ].world().empty() ||
-        !importers_[NVVM  ].world().empty() ||
-        !importers_[OpenCL].world().empty() ||
-        !importers_[AMDGPU].world().empty()) {
-        auto get_gpu_config = [&] (Continuation* use, Continuation* /* imported */) {
-            // determine whether or not this kernel uses restrict pointers
-            bool has_restrict = true;
-            DefSet allocs;
-            for (size_t i = LaunchArgs::Num, e = use->num_args(); has_restrict && i != e; ++i) {
-                auto arg = use->arg(i);
-                if (!arg->type()->isa<PtrType>()) continue;
-                auto alloc = get_alloc_call(arg);
-                if (!alloc) has_restrict = false;
-                auto p = allocs.insert(alloc);
-                has_restrict &= p.second;
-            }
+    for (auto backend : gpu_compute_backends) {
+        if (!importers_[backend].world().empty()) {
+            auto get_gpu_config = [&](Continuation *use, Continuation * /* imported */) {
+                // determine whether or not this kernel uses restrict pointers
+                bool has_restrict = true;
+                DefSet allocs;
+                for (size_t i = LaunchArgs::Num, e = use->num_args(); has_restrict && i != e; ++i) {
+                    auto arg = use->arg(i);
+                    if (!arg->type()->isa<PtrType>()) continue;
+                    auto alloc = get_alloc_call(arg);
+                    if (!alloc) has_restrict = false;
+                    auto p = allocs.insert(alloc);
+                    has_restrict &= p.second;
+                }
 
-            auto it_config = use->arg(LaunchArgs::Config)->as<Tuple>();
-            if (it_config->op(0)->isa<PrimLit>() &&
-                it_config->op(1)->isa<PrimLit>() &&
-                it_config->op(2)->isa<PrimLit>()) {
-                return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> {
-                it_config->op(0)->as<PrimLit>()->qu32_value().data(),
-                it_config->op(1)->as<PrimLit>()->qu32_value().data(),
-                it_config->op(2)->as<PrimLit>()->qu32_value().data()
-                }, has_restrict);
-            }
-            return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int> { -1, -1, -1 }, has_restrict);
-        };
-        get_kernel_configs(importers_[Cuda  ], kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(importers_[NVVM  ], kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(importers_[OpenCL], kernels, kernel_config, get_gpu_config);
-        get_kernel_configs(importers_[AMDGPU], kernels, kernel_config, get_gpu_config);
+                auto it_config = use->arg(LaunchArgs::Config)->as<Tuple>();
+                if (it_config->op(0)->isa<PrimLit>() &&
+                    it_config->op(1)->isa<PrimLit>() &&
+                    it_config->op(2)->isa<PrimLit>()) {
+                    return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{
+                            it_config->op(0)->as<PrimLit>()->qu32_value().data(),
+                            it_config->op(1)->as<PrimLit>()->qu32_value().data(),
+                            it_config->op(2)->as<PrimLit>()->qu32_value().data()
+                    }, has_restrict);
+                }
+                return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{-1, -1, -1}, has_restrict);
+            };
+            get_kernel_configs(importers_[backend], kernels, kernel_config, get_gpu_config);
+        }
     }
 
     // get the HLS kernel configurations
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index 0879b4426..d774db009 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -45,6 +45,9 @@ struct Backends {
 
     enum { Cuda, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
+
+    /// Backends that need GPUKernelConfig
+    static constexpr auto gpu_compute_backends = { Cuda, NVVM, OpenCL, AMDGPU };
 private:
     std::vector<Importer> importers_;
 };

From 97196bb6af50673ee3e13c81e6a64850a33fa1f2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 19:11:07 +0100
Subject: [PATCH 008/342] Fix #ifdefs

---
 src/thorin/be/backends.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index a7b5d0eab..9982da433 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -2,7 +2,7 @@
 
 #include "thorin/analyses/scope.h"
 
-#ifdef THORIN_ENABLE_LLVM
+#if THORIN_ENABLE_LLVM
 #include "thorin/be/llvm/cpu.h"
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu.h"
@@ -165,7 +165,7 @@ Backends::Backends(World& world, int opt, bool debug)
         get_kernel_configs(importers_[HLS], kernels, kernel_config, get_hls_config);
     }
 
-#ifdef THORIN_ENABLE_LLVM
+#if THORIN_ENABLE_LLVM
     cpu_cg = std::make_unique<llvm::CPUCodeGen>(world, opt, debug);
 
     if (!importers_[NVVM  ].world().empty()) device_cgs[NVVM  ] = std::make_unique<llvm::NVVMCodeGen  >(importers_[NVVM  ].world(), kernel_config,      debug);

From fc8e684e7cf81a0a40a0cd216b31f2ef244c3fba Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 19:11:31 +0100
Subject: [PATCH 009/342] Inline lambdas that are used once

---
 src/thorin/be/backends.cpp | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 9982da433..df4c6e9d3 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -101,7 +101,7 @@ Backends::Backends(World& world, int opt, bool debug)
 
     for (auto backend : gpu_compute_backends) {
         if (!importers_[backend].world().empty()) {
-            auto get_gpu_config = [&](Continuation *use, Continuation * /* imported */) {
+            get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 // determine whether or not this kernel uses restrict pointers
                 bool has_restrict = true;
                 DefSet allocs;
@@ -125,14 +125,13 @@ Backends::Backends(World& world, int opt, bool debug)
                     }, has_restrict);
                 }
                 return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{-1, -1, -1}, has_restrict);
-            };
-            get_kernel_configs(importers_[backend], kernels, kernel_config, get_gpu_config);
+            });
         }
     }
 
     // get the HLS kernel configurations
     if (!importers_[HLS].world().empty()) {
-        auto get_hls_config = [&] (Continuation* use, Continuation* imported) {
+        get_kernel_configs(importers_[HLS], kernels, kernel_config, [&] (Continuation* use, Continuation* imported) {
             HLSKernelConfig::Param2Size param_sizes;
             for (size_t i = 3, e = use->num_args(); i != e; ++i) {
                 auto arg = use->arg(i);
@@ -161,8 +160,7 @@ Backends::Backends(World& world, int opt, bool debug)
                 param_sizes.emplace(imported->param(i - 3 + 2), num_elems);
             }
             return std::make_unique<HLSKernelConfig>(param_sizes);
-        };
-        get_kernel_configs(importers_[HLS], kernels, kernel_config, get_hls_config);
+        });
     }
 
 #if THORIN_ENABLE_LLVM

From 9665f670ebe661173123444af876faf13c6fe3a2 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 15 Feb 2021 19:17:46 +0100
Subject: [PATCH 010/342] add file extensions for backends

---
 src/thorin/be/backends.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index d774db009..dce12fd55 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -48,6 +48,8 @@ struct Backends {
 
     /// Backends that need GPUKernelConfig
     static constexpr auto gpu_compute_backends = { Cuda, NVVM, OpenCL, AMDGPU };
+
+    static constexpr std::array<const char*, BackendCount> backends_extensions = { ".cu", ".nvvm", ".cl", ".amdgpu", ".hls" };
 private:
     std::vector<Importer> importers_;
 };

From b0a7af5a9512121a9b19fae78a73291d7e14b27a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 19:25:19 +0100
Subject: [PATCH 011/342] Cleanup

---
 src/thorin/be/backends.cpp | 2 +-
 src/thorin/be/backends.h   | 5 +----
 2 files changed, 2 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index df4c6e9d3..8a838f203 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -99,7 +99,7 @@ Backends::Backends(World& world, int opt, bool debug)
         kernels.emplace_back(continuation);
     });
 
-    for (auto backend : gpu_compute_backends) {
+    for (auto backend : std::array { Cuda, NVVM, OpenCL, AMDGPU }) {
         if (!importers_[backend].world().empty()) {
             get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 // determine whether or not this kernel uses restrict pointers
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index dce12fd55..d0093e9e0 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -46,10 +46,7 @@ struct Backends {
     enum { Cuda, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
 
-    /// Backends that need GPUKernelConfig
-    static constexpr auto gpu_compute_backends = { Cuda, NVVM, OpenCL, AMDGPU };
-
-    static constexpr std::array<const char*, BackendCount> backends_extensions = { ".cu", ".nvvm", ".cl", ".amdgpu", ".hls" };
+    static constexpr const char* file_exts[] = { ".cu", ".nvvm", ".cl", ".amdgpu", ".hls" };
 private:
     std::vector<Importer> importers_;
 };

From 8c49aaed74b35f43e85cfe02000375fa47b2b1a3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 19:26:20 +0100
Subject: [PATCH 012/342] Indentation

---
 src/thorin/be/backends.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 8a838f203..10904cd80 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -119,9 +119,9 @@ Backends::Backends(World& world, int opt, bool debug)
                     it_config->op(1)->isa<PrimLit>() &&
                     it_config->op(2)->isa<PrimLit>()) {
                     return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{
-                            it_config->op(0)->as<PrimLit>()->qu32_value().data(),
-                            it_config->op(1)->as<PrimLit>()->qu32_value().data(),
-                            it_config->op(2)->as<PrimLit>()->qu32_value().data()
+                        it_config->op(0)->as<PrimLit>()->qu32_value().data(),
+                        it_config->op(1)->as<PrimLit>()->qu32_value().data(),
+                        it_config->op(2)->as<PrimLit>()->qu32_value().data()
                     }, has_restrict);
                 }
                 return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{-1, -1, -1}, has_restrict);

From 76314568aede9aa7119a7caf91ee973052be5a46 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 19:31:43 +0100
Subject: [PATCH 013/342] Simplify

---
 src/thorin/be/backends.cpp | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 10904cd80..df31fbb36 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -171,9 +171,8 @@ Backends::Backends(World& world, int opt, bool debug)
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
-    if (!importers_[Cuda  ].world().empty()) device_cgs[Cuda  ] = std::make_unique<c::CodeGen>(importers_[Cuda  ].world(), kernel_config, c::Lang::CUDA  , debug);
-    if (!importers_[OpenCL].world().empty()) device_cgs[OpenCL] = std::make_unique<c::CodeGen>(importers_[OpenCL].world(), kernel_config, c::Lang::OPENCL, debug);
-    if (!importers_[HLS   ].world().empty()) device_cgs[HLS   ] = std::make_unique<c::CodeGen>(importers_[HLS   ].world(), kernel_config, c::Lang::HLS   , debug);
+    for (auto [backend, lang] : std::array { std::pair { Cuda, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OPENCL }, std::pair { HLS, c::Lang::HLS } })
+        if (!importers_[backend].world().empty()) device_cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug);
 }
 
 CodeGen::CodeGen(World& world, bool debug)

From ea1e8b285a113e4e6ea1b02d9b29b9ad15df60f6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 19:37:18 +0100
Subject: [PATCH 014/342] Remove file extensions

---
 src/thorin/be/backends.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index d0093e9e0..0879b4426 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -45,8 +45,6 @@ struct Backends {
 
     enum { Cuda, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
-
-    static constexpr const char* file_exts[] = { ".cu", ".nvvm", ".cl", ".amdgpu", ".hls" };
 private:
     std::vector<Importer> importers_;
 };

From eaf4555e83d17c78a5587f14f751853214f82b06 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ars=C3=A8ne=20P=C3=A9rard-Gayot?=
 <arsene.perard@laposte.net>
Date: Mon, 15 Feb 2021 20:05:46 +0100
Subject: [PATCH 015/342] Fix connection between intrinsic name and backend
 name

---
 src/thorin/be/backends.cpp | 21 ++++++++++++++-------
 src/thorin/be/backends.h   |  2 +-
 2 files changed, 15 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index df31fbb36..b843e96cc 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -72,18 +72,25 @@ static uint64_t get_alloc_size(const Def* def) {
 }
 
 Backends::Backends(World& world, int opt, bool debug)
-    : device_cgs({})
+    : device_cgs {}
 {
-    for (int backend = 0; backend < BackendCount; backend++)
+    for (size_t i = 0; i < device_cgs.size(); ++i)
         importers_.emplace_back(world);
 
     // determine different parts of the world which need to be compiled differently
     Scope::for_each(world, [&] (const Scope& scope) {
         auto continuation = scope.entry();
         Continuation* imported = nullptr;
-        for (int backend = 0; backend <= BackendCount; backend++) {
-            if (backend == BackendCount) return;
-            if (is_passed_to_intrinsic(continuation, Intrinsic(int(Intrinsic::AcceleratorBegin) + backend))) {
+
+        static const auto backend_intrinsics = std::array {
+            std::pair { CUDA,   Intrinsic::CUDA   },
+            std::pair { NVVM,   Intrinsic::NVVM   },
+            std::pair { OpenCL, Intrinsic::OpenCL },
+            std::pair { AMDGPU, Intrinsic::AMDGPU },
+            std::pair { HLS,    Intrinsic::HLS    }
+        };
+        for (auto [backend, intrinsic] : backend_intrinsics) {
+            if (is_passed_to_intrinsic(continuation, intrinsic)) {
                 imported = importers_[backend].import(continuation)->as_continuation();
                 break;
             }
@@ -99,7 +106,7 @@ Backends::Backends(World& world, int opt, bool debug)
         kernels.emplace_back(continuation);
     });
 
-    for (auto backend : std::array { Cuda, NVVM, OpenCL, AMDGPU }) {
+    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU }) {
         if (!importers_[backend].world().empty()) {
             get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 // determine whether or not this kernel uses restrict pointers
@@ -171,7 +178,7 @@ Backends::Backends(World& world, int opt, bool debug)
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
-    for (auto [backend, lang] : std::array { std::pair { Cuda, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OPENCL }, std::pair { HLS, c::Lang::HLS } })
+    for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OPENCL }, std::pair { HLS, c::Lang::HLS } })
         if (!importers_[backend].world().empty()) device_cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug);
 }
 
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index 0879b4426..fa45d52f1 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -43,7 +43,7 @@ struct Backends {
 
     std::unique_ptr<CodeGen> cpu_cg;
 
-    enum { Cuda, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
+    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
 private:
     std::vector<Importer> importers_;

From 70fef148f6d67866b8b35128829500f16300c064 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Tue, 16 Feb 2021 11:44:55 +0100
Subject: [PATCH 016/342] fix nullptr dereference

---
 src/thorin/be/backends.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index b843e96cc..369b14880 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -96,6 +96,9 @@ Backends::Backends(World& world, int opt, bool debug)
             }
         }
 
+        if (imported == nullptr)
+            return;
+
         imported->set_name(continuation->unique_name());
         imported->make_exported();
         continuation->set_name(continuation->unique_name());

From feb622fa6b35e2df0b8b377a6ff071df0b41d8d0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 12 Feb 2021 11:50:31 +0100
Subject: [PATCH 017/342] initial stub

---
 CMakeLists.txt                |  4 +++-
 src/thorin/CMakeLists.txt     |  7 +++++++
 src/thorin/be/spirv/spirv.cpp | 37 +++++++++++++++++++++++++++++++++++
 src/thorin/be/spirv/spirv.h   | 20 +++++++++++++++++++
 4 files changed, 67 insertions(+), 1 deletion(-)
 create mode 100644 src/thorin/be/spirv/spirv.cpp
 create mode 100644 src/thorin/be/spirv/spirv.h

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4c7d8b0f7..3bcb33f56 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -10,7 +10,7 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 option(BUILD_SHARED_LIBS "Build shared libraries" ON)
 option(THORIN_PROFILE "profile complexity in thorin::HashTable - only works in Debug build" ON)
-
+option(SPIRV_ENABLED "Enable spir-v backend" ON)
 
 if(CMAKE_BUILD_TYPE STREQUAL "")
     set(CMAKE_BUILD_TYPE Debug CACHE STRING "Debug or Release" FORCE)
@@ -41,6 +41,8 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
+find_package(SPIRV-Headers)
+
 message(STATUS "Using Debug flags: ${CMAKE_CXX_FLAGS_DEBUG}")
 message(STATUS "Using Release flags: ${CMAKE_CXX_FLAGS_RELEASE}")
 if(DEFINED CMAKE_BUILD_TYPE)
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 81bf603a8..c1c13a5ec 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -104,6 +104,13 @@ if(LLVM_FOUND)
     )
 endif()
 
+if(SPIRV_ENABLED)
+    list(APPEND THORIN_SOURCES
+        be/spirv/spirv.cpp
+        be/spirv/spirv.h
+    )
+endif()
+
 add_library(thorin ${THORIN_SOURCES})
 
 if(LLVM_FOUND)
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
new file mode 100644
index 000000000..33940c27e
--- /dev/null
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -0,0 +1,37 @@
+#include "thorin/be/spirv/spirv.h"
+
+#include <spirv/unified1/spirv.hpp>
+
+#include <iostream>
+#include <fstream>
+
+struct SpvFileBuilder {
+    SpvFileBuilder(std::ostream& output) : output(output) {}
+
+    std::ostream& output;
+    uint32_t bound = 0;
+
+    void finish() {
+        output << spv::MagicNumber;
+        output << spv::Version; // TODO: target a specific spirv version
+        output << uint32_t(0); // TODO get a magic number ?
+        output << bound;
+        output << uint32_t(0); // instruction schema padding
+    }
+};
+
+struct SpvMethodBuilder {
+
+};
+
+thorin::SpirVCodeGen::SpirVCodeGen(thorin::World& world)
+    : world_(world)
+{}
+
+void thorin::SpirVCodeGen::emit() {
+    std::ofstream myfile;
+    myfile.open ("test.spv");
+    auto builder = SpvFileBuilder(myfile);
+    builder.finish();
+    myfile.close();
+}
\ No newline at end of file
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
new file mode 100644
index 000000000..419953f5c
--- /dev/null
+++ b/src/thorin/be/spirv/spirv.h
@@ -0,0 +1,20 @@
+#ifndef THORIN_SPIRV_H
+#define THORIN_SPIRV_H
+
+#include "thorin/continuation.h"
+
+namespace thorin {
+
+class SpirVCodeGen {
+public:
+    SpirVCodeGen(World& world);
+
+    void emit();
+
+protected:
+    World& world_;
+};
+
+}
+
+#endif //THORIN_SPIRV_H

From 56cdbfbc0b6808813e72b3e5d3c4ca6d8774be3e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 12 Feb 2021 14:19:54 +0100
Subject: [PATCH 018/342] builds legal empty spir-v module

---
 src/thorin/be/spirv/spirv.cpp | 159 ++++++++++++++++++++++++++++++----
 src/thorin/be/spirv/spirv.h   |   3 +-
 2 files changed, 146 insertions(+), 16 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 33940c27e..10f21c0a9 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -1,37 +1,166 @@
 #include "thorin/be/spirv/spirv.h"
+#include "thorin/analyses/scope.h"
 
 #include <spirv/unified1/spirv.hpp>
 
 #include <iostream>
 #include <fstream>
 
-struct SpvFileBuilder {
-    SpvFileBuilder(std::ostream& output) : output(output) {}
+ int div_roundup(int a, int b) {
+    if (a % b == 0)
+        return a / b;
+    else
+        return (a / b) + 1;
+}
 
-    std::ostream& output;
-    uint32_t bound = 0;
+namespace thorin {
 
-    void finish() {
-        output << spv::MagicNumber;
-        output << spv::Version; // TODO: target a specific spirv version
-        output << uint32_t(0); // TODO get a magic number ?
-        output << bound;
-        output << uint32_t(0); // instruction schema padding
+struct SpvId { uint32_t id; };
+
+struct SpvSectionBuilder {
+    std::vector<uint32_t> data_;
+
+private:
+    void output_word(uint32_t word) {
+        data_.push_back(word);
+    }
+public:
+    void op(spv::Op op, int ops_size) {
+        uint32_t lower = op & 0xFFFFu;
+        uint32_t upper = (ops_size << 16) & 0xFFFF0000u;
+        output_word(lower | upper);
+    }
+
+    void ref_id(SpvId id) {
+        output_word(id.id);
+    }
+
+    void literal_name(std::string_view str) {
+        int i = 0;
+        uint32_t cword = 0;
+        for (char c : str) {
+            cword = cword | (c & 0xFF) << (i * 8);
+            i++;
+            if (i == 4) {
+                output_word(cword);
+                cword = 0;
+                i = 0;
+            }
+        }
+        output_word(cword);
     }
 };
 
-struct SpvMethodBuilder {
+struct SpvFileBuilder {
+    SpvFileBuilder(std::ostream& output) : output_(output) {}
+
+    SpvId fresh_id() { return { bound++ }; }
+
+    void name(SpvId id, std::string_view str) {
+        assert(id.id < bound);
+        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
+        debug_names.ref_id(id);
+        debug_names.literal_name(str);
+    }
+
+    SpvId declare_bool_type() {
+        types_constants.op(spv::Op::OpTypeBool, 2);
+        auto id = fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    void capability(spv::Capability cap) {
+        capabilities.op(spv::Op::OpCapability, 2);
+        capabilities.data_.push_back(cap);
+    }
+
+    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
+
+private:
+    std::ostream& output_;
+    uint32_t bound = 1;
+
+    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
+    SpvSectionBuilder capabilities;
+    SpvSectionBuilder extensions;
+    SpvSectionBuilder ext_inst_import;
+    SpvSectionBuilder entry_points;
+    SpvSectionBuilder execution_modes;
+    SpvSectionBuilder debug_string_source;
+    SpvSectionBuilder debug_names;
+    SpvSectionBuilder debug_module_processed;
+    SpvSectionBuilder annotations;
+    SpvSectionBuilder types_constants;
+    std::vector<SpvSectionBuilder> fn_decls;
+    std::vector<SpvSectionBuilder> fn_defs;
+
+    void output_word_le(uint32_t word) {
+        output_.put((word >> 0) & 0xFFu);
+        output_.put((word >> 8) & 0xFFu);
+        output_.put((word >> 16) & 0xFFu);
+        output_.put((word >> 24) & 0xFFu);
+    }
+
+    void output_section(SpvSectionBuilder& section) {
+        for (auto& word : section.data_) {
+            output_word_le(word);
+        }
+    }
+public:
+    void finish() {
+        SpvSectionBuilder memory_model_section;
+        memory_model_section.op(spv::Op::OpMemoryModel, 3);
+        memory_model_section.data_.push_back(addressing_model);
+        memory_model_section.data_.push_back(memory_model);
+
+        output_word_le(spv::MagicNumber);
+        output_word_le(spv::Version); // TODO: target a specific spirv version
+        output_word_le(uint32_t(0)); // TODO get a magic number ?
+        output_word_le(bound);
+        output_word_le(uint32_t(0)); // instruction schema padding
 
+        output_section(capabilities);
+        output_section(extensions);
+        output_section(ext_inst_import);
+        output_section(memory_model_section);
+        output_section(entry_points);
+        output_section(execution_modes);
+        output_section(debug_string_source);
+        output_section(debug_names);
+        output_section(debug_module_processed);
+        output_section(annotations);
+        output_section(types_constants);
+        for (auto& decl : fn_decls)
+            output_section(decl);
+        for (auto& def : fn_defs)
+            output_section(def);
+    }
 };
 
 thorin::SpirVCodeGen::SpirVCodeGen(thorin::World& world)
-    : world_(world)
-{}
+: world_(world) {}
 
 void thorin::SpirVCodeGen::emit() {
     std::ofstream myfile;
-    myfile.open ("test.spv");
+    myfile.open("test.spv");
     auto builder = SpvFileBuilder(myfile);
+
+    builder.capability(spv::Capability::CapabilityShader);
+    builder.capability(spv::Capability::CapabilityLinkage);
+
+    builder.name(builder.declare_bool_type(), "test");
+
+    Scope::for_each(world_, [&](const Scope& scope) { emit(scope); });
+
     builder.finish();
+    myfile.flush();
     myfile.close();
-}
\ No newline at end of file
+}
+
+void thorin::SpirVCodeGen::emit(const thorin::Scope& scope) {
+
+}
+
+}
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 419953f5c..ffebd6735 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -10,9 +10,10 @@ class SpirVCodeGen {
     SpirVCodeGen(World& world);
 
     void emit();
-
 protected:
     World& world_;
+
+    void emit(const Scope& scope);
 };
 
 }

From 3a91cb71725a90939e95c04bac6539a1fb87249c Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 12 Feb 2021 14:44:30 +0100
Subject: [PATCH 019/342] report spirv configuration to cmake

---
 CMakeLists.txt               | 5 ++++-
 cmake/thorin-config.cmake.in | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3bcb33f56..c1e89e2ec 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -41,7 +41,10 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
-find_package(SPIRV-Headers)
+if (SPIRV_ENABLED)
+    find_package(SPIRV-Headers REQUIRED)
+    message(STATUS "Enabled SPIR-V backend")
+endif()
 
 message(STATUS "Using Debug flags: ${CMAKE_CXX_FLAGS_DEBUG}")
 message(STATUS "Using Release flags: ${CMAKE_CXX_FLAGS_RELEASE}")
diff --git a/cmake/thorin-config.cmake.in b/cmake/thorin-config.cmake.in
index fe7e3220c..f1dc27bd6 100644
--- a/cmake/thorin-config.cmake.in
+++ b/cmake/thorin-config.cmake.in
@@ -30,6 +30,7 @@ find_package(Half REQUIRED)
 
 set(Thorin_HAS_LLVM_SUPPORT @LLVM_FOUND@)
 set(Thorin_HAS_RV_SUPPORT @RV_FOUND@)
+set(Thorin_HAS_SPIRV_SUPPORT @SPIRV_ENABLED@)
 set(AnyDSL_LLVM_LINK_SHARED @AnyDSL_LLVM_LINK_SHARED@)
 
 if(Thorin_HAS_LLVM_SUPPORT)

From e2008654f0a44beb2f2363522400bf3fce001db0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 12 Feb 2021 15:28:55 +0100
Subject: [PATCH 020/342] add vk_compute intrinsic

---
 src/thorin/continuation.cpp | 1 +
 src/thorin/continuation.h   | 1 +
 2 files changed, 2 insertions(+)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 81e26b025..0797dbe5e 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -167,6 +167,7 @@ void Continuation::set_intrinsic() {
     else if (name() == "nvvm")                 attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")               attributes().intrinsic = Intrinsic::OpenCL;
     else if (name() == "amdgpu")               attributes().intrinsic = Intrinsic::AMDGPU;
+    else if (name() == "vk_compute")           attributes().intrinsic = Intrinsic::VulkanCompute;
     else if (name() == "hls")                  attributes().intrinsic = Intrinsic::HLS;
     else if (name() == "parallel")             attributes().intrinsic = Intrinsic::Parallel;
     else if (name() == "fibers")               attributes().intrinsic = Intrinsic::Fibers;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 978c39f15..8b3085319 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -62,6 +62,7 @@ enum class Intrinsic : uint8_t {
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
     AMDGPU,                     ///< Internal AMDGPU-Backend.
+    VulkanCompute,              ///< Internal Vulkan-Compute-Shader-Backend.
     HLS,                        ///< Internal HLS-Backend.
     Parallel,                   ///< Internal Parallel-CPU-Backend.
     Fibers,                     ///< Internal Parallel-CPU-Backend using resumable fibers.

From 24cc66c23ed5577cbd8dfbc3ff06c4aa656a47f4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Tue, 16 Feb 2021 09:56:52 +0100
Subject: [PATCH 021/342] adapt to backends refactor

---
 CMakeLists.txt                |  3 +++
 src/thorin/be/backends.cpp    | 17 ++++++++++++-----
 src/thorin/be/backends.h      |  2 +-
 src/thorin/be/spirv/spirv.cpp | 22 +++++++++-------------
 src/thorin/be/spirv/spirv.h   | 12 +++++-------
 src/thorin/config.h.in        |  1 +
 src/thorin/continuation.cpp   |  2 +-
 src/thorin/continuation.h     |  2 +-
 8 files changed, 33 insertions(+), 28 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c1e89e2ec..58177d42c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -65,6 +65,9 @@ endif()
 if(RV_FOUND)
     set(THORIN_ENABLE_RV TRUE)
 endif()
+if(LLVM_FOUND)
+    set(THORIN_ENABLE_SPIRV TRUE)
+endif()
 configure_file(src/thorin/config.h.in ${CMAKE_BINARY_DIR}/include/thorin/config.h @ONLY)
 include_directories(${CMAKE_BINARY_DIR}/include)
 
diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 369b14880..2808c74e3 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -7,6 +7,9 @@
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu.h"
 #endif
+#if THORIN_ENABLE_SPIRV
+#include "thorin/be/spirv/spirv.h"
+#endif
 #include "thorin/be/c/c.h"
 
 namespace thorin {
@@ -83,11 +86,12 @@ Backends::Backends(World& world, int opt, bool debug)
         Continuation* imported = nullptr;
 
         static const auto backend_intrinsics = std::array {
-            std::pair { CUDA,   Intrinsic::CUDA   },
-            std::pair { NVVM,   Intrinsic::NVVM   },
-            std::pair { OpenCL, Intrinsic::OpenCL },
-            std::pair { AMDGPU, Intrinsic::AMDGPU },
-            std::pair { HLS,    Intrinsic::HLS    }
+            std::pair { CUDA,      Intrinsic::CUDA      },
+            std::pair { NVVM,      Intrinsic::NVVM      },
+            std::pair { OpenCL,    Intrinsic::OpenCL    },
+            std::pair { AMDGPU,    Intrinsic::AMDGPU    },
+            std::pair { HLS,       Intrinsic::HLS       },
+            std::pair { VkCompute, Intrinsic::VkCompute }
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
@@ -180,6 +184,9 @@ Backends::Backends(World& world, int opt, bool debug)
     if (!importers_[AMDGPU].world().empty()) device_cgs[AMDGPU] = std::make_unique<llvm::AMDGPUCodeGen>(importers_[AMDGPU].world(), kernel_config, opt, debug);
 #else
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
+#endif
+#if THORIN_ENABLE_SPIRV
+    if (!importers_[VkCompute].world().empty()) device_cgs[VkCompute] = std::make_unique<spirv::CodeGen>(importers_[VkCompute].world(), kernel_config, debug);
 #endif
     for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OPENCL }, std::pair { HLS, c::Lang::HLS } })
         if (!importers_[backend].world().empty()) device_cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug);
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index fa45d52f1..2d366b23a 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -43,7 +43,7 @@ struct Backends {
 
     std::unique_ptr<CodeGen> cpu_cg;
 
-    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
+    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, VkCompute, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
 private:
     std::vector<Importer> importers_;
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 10f21c0a9..34c1bed97 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -4,16 +4,15 @@
 #include <spirv/unified1/spirv.hpp>
 
 #include <iostream>
-#include <fstream>
 
- int div_roundup(int a, int b) {
+int div_roundup(int a, int b) {
     if (a % b == 0)
         return a / b;
     else
         return (a / b) + 1;
 }
 
-namespace thorin {
+namespace thorin::spirv {
 
 struct SpvId { uint32_t id; };
 
@@ -139,27 +138,24 @@ struct SpvFileBuilder {
     }
 };
 
-thorin::SpirVCodeGen::SpirVCodeGen(thorin::World& world)
-: world_(world) {}
+CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
+    : thorin::CodeGen(world, debug)
+{}
 
-void thorin::SpirVCodeGen::emit() {
-    std::ofstream myfile;
-    myfile.open("test.spv");
-    auto builder = SpvFileBuilder(myfile);
+void CodeGen::emit(std::ostream& out) {
+    auto builder = SpvFileBuilder(out);
 
     builder.capability(spv::Capability::CapabilityShader);
     builder.capability(spv::Capability::CapabilityLinkage);
 
     builder.name(builder.declare_bool_type(), "test");
 
-    Scope::for_each(world_, [&](const Scope& scope) { emit(scope); });
+    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
     builder.finish();
-    myfile.flush();
-    myfile.close();
 }
 
-void thorin::SpirVCodeGen::emit(const thorin::Scope& scope) {
+void CodeGen::emit(const thorin::Scope& scope) {
 
 }
 
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index ffebd6735..e7b309f01 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -1,18 +1,16 @@
 #ifndef THORIN_SPIRV_H
 #define THORIN_SPIRV_H
 
-#include "thorin/continuation.h"
+#include "thorin/be/backends.h"
 
-namespace thorin {
+namespace thorin::spirv {
 
-class SpirVCodeGen {
+class CodeGen : public thorin::CodeGen {
 public:
-    SpirVCodeGen(World& world);
+    CodeGen(World&, Cont2Config&, bool debug);
 
-    void emit();
+    void emit(std::ostream& stream) override;
 protected:
-    World& world_;
-
     void emit(const Scope& scope);
 };
 
diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index 5e061fe28..c594de907 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -5,5 +5,6 @@
 #cmakedefine01 THORIN_ENABLE_PROFILING
 #cmakedefine01 THORIN_ENABLE_LLVM
 #cmakedefine01 THORIN_ENABLE_RV
+#cmakedefine01 THORIN_ENABLE_SPIRV
 
 #endif
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 0797dbe5e..b354e8647 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -167,7 +167,7 @@ void Continuation::set_intrinsic() {
     else if (name() == "nvvm")                 attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")               attributes().intrinsic = Intrinsic::OpenCL;
     else if (name() == "amdgpu")               attributes().intrinsic = Intrinsic::AMDGPU;
-    else if (name() == "vk_compute")           attributes().intrinsic = Intrinsic::VulkanCompute;
+    else if (name() == "vk_compute")           attributes().intrinsic = Intrinsic::VkCompute;
     else if (name() == "hls")                  attributes().intrinsic = Intrinsic::HLS;
     else if (name() == "parallel")             attributes().intrinsic = Intrinsic::Parallel;
     else if (name() == "fibers")               attributes().intrinsic = Intrinsic::Fibers;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 8b3085319..3f6697d1b 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -62,7 +62,7 @@ enum class Intrinsic : uint8_t {
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
     AMDGPU,                     ///< Internal AMDGPU-Backend.
-    VulkanCompute,              ///< Internal Vulkan-Compute-Shader-Backend.
+    VkCompute,                  ///< Internal Vulkan-Compute-Shader-Backend.
     HLS,                        ///< Internal HLS-Backend.
     Parallel,                   ///< Internal Parallel-CPU-Backend.
     Fibers,                     ///< Internal Parallel-CPU-Backend using resumable fibers.

From 4a0db929f3e9aa9e6dfb75d1770aea855f6e77f0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Feb 2021 12:48:49 +0100
Subject: [PATCH 022/342] working stub for function definition

---
 src/thorin/be/llvm/llvm.cpp   |   1 +
 src/thorin/be/spirv/spirv.cpp | 468 ++++++++++++++++++++++++++++++++--
 src/thorin/be/spirv/spirv.h   |  15 ++
 3 files changed, 461 insertions(+), 23 deletions(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 5a2c867ce..e11e5d43a 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1115,6 +1115,7 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
         case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
         case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
         case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
+        case Intrinsic::VkCompute:   return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv",    continuation); // TODO have a real runtime component
         case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
         case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
         case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 34c1bed97..1e5bf883a 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -4,6 +4,7 @@
 #include <spirv/unified1/spirv.hpp>
 
 #include <iostream>
+#include <thorin/analyses/schedule.h>
 
 int div_roundup(int a, int b) {
     if (a % b == 0)
@@ -14,8 +15,6 @@ int div_roundup(int a, int b) {
 
 namespace thorin::spirv {
 
-struct SpvId { uint32_t id; };
-
 struct SpvSectionBuilder {
     std::vector<uint32_t> data_;
 
@@ -31,6 +30,7 @@ struct SpvSectionBuilder {
     }
 
     void ref_id(SpvId id) {
+        assert(id.id != 0);
         output_word(id.id);
     }
 
@@ -48,12 +48,46 @@ struct SpvSectionBuilder {
         }
         output_word(cword);
     }
+
+    void literal_int(uint32_t i) {
+        output_word(i);
+    }
+};
+
+struct SpvBasicBlockBuilder : public SpvSectionBuilder {
+    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
+        : file_builder(file_builder)
+    {}
+
+    SpvFileBuilder& file_builder;
+
+    SpvId label() {
+        op(spv::Op::OpLabel, 2);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        return id;
+    }
+
+    void return_void() {
+        op(spv::Op::OpReturn, 1);
+    }
+private:
+    SpvId generate_fresh_id();
+};
+
+struct SpvFnBuilder {
+    SpvId fn_type;
+    SpvId fn_ret_type;
+    ContinuationMap<std::unique_ptr<SpvBasicBlockBuilder>> bbs;
+    ContinuationMap<SpvId> labels;
 };
 
 struct SpvFileBuilder {
-    SpvFileBuilder(std::ostream& output) : output_(output) {}
+    SpvFileBuilder()
+        : void_type(declare_void_type())
+    {}
 
-    SpvId fresh_id() { return { bound++ }; }
+    SpvId generate_fresh_id() { return {bound++ }; }
 
     void name(SpvId id, std::string_view str) {
         assert(id.id < bound);
@@ -64,11 +98,56 @@ struct SpvFileBuilder {
 
     SpvId declare_bool_type() {
         types_constants.op(spv::Op::OpTypeBool, 2);
-        auto id = fresh_id();
+        auto id = generate_fresh_id();
         types_constants.ref_id(id);
         return id;
     }
 
+    SpvId declare_int_type(int width, bool signed_) {
+        types_constants.op(spv::Op::OpTypeInt, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        types_constants.literal_int(signed_ ? 1 : 0);
+        return id;
+    }
+
+    SpvId declare_float_type(int width) {
+        types_constants.op(spv::Op::OpTypeFloat, 3);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        return id;
+    }
+
+    SpvId declare_fn_type(std::vector<SpvId>& dom, SpvId codom) {
+        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(codom);
+        for (auto arg : dom)
+            types_constants.ref_id(arg);
+        return id;
+    }
+
+    SpvId define_function(SpvFnBuilder& fn_builder) {
+        fn_defs.op(spv::Op::OpFunction, 5);
+        fn_defs.ref_id(fn_builder.fn_ret_type);
+        auto id = generate_fresh_id();
+        fn_defs.ref_id(id);
+        fn_defs.data_.push_back(spv::FunctionControlMaskNone);
+        fn_defs.ref_id(fn_builder.fn_type);
+
+        // TODO OpFunctionParameters
+        for (auto& [cont, bb] : fn_builder.bbs) {
+            for (auto w : bb->data_)
+                fn_defs.data_.push_back(w);
+        }
+
+        fn_defs.op(spv::Op::OpFunctionEnd, 1);
+        return id;
+    }
+
     void capability(spv::Capability cap) {
         capabilities.op(spv::Op::OpCapability, 2);
         capabilities.data_.push_back(cap);
@@ -78,7 +157,7 @@ struct SpvFileBuilder {
     spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
 
 private:
-    std::ostream& output_;
+    std::ostream* output_ = nullptr;
     uint32_t bound = 1;
 
     // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
@@ -92,14 +171,21 @@ struct SpvFileBuilder {
     SpvSectionBuilder debug_module_processed;
     SpvSectionBuilder annotations;
     SpvSectionBuilder types_constants;
-    std::vector<SpvSectionBuilder> fn_decls;
-    std::vector<SpvSectionBuilder> fn_defs;
+    SpvSectionBuilder fn_decls;
+    SpvSectionBuilder fn_defs;
+
+    SpvId declare_void_type() {
+        types_constants.op(spv::Op::OpTypeVoid, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
 
     void output_word_le(uint32_t word) {
-        output_.put((word >> 0) & 0xFFu);
-        output_.put((word >> 8) & 0xFFu);
-        output_.put((word >> 16) & 0xFFu);
-        output_.put((word >> 24) & 0xFFu);
+        output_->put((word >> 0) & 0xFFu);
+        output_->put((word >> 8) & 0xFFu);
+        output_->put((word >> 16) & 0xFFu);
+        output_->put((word >> 24) & 0xFFu);
     }
 
     void output_section(SpvSectionBuilder& section) {
@@ -108,7 +194,10 @@ struct SpvFileBuilder {
         }
     }
 public:
-    void finish() {
+    const SpvId void_type;
+
+    void finish(std::ostream& output) {
+        output_ = &output;
         SpvSectionBuilder memory_model_section;
         memory_model_section.op(spv::Op::OpMemoryModel, 3);
         memory_model_section.data_.push_back(addressing_model);
@@ -131,32 +220,365 @@ struct SpvFileBuilder {
         output_section(debug_module_processed);
         output_section(annotations);
         output_section(types_constants);
-        for (auto& decl : fn_decls)
-            output_section(decl);
-        for (auto& def : fn_defs)
-            output_section(def);
+        output_section(fn_decls);
+        output_section(fn_defs);
     }
 };
 
+SpvId SpvBasicBlockBuilder::generate_fresh_id() {
+    return file_builder.generate_fresh_id();
+}
+
 CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
     : thorin::CodeGen(world, debug)
 {}
 
 void CodeGen::emit(std::ostream& out) {
-    auto builder = SpvFileBuilder(out);
+    SpvFileBuilder builder;
+    builder_ = &builder;
+    builder_->capability(spv::Capability::CapabilityShader);
+    builder_->capability(spv::Capability::CapabilityLinkage);
+
+    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
-    builder.capability(spv::Capability::CapabilityShader);
-    builder.capability(spv::Capability::CapabilityLinkage);
+    builder_->finish(out);
+    builder_ = nullptr;
+}
 
-    builder.name(builder.declare_bool_type(), "test");
+SpvId CodeGen::convert(const Type* type) {
+    if (auto llvm_type = types_.lookup(type)) return *llvm_type;
 
-    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
+    assert(!type->isa<MemType>());
+    SpvId spv_type;
+    switch (type->tag()) {
+        case PrimType_bool:                                                             spv_type = builder_->declare_bool_type(); break;
+        case PrimType_ps8:  case PrimType_qs8:  case PrimType_pu8:  case PrimType_qu8:  assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps16: case PrimType_qs16: case PrimType_pu16: case PrimType_qu16: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps32: case PrimType_qs32:                                         spv_type = builder_->declare_int_type(32, true ); break;
+        case PrimType_pu32: case PrimType_qu32:                                         spv_type = builder_->declare_int_type(32, false); break;
+        case PrimType_ps64: case PrimType_qs64: case PrimType_pu64: case PrimType_qu64: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf16: case PrimType_qf16:                                         assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf32: case PrimType_qf32:                                         spv_type = builder_->declare_float_type(32); break;
+        case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
+        case Node_PtrType: {
+            auto ptr = type->as<PtrType>();
+            assert(false && "TODO");
+            //llvm_type = llvm::PointerType::get(convert(ptr->pointee()), convert_addr_space(ptr->addr_space()));
+            break;
+        }
+        case Node_IndefiniteArrayType: {
+            assert(false && "TODO");
+            //llvm_type = llvm::ArrayType::get(convert(type->as<ArrayType>()->elem_type()), 0);
+            //return types_[type] = llvm_type;
+        }
+        case Node_DefiniteArrayType: {
+            assert(false && "TODO");
+            auto array = type->as<DefiniteArrayType>();
+            //llvm_type = llvm::ArrayType::get(convert(array->elem_type()), array->dim());
+            //return types_[type] = llvm_type;
+        }
+
+        case Node_ClosureType:
+        case Node_FnType: {
+            // extract "return" type, collect all other types
+            auto fn = type->as<FnType>();
+            std::unique_ptr<SpvId> ret;
+            std::vector<SpvId> ops;
+            for (auto op : fn->ops()) {
+                if (op->isa<MemType>() || op == world().unit()) continue;
+                auto fn = op->isa<FnType>();
+                if (fn && !op->isa<ClosureType>()) {
+                    assert(!ret && "only one 'return' supported");
+                    std::vector<SpvId> ret_types;
+                    for (auto fn_op : fn->ops()) {
+                        if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
+                        ret_types.push_back(convert(fn_op));
+                    }
+                    if (ret_types.size() == 0)      ret = std::make_unique<SpvId>(builder_->void_type);
+                    else if (ret_types.size() == 1) ret = std::make_unique<SpvId>(ret_types.back());
+                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?"); //ret = llvm::StructType::get(context(), ret_types);
+                } else
+                    ops.push_back(convert(op));
+            }
+            assert(ret);
+
+            if (type->tag() == Node_FnType) {
+                return types_[type] = builder_->declare_fn_type(ops, *ret);
+            }
+
+            assert(false && "TODO: handle closure mess");
+            /* auto env_type = convert(Closure::environment_type(world()));
+            ops.push_back(env_type);
+            auto fn_type = llvm::FunctionType::get(ret, ops, false);
+            auto ptr_type = llvm::PointerType::get(fn_type, 0);
+            llvm_type = llvm::StructType::get(context(), { ptr_type, env_type });
+            return types_[type] = llvm_type;*/
+        }
+
+        case Node_StructType: {
+            assert(false && "TODO");
+            /*auto struct_type = type->as<StructType>();
+            auto llvm_struct = llvm::StructType::create(context());
 
-    builder.finish();
+            // important: memoize before recursing into element types to avoid endless recursion
+            assert(!types_.contains(struct_type) && "type already converted");
+            types_[struct_type] = llvm_struct;
+
+            Array<llvm::Type*> llvm_types(struct_type->num_ops());
+            for (size_t i = 0, e = llvm_types.size(); i != e; ++i)
+                llvm_types[i] = convert(struct_type->op(i));
+            llvm_struct->setBody(llvm_ref(llvm_types));
+            return llvm_struct;*/
+        }
+
+        case Node_TupleType: {
+            assert(false && "TODO");
+            /*auto tuple = type->as<TupleType>();
+            Array<llvm::Type*> llvm_types(tuple->num_ops());
+            for (size_t i = 0, e = llvm_types.size(); i != e; ++i)
+                llvm_types[i] = convert(tuple->op(i));
+            llvm_type = llvm::StructType::get(context(), llvm_ref(llvm_types));
+            return types_[tuple] = llvm_type;*/
+        }
+
+        case Node_VariantType: {
+            assert(false && "TODO");
+            /*assert(type->num_ops() > 0);
+
+            // Max alignment/size constraints respectively in the variant type alternatives dictate the ones to use for the overall type
+            size_t max_align = 0, max_size = 0;
+
+            auto layout = module().getDataLayout();
+            llvm::Type* max_align_type;
+            for (auto op : type->ops()) {
+                auto op_type = convert(op);
+                size_t size  = layout.getTypeAllocSize(op_type);
+                size_t align = layout.getABITypeAlignment(op_type);
+                // Favor types that are not empty
+                if (align > max_align || (align == max_align && max_align_type->isEmptyTy())) {
+                    max_align_type = op_type;
+                    max_align = align;
+                }
+                max_size = std::max(max_size, size);
+            }
+
+            auto rem_size = max_size - layout.getTypeAllocSize(max_align_type);
+            auto union_type = rem_size > 0
+                              ? llvm::StructType::get(context(), llvm::ArrayRef<llvm::Type*> { max_align_type, llvm::ArrayType::get(llvm::Type::getInt8Ty(context()), rem_size)})
+                              : llvm::StructType::get(context(), llvm::ArrayRef<llvm::Type*> { max_align_type });
+
+            auto tag_type = type->num_ops() < (1_u64 <<  8) ? llvm::Type::getInt8Ty (context()) :
+                            type->num_ops() < (1_u64 << 16) ? llvm::Type::getInt16Ty(context()) :
+                            type->num_ops() < (1_u64 << 32) ? llvm::Type::getInt32Ty(context()) :
+                            llvm::Type::getInt64Ty(context());
+
+            return llvm::StructType::get(context(), { union_type, tag_type });*/
+        }
+
+        default:
+            THORIN_UNREACHABLE;
+    }
+
+    return types_[type] = spv_type;
 }
 
 void CodeGen::emit(const thorin::Scope& scope) {
+    entry_ = scope.entry();
+    assert(entry_->is_returning());
+    //auto fct = llvm::cast<llvm::Function>(emit(entry_));
+    auto fn = SpvFnBuilder { };
+    fn.fn_type = convert(entry_->type());
+    fn.fn_ret_type = builder_->void_type; // TODO !!!
+
+    current_fn_ = &fn;
+
+    //cont2llvm_.clear();
+    auto conts = schedule(scope);
+
+    // map all bb-like continuations to llvm bb stubs and handle params/phis
+    for (auto cont : conts) {
+        if (cont->intrinsic() == Intrinsic::EndScope) continue;
+
+        auto [i, b] = fn.bbs.emplace(cont, std::make_unique<SpvBasicBlockBuilder>(*builder_));
+        SpvBasicBlockBuilder& bb = *i->second;
+        SpvId label = bb.label();
+        builder_->name(label, cont->name().c_str());
+        fn.labels.emplace(cont, label);
+        //auto bb = llvm::BasicBlock::Create(context(), cont->name().c_str(), fct);
+        //auto [i, succ] = cont2llvm_.emplace(cont, std::pair(bb, std::make_unique<llvm::IRBuilder<>>(context())));
+        assert(b);
+        // auto& irbuilder = *i->second.second;
+        // irbuilder.SetInsertPoint(bb);
+
+        //if (debug())
+        //    irbuilder.SetCurrentDebugLocation(llvm::DebugLoc::get(cont->loc().begin.row, cont->loc().begin.row, discope));
+
+        /*if (entry_ == cont) {
+            auto arg = fct->arg_begin();
+            for (auto param : entry_->params()) {
+                if (is_mem(param) || is_unit(param)) {
+                    def2llvm_[param] = nullptr;
+                } else if (param->order() == 0) {
+                    auto argv = &*arg;
+                    auto value = map_param(fct, argv, param);
+                    if (value == argv) {
+                        arg->setName(param->unique_name()); // use param
+                        def2llvm_[param] = &*arg++;
+                    } else {
+                        def2llvm_[param] = value;           // use provided value
+                    }
+                }
+            }
+        } else {
+            for (auto param : cont->params()) {
+                if (is_mem(param) || is_unit(param)) {
+                    def2llvm_[param] = nullptr;
+                } else {
+                    // do not bother reserving anything (the 0 below) - it's a tiny optimization nobody cares about
+                    auto phi = irbuilder.CreatePHI(convert(param->type()), 0, param->name().c_str());
+                    def2llvm_[param] = phi;
+                }
+            }
+        }*/
+    }
+
+    Scheduler new_scheduler(scope);
+    swap(scheduler_, new_scheduler);
+
+    for (auto cont : conts) {
+        if (cont->intrinsic() == Intrinsic::EndScope) continue;
+        assert(cont == entry_ || cont->is_basicblock());
+        emit_epilogue(cont, *fn.bbs[cont]->get());
+    }
+
+    builder_->define_function(fn);
+}
+
+void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb) {
+    /*auto&& bb_ib = cont2llvm_[continuation];
+    auto bb = bb_ib->first;
+    auto& irbuilder = *bb_ib->second;*/
+
+    if (continuation->callee() == entry_->ret_param()) { // return
+        std::vector<SpvId> values;
+        std::vector<SpvId> types;
+
+        for (auto arg : continuation->args()) {
+            /*if (auto val = emit_unsafe(arg)) {
+                values.emplace_back(val);
+                types.emplace_back(val->getType());
+            }*/
+        }
+
+        switch (values.size()) {
+            case 0:  bb.return_void();      break;
+            //case 1:  irbuilder.CreateRet(values[0]); break;
+            default:
+                assert(false && "TODO handle non-void returns");
+                /*llvm::Value* agg = llvm::UndefValue::get(llvm::StructType::get(context(), types));
+
+                for (size_t i = 0, e = values.size(); i != e; ++i)
+                    agg = irbuilder.CreateInsertValue(agg, values[i], { unsigned(i) });
+
+                irbuilder.CreateRet(agg);*/
+        }
+    } /*else if (continuation->callee() == world().branch()) {
+        auto cond = emit(continuation->arg(0));
+        auto tbb = cont2bb(continuation->arg(1)->as_continuation());
+        auto fbb = cont2bb(continuation->arg(2)->as_continuation());
+        irbuilder.CreateCondBr(cond, tbb, fbb);
+    } else if (continuation->callee()->isa<Continuation>() &&
+               continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
+        auto val = emit(continuation->arg(0));
+        auto otherwise_bb = cont2bb(continuation->arg(1)->as_continuation());
+        auto match = irbuilder.CreateSwitch(val, otherwise_bb, continuation->num_args() - 2);
+        for (size_t i = 2; i < continuation->num_args(); i++) {
+            auto arg = continuation->arg(i)->as<Tuple>();
+            auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
+            auto case_bb    = cont2bb(arg->op(1)->as_continuation());
+            match->addCase(case_const, case_bb);
+        }
+    } else if (continuation->callee()->isa<Bottom>()) {
+        irbuilder.CreateUnreachable();
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+        for (size_t i = 0, e = continuation->num_args(); i != e; ++i) {
+            if (auto val = emit_unsafe(continuation->arg(i))) emit_phi_arg(irbuilder, callee->param(i), val);
+        }
+        irbuilder.CreateBr(cont2bb(callee));
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+        auto ret_continuation = emit_intrinsic(irbuilder, continuation);
+        irbuilder.CreateBr(cont2bb(ret_continuation));
+    } else { // function/closure call
+        // put all first-order args into an array
+        std::vector<llvm::Value*> args;
+        const Def* ret_arg = nullptr;
+        for (auto arg : continuation->args()) {
+            if (arg->order() == 0) {
+                if (auto val = emit_unsafe(arg))
+                    args.push_back(val);
+            } else {
+                assert(!ret_arg);
+                ret_arg = arg;
+            }
+        }
+
+        llvm::CallInst* call = nullptr;
+        if (auto callee = continuation->callee()->isa_continuation()) {
+            call = irbuilder.CreateCall(emit(callee), args);
+            if (callee->is_exported())
+                call->setCallingConv(kernel_calling_convention_);
+            else if (callee->cc() == CC::Device)
+                call->setCallingConv(device_calling_convention_);
+            else
+                call->setCallingConv(function_calling_convention_);
+        } else {
+            // must be a closure
+            auto closure = emit(callee);
+            args.push_back(irbuilder.CreateExtractValue(closure, 1));
+            call = irbuilder.CreateCall(irbuilder.CreateExtractValue(closure, 0), args);
+        }
+
+        // must be call + continuation --- call + return has been removed by codegen_prepare
+        auto succ = ret_arg->as_continuation();
+
+        size_t n = 0;
+        const Param* last_param = nullptr;
+        for (auto param : succ->params()) {
+            if (is_mem(param) || is_unit(param))
+                continue;
+            last_param = param;
+            n++;
+        }
+
+        if (n == 0) {
+            irbuilder.CreateBr(cont2bb(succ));
+        } else if (n == 1) {
+            irbuilder.CreateBr(cont2bb(succ));
+            emit_phi_arg(irbuilder, last_param, call);
+        } else {
+            Array<llvm::Value*> extracts(n);
+            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+                auto param = succ->param(i);
+                if (is_mem(param) || is_unit(param))
+                    continue;
+                extracts[j] = irbuilder.CreateExtractValue(call, unsigned(j));
+                j++;
+            }
+
+            irbuilder.CreateBr(cont2bb(succ));
+
+            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+                auto param = succ->param(i);
+                if (is_mem(param) || is_unit(param))
+                    continue;
+                emit_phi_arg(irbuilder, param, extracts[j]);
+                j++;
+            }
+        }
+    }*/
 
+    // new insert point is just before the terminator for all other instructions we have to add later on
+    // irbuilder.SetInsertPoint(bb->getTerminator());
 }
 
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index e7b309f01..903921be4 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -1,17 +1,32 @@
 #ifndef THORIN_SPIRV_H
 #define THORIN_SPIRV_H
 
+#include <thorin/analyses/schedule.h>
 #include "thorin/be/backends.h"
 
 namespace thorin::spirv {
 
+struct SpvSectionBuilder;
+struct SpvBasicBlockBuilder;
+struct SpvFnBuilder;
+struct SpvFileBuilder;
+struct SpvId { uint32_t id; };
+
 class CodeGen : public thorin::CodeGen {
 public:
     CodeGen(World&, Cont2Config&, bool debug);
 
     void emit(std::ostream& stream) override;
 protected:
+    SpvId convert(const Type*);
     void emit(const Scope& scope);
+    void emit_epilogue(Continuation*, SpvBasicBlockBuilder& bb);
+
+    SpvFileBuilder* builder_ = nullptr;
+    Continuation* entry_ = nullptr;
+    SpvFnBuilder* current_fn_ = nullptr;
+    Scheduler scheduler_;
+    TypeMap<SpvId> types_;
 };
 
 }

From e5021342a5ce4e44025fb2fb821639dbcb250299 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Feb 2021 14:36:16 +0100
Subject: [PATCH 023/342] remove hack for function codom type

---
 src/thorin/be/spirv/spirv.cpp | 149 ++++++++++------------------------
 src/thorin/be/spirv/spirv.h   |   2 +
 2 files changed, 43 insertions(+), 108 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 1e5bf883a..b08e71ebb 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -130,6 +130,15 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId declare_struct_type(std::vector<SpvId>& elements) {
+        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        for (auto arg : elements)
+            types_constants.ref_id(arg);
+        return id;
+    }
+
     SpvId define_function(SpvFnBuilder& fn_builder) {
         fn_defs.op(spv::Op::OpFunction, 5);
         fn_defs.ref_id(fn_builder.fn_ret_type);
@@ -246,7 +255,7 @@ void CodeGen::emit(std::ostream& out) {
 }
 
 SpvId CodeGen::convert(const Type* type) {
-    if (auto llvm_type = types_.lookup(type)) return *llvm_type;
+    if (auto spv_type = types_.lookup(type)) return *spv_type;
 
     assert(!type->isa<MemType>());
     SpvId spv_type;
@@ -263,19 +272,17 @@ SpvId CodeGen::convert(const Type* type) {
         case Node_PtrType: {
             auto ptr = type->as<PtrType>();
             assert(false && "TODO");
-            //llvm_type = llvm::PointerType::get(convert(ptr->pointee()), convert_addr_space(ptr->addr_space()));
             break;
         }
         case Node_IndefiniteArrayType: {
             assert(false && "TODO");
-            //llvm_type = llvm::ArrayType::get(convert(type->as<ArrayType>()->elem_type()), 0);
-            //return types_[type] = llvm_type;
+            auto array = type->as<IndefiniteArrayType>();
+            //return types_[type] = spv_type;
         }
         case Node_DefiniteArrayType: {
             assert(false && "TODO");
             auto array = type->as<DefiniteArrayType>();
-            //llvm_type = llvm::ArrayType::get(convert(array->elem_type()), array->dim());
-            //return types_[type] = llvm_type;
+            //return types_[type] = spv_type;
         }
 
         case Node_ClosureType:
@@ -296,7 +303,7 @@ SpvId CodeGen::convert(const Type* type) {
                     }
                     if (ret_types.size() == 0)      ret = std::make_unique<SpvId>(builder_->void_type);
                     else if (ret_types.size() == 1) ret = std::make_unique<SpvId>(ret_types.back());
-                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?"); //ret = llvm::StructType::get(context(), ret_types);
+                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
                 } else
                     ops.push_back(convert(op));
             }
@@ -307,72 +314,18 @@ SpvId CodeGen::convert(const Type* type) {
             }
 
             assert(false && "TODO: handle closure mess");
-            /* auto env_type = convert(Closure::environment_type(world()));
-            ops.push_back(env_type);
-            auto fn_type = llvm::FunctionType::get(ret, ops, false);
-            auto ptr_type = llvm::PointerType::get(fn_type, 0);
-            llvm_type = llvm::StructType::get(context(), { ptr_type, env_type });
-            return types_[type] = llvm_type;*/
         }
 
         case Node_StructType: {
             assert(false && "TODO");
-            /*auto struct_type = type->as<StructType>();
-            auto llvm_struct = llvm::StructType::create(context());
-
-            // important: memoize before recursing into element types to avoid endless recursion
-            assert(!types_.contains(struct_type) && "type already converted");
-            types_[struct_type] = llvm_struct;
-
-            Array<llvm::Type*> llvm_types(struct_type->num_ops());
-            for (size_t i = 0, e = llvm_types.size(); i != e; ++i)
-                llvm_types[i] = convert(struct_type->op(i));
-            llvm_struct->setBody(llvm_ref(llvm_types));
-            return llvm_struct;*/
         }
 
         case Node_TupleType: {
             assert(false && "TODO");
-            /*auto tuple = type->as<TupleType>();
-            Array<llvm::Type*> llvm_types(tuple->num_ops());
-            for (size_t i = 0, e = llvm_types.size(); i != e; ++i)
-                llvm_types[i] = convert(tuple->op(i));
-            llvm_type = llvm::StructType::get(context(), llvm_ref(llvm_types));
-            return types_[tuple] = llvm_type;*/
         }
 
         case Node_VariantType: {
             assert(false && "TODO");
-            /*assert(type->num_ops() > 0);
-
-            // Max alignment/size constraints respectively in the variant type alternatives dictate the ones to use for the overall type
-            size_t max_align = 0, max_size = 0;
-
-            auto layout = module().getDataLayout();
-            llvm::Type* max_align_type;
-            for (auto op : type->ops()) {
-                auto op_type = convert(op);
-                size_t size  = layout.getTypeAllocSize(op_type);
-                size_t align = layout.getABITypeAlignment(op_type);
-                // Favor types that are not empty
-                if (align > max_align || (align == max_align && max_align_type->isEmptyTy())) {
-                    max_align_type = op_type;
-                    max_align = align;
-                }
-                max_size = std::max(max_size, size);
-            }
-
-            auto rem_size = max_size - layout.getTypeAllocSize(max_align_type);
-            auto union_type = rem_size > 0
-                              ? llvm::StructType::get(context(), llvm::ArrayRef<llvm::Type*> { max_align_type, llvm::ArrayType::get(llvm::Type::getInt8Ty(context()), rem_size)})
-                              : llvm::StructType::get(context(), llvm::ArrayRef<llvm::Type*> { max_align_type });
-
-            auto tag_type = type->num_ops() < (1_u64 <<  8) ? llvm::Type::getInt8Ty (context()) :
-                            type->num_ops() < (1_u64 << 16) ? llvm::Type::getInt16Ty(context()) :
-                            type->num_ops() < (1_u64 << 32) ? llvm::Type::getInt32Ty(context()) :
-                            llvm::Type::getInt64Ty(context());
-
-            return llvm::StructType::get(context(), { union_type, tag_type });*/
         }
 
         default:
@@ -385,61 +338,29 @@ SpvId CodeGen::convert(const Type* type) {
 void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
-    //auto fct = llvm::cast<llvm::Function>(emit(entry_));
+
     auto fn = SpvFnBuilder { };
     fn.fn_type = convert(entry_->type());
-    fn.fn_ret_type = builder_->void_type; // TODO !!!
+    fn.fn_ret_type = get_codom_type(entry_);
 
     current_fn_ = &fn;
 
-    //cont2llvm_.clear();
     auto conts = schedule(scope);
 
-    // map all bb-like continuations to llvm bb stubs and handle params/phis
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
 
         auto [i, b] = fn.bbs.emplace(cont, std::make_unique<SpvBasicBlockBuilder>(*builder_));
+        assert(b);
+
         SpvBasicBlockBuilder& bb = *i->second;
         SpvId label = bb.label();
-        builder_->name(label, cont->name().c_str());
+
+        if (debug())
+            builder_->name(label, cont->name().c_str());
         fn.labels.emplace(cont, label);
-        //auto bb = llvm::BasicBlock::Create(context(), cont->name().c_str(), fct);
-        //auto [i, succ] = cont2llvm_.emplace(cont, std::pair(bb, std::make_unique<llvm::IRBuilder<>>(context())));
-        assert(b);
-        // auto& irbuilder = *i->second.second;
-        // irbuilder.SetInsertPoint(bb);
-
-        //if (debug())
-        //    irbuilder.SetCurrentDebugLocation(llvm::DebugLoc::get(cont->loc().begin.row, cont->loc().begin.row, discope));
-
-        /*if (entry_ == cont) {
-            auto arg = fct->arg_begin();
-            for (auto param : entry_->params()) {
-                if (is_mem(param) || is_unit(param)) {
-                    def2llvm_[param] = nullptr;
-                } else if (param->order() == 0) {
-                    auto argv = &*arg;
-                    auto value = map_param(fct, argv, param);
-                    if (value == argv) {
-                        arg->setName(param->unique_name()); // use param
-                        def2llvm_[param] = &*arg++;
-                    } else {
-                        def2llvm_[param] = value;           // use provided value
-                    }
-                }
-            }
-        } else {
-            for (auto param : cont->params()) {
-                if (is_mem(param) || is_unit(param)) {
-                    def2llvm_[param] = nullptr;
-                } else {
-                    // do not bother reserving anything (the 0 below) - it's a tiny optimization nobody cares about
-                    auto phi = irbuilder.CreatePHI(convert(param->type()), 0, param->name().c_str());
-                    def2llvm_[param] = phi;
-                }
-            }
-        }*/
+
+        // TODO prepare phis/params
     }
 
     Scheduler new_scheduler(scope);
@@ -454,14 +375,26 @@ void CodeGen::emit(const thorin::Scope& scope) {
     builder_->define_function(fn);
 }
 
-void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb) {
-    /*auto&& bb_ib = cont2llvm_[continuation];
-    auto bb = bb_ib->first;
-    auto& irbuilder = *bb_ib->second;*/
+SpvId CodeGen::get_codom_type(const Continuation* fn) {
+    auto ret_cont_type = fn->ret_param()->type();
+    std::vector<SpvId> types;
+    for (auto& op : ret_cont_type->ops()) {
+        if (op->isa<MemType>() || is_type_unit(op))
+            continue;
+        assert(op->order() == 0);
+        types.push_back(convert(op));
+    }
+    if (types.empty())
+        return builder_->void_type;
+    if (types.size() == 1)
+        return types[0];
+    return builder_->declare_struct_type(types);
+}
 
+void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb) {
     if (continuation->callee() == entry_->ret_param()) { // return
-        std::vector<SpvId> values;
         std::vector<SpvId> types;
+        std::vector<SpvId> values;
 
         for (auto arg : continuation->args()) {
             /*if (auto val = emit_unsafe(arg)) {
@@ -472,7 +405,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
 
         switch (values.size()) {
             case 0:  bb.return_void();      break;
-            //case 1:  irbuilder.CreateRet(values[0]); break;
+                //case 1:  irbuilder.CreateRet(values[0]); break;
             default:
                 assert(false && "TODO handle non-void returns");
                 /*llvm::Value* agg = llvm::UndefValue::get(llvm::StructType::get(context(), types));
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 903921be4..13de7ae12 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -22,6 +22,8 @@ class CodeGen : public thorin::CodeGen {
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, SpvBasicBlockBuilder& bb);
 
+    SpvId get_codom_type(const Continuation* fn);
+
     SpvFileBuilder* builder_ = nullptr;
     Continuation* entry_ = nullptr;
     SpvFnBuilder* current_fn_ = nullptr;

From 4a4dbb3cd8ac3556cef36efeedbf4546ebe33ea9 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Feb 2021 15:21:12 +0100
Subject: [PATCH 024/342] handle function arguments/return properly

---
 src/thorin/be/spirv/spirv.cpp | 95 ++++++++++++++++++++++++++++-------
 src/thorin/be/spirv/spirv.h   |  2 +
 2 files changed, 78 insertions(+), 19 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index b08e71ebb..2957892cc 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -68,9 +68,25 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
+        op(spv::Op::OpLabel, 3 + elements.size());
+        ref_id(aggregate_t);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        for (auto e : elements)
+            ref_id(e);
+        return id;
+    }
+
     void return_void() {
         op(spv::Op::OpReturn, 1);
     }
+
+    void return_value(SpvId value) {
+        op(spv::Op::OpReturnValue, 2);
+        ref_id(value);
+    }
+
 private:
     SpvId generate_fresh_id();
 };
@@ -80,6 +96,10 @@ struct SpvFnBuilder {
     SpvId fn_ret_type;
     ContinuationMap<std::unique_ptr<SpvBasicBlockBuilder>> bbs;
     ContinuationMap<SpvId> labels;
+    DefMap<SpvId> params;
+
+    // Contains OpFunctionParams
+    SpvSectionBuilder header;
 };
 
 struct SpvFileBuilder {
@@ -147,6 +167,9 @@ struct SpvFileBuilder {
         fn_defs.data_.push_back(spv::FunctionControlMaskNone);
         fn_defs.ref_id(fn_builder.fn_type);
 
+        for (auto w : fn_builder.header.data_)
+            fn_defs.data_.push_back(w);
+
         // TODO OpFunctionParameters
         for (auto& [cont, bb] : fn_builder.bbs) {
             for (auto w : bb->data_)
@@ -314,14 +337,25 @@ SpvId CodeGen::convert(const Type* type) {
             }
 
             assert(false && "TODO: handle closure mess");
+            break;
         }
 
         case Node_StructType: {
-            assert(false && "TODO");
+            std::vector<SpvId> types;
+            for (auto elem : type->as<StructType>()->ops())
+                types.push_back(convert(elem));
+            spv_type = builder_->declare_struct_type(types);
+            // TODO debug info
+            break;
         }
 
         case Node_TupleType: {
-            assert(false && "TODO");
+            std::vector<SpvId> types;
+            for (auto elem : type->as<TupleType>()->ops())
+                types.push_back(convert(elem));
+            spv_type = builder_->declare_struct_type(types);
+            // TODO debug info
+            break;
         }
 
         case Node_VariantType: {
@@ -360,7 +394,30 @@ void CodeGen::emit(const thorin::Scope& scope) {
             builder_->name(label, cont->name().c_str());
         fn.labels.emplace(cont, label);
 
-        // TODO prepare phis/params
+        if (entry_ == cont) {
+            int arg = 0;
+            for (auto param : entry_->params()) {
+                if (is_mem(param) || is_unit(param)) {
+                    // Nothing
+                } else if (param->order() == 0) {
+                    auto param_t = convert(param->type());
+                    fn.header.op(spv::Op::OpFunctionParameter, 3);
+                    auto id = builder_->generate_fresh_id();
+                    fn.header.ref_id(param_t);
+                    fn.header.ref_id(id);
+                    fn.params[param] = id;
+                }
+            }
+        } else {
+            for (auto param : cont->params()) {
+                if (is_mem(param) || is_unit(param)) {
+                    // Nothing
+                } else {
+                    // TODO OpPhi requires the full list of predecessors when emitting
+                    assert(false);
+                }
+            }
+        }
     }
 
     Scheduler new_scheduler(scope);
@@ -392,30 +449,26 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
 }
 
 void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb) {
-    if (continuation->callee() == entry_->ret_param()) { // return
-        std::vector<SpvId> types;
+    if (continuation->callee() == entry_->ret_param()) {
         std::vector<SpvId> values;
 
         for (auto arg : continuation->args()) {
-            /*if (auto val = emit_unsafe(arg)) {
-                values.emplace_back(val);
-                types.emplace_back(val->getType());
-            }*/
+            assert(arg->order() == 0);
+            if (is_mem(arg) || is_unit(arg))
+                continue;
+            auto val = emit(arg);
+            values.emplace_back(val);
         }
 
         switch (values.size()) {
             case 0:  bb.return_void();      break;
-                //case 1:  irbuilder.CreateRet(values[0]); break;
-            default:
-                assert(false && "TODO handle non-void returns");
-                /*llvm::Value* agg = llvm::UndefValue::get(llvm::StructType::get(context(), types));
-
-                for (size_t i = 0, e = values.size(); i != e; ++i)
-                    agg = irbuilder.CreateInsertValue(agg, values[i], { unsigned(i) });
-
-                irbuilder.CreateRet(agg);*/
+            case 1:  bb.return_value(values[0]); break;
+            default: bb.return_value(bb.composite(current_fn_->fn_ret_type, values));
         }
-    } /*else if (continuation->callee() == world().branch()) {
+    } else {
+        assert(false && "epilogue not implemented");
+    }
+    /*else if (continuation->callee() == world().branch()) {
         auto cond = emit(continuation->arg(0));
         auto tbb = cont2bb(continuation->arg(1)->as_continuation());
         auto fbb = cont2bb(continuation->arg(2)->as_continuation());
@@ -514,4 +567,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
     // irbuilder.SetInsertPoint(bb->getTerminator());
 }
 
+SpvId CodeGen::emit(const Def* def) {
+    return SpvId();
+}
+
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 13de7ae12..ff2288f9b 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -21,6 +21,7 @@ class CodeGen : public thorin::CodeGen {
     SpvId convert(const Type*);
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, SpvBasicBlockBuilder& bb);
+    SpvId emit(const Def* def);
 
     SpvId get_codom_type(const Continuation* fn);
 
@@ -29,6 +30,7 @@ class CodeGen : public thorin::CodeGen {
     SpvFnBuilder* current_fn_ = nullptr;
     Scheduler scheduler_;
     TypeMap<SpvId> types_;
+    DefMap<SpvId> defs_;
 };
 
 }

From cfb5a2b162d08f543e2d5c9435d16f921336d2d1 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Feb 2021 16:30:50 +0100
Subject: [PATCH 025/342] infrastructure for branching

---
 src/thorin/be/spirv/spirv.cpp | 89 ++++++++++++++++++++++++++---------
 src/thorin/be/spirv/spirv.h   |  2 +-
 2 files changed, 67 insertions(+), 24 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 2957892cc..81e31fa5d 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -61,6 +61,14 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     SpvFileBuilder& file_builder;
 
+    struct Phi {
+        SpvId type;
+        SpvId value;
+        std::vector<std::pair<SpvId, SpvId>> preds;
+    };
+    GIDMap<const Param*, Phi> phis;
+    DefMap<SpvId> args;
+
     SpvId label() {
         op(spv::Op::OpLabel, 2);
         auto id = generate_fresh_id();
@@ -78,6 +86,18 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    void branch(SpvId target) {
+        op(spv::Op::OpBranch, 2);
+        ref_id(target);
+    }
+
+    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
+        op(spv::Op::OpBranchConditional, 4);
+        ref_id(condition);
+        ref_id(true_target);
+        ref_id(false_target);
+    }
+
     void return_void() {
         op(spv::Op::OpReturn, 1);
     }
@@ -167,11 +187,20 @@ struct SpvFileBuilder {
         fn_defs.data_.push_back(spv::FunctionControlMaskNone);
         fn_defs.ref_id(fn_builder.fn_type);
 
+        // Includes stuff like OpFunctionParameters
         for (auto w : fn_builder.header.data_)
             fn_defs.data_.push_back(w);
 
-        // TODO OpFunctionParameters
         for (auto& [cont, bb] : fn_builder.bbs) {
+            for (auto [param, phi] : bb->phis) {
+                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi.preds.size());
+                fn_defs.ref_id(phi.type);
+                fn_defs.ref_id(phi.value);
+                for (auto& [pred_value, pred_label] : phi.preds) {
+                    fn_defs.ref_id(pred_value);
+                    fn_defs.ref_id(pred_label);
+                }
+            }
             for (auto w : bb->data_)
                 fn_defs.data_.push_back(w);
         }
@@ -395,7 +424,6 @@ void CodeGen::emit(const thorin::Scope& scope) {
         fn.labels.emplace(cont, label);
 
         if (entry_ == cont) {
-            int arg = 0;
             for (auto param : entry_->params()) {
                 if (is_mem(param) || is_unit(param)) {
                     // Nothing
@@ -413,8 +441,10 @@ void CodeGen::emit(const thorin::Scope& scope) {
                 if (is_mem(param) || is_unit(param)) {
                     // Nothing
                 } else {
-                    // TODO OpPhi requires the full list of predecessors when emitting
-                    assert(false);
+                    // OpPhi requires the full list of predecessors (values, labels)
+                    // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
+                    // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
+                    bb.phis[param] = { convert(param->type()), builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -429,6 +459,18 @@ void CodeGen::emit(const thorin::Scope& scope) {
         emit_epilogue(cont, *fn.bbs[cont]->get());
     }
 
+    // Wire up Phi nodes
+    for (auto& [cont, bb]: fn.bbs) {
+        for (auto [param, phi] : bb->phis) {
+            assert(param->order() == 0);
+            for (auto pred : cont->preds()) {
+                auto& pred_bb = *fn.bbs[pred];
+                auto arg = pred->arg(param->index());
+                phi.preds.emplace_back(*pred_bb->args[arg], *fn.labels[pred]);
+            }
+        }
+    }
+
     builder_->define_function(fn);
 }
 
@@ -456,7 +498,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
             assert(arg->order() == 0);
             if (is_mem(arg) || is_unit(arg))
                 continue;
-            auto val = emit(arg);
+            auto val = emit(arg, bb);
             values.emplace_back(val);
         }
 
@@ -465,15 +507,14 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
             case 1:  bb.return_value(values[0]); break;
             default: bb.return_value(bb.composite(current_fn_->fn_ret_type, values));
         }
-    } else {
-        assert(false && "epilogue not implemented");
     }
-    /*else if (continuation->callee() == world().branch()) {
-        auto cond = emit(continuation->arg(0));
-        auto tbb = cont2bb(continuation->arg(1)->as_continuation());
-        auto fbb = cont2bb(continuation->arg(2)->as_continuation());
-        irbuilder.CreateCondBr(cond, tbb, fbb);
-    } else if (continuation->callee()->isa<Continuation>() &&
+    else if (continuation->callee() == world().branch()) {
+        auto cond = emit(continuation->arg(0), bb);
+        bb.args[continuation->arg(0)] = cond;
+        auto tbb = *current_fn_->labels[continuation->arg(1)->as_continuation()];
+        auto fbb = *current_fn_->labels[continuation->arg(2)->as_continuation()];
+        bb.branch_conditional(cond, tbb, fbb);
+    } /*else if (continuation->callee()->isa<Continuation>() &&
                continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         auto val = emit(continuation->arg(0));
         auto otherwise_bb = cont2bb(continuation->arg(1)->as_continuation());
@@ -486,12 +527,14 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
         }
     } else if (continuation->callee()->isa<Bottom>()) {
         irbuilder.CreateUnreachable();
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
-        for (size_t i = 0, e = continuation->num_args(); i != e; ++i) {
-            if (auto val = emit_unsafe(continuation->arg(i))) emit_phi_arg(irbuilder, callee->param(i), val);
+    } */
+    else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+        for (auto& arg : continuation->args()) {
+            if (is_mem(arg) || is_unit(arg)) continue;
+            bb.args[arg] = emit(arg, bb);
         }
-        irbuilder.CreateBr(cont2bb(callee));
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+        bb.branch(*current_fn_->labels[callee]);
+    } /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         auto ret_continuation = emit_intrinsic(irbuilder, continuation);
         irbuilder.CreateBr(cont2bb(ret_continuation));
     } else { // function/closure call
@@ -562,13 +605,13 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
             }
         }
     }*/
-
-    // new insert point is just before the terminator for all other instructions we have to add later on
-    // irbuilder.SetInsertPoint(bb->getTerminator());
+    else {
+        assert(false && "epilogue not implemented for this");
+    }
 }
 
-SpvId CodeGen::emit(const Def* def) {
-    return SpvId();
+SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
+    assertf(false, "Incomplete emit(def) definition");
 }
 
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index ff2288f9b..a2535fdce 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -21,7 +21,7 @@ class CodeGen : public thorin::CodeGen {
     SpvId convert(const Type*);
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, SpvBasicBlockBuilder& bb);
-    SpvId emit(const Def* def);
+    SpvId emit(const Def* def, SpvBasicBlockBuilder& bb);
 
     SpvId get_codom_type(const Continuation* fn);
 

From 7eca24416b9c9e4a2f1e112edcdd1d61dcc407df Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Feb 2021 17:23:43 +0100
Subject: [PATCH 026/342] added codegen for most binops

---
 src/thorin/be/spirv/spirv.cpp | 154 +++++++++++++++++++++++++++++++++-
 1 file changed, 153 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 81e31fa5d..aec18d3f4 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -86,6 +86,16 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
+        op(op_, 5);
+        auto id = generate_fresh_id();
+        ref_id(result_type);
+        ref_id(id);
+        ref_id(lhs);
+        ref_id(rhs);
+        return id;
+    }
+
     void branch(SpvId target) {
         op(spv::Op::OpBranch, 2);
         ref_id(target);
@@ -179,6 +189,24 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId bool_constant(SpvId type, bool value) {
+        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    SpvId constant(SpvId type, std::vector<uint32_t>&& bit_pattern) {
+        types_constants.op(spv::Op::OpConstant, 4 + bit_pattern.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto arg : bit_pattern)
+            types_constants.data_.push_back(arg);
+        return id;
+    }
+
     SpvId define_function(SpvFnBuilder& fn_builder) {
         fn_defs.op(spv::Op::OpFunction, 5);
         fn_defs.ref_id(fn_builder.fn_ret_type);
@@ -290,7 +318,7 @@ SpvId SpvBasicBlockBuilder::generate_fresh_id() {
     return file_builder.generate_fresh_id();
 }
 
-CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
+CodeGen::CodeGen(thorin::World& world, Cont2Config&, bool debug)
     : thorin::CodeGen(world, debug)
 {}
 
@@ -611,6 +639,130 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
 }
 
 SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
+    if (auto bin = def->isa<BinOp>()) {
+        SpvId lhs = emit(bin->lhs(), bb);
+        SpvId rhs = emit(bin->rhs(), bb);
+        SpvId result_type = convert(def->type());
+
+        if (auto cmp = bin->isa<Cmp>()) {
+            auto type = cmp->lhs()->type();
+            if (is_type_s(type)) {
+                switch (cmp->cmp_tag()) {
+                    case Cmp_eq: return bb.binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb.binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb.binop(spv::Op::OpSGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb.binop(spv::Op::OpSGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb.binop(spv::Op::OpSLessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb.binop(spv::Op::OpSLessThanEqual    , result_type, lhs, rhs);
+                }
+            } else if (is_type_u(type)) {
+                switch (cmp->cmp_tag()) {
+                    case Cmp_eq: return bb.binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb.binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb.binop(spv::Op::OpUGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb.binop(spv::Op::OpUGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb.binop(spv::Op::OpULessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb.binop(spv::Op::OpULessThanEqual    , result_type, lhs, rhs);
+                }
+            } else if (is_type_f(type)) {
+                switch (cmp->cmp_tag()) {
+                    // TODO look into the NaN story
+                    case Cmp_eq: return bb.binop(spv::Op::OpFOrdEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb.binop(spv::Op::OpFOrdNotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb.binop(spv::Op::OpFOrdGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb.binop(spv::Op::OpFOrdGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb.binop(spv::Op::OpFOrdLessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb.binop(spv::Op::OpFOrdLessThanEqual    , result_type, lhs, rhs);
+                }
+            } else if (type->isa<PtrType>()) {
+                assertf(false, "Physical pointers are unsupported");
+            } else if(is_type_bool(type)) {
+                switch (cmp->cmp_tag()) {
+                    // TODO look into the NaN story
+                    case Cmp_eq: return bb.binop(spv::Op::OpLogicalEqual    , result_type, lhs, rhs);
+                    case Cmp_ne: return bb.binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
+                    default: THORIN_UNREACHABLE;
+                }
+                assertf(false, "TODO: should we emulate the other comparison ops ?");
+            }
+        }
+
+        if (auto arithop = bin->isa<ArithOp>()) {
+            auto type = arithop->type();
+
+            if (is_type_f(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_add: return bb.binop(spv::Op::OpFAdd, result_type, lhs, rhs);
+                    case ArithOp_sub: return bb.binop(spv::Op::OpFSub, result_type, lhs, rhs);
+                    case ArithOp_mul: return bb.binop(spv::Op::OpFMul, result_type, lhs, rhs);
+                    case ArithOp_div: return bb.binop(spv::Op::OpFDiv, result_type, lhs, rhs);
+                    case ArithOp_rem: return bb.binop(spv::Op::OpFRem, result_type, lhs, rhs);
+                    case ArithOp_and:
+                    case ArithOp_or:
+                    case ArithOp_xor:
+                    case ArithOp_shl:
+                    case ArithOp_shr: THORIN_UNREACHABLE;
+                }
+            }
+
+            if (is_type_s(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_add: return bb.binop(spv::Op::OpIAdd                 , result_type, lhs, rhs);
+                    case ArithOp_sub: return bb.binop(spv::Op::OpISub                 , result_type, lhs, rhs);
+                    case ArithOp_mul: return bb.binop(spv::Op::OpIMul                 , result_type, lhs, rhs);
+                    case ArithOp_div: return bb.binop(spv::Op::OpSDiv                 , result_type, lhs, rhs);
+                    case ArithOp_rem: return bb.binop(spv::Op::OpSRem                 , result_type, lhs, rhs);
+                    case ArithOp_and: return bb.binop(spv::Op::OpBitwiseAnd           , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb.binop(spv::Op::OpBitwiseOr            , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb.binop(spv::Op::OpBitwiseXor           , result_type, lhs, rhs);
+                    case ArithOp_shl: return bb.binop(spv::Op::OpShiftLeftLogical     , result_type, lhs, rhs);
+                    case ArithOp_shr: return bb.binop(spv::Op::OpShiftRightArithmetic , result_type, lhs, rhs);
+                }
+            } else if (is_type_u(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_add: return bb.binop(spv::Op::OpIAdd              , result_type, lhs, rhs);
+                    case ArithOp_sub: return bb.binop(spv::Op::OpISub              , result_type, lhs, rhs);
+                    case ArithOp_mul: return bb.binop(spv::Op::OpIMul              , result_type, lhs, rhs);
+                    case ArithOp_div: return bb.binop(spv::Op::OpUDiv              , result_type, lhs, rhs);
+                    case ArithOp_rem: return bb.binop(spv::Op::OpUMod              , result_type, lhs, rhs);
+                    case ArithOp_and: return bb.binop(spv::Op::OpBitwiseAnd        , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb.binop(spv::Op::OpBitwiseOr         , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb.binop(spv::Op::OpBitwiseXor        , result_type, lhs, rhs);
+                    case ArithOp_shl: return bb.binop(spv::Op::OpShiftLeftLogical  , result_type, lhs, rhs);
+                    case ArithOp_shr: return bb.binop(spv::Op::OpShiftRightLogical , result_type, lhs, rhs);
+                }
+            } else if(is_type_bool(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_and: return bb.binop(spv::Op::OpLogicalAnd      , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb.binop(spv::Op::OpLogicalOr       , result_type, lhs, rhs);
+                    // Note: there is no OpLogicalXor
+                    case ArithOp_xor: return bb.binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
+                    default: THORIN_UNREACHABLE;
+                }
+            }
+            THORIN_UNREACHABLE;
+        }
+    }
+    if (auto primlit = def->isa<PrimLit>()) {
+        Box box = primlit->value();
+        auto type = convert(def->type());
+        SpvId constant;
+        switch (primlit->primtype_tag()) {
+            case PrimType_bool:                     constant = bb.file_builder.bool_constant(type, box.get_bool()); break;
+            case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
+            case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
+            case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
+            case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
+            case PrimType_ps32: case PrimType_qs32: constant = bb.file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
+            case PrimType_pu32: case PrimType_qu32: constant = bb.file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
+            case PrimType_ps64: case PrimType_qs64: assertf(false, "not implemented yet");
+            case PrimType_pu64: case PrimType_qu64: assertf(false, "not implemented yet");
+            case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
+            case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
+            case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
+        }
+        return constant;
+    }
     assertf(false, "Incomplete emit(def) definition");
 }
 

From e0bd48b5cacde56dce3f650fd81821ab2c0cdf7e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Feb 2021 20:43:42 +0100
Subject: [PATCH 027/342] ordering, fixing phi nodes ...

---
 src/thorin/be/spirv/spirv.cpp | 189 +++++++++++++++++-----------------
 src/thorin/be/spirv/spirv.h   |   4 +-
 2 files changed, 96 insertions(+), 97 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index aec18d3f4..314afee95 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -66,15 +66,9 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         SpvId value;
         std::vector<std::pair<SpvId, SpvId>> preds;
     };
-    GIDMap<const Param*, Phi> phis;
+    std::unordered_map<const Param*, Phi> phis;
     DefMap<SpvId> args;
-
-    SpvId label() {
-        op(spv::Op::OpLabel, 2);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        return id;
-    }
+    SpvId label;
 
     SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
         op(spv::Op::OpLabel, 3 + elements.size());
@@ -124,7 +118,8 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 struct SpvFnBuilder {
     SpvId fn_type;
     SpvId fn_ret_type;
-    ContinuationMap<std::unique_ptr<SpvBasicBlockBuilder>> bbs;
+    std::vector<SpvBasicBlockBuilder> bbs;
+    std::unordered_map<Continuation*, SpvBasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
 
@@ -198,7 +193,7 @@ struct SpvFileBuilder {
     }
 
     SpvId constant(SpvId type, std::vector<uint32_t>&& bit_pattern) {
-        types_constants.op(spv::Op::OpConstant, 4 + bit_pattern.size());
+        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(type);
         types_constants.ref_id(id);
@@ -219,8 +214,11 @@ struct SpvFileBuilder {
         for (auto w : fn_builder.header.data_)
             fn_defs.data_.push_back(w);
 
-        for (auto& [cont, bb] : fn_builder.bbs) {
-            for (auto [param, phi] : bb->phis) {
+        for (auto& bb : fn_builder.bbs) {
+            fn_defs.op(spv::Op::OpLabel, 2);
+            fn_defs.ref_id(bb.label);
+
+            for (auto [param, phi] : bb.phis) {
                 fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi.preds.size());
                 fn_defs.ref_id(phi.type);
                 fn_defs.ref_id(phi.value);
@@ -229,7 +227,8 @@ struct SpvFileBuilder {
                     fn_defs.ref_id(pred_label);
                 }
             }
-            for (auto w : bb->data_)
+
+            for (auto w : bb.data_)
                 fn_defs.data_.push_back(w);
         }
 
@@ -438,18 +437,20 @@ void CodeGen::emit(const thorin::Scope& scope) {
 
     auto conts = schedule(scope);
 
+    fn.bbs.reserve(conts.size());
+
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
 
-        auto [i, b] = fn.bbs.emplace(cont, std::make_unique<SpvBasicBlockBuilder>(*builder_));
+        SpvBasicBlockBuilder* bb = &fn.bbs.emplace_back(*builder_);
+        auto [i, b] = fn.bbs_map.emplace(cont, bb);
         assert(b);
 
-        SpvBasicBlockBuilder& bb = *i->second;
-        SpvId label = bb.label();
+        bb->label = builder_->generate_fresh_id();
 
         if (debug())
-            builder_->name(label, cont->name().c_str());
-        fn.labels.emplace(cont, label);
+            builder_->name(bb->label, cont->name().c_str());
+        fn.labels.emplace(cont, bb->label);
 
         if (entry_ == cont) {
             for (auto param : entry_->params()) {
@@ -472,7 +473,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    bb.phis[param] = { convert(param->type()), builder_->generate_fresh_id(), {} };
+                    bb->phis[param] = { convert(param->type()), builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -484,21 +485,9 @@ void CodeGen::emit(const thorin::Scope& scope) {
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
         assert(cont == entry_ || cont->is_basicblock());
-        emit_epilogue(cont, *fn.bbs[cont]->get());
+        emit_epilogue(cont, fn.bbs_map[cont]);
     }
-
-    // Wire up Phi nodes
-    for (auto& [cont, bb]: fn.bbs) {
-        for (auto [param, phi] : bb->phis) {
-            assert(param->order() == 0);
-            for (auto pred : cont->preds()) {
-                auto& pred_bb = *fn.bbs[pred];
-                auto arg = pred->arg(param->index());
-                phi.preds.emplace_back(*pred_bb->args[arg], *fn.labels[pred]);
-            }
-        }
-    }
-
+    
     builder_->define_function(fn);
 }
 
@@ -518,7 +507,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
     return builder_->declare_struct_type(types);
 }
 
-void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb) {
+void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder* bb) {
     if (continuation->callee() == entry_->ret_param()) {
         std::vector<SpvId> values;
 
@@ -531,17 +520,17 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
         }
 
         switch (values.size()) {
-            case 0:  bb.return_void();      break;
-            case 1:  bb.return_value(values[0]); break;
-            default: bb.return_value(bb.composite(current_fn_->fn_ret_type, values));
+            case 0:  bb->return_void();      break;
+            case 1:  bb->return_value(values[0]); break;
+            default: bb->return_value(bb->composite(current_fn_->fn_ret_type, values));
         }
     }
     else if (continuation->callee() == world().branch()) {
         auto cond = emit(continuation->arg(0), bb);
-        bb.args[continuation->arg(0)] = cond;
+        bb->args[continuation->arg(0)] = cond;
         auto tbb = *current_fn_->labels[continuation->arg(1)->as_continuation()];
         auto fbb = *current_fn_->labels[continuation->arg(2)->as_continuation()];
-        bb.branch_conditional(cond, tbb, fbb);
+        bb->branch_conditional(cond, tbb, fbb);
     } /*else if (continuation->callee()->isa<Continuation>() &&
                continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         auto val = emit(continuation->arg(0));
@@ -557,11 +546,16 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
         irbuilder.CreateUnreachable();
     } */
     else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+        int index = -1;
         for (auto& arg : continuation->args()) {
+            index++;
             if (is_mem(arg) || is_unit(arg)) continue;
-            bb.args[arg] = emit(arg, bb);
+            bb->args[arg] = emit(arg, bb);
+            auto* param = callee->param(index);
+            auto& phi = current_fn_->bbs_map[callee]->phis[param];
+            phi.preds.emplace_back(*bb->args[arg], *current_fn_->labels[continuation]);
         }
-        bb.branch(*current_fn_->labels[callee]);
+        bb->branch(*current_fn_->labels[callee]);
     } /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         auto ret_continuation = emit_intrinsic(irbuilder, continuation);
         irbuilder.CreateBr(cont2bb(ret_continuation));
@@ -638,7 +632,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder& bb
     }
 }
 
-SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
+SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
         SpvId rhs = emit(bin->rhs(), bb);
@@ -648,39 +642,39 @@ SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
             auto type = cmp->lhs()->type();
             if (is_type_s(type)) {
                 switch (cmp->cmp_tag()) {
-                    case Cmp_eq: return bb.binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
-                    case Cmp_ne: return bb.binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
-                    case Cmp_gt: return bb.binop(spv::Op::OpSGreaterThan      , result_type, lhs, rhs);
-                    case Cmp_ge: return bb.binop(spv::Op::OpSGreaterThanEqual , result_type, lhs, rhs);
-                    case Cmp_lt: return bb.binop(spv::Op::OpSLessThan         , result_type, lhs, rhs);
-                    case Cmp_le: return bb.binop(spv::Op::OpSLessThanEqual    , result_type, lhs, rhs);
+                    case Cmp_eq: return bb->binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb->binop(spv::Op::OpSGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb->binop(spv::Op::OpSGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb->binop(spv::Op::OpSLessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb->binop(spv::Op::OpSLessThanEqual    , result_type, lhs, rhs);
                 }
             } else if (is_type_u(type)) {
                 switch (cmp->cmp_tag()) {
-                    case Cmp_eq: return bb.binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
-                    case Cmp_ne: return bb.binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
-                    case Cmp_gt: return bb.binop(spv::Op::OpUGreaterThan      , result_type, lhs, rhs);
-                    case Cmp_ge: return bb.binop(spv::Op::OpUGreaterThanEqual , result_type, lhs, rhs);
-                    case Cmp_lt: return bb.binop(spv::Op::OpULessThan         , result_type, lhs, rhs);
-                    case Cmp_le: return bb.binop(spv::Op::OpULessThanEqual    , result_type, lhs, rhs);
+                    case Cmp_eq: return bb->binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb->binop(spv::Op::OpUGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb->binop(spv::Op::OpUGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb->binop(spv::Op::OpULessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb->binop(spv::Op::OpULessThanEqual    , result_type, lhs, rhs);
                 }
             } else if (is_type_f(type)) {
                 switch (cmp->cmp_tag()) {
                     // TODO look into the NaN story
-                    case Cmp_eq: return bb.binop(spv::Op::OpFOrdEqual            , result_type, lhs, rhs);
-                    case Cmp_ne: return bb.binop(spv::Op::OpFOrdNotEqual         , result_type, lhs, rhs);
-                    case Cmp_gt: return bb.binop(spv::Op::OpFOrdGreaterThan      , result_type, lhs, rhs);
-                    case Cmp_ge: return bb.binop(spv::Op::OpFOrdGreaterThanEqual , result_type, lhs, rhs);
-                    case Cmp_lt: return bb.binop(spv::Op::OpFOrdLessThan         , result_type, lhs, rhs);
-                    case Cmp_le: return bb.binop(spv::Op::OpFOrdLessThanEqual    , result_type, lhs, rhs);
+                    case Cmp_eq: return bb->binop(spv::Op::OpFOrdEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpFOrdNotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb->binop(spv::Op::OpFOrdGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb->binop(spv::Op::OpFOrdGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb->binop(spv::Op::OpFOrdLessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb->binop(spv::Op::OpFOrdLessThanEqual    , result_type, lhs, rhs);
                 }
             } else if (type->isa<PtrType>()) {
                 assertf(false, "Physical pointers are unsupported");
             } else if(is_type_bool(type)) {
                 switch (cmp->cmp_tag()) {
                     // TODO look into the NaN story
-                    case Cmp_eq: return bb.binop(spv::Op::OpLogicalEqual    , result_type, lhs, rhs);
-                    case Cmp_ne: return bb.binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
+                    case Cmp_eq: return bb->binop(spv::Op::OpLogicalEqual    , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
                     default: THORIN_UNREACHABLE;
                 }
                 assertf(false, "TODO: should we emulate the other comparison ops ?");
@@ -692,11 +686,11 @@ SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
 
             if (is_type_f(type)) {
                 switch (arithop->arithop_tag()) {
-                    case ArithOp_add: return bb.binop(spv::Op::OpFAdd, result_type, lhs, rhs);
-                    case ArithOp_sub: return bb.binop(spv::Op::OpFSub, result_type, lhs, rhs);
-                    case ArithOp_mul: return bb.binop(spv::Op::OpFMul, result_type, lhs, rhs);
-                    case ArithOp_div: return bb.binop(spv::Op::OpFDiv, result_type, lhs, rhs);
-                    case ArithOp_rem: return bb.binop(spv::Op::OpFRem, result_type, lhs, rhs);
+                    case ArithOp_add: return bb->binop(spv::Op::OpFAdd, result_type, lhs, rhs);
+                    case ArithOp_sub: return bb->binop(spv::Op::OpFSub, result_type, lhs, rhs);
+                    case ArithOp_mul: return bb->binop(spv::Op::OpFMul, result_type, lhs, rhs);
+                    case ArithOp_div: return bb->binop(spv::Op::OpFDiv, result_type, lhs, rhs);
+                    case ArithOp_rem: return bb->binop(spv::Op::OpFRem, result_type, lhs, rhs);
                     case ArithOp_and:
                     case ArithOp_or:
                     case ArithOp_xor:
@@ -707,54 +701,53 @@ SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
 
             if (is_type_s(type)) {
                 switch (arithop->arithop_tag()) {
-                    case ArithOp_add: return bb.binop(spv::Op::OpIAdd                 , result_type, lhs, rhs);
-                    case ArithOp_sub: return bb.binop(spv::Op::OpISub                 , result_type, lhs, rhs);
-                    case ArithOp_mul: return bb.binop(spv::Op::OpIMul                 , result_type, lhs, rhs);
-                    case ArithOp_div: return bb.binop(spv::Op::OpSDiv                 , result_type, lhs, rhs);
-                    case ArithOp_rem: return bb.binop(spv::Op::OpSRem                 , result_type, lhs, rhs);
-                    case ArithOp_and: return bb.binop(spv::Op::OpBitwiseAnd           , result_type, lhs, rhs);
-                    case ArithOp_or:  return bb.binop(spv::Op::OpBitwiseOr            , result_type, lhs, rhs);
-                    case ArithOp_xor: return bb.binop(spv::Op::OpBitwiseXor           , result_type, lhs, rhs);
-                    case ArithOp_shl: return bb.binop(spv::Op::OpShiftLeftLogical     , result_type, lhs, rhs);
-                    case ArithOp_shr: return bb.binop(spv::Op::OpShiftRightArithmetic , result_type, lhs, rhs);
+                    case ArithOp_add: return bb->binop(spv::Op::OpIAdd                 , result_type, lhs, rhs);
+                    case ArithOp_sub: return bb->binop(spv::Op::OpISub                 , result_type, lhs, rhs);
+                    case ArithOp_mul: return bb->binop(spv::Op::OpIMul                 , result_type, lhs, rhs);
+                    case ArithOp_div: return bb->binop(spv::Op::OpSDiv                 , result_type, lhs, rhs);
+                    case ArithOp_rem: return bb->binop(spv::Op::OpSRem                 , result_type, lhs, rhs);
+                    case ArithOp_and: return bb->binop(spv::Op::OpBitwiseAnd           , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb->binop(spv::Op::OpBitwiseOr            , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb->binop(spv::Op::OpBitwiseXor           , result_type, lhs, rhs);
+                    case ArithOp_shl: return bb->binop(spv::Op::OpShiftLeftLogical     , result_type, lhs, rhs);
+                    case ArithOp_shr: return bb->binop(spv::Op::OpShiftRightArithmetic , result_type, lhs, rhs);
                 }
             } else if (is_type_u(type)) {
                 switch (arithop->arithop_tag()) {
-                    case ArithOp_add: return bb.binop(spv::Op::OpIAdd              , result_type, lhs, rhs);
-                    case ArithOp_sub: return bb.binop(spv::Op::OpISub              , result_type, lhs, rhs);
-                    case ArithOp_mul: return bb.binop(spv::Op::OpIMul              , result_type, lhs, rhs);
-                    case ArithOp_div: return bb.binop(spv::Op::OpUDiv              , result_type, lhs, rhs);
-                    case ArithOp_rem: return bb.binop(spv::Op::OpUMod              , result_type, lhs, rhs);
-                    case ArithOp_and: return bb.binop(spv::Op::OpBitwiseAnd        , result_type, lhs, rhs);
-                    case ArithOp_or:  return bb.binop(spv::Op::OpBitwiseOr         , result_type, lhs, rhs);
-                    case ArithOp_xor: return bb.binop(spv::Op::OpBitwiseXor        , result_type, lhs, rhs);
-                    case ArithOp_shl: return bb.binop(spv::Op::OpShiftLeftLogical  , result_type, lhs, rhs);
-                    case ArithOp_shr: return bb.binop(spv::Op::OpShiftRightLogical , result_type, lhs, rhs);
+                    case ArithOp_add: return bb->binop(spv::Op::OpIAdd              , result_type, lhs, rhs);
+                    case ArithOp_sub: return bb->binop(spv::Op::OpISub              , result_type, lhs, rhs);
+                    case ArithOp_mul: return bb->binop(spv::Op::OpIMul              , result_type, lhs, rhs);
+                    case ArithOp_div: return bb->binop(spv::Op::OpUDiv              , result_type, lhs, rhs);
+                    case ArithOp_rem: return bb->binop(spv::Op::OpUMod              , result_type, lhs, rhs);
+                    case ArithOp_and: return bb->binop(spv::Op::OpBitwiseAnd        , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb->binop(spv::Op::OpBitwiseOr         , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb->binop(spv::Op::OpBitwiseXor        , result_type, lhs, rhs);
+                    case ArithOp_shl: return bb->binop(spv::Op::OpShiftLeftLogical  , result_type, lhs, rhs);
+                    case ArithOp_shr: return bb->binop(spv::Op::OpShiftRightLogical , result_type, lhs, rhs);
                 }
             } else if(is_type_bool(type)) {
                 switch (arithop->arithop_tag()) {
-                    case ArithOp_and: return bb.binop(spv::Op::OpLogicalAnd      , result_type, lhs, rhs);
-                    case ArithOp_or:  return bb.binop(spv::Op::OpLogicalOr       , result_type, lhs, rhs);
+                    case ArithOp_and: return bb->binop(spv::Op::OpLogicalAnd      , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb->binop(spv::Op::OpLogicalOr       , result_type, lhs, rhs);
                     // Note: there is no OpLogicalXor
-                    case ArithOp_xor: return bb.binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb->binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
                     default: THORIN_UNREACHABLE;
                 }
             }
             THORIN_UNREACHABLE;
         }
-    }
-    if (auto primlit = def->isa<PrimLit>()) {
+    } else if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
         auto type = convert(def->type());
         SpvId constant;
         switch (primlit->primtype_tag()) {
-            case PrimType_bool:                     constant = bb.file_builder.bool_constant(type, box.get_bool()); break;
+            case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
             case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
             case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
             case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
             case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
-            case PrimType_ps32: case PrimType_qs32: constant = bb.file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
-            case PrimType_pu32: case PrimType_qu32: constant = bb.file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
+            case PrimType_ps32: case PrimType_qs32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
+            case PrimType_pu32: case PrimType_qu32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
             case PrimType_ps64: case PrimType_qs64: assertf(false, "not implemented yet");
             case PrimType_pu64: case PrimType_qu64: assertf(false, "not implemented yet");
             case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
@@ -762,6 +755,12 @@ SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder& bb) {
             case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
         }
         return constant;
+    } else if(auto param = def->isa<Param>()) {
+        if (auto param_id = current_fn_->params.lookup(param)) {
+            return *param_id;
+        } else {
+            return (*current_fn_->bbs_map[param->continuation()]).phis[param].value;
+        }
     }
     assertf(false, "Incomplete emit(def) definition");
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index a2535fdce..8cbc42072 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -20,8 +20,8 @@ class CodeGen : public thorin::CodeGen {
 protected:
     SpvId convert(const Type*);
     void emit(const Scope& scope);
-    void emit_epilogue(Continuation*, SpvBasicBlockBuilder& bb);
-    SpvId emit(const Def* def, SpvBasicBlockBuilder& bb);
+    void emit_epilogue(Continuation*, SpvBasicBlockBuilder* bb);
+    SpvId emit(const Def* def, SpvBasicBlockBuilder* bb);
 
     SpvId get_codom_type(const Continuation* fn);
 

From 4d70f1384f4430c21e383a0c1b7699f98c12dfdf Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 18 Feb 2021 10:49:22 +0100
Subject: [PATCH 028/342] renamed vkcompute to spirv

---
 src/thorin/be/backends.cpp  | 14 +++++++-------
 src/thorin/be/backends.h    |  2 +-
 src/thorin/be/llvm/llvm.cpp |  2 +-
 src/thorin/continuation.cpp |  2 +-
 src/thorin/continuation.h   |  2 +-
 5 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/src/thorin/be/backends.cpp b/src/thorin/be/backends.cpp
index 2808c74e3..ed0de4c5d 100644
--- a/src/thorin/be/backends.cpp
+++ b/src/thorin/be/backends.cpp
@@ -86,12 +86,12 @@ Backends::Backends(World& world, int opt, bool debug)
         Continuation* imported = nullptr;
 
         static const auto backend_intrinsics = std::array {
-            std::pair { CUDA,      Intrinsic::CUDA      },
-            std::pair { NVVM,      Intrinsic::NVVM      },
-            std::pair { OpenCL,    Intrinsic::OpenCL    },
-            std::pair { AMDGPU,    Intrinsic::AMDGPU    },
-            std::pair { HLS,       Intrinsic::HLS       },
-            std::pair { VkCompute, Intrinsic::VkCompute }
+            std::pair { CUDA,      Intrinsic::CUDA   },
+            std::pair { NVVM,      Intrinsic::NVVM   },
+            std::pair { OpenCL,    Intrinsic::OpenCL },
+            std::pair { AMDGPU,    Intrinsic::AMDGPU },
+            std::pair { HLS,       Intrinsic::HLS    },
+            std::pair { SpirV    , Intrinsic::SpirV  }
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
@@ -186,7 +186,7 @@ Backends::Backends(World& world, int opt, bool debug)
     // TODO: maybe use the C backend as a fallback when LLVM is not present for host codegen ?
 #endif
 #if THORIN_ENABLE_SPIRV
-    if (!importers_[VkCompute].world().empty()) device_cgs[VkCompute] = std::make_unique<spirv::CodeGen>(importers_[VkCompute].world(), kernel_config, debug);
+    if (!importers_[SpirV].world().empty()) device_cgs[SpirV] = std::make_unique<spirv::CodeGen>(importers_[SpirV].world(), kernel_config, debug);
 #endif
     for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OPENCL }, std::pair { HLS, c::Lang::HLS } })
         if (!importers_[backend].world().empty()) device_cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug);
diff --git a/src/thorin/be/backends.h b/src/thorin/be/backends.h
index 2d366b23a..c369606e5 100644
--- a/src/thorin/be/backends.h
+++ b/src/thorin/be/backends.h
@@ -43,7 +43,7 @@ struct Backends {
 
     std::unique_ptr<CodeGen> cpu_cg;
 
-    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, VkCompute, BackendCount };
+    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, SpirV, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> device_cgs;
 private:
     std::vector<Importer> importers_;
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index e11e5d43a..c1a365b19 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1115,7 +1115,7 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
         case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
         case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
         case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
-        case Intrinsic::VkCompute:   return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv",    continuation); // TODO have a real runtime component
+        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv", continuation); // TODO have a real runtime component
         case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
         case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
         case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index b354e8647..75f298991 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -167,7 +167,7 @@ void Continuation::set_intrinsic() {
     else if (name() == "nvvm")                 attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")               attributes().intrinsic = Intrinsic::OpenCL;
     else if (name() == "amdgpu")               attributes().intrinsic = Intrinsic::AMDGPU;
-    else if (name() == "vk_compute")           attributes().intrinsic = Intrinsic::VkCompute;
+    else if (name() == "spirv")                attributes().intrinsic = Intrinsic::SpirV;
     else if (name() == "hls")                  attributes().intrinsic = Intrinsic::HLS;
     else if (name() == "parallel")             attributes().intrinsic = Intrinsic::Parallel;
     else if (name() == "fibers")               attributes().intrinsic = Intrinsic::Fibers;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 3f6697d1b..bff19a11e 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -62,7 +62,7 @@ enum class Intrinsic : uint8_t {
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
     AMDGPU,                     ///< Internal AMDGPU-Backend.
-    VkCompute,                  ///< Internal Vulkan-Compute-Shader-Backend.
+    SpirV,                      ///< Internal Vulkan-Compute-Shader-Backend.
     HLS,                        ///< Internal HLS-Backend.
     Parallel,                   ///< Internal Parallel-CPU-Backend.
     Fibers,                     ///< Internal Parallel-CPU-Backend using resumable fibers.

From 2b166ca4eadddf586f633c0d949fcb11b45f12ba Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 10 Mar 2021 15:57:40 +0100
Subject: [PATCH 029/342] fix whitespace

---
 src/thorin/be/llvm/llvm.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index c1a365b19..0676f1bcd 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1115,7 +1115,7 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
         case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
         case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
         case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
-        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv", continuation); // TODO have a real runtime component
+        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv",    continuation); // TODO have a real runtime component
         case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
         case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
         case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);

From 282ab46df98dc1cb0f005f994329d8065a508432 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 12 Mar 2021 11:38:19 +0100
Subject: [PATCH 030/342] move builder stuff out

---
 src/thorin/CMakeLists.txt               |   3 +-
 src/thorin/be/spirv/spirv.cpp           | 333 ++----------------------
 src/thorin/be/spirv/spirv.h             |  36 ++-
 src/thorin/be/spirv/spirv_builder.hpp   | 325 +++++++++++++++++++++++
 src/thorin/be/spirv/spirv_transform.cpp |  14 +
 5 files changed, 383 insertions(+), 328 deletions(-)
 create mode 100644 src/thorin/be/spirv/spirv_builder.hpp
 create mode 100644 src/thorin/be/spirv/spirv_transform.cpp

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index c1c13a5ec..e38158fac 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -84,7 +84,8 @@ set(THORIN_SOURCES
     util/utility.h
     be/backends.cpp
     be/backends.h
-    )
+    be/spirv/spirv_transform.cpp
+)
 
 if(LLVM_FOUND)
     list(APPEND THORIN_SOURCES
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 314afee95..6a3de9306 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -1,332 +1,25 @@
 #include "thorin/be/spirv/spirv.h"
 #include "thorin/analyses/scope.h"
+#include "thorin/analyses/schedule.h"
 
-#include <spirv/unified1/spirv.hpp>
+#include "thorin/transform/cleanup_world.h"
 
 #include <iostream>
-#include <thorin/analyses/schedule.h>
-
-int div_roundup(int a, int b) {
-    if (a % b == 0)
-        return a / b;
-    else
-        return (a / b) + 1;
-}
 
 namespace thorin::spirv {
 
-struct SpvSectionBuilder {
-    std::vector<uint32_t> data_;
-
-private:
-    void output_word(uint32_t word) {
-        data_.push_back(word);
-    }
-public:
-    void op(spv::Op op, int ops_size) {
-        uint32_t lower = op & 0xFFFFu;
-        uint32_t upper = (ops_size << 16) & 0xFFFF0000u;
-        output_word(lower | upper);
-    }
-
-    void ref_id(SpvId id) {
-        assert(id.id != 0);
-        output_word(id.id);
-    }
-
-    void literal_name(std::string_view str) {
-        int i = 0;
-        uint32_t cword = 0;
-        for (char c : str) {
-            cword = cword | (c & 0xFF) << (i * 8);
-            i++;
-            if (i == 4) {
-                output_word(cword);
-                cword = 0;
-                i = 0;
-            }
-        }
-        output_word(cword);
-    }
-
-    void literal_int(uint32_t i) {
-        output_word(i);
-    }
-};
-
-struct SpvBasicBlockBuilder : public SpvSectionBuilder {
-    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
-        : file_builder(file_builder)
-    {}
-
-    SpvFileBuilder& file_builder;
-
-    struct Phi {
-        SpvId type;
-        SpvId value;
-        std::vector<std::pair<SpvId, SpvId>> preds;
-    };
-    std::unordered_map<const Param*, Phi> phis;
-    DefMap<SpvId> args;
-    SpvId label;
-
-    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
-        op(spv::Op::OpLabel, 3 + elements.size());
-        ref_id(aggregate_t);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        for (auto e : elements)
-            ref_id(e);
-        return id;
-    }
-
-    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
-        op(op_, 5);
-        auto id = generate_fresh_id();
-        ref_id(result_type);
-        ref_id(id);
-        ref_id(lhs);
-        ref_id(rhs);
-        return id;
-    }
-
-    void branch(SpvId target) {
-        op(spv::Op::OpBranch, 2);
-        ref_id(target);
-    }
-
-    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
-        op(spv::Op::OpBranchConditional, 4);
-        ref_id(condition);
-        ref_id(true_target);
-        ref_id(false_target);
-    }
-
-    void return_void() {
-        op(spv::Op::OpReturn, 1);
-    }
-
-    void return_value(SpvId value) {
-        op(spv::Op::OpReturnValue, 2);
-        ref_id(value);
-    }
-
-private:
-    SpvId generate_fresh_id();
-};
-
-struct SpvFnBuilder {
-    SpvId fn_type;
-    SpvId fn_ret_type;
-    std::vector<SpvBasicBlockBuilder> bbs;
-    std::unordered_map<Continuation*, SpvBasicBlockBuilder*> bbs_map;
-    ContinuationMap<SpvId> labels;
-    DefMap<SpvId> params;
-
-    // Contains OpFunctionParams
-    SpvSectionBuilder header;
-};
-
-struct SpvFileBuilder {
-    SpvFileBuilder()
-        : void_type(declare_void_type())
-    {}
-
-    SpvId generate_fresh_id() { return {bound++ }; }
-
-    void name(SpvId id, std::string_view str) {
-        assert(id.id < bound);
-        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
-        debug_names.ref_id(id);
-        debug_names.literal_name(str);
-    }
-
-    SpvId declare_bool_type() {
-        types_constants.op(spv::Op::OpTypeBool, 2);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    SpvId declare_int_type(int width, bool signed_) {
-        types_constants.op(spv::Op::OpTypeInt, 4);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(width);
-        types_constants.literal_int(signed_ ? 1 : 0);
-        return id;
-    }
-
-    SpvId declare_float_type(int width) {
-        types_constants.op(spv::Op::OpTypeFloat, 3);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(width);
-        return id;
-    }
-
-    SpvId declare_fn_type(std::vector<SpvId>& dom, SpvId codom) {
-        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(codom);
-        for (auto arg : dom)
-            types_constants.ref_id(arg);
-        return id;
-    }
-
-    SpvId declare_struct_type(std::vector<SpvId>& elements) {
-        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        for (auto arg : elements)
-            types_constants.ref_id(arg);
-        return id;
-    }
-
-    SpvId bool_constant(SpvId type, bool value) {
-        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    SpvId constant(SpvId type, std::vector<uint32_t>&& bit_pattern) {
-        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        for (auto arg : bit_pattern)
-            types_constants.data_.push_back(arg);
-        return id;
-    }
-
-    SpvId define_function(SpvFnBuilder& fn_builder) {
-        fn_defs.op(spv::Op::OpFunction, 5);
-        fn_defs.ref_id(fn_builder.fn_ret_type);
-        auto id = generate_fresh_id();
-        fn_defs.ref_id(id);
-        fn_defs.data_.push_back(spv::FunctionControlMaskNone);
-        fn_defs.ref_id(fn_builder.fn_type);
-
-        // Includes stuff like OpFunctionParameters
-        for (auto w : fn_builder.header.data_)
-            fn_defs.data_.push_back(w);
-
-        for (auto& bb : fn_builder.bbs) {
-            fn_defs.op(spv::Op::OpLabel, 2);
-            fn_defs.ref_id(bb.label);
-
-            for (auto [param, phi] : bb.phis) {
-                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi.preds.size());
-                fn_defs.ref_id(phi.type);
-                fn_defs.ref_id(phi.value);
-                for (auto& [pred_value, pred_label] : phi.preds) {
-                    fn_defs.ref_id(pred_value);
-                    fn_defs.ref_id(pred_label);
-                }
-            }
-
-            for (auto w : bb.data_)
-                fn_defs.data_.push_back(w);
-        }
-
-        fn_defs.op(spv::Op::OpFunctionEnd, 1);
-        return id;
-    }
-
-    void capability(spv::Capability cap) {
-        capabilities.op(spv::Op::OpCapability, 2);
-        capabilities.data_.push_back(cap);
-    }
-
-    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
-    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
-
-private:
-    std::ostream* output_ = nullptr;
-    uint32_t bound = 1;
-
-    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
-    SpvSectionBuilder capabilities;
-    SpvSectionBuilder extensions;
-    SpvSectionBuilder ext_inst_import;
-    SpvSectionBuilder entry_points;
-    SpvSectionBuilder execution_modes;
-    SpvSectionBuilder debug_string_source;
-    SpvSectionBuilder debug_names;
-    SpvSectionBuilder debug_module_processed;
-    SpvSectionBuilder annotations;
-    SpvSectionBuilder types_constants;
-    SpvSectionBuilder fn_decls;
-    SpvSectionBuilder fn_defs;
-
-    SpvId declare_void_type() {
-        types_constants.op(spv::Op::OpTypeVoid, 2);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    void output_word_le(uint32_t word) {
-        output_->put((word >> 0) & 0xFFu);
-        output_->put((word >> 8) & 0xFFu);
-        output_->put((word >> 16) & 0xFFu);
-        output_->put((word >> 24) & 0xFFu);
-    }
-
-    void output_section(SpvSectionBuilder& section) {
-        for (auto& word : section.data_) {
-            output_word_le(word);
-        }
-    }
-public:
-    const SpvId void_type;
-
-    void finish(std::ostream& output) {
-        output_ = &output;
-        SpvSectionBuilder memory_model_section;
-        memory_model_section.op(spv::Op::OpMemoryModel, 3);
-        memory_model_section.data_.push_back(addressing_model);
-        memory_model_section.data_.push_back(memory_model);
-
-        output_word_le(spv::MagicNumber);
-        output_word_le(spv::Version); // TODO: target a specific spirv version
-        output_word_le(uint32_t(0)); // TODO get a magic number ?
-        output_word_le(bound);
-        output_word_le(uint32_t(0)); // instruction schema padding
-
-        output_section(capabilities);
-        output_section(extensions);
-        output_section(ext_inst_import);
-        output_section(memory_model_section);
-        output_section(entry_points);
-        output_section(execution_modes);
-        output_section(debug_string_source);
-        output_section(debug_names);
-        output_section(debug_module_processed);
-        output_section(annotations);
-        output_section(types_constants);
-        output_section(fn_decls);
-        output_section(fn_defs);
-    }
-};
-
-SpvId SpvBasicBlockBuilder::generate_fresh_id() {
-    return file_builder.generate_fresh_id();
-}
-
 CodeGen::CodeGen(thorin::World& world, Cont2Config&, bool debug)
     : thorin::CodeGen(world, debug)
 {}
 
 void CodeGen::emit(std::ostream& out) {
-    SpvFileBuilder builder;
+    builder::SpvFileBuilder builder;
     builder_ = &builder;
     builder_->capability(spv::Capability::CapabilityShader);
     builder_->capability(spv::Capability::CapabilityLinkage);
 
+    cleanup_world(world());
+
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
     builder_->finish(out);
@@ -429,7 +122,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
-    auto fn = SpvFnBuilder { };
+    FnBuilder fn;
     fn.fn_type = convert(entry_->type());
     fn.fn_ret_type = get_codom_type(entry_);
 
@@ -438,11 +131,14 @@ void CodeGen::emit(const thorin::Scope& scope) {
     auto conts = schedule(scope);
 
     fn.bbs.reserve(conts.size());
+    std::vector<BasicBlockBuilder> bbs;
+    bbs.reserve(conts.size());
 
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
 
-        SpvBasicBlockBuilder* bb = &fn.bbs.emplace_back(*builder_);
+        BasicBlockBuilder* bb = &bbs.emplace_back(BasicBlockBuilder(*builder_));
+        fn.bbs.emplace_back(bb);
         auto [i, b] = fn.bbs_map.emplace(cont, bb);
         assert(b);
 
@@ -507,7 +203,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
     return builder_->declare_struct_type(types);
 }
 
-void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder* bb) {
+void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     if (continuation->callee() == entry_->ret_param()) {
         std::vector<SpvId> values;
 
@@ -632,7 +328,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, SpvBasicBlockBuilder* bb
     }
 }
 
-SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder* bb) {
+SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
         SpvId rhs = emit(bin->rhs(), bb);
@@ -757,9 +453,12 @@ SpvId CodeGen::emit(const Def* def, SpvBasicBlockBuilder* bb) {
         return constant;
     } else if(auto param = def->isa<Param>()) {
         if (auto param_id = current_fn_->params.lookup(param)) {
+            assert((*param_id).id != 0);
             return *param_id;
         } else {
-            return (*current_fn_->bbs_map[param->continuation()]).phis[param].value;
+            auto val = (*current_fn_->bbs_map[param->continuation()]).phis[param].value;
+            assert(val.id != 0);
+            return val;
         }
     }
     assertf(false, "Incomplete emit(def) definition");
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 8cbc42072..89d4a0451 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -1,16 +1,29 @@
 #ifndef THORIN_SPIRV_H
 #define THORIN_SPIRV_H
 
-#include <thorin/analyses/schedule.h>
+#include "thorin/be/spirv/spirv_builder.hpp"
 #include "thorin/be/backends.h"
 
+#include "thorin/analyses/schedule.h"
+
 namespace thorin::spirv {
 
-struct SpvSectionBuilder;
-struct SpvBasicBlockBuilder;
-struct SpvFnBuilder;
-struct SpvFileBuilder;
-struct SpvId { uint32_t id; };
+using SpvId = builder::SpvId;
+
+struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
+    explicit BasicBlockBuilder(builder::SpvFileBuilder& file_builder)
+    : builder::SpvBasicBlockBuilder(file_builder)
+    {}
+
+    std::unordered_map<const Param*, Phi> phis;
+    DefMap<SpvId> args;
+};
+
+struct FnBuilder : public builder::SpvFnBuilder {
+    std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
+    ContinuationMap<SpvId> labels;
+    DefMap<SpvId> params;
+};
 
 class CodeGen : public thorin::CodeGen {
 public:
@@ -18,16 +31,19 @@ class CodeGen : public thorin::CodeGen {
 
     void emit(std::ostream& stream) override;
 protected:
+    void structure_loops();
+    void structure_flow();
+
     SpvId convert(const Type*);
     void emit(const Scope& scope);
-    void emit_epilogue(Continuation*, SpvBasicBlockBuilder* bb);
-    SpvId emit(const Def* def, SpvBasicBlockBuilder* bb);
+    void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
+    SpvId emit(const Def* def, BasicBlockBuilder* bb);
 
     SpvId get_codom_type(const Continuation* fn);
 
-    SpvFileBuilder* builder_ = nullptr;
+    builder::SpvFileBuilder* builder_ = nullptr;
     Continuation* entry_ = nullptr;
-    SpvFnBuilder* current_fn_ = nullptr;
+    FnBuilder* current_fn_ = nullptr;
     Scheduler scheduler_;
     TypeMap<SpvId> types_;
     DefMap<SpvId> defs_;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
new file mode 100644
index 000000000..fb19b4af6
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -0,0 +1,325 @@
+#include <spirv/unified1/spirv.hpp>
+
+#include <string>
+#include <cassert>
+#include <vector>
+#include <unordered_map>
+#include <ostream>
+#include <memory>
+
+namespace thorin::spirv::builder {
+
+struct SpvId { uint32_t id; };
+
+struct SpvSectionBuilder;
+struct SpvBasicBlockBuilder;
+struct SpvFnBuilder;
+struct SpvFileBuilder;
+
+inline int div_roundup(int a, int b) {
+    if (a % b == 0)
+        return a / b;
+    else
+        return (a / b) + 1;
+}
+
+struct SpvSectionBuilder {
+    std::vector<uint32_t> data_;
+
+private:
+    void output_word(uint32_t word) {
+        data_.push_back(word);
+    }
+public:
+    void op(spv::Op op, int ops_size) {
+        uint32_t lower = op & 0xFFFFu;
+        uint32_t upper = (ops_size << 16) & 0xFFFF0000u;
+        output_word(lower | upper);
+    }
+
+    void ref_id(SpvId id) {
+        assert(id.id != 0);
+        output_word(id.id);
+    }
+
+    void literal_name(std::string_view str) {
+        int i = 0;
+        uint32_t cword = 0;
+        for (char c : str) {
+            cword = cword | (c & 0xFF) << (i * 8);
+            i++;
+            if (i == 4) {
+                output_word(cword);
+                cword = 0;
+                i = 0;
+            }
+        }
+        output_word(cword);
+    }
+
+    void literal_int(uint32_t i) {
+        output_word(i);
+    }
+};
+
+struct SpvBasicBlockBuilder : public SpvSectionBuilder {
+    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
+    : file_builder(file_builder)
+    {}
+
+    SpvFileBuilder& file_builder;
+
+    struct Phi {
+        SpvId type;
+        SpvId value;
+        std::vector<std::pair<SpvId, SpvId>> preds;
+    };
+    std::vector<Phi> phis;
+    SpvId label;
+
+    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
+        op(spv::Op::OpLabel, 3 + elements.size());
+        ref_id(aggregate_t);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        for (auto e : elements)
+            ref_id(e);
+        return id;
+    }
+
+    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
+        op(op_, 5);
+        auto id = generate_fresh_id();
+        ref_id(result_type);
+        ref_id(id);
+        ref_id(lhs);
+        ref_id(rhs);
+        return id;
+    }
+
+    void branch(SpvId target) {
+        op(spv::Op::OpBranch, 2);
+        ref_id(target);
+    }
+
+    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
+        op(spv::Op::OpBranchConditional, 4);
+        ref_id(condition);
+        ref_id(true_target);
+        ref_id(false_target);
+    }
+
+    void return_void() {
+        op(spv::Op::OpReturn, 1);
+    }
+
+    void return_value(SpvId value) {
+        op(spv::Op::OpReturnValue, 2);
+        ref_id(value);
+    }
+
+private:
+    SpvId generate_fresh_id();
+};
+
+struct SpvFnBuilder {
+public:
+    SpvId fn_type;
+    SpvId fn_ret_type;
+    std::vector<SpvBasicBlockBuilder*> bbs;
+
+    // Contains OpFunctionParams
+    SpvSectionBuilder header;
+};
+
+struct SpvFileBuilder {
+    SpvFileBuilder()
+    : void_type(declare_void_type())
+    {}
+
+    SpvId generate_fresh_id() { return {bound++ }; }
+
+    void name(SpvId id, std::string_view str) {
+        assert(id.id < bound);
+        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
+        debug_names.ref_id(id);
+        debug_names.literal_name(str);
+    }
+
+    SpvId declare_bool_type() {
+        types_constants.op(spv::Op::OpTypeBool, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    SpvId declare_int_type(int width, bool signed_) {
+        types_constants.op(spv::Op::OpTypeInt, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        types_constants.literal_int(signed_ ? 1 : 0);
+        return id;
+    }
+
+    SpvId declare_float_type(int width) {
+        types_constants.op(spv::Op::OpTypeFloat, 3);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        return id;
+    }
+
+    SpvId declare_fn_type(std::vector<SpvId>& dom, SpvId codom) {
+        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(codom);
+        for (auto arg : dom)
+            types_constants.ref_id(arg);
+        return id;
+    }
+
+    SpvId declare_struct_type(std::vector<SpvId>& elements) {
+        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        for (auto arg : elements)
+            types_constants.ref_id(arg);
+        return id;
+    }
+
+    SpvId bool_constant(SpvId type, bool value) {
+        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    SpvId constant(SpvId type, std::vector<uint32_t>&& bit_pattern) {
+        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto arg : bit_pattern)
+            types_constants.data_.push_back(arg);
+        return id;
+    }
+
+    SpvId define_function(SpvFnBuilder& fn_builder) {
+        fn_defs.op(spv::Op::OpFunction, 5);
+        fn_defs.ref_id(fn_builder.fn_ret_type);
+        auto id = generate_fresh_id();
+        fn_defs.ref_id(id);
+        fn_defs.data_.push_back(spv::FunctionControlMaskNone);
+        fn_defs.ref_id(fn_builder.fn_type);
+
+        // Includes stuff like OpFunctionParameters
+        for (auto w : fn_builder.header.data_)
+            fn_defs.data_.push_back(w);
+
+        for (auto& bb : fn_builder.bbs) {
+            fn_defs.op(spv::Op::OpLabel, 2);
+            fn_defs.ref_id(bb->label);
+
+            for (auto& phi : bb->phis) {
+                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi.preds.size());
+                fn_defs.ref_id(phi.type);
+                fn_defs.ref_id(phi.value);
+                for (auto& [pred_value, pred_label] : phi.preds) {
+                    fn_defs.ref_id(pred_value);
+                    fn_defs.ref_id(pred_label);
+                }
+            }
+
+            for (auto w : bb->data_)
+                fn_defs.data_.push_back(w);
+        }
+
+        fn_defs.op(spv::Op::OpFunctionEnd, 1);
+        return id;
+    }
+
+    void capability(spv::Capability cap) {
+        capabilities.op(spv::Op::OpCapability, 2);
+        capabilities.data_.push_back(cap);
+    }
+
+    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
+
+private:
+    std::ostream* output_ = nullptr;
+    uint32_t bound = 1;
+
+    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
+    SpvSectionBuilder capabilities;
+    SpvSectionBuilder extensions;
+    SpvSectionBuilder ext_inst_import;
+    SpvSectionBuilder entry_points;
+    SpvSectionBuilder execution_modes;
+    SpvSectionBuilder debug_string_source;
+    SpvSectionBuilder debug_names;
+    SpvSectionBuilder debug_module_processed;
+    SpvSectionBuilder annotations;
+    SpvSectionBuilder types_constants;
+    SpvSectionBuilder fn_decls;
+    SpvSectionBuilder fn_defs;
+
+    SpvId declare_void_type() {
+        types_constants.op(spv::Op::OpTypeVoid, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    void output_word_le(uint32_t word) {
+        output_->put((word >> 0) & 0xFFu);
+        output_->put((word >> 8) & 0xFFu);
+        output_->put((word >> 16) & 0xFFu);
+        output_->put((word >> 24) & 0xFFu);
+    }
+
+    void output_section(SpvSectionBuilder& section) {
+        for (auto& word : section.data_) {
+            output_word_le(word);
+        }
+    }
+public:
+    const SpvId void_type;
+
+    void finish(std::ostream& output) {
+        output_ = &output;
+        SpvSectionBuilder memory_model_section;
+        memory_model_section.op(spv::Op::OpMemoryModel, 3);
+        memory_model_section.data_.push_back(addressing_model);
+        memory_model_section.data_.push_back(memory_model);
+
+        output_word_le(spv::MagicNumber);
+        output_word_le(spv::Version); // TODO: target a specific spirv version
+        output_word_le(uint32_t(0)); // TODO get a magic number ?
+        output_word_le(bound);
+        output_word_le(uint32_t(0)); // instruction schema padding
+
+        output_section(capabilities);
+        output_section(extensions);
+        output_section(ext_inst_import);
+        output_section(memory_model_section);
+        output_section(entry_points);
+        output_section(execution_modes);
+        output_section(debug_string_source);
+        output_section(debug_names);
+        output_section(debug_module_processed);
+        output_section(annotations);
+        output_section(types_constants);
+        output_section(fn_decls);
+        output_section(fn_defs);
+    }
+};
+
+inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
+    return file_builder.generate_fresh_id();
+}
+
+}
\ No newline at end of file
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
new file mode 100644
index 000000000..85e7ac21f
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -0,0 +1,14 @@
+#include "thorin/be/spirv/spirv.h"
+#include "thorin/analyses/scope.h"
+
+namespace thorin::spirv {
+
+void CodeGen::structure_loops() {
+    // TODO
+}
+
+void CodeGen::structure_flow() {
+    // TODO
+}
+
+}
\ No newline at end of file

From b8c2c705b76da0e74dfe615e65b065aeec62cd09 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 15 Mar 2021 15:10:13 +0100
Subject: [PATCH 031/342] (initial stuff (does not work))

---
 src/thorin/CMakeLists.txt               |   1 +
 src/thorin/be/spirv/spirv.cpp           |   1 +
 src/thorin/be/spirv/spirv_transform.cpp | 126 +++++++++++++++++++++++-
 3 files changed, 127 insertions(+), 1 deletion(-)

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index e38158fac..5f55499a6 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -82,6 +82,7 @@ set(THORIN_SOURCES
     util/symbol.h
     util/types.h
     util/utility.h
+    util/dot_dump.cpp
     be/backends.cpp
     be/backends.h
     be/spirv/spirv_transform.cpp
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 6a3de9306..746c034d7 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -18,6 +18,7 @@ void CodeGen::emit(std::ostream& out) {
     builder_->capability(spv::Capability::CapabilityShader);
     builder_->capability(spv::Capability::CapabilityLinkage);
 
+    structure_loops();
     cleanup_world(world());
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 85e7ac21f..5ceb4185c 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -1,10 +1,134 @@
+#include <thorin/analyses/looptree.h>
 #include "thorin/be/spirv/spirv.h"
 #include "thorin/analyses/scope.h"
+#include "thorin/analyses/cfg.h"
+
+#include <algorithm>
 
 namespace thorin::spirv {
 
+using Head = LoopTree<true>::Head;
+using Base = LoopTree<true>::Base;
+using Leaf = LoopTree<true>::Leaf;
+
+struct StructuredLoop {
+    const Head* old_head;
+    const std::string name;
+    Continuation* new_header;
+    Continuation* inner_dispatch;
+    Continuation* outer_dispatch;
+
+    std::vector<Continuation*> inner_destinations;
+    std::vector<Continuation*> outer_destinations;
+};
+
+inline int get_or_create_destination(std::vector<Continuation*>& vec, Continuation* destination) {
+    if (auto i = std::find(vec.begin(), vec.end(), destination); i != vec.end())
+        return i - vec.begin();
+    vec.emplace_back(destination);
+    return vec.size() - 1;
+}
+
+struct ScopeContext {
+    explicit ScopeContext(const Scope& scope)
+    : cfa(scope)
+    {}
+
+    CFA cfa;
+    ContinuationMap<const Head*> def2loop;
+    std::unordered_map<const Head*, StructuredLoop> rewritten_loops;
+};
+
+/// Visits the forest and fills def2loop
+inline void tagContinuations(ScopeContext& ctx, const Base* base, const Head* parent) {
+    for (int i = 0; i < base->depth(); i++)
+        printf(" ");
+    if (auto* head = base->isa<Head>()) {
+        printf("loop: header = ");
+    }
+    for (auto& node : base->cf_nodes()) {
+        printf("%s ", node->continuation()->to_string().c_str());
+    }
+    printf("\n");
+
+    if (auto* head = base->isa<Head>()) {
+        for (auto& children : head->children()) {
+            tagContinuations(ctx, &*children, head);
+        }
+    } else if(auto* leaf = base->isa<Leaf>()) {
+        for (auto& node : base->cf_nodes()) {
+            auto[i, result] = ctx.def2loop.emplace(node->continuation(), parent);
+            assert(result);
+        }
+    } else {
+        assert(false);
+    }
+}
+
+inline std::string safe_name(const Head* head) {
+    if (head == nullptr || head->is_root()) {
+        return "root";
+    } else {
+        std::stringstream s;
+        s << "loop_";
+        for (auto& node : head->cf_nodes()) {
+            s << node->continuation()->to_string();
+            s << "_";
+        }
+        return s.str();
+    }
+}
+
+inline void augmentLoops(World& world, ScopeContext& ctx, const Base* base) {
+    if (const Head* head = base->isa<Head>()) {
+        StructuredLoop loop {
+            head,
+            safe_name(head),
+            world.continuation({ safe_name(head) + "_header"}),
+            world.continuation({ safe_name(head) + "_inner_dispatch"}),
+            world.continuation({ safe_name(head) + "_outer_dispatch"}),
+            {},
+            {},
+        };
+
+        for (auto& header_node : base->cf_nodes()) {
+            const Head* dest_loop = *ctx.def2loop[header_node->continuation()];
+            for (auto& pred : header_node->continuation()->preds()) {
+                const Head* src_loop = *ctx.def2loop[pred];
+                printf("%s -> %s\n", safe_name(src_loop).c_str(), safe_name(dest_loop).c_str());
+
+                // Backedge !
+                if (src_loop == dest_loop) {
+                    get_or_create_destination(loop.inner_destinations, header_node->continuation());
+
+                    // point backedge to loop header
+                    for (int i = 0; i < pred->num_ops(); i++) {
+                        if (pred->op(i) == header_node->continuation()) {
+                            pred->unset_op(i);
+                            pred->set_op(i, loop.new_header);
+                        }
+                    }
+                }
+            }
+        }
+
+        ctx.rewritten_loops.emplace(head, loop);
+
+        for (auto& children : head->children()) {
+            augmentLoops(world, ctx, &*children);
+        }
+    }
+}
+
 void CodeGen::structure_loops() {
-    // TODO
+    Scope::for_each(world(), [&](const Scope& scope) {
+        ScopeContext context(scope);
+
+        const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
+        tagContinuations(context, looptree.root(), nullptr);
+
+        augmentLoops(world(), context, looptree.root());
+    });
 }
 
 void CodeGen::structure_flow() {

From 3e70c3024ec6b1a80af3f0c8be0640b92a590be3 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 15 Mar 2021 15:59:23 +0100
Subject: [PATCH 032/342] more nonsense

---
 src/thorin/be/spirv/spirv_transform.cpp | 50 +++++++++++++++++++++----
 1 file changed, 42 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 5ceb4185c..62b747768 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -79,33 +79,67 @@ inline std::string safe_name(const Head* head) {
     }
 }
 
+inline const Type* dom_to_tuple(World& world, const thorin::FnType* fn_type) {
+    std::vector<const thorin::Type*> t;
+    t.resize(fn_type->num_ops());
+    for (size_t i = 0; i < fn_type->num_ops(); i++)
+        t[i] = fn_type->op(i);
+    return world.tuple_type(t);
+}
+
 inline void augmentLoops(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
+        auto name = safe_name(head);
+
+        const thorin::Type* unified_heads_param;
+        const thorin::VariantType* header_variant_type = nullptr;
+        if (base->cf_nodes().size() >= 1) {
+            header_variant_type = world.variant_type(name, base->cf_nodes().size());
+            int i = 0;
+            for (auto& header_node : base->cf_nodes()) {
+                header_variant_type->set(i++, dom_to_tuple(world, header_node->continuation()->type()));
+            }
+            unified_heads_param = header_variant_type;
+        //} else if (base->cf_nodes().size() == 1) {
+        //    unified_heads_param = dom_to_tuple(world, base->cf_nodes()[0]->continuation()->type());
+        } else {
+            unified_heads_param = world.unit();
+        }
+
+        auto fn_type = world.fn_type( { unified_heads_param } );
+
         StructuredLoop loop {
             head,
-            safe_name(head),
-            world.continuation({ safe_name(head) + "_header"}),
-            world.continuation({ safe_name(head) + "_inner_dispatch"}),
-            world.continuation({ safe_name(head) + "_outer_dispatch"}),
+            name,
+            world.continuation(fn_type, { name + "_header"}),
+            world.continuation({ name + "_inner_dispatch"}),
+            world.continuation({ name + "_outer_dispatch"}),
             {},
             {},
         };
 
-        for (auto& header_node : base->cf_nodes()) {
+        // Handle internal backedges: re-wire them to go through header
+        for (size_t header_index = 0; header_index < base->num_cf_nodes(); header_index++) {
+            auto& header_node = base->cf_nodes()[header_index];
+
             const Head* dest_loop = *ctx.def2loop[header_node->continuation()];
             for (auto& pred : header_node->continuation()->preds()) {
                 const Head* src_loop = *ctx.def2loop[pred];
                 printf("%s -> %s\n", safe_name(src_loop).c_str(), safe_name(dest_loop).c_str());
 
-                // Backedge !
                 if (src_loop == dest_loop) {
                     get_or_create_destination(loop.inner_destinations, header_node->continuation());
 
-                    // point backedge to loop header
                     for (int i = 0; i < pred->num_ops(); i++) {
                         if (pred->op(i) == header_node->continuation()) {
                             pred->unset_op(i);
-                            pred->set_op(i, loop.new_header);
+
+                            // Oops except we can't do that! The new header has a mismatched signature, so we must go through a synthetic wrapper
+                            //pred->set_op(i, loop.new_header);
+
+                            auto wrapper = world.continuation(fn_type, {"synthetic_backedge_wrapper"});
+                            wrapper->jump(loop.new_header, { world.variant(header_variant_type, world.tuple(pred->args()), header_index) });
+                            pred->set_op(i, wrapper);
                         }
                     }
                 }

From 1582a65efbd8c7f87e4e8c0f7355c9257656b89d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Mar 2021 10:38:32 +0100
Subject: [PATCH 033/342] proper types for headers/epilogues

---
 src/thorin/be/spirv/spirv_transform.cpp | 290 +++++++++++++++++++-----
 1 file changed, 236 insertions(+), 54 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 62b747768..8b8373c72 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -11,23 +11,35 @@ using Head = LoopTree<true>::Head;
 using Base = LoopTree<true>::Base;
 using Leaf = LoopTree<true>::Leaf;
 
+struct StructuredLoop;
+
+// Dispatch targets may dispatch to other dispatching nodes, and we can't actually create those until we know all their destinations,
+// because their fn type takes a variant type with a case for each target. So we symbolically refer to these yet-to-be dispatch nodes via their loop
+struct DispatchTarget {
+    bool operator==(const DispatchTarget& rhs) const {
+        return cont == rhs.cont &&
+               entry == rhs.entry &&
+               exit == rhs.exit;
+    }
+
+    Continuation* cont = nullptr;
+    StructuredLoop* entry = nullptr;
+    StructuredLoop* exit = nullptr;
+};
+
+// Represents one loop in the loop forest, that we then augment with a new loop header and epilogue, each dispatching
+// respectively to nodes inside of the loop, and nodes outside of the loop once we break out
 struct StructuredLoop {
-    const Head* old_head;
+    const Head* parent_head;
+    const Head* head;
     const std::string name;
-    Continuation* new_header;
-    Continuation* inner_dispatch;
-    Continuation* outer_dispatch;
 
-    std::vector<Continuation*> inner_destinations;
-    std::vector<Continuation*> outer_destinations;
-};
+    std::vector<DispatchTarget> inner_destinations = {};
+    std::vector<DispatchTarget> outer_destinations = {};
 
-inline int get_or_create_destination(std::vector<Continuation*>& vec, Continuation* destination) {
-    if (auto i = std::find(vec.begin(), vec.end(), destination); i != vec.end())
-        return i - vec.begin();
-    vec.emplace_back(destination);
-    return vec.size() - 1;
-}
+    Continuation* new_header = nullptr;
+    Continuation* new_epilogue = nullptr;
+};
 
 struct ScopeContext {
     explicit ScopeContext(const Scope& scope)
@@ -39,8 +51,22 @@ struct ScopeContext {
     std::unordered_map<const Head*, StructuredLoop> rewritten_loops;
 };
 
+inline std::string safe_name(const Head* head) {
+    if (head == nullptr || head->is_root()) {
+        return "root";
+    } else {
+        std::stringstream s;
+        s << "loop_";
+        for (auto& node : head->cf_nodes()) {
+            s << node->continuation()->to_string();
+            s << "_";
+        }
+        return s.str();
+    }
+}
+
 /// Visits the forest and fills def2loop
-inline void tagContinuations(ScopeContext& ctx, const Base* base, const Head* parent) {
+inline void tag_continuations(ScopeContext& ctx, const Base* base, const Head* parent) {
     for (int i = 0; i < base->depth(); i++)
         printf(" ");
     if (auto* head = base->isa<Head>()) {
@@ -51,10 +77,18 @@ inline void tagContinuations(ScopeContext& ctx, const Base* base, const Head* pa
     }
     printf("\n");
 
-    if (auto* head = base->isa<Head>()) {
+    if (const Head* head = base->isa<Head>()) {
+        auto name = safe_name(head);
+
         for (auto& children : head->children()) {
-            tagContinuations(ctx, &*children, head);
+            tag_continuations(ctx, &*children, head);
         }
+
+        if (parent != nullptr) {
+            StructuredLoop loop{parent, head, name};
+            ctx.rewritten_loops.emplace(head, loop);
+        }
+
     } else if(auto* leaf = base->isa<Leaf>()) {
         for (auto& node : base->cf_nodes()) {
             auto[i, result] = ctx.def2loop.emplace(node->continuation(), parent);
@@ -65,17 +99,117 @@ inline void tagContinuations(ScopeContext& ctx, const Base* base, const Head* pa
     }
 }
 
-inline std::string safe_name(const Head* head) {
-    if (head == nullptr || head->is_root()) {
-        return "root";
+inline int record_destination(std::vector<DispatchTarget>& vec, DispatchTarget dest) {
+    auto i = std::find(vec.begin(), vec.end(), dest);
+    if (i == vec.end()) {
+        vec.emplace_back(dest);
+        return vec.size() - 1;
+    } else return i - vec.begin();
+}
+
+inline std::vector<StructuredLoop*>&& get_path(ScopeContext& ctx, const Head* head) {
+    std::vector<StructuredLoop*> path;
+    do {
+        if (head != nullptr) {
+            auto* loop = &ctx.rewritten_loops[head];
+            path.emplace(path.begin(), loop);
+            head = loop->parent_head;
+        } else {
+            path.emplace(path.begin(), nullptr);
+            head = nullptr;
+        }
+    } while (head != nullptr);
+    return std::move(path);
+}
+
+inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base* base, const Head* parent) {
+    if (const Head* head = base->isa<Head>()) {
+        for (auto& children : head->children()) {
+            collect_dispatch_targets(world, ctx, &*children, head);
+        }
     } else {
-        std::stringstream s;
-        s << "loop_";
-        for (auto& node : head->cf_nodes()) {
-            s << node->continuation()->to_string();
-            s << "_";
+        const Leaf* leaf = base->as<Leaf>();
+        auto cont = leaf->cf_node()->continuation();
+        for (size_t i = 0; i < cont->num_ops(); i++) {
+            auto def = cont->op(i);
+            if (auto dest = def->isa_continuation()) {
+                const Head* source_loop = *ctx.def2loop[cont];
+                const Head* dest_loop = *ctx.def2loop[cont];
+
+                if (source_loop != dest_loop) {
+                    // We found a non-local jump
+
+                    auto source_path = get_path(ctx, source_loop);
+                    auto dest_path = get_path(ctx, dest_loop);
+                    size_t bi = 0;
+                    while (bi < std::min(source_path.size(), dest_path.size())) {
+                        if (source_path[bi] == dest_path[bi])
+                            bi++;
+                        else break;
+                    }
+
+                    // The path is made out of a sequence of loops to break out of, and a sequence of loops to jump into
+                    // these two sequences cannot be both empty (that wouldn't be a non-local jump then!)
+                    std::vector<StructuredLoop*> leave;
+                    std::vector<StructuredLoop*> enter;
+                    for (size_t j = source_path.size() - 1; j >= bi; j--)
+                        leave.emplace_back(source_path[j]);
+                    for (size_t j = bi; j < dest_path.size(); j++)
+                        enter.emplace_back(dest_path[j]);
+
+                    // 0 = this is the first step of the path
+                    // 1 = last step was to break out of a loop
+                    // 2 = last step was to enter a loop
+                    int last = 0;
+                    StructuredLoop* prev;
+
+                    auto record_step = [&](StructuredLoop* loop, DispatchTarget destination) {
+                        if (last == 0) {
+                            // nothing to do, this node isn't a dispatching one
+                        } else {
+                            if (last == 1)
+                                record_destination(prev->outer_destinations, destination);
+                            else
+                                record_destination(prev->inner_destinations, destination);
+                        }
+                    };
+
+                    for (auto loop : leave) {
+                        DispatchTarget destination;
+                        destination.exit = loop;
+
+                        record_step(loop, destination);
+                        last = 1;
+                        prev = loop;
+                    }
+                    for (auto loop : enter) {
+                        DispatchTarget destination;
+                        destination.entry = loop;
+
+                        record_step(loop, destination);
+                        last = 2;
+                        prev = loop;
+                    }
+
+                    assert(last != 0);
+                    DispatchTarget destination;
+                    destination.cont = dest;
+                    record_step(prev, destination);
+                } else if (source_loop == dest_loop && source_loop != nullptr) {
+                    for (auto head : source_loop->cf_nodes()) {
+                        if (head->continuation() == dest) {
+                            // We found a backedge
+                            auto loop = ctx.rewritten_loops[source_loop];
+
+                            DispatchTarget destination;
+                            destination.cont = dest;
+                            record_destination(loop.inner_destinations, destination);
+                            break;
+                        }
+                    }
+                }
+            }
         }
-        return s.str();
     }
 }
 
@@ -87,37 +221,83 @@ inline const Type* dom_to_tuple(World& world, const thorin::FnType* fn_type) {
     return world.tuple_type(t);
 }
 
-inline void augmentLoops(World& world, ScopeContext& ctx, const Base* base) {
+inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
+    if (const Head* head = base->isa<Head>()) {
+        for (auto& children : head->children()) {
+            create_headers(world, ctx, &*children);
+        }
+
+        if (head->num_cf_nodes() == 0)
+            return;
+        StructuredLoop& loop = ctx.rewritten_loops[head];
+
+        // here, parent headers need to know what they're jumping *into*
+        std::vector<const Type*> dest_types;
+        for (auto& target : loop.inner_destinations) {
+            const thorin::FnType* target_type;
+            if (target.cont != nullptr) {
+                target_type = target.cont->type();
+            } else if (target.entry != nullptr) {
+                assert(target.entry->new_header != nullptr);
+                target_type = target.entry->new_header->type();
+            } else {
+                assert(false && "Header dispatches may not exit loops");
+            }
+            dest_types.emplace_back(dom_to_tuple(world, target_type));
+        }
+        auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
+        for (size_t i = 0; i < dest_types.size(); i++)
+            variant_type->set(i, dest_types[i]);
+        auto fn_type = world.fn_type( { variant_type } );
+        loop.new_header = world.continuation(fn_type, { loop.name + "_new_header"});
+    }
+}
+
+inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
+        StructuredLoop& loop = ctx.rewritten_loops[head];
+
+        if (head->num_cf_nodes() > 0) {
+            // here, children epilogues need to know what they're jumping *out to*
+            std::vector<const Type*> dest_types;
+            for (auto& target : loop.inner_destinations) {
+                const thorin::FnType* target_type;
+                if (target.cont != nullptr) {
+                    target_type = target.cont->type();
+                } else if (target.entry != nullptr) {
+                    assert(target.entry->new_header != nullptr);
+                    target_type = target.entry->new_header->type();
+                } else {
+                    assert(target.exit != nullptr);
+                    assert(target.exit->new_epilogue != nullptr);
+                    target_type = target.exit->new_epilogue->type();
+                }
+                dest_types.emplace_back(dom_to_tuple(world, target_type));
+            }
+            auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
+            for (size_t i = 0; i < dest_types.size(); i++)
+                variant_type->set(i, dest_types[i]);
+            auto fn_type = world.fn_type({variant_type});
+            loop.new_header = world.continuation(fn_type, {loop.name + "_new_epilogue"});
+        }
+
+        for (auto& children : head->children()) {
+            create_epilogues(world, ctx, &*children);
+        }
+    }
+}
+
+// This creates the header/epilogue nodes for the loops
+inline void augment_loops(World& world, ScopeContext& ctx, const Base* base) {
+    if (const Head* head = base->isa<Head>()) {
+        auto& loop = ctx.rewritten_loops[head];
         auto name = safe_name(head);
 
         const thorin::Type* unified_heads_param;
         const thorin::VariantType* header_variant_type = nullptr;
-        if (base->cf_nodes().size() >= 1) {
-            header_variant_type = world.variant_type(name, base->cf_nodes().size());
-            int i = 0;
-            for (auto& header_node : base->cf_nodes()) {
-                header_variant_type->set(i++, dom_to_tuple(world, header_node->continuation()->type()));
-            }
-            unified_heads_param = header_variant_type;
-        //} else if (base->cf_nodes().size() == 1) {
-        //    unified_heads_param = dom_to_tuple(world, base->cf_nodes()[0]->continuation()->type());
-        } else {
-            unified_heads_param = world.unit();
-        }
 
         auto fn_type = world.fn_type( { unified_heads_param } );
 
-        StructuredLoop loop {
-            head,
-            name,
-            world.continuation(fn_type, { name + "_header"}),
-            world.continuation({ name + "_inner_dispatch"}),
-            world.continuation({ name + "_outer_dispatch"}),
-            {},
-            {},
-        };
-
         // Handle internal backedges: re-wire them to go through header
         for (size_t header_index = 0; header_index < base->num_cf_nodes(); header_index++) {
             auto& header_node = base->cf_nodes()[header_index];
@@ -128,7 +308,7 @@ inline void augmentLoops(World& world, ScopeContext& ctx, const Base* base) {
                 printf("%s -> %s\n", safe_name(src_loop).c_str(), safe_name(dest_loop).c_str());
 
                 if (src_loop == dest_loop) {
-                    get_or_create_destination(loop.inner_destinations, header_node->continuation());
+                    // get_or_create_destination(loop.inner_destinations, header_node->continuation());
 
                     for (int i = 0; i < pred->num_ops(); i++) {
                         if (pred->op(i) == header_node->continuation()) {
@@ -146,10 +326,8 @@ inline void augmentLoops(World& world, ScopeContext& ctx, const Base* base) {
             }
         }
 
-        ctx.rewritten_loops.emplace(head, loop);
-
         for (auto& children : head->children()) {
-            augmentLoops(world, ctx, &*children);
+            augment_loops(world, ctx, &*children);
         }
     }
 }
@@ -159,9 +337,13 @@ void CodeGen::structure_loops() {
         ScopeContext context(scope);
 
         const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
-        tagContinuations(context, looptree.root(), nullptr);
+        tag_continuations(context, looptree.root(), nullptr);
+        collect_dispatch_targets(world(), context, looptree.root(), nullptr);
+
+        create_headers(world(), context, looptree.root());
+        create_epilogues(world(), context, looptree.root());
 
-        augmentLoops(world(), context, looptree.root());
+        augment_loops(world(), context, looptree.root());
     });
 }
 

From 739a23fb2680bf5d7fe4639c5227d33efeef9ce9 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 17 Mar 2021 16:31:14 +0100
Subject: [PATCH 034/342] handle NLJs

---
 src/thorin/be/spirv/spirv_transform.cpp | 234 ++++++++++++++++++------
 1 file changed, 178 insertions(+), 56 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 8b8373c72..a6687fb81 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -27,6 +27,19 @@ struct DispatchTarget {
     StructuredLoop* exit = nullptr;
 };
 
+struct RewireMe {
+    RewireMe(Continuation* cont, int op) : cont(cont), op(op) {}
+    Continuation* cont;
+    int op;
+
+    Continuation* backedge = nullptr;
+    struct {
+        std::vector<StructuredLoop*> exits;
+        std::vector<StructuredLoop*> enters;
+        Continuation* final_destination = nullptr;
+    } non_local_jump;
+};
+
 // Represents one loop in the loop forest, that we then augment with a new loop header and epilogue, each dispatching
 // respectively to nodes inside of the loop, and nodes outside of the loop once we break out
 struct StructuredLoop {
@@ -39,6 +52,8 @@ struct StructuredLoop {
 
     Continuation* new_header = nullptr;
     Continuation* new_epilogue = nullptr;
+
+    std::vector<RewireMe> rewire;
 };
 
 struct ScopeContext {
@@ -84,10 +99,8 @@ inline void tag_continuations(ScopeContext& ctx, const Base* base, const Head* p
             tag_continuations(ctx, &*children, head);
         }
 
-        if (parent != nullptr) {
-            StructuredLoop loop{parent, head, name};
-            ctx.rewritten_loops.emplace(head, loop);
-        }
+        StructuredLoop loop{parent, head, name};
+        ctx.rewritten_loops.emplace(head, loop);
 
     } else if(auto* leaf = base->isa<Leaf>()) {
         for (auto& node : base->cf_nodes()) {
@@ -107,19 +120,23 @@ inline int record_destination(std::vector<DispatchTarget>& vec, DispatchTarget d
     } else return i - vec.begin();
 }
 
-inline std::vector<StructuredLoop*>&& get_path(ScopeContext& ctx, const Head* head) {
-    std::vector<StructuredLoop*> path;
-    do {
-        if (head != nullptr) {
-            auto* loop = &ctx.rewritten_loops[head];
-            path.emplace(path.begin(), loop);
-            head = loop->parent_head;
-        } else {
-            path.emplace(path.begin(), nullptr);
-            head = nullptr;
-        }
-    } while (head != nullptr);
-    return std::move(path);
+inline int index_of_destination(std::vector<DispatchTarget>& vec, DispatchTarget dest) {
+    auto i = std::find(vec.begin(), vec.end(), dest);
+    if (i == vec.end()) {
+        assert(false && "Missing destination");
+    } else return i - vec.begin();
+}
+
+inline std::vector<StructuredLoop*> get_path(ScopeContext& ctx, const Head* head) {
+    std::vector<StructuredLoop*> path = {};
+    assert(head != nullptr);
+    while (head != nullptr) {
+        auto* loop = &ctx.rewritten_loops[head];
+        assert(loop != nullptr);
+        path.emplace(path.begin(), loop);
+        head = loop->parent_head;
+    }
+    return path;
 }
 
 inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base* base, const Head* parent) {
@@ -134,14 +151,22 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
             auto def = cont->op(i);
             if (auto dest = def->isa_continuation()) {
                 const Head* source_loop = *ctx.def2loop[cont];
-                const Head* dest_loop = *ctx.def2loop[cont];
+
+                if (dest->intrinsic() == Intrinsic::Branch) {
+                    continue;
+                }
+
+                assert(ctx.def2loop.find(dest) != ctx.def2loop.end());
+                const Head* dest_loop = *ctx.def2loop[dest];
 
                 if (source_loop != dest_loop) {
                     // We found a non-local jump
+                    assert(ctx.rewritten_loops.find(source_loop) != ctx.rewritten_loops.end());
+                    auto& loop = ctx.rewritten_loops[source_loop];
 
-                    auto source_path = get_path(ctx, source_loop);
-                    auto dest_path = get_path(ctx, dest_loop);
-                    size_t bi = 0;
+                    std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop);
+                    std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop);
+                    int bi = 0;
                     while (bi < std::min(source_path.size(), dest_path.size())) {
                         if (source_path[bi] == dest_path[bi])
                             bi++;
@@ -152,9 +177,9 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                     // these two sequences cannot be both empty (that wouldn't be a non-local jump then!)
                     std::vector<StructuredLoop*> leave;
                     std::vector<StructuredLoop*> enter;
-                    for (size_t j = source_path.size() - 1; j >= bi; j--)
+                    for (int j = source_path.size() - 1; j >= bi; j--)
                         leave.emplace_back(source_path[j]);
-                    for (size_t j = bi; j < dest_path.size(); j++)
+                    for (int j = bi; j < dest_path.size(); j++)
                         enter.emplace_back(dest_path[j]);
 
                     // 0 = this is the first step of the path
@@ -181,6 +206,7 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                         record_step(loop, destination);
                         last = 1;
                         prev = loop;
+                        assert(prev != nullptr);
                     }
                     for (auto loop : enter) {
                         DispatchTarget destination;
@@ -189,21 +215,37 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                         record_step(loop, destination);
                         last = 2;
                         prev = loop;
+                        assert(prev != nullptr);
                     }
 
                     assert(last != 0);
                     DispatchTarget destination;
                     destination.cont = dest;
                     record_step(prev, destination);
+
+                    RewireMe rewire(cont, i);
+                    rewire.non_local_jump = {
+                        std::move(leave),
+                        std::move(enter),
+                        dest
+                    };
+                    loop.rewire.emplace_back(rewire);
+                    printf("nlj %s %d!\n", loop.name.c_str(), loop.rewire.size());
                 } else if (source_loop == dest_loop && source_loop != nullptr) {
-                    for (auto head : source_loop->cf_nodes()) {
+                    for (auto& head : source_loop->cf_nodes()) {
                         if (head->continuation() == dest) {
                             // We found a backedge
-                            auto loop = ctx.rewritten_loops[source_loop];
+                            assert(ctx.rewritten_loops.find(source_loop) != ctx.rewritten_loops.end());
+                            auto& loop = ctx.rewritten_loops[source_loop];
 
                             DispatchTarget destination;
                             destination.cont = dest;
                             record_destination(loop.inner_destinations, destination);
+
+                            RewireMe rewire(cont, i);
+                            rewire.backedge = head->continuation();
+                            loop.rewire.emplace_back(rewire);
+                            printf("backedge %s %d!\n", loop.name.c_str(), loop.rewire.size());
                             break;
                         }
                     }
@@ -221,6 +263,14 @@ inline const Type* dom_to_tuple(World& world, const thorin::FnType* fn_type) {
     return world.tuple_type(t);
 }
 
+inline const Def* tuple_from_params(World& world, const ArrayRef<const Param*> params) {
+    std::vector<const thorin::Def*> t;
+    t.resize(params.size());
+    for (size_t i = 0; i < params.size(); i++)
+        t[i] = params[i];
+    return world.tuple(t);
+}
+
 inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
         for (auto& children : head->children()) {
@@ -260,7 +310,7 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
         if (head->num_cf_nodes() > 0) {
             // here, children epilogues need to know what they're jumping *out to*
             std::vector<const Type*> dest_types;
-            for (auto& target : loop.inner_destinations) {
+            for (auto& target : loop.outer_destinations) {
                 const thorin::FnType* target_type;
                 if (target.cont != nullptr) {
                     target_type = target.cont->type();
@@ -278,7 +328,7 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
             for (size_t i = 0; i < dest_types.size(); i++)
                 variant_type->set(i, dest_types[i]);
             auto fn_type = world.fn_type({variant_type});
-            loop.new_header = world.continuation(fn_type, {loop.name + "_new_epilogue"});
+            loop.new_epilogue = world.continuation(fn_type, {loop.name + "_new_epilogue"});
         }
 
         for (auto& children : head->children()) {
@@ -287,47 +337,118 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
     }
 }
 
-// This creates the header/epilogue nodes for the loops
-inline void augment_loops(World& world, ScopeContext& ctx, const Base* base) {
+inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
+        assert(ctx.rewritten_loops.find(head) != ctx.rewritten_loops.end());
         auto& loop = ctx.rewritten_loops[head];
-        auto name = safe_name(head);
 
-        const thorin::Type* unified_heads_param;
-        const thorin::VariantType* header_variant_type = nullptr;
+        for (auto& children : head->children()) {
+            rewire_loops(world, ctx, &*children);
+        }
 
-        auto fn_type = world.fn_type( { unified_heads_param } );
+        printf("rewires %s %d!\n", loop.name.c_str(), loop.rewire.size());
+        for (auto& rewire : loop.rewire) {
+            printf("rewire!\n");
+            if (rewire.backedge != nullptr) {
+                printf("handling BE!\n");
+                DispatchTarget destination;
+                destination.cont = rewire.backedge;
+                auto variant_index = index_of_destination(loop.inner_destinations, destination);
 
-        // Handle internal backedges: re-wire them to go through header
-        for (size_t header_index = 0; header_index < base->num_cf_nodes(); header_index++) {
-            auto& header_node = base->cf_nodes()[header_index];
+                auto old_fn_type = rewire.backedge->type();
+                auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper"});
 
-            const Head* dest_loop = *ctx.def2loop[header_node->continuation()];
-            for (auto& pred : header_node->continuation()->preds()) {
-                const Head* src_loop = *ctx.def2loop[pred];
-                printf("%s -> %s\n", safe_name(src_loop).c_str(), safe_name(dest_loop).c_str());
+                auto header_variant_type = loop.new_header->type()->op(0)->as<VariantType>();
+                wrapper->jump(loop.new_header, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
 
-                if (src_loop == dest_loop) {
-                    // get_or_create_destination(loop.inner_destinations, header_node->continuation());
+                rewire.cont->unset_op(rewire.op);
+                rewire.cont->set_op(rewire.op, wrapper);
+            } else {
+                printf("handling NLJ!\n");
 
-                    for (int i = 0; i < pred->num_ops(); i++) {
-                        if (pred->op(i) == header_node->continuation()) {
-                            pred->unset_op(i);
+                auto& nlj = rewire.non_local_jump;
+                auto old_fn_type = nlj.final_destination->type();
+                auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper"});
 
-                            // Oops except we can't do that! The new header has a mismatched signature, so we must go through a synthetic wrapper
-                            //pred->set_op(i, loop.new_header);
+                const Def* argument = tuple_from_params(world, wrapper->params());
+                Continuation* first_jump = nullptr;
 
-                            auto wrapper = world.continuation(fn_type, {"synthetic_backedge_wrapper"});
-                            wrapper->jump(loop.new_header, { world.variant(header_variant_type, world.tuple(pred->args()), header_index) });
-                            pred->set_op(i, wrapper);
-                        }
+                DispatchTarget destination;
+                destination.cont = nlj.final_destination;
+
+                for (int i = nlj.enters.size() - 1; i >= 0; i--) {
+                    StructuredLoop* loop_to_enter = nlj.enters[i];
+
+                    auto variant_index = index_of_destination(loop_to_enter->inner_destinations, destination);
+                    auto header_variant_type = loop_to_enter->new_header->type()->op(0)->as<VariantType>();
+                    argument = world.variant(header_variant_type, argument, variant_index);
+
+                    first_jump = loop_to_enter->new_header;
+                    destination = {};
+                    destination.entry = loop_to_enter;
+                }
+
+                for (int i = nlj.exits.size() - 1; i >= 0; i--) {
+                    StructuredLoop* loop_to_exit = nlj.exits[i];
+
+                    auto variant_index = index_of_destination(loop_to_exit->outer_destinations, destination);
+                    auto header_variant_type = loop_to_exit->new_epilogue->type()->op(0)->as<VariantType>();
+                    argument = world.variant(header_variant_type, argument, variant_index);
+
+                    first_jump = loop_to_exit->new_epilogue;
+                    destination = {};
+                    destination.exit = loop_to_exit;
+                }
+
+                assert(first_jump != nullptr);
+                wrapper->jump(first_jump, { argument });
+
+                rewire.cont->unset_op(rewire.op);
+                rewire.cont->set_op(rewire.op, wrapper);
+
+                // ---------------------------------------------------------------------------------------
+
+                /*// 0 = this is the first step of the path
+                // 1 = last step was to break out of a loop
+                // 2 = last step was to enter a loop
+                int last = 0;
+                StructuredLoop* prev;
+
+                auto record_step = [&](StructuredLoop* loop, DispatchTarget destination) {
+                    if (last == 0) {
+                        // nothing to do, this node isn't a dispatching one
+                    } else {
+                        if (last == 1)
+                            record_destination(prev->outer_destinations, destination);
+                        else
+                            record_destination(prev->inner_destinations, destination);
                     }
+                };
+
+                for (auto loop : leave) {
+                    DispatchTarget destination;
+                    destination.exit = loop;
+
+                    record_step(loop, destination);
+                    last = 1;
+                    prev = loop;
+                    assert(prev != nullptr);
                 }
-            }
-        }
+                for (auto loop : enter) {
+                    DispatchTarget destination;
+                    destination.entry = loop;
 
-        for (auto& children : head->children()) {
-            augment_loops(world, ctx, &*children);
+                    record_step(loop, destination);
+                    last = 2;
+                    prev = loop;
+                    assert(prev != nullptr);
+                }
+
+                assert(last != 0);
+                DispatchTarget destination;
+                destination.cont = dest;
+                record_step(prev, destination);*/
+            }
         }
     }
 }
@@ -343,7 +464,8 @@ void CodeGen::structure_loops() {
         create_headers(world(), context, looptree.root());
         create_epilogues(world(), context, looptree.root());
 
-        augment_loops(world(), context, looptree.root());
+        rewire_loops(world(), context, looptree.root());
+        printf("done\n");
     });
 }
 

From cbdfe3ac5209c1e07915b156bb27c1b9615535ba Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 18 Mar 2021 13:06:51 +0100
Subject: [PATCH 035/342] use intrinsics

---
 src/thorin/be/spirv/spirv_transform.cpp | 95 +++++++++++--------------
 src/thorin/continuation.cpp             | 25 +++++++
 src/thorin/continuation.h               |  6 ++
 3 files changed, 72 insertions(+), 54 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index a6687fb81..5be442d66 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -17,9 +17,7 @@ struct StructuredLoop;
 // because their fn type takes a variant type with a case for each target. So we symbolically refer to these yet-to-be dispatch nodes via their loop
 struct DispatchTarget {
     bool operator==(const DispatchTarget& rhs) const {
-        return cont == rhs.cont &&
-               entry == rhs.entry &&
-               exit == rhs.exit;
+        return cont == rhs.cont &&entry == rhs.entry &&exit == rhs.exit;
     }
 
     Continuation* cont = nullptr;
@@ -50,8 +48,14 @@ struct StructuredLoop {
     std::vector<DispatchTarget> inner_destinations = {};
     std::vector<DispatchTarget> outer_destinations = {};
 
+    // Created to serve as codegen helpers
     Continuation* new_header = nullptr;
     Continuation* new_epilogue = nullptr;
+    Continuation* new_continue = nullptr;
+
+    // Same as the inner/outer destinations, but entry/exits instead now point to the corresponding header/epilogue nodes
+    std::vector<const Continuation*> header_destination_conts;
+    std::vector<const Continuation*> epilogue_destination_conts;
 
     std::vector<RewireMe> rewire;
 };
@@ -284,15 +288,17 @@ inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
         // here, parent headers need to know what they're jumping *into*
         std::vector<const Type*> dest_types;
         for (auto& target : loop.inner_destinations) {
-            const thorin::FnType* target_type;
+            const thorin::Continuation* target_cont;
             if (target.cont != nullptr) {
-                target_type = target.cont->type();
+                target_cont = target.cont;
             } else if (target.entry != nullptr) {
                 assert(target.entry->new_header != nullptr);
-                target_type = target.entry->new_header->type();
+                target_cont = target.entry->new_header;
             } else {
                 assert(false && "Header dispatches may not exit loops");
             }
+            loop.header_destination_conts.push_back(target_cont);
+            const thorin::FnType* target_type = target_cont->type();
             dest_types.emplace_back(dom_to_tuple(world, target_type));
         }
         auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
@@ -300,6 +306,7 @@ inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
             variant_type->set(i, dest_types[i]);
         auto fn_type = world.fn_type( { variant_type } );
         loop.new_header = world.continuation(fn_type, { loop.name + "_new_header"});
+        loop.new_continue = world.continuation(fn_type, { loop.name + "_new_continue"});
     }
 }
 
@@ -311,17 +318,19 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
             // here, children epilogues need to know what they're jumping *out to*
             std::vector<const Type*> dest_types;
             for (auto& target : loop.outer_destinations) {
-                const thorin::FnType* target_type;
+                const thorin::Continuation* target_cont;
                 if (target.cont != nullptr) {
-                    target_type = target.cont->type();
+                    target_cont = target.cont;
                 } else if (target.entry != nullptr) {
                     assert(target.entry->new_header != nullptr);
-                    target_type = target.entry->new_header->type();
+                    target_cont = target.entry->new_header;
                 } else {
                     assert(target.exit != nullptr);
                     assert(target.exit->new_epilogue != nullptr);
-                    target_type = target.exit->new_epilogue->type();
+                    target_cont = target.exit->new_epilogue;
                 }
+                loop.epilogue_destination_conts.push_back(target_cont);
+                const thorin::FnType* target_type = target_cont->type();
                 dest_types.emplace_back(dom_to_tuple(world, target_type));
             }
             auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
@@ -337,11 +346,23 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
     }
 }
 
+// Finishes loop headers & epilogues, and re-wires backedges and non-local jumps to go through structured CF intrinsics
 inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
         assert(ctx.rewritten_loops.find(head) != ctx.rewritten_loops.end());
         auto& loop = ctx.rewritten_loops[head];
 
+        if (head->num_cf_nodes() > 0) {
+            loop.new_epilogue->structured_loop_epilogue(loop.new_header, loop.epilogue_destination_conts);
+            loop.new_continue->structured_loop_continue(loop.new_header);
+            loop.new_header->structured_loop_header(loop.new_epilogue, loop.new_continue, loop.header_destination_conts);
+            printf("Loop %s!\n", loop.name.c_str());
+            for (auto c : loop.header_destination_conts)
+                printf("  header target: %s!\n", c->unique_name().c_str());
+            for (auto c : loop.epilogue_destination_conts)
+                printf("  epilogue target: %s!\n", c->unique_name().c_str());
+        }
+
         for (auto& children : head->children()) {
             rewire_loops(world, ctx, &*children);
         }
@@ -359,7 +380,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper"});
 
                 auto header_variant_type = loop.new_header->type()->op(0)->as<VariantType>();
-                wrapper->jump(loop.new_header, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
+                wrapper->jump(loop.new_continue, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
 
                 rewire.cont->unset_op(rewire.op);
                 rewire.cont->set_op(rewire.op, wrapper);
@@ -367,9 +388,18 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 printf("handling NLJ!\n");
 
                 auto& nlj = rewire.non_local_jump;
+
                 auto old_fn_type = nlj.final_destination->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper"});
 
+                printf("nlj = %s\n", wrapper->unique_name().c_str());
+                printf("src = %s\n", rewire.cont->name().c_str());
+                for (auto exit : nlj.exits)
+                    printf("exit = %s\n", exit->name.c_str());
+                for (auto enter : nlj.enters)
+                    printf("enter = %s\n", enter->name.c_str());
+                printf("dst = %s\n", nlj.final_destination->name().c_str());
+
                 const Def* argument = tuple_from_params(world, wrapper->params());
                 Continuation* first_jump = nullptr;
 
@@ -405,49 +435,6 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
 
                 rewire.cont->unset_op(rewire.op);
                 rewire.cont->set_op(rewire.op, wrapper);
-
-                // ---------------------------------------------------------------------------------------
-
-                /*// 0 = this is the first step of the path
-                // 1 = last step was to break out of a loop
-                // 2 = last step was to enter a loop
-                int last = 0;
-                StructuredLoop* prev;
-
-                auto record_step = [&](StructuredLoop* loop, DispatchTarget destination) {
-                    if (last == 0) {
-                        // nothing to do, this node isn't a dispatching one
-                    } else {
-                        if (last == 1)
-                            record_destination(prev->outer_destinations, destination);
-                        else
-                            record_destination(prev->inner_destinations, destination);
-                    }
-                };
-
-                for (auto loop : leave) {
-                    DispatchTarget destination;
-                    destination.exit = loop;
-
-                    record_step(loop, destination);
-                    last = 1;
-                    prev = loop;
-                    assert(prev != nullptr);
-                }
-                for (auto loop : enter) {
-                    DispatchTarget destination;
-                    destination.entry = loop;
-
-                    record_step(loop, destination);
-                    last = 2;
-                    prev = loop;
-                    assert(prev != nullptr);
-                }
-
-                assert(last != 0);
-                DispatchTarget destination;
-                destination.cont = dest;
-                record_step(prev, destination);*/
             }
         }
     }
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 75f298991..b52e5535e 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -252,6 +252,31 @@ void Continuation::match(const Def* val, Continuation* otherwise, Defs patterns,
     return jump(world().match(val->type(), patterns.size()), args, dbg);
 }
 
+void Continuation::structured_loop_epilogue(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
+    attributes_.intrinsic = Intrinsic::StructuredLoopMerge;
+    resize(1 + targets.size());
+    set_op(0, loop_header);
+    size_t x = 1;
+    for (auto target : targets)
+        set_op(x++, target);
+}
+
+void Continuation::structured_loop_continue(const Continuation* loop_header) {
+    attributes_.intrinsic = Intrinsic::StructuredLoopContinue;
+    resize(1);
+    set_op(0, loop_header);
+}
+
+void Continuation::structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets) {
+    attributes_.intrinsic = Intrinsic::StructuredLoopHeader;
+    resize(2 + targets.size());
+    set_op(0, loop_epilogue);
+    set_op(1, loop_continue);
+    size_t x = 2;
+    for (auto target : targets)
+        set_op(x++, target);
+}
+
 void jump_to_dropped_call(Continuation* src, Continuation* dst, const Call& call) {
     std::vector<const Def*> nargs;
     for (size_t i = 0, e = src->num_args(); i != e; ++i) {
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index bff19a11e..f7e481da7 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -84,6 +84,9 @@ enum class Intrinsic : uint8_t {
     Pipeline,                   ///< Intrinsic loop-pipelining-HLS-Backend
     Branch,                     ///< branch(cond, T, F).
     Match,                      ///< match(val, otherwise, (case1, cont1), (case2, cont2), ...)
+    StructuredLoopHeader,       ///< A header for a structured loop,
+    StructuredLoopMerge,        ///< A merge block for a structured loop,
+    StructuredLoopContinue,     ///< A continue block in a structured loop,
     PeInfo,                     ///< Partial evaluation debug info.
     EndScope                    ///< Dummy function which marks the end of a @p Scope.
 };
@@ -154,6 +157,9 @@ class Continuation : public Def {
     void jump(const Def* callee, Defs args, Debug dbg = {});
     void branch(const Def* cond, const Def* t, const Def* f, Debug dbg = {});
     void match(const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
+    void structured_loop_epilogue(const Continuation* loop_header, ArrayRef<const Continuation*> targets);
+    void structured_loop_continue(const Continuation* loop_header);
+    void structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets);
     void verify() const {
 #if THORIN_ENABLE_CHECKS
         auto c = callee_fn_type();

From 76970827fe47c2859a22bbf71b1a8ea5dfea7def Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 18 Mar 2021 16:07:58 +0100
Subject: [PATCH 036/342] more intrinsics

---
 src/thorin/be/spirv/spirv.cpp           |  2 +-
 src/thorin/be/spirv/spirv_transform.cpp |  8 +++++++-
 src/thorin/continuation.cpp             |  8 ++++----
 src/thorin/continuation.h               | 14 +++++++++-----
 4 files changed, 21 insertions(+), 11 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 746c034d7..d34a33f78 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -19,7 +19,7 @@ void CodeGen::emit(std::ostream& out) {
     builder_->capability(spv::Capability::CapabilityLinkage);
 
     structure_loops();
-    cleanup_world(world());
+    // cleanup_world(world());
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 5be442d66..6969980fb 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -151,6 +151,9 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
     } else {
         const Leaf* leaf = base->as<Leaf>();
         auto cont = leaf->cf_node()->continuation();
+        // For some nonsense reason, synthetic nodes created during scopes iteration leak in next iterations >:(
+        if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
+            return;
         for (size_t i = 0; i < cont->num_ops(); i++) {
             auto def = cont->op(i);
             if (auto dest = def->isa_continuation()) {
@@ -353,7 +356,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
         auto& loop = ctx.rewritten_loops[head];
 
         if (head->num_cf_nodes() > 0) {
-            loop.new_epilogue->structured_loop_epilogue(loop.new_header, loop.epilogue_destination_conts);
+            loop.new_epilogue->structured_loop_merge(loop.new_header, loop.epilogue_destination_conts);
             loop.new_continue->structured_loop_continue(loop.new_header);
             loop.new_header->structured_loop_header(loop.new_epilogue, loop.new_continue, loop.header_destination_conts);
             printf("Loop %s!\n", loop.name.c_str());
@@ -378,6 +381,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
 
                 auto old_fn_type = rewire.backedge->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper"});
+                wrapper->attributes_.intrinsic = Intrinsic::SCFBackEdge;
 
                 auto header_variant_type = loop.new_header->type()->op(0)->as<VariantType>();
                 wrapper->jump(loop.new_continue, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
@@ -391,6 +395,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
 
                 auto old_fn_type = nlj.final_destination->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper"});
+                wrapper->attributes_.intrinsic = Intrinsic::SCFNonLocalJump;
 
                 printf("nlj = %s\n", wrapper->unique_name().c_str());
                 printf("src = %s\n", rewire.cont->name().c_str());
@@ -443,6 +448,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
 void CodeGen::structure_loops() {
     Scope::for_each(world(), [&](const Scope& scope) {
         ScopeContext context(scope);
+        printf("top: %d\n", scope.has_free_params());
 
         const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
         tag_continuations(context, looptree.root(), nullptr);
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index b52e5535e..bcdcdec9e 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -252,8 +252,8 @@ void Continuation::match(const Def* val, Continuation* otherwise, Defs patterns,
     return jump(world().match(val->type(), patterns.size()), args, dbg);
 }
 
-void Continuation::structured_loop_epilogue(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
-    attributes_.intrinsic = Intrinsic::StructuredLoopMerge;
+void Continuation::structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
+    attributes_.intrinsic = Intrinsic::SCFLoopMerge;
     resize(1 + targets.size());
     set_op(0, loop_header);
     size_t x = 1;
@@ -262,13 +262,13 @@ void Continuation::structured_loop_epilogue(const Continuation* loop_header, Arr
 }
 
 void Continuation::structured_loop_continue(const Continuation* loop_header) {
-    attributes_.intrinsic = Intrinsic::StructuredLoopContinue;
+    attributes_.intrinsic = Intrinsic::SCFLoopContinue;
     resize(1);
     set_op(0, loop_header);
 }
 
 void Continuation::structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets) {
-    attributes_.intrinsic = Intrinsic::StructuredLoopHeader;
+    attributes_.intrinsic = Intrinsic::SCFLoopHeader;
     resize(2 + targets.size());
     set_op(0, loop_epilogue);
     set_op(1, loop_continue);
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index f7e481da7..9124536ed 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -84,10 +84,14 @@ enum class Intrinsic : uint8_t {
     Pipeline,                   ///< Intrinsic loop-pipelining-HLS-Backend
     Branch,                     ///< branch(cond, T, F).
     Match,                      ///< match(val, otherwise, (case1, cont1), (case2, cont2), ...)
-    StructuredLoopHeader,       ///< A header for a structured loop,
-    StructuredLoopMerge,        ///< A merge block for a structured loop,
-    StructuredLoopContinue,     ///< A continue block in a structured loop,
-    PeInfo,                     ///< Partial evaluation debug info.
+    SCFBegin,
+    SCFLoopHeader = SCFBegin,   ///< A header for a structured loop
+    SCFLoopMerge,               ///< A merge block for a structured loop
+    SCFLoopContinue,            ///< A continue block in a structured loop
+    SCFNonLocalJump,            ///< A non-local jump in a structured control flow graph
+    SCFBackEdge,                ///< A back edge a structured loop,
+    SCFEnd,
+    PeInfo = SCFEnd,            ///< Partial evaluation debug info.
     EndScope                    ///< Dummy function which marks the end of a @p Scope.
 };
 
@@ -157,7 +161,7 @@ class Continuation : public Def {
     void jump(const Def* callee, Defs args, Debug dbg = {});
     void branch(const Def* cond, const Def* t, const Def* f, Debug dbg = {});
     void match(const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
-    void structured_loop_epilogue(const Continuation* loop_header, ArrayRef<const Continuation*> targets);
+    void structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets);
     void structured_loop_continue(const Continuation* loop_header);
     void structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets);
     void verify() const {

From c274cbe919eb24012dd6bb686d198f6b9ca54077 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 18 Mar 2021 17:09:49 +0100
Subject: [PATCH 037/342] assert structured CF

---
 src/thorin/be/spirv/spirv.cpp           |  1 +
 src/thorin/be/spirv/spirv_transform.cpp | 27 ++++++++++++++++++++++++-
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index d34a33f78..4902750ee 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -19,6 +19,7 @@ void CodeGen::emit(std::ostream& out) {
     builder_->capability(spv::Capability::CapabilityLinkage);
 
     structure_loops();
+    structure_flow();
     // cleanup_world(world());
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 6969980fb..873e3991a 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -4,6 +4,7 @@
 #include "thorin/analyses/cfg.h"
 
 #include <algorithm>
+#include <thorin/analyses/domtree.h>
 
 namespace thorin::spirv {
 
@@ -463,7 +464,31 @@ void CodeGen::structure_loops() {
 }
 
 void CodeGen::structure_flow() {
-    // TODO
+    Scope::for_each(world(), [&](const Scope& scope) {
+        CFA cfa(scope);
+        auto& post_dom_tree = cfa.b_cfg().domtree();
+
+        for (auto def : scope.defs()) {
+            if (auto cont = def->isa_continuation()) {
+                auto cfn = cfa[cont];
+                /*if (cont->callee() == world().branch()) {
+                    printf("xd: %s\n", cont->unique_name().c_str());
+                }*/
+
+                if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
+                    continue;
+                if (cont->preds().size() <= 1)
+                    continue;
+
+                for (auto pred : cont->preds()) {
+                    auto& pred_dominators = post_dom_tree.children(cfa[cont]);
+                    // TODO insert join nodes when this assert breaks
+                    // TODO inspect postdom tree recursively
+                    assert(std::find(pred_dominators.begin(), pred_dominators.end(), cfa[pred]) != pred_dominators.end());
+                }
+            }
+        }
+    });
 }
 
 }
\ No newline at end of file

From 5d16699bf34314d72ac1109b865772b96d1e7a78 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 18 Mar 2021 18:12:59 +0100
Subject: [PATCH 038/342] improved convert()

---
 src/thorin/be/spirv/spirv.cpp         | 99 +++++++++++++++++++--------
 src/thorin/be/spirv/spirv.h           | 11 ++-
 src/thorin/be/spirv/spirv_builder.hpp |  9 +++
 3 files changed, 88 insertions(+), 31 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 4902750ee..d2417c3d0 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -28,20 +28,23 @@ void CodeGen::emit(std::ostream& out) {
     builder_ = nullptr;
 }
 
-SpvId CodeGen::convert(const Type* type) {
+SpvType CodeGen::convert(const Type* type) {
     if (auto spv_type = types_.lookup(type)) return *spv_type;
 
     assert(!type->isa<MemType>());
-    SpvId spv_type;
+    SpvType spv_type;
     switch (type->tag()) {
-        case PrimType_bool:                                                             spv_type = builder_->declare_bool_type(); break;
+        // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
+        // Using extensions, we could use 16 or 8-bit ints instead
+        // We can also pack them inside structures using bit-twiddling tricks, if the need arises
+        case PrimType_bool:                                                             spv_type.id = builder_->declare_bool_type(); spv_type.size = 4; spv_type.alignment = 4; break;
         case PrimType_ps8:  case PrimType_qs8:  case PrimType_pu8:  case PrimType_qu8:  assert(false && "TODO: look into capabilities to enable this");
         case PrimType_ps16: case PrimType_qs16: case PrimType_pu16: case PrimType_qu16: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps32: case PrimType_qs32:                                         spv_type = builder_->declare_int_type(32, true ); break;
-        case PrimType_pu32: case PrimType_qu32:                                         spv_type = builder_->declare_int_type(32, false); break;
+        case PrimType_ps32: case PrimType_qs32:                                         spv_type.id = builder_->declare_int_type(32, true ); spv_type.size = 4; spv_type.alignment = 4; break;
+        case PrimType_pu32: case PrimType_qu32:                                         spv_type.id = builder_->declare_int_type(32, false); spv_type.size = 4; spv_type.alignment = 4; break;
         case PrimType_ps64: case PrimType_qs64: case PrimType_pu64: case PrimType_qu64: assert(false && "TODO: look into capabilities to enable this");
         case PrimType_pf16: case PrimType_qf16:                                         assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pf32: case PrimType_qf32:                                         spv_type = builder_->declare_float_type(32); break;
+        case PrimType_pf32: case PrimType_qf32:                                         spv_type.id = builder_->declare_float_type(32); spv_type.size = 4; spv_type.alignment = 4; break;
         case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
         case Node_PtrType: {
             auto ptr = type->as<PtrType>();
@@ -54,37 +57,41 @@ SpvId CodeGen::convert(const Type* type) {
             //return types_[type] = spv_type;
         }
         case Node_DefiniteArrayType: {
-            assert(false && "TODO");
             auto array = type->as<DefiniteArrayType>();
-            //return types_[type] = spv_type;
+            SpvType element = convert(array->elem_type());
+            SpvId size = builder_->constant(convert(world().type_pu32()).id, { (uint32_t) array->dim() });
+            spv_type.id = builder_->declare_array_type(element.id, size);
+            spv_type.size = element.size * array->dim();
+            spv_type.alignment = element.alignment;
+            break;
         }
 
         case Node_ClosureType:
         case Node_FnType: {
             // extract "return" type, collect all other types
             auto fn = type->as<FnType>();
-            std::unique_ptr<SpvId> ret;
+            std::unique_ptr<SpvType> ret;
             std::vector<SpvId> ops;
             for (auto op : fn->ops()) {
                 if (op->isa<MemType>() || op == world().unit()) continue;
                 auto fn = op->isa<FnType>();
                 if (fn && !op->isa<ClosureType>()) {
                     assert(!ret && "only one 'return' supported");
-                    std::vector<SpvId> ret_types;
+                    std::vector<SpvType> ret_types;
                     for (auto fn_op : fn->ops()) {
                         if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
                         ret_types.push_back(convert(fn_op));
                     }
-                    if (ret_types.size() == 0)      ret = std::make_unique<SpvId>(builder_->void_type);
-                    else if (ret_types.size() == 1) ret = std::make_unique<SpvId>(ret_types.back());
+                    if (ret_types.empty())          ret = std::make_unique<SpvType>( SpvType { { builder_->void_type }, 0, 1} );
+                    else if (ret_types.size() == 1) ret = std::make_unique<SpvType>(ret_types.back());
                     else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
                 } else
-                    ops.push_back(convert(op));
+                    ops.push_back(convert(op).id);
             }
             assert(ret);
 
             if (type->tag() == Node_FnType) {
-                return types_[type] = builder_->declare_fn_type(ops, *ret);
+                return types_[type] = { builder_->declare_fn_type(ops, ret->id), 0, 0 };
             }
 
             assert(false && "TODO: handle closure mess");
@@ -93,24 +100,57 @@ SpvId CodeGen::convert(const Type* type) {
 
         case Node_StructType: {
             std::vector<SpvId> types;
-            for (auto elem : type->as<StructType>()->ops())
-                types.push_back(convert(elem));
-            spv_type = builder_->declare_struct_type(types);
-            // TODO debug info
+            for (auto elem : type->as<StructType>()->ops()) {
+                auto member_type = convert(elem);
+                types.push_back(member_type.id);
+                spv_type.size += member_type.size;
+
+                // TODO handle alignment for real
+                assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
+                spv_type.alignment = 4;
+            }
+            if (spv_type.size == 0)
+                spv_type.alignment = 1;
+            spv_type.id = builder_->declare_struct_type(types);
+            builder_->name(spv_type.id, type->to_string());
             break;
         }
 
         case Node_TupleType: {
             std::vector<SpvId> types;
-            for (auto elem : type->as<TupleType>()->ops())
-                types.push_back(convert(elem));
-            spv_type = builder_->declare_struct_type(types);
-            // TODO debug info
+            for (auto elem : type->as<TupleType>()->ops()){
+                auto member_type = convert(elem);
+                types.push_back(member_type.id);
+                spv_type.size += member_type.size;
+
+                // TODO handle alignment for real
+                assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
+                spv_type.alignment = 4;
+            }
+            if (spv_type.size == 0)
+                spv_type.alignment = 1;
+            spv_type.id = builder_->declare_struct_type(types);
+            builder_->name(spv_type.id, type->to_string());
             break;
         }
 
         case Node_VariantType: {
-            assert(false && "TODO");
+            std::vector<SpvId> types;
+            types.push_back(convert(world().type_pu32()).id);
+            for (auto elem : type->as<VariantType>()->ops()){
+                auto member_type = convert(elem);
+                spv_type.size = std::max(spv_type.size, member_type.size);
+
+                // TODO handle alignment for real
+                assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
+                spv_type.alignment = 4;
+            }
+            types.push_back(convert(world().definite_array_type(world().type_pu32(), (spv_type.size + 3) / 4)).id);
+            if (spv_type.size == 0)
+                spv_type.alignment = 1;
+            spv_type.id = builder_->declare_struct_type(types);
+            builder_->name(spv_type.id, type->to_string());
+            break;
         }
 
         default:
@@ -125,7 +165,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     assert(entry_->is_returning());
 
     FnBuilder fn;
-    fn.fn_type = convert(entry_->type());
+    fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
 
     current_fn_ = &fn;
@@ -158,7 +198,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     auto param_t = convert(param->type());
                     fn.header.op(spv::Op::OpFunctionParameter, 3);
                     auto id = builder_->generate_fresh_id();
-                    fn.header.ref_id(param_t);
+                    fn.header.ref_id(param_t.id);
                     fn.header.ref_id(id);
                     fn.params[param] = id;
                 }
@@ -171,7 +211,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    bb->phis[param] = { convert(param->type()), builder_->generate_fresh_id(), {} };
+                    bb->phis[param] = { convert(param->type()).id, builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -196,7 +236,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
         if (op->isa<MemType>() || is_type_unit(op))
             continue;
         assert(op->order() == 0);
-        types.push_back(convert(op));
+        types.push_back(convert(op).id);
     }
     if (types.empty())
         return builder_->void_type;
@@ -334,7 +374,8 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
         SpvId rhs = emit(bin->rhs(), bb);
-        SpvId result_type = convert(def->type());
+        SpvType result_types = convert(def->type());
+        SpvId result_type = result_types.id;
 
         if (auto cmp = bin->isa<Cmp>()) {
             auto type = cmp->lhs()->type();
@@ -436,7 +477,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
-        auto type = convert(def->type());
+        auto type = convert(def->type()).id;
         SpvId constant;
         switch (primlit->primtype_tag()) {
             case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 89d4a0451..61dae73de 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -10,6 +10,13 @@ namespace thorin::spirv {
 
 using SpvId = builder::SpvId;
 
+struct SpvType {
+    SpvId id;
+    size_t size = 0;
+    // TODO: Alignment rules are complicated and client API dependant
+    size_t alignment = 0;
+};
+
 struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     explicit BasicBlockBuilder(builder::SpvFileBuilder& file_builder)
     : builder::SpvBasicBlockBuilder(file_builder)
@@ -34,7 +41,7 @@ class CodeGen : public thorin::CodeGen {
     void structure_loops();
     void structure_flow();
 
-    SpvId convert(const Type*);
+    SpvType convert(const Type*);
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     SpvId emit(const Def* def, BasicBlockBuilder* bb);
@@ -45,7 +52,7 @@ class CodeGen : public thorin::CodeGen {
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
     Scheduler scheduler_;
-    TypeMap<SpvId> types_;
+    TypeMap<SpvType> types_;
     DefMap<SpvId> defs_;
 };
 
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index fb19b4af6..725ec466d 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -170,6 +170,15 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId declare_array_type(SpvId element_type, SpvId dim) {
+        types_constants.op(spv::Op::OpTypeArray, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(element_type);
+        types_constants.ref_id(dim);
+        return id;
+    }
+
     SpvId declare_fn_type(std::vector<SpvId>& dom, SpvId codom) {
         types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
         auto id = generate_fresh_id();

From ea29fd7c9218e83861e1c07e5e0c0a3075db6804 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 19 Mar 2021 16:09:27 +0100
Subject: [PATCH 039/342] WIP loop emission

---
 src/thorin/be/spirv/spirv.cpp         | 122 +++++++++++++++++++++++---
 src/thorin/be/spirv/spirv.h           |  15 +++-
 src/thorin/be/spirv/spirv_builder.hpp |  31 +++++--
 3 files changed, 142 insertions(+), 26 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index d2417c3d0..89d60f405 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -135,6 +135,7 @@ SpvType CodeGen::convert(const Type* type) {
         }
 
         case Node_VariantType: {
+            assert(type->num_ops() > 0 && "empty variants not supported");
             std::vector<SpvId> types;
             types.push_back(convert(world().type_pu32()).id);
             for (auto elem : type->as<VariantType>()->ops()){
@@ -145,9 +146,12 @@ SpvType CodeGen::convert(const Type* type) {
                 assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
                 spv_type.alignment = 4;
             }
-            types.push_back(convert(world().definite_array_type(world().type_pu32(), (spv_type.size + 3) / 4)).id);
+            spv_type.variant_data_size = (spv_type.size + 3) / 4;
             if (spv_type.size == 0)
                 spv_type.alignment = 1;
+            else
+                types.push_back(convert(world().definite_array_type(world().type_pu32(), spv_type.variant_data_size)).id);
+            spv_type.variant_trivial = type->num_ops() == 1;
             spv_type.id = builder_->declare_struct_type(types);
             builder_->name(spv_type.id, type->to_string());
             break;
@@ -165,6 +169,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     assert(entry_->is_returning());
 
     FnBuilder fn;
+    fn.file_builder = builder_;
     fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
 
@@ -172,20 +177,18 @@ void CodeGen::emit(const thorin::Scope& scope) {
 
     auto conts = schedule(scope);
 
+    fn.bbs_to_emit.reserve(conts.size());
     fn.bbs.reserve(conts.size());
-    std::vector<BasicBlockBuilder> bbs;
-    bbs.reserve(conts.size());
+    auto& bbs = fn.bbs;
 
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
 
-        BasicBlockBuilder* bb = &bbs.emplace_back(BasicBlockBuilder(*builder_));
-        fn.bbs.emplace_back(bb);
+        BasicBlockBuilder* bb = &bbs.emplace_back(fn);
+        fn.bbs_to_emit.emplace_back(bb);
         auto [i, b] = fn.bbs_map.emplace(cont, bb);
         assert(b);
 
-        bb->label = builder_->generate_fresh_id();
-
         if (debug())
             builder_->name(bb->label, cont->name().c_str());
         fn.labels.emplace(cont, bb->label);
@@ -211,7 +214,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    bb->phis[param] = { convert(param->type()).id, builder_->generate_fresh_id(), {} };
+                    bb->phis_map[param] = { convert(param->type()).id, builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -225,6 +228,12 @@ void CodeGen::emit(const thorin::Scope& scope) {
         assert(cont == entry_ || cont->is_basicblock());
         emit_epilogue(cont, fn.bbs_map[cont]);
     }
+
+    for(auto& bb : fn.bbs) {
+        for (auto& [param, phi] : bb.phis_map) {
+            bb.phis.emplace_back(&phi);
+        }
+    }
     
     builder_->define_function(fn);
 }
@@ -283,18 +292,76 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     } else if (continuation->callee()->isa<Bottom>()) {
         irbuilder.CreateUnreachable();
     } */
-    else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+    else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
+        auto merge_label = current_fn_->bbs_map[continuation->op(0)->as_continuation()]->label;
+        auto continue_label = current_fn_->bbs_map[continuation->op(1)->as_continuation()]->label;
+        bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
+
+        BasicBlockBuilder* dispatch_bb = &current_fn_->bbs.emplace_back(*current_fn_);
+        current_fn_->bbs_to_emit.emplace_back(dispatch_bb);
+        builder_->name(dispatch_bb->label, "inner_dispatch");
+
+        bb->branch(dispatch_bb->label);
+        int targets = continuation->num_ops() - 2;
+        assert(targets > 0);
+
+        assert(targets == 1);
+        auto callee = continuation->op(2)->as_continuation();
+        // Extract the relevant variant & expand the tuple if necessary
+        auto arg = world().variant_extract(continuation->param(0), 0);
+        bb->args[arg] = emit(arg, bb);
+
+        if (callee->param(0)->type()->equal(arg->type())) {
+            auto* param = callee->param(0);
+            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
+            phi.preds.emplace_back(*bb->args[arg], bb->label);
+        } else {
+            assert(false && "TODO destructure argument");
+        }
+
+        dispatch_bb->branch(current_fn_->bbs_map[callee]->label);
+
+    } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
+        auto loop_header =continuation->op(0)->as_continuation();
+        auto header_label = current_fn_->bbs_map[loop_header]->label;
+
+        auto arg = continuation->param(0);
+        bb->args[arg] = emit(arg, bb);
+        auto* param = loop_header->param(0);
+        auto& phi = current_fn_->bbs_map[loop_header]->phis_map[param];
+        phi.preds.emplace_back(*bb->args[arg], *current_fn_->labels[continuation]);
+
+        bb->branch(header_label);
+    } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
+        auto header_cont = continuation->op(0)->as_continuation();
+
+        int targets = continuation->num_ops() - 1;
+        assert(targets > 0);
+
+        assert(targets == 1);
+        auto callee = continuation->op(1)->as_continuation();
+        // TODO phis
+        bb->branch(current_fn_->bbs_map[callee]->label);
+    } /*else if (continuation->intrinsic() == Intrinsic::SCFNonLocalJump) {
+        auto header_cont = continuation->op(0)->as_continuation();
+        // TODO setup arguments & stuff
+        bb->branch(current_fn_->bbs_map[continuation->op(0)->as_continuation()]->label);
+    } else if (continuation->intrinsic() == Intrinsic::SCFBackEdge) {
+        // TODO setup arguments & stuff
+        bb->branch(current_fn_->bbs_map[continuation->op(0)->as_continuation()]->label);
+    } */ else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
         int index = -1;
         for (auto& arg : continuation->args()) {
             index++;
             if (is_mem(arg) || is_unit(arg)) continue;
             bb->args[arg] = emit(arg, bb);
             auto* param = callee->param(index);
-            auto& phi = current_fn_->bbs_map[callee]->phis[param];
+            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
             phi.preds.emplace_back(*bb->args[arg], *current_fn_->labels[continuation]);
         }
         bb->branch(*current_fn_->labels[callee]);
-    } /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+    }
+    /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         auto ret_continuation = emit_intrinsic(irbuilder, continuation);
         irbuilder.CreateBr(cont2bb(ret_continuation));
     } else { // function/closure call
@@ -494,17 +561,46 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
         }
         return constant;
-    } else if(auto param = def->isa<Param>()) {
+    } else if (auto param = def->isa<Param>()) {
         if (auto param_id = current_fn_->params.lookup(param)) {
             assert((*param_id).id != 0);
             return *param_id;
         } else {
-            auto val = (*current_fn_->bbs_map[param->continuation()]).phis[param].value;
+            auto val = (*current_fn_->bbs_map[param->continuation()]).phis_map[param].value;
             assert(val.id != 0);
             return val;
         }
+    } else if (auto variant = def->isa<Variant>()) {
+        auto type = convert(def->type());
+        assert(type.variant_trivial); // TODO !
+        return emit(variant->value(), bb);
+    } else if (auto vextract = def->isa<VariantExtract>()) {
+        auto type = convert(def->op(0)->type());
+        assert(type.variant_trivial); // TODO !
+        return emit(vextract->value(), bb);
+    } else if (auto vindex = def->isa<VariantIndex>()) {
+        assert(false && "Missing variant index");
+    } else if (auto tuple = def->isa<Tuple>()) {
+        std::vector<SpvId> elements;
+        size_t x = 0;
+        for (auto& e : tuple->ops()) {
+            elements[x++] = emit(e, bb);
+        }
+        return bb->composite(convert(tuple->type()).id, elements);
+    } else if (auto structagg = def->isa<StructAgg>()) {
+        std::vector<SpvId> elements;
+        size_t x = 0;
+        for (auto& e : structagg->ops()) {
+            elements[x++] = emit(e, bb);
+        }
+        return bb->composite(convert(structagg->type()).id, elements);
     }
     assertf(false, "Incomplete emit(def) definition");
 }
 
+BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
+: builder::SpvBasicBlockBuilder(*fn_builder.file_builder) {
+    label = file_builder.generate_fresh_id();
+}
+
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 61dae73de..f552e75de 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -13,20 +13,27 @@ using SpvId = builder::SpvId;
 struct SpvType {
     SpvId id;
     size_t size = 0;
+
     // TODO: Alignment rules are complicated and client API dependant
     size_t alignment = 0;
+
+    // Only set for variant types
+    bool variant_trivial = false;
+    size_t variant_data_size = -1;
 };
 
+struct FnBuilder;
+
 struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
-    explicit BasicBlockBuilder(builder::SpvFileBuilder& file_builder)
-    : builder::SpvBasicBlockBuilder(file_builder)
-    {}
+    explicit BasicBlockBuilder(FnBuilder& fn_builder);
 
-    std::unordered_map<const Param*, Phi> phis;
+    std::unordered_map<const Param*, Phi> phis_map;
     DefMap<SpvId> args;
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {
+    builder::SpvFileBuilder* file_builder;
+    std::vector<BasicBlockBuilder> bbs;
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 725ec466d..ca0168938 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -74,11 +74,11 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         SpvId value;
         std::vector<std::pair<SpvId, SpvId>> preds;
     };
-    std::vector<Phi> phis;
+    std::vector<Phi*> phis;
     SpvId label;
 
     SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
-        op(spv::Op::OpLabel, 3 + elements.size());
+        op(spv::Op::OpCompositeConstruct, 3 + elements.size());
         ref_id(aggregate_t);
         auto id = generate_fresh_id();
         ref_id(id);
@@ -109,6 +109,16 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         ref_id(false_target);
     }
 
+    void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
+        op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
+        ref_id(merge_bb);
+        ref_id(continue_bb);
+        literal_int(loop_control);
+
+        for (auto e : loop_control_ops)
+            literal_int(e);
+    }
+
     void return_void() {
         op(spv::Op::OpReturn, 1);
     }
@@ -126,7 +136,7 @@ struct SpvFnBuilder {
 public:
     SpvId fn_type;
     SpvId fn_ret_type;
-    std::vector<SpvBasicBlockBuilder*> bbs;
+    std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
 
     // Contains OpFunctionParams
     SpvSectionBuilder header;
@@ -136,8 +146,9 @@ struct SpvFileBuilder {
     SpvFileBuilder()
     : void_type(declare_void_type())
     {}
+    SpvFileBuilder(const SpvFileBuilder&) = delete;
 
-    SpvId generate_fresh_id() { return {bound++ }; }
+    SpvId generate_fresh_id() { return { bound++ }; }
 
     void name(SpvId id, std::string_view str) {
         assert(id.id < bound);
@@ -228,15 +239,17 @@ struct SpvFileBuilder {
         for (auto w : fn_builder.header.data_)
             fn_defs.data_.push_back(w);
 
-        for (auto& bb : fn_builder.bbs) {
+        for (auto& bb : fn_builder.bbs_to_emit) {
             fn_defs.op(spv::Op::OpLabel, 2);
             fn_defs.ref_id(bb->label);
 
             for (auto& phi : bb->phis) {
-                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi.preds.size());
-                fn_defs.ref_id(phi.type);
-                fn_defs.ref_id(phi.value);
-                for (auto& [pred_value, pred_label] : phi.preds) {
+                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
+                fn_defs.ref_id(phi->type);
+                fn_defs.ref_id(phi->value);
+                printf("Phi %d\n", phi->value);
+                assert(phi->preds.size() > 0);
+                for (auto& [pred_value, pred_label] : phi->preds) {
                     fn_defs.ref_id(pred_value);
                     fn_defs.ref_id(pred_label);
                 }

From f2416ee57338a23c7332823c193b385632bf6dfc Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 19 Mar 2021 18:02:05 +0100
Subject: [PATCH 040/342] spirv doesn't like bitcasted allocas

---
 src/thorin/be/spirv/spirv.cpp         | 55 ++++++++++++++++++++++-----
 src/thorin/be/spirv/spirv_builder.hpp | 54 ++++++++++++++++++++++++++
 src/thorin/type.h                     |  2 +
 3 files changed, 101 insertions(+), 10 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 89d60f405..e63e2f561 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -48,7 +48,16 @@ SpvType CodeGen::convert(const Type* type) {
         case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
         case Node_PtrType: {
             auto ptr = type->as<PtrType>();
-            assert(false && "TODO");
+            spv::StorageClass storage_class;
+            switch (ptr->addr_space()) {
+                case AddrSpace::Function: storage_class = spv::StorageClassFunction; break;
+                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;  break;
+                default:
+                    assert(false && "This address space is not supported");
+                    break;
+            }
+            SpvType element = convert(ptr->pointee());
+            spv_type.id = builder_->declare_ptr_type(storage_class, element.id);
             break;
         }
         case Node_IndefiniteArrayType: {
@@ -300,7 +309,6 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         BasicBlockBuilder* dispatch_bb = &current_fn_->bbs.emplace_back(*current_fn_);
         current_fn_->bbs_to_emit.emplace_back(dispatch_bb);
         builder_->name(dispatch_bb->label, "inner_dispatch");
-
         bb->branch(dispatch_bb->label);
         int targets = continuation->num_ops() - 2;
         assert(targets > 0);
@@ -309,12 +317,12 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto callee = continuation->op(2)->as_continuation();
         // Extract the relevant variant & expand the tuple if necessary
         auto arg = world().variant_extract(continuation->param(0), 0);
-        bb->args[arg] = emit(arg, bb);
+        auto extracted = emit(arg, dispatch_bb);
 
         if (callee->param(0)->type()->equal(arg->type())) {
             auto* param = callee->param(0);
             auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
-            phi.preds.emplace_back(*bb->args[arg], bb->label);
+            phi.preds.emplace_back(extracted, dispatch_bb->label);
         } else {
             assert(false && "TODO destructure argument");
         }
@@ -572,14 +580,41 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto variant = def->isa<Variant>()) {
         auto type = convert(def->type());
-        assert(type.variant_trivial); // TODO !
-        return emit(variant->value(), bb);
+        auto value = emit(variant->value(), bb);
+        std::vector<SpvId> elements;
+        elements.emplace_back(builder_->constant(convert(world().type_pu32()).id, { (uint32_t) variant->index() }));
+        if (type.variant_data_size > 0) {
+            auto variant_payload_type = world().definite_array_type(world().type_pu32(), type.variant_data_size);
+            auto payload_ptr_type = world().ptr_type(variant_payload_type, 1, -1, AddrSpace::Function);
+            auto alloca = bb->variable(convert(payload_ptr_type).id, spv::StorageClass::StorageClassFunction);
+
+            auto casted_ptr_type = world().ptr_type(variant->value()->type(), 1, -1, AddrSpace::Function);
+            auto casted = bb->bitcast(convert(casted_ptr_type).id, alloca);
+
+            bb->store(value, casted);
+            elements.push_back(bb->load(convert(variant_payload_type).id, alloca));
+        }
+        return bb->composite(convert(variant->type()).id, elements);
     } else if (auto vextract = def->isa<VariantExtract>()) {
-        auto type = convert(def->op(0)->type());
-        assert(type.variant_trivial); // TODO !
-        return emit(vextract->value(), bb);
+        auto type = convert(def->type());
+        assert(type.variant_data_size > 0); // TODO is it legal to extract () ?
+        if (type.variant_data_size > 0) {
+            auto payload = bb->extract(convert(world().type_pu32()).id, emit(vextract->value(), bb), {1});
+
+            auto variant_payload_type = world().definite_array_type(world().type_pu32(), type.variant_data_size);
+            auto payload_ptr_type = world().ptr_type(variant_payload_type, 1, -1, AddrSpace::Function);
+            auto alloca = bb->variable(convert(payload_ptr_type).id, spv::StorageClass::StorageClassFunction);
+
+            auto casted_ptr_type = world().ptr_type(def->type(), 1, -1, AddrSpace::Function);
+            auto casted = bb->bitcast(convert(casted_ptr_type).id, alloca);
+
+            bb->store(payload, alloca);
+            return bb->load(convert(def->type()).id, casted);
+        }
+        THORIN_UNREACHABLE;
     } else if (auto vindex = def->isa<VariantIndex>()) {
-        assert(false && "Missing variant index");
+        auto value = emit(vindex->op(0), bb);
+        return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
         std::vector<SpvId> elements;
         size_t x = 0;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index ca0168938..adc34fe01 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -87,6 +87,51 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
+        op(spv::Op::OpCompositeExtract, 4 + indices.size());
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(composite);
+        for (auto i : indices)
+            literal_int(i);
+        return id;
+    }
+
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        op(spv::Op::OpVariable, 4);
+        ref_id(type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        literal_int(storage_class);
+        return id;
+    }
+
+    SpvId bitcast(SpvId target_type, SpvId value) {
+        op(spv::Op::OpBitcast, 4);
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(value);
+        return id;
+    }
+
+    SpvId load(SpvId target_type, SpvId pointer) {
+        op(spv::Op::OpLoad, 4);
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(pointer);
+        return id;
+    }
+
+    void store(SpvId value, SpvId pointer) {
+        op(spv::Op::OpStore, 3);
+        auto id = generate_fresh_id();
+        ref_id(pointer);
+        ref_id(value);
+    }
+
     SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
         op(op_, 5);
         auto id = generate_fresh_id();
@@ -181,6 +226,15 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
+        types_constants.op(spv::Op::OpTypePointer, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
+        types_constants.ref_id(element_type);
+        return id;
+    }
+
     SpvId declare_array_type(SpvId element_type, SpvId dim) {
         types_constants.op(spv::Op::OpTypeArray, 4);
         auto id = generate_fresh_id();
diff --git a/src/thorin/type.h b/src/thorin/type.h
index d8fbe7138..815f3e902 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -233,6 +233,8 @@ enum class AddrSpace : uint32_t {
     Texture  = 2,
     Shared   = 3,
     Constant = 4,
+    Private =  5, // Corresponds to the 'private' storage class in SPIR-V
+    Function = 6, // Corresponds to the 'function' storage class in SPIR-V
 };
 
 /// Pointer type.

From 456c3eb1da51397947f038a515f3dfdd62f9c248 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 22 Mar 2021 17:56:04 +0100
Subject: [PATCH 041/342] doing variants the stupid way for now

---
 src/thorin/be/spirv/spirv.cpp         | 64 ++++++++++++---------------
 src/thorin/be/spirv/spirv.h           |  4 +-
 src/thorin/be/spirv/spirv_builder.hpp |  8 ++++
 3 files changed, 37 insertions(+), 39 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index e63e2f561..3754b4646 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -145,23 +145,23 @@ SpvType CodeGen::convert(const Type* type) {
 
         case Node_VariantType: {
             assert(type->num_ops() > 0 && "empty variants not supported");
-            std::vector<SpvId> types;
-            types.push_back(convert(world().type_pu32()).id);
+            std::vector<SpvId> payload_type;
             for (auto elem : type->as<VariantType>()->ops()){
                 auto member_type = convert(elem);
-                spv_type.size = std::max(spv_type.size, member_type.size);
+                payload_type.push_back(member_type.id);
+                spv_type.size += member_type.size;
 
                 // TODO handle alignment for real
                 assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
                 spv_type.alignment = 4;
             }
-            spv_type.variant_data_size = (spv_type.size + 3) / 4;
             if (spv_type.size == 0)
                 spv_type.alignment = 1;
-            else
-                types.push_back(convert(world().definite_array_type(world().type_pu32(), spv_type.variant_data_size)).id);
-            spv_type.variant_trivial = type->num_ops() == 1;
-            spv_type.id = builder_->declare_struct_type(types);
+            spv_type.payload_id = builder_->declare_struct_type(payload_type);
+            builder_->name(spv_type.payload_id, type->to_string() + "_payload");
+
+            std::vector<SpvId> with_tag = { convert(world().type_pu32()).id, spv_type.payload_id};
+            spv_type.id = builder_->declare_struct_type(with_tag);
             builder_->name(spv_type.id, type->to_string());
             break;
         }
@@ -579,44 +579,35 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             return val;
         }
     } else if (auto variant = def->isa<Variant>()) {
-        auto type = convert(def->type());
-        auto value = emit(variant->value(), bb);
+        auto struct_type = def->type()->as<VariantType>();
+        auto type = convert(struct_type);
         std::vector<SpvId> elements;
-        elements.emplace_back(builder_->constant(convert(world().type_pu32()).id, { (uint32_t) variant->index() }));
-        if (type.variant_data_size > 0) {
-            auto variant_payload_type = world().definite_array_type(world().type_pu32(), type.variant_data_size);
-            auto payload_ptr_type = world().ptr_type(variant_payload_type, 1, -1, AddrSpace::Function);
-            auto alloca = bb->variable(convert(payload_ptr_type).id, spv::StorageClass::StorageClassFunction);
-
-            auto casted_ptr_type = world().ptr_type(variant->value()->type(), 1, -1, AddrSpace::Function);
-            auto casted = bb->bitcast(convert(casted_ptr_type).id, alloca);
-
-            bb->store(value, casted);
-            elements.push_back(bb->load(convert(variant_payload_type).id, alloca));
+        elements.resize(struct_type->num_ops());
+        size_t x = 0;
+        for (auto& e : struct_type->ops()) {
+            if (x == variant->index())
+                elements[x] = emit(variant->value(), bb);
+            else
+                elements[x] = bb->undef(convert(e).id);
+            x++;
         }
-        return bb->composite(convert(variant->type()).id, elements);
+        auto payload = bb->composite(convert(variant->type()).payload_id, elements);
+        auto tag = builder_->constant(convert(world().type_pu32()).id, { static_cast<uint32_t>(variant->index()) });
+        std::vector<SpvId> with_tag = { tag, payload };
+        return bb->composite(convert(variant->type()).id, with_tag);
     } else if (auto vextract = def->isa<VariantExtract>()) {
-        auto type = convert(def->type());
-        assert(type.variant_data_size > 0); // TODO is it legal to extract () ?
-        if (type.variant_data_size > 0) {
-            auto payload = bb->extract(convert(world().type_pu32()).id, emit(vextract->value(), bb), {1});
+        auto variant_type = vextract->value()->type()->as<VariantType>();
 
-            auto variant_payload_type = world().definite_array_type(world().type_pu32(), type.variant_data_size);
-            auto payload_ptr_type = world().ptr_type(variant_payload_type, 1, -1, AddrSpace::Function);
-            auto alloca = bb->variable(convert(payload_ptr_type).id, spv::StorageClass::StorageClassFunction);
+        auto target_type = convert(def->type());
+        auto payload = bb->extract(convert(variant_type).payload_id, emit(vextract->value(), bb), {1});
 
-            auto casted_ptr_type = world().ptr_type(def->type(), 1, -1, AddrSpace::Function);
-            auto casted = bb->bitcast(convert(casted_ptr_type).id, alloca);
-
-            bb->store(payload, alloca);
-            return bb->load(convert(def->type()).id, casted);
-        }
-        THORIN_UNREACHABLE;
+        return bb->extract(target_type.id, payload, { static_cast<uint32_t>(vextract->index()) });
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
         return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
         std::vector<SpvId> elements;
+        elements.resize(tuple->num_ops());
         size_t x = 0;
         for (auto& e : tuple->ops()) {
             elements[x++] = emit(e, bb);
@@ -624,6 +615,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         return bb->composite(convert(tuple->type()).id, elements);
     } else if (auto structagg = def->isa<StructAgg>()) {
         std::vector<SpvId> elements;
+        elements.resize(structagg->num_ops());
         size_t x = 0;
         for (auto& e : structagg->ops()) {
             elements[x++] = emit(e, bb);
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index f552e75de..a464a9905 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -17,9 +17,7 @@ struct SpvType {
     // TODO: Alignment rules are complicated and client API dependant
     size_t alignment = 0;
 
-    // Only set for variant types
-    bool variant_trivial = false;
-    size_t variant_data_size = -1;
+    SpvId payload_id;
 };
 
 struct FnBuilder;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index adc34fe01..cd6318971 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -77,6 +77,14 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     std::vector<Phi*> phis;
     SpvId label;
 
+    SpvId undef(SpvId type) {
+        op(spv::Op::OpUndef, 3);
+        ref_id(type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        return id;
+    }
+
     SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
         op(spv::Op::OpCompositeConstruct, 3 + elements.size());
         ref_id(aggregate_t);

From e6bffda09f3b77c108a5726894aed2ab80f8fa6e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Wed, 24 Mar 2021 12:19:01 +0100
Subject: [PATCH 042/342] hacky but passes validation (needs cleanup)

---
 src/thorin/analyses/cfg.cpp             |  6 +++
 src/thorin/be/spirv/spirv.cpp           | 50 +++++++++++++++++++++----
 src/thorin/be/spirv/spirv.h             |  4 +-
 src/thorin/be/spirv/spirv_builder.hpp   |  6 +++
 src/thorin/be/spirv/spirv_transform.cpp |  6 ++-
 src/thorin/continuation.cpp             | 12 ++++++
 src/thorin/continuation.h               |  1 +
 7 files changed, 73 insertions(+), 12 deletions(-)

diff --git a/src/thorin/analyses/cfg.cpp b/src/thorin/analyses/cfg.cpp
index 8f5361c5f..636232734 100644
--- a/src/thorin/analyses/cfg.cpp
+++ b/src/thorin/analyses/cfg.cpp
@@ -61,6 +61,12 @@ CFA::CFA(const Scope& scope)
 
         while (!queue.empty()) {
             auto def = pop(queue);
+            // Hacky ?
+            if (auto cont = def->isa_continuation()) {
+                for (auto op : cont->potential_succs())
+                    enqueue(op);
+                continue;
+            }
             for (auto op : def->ops())
                 enqueue(op);
         }
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3754b4646..3cca081bb 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -1,7 +1,8 @@
 #include "thorin/be/spirv/spirv.h"
+
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/schedule.h"
-
+#include "thorin/analyses/domtree.h"
 #include "thorin/transform/cleanup_world.h"
 
 #include <iostream>
@@ -173,11 +174,36 @@ SpvType CodeGen::convert(const Type* type) {
     return types_[type] = spv_type;
 }
 
+inline Schedule schedule_structured(const Scope& scope) {
+    // until we have sth better simply use the RPO of the CFG
+    Schedule result;
+    for (auto n : scope.f_cfg().reverse_post_order())
+        result.emplace_back(n->continuation());
+
+    auto schedule = [&](const Continuation* cont) {
+        printf("scheduled: %s\n", cont->unique_name().c_str());
+    };
+
+    auto visit = [&](const Continuation* cont) {
+        if (cont->intrinsic() == Intrinsic::SCFLoopHeader) {
+            // Write continue block FIRST
+            schedule(cont->op(1)->as_continuation());
+            // Then write the header
+            schedule(cont);
+
+            schedule(cont->op(0)->as_continuation());
+        }
+    };
+
+    return result;
+}
+
 void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
     FnBuilder fn;
+    fn.scope = &scope;
     fn.file_builder = builder_;
     fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
@@ -229,9 +255,6 @@ void CodeGen::emit(const thorin::Scope& scope) {
         }
     }
 
-    Scheduler new_scheduler(scope);
-    swap(scheduler_, new_scheduler);
-
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
         assert(cont == entry_ || cont->is_basicblock());
@@ -282,10 +305,20 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
     }
     else if (continuation->callee() == world().branch()) {
+        auto& domtree = current_fn_->scope->b_cfg().domtree();
+        auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
+
+        printf("Merge @%s\n", merge_cont->unique_name().c_str());
+        /*BasicBlockBuilder* merge_bb = &current_fn_->bbs.emplace_back(*current_fn_);
+        auto merge_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), merge_cont);
+        current_fn_->bbs_to_emit.emplace(merge_bb_location + 1, merge_bb);
+        builder_->name(merge_bb->label, "merge_" + merge_cont->name());*/
+
         auto cond = emit(continuation->arg(0), bb);
         bb->args[continuation->arg(0)] = cond;
         auto tbb = *current_fn_->labels[continuation->arg(1)->as_continuation()];
         auto fbb = *current_fn_->labels[continuation->arg(2)->as_continuation()];
+        bb->selection_merge(*current_fn_->labels[merge_cont],spv::SelectionControlMaskNone);
         bb->branch_conditional(cond, tbb, fbb);
     } /*else if (continuation->callee()->isa<Continuation>() &&
                continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
@@ -307,8 +340,11 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
 
         BasicBlockBuilder* dispatch_bb = &current_fn_->bbs.emplace_back(*current_fn_);
-        current_fn_->bbs_to_emit.emplace_back(dispatch_bb);
-        builder_->name(dispatch_bb->label, "inner_dispatch");
+
+        auto header_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), bb);
+
+        current_fn_->bbs_to_emit.emplace(header_bb_location + 1, dispatch_bb);
+        builder_->name(dispatch_bb->label, "dispatch_" + continuation->name());
         bb->branch(dispatch_bb->label);
         int targets = continuation->num_ops() - 2;
         assert(targets > 0);
@@ -330,7 +366,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         dispatch_bb->branch(current_fn_->bbs_map[callee]->label);
 
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
-        auto loop_header =continuation->op(0)->as_continuation();
+        auto loop_header = continuation->op(0)->as_continuation();
         auto header_label = current_fn_->bbs_map[loop_header]->label;
 
         auto arg = continuation->param(0);
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index a464a9905..4e85a426b 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -4,8 +4,6 @@
 #include "thorin/be/spirv/spirv_builder.hpp"
 #include "thorin/be/backends.h"
 
-#include "thorin/analyses/schedule.h"
-
 namespace thorin::spirv {
 
 using SpvId = builder::SpvId;
@@ -30,6 +28,7 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {
+    const Scope* scope;
     builder::SpvFileBuilder* file_builder;
     std::vector<BasicBlockBuilder> bbs;
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
@@ -56,7 +55,6 @@ class CodeGen : public thorin::CodeGen {
     builder::SpvFileBuilder* builder_ = nullptr;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
-    Scheduler scheduler_;
     TypeMap<SpvType> types_;
     DefMap<SpvId> defs_;
 };
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index cd6318971..cacb9f2b3 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -162,6 +162,12 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         ref_id(false_target);
     }
 
+    void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
+        op(spv::Op::OpSelectionMerge, 3);
+        ref_id(merge_bb);
+        literal_int(selection_control);
+    }
+
     void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
         op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
         ref_id(merge_bb);
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 873e3991a..43ebcfe1b 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -381,7 +381,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 auto variant_index = index_of_destination(loop.inner_destinations, destination);
 
                 auto old_fn_type = rewire.backedge->type();
-                auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper"});
+                auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper_to" + destination.cont->unique_name() });
                 wrapper->attributes_.intrinsic = Intrinsic::SCFBackEdge;
 
                 auto header_variant_type = loop.new_header->type()->op(0)->as<VariantType>();
@@ -395,7 +395,7 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 auto& nlj = rewire.non_local_jump;
 
                 auto old_fn_type = nlj.final_destination->type();
-                auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper"});
+                auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper_to" + nlj.final_destination->unique_name() });
                 wrapper->attributes_.intrinsic = Intrinsic::SCFNonLocalJump;
 
                 printf("nlj = %s\n", wrapper->unique_name().c_str());
@@ -480,6 +480,8 @@ void CodeGen::structure_flow() {
                 if (cont->preds().size() <= 1)
                     continue;
 
+                printf("has more than 1 incoming branch: %s\n", cont->unique_name().c_str());
+
                 for (auto pred : cont->preds()) {
                     auto& pred_dominators = post_dom_tree.children(cfa[cont]);
                     // TODO insert join nodes when this assert breaks
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index bcdcdec9e..38d1f9663 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -15,6 +15,18 @@ const Def* Continuation::callee() const {
     return empty() ? world().bottom(world().fn_type(), debug()) : op(0);
 }
 
+Defs Continuation::potential_succs() const {
+    if (intrinsic() == Intrinsic::SCFLoopHeader)
+        return ops().skip_front(2);
+    else if (intrinsic() == Intrinsic::SCFLoopMerge)
+        return ops().skip_front();
+
+    //else if (intrinsic() == Intrinsic::SCFLoopContinue)
+    //return std::vector<const Def*> { op(0) };
+
+    return ops();
+}
+
 Continuation* Continuation::stub() const {
     Rewriter rewriter;
 
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 9124536ed..3770c63ee 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -133,6 +133,7 @@ class Continuation : public Def {
     const Param* ret_param() const;
     const Def* callee() const;
     Defs args() const { return num_ops() == 0 ? Defs(0, 0) : ops().skip_front(); }
+    Defs potential_succs() const;
     const Def* arg(size_t i) const { return args()[i]; }
     const FnType* type() const { return Def::type()->as<FnType>(); }
     const FnType* callee_fn_type() const { return callee()->type()->as<FnType>(); }

From f7e44007adf037b2a7e40d49bc9235d9c8c89664 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 25 Mar 2021 09:31:57 +0100
Subject: [PATCH 043/342] including debug stuff, commit history will be cleaned
 up afterwards

---
 src/thorin/be/spirv/spirv.cpp |   5 +
 src/thorin/util/dot_dump.cpp  | 235 ++++++++++++++++++++++++++++++++++
 2 files changed, 240 insertions(+)
 create mode 100644 src/thorin/util/dot_dump.cpp

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3cca081bb..c4b40f949 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -7,6 +7,10 @@
 
 #include <iostream>
 
+namespace thorin {
+    void dump_dot(thorin::World &world);
+}
+
 namespace thorin::spirv {
 
 CodeGen::CodeGen(thorin::World& world, Cont2Config&, bool debug)
@@ -22,6 +26,7 @@ void CodeGen::emit(std::ostream& out) {
     structure_loops();
     structure_flow();
     // cleanup_world(world());
+    dump_dot(world());
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
diff --git a/src/thorin/util/dot_dump.cpp b/src/thorin/util/dot_dump.cpp
new file mode 100644
index 000000000..21ebb0bcf
--- /dev/null
+++ b/src/thorin/util/dot_dump.cpp
@@ -0,0 +1,235 @@
+#ifndef DOT_DUMP_H
+#define DOT_DUMP_H
+
+#include "thorin/world.h"
+
+namespace thorin {
+
+/// Outputs the raw thorin IR as a graph without performing any scope or scheduling analysis
+struct DotPrinter {
+    DotPrinter(World& world, const char* filename = "world.dot") : world(world) {
+        file = std::ofstream(filename);
+    }
+
+    private:
+        void dump_def(const Def* def);
+        void dump_def_generic(const Def* def, const char* color, const char* shape);
+        void dump_literal(const Literal* cont);
+        void dump_primop(const PrimOp* cont);
+        void dump_continuation(const Continuation* cont);
+
+    #define up ""
+    #define down ""
+    #define endl "\n"
+
+    public:
+        void print() {
+            file << "digraph " << world.name() << " {" << up;
+            file << endl << "bgcolor=transparent;";
+            for (Continuation* continuation : world.continuations()) {
+                // Ignore those if they are not referenced elsewhere...
+                if (continuation == world.branch() || continuation == world.end_scope())
+                    continue;
+                dump_def(continuation);
+            }
+            file << down << endl << "}" << endl;
+        }
+
+    private:
+        thorin::World& world;
+
+        DefSet done;
+        std::ofstream file;
+};
+
+void DotPrinter::dump_def(const Def* def) {
+    if (done.contains(def))
+        return;
+
+    if (def->isa_continuation())
+        dump_continuation(def->as_continuation());
+    else if (def->isa<Literal>())
+        dump_literal(def->as<Literal>());
+    else if (def->isa<PrimOp>())
+        dump_primop(def->as<PrimOp>());
+    else if (def->isa<Param>())
+        dump_def_generic(def, "grey", "oval");
+    else
+        dump_def_generic(def, "red", "star");
+}
+
+void DotPrinter::dump_def_generic(const Def* def, const char* color, const char* shape) {
+    file << endl << def->unique_name() << " [" << up;
+
+    file << endl << "label = \"";
+
+    file << def->unique_name() << " : " << def->type()->to_string();
+
+    file << "\";";
+
+    file << endl << "shape = " << shape << ";";
+    file << endl << "color = " << color << ";";
+
+    file << down << endl << "]";
+
+    done.emplace(def);
+
+    for (size_t i = 0; i < def->num_ops(); i++) {
+        const auto& op = def->op(i);
+        dump_def(op);
+        file << endl << def->unique_name() << " -> " << op->unique_name() << " [arrowhead=vee,label=\"o" << i << "\",fontsize=8,fontcolor=grey];";
+    }
+}
+
+void DotPrinter::dump_literal(const Literal* def) {
+    file << endl << def->unique_name() << " [" << up;
+
+    file << endl << "label = \"";
+    file << def->to_string();
+    file << "\";";
+
+    file << endl << "style = dotted;";
+
+    file << down << endl << "]";
+
+    done.emplace(def);
+
+    assert(def->num_ops() == 0);
+}
+
+void DotPrinter::dump_primop(const PrimOp* def) {
+    file << endl << def->unique_name() << " [" << up;
+
+    file << endl << "label = \"";
+
+    file << def->op_name();
+
+    auto variant = def->isa<Variant>();
+    auto variant_extract = def->isa<VariantExtract>();
+    if (variant || variant_extract )
+        file << "(" << (variant ? variant->index() : variant_extract->index()) << ")";
+
+    file << "\";";
+
+    file << endl << "color = darkseagreen1;";
+    file << endl << "style = filled;";
+
+    file << down << endl << "]";
+
+    done.emplace(def);
+
+    for (size_t i = 0; i < def->num_ops(); i++) {
+        const auto& op = def->op(i);
+        dump_def(op);
+        file << endl << def->unique_name() << " -> " << op->unique_name() << " [arrowhead=vee,label=\"o" << i << "\",fontsize=8,fontcolor=grey];";
+    }
+}
+
+void DotPrinter::dump_continuation(const Continuation* cont) {
+    done.emplace(cont);
+    auto intrinsic = cont->intrinsic();
+    file << endl << cont->unique_name() << " [" << up;
+
+    file << endl << "label = \"";
+    if (cont->is_exported())
+        file << "[extern]\\n";
+    auto name = cont->name();
+    if (!cont->is_exported())
+        name = cont->unique_name();
+    file << name << "(";
+    for (size_t i = 0; i < cont->num_params(); i++) {
+        file << cont->param(i)->type()->to_string() << (i + 1 == cont->num_params() ? "" : ", ");
+    }
+    file << ")";
+    file << "\";";
+
+    file << endl << "shape = rectangle;";
+    if (intrinsic != Intrinsic::None) {
+        file << endl << "color = lightblue;";
+        file << endl << "style = filled;";
+    }
+    if (cont->is_exported()) {
+        file << endl << "color = pink;";
+        file << endl << "style = filled;";
+    }
+
+    file << down << endl << "]";
+
+    int x = 1;
+    switch (intrinsic) {
+        case Intrinsic::SCFLoopHeader:
+            dump_def(cont->op(1));
+            file << endl << cont->unique_name() << " -> " << cont->op(1)->unique_name() << " [arrowhead=none];";
+            x = 2;
+        case Intrinsic::SCFLoopContinue:
+        case Intrinsic::SCFLoopMerge:
+            dump_def(cont->op(0));
+            file << endl << cont->unique_name() << " -> " << cont->op(0)->unique_name() << " [arrowhead=none];";
+            for (size_t i = x; i < cont->num_ops(); i++) {
+                auto op = cont->op(i);
+                dump_def(op);
+                file << endl << cont->unique_name() << " -> " << op->unique_name() << " [arrowhead=normal];";
+            }
+            return;
+
+        default: break;
+    }
+
+    if (auto callee_cont = cont->callee()->isa_continuation()) {
+        switch (callee_cont->intrinsic()) {
+            case Intrinsic::Branch: {
+                auto condition = cont->arg(0);
+                dump_def(condition);
+                file << endl << condition->unique_name() << " -> " << cont->unique_name()
+                     << " [arrowhead=onormal,label=\"condition\",fontsize=8,fontcolor=grey];";
+                auto if_true = cont->arg(1);
+                dump_def(if_true);
+                file << endl << cont->unique_name() << " -> " << if_true->unique_name()
+                     << " [arrowhead=normal,label=\"if_true\",fontsize=8,fontcolor=grey];";
+                auto if_false = cont->arg(2);
+                dump_def(if_false);
+                file << endl << cont->unique_name() << " -> " << if_false->unique_name()
+                     << " [arrowhead=normal,label=\"if_false\",fontsize=8,fontcolor=grey];";
+                return;
+            }
+            default:
+                break;
+        }
+    }
+
+    for (size_t i = 0; i < cont->num_args(); i++) {
+        auto arg = cont->arg(i);
+        dump_def(arg);
+
+        if (cont->callee()->uses().size() > 1)
+            file << endl << arg->unique_name() << " -> " << cont->callee()->unique_name() << " [arrowhead=onormal,label=\"a" << i << " from " << cont->unique_name() << "\",fontsize=8,fontcolor=grey];";
+        else
+            file << endl << arg->unique_name() << " -> " << cont->callee()->unique_name() << " [arrowhead=onormal,label=\"a" << i << "\",fontsize=8,fontcolor=grey];";
+    }
+
+    switch (intrinsic) {
+        // We don't care about the params for these, or the callee
+        case Intrinsic::Match:
+        case Intrinsic::Branch:
+            return;
+        default:
+            break;
+    }
+
+    for (size_t i = 0; i < cont->num_params(); i++) {
+        auto param = cont->param(i);
+        dump_def(param);
+        file << endl << param->unique_name() << " -> " << cont->unique_name() << " [arrowhead=none,label=\"p" << i << "\",fontsize=8,fontcolor=grey];";
+    }
+    dump_def(cont->callee());
+    file << endl << cont->unique_name() << " -> " << cont->callee()->unique_name() << " [arrowhead=normal];";
+}
+
+void dump_dot(World& world) {
+    auto p = DotPrinter(world);
+    p.print();
+}
+
+}
+
+#endif //DOT_DUMP_H

From 0f585b4a7b94cd2ed32981ae3e241ea6215d3524 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 25 Mar 2021 09:48:32 +0100
Subject: [PATCH 044/342] ancestors

---
 src/thorin/be/spirv/spirv_transform.cpp | 33 ++++++++++++++++++++++---
 1 file changed, 29 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 43ebcfe1b..db0b0887d 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -463,6 +463,28 @@ void CodeGen::structure_loops() {
     });
 }
 
+template<typename Fn>
+inline void iterate_ancestors(Continuation* cont, Fn fn) {
+    ContinuationSet done;
+
+    Continuations stack;
+    stack.push_back(cont);
+    while (!stack.empty()) {
+        Continuation* top = stack.back();
+        stack.pop_back();
+        if (done.contains(top)) continue;
+        if (top != cont) fn(top);
+        done.insert(top);
+        for (auto pred : top->preds()) {
+            auto pred_cont = pred->isa_continuation();
+            if (!pred_cont) continue;
+            if (!done.contains(pred_cont)) {
+                stack.push_back(pred_cont);
+            }
+        }
+    }
+}
+
 void CodeGen::structure_flow() {
     Scope::for_each(world(), [&](const Scope& scope) {
         CFA cfa(scope);
@@ -475,19 +497,22 @@ void CodeGen::structure_flow() {
                     printf("xd: %s\n", cont->unique_name().c_str());
                 }*/
 
-                if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
-                    continue;
+                //if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
+                //    continue;
                 if (cont->preds().size() <= 1)
                     continue;
 
                 printf("has more than 1 incoming branch: %s\n", cont->unique_name().c_str());
 
-                for (auto pred : cont->preds()) {
+                /*for (auto pred : cont->preds()) {
                     auto& pred_dominators = post_dom_tree.children(cfa[cont]);
                     // TODO insert join nodes when this assert breaks
                     // TODO inspect postdom tree recursively
                     assert(std::find(pred_dominators.begin(), pred_dominators.end(), cfa[pred]) != pred_dominators.end());
-                }
+                }*/
+                iterate_ancestors(cont, [&](Continuation* ancestor) {
+                    printf("  ancestor: %s\n", ancestor->unique_name().c_str());
+                });
             }
         }
     });

From 270147c07338ba433cd6dcb603c5ee1257846aef Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 29 Mar 2021 10:30:44 +0200
Subject: [PATCH 045/342] some exp stuff

---
 src/thorin/analyses/cfg.cpp             |   6 --
 src/thorin/be/spirv/spirv.cpp           |  14 +--
 src/thorin/be/spirv/spirv_transform.cpp | 110 +++++++++++++++++++++---
 src/thorin/continuation.cpp             |  26 ++----
 src/thorin/continuation.h               |  16 +++-
 src/thorin/util/dot_dump.cpp            |  12 ++-
 6 files changed, 132 insertions(+), 52 deletions(-)

diff --git a/src/thorin/analyses/cfg.cpp b/src/thorin/analyses/cfg.cpp
index 636232734..8f5361c5f 100644
--- a/src/thorin/analyses/cfg.cpp
+++ b/src/thorin/analyses/cfg.cpp
@@ -61,12 +61,6 @@ CFA::CFA(const Scope& scope)
 
         while (!queue.empty()) {
             auto def = pop(queue);
-            // Hacky ?
-            if (auto cont = def->isa_continuation()) {
-                for (auto op : cont->potential_succs())
-                    enqueue(op);
-                continue;
-            }
             for (auto op : def->ops())
                 enqueue(op);
         }
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c4b40f949..324b41133 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -340,8 +340,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         irbuilder.CreateUnreachable();
     } */
     else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
-        auto merge_label = current_fn_->bbs_map[continuation->op(0)->as_continuation()]->label;
-        auto continue_label = current_fn_->bbs_map[continuation->op(1)->as_continuation()]->label;
+        auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
+        auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
         bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
 
         BasicBlockBuilder* dispatch_bb = &current_fn_->bbs.emplace_back(*current_fn_);
@@ -351,11 +351,11 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         current_fn_->bbs_to_emit.emplace(header_bb_location + 1, dispatch_bb);
         builder_->name(dispatch_bb->label, "dispatch_" + continuation->name());
         bb->branch(dispatch_bb->label);
-        int targets = continuation->num_ops() - 2;
+        int targets = continuation->num_ops();
         assert(targets > 0);
 
         assert(targets == 1);
-        auto callee = continuation->op(2)->as_continuation();
+        auto callee = continuation->op(0)->as_continuation();
         // Extract the relevant variant & expand the tuple if necessary
         auto arg = world().variant_extract(continuation->param(0), 0);
         auto extracted = emit(arg, dispatch_bb);
@@ -382,13 +382,13 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         bb->branch(header_label);
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
-        auto header_cont = continuation->op(0)->as_continuation();
+        // auto header_cont = continuation->op(0)->as_continuation();
 
-        int targets = continuation->num_ops() - 1;
+        int targets = continuation->num_ops();
         assert(targets > 0);
 
         assert(targets == 1);
-        auto callee = continuation->op(1)->as_continuation();
+        auto callee = continuation->op(0)->as_continuation();
         // TODO phis
         bb->branch(current_fn_->bbs_map[callee]->label);
     } /*else if (continuation->intrinsic() == Intrinsic::SCFNonLocalJump) {
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index db0b0887d..413ede8ce 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -473,7 +473,7 @@ inline void iterate_ancestors(Continuation* cont, Fn fn) {
         Continuation* top = stack.back();
         stack.pop_back();
         if (done.contains(top)) continue;
-        if (top != cont) fn(top);
+        if (top != cont && fn(top)) return;
         done.insert(top);
         for (auto pred : top->preds()) {
             auto pred_cont = pred->isa_continuation();
@@ -484,35 +484,121 @@ inline void iterate_ancestors(Continuation* cont, Fn fn) {
         }
     }
 }
+template<bool b, typename Fn>
+inline void visit_children(const DomTreeBase<b>& tree, const CFNode* n, Fn fn, bool is_children = false) {
+    if (is_children)
+        fn(n);
+    for (auto children : tree.children(n)) {
+        visit_children(tree, children, fn, true);
+    }
+}
 
 void CodeGen::structure_flow() {
     Scope::for_each(world(), [&](const Scope& scope) {
         CFA cfa(scope);
+        auto& dom_tree = cfa.f_cfg().domtree();
         auto& post_dom_tree = cfa.b_cfg().domtree();
 
         for (auto def : scope.defs()) {
             if (auto cont = def->isa_continuation()) {
-                auto cfn = cfa[cont];
-                /*if (cont->callee() == world().branch()) {
-                    printf("xd: %s\n", cont->unique_name().c_str());
-                }*/
-
                 //if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
                 //    continue;
                 if (cont->preds().size() <= 1)
                     continue;
 
+                auto dominator = dom_tree.idom(cfa[cont]);
+
                 printf("has more than 1 incoming branch: %s\n", cont->unique_name().c_str());
+                printf(" dominator: %s\n", dominator->continuation()->unique_name().c_str());
+
+                auto dominator_post_dominator = post_dom_tree.idom(dominator);
+                if (dominator_post_dominator != nullptr)
+                    printf(" dominator post dominator: %s\n", dominator_post_dominator->continuation()->unique_name().c_str());
+                else
+                    printf(" dominator post dominator: NONE lmao\n");
 
-                /*for (auto pred : cont->preds()) {
-                    auto& pred_dominators = post_dom_tree.children(cfa[cont]);
-                    // TODO insert join nodes when this assert breaks
-                    // TODO inspect postdom tree recursively
-                    assert(std::find(pred_dominators.begin(), pred_dominators.end(), cfa[pred]) != pred_dominators.end());
-                }*/
+                visit_children(dom_tree, dominator, [&](const CFNode* n) {
+                    printf("  dominator child: %s\n", n->continuation()->unique_name().c_str());
+                });
+
+                bool needs_join = false;
+                Continuation* selection_dominator = nullptr;
                 iterate_ancestors(cont, [&](Continuation* ancestor) {
                     printf("  ancestor: %s\n", ancestor->unique_name().c_str());
+
+                    /*for (auto post_dom : post_dom_tree.children(cfa[ancestor])) {
+                        printf("    post-dominator: %s\n", post_dom->continuation()->unique_name().c_str());
+                    }*/
+                    auto post_dom = cfa[ancestor];
+                    while(true) {
+                        post_dom = post_dom_tree.idom(post_dom);
+                        if (post_dom == nullptr) break;
+                        printf("    post-dominator: %s\n", post_dom->continuation()->unique_name().c_str());
+                    }
+
+                    /*bool ancestor_post_dominated = false;
+                    Continuation* post_dom = ancestor;
+                    while (true) {
+                        auto dom_cfn = post_dom_tree.idom(cfa[post_dom]);
+                        if (dom_cfn == nullptr) break;
+                        post_dom = dom_cfn->continuation();
+                        printf("    post-dominator: %s\n", post_dom->unique_name().c_str());
+                        if (post_dom == cont) {
+                            // Wrong.
+                            ancestor_post_dominated = true;
+                            continue;
+                        }
+                    }
+                    needs_join |= !ancestor_post_dominated;
+                    if (needs_join)
+                        return true;
+
+                    bool dominate_all_preds = true;
+                    for (auto pred : cont->preds()) {
+                        printf("    pred: %s\n", ancestor->unique_name().c_str());
+                        bool dominated = false;
+                        Continuation* dom = pred;
+                        while (true) {
+                            auto dom_cfn = dom_tree.idom(cfa[dom]);
+                            if (dom_cfn == nullptr) break;
+                            dom = dom_cfn->continuation();
+                            printf("        pred dominator: %s\n", dom->unique_name().c_str());
+                            if (dom == ancestor) {
+                                dominated = true;
+                                break;
+                            }
+                        }
+                        dominate_all_preds &= dominated;
+                    }
+                    if (dominate_all_preds) {
+                        selection_dominator = ancestor;
+                        printf("This one dominates all preds: %s!\n", selection_dominator->unique_name().c_str());
+                        return true;
+                    }*/
+
+                    /*
+                    ContinuationSet preds;
+                    for (auto pred : cont->preds())
+                        preds.insert(pred);
+                    Continuation* dom = ancestor;
+                    while (true) {
+                        auto dom_cfn = dom_tree.idom(cfa[dom]);
+                        if (dom_cfn == nullptr) break;
+                        dom = dom_cfn->continuation();
+                        printf("    dominator: %s\n", dom->unique_name().c_str());
+                        if (preds.contains(dom)) {
+                            preds.erase(dom);
+                        }
+                    }
+                    printf("%d\n", preds.size());
+                    if (preds.empty()) {
+                        printf("This one dominates all preds!\n");
+                    }*/
+
+                    return false;
                 });
+
+                assert(!needs_join);
             }
         }
     });
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 38d1f9663..1da0c5194 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -15,18 +15,6 @@ const Def* Continuation::callee() const {
     return empty() ? world().bottom(world().fn_type(), debug()) : op(0);
 }
 
-Defs Continuation::potential_succs() const {
-    if (intrinsic() == Intrinsic::SCFLoopHeader)
-        return ops().skip_front(2);
-    else if (intrinsic() == Intrinsic::SCFLoopMerge)
-        return ops().skip_front();
-
-    //else if (intrinsic() == Intrinsic::SCFLoopContinue)
-    //return std::vector<const Def*> { op(0) };
-
-    return ops();
-}
-
 Continuation* Continuation::stub() const {
     Rewriter rewriter;
 
@@ -266,9 +254,9 @@ void Continuation::match(const Def* val, Continuation* otherwise, Defs patterns,
 
 void Continuation::structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
     attributes_.intrinsic = Intrinsic::SCFLoopMerge;
-    resize(1 + targets.size());
-    set_op(0, loop_header);
-    size_t x = 1;
+    attributes_.scf_metadata.loop_epilogue.loop_header = loop_header;
+    resize(targets.size());
+    size_t x = 0;
     for (auto target : targets)
         set_op(x++, target);
 }
@@ -281,10 +269,10 @@ void Continuation::structured_loop_continue(const Continuation* loop_header) {
 
 void Continuation::structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets) {
     attributes_.intrinsic = Intrinsic::SCFLoopHeader;
-    resize(2 + targets.size());
-    set_op(0, loop_epilogue);
-    set_op(1, loop_continue);
-    size_t x = 2;
+    resize(targets.size());
+    attributes_.scf_metadata.loop_header.continue_target = loop_continue;
+    attributes_.scf_metadata.loop_header.merge_target = loop_epilogue;
+    size_t x = 0;
     for (auto target : targets)
         set_op(x++, target);
 }
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 3770c63ee..8f6ea6c50 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -102,10 +102,25 @@ enum class Intrinsic : uint8_t {
  */
 class Continuation : public Def {
 public:
+    /// Stores information about structured control flow that should not be encoded in ops, as ops encode control flow
+    union SCFMetadata {
+        struct {
+            const Continuation* continue_target;
+            const Continuation* merge_target;
+        } loop_header;
+        struct {
+            const Continuation* loop_header;
+        } loop_epilogue;
+        struct {
+            const Continuation* merge_target;
+        } selection_header;
+    };
+
     struct Attributes {
         Intrinsic intrinsic = Intrinsic::None;
         Visibility visibility = Visibility::Internal;
         CC cc = CC::C;
+        SCFMetadata scf_metadata = {};
 
         Attributes() = default;
         Attributes(Intrinsic intrinsic) : intrinsic(intrinsic) {}
@@ -133,7 +148,6 @@ class Continuation : public Def {
     const Param* ret_param() const;
     const Def* callee() const;
     Defs args() const { return num_ops() == 0 ? Defs(0, 0) : ops().skip_front(); }
-    Defs potential_succs() const;
     const Def* arg(size_t i) const { return args()[i]; }
     const FnType* type() const { return Def::type()->as<FnType>(); }
     const FnType* callee_fn_type() const { return callee()->type()->as<FnType>(); }
diff --git a/src/thorin/util/dot_dump.cpp b/src/thorin/util/dot_dump.cpp
index 21ebb0bcf..6fd419517 100644
--- a/src/thorin/util/dot_dump.cpp
+++ b/src/thorin/util/dot_dump.cpp
@@ -155,16 +155,14 @@ void DotPrinter::dump_continuation(const Continuation* cont) {
 
     file << down << endl << "]";
 
-    int x = 1;
+    int x = 0;
     switch (intrinsic) {
-        case Intrinsic::SCFLoopHeader:
-            dump_def(cont->op(1));
-            file << endl << cont->unique_name() << " -> " << cont->op(1)->unique_name() << " [arrowhead=none];";
-            x = 2;
         case Intrinsic::SCFLoopContinue:
-        case Intrinsic::SCFLoopMerge:
             dump_def(cont->op(0));
-            file << endl << cont->unique_name() << " -> " << cont->op(0)->unique_name() << " [arrowhead=none];";
+            file << endl << cont->unique_name() << " -> " << cont->op(0)->unique_name() << " [arrowhead=normal];";
+            return;
+        case Intrinsic::SCFLoopHeader:
+        case Intrinsic::SCFLoopMerge:
             for (size_t i = x; i < cont->num_ops(); i++) {
                 auto op = cont->op(i);
                 dump_def(op);

From 45e0f9b19196e292dad103da67dd283537bdac63 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 29 Mar 2021 11:31:33 +0200
Subject: [PATCH 046/342] cleanup

---
 src/thorin/be/spirv/spirv_transform.cpp | 229 +++++-------------------
 1 file changed, 43 insertions(+), 186 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 413ede8ce..bbdfc4018 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -45,6 +45,8 @@ struct StructuredLoop {
     const Head* parent_head;
     const Head* head;
     const std::string name;
+    StructuredLoop(const Head* parent_head, const Head* head, std::string&& name)
+    : parent_head(parent_head), head(head), name(name) {}
 
     std::vector<DispatchTarget> inner_destinations = {};
     std::vector<DispatchTarget> outer_destinations = {};
@@ -71,7 +73,7 @@ struct ScopeContext {
     std::unordered_map<const Head*, StructuredLoop> rewritten_loops;
 };
 
-inline std::string safe_name(const Head* head) {
+inline std::string loop_name(const Head* head) {
     if (head == nullptr || head->is_root()) {
         return "root";
     } else {
@@ -87,27 +89,17 @@ inline std::string safe_name(const Head* head) {
 
 /// Visits the forest and fills def2loop
 inline void tag_continuations(ScopeContext& ctx, const Base* base, const Head* parent) {
-    for (int i = 0; i < base->depth(); i++)
-        printf(" ");
-    if (auto* head = base->isa<Head>()) {
-        printf("loop: header = ");
-    }
-    for (auto& node : base->cf_nodes()) {
-        printf("%s ", node->continuation()->to_string().c_str());
-    }
-    printf("\n");
-
     if (const Head* head = base->isa<Head>()) {
-        auto name = safe_name(head);
+        auto name = loop_name(head);
 
         for (auto& children : head->children()) {
             tag_continuations(ctx, &*children, head);
         }
 
-        StructuredLoop loop{parent, head, name};
+        StructuredLoop loop(parent, head, std::move(name));
         ctx.rewritten_loops.emplace(head, loop);
 
-    } else if(auto* leaf = base->isa<Leaf>()) {
+    } else if(base->isa<Leaf>()) {
         for (auto& node : base->cf_nodes()) {
             auto[i, result] = ctx.def2loop.emplace(node->continuation(), parent);
             assert(result);
@@ -121,7 +113,7 @@ inline int record_destination(std::vector<DispatchTarget>& vec, DispatchTarget d
     auto i = std::find(vec.begin(), vec.end(), dest);
     if (i == vec.end()) {
         vec.emplace_back(dest);
-        return vec.size() - 1;
+        return static_cast<int>(vec.size()) - 1;
     } else return i - vec.begin();
 }
 
@@ -136,7 +128,7 @@ inline std::vector<StructuredLoop*> get_path(ScopeContext& ctx, const Head* head
     std::vector<StructuredLoop*> path = {};
     assert(head != nullptr);
     while (head != nullptr) {
-        auto* loop = &ctx.rewritten_loops[head];
+        StructuredLoop* loop = &ctx.rewritten_loops.find(head)->second;
         assert(loop != nullptr);
         path.emplace(path.begin(), loop);
         head = loop->parent_head;
@@ -158,22 +150,22 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
         for (size_t i = 0; i < cont->num_ops(); i++) {
             auto def = cont->op(i);
             if (auto dest = def->isa_continuation()) {
-                const Head* source_loop = *ctx.def2loop[cont];
+                const Head* source_loop_head = *ctx.def2loop[cont];
 
                 if (dest->intrinsic() == Intrinsic::Branch) {
                     continue;
                 }
 
                 assert(ctx.def2loop.find(dest) != ctx.def2loop.end());
-                const Head* dest_loop = *ctx.def2loop[dest];
+                const Head* dest_loop_head = *ctx.def2loop[dest];
 
-                if (source_loop != dest_loop) {
+                if (source_loop_head != dest_loop_head) {
                     // We found a non-local jump
-                    assert(ctx.rewritten_loops.find(source_loop) != ctx.rewritten_loops.end());
-                    auto& loop = ctx.rewritten_loops[source_loop];
+                    assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
+                    auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
 
-                    std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop);
-                    std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop);
+                    std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop_head);
+                    std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop_head);
                     int bi = 0;
                     while (bi < std::min(source_path.size(), dest_path.size())) {
                         if (source_path[bi] == dest_path[bi])
@@ -207,22 +199,22 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                         }
                     };
 
-                    for (auto loop : leave) {
+                    for (auto dest_loop : leave) {
                         DispatchTarget destination;
-                        destination.exit = loop;
+                        destination.exit = dest_loop;
 
-                        record_step(loop, destination);
+                        record_step(dest_loop, destination);
                         last = 1;
-                        prev = loop;
+                        prev = dest_loop;
                         assert(prev != nullptr);
                     }
-                    for (auto loop : enter) {
+                    for (auto dest_loop : enter) {
                         DispatchTarget destination;
-                        destination.entry = loop;
+                        destination.entry = dest_loop;
 
-                        record_step(loop, destination);
+                        record_step(dest_loop, destination);
                         last = 2;
-                        prev = loop;
+                        prev = dest_loop;
                         assert(prev != nullptr);
                     }
 
@@ -238,22 +230,20 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                         dest
                     };
                     loop.rewire.emplace_back(rewire);
-                    printf("nlj %s %d!\n", loop.name.c_str(), loop.rewire.size());
-                } else if (source_loop == dest_loop && source_loop != nullptr) {
-                    for (auto& head : source_loop->cf_nodes()) {
-                        if (head->continuation() == dest) {
+                } else if (source_loop_head == dest_loop_head && source_loop_head != nullptr) {
+                    for (auto& cf_node : source_loop_head->cf_nodes()) {
+                        if (cf_node->continuation() == dest) {
                             // We found a backedge
-                            assert(ctx.rewritten_loops.find(source_loop) != ctx.rewritten_loops.end());
-                            auto& loop = ctx.rewritten_loops[source_loop];
+                            assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
+                            auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
 
                             DispatchTarget destination;
                             destination.cont = dest;
                             record_destination(loop.inner_destinations, destination);
 
                             RewireMe rewire(cont, i);
-                            rewire.backedge = head->continuation();
+                            rewire.backedge = cf_node->continuation();
                             loop.rewire.emplace_back(rewire);
-                            printf("backedge %s %d!\n", loop.name.c_str(), loop.rewire.size());
                             break;
                         }
                     }
@@ -287,7 +277,7 @@ inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
 
         if (head->num_cf_nodes() == 0)
             return;
-        StructuredLoop& loop = ctx.rewritten_loops[head];
+        StructuredLoop& loop = ctx.rewritten_loops.find(head)->second;
 
         // here, parent headers need to know what they're jumping *into*
         std::vector<const Type*> dest_types;
@@ -316,7 +306,7 @@ inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
 
 inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
-        StructuredLoop& loop = ctx.rewritten_loops[head];
+        StructuredLoop& loop = ctx.rewritten_loops.find(head)->second;
 
         if (head->num_cf_nodes() > 0) {
             // here, children epilogues need to know what they're jumping *out to*
@@ -354,34 +344,29 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
 inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
         assert(ctx.rewritten_loops.find(head) != ctx.rewritten_loops.end());
-        auto& loop = ctx.rewritten_loops[head];
+        auto& loop = ctx.rewritten_loops.find(head)->second;
 
         if (head->num_cf_nodes() > 0) {
             loop.new_epilogue->structured_loop_merge(loop.new_header, loop.epilogue_destination_conts);
             loop.new_continue->structured_loop_continue(loop.new_header);
             loop.new_header->structured_loop_header(loop.new_epilogue, loop.new_continue, loop.header_destination_conts);
-            printf("Loop %s!\n", loop.name.c_str());
-            for (auto c : loop.header_destination_conts)
-                printf("  header target: %s!\n", c->unique_name().c_str());
-            for (auto c : loop.epilogue_destination_conts)
-                printf("  epilogue target: %s!\n", c->unique_name().c_str());
         }
 
         for (auto& children : head->children()) {
             rewire_loops(world, ctx, &*children);
         }
 
-        printf("rewires %s %d!\n", loop.name.c_str(), loop.rewire.size());
         for (auto& rewire : loop.rewire) {
-            printf("rewire!\n");
+            assert(loop.head != nullptr);
             if (rewire.backedge != nullptr) {
-                printf("handling BE!\n");
+
                 DispatchTarget destination;
                 destination.cont = rewire.backedge;
                 auto variant_index = index_of_destination(loop.inner_destinations, destination);
 
                 auto old_fn_type = rewire.backedge->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper_to" + destination.cont->unique_name() });
+                ctx.def2loop[wrapper] = loop.head;
                 wrapper->attributes_.intrinsic = Intrinsic::SCFBackEdge;
 
                 auto header_variant_type = loop.new_header->type()->op(0)->as<VariantType>();
@@ -390,22 +375,13 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 rewire.cont->unset_op(rewire.op);
                 rewire.cont->set_op(rewire.op, wrapper);
             } else {
-                printf("handling NLJ!\n");
-
                 auto& nlj = rewire.non_local_jump;
 
                 auto old_fn_type = nlj.final_destination->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper_to" + nlj.final_destination->unique_name() });
+                ctx.def2loop[wrapper] = loop.head;
                 wrapper->attributes_.intrinsic = Intrinsic::SCFNonLocalJump;
 
-                printf("nlj = %s\n", wrapper->unique_name().c_str());
-                printf("src = %s\n", rewire.cont->name().c_str());
-                for (auto exit : nlj.exits)
-                    printf("exit = %s\n", exit->name.c_str());
-                for (auto enter : nlj.enters)
-                    printf("enter = %s\n", enter->name.c_str());
-                printf("dst = %s\n", nlj.final_destination->name().c_str());
-
                 const Def* argument = tuple_from_params(world, wrapper->params());
                 Continuation* first_jump = nullptr;
 
@@ -447,9 +423,8 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
 }
 
 void CodeGen::structure_loops() {
-    Scope::for_each(world(), [&](const Scope& scope) {
+    Scope::for_each(world(), [&](Scope& scope) {
         ScopeContext context(scope);
-        printf("top: %d\n", scope.has_free_params());
 
         const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
         tag_continuations(context, looptree.root(), nullptr);
@@ -459,40 +434,10 @@ void CodeGen::structure_loops() {
         create_epilogues(world(), context, looptree.root());
 
         rewire_loops(world(), context, looptree.root());
-        printf("done\n");
+        scope.update();
     });
 }
 
-template<typename Fn>
-inline void iterate_ancestors(Continuation* cont, Fn fn) {
-    ContinuationSet done;
-
-    Continuations stack;
-    stack.push_back(cont);
-    while (!stack.empty()) {
-        Continuation* top = stack.back();
-        stack.pop_back();
-        if (done.contains(top)) continue;
-        if (top != cont && fn(top)) return;
-        done.insert(top);
-        for (auto pred : top->preds()) {
-            auto pred_cont = pred->isa_continuation();
-            if (!pred_cont) continue;
-            if (!done.contains(pred_cont)) {
-                stack.push_back(pred_cont);
-            }
-        }
-    }
-}
-template<bool b, typename Fn>
-inline void visit_children(const DomTreeBase<b>& tree, const CFNode* n, Fn fn, bool is_children = false) {
-    if (is_children)
-        fn(n);
-    for (auto children : tree.children(n)) {
-        visit_children(tree, children, fn, true);
-    }
-}
-
 void CodeGen::structure_flow() {
     Scope::for_each(world(), [&](const Scope& scope) {
         CFA cfa(scope);
@@ -501,104 +446,16 @@ void CodeGen::structure_flow() {
 
         for (auto def : scope.defs()) {
             if (auto cont = def->isa_continuation()) {
-                //if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
-                //    continue;
                 if (cont->preds().size() <= 1)
                     continue;
 
                 auto dominator = dom_tree.idom(cfa[cont]);
-
-                printf("has more than 1 incoming branch: %s\n", cont->unique_name().c_str());
-                printf(" dominator: %s\n", dominator->continuation()->unique_name().c_str());
-
                 auto dominator_post_dominator = post_dom_tree.idom(dominator);
-                if (dominator_post_dominator != nullptr)
-                    printf(" dominator post dominator: %s\n", dominator_post_dominator->continuation()->unique_name().c_str());
-                else
-                    printf(" dominator post dominator: NONE lmao\n");
-
-                visit_children(dom_tree, dominator, [&](const CFNode* n) {
-                    printf("  dominator child: %s\n", n->continuation()->unique_name().c_str());
-                });
-
-                bool needs_join = false;
-                Continuation* selection_dominator = nullptr;
-                iterate_ancestors(cont, [&](Continuation* ancestor) {
-                    printf("  ancestor: %s\n", ancestor->unique_name().c_str());
-
-                    /*for (auto post_dom : post_dom_tree.children(cfa[ancestor])) {
-                        printf("    post-dominator: %s\n", post_dom->continuation()->unique_name().c_str());
-                    }*/
-                    auto post_dom = cfa[ancestor];
-                    while(true) {
-                        post_dom = post_dom_tree.idom(post_dom);
-                        if (post_dom == nullptr) break;
-                        printf("    post-dominator: %s\n", post_dom->continuation()->unique_name().c_str());
-                    }
-
-                    /*bool ancestor_post_dominated = false;
-                    Continuation* post_dom = ancestor;
-                    while (true) {
-                        auto dom_cfn = post_dom_tree.idom(cfa[post_dom]);
-                        if (dom_cfn == nullptr) break;
-                        post_dom = dom_cfn->continuation();
-                        printf("    post-dominator: %s\n", post_dom->unique_name().c_str());
-                        if (post_dom == cont) {
-                            // Wrong.
-                            ancestor_post_dominated = true;
-                            continue;
-                        }
-                    }
-                    needs_join |= !ancestor_post_dominated;
-                    if (needs_join)
-                        return true;
-
-                    bool dominate_all_preds = true;
-                    for (auto pred : cont->preds()) {
-                        printf("    pred: %s\n", ancestor->unique_name().c_str());
-                        bool dominated = false;
-                        Continuation* dom = pred;
-                        while (true) {
-                            auto dom_cfn = dom_tree.idom(cfa[dom]);
-                            if (dom_cfn == nullptr) break;
-                            dom = dom_cfn->continuation();
-                            printf("        pred dominator: %s\n", dom->unique_name().c_str());
-                            if (dom == ancestor) {
-                                dominated = true;
-                                break;
-                            }
-                        }
-                        dominate_all_preds &= dominated;
-                    }
-                    if (dominate_all_preds) {
-                        selection_dominator = ancestor;
-                        printf("This one dominates all preds: %s!\n", selection_dominator->unique_name().c_str());
-                        return true;
-                    }*/
-
-                    /*
-                    ContinuationSet preds;
-                    for (auto pred : cont->preds())
-                        preds.insert(pred);
-                    Continuation* dom = ancestor;
-                    while (true) {
-                        auto dom_cfn = dom_tree.idom(cfa[dom]);
-                        if (dom_cfn == nullptr) break;
-                        dom = dom_cfn->continuation();
-                        printf("    dominator: %s\n", dom->unique_name().c_str());
-                        if (preds.contains(dom)) {
-                            preds.erase(dom);
-                        }
-                    }
-                    printf("%d\n", preds.size());
-                    if (preds.empty()) {
-                        printf("This one dominates all preds!\n");
-                    }*/
-
-                    return false;
-                });
-
-                assert(!needs_join);
+                bool needs_join = dominator_post_dominator->continuation() != cont;
+                if (needs_join) {
+                    assert(false && "Not structured CF !");
+                    // TODO: insert join node into dominator and redirect dominated nodes to take it
+                }
             }
         }
     });

From 6f722e51a394533afb079e22e4d61b97ab562c64 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 29 Mar 2021 11:42:12 +0200
Subject: [PATCH 047/342] create dummy merge block when needed

---
 src/thorin/be/spirv/spirv.cpp         | 21 +++++++++++++--------
 src/thorin/be/spirv/spirv_builder.hpp |  8 +++++---
 2 files changed, 18 insertions(+), 11 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 324b41133..37de4d114 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -312,18 +312,23 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     else if (continuation->callee() == world().branch()) {
         auto& domtree = current_fn_->scope->b_cfg().domtree();
         auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
-
-        printf("Merge @%s\n", merge_cont->unique_name().c_str());
-        /*BasicBlockBuilder* merge_bb = &current_fn_->bbs.emplace_back(*current_fn_);
-        auto merge_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), merge_cont);
-        current_fn_->bbs_to_emit.emplace(merge_bb_location + 1, merge_bb);
-        builder_->name(merge_bb->label, "merge_" + merge_cont->name());*/
+        SpvId merge_bb;
+        if (merge_cont == current_fn_->scope->exit()) {
+            BasicBlockBuilder* unreachable_merge_bb = &current_fn_->bbs.emplace_back(*current_fn_);
+            current_fn_->bbs_to_emit.emplace_back(unreachable_merge_bb);
+            builder_->name(unreachable_merge_bb->label, "merge_unreachable" + continuation->name());
+            unreachable_merge_bb->unreachable();
+            merge_bb = unreachable_merge_bb->label;
+        } else {
+            // TODO create a dedicated merge bb if this one is the merge blocks for more than 1 selection construct
+            *current_fn_->labels[merge_cont];
+        }
 
         auto cond = emit(continuation->arg(0), bb);
-        bb->args[continuation->arg(0)] = cond;
+        bb->args.emplace(continuation->arg(0), cond);
         auto tbb = *current_fn_->labels[continuation->arg(1)->as_continuation()];
         auto fbb = *current_fn_->labels[continuation->arg(2)->as_continuation()];
-        bb->selection_merge(*current_fn_->labels[merge_cont],spv::SelectionControlMaskNone);
+        bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
         bb->branch_conditional(cond, tbb, fbb);
     } /*else if (continuation->callee()->isa<Continuation>() &&
                continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index cacb9f2b3..773af85a2 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -135,7 +135,6 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     void store(SpvId value, SpvId pointer) {
         op(spv::Op::OpStore, 3);
-        auto id = generate_fresh_id();
         ref_id(pointer);
         ref_id(value);
     }
@@ -187,6 +186,10 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         ref_id(value);
     }
 
+    void unreachable() {
+        op(spv::Op::OpUnreachable, 1);
+    }
+
 private:
     SpvId generate_fresh_id();
 };
@@ -315,8 +318,7 @@ struct SpvFileBuilder {
                 fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
                 fn_defs.ref_id(phi->type);
                 fn_defs.ref_id(phi->value);
-                printf("Phi %d\n", phi->value);
-                assert(phi->preds.size() > 0);
+                assert(!phi->preds.empty());
                 for (auto& [pred_value, pred_label] : phi->preds) {
                     fn_defs.ref_id(pred_value);
                     fn_defs.ref_id(pred_label);

From 6f37d8869642bb61f34277030ad58fc77b076548 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 29 Mar 2021 12:02:33 +0200
Subject: [PATCH 048/342] more cleanup

---
 src/thorin/be/spirv/spirv.cpp           | 51 +++++--------------------
 src/thorin/be/spirv/spirv.h             |  4 +-
 src/thorin/be/spirv/spirv_transform.cpp | 20 +++++-----
 3 files changed, 22 insertions(+), 53 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 37de4d114..66fa5b9df 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -68,8 +68,9 @@ SpvType CodeGen::convert(const Type* type) {
         }
         case Node_IndefiniteArrayType: {
             assert(false && "TODO");
-            auto array = type->as<IndefiniteArrayType>();
-            //return types_[type] = spv_type;
+            // auto array = type->as<IndefiniteArrayType>();
+            // return types_[type] = spv_type;
+            THORIN_UNREACHABLE;
         }
         case Node_DefiniteArrayType: {
             auto array = type->as<DefiniteArrayType>();
@@ -110,7 +111,7 @@ SpvType CodeGen::convert(const Type* type) {
             }
 
             assert(false && "TODO: handle closure mess");
-            break;
+            THORIN_UNREACHABLE;
         }
 
         case Node_StructType: {
@@ -179,30 +180,6 @@ SpvType CodeGen::convert(const Type* type) {
     return types_[type] = spv_type;
 }
 
-inline Schedule schedule_structured(const Scope& scope) {
-    // until we have sth better simply use the RPO of the CFG
-    Schedule result;
-    for (auto n : scope.f_cfg().reverse_post_order())
-        result.emplace_back(n->continuation());
-
-    auto schedule = [&](const Continuation* cont) {
-        printf("scheduled: %s\n", cont->unique_name().c_str());
-    };
-
-    auto visit = [&](const Continuation* cont) {
-        if (cont->intrinsic() == Intrinsic::SCFLoopHeader) {
-            // Write continue block FIRST
-            schedule(cont->op(1)->as_continuation());
-            // Then write the header
-            schedule(cont);
-
-            schedule(cont->op(0)->as_continuation());
-        }
-    };
-
-    return result;
-}
-
 void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
@@ -224,7 +201,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     for (auto cont : conts) {
         if (cont->intrinsic() == Intrinsic::EndScope) continue;
 
-        BasicBlockBuilder* bb = &bbs.emplace_back(fn);
+        BasicBlockBuilder* bb = bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn)).get();
         fn.bbs_to_emit.emplace_back(bb);
         auto [i, b] = fn.bbs_map.emplace(cont, bb);
         assert(b);
@@ -267,8 +244,8 @@ void CodeGen::emit(const thorin::Scope& scope) {
     }
 
     for(auto& bb : fn.bbs) {
-        for (auto& [param, phi] : bb.phis_map) {
-            bb.phis.emplace_back(&phi);
+        for (auto& [param, phi] : bb->phis_map) {
+            bb->phis.emplace_back(&phi);
         }
     }
     
@@ -314,7 +291,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
         SpvId merge_bb;
         if (merge_cont == current_fn_->scope->exit()) {
-            BasicBlockBuilder* unreachable_merge_bb = &current_fn_->bbs.emplace_back(*current_fn_);
+            BasicBlockBuilder* unreachable_merge_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
             current_fn_->bbs_to_emit.emplace_back(unreachable_merge_bb);
             builder_->name(unreachable_merge_bb->label, "merge_unreachable" + continuation->name());
             unreachable_merge_bb->unreachable();
@@ -349,7 +326,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
         bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
 
-        BasicBlockBuilder* dispatch_bb = &current_fn_->bbs.emplace_back(*current_fn_);
+        BasicBlockBuilder* dispatch_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
 
         auto header_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), bb);
 
@@ -396,14 +373,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto callee = continuation->op(0)->as_continuation();
         // TODO phis
         bb->branch(current_fn_->bbs_map[callee]->label);
-    } /*else if (continuation->intrinsic() == Intrinsic::SCFNonLocalJump) {
-        auto header_cont = continuation->op(0)->as_continuation();
-        // TODO setup arguments & stuff
-        bb->branch(current_fn_->bbs_map[continuation->op(0)->as_continuation()]->label);
-    } else if (continuation->intrinsic() == Intrinsic::SCFBackEdge) {
-        // TODO setup arguments & stuff
-        bb->branch(current_fn_->bbs_map[continuation->op(0)->as_continuation()]->label);
-    } */ else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
         int index = -1;
         for (auto& arg : continuation->args()) {
             index++;
@@ -626,7 +596,6 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto variant = def->isa<Variant>()) {
         auto struct_type = def->type()->as<VariantType>();
-        auto type = convert(struct_type);
         std::vector<SpvId> elements;
         elements.resize(struct_type->num_ops());
         size_t x = 0;
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 4e85a426b..074876e26 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -15,7 +15,7 @@ struct SpvType {
     // TODO: Alignment rules are complicated and client API dependant
     size_t alignment = 0;
 
-    SpvId payload_id;
+    SpvId payload_id = { 0 };
 };
 
 struct FnBuilder;
@@ -30,7 +30,7 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
 struct FnBuilder : public builder::SpvFnBuilder {
     const Scope* scope;
     builder::SpvFileBuilder* file_builder;
-    std::vector<BasicBlockBuilder> bbs;
+    std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index bbdfc4018..392baa92d 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -136,10 +136,10 @@ inline std::vector<StructuredLoop*> get_path(ScopeContext& ctx, const Head* head
     return path;
 }
 
-inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base* base, const Head* parent) {
+inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base* base) {
     if (const Head* head = base->isa<Head>()) {
         for (auto& children : head->children()) {
-            collect_dispatch_targets(world, ctx, &*children, head);
+            collect_dispatch_targets(world, ctx, &*children);
         }
     } else {
         const Leaf* leaf = base->as<Leaf>();
@@ -167,7 +167,7 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                     std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop_head);
                     std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop_head);
                     int bi = 0;
-                    while (bi < std::min(source_path.size(), dest_path.size())) {
+                    while (bi < static_cast<int>(std::min(source_path.size(), dest_path.size()))) {
                         if (source_path[bi] == dest_path[bi])
                             bi++;
                         else break;
@@ -177,9 +177,9 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                     // these two sequences cannot be both empty (that wouldn't be a non-local jump then!)
                     std::vector<StructuredLoop*> leave;
                     std::vector<StructuredLoop*> enter;
-                    for (int j = source_path.size() - 1; j >= bi; j--)
+                    for (int j = static_cast<int>(source_path.size()) - 1; j >= bi; j--)
                         leave.emplace_back(source_path[j]);
-                    for (int j = bi; j < dest_path.size(); j++)
+                    for (int j = bi; j < static_cast<int>(dest_path.size()); j++)
                         enter.emplace_back(dest_path[j]);
 
                     // 0 = this is the first step of the path
@@ -188,7 +188,7 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                     int last = 0;
                     StructuredLoop* prev;
 
-                    auto record_step = [&](StructuredLoop* loop, DispatchTarget destination) {
+                    auto record_step = [&](DispatchTarget destination) {
                         if (last == 0) {
                             // nothing to do, this node isn't a dispatching one
                         } else {
@@ -203,7 +203,7 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                         DispatchTarget destination;
                         destination.exit = dest_loop;
 
-                        record_step(dest_loop, destination);
+                        record_step(destination);
                         last = 1;
                         prev = dest_loop;
                         assert(prev != nullptr);
@@ -212,7 +212,7 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                         DispatchTarget destination;
                         destination.entry = dest_loop;
 
-                        record_step(dest_loop, destination);
+                        record_step(destination);
                         last = 2;
                         prev = dest_loop;
                         assert(prev != nullptr);
@@ -221,7 +221,7 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
                     assert(last != 0);
                     DispatchTarget destination;
                     destination.cont = dest;
-                    record_step(prev, destination);
+                    record_step(destination);
 
                     RewireMe rewire(cont, i);
                     rewire.non_local_jump = {
@@ -428,7 +428,7 @@ void CodeGen::structure_loops() {
 
         const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
         tag_continuations(context, looptree.root(), nullptr);
-        collect_dispatch_targets(world(), context, looptree.root(), nullptr);
+        collect_dispatch_targets(world(), context, looptree.root());
 
         create_headers(world(), context, looptree.root());
         create_epilogues(world(), context, looptree.root());

From e34254966157c754d728ff6cdc48286ace66a967 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 30 Mar 2021 11:05:14 +0200
Subject: [PATCH 049/342] merge llvm_rewrite

---
 src/thorin/be/codegen.cpp               | 11 +++++++++--
 src/thorin/be/spirv/spirv.cpp           | 14 +++++++-------
 src/thorin/be/spirv/spirv.h             |  5 +++--
 src/thorin/be/spirv/spirv_transform.cpp |  4 ++--
 4 files changed, 21 insertions(+), 13 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index fa8078175..f6bec2238 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -6,6 +6,9 @@
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu.h"
 #endif
+#if THORIN_ENABLE_SPIRV
+#include "thorin/be/spirv/spirv.h"
+#endif
 #include "thorin/be/c/c.h"
 
 namespace thorin {
@@ -87,7 +90,8 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug)
             std::pair { NVVM,   Intrinsic::NVVM   },
             std::pair { OpenCL, Intrinsic::OpenCL },
             std::pair { AMDGPU, Intrinsic::AMDGPU },
-            std::pair { HLS,    Intrinsic::HLS    }
+            std::pair { HLS,    Intrinsic::HLS    },
+            std::pair { SpirV,  Intrinsic::SpirV  }
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
@@ -109,7 +113,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug)
         kernels.emplace_back(continuation);
     });
 
-    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU }) {
+    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU, SpirV }) {
         if (!importers_[backend].world().empty()) {
             get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 // determine whether or not this kernel uses restrict pointers
@@ -178,6 +182,9 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug)
     if (!importers_[AMDGPU].world().empty()) cgs[AMDGPU] = std::make_unique<llvm::AMDGPUCodeGen>(importers_[AMDGPU].world(), kernel_config, opt, debug);
 #else
     (void)opt;
+#endif
+#if THORIN_ENABLE_SPIRV
+    if (!importers_[SpirV].world().empty()) cgs[SpirV] = std::make_unique<spirv::CodeGen>(importers_[SpirV].world(), kernel_config, debug);
 #endif
     for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OpenCL }, std::pair { HLS, c::Lang::HLS } })
         if (!importers_[backend].world().empty()) cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug);
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 66fa5b9df..6b2bd7943 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -17,7 +17,7 @@ CodeGen::CodeGen(thorin::World& world, Cont2Config&, bool debug)
     : thorin::CodeGen(world, debug)
 {}
 
-void CodeGen::emit(std::ostream& out) {
+void CodeGen::emit_stream(std::ostream& out) {
     builder::SpvFileBuilder builder;
     builder_ = &builder;
     builder_->capability(spv::Capability::CapabilityShader);
@@ -298,13 +298,13 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             merge_bb = unreachable_merge_bb->label;
         } else {
             // TODO create a dedicated merge bb if this one is the merge blocks for more than 1 selection construct
-            *current_fn_->labels[merge_cont];
+            merge_bb = current_fn_->labels[merge_cont];
         }
 
         auto cond = emit(continuation->arg(0), bb);
         bb->args.emplace(continuation->arg(0), cond);
-        auto tbb = *current_fn_->labels[continuation->arg(1)->as_continuation()];
-        auto fbb = *current_fn_->labels[continuation->arg(2)->as_continuation()];
+        auto tbb = current_fn_->labels[continuation->arg(1)->as_continuation()];
+        auto fbb = current_fn_->labels[continuation->arg(2)->as_continuation()];
         bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
         bb->branch_conditional(cond, tbb, fbb);
     } /*else if (continuation->callee()->isa<Continuation>() &&
@@ -360,7 +360,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         bb->args[arg] = emit(arg, bb);
         auto* param = loop_header->param(0);
         auto& phi = current_fn_->bbs_map[loop_header]->phis_map[param];
-        phi.preds.emplace_back(*bb->args[arg], *current_fn_->labels[continuation]);
+        phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
 
         bb->branch(header_label);
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
@@ -381,9 +381,9 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             bb->args[arg] = emit(arg, bb);
             auto* param = callee->param(index);
             auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
-            phi.preds.emplace_back(*bb->args[arg], *current_fn_->labels[continuation]);
+            phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
         }
-        bb->branch(*current_fn_->labels[callee]);
+        bb->branch(current_fn_->labels[callee]);
     }
     /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         auto ret_continuation = emit_intrinsic(irbuilder, continuation);
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 074876e26..fda32aa2d 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -2,7 +2,7 @@
 #define THORIN_SPIRV_H
 
 #include "thorin/be/spirv/spirv_builder.hpp"
-#include "thorin/be/backends.h"
+#include "thorin/be/codegen.h"
 
 namespace thorin::spirv {
 
@@ -40,7 +40,8 @@ class CodeGen : public thorin::CodeGen {
 public:
     CodeGen(World&, Cont2Config&, bool debug);
 
-    void emit(std::ostream& stream) override;
+    void emit_stream(std::ostream& stream) override;
+    const char* file_ext() const override { return ".spv"; }
 protected:
     void structure_loops();
     void structure_flow();
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 392baa92d..51f613f2a 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -150,14 +150,14 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
         for (size_t i = 0; i < cont->num_ops(); i++) {
             auto def = cont->op(i);
             if (auto dest = def->isa_continuation()) {
-                const Head* source_loop_head = *ctx.def2loop[cont];
+                const Head* source_loop_head = ctx.def2loop[cont];
 
                 if (dest->intrinsic() == Intrinsic::Branch) {
                     continue;
                 }
 
                 assert(ctx.def2loop.find(dest) != ctx.def2loop.end());
-                const Head* dest_loop_head = *ctx.def2loop[dest];
+                const Head* dest_loop_head = ctx.def2loop[dest];
 
                 if (source_loop_head != dest_loop_head) {
                     // We found a non-local jump

From 49135bc57e85aff1f7d3a02d34aed2dc83b87506 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 1 Apr 2021 16:15:19 +0200
Subject: [PATCH 050/342] remember to delete later

---
 src/thorin/CMakeLists.txt | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index c7dd05f05..0e9915a48 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -82,6 +82,7 @@ set(THORIN_SOURCES
     util/symbol.h
     util/types.h
     util/utility.h
+    util/dot_dump.cpp
 )
 
 if(LLVM_FOUND)

From 0f74c581cd970882885d1940bbfaf310e7533a6b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 15 Apr 2021 15:10:18 +0200
Subject: [PATCH 051/342] datatypes refactor

---
 src/thorin/CMakeLists.txt               |   1 +
 src/thorin/be/spirv/spirv.cpp           | 208 ++++------------------
 src/thorin/be/spirv/spirv.h             |  70 +++++++-
 src/thorin/be/spirv/spirv_builder.hpp   |  12 ++
 src/thorin/be/spirv/spirv_datatypes.cpp | 219 ++++++++++++++++++++++++
 5 files changed, 329 insertions(+), 181 deletions(-)
 create mode 100644 src/thorin/be/spirv/spirv_datatypes.cpp

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 0e9915a48..8fbcec8fd 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -108,6 +108,7 @@ if(SPIRV_ENABLED)
         be/spirv/spirv.cpp
         be/spirv/spirv.h
         be/spirv/spirv_transform.cpp
+        be/spirv/spirv_datatypes.cpp
     )
 endif()
 
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 6b2bd7943..15a1cdb05 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -22,6 +22,8 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_ = &builder;
     builder_->capability(spv::Capability::CapabilityShader);
     builder_->capability(spv::Capability::CapabilityLinkage);
+    builder_->capability(spv::Capability::CapabilityVariablePointers);
+    builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
 
     structure_loops();
     structure_flow();
@@ -30,154 +32,14 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
-    builder_->finish(out);
-    builder_ = nullptr;
-}
-
-SpvType CodeGen::convert(const Type* type) {
-    if (auto spv_type = types_.lookup(type)) return *spv_type;
-
-    assert(!type->isa<MemType>());
-    SpvType spv_type;
-    switch (type->tag()) {
-        // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
-        // Using extensions, we could use 16 or 8-bit ints instead
-        // We can also pack them inside structures using bit-twiddling tricks, if the need arises
-        case PrimType_bool:                                                             spv_type.id = builder_->declare_bool_type(); spv_type.size = 4; spv_type.alignment = 4; break;
-        case PrimType_ps8:  case PrimType_qs8:  case PrimType_pu8:  case PrimType_qu8:  assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps16: case PrimType_qs16: case PrimType_pu16: case PrimType_qu16: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps32: case PrimType_qs32:                                         spv_type.id = builder_->declare_int_type(32, true ); spv_type.size = 4; spv_type.alignment = 4; break;
-        case PrimType_pu32: case PrimType_qu32:                                         spv_type.id = builder_->declare_int_type(32, false); spv_type.size = 4; spv_type.alignment = 4; break;
-        case PrimType_ps64: case PrimType_qs64: case PrimType_pu64: case PrimType_qu64: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pf16: case PrimType_qf16:                                         assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pf32: case PrimType_qf32:                                         spv_type.id = builder_->declare_float_type(32); spv_type.size = 4; spv_type.alignment = 4; break;
-        case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
-        case Node_PtrType: {
-            auto ptr = type->as<PtrType>();
-            spv::StorageClass storage_class;
-            switch (ptr->addr_space()) {
-                case AddrSpace::Function: storage_class = spv::StorageClassFunction; break;
-                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;  break;
-                default:
-                    assert(false && "This address space is not supported");
-                    break;
-            }
-            SpvType element = convert(ptr->pointee());
-            spv_type.id = builder_->declare_ptr_type(storage_class, element.id);
-            break;
-        }
-        case Node_IndefiniteArrayType: {
-            assert(false && "TODO");
-            // auto array = type->as<IndefiniteArrayType>();
-            // return types_[type] = spv_type;
-            THORIN_UNREACHABLE;
-        }
-        case Node_DefiniteArrayType: {
-            auto array = type->as<DefiniteArrayType>();
-            SpvType element = convert(array->elem_type());
-            SpvId size = builder_->constant(convert(world().type_pu32()).id, { (uint32_t) array->dim() });
-            spv_type.id = builder_->declare_array_type(element.id, size);
-            spv_type.size = element.size * array->dim();
-            spv_type.alignment = element.alignment;
-            break;
-        }
-
-        case Node_ClosureType:
-        case Node_FnType: {
-            // extract "return" type, collect all other types
-            auto fn = type->as<FnType>();
-            std::unique_ptr<SpvType> ret;
-            std::vector<SpvId> ops;
-            for (auto op : fn->ops()) {
-                if (op->isa<MemType>() || op == world().unit()) continue;
-                auto fn = op->isa<FnType>();
-                if (fn && !op->isa<ClosureType>()) {
-                    assert(!ret && "only one 'return' supported");
-                    std::vector<SpvType> ret_types;
-                    for (auto fn_op : fn->ops()) {
-                        if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
-                        ret_types.push_back(convert(fn_op));
-                    }
-                    if (ret_types.empty())          ret = std::make_unique<SpvType>( SpvType { { builder_->void_type }, 0, 1} );
-                    else if (ret_types.size() == 1) ret = std::make_unique<SpvType>(ret_types.back());
-                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
-                } else
-                    ops.push_back(convert(op).id);
-            }
-            assert(ret);
-
-            if (type->tag() == Node_FnType) {
-                return types_[type] = { builder_->declare_fn_type(ops, ret->id), 0, 0 };
-            }
-
-            assert(false && "TODO: handle closure mess");
-            THORIN_UNREACHABLE;
-        }
-
-        case Node_StructType: {
-            std::vector<SpvId> types;
-            for (auto elem : type->as<StructType>()->ops()) {
-                auto member_type = convert(elem);
-                types.push_back(member_type.id);
-                spv_type.size += member_type.size;
-
-                // TODO handle alignment for real
-                assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
-                spv_type.alignment = 4;
-            }
-            if (spv_type.size == 0)
-                spv_type.alignment = 1;
-            spv_type.id = builder_->declare_struct_type(types);
-            builder_->name(spv_type.id, type->to_string());
-            break;
+    for (auto& cont : world().continuations()) {
+        if (cont->is_exported()) {
+            // TODO create entry point
         }
-
-        case Node_TupleType: {
-            std::vector<SpvId> types;
-            for (auto elem : type->as<TupleType>()->ops()){
-                auto member_type = convert(elem);
-                types.push_back(member_type.id);
-                spv_type.size += member_type.size;
-
-                // TODO handle alignment for real
-                assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
-                spv_type.alignment = 4;
-            }
-            if (spv_type.size == 0)
-                spv_type.alignment = 1;
-            spv_type.id = builder_->declare_struct_type(types);
-            builder_->name(spv_type.id, type->to_string());
-            break;
-        }
-
-        case Node_VariantType: {
-            assert(type->num_ops() > 0 && "empty variants not supported");
-            std::vector<SpvId> payload_type;
-            for (auto elem : type->as<VariantType>()->ops()){
-                auto member_type = convert(elem);
-                payload_type.push_back(member_type.id);
-                spv_type.size += member_type.size;
-
-                // TODO handle alignment for real
-                assert(member_type.alignment == 4 || (member_type.size == 0 && member_type.alignment == 1));
-                spv_type.alignment = 4;
-            }
-            if (spv_type.size == 0)
-                spv_type.alignment = 1;
-            spv_type.payload_id = builder_->declare_struct_type(payload_type);
-            builder_->name(spv_type.payload_id, type->to_string() + "_payload");
-
-            std::vector<SpvId> with_tag = { convert(world().type_pu32()).id, spv_type.payload_id};
-            spv_type.id = builder_->declare_struct_type(with_tag);
-            builder_->name(spv_type.id, type->to_string());
-            break;
-        }
-
-        default:
-            THORIN_UNREACHABLE;
     }
 
-    return types_[type] = spv_type;
+    builder_->finish(out);
+    builder_ = nullptr;
 }
 
 void CodeGen::emit(const thorin::Scope& scope) {
@@ -187,7 +49,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     FnBuilder fn;
     fn.scope = &scope;
     fn.file_builder = builder_;
-    fn.fn_type = convert(entry_->type()).id;
+    fn.fn_type = convert(entry_->type()).type_id;
     fn.fn_ret_type = get_codom_type(entry_);
 
     current_fn_ = &fn;
@@ -215,10 +77,10 @@ void CodeGen::emit(const thorin::Scope& scope) {
                 if (is_mem(param) || is_unit(param)) {
                     // Nothing
                 } else if (param->order() == 0) {
-                    auto param_t = convert(param->type());
+                    auto& param_t = convert(param->type());
                     fn.header.op(spv::Op::OpFunctionParameter, 3);
                     auto id = builder_->generate_fresh_id();
-                    fn.header.ref_id(param_t.id);
+                    fn.header.ref_id(param_t.type_id);
                     fn.header.ref_id(id);
                     fn.params[param] = id;
                 }
@@ -231,7 +93,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    bb->phis_map[param] = { convert(param->type()).id, builder_->generate_fresh_id(), {} };
+                    bb->phis_map[param] = {convert(param->type()).type_id, builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -259,7 +121,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
         if (op->isa<MemType>() || is_type_unit(op))
             continue;
         assert(op->order() == 0);
-        types.push_back(convert(op).id);
+        types.push_back(convert(op).type_id);
     }
     if (types.empty())
         return builder_->void_type;
@@ -465,8 +327,8 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
         SpvId rhs = emit(bin->rhs(), bb);
-        SpvType result_types = convert(def->type());
-        SpvId result_type = result_types.id;
+        ConvertedType& result_types = convert(def->type());
+        SpvId result_type = result_types.type_id;
 
         if (auto cmp = bin->isa<Cmp>()) {
             auto type = cmp->lhs()->type();
@@ -568,7 +430,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
-        auto type = convert(def->type()).id;
+        auto type = convert(def->type()).type_id;
         SpvId constant;
         switch (primlit->primtype_tag()) {
             case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
@@ -595,31 +457,31 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             return val;
         }
     } else if (auto variant = def->isa<Variant>()) {
-        auto struct_type = def->type()->as<VariantType>();
-        std::vector<SpvId> elements;
-        elements.resize(struct_type->num_ops());
-        size_t x = 0;
-        for (auto& e : struct_type->ops()) {
-            if (x == variant->index())
-                elements[x] = emit(variant->value(), bb);
-            else
-                elements[x] = bb->undef(convert(e).id);
-            x++;
-        }
-        auto payload = bb->composite(convert(variant->type()).payload_id, elements);
-        auto tag = builder_->constant(convert(world().type_pu32()).id, { static_cast<uint32_t>(variant->index()) });
+        auto variant_type = def->type()->as<VariantType>();
+        auto& variant_datatype = (ProductDatatype&) convert(variant_type).datatype;
+
+        auto payload_arr = bb->variable(variant_datatype.elements_types[1]->type_id, spv::StorageClassFunction);
+        auto& converted_payload_type = convert(variant_type->op(variant->index()));
+        converted_payload_type.datatype->emit_serialization(*bb, payload_arr, emit(variant->value(), bb));
+        auto payload = bb->load(variant_datatype.elements_types[1]->type_id, payload_arr);
+
+        auto tag = builder_->constant(convert(world().type_pu32()).type_id, {static_cast<uint32_t>(variant->index()) });
         std::vector<SpvId> with_tag = { tag, payload };
-        return bb->composite(convert(variant->type()).id, with_tag);
+        return bb->composite(convert(variant->type()).type_id, with_tag);
     } else if (auto vextract = def->isa<VariantExtract>()) {
         auto variant_type = vextract->value()->type()->as<VariantType>();
+        auto& variant_datatype = (ProductDatatype&) convert(variant_type).datatype;
+
+        auto& target_type = convert(def->type());
 
-        auto target_type = convert(def->type());
-        auto payload = bb->extract(convert(variant_type).payload_id, emit(vextract->value(), bb), {1});
+        auto payload_arr = bb->variable(variant_datatype.elements_types[1]->type_id, spv::StorageClassFunction);
+        auto payload = bb->extract(variant_datatype.elements_types[1]->type_id, emit(vextract->value(), bb), {1});
+        bb->store(payload, payload_arr);
 
-        return bb->extract(target_type.id, payload, { static_cast<uint32_t>(vextract->index()) });
+        return target_type.datatype->emit_deserialization(*bb, payload_arr);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
-        return bb->extract(convert(world().type_pu32()).id, value, { 0 });
+        return bb->extract(convert(world().type_pu32()).type_id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
         std::vector<SpvId> elements;
         elements.resize(tuple->num_ops());
@@ -627,7 +489,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         for (auto& e : tuple->ops()) {
             elements[x++] = emit(e, bb);
         }
-        return bb->composite(convert(tuple->type()).id, elements);
+        return bb->composite(convert(tuple->type()).type_id, elements);
     } else if (auto structagg = def->isa<StructAgg>()) {
         std::vector<SpvId> elements;
         elements.resize(structagg->num_ops());
@@ -635,7 +497,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         for (auto& e : structagg->ops()) {
             elements[x++] = emit(e, bb);
         }
-        return bb->composite(convert(structagg->type()).id, elements);
+        return bb->composite(convert(structagg->type()).type_id, elements);
     }
     assertf(false, "Incomplete emit(def) definition");
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index fda32aa2d..07765d5ee 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -8,14 +8,18 @@ namespace thorin::spirv {
 
 using SpvId = builder::SpvId;
 
-struct SpvType {
-    SpvId id;
-    size_t size = 0;
+class CodeGen;
+struct Datatype;
 
-    // TODO: Alignment rules are complicated and client API dependant
-    size_t alignment = 0;
+struct ConvertedType {
+    CodeGen* code_gen;
+    SpvId type_id { 0 };
+    std::unique_ptr<Datatype> datatype;
 
-    SpvId payload_id = { 0 };
+    // TODO: delete
+    // SpvId payload_id { 0 };
+
+    bool is_known_size() { return datatype != nullptr; }
 };
 
 struct FnBuilder;
@@ -42,11 +46,12 @@ class CodeGen : public thorin::CodeGen {
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
+
+    ConvertedType& convert(const Type*);
 protected:
     void structure_loops();
     void structure_flow();
 
-    SpvType convert(const Type*);
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     SpvId emit(const Def* def, BasicBlockBuilder* bb);
@@ -56,8 +61,57 @@ class CodeGen : public thorin::CodeGen {
     builder::SpvFileBuilder* builder_ = nullptr;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
-    TypeMap<SpvType> types_;
+    TypeMap<ConvertedType> types_;
     DefMap<SpvId> defs_;
+
+};
+
+/// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
+/// potentially different mappings, depending on the capabilities of the target platform. The serdes code deals with pointers
+/// in arrays of unsigned 32 bit words, and is there to get around the limitation of not being able to bitcast pointers in the
+/// logical addressing mode.
+struct Datatype {
+public:
+    ConvertedType& type;
+    Datatype(ConvertedType& type) : type(type) {}
+
+    virtual size_t serialized_size() = 0;
+    virtual void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) = 0;
+    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) = 0;
+};
+
+/// For scalar datatypes
+struct ScalarDatatype : public Datatype {
+    int type_tag;
+    size_t size_in_bytes;
+    size_t alignment;
+    ScalarDatatype(ConvertedType& type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
+
+    size_t serialized_size() override { return size_in_bytes / 4; };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
+};
+
+struct DefiniteArrayDatatype : public Datatype {
+    ConvertedType& element_type;
+    size_t length;
+
+    DefiniteArrayDatatype(ConvertedType& type, ConvertedType& element_type, size_t length);
+
+    size_t serialized_size() override { return element_type.datatype->serialized_size(); };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
+};
+
+struct ProductDatatype : public Datatype {
+    std::vector<ConvertedType*> elements_types;
+    size_t total_size = 0;
+
+    ProductDatatype(ConvertedType& type, const std::vector<ConvertedType*>&& elements_types);
+
+    size_t serialized_size() override { return total_size; };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
 };
 
 }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 773af85a2..da1cc6eca 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -124,6 +124,18 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId>& indexes) {
+        op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(base);
+        ref_id(element);
+        for (auto index : indexes)
+            ref_id(index);
+        return id;
+    }
+
     SpvId load(SpvId target_type, SpvId pointer) {
         op(spv::Op::OpLoad, 4);
         auto id = generate_fresh_id();
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
new file mode 100644
index 000000000..60bf2c013
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -0,0 +1,219 @@
+#include "thorin/be/spirv/spirv.h"
+
+namespace thorin::spirv {
+
+ScalarDatatype::ScalarDatatype(ConvertedType& type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes)
+: Datatype(type), type_tag(type_tag), size_in_bytes(size_in_bytes), alignment(alignment_in_bytes)
+{
+    /// currently limited to 32-bit
+    assert(size_in_bytes == 4);
+}
+
+SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+    auto loaded = bb.load(type.type_id, input);
+    return bb.bitcast(type.type_id, loaded);
+}
+
+void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+    SpvId u32_tid = type.code_gen->convert(type.code_gen->world().type_pu32()).type_id;
+    auto casted = bb.bitcast(u32_tid, data);
+    bb.store(casted, output);
+}
+
+DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType& type, ConvertedType& element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
+    assert(element_type.datatype.get() != nullptr);
+}
+
+SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+    SpvId i32_tid = type.code_gen->convert(type.code_gen->world().type_ps32()).type_id;
+    std::vector<SpvId> indices;
+    std::vector<SpvId> elements;
+    for (size_t i = 0; i < length; i++) {
+        SpvId element_ptr = bb.ptr_access_chain(element_type.type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type.datatype->serialized_size()) }), indices);
+        SpvId element = element_type.datatype->emit_deserialization(bb, element_ptr);
+        elements.push_back(element);
+    }
+    return bb.composite(type.type_id, elements);
+}
+void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+    THORIN_UNREACHABLE;
+}
+
+ProductDatatype::ProductDatatype(ConvertedType& type, const std::vector<ConvertedType*>&& elements_types) : Datatype(type), elements_types(elements_types) {
+    for (auto& element_type : elements_types) {
+        total_size += element_type->datatype->serialized_size();
+    }
+}
+
+SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+    SpvId i32_tid = type.code_gen->convert(type.code_gen->world().type_pu32()).type_id;
+    std::vector<SpvId> indices;
+    std::vector<SpvId> elements;
+    size_t offset = 0;
+    for (auto& element_type : elements_types) {
+        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) offset }), indices);
+        SpvId element = element_type->datatype->emit_deserialization(bb, element_ptr);
+        offset += element_type->datatype->serialized_size();
+        elements.push_back(element);
+    }
+    return bb.composite(type.type_id, elements);
+}
+void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+    THORIN_UNREACHABLE;
+}
+
+ConvertedType& CodeGen::convert(const Type* type) {
+    if (auto iter = types_.find(type); iter != types_.end()) return iter->second;
+
+    assert(!type->isa<MemType>());
+    ConvertedType& converted = types_.emplace(type, ConvertedType {this } ).first->second;
+    switch (type->tag()) {
+        // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
+        // Using extensions, we could use 16 or 8-bit ints instead
+        // We can also pack them inside structures using bit-twiddling tricks, if the need arises
+        case PrimType_bool:
+            converted.type_id = builder_->declare_bool_type();
+            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 1, 4);
+            break;
+        case PrimType_ps8:  case PrimType_qs8:  case PrimType_pu8:  case PrimType_qu8:  assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps16: case PrimType_qs16: case PrimType_pu16: case PrimType_qu16: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps32: case PrimType_qs32:
+            converted.type_id = builder_->declare_int_type(32, true );
+            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            break;
+        case PrimType_pu32: case PrimType_qu32:
+            converted.type_id = builder_->declare_int_type(32, false);
+            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            break;
+        case PrimType_ps64: case PrimType_qs64: case PrimType_pu64: case PrimType_qu64: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf16: case PrimType_qf16:                                         assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf32: case PrimType_qf32:
+            converted.type_id = builder_->declare_float_type(32);
+            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            break;
+        case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
+        case Node_PtrType: {
+            auto ptr = type->as<PtrType>();
+            spv::StorageClass storage_class;
+            switch (ptr->addr_space()) {
+                case AddrSpace::Function: storage_class = spv::StorageClassFunction; break;
+                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;  break;
+                default:
+                    assert(false && "This address space is not supported");
+                    break;
+            }
+            ConvertedType& element = convert(ptr->pointee());
+            converted.type_id = builder_->declare_ptr_type(storage_class, element.type_id);
+            break;
+        }
+        case Node_IndefiniteArrayType: {
+            assert(false && "TODO");
+            // auto array = type->as<IndefiniteArrayType>();
+            // return types_[type] = spv_type;
+            THORIN_UNREACHABLE;
+        }
+        case Node_DefiniteArrayType: {
+            auto array = type->as<DefiniteArrayType>();
+            ConvertedType& element = convert(array->elem_type());
+            SpvId size = builder_->constant(convert(world().type_pu32()).type_id, {(uint32_t) array->dim() });
+            converted.type_id = builder_->declare_array_type(element.type_id, size);
+            converted.datatype = std::make_unique<DefiniteArrayDatatype>(converted, element, array->dim());
+            break;
+        }
+
+        case Node_ClosureType:
+        case Node_FnType: {
+            // extract "return" type, collect all other types
+            auto fn = type->as<FnType>();
+            ConvertedType* ret = nullptr;
+            std::vector<SpvId> ops;
+            for (auto op : fn->ops()) {
+                if (op->isa<MemType>() || op == world().unit()) continue;
+                auto fn_type = op->isa<FnType>();
+                if (fn_type && !op->isa<ClosureType>()) {
+                    assert(!ret && "only one 'return' supported");
+                    std::vector<ConvertedType*> ret_types;
+                    for (auto fn_op : fn_type->ops()) {
+                        if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
+                        ret_types.push_back(&convert(fn_op));
+                    }
+                    if (ret_types.empty())          ret = &convert(world().tuple_type({}));
+                    else if (ret_types.size() == 1) ret = ret_types.back();
+                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
+                } else
+                    ops.push_back(convert(op).type_id);
+            }
+            assert(ret);
+
+            if (type->tag() == Node_FnType) {
+                converted.type_id = builder_->declare_fn_type(ops, ret->type_id);
+            } else {
+                assert(false && "TODO: handle closure mess");
+                THORIN_UNREACHABLE;
+            }
+            break;
+        }
+
+        case Node_StructType: {
+            std::vector<ConvertedType*> types;
+            std::vector<SpvId> spv_types;
+            for (auto elem : type->as<StructType>()->ops()) {
+                auto& member_type = convert(elem);
+                types.push_back(&member_type);
+                spv_types.push_back(member_type.type_id);
+            }
+            converted.type_id = builder_->declare_struct_type(spv_types);
+            builder_->name(converted.type_id, type->to_string());
+            converted.datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
+            break;
+        }
+
+        case Node_TupleType: {
+            std::vector<ConvertedType*> types;
+            std::vector<SpvId> spv_types;
+            for (auto elem : type->as<TupleType>()->ops()){
+                auto& member_type = convert(elem);
+                types.push_back(&member_type);
+                spv_types.push_back(member_type.type_id);
+            }
+            converted.type_id = builder_->declare_struct_type(spv_types);
+            builder_->name(converted.type_id, type->to_string());
+            converted.datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
+            break;
+        }
+
+        case Node_VariantType: {
+            assert(type->num_ops() > 0 && "empty variants not supported");
+            auto tag_type = world().type_pu32();
+            ConvertedType& converted_tag_type = convert(tag_type);
+
+            size_t max_serialized_size = 0;
+            //std::vector<ConvertedType*> types;
+            //std::vector<SpvId> spv_types;
+            for (auto elem : type->as<VariantType>()->ops()){
+                auto& member_type = convert(elem);
+                //types.push_back(&member_type);
+                //spv_types.push_back(member_type.type_id);
+
+                if (member_type.datatype->serialized_size() > max_serialized_size)
+                    max_serialized_size = member_type.datatype->serialized_size();
+            }
+
+            auto payload_type = world().definite_array_type(world().type_pu32(), max_serialized_size);
+            auto& converted_payload_type = convert(payload_type);
+
+            std::vector<SpvId> spv_pair = {converted_tag_type.type_id, converted_payload_type.type_id };
+            converted.type_id = builder_->declare_struct_type(spv_pair);
+            converted.datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*> { &converted_tag_type, &converted_payload_type });
+            builder_->name(converted.type_id, type->to_string());
+            break;
+        }
+
+        default:
+            THORIN_UNREACHABLE;
+    }
+
+    return converted;
+}
+
+}
\ No newline at end of file

From 5e1b55ac64f46d88359a8d03801d4370510aa5a4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 16 Apr 2021 14:19:52 +0200
Subject: [PATCH 052/342] use pointers for ConvertedType

---
 src/thorin/be/spirv/spirv.cpp           |  46 ++++-----
 src/thorin/be/spirv/spirv.h             |  22 ++--
 src/thorin/be/spirv/spirv_datatypes.cpp | 131 +++++++++++++-----------
 3 files changed, 105 insertions(+), 94 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 15a1cdb05..20207ea06 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -49,7 +49,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     FnBuilder fn;
     fn.scope = &scope;
     fn.file_builder = builder_;
-    fn.fn_type = convert(entry_->type()).type_id;
+    fn.fn_type = convert(entry_->type())->type_id;
     fn.fn_ret_type = get_codom_type(entry_);
 
     current_fn_ = &fn;
@@ -77,10 +77,10 @@ void CodeGen::emit(const thorin::Scope& scope) {
                 if (is_mem(param) || is_unit(param)) {
                     // Nothing
                 } else if (param->order() == 0) {
-                    auto& param_t = convert(param->type());
+                    auto param_t = convert(param->type());
                     fn.header.op(spv::Op::OpFunctionParameter, 3);
                     auto id = builder_->generate_fresh_id();
-                    fn.header.ref_id(param_t.type_id);
+                    fn.header.ref_id(param_t->type_id);
                     fn.header.ref_id(id);
                     fn.params[param] = id;
                 }
@@ -93,7 +93,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    bb->phis_map[param] = {convert(param->type()).type_id, builder_->generate_fresh_id(), {} };
+                    bb->phis_map[param] = {convert(param->type())->type_id, builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -121,7 +121,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
         if (op->isa<MemType>() || is_type_unit(op))
             continue;
         assert(op->order() == 0);
-        types.push_back(convert(op).type_id);
+        types.push_back(convert(op)->type_id);
     }
     if (types.empty())
         return builder_->void_type;
@@ -327,8 +327,8 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
         SpvId rhs = emit(bin->rhs(), bb);
-        ConvertedType& result_types = convert(def->type());
-        SpvId result_type = result_types.type_id;
+        ConvertedType* result_types = convert(def->type());
+        SpvId result_type = result_types->type_id;
 
         if (auto cmp = bin->isa<Cmp>()) {
             auto type = cmp->lhs()->type();
@@ -430,7 +430,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
-        auto type = convert(def->type()).type_id;
+        auto type = convert(def->type())->type_id;
         SpvId constant;
         switch (primlit->primtype_tag()) {
             case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
@@ -458,30 +458,30 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto variant = def->isa<Variant>()) {
         auto variant_type = def->type()->as<VariantType>();
-        auto& variant_datatype = (ProductDatatype&) convert(variant_type).datatype;
+        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
-        auto payload_arr = bb->variable(variant_datatype.elements_types[1]->type_id, spv::StorageClassFunction);
-        auto& converted_payload_type = convert(variant_type->op(variant->index()));
-        converted_payload_type.datatype->emit_serialization(*bb, payload_arr, emit(variant->value(), bb));
-        auto payload = bb->load(variant_datatype.elements_types[1]->type_id, payload_arr);
+        auto payload_arr = bb->variable(variant_datatype->elements_types[1]->type_id, spv::StorageClassFunction);
+        auto converted_payload_type = convert(variant_type->op(variant->index()));
+        converted_payload_type->datatype->emit_serialization(*bb, payload_arr, emit(variant->value(), bb));
+        auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
 
-        auto tag = builder_->constant(convert(world().type_pu32()).type_id, {static_cast<uint32_t>(variant->index()) });
+        auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index()) });
         std::vector<SpvId> with_tag = { tag, payload };
-        return bb->composite(convert(variant->type()).type_id, with_tag);
+        return bb->composite(convert(variant->type())->type_id, with_tag);
     } else if (auto vextract = def->isa<VariantExtract>()) {
         auto variant_type = vextract->value()->type()->as<VariantType>();
-        auto& variant_datatype = (ProductDatatype&) convert(variant_type).datatype;
+        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
-        auto& target_type = convert(def->type());
+        auto target_type = convert(def->type());
 
-        auto payload_arr = bb->variable(variant_datatype.elements_types[1]->type_id, spv::StorageClassFunction);
-        auto payload = bb->extract(variant_datatype.elements_types[1]->type_id, emit(vextract->value(), bb), {1});
+        auto payload_arr = bb->variable(variant_datatype->elements_types[1]->type_id, spv::StorageClassFunction);
+        auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
-        return target_type.datatype->emit_deserialization(*bb, payload_arr);
+        return target_type->datatype->emit_deserialization(*bb, payload_arr);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
-        return bb->extract(convert(world().type_pu32()).type_id, value, { 0 });
+        return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
         std::vector<SpvId> elements;
         elements.resize(tuple->num_ops());
@@ -489,7 +489,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         for (auto& e : tuple->ops()) {
             elements[x++] = emit(e, bb);
         }
-        return bb->composite(convert(tuple->type()).type_id, elements);
+        return bb->composite(convert(tuple->type())->type_id, elements);
     } else if (auto structagg = def->isa<StructAgg>()) {
         std::vector<SpvId> elements;
         elements.resize(structagg->num_ops());
@@ -497,7 +497,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         for (auto& e : structagg->ops()) {
             elements[x++] = emit(e, bb);
         }
-        return bb->composite(convert(structagg->type()).type_id, elements);
+        return bb->composite(convert(structagg->type())->type_id, elements);
     }
     assertf(false, "Incomplete emit(def) definition");
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 07765d5ee..baf2c4909 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -16,9 +16,7 @@ struct ConvertedType {
     SpvId type_id { 0 };
     std::unique_ptr<Datatype> datatype;
 
-    // TODO: delete
-    // SpvId payload_id { 0 };
-
+    ConvertedType(CodeGen* cg) : code_gen(cg) {}
     bool is_known_size() { return datatype != nullptr; }
 };
 
@@ -47,7 +45,7 @@ class CodeGen : public thorin::CodeGen {
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
 
-    ConvertedType& convert(const Type*);
+    ConvertedType* convert(const Type*);
 protected:
     void structure_loops();
     void structure_flow();
@@ -61,7 +59,7 @@ class CodeGen : public thorin::CodeGen {
     builder::SpvFileBuilder* builder_ = nullptr;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
-    TypeMap<ConvertedType> types_;
+    TypeMap<std::unique_ptr<ConvertedType>> types_;
     DefMap<SpvId> defs_;
 
 };
@@ -72,8 +70,8 @@ class CodeGen : public thorin::CodeGen {
 /// logical addressing mode.
 struct Datatype {
 public:
-    ConvertedType& type;
-    Datatype(ConvertedType& type) : type(type) {}
+    ConvertedType* type;
+    Datatype(ConvertedType* type) : type(type) {}
 
     virtual size_t serialized_size() = 0;
     virtual void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) = 0;
@@ -85,7 +83,7 @@ struct ScalarDatatype : public Datatype {
     int type_tag;
     size_t size_in_bytes;
     size_t alignment;
-    ScalarDatatype(ConvertedType& type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
+    ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
 
     size_t serialized_size() override { return size_in_bytes / 4; };
     SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
@@ -93,12 +91,12 @@ struct ScalarDatatype : public Datatype {
 };
 
 struct DefiniteArrayDatatype : public Datatype {
-    ConvertedType& element_type;
+    ConvertedType* element_type;
     size_t length;
 
-    DefiniteArrayDatatype(ConvertedType& type, ConvertedType& element_type, size_t length);
+    DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length);
 
-    size_t serialized_size() override { return element_type.datatype->serialized_size(); };
+    size_t serialized_size() override { return element_type->datatype->serialized_size(); };
     SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
     void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
 };
@@ -107,7 +105,7 @@ struct ProductDatatype : public Datatype {
     std::vector<ConvertedType*> elements_types;
     size_t total_size = 0;
 
-    ProductDatatype(ConvertedType& type, const std::vector<ConvertedType*>&& elements_types);
+    ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types);
 
     size_t serialized_size() override { return total_size; };
     SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 60bf2c013..04c5c18f0 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -2,7 +2,7 @@
 
 namespace thorin::spirv {
 
-ScalarDatatype::ScalarDatatype(ConvertedType& type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes)
+ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes)
 : Datatype(type), type_tag(type_tag), size_in_bytes(size_in_bytes), alignment(alignment_in_bytes)
 {
     /// currently limited to 32-bit
@@ -10,43 +10,48 @@ ScalarDatatype::ScalarDatatype(ConvertedType& type, int type_tag, size_t size_in
 }
 
 SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
-    auto loaded = bb.load(type.type_id, input);
-    return bb.bitcast(type.type_id, loaded);
+    auto loaded = bb.load(type->type_id, input);
+    return bb.bitcast(type->type_id, loaded);
 }
 
 void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
-    SpvId u32_tid = type.code_gen->convert(type.code_gen->world().type_pu32()).type_id;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     auto casted = bb.bitcast(u32_tid, data);
     bb.store(casted, output);
 }
 
-DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType& type, ConvertedType& element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
-    assert(element_type.datatype.get() != nullptr);
+DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
+    assert(element_type->datatype.get() != nullptr);
 }
 
 SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
-    SpvId i32_tid = type.code_gen->convert(type.code_gen->world().type_ps32()).type_id;
+    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_ps32())->type_id;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     for (size_t i = 0; i < length; i++) {
-        SpvId element_ptr = bb.ptr_access_chain(element_type.type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type.datatype->serialized_size()) }), indices);
-        SpvId element = element_type.datatype->emit_deserialization(bb, element_ptr);
+        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type->datatype->serialized_size()) }), indices);
+        SpvId element = element_type->datatype->emit_deserialization(bb, element_ptr);
         elements.push_back(element);
     }
-    return bb.composite(type.type_id, elements);
+    return bb.composite(type->type_id, elements);
 }
 void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
-    THORIN_UNREACHABLE;
+    std::vector<SpvId> indices;
+    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_ps32())->type_id;
+    for (size_t i = 0; i < length; i++) {
+        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, output, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type->datatype->serialized_size()) }), indices);
+        element_type->datatype->emit_serialization(bb, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i }));
+    }
 }
 
-ProductDatatype::ProductDatatype(ConvertedType& type, const std::vector<ConvertedType*>&& elements_types) : Datatype(type), elements_types(elements_types) {
+ProductDatatype::ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types) : Datatype(type), elements_types(elements_types) {
     for (auto& element_type : elements_types) {
         total_size += element_type->datatype->serialized_size();
     }
 }
 
 SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
-    SpvId i32_tid = type.code_gen->convert(type.code_gen->world().type_pu32()).type_id;
+    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     size_t offset = 0;
@@ -56,40 +61,47 @@ SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input)
         offset += element_type->datatype->serialized_size();
         elements.push_back(element);
     }
-    return bb.composite(type.type_id, elements);
+    return bb.composite(type->type_id, elements);
 }
 void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
-    THORIN_UNREACHABLE;
+    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    std::vector<SpvId> indices;
+    size_t offset = 0;
+    int i = 0;
+    for (auto& element_type : elements_types) {
+        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, output, bb.file_builder.constant(i32_tid, { (uint32_t) offset }), indices);
+        element_type->datatype->emit_serialization(bb, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
+    }
 }
 
-ConvertedType& CodeGen::convert(const Type* type) {
-    if (auto iter = types_.find(type); iter != types_.end()) return iter->second;
+ConvertedType* CodeGen::convert(const Type* type) {
+    if (auto iter = types_.find(type); iter != types_.end()) return iter->second.get();
 
     assert(!type->isa<MemType>());
-    ConvertedType& converted = types_.emplace(type, ConvertedType {this } ).first->second;
+    ConvertedType* converted = types_.emplace(type, std::make_unique<ConvertedType>(this) ).first->second.get();
     switch (type->tag()) {
         // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
         // Using extensions, we could use 16 or 8-bit ints instead
         // We can also pack them inside structures using bit-twiddling tricks, if the need arises
         case PrimType_bool:
-            converted.type_id = builder_->declare_bool_type();
-            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 1, 4);
+            converted->type_id = builder_->declare_bool_type();
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
         case PrimType_ps8:  case PrimType_qs8:  case PrimType_pu8:  case PrimType_qu8:  assert(false && "TODO: look into capabilities to enable this");
         case PrimType_ps16: case PrimType_qs16: case PrimType_pu16: case PrimType_qu16: assert(false && "TODO: look into capabilities to enable this");
         case PrimType_ps32: case PrimType_qs32:
-            converted.type_id = builder_->declare_int_type(32, true );
-            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            converted->type_id = builder_->declare_int_type(32, true );
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
         case PrimType_pu32: case PrimType_qu32:
-            converted.type_id = builder_->declare_int_type(32, false);
-            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            converted->type_id = builder_->declare_int_type(32, false);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
         case PrimType_ps64: case PrimType_qs64: case PrimType_pu64: case PrimType_qu64: assert(false && "TODO: look into capabilities to enable this");
         case PrimType_pf16: case PrimType_qf16:                                         assert(false && "TODO: look into capabilities to enable this");
         case PrimType_pf32: case PrimType_qf32:
-            converted.type_id = builder_->declare_float_type(32);
-            converted.datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            converted->type_id = builder_->declare_float_type(32);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
         case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
         case Node_PtrType: {
@@ -102,8 +114,8 @@ ConvertedType& CodeGen::convert(const Type* type) {
                     assert(false && "This address space is not supported");
                     break;
             }
-            ConvertedType& element = convert(ptr->pointee());
-            converted.type_id = builder_->declare_ptr_type(storage_class, element.type_id);
+            ConvertedType* element = convert(ptr->pointee());
+            converted->type_id = builder_->declare_ptr_type(storage_class, element->type_id);
             break;
         }
         case Node_IndefiniteArrayType: {
@@ -114,10 +126,10 @@ ConvertedType& CodeGen::convert(const Type* type) {
         }
         case Node_DefiniteArrayType: {
             auto array = type->as<DefiniteArrayType>();
-            ConvertedType& element = convert(array->elem_type());
-            SpvId size = builder_->constant(convert(world().type_pu32()).type_id, {(uint32_t) array->dim() });
-            converted.type_id = builder_->declare_array_type(element.type_id, size);
-            converted.datatype = std::make_unique<DefiniteArrayDatatype>(converted, element, array->dim());
+            ConvertedType* element = convert(array->elem_type());
+            SpvId size = builder_->constant(convert(world().type_pu32())->type_id, {(uint32_t) array->dim() });
+            converted->type_id = builder_->declare_array_type(element->type_id, size);
+            converted->datatype = std::make_unique<DefiniteArrayDatatype>(converted, element, array->dim());
             break;
         }
 
@@ -135,18 +147,18 @@ ConvertedType& CodeGen::convert(const Type* type) {
                     std::vector<ConvertedType*> ret_types;
                     for (auto fn_op : fn_type->ops()) {
                         if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
-                        ret_types.push_back(&convert(fn_op));
+                        ret_types.push_back(convert(fn_op));
                     }
-                    if (ret_types.empty())          ret = &convert(world().tuple_type({}));
+                    if (ret_types.empty())          ret = convert(world().tuple_type({}));
                     else if (ret_types.size() == 1) ret = ret_types.back();
                     else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
                 } else
-                    ops.push_back(convert(op).type_id);
+                    ops.push_back(convert(op)->type_id);
             }
             assert(ret);
 
             if (type->tag() == Node_FnType) {
-                converted.type_id = builder_->declare_fn_type(ops, ret->type_id);
+                converted->type_id = builder_->declare_fn_type(ops, ret->type_id);
             } else {
                 assert(false && "TODO: handle closure mess");
                 THORIN_UNREACHABLE;
@@ -158,13 +170,13 @@ ConvertedType& CodeGen::convert(const Type* type) {
             std::vector<ConvertedType*> types;
             std::vector<SpvId> spv_types;
             for (auto elem : type->as<StructType>()->ops()) {
-                auto& member_type = convert(elem);
-                types.push_back(&member_type);
-                spv_types.push_back(member_type.type_id);
+                auto member_type = convert(elem);
+                types.push_back(member_type);
+                spv_types.push_back(member_type->type_id);
             }
-            converted.type_id = builder_->declare_struct_type(spv_types);
-            builder_->name(converted.type_id, type->to_string());
-            converted.datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
+            converted->type_id = builder_->declare_struct_type(spv_types);
+            builder_->name(converted->type_id, type->to_string());
+            converted->datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
             break;
         }
 
@@ -172,40 +184,41 @@ ConvertedType& CodeGen::convert(const Type* type) {
             std::vector<ConvertedType*> types;
             std::vector<SpvId> spv_types;
             for (auto elem : type->as<TupleType>()->ops()){
-                auto& member_type = convert(elem);
-                types.push_back(&member_type);
-                spv_types.push_back(member_type.type_id);
+                auto member_type = convert(elem);
+                types.push_back(member_type);
+                spv_types.push_back(member_type->type_id);
             }
-            converted.type_id = builder_->declare_struct_type(spv_types);
-            builder_->name(converted.type_id, type->to_string());
-            converted.datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
+            converted->type_id = builder_->declare_struct_type(spv_types);
+            builder_->name(converted->type_id, type->to_string());
+            converted->datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
             break;
         }
 
         case Node_VariantType: {
             assert(type->num_ops() > 0 && "empty variants not supported");
             auto tag_type = world().type_pu32();
-            ConvertedType& converted_tag_type = convert(tag_type);
+            ConvertedType* converted_tag_type = convert(tag_type);
 
             size_t max_serialized_size = 0;
             //std::vector<ConvertedType*> types;
             //std::vector<SpvId> spv_types;
             for (auto elem : type->as<VariantType>()->ops()){
-                auto& member_type = convert(elem);
-                //types.push_back(&member_type);
-                //spv_types.push_back(member_type.type_id);
+                auto member_type = convert(elem);
 
-                if (member_type.datatype->serialized_size() > max_serialized_size)
-                    max_serialized_size = member_type.datatype->serialized_size();
+                if (member_type->datatype->serialized_size() > max_serialized_size)
+                    max_serialized_size = member_type->datatype->serialized_size();
             }
 
             auto payload_type = world().definite_array_type(world().type_pu32(), max_serialized_size);
-            auto& converted_payload_type = convert(payload_type);
+            auto* converted_payload_type = convert(payload_type);
+
+            std::vector<SpvId> spv_pair = {converted_tag_type->type_id, converted_payload_type->type_id };
+            converted->type_id = builder_->declare_struct_type(spv_pair);
+
+            // auto oh_god_why = std::vector<ConvertedType*> ( &converted_tag_type, &converted_payload_type );
 
-            std::vector<SpvId> spv_pair = {converted_tag_type.type_id, converted_payload_type.type_id };
-            converted.type_id = builder_->declare_struct_type(spv_pair);
-            converted.datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*> { &converted_tag_type, &converted_payload_type });
-            builder_->name(converted.type_id, type->to_string());
+            converted->datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*> { converted_tag_type, converted_payload_type });
+            builder_->name(converted->type_id, type->to_string());
             break;
         }
 

From 93ffecae1189dbba7f2505a66054e2c19a1f3ead Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 19 Apr 2021 09:03:38 +0200
Subject: [PATCH 053/342] fixup serialization

---
 src/thorin/be/spirv/spirv.cpp           | 36 +++++++----
 src/thorin/be/spirv/spirv.h             |  3 +-
 src/thorin/be/spirv/spirv_builder.hpp   |  2 +-
 src/thorin/be/spirv/spirv_datatypes.cpp | 79 ++++++++++++++-----------
 4 files changed, 75 insertions(+), 45 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 20207ea06..257220349 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -460,25 +460,41 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto variant_type = def->type()->as<VariantType>();
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
-        auto payload_arr = bb->variable(variant_datatype->elements_types[1]->type_id, spv::StorageClassFunction);
-        auto converted_payload_type = convert(variant_type->op(variant->index()));
-        converted_payload_type->datatype->emit_serialization(*bb, payload_arr, emit(variant->value(), bb));
-        auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
-
-        auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index()) });
-        std::vector<SpvId> with_tag = { tag, payload };
-        return bb->composite(convert(variant->type())->type_id, with_tag);
+        if (variant_datatype->elements_types.size() > 1) {
+            auto ptr_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+            auto payload_arr = bb->variable(ptr_type, spv::StorageClassFunction);
+            auto converted_payload_type = convert(variant_type->op(variant->index()));
+
+            auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
+            auto ptr_arr = bb->ptr_access_chain(convert(world().type_pu32())->type_id, payload_arr, zero, { zero });
+
+            converted_payload_type->datatype->emit_serialization(*bb, ptr_arr, emit(variant->value(), bb));
+            auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
+
+            auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index())});
+            std::vector<SpvId> with_tag = {tag, payload};
+            return bb->composite(convert(variant->type())->type_id, with_tag);
+        } else {
+            // Zero-sized payload case
+            auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index())});
+            std::vector<SpvId> with_tag = { tag };
+            return bb->composite(convert(variant->type())->type_id, with_tag);
+        }
     } else if (auto vextract = def->isa<VariantExtract>()) {
         auto variant_type = vextract->value()->type()->as<VariantType>();
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
         auto target_type = convert(def->type());
 
-        auto payload_arr = bb->variable(variant_datatype->elements_types[1]->type_id, spv::StorageClassFunction);
+        assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
+        auto ptr_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+        auto payload_arr = bb->variable(ptr_type, spv::StorageClassFunction);
         auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
-        return target_type->datatype->emit_deserialization(*bb, payload_arr);
+        auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
+        auto ptr_arr = bb->ptr_access_chain(convert(world().type_pu32())->type_id, payload_arr, zero, { zero });
+        return target_type->datatype->emit_deserialization(*bb, ptr_arr);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
         return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index baf2c4909..c45232939 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -12,7 +12,8 @@ class CodeGen;
 struct Datatype;
 
 struct ConvertedType {
-    CodeGen* code_gen;
+    spirv::CodeGen* code_gen;
+    const thorin::Type* src_type;
     SpvId type_id { 0 };
     std::unique_ptr<Datatype> datatype;
 
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index da1cc6eca..9454dc618 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -124,7 +124,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId>& indexes) {
+    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
         op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
         auto id = generate_fresh_id();
         ref_id(target_type);
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 04c5c18f0..85763df97 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -1,4 +1,5 @@
 #include "thorin/be/spirv/spirv.h"
+#include "thorin/util/stream.h"
 
 namespace thorin::spirv {
 
@@ -22,6 +23,7 @@ void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, Spv
 
 DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
     assert(element_type->datatype.get() != nullptr);
+    assert(length > 0 && "Array lengths of zero are not supported");
 }
 
 SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
@@ -51,6 +53,7 @@ ProductDatatype::ProductDatatype(ConvertedType* type, const std::vector<Converte
 }
 
 SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+    assert(total_size > 0 && "It doesn't make sense to de-serialize Unit!");
     SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
@@ -64,6 +67,7 @@ SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input)
     return bb.composite(type->type_id, elements);
 }
 void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+    assert(total_size > 0 && "It doesn't make sense to serialize Unit!");
     SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     size_t offset = 0;
@@ -79,6 +83,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
 
     assert(!type->isa<MemType>());
     ConvertedType* converted = types_.emplace(type, std::make_unique<ConvertedType>(this) ).first->second.get();
+    converted->src_type = type;
     switch (type->tag()) {
         // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
         // Using extensions, we could use 16 or 8-bit ints instead
@@ -166,28 +171,27 @@ ConvertedType* CodeGen::convert(const Type* type) {
             break;
         }
 
-        case Node_StructType: {
-            std::vector<ConvertedType*> types;
-            std::vector<SpvId> spv_types;
-            for (auto elem : type->as<StructType>()->ops()) {
-                auto member_type = convert(elem);
-                types.push_back(member_type);
-                spv_types.push_back(member_type->type_id);
-            }
-            converted->type_id = builder_->declare_struct_type(spv_types);
-            builder_->name(converted->type_id, type->to_string());
-            converted->datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
-            break;
-        }
-
+        case Node_StructType:
         case Node_TupleType: {
             std::vector<ConvertedType*> types;
             std::vector<SpvId> spv_types;
-            for (auto elem : type->as<TupleType>()->ops()){
-                auto member_type = convert(elem);
-                types.push_back(member_type);
-                spv_types.push_back(member_type->type_id);
+            size_t total_serialized_size = 0;
+            for (auto member_type : type->ops()) {
+                if (member_type == world().unit() || member_type == world().mem_type()) {
+                    outf("skipped one");
+                    continue;
+                }
+                auto converted_member_type = convert(member_type);
+                types.push_back(converted_member_type);
+                spv_types.push_back(converted_member_type->type_id);
+                total_serialized_size = converted_member_type->datatype->serialized_size();
             }
+            if (total_serialized_size == 0) {
+                outf("this one is void");
+                converted->type_id = builder_->void_type;
+                break;
+            }
+
             converted->type_id = builder_->declare_struct_type(spv_types);
             builder_->name(converted->type_id, type->to_string());
             converted->datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
@@ -200,28 +204,37 @@ ConvertedType* CodeGen::convert(const Type* type) {
             ConvertedType* converted_tag_type = convert(tag_type);
 
             size_t max_serialized_size = 0;
-            //std::vector<ConvertedType*> types;
-            //std::vector<SpvId> spv_types;
-            for (auto elem : type->as<VariantType>()->ops()){
-                auto member_type = convert(elem);
-
-                if (member_type->datatype->serialized_size() > max_serialized_size)
-                    max_serialized_size = member_type->datatype->serialized_size();
+            for (auto member_type : type->as<VariantType>()->ops()) {
+                if (member_type == world().unit() || member_type == world().mem_type()) {
+                    outf("skipped one");
+                    continue;
+                }
+                auto converted_member_type = convert(member_type);
+                if (converted_member_type->datatype->serialized_size() > max_serialized_size)
+                    max_serialized_size = converted_member_type->datatype->serialized_size();
             }
 
-            auto payload_type = world().definite_array_type(world().type_pu32(), max_serialized_size);
-            auto* converted_payload_type = convert(payload_type);
-
-            std::vector<SpvId> spv_pair = {converted_tag_type->type_id, converted_payload_type->type_id };
-            converted->type_id = builder_->declare_struct_type(spv_pair);
-
-            // auto oh_god_why = std::vector<ConvertedType*> ( &converted_tag_type, &converted_payload_type );
+            if (max_serialized_size > 0) {
+                auto payload_type = world().definite_array_type(world().type_pu32(), max_serialized_size);
+                auto* converted_payload_type = convert(payload_type);
 
-            converted->datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*> { converted_tag_type, converted_payload_type });
+                std::vector<SpvId> spv_pair = {converted_tag_type->type_id, converted_payload_type->type_id};
+                converted->type_id = builder_->declare_struct_type(spv_pair);
+                converted->datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*>{ converted_tag_type, converted_payload_type });
+            } else {
+                // We keep this useless level of struct so the rest of the code doesn't need a special path to extract the tag
+                std::vector<SpvId> spv_singleton = { converted_tag_type->type_id };
+                converted->type_id = builder_->declare_struct_type(spv_singleton);
+                converted->datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*>{ converted_tag_type });
+            }
             builder_->name(converted->type_id, type->to_string());
             break;
         }
 
+        case Node_MemType: {
+            assert(false && "TODO: get arround this");
+        }
+
         default:
             THORIN_UNREACHABLE;
     }

From 76652387aab8000e7c08a4cf22b105ebbdf76cdb Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 19 Apr 2021 09:14:36 +0200
Subject: [PATCH 054/342] map precise/quick types to the same thing

---
 src/thorin/be/spirv/spirv_datatypes.cpp | 32 ++++++++++++++++++-------
 1 file changed, 24 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 85763df97..35e1b8c54 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -79,6 +79,18 @@ void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, Sp
 }
 
 ConvertedType* CodeGen::convert(const Type* type) {
+    // Spir-V requires each primitive type to be "unique", it doesn't allow for example two 32-bit signed integer types.
+    // Therefore we must enforce that precise/quick types map to the same thing.
+    switch (type->tag()) {
+#define THORIN_Q_TYPE(T, M) \
+    case PrimType_##T: \
+        type = world().prim_type(PrimType_p##M, 1); \
+        break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_Q_TYPE
+        default: break;
+    }
+
     if (auto iter = types_.find(type); iter != types_.end()) return iter->second.get();
 
     assert(!type->isa<MemType>());
@@ -88,27 +100,31 @@ ConvertedType* CodeGen::convert(const Type* type) {
         // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
         // Using extensions, we could use 16 or 8-bit ints instead
         // We can also pack them inside structures using bit-twiddling tricks, if the need arises
+        // Note: this only affects storing booleans inside structures, for regular variables the actual spir-v bool type is used.
         case PrimType_bool:
             converted->type_id = builder_->declare_bool_type();
             converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
-        case PrimType_ps8:  case PrimType_qs8:  case PrimType_pu8:  case PrimType_qu8:  assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps16: case PrimType_qs16: case PrimType_pu16: case PrimType_qu16: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps32: case PrimType_qs32:
+        case PrimType_ps8:  assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pu8:  assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps16: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pu16: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps32:
             converted->type_id = builder_->declare_int_type(32, true );
             converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
-        case PrimType_pu32: case PrimType_qu32:
+        case PrimType_pu32:
             converted->type_id = builder_->declare_int_type(32, false);
             converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
-        case PrimType_ps64: case PrimType_qs64: case PrimType_pu64: case PrimType_qu64: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pf16: case PrimType_qf16:                                         assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pf32: case PrimType_qf32:
+        case PrimType_ps64: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pu64: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf16: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf32:
             converted->type_id = builder_->declare_float_type(32);
             converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
-        case PrimType_pf64: case PrimType_qf64:                                         assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_pf64: assert(false && "TODO: look into capabilities to enable this");
         case Node_PtrType: {
             auto ptr = type->as<PtrType>();
             spv::StorageClass storage_class;

From 4bc32dd4ba709009c780b9dbc6f42d22aa220a5d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 19 Apr 2021 09:19:00 +0200
Subject: [PATCH 055/342] more serdes fixes

---
 src/thorin/be/spirv/spirv.cpp           | 14 ++++++++------
 src/thorin/be/spirv/spirv_datatypes.cpp |  3 ++-
 2 files changed, 10 insertions(+), 7 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 257220349..a6ec26b99 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -461,12 +461,13 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
         if (variant_datatype->elements_types.size() > 1) {
-            auto ptr_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
-            auto payload_arr = bb->variable(ptr_type, spv::StorageClassFunction);
+            auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
+            auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+            auto payload_arr = bb->variable(alloc_type, spv::StorageClassFunction);
             auto converted_payload_type = convert(variant_type->op(variant->index()));
 
             auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
-            auto ptr_arr = bb->ptr_access_chain(convert(world().type_pu32())->type_id, payload_arr, zero, { zero });
+            auto ptr_arr = bb->ptr_access_chain(ptr_type, payload_arr, zero, { zero });
 
             converted_payload_type->datatype->emit_serialization(*bb, ptr_arr, emit(variant->value(), bb));
             auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
@@ -487,13 +488,14 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto target_type = convert(def->type());
 
         assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
-        auto ptr_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
-        auto payload_arr = bb->variable(ptr_type, spv::StorageClassFunction);
+        auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
+        auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+        auto payload_arr = bb->variable(alloc_type, spv::StorageClassFunction);
         auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
         auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
-        auto ptr_arr = bb->ptr_access_chain(convert(world().type_pu32())->type_id, payload_arr, zero, { zero });
+        auto ptr_arr = bb->ptr_access_chain(ptr_type, payload_arr, zero, { zero });
         return target_type->datatype->emit_deserialization(*bb, ptr_arr);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 35e1b8c54..2d2badffb 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -11,7 +11,8 @@ ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in
 }
 
 SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
-    auto loaded = bb.load(type->type_id, input);
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    auto loaded = bb.load(u32_tid, input);
     return bb.bitcast(type->type_id, loaded);
 }
 

From 69e5e89f23249459b4b71ac92326abbd587b4e40 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 19 Apr 2021 09:33:31 +0200
Subject: [PATCH 056/342] moves variables decls to function start

---
 src/thorin/be/spirv/spirv.cpp         |  6 ++--
 src/thorin/be/spirv/spirv.h           |  2 ++
 src/thorin/be/spirv/spirv_builder.hpp | 41 ++++++++++++++++++++-------
 3 files changed, 36 insertions(+), 13 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index a6ec26b99..9c16c9e3f 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -46,7 +46,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
-    FnBuilder fn;
+    FnBuilder fn(*builder_);
     fn.scope = &scope;
     fn.file_builder = builder_;
     fn.fn_type = convert(entry_->type())->type_id;
@@ -463,7 +463,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         if (variant_datatype->elements_types.size() > 1) {
             auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
             auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
-            auto payload_arr = bb->variable(alloc_type, spv::StorageClassFunction);
+            auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
             auto converted_payload_type = convert(variant_type->op(variant->index()));
 
             auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
@@ -490,7 +490,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
         auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
         auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
-        auto payload_arr = bb->variable(alloc_type, spv::StorageClassFunction);
+        auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
         auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index c45232939..500c41499 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -37,6 +37,8 @@ struct FnBuilder : public builder::SpvFnBuilder {
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
+
+    explicit FnBuilder(builder::SpvFileBuilder& file_builder) : builder::SpvFnBuilder(file_builder) {}
 };
 
 class CodeGen : public thorin::CodeGen {
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 9454dc618..953ea0cfb 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -106,15 +106,6 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        op(spv::Op::OpVariable, 4);
-        ref_id(type);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        literal_int(storage_class);
-        return id;
-    }
-
     SpvId bitcast(SpvId target_type, SpvId value) {
         op(spv::Op::OpBitcast, 4);
         auto id = generate_fresh_id();
@@ -207,13 +198,32 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 };
 
 struct SpvFnBuilder {
-public:
+    explicit SpvFnBuilder(SpvFileBuilder& file_builder)
+    : file_builder(file_builder)
+    {}
+
+    SpvFileBuilder& file_builder;
+
     SpvId fn_type;
     SpvId fn_ret_type;
     std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
 
     // Contains OpFunctionParams
     SpvSectionBuilder header;
+
+    SpvSectionBuilder variables;
+
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        variables.op(spv::Op::OpVariable, 4);
+        variables.ref_id(type);
+        auto id = generate_fresh_id();
+        variables.ref_id(id);
+        variables.literal_int(storage_class);
+        return id;
+    }
+
+private:
+    SpvId generate_fresh_id();
 };
 
 struct SpvFileBuilder {
@@ -322,10 +332,17 @@ struct SpvFileBuilder {
         for (auto w : fn_builder.header.data_)
             fn_defs.data_.push_back(w);
 
+        bool first = true;
         for (auto& bb : fn_builder.bbs_to_emit) {
             fn_defs.op(spv::Op::OpLabel, 2);
             fn_defs.ref_id(bb->label);
 
+            if (first) {
+                for (auto w : fn_builder.variables.data_)
+                    fn_defs.data_.push_back(w);
+                first = false;
+            }
+
             for (auto& phi : bb->phis) {
                 fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
                 fn_defs.ref_id(phi->type);
@@ -426,4 +443,8 @@ inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
     return file_builder.generate_fresh_id();
 }
 
+inline SpvId SpvFnBuilder::generate_fresh_id() {
+    return file_builder.generate_fresh_id();
+}
+
 }
\ No newline at end of file

From 939c1ba631916d44fd3c86590424958441dff71b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 22 Apr 2021 13:48:17 +0200
Subject: [PATCH 057/342] push-constant reading entry points

---
 src/thorin/be/spirv/spirv.cpp         | 44 +++++++++++++++++++++++++--
 src/thorin/be/spirv/spirv.h           |  5 ++-
 src/thorin/be/spirv/spirv_builder.hpp | 37 +++++++++++++++++-----
 3 files changed, 73 insertions(+), 13 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 9c16c9e3f..09f09cac3 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -32,9 +32,47 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
+    auto push_constant_arr_type = world().definite_array_type(world().type_pu32(), 128);
+    auto push_constant_ptr_type = builder.declare_ptr_type(spv::StorageClassPushConstant, convert(push_constant_arr_type)->type_id);
+    auto push_constant_ptr = builder_->variable(push_constant_ptr_type, spv::StorageClassPushConstant);
+
+    auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
     for (auto& cont : world().continuations()) {
         if (cont->is_exported()) {
-            // TODO create entry point
+            assert(defs_.contains(cont));
+            SpvId callee = defs_[cont];
+
+            // TODO name entry points
+            FnBuilder fn_builder(builder_);
+            fn_builder.fn_type = entry_pt_signature;
+            fn_builder.fn_ret_type = builder_->void_type;
+
+            BasicBlockBuilder* bb = fn_builder.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn_builder)).get();
+            fn_builder.bbs_to_emit.push_back(bb);
+
+            // iterate on cont type and extract the
+            auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
+            auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
+            auto ptr_arr = bb->ptr_access_chain(ptr_type, push_constant_ptr, zero, { zero });
+            size_t offset = 0;
+            std::vector<SpvId> args;
+            for (size_t i = 0; i < cont->num_ops(); i++) {
+                auto op = cont->op(i);
+                auto op_type = op->type();
+                if (op_type == world().unit() || op_type == world().mem_type() || op_type->isa<FnType>()) continue;
+                assert(op_type->order() == 0);
+                auto converted = convert(op_type);
+                assert(converted->datatype != nullptr);
+                SpvId arg = converted->datatype->emit_deserialization(*bb, ptr_arr);
+                args.push_back(arg);
+                bb->ptr_access_chain(ptr_type, push_constant_ptr, bb->file_builder.constant(convert(world().type_ps32())->type_id, { (uint32_t) offset }), { });
+                offset += converted->datatype->serialized_size();
+            }
+
+            bb->call(builder_->void_type, callee, args);
+            bb->return_void();
+
+            builder_->define_function(fn_builder);
         }
     }
 
@@ -46,11 +84,11 @@ void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
-    FnBuilder fn(*builder_);
+    FnBuilder fn(builder_);
     fn.scope = &scope;
-    fn.file_builder = builder_;
     fn.fn_type = convert(entry_->type())->type_id;
     fn.fn_ret_type = get_codom_type(entry_);
+    defs_.emplace(scope.entry(), fn.function_id);
 
     current_fn_ = &fn;
 
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 500c41499..3d97a6475 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -31,14 +31,13 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {
-    const Scope* scope;
-    builder::SpvFileBuilder* file_builder;
+    const Scope* scope = nullptr;
     std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
 
-    explicit FnBuilder(builder::SpvFileBuilder& file_builder) : builder::SpvFnBuilder(file_builder) {}
+    explicit FnBuilder(builder::SpvFileBuilder* file_builder) : builder::SpvFnBuilder(file_builder) {}
 };
 
 class CodeGen : public thorin::CodeGen {
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 953ea0cfb..5b23e1d91 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -180,6 +180,18 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
             literal_int(e);
     }
 
+    SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
+        op(spv::Op::OpFunctionCall, 4 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(callee);
+
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
+    }
+
     void return_void() {
         op(spv::Op::OpReturn, 1);
     }
@@ -198,11 +210,14 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 };
 
 struct SpvFnBuilder {
-    explicit SpvFnBuilder(SpvFileBuilder& file_builder)
+    explicit SpvFnBuilder(SpvFileBuilder* file_builder)
     : file_builder(file_builder)
-    {}
+    {
+        function_id = generate_fresh_id();
+    }
 
-    SpvFileBuilder& file_builder;
+    SpvFileBuilder* file_builder;
+    SpvId function_id;
 
     SpvId fn_type;
     SpvId fn_ret_type;
@@ -283,7 +298,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_fn_type(std::vector<SpvId>& dom, SpvId codom) {
+    SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
         types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -323,8 +338,7 @@ struct SpvFileBuilder {
     SpvId define_function(SpvFnBuilder& fn_builder) {
         fn_defs.op(spv::Op::OpFunction, 5);
         fn_defs.ref_id(fn_builder.fn_ret_type);
-        auto id = generate_fresh_id();
-        fn_defs.ref_id(id);
+        fn_defs.ref_id(fn_builder.function_id);
         fn_defs.data_.push_back(spv::FunctionControlMaskNone);
         fn_defs.ref_id(fn_builder.fn_type);
 
@@ -359,6 +373,15 @@ struct SpvFileBuilder {
         }
 
         fn_defs.op(spv::Op::OpFunctionEnd, 1);
+        return fn_builder.function_id;
+    }
+
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        types_constants.op(spv::Op::OpVariable, 4);
+        types_constants.ref_id(type);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
         return id;
     }
 
@@ -444,7 +467,7 @@ inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
 }
 
 inline SpvId SpvFnBuilder::generate_fresh_id() {
-    return file_builder.generate_fresh_id();
+    return file_builder->generate_fresh_id();
 }
 
 }
\ No newline at end of file

From 93cae89069d9103f677a7af4b1f19050de102fd4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 22 Apr 2021 14:25:12 +0200
Subject: [PATCH 058/342] avoid duplicate definitions

---
 src/thorin/be/spirv/spirv_builder.hpp | 33 +++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 5b23e1d91..921ac936c 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -241,7 +241,32 @@ struct SpvFnBuilder {
     SpvId generate_fresh_id();
 };
 
+inline bool operator==(const SpvId &a, const SpvId &b) { return a.id == b.id; }
+
 struct SpvFileBuilder {
+    enum UniqueTypeTag {
+        NONE,
+        FN_TYPE
+    };
+
+    struct UniqueTypeKey {
+        UniqueTypeTag tag;
+        std::vector<SpvId> members;
+
+        bool operator==(const UniqueTypeKey &b) const {
+            return tag == b.tag && members == b.members;
+        }
+    };
+
+    struct UniqueTypeKeyHasher {
+        size_t operator() (const UniqueTypeKey& key) const {
+            size_t acc = 0;
+            for (auto id : key.members)
+                acc ^= std::hash<uint32_t>{}(id.id);
+            return std::hash<size_t>{}(key.tag) ^ acc;
+        }
+    };
+
     SpvFileBuilder()
     : void_type(declare_void_type())
     {}
@@ -299,12 +324,17 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
+        auto key = UniqueTypeKey { FN_TYPE, dom };
+        key.members.push_back(codom);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+
         types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.ref_id(codom);
         for (auto arg : dom)
             types_constants.ref_id(arg);
+        unique_decls[key] = id;
         return id;
     }
 
@@ -411,6 +441,9 @@ struct SpvFileBuilder {
     SpvSectionBuilder fn_decls;
     SpvSectionBuilder fn_defs;
 
+    // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
+    std::unordered_map<UniqueTypeKey, SpvId, UniqueTypeKeyHasher> unique_decls;
+
     SpvId declare_void_type() {
         types_constants.op(spv::Op::OpTypeVoid, 2);
         auto id = generate_fresh_id();

From 896277fe8d87d069e39a62b3358a8094b5e01f3b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 22 Apr 2021 15:54:27 +0200
Subject: [PATCH 059/342] valid entry points

---
 src/thorin/be/spirv/spirv.cpp           | 28 +++++++++++---------
 src/thorin/be/spirv/spirv.h             | 16 +++++------
 src/thorin/be/spirv/spirv_builder.hpp   | 23 +++++++++++-----
 src/thorin/be/spirv/spirv_datatypes.cpp | 35 ++++++++++++++-----------
 src/thorin/type.h                       |  1 +
 5 files changed, 60 insertions(+), 43 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 09f09cac3..f44b0429a 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -21,7 +21,7 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder::SpvFileBuilder builder;
     builder_ = &builder;
     builder_->capability(spv::Capability::CapabilityShader);
-    builder_->capability(spv::Capability::CapabilityLinkage);
+    // builder_->capability(spv::Capability::CapabilityLinkage);
     builder_->capability(spv::Capability::CapabilityVariablePointers);
     builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
 
@@ -51,28 +51,31 @@ void CodeGen::emit_stream(std::ostream& out) {
             fn_builder.bbs_to_emit.push_back(bb);
 
             // iterate on cont type and extract the
-            auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
+            auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Push))->type_id;
             auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
             auto ptr_arr = bb->ptr_access_chain(ptr_type, push_constant_ptr, zero, { zero });
             size_t offset = 0;
             std::vector<SpvId> args;
-            for (size_t i = 0; i < cont->num_ops(); i++) {
-                auto op = cont->op(i);
-                auto op_type = op->type();
-                if (op_type == world().unit() || op_type == world().mem_type() || op_type->isa<FnType>()) continue;
-                assert(op_type->order() == 0);
-                auto converted = convert(op_type);
+            for (size_t i = 0; i < cont->num_params(); i++) {
+                auto param = cont->param(i);
+                auto param_type = param->type();
+                if (param_type == world().unit() || param_type == world().mem_type() || param_type->isa<FnType>()) continue;
+                assert(param_type->order() == 0);
+                auto converted = convert(param_type);
                 assert(converted->datatype != nullptr);
-                SpvId arg = converted->datatype->emit_deserialization(*bb, ptr_arr);
+                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, ptr_arr);
                 args.push_back(arg);
-                bb->ptr_access_chain(ptr_type, push_constant_ptr, bb->file_builder.constant(convert(world().type_ps32())->type_id, { (uint32_t) offset }), { });
                 offset += converted->datatype->serialized_size();
+                ptr_arr = bb->ptr_access_chain(ptr_type, ptr_arr, bb->file_builder.constant(convert(world().type_ps32())->type_id, { (uint32_t) offset }), { });
             }
 
             bb->call(builder_->void_type, callee, args);
             bb->return_void();
 
             builder_->define_function(fn_builder);
+            builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
+
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "main", { push_constant_ptr });
         }
     }
 
@@ -150,6 +153,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     }
     
     builder_->define_function(fn);
+    builder_->name(fn.function_id, scope.entry()->name());
 }
 
 SpvId CodeGen::get_codom_type(const Continuation* fn) {
@@ -507,7 +511,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
             auto ptr_arr = bb->ptr_access_chain(ptr_type, payload_arr, zero, { zero });
 
-            converted_payload_type->datatype->emit_serialization(*bb, ptr_arr, emit(variant->value(), bb));
+            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, ptr_arr, emit(variant->value(), bb));
             auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
 
             auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index())});
@@ -534,7 +538,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
 
         auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
         auto ptr_arr = bb->ptr_access_chain(ptr_type, payload_arr, zero, { zero });
-        return target_type->datatype->emit_deserialization(*bb, ptr_arr);
+        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, ptr_arr);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
         return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 3d97a6475..1c3718070 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -76,8 +76,8 @@ struct Datatype {
     Datatype(ConvertedType* type) : type(type) {}
 
     virtual size_t serialized_size() = 0;
-    virtual void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) = 0;
-    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) = 0;
+    virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) = 0;
+    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) = 0;
 };
 
 /// For scalar datatypes
@@ -88,8 +88,8 @@ struct ScalarDatatype : public Datatype {
     ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
 
     size_t serialized_size() override { return size_in_bytes / 4; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
 };
 
 struct DefiniteArrayDatatype : public Datatype {
@@ -99,8 +99,8 @@ struct DefiniteArrayDatatype : public Datatype {
     DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length);
 
     size_t serialized_size() override { return element_type->datatype->serialized_size(); };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
 };
 
 struct ProductDatatype : public Datatype {
@@ -110,8 +110,8 @@ struct ProductDatatype : public Datatype {
     ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types);
 
     size_t serialized_size() override { return total_size; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
 };
 
 }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 921ac936c..069df70d2 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -365,6 +365,15 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        types_constants.op(spv::Op::OpVariable, 4);
+        types_constants.ref_id(type);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
+        return id;
+    }
+
     SpvId define_function(SpvFnBuilder& fn_builder) {
         fn_defs.op(spv::Op::OpFunction, 5);
         fn_defs.ref_id(fn_builder.fn_ret_type);
@@ -406,13 +415,13 @@ struct SpvFileBuilder {
         return fn_builder.function_id;
     }
 
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        types_constants.op(spv::Op::OpVariable, 4);
-        types_constants.ref_id(type);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(storage_class);
-        return id;
+    void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
+        entry_points.op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
+        entry_points.literal_int(execution_model);
+        entry_points.ref_id(entry_point);
+        entry_points.literal_name(name);
+        for (auto i : interface)
+            entry_points.ref_id(i);
     }
 
     void capability(spv::Capability cap) {
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 2d2badffb..2ba0c6974 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -10,13 +10,13 @@ ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in
     assert(size_in_bytes == 4);
 }
 
-SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
     SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     auto loaded = bb.load(u32_tid, input);
     return bb.bitcast(type->type_id, loaded);
 }
 
-void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
     SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     auto casted = bb.bitcast(u32_tid, data);
     bb.store(casted, output);
@@ -27,23 +27,23 @@ DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType*
     assert(length > 0 && "Array lengths of zero are not supported");
 }
 
-SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
     SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_ps32())->type_id;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     for (size_t i = 0; i < length; i++) {
         SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type->datatype->serialized_size()) }), indices);
-        SpvId element = element_type->datatype->emit_deserialization(bb, element_ptr);
+        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, element_ptr);
         elements.push_back(element);
     }
     return bb.composite(type->type_id, elements);
 }
-void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
     std::vector<SpvId> indices;
     SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_ps32())->type_id;
     for (size_t i = 0; i < length; i++) {
         SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, output, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type->datatype->serialized_size()) }), indices);
-        element_type->datatype->emit_serialization(bb, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i }));
+        element_type->datatype->emit_serialization(bb, storage_class, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i }));
     }
 }
 
@@ -53,29 +53,31 @@ ProductDatatype::ProductDatatype(ConvertedType* type, const std::vector<Converte
     }
 }
 
-SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, SpvId input) {
+SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
     assert(total_size > 0 && "It doesn't make sense to de-serialize Unit!");
-    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     size_t offset = 0;
     for (auto& element_type : elements_types) {
-        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) offset }), indices);
-        SpvId element = element_type->datatype->emit_deserialization(bb, element_ptr);
+        SpvId element_ptr = bb.ptr_access_chain(arr_cell_tid, input, bb.file_builder.constant(u32_tid, { (uint32_t) offset }), indices);
+        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, element_ptr);
         offset += element_type->datatype->serialized_size();
         elements.push_back(element);
     }
     return bb.composite(type->type_id, elements);
 }
-void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, SpvId output, SpvId data) {
+void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
     assert(total_size > 0 && "It doesn't make sense to serialize Unit!");
-    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
     std::vector<SpvId> indices;
     size_t offset = 0;
     int i = 0;
     for (auto& element_type : elements_types) {
-        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, output, bb.file_builder.constant(i32_tid, { (uint32_t) offset }), indices);
-        element_type->datatype->emit_serialization(bb, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
+        SpvId element_ptr = bb.ptr_access_chain(arr_cell_tid, output, bb.file_builder.constant(u32_tid, { (uint32_t) offset }), indices);
+        element_type->datatype->emit_serialization(bb, storage_class, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
     }
 }
 
@@ -130,8 +132,9 @@ ConvertedType* CodeGen::convert(const Type* type) {
             auto ptr = type->as<PtrType>();
             spv::StorageClass storage_class;
             switch (ptr->addr_space()) {
-                case AddrSpace::Function: storage_class = spv::StorageClassFunction; break;
-                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;  break;
+                case AddrSpace::Function: storage_class = spv::StorageClassFunction;     break;
+                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;      break;
+                case AddrSpace::Push   :  storage_class = spv::StorageClassPushConstant; break;
                 default:
                     assert(false && "This address space is not supported");
                     break;
diff --git a/src/thorin/type.h b/src/thorin/type.h
index b5916c7f5..97baf4d39 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -235,6 +235,7 @@ enum class AddrSpace : uint32_t {
     Constant = 4,
     Private =  5, // Corresponds to the 'private' storage class in SPIR-V
     Function = 6, // Corresponds to the 'function' storage class in SPIR-V
+    Push     = 7, // Corresponds to the 'push constant' storage class in SPIR-V
 };
 
 /// Pointer type.

From 135cc8939621e9ea429f1270221cf1a65b6d6a83 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 22 Apr 2021 16:55:05 +0200
Subject: [PATCH 060/342] "support" generic pointers in spir-v

---
 src/thorin/be/spirv/spirv_datatypes.cpp | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 2ba0c6974..2a8b9ace5 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -134,13 +134,27 @@ ConvertedType* CodeGen::convert(const Type* type) {
             switch (ptr->addr_space()) {
                 case AddrSpace::Function: storage_class = spv::StorageClassFunction;     break;
                 case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;      break;
-                case AddrSpace::Push   :  storage_class = spv::StorageClassPushConstant; break;
+                case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant; break;
+                case AddrSpace::Global: {
+                    storage_class = spv::StorageClassPhysicalStorageBuffer;
+                    // TODO datatype code for stuffing into push constants
+                    break;
+                }
+                case AddrSpace::Generic: {
+                    world().WLOG("Passing a generic pointer to a SPIR-V module. SpirV doesn't know about these, and so this will be passed as a 64 bit integer. Tread carefully !");
+                    ConvertedType* conv_u64 = convert(world().type_pu64());
+                    converted->type_id = conv_u64->type_id;
+                    goto ptr_done;
+                }
                 default:
                     assert(false && "This address space is not supported");
                     break;
             }
-            ConvertedType* element = convert(ptr->pointee());
-            converted->type_id = builder_->declare_ptr_type(storage_class, element->type_id);
+            {
+                ConvertedType* element = convert(ptr->pointee());
+                converted->type_id = builder_->declare_ptr_type(storage_class, element->type_id);
+            }
+            ptr_done:
             break;
         }
         case Node_IndefiniteArrayType: {

From 368815158bfe79f96a5132c6d2c0cb890e6c47be Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 23 Apr 2021 13:34:56 +0200
Subject: [PATCH 061/342] reading pointers

---
 src/thorin/be/spirv/spirv.h             |  9 +++++++
 src/thorin/be/spirv/spirv_builder.hpp   | 29 ++++++++++++++++++++++
 src/thorin/be/spirv/spirv_datatypes.cpp | 33 ++++++++++++++++++++-----
 3 files changed, 65 insertions(+), 6 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 1c3718070..78a22536e 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -92,6 +92,15 @@ struct ScalarDatatype : public Datatype {
     void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
 };
 
+struct PtrDatatype : public Datatype {
+    static constexpr size_t bitwidth = 64;
+    PtrDatatype(ConvertedType* type) : Datatype(type) {}
+
+    size_t serialized_size() override { return bitwidth / 32; };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
+};
+
 struct DefiniteArrayDatatype : public Datatype {
     ConvertedType* element_type;
     size_t length;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 069df70d2..893075916 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -115,6 +115,35 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    /// Change bit-width
+    SpvId u_convert(SpvId target_type, SpvId value) {
+        op(spv::Op::OpUConvert, 4);
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(value);
+        return id;
+    }
+
+    /// Change bit-width
+    SpvId s_convert(SpvId target_type, SpvId value) {
+        op(spv::Op::OpSConvert, 4);
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(value);
+        return id;
+    }
+
+    SpvId convert_u_ptr(SpvId target_type, SpvId value) {
+        op(spv::Op::OpConvertUToPtr, 4);
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(value);
+        return id;
+    }
+
     SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
         op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
         auto id = generate_fresh_id();
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 2a8b9ace5..7e8443b62 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -22,6 +22,26 @@ void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass
     bb.store(casted, output);
 }
 
+SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
+    assert(type->src_type->as<PtrType>()->addr_space() == AddrSpace::Global && "Only buffer device address (global memory) pointers supported");
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    SpvId u64_tid = type->code_gen->convert(type->code_gen->world().type_pu64())->type_id;
+    SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
+
+    auto input2 = bb.ptr_access_chain(arr_cell_tid, input, bb.file_builder.constant(u32_tid, { (uint32_t) 1 }), { });
+    auto upper = bb.u_convert(u64_tid, bb.load(u32_tid, input));
+    auto lower = bb.u_convert(u64_tid, bb.load(u32_tid, input2));
+
+    SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
+    auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32), lower);
+
+    return bb.convert_u_ptr(type->type_id, merged);
+}
+
+void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
+    assert(false && "TODO");
+}
+
 DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
     assert(element_type->datatype.get() != nullptr);
     assert(length > 0 && "Array lengths of zero are not supported");
@@ -137,7 +157,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
                 case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant; break;
                 case AddrSpace::Global: {
                     storage_class = spv::StorageClassPhysicalStorageBuffer;
-                    // TODO datatype code for stuffing into push constants
+                    converted->datatype = std::make_unique<PtrDatatype>(converted);
                     break;
                 }
                 case AddrSpace::Generic: {
@@ -151,17 +171,18 @@ ConvertedType* CodeGen::convert(const Type* type) {
                     break;
             }
             {
-                ConvertedType* element = convert(ptr->pointee());
+                const Type* pointee = ptr->pointee();
+                while (auto arr = pointee->isa<IndefiniteArrayType>())
+                    pointee = arr->elem_type();
+                ConvertedType* element = convert(pointee);
                 converted->type_id = builder_->declare_ptr_type(storage_class, element->type_id);
             }
             ptr_done:
             break;
         }
         case Node_IndefiniteArrayType: {
-            assert(false && "TODO");
-            // auto array = type->as<IndefiniteArrayType>();
-            // return types_[type] = spv_type;
-            THORIN_UNREACHABLE;
+            world().ELOG("Using indefinite types directly is not permitted - they may only be pointed to");
+            std::abort();
         }
         case Node_DefiniteArrayType: {
             auto array = type->as<DefiniteArrayType>();

From 9368bbb82819e533361b6c9ab71bd92cd3631761 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 23 Apr 2021 13:56:36 +0200
Subject: [PATCH 062/342] loads 64b bda pointers

---
 src/thorin/be/spirv/spirv.cpp           | 13 ++++++-----
 src/thorin/be/spirv/spirv.h             |  2 +-
 src/thorin/be/spirv/spirv_builder.hpp   | 14 ++++++++++++
 src/thorin/be/spirv/spirv_datatypes.cpp | 29 +++++++++++++++++++------
 4 files changed, 45 insertions(+), 13 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index f44b0429a..152ff40fe 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -21,9 +21,12 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder::SpvFileBuilder builder;
     builder_ = &builder;
     builder_->capability(spv::Capability::CapabilityShader);
-    // builder_->capability(spv::Capability::CapabilityLinkage);
     builder_->capability(spv::Capability::CapabilityVariablePointers);
     builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
+    builder_->capability(spv::Capability::CapabilityInt16);
+    builder_->capability(spv::Capability::CapabilityInt64);
+
+    builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
 
     structure_loops();
     structure_flow();
@@ -119,11 +122,11 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // Nothing
                 } else if (param->order() == 0) {
                     auto param_t = convert(param->type());
-                    fn.header.op(spv::Op::OpFunctionParameter, 3);
-                    auto id = builder_->generate_fresh_id();
-                    fn.header.ref_id(param_t->type_id);
-                    fn.header.ref_id(id);
+                    auto id = fn.parameter(param_t->type_id);
                     fn.params[param] = id;
+                    if (param->type()->isa<PtrType>()) {
+                        builder_->decorate(id, spv::DecorationAliased);
+                    }
                 }
             }
         } else {
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 78a22536e..56568a9bc 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -87,7 +87,7 @@ struct ScalarDatatype : public Datatype {
     size_t alignment;
     ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
 
-    size_t serialized_size() override { return size_in_bytes / 4; };
+    size_t serialized_size() override { return (size_in_bytes + 3) / 4; };
     SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
     void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
 };
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 893075916..8a5f38ae0 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -257,6 +257,14 @@ struct SpvFnBuilder {
 
     SpvSectionBuilder variables;
 
+    SpvId parameter(SpvId param_type) {
+        header.op(spv::Op::OpFunctionParameter, 3);
+        auto id = generate_fresh_id();
+        header.ref_id(param_type);
+        header.ref_id(id);
+        return id;
+    }
+
     SpvId variable(SpvId type, spv::StorageClass storage_class) {
         variables.op(spv::Op::OpVariable, 4);
         variables.ref_id(type);
@@ -376,6 +384,12 @@ struct SpvFileBuilder {
         return id;
     }
 
+    void decorate(SpvId target, spv::Decoration decoration) {
+        annotations.op(spv::Op::OpDecorate, 3);
+        annotations.ref_id(target);
+        annotations.literal_int(decoration);
+    }
+
     SpvId bool_constant(SpvId type, bool value) {
         types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
         auto id = generate_fresh_id();
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 7e8443b62..ec9247fb7 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -6,17 +6,20 @@ namespace thorin::spirv {
 ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes)
 : Datatype(type), type_tag(type_tag), size_in_bytes(size_in_bytes), alignment(alignment_in_bytes)
 {
-    /// currently limited to 32-bit
-    assert(size_in_bytes == 4);
+
 }
 
 SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
+    /// currently limited to 32-bit
+    assert(size_in_bytes == 4);
     SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     auto loaded = bb.load(u32_tid, input);
     return bb.bitcast(type->type_id, loaded);
 }
 
 void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
+    /// currently limited to 32-bit
+    assert(size_in_bytes == 4);
     SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     auto casted = bb.bitcast(u32_tid, data);
     bb.store(casted, output);
@@ -126,12 +129,18 @@ ConvertedType* CodeGen::convert(const Type* type) {
         // Note: this only affects storing booleans inside structures, for regular variables the actual spir-v bool type is used.
         case PrimType_bool:
             converted->type_id = builder_->declare_bool_type();
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 1, 1);
             break;
         case PrimType_ps8:  assert(false && "TODO: look into capabilities to enable this");
         case PrimType_pu8:  assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps16: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pu16: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps16:
+            converted->type_id = builder_->declare_int_type(16, true);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 2, 2);
+            break;
+        case PrimType_pu16:
+            converted->type_id = builder_->declare_int_type(16, false);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 2, 2);
+            break;
         case PrimType_ps32:
             converted->type_id = builder_->declare_int_type(32, true );
             converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
@@ -140,8 +149,14 @@ ConvertedType* CodeGen::convert(const Type* type) {
             converted->type_id = builder_->declare_int_type(32, false);
             converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
             break;
-        case PrimType_ps64: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pu64: assert(false && "TODO: look into capabilities to enable this");
+        case PrimType_ps64:
+            converted->type_id = builder_->declare_int_type(64, true);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 8, 8);
+            break;
+        case PrimType_pu64:
+            converted->type_id = builder_->declare_int_type(64, false);
+            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 8, 8);
+            break;
         case PrimType_pf16: assert(false && "TODO: look into capabilities to enable this");
         case PrimType_pf32:
             converted->type_id = builder_->declare_float_type(32);

From 0df66859247b3fdb0ec277bb87230568ff6435b0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 23 Apr 2021 14:10:38 +0200
Subject: [PATCH 063/342] do emit mem/unit arguments!

---
 src/thorin/be/spirv/spirv.cpp | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 152ff40fe..f6a239678 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -181,9 +181,9 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         for (auto arg : continuation->args()) {
             assert(arg->order() == 0);
+            auto val = emit(arg, bb);
             if (is_mem(arg) || is_unit(arg))
                 continue;
-            auto val = emit(arg, bb);
             values.emplace_back(val);
         }
 
@@ -284,8 +284,9 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         int index = -1;
         for (auto& arg : continuation->args()) {
             index++;
+            auto val = emit(arg, bb);
             if (is_mem(arg) || is_unit(arg)) continue;
-            bb->args[arg] = emit(arg, bb);
+            bb->args[arg] = val;
             auto* param = callee->param(index);
             auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
             phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);

From d7ee73e218aa0a658f1da5a675dfa2aec895876e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 23 Apr 2021 15:19:01 +0200
Subject: [PATCH 064/342] handle lea/store/load

---
 src/thorin/be/spirv/spirv.cpp         | 19 +++++++++++++++++++
 src/thorin/be/spirv/spirv_builder.hpp | 12 ++++++++----
 2 files changed, 27 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index f6a239678..3953e62c7 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -369,6 +369,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     }
 }
 
+constexpr SpvId spv_none { 0 };
+
 SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
@@ -562,6 +564,23 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             elements[x++] = emit(e, bb);
         }
         return bb->composite(convert(structagg->type())->type_id, elements);
+    } else if (auto access = def->isa<Access>()) {
+        std::vector<uint32_t> operands;
+        auto ptr_type = access->ptr()->type()->as<PtrType>();
+        if (ptr_type->addr_space() == AddrSpace::Global) {
+            operands.push_back(spv::MemoryAccessAlignedMask);
+            operands.push_back( 4 ); // TODO: SPIR-V docs say to consult client API for valid values.
+        }
+        if (auto load = def->isa<Load>()) {
+            return bb->load(convert(load->out_val_type())->type_id, emit(load->ptr(), bb), operands);
+        } else if (auto store = def->isa<Store>()) {
+            bb->store(emit(store->val(), bb), emit(store->ptr(), bb), operands);
+            return spv_none;
+        } else THORIN_UNREACHABLE;
+    } else if (auto lea = def->isa<LEA>()) {
+        auto type = convert(lea->ptr_type());
+        auto offset = emit(lea->index(), bb);
+        return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
     }
     assertf(false, "Incomplete emit(def) definition");
 }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 8a5f38ae0..93aa72ba2 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -156,19 +156,23 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId load(SpvId target_type, SpvId pointer) {
-        op(spv::Op::OpLoad, 4);
+    SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
+        op(spv::Op::OpLoad, 4 + operands.size());
         auto id = generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(pointer);
+        for (auto op : operands)
+            literal_int(op);
         return id;
     }
 
-    void store(SpvId value, SpvId pointer) {
-        op(spv::Op::OpStore, 3);
+    void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
+        op(spv::Op::OpStore, 3 + operands.size());
         ref_id(pointer);
         ref_id(value);
+        for (auto op : operands)
+            literal_int(op);
     }
 
     SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {

From 81d27bcf4235614bbbd6cecff03f7f5a20a3b444 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 23 Apr 2021 19:07:35 +0200
Subject: [PATCH 065/342] expose workgroup size in entry point

---
 src/thorin/be/llvm/llvm.cpp           |  2 +-
 src/thorin/be/llvm/runtime.h          |  3 ++-
 src/thorin/be/spirv/spirv.cpp         | 19 ++++++++++++++-----
 src/thorin/be/spirv/spirv.h           |  2 +-
 src/thorin/be/spirv/spirv_builder.hpp |  8 ++++++++
 5 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index e36a33267..028d9edf0 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1082,7 +1082,7 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
         case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
         case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
         case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
-        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv",    continuation); // TODO have a real runtime component
+        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::VULKAN_PLATFORM, ".spv",    continuation);
         case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
         case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
         case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index 616595415..9ecf7746b 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -22,7 +22,8 @@ class Runtime {
         CPU_PLATFORM,
         CUDA_PLATFORM,
         OPENCL_PLATFORM,
-        HSA_PLATFORM
+        HSA_PLATFORM,
+        VULKAN_PLATFORM,
     };
 
     /// Emits a call to anydsl_launch_kernel.
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3953e62c7..751415c84 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -13,8 +13,8 @@ namespace thorin {
 
 namespace thorin::spirv {
 
-CodeGen::CodeGen(thorin::World& world, Cont2Config&, bool debug)
-    : thorin::CodeGen(world, debug)
+CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
+    : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
 {}
 
 void CodeGen::emit_stream(std::ostream& out) {
@@ -42,10 +42,11 @@ void CodeGen::emit_stream(std::ostream& out) {
     auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
     for (auto& cont : world().continuations()) {
         if (cont->is_exported()) {
-            assert(defs_.contains(cont));
+            assert(defs_.contains(cont) && kernel_config_.contains(cont));
+            auto config = kernel_config_.find(cont);
+
             SpvId callee = defs_[cont];
 
-            // TODO name entry points
             FnBuilder fn_builder(builder_);
             fn_builder.fn_type = entry_pt_signature;
             fn_builder.fn_ret_type = builder_->void_type;
@@ -78,7 +79,15 @@ void CodeGen::emit_stream(std::ostream& out) {
             builder_->define_function(fn_builder);
             builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
 
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "main", { push_constant_ptr });
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", { push_constant_ptr });
+
+            auto block = config->second->as<GPUKernelConfig>()->block_size();
+            std::vector<uint32_t> local_size = {
+                (uint32_t) std::get<0>(block),
+                (uint32_t) std::get<1>(block),
+                (uint32_t) std::get<2>(block),
+            };
+            builder_->execution_mode(fn_builder.function_id, spv::ExecutionModeLocalSize, local_size);
         }
     }
 
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 56568a9bc..d5a0d5649 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -63,7 +63,7 @@ class CodeGen : public thorin::CodeGen {
     FnBuilder* current_fn_ = nullptr;
     TypeMap<std::unique_ptr<ConvertedType>> types_;
     DefMap<SpvId> defs_;
-
+    const Cont2Config& kernel_config_;
 };
 
 /// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 93aa72ba2..7bfbd4451 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -471,6 +471,14 @@ struct SpvFileBuilder {
             entry_points.ref_id(i);
     }
 
+    void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
+        entry_points.op(spv::Op::OpExecutionMode, 3 + payloads.size());
+        entry_points.ref_id(entry_point);
+        entry_points.literal_int(execution_mode);
+        for (auto d : payloads)
+            entry_points.literal_int(d);
+    }
+
     void capability(spv::Capability cap) {
         capabilities.op(spv::Op::OpCapability, 2);
         capabilities.data_.push_back(cap);

From 743961c1040e2baafb0edc1a112ac5b160099438 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 23 Apr 2021 19:40:18 +0200
Subject: [PATCH 066/342] success: validation passes, shader kills NIR

---
 src/thorin/be/spirv/spirv.cpp         | 18 ++++++++++++------
 src/thorin/be/spirv/spirv_builder.hpp | 17 ++++++++++++++---
 2 files changed, 26 insertions(+), 9 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 751415c84..d7f6c765d 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -35,9 +35,15 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
-    auto push_constant_arr_type = world().definite_array_type(world().type_pu32(), 128);
-    auto push_constant_ptr_type = builder.declare_ptr_type(spv::StorageClassPushConstant, convert(push_constant_arr_type)->type_id);
-    auto push_constant_ptr = builder_->variable(push_constant_ptr_type, spv::StorageClassPushConstant);
+    auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
+    auto push_constant_struct_type = builder.declare_struct_type({ push_constant_arr_type });
+    auto push_constant_struct_ptr_type = builder.declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
+    builder.name(push_constant_struct_type, "[i32 * 128]");
+    builder.decorate(push_constant_struct_type, spv::DecorationBlock);
+    builder.decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
+    builder.decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
+    auto push_constant_struct_ptr = builder_->variable(push_constant_struct_ptr_type, spv::StorageClassPushConstant);
+    builder.name(push_constant_struct_ptr, "push_constant_data");
 
     auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
     for (auto& cont : world().continuations()) {
@@ -54,10 +60,10 @@ void CodeGen::emit_stream(std::ostream& out) {
             BasicBlockBuilder* bb = fn_builder.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn_builder)).get();
             fn_builder.bbs_to_emit.push_back(bb);
 
-            // iterate on cont type and extract the
+            // iterate on cont type and extract the arguments
             auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Push))->type_id;
             auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
-            auto ptr_arr = bb->ptr_access_chain(ptr_type, push_constant_ptr, zero, { zero });
+            auto ptr_arr = bb->ptr_access_chain(ptr_type, push_constant_struct_ptr, zero, { zero, zero });
             size_t offset = 0;
             std::vector<SpvId> args;
             for (size_t i = 0; i < cont->num_params(); i++) {
@@ -79,7 +85,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             builder_->define_function(fn_builder);
             builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
 
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", { push_constant_ptr });
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", { push_constant_struct_ptr });
 
             auto block = config->second->as<GPUKernelConfig>()->block_size();
             std::vector<uint32_t> local_size = {
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 7bfbd4451..b52ae61d8 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -379,7 +379,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_struct_type(std::vector<SpvId>& elements) {
+    SpvId declare_struct_type(std::vector<SpvId> elements) {
         types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -388,10 +388,21 @@ struct SpvFileBuilder {
         return id;
     }
 
-    void decorate(SpvId target, spv::Decoration decoration) {
-        annotations.op(spv::Op::OpDecorate, 3);
+    void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+        annotations.op(spv::Op::OpDecorate, 3 + extra.size());
         annotations.ref_id(target);
         annotations.literal_int(decoration);
+        for (auto e : extra)
+            annotations.literal_int(e);
+    }
+
+    void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+        annotations.op(spv::Op::OpMemberDecorate, 4 + extra.size());
+        annotations.ref_id(target);
+        annotations.literal_int(member);
+        annotations.literal_int(decoration);
+        for (auto e : extra)
+            annotations.literal_int(e);
     }
 
     SpvId bool_constant(SpvId type, bool value) {

From 81f84ea8822510f1d3748238a562f164b098610d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 27 Apr 2021 16:35:50 +0200
Subject: [PATCH 067/342] limit usage of OpPtrAccessChain as much as possible

---
 src/thorin/be/spirv/spirv.cpp           | 33 ++++++-----
 src/thorin/be/spirv/spirv.h             | 20 +++----
 src/thorin/be/spirv/spirv_builder.hpp   | 11 ++++
 src/thorin/be/spirv/spirv_datatypes.cpp | 79 ++++++++++++++-----------
 4 files changed, 85 insertions(+), 58 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index d7f6c765d..a853bd0f1 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -38,12 +38,12 @@ void CodeGen::emit_stream(std::ostream& out) {
     auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
     auto push_constant_struct_type = builder.declare_struct_type({ push_constant_arr_type });
     auto push_constant_struct_ptr_type = builder.declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
-    builder.name(push_constant_struct_type, "[i32 * 128]");
+    builder.name(push_constant_struct_type, "ThorinPushConstant");
     builder.decorate(push_constant_struct_type, spv::DecorationBlock);
     builder.decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
     builder.decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
     auto push_constant_struct_ptr = builder_->variable(push_constant_struct_ptr_type, spv::StorageClassPushConstant);
-    builder.name(push_constant_struct_ptr, "push_constant_data");
+    builder.name(push_constant_struct_ptr, "thorin_push_constant_data");
 
     auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
     for (auto& cont : world().continuations()) {
@@ -61,10 +61,10 @@ void CodeGen::emit_stream(std::ostream& out) {
             fn_builder.bbs_to_emit.push_back(bb);
 
             // iterate on cont type and extract the arguments
-            auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Push))->type_id;
-            auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
-            auto ptr_arr = bb->ptr_access_chain(ptr_type, push_constant_struct_ptr, zero, { zero, zero });
-            size_t offset = 0;
+            auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push))->type_id;
+            auto zero = bb->file_builder.constant(convert(world().type_pu32())->type_id, { 0 });
+            auto arr_ref = bb->access_chain(ptr_type, push_constant_struct_ptr, { zero });
+            uint32_t offset = 0;
             std::vector<SpvId> args;
             for (size_t i = 0; i < cont->num_params(); i++) {
                 auto param = cont->param(i);
@@ -73,10 +73,9 @@ void CodeGen::emit_stream(std::ostream& out) {
                 assert(param_type->order() == 0);
                 auto converted = convert(param_type);
                 assert(converted->datatype != nullptr);
-                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, ptr_arr);
+                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.constant(convert(world().type_pu32())->type_id, { offset }));
                 args.push_back(arg);
                 offset += converted->datatype->serialized_size();
-                ptr_arr = bb->ptr_access_chain(ptr_type, ptr_arr, bb->file_builder.constant(convert(world().type_ps32())->type_id, { (uint32_t) offset }), { });
             }
 
             bb->call(builder_->void_type, callee, args);
@@ -529,10 +528,9 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
             auto converted_payload_type = convert(variant_type->op(variant->index()));
 
-            auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
-            auto ptr_arr = bb->ptr_access_chain(ptr_type, payload_arr, zero, { zero });
+            auto zero = bb->file_builder.constant(convert(world().type_pu32())->type_id, { 0 });
 
-            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, ptr_arr, emit(variant->value(), bb));
+            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, zero, emit(variant->value(), bb));
             auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
 
             auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index())});
@@ -557,9 +555,8 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
-        auto zero = bb->file_builder.constant(convert(world().type_ps32())->type_id, { 0 });
-        auto ptr_arr = bb->ptr_access_chain(ptr_type, payload_arr, zero, { zero });
-        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, ptr_arr);
+        auto zero = bb->file_builder.constant(convert(world().type_pu32())->type_id, { 0 });
+        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, zero);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
         return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
@@ -593,6 +590,14 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             return spv_none;
         } else THORIN_UNREACHABLE;
     } else if (auto lea = def->isa<LEA>()) {
+        switch (lea->ptr_type()->addr_space()) {
+            case AddrSpace::Global:
+            case AddrSpace::Shared:
+                break;
+            default:
+                world().ELOG("LEA is only allowed in global & shared address spaces");
+                break;
+        }
         auto type = convert(lea->ptr_type());
         auto offset = emit(lea->index(), bb);
         return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index d5a0d5649..630ef6153 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -76,8 +76,8 @@ struct Datatype {
     Datatype(ConvertedType* type) : type(type) {}
 
     virtual size_t serialized_size() = 0;
-    virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) = 0;
-    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) = 0;
+    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) = 0;
+    virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) = 0;
 };
 
 /// For scalar datatypes
@@ -88,8 +88,8 @@ struct ScalarDatatype : public Datatype {
     ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
 
     size_t serialized_size() override { return (size_in_bytes + 3) / 4; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
 };
 
 struct PtrDatatype : public Datatype {
@@ -97,8 +97,8 @@ struct PtrDatatype : public Datatype {
     PtrDatatype(ConvertedType* type) : Datatype(type) {}
 
     size_t serialized_size() override { return bitwidth / 32; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
 };
 
 struct DefiniteArrayDatatype : public Datatype {
@@ -108,8 +108,8 @@ struct DefiniteArrayDatatype : public Datatype {
     DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length);
 
     size_t serialized_size() override { return element_type->datatype->serialized_size(); };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
 };
 
 struct ProductDatatype : public Datatype {
@@ -119,8 +119,8 @@ struct ProductDatatype : public Datatype {
     ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types);
 
     size_t serialized_size() override { return total_size; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) override;
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
 };
 
 }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index b52ae61d8..3aa1bd5ae 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -144,6 +144,17 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
+        op(spv::Op::OpAccessChain, 4 + indexes.size());
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(element);
+        for (auto index : indexes)
+            ref_id(index);
+        return id;
+    }
+
     SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
         op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
         auto id = generate_fresh_id();
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index ec9247fb7..4b52f79ce 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -9,31 +9,39 @@ ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in
 
 }
 
-SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
+/// All serialization/deserialization methods use this so into a macro it goes
+#define serialization_types \
+SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id; \
+SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
+
+SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
     /// currently limited to 32-bit
     assert(size_in_bytes == 4);
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    auto loaded = bb.load(u32_tid, input);
+    serialization_types;
+    auto cell = bb.access_chain(arr_cell_tid, array, { base_offset });
+    auto loaded = bb.load(u32_tid, cell);
     return bb.bitcast(type->type_id, loaded);
 }
 
-void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
+void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
     /// currently limited to 32-bit
     assert(size_in_bytes == 4);
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    serialization_types;
+    auto cell = bb.access_chain(arr_cell_tid, array, { base_offset });
     auto casted = bb.bitcast(u32_tid, data);
-    bb.store(casted, output);
+    bb.store(casted, cell);
 }
 
-SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
+SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
     assert(type->src_type->as<PtrType>()->addr_space() == AddrSpace::Global && "Only buffer device address (global memory) pointers supported");
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
+    serialization_types;
     SpvId u64_tid = type->code_gen->convert(type->code_gen->world().type_pu64())->type_id;
-    SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
 
-    auto input2 = bb.ptr_access_chain(arr_cell_tid, input, bb.file_builder.constant(u32_tid, { (uint32_t) 1 }), { });
-    auto upper = bb.u_convert(u64_tid, bb.load(u32_tid, input));
-    auto lower = bb.u_convert(u64_tid, bb.load(u32_tid, input2));
+    auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
+    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.constant(u32_tid, { (uint32_t) 1 })) });
+
+    auto upper = bb.u_convert(u64_tid, bb.load(u32_tid, cell0));
+    auto lower = bb.u_convert(u64_tid, bb.load(u32_tid, cell1));
 
     SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
     auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32), lower);
@@ -41,7 +49,7 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
     return bb.convert_u_ptr(type->type_id, merged);
 }
 
-void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
+void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
     assert(false && "TODO");
 }
 
@@ -50,57 +58,60 @@ DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType*
     assert(length > 0 && "Array lengths of zero are not supported");
 }
 
-SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
-    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_ps32())->type_id;
+SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
+    serialization_types;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
+    SpvId offset = base_offset;
+    SpvId stride = bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() });
     for (size_t i = 0; i < length; i++) {
-        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, input, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type->datatype->serialized_size()) }), indices);
-        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, element_ptr);
+        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, array, offset);
         elements.push_back(element);
+        offset = bb.binop(spv::OpIAdd, u32_tid, offset, stride);
     }
     return bb.composite(type->type_id, elements);
 }
-void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
+void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
+    serialization_types;
     std::vector<SpvId> indices;
-    SpvId i32_tid = type->code_gen->convert(type->code_gen->world().type_ps32())->type_id;
+    SpvId offset = base_offset;
+    SpvId stride = bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() });
     for (size_t i = 0; i < length; i++) {
-        SpvId element_ptr = bb.ptr_access_chain(element_type->type_id, output, bb.file_builder.constant(i32_tid, { (uint32_t) (i * element_type->datatype->serialized_size()) }), indices);
-        element_type->datatype->emit_serialization(bb, storage_class, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i }));
+        element_type->datatype->emit_serialization(bb, storage_class, array, offset, bb.extract(element_type->type_id, data, { (uint32_t) i }));
+        offset = bb.binop(spv::OpIAdd, u32_tid, offset, stride);
     }
 }
 
 ProductDatatype::ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types) : Datatype(type), elements_types(elements_types) {
+    // Unit datatype is acceptable, but serdes methods should never be invoked.
     for (auto& element_type : elements_types) {
+        assert(element_type->datatype != nullptr);
         total_size += element_type->datatype->serialized_size();
     }
 }
 
-SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId input) {
+SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
     assert(total_size > 0 && "It doesn't make sense to de-serialize Unit!");
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
+    serialization_types;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
-    size_t offset = 0;
+    SpvId offset = base_offset;
     for (auto& element_type : elements_types) {
-        SpvId element_ptr = bb.ptr_access_chain(arr_cell_tid, input, bb.file_builder.constant(u32_tid, { (uint32_t) offset }), indices);
-        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, element_ptr);
-        offset += element_type->datatype->serialized_size();
+        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, array, offset);
+        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() }));
         elements.push_back(element);
     }
     return bb.composite(type->type_id, elements);
 }
-void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId output, SpvId data) {
+void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
     assert(total_size > 0 && "It doesn't make sense to serialize Unit!");
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
+    serialization_types;
     std::vector<SpvId> indices;
-    size_t offset = 0;
+    SpvId offset = base_offset;
     int i = 0;
     for (auto& element_type : elements_types) {
-        SpvId element_ptr = bb.ptr_access_chain(arr_cell_tid, output, bb.file_builder.constant(u32_tid, { (uint32_t) offset }), indices);
-        element_type->datatype->emit_serialization(bb, storage_class, element_ptr, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
+        element_type->datatype->emit_serialization(bb, storage_class, array, offset, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
+        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() }));
     }
 }
 

From 54749f8a771d508f346e51fd326981a8c2882711 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 28 Apr 2021 12:42:56 +0200
Subject: [PATCH 068/342] ignore imported defs in transformations

---
 src/thorin/be/spirv/spirv_transform.cpp | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/be/spirv/spirv_transform.cpp
index 51f613f2a..f210a8a7a 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/be/spirv/spirv_transform.cpp
@@ -150,12 +150,11 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
         for (size_t i = 0; i < cont->num_ops(); i++) {
             auto def = cont->op(i);
             if (auto dest = def->isa_continuation()) {
-                const Head* source_loop_head = ctx.def2loop[cont];
-
-                if (dest->intrinsic() == Intrinsic::Branch) {
+                if (dest->intrinsic() == Intrinsic::Branch || dest->is_imported()) {
                     continue;
                 }
 
+                const Head* source_loop_head = ctx.def2loop[cont];
                 assert(ctx.def2loop.find(dest) != ctx.def2loop.end());
                 const Head* dest_loop_head = ctx.def2loop[dest];
 

From d36ada78175cda15de7f0374377929c1237ae260 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 28 Apr 2021 13:28:34 +0200
Subject: [PATCH 069/342] added spirv.nonsemantic.printf

---
 src/thorin/be/spirv/spirv.cpp         | 25 ++++++++++++++++++++
 src/thorin/be/spirv/spirv.h           |  2 ++
 src/thorin/be/spirv/spirv_builder.hpp | 33 +++++++++++++++++++++++++++
 3 files changed, 60 insertions(+)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index a853bd0f1..8c88d2916 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -28,6 +28,9 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
 
+    builder.extension("SPV_KHR_non_semantic_info");
+    non_semantic_info = builder_->extended_import("NonSemantic.DebugPrintf");
+
     structure_loops();
     structure_flow();
     // cleanup_world(world());
@@ -306,6 +309,28 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
         }
         bb->branch(current_fn_->labels[callee]);
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee->is_imported()) {
+        if (callee->name() == "spirv.nonsemantic.printf") {
+            std::vector<SpvId> args;
+            auto string = continuation->arg(1);
+            if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
+                auto arr = string->as<DefiniteArray>();
+                std::vector<char> the_string;
+                for (int i = 0; i < arr_type->dim(); i++)
+                    the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
+                the_string.push_back('\0');
+                args.push_back(builder_->debug_string(the_string.data()));
+            } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
+
+            // TODO handle printing values
+
+            auto values = continuation->arg(2);
+            bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, args);
+        } else {
+            world().ELOG("This spir-v builtin isn't recognised: %s", callee->name());
+        }
+        auto next = continuation->args().back()->as_continuation();
+        emit_epilogue(next, bb);
     }
     /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         auto ret_continuation = emit_intrinsic(irbuilder, continuation);
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 630ef6153..73f1671f3 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -64,6 +64,8 @@ class CodeGen : public thorin::CodeGen {
     TypeMap<std::unique_ptr<ConvertedType>> types_;
     DefMap<SpvId> defs_;
     const Cont2Config& kernel_config_;
+
+    SpvId non_semantic_info;
 };
 
 /// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 3aa1bd5ae..2c5a83529 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -236,6 +236,18 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
+        op(spv::Op::OpExtInst, 5 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(set);
+        literal_int(instruction);
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
+    }
+
     void return_void() {
         op(spv::Op::OpReturn, 1);
     }
@@ -416,6 +428,14 @@ struct SpvFileBuilder {
             annotations.literal_int(e);
     }
 
+    SpvId debug_string(std::string string) {
+        debug_string_source.op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
+        auto id = generate_fresh_id();
+        debug_string_source.ref_id(id);
+        debug_string_source.literal_name(string);
+        return id;
+    }
+
     SpvId bool_constant(SpvId type, bool value) {
         types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
         auto id = generate_fresh_id();
@@ -506,6 +526,19 @@ struct SpvFileBuilder {
         capabilities.data_.push_back(cap);
     }
 
+    void extension(std::string name) {
+        extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
+        extensions.literal_name(name);
+    }
+
+    SpvId extended_import(std::string name) {
+        ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
+        auto id = generate_fresh_id();
+        ext_inst_import.ref_id(id);
+        ext_inst_import.literal_name(name);
+        return id;
+    }
+
     spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
     spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
 

From b1fc726d8b14fbf952f038970f88efb6996aad04 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 28 Apr 2021 20:53:09 +0200
Subject: [PATCH 070/342] messing with caps

---
 src/thorin/be/spirv/spirv.cpp         | 3 ++-
 src/thorin/be/spirv/spirv_builder.hpp | 2 +-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 8c88d2916..c23d58f93 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -23,7 +23,7 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_->capability(spv::Capability::CapabilityShader);
     builder_->capability(spv::Capability::CapabilityVariablePointers);
     builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
-    builder_->capability(spv::Capability::CapabilityInt16);
+    // builder_->capability(spv::Capability::CapabilityInt16);
     builder_->capability(spv::Capability::CapabilityInt64);
 
     builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
@@ -535,6 +535,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
         return constant;
     } else if (auto param = def->isa<Param>()) {
+        if (is_mem(param)) return spv_none;
         if (auto param_id = current_fn_->params.lookup(param)) {
             assert((*param_id).id != 0);
             return *param_id;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 2c5a83529..afefcb984 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -540,7 +540,7 @@ struct SpvFileBuilder {
     }
 
     spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
-    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelGLSL450    ;
 
 private:
     std::ostream* output_ = nullptr;

From 557aa2f893fb485cc67f7c1853297ff227d28975 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 28 Apr 2021 20:58:19 +0200
Subject: [PATCH 071/342] handles printing values

---
 src/thorin/be/spirv/spirv.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c23d58f93..e6178180c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -322,7 +322,9 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
                 args.push_back(builder_->debug_string(the_string.data()));
             } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
 
-            // TODO handle printing values
+            for (int i = 2; i < continuation->num_args() - 1; i++) {
+                args.push_back(emit(continuation->arg(i), bb));
+            }
 
             auto values = continuation->arg(2);
             bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, args);

From 85cfd46878bbe33d7e706310330cd8282e849ac4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 28 Apr 2021 22:13:29 +0200
Subject: [PATCH 072/342] u64 literals

---
 src/thorin/be/spirv/spirv.cpp           | 16 ++++++++++++++--
 src/thorin/be/spirv/spirv_datatypes.cpp |  6 +++---
 2 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index e6178180c..6dfa3f244 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -77,6 +77,10 @@ void CodeGen::emit_stream(std::ostream& out) {
                 auto converted = convert(param_type);
                 assert(converted->datatype != nullptr);
                 SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.constant(convert(world().type_pu32())->type_id, { offset }));
+                std::vector<SpvId> printf_args;
+                printf_args.push_back(builder_->debug_string("arg " + std::to_string((int)i) + " = %ul\n"));
+                printf_args.push_back(arg);
+                bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, printf_args);
                 args.push_back(arg);
                 offset += converted->datatype->serialized_size();
             }
@@ -529,8 +533,14 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
             case PrimType_ps32: case PrimType_qs32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
             case PrimType_pu32: case PrimType_qu32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
-            case PrimType_ps64: case PrimType_qs64: assertf(false, "not implemented yet");
-            case PrimType_pu64: case PrimType_qu64: assertf(false, "not implemented yet");
+            case PrimType_ps64: case PrimType_qs64:
+            case PrimType_pu64: case PrimType_qu64: {
+                uint64_t value = static_cast<uint64_t>(box.get_u64());
+                uint64_t upper = value >> 32U;
+                uint64_t lower = value & 0xFFFFFFFFU;
+                constant = bb->file_builder.constant(type, { (uint32_t) lower, (uint32_t) upper });
+                break;
+            }
             case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
             case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
             case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
@@ -629,6 +639,8 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto type = convert(lea->ptr_type());
         auto offset = emit(lea->index(), bb);
         return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
+    } else if (auto bitcast = def->isa<Bitcast>()) {
+        return bb->bitcast(convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
     }
     assertf(false, "Incomplete emit(def) definition");
 }
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 4b52f79ce..ae9eca006 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -40,11 +40,11 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
     auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
     auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.constant(u32_tid, { (uint32_t) 1 })) });
 
-    auto upper = bb.u_convert(u64_tid, bb.load(u32_tid, cell0));
-    auto lower = bb.u_convert(u64_tid, bb.load(u32_tid, cell1));
+    auto lower = bb.u_convert(u64_tid, bb.load(u32_tid, cell0));
+    auto upper = bb.u_convert(u64_tid, bb.load(u32_tid, cell1));
 
     SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
-    auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32), lower);
+    auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32));
 
     return bb.convert_u_ptr(type->type_id, merged);
 }

From 3678dfb9a7a8564ce5c9ca3e71f8cf7945d873cd Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 29 Apr 2021 10:57:07 +0200
Subject: [PATCH 073/342] implement extract and fix up codegen

---
 src/thorin/be/spirv/spirv.cpp           | 88 ++++++++++++++++++++++---
 src/thorin/be/spirv/spirv.h             |  7 +-
 src/thorin/be/spirv/spirv_datatypes.cpp |  7 ++
 3 files changed, 93 insertions(+), 9 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 6dfa3f244..7d09ab60b 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -56,7 +56,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
             SpvId callee = defs_[cont];
 
-            FnBuilder fn_builder(builder_);
+            FnBuilder fn_builder(this, builder_);
             fn_builder.fn_type = entry_pt_signature;
             fn_builder.fn_ret_type = builder_->void_type;
 
@@ -77,10 +77,6 @@ void CodeGen::emit_stream(std::ostream& out) {
                 auto converted = convert(param_type);
                 assert(converted->datatype != nullptr);
                 SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.constant(convert(world().type_pu32())->type_id, { offset }));
-                std::vector<SpvId> printf_args;
-                printf_args.push_back(builder_->debug_string("arg " + std::to_string((int)i) + " = %ul\n"));
-                printf_args.push_back(arg);
-                bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, printf_args);
                 args.push_back(arg);
                 offset += converted->datatype->serialized_size();
             }
@@ -111,7 +107,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
-    FnBuilder fn(builder_);
+    FnBuilder fn(this, builder_);
     fn.scope = &scope;
     fn.fn_type = convert(entry_->type())->type_id;
     fn.fn_ret_type = get_codom_type(entry_);
@@ -336,7 +332,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             world().ELOG("This spir-v builtin isn't recognised: %s", callee->name());
         }
         auto next = continuation->args().back()->as_continuation();
-        emit_epilogue(next, bb);
+        bb->branch(current_fn_->bbs_map[next]->label);
+        //emit_epilogue(next, bb);
     }
     /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         auto ret_continuation = emit_intrinsic(irbuilder, continuation);
@@ -615,6 +612,9 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
         return bb->composite(convert(structagg->type())->type_id, elements);
     } else if (auto access = def->isa<Access>()) {
+        // emit dependent operations first
+        emit(access->mem(), bb);
+
         std::vector<uint32_t> operands;
         auto ptr_type = access->ptr()->type()->as<PtrType>();
         if (ptr_type->addr_space() == AddrSpace::Global) {
@@ -641,12 +641,84 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
     } else if (auto bitcast = def->isa<Bitcast>()) {
         return bb->bitcast(convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
+    } else if (auto aggop = def->isa<AggOp>()) {
+        auto spv_agg = emit(aggop->agg(), bb);
+        auto agg_type = convert(aggop->agg()->type())->type_id;
+
+        bool mem = false;
+        if (auto tt = aggop->agg()->type()->isa<TupleType>(); tt && tt->op(0)->isa<MemType>()) mem = true;
+
+        auto copy_to_alloca = [&] (SpvId target_type) {
+            world().wdef(def, "slow: alloca and loads/stores needed for aggregate '{}'", def);
+            auto agg_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, agg_type);
+
+            auto variable = bb->fn_builder.variable(agg_ptr_type, spv::StorageClassFunction);
+            bb->store(spv_agg, variable);
+
+            auto cell_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, target_type);
+            auto cell = bb->access_chain(cell_ptr_type, variable, { emit(aggop->index(), bb)} );
+            return std::make_pair(variable, cell);
+        };
+        /*auto copy_to_alloca_or_global = [&] () -> llvm::Value* {
+            if (auto constant = llvm::dyn_cast<llvm::Constant>(llvm_agg)) {
+                auto global = llvm::cast<llvm::GlobalVariable>(module().getOrInsertGlobal(aggop->agg()->unique_name().c_str(), llvm_agg->getType()));
+                global->setLinkage(llvm::GlobalValue::InternalLinkage);
+                global->setInitializer(constant);
+                return irbuilder.CreateInBoundsGEP(global, { irbuilder.getInt64(0), llvm_idx });
+            }
+            return copy_to_alloca().second;
+        };*/
+
+        if (auto extract = aggop->isa<Extract>()) {
+            auto target_type = convert(extract->type())->type_id;
+            auto constant_index = aggop->index()->isa<PrimLit>();
+
+            // skip if the index is a constant
+            if (aggop->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
+                return bb->load(target_type, copy_to_alloca(target_type).second);
+            }
+
+            // TODO: evaluate what to do with those
+            // if (extract->agg()->type()->isa<VectorType>())
+            //     return irbuilder.CreateExtractElement(llvm_agg, llvm_idx);
+
+            // tuple/struct
+            if (is_mem(extract)) return spv_none;
+
+            // index *must* be constant
+            assert(constant_index != nullptr);
+            uint32_t index = constant_index->value().get_u32();
+
+            unsigned offset = 0;
+            if (mem) {
+                if (aggop->agg()->type()->num_ops() == 2) return spv_agg;
+                offset = 1;
+            }
+
+            return bb->extract(target_type, spv_agg, { index - offset });
+        }
+
+        THORIN_UNREACHABLE;
+        /*auto insert = def->as<Insert>();
+        auto value = emit(insert->value());
+
+        // TODO deal with mem - but I think for now this case shouldn't happen
+
+        if (insert->agg()->type()->isa<ArrayType>()) {
+            auto p = copy_to_alloca();
+            irbuilder.CreateStore(emit(aggop->as<Insert>()->value()), p.second);
+            return irbuilder.CreateLoad(p.first);
+        }
+        if (insert->agg()->type()->isa<VectorType>())
+            return irbuilder.CreateInsertElement(llvm_agg, emit(aggop->as<Insert>()->value()), llvm_idx);
+        // tuple/struct
+        return irbuilder.CreateInsertValue(llvm_agg, value, {primlit_value<unsigned>(aggop->index())});*/
     }
     assertf(false, "Incomplete emit(def) definition");
 }
 
 BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
-: builder::SpvBasicBlockBuilder(*fn_builder.file_builder) {
+: builder::SpvBasicBlockBuilder(*fn_builder.file_builder), fn_builder(fn_builder) {
     label = file_builder.generate_fresh_id();
 }
 
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 73f1671f3..9ecc0f8e0 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -10,6 +10,7 @@ using SpvId = builder::SpvId;
 
 class CodeGen;
 struct Datatype;
+struct PtrDatatype;
 
 struct ConvertedType {
     spirv::CodeGen* code_gen;
@@ -26,18 +27,20 @@ struct FnBuilder;
 struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     explicit BasicBlockBuilder(FnBuilder& fn_builder);
 
+    FnBuilder& fn_builder;
     std::unordered_map<const Param*, Phi> phis_map;
     DefMap<SpvId> args;
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {
+    CodeGen* cg;
     const Scope* scope = nullptr;
     std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
 
-    explicit FnBuilder(builder::SpvFileBuilder* file_builder) : builder::SpvFnBuilder(file_builder) {}
+    explicit FnBuilder(CodeGen* cg, builder::SpvFileBuilder* file_builder) : builder::SpvFnBuilder(file_builder), cg(cg) {}
 };
 
 class CodeGen : public thorin::CodeGen {
@@ -66,6 +69,8 @@ class CodeGen : public thorin::CodeGen {
     const Cont2Config& kernel_config_;
 
     SpvId non_semantic_info;
+
+    friend PtrDatatype;
 };
 
 /// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index ae9eca006..3f9039425 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -46,6 +46,13 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
     SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
     auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32));
 
+    std::vector<SpvId> printf_args;
+    printf_args.push_back(bb.file_builder.debug_string("lower  = %ul\nupper = %ul\nmerged = %ul\n"));
+    printf_args.push_back(lower);
+    printf_args.push_back(upper);
+    printf_args.push_back(merged);
+    bb.ext_instruction(bb.file_builder.void_type, bb.fn_builder.cg->non_semantic_info, 1, printf_args);
+
     return bb.convert_u_ptr(type->type_id, merged);
 }
 

From 5785ce533ddd1cb058d93180c5698260c1e583e0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 29 Apr 2021 12:43:24 +0200
Subject: [PATCH 074/342] removing debug print from PtrDatatype

---
 src/thorin/be/spirv/spirv_datatypes.cpp | 7 -------
 1 file changed, 7 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 3f9039425..ae9eca006 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -46,13 +46,6 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
     SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
     auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32));
 
-    std::vector<SpvId> printf_args;
-    printf_args.push_back(bb.file_builder.debug_string("lower  = %ul\nupper = %ul\nmerged = %ul\n"));
-    printf_args.push_back(lower);
-    printf_args.push_back(upper);
-    printf_args.push_back(merged);
-    bb.ext_instruction(bb.file_builder.void_type, bb.fn_builder.cg->non_semantic_info, 1, printf_args);
-
     return bb.convert_u_ptr(type->type_id, merged);
 }
 

From c820a6814a92fb21ee37db386e5f5ff7db1f03b7 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 3 May 2021 15:33:18 +0200
Subject: [PATCH 075/342] handle Bottom

---
 src/thorin/be/spirv/spirv.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 7d09ab60b..87a3abfe8 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -713,6 +713,8 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             return irbuilder.CreateInsertElement(llvm_agg, emit(aggop->as<Insert>()->value()), llvm_idx);
         // tuple/struct
         return irbuilder.CreateInsertValue(llvm_agg, value, {primlit_value<unsigned>(aggop->index())});*/
+    } else if (def->isa<Bottom>()) {
+        return bb->undef(convert(def->type())->type_id);
     }
     assertf(false, "Incomplete emit(def) definition");
 }

From c536b1523db0d702602bd43f23887ab591eee9c8 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 3 May 2021 16:25:22 +0200
Subject: [PATCH 076/342] support Insert

---
 src/thorin/be/spirv/spirv.cpp           | 50 +++++++++++++------------
 src/thorin/be/spirv/spirv_builder.hpp   | 12 ++++++
 src/thorin/be/spirv/spirv_datatypes.cpp |  4 +-
 3 files changed, 39 insertions(+), 27 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 87a3abfe8..0054c9b36 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -659,22 +659,17 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             auto cell = bb->access_chain(cell_ptr_type, variable, { emit(aggop->index(), bb)} );
             return std::make_pair(variable, cell);
         };
-        /*auto copy_to_alloca_or_global = [&] () -> llvm::Value* {
-            if (auto constant = llvm::dyn_cast<llvm::Constant>(llvm_agg)) {
-                auto global = llvm::cast<llvm::GlobalVariable>(module().getOrInsertGlobal(aggop->agg()->unique_name().c_str(), llvm_agg->getType()));
-                global->setLinkage(llvm::GlobalValue::InternalLinkage);
-                global->setInitializer(constant);
-                return irbuilder.CreateInBoundsGEP(global, { irbuilder.getInt64(0), llvm_idx });
-            }
-            return copy_to_alloca().second;
-        };*/
 
         if (auto extract = aggop->isa<Extract>()) {
+            if (is_mem(extract)) return spv_none;
+
             auto target_type = convert(extract->type())->type_id;
             auto constant_index = aggop->index()->isa<PrimLit>();
 
             // skip if the index is a constant
             if (aggop->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
+                assert(aggop->agg()->type()->isa<DefiniteArrayType>());
+                assert(!is_mem(extract));
                 return bb->load(target_type, copy_to_alloca(target_type).second);
             }
 
@@ -683,7 +678,6 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             //     return irbuilder.CreateExtractElement(llvm_agg, llvm_idx);
 
             // tuple/struct
-            if (is_mem(extract)) return spv_none;
 
             // index *must* be constant
             assert(constant_index != nullptr);
@@ -696,23 +690,31 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             }
 
             return bb->extract(target_type, spv_agg, { index - offset });
-        }
+        } else if (auto insert = def->isa<Insert>()) {
+            auto value = emit(insert->value(), bb);
+            auto constant_index = aggop->index()->isa<PrimLit>();
 
-        THORIN_UNREACHABLE;
-        /*auto insert = def->as<Insert>();
-        auto value = emit(insert->value());
+            // TODO deal with mem - but I think for now this case shouldn't happen
 
-        // TODO deal with mem - but I think for now this case shouldn't happen
+            if (insert->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
+                assert(aggop->agg()->type()->isa<DefiniteArrayType>());
+                auto [variable, cell] = copy_to_alloca(agg_type);
+                bb->store(value, cell);
+                return bb->load(agg_type, variable);
+            }
 
-        if (insert->agg()->type()->isa<ArrayType>()) {
-            auto p = copy_to_alloca();
-            irbuilder.CreateStore(emit(aggop->as<Insert>()->value()), p.second);
-            return irbuilder.CreateLoad(p.first);
-        }
-        if (insert->agg()->type()->isa<VectorType>())
-            return irbuilder.CreateInsertElement(llvm_agg, emit(aggop->as<Insert>()->value()), llvm_idx);
-        // tuple/struct
-        return irbuilder.CreateInsertValue(llvm_agg, value, {primlit_value<unsigned>(aggop->index())});*/
+            // TODO: evaluate what to do with those
+            //if (insert->agg()->type()->isa<VectorType>())
+            //    return irbuilder.CreateInsertElement(llvm_agg, emit(aggop->as<Insert>()->value()), llvm_idx);
+
+            // tuple/struct
+
+            // index *must* be constant
+            assert(constant_index != nullptr);
+            uint32_t index = constant_index->value().get_u32();
+
+            return bb->insert(agg_type, value, spv_agg, { index });
+        } else THORIN_UNREACHABLE;
     } else if (def->isa<Bottom>()) {
         return bb->undef(convert(def->type())->type_id);
     }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index afefcb984..254fb7cfb 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -106,6 +106,18 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
+        op(spv::Op::OpCompositeInsert, 5 + indices.size());
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(object);
+        ref_id(composite);
+        for (auto i : indices)
+            literal_int(i);
+        return id;
+    }
+
     SpvId bitcast(SpvId target_type, SpvId value) {
         op(spv::Op::OpBitcast, 4);
         auto id = generate_fresh_id();
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index ae9eca006..60d037388 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -129,8 +129,6 @@ ConvertedType* CodeGen::convert(const Type* type) {
     }
 
     if (auto iter = types_.find(type); iter != types_.end()) return iter->second.get();
-
-    assert(!type->isa<MemType>());
     ConvertedType* converted = types_.emplace(type, std::make_unique<ConvertedType>(this) ).first->second.get();
     converted->src_type = type;
     switch (type->tag()) {
@@ -313,7 +311,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
         }
 
         case Node_MemType: {
-            assert(false && "TODO: get arround this");
+            assert(false && "MemType cannot be converted to SPIR-V");
         }
 
         default:

From 34f3da58810e9ab8159eca87af427294b9899a8a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 3 May 2021 16:58:23 +0200
Subject: [PATCH 077/342] naive fn call implem

---
 src/thorin/be/spirv/spirv.cpp           | 71 ++++++++++++-------------
 src/thorin/be/spirv/spirv_datatypes.cpp | 10 +---
 2 files changed, 37 insertions(+), 44 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 0054c9b36..65a9cbc1e 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -231,9 +231,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto fbb = current_fn_->labels[continuation->arg(2)->as_continuation()];
         bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
         bb->branch_conditional(cond, tbb, fbb);
-    } /*else if (continuation->callee()->isa<Continuation>() &&
-               continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
-        auto val = emit(continuation->arg(0));
+    } else if (continuation->callee()->isa<Continuation>() && continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
+        /*auto val = emit(continuation->arg(0));
         auto otherwise_bb = cont2bb(continuation->arg(1)->as_continuation());
         auto match = irbuilder.CreateSwitch(val, otherwise_bb, continuation->num_args() - 2);
         for (size_t i = 2; i < continuation->num_args(); i++) {
@@ -241,11 +240,11 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
             auto case_bb    = cont2bb(arg->op(1)->as_continuation());
             match->addCase(case_const, case_bb);
-        }
+        }*/
+        THORIN_UNREACHABLE;
     } else if (continuation->callee()->isa<Bottom>()) {
-        irbuilder.CreateUnreachable();
-    } */
-    else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
+        bb->unreachable();
+    }  else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
         auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
         auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
         bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
@@ -333,39 +332,37 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
         auto next = continuation->args().back()->as_continuation();
         bb->branch(current_fn_->bbs_map[next]->label);
-        //emit_epilogue(next, bb);
-    }
-    /*else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
-        auto ret_continuation = emit_intrinsic(irbuilder, continuation);
-        irbuilder.CreateBr(cont2bb(ret_continuation));
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+        THORIN_UNREACHABLE;
+        //auto ret_continuation = emit_intrinsic(irbuilder, continuation);
+        //irbuilder.CreateBr(cont2bb(ret_continuation));
     } else { // function/closure call
         // put all first-order args into an array
-        std::vector<llvm::Value*> args;
+        std::vector<SpvId> args;
         const Def* ret_arg = nullptr;
         for (auto arg : continuation->args()) {
             if (arg->order() == 0) {
-                if (auto val = emit_unsafe(arg))
-                    args.push_back(val);
+                auto arg_type = arg->type();
+                if (arg_type == world().unit() || arg_type == world().mem_type()) continue;
+                args.push_back(emit(arg, bb));
             } else {
                 assert(!ret_arg);
                 ret_arg = arg;
             }
         }
 
-        llvm::CallInst* call = nullptr;
+        auto ret_type = get_codom_type(continuation);
+
+        SpvId call_result;
         if (auto callee = continuation->callee()->isa_continuation()) {
-            call = irbuilder.CreateCall(emit(callee), args);
-            if (callee->is_exported())
-                call->setCallingConv(kernel_calling_convention_);
-            else if (callee->cc() == CC::Device)
-                call->setCallingConv(device_calling_convention_);
-            else
-                call->setCallingConv(function_calling_convention_);
+            call_result = bb->call(ret_type, emit(callee, bb), args);
         } else {
             // must be a closure
-            auto closure = emit(callee);
-            args.push_back(irbuilder.CreateExtractValue(closure, 1));
-            call = irbuilder.CreateCall(irbuilder.CreateExtractValue(closure, 0), args);
+            THORIN_UNREACHABLE;
+
+            // auto closure = emit(callee);
+            // args.push_back(irbuilder.CreateExtractValue(closure, 1));
+            // call = irbuilder.CreateCall(irbuilder.CreateExtractValue(closure, 0), args);
         }
 
         // must be call + continuation --- call + return has been removed by codegen_prepare
@@ -381,33 +378,35 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
 
         if (n == 0) {
-            irbuilder.CreateBr(cont2bb(succ));
+            bb->branch(current_fn_->labels[succ]);
         } else if (n == 1) {
-            irbuilder.CreateBr(cont2bb(succ));
-            emit_phi_arg(irbuilder, last_param, call);
+            bb->branch(current_fn_->labels[succ]);
+
+            auto& phi = current_fn_->bbs_map[succ]->phis_map[last_param];
+            phi.preds.emplace_back(call_result, current_fn_->labels[continuation]);
         } else {
-            Array<llvm::Value*> extracts(n);
+            Array<SpvId> extracts(n);
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
                 auto param = succ->param(i);
                 if (is_mem(param) || is_unit(param))
                     continue;
-                extracts[j] = irbuilder.CreateExtractValue(call, unsigned(j));
+                extracts[j] = bb->extract(convert(param->type())->type_id, call_result, { (uint32_t) j });
                 j++;
             }
 
-            irbuilder.CreateBr(cont2bb(succ));
+            bb->branch(current_fn_->labels[succ]);
 
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
                 auto param = succ->param(i);
                 if (is_mem(param) || is_unit(param))
                     continue;
-                emit_phi_arg(irbuilder, param, extracts[j]);
+
+                auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
+                phi.preds.emplace_back(extracts[j], current_fn_->labels[continuation]);
+
                 j++;
             }
         }
-    }*/
-    else {
-        assert(false && "epilogue not implemented for this");
     }
 }
 
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 60d037388..4a11b756c 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -256,10 +256,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
             std::vector<SpvId> spv_types;
             size_t total_serialized_size = 0;
             for (auto member_type : type->ops()) {
-                if (member_type == world().unit() || member_type == world().mem_type()) {
-                    outf("skipped one");
-                    continue;
-                }
+                if (member_type == world().unit() || member_type == world().mem_type()) continue;
                 auto converted_member_type = convert(member_type);
                 types.push_back(converted_member_type);
                 spv_types.push_back(converted_member_type->type_id);
@@ -284,10 +281,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
 
             size_t max_serialized_size = 0;
             for (auto member_type : type->as<VariantType>()->ops()) {
-                if (member_type == world().unit() || member_type == world().mem_type()) {
-                    outf("skipped one");
-                    continue;
-                }
+                if (member_type == world().unit() || member_type == world().mem_type()) continue;
                 auto converted_member_type = convert(member_type);
                 if (converted_member_type->datatype->serialized_size() > max_serialized_size)
                     max_serialized_size = converted_member_type->datatype->serialized_size();

From eae1f8e776dc3850ff2e56a38ef4e08bb8e87c92 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 5 May 2021 11:01:05 +0200
Subject: [PATCH 078/342] moved builtins to their own function

---
 src/thorin/be/spirv/spirv.cpp | 52 +++++++++++++++++++----------------
 src/thorin/be/spirv/spirv.h   |  1 +
 2 files changed, 29 insertions(+), 24 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 65a9cbc1e..a652631f8 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -308,30 +308,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
         }
         bb->branch(current_fn_->labels[callee]);
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee->is_imported()) {
-        if (callee->name() == "spirv.nonsemantic.printf") {
-            std::vector<SpvId> args;
-            auto string = continuation->arg(1);
-            if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
-                auto arr = string->as<DefiniteArray>();
-                std::vector<char> the_string;
-                for (int i = 0; i < arr_type->dim(); i++)
-                    the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
-                the_string.push_back('\0');
-                args.push_back(builder_->debug_string(the_string.data()));
-            } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
-
-            for (int i = 2; i < continuation->num_args() - 1; i++) {
-                args.push_back(emit(continuation->arg(i), bb));
-            }
-
-            auto values = continuation->arg(2);
-            bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, args);
-        } else {
-            world().ELOG("This spir-v builtin isn't recognised: %s", callee->name());
-        }
-        auto next = continuation->args().back()->as_continuation();
-        bb->branch(current_fn_->bbs_map[next]->label);
+    } else if (auto builtin = continuation->callee()->isa_continuation(); builtin->is_imported()) {
+        emit_builtin(continuation, builtin, bb);
     } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         THORIN_UNREACHABLE;
         //auto ret_continuation = emit_intrinsic(irbuilder, continuation);
@@ -720,6 +698,32 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
+void CodeGen::emit_builtin(const Continuation* source_cont, const Continuation* builtin, BasicBlockBuilder* bb) {
+    if (builtin->name() == "spirv.nonsemantic.printf") {
+        std::vector<SpvId> args;
+        auto string = source_cont->arg(1);
+        if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
+            auto arr = string->as<DefiniteArray>();
+            std::vector<char> the_string;
+            for (int i = 0; i < arr_type->dim(); i++)
+                the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
+            the_string.push_back('\0');
+            args.push_back(builder_->debug_string(the_string.data()));
+        } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
+
+        for (int i = 2; i < source_cont->num_args() - 1; i++) {
+            args.push_back(emit(source_cont->arg(i), bb));
+        }
+
+        auto values = source_cont->arg(2);
+        bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, args);
+    } else {
+        world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
+    }
+    auto next = source_cont->args().back()->as_continuation();
+    bb->branch(current_fn_->bbs_map[next]->label);
+}
+
 BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
 : builder::SpvBasicBlockBuilder(*fn_builder.file_builder), fn_builder(fn_builder) {
     label = file_builder.generate_fresh_id();
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 9ecc0f8e0..201cfae33 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -58,6 +58,7 @@ class CodeGen : public thorin::CodeGen {
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     SpvId emit(const Def* def, BasicBlockBuilder* bb);
+    void emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
 
     SpvId get_codom_type(const Continuation* fn);
 

From 7fc55fb7f349e815d53a6c3d9365231d0665146b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 5 May 2021 11:20:50 +0200
Subject: [PATCH 079/342] move to a FileBuilder

---
 src/thorin/be/spirv/spirv.cpp         | 33 ++++++++++++++++---------
 src/thorin/be/spirv/spirv.h           | 35 +++++++++++++++++++++++----
 src/thorin/be/spirv/spirv_builder.hpp |  2 +-
 3 files changed, 53 insertions(+), 17 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index a652631f8..3d2f65803 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -17,19 +17,28 @@ CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
     : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
 {}
 
-void CodeGen::emit_stream(std::ostream& out) {
-    builder::SpvFileBuilder builder;
-    builder_ = &builder;
-    builder_->capability(spv::Capability::CapabilityShader);
-    builder_->capability(spv::Capability::CapabilityVariablePointers);
-    builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
-    // builder_->capability(spv::Capability::CapabilityInt16);
-    builder_->capability(spv::Capability::CapabilityInt64);
+FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg), builtins(*this), imported_instrs(*this) {
+    capability(spv::Capability::CapabilityShader);
+    capability(spv::Capability::CapabilityVariablePointers);
+    capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
+    // capability(spv::Capability::CapabilityInt16);
+    capability(spv::Capability::CapabilityInt64);
+
+    addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
+    memory_model = spv::MemoryModel::MemoryModelGLSL450;
+}
 
-    builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
+Builtins::Builtins(FileBuilder&) {
+}
 
+ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
     builder.extension("SPV_KHR_non_semantic_info");
-    non_semantic_info = builder_->extended_import("NonSemantic.DebugPrintf");
+    shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
+}
+
+void CodeGen::emit_stream(std::ostream& out) {
+    FileBuilder builder(this);
+    builder_ = &builder;
 
     structure_loops();
     structure_flow();
@@ -716,7 +725,9 @@ void CodeGen::emit_builtin(const Continuation* source_cont, const Continuation*
         }
 
         auto values = source_cont->arg(2);
-        bb->ext_instruction(bb->file_builder.void_type, non_semantic_info, 1, args);
+        bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs.shader_printf, 1, args);
+    } if (builtin->name() == "get_local_id") {
+
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
     }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 201cfae33..a3d947d87 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -12,6 +12,9 @@ class CodeGen;
 struct Datatype;
 struct PtrDatatype;
 
+struct FileBuilder;
+struct FnBuilder;
+
 struct ConvertedType {
     spirv::CodeGen* code_gen;
     const thorin::Type* src_type;
@@ -22,8 +25,6 @@ struct ConvertedType {
     bool is_known_size() { return datatype != nullptr; }
 };
 
-struct FnBuilder;
-
 struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     explicit BasicBlockBuilder(FnBuilder& fn_builder);
 
@@ -43,6 +44,32 @@ struct FnBuilder : public builder::SpvFnBuilder {
     explicit FnBuilder(CodeGen* cg, builder::SpvFileBuilder* file_builder) : builder::SpvFnBuilder(file_builder), cg(cg) {}
 };
 
+struct Builtins {
+    SpvId workgroup_size;
+    SpvId num_workgroups;
+    SpvId workgroup_id;
+    SpvId local_id;
+    SpvId global_id;
+    SpvId local_invocation_index;
+
+    explicit Builtins(FileBuilder&);
+};
+
+struct ImportedInstructions {
+    SpvId shader_printf;
+
+    explicit ImportedInstructions(FileBuilder&);
+};
+
+struct FileBuilder : public builder::SpvFileBuilder {
+    CodeGen* cg;
+
+    Builtins builtins;
+    ImportedInstructions imported_instrs;
+
+    explicit FileBuilder(CodeGen* cg);
+};
+
 class CodeGen : public thorin::CodeGen {
 public:
     CodeGen(World&, Cont2Config&, bool debug);
@@ -62,15 +89,13 @@ class CodeGen : public thorin::CodeGen {
 
     SpvId get_codom_type(const Continuation* fn);
 
-    builder::SpvFileBuilder* builder_ = nullptr;
+    FileBuilder* builder_ = nullptr;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
     TypeMap<std::unique_ptr<ConvertedType>> types_;
     DefMap<SpvId> defs_;
     const Cont2Config& kernel_config_;
 
-    SpvId non_semantic_info;
-
     friend PtrDatatype;
 };
 
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 254fb7cfb..371ab53ab 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -552,7 +552,7 @@ struct SpvFileBuilder {
     }
 
     spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
-    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelGLSL450    ;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
 
 private:
     std::ostream* output_ = nullptr;

From 1555d263c426cf8d82e12e8dd853749397df52e4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 5 May 2021 13:07:51 +0200
Subject: [PATCH 080/342] start implementing intrinsics properly

---
 src/thorin/be/spirv/spirv.cpp           | 113 +++++++++++++++++++-----
 src/thorin/be/spirv/spirv.h             |   8 +-
 src/thorin/be/spirv/spirv_builder.hpp   |  30 +++++++
 src/thorin/be/spirv/spirv_datatypes.cpp |  10 ++-
 4 files changed, 135 insertions(+), 26 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3d2f65803..c80bafe38 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -17,7 +17,7 @@ CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
     : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
 {}
 
-FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg), builtins(*this), imported_instrs(*this) {
+FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
     capability(spv::Capability::CapabilityShader);
     capability(spv::Capability::CapabilityVariablePointers);
     capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
@@ -28,7 +28,36 @@ FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg), built
     memory_model = spv::MemoryModel::MemoryModelGLSL450;
 }
 
-Builtins::Builtins(FileBuilder&) {
+Builtins::Builtins(FileBuilder& builder) {
+    auto& world = builder.cg->world();
+    auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
+    auto spv_uint_t = builder.cg->convert(world.type_pu32());
+    auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t->type_id);
+    auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t->type_id);
+
+    // workgroup_size = builder.constant(spv_uvec3_pt, spv::StorageClassInput);
+    // builder.decorate(workgroup_size, spv::DecorationBuiltIn, { spv::BuiltInWorkgroupSize });
+    // builder.name(workgroup_size, "BuiltInWorkgroupSize");
+
+    num_workgroups = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(num_workgroups, spv::DecorationBuiltIn, { spv::BuiltInNumWorkgroups });
+    builder.name(num_workgroups, "BuiltInNumWorkgroups");
+
+    workgroup_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(workgroup_id, spv::DecorationBuiltIn, { spv::BuiltInWorkgroupId });
+    builder.name(workgroup_id, "BuiltInWorkgroupId");
+
+    local_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(local_id, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationId });
+    builder.name(local_id, "BuiltInLocalInvocationId");
+
+    global_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(global_id, spv::DecorationBuiltIn, { spv::BuiltInGlobalInvocationId });
+    builder.name(global_id, "BuiltInGlobalInvocationId");
+
+    local_invocation_index = builder.variable(spv_uint_pt, spv::StorageClassInput);
+    builder.decorate(local_invocation_index, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationIndex });
+    builder.name(local_invocation_index, "BuiltInLocalInvocationIndex");
 }
 
 ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
@@ -37,8 +66,10 @@ ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
 }
 
 void CodeGen::emit_stream(std::ostream& out) {
-    FileBuilder builder(this);
-    builder_ = &builder;
+    builder_ = std::make_unique<FileBuilder>(this);
+
+    builder_->builtins = std::make_unique<Builtins>(*builder_);
+    builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
 
     structure_loops();
     structure_flow();
@@ -48,14 +79,14 @@ void CodeGen::emit_stream(std::ostream& out) {
     Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
 
     auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
-    auto push_constant_struct_type = builder.declare_struct_type({ push_constant_arr_type });
-    auto push_constant_struct_ptr_type = builder.declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
-    builder.name(push_constant_struct_type, "ThorinPushConstant");
-    builder.decorate(push_constant_struct_type, spv::DecorationBlock);
-    builder.decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
-    builder.decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
+    auto push_constant_struct_type = builder_->declare_struct_type({ push_constant_arr_type });
+    auto push_constant_struct_ptr_type = builder_->declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
+    builder_->name(push_constant_struct_type, "ThorinPushConstant");
+    builder_->decorate(push_constant_struct_type, spv::DecorationBlock);
+    builder_->decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
+    builder_->decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
     auto push_constant_struct_ptr = builder_->variable(push_constant_struct_ptr_type, spv::StorageClassPushConstant);
-    builder.name(push_constant_struct_ptr, "thorin_push_constant_data");
+    builder_->name(push_constant_struct_ptr, "thorin_push_constant_data");
 
     auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
     for (auto& cont : world().continuations()) {
@@ -65,7 +96,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
             SpvId callee = defs_[cont];
 
-            FnBuilder fn_builder(this, builder_);
+            FnBuilder fn_builder(this, builder_.get());
             fn_builder.fn_type = entry_pt_signature;
             fn_builder.fn_ret_type = builder_->void_type;
 
@@ -96,7 +127,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             builder_->define_function(fn_builder);
             builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
 
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", { push_constant_struct_ptr });
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", { push_constant_struct_ptr, builder_->builtins->local_id });
 
             auto block = config->second->as<GPUKernelConfig>()->block_size();
             std::vector<uint32_t> local_size = {
@@ -116,7 +147,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
-    FnBuilder fn(this, builder_);
+    FnBuilder fn(this, builder_.get());
     fn.scope = &scope;
     fn.fn_type = convert(entry_->type())->type_id;
     fn.fn_ret_type = get_codom_type(entry_);
@@ -163,7 +194,9 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    bb->phis_map[param] = {convert(param->type())->type_id, builder_->generate_fresh_id(), {} };
+                    auto type = convert(param->type())->type_id;
+                    assert(type.id != 0);
+                    bb->phis_map[param] = { type, builder_->generate_fresh_id(), {} };
                 }
             }
         }
@@ -202,6 +235,37 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
 }
 
 void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
+    // Handles the potential nuances of jumping to another continuation
+    auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
+        if (args.empty()) {
+            bb->branch(current_fn_->labels[succ]);
+        } else if (args.size() == 1) {
+            bb->branch(current_fn_->labels[succ]);
+
+            int i = 0;
+            while (is_mem(succ->param(i)) || is_unit(succ->param(i))) {
+                i++;
+                assert(i < succ->num_params());
+            }
+
+            auto& phi = current_fn_->bbs_map[succ]->phis_map[succ->param(i)];
+            phi.preds.emplace_back(args[0], current_fn_->labels[continuation]);
+        } else {
+            bb->branch(current_fn_->labels[succ]);
+
+            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+                auto param = succ->param(i);
+                if (is_mem(param) || is_unit(param))
+                    continue;
+
+                auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
+                phi.preds.emplace_back(args[j], current_fn_->labels[continuation]);
+
+                j++;
+            }
+        }
+    };
+
     if (continuation->callee() == entry_->ret_param()) {
         std::vector<SpvId> values;
 
@@ -318,7 +382,9 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
         bb->branch(current_fn_->labels[callee]);
     } else if (auto builtin = continuation->callee()->isa_continuation(); builtin->is_imported()) {
-        emit_builtin(continuation, builtin, bb);
+        auto productions = emit_builtin(continuation, builtin, bb);
+        auto succ = continuation->args().back()->as_continuation();
+        jump_to_next_cont_with_args(succ, productions);
     } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         THORIN_UNREACHABLE;
         //auto ret_continuation = emit_intrinsic(irbuilder, continuation);
@@ -707,7 +773,11 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-void CodeGen::emit_builtin(const Continuation* source_cont, const Continuation* builtin, BasicBlockBuilder* bb) {
+std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const Continuation* builtin, BasicBlockBuilder* bb) {
+    std::vector<SpvId> productions;
+    auto uvec3_t = convert(world().type_pu32(3));
+    auto u32_t = convert(world().type_pu32());
+    auto i32_t = convert(world().type_ps32());
     if (builtin->name() == "spirv.nonsemantic.printf") {
         std::vector<SpvId> args;
         auto string = source_cont->arg(1);
@@ -725,14 +795,15 @@ void CodeGen::emit_builtin(const Continuation* source_cont, const Continuation*
         }
 
         auto values = source_cont->arg(2);
-        bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs.shader_printf, 1, args);
+        bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
     } if (builtin->name() == "get_local_id") {
-
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->bitcast(i32_t->type_id, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
     }
-    auto next = source_cont->args().back()->as_continuation();
-    bb->branch(current_fn_->bbs_map[next]->label);
+    return productions;
 }
 
 BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index a3d947d87..91f5b6533 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -64,8 +64,8 @@ struct ImportedInstructions {
 struct FileBuilder : public builder::SpvFileBuilder {
     CodeGen* cg;
 
-    Builtins builtins;
-    ImportedInstructions imported_instrs;
+    std::unique_ptr<Builtins> builtins;
+    std::unique_ptr<ImportedInstructions> imported_instrs;
 
     explicit FileBuilder(CodeGen* cg);
 };
@@ -85,11 +85,11 @@ class CodeGen : public thorin::CodeGen {
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     SpvId emit(const Def* def, BasicBlockBuilder* bb);
-    void emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
+    std::vector<SpvId> emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
 
     SpvId get_codom_type(const Continuation* fn);
 
-    FileBuilder* builder_ = nullptr;
+    std::unique_ptr<FileBuilder> builder_;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
     TypeMap<std::unique_ptr<ConvertedType>> types_;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 371ab53ab..8881bc369 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -118,6 +118,27 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
+    SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
+        op(spv::Op::OpVectorExtractDynamic, 5);
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(vector);
+        ref_id(index);
+        return id;
+    }
+
+    SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
+        op(spv::Op::OpVectorInsertDynamic, 6);
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(vector);
+        ref_id(component);
+        ref_id(index);
+        return id;
+    }
+
     SpvId bitcast(SpvId target_type, SpvId value) {
         op(spv::Op::OpBitcast, 4);
         auto id = generate_fresh_id();
@@ -423,6 +444,15 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
+        types_constants.op(spv::Op::OpTypeVector, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(component_type);
+        types_constants.literal_int(dim);
+        return id;
+    }
+
     void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
         annotations.op(spv::Op::OpDecorate, 3 + extra.size());
         annotations.ref_id(target);
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 4a11b756c..3e6b27d7b 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -121,7 +121,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
     switch (type->tag()) {
 #define THORIN_Q_TYPE(T, M) \
     case PrimType_##T: \
-        type = world().prim_type(PrimType_p##M, 1); \
+        type = world().prim_type(PrimType_p##M, type->as<VectorType>()->length()); \
         break;
 #include "thorin/tables/primtypetable.h"
 #undef THORIN_Q_TYPE
@@ -131,6 +131,14 @@ ConvertedType* CodeGen::convert(const Type* type) {
     if (auto iter = types_.find(type); iter != types_.end()) return iter->second.get();
     ConvertedType* converted = types_.emplace(type, std::make_unique<ConvertedType>(this) ).first->second.get();
     converted->src_type = type;
+
+    if (auto vec = type->isa<VectorType>(); vec && vec->length() > 1) {
+        auto component = vec->scalarize();
+        auto conv_comp = convert(component);
+        converted->type_id = builder_->declare_vector_type(conv_comp->type_id, (uint32_t)vec->length());
+        return converted;
+    }
+
     switch (type->tag()) {
         // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
         // Using extensions, we could use 16 or 8-bit ints instead

From 96a346aefb6c4a8a952e7866e47b115ec3c56a49 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 5 May 2021 14:59:14 +0200
Subject: [PATCH 081/342] cleanup

---
 src/thorin/be/spirv/spirv.cpp | 76 ++++++++++++++---------------------
 1 file changed, 31 insertions(+), 45 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c80bafe38..40de8c2af 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -237,32 +237,14 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
 void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     // Handles the potential nuances of jumping to another continuation
     auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
-        if (args.empty()) {
-            bb->branch(current_fn_->labels[succ]);
-        } else if (args.size() == 1) {
-            bb->branch(current_fn_->labels[succ]);
-
-            int i = 0;
-            while (is_mem(succ->param(i)) || is_unit(succ->param(i))) {
-                i++;
-                assert(i < succ->num_params());
-            }
-
-            auto& phi = current_fn_->bbs_map[succ]->phis_map[succ->param(i)];
-            phi.preds.emplace_back(args[0], current_fn_->labels[continuation]);
-        } else {
-            bb->branch(current_fn_->labels[succ]);
-
-            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-                auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
-                    continue;
-
-                auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
-                phi.preds.emplace_back(args[j], current_fn_->labels[continuation]);
-
-                j++;
-            }
+        bb->branch(current_fn_->labels[succ]);
+        for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+            auto param = succ->param(i);
+            if (is_mem(param) || is_unit(param))
+                continue;
+            auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
+            phi.preds.emplace_back(args[j], current_fn_->labels[continuation]);
+            j++;
         }
     };
 
@@ -282,8 +264,19 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             case 1:  bb->return_value(values[0]); break;
             default: bb->return_value(bb->composite(current_fn_->fn_ret_type, values));
         }
-    }
-    else if (continuation->callee() == world().branch()) {
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+        int index = -1;
+        for (auto& arg : continuation->args()) {
+            index++;
+            auto val = emit(arg, bb);
+            if (is_mem(arg) || is_unit(arg)) continue;
+            bb->args[arg] = val;
+            auto* param = callee->param(index);
+            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
+            phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
+        }
+        bb->branch(current_fn_->labels[callee]);
+    } else if (continuation->callee() == world().branch()) {
         auto& domtree = current_fn_->scope->b_cfg().domtree();
         auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
         SpvId merge_bb;
@@ -369,35 +362,28 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto callee = continuation->op(0)->as_continuation();
         // TODO phis
         bb->branch(current_fn_->bbs_map[callee]->label);
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
-        int index = -1;
-        for (auto& arg : continuation->args()) {
-            index++;
-            auto val = emit(arg, bb);
-            if (is_mem(arg) || is_unit(arg)) continue;
-            bb->args[arg] = val;
-            auto* param = callee->param(index);
-            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
-            phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
-        }
-        bb->branch(current_fn_->labels[callee]);
     } else if (auto builtin = continuation->callee()->isa_continuation(); builtin->is_imported()) {
+        // Ensure we emit previous memory operations
+        assert(is_mem(continuation->arg(0)));
+        emit(continuation->arg(0), bb);
+
         auto productions = emit_builtin(continuation, builtin, bb);
         auto succ = continuation->args().back()->as_continuation();
         jump_to_next_cont_with_args(succ, productions);
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+    } else if (auto intrinsic = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         THORIN_UNREACHABLE;
         //auto ret_continuation = emit_intrinsic(irbuilder, continuation);
         //irbuilder.CreateBr(cont2bb(ret_continuation));
     } else { // function/closure call
         // put all first-order args into an array
-        std::vector<SpvId> args;
+        std::vector<SpvId> call_args;
         const Def* ret_arg = nullptr;
         for (auto arg : continuation->args()) {
             if (arg->order() == 0) {
                 auto arg_type = arg->type();
+                auto arg_val = emit(arg, bb);
                 if (arg_type == world().unit() || arg_type == world().mem_type()) continue;
-                args.push_back(emit(arg, bb));
+                call_args.push_back(arg_val);
             } else {
                 assert(!ret_arg);
                 ret_arg = arg;
@@ -407,8 +393,8 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto ret_type = get_codom_type(continuation);
 
         SpvId call_result;
-        if (auto callee = continuation->callee()->isa_continuation()) {
-            call_result = bb->call(ret_type, emit(callee, bb), args);
+        if (auto called_continuation = continuation->callee()->isa_continuation()) {
+            call_result = bb->call(ret_type, emit(called_continuation, bb), call_args);
         } else {
             // must be a closure
             THORIN_UNREACHABLE;

From 1fcc707e6b656a431a7339e87757ebb64182e808 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 5 May 2021 15:14:01 +0200
Subject: [PATCH 082/342] make a few more declaration types unique

---
 src/thorin/be/spirv/spirv_builder.hpp | 44 ++++++++++++++++++---------
 1 file changed, 29 insertions(+), 15 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 8881bc369..921b80189 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -338,28 +338,31 @@ struct SpvFnBuilder {
     SpvId generate_fresh_id();
 };
 
-inline bool operator==(const SpvId &a, const SpvId &b) { return a.id == b.id; }
-
 struct SpvFileBuilder {
-    enum UniqueTypeTag {
+
+    enum UniqueDeclTag {
         NONE,
-        FN_TYPE
+        FN_TYPE,
+        PTR_TYPE,
+        DEF_ARR_TYPE,
+        CONSTANT,
     };
 
-    struct UniqueTypeKey {
-        UniqueTypeTag tag;
-        std::vector<SpvId> members;
+    /// Prevents duplicate declarations
+    struct UniqueDeclKey {
+        UniqueDeclTag tag;
+        std::vector<uint32_t> members;
 
-        bool operator==(const UniqueTypeKey &b) const {
+        bool operator==(const UniqueDeclKey &b) const {
             return tag == b.tag && members == b.members;
         }
     };
 
-    struct UniqueTypeKeyHasher {
-        size_t operator() (const UniqueTypeKey& key) const {
+    struct UniqueDeclKeyHasher {
+        size_t operator() (const UniqueDeclKey& key) const {
             size_t acc = 0;
             for (auto id : key.members)
-                acc ^= std::hash<uint32_t>{}(id.id);
+                acc ^= std::hash<uint32_t>{}(id);
             return std::hash<size_t>{}(key.tag) ^ acc;
         }
     };
@@ -403,26 +406,33 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
+        auto key = UniqueDeclKey { PTR_TYPE, { element_type.id, (uint32_t) storage_class } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpTypePointer, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.literal_int(storage_class);
         types_constants.ref_id(element_type);
+        unique_decls[key] = id;
         return id;
     }
 
     SpvId declare_array_type(SpvId element_type, SpvId dim) {
+        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type.id, dim.id } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpTypeArray, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.ref_id(element_type);
         types_constants.ref_id(dim);
+        unique_decls[key] = id;
         return id;
     }
 
     SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
-        auto key = UniqueTypeKey { FN_TYPE, dom };
-        key.members.push_back(codom);
+        auto key = UniqueDeclKey { FN_TYPE, {} };
+        for (auto d : dom) key.members.push_back(d.id);
+        key.members.push_back(codom.id);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
 
         types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
@@ -486,13 +496,17 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId constant(SpvId type, std::vector<uint32_t>&& bit_pattern) {
+    SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
+        auto key = UniqueDeclKey { CONSTANT, bit_pattern };
+        key.members.push_back(type.id);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(type);
         types_constants.ref_id(id);
         for (auto arg : bit_pattern)
             types_constants.data_.push_back(arg);
+        unique_decls[key] = id;
         return id;
     }
 
@@ -603,7 +617,7 @@ struct SpvFileBuilder {
     SpvSectionBuilder fn_defs;
 
     // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
-    std::unordered_map<UniqueTypeKey, SpvId, UniqueTypeKeyHasher> unique_decls;
+    std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
 
     SpvId declare_void_type() {
         types_constants.op(spv::Op::OpTypeVoid, 2);

From 3430fe44acd2e88d1b4bf0aae3f1f3dfccbeda35 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 5 May 2021 20:42:04 +0200
Subject: [PATCH 083/342] decorate datatype pointers for GEP

---
 src/thorin/be/spirv/spirv_datatypes.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 3e6b27d7b..7d6d98e9f 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -208,6 +208,11 @@ ConvertedType* CodeGen::convert(const Type* type) {
                     pointee = arr->elem_type();
                 ConvertedType* element = convert(pointee);
                 converted->type_id = builder_->declare_ptr_type(storage_class, element->type_id);
+
+                if (ptr->addr_space() == AddrSpace::Global) {
+                    assert(element->datatype && "Can only have physical pointers to known-size types");
+                    builder_->decorate(converted->type_id, spv::DecorationArrayStride, {(uint32_t) element->datatype->serialized_size()});
+                }
             }
             ptr_done:
             break;

From ec1c6fce3e9a82b5c19e845ca832068bc19652e6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 12:56:58 +0200
Subject: [PATCH 084/342] fancy casting

---
 src/thorin/be/spirv/spirv.cpp           | 157 +++++++++++++++++++++++-
 src/thorin/be/spirv/spirv_builder.hpp   |  34 +----
 src/thorin/be/spirv/spirv_datatypes.cpp |  10 +-
 3 files changed, 159 insertions(+), 42 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 40de8c2af..b26f8ed98 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -13,6 +13,73 @@ namespace thorin {
 
 namespace thorin::spirv {
 
+/// Used as a dummy SSA value for emitting things like mem/unit
+/// Should never make it in the binary files !
+constexpr SpvId spv_none { 0 };
+
+// SPIR-V has 3 "kinds" of primitives, and the user may declare arbitrary bitwidths, the following helps in translation:
+enum class PrimTypeKind {
+    Signed, Unsigned, Float
+};
+inline PrimTypeKind classify_primtype(const PrimType* type) {
+    switch (type->tag()) {
+#define THORIN_QS_TYPE(T, M) THORIN_PS_TYPE(T, M)
+#define THORIN_PS_TYPE(T, M) \
+case PrimType_##T: \
+    return PrimTypeKind::Signed; \
+    break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_QS_TYPE
+#undef THORIN_PS_TYPE
+
+#define THORIN_QU_TYPE(T, M) THORIN_PU_TYPE(T, M)
+#define THORIN_PU_TYPE(T, M) \
+case PrimType_##T: \
+    return PrimTypeKind::Unsigned; \
+    break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_QU_TYPE
+#undef THORIN_PU_TYPE
+
+#define THORIN_QF_TYPE(T, M) THORIN_PF_TYPE(T, M)
+#define THORIN_PF_TYPE(T, M) \
+case PrimType_##T: \
+    return PrimTypeKind::Float; \
+    break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_QF_TYPE
+#undef THORIN_PF_TYPE
+        default: THORIN_UNREACHABLE;
+    }
+}
+inline const PrimType* get_primtype(World& world, PrimTypeKind kind, int bitwidth, int length) {
+#define GET_PRIMTYPE_WITH_KIND(kind) \
+switch (bitwidth) { \
+    case 8:  return world.type_p##kind##8(); \
+    case 16: return world.type_p##kind##16(); \
+    case 32: return world.type_p##kind##32(); \
+    case 64: return world.type_p##kind##64(); \
+}
+
+#define GET_PRIMTYPE_WITH_KIND_F(kind) \
+switch (bitwidth) { \
+    case 8: world.ELOG("8-bit floats do not exist"); \
+    case 16: return world.type_p##kind##16(); \
+    case 32: return world.type_p##kind##32(); \
+    case 64: return world.type_p##kind##64(); \
+}
+
+    switch (kind) {
+        case PrimTypeKind::Signed:   GET_PRIMTYPE_WITH_KIND(s);   THORIN_UNREACHABLE;
+        case PrimTypeKind::Unsigned: GET_PRIMTYPE_WITH_KIND(u);   THORIN_UNREACHABLE;
+        case PrimTypeKind::Float:    GET_PRIMTYPE_WITH_KIND_F(f); THORIN_UNREACHABLE;
+        default: THORIN_UNREACHABLE;
+    }
+
+#undef GET_PRIMTYPE_WITH_KIND
+#undef GET_PRIMTYPE_WITH_KIND_F
+}
+
 CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
     : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
 {}
@@ -449,8 +516,6 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     }
 }
 
-constexpr SpvId spv_none { 0 };
-
 SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
@@ -677,8 +742,6 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto type = convert(lea->ptr_type());
         auto offset = emit(lea->index(), bb);
         return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
-    } else if (auto bitcast = def->isa<Bitcast>()) {
-        return bb->bitcast(convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
     } else if (auto aggop = def->isa<AggOp>()) {
         auto spv_agg = emit(aggop->agg(), bb);
         auto agg_type = convert(aggop->agg()->type())->type_id;
@@ -753,6 +816,88 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
 
             return bb->insert(agg_type, value, spv_agg, { index });
         } else THORIN_UNREACHABLE;
+    } else if (auto conv = def->isa<ConvOp>()) {
+        auto src_type = conv->from()->type();
+        auto dst_type = conv->type();
+
+        auto conv_src_type = convert(src_type);
+        auto conv_dst_type = convert(dst_type);
+
+        if (auto bitcast = def->isa<Bitcast>()) {
+            if (conv_src_type->datatype->serialized_size() != conv_dst_type->datatype->serialized_size())
+                world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type->datatype->serialized_size(), conv_dst_type->datatype->serialized_size());
+
+            return bb->convert(spv::OpBitcast, convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
+        } else if (auto cast = def->isa<Cast>()) {
+            // NB: all ops used here are scalar/vector agnostic
+            auto src_prim = src_type->isa<PrimType>();
+            auto dst_prim = dst_type->isa<PrimType>();
+            if (!src_prim || !dst_prim || src_prim->length() != dst_prim->length())
+                world().ELOG("Illegal cast: % to %, casts are only supported between primitives with identical vector length", src_type->to_string(), dst_type->to_string());
+
+            auto length = src_prim->length();
+
+            auto src_kind = classify_primtype(src_prim);
+            auto dst_kind = classify_primtype(dst_prim);
+            size_t src_bitwidth = conv_src_type->datatype->serialized_size();
+            size_t dst_bitwidth = conv_src_type->datatype->serialized_size();
+
+            SpvId data = emit(cast->from(), bb);
+
+            // If floating point is involved (src or dst), OpConvert*ToF and OpConvertFTo* can take care of the bit width transformation so no need for any chopping/expanding
+            if (src_kind == PrimTypeKind::Float || dst_kind == PrimTypeKind::Float) {
+                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
+                switch (src_kind) {
+                    case PrimTypeKind::Signed:    data = bb->convert(spv::OpConvertSToF, target_type, data); break;
+                    case PrimTypeKind::Unsigned:  data = bb->convert(spv::OpConvertUToF, target_type, data); break;
+                    case PrimTypeKind::Float:
+                        switch (dst_kind) {
+                            case PrimTypeKind::Signed:   data = bb->convert(spv::OpConvertFToS, target_type, data); break;
+                            case PrimTypeKind::Unsigned: data = bb->convert(spv::OpConvertFToU, target_type, data); break;
+                            default: THORIN_UNREACHABLE;
+                        }
+                        break;
+                }
+            } else {
+                // we expand first and shrink last to minimize precision losses, with bitcast in the middle
+                bool needs_chopping = src_bitwidth > dst_bitwidth;
+                bool needs_expanding = src_bitwidth < dst_bitwidth;
+
+                if (needs_expanding) {
+                    auto target_type = convert(get_primtype(world(), src_kind, src_bitwidth, length))->type_id;
+                    switch (src_kind) {
+                        case PrimTypeKind::Signed:
+                            data = bb->convert(spv::OpSConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Unsigned:
+                            data = bb->convert(spv::OpUConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Float:
+                            data = bb->convert(spv::OpFConvert, target_type, data);
+                            break;
+                    }
+                }
+
+                auto expanded_bitwidth = needs_expanding ? dst_bitwidth : src_bitwidth;
+                auto bitcast_target_type = convert(get_primtype(world(), dst_kind, expanded_bitwidth, length))->type_id;
+                data = bb->convert(spv::OpBitcast, bitcast_target_type, data);
+
+                if (needs_chopping) {
+                    auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
+                    switch (dst_kind) {
+                        case PrimTypeKind::Signed:
+                            data = bb->convert(spv::OpSConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Unsigned:
+                            data = bb->convert(spv::OpUConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Float:
+                            data = bb->convert(spv::OpFConvert, target_type, data);
+                            break;
+                    }
+                }
+            }
+        } else THORIN_UNREACHABLE;
     } else if (def->isa<Bottom>()) {
         return bb->undef(convert(def->type())->type_id);
     }
@@ -782,10 +927,10 @@ std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const
 
         auto values = source_cont->arg(2);
         bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
-    } if (builtin->name() == "get_local_id") {
+    } else if (builtin->name() == "get_local_id") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
         auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
-        productions.push_back(bb->bitcast(i32_t->type_id, extracted));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
     }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 921b80189..44a2f287f 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -139,37 +139,9 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId bitcast(SpvId target_type, SpvId value) {
-        op(spv::Op::OpBitcast, 4);
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(value);
-        return id;
-    }
-
-    /// Change bit-width
-    SpvId u_convert(SpvId target_type, SpvId value) {
-        op(spv::Op::OpUConvert, 4);
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(value);
-        return id;
-    }
-
-    /// Change bit-width
-    SpvId s_convert(SpvId target_type, SpvId value) {
-        op(spv::Op::OpSConvert, 4);
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(value);
-        return id;
-    }
-
-    SpvId convert_u_ptr(SpvId target_type, SpvId value) {
-        op(spv::Op::OpConvertUToPtr, 4);
+    // Used for almost all conversion operations
+    SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
+        op(op_, 4);
         auto id = generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 7d6d98e9f..aaa255844 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -20,7 +20,7 @@ SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageCl
     serialization_types;
     auto cell = bb.access_chain(arr_cell_tid, array, { base_offset });
     auto loaded = bb.load(u32_tid, cell);
-    return bb.bitcast(type->type_id, loaded);
+    return bb.convert(spv::OpBitcast, type->type_id, loaded);
 }
 
 void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
@@ -28,7 +28,7 @@ void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass
     assert(size_in_bytes == 4);
     serialization_types;
     auto cell = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto casted = bb.bitcast(u32_tid, data);
+    auto casted = bb.convert(spv::OpBitcast, u32_tid, data);
     bb.store(casted, cell);
 }
 
@@ -40,13 +40,13 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
     auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
     auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.constant(u32_tid, { (uint32_t) 1 })) });
 
-    auto lower = bb.u_convert(u64_tid, bb.load(u32_tid, cell0));
-    auto upper = bb.u_convert(u64_tid, bb.load(u32_tid, cell1));
+    auto lower = bb.convert(spv::OpUConvert, u64_tid, bb.load(u32_tid, cell0));
+    auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.load(u32_tid, cell1));
 
     SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
     auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32));
 
-    return bb.convert_u_ptr(type->type_id, merged);
+    return bb.convert(spv::OpConvertUToPtr, type->type_id, merged);
 }
 
 void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {

From 2c187dc68d3bfbda4ececbb5bf004ec8dd24a11f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 13:07:33 +0200
Subject: [PATCH 085/342] implement insert/extract for vectors

---
 src/thorin/be/spirv/spirv.cpp | 20 +++++++-------------
 1 file changed, 7 insertions(+), 13 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index b26f8ed98..c91fe17f7 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -767,20 +767,17 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             auto target_type = convert(extract->type())->type_id;
             auto constant_index = aggop->index()->isa<PrimLit>();
 
-            // skip if the index is a constant
+            // We have a fast-path: if the index is constant, we can simply use OpCompositeExtract
             if (aggop->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
                 assert(aggop->agg()->type()->isa<DefiniteArrayType>());
                 assert(!is_mem(extract));
                 return bb->load(target_type, copy_to_alloca(target_type).second);
             }
 
-            // TODO: evaluate what to do with those
-            // if (extract->agg()->type()->isa<VectorType>())
-            //     return irbuilder.CreateExtractElement(llvm_agg, llvm_idx);
+            if (extract->agg()->type()->isa<VectorType>())
+                return bb->vector_extract_dynamic(target_type, spv_agg, emit(extract->index(), bb));
 
-            // tuple/struct
-
-            // index *must* be constant
+            // index *must* be constant for the remaining possible cases
             assert(constant_index != nullptr);
             uint32_t index = constant_index->value().get_u32();
 
@@ -804,13 +801,10 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
                 return bb->load(agg_type, variable);
             }
 
-            // TODO: evaluate what to do with those
-            //if (insert->agg()->type()->isa<VectorType>())
-            //    return irbuilder.CreateInsertElement(llvm_agg, emit(aggop->as<Insert>()->value()), llvm_idx);
-
-            // tuple/struct
+            if (insert->agg()->type()->isa<VectorType>())
+                return bb->vector_insert_dynamic(agg_type, spv_agg, value, emit(insert->index(), bb));
 
-            // index *must* be constant
+            // index *must* be constant for the remaining possible cases
             assert(constant_index != nullptr);
             uint32_t index = constant_index->value().get_u32();
 

From 62b4cb841ad57b92b2f00f882f3506b563e9d1bb Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 13:08:25 +0200
Subject: [PATCH 086/342] fix get_primtype with vectors

---
 src/thorin/be/spirv/spirv.cpp | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c91fe17f7..032be25db 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -55,18 +55,18 @@ case PrimType_##T: \
 inline const PrimType* get_primtype(World& world, PrimTypeKind kind, int bitwidth, int length) {
 #define GET_PRIMTYPE_WITH_KIND(kind) \
 switch (bitwidth) { \
-    case 8:  return world.type_p##kind##8(); \
-    case 16: return world.type_p##kind##16(); \
-    case 32: return world.type_p##kind##32(); \
-    case 64: return world.type_p##kind##64(); \
+    case 8:  return world.type_p##kind##8 (length); \
+    case 16: return world.type_p##kind##16(length); \
+    case 32: return world.type_p##kind##32(length); \
+    case 64: return world.type_p##kind##64(length); \
 }
 
 #define GET_PRIMTYPE_WITH_KIND_F(kind) \
 switch (bitwidth) { \
     case 8: world.ELOG("8-bit floats do not exist"); \
-    case 16: return world.type_p##kind##16(); \
-    case 32: return world.type_p##kind##32(); \
-    case 64: return world.type_p##kind##64(); \
+    case 16: return world.type_p##kind##16(length); \
+    case 32: return world.type_p##kind##32(length); \
+    case 64: return world.type_p##kind##64(length); \
 }
 
     switch (kind) {

From a0a2b004722a3232521d88f7eea4a925b4525b33 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 13:11:30 +0200
Subject: [PATCH 087/342] small cleanup

---
 src/thorin/be/spirv/spirv.cpp | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 032be25db..23149d150 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -661,7 +661,6 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
         if (variant_datatype->elements_types.size() > 1) {
-            auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
             auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
             auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
             auto converted_payload_type = convert(variant_type->op(variant->index()));
@@ -687,7 +686,6 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto target_type = convert(def->type());
 
         assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
-        auto ptr_type = convert(world().ptr_type(world().type_pu32(), 1, 4, AddrSpace::Function))->type_id;
         auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
         auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
         auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
@@ -909,17 +907,16 @@ std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const
         if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
             auto arr = string->as<DefiniteArray>();
             std::vector<char> the_string;
-            for (int i = 0; i < arr_type->dim(); i++)
+            for (size_t i = 0; i < arr_type->dim(); i++)
                 the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
             the_string.push_back('\0');
             args.push_back(builder_->debug_string(the_string.data()));
         } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
 
-        for (int i = 2; i < source_cont->num_args() - 1; i++) {
+        for (size_t i = 2; i < source_cont->num_args() - 1; i++) {
             args.push_back(emit(source_cont->arg(i), bb));
         }
 
-        auto values = source_cont->arg(2);
         bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
     } else if (builtin->name() == "get_local_id") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);

From c34cdb913c86711831d67cfdca0bf18b4d2b63b7 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 13:19:24 +0200
Subject: [PATCH 088/342] implement ptr serialization

---
 src/thorin/be/spirv/spirv_datatypes.cpp | 25 ++++++++++++++++++++-----
 1 file changed, 20 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index aaa255844..5260b0dae 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -50,7 +50,22 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
 }
 
 void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
-    assert(false && "TODO");
+    assert(type->src_type->as<PtrType>()->addr_space() == AddrSpace::Global && "Only buffer device address (global memory) pointers supported");
+    serialization_types;
+    SpvId u64_tid = type->code_gen->convert(type->code_gen->world().type_pu64())->type_id;
+
+    auto u64_ptr = bb.convert(spv::OpConvertPtrToU, u64_tid, data);
+
+    auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
+    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.constant(u32_tid, { (uint32_t) 1 })) });
+
+    SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
+
+    auto lower = bb.convert(spv::OpUConvert, u64_tid, u64_ptr);
+    auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.binop(spv::OpShiftRightLogical, u64_tid, u64_ptr, c32));
+
+    bb.store(lower, cell0);
+    bb.store(upper, cell1);
 }
 
 DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
@@ -59,7 +74,7 @@ DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType*
 }
 
 SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
-    serialization_types;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     SpvId offset = base_offset;
@@ -72,7 +87,7 @@ SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::St
     return bb.composite(type->type_id, elements);
 }
 void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
-    serialization_types;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     SpvId offset = base_offset;
     SpvId stride = bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() });
@@ -92,7 +107,7 @@ ProductDatatype::ProductDatatype(ConvertedType* type, const std::vector<Converte
 
 SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
     assert(total_size > 0 && "It doesn't make sense to de-serialize Unit!");
-    serialization_types;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     SpvId offset = base_offset;
@@ -105,7 +120,7 @@ SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageC
 }
 void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
     assert(total_size > 0 && "It doesn't make sense to serialize Unit!");
-    serialization_types;
+    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     SpvId offset = base_offset;
     int i = 0;

From f34532a845d40f2fa79ae5cd087935862e2c733f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 13:42:52 +0200
Subject: [PATCH 089/342] add helper u32_t and u32_constant methods

---
 src/thorin/be/spirv/spirv.cpp           | 46 +++++++++++++++----------
 src/thorin/be/spirv/spirv.h             | 29 +++++++++++++---
 src/thorin/be/spirv/spirv_datatypes.cpp | 22 +++++-------
 3 files changed, 61 insertions(+), 36 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 23149d150..9fda1080c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -80,9 +80,12 @@ switch (bitwidth) { \
 #undef GET_PRIMTYPE_WITH_KIND_F
 }
 
-CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
-    : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
-{}
+BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
+        : builder::SpvBasicBlockBuilder(fn_builder.file_builder), fn_builder(fn_builder), file_builder(fn_builder.file_builder) {
+    label = file_builder.generate_fresh_id();
+}
+
+FnBuilder::FnBuilder(CodeGen* cg, FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), cg(cg), file_builder(file_builder) {}
 
 FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
     capability(spv::Capability::CapabilityShader);
@@ -95,6 +98,16 @@ FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
     memory_model = spv::MemoryModel::MemoryModelGLSL450;
 }
 
+SpvId FileBuilder::u32_t() {
+    if (u32_t_.id == 0)
+        u32_t_ = cg->convert(cg->world().type_pu32())->type_id;
+    return u32_t_;
+}
+
+SpvId FileBuilder::u32_constant(uint32_t pattern) {
+    return constant(u32_t(), { pattern });
+}
+
 Builtins::Builtins(FileBuilder& builder) {
     auto& world = builder.cg->world();
     auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
@@ -132,6 +145,10 @@ ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
     shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
 }
 
+CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
+        : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
+{}
+
 void CodeGen::emit_stream(std::ostream& out) {
     builder_ = std::make_unique<FileBuilder>(this);
 
@@ -163,7 +180,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
             SpvId callee = defs_[cont];
 
-            FnBuilder fn_builder(this, builder_.get());
+            FnBuilder fn_builder(this, *builder_.get());
             fn_builder.fn_type = entry_pt_signature;
             fn_builder.fn_ret_type = builder_->void_type;
 
@@ -172,7 +189,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
             // iterate on cont type and extract the arguments
             auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push))->type_id;
-            auto zero = bb->file_builder.constant(convert(world().type_pu32())->type_id, { 0 });
+            auto zero = bb->file_builder.u32_constant(0);
             auto arr_ref = bb->access_chain(ptr_type, push_constant_struct_ptr, { zero });
             uint32_t offset = 0;
             std::vector<SpvId> args;
@@ -183,7 +200,7 @@ void CodeGen::emit_stream(std::ostream& out) {
                 assert(param_type->order() == 0);
                 auto converted = convert(param_type);
                 assert(converted->datatype != nullptr);
-                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.constant(convert(world().type_pu32())->type_id, { offset }));
+                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.u32_constant(offset));
                 args.push_back(arg);
                 offset += converted->datatype->serialized_size();
             }
@@ -214,7 +231,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
     entry_ = scope.entry();
     assert(entry_->is_returning());
 
-    FnBuilder fn(this, builder_.get());
+    FnBuilder fn(this, *builder_.get());
     fn.scope = &scope;
     fn.fn_type = convert(entry_->type())->type_id;
     fn.fn_ret_type = get_codom_type(entry_);
@@ -659,23 +676,20 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     } else if (auto variant = def->isa<Variant>()) {
         auto variant_type = def->type()->as<VariantType>();
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
+        auto tag = builder_->u32_constant(variant->index());
 
         if (variant_datatype->elements_types.size() > 1) {
             auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
             auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
             auto converted_payload_type = convert(variant_type->op(variant->index()));
 
-            auto zero = bb->file_builder.constant(convert(world().type_pu32())->type_id, { 0 });
-
-            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, zero, emit(variant->value(), bb));
+            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0), emit(variant->value(), bb));
             auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
 
-            auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index())});
             std::vector<SpvId> with_tag = {tag, payload};
             return bb->composite(convert(variant->type())->type_id, with_tag);
         } else {
             // Zero-sized payload case
-            auto tag = builder_->constant(convert(world().type_pu32())->type_id, {static_cast<uint32_t>(variant->index())});
             std::vector<SpvId> with_tag = { tag };
             return bb->composite(convert(variant->type())->type_id, with_tag);
         }
@@ -691,8 +705,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
-        auto zero = bb->file_builder.constant(convert(world().type_pu32())->type_id, { 0 });
-        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, zero);
+        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
         return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
@@ -928,9 +941,4 @@ std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const
     return productions;
 }
 
-BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
-: builder::SpvBasicBlockBuilder(*fn_builder.file_builder), fn_builder(fn_builder) {
-    label = file_builder.generate_fresh_id();
-}
-
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 91f5b6533..938c41193 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -16,32 +16,39 @@ struct FileBuilder;
 struct FnBuilder;
 
 struct ConvertedType {
+    ConvertedType(CodeGen* cg) : code_gen(cg) {}
+    ConvertedType(const ConvertedType&) = delete;
+
     spirv::CodeGen* code_gen;
     const thorin::Type* src_type;
     SpvId type_id { 0 };
     std::unique_ptr<Datatype> datatype;
 
-    ConvertedType(CodeGen* cg) : code_gen(cg) {}
     bool is_known_size() { return datatype != nullptr; }
 };
 
 struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     explicit BasicBlockBuilder(FnBuilder& fn_builder);
+    BasicBlockBuilder(const BasicBlockBuilder&) = delete;
 
     FnBuilder& fn_builder;
+    FileBuilder& file_builder;
     std::unordered_map<const Param*, Phi> phis_map;
     DefMap<SpvId> args;
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {
+    explicit FnBuilder(CodeGen* cg, FileBuilder& file_builder);
+    FnBuilder(const FnBuilder&) = delete;
+
     CodeGen* cg;
+    FileBuilder& file_builder;
+
     const Scope* scope = nullptr;
     std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
     std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
     ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
-
-    explicit FnBuilder(CodeGen* cg, builder::SpvFileBuilder* file_builder) : builder::SpvFnBuilder(file_builder), cg(cg) {}
 };
 
 struct Builtins {
@@ -62,12 +69,26 @@ struct ImportedInstructions {
 };
 
 struct FileBuilder : public builder::SpvFileBuilder {
+    explicit FileBuilder(CodeGen* cg);
+    FileBuilder(const FileBuilder&) = delete;
+
     CodeGen* cg;
 
     std::unique_ptr<Builtins> builtins;
     std::unique_ptr<ImportedInstructions> imported_instrs;
 
-    explicit FileBuilder(CodeGen* cg);
+    SpvId u32_t();
+    SpvId u32_constant(uint32_t);
+
+private:
+    SpvId u32_t_ { 0 };
+    /*SpvId i32_t;
+    SpvId u32_t;
+    SpvId i64_t;
+    SpvId u64_t;
+    SpvId i32_constant(int32_t);
+    SpvId i64_constant(int64_t);
+    SpvId u64_constant(uint64_t);*/
 };
 
 class CodeGen : public thorin::CodeGen {
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index 5260b0dae..af4479d9d 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -38,13 +38,12 @@ SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass
     SpvId u64_tid = type->code_gen->convert(type->code_gen->world().type_pu64())->type_id;
 
     auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.constant(u32_tid, { (uint32_t) 1 })) });
+    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.u32_constant(1)) });
 
     auto lower = bb.convert(spv::OpUConvert, u64_tid, bb.load(u32_tid, cell0));
     auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.load(u32_tid, cell1));
 
-    SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
-    auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, c32));
+    auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, bb.file_builder.u32_constant(32)));
 
     return bb.convert(spv::OpConvertUToPtr, type->type_id, merged);
 }
@@ -57,12 +56,10 @@ void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass st
     auto u64_ptr = bb.convert(spv::OpConvertPtrToU, u64_tid, data);
 
     auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.constant(u32_tid, { (uint32_t) 1 })) });
-
-    SpvId c32 = bb.file_builder.constant(u32_tid, { 32 });
+    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.u32_constant(1)) });
 
     auto lower = bb.convert(spv::OpUConvert, u64_tid, u64_ptr);
-    auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.binop(spv::OpShiftRightLogical, u64_tid, u64_ptr, c32));
+    auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.binop(spv::OpShiftRightLogical, u64_tid, u64_ptr, bb.file_builder.u32_constant(32)));
 
     bb.store(lower, cell0);
     bb.store(upper, cell1);
@@ -78,7 +75,7 @@ SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::St
     std::vector<SpvId> indices;
     std::vector<SpvId> elements;
     SpvId offset = base_offset;
-    SpvId stride = bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() });
+    SpvId stride = bb.file_builder.u32_constant(element_type->datatype->serialized_size());
     for (size_t i = 0; i < length; i++) {
         SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, array, offset);
         elements.push_back(element);
@@ -90,7 +87,7 @@ void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, spv::Stora
     SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
     std::vector<SpvId> indices;
     SpvId offset = base_offset;
-    SpvId stride = bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() });
+    SpvId stride = bb.file_builder.u32_constant(element_type->datatype->serialized_size());
     for (size_t i = 0; i < length; i++) {
         element_type->datatype->emit_serialization(bb, storage_class, array, offset, bb.extract(element_type->type_id, data, { (uint32_t) i }));
         offset = bb.binop(spv::OpIAdd, u32_tid, offset, stride);
@@ -113,7 +110,7 @@ SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageC
     SpvId offset = base_offset;
     for (auto& element_type : elements_types) {
         SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, array, offset);
-        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() }));
+        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.u32_constant(element_type->datatype->serialized_size()));
         elements.push_back(element);
     }
     return bb.composite(type->type_id, elements);
@@ -126,7 +123,7 @@ void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClas
     int i = 0;
     for (auto& element_type : elements_types) {
         element_type->datatype->emit_serialization(bb, storage_class, array, offset, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
-        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.constant(u32_tid, { (uint32_t) element_type->datatype->serialized_size() }));
+        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.u32_constant(element_type->datatype->serialized_size()));
     }
 }
 
@@ -239,8 +236,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
         case Node_DefiniteArrayType: {
             auto array = type->as<DefiniteArrayType>();
             ConvertedType* element = convert(array->elem_type());
-            SpvId size = builder_->constant(convert(world().type_pu32())->type_id, {(uint32_t) array->dim() });
-            converted->type_id = builder_->declare_array_type(element->type_id, size);
+            converted->type_id = builder_->declare_array_type(element->type_id, builder_->u32_constant(array->dim()));
             converted->datatype = std::make_unique<DefiniteArrayDatatype>(converted, element, array->dim());
             break;
         }

From d319a06aa367c75dc965db625a931ee05ff88a34 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 13:53:01 +0200
Subject: [PATCH 090/342] minor cleanup

---
 src/thorin/be/spirv/spirv.cpp           | 17 ++++++++---------
 src/thorin/be/spirv/spirv_datatypes.cpp |  5 +----
 2 files changed, 9 insertions(+), 13 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 9fda1080c..612c99744 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -394,7 +394,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         THORIN_UNREACHABLE;
     } else if (continuation->callee()->isa<Bottom>()) {
         bb->unreachable();
-    }  else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
+    } else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
         auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
         auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
         bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
@@ -409,21 +409,22 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         int targets = continuation->num_ops();
         assert(targets > 0);
 
+        // TODO handle dispatching to multiple targets
         assert(targets == 1);
-        auto callee = continuation->op(0)->as_continuation();
+        auto dispatch_target = continuation->op(0)->as_continuation();
         // Extract the relevant variant & expand the tuple if necessary
         auto arg = world().variant_extract(continuation->param(0), 0);
         auto extracted = emit(arg, dispatch_bb);
 
-        if (callee->param(0)->type()->equal(arg->type())) {
-            auto* param = callee->param(0);
-            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
+        if (dispatch_target->param(0)->type()->equal(arg->type())) {
+            auto* param = dispatch_target->param(0);
+            auto& phi = current_fn_->bbs_map[dispatch_target]->phis_map[param];
             phi.preds.emplace_back(extracted, dispatch_bb->label);
         } else {
             assert(false && "TODO destructure argument");
         }
 
-        dispatch_bb->branch(current_fn_->bbs_map[callee]->label);
+        dispatch_bb->branch(current_fn_->bbs_map[dispatch_target]->label);
 
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
         auto loop_header = continuation->op(0)->as_continuation();
@@ -437,11 +438,11 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         bb->branch(header_label);
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
-        // auto header_cont = continuation->op(0)->as_continuation();
 
         int targets = continuation->num_ops();
         assert(targets > 0);
 
+        // TODO handle dispatching to multiple targets
         assert(targets == 1);
         auto callee = continuation->op(0)->as_continuation();
         // TODO phis
@@ -456,8 +457,6 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         jump_to_next_cont_with_args(succ, productions);
     } else if (auto intrinsic = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
         THORIN_UNREACHABLE;
-        //auto ret_continuation = emit_intrinsic(irbuilder, continuation);
-        //irbuilder.CreateBr(cont2bb(ret_continuation));
     } else { // function/closure call
         // put all first-order args into an array
         std::vector<SpvId> call_args;
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index af4479d9d..f807415ae 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -4,10 +4,7 @@
 namespace thorin::spirv {
 
 ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes)
-: Datatype(type), type_tag(type_tag), size_in_bytes(size_in_bytes), alignment(alignment_in_bytes)
-{
-
-}
+: Datatype(type), type_tag(type_tag), size_in_bytes(size_in_bytes), alignment(alignment_in_bytes) {}
 
 /// All serialization/deserialization methods use this so into a macro it goes
 #define serialization_types \

From f09de92de067c88fe8cd4b4c9a28407e5658abf5 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 6 May 2021 14:14:15 +0200
Subject: [PATCH 091/342] implement most invocation id intrinsics

---
 src/thorin/be/spirv/spirv.cpp         | 58 ++++++++++++++++++++++-----
 src/thorin/be/spirv/spirv_builder.hpp | 18 ++++++++-
 2 files changed, 65 insertions(+), 11 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 612c99744..928d33812 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -113,11 +113,13 @@ Builtins::Builtins(FileBuilder& builder) {
     auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
     auto spv_uint_t = builder.cg->convert(world.type_pu32());
     auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t->type_id);
+    auto spv_uvec3_ptp = builder.declare_ptr_type(spv::StorageClassPrivate, spv_uvec3_t->type_id);
     auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t->type_id);
 
-    // workgroup_size = builder.constant(spv_uvec3_pt, spv::StorageClassInput);
-    // builder.decorate(workgroup_size, spv::DecorationBuiltIn, { spv::BuiltInWorkgroupSize });
-    // builder.name(workgroup_size, "BuiltInWorkgroupSize");
+    // Because we technically can have multiple entry points, we take the easy way out and make each entry point
+    // write to a private variable the actual workgroup size for that specific kernel. Dirty, but simple.
+    workgroup_size = builder.variable(spv_uvec3_ptp, spv::StorageClassPrivate);
+    builder.name(workgroup_size, "BuiltInWorkgroupSize");
 
     num_workgroups = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
     builder.decorate(num_workgroups, spv::DecorationBuiltIn, { spv::BuiltInNumWorkgroups });
@@ -187,6 +189,21 @@ void CodeGen::emit_stream(std::ostream& out) {
             BasicBlockBuilder* bb = fn_builder.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn_builder)).get();
             fn_builder.bbs_to_emit.push_back(bb);
 
+            auto block = config->second->as<GPUKernelConfig>()->block_size();
+            std::vector<uint32_t> local_size = {
+                    (uint32_t) std::get<0>(block),
+                    (uint32_t) std::get<1>(block),
+                    (uint32_t) std::get<2>(block),
+            };
+
+            auto spv_uvec3_t = convert(world().type_pu32(3));
+            SpvId wg_size_constant = builder_->constant_composite(spv_uvec3_t->type_id, {
+                builder_->u32_constant(local_size[0]),
+                builder_->u32_constant(local_size[1]),
+                builder_->u32_constant(local_size[2]),
+            });
+            bb->store(wg_size_constant, builder_->builtins->workgroup_size);
+
             // iterate on cont type and extract the arguments
             auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push))->type_id;
             auto zero = bb->file_builder.u32_constant(0);
@@ -211,14 +228,17 @@ void CodeGen::emit_stream(std::ostream& out) {
             builder_->define_function(fn_builder);
             builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
 
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", { push_constant_struct_ptr, builder_->builtins->local_id });
-
-            auto block = config->second->as<GPUKernelConfig>()->block_size();
-            std::vector<uint32_t> local_size = {
-                (uint32_t) std::get<0>(block),
-                (uint32_t) std::get<1>(block),
-                (uint32_t) std::get<2>(block),
+            std::vector<SpvId> interface = {
+                push_constant_struct_ptr,
+                builder_->builtins->workgroup_size,
+                builder_->builtins->num_workgroups,
+                builder_->builtins->workgroup_id,
+                builder_->builtins->local_id,
+                builder_->builtins->global_id,
+                builder_->builtins->local_invocation_index,
             };
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", interface);
+
             builder_->execution_mode(fn_builder.function_id, spv::ExecutionModeLocalSize, local_size);
         }
     }
@@ -930,10 +950,28 @@ std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const
         }
 
         bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
+    } else if (builtin->name() == "get_work_dim") {
+        THORIN_UNREACHABLE;
+    } else if (builtin->name() == "get_global_id") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->global_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_local_size") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_size);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else if (builtin->name() == "get_local_id") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
         auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_num_groups") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->num_workgroups);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_group_id") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
     }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 44a2f287f..7c1840cf9 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -318,6 +318,7 @@ struct SpvFileBuilder {
         PTR_TYPE,
         DEF_ARR_TYPE,
         CONSTANT,
+        CONSTANT_COMPOSITE,
     };
 
     /// Prevents duplicate declarations
@@ -477,7 +478,22 @@ struct SpvFileBuilder {
         types_constants.ref_id(type);
         types_constants.ref_id(id);
         for (auto arg : bit_pattern)
-            types_constants.data_.push_back(arg);
+            types_constants.literal_int(arg);
+        unique_decls[key] = id;
+        return id;
+    }
+
+    SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
+        auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
+        key.members.push_back(type.id);
+        for (auto op : ops) key.members.push_back(op.id);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto op : ops)
+            types_constants.ref_id(op);
         unique_decls[key] = id;
         return id;
     }

From cd483a1d08a4beb7b2c0aacb00d039b678729b93 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 11 May 2021 10:45:55 +0200
Subject: [PATCH 092/342] fix incorrect stride

---
 src/thorin/be/spirv/spirv.h             | 4 ++++
 src/thorin/be/spirv/spirv_datatypes.cpp | 2 +-
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 938c41193..7d31df5de 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -129,6 +129,10 @@ struct Datatype {
     ConvertedType* type;
     Datatype(ConvertedType* type) : type(type) {}
 
+    // Datatypes are serialized using a base element, for now it is hardcoded to use 32-bit scalar unsigned integers
+    static constexpr size_t base_element_bitwidth = 32;
+    static constexpr size_t base_element_bytes = base_element_bitwidth / 8;
+
     virtual size_t serialized_size() = 0;
     virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) = 0;
     virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) = 0;
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
index f807415ae..7c11a56c4 100644
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ b/src/thorin/be/spirv/spirv_datatypes.cpp
@@ -220,7 +220,7 @@ ConvertedType* CodeGen::convert(const Type* type) {
 
                 if (ptr->addr_space() == AddrSpace::Global) {
                     assert(element->datatype && "Can only have physical pointers to known-size types");
-                    builder_->decorate(converted->type_id, spv::DecorationArrayStride, {(uint32_t) element->datatype->serialized_size()});
+                    builder_->decorate(converted->type_id, spv::DecorationArrayStride, {(uint32_t) (element->datatype->serialized_size() * Datatype::base_element_bytes)});
                 }
             }
             ptr_done:

From b37998f256cd38b20ec329e109aae6597f2959db Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 5 Oct 2021 15:34:13 +0200
Subject: [PATCH 093/342] enable spirv backend iff headers are found

---
 CMakeLists.txt | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index e59c13a86..9ad593a62 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -10,7 +10,6 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 option(BUILD_SHARED_LIBS "Build shared libraries" ON)
 option(THORIN_PROFILE "profile complexity in thorin::HashTable - only works in Debug build" ON)
-option(SPIRV_ENABLED "Enable spir-v backend" ON)
 
 if(CMAKE_BUILD_TYPE STREQUAL "")
     set(CMAKE_BUILD_TYPE Debug CACHE STRING "Debug or Release" FORCE)
@@ -41,9 +40,10 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
-if (SPIRV_ENABLED)
-    find_package(SPIRV-Headers REQUIRED)
-    message(STATUS "Enabled SPIR-V backend")
+find_package(SPIRV-Headers QUIET CONFIG)
+if (SPIRV-Headers_FOUND)
+    message(STATUS "Found SPIRV-Headers at ${SPIRV-Headers_DIR}")
+    set(THORIN_ENABLE_SPIRV TRUE)
 endif()
 
 message(STATUS "Using Debug flags: ${CMAKE_CXX_FLAGS_DEBUG}")
@@ -65,9 +65,6 @@ endif()
 if(RV_FOUND)
     set(THORIN_ENABLE_RV TRUE)
 endif()
-if(LLVM_FOUND)
-    set(THORIN_ENABLE_SPIRV TRUE)
-endif()
 configure_file(src/thorin/config.h.in ${CMAKE_BINARY_DIR}/include/thorin/config.h @ONLY)
 include_directories(${CMAKE_BINARY_DIR}/include)
 

From 3d19b722dd571c571b938e4876498568c3232716 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Apr 2022 11:42:32 +0200
Subject: [PATCH 094/342] renamed spirv_transform to structurize

---
 src/thorin/CMakeLists.txt                     |   2 +
 src/thorin/be/shady/shady.cpp                 |   0
 src/thorin/be/shady/shady.h                   |   0
 src/thorin/be/spirv/spirv_builder.hpp         | 669 ------------------
 .../structurize.cpp}                          |  27 +-
 src/thorin/transform/structurize.h            |   8 +
 6 files changed, 23 insertions(+), 683 deletions(-)
 create mode 100644 src/thorin/be/shady/shady.cpp
 create mode 100644 src/thorin/be/shady/shady.h
 delete mode 100644 src/thorin/be/spirv/spirv_builder.hpp
 rename src/thorin/{be/spirv/spirv_transform.cpp => transform/structurize.cpp} (96%)
 create mode 100644 src/thorin/transform/structurize.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 5b4704492..5791dd247 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -70,6 +70,8 @@ set(THORIN_SOURCES
     transform/partial_evaluation.h
     transform/split_slots.cpp
     transform/split_slots.h
+    transform/structurize.cpp
+    transform/structurize.h
     util/array.h
     util/cast.h
     util/hash.h
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
new file mode 100644
index 000000000..e69de29bb
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
new file mode 100644
index 000000000..e69de29bb
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
deleted file mode 100644
index 7c1840cf9..000000000
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ /dev/null
@@ -1,669 +0,0 @@
-#include <spirv/unified1/spirv.hpp>
-
-#include <string>
-#include <cassert>
-#include <vector>
-#include <unordered_map>
-#include <ostream>
-#include <memory>
-
-namespace thorin::spirv::builder {
-
-struct SpvId { uint32_t id; };
-
-struct SpvSectionBuilder;
-struct SpvBasicBlockBuilder;
-struct SpvFnBuilder;
-struct SpvFileBuilder;
-
-inline int div_roundup(int a, int b) {
-    if (a % b == 0)
-        return a / b;
-    else
-        return (a / b) + 1;
-}
-
-struct SpvSectionBuilder {
-    std::vector<uint32_t> data_;
-
-private:
-    void output_word(uint32_t word) {
-        data_.push_back(word);
-    }
-public:
-    void op(spv::Op op, int ops_size) {
-        uint32_t lower = op & 0xFFFFu;
-        uint32_t upper = (ops_size << 16) & 0xFFFF0000u;
-        output_word(lower | upper);
-    }
-
-    void ref_id(SpvId id) {
-        assert(id.id != 0);
-        output_word(id.id);
-    }
-
-    void literal_name(std::string_view str) {
-        int i = 0;
-        uint32_t cword = 0;
-        for (char c : str) {
-            cword = cword | (c & 0xFF) << (i * 8);
-            i++;
-            if (i == 4) {
-                output_word(cword);
-                cword = 0;
-                i = 0;
-            }
-        }
-        output_word(cword);
-    }
-
-    void literal_int(uint32_t i) {
-        output_word(i);
-    }
-};
-
-struct SpvBasicBlockBuilder : public SpvSectionBuilder {
-    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
-    : file_builder(file_builder)
-    {}
-
-    SpvFileBuilder& file_builder;
-
-    struct Phi {
-        SpvId type;
-        SpvId value;
-        std::vector<std::pair<SpvId, SpvId>> preds;
-    };
-    std::vector<Phi*> phis;
-    SpvId label;
-
-    SpvId undef(SpvId type) {
-        op(spv::Op::OpUndef, 3);
-        ref_id(type);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        return id;
-    }
-
-    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
-        op(spv::Op::OpCompositeConstruct, 3 + elements.size());
-        ref_id(aggregate_t);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        for (auto e : elements)
-            ref_id(e);
-        return id;
-    }
-
-    SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
-        op(spv::Op::OpCompositeExtract, 4 + indices.size());
-        ref_id(target_type);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(composite);
-        for (auto i : indices)
-            literal_int(i);
-        return id;
-    }
-
-    SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
-        op(spv::Op::OpCompositeInsert, 5 + indices.size());
-        ref_id(target_type);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(object);
-        ref_id(composite);
-        for (auto i : indices)
-            literal_int(i);
-        return id;
-    }
-
-    SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
-        op(spv::Op::OpVectorExtractDynamic, 5);
-        ref_id(target_type);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(vector);
-        ref_id(index);
-        return id;
-    }
-
-    SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
-        op(spv::Op::OpVectorInsertDynamic, 6);
-        ref_id(target_type);
-        auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(vector);
-        ref_id(component);
-        ref_id(index);
-        return id;
-    }
-
-    // Used for almost all conversion operations
-    SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
-        op(op_, 4);
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(value);
-        return id;
-    }
-
-    SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
-        op(spv::Op::OpAccessChain, 4 + indexes.size());
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(element);
-        for (auto index : indexes)
-            ref_id(index);
-        return id;
-    }
-
-    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
-        op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(base);
-        ref_id(element);
-        for (auto index : indexes)
-            ref_id(index);
-        return id;
-    }
-
-    SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
-        op(spv::Op::OpLoad, 4 + operands.size());
-        auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(pointer);
-        for (auto op : operands)
-            literal_int(op);
-        return id;
-    }
-
-    void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
-        op(spv::Op::OpStore, 3 + operands.size());
-        ref_id(pointer);
-        ref_id(value);
-        for (auto op : operands)
-            literal_int(op);
-    }
-
-    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
-        op(op_, 5);
-        auto id = generate_fresh_id();
-        ref_id(result_type);
-        ref_id(id);
-        ref_id(lhs);
-        ref_id(rhs);
-        return id;
-    }
-
-    void branch(SpvId target) {
-        op(spv::Op::OpBranch, 2);
-        ref_id(target);
-    }
-
-    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
-        op(spv::Op::OpBranchConditional, 4);
-        ref_id(condition);
-        ref_id(true_target);
-        ref_id(false_target);
-    }
-
-    void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
-        op(spv::Op::OpSelectionMerge, 3);
-        ref_id(merge_bb);
-        literal_int(selection_control);
-    }
-
-    void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
-        op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
-        ref_id(merge_bb);
-        ref_id(continue_bb);
-        literal_int(loop_control);
-
-        for (auto e : loop_control_ops)
-            literal_int(e);
-    }
-
-    SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
-        op(spv::Op::OpFunctionCall, 4 + arguments.size());
-        auto id = generate_fresh_id();
-        ref_id(return_type);
-        ref_id(id);
-        ref_id(callee);
-
-        for (auto a : arguments)
-            ref_id(a);
-        return id;
-    }
-
-    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
-        op(spv::Op::OpExtInst, 5 + arguments.size());
-        auto id = generate_fresh_id();
-        ref_id(return_type);
-        ref_id(id);
-        ref_id(set);
-        literal_int(instruction);
-        for (auto a : arguments)
-            ref_id(a);
-        return id;
-    }
-
-    void return_void() {
-        op(spv::Op::OpReturn, 1);
-    }
-
-    void return_value(SpvId value) {
-        op(spv::Op::OpReturnValue, 2);
-        ref_id(value);
-    }
-
-    void unreachable() {
-        op(spv::Op::OpUnreachable, 1);
-    }
-
-private:
-    SpvId generate_fresh_id();
-};
-
-struct SpvFnBuilder {
-    explicit SpvFnBuilder(SpvFileBuilder* file_builder)
-    : file_builder(file_builder)
-    {
-        function_id = generate_fresh_id();
-    }
-
-    SpvFileBuilder* file_builder;
-    SpvId function_id;
-
-    SpvId fn_type;
-    SpvId fn_ret_type;
-    std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
-
-    // Contains OpFunctionParams
-    SpvSectionBuilder header;
-
-    SpvSectionBuilder variables;
-
-    SpvId parameter(SpvId param_type) {
-        header.op(spv::Op::OpFunctionParameter, 3);
-        auto id = generate_fresh_id();
-        header.ref_id(param_type);
-        header.ref_id(id);
-        return id;
-    }
-
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        variables.op(spv::Op::OpVariable, 4);
-        variables.ref_id(type);
-        auto id = generate_fresh_id();
-        variables.ref_id(id);
-        variables.literal_int(storage_class);
-        return id;
-    }
-
-private:
-    SpvId generate_fresh_id();
-};
-
-struct SpvFileBuilder {
-
-    enum UniqueDeclTag {
-        NONE,
-        FN_TYPE,
-        PTR_TYPE,
-        DEF_ARR_TYPE,
-        CONSTANT,
-        CONSTANT_COMPOSITE,
-    };
-
-    /// Prevents duplicate declarations
-    struct UniqueDeclKey {
-        UniqueDeclTag tag;
-        std::vector<uint32_t> members;
-
-        bool operator==(const UniqueDeclKey &b) const {
-            return tag == b.tag && members == b.members;
-        }
-    };
-
-    struct UniqueDeclKeyHasher {
-        size_t operator() (const UniqueDeclKey& key) const {
-            size_t acc = 0;
-            for (auto id : key.members)
-                acc ^= std::hash<uint32_t>{}(id);
-            return std::hash<size_t>{}(key.tag) ^ acc;
-        }
-    };
-
-    SpvFileBuilder()
-    : void_type(declare_void_type())
-    {}
-    SpvFileBuilder(const SpvFileBuilder&) = delete;
-
-    SpvId generate_fresh_id() { return { bound++ }; }
-
-    void name(SpvId id, std::string_view str) {
-        assert(id.id < bound);
-        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
-        debug_names.ref_id(id);
-        debug_names.literal_name(str);
-    }
-
-    SpvId declare_bool_type() {
-        types_constants.op(spv::Op::OpTypeBool, 2);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    SpvId declare_int_type(int width, bool signed_) {
-        types_constants.op(spv::Op::OpTypeInt, 4);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(width);
-        types_constants.literal_int(signed_ ? 1 : 0);
-        return id;
-    }
-
-    SpvId declare_float_type(int width) {
-        types_constants.op(spv::Op::OpTypeFloat, 3);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(width);
-        return id;
-    }
-
-    SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
-        auto key = UniqueDeclKey { PTR_TYPE, { element_type.id, (uint32_t) storage_class } };
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpTypePointer, 4);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(storage_class);
-        types_constants.ref_id(element_type);
-        unique_decls[key] = id;
-        return id;
-    }
-
-    SpvId declare_array_type(SpvId element_type, SpvId dim) {
-        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type.id, dim.id } };
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpTypeArray, 4);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(element_type);
-        types_constants.ref_id(dim);
-        unique_decls[key] = id;
-        return id;
-    }
-
-    SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
-        auto key = UniqueDeclKey { FN_TYPE, {} };
-        for (auto d : dom) key.members.push_back(d.id);
-        key.members.push_back(codom.id);
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-
-        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(codom);
-        for (auto arg : dom)
-            types_constants.ref_id(arg);
-        unique_decls[key] = id;
-        return id;
-    }
-
-    SpvId declare_struct_type(std::vector<SpvId> elements) {
-        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        for (auto arg : elements)
-            types_constants.ref_id(arg);
-        return id;
-    }
-
-    SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
-        types_constants.op(spv::Op::OpTypeVector, 4);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(component_type);
-        types_constants.literal_int(dim);
-        return id;
-    }
-
-    void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
-        annotations.op(spv::Op::OpDecorate, 3 + extra.size());
-        annotations.ref_id(target);
-        annotations.literal_int(decoration);
-        for (auto e : extra)
-            annotations.literal_int(e);
-    }
-
-    void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
-        annotations.op(spv::Op::OpMemberDecorate, 4 + extra.size());
-        annotations.ref_id(target);
-        annotations.literal_int(member);
-        annotations.literal_int(decoration);
-        for (auto e : extra)
-            annotations.literal_int(e);
-    }
-
-    SpvId debug_string(std::string string) {
-        debug_string_source.op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
-        auto id = generate_fresh_id();
-        debug_string_source.ref_id(id);
-        debug_string_source.literal_name(string);
-        return id;
-    }
-
-    SpvId bool_constant(SpvId type, bool value) {
-        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
-        auto key = UniqueDeclKey { CONSTANT, bit_pattern };
-        key.members.push_back(type.id);
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        for (auto arg : bit_pattern)
-            types_constants.literal_int(arg);
-        unique_decls[key] = id;
-        return id;
-    }
-
-    SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
-        auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
-        key.members.push_back(type.id);
-        for (auto op : ops) key.members.push_back(op.id);
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
-        auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        for (auto op : ops)
-            types_constants.ref_id(op);
-        unique_decls[key] = id;
-        return id;
-    }
-
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        types_constants.op(spv::Op::OpVariable, 4);
-        types_constants.ref_id(type);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(storage_class);
-        return id;
-    }
-
-    SpvId define_function(SpvFnBuilder& fn_builder) {
-        fn_defs.op(spv::Op::OpFunction, 5);
-        fn_defs.ref_id(fn_builder.fn_ret_type);
-        fn_defs.ref_id(fn_builder.function_id);
-        fn_defs.data_.push_back(spv::FunctionControlMaskNone);
-        fn_defs.ref_id(fn_builder.fn_type);
-
-        // Includes stuff like OpFunctionParameters
-        for (auto w : fn_builder.header.data_)
-            fn_defs.data_.push_back(w);
-
-        bool first = true;
-        for (auto& bb : fn_builder.bbs_to_emit) {
-            fn_defs.op(spv::Op::OpLabel, 2);
-            fn_defs.ref_id(bb->label);
-
-            if (first) {
-                for (auto w : fn_builder.variables.data_)
-                    fn_defs.data_.push_back(w);
-                first = false;
-            }
-
-            for (auto& phi : bb->phis) {
-                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
-                fn_defs.ref_id(phi->type);
-                fn_defs.ref_id(phi->value);
-                assert(!phi->preds.empty());
-                for (auto& [pred_value, pred_label] : phi->preds) {
-                    fn_defs.ref_id(pred_value);
-                    fn_defs.ref_id(pred_label);
-                }
-            }
-
-            for (auto w : bb->data_)
-                fn_defs.data_.push_back(w);
-        }
-
-        fn_defs.op(spv::Op::OpFunctionEnd, 1);
-        return fn_builder.function_id;
-    }
-
-    void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
-        entry_points.op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
-        entry_points.literal_int(execution_model);
-        entry_points.ref_id(entry_point);
-        entry_points.literal_name(name);
-        for (auto i : interface)
-            entry_points.ref_id(i);
-    }
-
-    void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
-        entry_points.op(spv::Op::OpExecutionMode, 3 + payloads.size());
-        entry_points.ref_id(entry_point);
-        entry_points.literal_int(execution_mode);
-        for (auto d : payloads)
-            entry_points.literal_int(d);
-    }
-
-    void capability(spv::Capability cap) {
-        capabilities.op(spv::Op::OpCapability, 2);
-        capabilities.data_.push_back(cap);
-    }
-
-    void extension(std::string name) {
-        extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
-        extensions.literal_name(name);
-    }
-
-    SpvId extended_import(std::string name) {
-        ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
-        auto id = generate_fresh_id();
-        ext_inst_import.ref_id(id);
-        ext_inst_import.literal_name(name);
-        return id;
-    }
-
-    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
-    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
-
-private:
-    std::ostream* output_ = nullptr;
-    uint32_t bound = 1;
-
-    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
-    SpvSectionBuilder capabilities;
-    SpvSectionBuilder extensions;
-    SpvSectionBuilder ext_inst_import;
-    SpvSectionBuilder entry_points;
-    SpvSectionBuilder execution_modes;
-    SpvSectionBuilder debug_string_source;
-    SpvSectionBuilder debug_names;
-    SpvSectionBuilder debug_module_processed;
-    SpvSectionBuilder annotations;
-    SpvSectionBuilder types_constants;
-    SpvSectionBuilder fn_decls;
-    SpvSectionBuilder fn_defs;
-
-    // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
-    std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
-
-    SpvId declare_void_type() {
-        types_constants.op(spv::Op::OpTypeVoid, 2);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    void output_word_le(uint32_t word) {
-        output_->put((word >> 0) & 0xFFu);
-        output_->put((word >> 8) & 0xFFu);
-        output_->put((word >> 16) & 0xFFu);
-        output_->put((word >> 24) & 0xFFu);
-    }
-
-    void output_section(SpvSectionBuilder& section) {
-        for (auto& word : section.data_) {
-            output_word_le(word);
-        }
-    }
-public:
-    const SpvId void_type;
-
-    void finish(std::ostream& output) {
-        output_ = &output;
-        SpvSectionBuilder memory_model_section;
-        memory_model_section.op(spv::Op::OpMemoryModel, 3);
-        memory_model_section.data_.push_back(addressing_model);
-        memory_model_section.data_.push_back(memory_model);
-
-        output_word_le(spv::MagicNumber);
-        output_word_le(spv::Version); // TODO: target a specific spirv version
-        output_word_le(uint32_t(0)); // TODO get a magic number ?
-        output_word_le(bound);
-        output_word_le(uint32_t(0)); // instruction schema padding
-
-        output_section(capabilities);
-        output_section(extensions);
-        output_section(ext_inst_import);
-        output_section(memory_model_section);
-        output_section(entry_points);
-        output_section(execution_modes);
-        output_section(debug_string_source);
-        output_section(debug_names);
-        output_section(debug_module_processed);
-        output_section(annotations);
-        output_section(types_constants);
-        output_section(fn_decls);
-        output_section(fn_defs);
-    }
-};
-
-inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
-    return file_builder.generate_fresh_id();
-}
-
-inline SpvId SpvFnBuilder::generate_fresh_id() {
-    return file_builder->generate_fresh_id();
-}
-
-}
\ No newline at end of file
diff --git a/src/thorin/be/spirv/spirv_transform.cpp b/src/thorin/transform/structurize.cpp
similarity index 96%
rename from src/thorin/be/spirv/spirv_transform.cpp
rename to src/thorin/transform/structurize.cpp
index f210a8a7a..ee72a64e9 100644
--- a/src/thorin/be/spirv/spirv_transform.cpp
+++ b/src/thorin/transform/structurize.cpp
@@ -1,10 +1,9 @@
-#include <thorin/analyses/looptree.h>
-#include "thorin/be/spirv/spirv.h"
-#include "thorin/analyses/scope.h"
-#include "thorin/analyses/cfg.h"
+#include "structurize.h"
 
-#include <algorithm>
-#include <thorin/analyses/domtree.h>
+//#include <algorithm>
+
+#include "thorin/analyses/domtree.h"
+#include "thorin/world.h"
 
 namespace thorin::spirv {
 
@@ -421,24 +420,24 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
     }
 }
 
-void CodeGen::structure_loops() {
-    Scope::for_each(world(), [&](Scope& scope) {
+void structure_loops(World& world) {
+    Scope::for_each(world, [&](Scope& scope) {
         ScopeContext context(scope);
 
         const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
         tag_continuations(context, looptree.root(), nullptr);
-        collect_dispatch_targets(world(), context, looptree.root());
+        collect_dispatch_targets(world, context, looptree.root());
 
-        create_headers(world(), context, looptree.root());
-        create_epilogues(world(), context, looptree.root());
+        create_headers(world, context, looptree.root());
+        create_epilogues(world, context, looptree.root());
 
-        rewire_loops(world(), context, looptree.root());
+        rewire_loops(world, context, looptree.root());
         scope.update();
     });
 }
 
-void CodeGen::structure_flow() {
-    Scope::for_each(world(), [&](const Scope& scope) {
+void structure_flow(World& world) {
+    Scope::for_each(world, [&](const Scope& scope) {
         CFA cfa(scope);
         auto& dom_tree = cfa.f_cfg().domtree();
         auto& post_dom_tree = cfa.b_cfg().domtree();
diff --git a/src/thorin/transform/structurize.h b/src/thorin/transform/structurize.h
new file mode 100644
index 000000000..6c6e23c70
--- /dev/null
+++ b/src/thorin/transform/structurize.h
@@ -0,0 +1,8 @@
+#include "thorin/analyses/looptree.h"
+#include "thorin/analyses/scope.h"
+#include "thorin/analyses/cfg.h"
+
+class World;
+
+void structure_loops(World& world);
+void structure_flow(World& world);

From 40f042c3b27fcbb8e5490bf9ec3a579df7cbcc55 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Apr 2022 12:08:01 +0200
Subject: [PATCH 095/342] stub of a shady backend replacing the spirv one

---
 CMakeLists.txt                          |  10 +-
 src/thorin/CMakeLists.txt               |  11 +-
 src/thorin/be/codegen.cpp               |   8 +-
 src/thorin/be/shady/shady.cpp           |  57 ++
 src/thorin/be/shady/shady.h             |  32 +
 src/thorin/be/spirv/spirv.cpp           | 976 ------------------------
 src/thorin/be/spirv/spirv.h             | 186 -----
 src/thorin/be/spirv/spirv_datatypes.cpp | 339 --------
 src/thorin/config.h.in                  |   2 +-
 src/thorin/transform/structurize.cpp    |   2 +-
 src/thorin/transform/structurize.h      |   4 +
 11 files changed, 111 insertions(+), 1516 deletions(-)
 delete mode 100644 src/thorin/be/spirv/spirv.cpp
 delete mode 100644 src/thorin/be/spirv/spirv.h
 delete mode 100644 src/thorin/be/spirv/spirv_datatypes.cpp

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9ad593a62..81074796d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -40,10 +40,12 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
-find_package(SPIRV-Headers QUIET CONFIG)
-if (SPIRV-Headers_FOUND)
-    message(STATUS "Found SPIRV-Headers at ${SPIRV-Headers_DIR}")
-    set(THORIN_ENABLE_SPIRV TRUE)
+find_package(shady REQUIRED CONFIG)
+if (shady_FOUND)
+    message(STATUS "Found shady at ${shady_DIR}")
+    message(STATUS "Found shady headers ${shady_INCLUDE_DIRS}")
+    include_directories(${shady_INCLUDE_DIRS})
+    set(THORIN_ENABLE_SHADY TRUE)
 endif()
 
 message(STATUS "Using Debug flags: ${CMAKE_CXX_FLAGS_DEBUG}")
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 5791dd247..c894813a3 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -105,12 +105,9 @@ if(LLVM_FOUND)
     )
 endif()
 
-if(SPIRV_ENABLED)
+if (shady_FOUND)
     list(APPEND THORIN_SOURCES
-        be/spirv/spirv.cpp
-        be/spirv/spirv.h
-        be/spirv/spirv_transform.cpp
-        be/spirv/spirv_datatypes.cpp
+        be/shady/shady.cpp
     )
 endif()
 
@@ -124,3 +121,7 @@ if(LLVM_FOUND)
     endif()
     llvm_config(thorin ${AnyDSL_LLVM_LINK_SHARED} ${Thorin_LLVM_COMPONENTS})
 endif()
+
+if (shady_FOUND)
+    target_link_libraries(thorin PRIVATE shady)
+endif()
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 6a4c08161..f15de6d9c 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -6,8 +6,8 @@
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu.h"
 #endif
-#if THORIN_ENABLE_SPIRV
-#include "thorin/be/spirv/spirv.h"
+#if THORIN_ENABLE_SHADY
+#include "thorin/be/shady/shady.h"
 #endif
 #include "thorin/be/c/c.h"
 
@@ -182,8 +182,8 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug)
 #else
     (void)opt;
 #endif
-#if THORIN_ENABLE_SPIRV
-    if (!importers_[SpirV].world().empty()) cgs[SpirV] = std::make_unique<spirv::CodeGen>(importers_[SpirV].world(), kernel_config, debug);
+#if THORIN_ENABLE_SHADY
+    if (!importers_[SpirV].world().empty()) cgs[SpirV] = std::make_unique<shady::CodeGen>(importers_[SpirV].world(), kernel_config, debug);
 #endif
     for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OpenCL }, std::pair { HLS, c::Lang::HLS } })
         if (!importers_[backend].world().empty()) cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug);
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index e69de29bb..9062753de 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -0,0 +1,57 @@
+#include "shady.h"
+
+#include "thorin/analyses/scope.h"
+#include "thorin/transform/structurize.h"
+
+namespace thorin::shady {
+
+CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
+        : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
+{}
+
+void CodeGen::emit_stream(std::ostream& out) {
+    assert(top_level.empty());
+
+    structure_loops(world());
+    structure_flow(world());
+
+    auto config = shady::IrConfig {
+        .check_types = true,
+    };
+    arena = shady::new_arena(config);
+
+    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
+
+    // build root node with the top level stuff that got emitted
+    auto defs = std::vector<shady::Node*>(top_level.size(), nullptr);
+    auto vars = std::vector<shady::Node*>(top_level.size(), nullptr);
+    for (size_t i = 0; i < top_level.size(); i++) {
+        defs[i] = top_level[i].first;
+        vars[i] = top_level[i].second;
+    }
+    auto root = shady::root(arena, (shady::Root) {
+        .variables = shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(vars.data())),
+        .definitions = shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(defs.data())),
+    });
+
+    shady::print_node(root);
+
+    out << "todo";
+
+    shady::destroy_arena(arena);
+    arena = nullptr;
+    top_level.clear();
+}
+
+void CodeGen::emit(const thorin::Scope& scope) {
+    entry_ = scope.entry();
+    assert(entry_->is_returning());
+
+    assert(false && "TODO");
+}
+
+shady::Type* CodeGen::convert(const Type *) {
+    return nullptr;
+}
+
+}
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
index e69de29bb..d1c2d9a70 100644
--- a/src/thorin/be/shady/shady.h
+++ b/src/thorin/be/shady/shady.h
@@ -0,0 +1,32 @@
+#include "thorin/be/codegen.h"
+
+namespace thorin::shady {
+
+#include <shady/ir.h>
+
+class CodeGen : public thorin::CodeGen {
+public:
+    CodeGen(World&, Cont2Config&, bool debug);
+
+    void emit_stream(std::ostream& stream) override;
+    const char* file_ext() const override { return ".shady"; }
+
+    shady::Type* convert(const Type*);
+protected:
+    void emit(const Scope& scope);
+    //void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
+    //shady::Node* emit(const Def* def, BasicBlockBuilder* bb);
+    //std::vector<SpvId> emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
+
+    //SpvId get_codom_type(const Continuation* fn);
+    shady::IrArena* arena = nullptr;
+    std::vector<std::pair<shady::Node*, shady::Node*>> top_level;
+
+    Continuation* entry_ = nullptr;
+    TypeMap<std::unique_ptr<shady::Type*>> types_;
+    DefMap<shady::Node*> defs_;
+    const Cont2Config& kernel_config_;
+
+};
+
+}
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
deleted file mode 100644
index da925e174..000000000
--- a/src/thorin/be/spirv/spirv.cpp
+++ /dev/null
@@ -1,976 +0,0 @@
-#include "thorin/be/spirv/spirv.h"
-
-#include "thorin/analyses/scope.h"
-#include "thorin/analyses/schedule.h"
-#include "thorin/analyses/domtree.h"
-#include "thorin/transform/cleanup_world.h"
-
-#include <iostream>
-
-namespace thorin::spirv {
-
-/// Used as a dummy SSA value for emitting things like mem/unit
-/// Should never make it in the binary files !
-constexpr SpvId spv_none { 0 };
-
-// SPIR-V has 3 "kinds" of primitives, and the user may declare arbitrary bitwidths, the following helps in translation:
-enum class PrimTypeKind {
-    Signed, Unsigned, Float
-};
-inline PrimTypeKind classify_primtype(const PrimType* type) {
-    switch (type->tag()) {
-#define THORIN_QS_TYPE(T, M) THORIN_PS_TYPE(T, M)
-#define THORIN_PS_TYPE(T, M) \
-case PrimType_##T: \
-    return PrimTypeKind::Signed; \
-    break;
-#include "thorin/tables/primtypetable.h"
-#undef THORIN_QS_TYPE
-#undef THORIN_PS_TYPE
-
-#define THORIN_QU_TYPE(T, M) THORIN_PU_TYPE(T, M)
-#define THORIN_PU_TYPE(T, M) \
-case PrimType_##T: \
-    return PrimTypeKind::Unsigned; \
-    break;
-#include "thorin/tables/primtypetable.h"
-#undef THORIN_QU_TYPE
-#undef THORIN_PU_TYPE
-
-#define THORIN_QF_TYPE(T, M) THORIN_PF_TYPE(T, M)
-#define THORIN_PF_TYPE(T, M) \
-case PrimType_##T: \
-    return PrimTypeKind::Float; \
-    break;
-#include "thorin/tables/primtypetable.h"
-#undef THORIN_QF_TYPE
-#undef THORIN_PF_TYPE
-        default: THORIN_UNREACHABLE;
-    }
-}
-inline const PrimType* get_primtype(World& world, PrimTypeKind kind, int bitwidth, int length) {
-#define GET_PRIMTYPE_WITH_KIND(kind) \
-switch (bitwidth) { \
-    case 8:  return world.type_p##kind##8 (length); \
-    case 16: return world.type_p##kind##16(length); \
-    case 32: return world.type_p##kind##32(length); \
-    case 64: return world.type_p##kind##64(length); \
-}
-
-#define GET_PRIMTYPE_WITH_KIND_F(kind) \
-switch (bitwidth) { \
-    case 8: world.ELOG("8-bit floats do not exist"); \
-    case 16: return world.type_p##kind##16(length); \
-    case 32: return world.type_p##kind##32(length); \
-    case 64: return world.type_p##kind##64(length); \
-}
-
-    switch (kind) {
-        case PrimTypeKind::Signed:   GET_PRIMTYPE_WITH_KIND(s);   THORIN_UNREACHABLE;
-        case PrimTypeKind::Unsigned: GET_PRIMTYPE_WITH_KIND(u);   THORIN_UNREACHABLE;
-        case PrimTypeKind::Float:    GET_PRIMTYPE_WITH_KIND_F(f); THORIN_UNREACHABLE;
-        default: THORIN_UNREACHABLE;
-    }
-
-#undef GET_PRIMTYPE_WITH_KIND
-#undef GET_PRIMTYPE_WITH_KIND_F
-}
-
-BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
-        : builder::SpvBasicBlockBuilder(fn_builder.file_builder), fn_builder(fn_builder), file_builder(fn_builder.file_builder) {
-    label = file_builder.generate_fresh_id();
-}
-
-FnBuilder::FnBuilder(CodeGen* cg, FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), cg(cg), file_builder(file_builder) {}
-
-FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
-    capability(spv::Capability::CapabilityShader);
-    capability(spv::Capability::CapabilityVariablePointers);
-    capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
-    // capability(spv::Capability::CapabilityInt16);
-    capability(spv::Capability::CapabilityInt64);
-
-    addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
-    memory_model = spv::MemoryModel::MemoryModelGLSL450;
-}
-
-SpvId FileBuilder::u32_t() {
-    if (u32_t_.id == 0)
-        u32_t_ = cg->convert(cg->world().type_pu32())->type_id;
-    return u32_t_;
-}
-
-SpvId FileBuilder::u32_constant(uint32_t pattern) {
-    return constant(u32_t(), { pattern });
-}
-
-Builtins::Builtins(FileBuilder& builder) {
-    auto& world = builder.cg->world();
-    auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
-    auto spv_uint_t = builder.cg->convert(world.type_pu32());
-    auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t->type_id);
-    auto spv_uvec3_ptp = builder.declare_ptr_type(spv::StorageClassPrivate, spv_uvec3_t->type_id);
-    auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t->type_id);
-
-    // Because we technically can have multiple entry points, we take the easy way out and make each entry point
-    // write to a private variable the actual workgroup size for that specific kernel. Dirty, but simple.
-    workgroup_size = builder.variable(spv_uvec3_ptp, spv::StorageClassPrivate);
-    builder.name(workgroup_size, "BuiltInWorkgroupSize");
-
-    num_workgroups = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(num_workgroups, spv::DecorationBuiltIn, { spv::BuiltInNumWorkgroups });
-    builder.name(num_workgroups, "BuiltInNumWorkgroups");
-
-    workgroup_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(workgroup_id, spv::DecorationBuiltIn, { spv::BuiltInWorkgroupId });
-    builder.name(workgroup_id, "BuiltInWorkgroupId");
-
-    local_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(local_id, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationId });
-    builder.name(local_id, "BuiltInLocalInvocationId");
-
-    global_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(global_id, spv::DecorationBuiltIn, { spv::BuiltInGlobalInvocationId });
-    builder.name(global_id, "BuiltInGlobalInvocationId");
-
-    local_invocation_index = builder.variable(spv_uint_pt, spv::StorageClassInput);
-    builder.decorate(local_invocation_index, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationIndex });
-    builder.name(local_invocation_index, "BuiltInLocalInvocationIndex");
-}
-
-ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
-    builder.extension("SPV_KHR_non_semantic_info");
-    shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
-}
-
-CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
-        : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
-{}
-
-void CodeGen::emit_stream(std::ostream& out) {
-    builder_ = std::make_unique<FileBuilder>(this);
-
-    builder_->builtins = std::make_unique<Builtins>(*builder_);
-    builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
-
-    structure_loops();
-    structure_flow();
-    // cleanup_world(world());
-
-    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
-
-    auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
-    auto push_constant_struct_type = builder_->declare_struct_type({ push_constant_arr_type });
-    auto push_constant_struct_ptr_type = builder_->declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
-    builder_->name(push_constant_struct_type, "ThorinPushConstant");
-    builder_->decorate(push_constant_struct_type, spv::DecorationBlock);
-    builder_->decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
-    builder_->decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
-    auto push_constant_struct_ptr = builder_->variable(push_constant_struct_ptr_type, spv::StorageClassPushConstant);
-    builder_->name(push_constant_struct_ptr, "thorin_push_constant_data");
-
-    auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
-    for (auto& cont : world().continuations()) {
-        if (cont->is_exported()) {
-            assert(defs_.contains(cont) && kernel_config_.contains(cont));
-            auto config = kernel_config_.find(cont);
-
-            SpvId callee = defs_[cont];
-
-            FnBuilder fn_builder(this, *builder_.get());
-            fn_builder.fn_type = entry_pt_signature;
-            fn_builder.fn_ret_type = builder_->void_type;
-
-            BasicBlockBuilder* bb = fn_builder.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn_builder)).get();
-            fn_builder.bbs_to_emit.push_back(bb);
-
-            auto block = config->second->as<GPUKernelConfig>()->block_size();
-            std::vector<uint32_t> local_size = {
-                    (uint32_t) std::get<0>(block),
-                    (uint32_t) std::get<1>(block),
-                    (uint32_t) std::get<2>(block),
-            };
-
-            auto spv_uvec3_t = convert(world().type_pu32(3));
-            SpvId wg_size_constant = builder_->constant_composite(spv_uvec3_t->type_id, {
-                builder_->u32_constant(local_size[0]),
-                builder_->u32_constant(local_size[1]),
-                builder_->u32_constant(local_size[2]),
-            });
-            bb->store(wg_size_constant, builder_->builtins->workgroup_size);
-
-            // iterate on cont type and extract the arguments
-            auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push))->type_id;
-            auto zero = bb->file_builder.u32_constant(0);
-            auto arr_ref = bb->access_chain(ptr_type, push_constant_struct_ptr, { zero });
-            uint32_t offset = 0;
-            std::vector<SpvId> args;
-            for (size_t i = 0; i < cont->num_params(); i++) {
-                auto param = cont->param(i);
-                auto param_type = param->type();
-                if (param_type == world().unit() || param_type == world().mem_type() || param_type->isa<FnType>()) continue;
-                assert(param_type->order() == 0);
-                auto converted = convert(param_type);
-                assert(converted->datatype != nullptr);
-                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.u32_constant(offset));
-                args.push_back(arg);
-                offset += converted->datatype->serialized_size();
-            }
-
-            bb->call(builder_->void_type, callee, args);
-            bb->return_void();
-
-            builder_->define_function(fn_builder);
-            builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
-
-            std::vector<SpvId> interface = {
-                push_constant_struct_ptr,
-                builder_->builtins->workgroup_size,
-                builder_->builtins->num_workgroups,
-                builder_->builtins->workgroup_id,
-                builder_->builtins->local_id,
-                builder_->builtins->global_id,
-                builder_->builtins->local_invocation_index,
-            };
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", interface);
-
-            builder_->execution_mode(fn_builder.function_id, spv::ExecutionModeLocalSize, local_size);
-        }
-    }
-
-    builder_->finish(out);
-    builder_ = nullptr;
-}
-
-void CodeGen::emit(const thorin::Scope& scope) {
-    entry_ = scope.entry();
-    assert(entry_->is_returning());
-
-    FnBuilder fn(this, *builder_.get());
-    fn.scope = &scope;
-    fn.fn_type = convert(entry_->type())->type_id;
-    fn.fn_ret_type = get_codom_type(entry_);
-    defs_.emplace(scope.entry(), fn.function_id);
-
-    current_fn_ = &fn;
-
-    auto conts = schedule(scope);
-
-    fn.bbs_to_emit.reserve(conts.size());
-    fn.bbs.reserve(conts.size());
-    auto& bbs = fn.bbs;
-
-    for (auto cont : conts) {
-        if (cont->intrinsic() == Intrinsic::EndScope) continue;
-
-        BasicBlockBuilder* bb = bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn)).get();
-        fn.bbs_to_emit.emplace_back(bb);
-        auto [i, b] = fn.bbs_map.emplace(cont, bb);
-        assert(b);
-
-        if (debug())
-            builder_->name(bb->label, cont->name().c_str());
-        fn.labels.emplace(cont, bb->label);
-
-        if (entry_ == cont) {
-            for (auto param : entry_->params()) {
-                if (is_mem(param) || is_unit(param)) {
-                    // Nothing
-                } else if (param->order() == 0) {
-                    auto param_t = convert(param->type());
-                    auto id = fn.parameter(param_t->type_id);
-                    fn.params[param] = id;
-                    if (param->type()->isa<PtrType>()) {
-                        builder_->decorate(id, spv::DecorationAliased);
-                    }
-                }
-            }
-        } else {
-            for (auto param : cont->params()) {
-                if (is_mem(param) || is_unit(param)) {
-                    // Nothing
-                } else {
-                    // OpPhi requires the full list of predecessors (values, labels)
-                    // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
-                    // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    auto type = convert(param->type())->type_id;
-                    assert(type.id != 0);
-                    bb->phis_map[param] = { type, builder_->generate_fresh_id(), {} };
-                }
-            }
-        }
-    }
-
-    for (auto cont : conts) {
-        if (cont->intrinsic() == Intrinsic::EndScope) continue;
-        assert(cont == entry_ || cont->is_basicblock());
-        emit_epilogue(cont, fn.bbs_map[cont]);
-    }
-
-    for(auto& bb : fn.bbs) {
-        for (auto& [param, phi] : bb->phis_map) {
-            bb->phis.emplace_back(&phi);
-        }
-    }
-    
-    builder_->define_function(fn);
-    builder_->name(fn.function_id, scope.entry()->name());
-}
-
-SpvId CodeGen::get_codom_type(const Continuation* fn) {
-    auto ret_cont_type = fn->ret_param()->type();
-    std::vector<SpvId> types;
-    for (auto& op : ret_cont_type->ops()) {
-        if (op->isa<MemType>() || is_type_unit(op))
-            continue;
-        assert(op->order() == 0);
-        types.push_back(convert(op)->type_id);
-    }
-    if (types.empty())
-        return builder_->void_type;
-    if (types.size() == 1)
-        return types[0];
-    return builder_->declare_struct_type(types);
-}
-
-void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
-    // Handles the potential nuances of jumping to another continuation
-    auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
-        bb->branch(current_fn_->labels[succ]);
-        for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-            auto param = succ->param(i);
-            if (is_mem(param) || is_unit(param))
-                continue;
-            auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
-            phi.preds.emplace_back(args[j], current_fn_->labels[continuation]);
-            j++;
-        }
-    };
-
-    if (continuation->callee() == entry_->ret_param()) {
-        std::vector<SpvId> values;
-
-        for (auto arg : continuation->args()) {
-            assert(arg->order() == 0);
-            auto val = emit(arg, bb);
-            if (is_mem(arg) || is_unit(arg))
-                continue;
-            values.emplace_back(val);
-        }
-
-        switch (values.size()) {
-            case 0:  bb->return_void();      break;
-            case 1:  bb->return_value(values[0]); break;
-            default: bb->return_value(bb->composite(current_fn_->fn_ret_type, values));
-        }
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
-        int index = -1;
-        for (auto& arg : continuation->args()) {
-            index++;
-            auto val = emit(arg, bb);
-            if (is_mem(arg) || is_unit(arg)) continue;
-            bb->args[arg] = val;
-            auto* param = callee->param(index);
-            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
-            phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
-        }
-        bb->branch(current_fn_->labels[callee]);
-    } else if (continuation->callee() == world().branch()) {
-        auto& domtree = current_fn_->scope->b_cfg().domtree();
-        auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
-        SpvId merge_bb;
-        if (merge_cont == current_fn_->scope->exit()) {
-            BasicBlockBuilder* unreachable_merge_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
-            current_fn_->bbs_to_emit.emplace_back(unreachable_merge_bb);
-            builder_->name(unreachable_merge_bb->label, "merge_unreachable" + continuation->name());
-            unreachable_merge_bb->unreachable();
-            merge_bb = unreachable_merge_bb->label;
-        } else {
-            // TODO create a dedicated merge bb if this one is the merge blocks for more than 1 selection construct
-            merge_bb = current_fn_->labels[merge_cont];
-        }
-
-        auto cond = emit(continuation->arg(0), bb);
-        bb->args.emplace(continuation->arg(0), cond);
-        auto tbb = current_fn_->labels[continuation->arg(1)->as_continuation()];
-        auto fbb = current_fn_->labels[continuation->arg(2)->as_continuation()];
-        bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
-        bb->branch_conditional(cond, tbb, fbb);
-    } else if (continuation->callee()->isa<Continuation>() && continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
-        /*auto val = emit(continuation->arg(0));
-        auto otherwise_bb = cont2bb(continuation->arg(1)->as_continuation());
-        auto match = irbuilder.CreateSwitch(val, otherwise_bb, continuation->num_args() - 2);
-        for (size_t i = 2; i < continuation->num_args(); i++) {
-            auto arg = continuation->arg(i)->as<Tuple>();
-            auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
-            auto case_bb    = cont2bb(arg->op(1)->as_continuation());
-            match->addCase(case_const, case_bb);
-        }*/
-        THORIN_UNREACHABLE;
-    } else if (continuation->callee()->isa<Bottom>()) {
-        bb->unreachable();
-    } else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
-        auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
-        auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
-        bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
-
-        BasicBlockBuilder* dispatch_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
-
-        auto header_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), bb);
-
-        current_fn_->bbs_to_emit.emplace(header_bb_location + 1, dispatch_bb);
-        builder_->name(dispatch_bb->label, "dispatch_" + continuation->name());
-        bb->branch(dispatch_bb->label);
-        int targets = continuation->num_ops();
-        assert(targets > 0);
-
-        // TODO handle dispatching to multiple targets
-        assert(targets == 1);
-        auto dispatch_target = continuation->op(0)->as_continuation();
-        // Extract the relevant variant & expand the tuple if necessary
-        auto arg = world().variant_extract(continuation->param(0), 0);
-        auto extracted = emit(arg, dispatch_bb);
-
-        if (dispatch_target->param(0)->type()->equal(arg->type())) {
-            auto* param = dispatch_target->param(0);
-            auto& phi = current_fn_->bbs_map[dispatch_target]->phis_map[param];
-            phi.preds.emplace_back(extracted, dispatch_bb->label);
-        } else {
-            assert(false && "TODO destructure argument");
-        }
-
-        dispatch_bb->branch(current_fn_->bbs_map[dispatch_target]->label);
-
-    } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
-        auto loop_header = continuation->op(0)->as_continuation();
-        auto header_label = current_fn_->bbs_map[loop_header]->label;
-
-        auto arg = continuation->param(0);
-        bb->args[arg] = emit(arg, bb);
-        auto* param = loop_header->param(0);
-        auto& phi = current_fn_->bbs_map[loop_header]->phis_map[param];
-        phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
-
-        bb->branch(header_label);
-    } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
-
-        int targets = continuation->num_ops();
-        assert(targets > 0);
-
-        // TODO handle dispatching to multiple targets
-        assert(targets == 1);
-        auto callee = continuation->op(0)->as_continuation();
-        // TODO phis
-        bb->branch(current_fn_->bbs_map[callee]->label);
-    } else if (auto builtin = continuation->callee()->isa_continuation(); builtin->is_imported()) {
-        // Ensure we emit previous memory operations
-        assert(is_mem(continuation->arg(0)));
-        emit(continuation->arg(0), bb);
-
-        auto productions = emit_builtin(continuation, builtin, bb);
-        auto succ = continuation->args().back()->as_continuation();
-        jump_to_next_cont_with_args(succ, productions);
-    } else if (auto intrinsic = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
-        THORIN_UNREACHABLE;
-    } else { // function/closure call
-        // put all first-order args into an array
-        std::vector<SpvId> call_args;
-        const Def* ret_arg = nullptr;
-        for (auto arg : continuation->args()) {
-            if (arg->order() == 0) {
-                auto arg_type = arg->type();
-                auto arg_val = emit(arg, bb);
-                if (arg_type == world().unit() || arg_type == world().mem_type()) continue;
-                call_args.push_back(arg_val);
-            } else {
-                assert(!ret_arg);
-                ret_arg = arg;
-            }
-        }
-
-        auto ret_type = get_codom_type(continuation);
-
-        SpvId call_result;
-        if (auto called_continuation = continuation->callee()->isa_continuation()) {
-            call_result = bb->call(ret_type, emit(called_continuation, bb), call_args);
-        } else {
-            // must be a closure
-            THORIN_UNREACHABLE;
-
-            // auto closure = emit(callee);
-            // args.push_back(irbuilder.CreateExtractValue(closure, 1));
-            // call = irbuilder.CreateCall(irbuilder.CreateExtractValue(closure, 0), args);
-        }
-
-        // must be call + continuation --- call + return has been removed by codegen_prepare
-        auto succ = ret_arg->as_continuation();
-
-        size_t n = 0;
-        const Param* last_param = nullptr;
-        for (auto param : succ->params()) {
-            if (is_mem(param) || is_unit(param))
-                continue;
-            last_param = param;
-            n++;
-        }
-
-        if (n == 0) {
-            bb->branch(current_fn_->labels[succ]);
-        } else if (n == 1) {
-            bb->branch(current_fn_->labels[succ]);
-
-            auto& phi = current_fn_->bbs_map[succ]->phis_map[last_param];
-            phi.preds.emplace_back(call_result, current_fn_->labels[continuation]);
-        } else {
-            Array<SpvId> extracts(n);
-            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-                auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
-                    continue;
-                extracts[j] = bb->extract(convert(param->type())->type_id, call_result, { (uint32_t) j });
-                j++;
-            }
-
-            bb->branch(current_fn_->labels[succ]);
-
-            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-                auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
-                    continue;
-
-                auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
-                phi.preds.emplace_back(extracts[j], current_fn_->labels[continuation]);
-
-                j++;
-            }
-        }
-    }
-}
-
-SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
-    if (auto bin = def->isa<BinOp>()) {
-        SpvId lhs = emit(bin->lhs(), bb);
-        SpvId rhs = emit(bin->rhs(), bb);
-        ConvertedType* result_types = convert(def->type());
-        SpvId result_type = result_types->type_id;
-
-        if (auto cmp = bin->isa<Cmp>()) {
-            auto type = cmp->lhs()->type();
-            if (is_type_s(type)) {
-                switch (cmp->cmp_tag()) {
-                    case Cmp_eq: return bb->binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
-                    case Cmp_ne: return bb->binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
-                    case Cmp_gt: return bb->binop(spv::Op::OpSGreaterThan      , result_type, lhs, rhs);
-                    case Cmp_ge: return bb->binop(spv::Op::OpSGreaterThanEqual , result_type, lhs, rhs);
-                    case Cmp_lt: return bb->binop(spv::Op::OpSLessThan         , result_type, lhs, rhs);
-                    case Cmp_le: return bb->binop(spv::Op::OpSLessThanEqual    , result_type, lhs, rhs);
-                }
-            } else if (is_type_u(type)) {
-                switch (cmp->cmp_tag()) {
-                    case Cmp_eq: return bb->binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
-                    case Cmp_ne: return bb->binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
-                    case Cmp_gt: return bb->binop(spv::Op::OpUGreaterThan      , result_type, lhs, rhs);
-                    case Cmp_ge: return bb->binop(spv::Op::OpUGreaterThanEqual , result_type, lhs, rhs);
-                    case Cmp_lt: return bb->binop(spv::Op::OpULessThan         , result_type, lhs, rhs);
-                    case Cmp_le: return bb->binop(spv::Op::OpULessThanEqual    , result_type, lhs, rhs);
-                }
-            } else if (is_type_f(type)) {
-                switch (cmp->cmp_tag()) {
-                    // TODO look into the NaN story
-                    case Cmp_eq: return bb->binop(spv::Op::OpFOrdEqual            , result_type, lhs, rhs);
-                    case Cmp_ne: return bb->binop(spv::Op::OpFOrdNotEqual         , result_type, lhs, rhs);
-                    case Cmp_gt: return bb->binop(spv::Op::OpFOrdGreaterThan      , result_type, lhs, rhs);
-                    case Cmp_ge: return bb->binop(spv::Op::OpFOrdGreaterThanEqual , result_type, lhs, rhs);
-                    case Cmp_lt: return bb->binop(spv::Op::OpFOrdLessThan         , result_type, lhs, rhs);
-                    case Cmp_le: return bb->binop(spv::Op::OpFOrdLessThanEqual    , result_type, lhs, rhs);
-                }
-            } else if (type->isa<PtrType>()) {
-                assertf(false, "Physical pointers are unsupported");
-            } else if(is_type_bool(type)) {
-                switch (cmp->cmp_tag()) {
-                    // TODO look into the NaN story
-                    case Cmp_eq: return bb->binop(spv::Op::OpLogicalEqual    , result_type, lhs, rhs);
-                    case Cmp_ne: return bb->binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
-                    default: THORIN_UNREACHABLE;
-                }
-                assertf(false, "TODO: should we emulate the other comparison ops ?");
-            }
-        }
-
-        if (auto arithop = bin->isa<ArithOp>()) {
-            auto type = arithop->type();
-
-            if (is_type_f(type)) {
-                switch (arithop->arithop_tag()) {
-                    case ArithOp_add: return bb->binop(spv::Op::OpFAdd, result_type, lhs, rhs);
-                    case ArithOp_sub: return bb->binop(spv::Op::OpFSub, result_type, lhs, rhs);
-                    case ArithOp_mul: return bb->binop(spv::Op::OpFMul, result_type, lhs, rhs);
-                    case ArithOp_div: return bb->binop(spv::Op::OpFDiv, result_type, lhs, rhs);
-                    case ArithOp_rem: return bb->binop(spv::Op::OpFRem, result_type, lhs, rhs);
-                    case ArithOp_and:
-                    case ArithOp_or:
-                    case ArithOp_xor:
-                    case ArithOp_shl:
-                    case ArithOp_shr: THORIN_UNREACHABLE;
-                }
-            }
-
-            if (is_type_s(type)) {
-                switch (arithop->arithop_tag()) {
-                    case ArithOp_add: return bb->binop(spv::Op::OpIAdd                 , result_type, lhs, rhs);
-                    case ArithOp_sub: return bb->binop(spv::Op::OpISub                 , result_type, lhs, rhs);
-                    case ArithOp_mul: return bb->binop(spv::Op::OpIMul                 , result_type, lhs, rhs);
-                    case ArithOp_div: return bb->binop(spv::Op::OpSDiv                 , result_type, lhs, rhs);
-                    case ArithOp_rem: return bb->binop(spv::Op::OpSRem                 , result_type, lhs, rhs);
-                    case ArithOp_and: return bb->binop(spv::Op::OpBitwiseAnd           , result_type, lhs, rhs);
-                    case ArithOp_or:  return bb->binop(spv::Op::OpBitwiseOr            , result_type, lhs, rhs);
-                    case ArithOp_xor: return bb->binop(spv::Op::OpBitwiseXor           , result_type, lhs, rhs);
-                    case ArithOp_shl: return bb->binop(spv::Op::OpShiftLeftLogical     , result_type, lhs, rhs);
-                    case ArithOp_shr: return bb->binop(spv::Op::OpShiftRightArithmetic , result_type, lhs, rhs);
-                }
-            } else if (is_type_u(type)) {
-                switch (arithop->arithop_tag()) {
-                    case ArithOp_add: return bb->binop(spv::Op::OpIAdd              , result_type, lhs, rhs);
-                    case ArithOp_sub: return bb->binop(spv::Op::OpISub              , result_type, lhs, rhs);
-                    case ArithOp_mul: return bb->binop(spv::Op::OpIMul              , result_type, lhs, rhs);
-                    case ArithOp_div: return bb->binop(spv::Op::OpUDiv              , result_type, lhs, rhs);
-                    case ArithOp_rem: return bb->binop(spv::Op::OpUMod              , result_type, lhs, rhs);
-                    case ArithOp_and: return bb->binop(spv::Op::OpBitwiseAnd        , result_type, lhs, rhs);
-                    case ArithOp_or:  return bb->binop(spv::Op::OpBitwiseOr         , result_type, lhs, rhs);
-                    case ArithOp_xor: return bb->binop(spv::Op::OpBitwiseXor        , result_type, lhs, rhs);
-                    case ArithOp_shl: return bb->binop(spv::Op::OpShiftLeftLogical  , result_type, lhs, rhs);
-                    case ArithOp_shr: return bb->binop(spv::Op::OpShiftRightLogical , result_type, lhs, rhs);
-                }
-            } else if(is_type_bool(type)) {
-                switch (arithop->arithop_tag()) {
-                    case ArithOp_and: return bb->binop(spv::Op::OpLogicalAnd      , result_type, lhs, rhs);
-                    case ArithOp_or:  return bb->binop(spv::Op::OpLogicalOr       , result_type, lhs, rhs);
-                    // Note: there is no OpLogicalXor
-                    case ArithOp_xor: return bb->binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
-                    default: THORIN_UNREACHABLE;
-                }
-            }
-            THORIN_UNREACHABLE;
-        }
-    } else if (auto primlit = def->isa<PrimLit>()) {
-        Box box = primlit->value();
-        auto type = convert(def->type())->type_id;
-        SpvId constant;
-        switch (primlit->primtype_tag()) {
-            case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
-            case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
-            case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
-            case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
-            case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
-            case PrimType_ps32: case PrimType_qs32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
-            case PrimType_pu32: case PrimType_qu32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
-            case PrimType_ps64: case PrimType_qs64:
-            case PrimType_pu64: case PrimType_qu64: {
-                uint64_t value = static_cast<uint64_t>(box.get_u64());
-                uint64_t upper = value >> 32U;
-                uint64_t lower = value & 0xFFFFFFFFU;
-                constant = bb->file_builder.constant(type, { (uint32_t) lower, (uint32_t) upper });
-                break;
-            }
-            case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
-            case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
-            case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
-        }
-        return constant;
-    } else if (auto param = def->isa<Param>()) {
-        if (is_mem(param)) return spv_none;
-        if (auto param_id = current_fn_->params.lookup(param)) {
-            assert((*param_id).id != 0);
-            return *param_id;
-        } else {
-            auto val = (*current_fn_->bbs_map[param->continuation()]).phis_map[param].value;
-            assert(val.id != 0);
-            return val;
-        }
-    } else if (auto variant = def->isa<Variant>()) {
-        auto variant_type = def->type()->as<VariantType>();
-        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
-        auto tag = builder_->u32_constant(variant->index());
-
-        if (variant_datatype->elements_types.size() > 1) {
-            auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
-            auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
-            auto converted_payload_type = convert(variant_type->op(variant->index()));
-
-            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0), emit(variant->value(), bb));
-            auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
-
-            std::vector<SpvId> with_tag = {tag, payload};
-            return bb->composite(convert(variant->type())->type_id, with_tag);
-        } else {
-            // Zero-sized payload case
-            std::vector<SpvId> with_tag = { tag };
-            return bb->composite(convert(variant->type())->type_id, with_tag);
-        }
-    } else if (auto vextract = def->isa<VariantExtract>()) {
-        auto variant_type = vextract->value()->type()->as<VariantType>();
-        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
-
-        auto target_type = convert(def->type());
-
-        assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
-        auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
-        auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
-        auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
-        bb->store(payload, payload_arr);
-
-        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));
-    } else if (auto vindex = def->isa<VariantIndex>()) {
-        auto value = emit(vindex->op(0), bb);
-        return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
-    } else if (auto tuple = def->isa<Tuple>()) {
-        std::vector<SpvId> elements;
-        elements.resize(tuple->num_ops());
-        size_t x = 0;
-        for (auto& e : tuple->ops()) {
-            elements[x++] = emit(e, bb);
-        }
-        return bb->composite(convert(tuple->type())->type_id, elements);
-    } else if (auto structagg = def->isa<StructAgg>()) {
-        std::vector<SpvId> elements;
-        elements.resize(structagg->num_ops());
-        size_t x = 0;
-        for (auto& e : structagg->ops()) {
-            elements[x++] = emit(e, bb);
-        }
-        return bb->composite(convert(structagg->type())->type_id, elements);
-    } else if (auto access = def->isa<Access>()) {
-        // emit dependent operations first
-        emit(access->mem(), bb);
-
-        std::vector<uint32_t> operands;
-        auto ptr_type = access->ptr()->type()->as<PtrType>();
-        if (ptr_type->addr_space() == AddrSpace::Global) {
-            operands.push_back(spv::MemoryAccessAlignedMask);
-            operands.push_back( 4 ); // TODO: SPIR-V docs say to consult client API for valid values.
-        }
-        if (auto load = def->isa<Load>()) {
-            return bb->load(convert(load->out_val_type())->type_id, emit(load->ptr(), bb), operands);
-        } else if (auto store = def->isa<Store>()) {
-            bb->store(emit(store->val(), bb), emit(store->ptr(), bb), operands);
-            return spv_none;
-        } else THORIN_UNREACHABLE;
-    } else if (auto lea = def->isa<LEA>()) {
-        switch (lea->ptr_type()->addr_space()) {
-            case AddrSpace::Global:
-            case AddrSpace::Shared:
-                break;
-            default:
-                world().ELOG("LEA is only allowed in global & shared address spaces");
-                break;
-        }
-        auto type = convert(lea->ptr_type());
-        auto offset = emit(lea->index(), bb);
-        return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
-    } else if (auto aggop = def->isa<AggOp>()) {
-        auto spv_agg = emit(aggop->agg(), bb);
-        auto agg_type = convert(aggop->agg()->type())->type_id;
-
-        bool mem = false;
-        if (auto tt = aggop->agg()->type()->isa<TupleType>(); tt && tt->op(0)->isa<MemType>()) mem = true;
-
-        auto copy_to_alloca = [&] (SpvId target_type) {
-            world().wdef(def, "slow: alloca and loads/stores needed for aggregate '{}'", def);
-            auto agg_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, agg_type);
-
-            auto variable = bb->fn_builder.variable(agg_ptr_type, spv::StorageClassFunction);
-            bb->store(spv_agg, variable);
-
-            auto cell_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, target_type);
-            auto cell = bb->access_chain(cell_ptr_type, variable, { emit(aggop->index(), bb)} );
-            return std::make_pair(variable, cell);
-        };
-
-        if (auto extract = aggop->isa<Extract>()) {
-            if (is_mem(extract)) return spv_none;
-
-            auto target_type = convert(extract->type())->type_id;
-            auto constant_index = aggop->index()->isa<PrimLit>();
-
-            // We have a fast-path: if the index is constant, we can simply use OpCompositeExtract
-            if (aggop->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
-                assert(aggop->agg()->type()->isa<DefiniteArrayType>());
-                assert(!is_mem(extract));
-                return bb->load(target_type, copy_to_alloca(target_type).second);
-            }
-
-            if (extract->agg()->type()->isa<VectorType>())
-                return bb->vector_extract_dynamic(target_type, spv_agg, emit(extract->index(), bb));
-
-            // index *must* be constant for the remaining possible cases
-            assert(constant_index != nullptr);
-            uint32_t index = constant_index->value().get_u32();
-
-            unsigned offset = 0;
-            if (mem) {
-                if (aggop->agg()->type()->num_ops() == 2) return spv_agg;
-                offset = 1;
-            }
-
-            return bb->extract(target_type, spv_agg, { index - offset });
-        } else if (auto insert = def->isa<Insert>()) {
-            auto value = emit(insert->value(), bb);
-            auto constant_index = aggop->index()->isa<PrimLit>();
-
-            // TODO deal with mem - but I think for now this case shouldn't happen
-
-            if (insert->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
-                assert(aggop->agg()->type()->isa<DefiniteArrayType>());
-                auto [variable, cell] = copy_to_alloca(agg_type);
-                bb->store(value, cell);
-                return bb->load(agg_type, variable);
-            }
-
-            if (insert->agg()->type()->isa<VectorType>())
-                return bb->vector_insert_dynamic(agg_type, spv_agg, value, emit(insert->index(), bb));
-
-            // index *must* be constant for the remaining possible cases
-            assert(constant_index != nullptr);
-            uint32_t index = constant_index->value().get_u32();
-
-            return bb->insert(agg_type, value, spv_agg, { index });
-        } else THORIN_UNREACHABLE;
-    } else if (auto conv = def->isa<ConvOp>()) {
-        auto src_type = conv->from()->type();
-        auto dst_type = conv->type();
-
-        auto conv_src_type = convert(src_type);
-        auto conv_dst_type = convert(dst_type);
-
-        if (auto bitcast = def->isa<Bitcast>()) {
-            if (conv_src_type->datatype->serialized_size() != conv_dst_type->datatype->serialized_size())
-                world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type->datatype->serialized_size(), conv_dst_type->datatype->serialized_size());
-
-            return bb->convert(spv::OpBitcast, convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
-        } else if (auto cast = def->isa<Cast>()) {
-            // NB: all ops used here are scalar/vector agnostic
-            auto src_prim = src_type->isa<PrimType>();
-            auto dst_prim = dst_type->isa<PrimType>();
-            if (!src_prim || !dst_prim || src_prim->length() != dst_prim->length())
-                world().ELOG("Illegal cast: % to %, casts are only supported between primitives with identical vector length", src_type->to_string(), dst_type->to_string());
-
-            auto length = src_prim->length();
-
-            auto src_kind = classify_primtype(src_prim);
-            auto dst_kind = classify_primtype(dst_prim);
-            size_t src_bitwidth = conv_src_type->datatype->serialized_size();
-            size_t dst_bitwidth = conv_src_type->datatype->serialized_size();
-
-            SpvId data = emit(cast->from(), bb);
-
-            // If floating point is involved (src or dst), OpConvert*ToF and OpConvertFTo* can take care of the bit width transformation so no need for any chopping/expanding
-            if (src_kind == PrimTypeKind::Float || dst_kind == PrimTypeKind::Float) {
-                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
-                switch (src_kind) {
-                    case PrimTypeKind::Signed:    data = bb->convert(spv::OpConvertSToF, target_type, data); break;
-                    case PrimTypeKind::Unsigned:  data = bb->convert(spv::OpConvertUToF, target_type, data); break;
-                    case PrimTypeKind::Float:
-                        switch (dst_kind) {
-                            case PrimTypeKind::Signed:   data = bb->convert(spv::OpConvertFToS, target_type, data); break;
-                            case PrimTypeKind::Unsigned: data = bb->convert(spv::OpConvertFToU, target_type, data); break;
-                            default: THORIN_UNREACHABLE;
-                        }
-                        break;
-                }
-            } else {
-                // we expand first and shrink last to minimize precision losses, with bitcast in the middle
-                bool needs_chopping = src_bitwidth > dst_bitwidth;
-                bool needs_expanding = src_bitwidth < dst_bitwidth;
-
-                if (needs_expanding) {
-                    auto target_type = convert(get_primtype(world(), src_kind, src_bitwidth, length))->type_id;
-                    switch (src_kind) {
-                        case PrimTypeKind::Signed:
-                            data = bb->convert(spv::OpSConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Unsigned:
-                            data = bb->convert(spv::OpUConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Float:
-                            data = bb->convert(spv::OpFConvert, target_type, data);
-                            break;
-                    }
-                }
-
-                auto expanded_bitwidth = needs_expanding ? dst_bitwidth : src_bitwidth;
-                auto bitcast_target_type = convert(get_primtype(world(), dst_kind, expanded_bitwidth, length))->type_id;
-                data = bb->convert(spv::OpBitcast, bitcast_target_type, data);
-
-                if (needs_chopping) {
-                    auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
-                    switch (dst_kind) {
-                        case PrimTypeKind::Signed:
-                            data = bb->convert(spv::OpSConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Unsigned:
-                            data = bb->convert(spv::OpUConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Float:
-                            data = bb->convert(spv::OpFConvert, target_type, data);
-                            break;
-                    }
-                }
-            }
-        } else THORIN_UNREACHABLE;
-    } else if (def->isa<Bottom>()) {
-        return bb->undef(convert(def->type())->type_id);
-    }
-    assertf(false, "Incomplete emit(def) definition");
-}
-
-std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const Continuation* builtin, BasicBlockBuilder* bb) {
-    std::vector<SpvId> productions;
-    auto uvec3_t = convert(world().type_pu32(3));
-    auto u32_t = convert(world().type_pu32());
-    auto i32_t = convert(world().type_ps32());
-    if (builtin->name() == "spirv.nonsemantic.printf") {
-        std::vector<SpvId> args;
-        auto string = source_cont->arg(1);
-        if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
-            auto arr = string->as<DefiniteArray>();
-            std::vector<char> the_string;
-            for (size_t i = 0; i < arr_type->dim(); i++)
-                the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
-            the_string.push_back('\0');
-            args.push_back(builder_->debug_string(the_string.data()));
-        } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
-
-        for (size_t i = 2; i < source_cont->num_args() - 1; i++) {
-            args.push_back(emit(source_cont->arg(i), bb));
-        }
-
-        bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
-    } else if (builtin->name() == "get_work_dim") {
-        THORIN_UNREACHABLE;
-    } else if (builtin->name() == "get_global_id") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->global_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
-    } else if (builtin->name() == "get_local_size") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_size);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
-    } else if (builtin->name() == "get_local_id") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
-    } else if (builtin->name() == "get_num_groups") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->num_workgroups);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
-    } else if (builtin->name() == "get_group_id") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
-    } else {
-        world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
-    }
-    return productions;
-}
-
-}
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
deleted file mode 100644
index 7d31df5de..000000000
--- a/src/thorin/be/spirv/spirv.h
+++ /dev/null
@@ -1,186 +0,0 @@
-#ifndef THORIN_SPIRV_H
-#define THORIN_SPIRV_H
-
-#include "thorin/be/spirv/spirv_builder.hpp"
-#include "thorin/be/codegen.h"
-
-namespace thorin::spirv {
-
-using SpvId = builder::SpvId;
-
-class CodeGen;
-struct Datatype;
-struct PtrDatatype;
-
-struct FileBuilder;
-struct FnBuilder;
-
-struct ConvertedType {
-    ConvertedType(CodeGen* cg) : code_gen(cg) {}
-    ConvertedType(const ConvertedType&) = delete;
-
-    spirv::CodeGen* code_gen;
-    const thorin::Type* src_type;
-    SpvId type_id { 0 };
-    std::unique_ptr<Datatype> datatype;
-
-    bool is_known_size() { return datatype != nullptr; }
-};
-
-struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
-    explicit BasicBlockBuilder(FnBuilder& fn_builder);
-    BasicBlockBuilder(const BasicBlockBuilder&) = delete;
-
-    FnBuilder& fn_builder;
-    FileBuilder& file_builder;
-    std::unordered_map<const Param*, Phi> phis_map;
-    DefMap<SpvId> args;
-};
-
-struct FnBuilder : public builder::SpvFnBuilder {
-    explicit FnBuilder(CodeGen* cg, FileBuilder& file_builder);
-    FnBuilder(const FnBuilder&) = delete;
-
-    CodeGen* cg;
-    FileBuilder& file_builder;
-
-    const Scope* scope = nullptr;
-    std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
-    std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
-    ContinuationMap<SpvId> labels;
-    DefMap<SpvId> params;
-};
-
-struct Builtins {
-    SpvId workgroup_size;
-    SpvId num_workgroups;
-    SpvId workgroup_id;
-    SpvId local_id;
-    SpvId global_id;
-    SpvId local_invocation_index;
-
-    explicit Builtins(FileBuilder&);
-};
-
-struct ImportedInstructions {
-    SpvId shader_printf;
-
-    explicit ImportedInstructions(FileBuilder&);
-};
-
-struct FileBuilder : public builder::SpvFileBuilder {
-    explicit FileBuilder(CodeGen* cg);
-    FileBuilder(const FileBuilder&) = delete;
-
-    CodeGen* cg;
-
-    std::unique_ptr<Builtins> builtins;
-    std::unique_ptr<ImportedInstructions> imported_instrs;
-
-    SpvId u32_t();
-    SpvId u32_constant(uint32_t);
-
-private:
-    SpvId u32_t_ { 0 };
-    /*SpvId i32_t;
-    SpvId u32_t;
-    SpvId i64_t;
-    SpvId u64_t;
-    SpvId i32_constant(int32_t);
-    SpvId i64_constant(int64_t);
-    SpvId u64_constant(uint64_t);*/
-};
-
-class CodeGen : public thorin::CodeGen {
-public:
-    CodeGen(World&, Cont2Config&, bool debug);
-
-    void emit_stream(std::ostream& stream) override;
-    const char* file_ext() const override { return ".spv"; }
-
-    ConvertedType* convert(const Type*);
-protected:
-    void structure_loops();
-    void structure_flow();
-
-    void emit(const Scope& scope);
-    void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
-    SpvId emit(const Def* def, BasicBlockBuilder* bb);
-    std::vector<SpvId> emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
-
-    SpvId get_codom_type(const Continuation* fn);
-
-    std::unique_ptr<FileBuilder> builder_;
-    Continuation* entry_ = nullptr;
-    FnBuilder* current_fn_ = nullptr;
-    TypeMap<std::unique_ptr<ConvertedType>> types_;
-    DefMap<SpvId> defs_;
-    const Cont2Config& kernel_config_;
-
-    friend PtrDatatype;
-};
-
-/// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
-/// potentially different mappings, depending on the capabilities of the target platform. The serdes code deals with pointers
-/// in arrays of unsigned 32 bit words, and is there to get around the limitation of not being able to bitcast pointers in the
-/// logical addressing mode.
-struct Datatype {
-public:
-    ConvertedType* type;
-    Datatype(ConvertedType* type) : type(type) {}
-
-    // Datatypes are serialized using a base element, for now it is hardcoded to use 32-bit scalar unsigned integers
-    static constexpr size_t base_element_bitwidth = 32;
-    static constexpr size_t base_element_bytes = base_element_bitwidth / 8;
-
-    virtual size_t serialized_size() = 0;
-    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) = 0;
-    virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) = 0;
-};
-
-/// For scalar datatypes
-struct ScalarDatatype : public Datatype {
-    int type_tag;
-    size_t size_in_bytes;
-    size_t alignment;
-    ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
-
-    size_t serialized_size() override { return (size_in_bytes + 3) / 4; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-struct PtrDatatype : public Datatype {
-    static constexpr size_t bitwidth = 64;
-    PtrDatatype(ConvertedType* type) : Datatype(type) {}
-
-    size_t serialized_size() override { return bitwidth / 32; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-struct DefiniteArrayDatatype : public Datatype {
-    ConvertedType* element_type;
-    size_t length;
-
-    DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length);
-
-    size_t serialized_size() override { return element_type->datatype->serialized_size(); };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-struct ProductDatatype : public Datatype {
-    std::vector<ConvertedType*> elements_types;
-    size_t total_size = 0;
-
-    ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types);
-
-    size_t serialized_size() override { return total_size; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-}
-
-#endif //THORIN_SPIRV_H
diff --git a/src/thorin/be/spirv/spirv_datatypes.cpp b/src/thorin/be/spirv/spirv_datatypes.cpp
deleted file mode 100644
index 7c11a56c4..000000000
--- a/src/thorin/be/spirv/spirv_datatypes.cpp
+++ /dev/null
@@ -1,339 +0,0 @@
-#include "thorin/be/spirv/spirv.h"
-#include "thorin/util/stream.h"
-
-namespace thorin::spirv {
-
-ScalarDatatype::ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes)
-: Datatype(type), type_tag(type_tag), size_in_bytes(size_in_bytes), alignment(alignment_in_bytes) {}
-
-/// All serialization/deserialization methods use this so into a macro it goes
-#define serialization_types \
-SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id; \
-SpvId arr_cell_tid = bb.file_builder.declare_ptr_type(storage_class, u32_tid);
-
-SpvId ScalarDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
-    /// currently limited to 32-bit
-    assert(size_in_bytes == 4);
-    serialization_types;
-    auto cell = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto loaded = bb.load(u32_tid, cell);
-    return bb.convert(spv::OpBitcast, type->type_id, loaded);
-}
-
-void ScalarDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
-    /// currently limited to 32-bit
-    assert(size_in_bytes == 4);
-    serialization_types;
-    auto cell = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto casted = bb.convert(spv::OpBitcast, u32_tid, data);
-    bb.store(casted, cell);
-}
-
-SpvId PtrDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
-    assert(type->src_type->as<PtrType>()->addr_space() == AddrSpace::Global && "Only buffer device address (global memory) pointers supported");
-    serialization_types;
-    SpvId u64_tid = type->code_gen->convert(type->code_gen->world().type_pu64())->type_id;
-
-    auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.u32_constant(1)) });
-
-    auto lower = bb.convert(spv::OpUConvert, u64_tid, bb.load(u32_tid, cell0));
-    auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.load(u32_tid, cell1));
-
-    auto merged = bb.binop(spv::OpBitwiseOr, u64_tid, lower, bb.binop(spv::OpShiftLeftLogical, u64_tid, upper, bb.file_builder.u32_constant(32)));
-
-    return bb.convert(spv::OpConvertUToPtr, type->type_id, merged);
-}
-
-void PtrDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
-    assert(type->src_type->as<PtrType>()->addr_space() == AddrSpace::Global && "Only buffer device address (global memory) pointers supported");
-    serialization_types;
-    SpvId u64_tid = type->code_gen->convert(type->code_gen->world().type_pu64())->type_id;
-
-    auto u64_ptr = bb.convert(spv::OpConvertPtrToU, u64_tid, data);
-
-    auto cell0 = bb.access_chain(arr_cell_tid, array, { base_offset });
-    auto cell1 = bb.access_chain(arr_cell_tid, array, { bb.binop(spv::OpIAdd, u32_tid, base_offset, bb.file_builder.u32_constant(1)) });
-
-    auto lower = bb.convert(spv::OpUConvert, u64_tid, u64_ptr);
-    auto upper = bb.convert(spv::OpUConvert, u64_tid, bb.binop(spv::OpShiftRightLogical, u64_tid, u64_ptr, bb.file_builder.u32_constant(32)));
-
-    bb.store(lower, cell0);
-    bb.store(upper, cell1);
-}
-
-DefiniteArrayDatatype::DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length) : Datatype(type), element_type(element_type), length(length) {
-    assert(element_type->datatype.get() != nullptr);
-    assert(length > 0 && "Array lengths of zero are not supported");
-}
-
-SpvId DefiniteArrayDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    std::vector<SpvId> indices;
-    std::vector<SpvId> elements;
-    SpvId offset = base_offset;
-    SpvId stride = bb.file_builder.u32_constant(element_type->datatype->serialized_size());
-    for (size_t i = 0; i < length; i++) {
-        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, array, offset);
-        elements.push_back(element);
-        offset = bb.binop(spv::OpIAdd, u32_tid, offset, stride);
-    }
-    return bb.composite(type->type_id, elements);
-}
-void DefiniteArrayDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    std::vector<SpvId> indices;
-    SpvId offset = base_offset;
-    SpvId stride = bb.file_builder.u32_constant(element_type->datatype->serialized_size());
-    for (size_t i = 0; i < length; i++) {
-        element_type->datatype->emit_serialization(bb, storage_class, array, offset, bb.extract(element_type->type_id, data, { (uint32_t) i }));
-        offset = bb.binop(spv::OpIAdd, u32_tid, offset, stride);
-    }
-}
-
-ProductDatatype::ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types) : Datatype(type), elements_types(elements_types) {
-    // Unit datatype is acceptable, but serdes methods should never be invoked.
-    for (auto& element_type : elements_types) {
-        assert(element_type->datatype != nullptr);
-        total_size += element_type->datatype->serialized_size();
-    }
-}
-
-SpvId ProductDatatype::emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) {
-    assert(total_size > 0 && "It doesn't make sense to de-serialize Unit!");
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    std::vector<SpvId> indices;
-    std::vector<SpvId> elements;
-    SpvId offset = base_offset;
-    for (auto& element_type : elements_types) {
-        SpvId element = element_type->datatype->emit_deserialization(bb, storage_class, array, offset);
-        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.u32_constant(element_type->datatype->serialized_size()));
-        elements.push_back(element);
-    }
-    return bb.composite(type->type_id, elements);
-}
-void ProductDatatype::emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) {
-    assert(total_size > 0 && "It doesn't make sense to serialize Unit!");
-    SpvId u32_tid = type->code_gen->convert(type->code_gen->world().type_pu32())->type_id;
-    std::vector<SpvId> indices;
-    SpvId offset = base_offset;
-    int i = 0;
-    for (auto& element_type : elements_types) {
-        element_type->datatype->emit_serialization(bb, storage_class, array, offset, bb.extract(element_type->type_id, data, { (uint32_t) i++ }));
-        offset = bb.binop(spv::OpIAdd, u32_tid, offset, bb.file_builder.u32_constant(element_type->datatype->serialized_size()));
-    }
-}
-
-ConvertedType* CodeGen::convert(const Type* type) {
-    // Spir-V requires each primitive type to be "unique", it doesn't allow for example two 32-bit signed integer types.
-    // Therefore we must enforce that precise/quick types map to the same thing.
-    switch (type->tag()) {
-#define THORIN_Q_TYPE(T, M) \
-    case PrimType_##T: \
-        type = world().prim_type(PrimType_p##M, type->as<VectorType>()->length()); \
-        break;
-#include "thorin/tables/primtypetable.h"
-#undef THORIN_Q_TYPE
-        default: break;
-    }
-
-    if (auto iter = types_.find(type); iter != types_.end()) return iter->second.get();
-    ConvertedType* converted = types_.emplace(type, std::make_unique<ConvertedType>(this) ).first->second.get();
-    converted->src_type = type;
-
-    if (auto vec = type->isa<VectorType>(); vec && vec->length() > 1) {
-        auto component = vec->scalarize();
-        auto conv_comp = convert(component);
-        converted->type_id = builder_->declare_vector_type(conv_comp->type_id, (uint32_t)vec->length());
-        return converted;
-    }
-
-    switch (type->tag()) {
-        // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
-        // Using extensions, we could use 16 or 8-bit ints instead
-        // We can also pack them inside structures using bit-twiddling tricks, if the need arises
-        // Note: this only affects storing booleans inside structures, for regular variables the actual spir-v bool type is used.
-        case PrimType_bool:
-            converted->type_id = builder_->declare_bool_type();
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 1, 1);
-            break;
-        case PrimType_ps8:  assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pu8:  assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_ps16:
-            converted->type_id = builder_->declare_int_type(16, true);
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 2, 2);
-            break;
-        case PrimType_pu16:
-            converted->type_id = builder_->declare_int_type(16, false);
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 2, 2);
-            break;
-        case PrimType_ps32:
-            converted->type_id = builder_->declare_int_type(32, true );
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
-            break;
-        case PrimType_pu32:
-            converted->type_id = builder_->declare_int_type(32, false);
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
-            break;
-        case PrimType_ps64:
-            converted->type_id = builder_->declare_int_type(64, true);
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 8, 8);
-            break;
-        case PrimType_pu64:
-            converted->type_id = builder_->declare_int_type(64, false);
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 8, 8);
-            break;
-        case PrimType_pf16: assert(false && "TODO: look into capabilities to enable this");
-        case PrimType_pf32:
-            converted->type_id = builder_->declare_float_type(32);
-            converted->datatype = std::make_unique<ScalarDatatype>(converted, type->tag(), 4, 4);
-            break;
-        case PrimType_pf64: assert(false && "TODO: look into capabilities to enable this");
-        case Node_PtrType: {
-            auto ptr = type->as<PtrType>();
-            spv::StorageClass storage_class;
-            switch (ptr->addr_space()) {
-                case AddrSpace::Function: storage_class = spv::StorageClassFunction;     break;
-                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;      break;
-                case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant; break;
-                case AddrSpace::Global: {
-                    storage_class = spv::StorageClassPhysicalStorageBuffer;
-                    converted->datatype = std::make_unique<PtrDatatype>(converted);
-                    break;
-                }
-                case AddrSpace::Generic: {
-                    world().WLOG("Passing a generic pointer to a SPIR-V module. SpirV doesn't know about these, and so this will be passed as a 64 bit integer. Tread carefully !");
-                    ConvertedType* conv_u64 = convert(world().type_pu64());
-                    converted->type_id = conv_u64->type_id;
-                    goto ptr_done;
-                }
-                default:
-                    assert(false && "This address space is not supported");
-                    break;
-            }
-            {
-                const Type* pointee = ptr->pointee();
-                while (auto arr = pointee->isa<IndefiniteArrayType>())
-                    pointee = arr->elem_type();
-                ConvertedType* element = convert(pointee);
-                converted->type_id = builder_->declare_ptr_type(storage_class, element->type_id);
-
-                if (ptr->addr_space() == AddrSpace::Global) {
-                    assert(element->datatype && "Can only have physical pointers to known-size types");
-                    builder_->decorate(converted->type_id, spv::DecorationArrayStride, {(uint32_t) (element->datatype->serialized_size() * Datatype::base_element_bytes)});
-                }
-            }
-            ptr_done:
-            break;
-        }
-        case Node_IndefiniteArrayType: {
-            world().ELOG("Using indefinite types directly is not permitted - they may only be pointed to");
-            std::abort();
-        }
-        case Node_DefiniteArrayType: {
-            auto array = type->as<DefiniteArrayType>();
-            ConvertedType* element = convert(array->elem_type());
-            converted->type_id = builder_->declare_array_type(element->type_id, builder_->u32_constant(array->dim()));
-            converted->datatype = std::make_unique<DefiniteArrayDatatype>(converted, element, array->dim());
-            break;
-        }
-
-        case Node_ClosureType:
-        case Node_FnType: {
-            // extract "return" type, collect all other types
-            auto fn = type->as<FnType>();
-            ConvertedType* ret = nullptr;
-            std::vector<SpvId> ops;
-            for (auto op : fn->ops()) {
-                if (op->isa<MemType>() || op == world().unit()) continue;
-                auto fn_type = op->isa<FnType>();
-                if (fn_type && !op->isa<ClosureType>()) {
-                    assert(!ret && "only one 'return' supported");
-                    std::vector<ConvertedType*> ret_types;
-                    for (auto fn_op : fn_type->ops()) {
-                        if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
-                        ret_types.push_back(convert(fn_op));
-                    }
-                    if (ret_types.empty())          ret = convert(world().tuple_type({}));
-                    else if (ret_types.size() == 1) ret = ret_types.back();
-                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
-                } else
-                    ops.push_back(convert(op)->type_id);
-            }
-            assert(ret);
-
-            if (type->tag() == Node_FnType) {
-                converted->type_id = builder_->declare_fn_type(ops, ret->type_id);
-            } else {
-                assert(false && "TODO: handle closure mess");
-                THORIN_UNREACHABLE;
-            }
-            break;
-        }
-
-        case Node_StructType:
-        case Node_TupleType: {
-            std::vector<ConvertedType*> types;
-            std::vector<SpvId> spv_types;
-            size_t total_serialized_size = 0;
-            for (auto member_type : type->ops()) {
-                if (member_type == world().unit() || member_type == world().mem_type()) continue;
-                auto converted_member_type = convert(member_type);
-                types.push_back(converted_member_type);
-                spv_types.push_back(converted_member_type->type_id);
-                total_serialized_size = converted_member_type->datatype->serialized_size();
-            }
-            if (total_serialized_size == 0) {
-                outf("this one is void");
-                converted->type_id = builder_->void_type;
-                break;
-            }
-
-            converted->type_id = builder_->declare_struct_type(spv_types);
-            builder_->name(converted->type_id, type->to_string());
-            converted->datatype = std::make_unique<ProductDatatype>(converted, std::move(types));
-            break;
-        }
-
-        case Node_VariantType: {
-            assert(type->num_ops() > 0 && "empty variants not supported");
-            auto tag_type = world().type_pu32();
-            ConvertedType* converted_tag_type = convert(tag_type);
-
-            size_t max_serialized_size = 0;
-            for (auto member_type : type->as<VariantType>()->ops()) {
-                if (member_type == world().unit() || member_type == world().mem_type()) continue;
-                auto converted_member_type = convert(member_type);
-                if (converted_member_type->datatype->serialized_size() > max_serialized_size)
-                    max_serialized_size = converted_member_type->datatype->serialized_size();
-            }
-
-            if (max_serialized_size > 0) {
-                auto payload_type = world().definite_array_type(world().type_pu32(), max_serialized_size);
-                auto* converted_payload_type = convert(payload_type);
-
-                std::vector<SpvId> spv_pair = {converted_tag_type->type_id, converted_payload_type->type_id};
-                converted->type_id = builder_->declare_struct_type(spv_pair);
-                converted->datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*>{ converted_tag_type, converted_payload_type });
-            } else {
-                // We keep this useless level of struct so the rest of the code doesn't need a special path to extract the tag
-                std::vector<SpvId> spv_singleton = { converted_tag_type->type_id };
-                converted->type_id = builder_->declare_struct_type(spv_singleton);
-                converted->datatype = std::make_unique<ProductDatatype>(converted, std::vector<ConvertedType*>{ converted_tag_type });
-            }
-            builder_->name(converted->type_id, type->to_string());
-            break;
-        }
-
-        case Node_MemType: {
-            assert(false && "MemType cannot be converted to SPIR-V");
-        }
-
-        default:
-            THORIN_UNREACHABLE;
-    }
-
-    return converted;
-}
-
-}
\ No newline at end of file
diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index c594de907..44750da4a 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -5,6 +5,6 @@
 #cmakedefine01 THORIN_ENABLE_PROFILING
 #cmakedefine01 THORIN_ENABLE_LLVM
 #cmakedefine01 THORIN_ENABLE_RV
-#cmakedefine01 THORIN_ENABLE_SPIRV
+#cmakedefine01 THORIN_ENABLE_SHADY
 
 #endif
diff --git a/src/thorin/transform/structurize.cpp b/src/thorin/transform/structurize.cpp
index ee72a64e9..230e26b80 100644
--- a/src/thorin/transform/structurize.cpp
+++ b/src/thorin/transform/structurize.cpp
@@ -5,7 +5,7 @@
 #include "thorin/analyses/domtree.h"
 #include "thorin/world.h"
 
-namespace thorin::spirv {
+namespace thorin {
 
 using Head = LoopTree<true>::Head;
 using Base = LoopTree<true>::Base;
diff --git a/src/thorin/transform/structurize.h b/src/thorin/transform/structurize.h
index 6c6e23c70..0f82cf9ae 100644
--- a/src/thorin/transform/structurize.h
+++ b/src/thorin/transform/structurize.h
@@ -2,7 +2,11 @@
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/cfg.h"
 
+namespace thorin {
+
 class World;
 
 void structure_loops(World& world);
 void structure_flow(World& world);
+
+}
\ No newline at end of file

From d37167051026dc326ff063737b07dd9b7d499fb7 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 13 May 2022 08:52:21 +0200
Subject: [PATCH 096/342] started reworking structuriser to use Loop intrinsic
 in shd

---
 src/thorin/continuation.cpp          |   4 +-
 src/thorin/continuation.h            |  18 +-
 src/thorin/transform/structurize.cpp | 256 ++++++++++++++-------------
 src/thorin/world.h                   |   4 +
 4 files changed, 143 insertions(+), 139 deletions(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index d86338edc..00ed20a0e 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -260,7 +260,7 @@ void Continuation::match(const Def* val, Continuation* otherwise, Defs patterns,
     verify();
 }
 
-void Continuation::structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
+/*void Continuation::structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
     attributes_.intrinsic = Intrinsic::SCFLoopMerge;
     attributes_.scf_metadata.loop_epilogue.loop_header = loop_header;
     resize(targets.size());
@@ -283,7 +283,7 @@ void Continuation::structured_loop_header(const Continuation* loop_epilogue, con
     size_t x = 0;
     for (auto target : targets)
         set_op(x++, target);
-}
+}*/
 
 void Continuation::verify() const {
     if (!has_body())
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 1a9e54d27..ed0ed113a 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -111,14 +111,10 @@ enum class Intrinsic : uint8_t {
     Pipeline,                   ///< Intrinsic loop-pipelining-HLS-Backend
     Branch,                     ///< branch(cond, T, F).
     Match,                      ///< match(val, otherwise, (case1, cont1), (case2, cont2), ...)
-    SCFBegin,
-    SCFLoopHeader = SCFBegin,   ///< A header for a structured loop
-    SCFLoopMerge,               ///< A merge block for a structured loop
-    SCFLoopContinue,            ///< A continue block in a structured loop
-    SCFNonLocalJump,            ///< A non-local jump in a structured control flow graph
-    SCFBackEdge,                ///< A back edge a structured loop,
-    SCFEnd,
-    PeInfo = SCFEnd,            ///< Partial evaluation debug info.
+    LoopBegin,                  ///<
+    LoopBreak,                  ///<
+    LoopContinue,               ///<
+    PeInfo,                     ///< Partial evaluation debug info.
     EndScope                    ///< Dummy function which marks the end of a @p Scope.
 };
 
@@ -211,9 +207,9 @@ class Continuation : public Def {
     void jump(const Def* callee, Defs args, Debug dbg = {});
     void branch(const Def* cond, const Def* t, const Def* f, Debug dbg = {});
     void match(const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
-    void structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets);
-    void structured_loop_continue(const Continuation* loop_header);
-    void structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets);
+    //void structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets);
+    //void structured_loop_continue(const Continuation* loop_header);
+    //void structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets);
     void verify() const;
 
     const Filter* filter() const { return op(1)->as<Filter>(); }
diff --git a/src/thorin/transform/structurize.cpp b/src/thorin/transform/structurize.cpp
index 4ebd018a9..38ca453bd 100644
--- a/src/thorin/transform/structurize.cpp
+++ b/src/thorin/transform/structurize.cpp
@@ -26,9 +26,8 @@ struct DispatchTarget {
 };
 
 struct RewireMe {
-    RewireMe(Continuation* cont, int op) : cont(cont), op(op) {}
+    explicit RewireMe(Continuation* cont) : cont(cont) {}
     Continuation* cont;
-    int op;
 
     Continuation* backedge = nullptr;
     struct {
@@ -50,10 +49,11 @@ struct StructuredLoop {
     std::vector<DispatchTarget> inner_destinations = {};
     std::vector<DispatchTarget> outer_destinations = {};
 
-    // Created to serve as codegen helpers
-    Continuation* new_header = nullptr;
-    Continuation* new_epilogue = nullptr;
-    Continuation* new_continue = nullptr;
+    /// Just a regular continuation that simply calls into the real header
+    Continuation* pre_header;
+    /// Calls the loop_enter intrinsic with the set of internal and external dispatch nodes
+    Continuation* real_header;
+    Continuation* exit;
 
     // Same as the inner/outer destinations, but entry/exits instead now point to the corresponding header/epilogue nodes
     std::vector<const Continuation*> header_destination_conts;
@@ -97,7 +97,6 @@ inline void tag_continuations(ScopeContext& ctx, const Base* base, const Head* p
 
         StructuredLoop loop(parent, head, std::move(name));
         ctx.rewritten_loops.emplace(head, loop);
-
     } else if(base->isa<Leaf>()) {
         for (auto& node : base->cf_nodes()) {
             auto[i, result] = ctx.def2loop.emplace(node->continuation(), parent);
@@ -144,109 +143,106 @@ inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base
         const Leaf* leaf = base->as<Leaf>();
         auto cont = leaf->cf_node()->continuation();
         // For some nonsense reason, synthetic nodes created during scopes iteration leak in next iterations >:(
-        if (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)
+        if (!cont->has_body() /*|| (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)*/)
             return;
-        for (size_t i = 0; i < cont->num_ops(); i++) {
-            auto def = cont->op(i);
-            if (auto dest = def->isa_nom<Continuation>()) {
-                if (dest->intrinsic() == Intrinsic::Branch || dest->is_imported()) {
-                    continue;
+        auto app = cont->body();
+        auto callee = app->callee()->isa_nom<Continuation>();
+        if (!callee || callee->intrinsic() == Intrinsic::Branch || callee->is_imported())
+            return;
+
+        const Head* source_loop_head = ctx.def2loop[cont];
+        assert(ctx.def2loop.find(callee) != ctx.def2loop.end());
+        const Head* dest_loop_head = ctx.def2loop[callee];
+
+        if (source_loop_head != dest_loop_head) {
+            // We found a non-local jump
+            assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
+            auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
+
+            std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop_head);
+            std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop_head);
+            int bi = 0;
+            while (bi < static_cast<int>(std::min(source_path.size(), dest_path.size()))) {
+                if (source_path[bi] == dest_path[bi])
+                    bi++;
+                else break;
+            }
+
+            // The path is made out of a sequence of loops to break out of, and a sequence of loops to jump into
+            // these two sequences cannot be both empty (that wouldn't be a non-local jump then!)
+            std::vector<StructuredLoop*> leave;
+            std::vector<StructuredLoop*> enter;
+            for (int j = static_cast<int>(source_path.size()) - 1; j >= bi; j--)
+                leave.emplace_back(source_path[j]);
+            for (int j = bi; j < static_cast<int>(dest_path.size()); j++)
+                enter.emplace_back(dest_path[j]);
+
+            // 0 = this is the first step of the path
+            // 1 = last step was to break out of a loop
+            // 2 = last step was to enter a loop
+            int last = 0;
+            StructuredLoop* prev;
+
+            auto record_step = [&](DispatchTarget destination) {
+                if (last == 0) {
+                    // nothing to do, this node isn't a dispatching one
+                } else {
+                    if (last == 1)
+                        record_destination(prev->outer_destinations, destination);
+                    else
+                        record_destination(prev->inner_destinations, destination);
                 }
+            };
+
+            for (auto dest_loop : leave) {
+                DispatchTarget destination;
+                destination.exit = dest_loop;
 
-                const Head* source_loop_head = ctx.def2loop[cont];
-                assert(ctx.def2loop.find(dest) != ctx.def2loop.end());
-                const Head* dest_loop_head = ctx.def2loop[dest];
+                record_step(destination);
+                last = 1;
+                prev = dest_loop;
+                assert(prev != nullptr);
+            }
+            for (auto dest_loop : enter) {
+                DispatchTarget destination;
+                destination.entry = dest_loop;
 
-                if (source_loop_head != dest_loop_head) {
-                    // We found a non-local jump
+                record_step(destination);
+                last = 2;
+                prev = dest_loop;
+                assert(prev != nullptr);
+            }
+
+            assert(last != 0);
+            DispatchTarget destination;
+            destination.cont = callee;
+            record_step(destination);
+
+            RewireMe rewire(cont);
+            rewire.non_local_jump = {
+                std::move(leave),
+                std::move(enter),
+                callee
+            };
+            loop.rewire.emplace_back(rewire);
+        } else if (source_loop_head == dest_loop_head && source_loop_head != nullptr) {
+            for (auto& cf_node : source_loop_head->cf_nodes()) {
+                if (cf_node->continuation() == callee) {
+                    // We found a backedge
                     assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
                     auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
 
-                    std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop_head);
-                    std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop_head);
-                    int bi = 0;
-                    while (bi < static_cast<int>(std::min(source_path.size(), dest_path.size()))) {
-                        if (source_path[bi] == dest_path[bi])
-                            bi++;
-                        else break;
-                    }
-
-                    // The path is made out of a sequence of loops to break out of, and a sequence of loops to jump into
-                    // these two sequences cannot be both empty (that wouldn't be a non-local jump then!)
-                    std::vector<StructuredLoop*> leave;
-                    std::vector<StructuredLoop*> enter;
-                    for (int j = static_cast<int>(source_path.size()) - 1; j >= bi; j--)
-                        leave.emplace_back(source_path[j]);
-                    for (int j = bi; j < static_cast<int>(dest_path.size()); j++)
-                        enter.emplace_back(dest_path[j]);
-
-                    // 0 = this is the first step of the path
-                    // 1 = last step was to break out of a loop
-                    // 2 = last step was to enter a loop
-                    int last = 0;
-                    StructuredLoop* prev;
-
-                    auto record_step = [&](DispatchTarget destination) {
-                        if (last == 0) {
-                            // nothing to do, this node isn't a dispatching one
-                        } else {
-                            if (last == 1)
-                                record_destination(prev->outer_destinations, destination);
-                            else
-                                record_destination(prev->inner_destinations, destination);
-                        }
-                    };
-
-                    for (auto dest_loop : leave) {
-                        DispatchTarget destination;
-                        destination.exit = dest_loop;
-
-                        record_step(destination);
-                        last = 1;
-                        prev = dest_loop;
-                        assert(prev != nullptr);
-                    }
-                    for (auto dest_loop : enter) {
-                        DispatchTarget destination;
-                        destination.entry = dest_loop;
-
-                        record_step(destination);
-                        last = 2;
-                        prev = dest_loop;
-                        assert(prev != nullptr);
-                    }
-
-                    assert(last != 0);
                     DispatchTarget destination;
-                    destination.cont = dest;
-                    record_step(destination);
-
-                    RewireMe rewire(cont, i);
-                    rewire.non_local_jump = {
-                        std::move(leave),
-                        std::move(enter),
-                        dest
-                    };
+                    destination.cont = callee;
+                    record_destination(loop.inner_destinations, destination);
+
+                    RewireMe rewire(cont);
+                    rewire.backedge = cf_node->continuation();
                     loop.rewire.emplace_back(rewire);
-                } else if (source_loop_head == dest_loop_head && source_loop_head != nullptr) {
-                    for (auto& cf_node : source_loop_head->cf_nodes()) {
-                        if (cf_node->continuation() == dest) {
-                            // We found a backedge
-                            assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
-                            auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
-
-                            DispatchTarget destination;
-                            destination.cont = dest;
-                            record_destination(loop.inner_destinations, destination);
-
-                            RewireMe rewire(cont, i);
-                            rewire.backedge = cf_node->continuation();
-                            loop.rewire.emplace_back(rewire);
-                            break;
-                        }
-                    }
+                    return;
                 }
             }
+            assert(false);
         }
     }
 }
@@ -284,8 +280,8 @@ inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
             if (target.cont != nullptr) {
                 target_cont = target.cont;
             } else if (target.entry != nullptr) {
-                assert(target.entry->new_header != nullptr);
-                target_cont = target.entry->new_header;
+                assert(target.entry->pre_header != nullptr);
+                target_cont = target.entry->pre_header;
             } else {
                 assert(false && "Header dispatches may not exit loops");
             }
@@ -296,9 +292,13 @@ inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
         auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
         for (size_t i = 0; i < dest_types.size(); i++)
             variant_type->set(i, dest_types[i]);
-        auto fn_type = world.fn_type( { variant_type } );
-        loop.new_header = world.continuation(fn_type, { loop.name + "_new_header"});
-        loop.new_continue = world.continuation(fn_type, { loop.name + "_new_continue"});
+        Types enter_intrinsic_types = { variant_type };
+
+        //loop.enter_intrinsic = world.loop_enter(enter_intrinsic_types);
+        //loop.continue_intrinsic = world.loop_continue(enter_intrinsic_types);
+
+        loop.pre_header = world.continuation(world.fn_type(), { loop.name + "_new_header"});
+        // loop.new_continue = world.continuation(fn_type, { loop.name + "_new_continue"});
     }
 }
 
@@ -314,12 +314,12 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
                 if (target.cont != nullptr) {
                     target_cont = target.cont;
                 } else if (target.entry != nullptr) {
-                    assert(target.entry->new_header != nullptr);
-                    target_cont = target.entry->new_header;
+                    assert(target.entry->pre_header != nullptr);
+                    target_cont = target.entry->pre_header;
                 } else {
                     assert(target.exit != nullptr);
-                    assert(target.exit->new_epilogue != nullptr);
-                    target_cont = target.exit->new_epilogue;
+                    assert(target.exit->exit != nullptr);
+                    target_cont = target.exit->exit;
                 }
                 loop.epilogue_destination_conts.push_back(target_cont);
                 const thorin::FnType* target_type = target_cont->type();
@@ -328,8 +328,10 @@ inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base)
             auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
             for (size_t i = 0; i < dest_types.size(); i++)
                 variant_type->set(i, dest_types[i]);
-            auto fn_type = world.fn_type({variant_type});
-            loop.new_epilogue = world.continuation(fn_type, {loop.name + "_new_epilogue"});
+
+            //Types break_types = {variant_type};
+            //loop.break_intrinsic = world.loop_break(break_types);
+            loop.exit = world.continuation(world.fn_type(), {loop.name + "_new_epilogue"});
         }
 
         for (auto& children : head->children()) {
@@ -345,9 +347,8 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
         auto& loop = ctx.rewritten_loops.find(head)->second;
 
         if (head->num_cf_nodes() > 0) {
-            loop.new_epilogue->structured_loop_merge(loop.new_header, loop.epilogue_destination_conts);
-            loop.new_continue->structured_loop_continue(loop.new_header);
-            loop.new_header->structured_loop_header(loop.new_epilogue, loop.new_continue, loop.header_destination_conts);
+            //loop.header->structured_loop_merge(loop.new_header, loop.epilogue_destination_conts);
+            //loop.exit->structured_loop_header(loop.new_epilogue, loop.new_continue, loop.header_destination_conts);
         }
 
         for (auto& children : head->children()) {
@@ -365,20 +366,22 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 auto old_fn_type = rewire.backedge->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper_to" + destination.cont->unique_name() });
                 ctx.def2loop[wrapper] = loop.head;
-                wrapper->attributes_.intrinsic = Intrinsic::SCFBackEdge;
+                //wrapper->attributes_.intrinsic = Intrinsic::SCFBackEdge;
 
-                auto header_variant_type = loop.new_header->type()->op(0)->as<VariantType>();
-                wrapper->jump(loop.new_continue, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
+                //TODO
+                //auto header_variant_type = loop.continue_intrinsic->type()->op(0)->as<VariantType>();
+                //wrapper->jump(loop.continue_intrinsic, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
 
-                rewire.cont->unset_op(rewire.op);
-                rewire.cont->set_op(rewire.op, wrapper);
+                auto old_app = rewire.cont->body();
+                assert(old_app);
+                rewire.cont->jump(wrapper, old_app->args(), old_app->debug());
             } else {
                 auto& nlj = rewire.non_local_jump;
 
                 auto old_fn_type = nlj.final_destination->type();
                 auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper_to" + nlj.final_destination->unique_name() });
                 ctx.def2loop[wrapper] = loop.head;
-                wrapper->attributes_.intrinsic = Intrinsic::SCFNonLocalJump;
+                // wrapper->attributes_.intrinsic = Intrinsic::SCFNonLocalJump;
 
                 const Def* argument = tuple_from_params(world, wrapper->params());
                 Continuation* first_jump = nullptr;
@@ -390,10 +393,10 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                     StructuredLoop* loop_to_enter = nlj.enters[i];
 
                     auto variant_index = index_of_destination(loop_to_enter->inner_destinations, destination);
-                    auto header_variant_type = loop_to_enter->new_header->type()->op(0)->as<VariantType>();
+                    auto header_variant_type = loop_to_enter->pre_header->type()->op(0)->as<VariantType>();
                     argument = world.variant(header_variant_type, argument, variant_index);
 
-                    first_jump = loop_to_enter->new_header;
+                    first_jump = loop_to_enter->pre_header;
                     destination = {};
                     destination.entry = loop_to_enter;
                 }
@@ -402,10 +405,10 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                     StructuredLoop* loop_to_exit = nlj.exits[i];
 
                     auto variant_index = index_of_destination(loop_to_exit->outer_destinations, destination);
-                    auto header_variant_type = loop_to_exit->new_epilogue->type()->op(0)->as<VariantType>();
+                    auto header_variant_type = loop_to_exit->exit->type()->op(0)->as<VariantType>();
                     argument = world.variant(header_variant_type, argument, variant_index);
 
-                    first_jump = loop_to_exit->new_epilogue;
+                    first_jump = loop_to_exit->exit;
                     destination = {};
                     destination.exit = loop_to_exit;
                 }
@@ -413,8 +416,9 @@ inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
                 assert(first_jump != nullptr);
                 wrapper->jump(first_jump, { argument });
 
-                rewire.cont->unset_op(rewire.op);
-                rewire.cont->set_op(rewire.op, wrapper);
+                auto old_app = rewire.cont->body();
+                assert(old_app);
+                rewire.cont->jump(wrapper, old_app->args(), old_app->debug());
             }
         }
     }
@@ -459,4 +463,4 @@ void structure_flow(World& world) {
     });
 }
 
-}
\ No newline at end of file
+}
diff --git a/src/thorin/world.h b/src/thorin/world.h
index bddecbd6c..32035bf40 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -245,6 +245,10 @@ class World : public TypeTable, public Streamable<World> {
     Continuation* end_scope() const { return data_.end_scope_; }
     const Filter* filter(const Defs, Debug dbg = {});
 
+    Continuation* loop_enter(Types types, Continuations, Continuations);
+    Continuation* loop_continue(Types types);
+    Continuation* loop_break(Types types);
+
     /// Performs dead code, unreachable code and unused type elimination.
     void cleanup();
     void opt();

From f5f34e544a736f6d8e186bdc064e95746b0b70e9 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 8 Mar 2022 14:11:01 +0100
Subject: [PATCH 097/342] Add mem to branch instructions.

Will break artic and impala as a result of added parameters to branch
and match intrinsics. There are patches for those as well that need to
be merged together with this change.
---
 src/thorin/be/c/c.cpp       | 15 +++++++++------
 src/thorin/be/llvm/llvm.cpp | 20 +++++++++++++-------
 src/thorin/continuation.cpp | 15 ++++++++-------
 src/thorin/continuation.h   |  8 ++++----
 src/thorin/world.cpp        | 34 ++++++++++++++++++----------------
 5 files changed, 52 insertions(+), 40 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 56009879e..b0f0e2fc2 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -712,19 +712,22 @@ void CCodeGen::emit_epilogue(Continuation* cont) {
                 break;
         }
     } else if (body->callee() == world().branch()) {
-        auto c = emit(body->arg(0));
-        auto t = label_name(body->arg(1));
-        auto f = label_name(body->arg(2));
+        emit_unsafe(body->arg(0));
+        auto c = emit(body->arg(1));
+        auto t = label_name(body->arg(2));
+        auto f = label_name(body->arg(3));
         bb.tail.fmt("if ({}) goto {}; else goto {};", c, t, f);
     } else if (auto callee = body->callee()->as_nom<Continuation>(); callee && callee->intrinsic() == Intrinsic::Match) {
-        bb.tail.fmt("switch ({}) {{\t\n", emit(body->arg(0)));
+        emit_unsafe(body->arg(0));
 
-        for (size_t i = 2; i < body->num_args(); i++) {
+        bb.tail.fmt("switch ({}) {{\t\n", emit(body->arg(1)));
+
+        for (size_t i = 3; i < body->num_args(); i++) {
             auto arg = body->arg(i)->as<Tuple>();
             bb.tail.fmt("case {}: goto {};\n", emit_constant(arg->op(0)), label_name(arg->op(1)));
         }
 
-        bb.tail.fmt("default: goto {};", label_name(body->arg(1)));
+        bb.tail.fmt("default: goto {};", label_name(body->arg(2)));
         bb.tail.fmt("\b\n}}");
     } else if (body->callee()->isa<Bottom>()) {
         bb.tail.fmt("return;  // bottom: unreachable");
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index daa03569d..287cffce3 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -436,15 +436,21 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
                 irbuilder.CreateRet(agg);
         }
     } else if (body->callee() == world().branch()) {
-        auto cond = emit(body->arg(0));
-        auto tbb = cont2bb(body->arg(1)->as_nom<Continuation>());
-        auto fbb = cont2bb(body->arg(2)->as_nom<Continuation>());
+        auto mem = body->arg(0);
+        emit_unsafe(mem);
+
+        auto cond = emit(body->arg(1));
+        auto tbb = cont2bb(body->arg(2)->as_nom<Continuation>());
+        auto fbb = cont2bb(body->arg(3)->as_nom<Continuation>());
         irbuilder.CreateCondBr(cond, tbb, fbb);
     } else if (body->callee()->isa<Continuation>() && body->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
-        auto val = emit(body->arg(0));
-        auto otherwise_bb = cont2bb(body->arg(1)->as_nom<Continuation>());
-        auto match = irbuilder.CreateSwitch(val, otherwise_bb, body->num_args() - 2);
-        for (size_t i = 2; i < body->num_args(); i++) {
+        auto mem = body->arg(0);
+        emit_unsafe(mem);
+
+        auto val = emit(body->arg(1));
+        auto otherwise_bb = cont2bb(body->arg(2)->as_nom<Continuation>());
+        auto match = irbuilder.CreateSwitch(val, otherwise_bb, body->num_args() - 3);
+        for (size_t i = 3; i < body->num_args(); i++) {
             auto arg = body->arg(i)->as<Tuple>();
             auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
             auto case_bb    = cont2bb(arg->op(1)->as_nom<Continuation>());
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 719bf3477..04ec649ea 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -241,19 +241,20 @@ void Continuation::jump(const Def* callee, Defs args, Debug dbg) {
     verify();
 }
 
-void Continuation::branch(const Def* cond, const Def* t, const Def* f, Debug dbg) {
-    set_body(world().app(world().branch(), {cond, t, f}, dbg));
+void Continuation::branch(const Def* mem, const Def* cond, const Def* t, const Def* f, Debug dbg) {
+    set_body(world().app(world().branch(), {mem, cond, t, f}, dbg));
     verify();
 }
 
-void Continuation::match(const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg) {
-    Array<const Def*> args(patterns.size() + 2);
+void Continuation::match(const Def* mem, const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg) {
+    Array<const Def*> args(patterns.size() + 3);
 
-    args[0] = val;
-    args[1] = otherwise;
+    args[0] = mem;
+    args[1] = val;
+    args[2] = otherwise;
     assert(patterns.size() == continuations.size());
     for (size_t i = 0; i < patterns.size(); i++)
-        args[i + 2] = world().tuple({patterns[i], continuations[i]}, dbg);
+        args[i + 3] = world().tuple({patterns[i], continuations[i]}, dbg);
 
     set_body(world().app(world().match(val->type(), patterns.size()), args, dbg));
     verify();
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 9f34fe992..02f203a03 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -108,8 +108,8 @@ enum class Intrinsic : uint8_t {
     Undef,                      ///< Intrinsic undef function
     PipelineContinue,           ///< Intrinsic loop-pipelining-HLS-Backend
     Pipeline,                   ///< Intrinsic loop-pipelining-HLS-Backend
-    Branch,                     ///< branch(cond, T, F).
-    Match,                      ///< match(val, otherwise, (case1, cont1), (case2, cont2), ...)
+    Branch,                     ///< branch(mem, cond, T, F).
+    Match,                      ///< match(mem, val, otherwise, (case1, cont1), (case2, cont2), ...)
     PeInfo,                     ///< Partial evaluation debug info.
     EndScope                    ///< Dummy function which marks the end of a @p Scope.
 };
@@ -186,8 +186,8 @@ class Continuation : public Def {
     void destroy(const char*);
 
     void jump(const Def* callee, Defs args, Debug dbg = {});
-    void branch(const Def* cond, const Def* t, const Def* f, Debug dbg = {});
-    void match(const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
+    void branch(const Def* mem, const Def* cond, const Def* t, const Def* f, Debug dbg = {});
+    void match(const Def* mem, const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
     void verify() const;
 
     const Filter* filter() const { return op(1)->as<Filter>(); }
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 47f34b6aa..eb05a4ff4 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -44,7 +44,7 @@ namespace thorin {
 
 World::World(const std::string& name) {
     data_.name_   = name;
-    data_.branch_ = continuation(fn_type({type_bool(), fn_type(), fn_type()}), Intrinsic::Branch, {"br"});
+    data_.branch_ = continuation(fn_type({mem_type(), type_bool(), fn_type(), fn_type()}), Intrinsic::Branch, {"br"});
     data_.end_scope_ = continuation(fn_type(), Intrinsic::EndScope, {"end_scope"});
 }
 
@@ -1114,11 +1114,13 @@ Continuation* World::continuation(const FnType* fn, Continuation::Attributes att
 }
 
 Continuation* World::match(const Type* type, size_t num_patterns) {
-    Array<const Type*> arg_types(num_patterns + 2);
-    arg_types[0] = type;
-    arg_types[1] = fn_type();
-    for (size_t i = 0; i < num_patterns; i++)
-        arg_types[i + 2] = tuple_type({type, fn_type()});
+    Array<const Type*> arg_types(num_patterns + 3);
+    arg_types[0] = mem_type();
+    arg_types[1] = type;
+    arg_types[2] = fn_type({mem_type()});
+    for (size_t i = 0; i < num_patterns; i++) {
+        arg_types[i + 3] = tuple_type({type, fn_type({mem_type()})});
+    }
     return continuation(fn_type(arg_types), Intrinsic::Match, {"match"});
 }
 
@@ -1139,26 +1141,26 @@ const App* World::app(const Def* callee, const Defs args, Debug dbg) {
     if (auto continuation = callee->isa<Continuation>()) {
         switch (continuation->intrinsic()) {
             case Intrinsic::Branch: {
-                assert(args.size() == 3);
-                auto cond = args[0], t = args[1], f = args[2];
+                assert(args.size() == 4);
+                auto mem = args[0], cond = args[1], t = args[2], f = args[3];
                 if (auto lit = cond->isa<PrimLit>())
-                    return app(lit->value().get_bool() ? t : f, {}, dbg);
+                    return app(lit->value().get_bool() ? t : f, { mem }, dbg);
                 if (t == f)
-                    return app(t, {}, dbg);
+                    return app(t, { mem }, dbg);
                 if (is_not(cond)) {
                     auto inverted = cond->as<ArithOp>()->rhs();
-                    return app(branch(), {inverted, f, t}, dbg);
+                    return app(branch(), {mem, inverted, f, t}, dbg);
                 }
                 break;
             }
             case Intrinsic::Match:
-                if (args.size() == 2) return app(args[1], {}, dbg);
-                if (auto lit = args[0]->isa<PrimLit>()) {
-                    for (size_t i = 2; i < args.size(); i++) {
+                if (args.size() == 3) return app(args[2], { args[0] }, dbg);
+                if (auto lit = args[1]->isa<PrimLit>()) {
+                    for (size_t i = 3; i < args.size(); i++) {
                         if (extract(args[i], 0_s)->as<PrimLit>() == lit)
-                            return app(extract(args[i], 1), {}, dbg);
+                            return app(extract(args[i], 1), { args[0] }, dbg);
                     }
-                    return app(args[1], {}, dbg);
+                    return app(args[2], { args[0] }, dbg);
                 }
                 break;
             default:

From bd1ffe341daab334c5d9189c6dbcd91eeda80800 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 14 Jun 2022 14:27:02 +0200
Subject: [PATCH 098/342] Fix branch continuation type.

---
 src/thorin/world.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index eb05a4ff4..605a924ca 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -44,7 +44,7 @@ namespace thorin {
 
 World::World(const std::string& name) {
     data_.name_   = name;
-    data_.branch_ = continuation(fn_type({mem_type(), type_bool(), fn_type(), fn_type()}), Intrinsic::Branch, {"br"});
+    data_.branch_ = continuation(fn_type({mem_type(), type_bool(), fn_type({mem_type()}), fn_type({mem_type()})}), Intrinsic::Branch, {"br"});
     data_.end_scope_ = continuation(fn_type(), Intrinsic::EndScope, {"end_scope"});
 }
 

From c90a46e870e526445e29ff54979833c10c09ad71 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 25 Aug 2022 13:58:14 +0200
Subject: [PATCH 099/342] follow changes in shady's cmake target

---
 CMakeLists.txt            | 2 --
 src/thorin/CMakeLists.txt | 2 +-
 2 files changed, 1 insertion(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index a6452c04c..c47d74578 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -47,8 +47,6 @@ endif()
 find_package(shady REQUIRED CONFIG)
 if (shady_FOUND)
     message(STATUS "Found shady at ${shady_DIR}")
-    message(STATUS "Found shady headers ${shady_INCLUDE_DIRS}")
-    include_directories(${shady_INCLUDE_DIRS})
     set(THORIN_ENABLE_SHADY TRUE)
 endif()
 
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index ae04e530d..de8280c81 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -131,5 +131,5 @@ if(LLVM_FOUND)
 endif()
 
 if (shady_FOUND)
-    target_link_libraries(thorin PRIVATE shady)
+    target_link_libraries(thorin PRIVATE shady::shady)
 endif()

From 694644707c1182528d1e4fd14b1cf48d028f068d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Thu, 25 Aug 2022 22:34:10 +0200
Subject: [PATCH 100/342] shady backend: convert types

---
 src/thorin/be/shady/shady.cpp | 67 +++++++++++++++++++++++++++++++++--
 src/thorin/be/shady/shady.h   | 14 +++++---
 src/thorin/type.h             |  2 --
 3 files changed, 73 insertions(+), 10 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 814caae36..b74544527 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -3,7 +3,7 @@
 #include "thorin/analyses/scope.h"
 #include "thorin/transform/structurize.h"
 
-namespace thorin::shady {
+namespace thorin::shady_be {
 
 CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
         : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
@@ -47,8 +47,69 @@ void CodeGen::emit(const thorin::Scope& scope) {
     assert(false && "TODO");
 }
 
-shady::Type* CodeGen::convert(const Type *) {
-    return nullptr;
+shady::AddressSpace CodeGen::convert_address_space(AddrSpace as) {
+    switch(as) {
+        case AddrSpace::Generic: assert(false); break;
+        case AddrSpace::Global: return shady::AsGlobalPhysical;
+        case AddrSpace::Texture: assert(false); break;
+        case AddrSpace::Shared: return shady::AsSharedPhysical;
+        case AddrSpace::Constant: assert(false); break;
+        case AddrSpace::Private: return shady::AsPrivatePhysical;
+    }
+    assert(false);
+}
+
+const shady::Type* CodeGen::convert(const Type* type) {
+    if (auto res = types_.lookup(type)) return *res;
+    const shady::Type* t;
+    if (auto prim = type->isa<PrimType>()) {
+        switch (prim->primtype_tag()) {
+            case PrimType_bool:                     t = shady::bool_type(arena); break;
+            case PrimType_ps8:  case PrimType_qs8:
+            case PrimType_pu8:  case PrimType_qu8:  t = shady::int8_type(arena); break;
+            case PrimType_ps16: case PrimType_qs16:
+            case PrimType_pu16: case PrimType_qu16: t = shady::int16_type(arena); break;
+            case PrimType_ps32: case PrimType_qs32:
+            case PrimType_pu32: case PrimType_qu32: t = shady::int32_type(arena); break;
+            case PrimType_ps64: case PrimType_qs64:
+            case PrimType_pu64: case PrimType_qu64: t = shady::int64_type(arena); break;
+            case PrimType_pf16: case PrimType_qf16: assert(false && "TODO");
+            case PrimType_pf32: case PrimType_qf32: t = shady::float_type(arena); break;
+            case PrimType_pf64: case PrimType_qf64: assert(false && "TODO");
+            default: THORIN_UNREACHABLE;
+        }
+    } else if (auto ptr = type->isa<PtrType>()) {
+        t = shady::ptr_type(arena, (shady::PtrType) {
+            convert_address_space(ptr->addr_space()),
+            convert(ptr->pointee())
+        });
+    } else if (auto arr = type->isa<ArrayType>()) {
+        shady::ArrType payload = {};
+        payload.element_type = convert(arr->elem_type());
+        payload.size = nullptr;
+        if (auto definite = arr->isa<DefiniteArrayType>()) {
+            payload.size = shady::int32_literal(arena, static_cast<uint32_t>(definite->dim()));
+        }
+        t = shady::arr_type(arena, payload);
+    } else if (auto strct = type->isa<StructType>()) {
+        auto members = std::vector<const shady::Type*>(strct->num_ops());
+        for (size_t i = 0; i < strct->num_ops(); i++) {
+            members[i] = convert(strct->op(i));
+        }
+        shady::RecordType payload = {};
+        payload.members = shady::nodes(arena, strct->num_ops(), members.data());
+        payload.names = shady::strings(arena, 0, nullptr);
+        payload.special = shady::RecordType::NotSpecial;
+        t = shady::record_type(arena, payload);
+    } else if (auto variant = type->isa<VariantType>()) {
+        assert(false && "TODO");
+    } else {
+        assert(false);
+    }
+
+    assert(t);
+    types_[type] = t;
+    return t;
 }
 
 }
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
index df4ca0d77..848f43a3a 100644
--- a/src/thorin/be/shady/shady.h
+++ b/src/thorin/be/shady/shady.h
@@ -1,10 +1,12 @@
 #include "thorin/be/codegen.h"
 
-namespace thorin::shady {
-
+namespace shady {
 extern "C" {
 #include <shady/ir.h>
 }
+}
+
+namespace thorin::shady_be {
 
 class CodeGen : public thorin::CodeGen {
 public:
@@ -13,8 +15,10 @@ class CodeGen : public thorin::CodeGen {
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".shady"; }
 
-    shady::Type* convert(const Type*);
+    const shady::Type* convert(const Type*);
 protected:
+    shady::AddressSpace convert_address_space(AddrSpace);
+
     void emit(const Scope& scope);
     //void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     //shady::Node* emit(const Def* def, BasicBlockBuilder* bb);
@@ -25,8 +29,8 @@ class CodeGen : public thorin::CodeGen {
     std::vector<std::pair<shady::Node*, shady::Node*>> top_level;
 
     Continuation* entry_ = nullptr;
-    TypeMap<std::unique_ptr<shady::Type*>> types_;
-    DefMap<shady::Node*> defs_;
+    TypeMap<const shady::Type*> types_;
+    DefMap<const shady::Node*> defs_;
     const Cont2Config& kernel_config_;
 
 };
diff --git a/src/thorin/type.h b/src/thorin/type.h
index cb93682c6..5b96929bc 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -236,8 +236,6 @@ enum class AddrSpace : uint32_t {
     Shared   = 3,
     Constant = 4,
     Private =  5, // Corresponds to the 'private' storage class in SPIR-V
-    Function = 6, // Corresponds to the 'function' storage class in SPIR-V
-    Push     = 7, // Corresponds to the 'push constant' storage class in SPIR-V
 };
 
 /// Pointer type.

From 03b25a3d0af0d4322d395d02320644ffbfa88562 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 26 Aug 2022 14:03:18 +0200
Subject: [PATCH 101/342] shady: can codegen an empty hello world

---
 src/thorin/be/codegen.cpp     |   6 +-
 src/thorin/be/codegen.h       |   2 +-
 src/thorin/be/shady/shady.cpp | 204 +++++++++++++++++++++++++++++++---
 src/thorin/be/shady/shady.h   |  49 ++++++--
 src/thorin/continuation.cpp   |   2 +-
 5 files changed, 231 insertions(+), 32 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 87edaff30..517cb0232 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -95,7 +95,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
             std::pair { OpenCL, Intrinsic::OpenCL },
             std::pair { AMDGPU, Intrinsic::AMDGPU },
             std::pair { HLS,    Intrinsic::HLS    },
-            std::pair { SpirV,  Intrinsic::SpirV  }
+            std::pair { Shady,  Intrinsic::SpirV  }
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
@@ -116,7 +116,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         kernels.emplace_back(continuation);
     });
 
-    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU, SpirV }) {
+    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU, Shady }) {
         if (!importers_[backend].world().empty()) {
             get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 auto app = use->body();
@@ -195,7 +195,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     (void)opt;
 #endif
 #if THORIN_ENABLE_SHADY
-    if (!importers_[SpirV].world().empty()) cgs[SpirV] = std::make_unique<shady::CodeGen>(importers_[SpirV].world(), kernel_config, debug);
+    if (!importers_[Shady].world().empty()) cgs[Shady] = std::make_unique<shady_be::CodeGen>(importers_[Shady].world(), kernel_config, debug);
 #endif
     for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OpenCL }, std::pair { HLS, c::Lang::HLS } })
         if (!importers_[backend].world().empty()) cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug, flags);
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index b7477ef2f..8123e29a3 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -44,7 +44,7 @@ struct DeviceBackends {
     Cont2Config kernel_config;
     std::vector<Continuation*> kernels;
 
-    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, SpirV, BackendCount };
+    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, Shady, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> cgs;
 private:
     std::vector<Importer> importers_;
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index b74544527..f60ae0ed8 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -20,33 +20,24 @@ void CodeGen::emit_stream(std::ostream& out) {
     };
     arena = shady::new_arena(config);
 
-    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
+    Scope::for_each(world(), [&](const Scope& scope) { emit_scope(scope); });
 
     // build root node with the top level stuff that got emitted
-    auto decls = std::vector<shady::Node*>(top_level.size(), nullptr);
-    for (size_t i = 0; i < top_level.size(); i++) {
-        decls[i] = top_level[i].first;
-    }
     auto root = shady::root(arena, (shady::Root) {
-        .declarations = shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(decls.data())),
+        .declarations = shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(top_level.data())),
     });
 
-    shady::print_node(root);
-
-    out << "todo";
+    char* bufptr;
+    size_t size;
+    shady::print_node_into_str(root, &bufptr, &size);
+    out.write(bufptr, static_cast<std::streamsize>(size));
+    free(bufptr);
 
     shady::destroy_arena(arena);
     arena = nullptr;
     top_level.clear();
 }
 
-void CodeGen::emit(const thorin::Scope& scope) {
-    entry_ = scope.entry();
-    assert(entry_->is_returning());
-
-    assert(false && "TODO");
-}
-
 shady::AddressSpace CodeGen::convert_address_space(AddrSpace as) {
     switch(as) {
         case AddrSpace::Generic: assert(false); break;
@@ -59,10 +50,22 @@ shady::AddressSpace CodeGen::convert_address_space(AddrSpace as) {
     assert(false);
 }
 
+static inline int find_return_parameter(const FnType* type) {
+    for (size_t i = 0; i < type->num_ops(); i++) {
+        auto t = type->op(i);
+        if (t->order() % 2 == 1)
+            return static_cast<int>(i);
+    }
+    return -1;
+}
+
 const shady::Type* CodeGen::convert(const Type* type) {
     if (auto res = types_.lookup(type)) return *res;
     const shady::Type* t;
-    if (auto prim = type->isa<PrimType>()) {
+    if (type == world().mem_type()) {
+        t = nullptr;
+        goto skip_check;
+    } else if (auto prim = type->isa<PrimType>()) {
         switch (prim->primtype_tag()) {
             case PrimType_bool:                     t = shady::bool_type(arena); break;
             case PrimType_ps8:  case PrimType_qs8:
@@ -103,13 +106,180 @@ const shady::Type* CodeGen::convert(const Type* type) {
         t = shady::record_type(arena, payload);
     } else if (auto variant = type->isa<VariantType>()) {
         assert(false && "TODO");
+    } else if (auto fn_type = type->isa<FnType>()) {
+        shady::FnType payload = {};
+        payload.is_basic_block = fn_type->is_basicblock();
+        NodeVec dom, codom;
+
+        int return_param_i = find_return_parameter(fn_type);
+        for (size_t i = 0; i < fn_type->num_ops(); i++) {
+            // Skip the return param
+            if (return_param_i != -1 && i == static_cast<size_t>(return_param_i)) continue;
+            auto converted = convert(fn_type->op(i));
+            if (!converted)
+                continue; // Eliminate mem params
+            dom.push_back(converted);
+        }
+
+        if (return_param_i != -1) {
+            auto ret_fn_type = fn_type->op(return_param_i);
+            for (size_t i = 0; i < ret_fn_type->num_ops(); i++) {
+                auto converted = convert(ret_fn_type->op(i));
+                if (!converted)
+                    continue; // Eliminate mem params
+                codom.push_back(converted);
+            }
+        }
+
+        if (fn_type->is_basicblock())
+            assert(codom.empty());
+        payload.param_types = vec2nodes(dom);
+        payload.return_types = vec2nodes(codom);
+        t = shady::fn_type(arena, payload);
     } else {
         assert(false);
     }
 
     assert(t);
+    skip_check:
     types_[type] = t;
     return t;
 }
 
+shady::Node* CodeGen::def_to_decl(Def* def) {
+    NodeVec annotations;
+    if (auto cont = def->isa_nom<Continuation>()) {
+        NodeVec params;
+        NodeVec returns;
+
+        int ret_param_i = find_return_parameter(cont->type());
+
+        for (size_t i = 0; i < cont->num_params(); i++) {
+            if (ret_param_i != -1 && i == static_cast<size_t>(ret_param_i))
+                continue; // Skip the return parameter
+            auto type = convert(cont->param(i)->type());
+            if (!type) continue; // Eliminate mem tokens
+            auto param = shady::var(arena, type, cont->param(i)->name().c_str());
+            defs_[cont->param(i)] = param; // Register the param as emitted already
+            params.push_back(param);
+        }
+
+        if (!cont->is_basicblock() && ret_param_i >= 0) {
+            auto ret_fn_type = cont->type()->op(ret_param_i);
+
+            for (auto t : ret_fn_type->ops()) {
+                auto ret_type = convert(t);
+                if (!ret_type)
+                    continue; // Eliminate mem types
+                returns.push_back(ret_type);
+            }
+        }
+
+        return shady::fn(arena, vec2nodes(annotations), def->unique_name().c_str(), cont->is_basicblock(), vec2nodes(params), vec2nodes(returns));
+    } else if (auto global = def->isa<Global>()) {
+        if (global->is_mutable()) {
+            return shady::global_var(arena, vec2nodes(annotations), convert(global->alloced_type()), global->unique_name().c_str(), convert_address_space(AddrSpace::Private));
+        } else {
+            // Tentatively make those things constants...
+            auto constant = shady::constant(arena, vec2nodes(annotations), global->unique_name().c_str());;
+            constant->payload.constant.type_hint = convert(global->alloced_type());
+            return constant;
+        }
+    } else {
+        assert(false && "This doesn't map to a decl !");
+    }
+}
+
+shady::Node* CodeGen::get_decl(Def* def) {
+    for (auto& e : top_level) {
+        if (shady::get_decl_name(e) == def->unique_name())
+            return e;
+    }
+
+    auto decl = def_to_decl(def);
+    top_level.push_back(decl);
+    return decl;
+}
+
+shady::Node* CodeGen::prepare(const Scope& scope) {
+    cont2bb_[scope.entry()].head = curr_fn = get_decl(scope.entry());
+}
+
+void CodeGen::prepare(Continuation* cont, shady::Node*) {
+    BB& bb = cont2bb_[cont];
+    if (cont->is_basicblock())
+        bb.head = emit_fun_decl(cont);
+    else
+        assert(bb.head);
+
+    // Register params
+    // for (size_t i = 0; i < cont->num_params(); i++)
+    //     defs_[cont->param(i)] = bb.head->payload.fn.params.nodes[i];
+
+    bb.builder = shady::begin_block(arena);
+}
+
+void CodeGen::emit_epilogue(Continuation* cont) {
+    BB& bb = cont2bb_[cont];
+    assert(cont->has_body());
+    auto body = cont->body();
+    NodeVec args;
+    for (auto& arg : body->args()) {
+        if (convert(arg->type()) == nullptr) continue;
+        args.push_back(emit(arg));
+    }
+
+    if (body->callee() == entry_->ret_param()) {
+        shady::Return payload = {};
+        payload.fn = curr_fn;
+        payload.values = vec2nodes(args);
+        bb.terminator = shady::fn_ret(arena, payload);
+    } else if (body->callee() == world().branch()) {
+        shady::Branch payload = {};
+        payload.branch_mode = shady::Branch::BrIfElse;
+        payload.args = shady::nodes(arena, 0, nullptr);
+        payload.branch_condition = args[0];
+        payload.true_target      = args[1];
+        payload.false_target     = args[2];
+        bb.terminator = shady::branch(arena, payload);
+    } else if (auto match = body->callee()->as_nom<Continuation>(); match && match->intrinsic() == Intrinsic::Match) {
+        assert(false);
+    } else if (auto destination = body->callee()->isa_nom<Continuation>(); destination && destination->is_basicblock()) {
+        shady::Branch payload = {};
+        payload.args = vec2nodes(args);
+        payload.branch_mode = shady::Branch_::BrJump;
+        bb.terminator = shady::branch(arena, payload);
+    } else if (auto intrinsic = body->callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
+        assert(false);
+    } else if (auto callee = body->callee()->isa_nom<Continuation>()) {
+        shady::Callc payload = {};
+        int ret_param = find_return_parameter(callee->type());
+        payload.ret_cont = args[ret_param];
+        args.erase(args.begin() + ret_param);
+        payload.args = vec2nodes(args);
+        payload.is_return_indirect = false;
+        payload.callee = emit(callee);
+        bb.terminator = shady::callc(arena, payload);
+    } else {
+        assert(false);
+    }
+}
+
+void CodeGen::finalize(Continuation* cont) {
+    BB& bb = cont2bb_[cont];
+    assert(bb.head && bb.builder && bb.terminator);
+    bb.block = shady::finish_block(bb.builder, bb.terminator);
+    bb.head->payload.fn.block = bb.block;
+}
+
+void CodeGen::finalize(const Scope& scope) {
+    BB& bb = cont2bb_[scope.entry()];
+    assert(bb.head->payload.fn.block != nullptr);
+    curr_fn = nullptr;
+}
+
+const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
+    assert("TODO");
+}
+
 }
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
index 848f43a3a..ece795be4 100644
--- a/src/thorin/be/shady/shady.h
+++ b/src/thorin/be/shady/shady.h
@@ -1,4 +1,6 @@
 #include "thorin/be/codegen.h"
+#include "thorin/analyses/schedule.h"
+#include "thorin/be/emitter.h"
 
 namespace shady {
 extern "C" {
@@ -8,29 +10,56 @@ extern "C" {
 
 namespace thorin::shady_be {
 
-class CodeGen : public thorin::CodeGen {
+// using BB = std::pair<shady::BlockBuilder*, shady::Node*>;
+
+struct BB {
+    /// For an entry BB, this will also be the head of the entire function
+    shady::Node* head;
+    shady::BlockBuilder* builder;
+    const shady::Node* terminator;
+    const shady::Node* block;
+};
+
+class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node*, const shady::Type*, BB, CodeGen> {
 public:
     CodeGen(World&, Cont2Config&, bool debug);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".shady"; }
 
+    shady::Node* prepare(const Scope&);
+    void prepare(Continuation*, shady::Node*);
+    void emit_epilogue(Continuation*);
+    const shady::Node* emit_(const Def* def);
+    void finalize(const Scope&);
+    void finalize(Continuation*);
+
     const shady::Type* convert(const Type*);
+    const shady::Node* emit_bb(BB&, const Def*);
+
+    bool is_valid(const shady::Node* n) {
+        return n;
+    }
+
+    shady::Node* emit_fun_decl(Def* def) {
+        return get_decl(def);
+    }
 protected:
     shady::AddressSpace convert_address_space(AddrSpace);
+    shady::Node* def_to_decl(Def*);
+    shady::Node* get_decl(Def*);
 
-    void emit(const Scope& scope);
-    //void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
-    //shady::Node* emit(const Def* def, BasicBlockBuilder* bb);
-    //std::vector<SpvId> emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
+    using NodeVec = std::vector<const shady::Node*>;
+
+    inline shady::Nodes vec2nodes(NodeVec& vec) {
+        return shady::nodes(arena, vec.size(), vec.data());
+    }
 
-    //SpvId get_codom_type(const Continuation* fn);
     shady::IrArena* arena = nullptr;
-    std::vector<std::pair<shady::Node*, shady::Node*>> top_level;
+    std::vector<shady::Node*> top_level;
+
+    shady::Node* curr_fn;
 
-    Continuation* entry_ = nullptr;
-    TypeMap<const shady::Type*> types_;
-    DefMap<const shady::Node*> defs_;
     const Cont2Config& kernel_config_;
 
 };
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 1fa4a85c4..e6429b480 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -215,7 +215,7 @@ void Continuation::set_intrinsic() {
     else if (name() == "nvvm")           attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")         attributes().intrinsic = Intrinsic::OpenCL;
     else if (name() == "amdgpu")         attributes().intrinsic = Intrinsic::AMDGPU;
-    else if (name() == "spirv")                attributes().intrinsic = Intrinsic::SpirV;
+    else if (name() == "spirv")          attributes().intrinsic = Intrinsic::SpirV;
     else if (name() == "hls")            attributes().intrinsic = Intrinsic::HLS;
     else if (name() == "parallel")       attributes().intrinsic = Intrinsic::Parallel;
     else if (name() == "fibers")         attributes().intrinsic = Intrinsic::Fibers;

From 8dcfdee2c0c54dcde7093890fc30a48526ebe02d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 26 Aug 2022 17:58:57 +0200
Subject: [PATCH 102/342] can emit a few literals

---
 src/thorin/be/shady/shady.cpp | 66 ++++++++++++++++++++++++++++++++---
 src/thorin/be/shady/shady.h   |  6 ++--
 2 files changed, 64 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index f60ae0ed8..a267388df 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -118,6 +118,10 @@ const shady::Type* CodeGen::convert(const Type* type) {
             auto converted = convert(fn_type->op(i));
             if (!converted)
                 continue; // Eliminate mem params
+            shady::QualifiedType qtype;
+            qtype.type = converted;
+            qtype.is_uniform = false;
+            converted = shady::qualified_type(arena, qtype);
             dom.push_back(converted);
         }
 
@@ -159,6 +163,10 @@ shady::Node* CodeGen::def_to_decl(Def* def) {
                 continue; // Skip the return parameter
             auto type = convert(cont->param(i)->type());
             if (!type) continue; // Eliminate mem tokens
+            shady::QualifiedType qtype;
+            qtype.type = type;
+            qtype.is_uniform = false;
+            type = shady::qualified_type(arena, qtype);
             auto param = shady::var(arena, type, cont->param(i)->name().c_str());
             defs_[cont->param(i)] = param; // Register the param as emitted already
             params.push_back(param);
@@ -208,7 +216,7 @@ shady::Node* CodeGen::prepare(const Scope& scope) {
 void CodeGen::prepare(Continuation* cont, shady::Node*) {
     BB& bb = cont2bb_[cont];
     if (cont->is_basicblock())
-        bb.head = emit_fun_decl(cont);
+        bb.head = def_to_decl(cont);
     else
         assert(bb.head);
 
@@ -225,13 +233,22 @@ void CodeGen::emit_epilogue(Continuation* cont) {
     auto body = cont->body();
     NodeVec args;
     for (auto& arg : body->args()) {
-        if (convert(arg->type()) == nullptr) continue;
-        args.push_back(emit(arg));
+        if (convert(arg->type()) == nullptr) {
+            args.push_back(nullptr);
+        } else if (auto callee = arg->isa_nom<Continuation>(); callee && callee->is_basicblock()) {
+            // Emitting basic blocks as values isn't legal - but for convenience we'll put them in our list.
+            BB& callee_bb = cont2bb_[callee];
+            assert(callee_bb.head && callee_bb.head->payload.fn.is_basic_block);
+            args.push_back(callee_bb.head);
+        } else {
+            args.push_back(emit(arg));
+        }
     }
 
     if (body->callee() == entry_->ret_param()) {
         shady::Return payload = {};
         payload.fn = curr_fn;
+        args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
         payload.values = vec2nodes(args);
         bb.terminator = shady::fn_ret(arena, payload);
     } else if (body->callee() == world().branch()) {
@@ -246,6 +263,7 @@ void CodeGen::emit_epilogue(Continuation* cont) {
         assert(false);
     } else if (auto destination = body->callee()->isa_nom<Continuation>(); destination && destination->is_basicblock()) {
         shady::Branch payload = {};
+        args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
         payload.args = vec2nodes(args);
         payload.branch_mode = shady::Branch_::BrJump;
         bb.terminator = shady::branch(arena, payload);
@@ -254,8 +272,10 @@ void CodeGen::emit_epilogue(Continuation* cont) {
     } else if (auto callee = body->callee()->isa_nom<Continuation>()) {
         shady::Callc payload = {};
         int ret_param = find_return_parameter(callee->type());
+        assert(ret_param >= 0);
         payload.ret_cont = args[ret_param];
         args.erase(args.begin() + ret_param);
+        args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
         payload.args = vec2nodes(args);
         payload.is_return_indirect = false;
         payload.callee = emit(callee);
@@ -278,8 +298,46 @@ void CodeGen::finalize(const Scope& scope) {
     curr_fn = nullptr;
 }
 
+const shady::Node* CodeGen::emit_fun_decl(Continuation* cont) {
+    assert(!cont->is_basicblock());
+    shady::FnAddr payload;
+    payload.fn = get_decl(cont);
+    return shady::fn_addr(arena, payload);
+}
+
 const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
-    assert("TODO");
+    const shady::Node* v = nullptr;
+    if (auto prim_lit = def->isa<PrimLit>()) {
+        const auto& box = prim_lit->value();
+        switch (prim_lit->primtype_tag()) {
+            case PrimType_bool:                     v = box.get_bool() ? shady::true_lit(arena) : shady::false_lit(arena); break;
+            case PrimType_ps8:  case PrimType_qs8:  v =  shady::int8_literal (arena,  box.get_s8()); break;
+            case PrimType_pu8:  case PrimType_qu8:  v = shady::uint8_literal (arena,  box.get_u8()); break;
+            case PrimType_ps16: case PrimType_qs16: v =  shady::int16_literal(arena, box.get_s16()); break;
+            case PrimType_pu16: case PrimType_qu16: v = shady::uint16_literal(arena, box.get_u16()); break;
+            case PrimType_ps32: case PrimType_qs32: v =  shady::int32_literal(arena, box.get_s32()); break;
+            case PrimType_pu32: case PrimType_qu32: v = shady::uint32_literal(arena, box.get_u32()); break;
+            case PrimType_ps64: case PrimType_qs64: v =  shady::int64_literal(arena, box.get_s64()); break;
+            case PrimType_pu64: case PrimType_qu64: v = shady::uint64_literal(arena, box.get_u64()); break;
+            case PrimType_pf16: case PrimType_qf16: assert(false && "TODO");
+            case PrimType_pf32: case PrimType_qf32: v = shady::float_type(arena); break;
+            case PrimType_pf64: case PrimType_qf64: assert(false && "TODO");
+            default: THORIN_UNREACHABLE;
+        }
+    } else if (auto arr = def->isa<DefiniteArray>()) {
+        NodeVec contents;
+        for (auto& e : arr->ops()) {
+            assert(emit(e));
+            contents.push_back(emit(e));
+        }
+        shady::ArrayLiteral payload;
+        payload.element_type = convert(arr->elem_type());
+        payload.contents = vec2nodes(contents);
+        v = shady::arr_lit(arena, payload);
+    }
+    assert(v && shady::is_value(v));
+    defs_[def] = v;
+    return v;
 }
 
 }
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
index ece795be4..bd4c8026b 100644
--- a/src/thorin/be/shady/shady.h
+++ b/src/thorin/be/shady/shady.h
@@ -30,7 +30,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node
     shady::Node* prepare(const Scope&);
     void prepare(Continuation*, shady::Node*);
     void emit_epilogue(Continuation*);
-    const shady::Node* emit_(const Def* def);
+    // const shady::Node* emit_(const Def* def);
     void finalize(const Scope&);
     void finalize(Continuation*);
 
@@ -41,9 +41,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node
         return n;
     }
 
-    shady::Node* emit_fun_decl(Def* def) {
-        return get_decl(def);
-    }
+    const shady::Node* emit_fun_decl(Continuation*);
 protected:
     shady::AddressSpace convert_address_space(AddrSpace);
     shady::Node* def_to_decl(Def*);

From 8903cb44bcaeb6c74d471d7acdb06618cb3fb30f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Tue, 30 Aug 2022 18:36:16 +0200
Subject: [PATCH 103/342] fixes for msvc

---
 src/thorin/be/shady/shady.cpp | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index a267388df..84136734a 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -15,17 +15,17 @@ void CodeGen::emit_stream(std::ostream& out) {
     structure_loops(world());
     structure_flow(world());
 
-    auto config = shady::ArenaConfig {
-        .check_types = true,
-    };
+    shady::ArenaConfig config = { 0 };
+    config.check_types = true;
     arena = shady::new_arena(config);
 
     Scope::for_each(world(), [&](const Scope& scope) { emit_scope(scope); });
 
     // build root node with the top level stuff that got emitted
-    auto root = shady::root(arena, (shady::Root) {
-        .declarations = shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(top_level.data())),
-    });
+    shady::Root root_payload = {
+        shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(top_level.data()))
+    };
+    auto root = shady::root(arena, root_payload);
 
     char* bufptr;
     size_t size;
@@ -82,10 +82,11 @@ const shady::Type* CodeGen::convert(const Type* type) {
             default: THORIN_UNREACHABLE;
         }
     } else if (auto ptr = type->isa<PtrType>()) {
-        t = shady::ptr_type(arena, (shady::PtrType) {
+        shady::PtrType payload = {
             convert_address_space(ptr->addr_space()),
             convert(ptr->pointee())
-        });
+        };
+        t = shady::ptr_type(arena, payload);
     } else if (auto arr = type->isa<ArrayType>()) {
         shady::ArrType payload = {};
         payload.element_type = convert(arr->elem_type());
@@ -210,7 +211,7 @@ shady::Node* CodeGen::get_decl(Def* def) {
 }
 
 shady::Node* CodeGen::prepare(const Scope& scope) {
-    cont2bb_[scope.entry()].head = curr_fn = get_decl(scope.entry());
+    return cont2bb_[scope.entry()].head = curr_fn = get_decl(scope.entry());
 }
 
 void CodeGen::prepare(Continuation* cont, shady::Node*) {
@@ -273,7 +274,7 @@ void CodeGen::emit_epilogue(Continuation* cont) {
         shady::Callc payload = {};
         int ret_param = find_return_parameter(callee->type());
         assert(ret_param >= 0);
-        payload.ret_cont = args[ret_param];
+        payload.join_at = args[ret_param];
         args.erase(args.begin() + ret_param);
         args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
         payload.args = vec2nodes(args);

From 2c0ce6c68db2ebd70ec40c4c3229cd04f2ea7ad1 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 13 Sep 2022 16:45:52 +0200
Subject: [PATCH 104/342] Remove recursion in schedule, hoist_enters and
 importer.

Using recursion in these passes can lead to stack overflows if large
programs are being compiled.
---
 src/thorin/analyses/schedule.cpp      |  67 ++++++++++++++--
 src/thorin/analyses/schedule.h        |   3 +
 src/thorin/transform/hoist_enters.cpp |  17 +++-
 src/thorin/transform/importer.cpp     | 108 +++++++++++++++++++++++---
 src/thorin/transform/importer.h       |   5 +-
 5 files changed, 180 insertions(+), 20 deletions(-)

diff --git a/src/thorin/analyses/schedule.cpp b/src/thorin/analyses/schedule.cpp
index 798de7255..03a6eb23f 100644
--- a/src/thorin/analyses/schedule.cpp
+++ b/src/thorin/analyses/schedule.cpp
@@ -9,6 +9,8 @@
 #include "thorin/analyses/looptree.h"
 #include "thorin/analyses/scope.h"
 
+#include <stack>
+
 namespace thorin {
 
 Scheduler::Scheduler(const Scope& s)
@@ -44,25 +46,73 @@ Scheduler::Scheduler(const Scope& s)
     }
 }
 
-Continuation* Scheduler::early(const Def* def) {
+std::stack<const Def*> early_todo;
+
+Continuation* Scheduler::early(const Def * def) {
     if (auto cont = early_.lookup(def)) return *cont;
-    if (auto param = def->isa<Param>()) return early_[def] = param->continuation();
+
+    early_todo.push(def);
+    Continuation *return_cont = nullptr;
+    while (!early_todo.empty()) {
+        return_cont = early_intern();
+    }
+    assert(return_cont);
+    return return_cont;
+}
+
+Continuation* Scheduler::early_intern() {
+    const Def* def = early_todo.top();
+
+    if (auto cont = early_.lookup(def)) {
+        early_todo.pop();
+        return *cont;
+    }
+    if (auto param = def->isa<Param>()) {
+        early_todo.pop();
+        return early_[def] = param->continuation();
+    }
 
     auto result = scope().entry();
     for (auto op : def->as_structural()->ops()) {
         if (!op->isa_nom<Continuation>() && def2uses_.find(op) != def2uses_.end()) {
-            auto cont = early(op);
+            Continuation *cont;
+            if (early_.lookup(op)) {
+                cont = *early_.lookup(op);
+            } else {
+                early_todo.push(op);
+                return nullptr;
+            }
             if (domtree().depth(cfg(cont)) > domtree().depth(cfg(result)))
                 result = cont;
         }
     }
 
+    early_todo.pop();
     return early_[def] = result;
 }
 
-Continuation* Scheduler::late(const Def* def) {
+std::stack<const Def*> late_todo;
+
+Continuation* Scheduler::late(const Def * def) {
     if (auto cont = late_.lookup(def)) return *cont;
 
+    late_todo.push(def);
+    Continuation *return_cont = nullptr;
+    while (!late_todo.empty()) {
+        return_cont = late_intern();
+    }
+    assert(return_cont);
+    return return_cont;
+}
+
+Continuation* Scheduler::late_intern() {
+    const Def* def = late_todo.top();
+
+    if (auto cont = late_.lookup(def)) {
+        late_todo.pop();
+        return *cont;
+    }
+
     Continuation* result = nullptr;
     if (auto continuation = def->isa_nom<Continuation>()) {
         result = continuation;
@@ -70,11 +120,18 @@ Continuation* Scheduler::late(const Def* def) {
         result = param->continuation();
     } else {
         for (auto use : uses(def)) {
-            auto cont = late(use);
+            Continuation* cont;
+            if (late_.lookup(use)) {
+                cont = *late_.lookup(use);
+            } else {
+                late_todo.push(use);
+                return nullptr;
+            }
             result = result ? domtree().least_common_ancestor(cfg(result), cfg(cont))->continuation() : cont;
         }
     }
 
+    late_todo.pop();
     return late_[def] = result;
 }
 
diff --git a/src/thorin/analyses/schedule.h b/src/thorin/analyses/schedule.h
index b2cfa65aa..57d874274 100644
--- a/src/thorin/analyses/schedule.h
+++ b/src/thorin/analyses/schedule.h
@@ -48,6 +48,9 @@ class Scheduler {
     DefMap<Continuation*> late_;
     DefMap<Continuation*> smart_;
     DefMap<Uses> def2uses_;
+
+    Continuation* early_intern();
+    Continuation* late_intern();
 };
 
 using Schedule = std::vector<Continuation*>;
diff --git a/src/thorin/transform/hoist_enters.cpp b/src/thorin/transform/hoist_enters.cpp
index c052498e2..b44a8e22a 100644
--- a/src/thorin/transform/hoist_enters.cpp
+++ b/src/thorin/transform/hoist_enters.cpp
@@ -4,8 +4,12 @@
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/verify.h"
 
+#include <stack>
+
 namespace thorin {
 
+std::stack<const Def*> todo;
+
 static void find_enters(std::deque<const Enter*>& enters, const Def* def) {
     if (auto enter = def->isa<Enter>())
         enters.push_back(enter);
@@ -15,13 +19,20 @@ static void find_enters(std::deque<const Enter*>& enters, const Def* def) {
 
     for (auto use : def->uses()) {
         if (auto memop = use->isa<MemOp>())
-            find_enters(enters, memop);
+            todo.push(memop);
     }
 }
 
 static void find_enters(std::deque<const Enter*>& enters, Continuation* continuation) {
-    if (auto mem_param = continuation->mem_param())
-        find_enters(enters, mem_param);
+    if (auto mem_param = continuation->mem_param()) {
+        todo.push(mem_param);
+        while (!todo.empty()) {
+            auto next_item = todo.top();
+            todo.pop();
+
+            find_enters(enters, next_item);
+        }
+    }
 }
 
 static void hoist_enters(const Scope& scope) {
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index f125585de..d476fb718 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -1,8 +1,12 @@
 #include "thorin/transform/importer.h"
 
+#include <set>
+#include <stack>
+#include <utility>
+
 namespace thorin {
 
-const Type* Importer::import(const Type* otype) {
+const Type* Importer::import_type(const Type* otype) {
     if (auto ntype = type_old2new_.lookup(otype)) {
         assert(&(*ntype)->table() == &world_);
         return *ntype;
@@ -13,13 +17,13 @@ const Type* Importer::import(const Type* otype) {
         auto ntype = nominal_type->stub(world_);
         type_old2new_[otype] = ntype;
         for (size_t i = 0; i != size; ++i)
-            ntype->set(i, import(otype->op(i)));
+            ntype->set(i, import_type(otype->op(i)));
         return ntype;
     }
 
     Array<const Type*> nops(size);
     for (size_t i = 0; i != size; ++i)
-        nops[i] = import(otype->op(i));
+        nops[i] = import_type(otype->op(i));
 
     auto ntype = otype->rebuild(world_, nops);
     type_old2new_[otype] = ntype;
@@ -28,38 +32,108 @@ const Type* Importer::import(const Type* otype) {
     return ntype;
 }
 
+std::stack<std::pair<const Def*, bool>> required_defs;
+std::set<const Def *> analyzed_conts;
+
+void enqueue(const Def* elem) {
+    if (elem->isa_nom<Continuation>()) {
+        if (analyzed_conts.find(elem) != analyzed_conts.end()) {
+            required_defs.push(std::pair(elem, false));
+        } else {
+            analyzed_conts.insert(elem);
+            required_defs.push(std::pair(elem, true));
+        }
+    } else {
+        required_defs.push(std::pair(elem, false));
+    }
+}
+
 const Def* Importer::import(const Def* odef) {
     if (auto ndef = def_old2new_.lookup(odef)) {
         assert(&(*ndef)->world() == &world_);
         return *ndef;
     }
 
-    auto ntype = import(odef->type());
+    assert(required_defs.empty());
+    enqueue(odef);
+
+    const Def* return_def = nullptr;
+    while (!required_defs.empty()) {
+        return_def = import_nonrecursive();
+    }
+
+    assert(return_def);
+    assert (return_def == def_old2new_.lookup(odef));
+
+    analyzed_conts.clear();
+
+    return return_def;
+}
+
+const Def* Importer::import_nonrecursive() {
+    const Def* odef = required_defs.top().first;
+    bool jump_to_analyze = required_defs.top().second;
+
+    Continuation* ncontinuation = nullptr;
+
+    if (auto ndef = def_old2new_.lookup(odef)) {
+        assert(&(*ndef)->world() == &world_);
+        if (odef->isa_nom<Continuation>()) {
+            if (!jump_to_analyze) {
+                required_defs.pop();
+                return *ndef;
+            }
+            ncontinuation = (*ndef)->as_nom<Continuation>();
+        } else {
+            required_defs.pop();
+            return *ndef;
+        }
+    }
+
+    auto ntype = import_type(odef->type());
 
     if (auto oparam = odef->isa<Param>()) {
+        if (!def_old2new_.lookup(oparam->continuation())) {
+            enqueue(oparam->continuation());
+            return nullptr;
+        }
         import(oparam->continuation())->as_nom<Continuation>();
         auto nparam = def_old2new_[oparam];
         assert(nparam && &nparam->world() == &world_);
-        return nparam;
+        required_defs.pop();
+        return def_old2new_[oparam] = nparam;
     }
 
     if (auto ofilter = odef->isa<Filter>()) {
         Array<const Def*> new_conditions(ofilter->num_ops());
+
+        bool unfinished_business = false;
+        for (size_t i = 0, e = ofilter->size(); i != e; ++i)
+            if (!def_old2new_.lookup(ofilter->condition(i))) {
+                enqueue(ofilter->condition(i));
+                unfinished_business = true;
+            }
+        if (unfinished_business)
+            return nullptr;
+
         for (size_t i = 0, e = ofilter->size(); i != e; ++i)
             new_conditions[i] = import(ofilter->condition(i));
         auto nfilter = world().filter(new_conditions, ofilter->debug());
-        return nfilter;
+        required_defs.pop();
+        return def_old2new_[ofilter] = nfilter;
     }
 
-    Continuation* ncontinuation = nullptr;
-    if (auto ocontinuation = odef->isa_nom<Continuation>()) { // create stub in new world
+    if (auto ocontinuation = odef->isa_nom<Continuation>(); ocontinuation && !ncontinuation) { // create stub in new world
         assert(!ocontinuation->dead_);
         // TODO maybe we want to deal with intrinsics in a more streamlined way
-        if (ocontinuation == ocontinuation->world().branch())
+        if (ocontinuation == ocontinuation->world().branch()) {
+            required_defs.pop();
             return def_old2new_[ocontinuation] = world().branch();
-        if (ocontinuation == ocontinuation->world().end_scope())
+        } else if (ocontinuation == ocontinuation->world().end_scope()) {
+            required_defs.pop();
             return def_old2new_[ocontinuation] = world().end_scope();
-        auto npi = import(ocontinuation->type())->as<FnType>();
+        }
+        auto npi = import_type(ocontinuation->type())->as<FnType>();
         ncontinuation = world().continuation(npi, ocontinuation->attributes(), ocontinuation->debug_history());
         assert(&ncontinuation->world() == &world());
         assert(&npi->table() == &world());
@@ -76,6 +150,16 @@ const Def* Importer::import(const Def* odef) {
 
     size_t size = odef->num_ops();
     Array<const Def*> nops(size);
+
+    bool unfinished = false;
+    for (size_t i = 0; i != size; ++i)
+        if (!def_old2new_.lookup(odef->op(i))) {
+            enqueue(odef->op(i));
+            unfinished = true;
+        }
+    if (unfinished)
+        return nullptr;
+
     for (size_t i = 0; i != size; ++i) {
         assert(odef->op(i) != odef);
         nops[i] = import(odef->op(i));
@@ -85,6 +169,7 @@ const Def* Importer::import(const Def* odef) {
     if (odef->isa_structural()) {
         auto ndef = odef->rebuild(world(), ntype, nops);
         todo_ |= odef->tag() != ndef->tag();
+        required_defs.pop();
         return def_old2new_[odef] = ndef;
     }
 
@@ -94,6 +179,7 @@ const Def* Importer::import(const Def* odef) {
         ncontinuation->set_body(napp);
     ncontinuation->set_filter(nops[1]->as<Filter>());
     ncontinuation->verify();
+    required_defs.pop();
     return ncontinuation;
 }
 
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index c34db9ef3..eb7d5bc42 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -20,10 +20,13 @@ class Importer {
     }
 
     World& world() { return world_; }
-    const Type* import(const Type*);
+    const Type* import_type(const Type*);
     const Def* import(const Def*);
     bool todo() const { return todo_; }
 
+private:
+    const Def* import_nonrecursive();
+
 public:
     Type2Type type_old2new_;
     Def2Def def_old2new_;

From 0374889ca135b5f0f5c099c4652d2a92a345dc61 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 16 Sep 2022 16:02:57 +0200
Subject: [PATCH 105/342] Initialize HashTable array before using it.

---
 src/thorin/util/hash.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/util/hash.h b/src/thorin/util/hash.h
index 32bf6d6bd..007c7fbb4 100644
--- a/src/thorin/util/hash.h
+++ b/src/thorin/util/hash.h
@@ -217,6 +217,7 @@ class HashTable {
     HashTable()
         : capacity_(StackCapacity)
         , size_(0)
+        , array_()
         , nodes_(array_.data())
 #if THORIN_ENABLE_CHECKS
         , id_(0)

From dc54ccaac495d898f0dfb6338953df2aa6d944a6 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 22 Sep 2022 16:18:56 +0200
Subject: [PATCH 106/342] Scheduler performance enhancement

---
 src/thorin/analyses/schedule.cpp | 33 +++++++++++++++++++++-----------
 1 file changed, 22 insertions(+), 11 deletions(-)

diff --git a/src/thorin/analyses/schedule.cpp b/src/thorin/analyses/schedule.cpp
index 03a6eb23f..38a67a4b6 100644
--- a/src/thorin/analyses/schedule.cpp
+++ b/src/thorin/analyses/schedule.cpp
@@ -72,16 +72,23 @@ Continuation* Scheduler::early_intern() {
         return early_[def] = param->continuation();
     }
 
-    auto result = scope().entry();
+    bool todo_empty = true;
     for (auto op : def->as_structural()->ops()) {
         if (!op->isa_nom<Continuation>() && def2uses_.find(op) != def2uses_.end()) {
-            Continuation *cont;
-            if (early_.lookup(op)) {
-                cont = *early_.lookup(op);
-            } else {
+            if (!early_.lookup(op)) {
                 early_todo.push(op);
-                return nullptr;
+                todo_empty = false;
             }
+        }
+    }
+    if (!todo_empty)
+        return nullptr;
+
+    auto result = scope().entry();
+    for (auto op : def->as_structural()->ops()) {
+        if (!op->isa_nom<Continuation>() && def2uses_.find(op) != def2uses_.end()) {
+            Continuation *cont = *early_.lookup(op);
+            assert(cont);
             if (domtree().depth(cfg(cont)) > domtree().depth(cfg(result)))
                 result = cont;
         }
@@ -119,14 +126,18 @@ Continuation* Scheduler::late_intern() {
     } else if (auto param = def->isa<Param>()) {
         result = param->continuation();
     } else {
+        bool todo_empty = true;
         for (auto use : uses(def)) {
-            Continuation* cont;
-            if (late_.lookup(use)) {
-                cont = *late_.lookup(use);
-            } else {
+            if (!late_.lookup(use)) {
                 late_todo.push(use);
-                return nullptr;
+                todo_empty = false;
             }
+        }
+        if (!todo_empty)
+            return nullptr;
+        for (auto use : uses(def)) {
+            Continuation* cont = *late_.lookup(use);
+            assert(cont);
             result = result ? domtree().least_common_ancestor(cfg(result), cfg(cont))->continuation() : cont;
         }
     }

From 624294b0b3ba0ec2d3091fbed46d7608bbcdb301 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 22 Sep 2022 16:22:43 +0200
Subject: [PATCH 107/342] Backend remove recursion on memory.

---
 src/thorin/be/emitter.h | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

diff --git a/src/thorin/be/emitter.h b/src/thorin/be/emitter.h
index 72300ba06..655e8f544 100644
--- a/src/thorin/be/emitter.h
+++ b/src/thorin/be/emitter.h
@@ -1,6 +1,9 @@
 #ifndef THORIN_BE_EMITTER_H
 #define THORIN_BE_EMITTER_H
 
+#include<stack>
+#include<queue>
+
 namespace thorin {
 
 template<class Value, class Type, class BB, class Child>
@@ -11,6 +14,32 @@ class Emitter {
 
     /// Internal wrapper for @p emit that checks and retrieves/puts the @c Value from @p defs_.
     Value emit_(const Def* def) {
+        std::stack<const Def*> required_defs;
+        std::queue<const Def*> todo;
+        todo.push(def);
+
+        while (!todo.empty()) {
+            auto def = todo.front();
+            todo.pop();
+            if (defs_.lookup(def)) continue;
+
+            if (auto memop = def->isa<MemOp>()) {
+                todo.push(memop->mem());
+                required_defs.push(memop->mem());
+            } else if (auto extract = def->isa<Extract>()) {
+                if (is_mem(extract)) {
+                    todo.push(extract->agg());
+                    required_defs.push(extract->agg());
+                }
+            }
+        }
+
+        while (!required_defs.empty()) {
+            auto r = required_defs.top();
+            required_defs.pop();
+            emit_unsafe(r);
+        }
+
         auto place = def->no_dep() ? entry_ : scheduler_.smart(def);
         auto& bb = cont2bb_[place];
         return child().emit_bb(bb, def);

From a9ad15284133a0d5b6a5594f71457c250b539337 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 7 Oct 2022 13:47:35 +0200
Subject: [PATCH 108/342] use pop(x) instead of x.top(); x.pop();

---
 src/thorin/be/emitter.h               | 3 +--
 src/thorin/transform/hoist_enters.cpp | 3 +--
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/emitter.h b/src/thorin/be/emitter.h
index 655e8f544..1b396804f 100644
--- a/src/thorin/be/emitter.h
+++ b/src/thorin/be/emitter.h
@@ -35,8 +35,7 @@ class Emitter {
         }
 
         while (!required_defs.empty()) {
-            auto r = required_defs.top();
-            required_defs.pop();
+            auto r = pop(required_defs);
             emit_unsafe(r);
         }
 
diff --git a/src/thorin/transform/hoist_enters.cpp b/src/thorin/transform/hoist_enters.cpp
index b44a8e22a..5cd7fcfb4 100644
--- a/src/thorin/transform/hoist_enters.cpp
+++ b/src/thorin/transform/hoist_enters.cpp
@@ -27,8 +27,7 @@ static void find_enters(std::deque<const Enter*>& enters, Continuation* continua
     if (auto mem_param = continuation->mem_param()) {
         todo.push(mem_param);
         while (!todo.empty()) {
-            auto next_item = todo.top();
-            todo.pop();
+            auto next_item = pop(todo);
 
             find_enters(enters, next_item);
         }

From 72d2deed2b07094d41b01d606ff5dce6450d6ddb Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 7 Oct 2022 13:48:32 +0200
Subject: [PATCH 109/342] Mangler remove recursion on body.

---
 src/thorin/transform/mangle.cpp | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index 0dcc0d909..629134fb7 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -5,6 +5,9 @@
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
 
+#include<queue>
+#include<utility>
+
 namespace thorin {
 
 const Def* Rewriter::instantiate(const Def* odef) {
@@ -54,6 +57,8 @@ Mangler::Mangler(const Scope& scope, Defs args, Defs lift)
     }
 }
 
+std::queue<std::pair<Continuation*, const Continuation*>> bodies_to_mangle;
+
 Continuation* Mangler::mangle() {
     // create new_entry - but first collect and specialize all param types
     std::vector<const Type*> param_types;
@@ -97,7 +102,17 @@ Continuation* Mangler::mangle() {
         new_entry()->set_filter(world().filter(new_conditions, old_entry()->filter()->debug()));
     }
 
-    new_entry()->set_body(mangle_body(old_entry()->body()));
+    bodies_to_mangle.push(std::pair(new_entry(), old_entry()));
+
+    while (!bodies_to_mangle.empty()) {
+        auto task = pop(bodies_to_mangle);
+
+        auto new_cont = task.first;
+        auto old_cont = task.second;
+
+        assert(!new_cont->has_body());
+        new_cont->set_body(mangle_body(old_cont->body()));
+    }
 
     new_entry()->verify();
 
@@ -155,7 +170,7 @@ const Def* Mangler::mangle(const Def* old_def) {
     else if (auto old_continuation = old_def->isa_nom<Continuation>()) {
         auto new_continuation = mangle_head(old_continuation);
         if (old_continuation->has_body())
-            new_continuation->set_body(mangle_body(old_continuation->body()));
+            bodies_to_mangle.push(std::pair(new_continuation, old_continuation));
         return new_continuation;
     } else if (auto param = old_def->isa<Param>()) {
         assert(within(param->continuation()));

From fcbd30423b842e9352138eef9a0f1192dbe95eef Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 18 Oct 2022 14:37:16 +0200
Subject: [PATCH 110/342] Improve verification.

---
 src/thorin/analyses/verify.cpp | 30 ++++++++++++++++++++++++------
 src/thorin/continuation.cpp    |  9 ++++++---
 src/thorin/continuation.h      |  4 ++--
 3 files changed, 32 insertions(+), 11 deletions(-)

diff --git a/src/thorin/analyses/verify.cpp b/src/thorin/analyses/verify.cpp
index 33ba03a19..bbd034e8c 100644
--- a/src/thorin/analyses/verify.cpp
+++ b/src/thorin/analyses/verify.cpp
@@ -2,16 +2,19 @@
 #include "thorin/type.h"
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
+#include "thorin/analyses/free_defs.h"
 
 namespace thorin {
 
 // TODO this needs serious rewriting
 
-static void verify_calls(World& world) {
+static bool verify_calls(World& world) {
+    bool ok = true;
     for (auto def : world.defs()) {
         if (auto cont = def->isa<Continuation>())
-            cont->verify();
+            ok &= cont->verify();
     }
+    return ok;
 }
 
 static bool verify_top_level(World& world) {
@@ -24,14 +27,29 @@ static bool verify_top_level(World& world) {
             ok = false;
         }
     });
-    if (!ok)
-        world.dump();
+    return ok;
+}
+
+static bool verify_mem(World& world) {
+    bool ok = true;
+    Scope::for_each(world, [&] (const Scope& scope) {
+        for (auto def : free_defs(scope)) {
+            if (is_mem(def)) {
+                world.ELOG("scope '{}' got free mem '{}' with {} uses", scope.entry(), def, def->num_uses());
+                ok = false;
+            }
+        }
+    });
     return ok;
 }
 
 void verify(World& world) {
-    verify_calls(world);
-    verify_top_level(world);
+    bool ok = true;
+    ok &= verify_calls(world);
+    ok &= verify_top_level(world);
+    ok &= verify_mem(world);
+    if (!ok)
+        world.dump();
 }
 
 }
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 048e0449b..a33183d17 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -28,7 +28,7 @@ App::App(const Defs ops, Debug dbg) : Def(Node_App, ops[0]->world().bottom_type(
 #endif
 }
 
-void App::verify() const {
+bool App::verify() const {
     auto callee_type = callee()->type()->isa<FnType>(); // works for closures too, no need for a special case
     assertf(callee_type, "callee type must be a FnType");
     assertf(callee_type->num_ops() == num_args(), "app node '{}' has fn type {} with {} parameters, but is supplied {} arguments", this, callee_type, callee_type->num_ops(), num_args());
@@ -37,6 +37,7 @@ void App::verify() const {
         auto at = arg(i)->type();
         assertf(pt == at, "app node argument {} has type {} but the callee was expecting {}", this, at, pt);
     }
+    return true;
 }
 
 //------------------------------------------------------------------------------
@@ -262,15 +263,17 @@ void Continuation::match(const Def* mem, const Def* val, Continuation* otherwise
     verify();
 }
 
-void Continuation::verify() const {
+bool Continuation::verify() const {
+    bool ok = true;
     if (!has_body())
         assertf(filter()->is_empty(), "continuations with no body should have an empty (no) filter");
     else {
-        body()->verify();
+        ok &= body()->verify();
         assert(!dead_); // destroy() should remove the body
         assert(intrinsic() == Intrinsic::None);
         assertf(filter()->is_empty() || num_params() == filter()->size(), "The filter needs to be either empty, or match the param count");
     }
+    return ok;
 }
 
 /// Rewrites the body to only keep the non-specialized arguments
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 02f203a03..ded1fd25d 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -72,7 +72,7 @@ class App : public Def {
     }
 
     void jump(const Def* callee, Defs args, Debug dbg = {});
-    void verify() const;
+    bool verify() const;
 
     friend class World;
 };
@@ -188,7 +188,7 @@ class Continuation : public Def {
     void jump(const Def* callee, Defs args, Debug dbg = {});
     void branch(const Def* mem, const Def* cond, const Def* t, const Def* f, Debug dbg = {});
     void match(const Def* mem, const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
-    void verify() const;
+    bool verify() const;
 
     const Filter* filter() const { return op(1)->as<Filter>(); }
     void set_filter(const Filter* f) {

From 3dd3f974a10ea916b3424c7797e97a1b6e0ffd51 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 18 Oct 2022 14:40:45 +0200
Subject: [PATCH 111/342] World::continuation: Add caller name to debug
 history, for debugging.

---
 src/thorin/debug.h        | 11 +++++++++++
 src/thorin/rec_stream.cpp |  5 ++++-
 src/thorin/type.cpp       |  4 ++--
 src/thorin/world.cpp      |  9 +++++++++
 4 files changed, 26 insertions(+), 3 deletions(-)

diff --git a/src/thorin/debug.h b/src/thorin/debug.h
index b4d19d14b..ef35f2a36 100644
--- a/src/thorin/debug.h
+++ b/src/thorin/debug.h
@@ -43,18 +43,29 @@ class Debug {
     Debug() = default; // TODO remove
     Debug(std::string name, Loc loc = {}, const Def* meta = nullptr)
         : name(name)
+        , creation_context("")
         , loc(loc)
         , meta(meta)
     {}
     Debug(const char* name, Loc loc = {}, const Def* meta = nullptr)
         : Debug(std::string(name), loc, meta)
     {}
+    Debug(std::string name, std::string creation_context, Loc loc = {}, const Def* meta = nullptr)
+        : name(name)
+        , creation_context(creation_context)
+        , loc(loc)
+        , meta(meta)
+    {}
+    Debug(const char* name, const char* creation_context, Loc loc = {}, const Def* meta = nullptr)
+        : Debug(std::string(name), std::string(creation_context), loc, meta)
+    {}
     Debug(Loc loc)
         : Debug("", loc)
     {}
     //Debug(const Def*);
 
     std::string name;
+    std::string creation_context;
     Loc loc;
     const Def* meta = nullptr;
 };
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 03d6a8c20..df983be34 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -99,7 +99,10 @@ Stream& Def::stream1(Stream& s) const {
     if (auto param = isa<Param>()) {
         return s.fmt("{}.{}", param->continuation(), param->unique_name());
     } else if (isa<Continuation>()) {
-        return s.fmt("cont {}", unique_name());
+        if (debug().creation_context != "")
+            return s.fmt("cont {} [{}]", unique_name(), debug().creation_context);
+        else
+            return s.fmt("cont {}", unique_name());
     } else if (auto app = isa<App>()) {
         return s.fmt("{}({, })", app->callee(), app->args());
     } else if (isa<Filter>()) {
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index f8cc33751..7918b0e45 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -143,10 +143,10 @@ Stream& Type::stream(Stream& s) const {
     else if (isa< FrameType>()) return s.fmt("frame");
     else if (auto t = isa<DefiniteArrayType>()) {
         return s.fmt("[{} x {}]", t->dim(), t->elem_type());
+    } else if (auto t = isa<ClosureType>()) {
+        return s.fmt("closure[{, }]", t->ops());
     } else if (auto t = isa<FnType>()) {
         return s.fmt("fn[{, }]", t->ops());
-    } else if (auto t = isa<ClosureType>()) {
-        return s.fmt("closure [{, }]", t->ops());
     } else if (auto t = isa<IndefiniteArrayType>()) {
         return s.fmt("[{}]", t->elem_type());
     } else if (auto t = isa<StructType>()) {
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 605a924ca..d2c9d17a2 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -10,6 +10,7 @@
 #endif
 
 #include <cmath>
+#include <execinfo.h>
 
 #include "thorin/def.h"
 #include "thorin/primop.h"
@@ -1102,7 +1103,15 @@ const Def* World::run(const Def* def, Debug dbg) {
  */
 
 Continuation* World::continuation(const FnType* fn, Continuation::Attributes attributes, Debug dbg) {
+    void *array[10];
+    size_t size = backtrace(array, 10);
+    assert(size >= 2);
+    char ** symbols = backtrace_symbols(array, 10);
+
+    dbg.creation_context = symbols[1];
+
     auto cont = put<Continuation>(fn, attributes, dbg);
+    free(symbols);
 
     size_t i = 0;
     for (auto op : fn->ops()) {

From da1fe76d0f03a5b04862dec11f97a3abe736c9c8 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 31 Oct 2022 11:13:18 +0100
Subject: [PATCH 112/342] adapted API

---
 src/thorin/be/shady/shady.cpp | 145 +++++++++++++++++++++-------------
 src/thorin/be/shady/shady.h   |   8 +-
 2 files changed, 93 insertions(+), 60 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 84136734a..4da148a44 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -10,32 +10,26 @@ CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
 {}
 
 void CodeGen::emit_stream(std::ostream& out) {
-    assert(top_level.empty());
+    // structure_loops(world());
+    // structure_flow(world());
 
-    structure_loops(world());
-    structure_flow(world());
+    assert(!module);
 
     shady::ArenaConfig config = { 0 };
     config.check_types = true;
-    arena = shady::new_arena(config);
+    arena = shady::new_ir_arena(config);
+    module = shady::new_module(arena, world().name().c_str());
 
     Scope::for_each(world(), [&](const Scope& scope) { emit_scope(scope); });
 
-    // build root node with the top level stuff that got emitted
-    shady::Root root_payload = {
-        shady::nodes(arena, top_level.size(), const_cast<const shady::Node**>(top_level.data()))
-    };
-    auto root = shady::root(arena, root_payload);
-
     char* bufptr;
     size_t size;
-    shady::print_node_into_str(root, &bufptr, &size);
+    shady::print_module_into_str(module, &bufptr, &size);
     out.write(bufptr, static_cast<std::streamsize>(size));
     free(bufptr);
 
-    shady::destroy_arena(arena);
+    shady::destroy_ir_arena(arena);
     arena = nullptr;
-    top_level.clear();
 }
 
 shady::AddressSpace CodeGen::convert_address_space(AddrSpace as) {
@@ -103,13 +97,12 @@ const shady::Type* CodeGen::convert(const Type* type) {
         shady::RecordType payload = {};
         payload.members = shady::nodes(arena, strct->num_ops(), members.data());
         payload.names = shady::strings(arena, 0, nullptr);
-        payload.special = shady::RecordType::NotSpecial;
+        payload.special = shady::NotSpecial;
         t = shady::record_type(arena, payload);
     } else if (auto variant = type->isa<VariantType>()) {
         assert(false && "TODO");
     } else if (auto fn_type = type->isa<FnType>()) {
         shady::FnType payload = {};
-        payload.is_basic_block = fn_type->is_basicblock();
         NodeVec dom, codom;
 
         int return_param_i = find_return_parameter(fn_type);
@@ -151,9 +144,10 @@ const shady::Type* CodeGen::convert(const Type* type) {
     return t;
 }
 
-shady::Node* CodeGen::def_to_decl(Def* def) {
+shady::Node* CodeGen::emit_decl_head(Def* def) {
     NodeVec annotations;
     if (auto cont = def->isa_nom<Continuation>()) {
+        assert(!cont->is_basicblock());
         NodeVec params;
         NodeVec returns;
 
@@ -173,26 +167,22 @@ shady::Node* CodeGen::def_to_decl(Def* def) {
             params.push_back(param);
         }
 
-        if (!cont->is_basicblock() && ret_param_i >= 0) {
-            auto ret_fn_type = cont->type()->op(ret_param_i);
+        auto ret_fn_type = cont->type()->op(ret_param_i);
 
-            for (auto t : ret_fn_type->ops()) {
-                auto ret_type = convert(t);
-                if (!ret_type)
-                    continue; // Eliminate mem types
-                returns.push_back(ret_type);
-            }
+        for (auto t : ret_fn_type->ops()) {
+            auto ret_type = convert(t);
+            if (!ret_type)
+                continue; // Eliminate mem types
+            returns.push_back(ret_type);
         }
 
-        return shady::fn(arena, vec2nodes(annotations), def->unique_name().c_str(), cont->is_basicblock(), vec2nodes(params), vec2nodes(returns));
+        return shady::function(module, vec2nodes(params), def->unique_name().c_str(), vec2nodes(annotations), vec2nodes(returns));
     } else if (auto global = def->isa<Global>()) {
         if (global->is_mutable()) {
-            return shady::global_var(arena, vec2nodes(annotations), convert(global->alloced_type()), global->unique_name().c_str(), convert_address_space(AddrSpace::Private));
+            return shady::global_var(module, vec2nodes(annotations), convert(global->alloced_type()), global->unique_name().c_str(), convert_address_space(AddrSpace::Private));
         } else {
             // Tentatively make those things constants...
-            auto constant = shady::constant(arena, vec2nodes(annotations), global->unique_name().c_str());;
-            constant->payload.constant.type_hint = convert(global->alloced_type());
-            return constant;
+            return shady::constant(module, vec2nodes(annotations), convert(global->alloced_type()), global->unique_name().c_str());
         }
     } else {
         assert(false && "This doesn't map to a decl !");
@@ -200,14 +190,14 @@ shady::Node* CodeGen::def_to_decl(Def* def) {
 }
 
 shady::Node* CodeGen::get_decl(Def* def) {
-    for (auto& e : top_level) {
+    shady::Nodes already_done = shady::get_module_declarations(module);
+    for (size_t i = 0; i < already_done.count; i++) {
+        auto& e = already_done.nodes[i];
         if (shady::get_decl_name(e) == def->unique_name())
-            return e;
+            return (shady::Node*) e;
     }
 
-    auto decl = def_to_decl(def);
-    top_level.push_back(decl);
-    return decl;
+    return emit_decl_head(def);
 }
 
 shady::Node* CodeGen::prepare(const Scope& scope) {
@@ -216,18 +206,38 @@ shady::Node* CodeGen::prepare(const Scope& scope) {
 
 void CodeGen::prepare(Continuation* cont, shady::Node*) {
     BB& bb = cont2bb_[cont];
-    if (cont->is_basicblock())
-        bb.head = def_to_decl(cont);
-    else
+    if (cont->is_basicblock()) {
+        NodeVec params;
+
+        for (size_t i = 0; i < cont->num_params(); i++) {
+            auto type = convert(cont->param(i)->type());
+            if (!type) continue; // Eliminate mem tokens
+            shady::QualifiedType qtype;
+            qtype.type = type;
+            qtype.is_uniform = false;
+            type = shady::qualified_type(arena, qtype);
+            auto param = shady::var(arena, type, cont->param(i)->name().c_str());
+            defs_[cont->param(i)] = param; // Register the param as emitted already
+            params.push_back(param);
+        }
+
+        bb.head = shady::basic_block(arena, curr_fn, vec2nodes(params), cont->name().c_str());
+    } else
         assert(bb.head);
 
-    // Register params
-    // for (size_t i = 0; i < cont->num_params(); i++)
-    //     defs_[cont->param(i)] = bb.head->payload.fn.params.nodes[i];
+    bb.builder = shady::begin_body(arena);
+}
 
-    bb.builder = shady::begin_block(arena);
+static std::optional<shady::Op> is_shady_prim_op(const Continuation* cont) {
+    for (int i = 0; i < shady::PRIMOPS_COUNT; i++) {
+        if (cont->name() == shady::primop_names[i])
+            return std::make_optional((shady::Op) i);
+    }
+    return std::nullopt;
 }
 
+static std::vector<const shady::Node*> emit_instruction(shady::BodyBuilder* builder, const shady::Node* instruction);
+
 void CodeGen::emit_epilogue(Continuation* cont) {
     BB& bb = cont2bb_[cont];
     assert(cont->has_body());
@@ -236,10 +246,11 @@ void CodeGen::emit_epilogue(Continuation* cont) {
     for (auto& arg : body->args()) {
         if (convert(arg->type()) == nullptr) {
             args.push_back(nullptr);
-        } else if (auto callee = arg->isa_nom<Continuation>(); callee && callee->is_basicblock()) {
+        } else if (auto target = arg->isa_nom<Continuation>(); target && target->is_basicblock()) {
             // Emitting basic blocks as values isn't legal - but for convenience we'll put them in our list.
-            BB& callee_bb = cont2bb_[callee];
-            assert(callee_bb.head && callee_bb.head->payload.fn.is_basic_block);
+            assert(cont2bb_.contains(target));
+            BB& callee_bb = cont2bb_[target];
+            assert(callee_bb.head && shady::is_basic_block(callee_bb.head));
             args.push_back(callee_bb.head);
         } else {
             args.push_back(emit(arg));
@@ -250,11 +261,10 @@ void CodeGen::emit_epilogue(Continuation* cont) {
         shady::Return payload = {};
         payload.fn = curr_fn;
         args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
-        payload.values = vec2nodes(args);
+        payload.args = vec2nodes(args);
         bb.terminator = shady::fn_ret(arena, payload);
     } else if (body->callee() == world().branch()) {
         shady::Branch payload = {};
-        payload.branch_mode = shady::Branch::BrIfElse;
         payload.args = shady::nodes(arena, 0, nullptr);
         payload.branch_condition = args[0];
         payload.true_target      = args[1];
@@ -263,24 +273,43 @@ void CodeGen::emit_epilogue(Continuation* cont) {
     } else if (auto match = body->callee()->as_nom<Continuation>(); match && match->intrinsic() == Intrinsic::Match) {
         assert(false);
     } else if (auto destination = body->callee()->isa_nom<Continuation>(); destination && destination->is_basicblock()) {
-        shady::Branch payload = {};
+        shady::Jump payload = {};
         args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
         payload.args = vec2nodes(args);
-        payload.branch_mode = shady::Branch_::BrJump;
-        bb.terminator = shady::branch(arena, payload);
+        payload.target = args[0];
+        bb.terminator = shady::jump(arena, payload);
     } else if (auto intrinsic = body->callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
         assert(false);
     } else if (auto callee = body->callee()->isa_nom<Continuation>()) {
-        shady::Callc payload = {};
         int ret_param = find_return_parameter(callee->type());
         assert(ret_param >= 0);
-        payload.join_at = args[ret_param];
+
         args.erase(args.begin() + ret_param);
         args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
+
+        // shady primop called as imported continuations look like continuation calls to thorin, but not to shady
+        // we just need to carefully emit the primop as an instruction, then jump to the target BB, passing the stuff as we do
+        if (auto op = is_shady_prim_op(callee); op.has_value()) {
+            shady::bind_instruction(bb.builder, shady::prim_op(arena, (shady::PrimOp) {
+                .op = op.value(),
+                .type_arguments = empty(arena),
+                .operands = vec2nodes(args),
+            }));
+            shady::Jump jump;
+            jump.target = args[ret_param];
+            bb.terminator = shady::jump(arena, jump);
+            return;
+        }
+
+        shady::Call payload;
         payload.args = vec2nodes(args);
-        payload.is_return_indirect = false;
         payload.callee = emit(callee);
-        bb.terminator = shady::callc(arena, payload);
+        auto call = shady::call_instr(arena, payload);
+
+        shady::LetInto payload2;
+        payload2.instruction = call;
+        payload2.tail = args[ret_param];
+        bb.terminator = shady::let_into(arena, payload2);
     } else {
         assert(false);
     }
@@ -289,13 +318,17 @@ void CodeGen::emit_epilogue(Continuation* cont) {
 void CodeGen::finalize(Continuation* cont) {
     BB& bb = cont2bb_[cont];
     assert(bb.head && bb.builder && bb.terminator);
-    bb.block = shady::finish_block(bb.builder, bb.terminator);
-    bb.head->payload.fn.block = bb.block;
+    if (shady::is_basic_block(bb.head))
+        bb.head->payload.basic_block.body = shady::finish_body(bb.builder, bb.terminator);
+    else if (shady::is_function(bb.head))
+        bb.head->payload.fun.body = shady::finish_body(bb.builder, bb.terminator);
+    else
+        assert(false);
 }
 
 void CodeGen::finalize(const Scope& scope) {
     BB& bb = cont2bb_[scope.entry()];
-    assert(bb.head->payload.fn.block != nullptr);
+    assert(bb.head->payload.fun.body != nullptr);
     curr_fn = nullptr;
 }
 
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
index bd4c8026b..c918f4fe5 100644
--- a/src/thorin/be/shady/shady.h
+++ b/src/thorin/be/shady/shady.h
@@ -15,9 +15,8 @@ namespace thorin::shady_be {
 struct BB {
     /// For an entry BB, this will also be the head of the entire function
     shady::Node* head;
-    shady::BlockBuilder* builder;
+    shady::BodyBuilder* builder;
     const shady::Node* terminator;
-    const shady::Node* block;
 };
 
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node*, const shady::Type*, BB, CodeGen> {
@@ -44,7 +43,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node
     const shady::Node* emit_fun_decl(Continuation*);
 protected:
     shady::AddressSpace convert_address_space(AddrSpace);
-    shady::Node* def_to_decl(Def*);
+    shady::Node* emit_decl_head(Def*);
     shady::Node* get_decl(Def*);
 
     using NodeVec = std::vector<const shady::Node*>;
@@ -54,7 +53,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node
     }
 
     shady::IrArena* arena = nullptr;
-    std::vector<shady::Node*> top_level;
+    shady::Module* module = nullptr;
+    //std::vector<shady::Node*> top_level;
 
     shady::Node* curr_fn;
 

From 9afa2d4114229e253c7dc00d639afaa91511b1e5 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 9 Nov 2022 14:31:24 +0100
Subject: [PATCH 113/342] Update closure conversion to maintain a "converted"
 mapping for already converted definitions.

---
 src/thorin/analyses/verify.cpp              | 16 +++--
 src/thorin/transform/closure_conversion.cpp | 80 +++++++++++++++------
 src/thorin/type.cpp                         |  2 +-
 src/thorin/world.cpp                        |  3 +
 4 files changed, 70 insertions(+), 31 deletions(-)

diff --git a/src/thorin/analyses/verify.cpp b/src/thorin/analyses/verify.cpp
index bbd034e8c..f68af86fe 100644
--- a/src/thorin/analyses/verify.cpp
+++ b/src/thorin/analyses/verify.cpp
@@ -30,16 +30,17 @@ static bool verify_top_level(World& world) {
     return ok;
 }
 
-static bool verify_mem(World& world) {
+static bool verify_param(World& world) {
     bool ok = true;
-    Scope::for_each(world, [&] (const Scope& scope) {
-        for (auto def : free_defs(scope)) {
-            if (is_mem(def)) {
-                world.ELOG("scope '{}' got free mem '{}' with {} uses", scope.entry(), def, def->num_uses());
+    for (auto def : world.defs()) {
+        if (auto param = def->isa<Param>()) {
+            auto cont = param->continuation();
+            if (cont->dead_) {
+                world.ELOG("param '{}' originates in dead continuation {}", param, cont);
                 ok = false;
             }
         }
-    });
+    }
     return ok;
 }
 
@@ -47,9 +48,10 @@ void verify(World& world) {
     bool ok = true;
     ok &= verify_calls(world);
     ok &= verify_top_level(world);
-    ok &= verify_mem(world);
+    ok &= verify_param(world);
     if (!ok)
         world.dump();
+    assert(ok);
 }
 
 }
diff --git a/src/thorin/transform/closure_conversion.cpp b/src/thorin/transform/closure_conversion.cpp
index bb118c725..7d9250589 100644
--- a/src/thorin/transform/closure_conversion.cpp
+++ b/src/thorin/transform/closure_conversion.cpp
@@ -26,27 +26,28 @@ class ClosureConversion {
 
             auto new_type = world_.fn_type(convert(continuation->type())->ops());
             if (new_type != continuation->type()) {
+                //The function type was changed, so the continuation takes another function as a parameter.
                 auto new_continuation = world_.continuation(new_type->as<FnType>(), continuation->debug());
                 if (continuation->is_intrinsic())
                     new_continuation->set_intrinsic();
 
+                converted.emplace_back(continuation, new_continuation);
+                std::cerr << "Creating new conversion: " << continuation->unique_name() << " to " << new_continuation->unique_name() << "\n";
+                continuation->type()->dump();
+                new_continuation->type()->dump();
                 new_defs_[continuation] = new_continuation;
-                if (continuation->has_body()) {
-                    auto body = continuation->body();
-                    for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
-                        new_defs_[continuation->param(i)] = new_continuation->param(i);
-                    // copy existing call from old continuation
-                    new_continuation->jump(body->callee(), body->args(), continuation->debug());
-                    converted.emplace_back(continuation, new_continuation);
-                }
-            } else if (continuation->has_body()) {
+                for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
+                    new_defs_[continuation->param(i)] = new_continuation->param(i);
+            } else {
+                //The type remains unchanged, do not generate a new continuation.
+                //We still need to add the continuation to converted, to ensure the jump will be converted later on.
                 converted.emplace_back(continuation, continuation);
             }
         }
 
         // convert the calls to each continuation
         for (auto pair : converted)
-            convert_jump(pair.second);
+            convert_jump(pair.first, pair.second);
 
         // remove old continuations
         for (auto pair : converted) {
@@ -57,31 +58,59 @@ class ClosureConversion {
         }
     }
 
-    void convert_jump(Continuation* continuation) {
-        assert(continuation->has_body());
-        auto body = continuation->body();
-        // prevent conversion of calls to vectorize() or cuda(), but allow graph intrinsics
+    //Convert jump and all arguments.
+    void convert_jump(Continuation* source, Continuation* target) {
+        assert(source);
+        assert(target);
+
+        if (!converted_.emplace(source).second)
+            return; //Was already converted once before.
+
+        assert(source->has_body());
+        auto body = source->body();
+
         auto callee = body->callee()->isa_nom<Continuation>();
-        if (callee == continuation) return;
+
+        if (callee == source) {
+            target->jump(callee, body->args(), source->debug());
+            return;
+        }
+
+        // prevent conversion of calls to vectorize() or cuda(), but allow graph intrinsics
         if (!callee || !callee->is_intrinsic()) {
             Array<const Def*> new_args(body->num_args());
             for (size_t i = 0, e = body->num_args(); i != e; ++i)
                 new_args[i] = convert(body->arg(i));
-            continuation->jump(convert(body->callee(), true), new_args, continuation->debug());
+            target->jump(convert(body->callee(), true), new_args, source->debug());
+        } else {
+            Array<const Def*> new_args(body->num_args());
+            for (size_t i = 0, e = body->num_args(); i != e; ++i) {
+                if (body->arg(i)->type()->isa<FnType>())
+                    new_args[i] = body->arg(i);
+                else if (callee->intrinsic() == Intrinsic::Match && i > 2)
+                    new_args[i] = body->arg(i);
+                else
+                    new_args[i] = convert(body->arg(i));
+            }
+            target->jump(callee, new_args, source->debug());
         }
     }
 
     const Def* convert(const Def* def, bool as_callee = false) {
-        if (new_defs_.count(def)) def = new_defs_[def];
-        if (def->order() <= 1)
-            return def;
+        if (auto * source = def->isa_nom<Continuation>()) {
+            if (new_defs_.count(def)) def = new_defs_[def];
+            auto continuation = def->isa_nom<Continuation>();
+            assert(continuation);
 
-        if (auto continuation = def->isa_nom<Continuation>()) {
-            if (!continuation->has_body())
-                return continuation;
-            convert_jump(continuation);
             if (as_callee)
                 return continuation;
+            if (!source->has_body())
+                return continuation;
+            if (continuation->order() <= 1)
+                return continuation;
+
+            convert_jump(source, continuation); //convert_jump must be executed so that continuation has a body.
+            assert(continuation->has_body());
 
             world_.WLOG("slow: closure generated for '{}'", continuation);
 
@@ -148,6 +177,10 @@ class ClosureConversion {
             auto closure_type = convert(continuation->type());
             return world_.closure(closure_type->as<ClosureType>(), wrapper, thin_env ? free_vars[0] : world_.tuple(free_vars), continuation->debug());
         } else {
+            if (new_defs_.count(def)) return new_defs_[def];
+            if (def->isa<Param>() || def->isa<Closure>())
+                return def;
+
             // TODO need to consider Params?
             Array<const Def*> ops(def->ops());
             for (auto& op : ops) op = convert(op);
@@ -201,6 +234,7 @@ class ClosureConversion {
     World& world_;
     Def2Def new_defs_;
     Type2Type new_types_;
+    ContinuationSet converted_;
 };
 
 
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index 7918b0e45..9b8f8b49f 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -199,7 +199,7 @@ TypeTable::TypeTable()
 }
 
 const Type* TypeTable::tuple_type(Types ops) {
-    return ops.size() == 1 ? ops.front() : insert<TupleType>(*this, ops);
+    return (ops.size() == 1 && is_thin(ops[0])) ? ops.front() : insert<TupleType>(*this, ops);
 }
 
 const StructType* TypeTable::struct_type(Symbol name, size_t size) {
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index d2c9d17a2..f465fba27 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -1299,10 +1299,13 @@ void World::opt() {
 
     RUN_PASS(cleanup())
     RUN_PASS(while (partial_evaluation(*this, true))); // lower2cff
+    RUN_PASS(verify(*this));
     RUN_PASS(flatten_tuples(*this))
     RUN_PASS(clone_bodies(*this))
     RUN_PASS(split_slots(*this))
+    RUN_PASS(verify(*this));
     RUN_PASS(closure_conversion(*this))
+    RUN_PASS(verify(*this));
     RUN_PASS(lift_builtins(*this))
     RUN_PASS(inliner(*this))
     RUN_PASS(hoist_enters(*this))

From b14f556af512c65b00ccd8c2edf472d4a64a3c06 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 16 Sep 2022 16:02:57 +0200
Subject: [PATCH 114/342] Initialize HashTable array before using it.

---
 src/thorin/util/hash.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/util/hash.h b/src/thorin/util/hash.h
index 32bf6d6bd..007c7fbb4 100644
--- a/src/thorin/util/hash.h
+++ b/src/thorin/util/hash.h
@@ -217,6 +217,7 @@ class HashTable {
     HashTable()
         : capacity_(StackCapacity)
         , size_(0)
+        , array_()
         , nodes_(array_.data())
 #if THORIN_ENABLE_CHECKS
         , id_(0)

From 1476d1d6a8084fbe6ea70f462c76673570805905 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 9 Nov 2022 15:44:26 +0100
Subject: [PATCH 115/342] Add json output backend module. No actual content
 generated for now.

---
 CMakeLists.txt               |  6 ++++++
 cmake/thorin-config.cmake.in |  1 +
 src/thorin/CMakeLists.txt    | 11 ++++++++++
 src/thorin/be/json/json.cpp  | 10 +++++++++
 src/thorin/be/json/json.h    | 39 ++++++++++++++++++++++++++++++++++++
 src/thorin/config.h.in       |  1 +
 6 files changed, 68 insertions(+)
 create mode 100644 src/thorin/be/json/json.cpp
 create mode 100644 src/thorin/be/json/json.h

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 60c708319..ea48e171c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -26,6 +26,12 @@ list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/modules")
 find_package(Half REQUIRED)
 message(STATUS "Building with Half library from ${Half_INCLUDE_DIRS}.")
 
+# find json package for json output support.
+find_package(nlohmann_json 3.2.0)
+if(nlohmann_json_FOUND)
+    set(THORIN_ENABLE_JSON TRUE)
+endif()
+
 # check for possible llvm extension
 find_package(LLVM QUIET CONFIG)
 if(LLVM_FOUND)
diff --git a/cmake/thorin-config.cmake.in b/cmake/thorin-config.cmake.in
index c279f3127..b5fcc4719 100644
--- a/cmake/thorin-config.cmake.in
+++ b/cmake/thorin-config.cmake.in
@@ -28,6 +28,7 @@ list(APPEND CMAKE_MODULE_PATH "${Thorin_ROOT_DIR}/cmake/modules")
 find_path(Half_DIR NAMES half.hpp PATHS ${Half_DIR} $ENV{Half_DIR} "@Half_DIR@" "@Half_INCLUDE_DIR@")
 find_package(Half REQUIRED)
 
+set(Thorin_HAS_JSON_SUPPORT @nlohmann_json_FOUND@)
 set(Thorin_HAS_LLVM_SUPPORT @LLVM_FOUND@)
 set(Thorin_HAS_RV_SUPPORT @RV_FOUND@)
 set(AnyDSL_LLVM_LINK_SHARED @AnyDSL_LLVM_LINK_SHARED@)
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 944c785d2..ba6a2a106 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -107,6 +107,13 @@ if(LLVM_FOUND)
     )
 endif()
 
+if(nlohmann_json_FOUND)
+    list(APPEND THORIN_SOURCES
+        be/json/json.cpp
+        be/json/json.h
+    )
+endif()
+
 add_library(thorin ${THORIN_SOURCES})
 target_include_directories(thorin PUBLIC ${Half_INCLUDE_DIRS} ${Thorin_ROOT_DIR}/src ${CMAKE_BINARY_DIR}/include)
 
@@ -121,3 +128,7 @@ if(LLVM_FOUND)
     endif()
     llvm_config(thorin ${AnyDSL_LLVM_LINK_SHARED} ${Thorin_LLVM_COMPONENTS})
 endif()
+
+if(nlohmann_json_FOUND)
+    target_link_libraries(thorin PRIVATE nlohmann_json::nlohmann_json)
+endif()
diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
new file mode 100644
index 000000000..a8d1b0f0d
--- /dev/null
+++ b/src/thorin/be/json/json.cpp
@@ -0,0 +1,10 @@
+#include "json.h"
+
+namespace thorin::json {
+
+void CodeGen::emit_stream(std::ostream& stream) {
+    Stream s(stream);
+    s << "Currently not implemented\n";
+}
+
+}
diff --git a/src/thorin/be/json/json.h b/src/thorin/be/json/json.h
new file mode 100644
index 000000000..ee221ea96
--- /dev/null
+++ b/src/thorin/be/json/json.h
@@ -0,0 +1,39 @@
+#ifndef THORIN_BE_JSON_H
+#define THORIN_BE_JSON_H
+
+#include <cstdint>
+#include <iostream>
+#include <nlohmann/json.hpp>
+
+#include "thorin/be/codegen.h"
+
+namespace thorin {
+
+class World;
+
+namespace json {
+
+using json = nlohmann::json;
+
+class CodeGen : public thorin::CodeGen {
+public:
+    CodeGen(World& world, const Cont2Config& kernel_config, bool debug)
+        : thorin::CodeGen(world, debug)
+        , kernel_config_(kernel_config)
+    {}
+    
+    void emit_stream(std::ostream& stream) override;
+
+    const char* file_ext() const override {
+        return ".thorin.json";
+    }
+
+private:
+    const Cont2Config& kernel_config_;
+};
+
+}
+
+}
+
+#endif
diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index 5e061fe28..9216ecd82 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -4,6 +4,7 @@
 #cmakedefine01 THORIN_ENABLE_CHECKS
 #cmakedefine01 THORIN_ENABLE_PROFILING
 #cmakedefine01 THORIN_ENABLE_LLVM
+#cmakedefine01 THORIN_ENABLE_JSON
 #cmakedefine01 THORIN_ENABLE_RV
 
 #endif

From cc290203bfa5bea826e86237b9b502198d0017c5 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 10 Nov 2022 17:19:13 +0100
Subject: [PATCH 116/342] Emit json files for some simple cases.

---
 src/thorin/be/json/json.cpp | 140 +++++++++++++++++++++++++++++++++++-
 1 file changed, 139 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index a8d1b0f0d..6cb8eb346 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -2,9 +2,147 @@
 
 namespace thorin::json {
 
+class TypeTable {
+public:
+    json type_table = json::array();
+
+    TypeMap<std::string> known_types;
+
+    std::string translate_type (const Type * type) {
+        auto it = known_types.find(type);
+        if (it != known_types.end()) {
+            return it->second;
+        }
+
+        json result;
+        if (type->isa<MemType>()) {
+            result["name"] = "mem_t";
+            result["type"] = "mem";
+        } else if (auto prim = type->isa<PrimType>()) {
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["length"] = prim->length();
+            result["type"] = "prim";
+            switch (prim->primtype_tag()) {
+#define THORIN_ALL_TYPE(T, M) case PrimTypeTag::PrimType_##T: { result["tag"] = #T; break; }
+#include <thorin/tables/primtypetable.h>
+            }
+        } else if (auto fntype = type->isa<FnType>()) {
+            json arg_types = json::array();
+            for (auto arg : fntype->ops()) {
+                arg_types.push_back(translate_type(arg));
+            }
+
+            result["type"] = "fn";
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["args"] = arg_types;
+        } else if (auto ptrtype = type->isa<PtrType>()) {
+            auto pointee_type = translate_type(ptrtype->pointee());
+
+            result["type"] = "ptr";
+            result["args"] = { pointee_type };
+            result["name"] = pointee_type + "_p";
+            result["length"] = ptrtype->length();
+        } else if (auto arr = type->isa<IndefiniteArrayType>()) {
+            auto elem_type = translate_type(arr->elem_type());
+
+            result["type"] = "indef_array";
+            result["args"] = { elem_type };
+            result["name"] = elem_type + "_iarr";
+        } else {
+            THORIN_UNREACHABLE;
+        }
+        known_types[type] = result["name"];
+        type_table.push_back(result);
+        return result["name"];
+    }
+};
+
+class DefTable {
+public:
+    DefTable(TypeTable& type_table) : type_table_(type_table) {}
+
+    json decl_table = json::array();
+    json def_table = json::array();
+    TypeTable& type_table_;
+
+    DefMap<std::string> known_defs;
+
+    std::string translate_def (const Def * def, std::string expected_name = "") {
+        json result;
+        if (auto cont = def->isa<Continuation>()) {
+            auto type = type_table_.translate_type(def->type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            json arg_names;
+            for (auto arg : cont->params()) {
+                arg_names.push_back(translate_def(arg));
+            }
+
+            json forward_decl;
+            forward_decl["name"] = name;
+            forward_decl["type"] = "continuation";
+            forward_decl["fn_type"] = type;
+            forward_decl["arg_names"] = arg_names;
+            forward_decl["external"] = cont->is_external();
+            decl_table.push_back(forward_decl); //TODO: should be pushed to the front instead.
+
+            assert(cont->has_body());
+            auto app = cont->body();
+            auto target = translate_def(app->callee());
+            json args = json::array();
+            for (auto arg : app->args()) {
+                args.push_back(translate_def(arg));
+            }
+
+            result["name"] = name;
+            result["type"] = "continuation";
+            result["app"] = {
+                {"target", target},
+                {"args", args}
+            };
+        } else if (auto lit = def->isa<PrimLit>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto type = type_table_.translate_type(def->type());
+
+            result["name"] = name;
+            result["type"] = "const";
+            result["const_type"] = type;
+            result["value"] = lit->value().get_s32(); //TODO: this looks wrong. What I get should depend on the lit type.
+        } else if (auto param = def->isa<Param>()) {
+            auto name = expected_name != "" ? expected_name : param->continuation()->unique_name() + "." + std::to_string(param->index());
+            known_defs[def] = name;
+            return name;
+        } else {
+            THORIN_UNREACHABLE;
+        }
+        known_defs[def] = result["name"];
+        def_table.push_back(result);
+        return result["name"];
+    }
+};
+
 void CodeGen::emit_stream(std::ostream& stream) {
+    json j;
+
+    j["module"] = world().name();
+
+    TypeTable type_table;
+    DefTable def_table(type_table);
+
+    for (auto external : world().externals()) {
+        const Continuation* continuation = external.second;
+        auto expected_name = continuation->name();
+        def_table.translate_def(continuation, expected_name);
+    }
+
+    j["type_table"] = type_table.type_table;
+    j["defs"] = def_table.decl_table;
+    for (auto it : def_table.def_table)
+        j["defs"] += it;
+    
+    std::cerr << j.dump(2) << std::endl;
+
     Stream s(stream);
-    s << "Currently not implemented\n";
+    s << j.dump(2) << "\n";
 }
 
 }

From 373a6e0e562ac10477d90f9fcc5b37117ee2f203 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 10 Nov 2022 18:00:08 +0100
Subject: [PATCH 117/342] [JSON]: Convert enought defs to convert a simple test
 program.

---
 src/thorin/be/json/json.cpp | 47 +++++++++++++++++++++++++++++++++++++
 1 file changed, 47 insertions(+)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 6cb8eb346..7eadba442 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -111,6 +111,53 @@ class DefTable {
             auto name = expected_name != "" ? expected_name : param->continuation()->unique_name() + "." + std::to_string(param->index());
             known_defs[def] = name;
             return name;
+        } else if (auto load = def->isa<Load>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            json args = json::array();
+            args.push_back(translate_def(load->mem()));
+            args.push_back(translate_def(load->ptr()));
+
+            result["name"] = name;
+            result["type"] = "load";
+            result["args"] = args;
+        } else if (auto cast = def->isa<Cast>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto source = translate_def(cast->from());
+            auto target_type = type_table_.translate_type(cast->type());
+
+            result["name"] = name;
+            result["type"] = "cast";
+            result["source"] = source;
+            result["target_type"] = target_type;
+        } else if (auto lea = def->isa<LEA>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            json args = json::array();
+            args.push_back(translate_def(lea->ptr()));
+            args.push_back(translate_def(lea->index()));
+
+            result["name"] = name;
+            result["type"] = "lea";
+            result["args"] = args;
+        } else if (auto extract = def->isa<Extract>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            json args = json::array();
+            args.push_back(translate_def(extract->agg()));
+            args.push_back(translate_def(extract->index()));
+
+            result["name"] = name;
+            result["type"] = "extract";
+            result["args"] = args;
+        } else if (auto arithop = def->isa<ArithOp>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto op = arithop->op_name();
+            json args = json::array();
+            args.push_back(translate_def(arithop->lhs()));
+            args.push_back(translate_def(arithop->rhs()));
+
+            result["name"] = name;
+            result["type"] = "arithop";
+            result["op"] = op;
+            result["args"] = args;
         } else {
             THORIN_UNREACHABLE;
         }

From 841c2086d37f627e23a0139f31ef82f6d1d3910b Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 14 Nov 2022 15:18:40 +0100
Subject: [PATCH 118/342] [Json] various improvements:

  * Avoid name collisions, generate names after args have been created.
  * Querry known_defs to avoid multiple instances of the same definition.
  * Deal with intrinsics. Limited to branch and named intrinsics for now.
  * Create Cmp defs.
---
 src/thorin/be/json/json.cpp | 94 +++++++++++++++++++++++--------------
 1 file changed, 59 insertions(+), 35 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 7eadba442..0cd4837a6 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -68,37 +68,52 @@ class DefTable {
     DefMap<std::string> known_defs;
 
     std::string translate_def (const Def * def, std::string expected_name = "") {
+        auto it = known_defs.find(def);
+        if (it != known_defs.end()) {
+            return it->second;
+        }
+
         json result;
         if (auto cont = def->isa<Continuation>()) {
-            auto type = type_table_.translate_type(def->type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
-            json arg_names;
-            for (auto arg : cont->params()) {
-                arg_names.push_back(translate_def(arg));
-            }
-
-            json forward_decl;
-            forward_decl["name"] = name;
-            forward_decl["type"] = "continuation";
-            forward_decl["fn_type"] = type;
-            forward_decl["arg_names"] = arg_names;
-            forward_decl["external"] = cont->is_external();
-            decl_table.push_back(forward_decl); //TODO: should be pushed to the front instead.
-
-            assert(cont->has_body());
-            auto app = cont->body();
-            auto target = translate_def(app->callee());
-            json args = json::array();
-            for (auto arg : app->args()) {
-                args.push_back(translate_def(arg));
+            if (cont->is_intrinsic()) {
+                assert(cont->intrinsic() == Intrinsic::Branch && "TODO: anything else is unsupported RN");
+
+                result["name"] = "branch";
+                result["type"] = "continuation";
+                result["intrinsic"] = "branch";
+            } else {
+                assert(cont->has_body());
+
+                auto type = type_table_.translate_type(def->type());
+                json arg_names = json::array();
+                for (auto arg : cont->params()) {
+                    arg_names.push_back(translate_def(arg));
+                }
+
+                auto name = expected_name != "" ? expected_name : "_cont_" + std::to_string(decl_table.size());
+
+                json forward_decl;
+                forward_decl["name"] = name;
+                forward_decl["type"] = "continuation";
+                forward_decl["fn_type"] = type;
+                forward_decl["arg_names"] = arg_names;
+                forward_decl["external"] = cont->is_external();
+                decl_table.push_back(forward_decl);
+
+                auto app = cont->body();
+                auto target = translate_def(app->callee());
+                json args = json::array();
+                for (auto arg : app->args()) {
+                    args.push_back(translate_def(arg));
+                }
+
+                result["name"] = name;
+                result["type"] = "continuation";
+                result["app"] = {
+                    {"target", target},
+                    {"args", args}
+                };
             }
-
-            result["name"] = name;
-            result["type"] = "continuation";
-            result["app"] = {
-                {"target", target},
-                {"args", args}
-            };
         } else if (auto lit = def->isa<PrimLit>()) {
             auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             auto type = type_table_.translate_type(def->type());
@@ -112,52 +127,63 @@ class DefTable {
             known_defs[def] = name;
             return name;
         } else if (auto load = def->isa<Load>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             json args = json::array();
             args.push_back(translate_def(load->mem()));
             args.push_back(translate_def(load->ptr()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "load";
             result["args"] = args;
         } else if (auto cast = def->isa<Cast>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             auto source = translate_def(cast->from());
             auto target_type = type_table_.translate_type(cast->type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "cast";
             result["source"] = source;
             result["target_type"] = target_type;
         } else if (auto lea = def->isa<LEA>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             json args = json::array();
             args.push_back(translate_def(lea->ptr()));
             args.push_back(translate_def(lea->index()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "lea";
             result["args"] = args;
         } else if (auto extract = def->isa<Extract>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             json args = json::array();
             args.push_back(translate_def(extract->agg()));
             args.push_back(translate_def(extract->index()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "extract";
             result["args"] = args;
         } else if (auto arithop = def->isa<ArithOp>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             auto op = arithop->op_name();
             json args = json::array();
             args.push_back(translate_def(arithop->lhs()));
             args.push_back(translate_def(arithop->rhs()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "arithop";
             result["op"] = op;
             result["args"] = args;
+        } else if (auto cmp = def->isa<Cmp>()) {
+            auto op = cmp->op_name();
+            json args = json::array();
+            args.push_back(translate_def(cmp->lhs()));
+            args.push_back(translate_def(cmp->rhs()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "cmp";
+            result["op"] = op;
+            result["args"] = args;
         } else {
             THORIN_UNREACHABLE;
         }
@@ -186,8 +212,6 @@ void CodeGen::emit_stream(std::ostream& stream) {
     for (auto it : def_table.def_table)
         j["defs"] += it;
     
-    std::cerr << j.dump(2) << std::endl;
-
     Stream s(stream);
     s << j.dump(2) << "\n";
 }

From acb6e139430a6ff535f1f6ab8ce55283339d2ed7 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 15 Nov 2022 19:29:03 +0100
Subject: [PATCH 119/342] [Json] create additional defs and types:

  * Run
  * Hlt
  * Store
  * Enter
  * Slot
  * Definite Array Type

Fix runaway recursion on continuation self refference.
---
 src/thorin/be/json/json.cpp | 64 +++++++++++++++++++++++++++++++++++++
 1 file changed, 64 insertions(+)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 0cd4837a6..0b9e42429 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -48,7 +48,16 @@ class TypeTable {
             result["type"] = "indef_array";
             result["args"] = { elem_type };
             result["name"] = elem_type + "_iarr";
+        } else if (auto arr = type->isa<DefiniteArrayType>()) {
+            auto elem_type = translate_type(arr->elem_type());
+
+            result["type"] = "def_array";
+            result["args"] = { elem_type };
+            result["length"] = arr->dim();
+            result["name"] = elem_type + "_iarr";
         } else {
+            std::cerr << "type cannot be translated\n";
+            type->dump();
             THORIN_UNREACHABLE;
         }
         known_types[type] = result["name"];
@@ -81,6 +90,15 @@ class DefTable {
                 result["name"] = "branch";
                 result["type"] = "continuation";
                 result["intrinsic"] = "branch";
+            } else if (cont->is_imported()) {
+                auto name = cont->name();
+                auto type = type_table_.translate_type(def->type());
+
+                result["name"] = name;
+                result["type"] = "continuation";
+                result["fn_type"] = type;
+                result["imported"] = true;
+                result["external"] = cont->is_external();
             } else {
                 assert(cont->has_body());
 
@@ -100,6 +118,8 @@ class DefTable {
                 forward_decl["external"] = cont->is_external();
                 decl_table.push_back(forward_decl);
 
+                known_defs[def] = name;
+
                 auto app = cont->body();
                 auto target = translate_def(app->callee());
                 json args = json::array();
@@ -135,6 +155,16 @@ class DefTable {
             result["name"] = name;
             result["type"] = "load";
             result["args"] = args;
+        } else if (auto store = def->isa<Store>()) {
+            json args = json::array();
+            args.push_back(translate_def(store->mem()));
+            args.push_back(translate_def(store->ptr()));
+            args.push_back(translate_def(store->val()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "store";
+            result["args"] = args;
         } else if (auto cast = def->isa<Cast>()) {
             auto source = translate_def(cast->from());
             auto target_type = type_table_.translate_type(cast->type());
@@ -184,7 +214,41 @@ class DefTable {
             result["type"] = "cmp";
             result["op"] = op;
             result["args"] = args;
+        } else if (auto run = def->isa<Run>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto target = translate_def(run->def());
+
+            result["name"] = name;
+            result["type"] = "run";
+            result["target"] = target;
+        } else if (auto hlt = def->isa<Hlt>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            json args = json::array();
+            auto target = translate_def(hlt->def());
+
+            result["name"] = name;
+            result["type"] = "hlt";
+            result["target"] = target;
+        } else if (auto enter = def->isa<Enter>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            json args = json::array();
+            auto mem = translate_def(enter->mem());
+
+            result["name"] = name;
+            result["type"] = "enter";
+            result["mem"] = mem;
+        } else if (auto slot = def->isa<Slot>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto frame = translate_def(slot->frame());
+            auto target_type = type_table_.translate_type(slot->alloced_type());
+
+            result["name"] = name;
+            result["type"] = "slot";
+            result["frame"] = frame;
+            result["target_type"] = target_type;
         } else {
+            def->dump(2);
+            std::cerr << "cannot be translated\n";
             THORIN_UNREACHABLE;
         }
         known_defs[def] = result["name"];

From 6c55fac7d5e0813190eba8da7c7e43c21a3cc471 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 17 Nov 2022 13:15:15 +0100
Subject: [PATCH 120/342] [Json] More defs added:

  * Insert
  * Bitcast
  * Global
  * Aggregates: Closure, StructAgg, Tuple, Vector
  * Filter
  * Known
  * Alloc
  * Select
  * Top, Bottom
  * SizeOf, AlignOf
  * DefiniteArray, IndefiniteArray
---
 src/thorin/be/json/json.cpp | 182 ++++++++++++++++++++++++++++++++++--
 1 file changed, 174 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 0b9e42429..e28531161 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -136,12 +136,34 @@ class DefTable {
             }
         } else if (auto lit = def->isa<PrimLit>()) {
             auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
-            auto type = type_table_.translate_type(def->type());
+            auto type = type_table_.translate_type(lit->type());
 
             result["name"] = name;
             result["type"] = "const";
             result["const_type"] = type;
-            result["value"] = lit->value().get_s32(); //TODO: this looks wrong. What I get should depend on the lit type.
+            //result["value"] = lit->value().get_s32(); //TODO: this looks wrong. What I get should depend on the lit type.
+            switch (lit->primtype_tag()) {
+#define THORIN_I_TYPE(T, M) case PrimType_##T: { result["value"] = lit->value().get_##M(); break; }
+#define THORIN_BOOL_TYPE(T, M) case PrimType_##T: { result["value"] = lit->value().get_##M(); break; }
+#define THORIN_F_TYPE(T, M) case PrimType_##T: { result["value"] = (double)lit->value().get_##M(); break; }
+#include <thorin/tables/primtypetable.h>
+            default:
+                assert(false && "not implemented");
+            }
+        } else if (def->isa<Top>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto type = type_table_.translate_type(def->type());
+
+            result["name"] = name;
+            result["type"] = "top";
+            result["const_type"] = type;
+        } else if (def->isa<Bottom>()) {
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto type = type_table_.translate_type(def->type());
+
+            result["name"] = name;
+            result["type"] = "bottom";
+            result["const_type"] = type;
         } else if (auto param = def->isa<Param>()) {
             auto name = expected_name != "" ? expected_name : param->continuation()->unique_name() + "." + std::to_string(param->index());
             known_defs[def] = name;
@@ -165,6 +187,20 @@ class DefTable {
             result["name"] = name;
             result["type"] = "store";
             result["args"] = args;
+        } else if (auto size_of = def->isa<SizeOf>()) {
+            auto target_type = type_table_.translate_type(size_of->of());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "sizeof";
+            result["target_type"] = target_type;
+        } else if (auto align_of = def->isa<AlignOf>()) {
+            auto target_type = type_table_.translate_type(align_of->of());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "alignof";
+            result["target_type"] = target_type;
         } else if (auto cast = def->isa<Cast>()) {
             auto source = translate_def(cast->from());
             auto target_type = type_table_.translate_type(cast->type());
@@ -174,6 +210,37 @@ class DefTable {
             result["type"] = "cast";
             result["source"] = source;
             result["target_type"] = target_type;
+         } else if (auto bitcast = def->isa<Bitcast>()) {
+            auto source = translate_def(bitcast->from());
+            auto target_type = type_table_.translate_type(bitcast->type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "bitcast";
+            result["source"] = source;
+            result["target_type"] = target_type;
+        } else if (auto indef_array = def->isa<IndefiniteArray>()) {
+            auto dim = translate_def(indef_array->op(0));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto element_type = type_table_.translate_type(indef_array->elem_type());
+
+            result["name"] = name;
+            result["type"] = "indef_array";
+            result["elem_type"] = element_type;
+            result["dim"] = dim;
+        } else if (auto def_array = def->isa<DefiniteArray>()) {
+            json args = json::array();
+            for (auto arg : def_array->ops()) {
+                args.push_back(translate_def(arg));
+            }
+
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto element_type = type_table_.translate_type(def_array->elem_type());
+
+            result["name"] = name;
+            result["type"] = "def_array";
+            result["elem_type"] = element_type;
+            result["args"] = args;
         } else if (auto lea = def->isa<LEA>()) {
             json args = json::array();
             args.push_back(translate_def(lea->ptr()));
@@ -192,6 +259,69 @@ class DefTable {
             result["name"] = name;
             result["type"] = "extract";
             result["args"] = args;
+        } else if (auto insert = def->isa<Insert>()) {
+            json args = json::array();
+            args.push_back(translate_def(insert->agg()));
+            args.push_back(translate_def(insert->index()));
+            args.push_back(translate_def(insert->value()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "insert";
+            result["args"] = args;
+        } else if (auto closure = def->isa<Closure>()) {
+            json args = json::array();
+            args.push_back(translate_def(closure->op(0)));
+            args.push_back(translate_def(closure->op(1)));
+            auto closure_type = type_table_.translate_type(closure->type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "closure";
+            result["args"] = args;
+            result["closure_type"] = closure_type;
+        } else if (auto struct_agg = def->isa<StructAgg>()) {
+            json args = json::array();
+            for (auto arg : struct_agg->ops()) {
+                args.push_back(translate_def(arg));
+            }
+            auto struct_type = type_table_.translate_type(struct_agg->type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "struct";
+            result["args"] = args;
+            result["struct_type"] = struct_type;
+        } else if (auto tuple = def->isa<Tuple>()) {
+            json args = json::array();
+            for (auto arg : tuple->ops()) {
+                args.push_back(translate_def(arg));
+            }
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "tuple";
+            result["args"] = args;
+        } else if (auto vector = def->isa<Vector>()) {
+            json args = json::array();
+            for (auto arg : vector->ops()) {
+                args.push_back(translate_def(arg));
+            }
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "vector";
+            result["args"] = args;
+        } else if (auto filter = def->isa<Filter>()) {
+            json args = json::array();
+            for (auto arg : filter->ops()) {
+                args.push_back(translate_def(arg));
+            }
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "filter";
+            result["args"] = args;
         } else if (auto arithop = def->isa<ArithOp>()) {
             auto op = arithop->op_name();
             json args = json::array();
@@ -203,6 +333,16 @@ class DefTable {
             result["type"] = "arithop";
             result["op"] = op;
             result["args"] = args;
+        } else if (auto select = def->isa<Select>()) {
+            json args = json::array();
+            args.push_back(translate_def(select->cond()));
+            args.push_back(translate_def(select->tval()));
+            args.push_back(translate_def(select->fval()));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "select";
+            result["args"] = args;
         } else if (auto cmp = def->isa<Cmp>()) {
             auto op = cmp->op_name();
             json args = json::array();
@@ -215,38 +355,64 @@ class DefTable {
             result["op"] = op;
             result["args"] = args;
         } else if (auto run = def->isa<Run>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             auto target = translate_def(run->def());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "run";
             result["target"] = target;
         } else if (auto hlt = def->isa<Hlt>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
-            json args = json::array();
             auto target = translate_def(hlt->def());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "hlt";
             result["target"] = target;
-        } else if (auto enter = def->isa<Enter>()) {
+        } else if (auto known = def->isa<Known>()) {
+            auto def = translate_def(known->def());
             auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
-            json args = json::array();
+
+            result["name"] = name;
+            result["type"] = "known";
+            result["def"] = def;
+        } else if (auto enter = def->isa<Enter>()) {
             auto mem = translate_def(enter->mem());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "enter";
             result["mem"] = mem;
         } else if (auto slot = def->isa<Slot>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
             auto frame = translate_def(slot->frame());
             auto target_type = type_table_.translate_type(slot->alloced_type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "slot";
             result["frame"] = frame;
             result["target_type"] = target_type;
+        } else if (auto alloc = def->isa<Alloc>()) {
+            json args = json::array();
+            args.push_back(translate_def(alloc->mem()));
+            args.push_back(translate_def(alloc->extra()));
+            auto target_type = type_table_.translate_type(alloc->alloced_type());
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "alloc";
+            result["args"] = args;
+            result["target_type"] = target_type;
+        } else if (auto global = def->isa<Global>()) {
+            auto init = translate_def(global->init());
+            bool is_mutable = global->is_mutable();
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "global";
+            result["init"] = init;
+            result["mutable"] = is_mutable;
         } else {
+            def->dump();
             def->dump(2);
             std::cerr << "cannot be translated\n";
             THORIN_UNREACHABLE;

From 2397aadd51121519f171ba369acc9c7205cd58f6 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 18 Nov 2022 17:41:13 +0100
Subject: [PATCH 121/342] [Json]: Export all the types! + Additional Defs.

Match, Mathop, Variant, Assembly export added.
Added export for multiple previously unsupported types.
---
 src/thorin/be/json/json.cpp | 206 +++++++++++++++++++++++++++++++-----
 1 file changed, 178 insertions(+), 28 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index e28531161..ff9905558 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -15,9 +15,81 @@ class TypeTable {
         }
 
         json result;
-        if (type->isa<MemType>()) {
+        if (auto arr = type->isa<DefiniteArrayType>()) {
+            auto elem_type = translate_type(arr->elem_type());
+
+            result["type"] = "def_array";
+            result["args"] = { elem_type };
+            result["length"] = arr->dim();
+            result["name"] = elem_type + "_darr";
+        } else if (auto arr = type->isa<IndefiniteArrayType>()) {
+            auto elem_type = translate_type(arr->elem_type());
+
+            result["type"] = "indef_array";
+            result["args"] = { elem_type };
+            result["name"] = elem_type + "_iarr";
+        } else if (type->isa<BottomType>()) {
+            result["name"] = "bottom_t";
+            result["type"] = "bottom";
+        } else if (auto fntype = type->isa<FnType>()) {
+            json arg_types = json::array();
+            for (auto arg : fntype->ops()) {
+                arg_types.push_back(translate_type(arg));
+            }
+
+            result["type"] = "function";
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["args"] = arg_types;
+        } else if (auto closuretype = type->isa<ClosureType>()) {
+            json args = json::array();
+            for (auto arg : closuretype->ops()) {
+                args.push_back(translate_type(arg));
+            }
+
+            result["type"] = "closure";
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["args"] = args;
+        } else if (type->isa<FrameType>()) {
+            result["name"] = "frame_t";
+            result["type"] = "frame";
+        } else if (type->isa<MemType>()) {
             result["name"] = "mem_t";
             result["type"] = "mem";
+        } else if (auto structtype = type->isa<StructType>()) {
+            json args = json::array();
+            json arg_names = json::array();
+            for (size_t i = 0; i < structtype->num_ops(); ++i) {
+                args.push_back(translate_type(structtype->op(i)));
+                arg_names.push_back(structtype->op_name(i).str());
+            }
+
+            result["type"] = "struct";
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["struct_name"] = structtype->name().str();
+            result["args"] = args;
+            result["arg_names"] = arg_names;
+        } else if (auto varianttype = type->isa<VariantType>()) {
+            json args = json::array();
+            json arg_names = json::array();
+            for (size_t i = 0; i < varianttype->num_ops(); ++i) {
+                args.push_back(translate_type(varianttype->op(i)));
+                arg_names.push_back(varianttype->op_name(i).str());
+            }
+
+            result["type"] = "variant";
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["variant_name"] = varianttype->name().str();
+            result["args"] = args;
+            result["arg_names"] = arg_names;
+        } else if (auto tupletype = type->isa<TupleType>()) {
+            json args = json::array();
+            for (size_t i = 0; i < tupletype->num_ops(); ++i) {
+                args.push_back(translate_type(tupletype->op(i)));
+            }
+
+            result["type"] = "tuple";
+            result["name"] = "_" + std::to_string(type_table.size());
+            result["args"] = args;
         } else if (auto prim = type->isa<PrimType>()) {
             result["name"] = "_" + std::to_string(type_table.size());
             result["length"] = prim->length();
@@ -26,15 +98,6 @@ class TypeTable {
 #define THORIN_ALL_TYPE(T, M) case PrimTypeTag::PrimType_##T: { result["tag"] = #T; break; }
 #include <thorin/tables/primtypetable.h>
             }
-        } else if (auto fntype = type->isa<FnType>()) {
-            json arg_types = json::array();
-            for (auto arg : fntype->ops()) {
-                arg_types.push_back(translate_type(arg));
-            }
-
-            result["type"] = "fn";
-            result["name"] = "_" + std::to_string(type_table.size());
-            result["args"] = arg_types;
         } else if (auto ptrtype = type->isa<PtrType>()) {
             auto pointee_type = translate_type(ptrtype->pointee());
 
@@ -42,19 +105,6 @@ class TypeTable {
             result["args"] = { pointee_type };
             result["name"] = pointee_type + "_p";
             result["length"] = ptrtype->length();
-        } else if (auto arr = type->isa<IndefiniteArrayType>()) {
-            auto elem_type = translate_type(arr->elem_type());
-
-            result["type"] = "indef_array";
-            result["args"] = { elem_type };
-            result["name"] = elem_type + "_iarr";
-        } else if (auto arr = type->isa<DefiniteArrayType>()) {
-            auto elem_type = translate_type(arr->elem_type());
-
-            result["type"] = "def_array";
-            result["args"] = { elem_type };
-            result["length"] = arr->dim();
-            result["name"] = elem_type + "_iarr";
         } else {
             std::cerr << "type cannot be translated\n";
             type->dump();
@@ -85,11 +135,26 @@ class DefTable {
         json result;
         if (auto cont = def->isa<Continuation>()) {
             if (cont->is_intrinsic()) {
-                assert(cont->intrinsic() == Intrinsic::Branch && "TODO: anything else is unsupported RN");
-
-                result["name"] = "branch";
-                result["type"] = "continuation";
-                result["intrinsic"] = "branch";
+                if (cont->intrinsic() == Intrinsic::Branch) {
+                    result["name"] = "branch";
+                    result["type"] = "continuation";
+                    result["intrinsic"] = "branch";
+                } else if (cont->intrinsic() == Intrinsic::Match) {
+                    //TODO: These will change in the memory branch!
+                    assert(!is_mem(cont->param(0)));
+
+                    size_t num_patterns = cont->num_params() - 2;
+                    auto variant_type = type_table_.translate_type(cont->param(0)->type());
+                    auto name = expected_name != "" ? expected_name : "_match_" + std::to_string(def_table.size());
+
+                    result["name"] = name;
+                    result["type"] = "continuation";
+                    result["intrinsic"] = "match";
+                    result["variant_type"] = variant_type;
+                    result["num_patterns"] = num_patterns;
+                } else {
+                    assert(false && "TODO: only Branch and Match supported RN");
+                }
             } else if (cont->is_imported()) {
                 auto name = cont->name();
                 auto type = type_table_.translate_type(def->type());
@@ -333,6 +398,18 @@ class DefTable {
             result["type"] = "arithop";
             result["op"] = op;
             result["args"] = args;
+        } else if (auto mathop = def->isa<MathOp>()) {
+            auto op = mathop->op_name();
+            json args = json::array();
+            for (auto arg : mathop->ops()) {
+                args.push_back(translate_def(arg));
+            }
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "mathop";
+            result["op"] = op;
+            result["args"] = args;
         } else if (auto select = def->isa<Select>()) {
             json args = json::array();
             args.push_back(translate_def(select->cond()));
@@ -411,6 +488,79 @@ class DefTable {
             result["type"] = "global";
             result["init"] = init;
             result["mutable"] = is_mutable;
+        } else if (auto variant = def->isa<Variant>()) {
+            auto variant_type = type_table_.translate_type(variant->type());
+            auto value = translate_def(variant->value());
+            size_t index = variant->index();
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "variant";
+            result["variant_type"] = variant_type;
+            result["value"] = value;
+            result["index"] = index;
+        } else if (auto variant_extract = def->isa<VariantExtract>()) {
+            auto value = translate_def(variant_extract->value());
+            size_t index = variant_extract->index();
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "variant_extract";
+            result["value"] = value;
+            result["index"] = index;
+        } else if (auto variant_index = def->isa<VariantIndex>()) {
+            auto value = translate_def(variant_index->op(0));
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "variant_index";
+            result["value"] = value;
+        } else if (auto assembly = def->isa<Assembly>()) {
+            auto asm_type = type_table_.translate_type(assembly->type());
+            json inputs = json::array();
+            for (auto input : assembly->inputs()) {
+                inputs.push_back(translate_def(input));
+            }
+            auto asm_template = assembly->asm_template();
+            json out_constraints = json::array();
+            for (auto constraint : assembly->output_constraints()) {
+                out_constraints.push_back(constraint);
+            }
+            json in_constraints = json::array();
+            for (auto constraint : assembly->input_constraints()) {
+                in_constraints.push_back(constraint);
+            }
+            json clobbers = json::array();
+            for (auto c : assembly->clobbers()) {
+                clobbers.push_back(c);
+            }
+
+            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+
+            result["name"] = name;
+            result["type"] = "assembly";
+
+            result["asm_type"] = asm_type;
+            result["inputs"] = inputs;
+            result["asm_template"] = asm_template;
+
+            result["output_constraints"] = out_constraints;
+            result["input_constraints"] = in_constraints;
+            result["clobbers"] = clobbers;
+            switch (assembly->flags()) {
+            case Assembly::Flags::NoFlag:
+                result["flags"] = "noflag";
+                break;
+            case Assembly::Flags::HasSideEffects:
+                result["flags"] = "hassideeffects";
+                break;
+            case Assembly::Flags::IsAlignStack:
+                result["flags"] = "isalignstack";
+                break;
+            case Assembly::Flags::IsIntelDialect:
+                result["flags"] = "isinteldialect";
+                break;
+            }
         } else {
             def->dump();
             def->dump(2);

From c7dc7b8d1ed8b53c9c492ef1565ac0550fcd83f5 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 23 Nov 2022 13:16:14 +0100
Subject: [PATCH 122/342] [Json]: Improved continuation generation:

  * Support named intrinsics.
  * Generate filters.
  * Improve support for external functions.
    * Remove expected_name from defs generation.
---
 src/thorin/be/json/json.cpp | 141 +++++++++++++++++++-----------------
 1 file changed, 75 insertions(+), 66 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index ff9905558..db13db2ce 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -126,7 +126,7 @@ class DefTable {
 
     DefMap<std::string> known_defs;
 
-    std::string translate_def (const Def * def, std::string expected_name = "") {
+    std::string translate_def (const Def * def) {
         auto it = known_defs.find(def);
         if (it != known_defs.end()) {
             return it->second;
@@ -145,7 +145,7 @@ class DefTable {
 
                     size_t num_patterns = cont->num_params() - 2;
                     auto variant_type = type_table_.translate_type(cont->param(0)->type());
-                    auto name = expected_name != "" ? expected_name : "_match_" + std::to_string(def_table.size());
+                    auto name = "_match_" + std::to_string(def_table.size());
 
                     result["name"] = name;
                     result["type"] = "continuation";
@@ -153,54 +153,64 @@ class DefTable {
                     result["variant_type"] = variant_type;
                     result["num_patterns"] = num_patterns;
                 } else {
-                    assert(false && "TODO: only Branch and Match supported RN");
+                    auto intrinsic_name = cont->name();
+                    auto intrinsic_type = type_table_.translate_type(cont->type());
+                    auto name = "_in_" + std::to_string(def_table.size());
+
+                    result["name"] = name;
+                    result["type"] = "continuation";
+                    result["intrinsic"] = intrinsic_name;
+                    result["fn_type"] = intrinsic_type;
                 }
-            } else if (cont->is_imported()) {
-                auto name = cont->name();
+                if (cont->filter() && !cont->filter()->empty())
+                    result["filter"] = translate_def(cont->filter());
+            } else {
                 auto type = type_table_.translate_type(def->type());
 
-                result["name"] = name;
-                result["type"] = "continuation";
-                result["fn_type"] = type;
-                result["imported"] = true;
-                result["external"] = cont->is_external();
-            } else {
-                assert(cont->has_body());
+                //Make the name available in known_defs as early as possible to prevent recursion issues.
+                auto name = "_cont_" + std::to_string(decl_table.size());
+                known_defs[def] = name;
 
-                auto type = type_table_.translate_type(def->type());
+                //TODO: Is this actually required for imported functions?
                 json arg_names = json::array();
                 for (auto arg : cont->params()) {
                     arg_names.push_back(translate_def(arg));
                 }
 
-                auto name = expected_name != "" ? expected_name : "_cont_" + std::to_string(decl_table.size());
-
                 json forward_decl;
                 forward_decl["name"] = name;
                 forward_decl["type"] = "continuation";
                 forward_decl["fn_type"] = type;
                 forward_decl["arg_names"] = arg_names;
-                forward_decl["external"] = cont->is_external();
+                if (cont->is_external())
+                    forward_decl["external"] = cont->name();
                 decl_table.push_back(forward_decl);
 
-                known_defs[def] = name;
+                if(cont->has_body()) {
+                    auto app = cont->body();
+                    auto target = translate_def(app->callee());
+                    json args = json::array();
+                    for (auto arg : app->args()) {
+                        args.push_back(translate_def(arg));
+                    }
 
-                auto app = cont->body();
-                auto target = translate_def(app->callee());
-                json args = json::array();
-                for (auto arg : app->args()) {
-                    args.push_back(translate_def(arg));
+                    result["name"] = name;
+                    result["type"] = "continuation";
+                    if (cont->filter() && !cont->filter()->empty())
+                        result["filter"] = translate_def(cont->filter());
+                    result["app"] = {
+                        {"target", target},
+                        {"args", args}
+                    };
+                } else {
+                    //Early return. We do not have a body, so there is no point in writing something to the def table.
+                    if (cont->filter() && !cont->filter()->empty())
+                        assert(false && "These filters cannot be generated RN");
+                    return name;
                 }
-
-                result["name"] = name;
-                result["type"] = "continuation";
-                result["app"] = {
-                    {"target", target},
-                    {"args", args}
-                };
             }
         } else if (auto lit = def->isa<PrimLit>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
             auto type = type_table_.translate_type(lit->type());
 
             result["name"] = name;
@@ -216,28 +226,28 @@ class DefTable {
                 assert(false && "not implemented");
             }
         } else if (def->isa<Top>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
             auto type = type_table_.translate_type(def->type());
 
             result["name"] = name;
             result["type"] = "top";
             result["const_type"] = type;
         } else if (def->isa<Bottom>()) {
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
             auto type = type_table_.translate_type(def->type());
 
             result["name"] = name;
             result["type"] = "bottom";
             result["const_type"] = type;
         } else if (auto param = def->isa<Param>()) {
-            auto name = expected_name != "" ? expected_name : param->continuation()->unique_name() + "." + std::to_string(param->index());
+            auto name = param->continuation()->unique_name() + "." + std::to_string(param->index());
             known_defs[def] = name;
             return name;
         } else if (auto load = def->isa<Load>()) {
             json args = json::array();
             args.push_back(translate_def(load->mem()));
             args.push_back(translate_def(load->ptr()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "load";
@@ -247,21 +257,21 @@ class DefTable {
             args.push_back(translate_def(store->mem()));
             args.push_back(translate_def(store->ptr()));
             args.push_back(translate_def(store->val()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "store";
             result["args"] = args;
         } else if (auto size_of = def->isa<SizeOf>()) {
             auto target_type = type_table_.translate_type(size_of->of());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "sizeof";
             result["target_type"] = target_type;
         } else if (auto align_of = def->isa<AlignOf>()) {
             auto target_type = type_table_.translate_type(align_of->of());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "alignof";
@@ -269,7 +279,7 @@ class DefTable {
         } else if (auto cast = def->isa<Cast>()) {
             auto source = translate_def(cast->from());
             auto target_type = type_table_.translate_type(cast->type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "cast";
@@ -278,7 +288,7 @@ class DefTable {
          } else if (auto bitcast = def->isa<Bitcast>()) {
             auto source = translate_def(bitcast->from());
             auto target_type = type_table_.translate_type(bitcast->type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "bitcast";
@@ -286,7 +296,7 @@ class DefTable {
             result["target_type"] = target_type;
         } else if (auto indef_array = def->isa<IndefiniteArray>()) {
             auto dim = translate_def(indef_array->op(0));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
             auto element_type = type_table_.translate_type(indef_array->elem_type());
 
             result["name"] = name;
@@ -299,7 +309,7 @@ class DefTable {
                 args.push_back(translate_def(arg));
             }
 
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
             auto element_type = type_table_.translate_type(def_array->elem_type());
 
             result["name"] = name;
@@ -310,7 +320,7 @@ class DefTable {
             json args = json::array();
             args.push_back(translate_def(lea->ptr()));
             args.push_back(translate_def(lea->index()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "lea";
@@ -319,7 +329,7 @@ class DefTable {
             json args = json::array();
             args.push_back(translate_def(extract->agg()));
             args.push_back(translate_def(extract->index()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "extract";
@@ -329,7 +339,7 @@ class DefTable {
             args.push_back(translate_def(insert->agg()));
             args.push_back(translate_def(insert->index()));
             args.push_back(translate_def(insert->value()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "insert";
@@ -339,7 +349,7 @@ class DefTable {
             args.push_back(translate_def(closure->op(0)));
             args.push_back(translate_def(closure->op(1)));
             auto closure_type = type_table_.translate_type(closure->type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "closure";
@@ -351,7 +361,7 @@ class DefTable {
                 args.push_back(translate_def(arg));
             }
             auto struct_type = type_table_.translate_type(struct_agg->type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "struct";
@@ -362,7 +372,7 @@ class DefTable {
             for (auto arg : tuple->ops()) {
                 args.push_back(translate_def(arg));
             }
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "tuple";
@@ -372,7 +382,7 @@ class DefTable {
             for (auto arg : vector->ops()) {
                 args.push_back(translate_def(arg));
             }
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "vector";
@@ -382,7 +392,7 @@ class DefTable {
             for (auto arg : filter->ops()) {
                 args.push_back(translate_def(arg));
             }
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "filter";
@@ -392,7 +402,7 @@ class DefTable {
             json args = json::array();
             args.push_back(translate_def(arithop->lhs()));
             args.push_back(translate_def(arithop->rhs()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "arithop";
@@ -404,7 +414,7 @@ class DefTable {
             for (auto arg : mathop->ops()) {
                 args.push_back(translate_def(arg));
             }
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "mathop";
@@ -415,7 +425,7 @@ class DefTable {
             args.push_back(translate_def(select->cond()));
             args.push_back(translate_def(select->tval()));
             args.push_back(translate_def(select->fval()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "select";
@@ -425,7 +435,7 @@ class DefTable {
             json args = json::array();
             args.push_back(translate_def(cmp->lhs()));
             args.push_back(translate_def(cmp->rhs()));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "cmp";
@@ -433,28 +443,28 @@ class DefTable {
             result["args"] = args;
         } else if (auto run = def->isa<Run>()) {
             auto target = translate_def(run->def());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "run";
             result["target"] = target;
         } else if (auto hlt = def->isa<Hlt>()) {
             auto target = translate_def(hlt->def());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "hlt";
             result["target"] = target;
         } else if (auto known = def->isa<Known>()) {
             auto def = translate_def(known->def());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "known";
             result["def"] = def;
         } else if (auto enter = def->isa<Enter>()) {
             auto mem = translate_def(enter->mem());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "enter";
@@ -462,7 +472,7 @@ class DefTable {
         } else if (auto slot = def->isa<Slot>()) {
             auto frame = translate_def(slot->frame());
             auto target_type = type_table_.translate_type(slot->alloced_type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "slot";
@@ -473,7 +483,7 @@ class DefTable {
             args.push_back(translate_def(alloc->mem()));
             args.push_back(translate_def(alloc->extra()));
             auto target_type = type_table_.translate_type(alloc->alloced_type());
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "alloc";
@@ -482,7 +492,7 @@ class DefTable {
         } else if (auto global = def->isa<Global>()) {
             auto init = translate_def(global->init());
             bool is_mutable = global->is_mutable();
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "global";
@@ -492,7 +502,7 @@ class DefTable {
             auto variant_type = type_table_.translate_type(variant->type());
             auto value = translate_def(variant->value());
             size_t index = variant->index();
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "variant";
@@ -502,7 +512,7 @@ class DefTable {
         } else if (auto variant_extract = def->isa<VariantExtract>()) {
             auto value = translate_def(variant_extract->value());
             size_t index = variant_extract->index();
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "variant_extract";
@@ -510,7 +520,7 @@ class DefTable {
             result["index"] = index;
         } else if (auto variant_index = def->isa<VariantIndex>()) {
             auto value = translate_def(variant_index->op(0));
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "variant_index";
@@ -535,7 +545,7 @@ class DefTable {
                 clobbers.push_back(c);
             }
 
-            auto name = expected_name != "" ? expected_name : "_" + std::to_string(def_table.size());
+            auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "assembly";
@@ -583,8 +593,7 @@ void CodeGen::emit_stream(std::ostream& stream) {
 
     for (auto external : world().externals()) {
         const Continuation* continuation = external.second;
-        auto expected_name = continuation->name();
-        def_table.translate_def(continuation, expected_name);
+        def_table.translate_def(continuation);
     }
 
     j["type_table"] = type_table.type_table;

From 09853d507d543dd312bf8a8fdee76fef228aa965 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 24 Nov 2022 17:47:12 +0100
Subject: [PATCH 123/342] [Json]: Bugfix: Assembly needs a memory input.

---
 src/thorin/be/json/json.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index db13db2ce..39d3ca96d 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -528,6 +528,7 @@ class DefTable {
         } else if (auto assembly = def->isa<Assembly>()) {
             auto asm_type = type_table_.translate_type(assembly->type());
             json inputs = json::array();
+            inputs.push_back(translate_def(assembly->mem()));
             for (auto input : assembly->inputs()) {
                 inputs.push_back(translate_def(input));
             }

From a44eff8d7d299be41a9d30e6d5712c794cbe4311 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 25 Nov 2022 14:27:51 +0100
Subject: [PATCH 124/342] [Json]: Add device and addrspace to pointer types.

---
 src/thorin/be/json/json.cpp | 22 +++++++++++++++++++++-
 1 file changed, 21 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 39d3ca96d..612e6d3ba 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -100,11 +100,31 @@ class TypeTable {
             }
         } else if (auto ptrtype = type->isa<PtrType>()) {
             auto pointee_type = translate_type(ptrtype->pointee());
+            auto device = ptrtype->device();
 
             result["type"] = "ptr";
             result["args"] = { pointee_type };
-            result["name"] = pointee_type + "_p";
+            result["name"] = pointee_type + "_p_" + std::to_string(type_table.size());
             result["length"] = ptrtype->length();
+            if (device != -1)
+                result["device"] = device;
+            switch (ptrtype->addr_space()) {
+            case AddrSpace::Generic:
+                //result["addrspace"] = "generic"; //Default
+                break;
+            case AddrSpace::Global:
+                result["addrspace"] = "global";
+                break;
+            case AddrSpace::Texture:
+                result["addrspace"] = "texture";
+                break;
+            case AddrSpace::Shared:
+                result["addrspace"] = "shared";
+                break;
+            case AddrSpace::Constant:
+                result["addrspace"] = "constant";
+                break;
+            }
         } else {
             std::cerr << "type cannot be translated\n";
             type->dump();

From 7667117ac71c9a54e144cb43b1defb9decd3d54d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Tue, 29 Nov 2022 14:29:11 +0100
Subject: [PATCH 125/342] updated to latest shady API

---
 src/thorin/be/codegen.cpp     |  2 ++
 src/thorin/be/shady/shady.cpp | 60 +++++++++++++++++++----------------
 2 files changed, 35 insertions(+), 27 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 517cb0232..b804fa6e6 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -10,6 +10,8 @@
 #endif
 #if THORIN_ENABLE_SHADY
 #include "thorin/be/shady/shady.h"
+#undef empty
+#undef nodes
 #endif
 #include "thorin/be/c/c.h"
 
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 4da148a44..7ca15a317 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -1,4 +1,5 @@
 #include "shady.h"
+#undef empty
 
 #include "thorin/analyses/scope.h"
 #include "thorin/transform/structurize.h"
@@ -15,7 +16,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     assert(!module);
 
-    shady::ArenaConfig config = { 0 };
+    shady::ArenaConfig config = { };
     config.check_types = true;
     arena = shady::new_ir_arena(config);
     module = shady::new_module(arena, world().name().c_str());
@@ -225,7 +226,7 @@ void CodeGen::prepare(Continuation* cont, shady::Node*) {
     } else
         assert(bb.head);
 
-    bb.builder = shady::begin_body(arena);
+    bb.builder = shady::begin_body(module);
 }
 
 static std::optional<shady::Op> is_shady_prim_op(const Continuation* cont) {
@@ -280,38 +281,43 @@ void CodeGen::emit_epilogue(Continuation* cont) {
         bb.terminator = shady::jump(arena, payload);
     } else if (auto intrinsic = body->callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
         assert(false);
-    } else if (auto callee = body->callee()->isa_nom<Continuation>()) {
-        int ret_param = find_return_parameter(callee->type());
+    } else {
+        int ret_param = find_return_parameter(body->callee()->type()->as<FnType>());
+        // TODO handle tail calls ?
         assert(ret_param >= 0);
 
         args.erase(args.begin() + ret_param);
         args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
 
-        // shady primop called as imported continuations look like continuation calls to thorin, but not to shady
-        // we just need to carefully emit the primop as an instruction, then jump to the target BB, passing the stuff as we do
-        if (auto op = is_shady_prim_op(callee); op.has_value()) {
-            shady::bind_instruction(bb.builder, shady::prim_op(arena, (shady::PrimOp) {
-                .op = op.value(),
-                .type_arguments = empty(arena),
-                .operands = vec2nodes(args),
-            }));
-            shady::Jump jump;
-            jump.target = args[ret_param];
-            bb.terminator = shady::jump(arena, jump);
-            return;
-        }
+        const shady::Node* call;
+        if (auto callee = body->callee()->isa_nom<Continuation>()) {
+            // shady primop called as imported continuations look like continuation calls to thorin, but not to shady
+            // we just need to carefully emit the primop as an instruction, then jump to the target BB, passing the stuff as we do
+            if (auto op = is_shady_prim_op(callee); op.has_value()) {
+                shady::bind_instruction(bb.builder, shady::prim_op(arena, (shady::PrimOp) {
+                        .op = op.value(),
+                        .type_arguments = shady::nodes(arena, 0, nullptr),
+                        .operands = vec2nodes(args),
+                }));
+                shady::Jump jump;
+                jump.target = args[ret_param];
+                bb.terminator = shady::jump(arena, jump);
+                return;
+            }
 
-        shady::Call payload;
-        payload.args = vec2nodes(args);
-        payload.callee = emit(callee);
-        auto call = shady::call_instr(arena, payload);
+            shady::LeafCall payload;
+            payload.args = vec2nodes(args);
+            payload.callee = emit(callee);
+            call = shady::leaf_call(arena, payload);
+        } else {
+            shady::IndirectCall payload;
+            payload.args = vec2nodes(args);
+            payload.callee = emit(callee);
+            call = shady::indirect_call(arena, payload);
+        }
 
-        shady::LetInto payload2;
-        payload2.instruction = call;
-        payload2.tail = args[ret_param];
-        bb.terminator = shady::let_into(arena, payload2);
-    } else {
-        assert(false);
+        assert(args[ret_param]->tag == shady::BasicBlock_TAG);
+        bb.terminator = shady::let_into(arena, call, args[ret_param]);
     }
 }
 

From 4357af61b41651ce2c5a6628f0e04b787650df6a Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 29 Nov 2022 16:02:46 +0100
Subject: [PATCH 126/342] Mark Generated Device Code with the correct name to
 call.

---
 src/thorin/be/codegen.cpp      |  4 +++-
 src/thorin/be/json/json.cpp    |  8 ++++++++
 src/thorin/be/json/json.h      | 11 +++++++----
 src/thorin/be/llvm/runtime.cpp |  3 ++-
 src/thorin/continuation.h      |  5 +++--
 5 files changed, 23 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index db6f6d802..0e82fe2ce 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -26,7 +26,7 @@ static void get_kernel_configs(
         Continuation* imported = nullptr;
         for (auto [_, exported] : externals) {
             if (!exported->has_body()) continue;
-            if (exported->name() == continuation->unique_name())
+            if (exported->name() == continuation->name())
                 imported = exported;
         }
         if (!imported) continue;
@@ -41,6 +41,7 @@ static void get_kernel_configs(
             return false;
         }, true);
 
+        continuation->attributes().cc = CC::DeviceHostCode;
         continuation->destroy("codegen");
     }
 }
@@ -105,6 +106,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
 
         // Necessary so that the names match in the original and imported worlds
         imported->set_name(continuation->unique_name());
+        continuation->set_name(continuation->unique_name());
         for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
             imported->param(i)->set_name(continuation->param(i)->name());
         imported->world().make_external(imported);
diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 612e6d3ba..0b120e621 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -204,6 +204,8 @@ class DefTable {
                 forward_decl["arg_names"] = arg_names;
                 if (cont->is_external())
                     forward_decl["external"] = cont->name();
+                if (cont->cc() == CC::DeviceHostCode)
+                    forward_decl["device"] = cont->name();
                 decl_table.push_back(forward_decl);
 
                 if(cont->has_body()) {
@@ -608,6 +610,12 @@ void CodeGen::emit_stream(std::ostream& stream) {
     json j;
 
     j["module"] = world().name();
+    if (target_triple != "")
+        j["target_triple"] = target_triple;
+    if (target_cpu != "")
+        j["target_cpu"] = target_cpu;
+    if (target_attr != "")
+        j["target_attr"] = target_attr;
 
     TypeTable type_table;
     DefTable def_table(type_table);
diff --git a/src/thorin/be/json/json.h b/src/thorin/be/json/json.h
index ee221ea96..23744480d 100644
--- a/src/thorin/be/json/json.h
+++ b/src/thorin/be/json/json.h
@@ -17,9 +17,11 @@ using json = nlohmann::json;
 
 class CodeGen : public thorin::CodeGen {
 public:
-    CodeGen(World& world, const Cont2Config& kernel_config, bool debug)
+    CodeGen(World& world, bool debug, std::string& target_triple, std::string& target_cpu, std::string& target_attr)
         : thorin::CodeGen(world, debug)
-        , kernel_config_(kernel_config)
+        , target_triple(target_triple)
+        , target_cpu(target_cpu)
+        , target_attr(target_attr)
     {}
     
     void emit_stream(std::ostream& stream) override;
@@ -27,9 +29,10 @@ class CodeGen : public thorin::CodeGen {
     const char* file_ext() const override {
         return ".thorin.json";
     }
-
 private:
-    const Cont2Config& kernel_config_;
+    std::string& target_triple;
+    std::string& target_cpu;
+    std::string& target_attr;
 };
 
 }
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index 47e7dda8f..070e0fb6c 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -79,7 +79,8 @@ Continuation* Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& buil
     auto kernel = body->arg(LaunchArgs::Body)->as<Global>()->init()->as<Continuation>();
 
     auto& world = continuation->world();
-    auto kernel_name = builder.CreateGlobalStringPtr(kernel->name() == "hls_top" ? kernel->name() : kernel->unique_name());
+    //auto kernel_name = builder.CreateGlobalStringPtr(kernel->name() == "hls_top" ? kernel->name() : kernel->name());
+    auto kernel_name = builder.CreateGlobalStringPtr(kernel->name());
     auto file_name = builder.CreateGlobalStringPtr(world.name() + ext);
     const size_t num_kernel_args = body->num_args() - LaunchArgs::Num;
 
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 9f34fe992..763dbcaad 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -80,8 +80,9 @@ class App : public Def {
 //------------------------------------------------------------------------------
 
 enum class CC : uint8_t {
-    C,          ///< C calling convention.
-    Device,     ///< Device calling convention. These are special functions only available on a particular device.
+    C,              ///< C calling convention.
+    Device,         ///< Device calling convention. These are special functions only available on a particular device.
+    DeviceHostCode, ///< Calling convention to denote continuations that are generated as device code.
 };
 
 enum class Intrinsic : uint8_t {

From b04b2faff74178f34abc64883adaa7c7cca574b4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Nov 2022 16:53:37 +0100
Subject: [PATCH 127/342] fixed undefined jump arguments in shady be

---
 src/thorin/be/shady/shady.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 7ca15a317..7f6e0a20e 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -301,6 +301,7 @@ void CodeGen::emit_epilogue(Continuation* cont) {
                 }));
                 shady::Jump jump;
                 jump.target = args[ret_param];
+                jump.args = shady::nodes(arena, 0, NULL);
                 bb.terminator = shady::jump(arena, jump);
                 return;
             }

From 00e50649f8ee5808bf73fbad55c00aa34c0da524 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Nov 2022 18:19:29 +0100
Subject: [PATCH 128/342] shady: gpu entry point annotations

---
 src/thorin/be/shady/shady.cpp | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 7f6e0a20e..b86fb5613 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -17,6 +17,7 @@ void CodeGen::emit_stream(std::ostream& out) {
     assert(!module);
 
     shady::ArenaConfig config = { };
+    config.name_bound = true;
     config.check_types = true;
     arena = shady::new_ir_arena(config);
     module = shady::new_module(arena, world().name().c_str());
@@ -177,6 +178,20 @@ shady::Node* CodeGen::emit_decl_head(Def* def) {
             returns.push_back(ret_type);
         }
 
+        auto config = kernel_config_.find(cont);
+        if (config != kernel_config_.end()) {
+            if (auto gpu_config = config->second->isa<GPUKernelConfig>()) {
+                annotations.push_back(shady::annotation_value(arena, { .name = "EntryPoint", .value = shady::string_lit(arena, { .string = "compute" })}));
+                std::vector<const shady::Node*> block_size;
+                block_size.emplace_back(shady::int32_literal(arena, get<0>(gpu_config->block_size())));
+                block_size.emplace_back(shady::int32_literal(arena, get<1>(gpu_config->block_size())));
+                block_size.emplace_back(shady::int32_literal(arena, get<2>(gpu_config->block_size())));
+                annotations.push_back(shady::annotation_values(arena, { .name = "WorkgroupSize", .values = vec2nodes(block_size) }));
+            } else {
+                assert(false && "Only GPU kernel configs are currently supported");
+            }
+        }
+
         return shady::function(module, vec2nodes(params), def->unique_name().c_str(), vec2nodes(annotations), vec2nodes(returns));
     } else if (auto global = def->isa<Global>()) {
         if (global->is_mutable()) {

From 9eb16a3a9ac1315ed56ae0a25f9fd84d5454a309 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 1 Dec 2022 14:53:50 +0100
Subject: [PATCH 129/342] Add support for external glbobals.

---
 src/thorin/analyses/scope.cpp               |  5 +--
 src/thorin/be/codegen.cpp                   |  4 ++-
 src/thorin/be/emitter.h                     |  9 ++++--
 src/thorin/be/json/json.cpp                 |  6 ++--
 src/thorin/be/llvm/llvm.cpp                 | 34 ++++++++++++++++++---
 src/thorin/be/llvm/llvm.h                   |  2 ++
 src/thorin/primop.cpp                       |  2 ++
 src/thorin/primop.h                         |  2 ++
 src/thorin/rec_stream.cpp                   | 16 ++++++++--
 src/thorin/transform/cleanup_world.cpp      | 10 +++---
 src/thorin/transform/importer.cpp           |  7 +++++
 src/thorin/transform/partial_evaluation.cpp |  4 ++-
 src/thorin/world.h                          | 10 +++---
 13 files changed, 86 insertions(+), 25 deletions(-)

diff --git a/src/thorin/analyses/scope.cpp b/src/thorin/analyses/scope.cpp
index 37d34114c..8fd518885 100644
--- a/src/thorin/analyses/scope.cpp
+++ b/src/thorin/analyses/scope.cpp
@@ -111,8 +111,9 @@ template<bool elide_empty>
 void Scope::for_each(const World& world, std::function<void(Scope&)> f) {
     unique_queue<ContinuationSet> continuation_queue;
 
-    for (auto&& [_, cont] : world.externals()) {
-        if (cont->has_body()) continuation_queue.push(cont);
+    for (auto&& [_, def] : world.externals()) {
+        if (auto cont = def->template isa<Continuation>())
+            if (cont->has_body()) continuation_queue.push(cont);
     }
 
     while (!continuation_queue.empty()) {
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 0e82fe2ce..5a0300eb4 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -24,7 +24,9 @@ static void get_kernel_configs(
     for (auto continuation : kernels) {
         // recover the imported continuation (lost after the call to opt)
         Continuation* imported = nullptr;
-        for (auto [_, exported] : externals) {
+        for (auto [_, def] : externals) {
+            auto exported = def->isa<Continuation>();
+            if (!exported) continue;
             if (!exported->has_body()) continue;
             if (exported->name() == continuation->name())
                 imported = exported;
diff --git a/src/thorin/be/emitter.h b/src/thorin/be/emitter.h
index 72300ba06..073c76f34 100644
--- a/src/thorin/be/emitter.h
+++ b/src/thorin/be/emitter.h
@@ -12,8 +12,13 @@ class Emitter {
     /// Internal wrapper for @p emit that checks and retrieves/puts the @c Value from @p defs_.
     Value emit_(const Def* def) {
         auto place = def->no_dep() ? entry_ : scheduler_.smart(def);
-        auto& bb = cont2bb_[place];
-        return child().emit_bb(bb, def);
+
+        if (place) {
+            auto& bb = cont2bb_[place];
+            return child().emit_bb(bb, def);
+        } else {
+            return child().emit_constant(def);
+        }
     }
 
 protected:
diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 0b120e621..ae9d9d1de 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -514,12 +514,15 @@ class DefTable {
         } else if (auto global = def->isa<Global>()) {
             auto init = translate_def(global->init());
             bool is_mutable = global->is_mutable();
+            bool is_external = global->is_external();
             auto name = "_" + std::to_string(def_table.size());
 
             result["name"] = name;
             result["type"] = "global";
             result["init"] = init;
             result["mutable"] = is_mutable;
+            if (is_external)
+                result["external"] = global->name();
         } else if (auto variant = def->isa<Variant>()) {
             auto variant_type = type_table_.translate_type(variant->type());
             auto value = translate_def(variant->value());
@@ -621,8 +624,7 @@ void CodeGen::emit_stream(std::ostream& stream) {
     DefTable def_table(type_table);
 
     for (auto external : world().externals()) {
-        const Continuation* continuation = external.second;
-        def_table.translate_def(continuation);
+        def_table.translate_def(external.second);
     }
 
     j["type_table"] = type_table.type_table;
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index ce9fb159e..977893f75 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -276,6 +276,12 @@ CodeGen::emit_module() {
         dicompile_unit_ = dibuilder_.createCompileUnit(llvm::dwarf::DW_LANG_C, dibuilder_.createFile(world().name(), llvm::StringRef()), "Impala", opt() > 0, llvm::StringRef(), 0);
     }
 
+    for (auto&& [_, def] : world().externals()) {
+        if (auto global = def->isa<Global>()) {
+            emit(global);
+        }
+    }
+
     Scope::for_each(world(), [&] (const Scope& scope) { emit_scope(scope); });
 
     if (debug()) dibuilder_.finalize();
@@ -538,9 +544,17 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
     irbuilder.SetInsertPoint(bb->getTerminator());
 }
 
+llvm::Value* CodeGen::emit_constant(const Def* def) {
+    auto irbuilder = llvm::IRBuilder(context());
+    return emit_builder(irbuilder, def);
+}
+
 llvm::Value* CodeGen::emit_bb(BB& bb, const Def* def) {
     auto& irbuilder = *bb.second;
+    return emit_builder(irbuilder, def);
+}
 
+llvm::Value* CodeGen::emit_builder(llvm::IRBuilder<>& irbuilder, const Def* def) {
     // TODO
     //if (debug())
         //irbuilder.SetCurrentDebugLocation(llvm::DILocation::get(discope_->getContext(), def->loc().begin.row, def->loc().begin.col, discope_));
@@ -955,13 +969,23 @@ llvm::Value* CodeGen::emit_global(const Global* global) {
         val = emit(continuation);
     else {
         auto llvm_type = convert(global->alloced_type());
-        auto var = llvm::cast<llvm::GlobalVariable>(module().getOrInsertGlobal(global->unique_name().c_str(), llvm_type));
+        auto var = llvm::cast<llvm::GlobalVariable>(module().getOrInsertGlobal(global->is_external() ? global->name().c_str() : global->unique_name().c_str(), llvm_type));
         var->setConstant(!global->is_mutable());
-        var->setLinkage(llvm::GlobalValue::InternalLinkage);
-        if (global->init()->isa<Bottom>())
-            var->setInitializer(llvm::Constant::getNullValue(llvm_type)); // HACK
-        else
+
+        if (global->init()->isa<Bottom>()) {
+            if (global->is_external())
+                var->setExternallyInitialized(true);
+            else
+                var->setInitializer(llvm::Constant::getNullValue(llvm_type)); // HACK
+        } else
             var->setInitializer(llvm::cast<llvm::Constant>(emit(global->init())));
+
+        if (global->is_external()) {
+            var->setAlignment(llvm::Align(4));
+            var->setDSOLocal(true);
+            var->setUnnamedAddr(llvm::GlobalVariable::UnnamedAddr::None);
+        } else
+            var->setLinkage(llvm::GlobalValue::InternalLinkage);
         val = var;
     }
     return val;
diff --git a/src/thorin/be/llvm/llvm.h b/src/thorin/be/llvm/llvm.h
index 93e9aeb7e..4c41ba40c 100644
--- a/src/thorin/be/llvm/llvm.h
+++ b/src/thorin/be/llvm/llvm.h
@@ -54,7 +54,9 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<llvm::Value*, llv
         std::unique_ptr<llvm::Module>> emit_module();
     llvm::Function* prepare(const Scope&);
     virtual void prepare(Continuation*, llvm::Function*);
+    llvm::Value* emit_constant(const Def* def);
     llvm::Value* emit_bb(BB&, const Def* def);
+    llvm::Value* emit_builder(llvm::IRBuilder<>&, const Def* def);
     virtual llvm::Function* emit_fun_decl(Continuation*);
     bool is_valid(llvm::Value* value) { return value != nullptr; }
     void finalize(const Scope&);
diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index ddad23575..453db0bc2 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -288,6 +288,8 @@ const char* Global::op_name() const { return is_mutable() ? "global_mutable" : "
  * misc
  */
 
+bool Global::is_external() const { return world().is_external(this); }
+
 std::string DefiniteArray::as_string() const {
     std::string res;
     for (auto op : ops()) {
diff --git a/src/thorin/primop.h b/src/thorin/primop.h
index ad4e2efcb..f12fc74c0 100644
--- a/src/thorin/primop.h
+++ b/src/thorin/primop.h
@@ -534,6 +534,8 @@ class Global : public Def {
     const Type* alloced_type() const { return type()->pointee(); }
     const char* op_name() const override;
 
+    bool is_external() const;
+
 private:
     hash_t vhash() const override { return murmur3(gid()); }
     bool equal(const Def* other) const override { return this == other; }
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 03d6a8c20..b021544bc 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -125,6 +125,11 @@ Stream& Def::stream1(Stream& s) const {
         s.fmt(": ({, })\n", ass->clobbers());
         s.fmt(": ({, })\b", ass->ops());
         return s;
+    } else if (auto global = isa<Global>()) {
+        if (global->is_external())
+            return s.fmt("{}", unique_name());
+        else
+            return s.fmt("{}({, }))", op_name(), ops());
     }
 
     return s.fmt("{}({, }))", op_name(), ops());
@@ -138,9 +143,14 @@ Stream& World::stream(Stream& s) const {
     RecStreamer rec(s, std::numeric_limits<size_t>::max());
     s << "module '" << name() << "'";
 
-    for (auto&& [_, cont] : externals()) {
-        rec.conts.push(cont);
-        rec.run();
+    for (auto&& [_, def] : externals()) {
+        auto cont = def->isa<Continuation>();
+        if (cont) {
+            rec.conts.push(cont);
+            rec.run();
+        } else {
+            s.fmt("\n{} = {}({, })\n", def->unique_name(), def->op_name(), def->ops());
+        }
     }
 
     return s.endl();
diff --git a/src/thorin/transform/cleanup_world.cpp b/src/thorin/transform/cleanup_world.cpp
index 2748d25d9..510c2139b 100644
--- a/src/thorin/transform/cleanup_world.cpp
+++ b/src/thorin/transform/cleanup_world.cpp
@@ -236,9 +236,11 @@ void Cleaner::rebuild() {
     importer.type_old2new_.rehash(world_.types().capacity());
     importer.def_old2new_.rehash(world_.defs().capacity());
 
-    for (auto&& [_, cont] : world().externals()) {
-        if (cont->is_exported())
+    for (auto&& [_, def] : world().externals()) {
+        if (auto cont = def->isa<Continuation>(); cont && cont->is_exported())
             importer.import(cont);
+        if (auto global = def->isa<Global>(); global && global->is_external())
+            importer.import(global);
     }
 
     swap(importer.world(), world_);
@@ -296,8 +298,8 @@ void Cleaner::clean_pe_infos() {
             queue.push(continuation);
     };
 
-    for (auto&& [_, cont] : world().externals())
-        if (cont->has_body()) enqueue(cont);
+    for (auto&& [_, def] : world().externals())
+        if (auto cont = def->isa<Continuation>(); cont && cont->has_body()) enqueue(cont);
 
     while (!queue.empty()) {
         auto continuation = pop(queue);
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index f125585de..72518b5a5 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -84,6 +84,13 @@ const Def* Importer::import(const Def* odef) {
 
     if (odef->isa_structural()) {
         auto ndef = odef->rebuild(world(), ntype, nops);
+
+        if (auto oglobal = odef->isa<Global>()) {
+            if (oglobal->is_external())
+                world().make_external(const_cast<Def*>(ndef));
+        }
+
+
         todo_ |= odef->tag() != ndef->tag();
         return def_old2new_[odef] = ndef;
     }
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 28d383a38..1b0739a7e 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -146,7 +146,9 @@ void PartialEvaluator::eat_pe_info(Continuation* cur) {
 bool PartialEvaluator::run() {
     bool todo = false;
 
-    for (auto&& [_, cont] : world().externals()) {
+    for (auto&& [_, def] : world().externals()) {
+        auto cont = def->isa<Continuation>();
+        if (!cont) continue;
         if (!cont->has_body()) continue;
         enqueue(cont);
         top_level_[cont] = true;
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 96e30af2b..311279a30 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -65,7 +65,7 @@ class World : public TypeTable, public Streamable<World> {
 
     using Sea         = HashSet<const Def*, SeaHash>;///< This @p HashSet contains Thorin's "sea of nodes".
     using Breakpoints = HashSet<size_t, BreakHash>;
-    using Externals   = HashMap<std::string, Continuation*, ExternalsHash>;
+    using Externals   = HashMap<std::string, Def*, ExternalsHash>;
 
     World(World&&) = delete;
     World& operator=(const World&) = delete;
@@ -90,10 +90,10 @@ class World : public TypeTable, public Streamable<World> {
     //@{
     bool empty() { return data_.externals_.empty(); }
     const Externals& externals() const { return data_.externals_; }
-    void make_external(Continuation* cont) { data_.externals_.emplace(cont->unique_name(), cont); }
-    void make_internal(Continuation* cont) { data_.externals_.erase(cont->unique_name()); }
-    bool is_external(const Continuation* cont) { return data_.externals_.contains(cont->unique_name()); }
-    Continuation* lookup(const std::string& name) { return data_.externals_.lookup(name).value_or(nullptr); }
+    void make_external(Def* cont) { data_.externals_.emplace(cont->unique_name(), cont).second; }
+    void make_internal(Def* cont) { data_.externals_.erase(cont->unique_name()); }
+    bool is_external(const Def* cont) { return data_.externals_.contains(cont->unique_name()); }
+    Def* lookup(const std::string& name) { return const_cast<Def*>(data_.externals_.lookup(name).value_or(nullptr)); }
     //@}
 
     // literals

From 2fbe0e7727a13047a8f7dda18612354618780a9e Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 1 Dec 2022 17:15:22 +0100
Subject: [PATCH 130/342] Do not evaluate external globals if the initializer
 is bottom.

---
 src/thorin/transform/resolve_loads.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/transform/resolve_loads.cpp b/src/thorin/transform/resolve_loads.cpp
index d0a687ce4..02b5bcc1b 100644
--- a/src/thorin/transform/resolve_loads.cpp
+++ b/src/thorin/transform/resolve_loads.cpp
@@ -95,7 +95,7 @@ class ResolveLoads {
             return it->second;
         if (auto global = alloc->isa<Global>()) {
             // Immutable globals will remain set to their initial value
-            if (!global->is_mutable())
+            if (!global->is_mutable() && (!global->is_external()) || !global->init()->isa<Bottom>())
                 return mapping[alloc] = global->init();
         }
         // Nothing is known about this allocation yet
@@ -233,7 +233,7 @@ public: \
         while (true) {
             while (auto bitcast = ptr->isa<Bitcast>())
                 ptr = bitcast->from();
-            if (ptr->isa<Global>() && !ptr->as<Global>()->is_mutable())
+            if (ptr->isa<Global>() && !ptr->as<Global>()->is_mutable() && (!ptr->as<Global>()->is_external() || !ptr->as<Global>()->init()->isa<Bottom>()))
                 return ptr;
             // If first == ptr, we are looking at the pointed value.
             // In that case, we need to make sure the pointer does not escape.

From 9d4b4afb2f62eafbe2e08b84ab42081ed89102d6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 2 Dec 2022 14:56:45 +0100
Subject: [PATCH 131/342] shady files are .shady

---
 src/thorin/be/llvm/llvm.cpp  | 2 +-
 src/thorin/be/llvm/runtime.h | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index ff2d95ec4..c1ea92dc4 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1157,7 +1157,7 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
         case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
         case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
         case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
-        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::VULKAN_PLATFORM, ".spv",    continuation);
+        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation);
         case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
         case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
         case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index 19779b6a3..d99de5c27 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -23,7 +23,7 @@ class Runtime {
         CUDA_PLATFORM,
         OPENCL_PLATFORM,
         HSA_PLATFORM,
-        VULKAN_PLATFORM,
+        SHADY_PLATFORM,
     };
 
     /// Emits a call to anydsl_launch_kernel.

From ee3f679bd7f411a69a33e86009d3ae4ecf0ad4d8 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 2 Dec 2022 15:01:25 +0100
Subject: [PATCH 132/342] Bugfix: bad parentheses in resolve_loads.

---
 src/thorin/transform/resolve_loads.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/transform/resolve_loads.cpp b/src/thorin/transform/resolve_loads.cpp
index 02b5bcc1b..ba916b837 100644
--- a/src/thorin/transform/resolve_loads.cpp
+++ b/src/thorin/transform/resolve_loads.cpp
@@ -95,7 +95,7 @@ class ResolveLoads {
             return it->second;
         if (auto global = alloc->isa<Global>()) {
             // Immutable globals will remain set to their initial value
-            if (!global->is_mutable() && (!global->is_external()) || !global->init()->isa<Bottom>())
+            if (!global->is_mutable() && (!global->is_external() || !global->init()->isa<Bottom>()))
                 return mapping[alloc] = global->init();
         }
         // Nothing is known about this allocation yet

From 21bc0f6f96e75645c85611eb24d38d17cb8a2dd8 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 2 Dec 2022 16:08:29 +0100
Subject: [PATCH 133/342] fix a few issues with the shady codegen

---
 src/thorin/be/shady/shady.cpp | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index b86fb5613..17589cea8 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -178,6 +178,8 @@ shady::Node* CodeGen::emit_decl_head(Def* def) {
             returns.push_back(ret_type);
         }
 
+        std::string name = def->unique_name();
+
         auto config = kernel_config_.find(cont);
         if (config != kernel_config_.end()) {
             if (auto gpu_config = config->second->isa<GPUKernelConfig>()) {
@@ -187,12 +189,13 @@ shady::Node* CodeGen::emit_decl_head(Def* def) {
                 block_size.emplace_back(shady::int32_literal(arena, get<1>(gpu_config->block_size())));
                 block_size.emplace_back(shady::int32_literal(arena, get<2>(gpu_config->block_size())));
                 annotations.push_back(shady::annotation_values(arena, { .name = "WorkgroupSize", .values = vec2nodes(block_size) }));
+                name = "main";
             } else {
                 assert(false && "Only GPU kernel configs are currently supported");
             }
         }
 
-        return shady::function(module, vec2nodes(params), def->unique_name().c_str(), vec2nodes(annotations), vec2nodes(returns));
+        return shady::function(module, vec2nodes(params), name.c_str(), vec2nodes(annotations), vec2nodes(returns));
     } else if (auto global = def->isa<Global>()) {
         if (global->is_mutable()) {
             return shady::global_var(module, vec2nodes(annotations), convert(global->alloced_type()), global->unique_name().c_str(), convert_address_space(AddrSpace::Private));
@@ -237,7 +240,7 @@ void CodeGen::prepare(Continuation* cont, shady::Node*) {
             params.push_back(param);
         }
 
-        bb.head = shady::basic_block(arena, curr_fn, vec2nodes(params), cont->name().c_str());
+        bb.head = shady::basic_block(arena, curr_fn, vec2nodes(params), cont->unique_name().c_str());
     } else
         assert(bb.head);
 
@@ -309,14 +312,13 @@ void CodeGen::emit_epilogue(Continuation* cont) {
             // shady primop called as imported continuations look like continuation calls to thorin, but not to shady
             // we just need to carefully emit the primop as an instruction, then jump to the target BB, passing the stuff as we do
             if (auto op = is_shady_prim_op(callee); op.has_value()) {
-                shady::bind_instruction(bb.builder, shady::prim_op(arena, (shady::PrimOp) {
+                shady::Jump jump;
+                jump.target = args[ret_param];
+                jump.args = shady::bind_instruction(bb.builder, shady::prim_op(arena, (shady::PrimOp) {
                         .op = op.value(),
                         .type_arguments = shady::nodes(arena, 0, nullptr),
                         .operands = vec2nodes(args),
                 }));
-                shady::Jump jump;
-                jump.target = args[ret_param];
-                jump.args = shady::nodes(arena, 0, NULL);
                 bb.terminator = shady::jump(arena, jump);
                 return;
             }

From e4073382d6198eefb653b9bb985eb971ae148102 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 2 Dec 2022 16:49:34 +0100
Subject: [PATCH 134/342] shady: cmp, arithm

---
 src/thorin/be/shady/shady.cpp | 43 ++++++++++++++++++++++++++++++++---
 1 file changed, 40 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 17589cea8..e1485bf46 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -315,9 +315,9 @@ void CodeGen::emit_epilogue(Continuation* cont) {
                 shady::Jump jump;
                 jump.target = args[ret_param];
                 jump.args = shady::bind_instruction(bb.builder, shady::prim_op(arena, (shady::PrimOp) {
-                        .op = op.value(),
-                        .type_arguments = shady::nodes(arena, 0, nullptr),
-                        .operands = vec2nodes(args),
+                    .op = op.value(),
+                    .type_arguments = shady::nodes(arena, 0, nullptr),
+                    .operands = vec2nodes(args),
                 }));
                 bb.terminator = shady::jump(arena, jump);
                 return;
@@ -365,6 +365,21 @@ const shady::Node* CodeGen::emit_fun_decl(Continuation* cont) {
 
 const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
     const shady::Node* v = nullptr;
+
+    auto mk_primop = [&](shady::Op op, std::vector<const Def*> args, std::vector<const Type*> types = {}) -> const shady::Node* {
+        shady::PrimOp payload = {};
+        payload.op = op;
+        std::vector<const shady::Node*> operands;
+        for (auto arg : args)
+            operands.push_back(emit_bb(bb, arg));
+        std::vector<const shady::Node*> type_arguments;
+        for (auto type_arg : types)
+            type_arguments.push_back(convert(type_arg));
+        payload.operands = vec2nodes(operands);
+        payload.type_arguments = vec2nodes(type_arguments);
+        return shady::first(shady::bind_instruction(bb.builder, shady::prim_op(arena, payload)));
+    };
+
     if (auto prim_lit = def->isa<PrimLit>()) {
         const auto& box = prim_lit->value();
         switch (prim_lit->primtype_tag()) {
@@ -392,6 +407,28 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
         payload.element_type = convert(arr->elem_type());
         payload.contents = vec2nodes(contents);
         v = shady::arr_lit(arena, payload);
+    } else if (auto cmp = def->isa<Cmp>()) {
+        switch (cmp->cmp_tag()) {
+            case Cmp_eq: v = mk_primop(shady::Op::eq_op,  { cmp->lhs(), cmp->rhs() }); break;
+            case Cmp_ne: v = mk_primop(shady::Op::neq_op, { cmp->lhs(), cmp->rhs() }); break;
+            case Cmp_gt: v = mk_primop(shady::Op::gt_op,  { cmp->lhs(), cmp->rhs() }); break;
+            case Cmp_ge: v = mk_primop(shady::Op::gte_op, { cmp->lhs(), cmp->rhs() }); break;
+            case Cmp_lt: v = mk_primop(shady::Op::lt_op,  { cmp->lhs(), cmp->rhs() }); break;
+            case Cmp_le: v = mk_primop(shady::Op::lte_op, { cmp->lhs(), cmp->rhs() }); break;
+        }
+    } else if (auto arith = def->isa<ArithOp>()) {
+        switch (arith->arithop_tag()) {
+            case ArithOp_add: v = mk_primop(shady::Op::add_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_sub: v = mk_primop(shady::Op::sub_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_mul: v = mk_primop(shady::Op::mul_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_div: v = mk_primop(shady::Op::div_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_rem: v = mk_primop(shady::Op::mod_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_and: v = mk_primop(shady::Op::and_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_or:  v = mk_primop(shady::Op::or_op,     { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_xor: v = mk_primop(shady::Op::xor_op,    { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_shl: v = mk_primop(shady::Op::lshift_op, { arith->lhs(), arith->rhs() }); break;
+            case ArithOp_shr: v = mk_primop(shady::Op::rshift_logical_op, { arith->lhs(), arith->rhs() }); break;
+        }
     }
     assert(v && shady::is_value(v));
     defs_[def] = v;

From 494598c4328e8eca093ca588b549b7b94652127e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Sat, 3 Dec 2022 12:13:47 +0100
Subject: [PATCH 135/342] fix some broken codegen

---
 src/thorin/be/shady/shady.cpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index e1485bf46..2f1d04a1d 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -371,7 +371,7 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
         payload.op = op;
         std::vector<const shady::Node*> operands;
         for (auto arg : args)
-            operands.push_back(emit_bb(bb, arg));
+            operands.push_back(emit(arg));
         std::vector<const shady::Node*> type_arguments;
         for (auto type_arg : types)
             type_arguments.push_back(convert(type_arg));
@@ -429,6 +429,10 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             case ArithOp_shl: v = mk_primop(shady::Op::lshift_op, { arith->lhs(), arith->rhs() }); break;
             case ArithOp_shr: v = mk_primop(shady::Op::rshift_logical_op, { arith->lhs(), arith->rhs() }); break;
         }
+    } else if (auto param = def->isa<Param>()) {
+        assert(param->type() == world().mem_type());
+        defs_[def] = nullptr;
+        return nullptr;
     }
     assert(v && shady::is_value(v));
     defs_[def] = v;

From f801e8812daf4d632f2014c597b9d059403f5a96 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Sat, 3 Dec 2022 16:30:41 +0100
Subject: [PATCH 136/342] changed references to 'spirv' to 'shady'

---
 src/thorin/be/codegen.cpp   | 12 ++++++------
 src/thorin/be/llvm/llvm.cpp | 34 +++++++++++++++++-----------------
 src/thorin/continuation.cpp |  2 +-
 src/thorin/continuation.h   |  2 +-
 src/thorin/type.h           |  2 +-
 5 files changed, 26 insertions(+), 26 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index b804fa6e6..fbbc29c4b 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -92,12 +92,12 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         Continuation* imported = nullptr;
 
         static const auto backend_intrinsics = std::array {
-            std::pair { CUDA,   Intrinsic::CUDA   },
-            std::pair { NVVM,   Intrinsic::NVVM   },
-            std::pair { OpenCL, Intrinsic::OpenCL },
-            std::pair { AMDGPU, Intrinsic::AMDGPU },
-            std::pair { HLS,    Intrinsic::HLS    },
-            std::pair { Shady,  Intrinsic::SpirV  }
+            std::pair { CUDA,   Intrinsic::CUDA         },
+            std::pair { NVVM,   Intrinsic::NVVM         },
+            std::pair { OpenCL, Intrinsic::OpenCL       },
+            std::pair { AMDGPU, Intrinsic::AMDGPU       },
+            std::pair { HLS,    Intrinsic::HLS          },
+            std::pair { Shady,  Intrinsic::ShadyCompute }
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index c1ea92dc4..df1f0e876 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1146,23 +1146,23 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
     }
 
     switch (callee->intrinsic()) {
-        case Intrinsic::Atomic:      return emit_atomic(irbuilder, continuation);
-        case Intrinsic::AtomicLoad:  return emit_atomic_load(irbuilder, continuation);
-        case Intrinsic::AtomicStore: return emit_atomic_store(irbuilder, continuation);
-        case Intrinsic::CmpXchg:     return emit_cmpxchg(irbuilder, continuation, false);
-        case Intrinsic::CmpXchgWeak: return emit_cmpxchg(irbuilder, continuation, true);
-        case Intrinsic::Fence:       return emit_fence(irbuilder, continuation);
-        case Intrinsic::Reserve:     return emit_reserve(irbuilder, continuation);
-        case Intrinsic::CUDA:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation);
-        case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
-        case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
-        case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
-        case Intrinsic::SpirV:       return runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation);
-        case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
-        case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
-        case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);
-        case Intrinsic::Spawn:       return emit_spawn(irbuilder, continuation);
-        case Intrinsic::Sync:        return emit_sync(irbuilder, continuation);
+        case Intrinsic::Atomic:       return emit_atomic(irbuilder, continuation);
+        case Intrinsic::AtomicLoad:   return emit_atomic_load(irbuilder, continuation);
+        case Intrinsic::AtomicStore:  return emit_atomic_store(irbuilder, continuation);
+        case Intrinsic::CmpXchg:      return emit_cmpxchg(irbuilder, continuation, false);
+        case Intrinsic::CmpXchgWeak:  return emit_cmpxchg(irbuilder, continuation, true);
+        case Intrinsic::Fence:        return emit_fence(irbuilder, continuation);
+        case Intrinsic::Reserve:      return emit_reserve(irbuilder, continuation);
+        case Intrinsic::CUDA:         return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation);
+        case Intrinsic::NVVM:         return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
+        case Intrinsic::OpenCL:       return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
+        case Intrinsic::AMDGPU:       return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
+        case Intrinsic::ShadyCompute: return runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation);
+        case Intrinsic::HLS:          return emit_hls(irbuilder, continuation);
+        case Intrinsic::Parallel:     return emit_parallel(irbuilder, continuation);
+        case Intrinsic::Fibers:       return emit_fibers(irbuilder, continuation);
+        case Intrinsic::Spawn:        return emit_spawn(irbuilder, continuation);
+        case Intrinsic::Sync:         return emit_sync(irbuilder, continuation);
 #if THORIN_ENABLE_RV
         case Intrinsic::Vectorize:   return emit_vectorize_continuation(irbuilder, continuation);
 #else
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index e6429b480..ca7675235 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -215,7 +215,7 @@ void Continuation::set_intrinsic() {
     else if (name() == "nvvm")           attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")         attributes().intrinsic = Intrinsic::OpenCL;
     else if (name() == "amdgpu")         attributes().intrinsic = Intrinsic::AMDGPU;
-    else if (name() == "spirv")          attributes().intrinsic = Intrinsic::SpirV;
+    else if (name() == "shady_compute")  attributes().intrinsic = Intrinsic::ShadyCompute;
     else if (name() == "hls")            attributes().intrinsic = Intrinsic::HLS;
     else if (name() == "parallel")       attributes().intrinsic = Intrinsic::Parallel;
     else if (name() == "fibers")         attributes().intrinsic = Intrinsic::Fibers;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index ed0ed113a..353d04576 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -91,7 +91,7 @@ enum class Intrinsic : uint8_t {
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
     AMDGPU,                     ///< Internal AMDGPU-Backend.
-    SpirV,                      ///< Internal Vulkan-Compute-Shader-Backend.
+    ShadyCompute,               ///< Internal Shady Compute Backend.
     HLS,                        ///< Internal HLS-Backend.
     Parallel,                   ///< Internal Parallel-CPU-Backend.
     Fibers,                     ///< Internal Parallel-CPU-Backend using resumable fibers.
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 5b96929bc..0d313d9f7 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -235,7 +235,7 @@ enum class AddrSpace : uint32_t {
     Texture  = 2,
     Shared   = 3,
     Constant = 4,
-    Private =  5, // Corresponds to the 'private' storage class in SPIR-V
+    Private =  5, // Corresponds to the 'private' storage class in compute kernels/shaders, as in thread-private
 };
 
 /// Pointer type.

From 0853d32d573d4d9ac9b3c4a8a0799b130df92f0e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Sun, 4 Dec 2022 18:14:47 +0100
Subject: [PATCH 137/342] qualify fn return types

---
 src/thorin/be/shady/shady.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 2f1d04a1d..0dc62c600 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -175,6 +175,11 @@ shady::Node* CodeGen::emit_decl_head(Def* def) {
             auto ret_type = convert(t);
             if (!ret_type)
                 continue; // Eliminate mem types
+
+            shady::QualifiedType qtype;
+            qtype.type = ret_type;
+            qtype.is_uniform = false;
+            ret_type = shady::qualified_type(arena, qtype);
             returns.push_back(ret_type);
         }
 

From f94c74ad052bc7a87846f51ecf40b7bb6672b601 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Sun, 4 Dec 2022 18:14:52 +0100
Subject: [PATCH 138/342] use a jump instead of let_into

---
 src/thorin/be/shady/shady.cpp | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 0dc62c600..80be8bab0 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -312,7 +312,6 @@ void CodeGen::emit_epilogue(Continuation* cont) {
         args.erase(args.begin() + ret_param);
         args.erase(std::remove_if(args.begin(), args.end(), [&](const auto& item){ return item == nullptr || !shady::is_value(item); }), args.end());
 
-        const shady::Node* call;
         if (auto callee = body->callee()->isa_nom<Continuation>()) {
             // shady primop called as imported continuations look like continuation calls to thorin, but not to shady
             // we just need to carefully emit the primop as an instruction, then jump to the target BB, passing the stuff as we do
@@ -327,20 +326,20 @@ void CodeGen::emit_epilogue(Continuation* cont) {
                 bb.terminator = shady::jump(arena, jump);
                 return;
             }
-
-            shady::LeafCall payload;
-            payload.args = vec2nodes(args);
-            payload.callee = emit(callee);
-            call = shady::leaf_call(arena, payload);
-        } else {
-            shady::IndirectCall payload;
-            payload.args = vec2nodes(args);
-            payload.callee = emit(callee);
-            call = shady::indirect_call(arena, payload);
         }
 
+        shady::BodyBuilder* builder = shady::begin_body(module);
+
+        shady::IndirectCall icall_payload;
+        icall_payload.args = vec2nodes(args);
+        icall_payload.callee = emit(body->callee());
+        shady::Nodes results = shady::bind_instruction(builder, shady::indirect_call(arena, icall_payload));
+
         assert(args[ret_param]->tag == shady::BasicBlock_TAG);
-        bb.terminator = shady::let_into(arena, call, args[ret_param]);
+        shady::Jump jump_payload;
+        jump_payload.target = args[ret_param];
+        jump_payload.args = results;
+        bb.terminator = shady::finish_body(builder, shady::jump(arena, jump_payload));
     }
 }
 

From e6bd868bb5fe72f8245fa9660a5376d7bc6ba0a2 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 8 Dec 2022 11:19:27 +0100
Subject: [PATCH 139/342] Allow changing the init definition of a Global.

---
 src/thorin/primop.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/primop.h b/src/thorin/primop.h
index f12fc74c0..db117779a 100644
--- a/src/thorin/primop.h
+++ b/src/thorin/primop.h
@@ -535,6 +535,7 @@ class Global : public Def {
     const char* op_name() const override;
 
     bool is_external() const;
+    void set_init(const Def* new_init) { unset_op(0); set_op(0, new_init); }
 
 private:
     hash_t vhash() const override { return murmur3(gid()); }

From c59f311abd0ada73fe3fd35c6aa04f03d7c7d6c8 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 8 Dec 2022 15:53:12 +0100
Subject: [PATCH 140/342] Resolve naming conflicts for definite and indefinite
 array types.

---
 src/thorin/be/json/json.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index ae9d9d1de..8e3665a65 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -21,13 +21,13 @@ class TypeTable {
             result["type"] = "def_array";
             result["args"] = { elem_type };
             result["length"] = arr->dim();
-            result["name"] = elem_type + "_darr";
+            result["name"] = elem_type + "_darr_" + std::to_string(type_table.size());
         } else if (auto arr = type->isa<IndefiniteArrayType>()) {
             auto elem_type = translate_type(arr->elem_type());
 
             result["type"] = "indef_array";
             result["args"] = { elem_type };
-            result["name"] = elem_type + "_iarr";
+            result["name"] = elem_type + "_iarr_" + std::to_string(type_table.size());
         } else if (type->isa<BottomType>()) {
             result["name"] = "bottom_t";
             result["type"] = "bottom";

From bda2713094147b4640c07e072ae2a43faf0849f6 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 8 Dec 2022 20:30:46 +0100
Subject: [PATCH 141/342] Json backend fix for newmem.

---
 src/thorin/be/json/json.cpp | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 8e3665a65..1ab646965 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -160,11 +160,8 @@ class DefTable {
                     result["type"] = "continuation";
                     result["intrinsic"] = "branch";
                 } else if (cont->intrinsic() == Intrinsic::Match) {
-                    //TODO: These will change in the memory branch!
-                    assert(!is_mem(cont->param(0)));
-
-                    size_t num_patterns = cont->num_params() - 2;
-                    auto variant_type = type_table_.translate_type(cont->param(0)->type());
+                    size_t num_patterns = cont->num_params() - 3;
+                    auto variant_type = type_table_.translate_type(cont->param(1)->type());
                     auto name = "_match_" + std::to_string(def_table.size());
 
                     result["name"] = name;

From 1b91c0379d6a9406e7260895d59ac1b9f208a2e9 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 15 Dec 2022 13:55:04 +0100
Subject: [PATCH 142/342] Some small mistakes in world.h.

---
 src/thorin/world.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/world.h b/src/thorin/world.h
index 311279a30..fbdb2c3c5 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -90,10 +90,10 @@ class World : public TypeTable, public Streamable<World> {
     //@{
     bool empty() { return data_.externals_.empty(); }
     const Externals& externals() const { return data_.externals_; }
-    void make_external(Def* cont) { data_.externals_.emplace(cont->unique_name(), cont).second; }
+    void make_external(Def* cont) { data_.externals_.emplace(cont->unique_name(), cont); }
     void make_internal(Def* cont) { data_.externals_.erase(cont->unique_name()); }
     bool is_external(const Def* cont) { return data_.externals_.contains(cont->unique_name()); }
-    Def* lookup(const std::string& name) { return const_cast<Def*>(data_.externals_.lookup(name).value_or(nullptr)); }
+    Def* lookup(const std::string& name) { return data_.externals_.lookup(name).value_or(nullptr); }
     //@}
 
     // literals

From 278c4fa2904e950cdd40e05e86104343d99a8e44 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 3 Jan 2023 15:43:53 +0100
Subject: [PATCH 143/342] [Json]: Forward declarations for nominal types.

---
 src/thorin/be/json/json.cpp | 46 ++++++++++++++++++++++++++++++-------
 1 file changed, 38 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 8e3665a65..44df89063 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -4,6 +4,7 @@ namespace thorin::json {
 
 class TypeTable {
 public:
+    json nominal_fwd_table = json::array();
     json type_table = json::array();
 
     TypeMap<std::string> known_types;
@@ -56,28 +57,54 @@ class TypeTable {
             result["name"] = "mem_t";
             result["type"] = "mem";
         } else if (auto structtype = type->isa<StructType>()) {
-            json args = json::array();
+            auto name = "_struct_" + std::to_string(nominal_fwd_table.size());
+            known_types[type] = name;
+
             json arg_names = json::array();
             for (size_t i = 0; i < structtype->num_ops(); ++i) {
-                args.push_back(translate_type(structtype->op(i)));
                 arg_names.push_back(structtype->op_name(i).str());
             }
 
+            json forward_decl;
+            forward_decl["name"] = name;
+            forward_decl["type"] = "struct";
+            forward_decl["struct_name"] = structtype->name().str();
+            forward_decl["arg_names"] = arg_names;
+            nominal_fwd_table.push_back(forward_decl);
+
+            json args = json::array();
+            for (size_t i = 0; i < structtype->num_ops(); ++i) {
+                args.push_back(translate_type(structtype->op(i)));
+            }
+
             result["type"] = "struct";
-            result["name"] = "_" + std::to_string(type_table.size());
+            result["name"] = name;
             result["struct_name"] = structtype->name().str();
-            result["args"] = args;
             result["arg_names"] = arg_names;
+            result["args"] = args;
         } else if (auto varianttype = type->isa<VariantType>()) {
-            json args = json::array();
+            auto name = "_variant_" + std::to_string(nominal_fwd_table.size());
+            known_types[type] = name;
+
             json arg_names = json::array();
             for (size_t i = 0; i < varianttype->num_ops(); ++i) {
-                args.push_back(translate_type(varianttype->op(i)));
                 arg_names.push_back(varianttype->op_name(i).str());
             }
 
+            json forward_decl;
+            forward_decl["name"] = name;
+            forward_decl["type"] = "variant";
+            forward_decl["variant_name"] = varianttype->name().str();
+            forward_decl["arg_names"] = arg_names;
+            nominal_fwd_table.push_back(forward_decl);
+
+            json args = json::array();
+            for (size_t i = 0; i < varianttype->num_ops(); ++i) {
+                args.push_back(translate_type(varianttype->op(i)));
+            }
+
             result["type"] = "variant";
-            result["name"] = "_" + std::to_string(type_table.size());
+            result["name"] = name;
             result["variant_name"] = varianttype->name().str();
             result["args"] = args;
             result["arg_names"] = arg_names;
@@ -627,7 +654,10 @@ void CodeGen::emit_stream(std::ostream& stream) {
         def_table.translate_def(external.second);
     }
 
-    j["type_table"] = type_table.type_table;
+    j["type_table"] = type_table.nominal_fwd_table;
+    for (auto it : type_table.type_table)
+        j["type_table"] += it;
+
     j["defs"] = def_table.decl_table;
     for (auto it : def_table.def_table)
         j["defs"] += it;

From 8fa238d591abdeda7c316b37843e494c99fae009 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 3 Jan 2023 17:30:06 +0100
Subject: [PATCH 144/342] Increase inliner limits to prevent global structs
 from ocurring in tests.

---
 src/thorin/transform/inliner.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/transform/inliner.cpp b/src/thorin/transform/inliner.cpp
index 175f2a518..37d858807 100644
--- a/src/thorin/transform/inliner.cpp
+++ b/src/thorin/transform/inliner.cpp
@@ -39,8 +39,8 @@ void force_inline(Scope& scope, int threshold) {
 void inliner(World& world) {
     world.VLOG("start inliner");
 
-    static const int factor = 4;
-    static const int offset = 4;
+    static const int factor = 8;
+    static const int offset = 8;
 
     ContinuationMap<std::unique_ptr<Scope>> continuation2scope;
 

From 6ec9748baba499f352dd58e62927514eba6aaba9 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 4 Jan 2023 17:18:18 +0100
Subject: [PATCH 145/342] Allow generation of global structs by constructing
 ConstantStruct when possible.

---
 src/thorin/be/llvm/llvm.cpp | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index b3967860e..a369de7b4 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -753,6 +753,25 @@ llvm::Value* CodeGen::emit_builder(llvm::IRBuilder<>& irbuilder, const Def* def)
         assert(def->isa<Tuple>() || def->isa<StructAgg>() || def->isa<Vector>() || def->isa<Closure>());
         if (is_unit(agg)) return nullptr;
 
+        if (def->isa<StructAgg>() || def->isa<Vector>()) {
+            // Try to emit it as a constant first
+            Array<llvm::Constant*> consts(agg->num_ops());
+            bool all_consts = true;
+            for (size_t i = 0, n = consts.size(); i != n; ++i) {
+                consts[i] = llvm::dyn_cast<llvm::Constant>(emit(agg->op(i)));
+                if (!consts[i]) {
+                    all_consts = false;
+                    break;
+                }
+            }
+            if (all_consts) {
+                if (def->isa<StructAgg>())
+                    return llvm::ConstantStruct::get(llvm::cast<llvm::StructType>(convert(agg->type())), llvm_ref(consts));
+                else
+                    return llvm::ConstantVector::get(llvm_ref(consts));
+            }
+        }
+
         llvm::Value* llvm_agg = llvm::UndefValue::get(convert(agg->type()));
         if (def->isa<Vector>()) {
             for (size_t i = 0, e = agg->num_ops(); i != e; ++i)
@@ -863,6 +882,11 @@ llvm::Value* CodeGen::emit_builder(llvm::IRBuilder<>& irbuilder, const Def* def)
         auto llvm_type = convert(variant_ctor->type());
         auto tag_value = irbuilder.getIntN(llvm_type->getStructElementType(1)->getScalarSizeInBits(), variant_ctor->index());
 
+        //Unit type variants can be emitted as constants.
+        if (is_type_unit(variant_ctor->op(0)->type())) {
+            return llvm::ConstantStruct::get(llvm::cast<llvm::StructType>(llvm_type), {llvm::UndefValue::get(llvm_type->getStructElementType(0)), tag_value});
+        }
+
         return create_tmp_alloca(irbuilder, llvm_type, [&] (llvm::AllocaInst* alloca) {
             auto tag_addr = irbuilder.CreateInBoundsGEP(llvm_type, alloca, { irbuilder.getInt32(0), irbuilder.getInt32(1) });
             irbuilder.CreateStore(tag_value, tag_addr);

From 9eaba5daceac48e191683268aa3d8a4ed28feeb9 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 4 Jan 2023 17:50:11 +0100
Subject: [PATCH 146/342] hoist_enters + importer cleanup.

---
 src/thorin/transform/hoist_enters.cpp | 10 +++++-----
 src/thorin/transform/importer.cpp     |  9 +--------
 src/thorin/transform/importer.h       |  8 ++++++++
 3 files changed, 14 insertions(+), 13 deletions(-)

diff --git a/src/thorin/transform/hoist_enters.cpp b/src/thorin/transform/hoist_enters.cpp
index 5cd7fcfb4..7e8ab216e 100644
--- a/src/thorin/transform/hoist_enters.cpp
+++ b/src/thorin/transform/hoist_enters.cpp
@@ -8,7 +8,7 @@
 
 namespace thorin {
 
-std::stack<const Def*> todo;
+static std::stack<const Def*> hoist_enters_todo;
 
 static void find_enters(std::deque<const Enter*>& enters, const Def* def) {
     if (auto enter = def->isa<Enter>())
@@ -19,15 +19,15 @@ static void find_enters(std::deque<const Enter*>& enters, const Def* def) {
 
     for (auto use : def->uses()) {
         if (auto memop = use->isa<MemOp>())
-            todo.push(memop);
+            hoist_enters_todo.push(memop);
     }
 }
 
 static void find_enters(std::deque<const Enter*>& enters, Continuation* continuation) {
     if (auto mem_param = continuation->mem_param()) {
-        todo.push(mem_param);
-        while (!todo.empty()) {
-            auto next_item = pop(todo);
+        hoist_enters_todo.push(mem_param);
+        while (!hoist_enters_todo.empty()) {
+            auto next_item = pop(hoist_enters_todo);
 
             find_enters(enters, next_item);
         }
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index d476fb718..976e1c77f 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -1,9 +1,5 @@
 #include "thorin/transform/importer.h"
 
-#include <set>
-#include <stack>
-#include <utility>
-
 namespace thorin {
 
 const Type* Importer::import_type(const Type* otype) {
@@ -32,10 +28,7 @@ const Type* Importer::import_type(const Type* otype) {
     return ntype;
 }
 
-std::stack<std::pair<const Def*, bool>> required_defs;
-std::set<const Def *> analyzed_conts;
-
-void enqueue(const Def* elem) {
+void Importer::enqueue(const Def* elem) {
     if (elem->isa_nom<Continuation>()) {
         if (analyzed_conts.find(elem) != analyzed_conts.end()) {
             required_defs.push(std::pair(elem, false));
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index eb7d5bc42..8d132a75b 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -4,6 +4,10 @@
 #include "thorin/world.h"
 #include "thorin/config.h"
 
+#include <set>
+#include <stack>
+#include <utility>
+
 namespace thorin {
 
 class Importer {
@@ -26,6 +30,10 @@ class Importer {
 
 private:
     const Def* import_nonrecursive();
+    void enqueue(const Def* elem);
+
+    std::stack<std::pair<const Def*, bool>> required_defs;
+    std::set<const Def*> analyzed_conts;
 
 public:
     Type2Type type_old2new_;

From 1de509b5486bcf5b9b33ce8c8737288a9f56aa56 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Tue, 10 Jan 2023 15:32:06 +0100
Subject: [PATCH 147/342] update to latest shady API

---
 src/thorin/be/shady/shady.cpp | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 80be8bab0..f3d060e43 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -407,10 +407,11 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             assert(emit(e));
             contents.push_back(emit(e));
         }
-        shady::ArrayLiteral payload;
+        shady::ArrType payload;
+        const shady::Type* arr_type = shady::arr_type(arena, payload);
         payload.element_type = convert(arr->elem_type());
-        payload.contents = vec2nodes(contents);
-        v = shady::arr_lit(arena, payload);
+        payload.size = shady::int32_literal(arena, contents.size());
+        v = shady::composite(arena, arr_type, vec2nodes(contents));
     } else if (auto cmp = def->isa<Cmp>()) {
         switch (cmp->cmp_tag()) {
             case Cmp_eq: v = mk_primop(shady::Op::eq_op,  { cmp->lhs(), cmp->rhs() }); break;

From c0961acab3fef38fb88139a219771943fdfbe144 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 11 Jan 2023 13:22:11 +0100
Subject: [PATCH 148/342] remove structurizer

---
 src/thorin/CMakeLists.txt            |   2 -
 src/thorin/be/shady/shady.cpp        |   6 +-
 src/thorin/transform/structurize.cpp | 466 ---------------------------
 src/thorin/transform/structurize.h   |  12 -
 4 files changed, 1 insertion(+), 485 deletions(-)
 delete mode 100644 src/thorin/transform/structurize.cpp
 delete mode 100644 src/thorin/transform/structurize.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index de8280c81..43537af49 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -70,8 +70,6 @@ set(THORIN_SOURCES
     transform/partial_evaluation.h
     transform/split_slots.cpp
     transform/split_slots.h
-    transform/structurize.cpp
-    transform/structurize.h
     transform/hls_channels.cpp
     transform/hls_channels.h
     transform/hls_kernel_launch.h
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index f3d060e43..89dab73f0 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -2,7 +2,6 @@
 #undef empty
 
 #include "thorin/analyses/scope.h"
-#include "thorin/transform/structurize.h"
 
 namespace thorin::shady_be {
 
@@ -11,12 +10,9 @@ CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
 {}
 
 void CodeGen::emit_stream(std::ostream& out) {
-    // structure_loops(world());
-    // structure_flow(world());
-
     assert(!module);
 
-    shady::ArenaConfig config = { };
+    shady::ArenaConfig config = shady::default_arena_config();
     config.name_bound = true;
     config.check_types = true;
     arena = shady::new_ir_arena(config);
diff --git a/src/thorin/transform/structurize.cpp b/src/thorin/transform/structurize.cpp
deleted file mode 100644
index 38ca453bd..000000000
--- a/src/thorin/transform/structurize.cpp
+++ /dev/null
@@ -1,466 +0,0 @@
-#include "structurize.h"
-
-#include <unordered_map>
-
-#include "thorin/analyses/domtree.h"
-#include "thorin/world.h"
-
-namespace thorin {
-
-using Head = LoopTree<true>::Head;
-using Base = LoopTree<true>::Base;
-using Leaf = LoopTree<true>::Leaf;
-
-struct StructuredLoop;
-
-// Dispatch targets may dispatch to other dispatching nodes, and we can't actually create those until we know all their destinations,
-// because their fn type takes a variant type with a case for each target. So we symbolically refer to these yet-to-be dispatch nodes via their loop
-struct DispatchTarget {
-    bool operator==(const DispatchTarget& rhs) const {
-        return cont == rhs.cont &&entry == rhs.entry &&exit == rhs.exit;
-    }
-
-    Continuation* cont = nullptr;
-    StructuredLoop* entry = nullptr;
-    StructuredLoop* exit = nullptr;
-};
-
-struct RewireMe {
-    explicit RewireMe(Continuation* cont) : cont(cont) {}
-    Continuation* cont;
-
-    Continuation* backedge = nullptr;
-    struct {
-        std::vector<StructuredLoop*> exits;
-        std::vector<StructuredLoop*> enters;
-        Continuation* final_destination = nullptr;
-    } non_local_jump;
-};
-
-// Represents one loop in the loop forest, that we then augment with a new loop header and epilogue, each dispatching
-// respectively to nodes inside of the loop, and nodes outside of the loop once we break out
-struct StructuredLoop {
-    const Head* parent_head;
-    const Head* head;
-    const std::string name;
-    StructuredLoop(const Head* parent_head, const Head* head, std::string&& name)
-    : parent_head(parent_head), head(head), name(name) {}
-
-    std::vector<DispatchTarget> inner_destinations = {};
-    std::vector<DispatchTarget> outer_destinations = {};
-
-    /// Just a regular continuation that simply calls into the real header
-    Continuation* pre_header;
-    /// Calls the loop_enter intrinsic with the set of internal and external dispatch nodes
-    Continuation* real_header;
-    Continuation* exit;
-
-    // Same as the inner/outer destinations, but entry/exits instead now point to the corresponding header/epilogue nodes
-    std::vector<const Continuation*> header_destination_conts;
-    std::vector<const Continuation*> epilogue_destination_conts;
-
-    std::vector<RewireMe> rewire;
-};
-
-struct ScopeContext {
-    explicit ScopeContext(const Scope& scope)
-    : cfa(scope)
-    {}
-
-    CFA cfa;
-    ContinuationMap<const Head*> def2loop;
-    std::unordered_map<const Head*, StructuredLoop> rewritten_loops;
-};
-
-inline std::string loop_name(const Head* head) {
-    if (head == nullptr || head->is_root()) {
-        return "root";
-    } else {
-        std::stringstream s;
-        s << "loop_";
-        for (auto& node : head->cf_nodes()) {
-            s << node->continuation()->to_string();
-            s << "_";
-        }
-        return s.str();
-    }
-}
-
-/// Visits the forest and fills def2loop
-inline void tag_continuations(ScopeContext& ctx, const Base* base, const Head* parent) {
-    if (const Head* head = base->isa<Head>()) {
-        auto name = loop_name(head);
-
-        for (auto& children : head->children()) {
-            tag_continuations(ctx, &*children, head);
-        }
-
-        StructuredLoop loop(parent, head, std::move(name));
-        ctx.rewritten_loops.emplace(head, loop);
-    } else if(base->isa<Leaf>()) {
-        for (auto& node : base->cf_nodes()) {
-            auto[i, result] = ctx.def2loop.emplace(node->continuation(), parent);
-            assert(result);
-        }
-    } else {
-        assert(false);
-    }
-}
-
-inline int record_destination(std::vector<DispatchTarget>& vec, DispatchTarget dest) {
-    auto i = std::find(vec.begin(), vec.end(), dest);
-    if (i == vec.end()) {
-        vec.emplace_back(dest);
-        return static_cast<int>(vec.size()) - 1;
-    } else return i - vec.begin();
-}
-
-inline int index_of_destination(std::vector<DispatchTarget>& vec, DispatchTarget dest) {
-    auto i = std::find(vec.begin(), vec.end(), dest);
-    if (i == vec.end()) {
-        assert(false && "Missing destination");
-    } else return i - vec.begin();
-}
-
-inline std::vector<StructuredLoop*> get_path(ScopeContext& ctx, const Head* head) {
-    std::vector<StructuredLoop*> path = {};
-    assert(head != nullptr);
-    while (head != nullptr) {
-        StructuredLoop* loop = &ctx.rewritten_loops.find(head)->second;
-        assert(loop != nullptr);
-        path.emplace(path.begin(), loop);
-        head = loop->parent_head;
-    }
-    return path;
-}
-
-inline void collect_dispatch_targets(World& world, ScopeContext& ctx, const Base* base) {
-    if (const Head* head = base->isa<Head>()) {
-        for (auto& children : head->children()) {
-            collect_dispatch_targets(world, ctx, &*children);
-        }
-    } else {
-        const Leaf* leaf = base->as<Leaf>();
-        auto cont = leaf->cf_node()->continuation();
-        // For some nonsense reason, synthetic nodes created during scopes iteration leak in next iterations >:(
-        if (!cont->has_body() /*|| (cont->intrinsic() >= Intrinsic::SCFBegin && cont->intrinsic() < Intrinsic::SCFEnd)*/)
-            return;
-        auto app = cont->body();
-        auto callee = app->callee()->isa_nom<Continuation>();
-        if (!callee || callee->intrinsic() == Intrinsic::Branch || callee->is_imported())
-            return;
-
-        const Head* source_loop_head = ctx.def2loop[cont];
-        assert(ctx.def2loop.find(callee) != ctx.def2loop.end());
-        const Head* dest_loop_head = ctx.def2loop[callee];
-
-        if (source_loop_head != dest_loop_head) {
-            // We found a non-local jump
-            assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
-            auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
-
-            std::vector<StructuredLoop*> source_path = get_path(ctx, source_loop_head);
-            std::vector<StructuredLoop*> dest_path = get_path(ctx, dest_loop_head);
-            int bi = 0;
-            while (bi < static_cast<int>(std::min(source_path.size(), dest_path.size()))) {
-                if (source_path[bi] == dest_path[bi])
-                    bi++;
-                else break;
-            }
-
-            // The path is made out of a sequence of loops to break out of, and a sequence of loops to jump into
-            // these two sequences cannot be both empty (that wouldn't be a non-local jump then!)
-            std::vector<StructuredLoop*> leave;
-            std::vector<StructuredLoop*> enter;
-            for (int j = static_cast<int>(source_path.size()) - 1; j >= bi; j--)
-                leave.emplace_back(source_path[j]);
-            for (int j = bi; j < static_cast<int>(dest_path.size()); j++)
-                enter.emplace_back(dest_path[j]);
-
-            // 0 = this is the first step of the path
-            // 1 = last step was to break out of a loop
-            // 2 = last step was to enter a loop
-            int last = 0;
-            StructuredLoop* prev;
-
-            auto record_step = [&](DispatchTarget destination) {
-                if (last == 0) {
-                    // nothing to do, this node isn't a dispatching one
-                } else {
-                    if (last == 1)
-                        record_destination(prev->outer_destinations, destination);
-                    else
-                        record_destination(prev->inner_destinations, destination);
-                }
-            };
-
-            for (auto dest_loop : leave) {
-                DispatchTarget destination;
-                destination.exit = dest_loop;
-
-                record_step(destination);
-                last = 1;
-                prev = dest_loop;
-                assert(prev != nullptr);
-            }
-            for (auto dest_loop : enter) {
-                DispatchTarget destination;
-                destination.entry = dest_loop;
-
-                record_step(destination);
-                last = 2;
-                prev = dest_loop;
-                assert(prev != nullptr);
-            }
-
-            assert(last != 0);
-            DispatchTarget destination;
-            destination.cont = callee;
-            record_step(destination);
-
-            RewireMe rewire(cont);
-            rewire.non_local_jump = {
-                std::move(leave),
-                std::move(enter),
-                callee
-            };
-            loop.rewire.emplace_back(rewire);
-        } else if (source_loop_head == dest_loop_head && source_loop_head != nullptr) {
-            for (auto& cf_node : source_loop_head->cf_nodes()) {
-                if (cf_node->continuation() == callee) {
-                    // We found a backedge
-                    assert(ctx.rewritten_loops.find(source_loop_head) != ctx.rewritten_loops.end());
-                    auto& loop = ctx.rewritten_loops.find(source_loop_head)->second;
-
-                    DispatchTarget destination;
-                    destination.cont = callee;
-                    record_destination(loop.inner_destinations, destination);
-
-                    RewireMe rewire(cont);
-                    rewire.backedge = cf_node->continuation();
-                    loop.rewire.emplace_back(rewire);
-                    return;
-                }
-            }
-            assert(false);
-        }
-    }
-}
-
-inline const Type* dom_to_tuple(World& world, const thorin::FnType* fn_type) {
-    std::vector<const thorin::Type*> t;
-    t.resize(fn_type->num_ops());
-    for (size_t i = 0; i < fn_type->num_ops(); i++)
-        t[i] = fn_type->op(i);
-    return world.tuple_type(t);
-}
-
-inline const Def* tuple_from_params(World& world, const ArrayRef<const Param*> params) {
-    std::vector<const thorin::Def*> t;
-    t.resize(params.size());
-    for (size_t i = 0; i < params.size(); i++)
-        t[i] = params[i];
-    return world.tuple(t);
-}
-
-inline void create_headers(World& world, ScopeContext& ctx, const Base* base) {
-    if (const Head* head = base->isa<Head>()) {
-        for (auto& children : head->children()) {
-            create_headers(world, ctx, &*children);
-        }
-
-        if (head->num_cf_nodes() == 0)
-            return;
-        StructuredLoop& loop = ctx.rewritten_loops.find(head)->second;
-
-        // here, parent headers need to know what they're jumping *into*
-        std::vector<const Type*> dest_types;
-        for (auto& target : loop.inner_destinations) {
-            const thorin::Continuation* target_cont;
-            if (target.cont != nullptr) {
-                target_cont = target.cont;
-            } else if (target.entry != nullptr) {
-                assert(target.entry->pre_header != nullptr);
-                target_cont = target.entry->pre_header;
-            } else {
-                assert(false && "Header dispatches may not exit loops");
-            }
-            loop.header_destination_conts.push_back(target_cont);
-            const thorin::FnType* target_type = target_cont->type();
-            dest_types.emplace_back(dom_to_tuple(world, target_type));
-        }
-        auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
-        for (size_t i = 0; i < dest_types.size(); i++)
-            variant_type->set(i, dest_types[i]);
-        Types enter_intrinsic_types = { variant_type };
-
-        //loop.enter_intrinsic = world.loop_enter(enter_intrinsic_types);
-        //loop.continue_intrinsic = world.loop_continue(enter_intrinsic_types);
-
-        loop.pre_header = world.continuation(world.fn_type(), { loop.name + "_new_header"});
-        // loop.new_continue = world.continuation(fn_type, { loop.name + "_new_continue"});
-    }
-}
-
-inline void create_epilogues(World& world, ScopeContext& ctx, const Base* base) {
-    if (const Head* head = base->isa<Head>()) {
-        StructuredLoop& loop = ctx.rewritten_loops.find(head)->second;
-
-        if (head->num_cf_nodes() > 0) {
-            // here, children epilogues need to know what they're jumping *out to*
-            std::vector<const Type*> dest_types;
-            for (auto& target : loop.outer_destinations) {
-                const thorin::Continuation* target_cont;
-                if (target.cont != nullptr) {
-                    target_cont = target.cont;
-                } else if (target.entry != nullptr) {
-                    assert(target.entry->pre_header != nullptr);
-                    target_cont = target.entry->pre_header;
-                } else {
-                    assert(target.exit != nullptr);
-                    assert(target.exit->exit != nullptr);
-                    target_cont = target.exit->exit;
-                }
-                loop.epilogue_destination_conts.push_back(target_cont);
-                const thorin::FnType* target_type = target_cont->type();
-                dest_types.emplace_back(dom_to_tuple(world, target_type));
-            }
-            auto variant_type = world.variant_type(loop.name + "_param", dest_types.size());
-            for (size_t i = 0; i < dest_types.size(); i++)
-                variant_type->set(i, dest_types[i]);
-
-            //Types break_types = {variant_type};
-            //loop.break_intrinsic = world.loop_break(break_types);
-            loop.exit = world.continuation(world.fn_type(), {loop.name + "_new_epilogue"});
-        }
-
-        for (auto& children : head->children()) {
-            create_epilogues(world, ctx, &*children);
-        }
-    }
-}
-
-// Finishes loop headers & epilogues, and re-wires backedges and non-local jumps to go through structured CF intrinsics
-inline void rewire_loops(World& world, ScopeContext& ctx, const Base* base) {
-    if (const Head* head = base->isa<Head>()) {
-        assert(ctx.rewritten_loops.find(head) != ctx.rewritten_loops.end());
-        auto& loop = ctx.rewritten_loops.find(head)->second;
-
-        if (head->num_cf_nodes() > 0) {
-            //loop.header->structured_loop_merge(loop.new_header, loop.epilogue_destination_conts);
-            //loop.exit->structured_loop_header(loop.new_epilogue, loop.new_continue, loop.header_destination_conts);
-        }
-
-        for (auto& children : head->children()) {
-            rewire_loops(world, ctx, &*children);
-        }
-
-        for (auto& rewire : loop.rewire) {
-            assert(loop.head != nullptr);
-            if (rewire.backedge != nullptr) {
-
-                DispatchTarget destination;
-                destination.cont = rewire.backedge;
-                auto variant_index = index_of_destination(loop.inner_destinations, destination);
-
-                auto old_fn_type = rewire.backedge->type();
-                auto wrapper = world.continuation(old_fn_type, {"synthetic_backedge_wrapper_to" + destination.cont->unique_name() });
-                ctx.def2loop[wrapper] = loop.head;
-                //wrapper->attributes_.intrinsic = Intrinsic::SCFBackEdge;
-
-                //TODO
-                //auto header_variant_type = loop.continue_intrinsic->type()->op(0)->as<VariantType>();
-                //wrapper->jump(loop.continue_intrinsic, { world.variant(header_variant_type, tuple_from_params(world, wrapper->params()), variant_index) });
-
-                auto old_app = rewire.cont->body();
-                assert(old_app);
-                rewire.cont->jump(wrapper, old_app->args(), old_app->debug());
-            } else {
-                auto& nlj = rewire.non_local_jump;
-
-                auto old_fn_type = nlj.final_destination->type();
-                auto wrapper = world.continuation(old_fn_type, {"synthetic_nlj_wrapper_to" + nlj.final_destination->unique_name() });
-                ctx.def2loop[wrapper] = loop.head;
-                // wrapper->attributes_.intrinsic = Intrinsic::SCFNonLocalJump;
-
-                const Def* argument = tuple_from_params(world, wrapper->params());
-                Continuation* first_jump = nullptr;
-
-                DispatchTarget destination;
-                destination.cont = nlj.final_destination;
-
-                for (int i = nlj.enters.size() - 1; i >= 0; i--) {
-                    StructuredLoop* loop_to_enter = nlj.enters[i];
-
-                    auto variant_index = index_of_destination(loop_to_enter->inner_destinations, destination);
-                    auto header_variant_type = loop_to_enter->pre_header->type()->op(0)->as<VariantType>();
-                    argument = world.variant(header_variant_type, argument, variant_index);
-
-                    first_jump = loop_to_enter->pre_header;
-                    destination = {};
-                    destination.entry = loop_to_enter;
-                }
-
-                for (int i = nlj.exits.size() - 1; i >= 0; i--) {
-                    StructuredLoop* loop_to_exit = nlj.exits[i];
-
-                    auto variant_index = index_of_destination(loop_to_exit->outer_destinations, destination);
-                    auto header_variant_type = loop_to_exit->exit->type()->op(0)->as<VariantType>();
-                    argument = world.variant(header_variant_type, argument, variant_index);
-
-                    first_jump = loop_to_exit->exit;
-                    destination = {};
-                    destination.exit = loop_to_exit;
-                }
-
-                assert(first_jump != nullptr);
-                wrapper->jump(first_jump, { argument });
-
-                auto old_app = rewire.cont->body();
-                assert(old_app);
-                rewire.cont->jump(wrapper, old_app->args(), old_app->debug());
-            }
-        }
-    }
-}
-
-void structure_loops(World& world) {
-    Scope::for_each(world, [&](Scope& scope) {
-        ScopeContext context(scope);
-
-        const LoopTree<true>& looptree = context.cfa.f_cfg().looptree();
-        tag_continuations(context, looptree.root(), nullptr);
-        collect_dispatch_targets(world, context, looptree.root());
-
-        create_headers(world, context, looptree.root());
-        create_epilogues(world, context, looptree.root());
-
-        rewire_loops(world, context, looptree.root());
-        scope.update();
-    });
-}
-
-void structure_flow(World& world) {
-    Scope::for_each(world, [&](const Scope& scope) {
-        CFA cfa(scope);
-        auto& dom_tree = cfa.f_cfg().domtree();
-        auto& post_dom_tree = cfa.b_cfg().domtree();
-
-        for (auto def : scope.defs()) {
-            if (auto cont = def->isa_nom<Continuation>()) {
-                if (cont->preds().size() <= 1)
-                    continue;
-
-                auto dominator = dom_tree.idom(cfa[cont]);
-                auto dominator_post_dominator = post_dom_tree.idom(dominator);
-                bool needs_join = dominator_post_dominator->continuation() != cont;
-                if (needs_join) {
-                    assert(false && "Not structured CF !");
-                    // TODO: insert join node into dominator and redirect dominated nodes to take it
-                }
-            }
-        }
-    });
-}
-
-}
diff --git a/src/thorin/transform/structurize.h b/src/thorin/transform/structurize.h
deleted file mode 100644
index 0f82cf9ae..000000000
--- a/src/thorin/transform/structurize.h
+++ /dev/null
@@ -1,12 +0,0 @@
-#include "thorin/analyses/looptree.h"
-#include "thorin/analyses/scope.h"
-#include "thorin/analyses/cfg.h"
-
-namespace thorin {
-
-class World;
-
-void structure_loops(World& world);
-void structure_flow(World& world);
-
-}
\ No newline at end of file

From 12890d82be31931f786851019695bc1316192bce Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 13 Jan 2023 13:46:43 +0100
Subject: [PATCH 149/342] Destroy obsolete continuations during partial
 evaluation.

---
 src/thorin/continuation.h                   | 12 ++++++++++++
 src/thorin/rec_stream.cpp                   |  1 +
 src/thorin/transform/partial_evaluation.cpp | 10 ++++++++++
 3 files changed, 23 insertions(+)

diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index ddfed253e..fc3be19c4 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -214,6 +214,18 @@ class Continuation : public Def {
         }
         return true;
     }
+    bool never_called() const {
+        for (auto use : uses()) {
+            if (auto app = use->isa<App>()) {
+                if (app->num_uses() != 0) {
+                    return false;
+                }
+            } else if (!use->isa<Param>()) {
+                return false;
+            }
+        }
+        return true;
+    }
 
     bool dead_ = false;
     std::vector<const Param*> params_;
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index d4507d775..fa4547eb6 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -55,6 +55,7 @@ void RecStreamer::run() {
             std::vector<std::string> param_names;
             for (auto param : cont->params()) param_names.push_back(param->unique_name());
             s.fmt("{}: {} = ({, }) => {{\t\n", cont->unique_name(), cont->type(), param_names);
+            run(cont->filter());
             run(cont->body()); // TODO app node
             s.fmt("\b\n}}");
         } else {
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 1b0739a7e..9f1124734 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -200,6 +200,16 @@ bool PartialEvaluator::run() {
 
                     jump_to_dropped_call(continuation, target, specialize);
 
+                    while (callee && callee->never_called()) {
+                        if (callee->has_body()) {
+                            auto new_callee = const_cast<Continuation*>(callee->body()->callee()->isa<Continuation>());
+                            callee->destroy("partial_evaluation");
+                            callee = new_callee;
+                        } else {
+                            callee = nullptr;
+                        }
+                    }
+
                     if (lower2cff_ && fold) {
                         // re-examine next iteration:
                         // maybe the specialization is not top-level anymore which might need further specialization

From 3053806191939678acb6947d68a033bf9ff59555 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 16 Jan 2023 16:59:02 +0100
Subject: [PATCH 150/342] Rewrite filter when stubs are created in mangler.

---
 src/thorin/continuation.cpp     | 3 +++
 src/thorin/continuation.h       | 2 ++
 src/thorin/transform/mangle.cpp | 3 ++-
 3 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index a33183d17..1f98867ec 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -61,7 +61,10 @@ Continuation::Continuation(const FnType* fn, const Attributes& attributes, Debug
 
 Continuation* Continuation::stub() const {
     Rewriter rewriter;
+    stub(rewriter);
+}
 
+Continuation* Continuation::stub(Rewriter& rewriter) const {
     auto result = world().continuation(type(), attributes(), debug_history());
     for (size_t i = 0, e = num_params(); i != e; ++i) {
         result->param(i)->set_name(debug_history().name);
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index fc3be19c4..6963a6672 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -13,6 +13,7 @@ namespace thorin {
 
 class Continuation;
 class Scope;
+struct Rewriter;
 
 typedef std::vector<Continuation*> Continuations;
 
@@ -138,6 +139,7 @@ class Continuation : public Def {
     const FnType* type() const { return Def::type()->as<FnType>(); }
 
     Continuation* stub() const;
+    Continuation* stub(Rewriter& rewriter) const;
     const Param* append_param(const Type* type, Debug dbg = {});
     Continuations preds() const;
     Continuations succs() const;
diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index 629134fb7..6ac5bda6e 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -122,7 +122,8 @@ Continuation* Mangler::mangle() {
 Continuation* Mangler::mangle_head(Continuation* old_continuation) {
     assert(!def2def_.contains(old_continuation));
     assert(old_continuation->has_body());
-    Continuation* new_continuation = old_continuation->stub();
+    Rewriter rewriter{def2def_};
+    Continuation* new_continuation = old_continuation->stub(rewriter);
     def2def_[old_continuation] = new_continuation;
 
     for (size_t i = 0, e = old_continuation->num_params(); i != e; ++i)

From d00f3bd1e954588b3e25df1715c00a4ae3fc8d2e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 17 Jan 2023 13:52:01 +0100
Subject: [PATCH 151/342] nuke last remnants of structured control flow support

---
 src/thorin/continuation.cpp | 25 -------------------------
 src/thorin/continuation.h   | 21 ---------------------
 src/thorin/world.h          |  4 ----
 3 files changed, 50 deletions(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index ca7675235..76a030c4a 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -262,31 +262,6 @@ void Continuation::match(const Def* val, Continuation* otherwise, Defs patterns,
     verify();
 }
 
-/*void Continuation::structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets) {
-    attributes_.intrinsic = Intrinsic::SCFLoopMerge;
-    attributes_.scf_metadata.loop_epilogue.loop_header = loop_header;
-    resize(targets.size());
-    size_t x = 0;
-    for (auto target : targets)
-        set_op(x++, target);
-}
-
-void Continuation::structured_loop_continue(const Continuation* loop_header) {
-    attributes_.intrinsic = Intrinsic::SCFLoopContinue;
-    resize(1);
-    set_op(0, loop_header);
-}
-
-void Continuation::structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets) {
-    attributes_.intrinsic = Intrinsic::SCFLoopHeader;
-    resize(targets.size());
-    attributes_.scf_metadata.loop_header.continue_target = loop_continue;
-    attributes_.scf_metadata.loop_header.merge_target = loop_epilogue;
-    size_t x = 0;
-    for (auto target : targets)
-        set_op(x++, target);
-}*/
-
 void Continuation::verify() const {
     if (!has_body())
         assertf(filter()->is_empty(), "continuations with no body should have an empty (no) filter");
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 353d04576..2bf830bc5 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -111,9 +111,6 @@ enum class Intrinsic : uint8_t {
     Pipeline,                   ///< Intrinsic loop-pipelining-HLS-Backend
     Branch,                     ///< branch(cond, T, F).
     Match,                      ///< match(val, otherwise, (case1, cont1), (case2, cont2), ...)
-    LoopBegin,                  ///<
-    LoopBreak,                  ///<
-    LoopContinue,               ///<
     PeInfo,                     ///< Partial evaluation debug info.
     EndScope                    ///< Dummy function which marks the end of a @p Scope.
 };
@@ -125,24 +122,9 @@ enum class Intrinsic : uint8_t {
  */
 class Continuation : public Def {
 public:
-    /// Stores information about structured control flow that should not be encoded in ops, as ops encode control flow
-    union SCFMetadata {
-        struct {
-            const Continuation* continue_target;
-            const Continuation* merge_target;
-        } loop_header;
-        struct {
-            const Continuation* loop_header;
-        } loop_epilogue;
-        struct {
-            const Continuation* merge_target;
-        } selection_header;
-    };
-
     struct Attributes {
         Intrinsic intrinsic = Intrinsic::None;
         CC cc = CC::C;
-        SCFMetadata scf_metadata = {};
 
         Attributes(Intrinsic intrinsic) : intrinsic(intrinsic) {}
         Attributes(CC cc = CC::C) : cc(cc) {}
@@ -207,9 +189,6 @@ class Continuation : public Def {
     void jump(const Def* callee, Defs args, Debug dbg = {});
     void branch(const Def* cond, const Def* t, const Def* f, Debug dbg = {});
     void match(const Def* val, Continuation* otherwise, Defs patterns, ArrayRef<Continuation*> continuations, Debug dbg = {});
-    //void structured_loop_merge(const Continuation* loop_header, ArrayRef<const Continuation*> targets);
-    //void structured_loop_continue(const Continuation* loop_header);
-    //void structured_loop_header(const Continuation* loop_epilogue, const Continuation* loop_continue, ArrayRef<const Continuation*> targets);
     void verify() const;
 
     const Filter* filter() const { return op(1)->as<Filter>(); }
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 8449a5758..96e30af2b 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -245,10 +245,6 @@ class World : public TypeTable, public Streamable<World> {
     Continuation* end_scope() const { return data_.end_scope_; }
     const Filter* filter(const Defs, Debug dbg = {});
 
-    Continuation* loop_enter(Types types, Continuations, Continuations);
-    Continuation* loop_continue(Types types);
-    Continuation* loop_break(Types types);
-
     /// Performs dead code, unreachable code and unused type elimination.
     void cleanup();
     void opt();

From a4f048d5f378cb2723a2f200dbf58e28a15efd01 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 17 Jan 2023 13:57:36 +0100
Subject: [PATCH 152/342] made shady dependency optional

---
 CMakeLists.txt               | 2 +-
 cmake/thorin-config.cmake.in | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c47d74578..67189ba7c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -44,7 +44,7 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
-find_package(shady REQUIRED CONFIG)
+find_package(shady CONFIG)
 if (shady_FOUND)
     message(STATUS "Found shady at ${shady_DIR}")
     set(THORIN_ENABLE_SHADY TRUE)
diff --git a/cmake/thorin-config.cmake.in b/cmake/thorin-config.cmake.in
index b55145178..833c1abbe 100644
--- a/cmake/thorin-config.cmake.in
+++ b/cmake/thorin-config.cmake.in
@@ -30,7 +30,7 @@ find_package(Half REQUIRED)
 
 set(Thorin_HAS_LLVM_SUPPORT @LLVM_FOUND@)
 set(Thorin_HAS_RV_SUPPORT @RV_FOUND@)
-set(Thorin_HAS_SPIRV_SUPPORT @SPIRV_ENABLED@)
+set(Thorin_HAS_SHADY_SUPPORT @shady_FOUND@)
 set(AnyDSL_LLVM_LINK_SHARED @AnyDSL_LLVM_LINK_SHARED@)
 
 if(Thorin_HAS_LLVM_SUPPORT)

From 698f559ef6442fc50e03942c2a7eb3639d495aad Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 17 Jan 2023 23:24:32 +0100
Subject: [PATCH 153/342] Small fixes.

---
 src/thorin/continuation.cpp |  2 +-
 src/thorin/rec_stream.cpp   | 12 ++++++++++--
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 1f98867ec..a7a56242d 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -61,7 +61,7 @@ Continuation::Continuation(const FnType* fn, const Attributes& attributes, Debug
 
 Continuation* Continuation::stub() const {
     Rewriter rewriter;
-    stub(rewriter);
+    return stub(rewriter);
 }
 
 Continuation* Continuation::stub(Rewriter& rewriter) const {
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index fa4547eb6..4a17429c8 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -56,7 +56,15 @@ void RecStreamer::run() {
             for (auto param : cont->params()) param_names.push_back(param->unique_name());
             s.fmt("{}: {} = ({, }) => {{\t\n", cont->unique_name(), cont->type(), param_names);
             run(cont->filter());
-            run(cont->body()); // TODO app node
+            if (defs.contains(cont->body())) {
+                auto body = cont->body();
+                if (auto cont2 = body->isa_nom<Continuation>()) {
+                    s.fmt("{}: {} = {}({, })", cont2->unique_name(), cont2->type(), cont2->body()->callee(), cont2->body()->args());
+                } else if (!body->no_dep() && !body->isa<Param>())
+                    body->stream_let(s);
+            } else {
+                run(cont->body()); // TODO app node
+            }
             s.fmt("\b\n}}");
         } else {
             s.fmt("{}: {} = {{ <no body> }}", cont->unique_name(), cont->type());
@@ -140,7 +148,7 @@ Stream& Def::stream1(Stream& s) const {
 }
 
 Stream& Def::stream_let(Stream& s) const {
-    return stream1(s.fmt("{}: {} = ", this, type())).endl();
+    return stream1(s.fmt("{}: {} = ", this->unique_name(), type())).endl();
 }
 
 Stream& World::stream(Stream& s) const {

From 87e7cdb8a66419a2641ccb38ca229a8420325937 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 18 Jan 2023 12:01:16 +0100
Subject: [PATCH 154/342] Guard continuation creation context, as it takes eons
 to compute.

---
 src/thorin/config.h.in    | 1 +
 src/thorin/debug.h        | 6 ++++++
 src/thorin/rec_stream.cpp | 2 ++
 src/thorin/world.cpp      | 5 +++++
 4 files changed, 14 insertions(+)

diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index 9216ecd82..78f9ea402 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -3,6 +3,7 @@
 
 #cmakedefine01 THORIN_ENABLE_CHECKS
 #cmakedefine01 THORIN_ENABLE_PROFILING
+#cmakedefine01 THORIN_ENABLE_CREATION_CONTEXT
 #cmakedefine01 THORIN_ENABLE_LLVM
 #cmakedefine01 THORIN_ENABLE_JSON
 #cmakedefine01 THORIN_ENABLE_RV
diff --git a/src/thorin/debug.h b/src/thorin/debug.h
index ef35f2a36..05613fccb 100644
--- a/src/thorin/debug.h
+++ b/src/thorin/debug.h
@@ -43,13 +43,16 @@ class Debug {
     Debug() = default; // TODO remove
     Debug(std::string name, Loc loc = {}, const Def* meta = nullptr)
         : name(name)
+#if THORIN_ENABLE_CREATION_CONTEXT
         , creation_context("")
+#endif
         , loc(loc)
         , meta(meta)
     {}
     Debug(const char* name, Loc loc = {}, const Def* meta = nullptr)
         : Debug(std::string(name), loc, meta)
     {}
+#if THORIN_ENABLE_CREATION_CONTEXT
     Debug(std::string name, std::string creation_context, Loc loc = {}, const Def* meta = nullptr)
         : name(name)
         , creation_context(creation_context)
@@ -59,13 +62,16 @@ class Debug {
     Debug(const char* name, const char* creation_context, Loc loc = {}, const Def* meta = nullptr)
         : Debug(std::string(name), std::string(creation_context), loc, meta)
     {}
+#endif
     Debug(Loc loc)
         : Debug("", loc)
     {}
     //Debug(const Def*);
 
     std::string name;
+#if THORIN_ENABLE_CREATION_CONTEXT
     std::string creation_context;
+#endif
     Loc loc;
     const Def* meta = nullptr;
 };
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 4a17429c8..9b800127e 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -108,9 +108,11 @@ Stream& Def::stream1(Stream& s) const {
     if (auto param = isa<Param>()) {
         return s.fmt("{}.{}", param->continuation(), param->unique_name());
     } else if (isa<Continuation>()) {
+#if THORIN_ENABLE_CREATION_CONTEXT
         if (debug().creation_context != "")
             return s.fmt("cont {} [{}]", unique_name(), debug().creation_context);
         else
+#endif
             return s.fmt("cont {}", unique_name());
     } else if (auto app = isa<App>()) {
         return s.fmt("{}({, })", app->callee(), app->args());
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index f465fba27..944dd1b83 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -1103,15 +1103,20 @@ const Def* World::run(const Def* def, Debug dbg) {
  */
 
 Continuation* World::continuation(const FnType* fn, Continuation::Attributes attributes, Debug dbg) {
+#if THORIN_ENABLE_CREATION_CONTEXT
     void *array[10];
     size_t size = backtrace(array, 10);
     assert(size >= 2);
     char ** symbols = backtrace_symbols(array, 10);
 
     dbg.creation_context = symbols[1];
+#endif
 
     auto cont = put<Continuation>(fn, attributes, dbg);
+
+#if THORIN_ENABLE_CREATION_CONTEXT
     free(symbols);
+#endif
 
     size_t i = 0;
     for (auto op : fn->ops()) {

From d4b30a103c50c52ebfc6ce3f10352f2688d7fa6b Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 18 Jan 2023 20:17:04 +0100
Subject: [PATCH 155/342] Fix: Tuple type with single element.

---
 src/thorin/type.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index 9b8f8b49f..7918b0e45 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -199,7 +199,7 @@ TypeTable::TypeTable()
 }
 
 const Type* TypeTable::tuple_type(Types ops) {
-    return (ops.size() == 1 && is_thin(ops[0])) ? ops.front() : insert<TupleType>(*this, ops);
+    return ops.size() == 1 ? ops.front() : insert<TupleType>(*this, ops);
 }
 
 const StructType* TypeTable::struct_type(Symbol name, size_t size) {

From 562992b5953d1dbe457ae2676e910c4d628a8aa0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Sun, 22 Jan 2023 13:23:59 +0100
Subject: [PATCH 156/342] added a World& field to every Def

---
 src/thorin/continuation.cpp |  12 ++--
 src/thorin/continuation.h   |   6 +-
 src/thorin/def.cpp          |   8 ++-
 src/thorin/def.h            |   5 +-
 src/thorin/primop.cpp       |  64 +++++++++----------
 src/thorin/primop.h         | 118 ++++++++++++++++++------------------
 src/thorin/world.cpp        |  52 ++++++++--------
 src/thorin/world.h          |  12 ++--
 8 files changed, 138 insertions(+), 139 deletions(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index f729937f8..c66302040 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -10,8 +10,8 @@ namespace thorin {
 
 //------------------------------------------------------------------------------
 
-Param::Param(const Type* type, Continuation* continuation, size_t index, Debug dbg)
-    : Def(Node_Param, type, 1, dbg)
+Param::Param(World& world, const Type* type, Continuation* continuation, size_t index, Debug dbg)
+    : Def(Node_Param, world, type, 1, dbg)
     , index_(index)
 {
     set_op(0, continuation);
@@ -19,7 +19,7 @@ Param::Param(const Type* type, Continuation* continuation, size_t index, Debug d
 
 //------------------------------------------------------------------------------
 
-App::App(const Defs ops, Debug dbg) : Def(Node_App, ops[0]->world().bottom_type(), ops, dbg) {
+App::App(World& world, const Defs ops, Debug dbg) : Def(Node_App, world, ops[0]->world().bottom_type(), ops, dbg) {
 #if THORIN_ENABLE_CHECKS
     verify();
     if (auto cont = callee()->isa_nom<Continuation>())
@@ -41,7 +41,7 @@ void App::verify() const {
 
 //------------------------------------------------------------------------------
 
-Filter::Filter(World& world, const Defs defs, Debug dbg) : Def(Node_Filter, world.bottom_type(), defs, dbg) {}
+Filter::Filter(World& world, const Defs defs, Debug dbg) : Def(Node_Filter, world, world.bottom_type(), defs, dbg) {}
 
 const Filter* Filter::cut(ArrayRef<size_t> indices) const {
     return world().filter(ops().cut(indices), debug());
@@ -49,8 +49,8 @@ const Filter* Filter::cut(ArrayRef<size_t> indices) const {
 
 //------------------------------------------------------------------------------
 
-Continuation::Continuation(const FnType* fn, const Attributes& attributes, Debug dbg)
-    : Def(Node_Continuation, fn, 2, dbg)
+Continuation::Continuation(World& w, const FnType* fn, const Attributes& attributes, Debug dbg)
+    : Def(Node_Continuation, w, fn, 2, dbg)
     , attributes_(attributes)
 {
     params_.reserve(fn->num_ops());
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 9f34fe992..f88e24885 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -24,7 +24,7 @@ typedef std::vector<Continuation*> Continuations;
  */
 class Param : public Def {
 private:
-    Param(const Type* type, Continuation* continuation, size_t index, Debug dbg);
+    Param(World&, const Type* type, Continuation* continuation, size_t index, Debug dbg);
 
 public:
     Continuation* continuation() const { return op(0)->as_nom<Continuation>(); }
@@ -53,7 +53,7 @@ class Filter : public Def {
 
 class App : public Def {
 private:
-    App(const Defs ops, Debug dbg);
+    App(World&, const Defs ops, Debug dbg);
 
 public:
     const Def* callee() const { return op(0); }
@@ -130,7 +130,7 @@ class Continuation : public Def {
     };
 
 private:
-    Continuation(const FnType* fn, const Attributes& attributes, Debug dbg);
+    Continuation(World&, const FnType* fn, const Attributes& attributes, Debug dbg);
     virtual ~Continuation() { for (auto param : params()) delete param; }
 
 public:
diff --git a/src/thorin/def.cpp b/src/thorin/def.cpp
index 2c01a687d..03516f1d9 100644
--- a/src/thorin/def.cpp
+++ b/src/thorin/def.cpp
@@ -14,9 +14,10 @@ namespace thorin {
 
 size_t Def::gid_counter_ = 1;
 
-Def::Def(NodeTag tag, const Type* type, Defs ops, Debug dbg)
+Def::Def(NodeTag tag, World& world, const Type* type, Defs ops, Debug dbg)
     : tag_(tag)
     , ops_(ops.size())
+    , world_(world)
     , type_(type)
     , debug_(dbg)
     , gid_(gid_counter_++)
@@ -29,9 +30,10 @@ Def::Def(NodeTag tag, const Type* type, Defs ops, Debug dbg)
         set_op(i, ops[i]);
 }
 
-Def::Def(NodeTag tag, const Type* type, size_t size, Debug dbg)
+Def::Def(NodeTag tag, World& world, const Type* type, size_t size, Debug dbg)
     : tag_(tag)
     , ops_(size)
+    , world_(world)
     , type_(type)
     , debug_(dbg)
     , gid_(gid_counter_++)
@@ -145,7 +147,7 @@ void Def::replace_uses(const Def* with) const {
     }
 }
 
-World& Def::world() const { return *static_cast<World*>(&type()->table()); }
+World& Def::world() const { return world_; }
 
 uint64_t UseHash::hash(Use use) {
     assert(use->gid() != uint32_t(-1));
diff --git a/src/thorin/def.h b/src/thorin/def.h
index a40a63738..18118e9d4 100644
--- a/src/thorin/def.h
+++ b/src/thorin/def.h
@@ -106,9 +106,9 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
 
 protected:
     /// Constructor for a @em structural Def.
-    Def(NodeTag tag, const Type* type, Defs args, Debug dbg);
+    Def(NodeTag tag, World&, const Type* type, Defs args, Debug dbg);
     /// Constructor for a @em nom Def.
-    Def(NodeTag tag, const Type* type, size_t size, Debug);
+    Def(NodeTag tag, World&, const Type* type, size_t size, Debug);
     virtual ~Def() {}
 
     void clear_type() { type_ = nullptr; }
@@ -230,6 +230,7 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
 private:
     const NodeTag tag_;
     std::vector<const Def*> ops_;
+    World& world_;
     const Type* type_;
     mutable Uses uses_;
     mutable Debug debug_;
diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index ddad23575..9a06a93a1 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -15,16 +15,16 @@ namespace thorin {
  */
 
 PrimLit::PrimLit(World& world, PrimTypeTag tag, Box box, Debug dbg)
-    : Literal((NodeTag) tag, world.prim_type(tag), dbg)
+    : Literal((NodeTag) tag, world, world.prim_type(tag), dbg)
     , box_(box)
 {}
 
-Cmp::Cmp(CmpTag tag, const Def* lhs, const Def* rhs, Debug dbg)
-    : BinOp((NodeTag) tag, lhs->world().type_bool(vector_length(lhs->type())), lhs, rhs, dbg)
+Cmp::Cmp(CmpTag tag, World& world, const Def* lhs, const Def* rhs, Debug dbg)
+    : BinOp((NodeTag) tag, world, world.type_bool(vector_length(lhs->type())), lhs, rhs, dbg)
 {}
 
 DefiniteArray::DefiniteArray(World& world, const Type* elem, Defs args, Debug dbg)
-    : Aggregate(Node_DefiniteArray, args, dbg)
+    : Aggregate(Node_DefiniteArray, world, args, dbg)
 {
     set_type(world.definite_array_type(elem, args.size()));
 #if THORIN_ENABLE_CHECKS
@@ -34,13 +34,13 @@ DefiniteArray::DefiniteArray(World& world, const Type* elem, Defs args, Debug db
 }
 
 IndefiniteArray::IndefiniteArray(World& world, const Type* elem, const Def* dim, Debug dbg)
-    : Aggregate(Node_IndefiniteArray, {dim}, dbg)
+    : Aggregate(Node_IndefiniteArray, world, {dim}, dbg)
 {
     set_type(world.indefinite_array_type(elem));
 }
 
 Tuple::Tuple(World& world, Defs args, Debug dbg)
-    : Aggregate(Node_Tuple, args, dbg)
+    : Aggregate(Node_Tuple, world, args, dbg)
 {
     Array<const Type*> elems(num_ops());
     for (size_t i = 0, e = num_ops(); i != e; ++i)
@@ -50,7 +50,7 @@ Tuple::Tuple(World& world, Defs args, Debug dbg)
 }
 
 Vector::Vector(World& world, Defs args, Debug dbg)
-    : Aggregate(Node_Vector, args, dbg)
+    : Aggregate(Node_Vector, world, args, dbg)
 {
     if (auto primtype = args.front()->type()->isa<PrimType>()) {
         assert(primtype->length() == 1);
@@ -62,10 +62,9 @@ Vector::Vector(World& world, Defs args, Debug dbg)
     }
 }
 
-LEA::LEA(const Def* ptr, const Def* index, Debug dbg)
-    : Def(Node_LEA, nullptr, {ptr, index}, dbg)
+LEA::LEA(World& world, const Def* ptr, const Def* index, Debug dbg)
+    : Def(Node_LEA, world, nullptr, {ptr, index}, dbg)
 {
-    auto& world = index->world();
     auto type = ptr_type();
     if (auto tuple = ptr_pointee()->isa<TupleType>()) {
         set_type(world.ptr_type(get(tuple->ops(), index), type->length(), type->device(), type->addr_space()));
@@ -81,54 +80,51 @@ LEA::LEA(const Def* ptr, const Def* index, Debug dbg)
     }
 }
 
-Known::Known(const Def* def, Debug dbg)
-    : Def(Node_Known, def->world().type_bool(), {def}, dbg)
+Known::Known(World& world, const Def* def, Debug dbg)
+    : Def(Node_Known, world, world.type_bool(), {def}, dbg)
 {}
 
-AlignOf::AlignOf(const Def* def, Debug dbg)
-    : Def(Node_AlignOf, def->world().type_qs64(), {def}, dbg)
+AlignOf::AlignOf(World& world, const Def* def, Debug dbg)
+    : Def(Node_AlignOf, world, world.type_qs64(), {def}, dbg)
 {}
 
-SizeOf::SizeOf(const Def* def, Debug dbg)
-    : Def(Node_SizeOf, def->world().type_qs64(), {def}, dbg)
+SizeOf::SizeOf(World& world, const Def* def, Debug dbg)
+    : Def(Node_SizeOf, world, world.type_qs64(), {def}, dbg)
 {}
 
-Slot::Slot(const Type* type, const Def* frame, Debug dbg)
-    : Def(Node_Slot, type->table().ptr_type(type), {frame}, dbg)
+Slot::Slot(World& world, const Type* type, const Def* frame, Debug dbg)
+    : Def(Node_Slot, world, type->table().ptr_type(type), {frame}, dbg)
 {
     assert(frame->type()->isa<FrameType>());
 }
 
-Global::Global(const Def* init, bool is_mutable, Debug dbg)
-    : Def(Node_Global, init->type()->table().ptr_type(init->type()), {init}, dbg)
+Global::Global(World& world, const Def* init, bool is_mutable, Debug dbg)
+    : Def(Node_Global, world, init->type()->table().ptr_type(init->type()), {init}, dbg)
     , is_mutable_(is_mutable)
 {
     assert(!init->has_dep(Dep::Param));
 }
 
-Alloc::Alloc(const Type* type, const Def* mem, const Def* extra, Debug dbg)
-    : MemOp(Node_Alloc, nullptr, {mem, extra}, dbg)
+Alloc::Alloc(World& world, const Type* type, const Def* mem, const Def* extra, Debug dbg)
+    : MemOp(Node_Alloc, world, nullptr, {mem, extra}, dbg)
 {
-    World& w = mem->world();
-    set_type(w.tuple_type({w.mem_type(), w.ptr_type(type)}));
+    set_type(world.tuple_type({world.mem_type(), world.ptr_type(type)}));
 }
 
-Load::Load(const Def* mem, const Def* ptr, Debug dbg)
-    : Access(Node_Load, nullptr, {mem, ptr}, dbg)
+Load::Load(World& world, const Def* mem, const Def* ptr, Debug dbg)
+    : Access(Node_Load, world, nullptr, {mem, ptr}, dbg)
 {
-    World& w = mem->world();
-    set_type(w.tuple_type({w.mem_type(), ptr->type()->as<PtrType>()->pointee()}));
+    set_type(world.tuple_type({world.mem_type(), ptr->type()->as<PtrType>()->pointee()}));
 }
 
-Enter::Enter(const Def* mem, Debug dbg)
-    : MemOp(Node_Enter, nullptr, {mem}, dbg)
+Enter::Enter(World& world, const Def* mem, Debug dbg)
+    : MemOp(Node_Enter, world, nullptr, {mem}, dbg)
 {
-    World& w = mem->world();
-    set_type(w.tuple_type({w.mem_type(), w.frame_type()}));
+    set_type(world.tuple_type({world.mem_type(), world.frame_type()}));
 }
 
-Assembly::Assembly(const Type *type, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints, ArrayRef<std::string> input_constraints, ArrayRef<std::string> clobbers, Flags flags, Debug dbg)
-    : MemOp(Node_Assembly, type, inputs, dbg)
+Assembly::Assembly(World& world, const Type *type, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints, ArrayRef<std::string> input_constraints, ArrayRef<std::string> clobbers, Flags flags, Debug dbg)
+    : MemOp(Node_Assembly, world, type, inputs, dbg)
     , asm_template_(asm_template)
     , output_constraints_(output_constraints)
     , input_constraints_(input_constraints)
diff --git a/src/thorin/primop.h b/src/thorin/primop.h
index ad4e2efcb..db2aaf175 100644
--- a/src/thorin/primop.h
+++ b/src/thorin/primop.h
@@ -10,16 +10,16 @@ namespace thorin {
 
 class Literal : public Def {
 protected:
-    Literal(NodeTag tag, const Type* type, Debug dbg)
-        : Def(tag, type, Defs{}, dbg)
+    Literal(NodeTag tag, World& world, const Type* type, Debug dbg)
+        : Def(tag, world, type, Defs{}, dbg)
     {}
 };
 
 /// This literal represents 'no value'.
 class Bottom : public Literal {
 private:
-    Bottom(const Type* type, Debug dbg)
-        : Literal(Node_Bottom, type, dbg)
+    Bottom(World& world, const Type* type, Debug dbg)
+        : Literal(Node_Bottom, world, type, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -30,8 +30,8 @@ class Bottom : public Literal {
 /// This literal represents 'any value'.
 class Top : public Literal {
 private:
-    Top(const Type* type, Debug dbg)
-        : Literal(Node_Top, type, dbg)
+    Top(World& world, const Type* type, Debug dbg)
+        : Literal(Node_Top, world, type, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -79,8 +79,8 @@ T get(ArrayRef<T> array, const Def* def) { return array[primlit_value<size_t>(de
 /// Akin to <tt>cond ? tval : fval</tt>.
 class Select : public Def {
 private:
-    Select(const Def* cond, const Def* tval, const Def* fval, Debug dbg)
-        : Def(Node_Select, tval->type(), {cond, tval, fval}, dbg)
+    Select(World& world, const Def* cond, const Def* tval, const Def* fval, Debug dbg)
+        : Def(Node_Select, world, tval->type(), {cond, tval, fval}, dbg)
     {
         assert(is_type_bool(cond->type()));
         assert(tval->type() == fval->type() && "types of both values must be equal");
@@ -100,7 +100,7 @@ class Select : public Def {
 /// Get the alignment in number of bytes needed for any value (including bottom) of a given @p Type.
 class AlignOf : public Def {
 private:
-    AlignOf(const Def* def, Debug dbg);
+    AlignOf(World& world, const Def* def, Debug dbg);
 
     const Def* rebuild(World&, const Type*, Defs) const override;
 
@@ -113,7 +113,7 @@ class AlignOf : public Def {
 /// Get number of bytes needed for any value (including bottom) of a given @p Type.
 class SizeOf : public Def {
 private:
-    SizeOf(const Def* def, Debug dbg);
+    SizeOf(World& world, const Def* def, Debug dbg);
 
     const Def* rebuild(World&, const Type*, Defs) const override;
 
@@ -126,8 +126,8 @@ class SizeOf : public Def {
 /// Base class for all side-effect free binary \p Def%s.
 class BinOp : public Def {
 protected:
-    BinOp(NodeTag tag, const Type* type, const Def* lhs, const Def* rhs, Debug dbg)
-        : Def(tag, type, {lhs, rhs}, dbg)
+    BinOp(NodeTag tag, World& world, const Type* type, const Def* lhs, const Def* rhs, Debug dbg)
+        : Def(tag, world, type, {lhs, rhs}, dbg)
     {
         assert(lhs->type() == rhs->type() && "types are not equal");
     }
@@ -140,8 +140,8 @@ class BinOp : public Def {
 /// One of \p ArithOpTag arithmetic operation.
 class ArithOp : public BinOp {
 private:
-    ArithOp(ArithOpTag tag, const Def* lhs, const Def* rhs, Debug dbg)
-        : BinOp((NodeTag) tag, lhs->type(), lhs, rhs, dbg)
+    ArithOp(ArithOpTag tag, World& world, const Def* lhs, const Def* rhs, Debug dbg)
+        : BinOp((NodeTag) tag, world, lhs->type(), lhs, rhs, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -157,7 +157,7 @@ class ArithOp : public BinOp {
 /// One of \p CmpTag compare.
 class Cmp : public BinOp {
 private:
-    Cmp(CmpTag tag, const Def* lhs, const Def* rhs, Debug dbg);
+    Cmp(CmpTag tag, World& world, const Def* lhs, const Def* rhs, Debug dbg);
 
     const Def* rebuild(World&, const Type*, Defs) const override;
 
@@ -172,8 +172,8 @@ class Cmp : public BinOp {
 /// Common mathematical function such as `sin()` or `cos()`.
 class MathOp : public Def {
 private:
-    MathOp(MathOpTag tag, const Type* type, Defs args, Debug dbg)
-        : Def((NodeTag)tag, type, args, dbg)
+    MathOp(MathOpTag tag, World& world, const Type* type, Defs args, Debug dbg)
+        : Def((NodeTag)tag, world, type, args, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -189,8 +189,8 @@ class MathOp : public Def {
 /// Base class for @p Bitcast and @p Cast.
 class ConvOp : public Def {
 protected:
-    ConvOp(NodeTag tag, const Def* from, const Type* to, Debug dbg)
-        : Def(tag, to, {from}, dbg)
+    ConvOp(NodeTag tag, World& world, const Def* from, const Type* to, Debug dbg)
+        : Def(tag, world, to, {from}, dbg)
     {}
 
 public:
@@ -200,8 +200,8 @@ class ConvOp : public Def {
 /// Converts <tt>from</tt> to type <tt>to</tt>.
 class Cast : public ConvOp {
 private:
-    Cast(const Type* to, const Def* from, Debug dbg)
-        : ConvOp(Node_Cast, from, to, dbg)
+    Cast(World& world, const Type* to, const Def* from, Debug dbg)
+        : ConvOp(Node_Cast, world, from, to, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -212,8 +212,8 @@ class Cast : public ConvOp {
 /// Reinterprets the bits of <tt>from</tt> as type <tt>to</tt>.
 class Bitcast : public ConvOp {
 private:
-    Bitcast(const Type* to, const Def* from, Debug dbg)
-        : ConvOp(Node_Bitcast, from, to, dbg)
+    Bitcast(World& world, const Type* to, const Def* from, Debug dbg)
+        : ConvOp(Node_Bitcast, world, from, to, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -224,8 +224,8 @@ class Bitcast : public ConvOp {
 /// Base class for all aggregate data constructers.
 class Aggregate : public Def {
 protected:
-    Aggregate(NodeTag tag, Defs args, Debug dbg)
-        : Def(tag, nullptr /*set later*/, args, dbg)
+    Aggregate(NodeTag tag, World& world, Defs args, Debug dbg)
+        : Def(tag, world, nullptr /*set later*/, args, dbg)
     {}
 };
 
@@ -274,8 +274,8 @@ class Tuple : public Aggregate {
 /// Data constructor for a @p VariantType.
 class Variant : public Def {
 private:
-    Variant(const VariantType* variant_type, const Def* value, size_t index, Debug dbg)
-        : Def(Node_Variant, variant_type, {value}, dbg), index_(index)
+    Variant(World& world, const VariantType* variant_type, const Def* value, size_t index, Debug dbg)
+        : Def(Node_Variant, world, variant_type, {value}, dbg), index_(index)
     {
         assert(variant_type->op(index) == value->type());
     }
@@ -297,8 +297,8 @@ class Variant : public Def {
 /// Yields the tag/index for this variant in the supplied integer type
 class VariantIndex : public Def {
 private:
-    VariantIndex(const Type* int_type, const Def* value, Debug dbg)
-        : Def(Node_VariantIndex, int_type, {value}, dbg)
+    VariantIndex(World& world, const Type* int_type, const Def* value, Debug dbg)
+        : Def(Node_VariantIndex, world, int_type, {value}, dbg)
     {
         assert(value->type()->isa<VariantType>());
         assert(is_type_s(int_type) || is_type_u(int_type));
@@ -311,8 +311,8 @@ class VariantIndex : public Def {
 
 class VariantExtract : public Def {
 private:
-    VariantExtract(const Type* type, const Def* value, size_t index, Debug dbg)
-        : Def(Node_VariantExtract, type, {value}, dbg), index_(index)
+    VariantExtract(World& world, const Type* type, const Def* value, size_t index, Debug dbg)
+        : Def(Node_VariantExtract, world, type, {value}, dbg), index_(index)
     {
         assert(value->type()->as<VariantType>()->op(index) == type);
     }
@@ -333,8 +333,8 @@ class VariantExtract : public Def {
 /// Data constructor for a @p ClosureType.
 class Closure : public Aggregate {
 private:
-    Closure(const ClosureType* closure_type, const Def* fn, const Def* env, Debug dbg)
-        : Aggregate(Node_Closure, {fn, env}, dbg)
+    Closure(World& world, const ClosureType* closure_type, const Def* fn, const Def* env, Debug dbg)
+        : Aggregate(Node_Closure, world, {fn, env}, dbg)
     {
         set_type(closure_type);
     }
@@ -351,8 +351,8 @@ class Closure : public Aggregate {
 /// Data constructor for a @p StructType.
 class StructAgg : public Aggregate {
 private:
-    StructAgg(const StructType* struct_type, Defs args, Debug dbg)
-        : Aggregate(Node_StructAgg, args, dbg)
+    StructAgg(World& world, const StructType* struct_type, Defs args, Debug dbg)
+        : Aggregate(Node_StructAgg, world, args, dbg)
     {
 #if THORIN_ENABLE_CHECKS
         assert(struct_type->num_ops() == args.size());
@@ -383,8 +383,8 @@ class Vector : public Aggregate {
 /// Base class for functional @p Insert and @p Extract.
 class AggOp : public Def {
 protected:
-    AggOp(NodeTag tag, const Type* type, Defs args, Debug dbg)
-        : Def(tag, type, args, dbg)
+    AggOp(NodeTag tag, World& world, const Type* type, Defs args, Debug dbg)
+        : Def(tag, world, type, args, dbg)
     {}
 
 public:
@@ -397,8 +397,8 @@ class AggOp : public Def {
 /// Extracts from aggregate <tt>agg</tt> the element at position <tt>index</tt>.
 class Extract : public AggOp {
 private:
-    Extract(const Def* agg, const Def* index, Debug dbg)
-        : AggOp(Node_Extract, extracted_type(agg, index), {agg, index}, dbg)
+    Extract(World& world, const Def* agg, const Def* index, Debug dbg)
+        : AggOp(Node_Extract, world, extracted_type(agg, index), {agg, index}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -417,8 +417,8 @@ class Extract : public AggOp {
  */
 class Insert : public AggOp {
 private:
-    Insert(const Def* agg, const Def* index, const Def* value, Debug dbg)
-        : AggOp(Node_Insert, agg->type(), {agg, index, value}, dbg)
+    Insert(World& world, const Def* agg, const Def* index, const Def* value, Debug dbg)
+        : AggOp(Node_Insert, world, agg->type(), {agg, index, value}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -437,7 +437,7 @@ class Insert : public AggOp {
  */
 class LEA : public Def {
 private:
-    LEA(const Def* ptr, const Def* index, Debug dbg);
+    LEA(World& world, const Def* ptr, const Def* index, Debug dbg);
 
     const Def* rebuild(World&, const Type*, Defs) const override;
 
@@ -454,8 +454,8 @@ class LEA : public Def {
 /// Casts the underlying @p def to a dynamic value during @p partial_evaluation.
 class Hlt : public Def {
 private:
-    Hlt(const Def* def, Debug dbg)
-        : Def(Node_Hlt, def->type(), {def}, dbg)
+    Hlt(World& world, const Def* def, Debug dbg)
+        : Def(Node_Hlt, world, def->type(), {def}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -469,7 +469,7 @@ class Hlt : public Def {
 /// Evaluates to @c true, if @p def is a literal.
 class Known : public Def {
 private:
-    Known(const Def* def, Debug dbg);
+    Known(World& world, const Def* def, Debug dbg);
 
     const Def* rebuild(World&, const Type*, Defs) const override;
 
@@ -485,8 +485,8 @@ class Known : public Def {
  */
 class Run : public Def {
 private:
-    Run(const Def* def, Debug dbg)
-        : Def(Node_Run, def->type(), {def}, dbg)
+    Run(World& world, const Def* def, Debug dbg)
+        : Def(Node_Run, world, def->type(), {def}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -504,7 +504,7 @@ class Run : public Def {
  */
 class Slot : public Def {
 private:
-    Slot(const Type* type, const Def* frame, Debug dbg);
+    Slot(World& world, const Type* type, const Def* frame, Debug dbg);
 
 public:
     const Def* frame() const { return op(0); }
@@ -525,7 +525,7 @@ class Slot : public Def {
  */
 class Global : public Def {
 private:
-    Global(const Def* init, bool is_mutable, Debug dbg);
+    Global(World& world, const Def* init, bool is_mutable, Debug dbg);
 
 public:
     const Def* init() const { return op(0); }
@@ -547,8 +547,8 @@ class Global : public Def {
 /// Base class for all \p Def%s taking and producing side-effects.
 class MemOp : public Def {
 protected:
-    MemOp(NodeTag tag, const Type* type, Defs args, Debug dbg)
-        : Def(tag, type, args, dbg)
+    MemOp(NodeTag tag, World& world, const Type* type, Defs args, Debug dbg)
+        : Def(tag, world, type, args, dbg)
     {
         assert(mem()->type()->isa<MemType>());
         assert(args.size() >= 1);
@@ -566,7 +566,7 @@ class MemOp : public Def {
 /// Allocates memory on the heap.
 class Alloc : public MemOp {
 private:
-    Alloc(const Type* type, const Def* mem, const Def* extra, Debug dbg);
+    Alloc(World& world, const Type* type, const Def* mem, const Def* extra, Debug dbg);
 
 public:
     const Def* extra() const { return op(1); }
@@ -585,8 +585,8 @@ class Alloc : public MemOp {
 /// Base class for @p Load and @p Store.
 class Access : public MemOp {
 protected:
-    Access(NodeTag tag, const Type* type, Defs args, Debug dbg)
-        : MemOp(tag, type, args, dbg)
+    Access(NodeTag tag, World& world, const Type* type, Defs args, Debug dbg)
+        : MemOp(tag, world, type, args, dbg)
     {
         assert(args.size() >= 2);
     }
@@ -598,7 +598,7 @@ class Access : public MemOp {
 /// Loads with current effect <tt>mem</tt> from <tt>ptr</tt> to produce a pair of a new effect and the loaded value.
 class Load : public Access {
 private:
-    Load(const Def* mem, const Def* ptr, Debug dbg);
+    Load(World& world, const Def* mem, const Def* ptr, Debug dbg);
 
 public:
     bool has_multiple_outs() const override { return true; }
@@ -615,8 +615,8 @@ class Load : public Access {
 /// Stores with current effect <tt>mem</tt> <tt>value</tt> into <tt>ptr</tt> while producing a new effect.
 class Store : public Access {
 private:
-    Store(const Def* mem, const Def* ptr, const Def* value, Debug dbg)
-        : Access(Node_Store, mem->type(), {mem, ptr, value}, dbg)
+    Store(World& world, const Def* mem, const Def* ptr, const Def* value, Debug dbg)
+        : Access(Node_Store, world, mem->type(), {mem, ptr, value}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -631,7 +631,7 @@ class Store : public Access {
 /// Creates a stack \p Frame with current effect <tt>mem</tt>.
 class Enter : public MemOp {
 private:
-    Enter(const Def* mem, Debug dbg);
+    Enter(World& world, const Def* mem, Debug dbg);
 
     const Def* rebuild(World&, const Type*, Defs) const override;
 
@@ -655,7 +655,7 @@ class Assembly : public MemOp {
     };
 
 private:
-    Assembly(const Type *type, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints,
+    Assembly(World& world, const Type *type, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints,
              ArrayRef<std::string> input_constraints, ArrayRef<std::string> clobbers, Flags flags, Debug dbg);
 
 public:
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 47f34b6aa..013d09544 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -55,14 +55,14 @@ World::~World() {
 const Def* World::variant_index(const Def* value, Debug dbg) {
     if (auto variant = value->isa<Variant>())
         return literal_qu64(variant->index(), dbg);
-    return cse(new VariantIndex(type_qu64(), value, dbg));
+    return cse(new VariantIndex(*this, type_qu64(), value, dbg));
 }
 
 const Def* World::variant_extract(const Def* value, size_t index, Debug dbg) {
     auto type = value->type()->as<VariantType>()->op(index);
     if (auto variant = value->isa<Variant>())
         return variant->index() == index ? variant->value() : bottom(type);
-    return cse(new VariantExtract(type, value, index, dbg));
+    return cse(new VariantExtract(*this, type, value, index, dbg));
 }
 
 /*
@@ -391,7 +391,7 @@ const Def* World::arithop(ArithOpTag tag, const Def* a, const Def* b, Debug dbg)
             return arithop(tag, a_lhs_lv, arithop(tag, a_same->rhs(), b, dbg), dbg);
     }
 
-    return cse(new ArithOp(tag, a, b, dbg));
+    return cse(new ArithOp(tag, *this, a, b, dbg));
 }
 
 const Def* World::arithop_not(const Def* def, Debug dbg) { return arithop_xor(allset(def->type(), dbg, vector_length(def)), def, dbg); }
@@ -483,7 +483,7 @@ const Def* World::cmp(CmpTag tag, const Def* a, const Def* b, Debug dbg) {
         }
     }
 
-    return cse(new Cmp(tag, a, b, dbg));
+    return cse(new Cmp(tag, *this, a, b, dbg));
 }
 
 /*
@@ -616,7 +616,7 @@ const Def* World::cast(const Type* to, const Def* from, Debug dbg) {
         }
     }
 
-    return cse(new Cast(to, from, dbg));
+    return cse(new Cast(*this, to, from, dbg));
 }
 
 const Def* World::bitcast(const Type* to, const Def* from, Debug dbg) {
@@ -655,7 +655,7 @@ const Def* World::bitcast(const Type* to, const Def* from, Debug dbg) {
         return vector(ops, dbg);
     }
 
-    return cse(new Bitcast(to, from, dbg));
+    return cse(new Bitcast(*this, to, from, dbg));
 }
 
 /*
@@ -712,7 +712,7 @@ const Def* World::extract(const Def* agg, const Def* index, Debug dbg) {
         }
     }
 
-    return cse(new Extract(agg, index, dbg));
+    return cse(new Extract(*this, agg, index, dbg));
 }
 
 const Def* World::insert(const Def* agg, const Def* index, const Def* value, Debug dbg) {
@@ -760,14 +760,14 @@ const Def* World::insert(const Def* agg, const Def* index, const Def* value, Deb
         }
     }
 
-    return cse(new Insert(agg, index, value, dbg));
+    return cse(new Insert(*this, agg, index, value, dbg));
 }
 
 const Def* World::lea(const Def* ptr, const Def* index, Debug dbg) {
     if (fold_1_tuple(ptr->type()->as<PtrType>()->pointee(), index))
         return ptr;
 
-    return cse(new LEA(ptr, index, dbg));
+    return cse(new LEA(*this, ptr, index, dbg));
 }
 
 const Def* World::select(const Def* cond, const Def* a, const Def* b, Debug dbg) {
@@ -785,21 +785,21 @@ const Def* World::select(const Def* cond, const Def* a, const Def* b, Debug dbg)
     if (a == b)
         return a;
 
-    return cse(new Select(cond, a, b, dbg));
+    return cse(new Select(*this, cond, a, b, dbg));
 }
 
 const Def* World::align_of(const Type* type, Debug dbg) {
     if (auto ptype = type->isa<PrimType>())
         return literal(qs64(num_bits(ptype->primtype_tag()) / 8), dbg);
 
-    return cse(new AlignOf(bottom(type, dbg), dbg));
+    return cse(new AlignOf(*this, bottom(type, dbg), dbg));
 }
 
 const Def* World::size_of(const Type* type, Debug dbg) {
     if (auto ptype = type->isa<PrimType>())
         return literal(qs64(num_bits(ptype->primtype_tag()) / 8), dbg);
 
-    return cse(new SizeOf(bottom(type, dbg), dbg));
+    return cse(new SizeOf(*this, bottom(type, dbg), dbg));
 }
 
 /*
@@ -824,7 +824,7 @@ const Def* World::transcendental(MathOpTag tag, const Def* arg, Debug dbg, F&& f
                 THORIN_UNREACHABLE;
         }
     }
-    return cse(new MathOp(tag, arg->type(), { arg }, dbg));
+    return cse(new MathOp(tag, *this, arg->type(), { arg }, dbg));
 }
 
 template <class F>
@@ -846,7 +846,7 @@ const Def* World::transcendental(MathOpTag tag, const Def* left, const Def* righ
                 THORIN_UNREACHABLE;
         }
     }
-    return cse(new MathOp(tag, left->type(), { left, right }, dbg));
+    return cse(new MathOp(tag, *this, left->type(), { left, right }, dbg));
 }
 
 template <class F>
@@ -1019,7 +1019,7 @@ const Def* World::load(const Def* mem, const Def* ptr, Debug dbg) {
             return tuple({mem, tuple({}, dbg)});
         }
     }
-    return cse(new Load(mem, ptr, dbg));
+    return cse(new Load(*this, mem, ptr, dbg));
 }
 
 bool is_agg_const(const Def* def) {
@@ -1029,7 +1029,7 @@ bool is_agg_const(const Def* def) {
 const Def* World::store(const Def* mem, const Def* ptr, const Def* value, Debug dbg) {
     if (value->isa<Bottom>())
         return mem;
-    return cse(new Store(mem, ptr, value, dbg));
+    return cse(new Store(*this, mem, ptr, value, dbg));
 }
 
 const Def* World::enter(const Def* mem, Debug dbg) {
@@ -1037,15 +1037,15 @@ const Def* World::enter(const Def* mem, Debug dbg) {
     // in order to simplify as we go and prevent code size from exploding
     if (auto e = Enter::is_out_mem(mem))
         return e;
-    return cse(new Enter(mem, dbg));
+    return cse(new Enter(*this, mem, dbg));
 }
 
 const Def* World::alloc(const Type* type, const Def* mem, const Def* extra, Debug dbg) {
-    return cse(new Alloc(type, mem, extra, dbg));
+    return cse(new Alloc(*this, type, mem, extra, dbg));
 }
 
 const Def* World::global(const Def* init, bool is_mutable, Debug dbg) {
-    return cse(new Global(init, is_mutable, dbg));
+    return cse(new Global(*this, init, is_mutable, dbg));
 }
 
 const Def* World::global_immutable_string(const std::string& str, Debug dbg) {
@@ -1060,7 +1060,7 @@ const Def* World::global_immutable_string(const std::string& str, Debug dbg) {
 }
 
 const Assembly* World::assembly(const Type* type, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints, ArrayRef<std::string> input_constraints, ArrayRef<std::string> clobbers, Assembly::Flags flags, Debug dbg) {
-    return cse(new Assembly(type, inputs, asm_template, output_constraints, input_constraints, clobbers, flags, dbg))->as<Assembly>();;
+    return cse(new Assembly(*this, type, inputs, asm_template, output_constraints, input_constraints, clobbers, flags, dbg))->as<Assembly>();;
 }
 
 const Assembly* World::assembly(Types types, const Def* mem, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints, ArrayRef<std::string> input_constraints, ArrayRef<std::string> clobbers, Assembly::Flags flags, Debug dbg) {
@@ -1081,7 +1081,7 @@ const Assembly* World::assembly(Types types, const Def* mem, Defs inputs, std::s
 
 const Def* World::hlt(const Def* def, Debug dbg) {
     if (is_pe_done()) return def;
-    return cse(new Hlt(def, dbg));
+    return cse(new Hlt(*this, def, dbg));
 }
 
 const Def* World::known(const Def* def, Debug dbg) {
@@ -1089,12 +1089,12 @@ const Def* World::known(const Def* def, Debug dbg) {
         return literal_bool(false, dbg);
     if (!def->has_dep(Dep::Param))
         return literal_bool(true, dbg);
-    return cse(new Known(def, dbg));
+    return cse(new Known(*this, def, dbg));
 }
 
 const Def* World::run(const Def* def, Debug dbg) {
     if (is_pe_done()) return def;
-    return cse(new Run(def, dbg));
+    return cse(new Run(*this, def, dbg));
 }
 
 /*
@@ -1102,7 +1102,7 @@ const Def* World::run(const Def* def, Debug dbg) {
  */
 
 Continuation* World::continuation(const FnType* fn, Continuation::Attributes attributes, Debug dbg) {
-    auto cont = put<Continuation>(fn, attributes, dbg);
+    auto cont = put<Continuation>(*this, fn, attributes, dbg);
 
     size_t i = 0;
     for (auto op : fn->ops()) {
@@ -1123,7 +1123,7 @@ Continuation* World::match(const Type* type, size_t num_patterns) {
 }
 
 const Param* World::param(const Type* type, Continuation* continuation, size_t index, Debug dbg) {
-    auto param = new Param(type, continuation, index, dbg);
+    auto param = new Param(*this, type, continuation, index, dbg);
 #if THORIN_ENABLE_CHECKS
     if (state_.breakpoints.contains(param->gid())) THORIN_BREAK;
 #endif
@@ -1171,7 +1171,7 @@ const App* World::app(const Def* callee, const Defs args, Debug dbg) {
     for (size_t i = 0; i < args.size(); i++)
         ops[i + 1] = args[i];
 
-    return cse(new App(ops, dbg));
+    return cse(new App(*this, ops, dbg));
 }
 
 /*
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 96e30af2b..940680ef7 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -110,8 +110,8 @@ class World : public TypeTable, public Streamable<World> {
     const Def* one(const Type* type, Debug dbg = {}, size_t length = 1) { return one(type->as<PrimType>()->primtype_tag(), dbg, length); }
     const Def* allset(PrimTypeTag tag, Debug dbg = {}, size_t length = 1);
     const Def* allset(const Type* type, Debug dbg = {}, size_t length = 1) { return allset(type->as<PrimType>()->primtype_tag(), dbg, length); }
-    const Def* top(const Type* type, Debug dbg = {}, size_t length = 1) { return splat(cse(new Top(type, dbg)), length); }
-    const Def* bottom(const Type* type, Debug dbg = {}, size_t length = 1) { return splat(cse(new Bottom(type, dbg)), length); }
+    const Def* top(const Type* type, Debug dbg = {}, size_t length = 1) { return splat(cse(new Top(*this, type, dbg)), length); }
+    const Def* bottom(const Type* type, Debug dbg = {}, size_t length = 1) { return splat(cse(new Bottom(*this, type, dbg)), length); }
     const Def* bottom(PrimTypeTag tag, Debug dbg = {}, size_t length = 1) { return bottom(prim_type(tag), dbg, length); }
 
     // arithops
@@ -156,15 +156,15 @@ class World : public TypeTable, public Streamable<World> {
         return cse(new IndefiniteArray(*this, elem, dim, dbg));
     }
     const Def* struct_agg(const StructType* struct_type, Defs args, Debug dbg = {}) {
-        return try_fold_aggregate(cse(new StructAgg(struct_type, args, dbg)));
+        return try_fold_aggregate(cse(new StructAgg(*this, struct_type, args, dbg)));
     }
     const Def* tuple(Defs args, Debug dbg = {}) { return args.size() == 1 ? args.front() : try_fold_aggregate(cse(new Tuple(*this, args, dbg))); }
 
-    const Def* variant(const VariantType* variant_type, const Def* value, size_t index, Debug dbg = {}) { return cse(new Variant(variant_type, value, index, dbg)); }
+    const Def* variant(const VariantType* variant_type, const Def* value, size_t index, Debug dbg = {}) { return cse(new Variant(*this, variant_type, value, index, dbg)); }
     const Def* variant_index  (const Def* value, Debug dbg = {});
     const Def* variant_extract(const Def* value, size_t index, Debug dbg = {});
 
-    const Def* closure(const ClosureType* closure_type, const Def* fn, const Def* env, Debug dbg = {}) { return cse(new Closure(closure_type, fn, env, dbg)); }
+    const Def* closure(const ClosureType* closure_type, const Def* fn, const Def* env, Debug dbg = {}) { return cse(new Closure(*this, closure_type, fn, env, dbg)); }
     const Def* vector(Defs args, Debug dbg = {}) {
         if (args.size() == 1) return args[0];
         return try_fold_aggregate(cse(new Vector(*this, args, dbg)));
@@ -216,7 +216,7 @@ class World : public TypeTable, public Streamable<World> {
     const Def* load(const Def* mem, const Def* ptr, Debug dbg = {});
     const Def* store(const Def* mem, const Def* ptr, const Def* val, Debug dbg = {});
     const Def* enter(const Def* mem, Debug dbg = {});
-    const Def* slot(const Type* type, const Def* frame, Debug dbg = {}) { return cse(new Slot(type, frame, dbg)); }
+    const Def* slot(const Type* type, const Def* frame, Debug dbg = {}) { return cse(new Slot(*this, type, frame, dbg)); }
     const Def* alloc(const Type* type, const Def* mem, const Def* extra, Debug dbg = {});
     const Def* alloc(const Type* type, const Def* mem, Debug dbg = {}) { return alloc(type, mem, literal_qu64(0, dbg), dbg); }
     const Def* global(const Def* init, bool is_mutable = true, Debug dbg = {});

From 654d93577e7820ec0aa822d41aa5de24acd48e79 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Sun, 22 Jan 2023 15:28:55 +0100
Subject: [PATCH 157/342] rebuild() uses new worlds instead of recycling the
 same one

---
 src/thorin/be/c/c.cpp                   | 26 ++++++++---------
 src/thorin/be/c/c.h                     |  6 ++--
 src/thorin/be/codegen.cpp               | 37 +++++++++++++------------
 src/thorin/be/codegen.h                 | 10 ++++---
 src/thorin/be/llvm/amdgpu.cpp           |  4 +--
 src/thorin/be/llvm/amdgpu.h             |  2 +-
 src/thorin/be/llvm/cpu.cpp              |  4 +--
 src/thorin/be/llvm/cpu.h                |  2 +-
 src/thorin/be/llvm/llvm.cpp             |  6 ++--
 src/thorin/be/llvm/llvm.h               |  2 +-
 src/thorin/be/llvm/nvvm.cpp             |  4 +--
 src/thorin/be/llvm/nvvm.h               |  2 +-
 src/thorin/def.cpp                      |  2 ++
 src/thorin/transform/cleanup_world.cpp  | 33 +++++++++++-----------
 src/thorin/transform/cleanup_world.h    |  2 +-
 src/thorin/transform/flatten_tuples.cpp | 16 +++++------
 src/thorin/transform/flatten_tuples.h   |  2 +-
 src/thorin/transform/hls_channels.cpp   |  8 +++---
 src/thorin/transform/hls_channels.h     |  2 +-
 src/thorin/transform/hoist_enters.cpp   |  6 ++--
 src/thorin/transform/hoist_enters.h     |  2 +-
 src/thorin/transform/importer.cpp       | 12 ++++----
 src/thorin/transform/importer.h         | 14 ++++++----
 src/thorin/transform/inliner.cpp        |  5 ++--
 src/thorin/transform/inliner.h          |  2 +-
 src/thorin/transform/lift_builtins.cpp  |  5 ++--
 src/thorin/transform/lift_builtins.h    |  2 +-
 src/thorin/transform/split_slots.cpp    |  6 ++--
 src/thorin/transform/split_slots.h      |  2 +-
 src/thorin/world.cpp                    | 22 +++++++++------
 src/thorin/world.h                      | 27 ++++++++++--------
 31 files changed, 147 insertions(+), 128 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 3f77582fd..33ac9405f 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -73,17 +73,17 @@ enum class HlsInterface : uint8_t {
 
 class CCodeGen : public thorin::Emitter<std::string, std::string, BB, CCodeGen> {
 public:
-    CCodeGen(World& world, const Cont2Config& kernel_config, Stream& stream, Lang lang, bool debug, std::string& flags)
-        : world_(world)
+    CCodeGen(Thorin& thorin, const Cont2Config& kernel_config, Stream& stream, Lang lang, bool debug, std::string& flags)
+        : thorin_(thorin)
         , kernel_config_(kernel_config)
         , lang_(lang)
-        , fn_mem_(world.fn_type({world.mem_type()}))
+        , fn_mem_(world().fn_type({world().mem_type()}))
         , debug_(debug)
         , flags_(flags)
         , stream_(stream)
     {}
 
-    World& world() const { return world_; }
+    World& world() const { return thorin_.world(); }
     void emit_module();
     void emit_c_int();
     void emit_epilogue(Continuation*);
@@ -115,7 +115,7 @@ class CCodeGen : public thorin::Emitter<std::string, std::string, BB, CCodeGen>
     std::string array_name(const DefiniteArrayType*);
     std::string tuple_name(const TupleType*);
 
-    World& world_;
+    Thorin& thorin_;
     const Cont2Config& kernel_config_;
     Lang lang_;
     const FnType* fn_mem_;
@@ -257,10 +257,10 @@ std::string CCodeGen::convert(const Type* type) {
     } else if (auto variant = type->isa<VariantType>()) {
         types_[variant] = name = variant->name().str();
         auto tag_type =
-            variant->num_ops() < (UINT64_C(1) <<  8u) ? world_.type_qu8()  :
-            variant->num_ops() < (UINT64_C(1) << 16u) ? world_.type_qu16() :
-            variant->num_ops() < (UINT64_C(1) << 32u) ? world_.type_qu32() :
-                                                        world_.type_qu64();
+            variant->num_ops() < (UINT64_C(1) <<  8u) ? world().type_qu8()  :
+            variant->num_ops() < (UINT64_C(1) << 16u) ? world().type_qu16() :
+            variant->num_ops() < (UINT64_C(1) << 32u) ? world().type_qu32() :
+                                                        world().type_qu64();
         s.fmt("typedef struct {{\t\n");
 
         // This is required because we have zero-sized types but C/C++ do not
@@ -1453,7 +1453,7 @@ Stream& CCodeGen::emit_debug_info(Stream& s, const Def* def) {
 
 void CCodeGen::emit_c_int() {
     // Do not emit C interfaces for definitions that are not used
-    world().cleanup();
+    thorin_.cleanup();
 
     for (auto def : world().defs()) {
         auto cont = def->isa_nom<Continuation>();
@@ -1586,12 +1586,12 @@ std::string CCodeGen::tuple_name(const TupleType* tuple_type) {
 
 void CodeGen::emit_stream(std::ostream& stream) {
     Stream s(stream);
-    CCodeGen(world(), kernel_config_, s, lang_, debug_, flags_).emit_module();
+    CCodeGen(thorin(), kernel_config_, s, lang_, debug_, flags_).emit_module();
 }
 
-void emit_c_int(World& world, Stream& stream) {
+void emit_c_int(Thorin& thorin, Stream& stream) {
     std::string flags;
-    CCodeGen(world, {}, stream, Lang::C99, false, flags).emit_c_int();
+    CCodeGen(thorin, {}, stream, Lang::C99, false, flags).emit_c_int();
 }
 
 //------------------------------------------------------------------------------
diff --git a/src/thorin/be/c/c.h b/src/thorin/be/c/c.h
index 6ebe5595f..f33781ec2 100644
--- a/src/thorin/be/c/c.h
+++ b/src/thorin/be/c/c.h
@@ -16,8 +16,8 @@ enum class Lang : uint8_t { C99, HLS, CUDA, OpenCL };
 
 class CodeGen : public thorin::CodeGen {
 public:
-    CodeGen(World& world, const Cont2Config& kernel_config, Lang lang, bool debug, std::string& flags)
-        : thorin::CodeGen(world, debug)
+    CodeGen(Thorin& thorin, const Cont2Config& kernel_config, Lang lang, bool debug, std::string& flags)
+        : thorin::CodeGen(thorin, debug)
         , kernel_config_(kernel_config)
         , lang_(lang)
         , debug_(debug)
@@ -43,7 +43,7 @@ class CodeGen : public thorin::CodeGen {
     std::string flags_;
 };
 
-void emit_c_int(World&, Stream& stream);
+void emit_c_int(Thorin&, Stream& stream);
 
 }
 
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index db6f6d802..afd8afd58 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -13,14 +13,14 @@
 namespace thorin {
 
 static void get_kernel_configs(
-    Importer& importer,
+    Thorin& thorin,
     const std::vector<Continuation*>& kernels,
     Cont2Config& kernel_configs,
     std::function<std::unique_ptr<KernelConfig> (Continuation*, Continuation*)> use_callback)
 {
-    importer.world().opt();
+    thorin.opt();
 
-    auto externals = importer.world().externals();
+    auto externals = thorin.world().externals();
     for (auto continuation : kernels) {
         // recover the imported continuation (lost after the call to opt)
         Continuation* imported = nullptr;
@@ -78,8 +78,11 @@ static uint64_t get_alloc_size(const Def* def) {
 DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& flags)
     : cgs {}
 {
-    for (size_t i = 0; i < cgs.size(); ++i)
-        importers_.emplace_back(world);
+    std::vector<Importer> importers;
+    for (auto& name : backend_names) {
+        accelerator_code.emplace_back(name);
+        importers.emplace_back(world, accelerator_code.back().world());
+    }
 
     // determine different parts of the world which need to be compiled differently
     Scope::for_each(world, [&] (const Scope& scope) {
@@ -95,7 +98,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
-                imported = importers_[backend].import(continuation)->as_nom<Continuation>();
+                imported = importers[backend].import(continuation)->as_nom<Continuation>();
                 break;
             }
         }
@@ -113,8 +116,8 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     });
 
     for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU }) {
-        if (!importers_[backend].world().empty()) {
-            get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
+        if (!accelerator_code[backend].world().empty()) {
+            get_kernel_configs(accelerator_code[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 auto app = use->body();
                 // determine whether or not this kernel uses restrict pointers
                 bool has_restrict = true;
@@ -146,10 +149,10 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     // get the HLS kernel configurations
     Top2Kernel top2kernel;
     DeviceParams hls_host_params;
-    if (!importers_[HLS].world().empty()) {
-        hls_host_params = hls_channels(importers_[HLS], top2kernel, world);
+    if (!accelerator_code[HLS].world().empty()) {
+        hls_host_params = hls_channels(accelerator_code[HLS], importers[HLS], top2kernel, world);
 
-        get_kernel_configs(importers_[HLS], kernels, kernel_config, [&] (Continuation* use, Continuation* imported) {
+        get_kernel_configs(accelerator_code[HLS], kernels, kernel_config, [&] (Continuation* use, Continuation* imported) {
             auto app = use->body();
             HLSKernelConfig::Param2Size param_sizes;
             for (size_t i = hls_free_vars_offset, e = app->num_args(); i != e; ++i) {
@@ -180,22 +183,22 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
             }
             return std::make_unique<HLSKernelConfig>(param_sizes);
         });
-        hls_annotate_top(importers_[HLS].world(), top2kernel, kernel_config);
+        hls_annotate_top(importers[HLS].world(), top2kernel, kernel_config);
     }
     hls_kernel_launch(world, hls_host_params);
 
 #if THORIN_ENABLE_LLVM
-    if (!importers_[NVVM  ].world().empty()) cgs[NVVM  ] = std::make_unique<llvm::NVVMCodeGen  >(importers_[NVVM  ].world(), kernel_config,      debug);
-    if (!importers_[AMDGPU].world().empty()) cgs[AMDGPU] = std::make_unique<llvm::AMDGPUCodeGen>(importers_[AMDGPU].world(), kernel_config, opt, debug);
+    if (!accelerator_code[NVVM  ].world().empty()) cgs[NVVM  ] = std::make_unique<llvm::NVVMCodeGen  >(accelerator_code[NVVM  ], kernel_config,      debug);
+    if (!accelerator_code[AMDGPU].world().empty()) cgs[AMDGPU] = std::make_unique<llvm::AMDGPUCodeGen>(accelerator_code[AMDGPU], kernel_config, opt, debug);
 #else
     (void)opt;
 #endif
     for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OpenCL }, std::pair { HLS, c::Lang::HLS } })
-        if (!importers_[backend].world().empty()) cgs[backend] = std::make_unique<c::CodeGen>(importers_[backend].world(), kernel_config, lang, debug, flags);
+        if (!accelerator_code[backend].world().empty()) cgs[backend] = std::make_unique<c::CodeGen>(accelerator_code[backend], kernel_config, lang, debug, flags);
 }
 
-CodeGen::CodeGen(World& world, bool debug)
-    : world_(world)
+CodeGen::CodeGen(Thorin& thorin, bool debug)
+    : thorin_(thorin)
     , debug_(debug)
 {}
 
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index cd7c5689f..e124f875d 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -8,7 +8,7 @@ namespace thorin {
 
 class CodeGen {
 protected:
-    CodeGen(World& world, bool debug);
+    CodeGen(Thorin& thorin, bool debug);
 public:
     virtual ~CodeGen() {}
 
@@ -17,12 +17,13 @@ class CodeGen {
 
     /// @name getters
     //@{
-    World& world() const { return world_; }
+    Thorin& thorin() const { return thorin_; }
+    World& world() const { return thorin().world(); }
     bool debug() const { return debug_; }
     //@}
 
 private:
-    World& world_;
+    Thorin& thorin_;
     bool debug_;
 };
 
@@ -47,7 +48,8 @@ struct DeviceBackends {
     enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> cgs;
 private:
-    std::vector<Importer> importers_;
+    std::array<const char*, BackendCount> backend_names = { "CUDA", "NVVM", "OpenCL", "AMDGPU", "HLS" };
+    std::vector<Thorin> accelerator_code;
 };
 
 }
diff --git a/src/thorin/be/llvm/amdgpu.cpp b/src/thorin/be/llvm/amdgpu.cpp
index 1d19ff7a9..1f515a1bb 100644
--- a/src/thorin/be/llvm/amdgpu.cpp
+++ b/src/thorin/be/llvm/amdgpu.cpp
@@ -7,8 +7,8 @@
 
 namespace thorin::llvm {
 
-AMDGPUCodeGen::AMDGPUCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
+AMDGPUCodeGen::AMDGPUCodeGen(Thorin& thorin, const Cont2Config& kernel_config, int opt, bool debug)
+    : CodeGen(thorin, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
     , kernel_config_(kernel_config)
 {
     module().setDataLayout("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-ni:7");
diff --git a/src/thorin/be/llvm/amdgpu.h b/src/thorin/be/llvm/amdgpu.h
index 640fc5b69..fd5bd5779 100644
--- a/src/thorin/be/llvm/amdgpu.h
+++ b/src/thorin/be/llvm/amdgpu.h
@@ -13,7 +13,7 @@ namespace llvm = ::llvm;
 
 class AMDGPUCodeGen : public CodeGen {
 public:
-    AMDGPUCodeGen(World& world, const Cont2Config&, int opt, bool debug);
+    AMDGPUCodeGen(Thorin&, const Cont2Config&, int opt, bool debug);
 
     const char* file_ext() const override { return ".amdgpu"; }
 
diff --git a/src/thorin/be/llvm/cpu.cpp b/src/thorin/be/llvm/cpu.cpp
index 237a42a99..b7f5dae3b 100644
--- a/src/thorin/be/llvm/cpu.cpp
+++ b/src/thorin/be/llvm/cpu.cpp
@@ -8,8 +8,8 @@
 
 namespace thorin::llvm {
 
-CPUCodeGen::CPUCodeGen(World& world, int opt, bool debug, std::string& target_triple, std::string& target_cpu, std::string& target_attr)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
+CPUCodeGen::CPUCodeGen(Thorin& thorin, int opt, bool debug, std::string& target_triple, std::string& target_cpu, std::string& target_attr)
+    : CodeGen(thorin, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::C, opt, debug)
 {
     llvm::InitializeNativeTarget();
     auto triple_str = llvm::sys::getDefaultTargetTriple();
diff --git a/src/thorin/be/llvm/cpu.h b/src/thorin/be/llvm/cpu.h
index 7724ccc91..78db51cd9 100644
--- a/src/thorin/be/llvm/cpu.h
+++ b/src/thorin/be/llvm/cpu.h
@@ -9,7 +9,7 @@ namespace llvm = ::llvm;
 
 class CPUCodeGen : public CodeGen {
 public:
-    CPUCodeGen(World& world, int opt, bool debug, std::string& target_triple, std::string& target_cpu, std::string& target_attr);
+    CPUCodeGen(Thorin&, int opt, bool debug, std::string& target_triple, std::string& target_cpu, std::string& target_attr);
 
 protected:
     std::string get_alloc_name() const override { return "anydsl_alloc"; }
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index ce9fb159e..9b04de2a1 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -42,14 +42,14 @@
 namespace thorin::llvm {
 
 CodeGen::CodeGen(
-    World& world,
+    Thorin& thorin,
     llvm::CallingConv::ID function_calling_convention,
     llvm::CallingConv::ID device_calling_convention,
     llvm::CallingConv::ID kernel_calling_convention,
     int opt, bool debug)
-    : thorin::CodeGen(world, debug)
+    : thorin::CodeGen(thorin, debug)
     , context_(std::make_unique<llvm::LLVMContext>())
-    , module_(std::make_unique<llvm::Module>(world.name(), context()))
+    , module_(std::make_unique<llvm::Module>(world().name(), context()))
     , opt_(opt)
     , dibuilder_(module())
     , function_calling_convention_(function_calling_convention)
diff --git a/src/thorin/be/llvm/llvm.h b/src/thorin/be/llvm/llvm.h
index 93e9aeb7e..7dc6f6b90 100644
--- a/src/thorin/be/llvm/llvm.h
+++ b/src/thorin/be/llvm/llvm.h
@@ -28,7 +28,7 @@ using BB = std::pair<llvm::BasicBlock*, std::unique_ptr<llvm::IRBuilder<>>>;
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<llvm::Value*, llvm::Type*, BB, CodeGen> {
 protected:
     CodeGen(
-        World& world,
+        Thorin&,
         llvm::CallingConv::ID function_calling_convention,
         llvm::CallingConv::ID device_calling_convention,
         llvm::CallingConv::ID kernel_calling_convention,
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index 8a80e9a07..180f3ac2e 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -19,8 +19,8 @@
 
 namespace thorin::llvm {
 
-NVVMCodeGen::NVVMCodeGen(World& world, const Cont2Config& kernel_config, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::PTX_Device, llvm::CallingConv::PTX_Kernel, 0, debug)
+NVVMCodeGen::NVVMCodeGen(Thorin& thorin, const Cont2Config& kernel_config, bool debug)
+    : CodeGen(thorin, llvm::CallingConv::C, llvm::CallingConv::PTX_Device, llvm::CallingConv::PTX_Kernel, 0, debug)
     , kernel_config_(kernel_config)
 {
     auto triple = llvm::Triple(llvm::sys::getDefaultTargetTriple());
diff --git a/src/thorin/be/llvm/nvvm.h b/src/thorin/be/llvm/nvvm.h
index ee6f5239f..204ce6648 100644
--- a/src/thorin/be/llvm/nvvm.h
+++ b/src/thorin/be/llvm/nvvm.h
@@ -13,7 +13,7 @@ namespace llvm = ::llvm;
 
 class NVVMCodeGen : public CodeGen {
 public:
-    NVVMCodeGen(World& world, const Cont2Config&, bool debug); // NVVM-specific optimizations are run in the runtime
+    NVVMCodeGen(Thorin&, const Cont2Config&, bool debug); // NVVM-specific optimizations are run in the runtime
 
     const char* file_ext() const override { return ".nvvm"; }
 
diff --git a/src/thorin/def.cpp b/src/thorin/def.cpp
index 03516f1d9..a0d14e59c 100644
--- a/src/thorin/def.cpp
+++ b/src/thorin/def.cpp
@@ -56,6 +56,7 @@ void Def::set_name(const std::string& name) const { debug_.name = name; }
 void Def::set_op(size_t i, const Def* def) {
     assert(!op(i) && "already set");
     assert(def && "setting null pointer");
+    assert(&def->world() == &world());
     ops_[i] = def;
     // A Param/Continuation should not have other bits than its own set.
     // (Right now, Param doesn't have ops, but this will change in the future).
@@ -136,6 +137,7 @@ bool is_minus_zero(const Def* def) {
 void Def::replace_uses(const Def* with) const {
     world().DLOG("replace uses: {} -> {}", this, with);
     if (this != with) {
+        assert(&with->world() == &this->world());
         for (auto& use : copy_uses()) {
             auto def = const_cast<Def*>(use.def());
             auto index = use.index();
diff --git a/src/thorin/transform/cleanup_world.cpp b/src/thorin/transform/cleanup_world.cpp
index 2748d25d9..041db668a 100644
--- a/src/thorin/transform/cleanup_world.cpp
+++ b/src/thorin/transform/cleanup_world.cpp
@@ -14,11 +14,11 @@ namespace thorin {
 
 class Cleaner {
 public:
-    Cleaner(World& world)
+    Cleaner(std::unique_ptr<World>& world)
         : world_(world)
     {}
 
-    World& world() { return world_; }
+    World& world() { return *world_; }
     void cleanup();
     void eliminate_tail_rec();
     void eta_conversion();
@@ -31,12 +31,12 @@ class Cleaner {
 private:
     void cleanup_fix_point();
     void clean_pe_info(std::queue<Continuation*>, Continuation*);
-    World& world_;
+    std::unique_ptr<World>& world_;
     bool todo_ = true;
 };
 
 void Cleaner::eliminate_tail_rec() {
-    Scope::for_each(world_, [&](Scope& scope) {
+    Scope::for_each(*world_, [&](Scope& scope) {
         auto entry = scope.entry();
 
         bool only_tail_calls = true;
@@ -232,16 +232,17 @@ next_continuation:;
 }
 
 void Cleaner::rebuild() {
-    Importer importer(world_);
-    importer.type_old2new_.rehash(world_.types().capacity());
-    importer.def_old2new_.rehash(world_.defs().capacity());
+    auto fresh_world = std::make_unique<World>(world());
+    Importer importer(*world_, *fresh_world);
+    importer.type_old2new_.rehash(world_->types().capacity());
+    importer.def_old2new_.rehash(world_->defs().capacity());
 
     for (auto&& [_, cont] : world().externals()) {
         if (cont->is_exported())
             importer.import(cont);
     }
 
-    swap(importer.world(), world_);
+    std::swap(world_, fresh_world);
 
     // verify(world());
 
@@ -279,7 +280,7 @@ void Cleaner::clean_pe_info(std::queue<Continuation*> queue, Continuation* cur)
     auto next = body->arg(3);
     auto msg = body->arg(1)->as<Bitcast>()->from()->as<Global>()->init()->as<DefiniteArray>();
 
-    world_.idef(body->callee(), "pe_info was not constant: {}: {}", msg->as_string(), body->arg(2));
+    world_->idef(body->callee(), "pe_info was not constant: {}: {}", msg->as_string(), body->arg(2));
     cur->jump(next, {body->arg(0)}, cur->debug()); // TODO debug
     todo_ = true;
 
@@ -288,7 +289,7 @@ void Cleaner::clean_pe_info(std::queue<Continuation*> queue, Continuation* cur)
 }
 
 void Cleaner::clean_pe_infos() {
-    world_.VLOG("cleaning remaining pe_infos");
+    world_->VLOG("cleaning remaining pe_infos");
     std::queue<Continuation*> queue;
     ContinuationSet done;
     auto enqueue = [&](Continuation* continuation) {
@@ -319,9 +320,9 @@ void Cleaner::clean_pe_infos() {
 void Cleaner::cleanup_fix_point() {
     int i = 0;
     for (; todo_; ++i) {
-        world_.VLOG("iteration: {}", i);
+        world_->VLOG("iteration: {}", i);
         todo_ = false;
-        if (world_.is_pe_done())
+        if (world_->is_pe_done())
             eliminate_tail_rec();
         eta_conversion();
         eliminate_params();
@@ -329,14 +330,14 @@ void Cleaner::cleanup_fix_point() {
         todo_ |= resolve_loads(world());
         rebuild();
         if (!world().is_pe_done())
-            todo_ |= partial_evaluation(world_);
+            todo_ |= partial_evaluation(*world_);
         else
             clean_pe_infos();
     }
 }
 
 void Cleaner::cleanup() {
-    world_.VLOG("start cleanup");
+    world_->VLOG("start cleanup");
     cleanup_fix_point();
 
     if (!world().is_pe_done()) {
@@ -350,13 +351,13 @@ void Cleaner::cleanup() {
         cleanup_fix_point();
     }
 
-    world_.VLOG("end cleanup");
+    world_->VLOG("end cleanup");
 #if THORIN_ENABLE_CHECKS
     verify_closedness();
     debug_verify(world());
 #endif
 }
 
-void cleanup_world(World& world) { Cleaner(world).cleanup(); }
+void cleanup_world(std::unique_ptr<World>& world) { Cleaner(world).cleanup(); }
 
 }
diff --git a/src/thorin/transform/cleanup_world.h b/src/thorin/transform/cleanup_world.h
index 22c91b0b6..afcb68b75 100644
--- a/src/thorin/transform/cleanup_world.h
+++ b/src/thorin/transform/cleanup_world.h
@@ -5,7 +5,7 @@ namespace thorin {
 
 class World;
 
-void cleanup_world(World& world);
+void cleanup_world(std::unique_ptr<World>& world);
 
 }
 
diff --git a/src/thorin/transform/flatten_tuples.cpp b/src/thorin/transform/flatten_tuples.cpp
index 0b4529c25..8f99bd220 100644
--- a/src/thorin/transform/flatten_tuples.cpp
+++ b/src/thorin/transform/flatten_tuples.cpp
@@ -166,7 +166,7 @@ static Continuation* unwrap_def(Def2Def& wrapped, Def2Def& unwrapped, const Def*
     return jump(old_cont, call_args);
 }
 
-static void flatten_tuples(World& world, size_t max_tuple_size) {
+static void flatten_tuples(Thorin& thorin, size_t max_tuple_size) {
     // flatten tuples passed as arguments to functions
     bool todo = true;
     Def2Def wrapped, unwrapped;
@@ -177,11 +177,11 @@ static void flatten_tuples(World& world, size_t max_tuple_size) {
 
         for (auto pair : unwrapped) unwrapped_codom.emplace(pair.second);
 
-        for (auto cont : world.copy_continuations()) {
+        for (auto cont : thorin.world().copy_continuations()) {
             // do not change the signature of intrinsic/external functions
             if (!cont->has_body() ||
                 cont->is_intrinsic() ||
-                world.is_external(cont) ||
+                thorin.world().is_external(cont) ||
                 is_passed_to_accelerator(cont))
                 continue;
 
@@ -196,7 +196,7 @@ static void flatten_tuples(World& world, size_t max_tuple_size) {
 
             todo = true;
 
-            world.DLOG("flattened {}", cont);
+            thorin.world().DLOG("flattened {}", cont);
         }
 
         // remove original versions of wrapped functions
@@ -218,12 +218,12 @@ static void flatten_tuples(World& world, size_t max_tuple_size) {
     for (auto unwrap_pair : unwrapped)
         inline_calls(unwrap_pair.second->as_nom<Continuation>());
 
-    world.cleanup();
-    debug_verify(world);
+    thorin.cleanup();
+    debug_verify(thorin.world());
 }
 
-void flatten_tuples(World& world) {
-    flatten_tuples(world, std::numeric_limits<size_t>::max());
+void flatten_tuples(Thorin& thorin) {
+    flatten_tuples(thorin, std::numeric_limits<size_t>::max());
 }
 
 }
diff --git a/src/thorin/transform/flatten_tuples.h b/src/thorin/transform/flatten_tuples.h
index 74b16c792..0e2cbb91d 100644
--- a/src/thorin/transform/flatten_tuples.h
+++ b/src/thorin/transform/flatten_tuples.h
@@ -2,6 +2,6 @@
 
 namespace thorin {
 
-void flatten_tuples(World& world);
+void flatten_tuples(Thorin& thorin);
 
 }
diff --git a/src/thorin/transform/hls_channels.cpp b/src/thorin/transform/hls_channels.cpp
index 954375b10..c778baf8f 100644
--- a/src/thorin/transform/hls_channels.cpp
+++ b/src/thorin/transform/hls_channels.cpp
@@ -145,14 +145,14 @@ bool dependency_resolver(Dependencies& dependencies, const size_t dependent_kern
 }
 
 /**
- * @param importer hls world
+ * @param thorin hls world
  * @param Top2Kernel annonating hls_top configuration
  * @param old_world to connect with runtime (host) world
  * @return corresponding hls_top parameter for hls_launch_kernel in another world (params before rewriting kernels)
  */
 
-DeviceParams hls_channels(Importer& importer, Top2Kernel& top2kernel, World& old_world) {
-    auto& world = importer.world();
+DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2kernel, World& old_world) {
+    auto& world = thorin.world();
     std::vector<Def2Mode> kernels_ch_modes; // vector of channel->mode maps for kernels which use channel(s)
     std::vector<Continuation*> new_kernels;
     Def2Def kernel_new2old;
@@ -403,7 +403,7 @@ DeviceParams hls_channels(Importer& importer, Top2Kernel& top2kernel, World& old
     world.make_external(hls_top);
 
     debug_verify(world);
-    world.cleanup();
+    thorin.cleanup();
 
     return old_kernels_params;
 }
diff --git a/src/thorin/transform/hls_channels.h b/src/thorin/transform/hls_channels.h
index 21a50f873..b2296fa44 100644
--- a/src/thorin/transform/hls_channels.h
+++ b/src/thorin/transform/hls_channels.h
@@ -23,7 +23,7 @@ class World;
  * resolves all dependency requirements between kernel calls
  * provides hls_top parameters for hls runtime
  */
-DeviceParams hls_channels(Importer&, Top2Kernel&, World&);
+DeviceParams hls_channels(Thorin&, Importer&, Top2Kernel&, World&);
 void hls_annotate_top(World&, const Top2Kernel&, Cont2Config&);
 
 }
diff --git a/src/thorin/transform/hoist_enters.cpp b/src/thorin/transform/hoist_enters.cpp
index c052498e2..a51cf8d64 100644
--- a/src/thorin/transform/hoist_enters.cpp
+++ b/src/thorin/transform/hoist_enters.cpp
@@ -57,9 +57,9 @@ static void hoist_enters(const Scope& scope) {
         entry_enter->out_mem()->replace_uses(entry_enter->mem());
 }
 
-void hoist_enters(World& world) {
-    Scope::for_each(world, [] (const Scope& scope) { hoist_enters(scope); });
-    world.cleanup();
+void hoist_enters(Thorin& thorin) {
+    Scope::for_each(thorin.world(), [] (const Scope& scope) { hoist_enters(scope); });
+    thorin.cleanup();
 }
 
 }
diff --git a/src/thorin/transform/hoist_enters.h b/src/thorin/transform/hoist_enters.h
index bff4d9492..5d3155f76 100644
--- a/src/thorin/transform/hoist_enters.h
+++ b/src/thorin/transform/hoist_enters.h
@@ -5,7 +5,7 @@ namespace thorin {
 
 class World;
 
-void hoist_enters(World&);
+void hoist_enters(Thorin&);
 
 }
 
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index f125585de..ceaabb6f2 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -4,13 +4,13 @@ namespace thorin {
 
 const Type* Importer::import(const Type* otype) {
     if (auto ntype = type_old2new_.lookup(otype)) {
-        assert(&(*ntype)->table() == &world_);
+        assert(&(*ntype)->table() == &world());
         return *ntype;
     }
     size_t size = otype->num_ops();
 
     if (auto nominal_type = otype->isa<NominalType>()) {
-        auto ntype = nominal_type->stub(world_);
+        auto ntype = nominal_type->stub(world());
         type_old2new_[otype] = ntype;
         for (size_t i = 0; i != size; ++i)
             ntype->set(i, import(otype->op(i)));
@@ -21,16 +21,16 @@ const Type* Importer::import(const Type* otype) {
     for (size_t i = 0; i != size; ++i)
         nops[i] = import(otype->op(i));
 
-    auto ntype = otype->rebuild(world_, nops);
+    auto ntype = otype->rebuild(world(), nops);
     type_old2new_[otype] = ntype;
-    assert(&ntype->table() == &world_);
+    assert(&ntype->table() == &world());
 
     return ntype;
 }
 
 const Def* Importer::import(const Def* odef) {
     if (auto ndef = def_old2new_.lookup(odef)) {
-        assert(&(*ndef)->world() == &world_);
+        assert(&(*ndef)->world() == &world());
         return *ndef;
     }
 
@@ -39,7 +39,7 @@ const Def* Importer::import(const Def* odef) {
     if (auto oparam = odef->isa<Param>()) {
         import(oparam->continuation())->as_nom<Continuation>();
         auto nparam = def_old2new_[oparam];
-        assert(nparam && &nparam->world() == &world_);
+        assert(nparam && &nparam->world() == &world());
         return nparam;
     }
 
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index c34db9ef3..db41b3161 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -8,18 +8,19 @@ namespace thorin {
 
 class Importer {
 public:
-    Importer(World& src)
-        : world_(src)
+    explicit Importer(World& src, World& dst)
+        : src(src)
+        , dst(dst)
     {
         if (src.is_pe_done())
-            world_.mark_pe_done();
+            world().mark_pe_done();
 #if THORIN_ENABLE_CHECKS
         if (src.track_history())
-            world_.enable_history(true);
+            world().enable_history(true);
 #endif
     }
 
-    World& world() { return world_; }
+    World& world() { return dst; }
     const Type* import(const Type*);
     const Def* import(const Def*);
     bool todo() const { return todo_; }
@@ -27,7 +28,8 @@ class Importer {
 public:
     Type2Type type_old2new_;
     Def2Def def_old2new_;
-    World world_;
+    World& src;
+    World& dst;
     bool todo_ = false;
 };
 
diff --git a/src/thorin/transform/inliner.cpp b/src/thorin/transform/inliner.cpp
index 175f2a518..2205ad4cc 100644
--- a/src/thorin/transform/inliner.cpp
+++ b/src/thorin/transform/inliner.cpp
@@ -36,7 +36,8 @@ void force_inline(Scope& scope, int threshold) {
     }
 }
 
-void inliner(World& world) {
+void inliner(Thorin& thorin) {
+    World& world = thorin.world();
     world.VLOG("start inliner");
 
     static const int factor = 4;
@@ -95,7 +96,7 @@ void inliner(World& world) {
 
     world.VLOG("stop inliner");
     debug_verify(world);
-    world.cleanup();
+    thorin.cleanup();
 
 }
 
diff --git a/src/thorin/transform/inliner.h b/src/thorin/transform/inliner.h
index 89b1dcfd4..dc22de086 100644
--- a/src/thorin/transform/inliner.h
+++ b/src/thorin/transform/inliner.h
@@ -11,7 +11,7 @@ class World;
  * If there still remain functions to be inlined, warnings will be emitted
  */
 void force_inline(Scope& scope, int threshold);
-void inliner(World& world);
+void inliner(Thorin&);
 
 }
 
diff --git a/src/thorin/transform/lift_builtins.cpp b/src/thorin/transform/lift_builtins.cpp
index 82dd10d43..cfc65a1de 100644
--- a/src/thorin/transform/lift_builtins.cpp
+++ b/src/thorin/transform/lift_builtins.cpp
@@ -63,8 +63,9 @@ void lift_pipeline(World& world) {
 
 }
 
-void lift_builtins(World& world) {
+void lift_builtins(Thorin& thorin) {
     // This must be run first
+    World& world = thorin.world();
     lift_pipeline(world);
 
     while (true) {
@@ -125,7 +126,7 @@ void lift_builtins(World& world) {
             }
         }
 
-        world.cleanup();
+        thorin.cleanup();
     }
 }
 
diff --git a/src/thorin/transform/lift_builtins.h b/src/thorin/transform/lift_builtins.h
index 7da4d5e05..e9c413e0b 100644
--- a/src/thorin/transform/lift_builtins.h
+++ b/src/thorin/transform/lift_builtins.h
@@ -5,7 +5,7 @@ namespace thorin {
 
 class World;
 
-void lift_builtins(World&);
+void lift_builtins(Thorin&);
 
 }
 
diff --git a/src/thorin/transform/split_slots.cpp b/src/thorin/transform/split_slots.cpp
index 5289777c9..1f81e763d 100644
--- a/src/thorin/transform/split_slots.cpp
+++ b/src/thorin/transform/split_slots.cpp
@@ -86,12 +86,12 @@ static bool split_slots(const Scope& scope) {
     return todo;
 }
 
-void split_slots(World& world) {
+void split_slots(Thorin& thorin) {
     bool todo = true;
     while (todo) {
         todo = false;
-        Scope::for_each(world, [&] (const Scope& scope) { todo |= split_slots(scope); });
-        world.cleanup();
+        Scope::for_each(thorin.world(), [&] (const Scope& scope) { todo |= split_slots(scope); });
+        thorin.cleanup();
     }
 }
 
diff --git a/src/thorin/transform/split_slots.h b/src/thorin/transform/split_slots.h
index dff363e43..767ac1ae8 100644
--- a/src/thorin/transform/split_slots.h
+++ b/src/thorin/transform/split_slots.h
@@ -8,7 +8,7 @@ class World;
 /**
  * Tries to split @p Slot%s that are accessed through constant @p LEA%s.
  */
-void split_slots(World&);
+void split_slots(Thorin&);
 
 }
 
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 013d09544..a45ac14a4 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -1276,28 +1276,32 @@ const Def* World::cse_base(const Def* def) {
  * optimizations
  */
 
-void World::cleanup() { cleanup_world(*this); }
+Thorin::Thorin(const std::string& name)
+    : world_(std::make_unique<World>(name))
+{}
 
-void World::opt() {
+void Thorin::cleanup() { cleanup_world(world_); }
+
+void Thorin::opt() {
 #define RUN_PASS(pass) \
 { \
-    VLOG("running pass {}", #pass); \
+    world().VLOG("running pass {}", #pass); \
     pass; \
-    debug_verify(*this); \
+    debug_verify(world()); \
 }
 
     RUN_PASS(cleanup())
-    RUN_PASS(while (partial_evaluation(*this, true))); // lower2cff
+    RUN_PASS(while (partial_evaluation(world(), true))); // lower2cff
     RUN_PASS(flatten_tuples(*this))
-    RUN_PASS(clone_bodies(*this))
+    RUN_PASS(clone_bodies(world()))
     RUN_PASS(split_slots(*this))
-    RUN_PASS(closure_conversion(*this))
+    RUN_PASS(closure_conversion(world()))
     RUN_PASS(lift_builtins(*this))
     RUN_PASS(inliner(*this))
     RUN_PASS(hoist_enters(*this))
-    RUN_PASS(dead_load_opt(*this))
+    RUN_PASS(dead_load_opt(world()))
     RUN_PASS(cleanup())
-    RUN_PASS(codegen_prepare(*this))
+    RUN_PASS(codegen_prepare(world()))
 }
 
 }
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 940680ef7..3a0d11be0 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -245,10 +245,6 @@ class World : public TypeTable, public Streamable<World> {
     Continuation* end_scope() const { return data_.end_scope_; }
     const Filter* filter(const Defs, Debug dbg = {});
 
-    /// Performs dead code, unreachable code and unused type elimination.
-    void cleanup();
-    void opt();
-
     // getters
 
     const std::string& name() const { return data_.name_; }
@@ -310,14 +306,6 @@ class World : public TypeTable, public Streamable<World> {
     static std::string colorize(const std::string& str, int color);
     //@}
 
-    friend void swap(World& w1, World& w2) {
-        using std::swap;
-        swap(static_cast<TypeTable&>(w1), static_cast<TypeTable&>(w2));
-        swap(w1.state_,  w2.state_);
-        swap(w1.data_,   w2.data_);
-        swap(w1.stream_, w2.stream_);
-    }
-
 private:
     const Param* param(const Type* type, Continuation* continuation, size_t index, Debug dbg);
     const App* app(const Def* callee, const Defs args, Debug dbg = {});
@@ -369,6 +357,21 @@ class World : public TypeTable, public Streamable<World> {
     friend class Filter;
     friend class App;
     friend class Importer;
+    friend class Thorin;
+};
+
+class Thorin {
+public:
+    /// Initial world constructor
+    explicit Thorin(const std::string& name);
+
+    World& world() { return *world_; };
+
+    /// Performs dead code, unreachable code and unused type elimination.
+    void cleanup();
+    void opt();
+private:
+    std::unique_ptr<World> world_;
 };
 
 }

From 9aeac583e614db344e52559e2fefef979d66ba70 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 23 Jan 2023 17:31:48 +0100
Subject: [PATCH 158/342] merged Type and Def

---
 src/thorin/be/c/c.cpp                       |  30 +-
 src/thorin/be/emitter.h                     |   2 +-
 src/thorin/be/llvm/llvm.cpp                 |  25 +-
 src/thorin/be/llvm/nvvm.cpp                 |   2 +-
 src/thorin/be/llvm/parallel.cpp             |   6 +-
 src/thorin/be/llvm/runtime.cpp              |   8 +-
 src/thorin/continuation.cpp                 |  76 ++++-
 src/thorin/continuation.h                   |  11 +-
 src/thorin/def.cpp                          |  18 +-
 src/thorin/def.h                            |  33 ++-
 src/thorin/primop.cpp                       |  40 +--
 src/thorin/primop.h                         |  67 ++---
 src/thorin/rec_stream.cpp                   |   2 -
 src/thorin/tables/nodetable.h               |   1 +
 src/thorin/transform/cleanup_world.cpp      |   7 +-
 src/thorin/transform/closure_conversion.cpp |  35 +--
 src/thorin/transform/flatten_tuples.cpp     |   6 +-
 src/thorin/transform/hls_channels.cpp       |   6 +-
 src/thorin/transform/importer.cpp           |  29 +-
 src/thorin/transform/importer.h             |   3 +-
 src/thorin/transform/mangle.cpp             |   2 +-
 src/thorin/transform/mangle.h               |   1 -
 src/thorin/type.cpp                         | 144 ++++------
 src/thorin/type.h                           | 304 ++++++++------------
 src/thorin/world.cpp                        |  32 +--
 src/thorin/world.h                          |  41 ++-
 26 files changed, 474 insertions(+), 457 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 33ac9405f..ba109b4af 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -218,7 +218,7 @@ std::string CCodeGen::convert(const Type* type) {
     StringStream s;
     std::string name;
 
-    if (type == world().unit() || type->isa<MemType>() || type->isa<FrameType>())
+    if (type == world().unit_type() || type->isa<MemType>() || type->isa<FrameType>())
         s << "void";
     else if (auto primtype = type->isa<PrimType>()) {
         switch (primtype->primtype_tag()) {
@@ -252,7 +252,7 @@ std::string CCodeGen::convert(const Type* type) {
     } else if (auto tuple = type->isa<TupleType>()) {
         name = tuple_name(tuple);
         s.fmt("typedef struct {{\t\n");
-        s.rangei(tuple->ops(), "\n", [&](size_t i) { s.fmt("{} e{};", convert(tuple->op(i)), i); });
+        s.rangei(tuple->ops(), "\n", [&](size_t i) { s.fmt("{} e{};", convert(tuple->types()[i]), i); });
         s.fmt("\b\n}} {};\n", name);
     } else if (auto variant = type->isa<VariantType>()) {
         types_[variant] = name = variant->name().str();
@@ -267,9 +267,9 @@ std::string CCodeGen::convert(const Type* type) {
         if (variant->has_payload()) {
             s.fmt("union {{\t\n");
             s.rangei(variant->ops(), "\n", [&] (size_t i) {
-                if (is_type_unit(variant->op(i)))
+                if (is_type_unit(variant->types()[i]))
                     s << "// ";
-                s.fmt("{} {};", convert(variant->op(i)), variant->op_name(i));
+                s.fmt("{} {};", convert(variant->types()[i]), variant->op_name(i));
             });
             s.fmt("\b\n}} data;\n");
         }
@@ -281,14 +281,14 @@ std::string CCodeGen::convert(const Type* type) {
         if ((lang_ == Lang::OpenCL || lang_ == Lang::HLS) && is_channel_type(struct_type))
             use_channels_ = true;
         if (lang_ == Lang::OpenCL && use_channels_) {
-            s.fmt("typedef {} {}_{};\n", convert(struct_type->op(0)), name, struct_type->gid());
+            s.fmt("typedef {} {}_{};\n", convert(struct_type->types()[0]), name, struct_type->gid());
             name = (struct_type->name().str() + "_" + std::to_string(type->gid()));
         } else if (is_channel_type(struct_type) && lang_ == Lang::HLS) {
-            s.fmt("typedef {} {}_{};\n", convert(struct_type->op(0)), name, struct_type->gid());
+            s.fmt("typedef {} {}_{};\n", convert(struct_type->types()[0]), name, struct_type->gid());
             name = ("hls::stream<" + name + "_" + std::to_string(type->gid()) + ">");
         } else {
             s.fmt("typedef struct {{\t\n");
-            s.rangei(struct_type->ops(), "\n", [&] (size_t i) { s.fmt("{} {};", convert(struct_type->op(i)), struct_type->op_name(i)); });
+            s.rangei(struct_type->ops(), "\n", [&] (size_t i) { s.fmt("{} {};", convert(struct_type->types()[i]), struct_type->op_name(i)); });
             s.fmt("\b\n}} {};\n", name);
         }
     } else {
@@ -526,15 +526,15 @@ static inline bool is_passed_via_buffer(const Param* param) {
 
 static inline const Type* ret_type(const FnType* fn_type) {
     auto ret_fn_type = (*std::find_if(
-        fn_type->ops().begin(), fn_type->ops().end(), [] (const Type* op) {
+        fn_type->types().begin(), fn_type->types().end(), [] (const Type* op) {
             return op->order() % 2 == 1;
         }))->as<FnType>();
     std::vector<const Type*> types;
-    for (auto op : ret_fn_type->ops()) {
+    for (auto op : ret_fn_type->types()) {
         if (op->isa<MemType>() || is_type_unit(op) || op->order() > 0) continue;
         types.push_back(op);
     }
-    return fn_type->table().tuple_type(types);
+    return fn_type->world().tuple_type(types);
 }
 
 static inline const Type* pointee_or_elem_type(const PtrType* ptr_type) {
@@ -913,12 +913,12 @@ std::string CCodeGen::emit_bottom(const Type* type) {
         StringStream s;
         s.fmt("{} ", constructor_prefix(type));
         s << "{ ";
-        s.range(type->ops(), ", ", [&] (const Type* op) { s << emit_bottom(op); });
+        s.range(type->ops(), ", ", [&] (const Def* op) { s << emit_bottom(op->as<Type>()); });
         s << " }";
         return s.str();
     } else if (auto variant_type = type->isa<VariantType>()) {
         if (variant_type->has_payload()) {
-            auto non_unit = *std::find_if(variant_type->ops().begin(), variant_type->ops().end(),
+            auto non_unit = *std::find_if(variant_type->types().begin(), variant_type->types().end(),
                 [] (const Type* op) { return !is_type_unit(op); });
             return constructor_prefix(type) + " { { " + emit_bottom(non_unit) + " }, 0 }";
         }
@@ -1252,7 +1252,7 @@ std::string CCodeGen::emit_def(BB* bb, const Def* def) {
         if (auto tup = ass->type()->isa<TupleType>()) {
             for (size_t i = 1, e = tup->num_ops(); i != e; ++i) {
                 auto name = ass->out(i)->unique_name();
-                func_impls_.fmt("{} {};\n", convert(tup->op(i)), name);
+                func_impls_.fmt("{} {};\n", convert(tup->types()[i]), name);
                 func_defs_.insert(ass->out(i));
                 outputs.emplace_back(name);
                 defs_[ass->out(i)] = name;
@@ -1465,10 +1465,10 @@ void CCodeGen::emit_c_int() {
             continue;
 
         // Generate C types for structs used by imported or exported functions
-        for (auto op : cont->type()->ops()) {
+        for (auto op : cont->type()->types()) {
             if (auto fn_type = op->isa<FnType>()) {
                 // Convert the return types as well (those are embedded in return continuations)
-                for (auto other_op : fn_type->ops()) {
+                for (auto other_op : fn_type->types()) {
                     if (!other_op->isa<FnType>())
                         convert(other_op);
                 }
diff --git a/src/thorin/be/emitter.h b/src/thorin/be/emitter.h
index 72300ba06..90caef62f 100644
--- a/src/thorin/be/emitter.h
+++ b/src/thorin/be/emitter.h
@@ -66,7 +66,7 @@ class Emitter {
 
     Scheduler scheduler_;
     DefMap<Value> defs_;
-    TypeMap<Type> types_;
+    DefMap<Type> types_;
     ContinuationMap<BB> cont2bb_;
     Continuation* entry_ = nullptr;
 };
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 9b04de2a1..cdcc260aa 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -144,14 +144,14 @@ llvm::Type* CodeGen::convert(const Type* type) {
             auto fn = type->as<FnType>();
             llvm::Type* ret = nullptr;
             std::vector<llvm::Type*> ops;
-            for (auto op : fn->ops()) {
-                if (op->isa<MemType>() || op == world().unit()) continue;
+            for (auto op : fn->types()) {
+                if (op->isa<MemType>() || op == world().unit_type()) continue;
                 auto fn = op->isa<FnType>();
                 if (fn && !op->isa<ClosureType>()) {
                     assert(!ret && "only one 'return' supported");
                     std::vector<llvm::Type*> ret_types;
-                    for (auto fn_op : fn->ops()) {
-                        if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
+                    for (auto fn_op : fn->types()) {
+                        if (fn_op->isa<MemType>() || fn_op == world().unit_type()) continue;
                         ret_types.push_back(convert(fn_op));
                     }
                     if (ret_types.size() == 0)      ret = llvm::Type::getVoidTy(context());
@@ -185,7 +185,7 @@ llvm::Type* CodeGen::convert(const Type* type) {
 
             Array<llvm::Type*> llvm_types(struct_type->num_ops());
             for (size_t i = 0, e = llvm_types.size(); i != e; ++i)
-                llvm_types[i] = convert(struct_type->op(i));
+                llvm_types[i] = convert(struct_type->types()[i]);
             llvm_struct->setBody(llvm_ref(llvm_types));
             return llvm_struct;
         }
@@ -194,19 +194,20 @@ llvm::Type* CodeGen::convert(const Type* type) {
             auto tuple = type->as<TupleType>();
             Array<llvm::Type*> llvm_types(tuple->num_ops());
             for (size_t i = 0, e = llvm_types.size(); i != e; ++i)
-                llvm_types[i] = convert(tuple->op(i));
+                llvm_types[i] = convert(tuple->types()[i]);
             llvm_type = llvm::StructType::get(context(), llvm_ref(llvm_types));
             return types_[tuple] = llvm_type;
         }
 
         case Node_VariantType: {
+            auto variant_type = type->as<VariantType>();
             assert(type->num_ops() > 0);
             // Max alignment/size constraints respectively in the variant type alternatives dictate the ones to use for the overall type
             size_t max_align = 0, max_size = 0;
 
             auto layout = module().getDataLayout();
             llvm::Type* max_align_type;
-            for (auto op : type->ops()) {
+            for (auto op : variant_type->types()) {
                 auto op_type = convert(op);
                 size_t size  = layout.getTypeAllocSize(op_type);
                 size_t align = layout.getABITypeAlignment(op_type);
@@ -828,7 +829,7 @@ llvm::Value* CodeGen::emit_bb(BB& bb, const Def* def) {
     } else if (auto variant_extract = def->isa<VariantExtract>()) {
         auto variant_value = variant_extract->op(0);
         auto llvm_value    = emit(variant_value);
-        auto target_type   = variant_value->type()->op(variant_extract->index());
+        auto target_type   = variant_value->type()->op(variant_extract->index())->as<Type>();
         if (is_type_unit(target_type))
             return nullptr;
 
@@ -1081,15 +1082,15 @@ llvm::Value* CodeGen::emit_lea(llvm::IRBuilder<>& irbuilder, const LEA* lea) {
 
 llvm::Value* CodeGen::emit_assembly(llvm::IRBuilder<>& irbuilder, const Assembly* assembly) {
     emit_unsafe(assembly->mem());
-    auto out_type = assembly->type();
+    auto out_type = assembly->type()->isa<TupleType>();
     llvm::Type* res_type;
     bool mem_only = false;
 
-    if (out_type->isa<TupleType>()) {
+    if (out_type) {
         if (out_type->num_ops() == 2)
-            res_type = convert(assembly->type()->op(1));
+            res_type = convert(out_type->types()[1]);
         else
-            res_type = convert(world().tuple_type(assembly->type()->ops().skip_front()));
+            res_type = convert(world().tuple_type(out_type->types().skip_front()));
     } else {
         res_type = llvm::Type::getVoidTy(context());
         mem_only = true;
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index 180f3ac2e..2374b48b8 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -54,7 +54,7 @@ static AddrSpace resolve_addr_space(const Def* def) {
 llvm::FunctionType* NVVMCodeGen::convert_fn_type(Continuation* continuation) {
     // skip non-global address-space parameters
     std::vector<const Type*> types;
-    for (auto type : continuation->type()->ops()) {
+    for (auto type : continuation->type()->types()) {
         if (auto ptr = type->isa<PtrType>())
             if (ptr->addr_space() == AddrSpace::Texture)
                 continue;
diff --git a/src/thorin/be/llvm/parallel.cpp b/src/thorin/be/llvm/parallel.cpp
index 5494ddbe1..ed3dac954 100644
--- a/src/thorin/be/llvm/parallel.cpp
+++ b/src/thorin/be/llvm/parallel.cpp
@@ -36,7 +36,7 @@ Continuation* CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation*
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->ops().skip_front(PAR_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(PAR_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -131,7 +131,7 @@ Continuation* CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* c
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->ops().skip_front(FIB_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(FIB_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -215,7 +215,7 @@ Continuation* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* co
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->ops().skip_front(SPAWN_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(SPAWN_NUM_ARGS)));
     llvm::Value* closure = nullptr;
     if (closure_type->isStructTy()) {
         closure = llvm::UndefValue::get(closure_type);
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index 47e7dda8f..04f9a6d0b 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -45,15 +45,15 @@ static bool contains_ptrtype(const Type* type) {
         case Node_StructType: {
             bool good = true;
             auto struct_type = type->as<StructType>();
-            for (size_t i = 0, e = struct_type->num_ops(); i != e; ++i)
-                good &= contains_ptrtype(struct_type->op(i));
+            for (auto& t : struct_type->types())
+                good &= contains_ptrtype(t);
             return good;
         }
         case Node_TupleType: {
             bool good = true;
             auto tuple = type->as<TupleType>();
-            for (size_t i = 0, e = tuple->num_ops(); i != e; ++i)
-                good &= contains_ptrtype(tuple->op(i));
+            for (auto& t : tuple->types())
+                good &= contains_ptrtype(t);
             return good;
         }
         default: return true;
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index c66302040..ccf1e6dc0 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -10,16 +10,27 @@ namespace thorin {
 
 //------------------------------------------------------------------------------
 
-Param::Param(World& world, const Type* type, Continuation* continuation, size_t index, Debug dbg)
-    : Def(Node_Param, world, type, 1, dbg)
+Param::Param(World& world, const Type* type, const Continuation* continuation, size_t index, Debug dbg)
+    : Def(world, Node_Param, type, { continuation }, dbg)
     , index_(index)
-{
-    set_op(0, continuation);
+{}
+
+const Def* Param::rebuild(World& world, const Type* t, Defs defs) const {
+    assert(defs.size() == 1 && defs[0]->isa<Continuation>());
+    return world.param(t, defs[0]->as<Continuation>(), index(), debug());
+}
+
+hash_t Param::vhash() const {
+    return hash_combine(Def::vhash(), (hash_t) index());
+}
+
+bool Param::equal(const Def* other) const {
+    return Def::equal(other) && other->as<Param>()->index() == index();
 }
 
 //------------------------------------------------------------------------------
 
-App::App(World& world, const Defs ops, Debug dbg) : Def(Node_App, world, ops[0]->world().bottom_type(), ops, dbg) {
+App::App(World& world, const Defs ops, Debug dbg) : Def(world, Node_App, ops[0]->world().bottom_type(), ops, dbg) {
 #if THORIN_ENABLE_CHECKS
     verify();
     if (auto cont = callee()->isa_nom<Continuation>())
@@ -41,7 +52,7 @@ void App::verify() const {
 
 //------------------------------------------------------------------------------
 
-Filter::Filter(World& world, const Defs defs, Debug dbg) : Def(Node_Filter, world, world.bottom_type(), defs, dbg) {}
+Filter::Filter(World& world, const Defs defs, Debug dbg) : Def(world, Node_Filter, world.bottom_type(), defs, dbg) {}
 
 const Filter* Filter::cut(ArrayRef<size_t> indices) const {
     return world().filter(ops().cut(indices), debug());
@@ -49,16 +60,23 @@ const Filter* Filter::cut(ArrayRef<size_t> indices) const {
 
 //------------------------------------------------------------------------------
 
-Continuation::Continuation(World& w, const FnType* fn, const Attributes& attributes, Debug dbg)
-    : Def(Node_Continuation, w, fn, 2, dbg)
+Continuation::Continuation(World& w, const FnType* pi, const Attributes& attributes, Debug dbg)
+    : Def(w, Node_Continuation, pi, 2, dbg)
     , attributes_(attributes)
 {
-    params_.reserve(fn->num_ops());
+    params_.reserve(pi->num_ops());
     set_op(0, world().bottom(world().bottom_type()));
     set_op(1, world().filter({}, dbg));
+
+    size_t i = 0;
+    for (auto op : pi->types()) {
+        auto p = w.param(op, this, i++, dbg);
+        params_.emplace_back(p);
+    }
 }
 
-Continuation* Continuation::stub() const {
+// TODO: merge with regular stub()
+Continuation* Continuation::mangle_stub() const {
     Rewriter rewriter;
 
     auto result = world().continuation(type(), attributes(), debug_history());
@@ -78,6 +96,40 @@ Continuation* Continuation::stub() const {
     return result;
 }
 
+Continuation* Continuation::stub(World& nworld, const Type* t) const {
+    assert(!dead_);
+    // TODO maybe we want to deal with intrinsics in a more streamlined way
+    if (this == world().branch())
+        return nworld.branch();
+    if (this == world().end_scope())
+        return nworld.end_scope();
+
+    auto npi = t->isa<FnType>();
+    assert(npi);
+    Continuation* ncontinuation = nworld.continuation(npi, attributes(), debug_history());
+    assert(&ncontinuation->world() == &nworld);
+    assert(&npi->world() == &nworld);
+    for (size_t i = 0, e = num_params(); i != e; ++i)
+        ncontinuation->param(i)->set_name(param(i)->debug_history().name);
+
+    if (is_external())
+        nworld.make_external(ncontinuation);
+    return ncontinuation;
+}
+
+void Continuation::rebuild_from(const Def*, Defs nops) {
+    if (this == world().branch())
+        return;
+    if (this == world().end_scope())
+        return;
+
+    auto napp = nops[0]->isa<App>();
+    if (napp)
+        set_body(napp);
+    set_filter(nops[1]->as<Filter>());
+    verify();
+}
+
 Array<const Def*> Continuation::params_as_defs() const {
     Array<const Def*> params(num_params());
     for (size_t i = 0, e = num_params(); i != e; ++i)
@@ -126,9 +178,9 @@ const FnType* Continuation::arg_fn_type() const {
 const Param* Continuation::append_param(const Type* param_type, Debug dbg) {
     size_t size = type()->num_ops();
     Array<const Type*> ops(size + 1);
-    *std::copy(type()->ops().begin(), type()->ops().end(), ops.begin()) = param_type;
+    *std::copy(type()->types().begin(), type()->types().end(), ops.begin()) = param_type;
     clear_type();
-    set_type(param_type->table().fn_type(ops));              // update type
+    set_type(world().fn_type(ops));              // update type
     auto param = world().param(param_type, this, size, dbg); // append new param
     params_.push_back(param);
 
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index f88e24885..32bbecd69 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -24,12 +24,15 @@ typedef std::vector<Continuation*> Continuations;
  */
 class Param : public Def {
 private:
-    Param(World&, const Type* type, Continuation* continuation, size_t index, Debug dbg);
+    Param(World&, const Type* type, const Continuation* continuation, size_t index, Debug dbg);
 
 public:
     Continuation* continuation() const { return op(0)->as_nom<Continuation>(); }
     size_t index() const { return index_; }
 
+    const Def * rebuild(World&, const Type*, Defs) const override;
+    bool equal(const Def*) const override;
+    hash_t vhash() const override;
 private:
     const size_t index_;
 
@@ -130,13 +133,15 @@ class Continuation : public Def {
     };
 
 private:
-    Continuation(World&, const FnType* fn, const Attributes& attributes, Debug dbg);
+    Continuation(World&, const FnType* pi, const Attributes& attributes, Debug dbg);
     virtual ~Continuation() { for (auto param : params()) delete param; }
 
 public:
     const FnType* type() const { return Def::type()->as<FnType>(); }
 
-    Continuation* stub() const;
+    Continuation* mangle_stub() const;
+    Continuation* stub(World&, const Type*) const override;
+    void rebuild_from(const Def* old, Defs new_ops) override;
     const Param* append_param(const Type* type, Debug dbg = {});
     Continuations preds() const;
     Continuations succs() const;
diff --git a/src/thorin/def.cpp b/src/thorin/def.cpp
index a0d14e59c..b3b84d731 100644
--- a/src/thorin/def.cpp
+++ b/src/thorin/def.cpp
@@ -14,7 +14,7 @@ namespace thorin {
 
 size_t Def::gid_counter_ = 1;
 
-Def::Def(NodeTag tag, World& world, const Type* type, Defs ops, Debug dbg)
+Def::Def(World& world, NodeTag tag, const Type* type, Defs ops, Debug dbg)
     : tag_(tag)
     , ops_(ops.size())
     , world_(world)
@@ -30,7 +30,7 @@ Def::Def(NodeTag tag, World& world, const Type* type, Defs ops, Debug dbg)
         set_op(i, ops[i]);
 }
 
-Def::Def(NodeTag tag, World& world, const Type* type, size_t size, Debug dbg)
+Def::Def(World& world, NodeTag tag, const Type* type, size_t size, Debug dbg)
     : tag_(tag)
     , ops_(size)
     , world_(world)
@@ -43,6 +43,10 @@ Def::Def(NodeTag tag, World& world, const Type* type, size_t size, Debug dbg)
                                       Dep::Bot   )
 {}
 
+int Def::order() const {
+    return type()->order();
+}
+
 Debug Def::debug_history() const {
 #if THORIN_ENABLE_CHECKS
     return world().track_history() ? Debug(unique_name(), loc()) : debug();
@@ -96,7 +100,7 @@ std::string Def::unique_name() const {
 }
 
 bool is_unit(const Def* def) {
-    return def->type() == def->world().unit();
+    return def->type() == def->world().unit_type();
 }
 
 size_t vector_length(const Def* def) { return def->type()->as<VectorType>()->length(); }
@@ -134,6 +138,12 @@ bool is_minus_zero(const Def* def) {
     return false;
 }
 
+void Def::rebuild_from(const Def* old, Defs new_ops) {
+    assert(new_ops.size() == num_ops());
+    for (size_t i = 0; i < num_ops(); i++)
+        set_op(i, new_ops[i]);
+}
+
 void Def::replace_uses(const Def* with) const {
     world().DLOG("replace uses: {} -> {}", this, with);
     if (this != with) {
@@ -149,8 +159,6 @@ void Def::replace_uses(const Def* with) const {
     }
 }
 
-World& Def::world() const { return world_; }
-
 uint64_t UseHash::hash(Use use) {
     assert(use->gid() != uint32_t(-1));
     hash_t seed = hash_begin(use.index());
diff --git a/src/thorin/def.h b/src/thorin/def.h
index 18118e9d4..b3254d704 100644
--- a/src/thorin/def.h
+++ b/src/thorin/def.h
@@ -5,8 +5,8 @@
 #include <vector>
 
 #include "thorin/enums.h"
-#include "thorin/type.h"
 #include "thorin/debug.h"
+#include "thorin/util/array.h"
 
 namespace thorin {
 
@@ -17,6 +17,7 @@ class Def;
 class Tracker;
 class Use;
 class World;
+class Type;
 
 typedef ArrayRef<const Def*> Defs;
 
@@ -76,6 +77,23 @@ struct UseHash {
 // using a StackCapacity of 8 covers almost 99% of all real-world use-lists
 typedef HashSet<Use, UseHash> Uses;
 
+template<class T>
+struct GIDLt {
+    bool operator()(T a, T b) const { return a->gid() < b->gid(); }
+};
+
+template<class T>
+struct GIDHash {
+    static hash_t hash(T n) { return thorin::murmur3(n->gid()); }
+    static bool eq(T a, T b) { return a == b; }
+    static T sentinel() { return T(1); }
+};
+
+template<class Key, class Value>
+using GIDMap = thorin::HashMap<Key, Value, GIDHash<Key>>;
+template<class Key>
+using GIDSet = thorin::HashSet<Key, GIDHash<Key>>;
+
 template<class To>
 using DefMap  = GIDMap<const Def*, To>;
 using DefSet  = GIDSet<const Def*>;
@@ -106,9 +124,9 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
 
 protected:
     /// Constructor for a @em structural Def.
-    Def(NodeTag tag, World&, const Type* type, Defs args, Debug dbg);
+    Def(World&, NodeTag tag, const Type* type, Defs args, Debug dbg);
     /// Constructor for a @em nom Def.
-    Def(NodeTag tag, World&, const Type* type, size_t size, Debug);
+    Def(World&, NodeTag tag, const Type* type, size_t size, Debug);
     virtual ~Def() {}
 
     void clear_type() { type_ = nullptr; }
@@ -122,7 +140,7 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
     //@{
     NodeTag tag() const { return tag_; }
     size_t gid() const { return gid_; }
-    World& world() const;
+    World& world() const { return world_; };
     //@}
 
     /// @name ops
@@ -155,7 +173,8 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
     /// @name type
     //@{
     const Type* type() const { return type_; }
-    int order() const { return type()->order(); }
+
+    virtual int order() const;
     //@}
 
     /// @name dependence checks
@@ -202,7 +221,8 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
     /// @name rebuild/stub
     //@{
     virtual const Def* rebuild(World&, const Type*, Defs) const { THORIN_UNREACHABLE; }
-    // TODO stub
+    virtual       Def* stub(World&, const Type*) const { THORIN_UNREACHABLE; }
+    virtual       void rebuild_from(const Def* old, Defs new_ops);
     //@}
 
     void replace_uses(const Def*) const;
@@ -251,7 +271,6 @@ size_t vector_length(const Def*);
 bool is_unit(const Def*);
 bool is_primlit(const Def*, int64_t);
 bool is_minus_zero(const Def*);
-inline bool is_mem        (const Def* def) { return def->type()->isa<MemType>(); }
 inline bool is_zero       (const Def* def) { return is_primlit(def, 0); }
 inline bool is_one        (const Def* def) { return is_primlit(def, 1); }
 inline bool is_allset     (const Def* def) { return is_primlit(def, -1); }
diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index 9a06a93a1..23d5ba859 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -15,16 +15,16 @@ namespace thorin {
  */
 
 PrimLit::PrimLit(World& world, PrimTypeTag tag, Box box, Debug dbg)
-    : Literal((NodeTag) tag, world, world.prim_type(tag), dbg)
+    : Literal(world, (NodeTag) tag, world.prim_type(tag), dbg)
     , box_(box)
 {}
 
 Cmp::Cmp(CmpTag tag, World& world, const Def* lhs, const Def* rhs, Debug dbg)
-    : BinOp((NodeTag) tag, world, world.type_bool(vector_length(lhs->type())), lhs, rhs, dbg)
+    : BinOp(world, (NodeTag) tag, world.type_bool(vector_length(lhs->type())), lhs, rhs, dbg)
 {}
 
 DefiniteArray::DefiniteArray(World& world, const Type* elem, Defs args, Debug dbg)
-    : Aggregate(Node_DefiniteArray, world, args, dbg)
+    : Aggregate(world, Node_DefiniteArray, args, dbg)
 {
     set_type(world.definite_array_type(elem, args.size()));
 #if THORIN_ENABLE_CHECKS
@@ -34,13 +34,13 @@ DefiniteArray::DefiniteArray(World& world, const Type* elem, Defs args, Debug db
 }
 
 IndefiniteArray::IndefiniteArray(World& world, const Type* elem, const Def* dim, Debug dbg)
-    : Aggregate(Node_IndefiniteArray, world, {dim}, dbg)
+    : Aggregate(world, Node_IndefiniteArray, {dim}, dbg)
 {
     set_type(world.indefinite_array_type(elem));
 }
 
 Tuple::Tuple(World& world, Defs args, Debug dbg)
-    : Aggregate(Node_Tuple, world, args, dbg)
+    : Aggregate(world, Node_Tuple, args, dbg)
 {
     Array<const Type*> elems(num_ops());
     for (size_t i = 0, e = num_ops(); i != e; ++i)
@@ -50,7 +50,7 @@ Tuple::Tuple(World& world, Defs args, Debug dbg)
 }
 
 Vector::Vector(World& world, Defs args, Debug dbg)
-    : Aggregate(Node_Vector, world, args, dbg)
+    : Aggregate(world, Node_Vector, args, dbg)
 {
     if (auto primtype = args.front()->type()->isa<PrimType>()) {
         assert(primtype->length() == 1);
@@ -63,15 +63,15 @@ Vector::Vector(World& world, Defs args, Debug dbg)
 }
 
 LEA::LEA(World& world, const Def* ptr, const Def* index, Debug dbg)
-    : Def(Node_LEA, world, nullptr, {ptr, index}, dbg)
+    : Def(world, Node_LEA, nullptr, {ptr, index}, dbg)
 {
     auto type = ptr_type();
     if (auto tuple = ptr_pointee()->isa<TupleType>()) {
-        set_type(world.ptr_type(get(tuple->ops(), index), type->length(), type->device(), type->addr_space()));
+        set_type(world.ptr_type(get(tuple->types(), index), type->length(), type->device(), type->addr_space()));
     } else if (auto array = ptr_pointee()->isa<ArrayType>()) {
         set_type(world.ptr_type(array->elem_type(), type->length(), type->device(), type->addr_space()));
     } else if (auto struct_type = ptr_pointee()->isa<StructType>()) {
-        set_type(world.ptr_type(get(struct_type->ops(), index), type->length(), type->device(), type->addr_space()));
+        set_type(world.ptr_type(get(struct_type->types(), index), type->length(), type->device(), type->addr_space()));
     } else if (auto prim_type = ptr_pointee()->isa<PrimType>()) {
         assert(prim_type->length() > 1);
         set_type(world.ptr_type(world.prim_type(prim_type->primtype_tag()), type->length(), type->device(), type->addr_space()));
@@ -81,50 +81,50 @@ LEA::LEA(World& world, const Def* ptr, const Def* index, Debug dbg)
 }
 
 Known::Known(World& world, const Def* def, Debug dbg)
-    : Def(Node_Known, world, world.type_bool(), {def}, dbg)
+    : Def(world, Node_Known, world.type_bool(), {def}, dbg)
 {}
 
 AlignOf::AlignOf(World& world, const Def* def, Debug dbg)
-    : Def(Node_AlignOf, world, world.type_qs64(), {def}, dbg)
+    : Def(world, Node_AlignOf, world.type_qs64(), {def}, dbg)
 {}
 
 SizeOf::SizeOf(World& world, const Def* def, Debug dbg)
-    : Def(Node_SizeOf, world, world.type_qs64(), {def}, dbg)
+    : Def(world, Node_SizeOf, world.type_qs64(), {def}, dbg)
 {}
 
 Slot::Slot(World& world, const Type* type, const Def* frame, Debug dbg)
-    : Def(Node_Slot, world, type->table().ptr_type(type), {frame}, dbg)
+    : Def(world, Node_Slot, world.ptr_type(type), {frame}, dbg)
 {
     assert(frame->type()->isa<FrameType>());
 }
 
 Global::Global(World& world, const Def* init, bool is_mutable, Debug dbg)
-    : Def(Node_Global, world, init->type()->table().ptr_type(init->type()), {init}, dbg)
+    : Def(world, Node_Global, world.ptr_type(init->type()), {init}, dbg)
     , is_mutable_(is_mutable)
 {
     assert(!init->has_dep(Dep::Param));
 }
 
 Alloc::Alloc(World& world, const Type* type, const Def* mem, const Def* extra, Debug dbg)
-    : MemOp(Node_Alloc, world, nullptr, {mem, extra}, dbg)
+    : MemOp(world, Node_Alloc, nullptr, {mem, extra}, dbg)
 {
     set_type(world.tuple_type({world.mem_type(), world.ptr_type(type)}));
 }
 
 Load::Load(World& world, const Def* mem, const Def* ptr, Debug dbg)
-    : Access(Node_Load, world, nullptr, {mem, ptr}, dbg)
+    : Access(world, Node_Load, nullptr, {mem, ptr}, dbg)
 {
     set_type(world.tuple_type({world.mem_type(), ptr->type()->as<PtrType>()->pointee()}));
 }
 
 Enter::Enter(World& world, const Def* mem, Debug dbg)
-    : MemOp(Node_Enter, world, nullptr, {mem}, dbg)
+    : MemOp(world, Node_Enter, nullptr, {mem}, dbg)
 {
     set_type(world.tuple_type({world.mem_type(), world.frame_type()}));
 }
 
 Assembly::Assembly(World& world, const Type *type, Defs inputs, std::string asm_template, ArrayRef<std::string> output_constraints, ArrayRef<std::string> input_constraints, ArrayRef<std::string> clobbers, Flags flags, Debug dbg)
-    : MemOp(Node_Assembly, world, type, inputs, dbg)
+    : MemOp(world, Node_Assembly, type, inputs, dbg)
     , asm_template_(asm_template)
     , output_constraints_(output_constraints)
     , input_constraints_(input_constraints)
@@ -301,13 +301,13 @@ const Def* Def::out(size_t i) const {
 
 const Type* Extract::extracted_type(const Def* agg, const Def* index) {
     if (auto tuple = agg->type()->isa<TupleType>())
-        return get(tuple->ops(), index);
+        return get(tuple->types(), index);
     else if (auto array = agg->type()->isa<ArrayType>())
         return array->elem_type();
     else if (auto vector = agg->type()->isa<VectorType>())
         return vector->scalarize();
     else if (auto struct_type = agg->type()->isa<StructType>())
-        return get(struct_type->ops(), index);
+        return get(struct_type->types(), index);
 
     THORIN_UNREACHABLE;
 }
diff --git a/src/thorin/primop.h b/src/thorin/primop.h
index db2aaf175..e82d2c218 100644
--- a/src/thorin/primop.h
+++ b/src/thorin/primop.h
@@ -3,6 +3,7 @@
 
 #include "thorin/config.h"
 #include "thorin/def.h"
+#include "thorin/type.h"
 #include "thorin/enums.h"
 #include "thorin/util/hash.h"
 
@@ -10,8 +11,8 @@ namespace thorin {
 
 class Literal : public Def {
 protected:
-    Literal(NodeTag tag, World& world, const Type* type, Debug dbg)
-        : Def(tag, world, type, Defs{}, dbg)
+    Literal(World& world, NodeTag tag, const Type* type, Debug dbg)
+        : Def(world, tag, type, Defs{}, dbg)
     {}
 };
 
@@ -19,7 +20,7 @@ class Literal : public Def {
 class Bottom : public Literal {
 private:
     Bottom(World& world, const Type* type, Debug dbg)
-        : Literal(Node_Bottom, world, type, dbg)
+        : Literal(world, Node_Bottom, type, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -31,7 +32,7 @@ class Bottom : public Literal {
 class Top : public Literal {
 private:
     Top(World& world, const Type* type, Debug dbg)
-        : Literal(Node_Top, world, type, dbg)
+        : Literal(world, Node_Top, type, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -80,7 +81,7 @@ T get(ArrayRef<T> array, const Def* def) { return array[primlit_value<size_t>(de
 class Select : public Def {
 private:
     Select(World& world, const Def* cond, const Def* tval, const Def* fval, Debug dbg)
-        : Def(Node_Select, world, tval->type(), {cond, tval, fval}, dbg)
+        : Def(world, Node_Select, tval->type(), {cond, tval, fval}, dbg)
     {
         assert(is_type_bool(cond->type()));
         assert(tval->type() == fval->type() && "types of both values must be equal");
@@ -126,8 +127,8 @@ class SizeOf : public Def {
 /// Base class for all side-effect free binary \p Def%s.
 class BinOp : public Def {
 protected:
-    BinOp(NodeTag tag, World& world, const Type* type, const Def* lhs, const Def* rhs, Debug dbg)
-        : Def(tag, world, type, {lhs, rhs}, dbg)
+    BinOp(World& world, NodeTag tag, const Type* type, const Def* lhs, const Def* rhs, Debug dbg)
+        : Def(world, tag, type, {lhs, rhs}, dbg)
     {
         assert(lhs->type() == rhs->type() && "types are not equal");
     }
@@ -141,7 +142,7 @@ class BinOp : public Def {
 class ArithOp : public BinOp {
 private:
     ArithOp(ArithOpTag tag, World& world, const Def* lhs, const Def* rhs, Debug dbg)
-        : BinOp((NodeTag) tag, world, lhs->type(), lhs, rhs, dbg)
+        : BinOp(world, (NodeTag) tag, lhs->type(), lhs, rhs, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -172,8 +173,8 @@ class Cmp : public BinOp {
 /// Common mathematical function such as `sin()` or `cos()`.
 class MathOp : public Def {
 private:
-    MathOp(MathOpTag tag, World& world, const Type* type, Defs args, Debug dbg)
-        : Def((NodeTag)tag, world, type, args, dbg)
+    MathOp(World& world, MathOpTag tag, const Type* type, Defs args, Debug dbg)
+        : Def(world, (NodeTag)tag, type, args, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -189,8 +190,8 @@ class MathOp : public Def {
 /// Base class for @p Bitcast and @p Cast.
 class ConvOp : public Def {
 protected:
-    ConvOp(NodeTag tag, World& world, const Def* from, const Type* to, Debug dbg)
-        : Def(tag, world, to, {from}, dbg)
+    ConvOp(World& world, NodeTag tag, const Def* from, const Type* to, Debug dbg)
+        : Def(world, tag, to, {from}, dbg)
     {}
 
 public:
@@ -201,7 +202,7 @@ class ConvOp : public Def {
 class Cast : public ConvOp {
 private:
     Cast(World& world, const Type* to, const Def* from, Debug dbg)
-        : ConvOp(Node_Cast, world, from, to, dbg)
+        : ConvOp(world, Node_Cast, from, to, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -213,7 +214,7 @@ class Cast : public ConvOp {
 class Bitcast : public ConvOp {
 private:
     Bitcast(World& world, const Type* to, const Def* from, Debug dbg)
-        : ConvOp(Node_Bitcast, world, from, to, dbg)
+        : ConvOp(world, Node_Bitcast, from, to, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -224,8 +225,8 @@ class Bitcast : public ConvOp {
 /// Base class for all aggregate data constructers.
 class Aggregate : public Def {
 protected:
-    Aggregate(NodeTag tag, World& world, Defs args, Debug dbg)
-        : Def(tag, world, nullptr /*set later*/, args, dbg)
+    Aggregate(World& world, NodeTag tag, Defs args, Debug dbg)
+        : Def(world, tag, nullptr /*set later*/, args, dbg)
     {}
 };
 
@@ -275,7 +276,7 @@ class Tuple : public Aggregate {
 class Variant : public Def {
 private:
     Variant(World& world, const VariantType* variant_type, const Def* value, size_t index, Debug dbg)
-        : Def(Node_Variant, world, variant_type, {value}, dbg), index_(index)
+        : Def(world, Node_Variant, variant_type, {value}, dbg), index_(index)
     {
         assert(variant_type->op(index) == value->type());
     }
@@ -298,7 +299,7 @@ class Variant : public Def {
 class VariantIndex : public Def {
 private:
     VariantIndex(World& world, const Type* int_type, const Def* value, Debug dbg)
-        : Def(Node_VariantIndex, world, int_type, {value}, dbg)
+        : Def(world, Node_VariantIndex, int_type, {value}, dbg)
     {
         assert(value->type()->isa<VariantType>());
         assert(is_type_s(int_type) || is_type_u(int_type));
@@ -312,7 +313,7 @@ class VariantIndex : public Def {
 class VariantExtract : public Def {
 private:
     VariantExtract(World& world, const Type* type, const Def* value, size_t index, Debug dbg)
-        : Def(Node_VariantExtract, world, type, {value}, dbg), index_(index)
+        : Def(world, Node_VariantExtract, type, {value}, dbg), index_(index)
     {
         assert(value->type()->as<VariantType>()->op(index) == type);
     }
@@ -334,7 +335,7 @@ class VariantExtract : public Def {
 class Closure : public Aggregate {
 private:
     Closure(World& world, const ClosureType* closure_type, const Def* fn, const Def* env, Debug dbg)
-        : Aggregate(Node_Closure, world, {fn, env}, dbg)
+        : Aggregate(world, Node_Closure, {fn, env}, dbg)
     {
         set_type(closure_type);
     }
@@ -352,7 +353,7 @@ class Closure : public Aggregate {
 class StructAgg : public Aggregate {
 private:
     StructAgg(World& world, const StructType* struct_type, Defs args, Debug dbg)
-        : Aggregate(Node_StructAgg, world, args, dbg)
+        : Aggregate(world, Node_StructAgg, args, dbg)
     {
 #if THORIN_ENABLE_CHECKS
         assert(struct_type->num_ops() == args.size());
@@ -383,8 +384,8 @@ class Vector : public Aggregate {
 /// Base class for functional @p Insert and @p Extract.
 class AggOp : public Def {
 protected:
-    AggOp(NodeTag tag, World& world, const Type* type, Defs args, Debug dbg)
-        : Def(tag, world, type, args, dbg)
+    AggOp(World& world, NodeTag tag, const Type* type, Defs args, Debug dbg)
+        : Def(world, tag, type, args, dbg)
     {}
 
 public:
@@ -398,7 +399,7 @@ class AggOp : public Def {
 class Extract : public AggOp {
 private:
     Extract(World& world, const Def* agg, const Def* index, Debug dbg)
-        : AggOp(Node_Extract, world, extracted_type(agg, index), {agg, index}, dbg)
+        : AggOp(world, Node_Extract, extracted_type(agg, index), {agg, index}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -418,7 +419,7 @@ class Extract : public AggOp {
 class Insert : public AggOp {
 private:
     Insert(World& world, const Def* agg, const Def* index, const Def* value, Debug dbg)
-        : AggOp(Node_Insert, world, agg->type(), {agg, index, value}, dbg)
+        : AggOp(world, Node_Insert, agg->type(), {agg, index, value}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -455,7 +456,7 @@ class LEA : public Def {
 class Hlt : public Def {
 private:
     Hlt(World& world, const Def* def, Debug dbg)
-        : Def(Node_Hlt, world, def->type(), {def}, dbg)
+        : Def(world, Node_Hlt, def->type(), {def}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -486,7 +487,7 @@ class Known : public Def {
 class Run : public Def {
 private:
     Run(World& world, const Def* def, Debug dbg)
-        : Def(Node_Run, world, def->type(), {def}, dbg)
+        : Def(world, Node_Run, def->type(), {def}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -547,8 +548,8 @@ class Global : public Def {
 /// Base class for all \p Def%s taking and producing side-effects.
 class MemOp : public Def {
 protected:
-    MemOp(NodeTag tag, World& world, const Type* type, Defs args, Debug dbg)
-        : Def(tag, world, type, args, dbg)
+    MemOp(World& world, NodeTag tag, const Type* type, Defs args, Debug dbg)
+        : Def(world, tag, type, args, dbg)
     {
         assert(mem()->type()->isa<MemType>());
         assert(args.size() >= 1);
@@ -585,8 +586,8 @@ class Alloc : public MemOp {
 /// Base class for @p Load and @p Store.
 class Access : public MemOp {
 protected:
-    Access(NodeTag tag, World& world, const Type* type, Defs args, Debug dbg)
-        : MemOp(tag, world, type, args, dbg)
+    Access(World& world, NodeTag tag, const Type* type, Defs args, Debug dbg)
+        : MemOp(world, tag, type, args, dbg)
     {
         assert(args.size() >= 2);
     }
@@ -604,7 +605,7 @@ class Load : public Access {
     bool has_multiple_outs() const override { return true; }
     const Def* out_val() const { return out(1); }
     const TupleType* type() const { return MemOp::type()->as<TupleType>(); }
-    const Type* out_val_type() const { return type()->op(1); }
+    const Type* out_val_type() const { return type()->op(1)->as<Type>(); }
 
 private:
     const Def* rebuild(World&, const Type*, Defs) const override;
@@ -616,7 +617,7 @@ class Load : public Access {
 class Store : public Access {
 private:
     Store(World& world, const Def* mem, const Def* ptr, const Def* value, Debug dbg)
-        : Access(Node_Store, world, mem->type(), {mem, ptr, value}, dbg)
+        : Access(world, Node_Store, mem->type(), {mem, ptr, value}, dbg)
     {}
 
     const Def* rebuild(World&, const Type*, Defs) const override;
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 03d6a8c20..4980cd91d 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -68,8 +68,6 @@ void RecStreamer::run() {
 void Def::dump() const { dump(0); }
 void Def::dump(size_t max) const { Stream s(std::cout); stream(s, max).endl(); }
 
-void Type::dump() const { Stream s(std::cout); stream(s).endl(); }
-
 Stream& Def::stream(Stream& s) const {
     if (isa<Param>() || isa<App>() || no_dep()) return stream1(s);
     return s << unique_name();
diff --git a/src/thorin/tables/nodetable.h b/src/thorin/tables/nodetable.h
index a302d270f..8c17d04dc 100644
--- a/src/thorin/tables/nodetable.h
+++ b/src/thorin/tables/nodetable.h
@@ -47,6 +47,7 @@
     THORIN_NODE(Filter, filter)
     // Type
         // PrimType
+        THORIN_NODE(Star, star)
         THORIN_NODE(App, app)
         THORIN_NODE(DefiniteArrayType, definite_array_type)
         THORIN_NODE(FnType, fn)
diff --git a/src/thorin/transform/cleanup_world.cpp b/src/thorin/transform/cleanup_world.cpp
index 041db668a..3ea2f26e2 100644
--- a/src/thorin/transform/cleanup_world.cpp
+++ b/src/thorin/transform/cleanup_world.cpp
@@ -46,7 +46,7 @@ void Cleaner::eliminate_tail_rec() {
                 if (use.index() == 0 && use->isa<App>()) {
                     recursive = true;
                     continue;
-                } else if (use->isa_nom<Param>())
+                } else if (use->isa<Param>())
                     continue; // ignore params
 
                 world().ELOG("non-recursive usage of {} index:{} use:{}", scope.entry()->name(), use.index(), use.def()->to_string());
@@ -202,7 +202,7 @@ void Cleaner::eliminate_params() {
 
             if (!proxy_idx.empty()) {
                 auto ncontinuation = world().continuation(
-                    world().fn_type(ocontinuation->type()->ops().cut(proxy_idx)),
+                    world().fn_type(ocontinuation->type()->types().cut(proxy_idx)),
                     ocontinuation->attributes(), ocontinuation->debug_history());
                 size_t j = 0;
                 for (auto i : param_idx) {
@@ -234,7 +234,6 @@ next_continuation:;
 void Cleaner::rebuild() {
     auto fresh_world = std::make_unique<World>(world());
     Importer importer(*world_, *fresh_world);
-    importer.type_old2new_.rehash(world_->types().capacity());
     importer.def_old2new_.rehash(world_->defs().capacity());
 
     for (auto&& [_, cont] : world().externals()) {
@@ -269,7 +268,7 @@ void Cleaner::verify_closedness() {
 
 void Cleaner::within(const Def* def) {
     if (def->isa<Param>()) return; // TODO remove once Params are within World's sea of nodes
-    assert(world().types().contains(def->type()));
+    assert(&def->type()->world() == &world());
     assert_unused(world().defs().contains(def));
 }
 
diff --git a/src/thorin/transform/closure_conversion.cpp b/src/thorin/transform/closure_conversion.cpp
index bb118c725..d50d6d1ff 100644
--- a/src/thorin/transform/closure_conversion.cpp
+++ b/src/thorin/transform/closure_conversion.cpp
@@ -24,7 +24,7 @@ class ClosureConversion {
                 continue;
             }
 
-            auto new_type = world_.fn_type(convert(continuation->type())->ops());
+            auto new_type = world_.fn_type(defs2types(convert_type(continuation->type())->ops()));
             if (new_type != continuation->type()) {
                 auto new_continuation = world_.continuation(new_type->as<FnType>(), continuation->debug());
                 if (continuation->is_intrinsic())
@@ -66,12 +66,15 @@ class ClosureConversion {
         if (!callee || !callee->is_intrinsic()) {
             Array<const Def*> new_args(body->num_args());
             for (size_t i = 0, e = body->num_args(); i != e; ++i)
-                new_args[i] = convert(body->arg(i));
-            continuation->jump(convert(body->callee(), true), new_args, continuation->debug());
+                new_args[i] = convert_def(body->arg(i));
+            continuation->jump(convert_def(body->callee(), true), new_args, continuation->debug());
         }
     }
 
-    const Def* convert(const Def* def, bool as_callee = false) {
+    const Def* convert_def(const Def* def, bool as_callee = false) {
+        if (auto t = def->isa<Type>())
+            return convert_type(t);
+
         if (new_defs_.count(def)) def = new_defs_[def];
         if (def->order() <= 1)
             return def;
@@ -145,13 +148,13 @@ class ClosureConversion {
             }
             wrapper->jump(lifted, wrapper_args);
 
-            auto closure_type = convert(continuation->type());
+            auto closure_type = convert_type(continuation->type());
             return world_.closure(closure_type->as<ClosureType>(), wrapper, thin_env ? free_vars[0] : world_.tuple(free_vars), continuation->debug());
         } else {
             // TODO need to consider Params?
             Array<const Def*> ops(def->ops());
-            for (auto& op : ops) op = convert(op);
-            return new_defs_[def] = def->rebuild(world_, convert(def->type()), ops);
+            for (auto& op : ops) op = convert_def(op);
+            return new_defs_[def] = def->rebuild(world_, convert_type(def->type()), ops);
         }
         THORIN_UNREACHABLE;
     }
@@ -160,10 +163,10 @@ class ClosureConversion {
     // - fn (A, B, fn(C), fn(D)) => closure(fn (A, B, fn(convert(C)), closure(fn(convert(D)))))
     // - struct S { fn (X, fn(Y)) } => struct T { closure(fn (X, fn(Y))) }
     // - ...
-    const Type* convert(const Type* type) {
+    const Type* convert_type(const Type* type) {
         if (new_types_.count(type)) return new_types_[type];
         if (type->order() <= 1) return type;
-        Array<const Type*> ops(type->ops());
+        Array<const Def*> ops(type->ops());
 
         const Type* new_type = nullptr;
         if (type->isa<StructType>()) {
@@ -175,32 +178,32 @@ class ClosureConversion {
         // accept one parameter of order 1 (the return continuation) for function types
         bool ret = !type->isa<FnType>();
         for (auto& op : ops) {
-            op = convert(op);
+            op = convert_def(op);
             if (!ret &&
                 op->isa<ClosureType>() &&
                 op->as<ClosureType>()->inner_order() == 1) {
                 ret = true;
-                op = world_.fn_type(op->ops());
+                op = world_.fn_type(defs2types(op->ops()));
             }
         }
 
         if (type->isa<StructType>()) {
-            auto struct_type = new_type->as<StructType>();
+            StructType* struct_type = const_cast<StructType*>(new_type->as<StructType>());
             for (size_t i = 0, e = ops.size(); i != e; ++i)
-                struct_type->set(i, ops[i]);
+                struct_type->set_op(i, ops[i]);
         } else {
-            new_type = type->rebuild(type->table(), ops);
+            new_type = type->rebuild(world_, type->type(), ops)->as<Type>();
         }
         if (new_type->order() <= 1)
             return new_types_[type] = new_type;
         else
-            return new_types_[type] = world_.closure_type(new_type->ops());
+            return new_types_[type] = world_.closure_type(defs2types(new_type->ops()));
     }
 
 private:
     World& world_;
     Def2Def new_defs_;
-    Type2Type new_types_;
+    DefMap<const Type*> new_types_;
 };
 
 
diff --git a/src/thorin/transform/flatten_tuples.cpp b/src/thorin/transform/flatten_tuples.cpp
index 8f99bd220..8187b1001 100644
--- a/src/thorin/transform/flatten_tuples.cpp
+++ b/src/thorin/transform/flatten_tuples.cpp
@@ -13,10 +13,10 @@ static Continuation* unwrap_def(Def2Def&, Def2Def&, const Def*, const FnType*, s
 // Computes the type of the wrapped function
 static const Type* wrapped_type(const FnType* fn_type, size_t max_tuple_size) {
     std::vector<const Type*> nops;
-    for (auto op : fn_type->ops()) {
+    for (auto op : fn_type->types()) {
         if (auto tuple_type = op->isa<TupleType>()) {
             if (tuple_type->num_ops() <= max_tuple_size) {
-                for (auto arg : tuple_type->ops())
+                for (auto arg : tuple_type->types())
                     nops.push_back(arg);
             } else
                 nops.push_back(op);
@@ -26,7 +26,7 @@ static const Type* wrapped_type(const FnType* fn_type, size_t max_tuple_size) {
             nops.push_back(op);
         }
     }
-    return fn_type->table().fn_type(nops);
+    return fn_type->world().fn_type(nops);
 }
 
 static Continuation* jump(Continuation* cont, Array<const Def*>& args) {
diff --git a/src/thorin/transform/hls_channels.cpp b/src/thorin/transform/hls_channels.cpp
index c778baf8f..bae0f5291 100644
--- a/src/thorin/transform/hls_channels.cpp
+++ b/src/thorin/transform/hls_channels.cpp
@@ -166,8 +166,8 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
             extract_kernel_channels(schedule(scope), def2mode);
 
             Array<const Type*> new_param_types(def2mode.size() + old_kernel->num_params());
-            std::copy(old_kernel->type()->ops().begin(),
-                    old_kernel->type()->ops().end(),
+            std::copy(old_kernel->type()->types().begin(),
+                    old_kernel->type()->types().end(),
                     new_param_types.begin());
             size_t i = old_kernel->num_params();
             // This vector records pairs containing:
@@ -205,7 +205,7 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
                 if (auto cont = def->isa_nom<Continuation>()) {
                     // Copy the basic block by calling stub
                     // Or reuse the newly created kernel copy if def is the old kernel
-                    auto new_cont = def == old_kernel ? new_kernel : cont->stub();
+                    auto new_cont = def == old_kernel ? new_kernel : cont->mangle_stub();
                     rewriter.old2new[cont] = new_cont;
                     for (size_t i = 0; i < cont->num_params(); ++i)
                         rewriter.old2new[cont->param(i)] = new_cont->param(i);
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index ceaabb6f2..a1ff8bffa 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -2,7 +2,7 @@
 
 namespace thorin {
 
-const Type* Importer::import(const Type* otype) {
+/*const Type* Importer::import(const Type* otype) {
     if (auto ntype = type_old2new_.lookup(otype)) {
         assert(&(*ntype)->table() == &world());
         return *ntype;
@@ -26,7 +26,7 @@ const Type* Importer::import(const Type* otype) {
     assert(&ntype->table() == &world());
 
     return ntype;
-}
+}*/
 
 const Def* Importer::import(const Def* odef) {
     if (auto ndef = def_old2new_.lookup(odef)) {
@@ -34,9 +34,14 @@ const Def* Importer::import(const Def* odef) {
         return *ndef;
     }
 
-    auto ntype = import(odef->type());
+    if (odef == odef->world().star()) {
+        def_old2new_[odef] = world().star();
+        return world().star();
+    }
+
+    auto ntype = import(odef->type())->as<Type>();
 
-    if (auto oparam = odef->isa<Param>()) {
+    /*if (auto oparam = odef->isa<Param>()) {
         import(oparam->continuation())->as_nom<Continuation>();
         auto nparam = def_old2new_[oparam];
         assert(nparam && &nparam->world() == &world());
@@ -62,7 +67,7 @@ const Def* Importer::import(const Def* odef) {
         auto npi = import(ocontinuation->type())->as<FnType>();
         ncontinuation = world().continuation(npi, ocontinuation->attributes(), ocontinuation->debug_history());
         assert(&ncontinuation->world() == &world());
-        assert(&npi->table() == &world());
+        assert(&npi->world() == &world());
         for (size_t i = 0, e = ocontinuation->num_params(); i != e; ++i) {
             ncontinuation->param(i)->set_name(ocontinuation->param(i)->debug_history().name);
             def_old2new_[ocontinuation->param(i)] = ncontinuation->param(i);
@@ -72,6 +77,12 @@ const Def* Importer::import(const Def* odef) {
 
         if (ocontinuation->is_external())
             world().make_external(ncontinuation);
+    }*/
+
+    Def* stub = nullptr;
+    if (odef->isa_nom()) {
+        stub = odef->stub(world(), ntype);
+        def_old2new_[odef] = stub;
     }
 
     size_t size = odef->num_ops();
@@ -86,15 +97,19 @@ const Def* Importer::import(const Def* odef) {
         auto ndef = odef->rebuild(world(), ntype, nops);
         todo_ |= odef->tag() != ndef->tag();
         return def_old2new_[odef] = ndef;
+    } else {
+        assert(odef->isa_nom() && stub);
+        stub->rebuild_from(odef, nops);
+        return stub;
     }
 
-    assert(ncontinuation && &ncontinuation->world() == &world());
+    /*assert(ncontinuation && &ncontinuation->world() == &world());
     auto napp = nops[0]->isa<App>();
     if (napp)
         ncontinuation->set_body(napp);
     ncontinuation->set_filter(nops[1]->as<Filter>());
     ncontinuation->verify();
-    return ncontinuation;
+    return ncontinuation;*/
 }
 
 }
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index db41b3161..279dadeaf 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -21,12 +21,11 @@ class Importer {
     }
 
     World& world() { return dst; }
-    const Type* import(const Type*);
+    //const Type* import(const Type*);
     const Def* import(const Def*);
     bool todo() const { return todo_; }
 
 public:
-    Type2Type type_old2new_;
     Def2Def def_old2new_;
     World& src;
     World& dst;
diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index 0dcc0d909..e6523f21f 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -107,7 +107,7 @@ Continuation* Mangler::mangle() {
 Continuation* Mangler::mangle_head(Continuation* old_continuation) {
     assert(!def2def_.contains(old_continuation));
     assert(old_continuation->has_body());
-    Continuation* new_continuation = old_continuation->stub();
+    Continuation* new_continuation = old_continuation->mangle_stub();
     def2def_[old_continuation] = new_continuation;
 
     for (size_t i = 0, e = old_continuation->num_params(); i != e; ++i)
diff --git a/src/thorin/transform/mangle.h b/src/thorin/transform/mangle.h
index 52621968d..162da6071 100644
--- a/src/thorin/transform/mangle.h
+++ b/src/thorin/transform/mangle.h
@@ -30,7 +30,6 @@ class Mangler {
     const Scope& scope_;
     Defs args_;
     Defs lift_;
-    Type2Type type2type_;
     Continuation* old_entry_;
     Continuation* new_entry_;
     DefSet defs_;
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index f8cc33751..b317806fe 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -11,15 +11,23 @@
 
 namespace thorin {
 
-Type::Type(TypeTable& table, int tag, Types ops)
-    : table_(&table)
-    , tag_(tag)
-    , ops_(ops.size())
-{
-    for (size_t i = 0, e = num_ops(); i != e; ++i) {
-        if (auto op = ops[i])
-            set(i, op);
-    }
+Type::Type(World& w, NodeTag tag, Defs args, Debug dbg) : Type(w, tag, w.star(), args, dbg) {}
+Type::Type(World& w, NodeTag tag, size_t size, Debug dbg) : Type(w, tag, w.star(), size, dbg) {}
+
+Array<const Def*> types2defs(ArrayRef<const Type*> types) {
+    Array<const Def*> defs(types.size());
+    size_t i = 0;
+    for (auto type : types)
+        defs[i++] = type->as<Def>();
+    return defs;
+}
+
+Array<const Type*> defs2types(ArrayRef<const Def*> defs) {
+    Array<const Type*> types(defs.size());
+    size_t i = 0;
+    for (auto type : defs)
+        types[i++] = type->as<Type>();
+    return types;
 }
 
 //------------------------------------------------------------------------------
@@ -28,34 +36,33 @@ Type::Type(TypeTable& table, int tag, Types ops)
  * rebuild
  */
 
-const Type* NominalType::rebuild(TypeTable&, Types) const {
+const Type* NominalType::rebuild(World& w, const Type* t, Defs o) const {
     THORIN_UNREACHABLE;
-    return this;
 }
 
-const Type* BottomType         ::rebuild(TypeTable& t, Types  ) const { return t.bottom_type(); }
-const Type* ClosureType        ::rebuild(TypeTable& t, Types o) const { return t.closure_type(o); }
-const Type* DefiniteArrayType  ::rebuild(TypeTable& t, Types o) const { return t.definite_array_type(o[0], dim()); }
-const Type* FnType             ::rebuild(TypeTable& t, Types o) const { return t.fn_type(o); }
-const Type* FrameType          ::rebuild(TypeTable& t, Types  ) const { return t.frame_type(); }
-const Type* IndefiniteArrayType::rebuild(TypeTable& t, Types o) const { return t.indefinite_array_type(o[0]); }
-const Type* MemType            ::rebuild(TypeTable& t, Types  ) const { return t.mem_type(); }
-const Type* PrimType           ::rebuild(TypeTable& t, Types  ) const { return t.prim_type(primtype_tag(), length()); }
-const Type* PtrType            ::rebuild(TypeTable& t, Types o) const { return t.ptr_type(o.front(), length(), device(), addr_space()); }
-const Type* TupleType          ::rebuild(TypeTable& t, Types o) const { return t.tuple_type(o); }
+const Type* BottomType         ::rebuild(World& w, const Type* t, Defs o) const { return w.bottom_type(); }
+const Type* ClosureType        ::rebuild(World& w, const Type* t, Defs o) const { return w.closure_type(defs2types(o)); }
+const Type* DefiniteArrayType  ::rebuild(World& w, const Type* t, Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
+const Type* FnType             ::rebuild(World& w, const Type* t, Defs o) const { return w.fn_type(defs2types(o)); }
+const Type* FrameType          ::rebuild(World& w, const Type* t, Defs o) const { return w.frame_type(); }
+const Type* IndefiniteArrayType::rebuild(World& w, const Type* t, Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
+const Type* MemType            ::rebuild(World& w, const Type* t, Defs o) const { return w.mem_type(); }
+const Type* PrimType           ::rebuild(World& w, const Type* t, Defs o) const { return w.prim_type(primtype_tag(), length()); }
+const Type* PtrType            ::rebuild(World& w, const Type* t, Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
+const Type* TupleType          ::rebuild(World& w, const Type* t, Defs o) const { return w.tuple_type(defs2types(o)); }
 
 /*
  * stub
  */
 
-const NominalType* StructType::stub(TypeTable& to) const {
-    auto type = to.struct_type(name(), num_ops());
+StructType* StructType::stub(World& world, const Type*) const {
+    auto type = world.struct_type(name(), num_ops());
     std::copy(op_names_.begin(), op_names_.end(), type->op_names().begin());
     return type;
 }
 
-const NominalType* VariantType::stub(TypeTable& to) const {
-    auto type = to.variant_type(name(), num_ops());
+VariantType* VariantType::stub(World& world, const Type*) const {
+    auto type = world.variant_type(name(), num_ops());
     std::copy(op_names_.begin(), op_names_.end(), type->op_names().begin());
     return type;
 }
@@ -64,8 +71,8 @@ const NominalType* VariantType::stub(TypeTable& to) const {
 
 const VectorType* VectorType::scalarize() const {
     if (auto ptr = isa<PtrType>())
-        return table().ptr_type(ptr->pointee());
-    return table().prim_type(as<PrimType>()->primtype_tag());
+        return world().ptr_type(ptr->pointee());
+    return world().prim_type(as<PrimType>()->primtype_tag());
 }
 
 bool FnType::is_returning() const {
@@ -85,7 +92,7 @@ bool FnType::is_returning() const {
 }
 
 bool VariantType::has_payload() const {
-    return !std::all_of(ops().begin(), ops().end(), is_type_unit);
+    return !std::all_of(types().begin(), types().end(), is_type_unit);
 }
 
 bool use_lea(const Type* type) { return type->isa<StructType>() || type->isa<ArrayType>(); }
@@ -96,15 +103,6 @@ bool use_lea(const Type* type) { return type->isa<StructType>() || type->isa<Arr
  * hash
  */
 
-hash_t Type::vhash() const {
-    if (is_nominal())
-        return thorin::murmur3(hash_t(tag()) << hash_t(32-8) | hash_t(gid()));
-    hash_t seed = thorin::hash_begin(uint8_t(tag()));
-    for (auto op : ops_)
-        seed = thorin::hash_combine(seed, uint32_t(op->gid()));
-    return seed;
-}
-
 hash_t PtrType::vhash() const {
     return hash_combine(VectorType::vhash(), (hash_t)device(), (hash_t)addr_space());
 }
@@ -115,15 +113,7 @@ hash_t PtrType::vhash() const {
  * equal
  */
 
-bool Type::equal(const Type* other) const {
-    if (is_nominal())
-        return this == other;
-    if (tag() == other->tag() && num_ops() == other->num_ops())
-        return std::equal(ops().begin(), ops().end(), other->ops().begin());
-    return false;
-}
-
-bool PtrType::equal(const Type* other) const {
+bool PtrType::equal(const Def* other) const {
     if (!VectorType::equal(other))
         return false;
     auto ptr = other->as<PtrType>();
@@ -186,62 +176,46 @@ Stream& Type::stream(Stream& s) const {
 
 //------------------------------------------------------------------------------
 
-TypeTable::TypeTable()
-    : unit_     (insert<TupleType >(*this, Types()))
-    , fn0_      (insert<FnType    >(*this, Types()))
-    , bottom_ty_(insert<BottomType>(*this))
-    , mem_      (insert<MemType   >(*this))
-    , frame_    (insert<FrameType >(*this))
+TypeTable::TypeTable(World& world)
+    : world_(world)
+    , star_     (world.put<Star>((world)))
+    , unit_     (world.put<TupleType>(world, Defs(), Debug()))
+    , fn0_      (world.put<FnType    >(world, Defs(), Node_FnType, Debug()))
+    , bottom_ty_(world.put<BottomType>(world, Debug()))
+    , mem_      (world.put<MemType   >(world, Debug()))
+    , frame_    (world.put<FrameType >(world, Debug()))
 {
 #define THORIN_ALL_TYPE(T, M) \
-    primtypes_[PrimType_##T - Begin_PrimType] = insert<PrimType>(*this, PrimType_##T, 1);
+    primtypes_[PrimType_##T - Begin_PrimType] = world.make<PrimType>(world, PrimType_##T, 1, Debug());
 #include "thorin/tables/primtypetable.h"
 }
 
-const Type* TypeTable::tuple_type(Types ops) {
-    return ops.size() == 1 ? ops.front() : insert<TupleType>(*this, ops);
+const Type* World::tuple_type(Types ops) {
+    return ops.size() == 1 ? ops.front()->as<Type>() : make<TupleType>(*this, types2defs(ops), Debug());
 }
 
-const StructType* TypeTable::struct_type(Symbol name, size_t size) {
-    auto type = new StructType(*this, name, size, types_.size());
-    const auto& p = types_.insert(type);
-    assert_unused(p.second && "hash/equal broken");
-    return type;
+StructType* World::struct_type(Symbol name, size_t size) {
+    return put<StructType>(*this, name, size, Debug());
 }
 
-const VariantType* TypeTable::variant_type(Symbol name, size_t size) {
-    auto type = new VariantType(*this, name, size, types_.size());
-    const auto& p = types_.insert(type);
-    assert_unused(p.second && "hash/equal broken");
-    return type;
+VariantType* World::variant_type(Symbol name, size_t size) {
+    return put<VariantType>(*this, name, size, Debug());
 }
 
-const PrimType* TypeTable::prim_type(PrimTypeTag tag, size_t length) {
+const PrimType* World::prim_type(PrimTypeTag tag, size_t length) {
     size_t i = tag - Begin_PrimType;
     assert(i < (size_t) Num_PrimTypes);
-    return length == 1 ? primtypes_[i] : insert<PrimType>(*this, tag, length);
+    return length == 1 ? types_.primtypes_[i] : make<PrimType>(*this, tag, length, Debug());
 }
 
-const PtrType* TypeTable::ptr_type(const Type* pointee, size_t length, int32_t device, AddrSpace addr_space) {
-    return insert<PtrType>(*this, pointee, length, device, addr_space);
+const PtrType* World::ptr_type(const Type* pointee, size_t length, int32_t device, AddrSpace addr_space) {
+    return make<PtrType>(*this, pointee, length, device, addr_space, Debug());
 }
 
-const FnType*              TypeTable::fn_type(Types args) { return insert<FnType>(*this, args); }
-const ClosureType*         TypeTable::closure_type(Types args) { return insert<ClosureType>(*this, args); }
-const DefiniteArrayType*   TypeTable::definite_array_type(const Type* elem, u64 dim) { return insert<DefiniteArrayType>(*this, elem, dim); }
-const IndefiniteArrayType* TypeTable::indefinite_array_type(const Type* elem) { return insert<IndefiniteArrayType>(*this, elem); }
-
-template <typename T, typename... Args>
-const T* TypeTable::insert(Args&&... args) {
-    T t(std::forward<Args&&>(args)...);
-    auto it = types_.find(&t);
-    if (it != types_.end())
-        return (*it)->template as<T>();
-    auto new_t = new T(std::move(t));
-    new_t->gid_ = types_.size();
-    types_.emplace(new_t);
-    return new_t;
-}
+const FnType*              World::fn_type(Types args) { return make<FnType>(*this, types2defs(args), Node_FnType, Debug()); }
+const ClosureType*         World::closure_type(Types args) { return make<ClosureType>(*this, types2defs(args), Debug()); }
+const DefiniteArrayType*   World::definite_array_type(const Type* elem, u64 dim) { return make<DefiniteArrayType>(*this, elem, dim, Debug()); }
+const IndefiniteArrayType* World::indefinite_array_type(const Type* elem) { return make<IndefiniteArrayType>(*this, elem, Debug()); }
 
 //------------------------------------------------------------------------------
 
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 2ce899290..84383e250 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -1,6 +1,7 @@
 #ifndef THORIN_TYPE_H
 #define THORIN_TYPE_H
 
+#include "thorin/def.h"
 #include "thorin/enums.h"
 #include "thorin/util/hash.h"
 #include "thorin/util/cast.h"
@@ -15,174 +16,168 @@ class Type;
 using Types = ArrayRef<const Type*>;
 
 /// Base class for all \p Type%s.
-class Type : public RuntimeCast<Type>, public Streamable<Type> {
+class Type : public Def, public Streamable<Type> {
 protected:
-    Type(TypeTable& table, int tag, Types ops);
-
-    void set(size_t i, const Type* type) {
-        ops_[i] = type;
-        order_ = std::max(order_, type->order());
-    }
+    /// Constructor for a @em structural Type.
+    Type(World& w, NodeTag tag, const Type* type, Defs args, Debug dbg) : Def(w, tag, type, args, dbg) {}
+    Type(World& w, NodeTag tag, Defs args, Debug dbg);
+    /// Constructor for a @em nom Type.
+    Type(World& w, NodeTag tag, const Type* type, size_t size, Debug dbg) : Def(w, tag, type, size, dbg) {}
+    Type(World& w, NodeTag tag, size_t size, Debug dbg);
 
 public:
-    int tag() const { return tag_; }
-    TypeTable& table() const { return *table_; }
-
-    Types ops() const { return ops_; }
-    const Type* op(size_t i) const { return ops()[i]; }
-    size_t num_ops() const { return ops_.size(); }
-    bool empty() const { return ops_.empty(); }
-
-    bool is_nominal() const { return nominal_; } ///< A nominal @p Type is always different from each other @p Type.
-    int order() const { return order_; }
-    size_t gid() const { return gid_; }
-    hash_t hash() const { return hash_ == 0 ? hash_ = vhash() : hash_; }
-    virtual bool equal(const Type*) const;
-    virtual const Type* rebuild(TypeTable&, Types) const = 0;
+    int order() const override { return order_; }
     Stream& stream(Stream&) const;
-    void dump() const;
 
+    std::vector<const Type*> filter_type_ops() const {
+        std::vector<const Type*> type_ops;
+        for (auto& op : ops()) {
+            if (auto t = op->isa<Type>())
+                type_ops.push_back(t);
+        }
+        return type_ops;
+    }
 protected:
-    virtual hash_t vhash() const;
-
-    mutable hash_t hash_ = 0;
-    mutable bool nominal_  = false;
     int order_ = 0;
-    size_t gid_;
+    friend class World;
+};
 
-private:
-    mutable TypeTable* table_;
+class Star : public Type {
+protected:
+    explicit Star(World& w) : Type(w, Node_Star, nullptr, 0, {}) {
+        set_type(this);
+    }
+
+    friend class World;
+};
 
-    int tag_;
-    thorin::Array<const Type*> ops_;
+Array<const Def*> types2defs(ArrayRef<const Type*> types);
+Array<const Type*> defs2types(ArrayRef<const Def*> types);
 
-    friend TypeTable;
+template<class T>
+class TypeOpsMixin {
+public:
+    Types types() const {
+        Defs defs = static_cast<const T*>(this)->ops();
+        const Def* const* ptr = defs.begin();
+        auto ptr2 = reinterpret_cast<const Type* const*>(ptr);
+        auto types = Types(ptr2, defs.size());
+        return types;
+    }
 };
 
 /// Type of a tuple (structurally typed).
-class TupleType : public Type {
+class TupleType : public Type, public TypeOpsMixin<TupleType> {
 private:
-    TupleType(TypeTable& table, Types ops)
-        : Type(table, Node_TupleType, ops)
+    TupleType(World& world, Defs ops, Debug dbg)
+        : Type(world, Node_TupleType, ops, dbg)
     {}
 
 public:
-    const Type* rebuild(TypeTable&, Types) const override;
-
-    friend class TypeTable;
+    const Type* rebuild(World&, const Type*, Defs) const override;
+    friend class World;
 };
 
 /// Base class for nominal types (types that have
 /// a name that uniquely identifies them).
 class NominalType : public Type {
 protected:
-    NominalType(TypeTable& table, int tag, Symbol name, size_t size, size_t gid)
-        : Type(table, tag, thorin::Array<const Type*>(size))
+    NominalType(World& world, NodeTag tag, Symbol name, size_t size, Debug dbg)
+        : Type(world, tag, size, dbg)
         , name_(name)
         , op_names_(size)
-    {
-        nominal_ = true;
-        gid_ = gid;
-    }
+    {}
 
     Symbol name_;
     Array<Symbol> op_names_;
 
 private:
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
 public:
     Symbol name() const { return name_; }
     Symbol op_name(size_t i) const { return op_names_[i]; }
-    void set(size_t i, const Type* type) const {
-        return const_cast<NominalType*>(this)->Type::set(i, type);
-    }
     void set_op_name(size_t i, Symbol name) const {
         const_cast<NominalType*>(this)->op_names_[i] = name;
     }
     Array<Symbol>& op_names() const {
         return const_cast<NominalType*>(this)->op_names_;
     }
-
-    /// Recreates a fresh new nominal type of the
-    /// same kind with the same number of operands,
-    /// initially all unset.
-    virtual const NominalType* stub(TypeTable&) const = 0;
 };
 
-class StructType : public NominalType {
+class StructType : public NominalType, public TypeOpsMixin<TupleType> {
 private:
-    StructType(TypeTable& table, Symbol name, size_t size, size_t gid)
-        : NominalType(table, Node_StructType, name, size, gid)
+    StructType(World& world, Symbol name, size_t size, Debug dbg)
+        : NominalType(world, Node_StructType, name, size, dbg)
     {}
 
 public:
-    const NominalType* stub(TypeTable&) const override;
+    virtual StructType* stub(World&, const Type*) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
-class VariantType : public NominalType {
+class VariantType : public NominalType, public TypeOpsMixin<TupleType> {
 private:
-    VariantType(TypeTable& table, Symbol name, size_t size, size_t gid)
-        : NominalType(table, Node_VariantType, name, size, gid)
+    VariantType(World& world, Symbol name, size_t size, Debug dbg)
+        : NominalType(world, Node_VariantType, name, size, dbg)
     {}
 
 public:
-    const NominalType* stub(TypeTable&) const override;
+    virtual VariantType* stub(World&, const Type*) const override;
 
     bool has_payload() const;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 /// The type of the memory monad.
 class MemType : public Type {
 private:
-    MemType(TypeTable& table)
-        : Type(table, Node_MemType, {})
+    MemType(World& world, Debug dbg)
+        : Type(world, Node_MemType, Defs(), dbg)
     {}
 
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 /// The type of App nodes.
 class BottomType : public Type {
 private:
-    BottomType(TypeTable& table)
-            : Type(table, Node_BotType, {})
+    BottomType(World& world, Debug dbg)
+        : Type(world, Node_BotType, Defs(), dbg)
     {}
 
-    const Type* rebuild(TypeTable& to, Types ops) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 /// The type of a stack frame.
 class FrameType : public Type {
 private:
-    FrameType(TypeTable& table)
-        : Type(table, Node_FrameType, {})
+    FrameType(World& world, Debug dbg)
+        : Type(world, Node_FrameType, Defs(), dbg)
     {}
 
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 /// Base class for all SIMD types.
 class VectorType : public Type {
 protected:
-    VectorType(TypeTable& table, int tag, Types ops, size_t length)
-        : Type(table, tag, ops)
+    VectorType(World& world, NodeTag tag, Defs ops, size_t length, Debug dbg)
+        : Type(world, tag, ops, dbg)
         , length_(length)
     {}
 
     hash_t vhash() const override { return hash_combine(Type::vhash(), length()); }
-    bool equal(const Type* other) const override {
-        return Type::equal(other) && this->length() == other->as<VectorType>()->length();
+    bool equal(const Def* other) const override {
+        return Def::equal(other) && this->length() == other->as<VectorType>()->length();
     }
 
 public:
@@ -202,15 +197,15 @@ inline size_t vector_length(const Type* type) { return type->as<VectorType>()->l
 /// Primitive type.
 class PrimType : public VectorType {
 private:
-    PrimType(TypeTable& table, PrimTypeTag tag, size_t length)
-        : VectorType(table, (int) tag, {}, length)
+    PrimType(World& world, PrimTypeTag tag, size_t length, Debug dbg)
+        : VectorType(world, (NodeTag) tag, Defs(), length, dbg)
     {}
 
 public:
     PrimTypeTag primtype_tag() const { return (PrimTypeTag) tag(); }
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 inline bool is_primtype (const Type* t) { return thorin::is_primtype(t->tag()); }
@@ -238,30 +233,30 @@ enum class AddrSpace : uint32_t {
 };
 
 /// Pointer type.
-class PtrType : public VectorType {
+class PtrType : public VectorType, public TypeOpsMixin<TupleType> {
 private:
-    PtrType(TypeTable& table, const Type* pointee, size_t length, int32_t device, AddrSpace addr_space)
-        : VectorType(table, Node_PtrType, {pointee}, length)
+    PtrType(World& world, const Type* pointee, size_t length, int32_t device, AddrSpace addr_space, Debug dbg)
+        : VectorType(world, Node_PtrType, {pointee}, length, dbg)
         , addr_space_(addr_space)
         , device_(device)
     {}
 
 public:
-    const Type* pointee() const { return op(0); }
+    const Type* pointee() const { return op(0)->as<Type>(); }
     AddrSpace addr_space() const { return addr_space_; }
     int32_t device() const { return device_; }
     bool is_host_device() const { return device_ == -1; }
 
     hash_t vhash() const override;
-    bool equal(const Type* other) const override;
+    bool equal(const Def* other) const override;
 
 private:
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
     AddrSpace addr_space_;
     int32_t device_;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 /// Returns true if the given type is small enough to fit in a closure environment
@@ -269,10 +264,10 @@ inline bool is_thin(const Type* type) {
     return type->isa<PrimType>() || type->isa<PtrType>() || is_type_unit(type);
 }
 
-class FnType : public Type {
+class FnType : public Type, public TypeOpsMixin<TupleType> {
 protected:
-    FnType(TypeTable& table, Types ops, int tag = Node_FnType)
-        : Type(table, tag, ops)
+    FnType(World& world, Defs ops, NodeTag tag, Debug dbg)
+        : Type(world, tag, ops, dbg)
     {
         ++order_;
     }
@@ -282,15 +277,15 @@ class FnType : public Type {
     bool is_returning() const;
 
 private:
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 class ClosureType : public FnType {
 private:
-    ClosureType(TypeTable& table, Types ops)
-        : FnType(table, ops, Node_ClosureType)
+    ClosureType(World& world, Defs ops, Debug dbg)
+        : FnType(world, ops, Node_ClosureType, dbg)
     {
         inner_order_ = order_;
         order_ = 0;
@@ -298,155 +293,84 @@ class ClosureType : public FnType {
 
 public:
     int inner_order() const { return inner_order_; }
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
 private:
     int inner_order_;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 //------------------------------------------------------------------------------
 
-class ArrayType : public Type {
+class ArrayType : public Type, public TypeOpsMixin<TupleType> {
 protected:
-    ArrayType(TypeTable& table, int tag, const Type* elem_type)
-        : Type(table, tag, {elem_type})
+    ArrayType(World& world, NodeTag tag, const Type* elem_type, Debug dbg)
+        : Type(world, tag, {elem_type}, dbg)
     {}
 
 public:
-    const Type* elem_type() const { return op(0); }
+    const Type* elem_type() const { return op(0)->as<Type>(); }
 };
 
 class IndefiniteArrayType : public ArrayType {
 public:
-    IndefiniteArrayType(TypeTable& table, const Type* elem_type)
-        : ArrayType(table, Node_IndefiniteArrayType, elem_type)
+    IndefiniteArrayType(World& world, const Type* elem_type, Debug dbg)
+        : ArrayType(world, Node_IndefiniteArrayType, elem_type, dbg)
     {}
 
 private:
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 class DefiniteArrayType : public ArrayType {
 public:
-    DefiniteArrayType(TypeTable& table, const Type* elem_type, u64 dim)
-        : ArrayType(table, Node_DefiniteArrayType, elem_type)
+    DefiniteArrayType(World& world, const Type* elem_type, u64 dim, Debug dbg)
+        : ArrayType(world, Node_DefiniteArrayType, elem_type, dbg)
         , dim_(dim)
     {}
 
     u64 dim() const { return dim_; }
     hash_t vhash() const override { return hash_combine(Type::vhash(), dim()); }
-    bool equal(const Type* other) const override {
-        return Type::equal(other) && this->dim() == other->as<DefiniteArrayType>()->dim();
+    bool equal(const Def* other) const override {
+        return Def::equal(other) && this->dim() == other->as<DefiniteArrayType>()->dim();
     }
 
 private:
-    const Type* rebuild(TypeTable&, Types) const override;
+    const Type* rebuild(World&, const Type*, Defs) const override;
 
     u64 dim_;
 
-    friend class TypeTable;
+    friend class World;
 };
 
 bool use_lea(const Type*);
 
 //------------------------------------------------------------------------------
 
-/// Container for all types. Types are hashed and can be compared using pointer equality.
 class TypeTable {
-private:
-    struct TypeHash {
-        static hash_t hash(const Type* t) { return t->hash(); }
-        static bool eq(const Type* t1, const Type* t2) { return t2->equal(t1); }
-        static const Type* sentinel() { return (const Type*)(1); }
-    };
-
-    typedef thorin::HashSet<const Type*, TypeHash> TypeSet;
-
 public:
-    TypeTable();
-
-    const Type* tuple_type(Types ops);
-    const TupleType* unit() { return unit_; } ///< Returns unit, i.e., an empty @p TupleType.
-    const VariantType* variant_type(Symbol name, size_t size);
-    const StructType* struct_type(Symbol name, size_t size);
-
-#define THORIN_ALL_TYPE(T, M) \
-    const PrimType* type_##T(size_t length = 1) { return prim_type(PrimType_##T, length); }
-#include "thorin/tables/primtypetable.h"
-    const PrimType* prim_type(PrimTypeTag tag, size_t length = 1);
-    const BottomType* bottom_type() const { return bottom_ty_; }
-    const MemType* mem_type() const { return mem_; }
-    const FrameType* frame_type() const { return frame_; }
-    const PtrType* ptr_type(const Type* pointee, size_t length = 1, int32_t device = -1, AddrSpace addr_space = AddrSpace::Generic);
-    const FnType* fn_type() { return fn0_; } ///< Returns an empty @p FnType.
-    const FnType* fn_type(Types args);
-    const ClosureType* closure_type(Types args);
-    const DefiniteArrayType*   definite_array_type(const Type* elem, u64 dim);
-    const IndefiniteArrayType* indefinite_array_type(const Type* elem);
-
-    const TypeSet& types() const { return types_; }
-
-    friend void swap(TypeTable& t1, TypeTable& t2) {
-        using std::swap;
-        swap(t1.types_, t2.types_);
-        swap(t1.unit_,  t2.unit_);
-        swap(t1.fn0_,   t2.fn0_);
-        swap(t1.bottom_ty_,   t2.bottom_ty_);
-        swap(t1.mem_,   t2.mem_);
-        swap(t1.frame_, t2.frame_);
-        std::swap_ranges(t1.primtypes_, t1.primtypes_ + Num_PrimTypes, t2.primtypes_);
-
-        t1.fix();
-        t2.fix();
-    }
-
-private:
-    void fix() {
-        for (auto type : types_)
-            type->table_ = this;
-    }
-
-    template <typename T, typename... Args>
-    const T* insert(Args&&... args);
+    explicit TypeTable(World& world);
 
 private:
-    TypeSet types_;
+    World& world_;
 
+    const Type* star_;
     const TupleType* unit_; ///< tuple().
     const FnType* fn0_;
     const BottomType* bottom_ty_;
     const MemType* mem_;
     const FrameType* frame_;
     const PrimType* primtypes_[Num_PrimTypes];
-};
-
-//------------------------------------------------------------------------------
-
-template<class T>
-struct GIDLt {
-    bool operator()(T a, T b) const { return a->gid() < b->gid(); }
-};
 
-template<class T>
-struct GIDHash {
-    static hash_t hash(T n) { return thorin::murmur3(n->gid()); }
-    static bool eq(T a, T b) { return a == b; }
-    static T sentinel() { return T(1); }
+    friend class World;
 };
 
-template<class Key, class Value>
-using GIDMap = thorin::HashMap<Key, Value, GIDHash<Key>>;
-template<class Key>
-using GIDSet = thorin::HashSet<Key, GIDHash<Key>>;
+//------------------------------------------------------------------------------
 
-template<class To>
-using TypeMap   = GIDMap<const Type*, To>;
-using Type2Type = TypeMap<const Type*>;
-using TypeSet   = GIDSet<const Type*>;
+inline bool is_mem        (const Def* def) { return def->type()->isa<MemType>(); }
 
 //------------------------------------------------------------------------------
 
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index a45ac14a4..1550493b3 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -42,16 +42,12 @@ namespace thorin {
  * constructor and destructor
  */
 
-World::World(const std::string& name) {
+World::World(const std::string& name) : types_(TypeTable(*this)) {
     data_.name_   = name;
     data_.branch_ = continuation(fn_type({type_bool(), fn_type(), fn_type()}), Intrinsic::Branch, {"br"});
     data_.end_scope_ = continuation(fn_type(), Intrinsic::EndScope, {"end_scope"});
 }
 
-World::~World() {
-    for (auto def : data_.defs_) delete def;
-}
-
 const Def* World::variant_index(const Def* value, Debug dbg) {
     if (auto variant = value->isa<Variant>())
         return literal_qu64(variant->index(), dbg);
@@ -59,7 +55,7 @@ const Def* World::variant_index(const Def* value, Debug dbg) {
 }
 
 const Def* World::variant_extract(const Def* value, size_t index, Debug dbg) {
-    auto type = value->type()->as<VariantType>()->op(index);
+    auto type = value->type()->as<VariantType>()->op(index)->as<Type>();
     if (auto variant = value->isa<Variant>())
         return variant->index() == index ? variant->value() : bottom(type);
     return cse(new VariantExtract(*this, type, value, index, dbg));
@@ -500,7 +496,7 @@ const Def* World::convert(const Type* dst_type, const Def* src, Debug dbg) {
 
         Array<const Def*> new_tuple(dst_tuple_type->num_ops());
         for (size_t i = 0, e = new_tuple.size(); i != e; ++i)
-            new_tuple[i] = convert(dst_type->op(i), extract(src, i, dbg), dbg);
+            new_tuple[i] = convert(dst_tuple_type->types()[i], extract(src, i, dbg), dbg);
 
         return tuple(new_tuple, dbg);
     }
@@ -730,13 +726,13 @@ const Def* World::insert(const Def* agg, const Def* index, const Def* value, Deb
         } else if (auto tuple_type = agg->type()->isa<TupleType>()) {
             Array<const Def*> args(tuple_type->num_ops());
             size_t i = 0;
-            for (auto type : tuple_type->ops())
+            for (auto type : tuple_type->types())
                 args[i++] = agg->isa<Bottom>() ? bottom(type, dbg) : top(type, dbg);
             agg = tuple(args, dbg);
         } else if (auto struct_type = agg->type()->isa<StructType>()) {
             Array<const Def*> args(struct_type->num_ops());
             size_t i = 0;
-            for (auto type : struct_type->ops())
+            for (auto type : struct_type->types())
                 args[i++] = agg->isa<Bottom>() ? bottom(type, dbg) : top(type, dbg);
             agg = struct_agg(struct_type, args, dbg);
         }
@@ -824,7 +820,7 @@ const Def* World::transcendental(MathOpTag tag, const Def* arg, Debug dbg, F&& f
                 THORIN_UNREACHABLE;
         }
     }
-    return cse(new MathOp(tag, *this, arg->type(), { arg }, dbg));
+    return cse(new MathOp(*this, tag, arg->type(), { arg }, dbg));
 }
 
 template <class F>
@@ -846,7 +842,7 @@ const Def* World::transcendental(MathOpTag tag, const Def* left, const Def* righ
                 THORIN_UNREACHABLE;
         }
     }
-    return cse(new MathOp(tag, *this, left->type(), { left, right }, dbg));
+    return cse(new MathOp(*this, tag, left->type(), { left, right }, dbg));
 }
 
 template <class F>
@@ -1102,15 +1098,7 @@ const Def* World::run(const Def* def, Debug dbg) {
  */
 
 Continuation* World::continuation(const FnType* fn, Continuation::Attributes attributes, Debug dbg) {
-    auto cont = put<Continuation>(*this, fn, attributes, dbg);
-
-    size_t i = 0;
-    for (auto op : fn->ops()) {
-        auto p = param(op, cont, i++, dbg);
-        cont->params_.emplace_back(p);
-    }
-
-    return cont;
+    return put<Continuation>(*this, fn, attributes, dbg);
 }
 
 Continuation* World::match(const Type* type, size_t num_patterns) {
@@ -1122,8 +1110,8 @@ Continuation* World::match(const Type* type, size_t num_patterns) {
     return continuation(fn_type(arg_types), Intrinsic::Match, {"match"});
 }
 
-const Param* World::param(const Type* type, Continuation* continuation, size_t index, Debug dbg) {
-    auto param = new Param(*this, type, continuation, index, dbg);
+const Param* World::param(const Type* type, const Continuation* continuation, size_t index, Debug dbg) {
+    auto param = cse(new Param(*this, type, continuation, index, dbg));
 #if THORIN_ENABLE_CHECKS
     if (state_.breakpoints.contains(param->gid())) THORIN_BREAK;
 #endif
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 3a0d11be0..e6f4d98f5 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -43,7 +43,7 @@ enum class LogLevel { Debug, Verbose, Info, Warn, Error };
  *  All worlds are completely independent from each other.
  *  This is particular useful for multi-threading.
  */
-class World : public TypeTable, public Streamable<World> {
+class World : public Streamable<World> {
 public:
     struct SeaHash {
         static hash_t hash(const Def* def) { return def->hash(); }
@@ -78,7 +78,6 @@ class World : public TypeTable, public Streamable<World> {
         stream_ = other.stream_;
         state_  = other.state_;
     }
-    ~World();
 
     /// @name manage global identifier - a unique number for each Def
     //@{
@@ -90,12 +89,35 @@ class World : public TypeTable, public Streamable<World> {
     //@{
     bool empty() { return data_.externals_.empty(); }
     const Externals& externals() const { return data_.externals_; }
-    void make_external(Continuation* cont) { data_.externals_.emplace(cont->unique_name(), cont); }
-    void make_internal(Continuation* cont) { data_.externals_.erase(cont->unique_name()); }
+    void make_external(Continuation* cont) { assert(&cont->world() == this); data_.externals_.emplace(cont->unique_name(), cont); }
+    void make_internal(Continuation* cont) { assert(&cont->world() == this); data_.externals_.erase(cont->unique_name()); }
     bool is_external(const Continuation* cont) { return data_.externals_.contains(cont->unique_name()); }
     Continuation* lookup(const std::string& name) { return data_.externals_.lookup(name).value_or(nullptr); }
     //@}
 
+    // types
+
+    const Type* star() { return types_.star_; }
+
+    const Type* tuple_type(Types ops);
+    const TupleType* unit_type() { return tuple_type({})->as<TupleType>(); } ///< Returns unit, i.e., an empty @p TupleType.
+    VariantType* variant_type(Symbol name, size_t size);
+    StructType* struct_type(Symbol name, size_t size);
+
+#define THORIN_ALL_TYPE(T, M) \
+    const PrimType* type_##T(size_t length = 1) { return prim_type(PrimType_##T, length); }
+#include "thorin/tables/primtypetable.h"
+    const PrimType* prim_type(PrimTypeTag tag, size_t length = 1);
+    const BottomType* bottom_type() { return make<BottomType>(*this, Debug()); }
+    const MemType* mem_type() { return make<MemType>(*this, Debug()); }
+    const FrameType* frame_type() { return make<FrameType>(*this, Debug()); }
+    const PtrType* ptr_type(const Type* pointee, size_t length = 1, int32_t device = -1, AddrSpace addr_space = AddrSpace::Generic);
+    const FnType* fn_type() { return fn_type({}); } ///< Returns an empty @p FnType.
+    const FnType* fn_type(Types args);
+    const ClosureType* closure_type(Types args);
+    const DefiniteArrayType*   definite_array_type(const Type* elem, u64 dim);
+    const IndefiniteArrayType* indefinite_array_type(const Type* elem);
+
     // literals
 
 #define THORIN_ALL_TYPE(T, M) \
@@ -307,7 +329,7 @@ class World : public TypeTable, public Streamable<World> {
     //@}
 
 private:
-    const Param* param(const Type* type, Continuation* continuation, size_t index, Debug dbg);
+    const Param* param(const Type* type, const Continuation*, size_t index, Debug dbg);
     const App* app(const Def* callee, const Defs args, Debug dbg = {});
     const Def* try_fold_aggregate(const Aggregate*);
     template <class F> const Def* transcendental(MathOpTag, const Def*, Debug, F&&);
@@ -317,6 +339,11 @@ class World : public TypeTable, public Streamable<World> {
     //@{
     template <class T> const T* cse(const T* primop) { return cse_base(primop)->template as<T>(); }
     const Def* cse_base(const Def*);
+    template <class T, class... Args>
+    const T* make(Args&&... args) {
+        auto def = new T(args...);
+        return cse<T>(def);
+    }
 
     template<class T, class... Args>
     T* put(Args&&... args) {
@@ -349,15 +376,19 @@ class World : public TypeTable, public Streamable<World> {
         Continuation* end_scope_;
     } data_;
 
+    TypeTable types_;
+
     std::shared_ptr<Stream> stream_;
 
     friend class Mangler;
     friend class Cleaner;
     friend class Continuation;
+    friend class Param;
     friend class Filter;
     friend class App;
     friend class Importer;
     friend class Thorin;
+    friend class TypeTable;
 };
 
 class Thorin {

From afd08d77ba6298e0311d7b1ce3a67f05311069eb Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 23 Jan 2023 17:48:05 +0100
Subject: [PATCH 159/342] eliminate commented out code from the importer

---
 src/thorin/transform/importer.cpp | 72 -------------------------------
 src/thorin/transform/importer.h   |  1 -
 2 files changed, 73 deletions(-)

diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index a1ff8bffa..4691e9795 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -2,32 +2,6 @@
 
 namespace thorin {
 
-/*const Type* Importer::import(const Type* otype) {
-    if (auto ntype = type_old2new_.lookup(otype)) {
-        assert(&(*ntype)->table() == &world());
-        return *ntype;
-    }
-    size_t size = otype->num_ops();
-
-    if (auto nominal_type = otype->isa<NominalType>()) {
-        auto ntype = nominal_type->stub(world());
-        type_old2new_[otype] = ntype;
-        for (size_t i = 0; i != size; ++i)
-            ntype->set(i, import(otype->op(i)));
-        return ntype;
-    }
-
-    Array<const Type*> nops(size);
-    for (size_t i = 0; i != size; ++i)
-        nops[i] = import(otype->op(i));
-
-    auto ntype = otype->rebuild(world(), nops);
-    type_old2new_[otype] = ntype;
-    assert(&ntype->table() == &world());
-
-    return ntype;
-}*/
-
 const Def* Importer::import(const Def* odef) {
     if (auto ndef = def_old2new_.lookup(odef)) {
         assert(&(*ndef)->world() == &world());
@@ -41,44 +15,6 @@ const Def* Importer::import(const Def* odef) {
 
     auto ntype = import(odef->type())->as<Type>();
 
-    /*if (auto oparam = odef->isa<Param>()) {
-        import(oparam->continuation())->as_nom<Continuation>();
-        auto nparam = def_old2new_[oparam];
-        assert(nparam && &nparam->world() == &world());
-        return nparam;
-    }
-
-    if (auto ofilter = odef->isa<Filter>()) {
-        Array<const Def*> new_conditions(ofilter->num_ops());
-        for (size_t i = 0, e = ofilter->size(); i != e; ++i)
-            new_conditions[i] = import(ofilter->condition(i));
-        auto nfilter = world().filter(new_conditions, ofilter->debug());
-        return nfilter;
-    }
-
-    Continuation* ncontinuation = nullptr;
-    if (auto ocontinuation = odef->isa_nom<Continuation>()) { // create stub in new world
-        assert(!ocontinuation->dead_);
-        // TODO maybe we want to deal with intrinsics in a more streamlined way
-        if (ocontinuation == ocontinuation->world().branch())
-            return def_old2new_[ocontinuation] = world().branch();
-        if (ocontinuation == ocontinuation->world().end_scope())
-            return def_old2new_[ocontinuation] = world().end_scope();
-        auto npi = import(ocontinuation->type())->as<FnType>();
-        ncontinuation = world().continuation(npi, ocontinuation->attributes(), ocontinuation->debug_history());
-        assert(&ncontinuation->world() == &world());
-        assert(&npi->world() == &world());
-        for (size_t i = 0, e = ocontinuation->num_params(); i != e; ++i) {
-            ncontinuation->param(i)->set_name(ocontinuation->param(i)->debug_history().name);
-            def_old2new_[ocontinuation->param(i)] = ncontinuation->param(i);
-        }
-
-        def_old2new_[ocontinuation] = ncontinuation;
-
-        if (ocontinuation->is_external())
-            world().make_external(ncontinuation);
-    }*/
-
     Def* stub = nullptr;
     if (odef->isa_nom()) {
         stub = odef->stub(world(), ntype);
@@ -102,14 +38,6 @@ const Def* Importer::import(const Def* odef) {
         stub->rebuild_from(odef, nops);
         return stub;
     }
-
-    /*assert(ncontinuation && &ncontinuation->world() == &world());
-    auto napp = nops[0]->isa<App>();
-    if (napp)
-        ncontinuation->set_body(napp);
-    ncontinuation->set_filter(nops[1]->as<Filter>());
-    ncontinuation->verify();
-    return ncontinuation;*/
 }
 
 }
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index 279dadeaf..10941b915 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -21,7 +21,6 @@ class Importer {
     }
 
     World& world() { return dst; }
-    //const Type* import(const Type*);
     const Def* import(const Def*);
     bool todo() const { return todo_; }
 

From fc366745a796ec687d957a0a0ce6372d4d0af48e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 23 Jan 2023 21:26:08 +0100
Subject: [PATCH 160/342] simplify op_name()

---
 src/thorin/primop.cpp | 25 +++----------------------
 src/thorin/primop.h   |  3 ---
 2 files changed, 3 insertions(+), 25 deletions(-)

diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index 23d5ba859..6c5ab6b26 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -246,32 +246,13 @@ const Def* IndefiniteArray::rebuild(World& w, const Type* t, Defs o) const {
 
 const char* Def::op_name() const {
     switch (tag()) {
+#define THORIN_GLUE(pre, next)
 #define THORIN_NODE(op, abbr) case Node_##op: return #abbr;
-#include "thorin/tables/nodetable.h"
-        default: THORIN_UNREACHABLE;
-    }
-}
-
-const char* ArithOp::op_name() const {
-    switch (tag()) {
+#define THORIN_PRIMTYPE(T) case Node_PrimType_##T: return #T;
 #define THORIN_ARITHOP(op) case ArithOp_##op: return #op;
-#include "thorin/tables/arithoptable.h"
-        default: THORIN_UNREACHABLE;
-    }
-}
-
-const char* Cmp::op_name() const {
-    switch (tag()) {
 #define THORIN_CMP(op) case Cmp_##op: return #op;
-#include "thorin/tables/cmptable.h"
-        default: THORIN_UNREACHABLE;
-    }
-}
-
-const char* MathOp::op_name() const {
-    switch (tag()) {
 #define THORIN_MATHOP(op) case MathOp_##op: return #op;
-#include "thorin/tables/mathoptable.h"
+#include "thorin/tables/allnodes.h"
         default: THORIN_UNREACHABLE;
     }
 }
diff --git a/src/thorin/primop.h b/src/thorin/primop.h
index e82d2c218..4f9ad4efe 100644
--- a/src/thorin/primop.h
+++ b/src/thorin/primop.h
@@ -150,7 +150,6 @@ class ArithOp : public BinOp {
 public:
     const PrimType* type() const { return BinOp::type()->as<PrimType>(); }
     ArithOpTag arithop_tag() const { return (ArithOpTag) tag(); }
-    const char* op_name() const override;
 
     friend class World;
 };
@@ -165,7 +164,6 @@ class Cmp : public BinOp {
 public:
     const PrimType* type() const { return BinOp::type()->as<PrimType>(); }
     CmpTag cmp_tag() const { return (CmpTag) tag(); }
-    const char* op_name() const override;
 
     friend class World;
 };
@@ -182,7 +180,6 @@ class MathOp : public Def {
 public:
     const PrimType* type() const { return Def::type()->as<PrimType>(); }
     MathOpTag mathop_tag() const { return (MathOpTag) tag(); }
-    const char* op_name() const override;
 
     friend class World;
 };

From 50cdbdf7cfb076f2c0417a7929a97f15304ce49d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 23 Jan 2023 21:59:53 +0100
Subject: [PATCH 161/342] fix order

---
 src/thorin/def.h    |  2 +-
 src/thorin/type.cpp | 12 +++++++++++-
 src/thorin/type.h   |  1 +
 3 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/src/thorin/def.h b/src/thorin/def.h
index b3254d704..a58b2a216 100644
--- a/src/thorin/def.h
+++ b/src/thorin/def.h
@@ -157,7 +157,7 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
     //@{
     const Def* out(size_t i) const;
     bool empty() const { return ops_.empty(); }
-    void set_op(size_t i, const Def* def);
+    virtual void set_op(size_t i, const Def* def);
     void unset_op(size_t i);
     void unset_ops();
     virtual bool has_multiple_outs() const { return false; }
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index b317806fe..76f393d93 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -11,9 +11,19 @@
 
 namespace thorin {
 
-Type::Type(World& w, NodeTag tag, Defs args, Debug dbg) : Type(w, tag, w.star(), args, dbg) {}
+Type::Type(World& w, NodeTag tag, Defs args, Debug dbg) : Type(w, tag, w.star(), args, dbg) {
+    // The overridden version of set_op is ignored in the Def ctor, because according to the C++ spec, virtuals are disabled in ctors (!)
+    // So this is not actually duplicate code - for the nominal types Type::set_op will do what we want.
+    for (auto& def : args)
+        order_ = std::max(order_, def->order());
+}
 Type::Type(World& w, NodeTag tag, size_t size, Debug dbg) : Type(w, tag, w.star(), size, dbg) {}
 
+void Type::set_op(size_t i, const Def* def) {
+    Def::set_op(i, def);
+    order_ = std::max(order_, def->order());
+}
+
 Array<const Def*> types2defs(ArrayRef<const Type*> types) {
     Array<const Def*> defs(types.size());
     size_t i = 0;
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 84383e250..8af1bfda2 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -27,6 +27,7 @@ class Type : public Def, public Streamable<Type> {
 
 public:
     int order() const override { return order_; }
+    void set_op(size_t i, const Def *def) override;
     Stream& stream(Stream&) const;
 
     std::vector<const Type*> filter_type_ops() const {

From f8f309ec7b15aadbb100699e723ceb8d2982f925 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 27 Jan 2023 16:15:22 +0100
Subject: [PATCH 162/342] fix issues with params

---
 src/thorin/analyses/free_defs.cpp           |  2 +-
 src/thorin/continuation.cpp                 | 10 +++++++---
 src/thorin/continuation.h                   |  2 +-
 src/thorin/transform/cleanup_world.cpp      |  5 ++---
 src/thorin/transform/lift_builtins.cpp      |  4 ++--
 src/thorin/transform/mangle.cpp             |  3 ++-
 src/thorin/transform/partial_evaluation.cpp |  2 +-
 7 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/src/thorin/analyses/free_defs.cpp b/src/thorin/analyses/free_defs.cpp
index 31bfd28e0..bef50546e 100644
--- a/src/thorin/analyses/free_defs.cpp
+++ b/src/thorin/analyses/free_defs.cpp
@@ -23,7 +23,7 @@ DefSet free_defs(const Scope& scope, bool include_closures) {
 
     while (!queue.empty()) {
         auto def = pop(queue);
-        if (def->isa_structural()) {
+        if (def->isa_structural() && !def->isa<Param>()) {
             if (!include_closures && def->isa<Closure>()) {
                 result.emplace(def);
                 queue.push(def->op(1));
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index ccf1e6dc0..7b5bad32b 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -12,12 +12,16 @@ namespace thorin {
 
 Param::Param(World& world, const Type* type, const Continuation* continuation, size_t index, Debug dbg)
     : Def(world, Node_Param, type, { continuation }, dbg)
+    //: Def(world, Node_Param, type, 1, dbg)
     , index_(index)
-{}
+{
+    //set_op(0, continuation);
+}
 
 const Def* Param::rebuild(World& world, const Type* t, Defs defs) const {
-    assert(defs.size() == 1 && defs[0]->isa<Continuation>());
-    return world.param(t, defs[0]->as<Continuation>(), index(), debug());
+    assert(defs.size() == 1);
+    auto cont = defs[0]->as<Continuation>();
+    return cont->param(index());
 }
 
 hash_t Param::vhash() const {
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 32bbecd69..7d1e58cd4 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -30,7 +30,7 @@ class Param : public Def {
     Continuation* continuation() const { return op(0)->as_nom<Continuation>(); }
     size_t index() const { return index_; }
 
-    const Def * rebuild(World&, const Type*, Defs) const override;
+    const Def* rebuild(World&, const Type*, Defs) const override;
     bool equal(const Def*) const override;
     hash_t vhash() const override;
 private:
diff --git a/src/thorin/transform/cleanup_world.cpp b/src/thorin/transform/cleanup_world.cpp
index 3ea2f26e2..84e995d91 100644
--- a/src/thorin/transform/cleanup_world.cpp
+++ b/src/thorin/transform/cleanup_world.cpp
@@ -36,7 +36,7 @@ class Cleaner {
 };
 
 void Cleaner::eliminate_tail_rec() {
-    Scope::for_each(*world_, [&](Scope& scope) {
+    Scope::for_each(world(), [&](Scope& scope) {
         auto entry = scope.entry();
 
         bool only_tail_calls = true;
@@ -233,7 +233,7 @@ next_continuation:;
 
 void Cleaner::rebuild() {
     auto fresh_world = std::make_unique<World>(world());
-    Importer importer(*world_, *fresh_world);
+    Importer importer(world(), *fresh_world);
     importer.def_old2new_.rehash(world_->defs().capacity());
 
     for (auto&& [_, cont] : world().externals()) {
@@ -267,7 +267,6 @@ void Cleaner::verify_closedness() {
 }
 
 void Cleaner::within(const Def* def) {
-    if (def->isa<Param>()) return; // TODO remove once Params are within World's sea of nodes
     assert(&def->type()->world() == &world());
     assert_unused(world().defs().contains(def));
 }
diff --git a/src/thorin/transform/lift_builtins.cpp b/src/thorin/transform/lift_builtins.cpp
index cfc65a1de..0ececb949 100644
--- a/src/thorin/transform/lift_builtins.cpp
+++ b/src/thorin/transform/lift_builtins.cpp
@@ -65,10 +65,10 @@ void lift_pipeline(World& world) {
 
 void lift_builtins(Thorin& thorin) {
     // This must be run first
-    World& world = thorin.world();
-    lift_pipeline(world);
+    lift_pipeline(thorin.world());
 
     while (true) {
+        World& world = thorin.world();
         Continuation* cur = nullptr;
         Scope::for_each(world, [&] (const Scope& scope) {
             if (cur) return;
diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index e6523f21f..4bc733c5c 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -10,7 +10,7 @@ namespace thorin {
 const Def* Rewriter::instantiate(const Def* odef) {
     if (auto ndef = old2new.lookup(odef)) return *ndef;
 
-    if (odef->isa_structural()) {
+    if (odef->isa_structural() && !odef->isa<Param>()) {
         Array<const Def*> nops(odef->num_ops());
         for (size_t i = 0; i != odef->num_ops(); ++i)
             nops[i] = instantiate(odef->op(i));
@@ -168,6 +168,7 @@ const Def* Mangler::mangle(const Def* old_def) {
             nops[i] = mangle(old_def->op(i));
 
         auto type = old_def->type(); // TODO reduce
+        assert(!old_def->isa<Type>());
         return def2def_[old_def] = old_def->rebuild(world(), type, nops);
     }
 }
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 28d383a38..2bbf8d049 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -57,7 +57,7 @@ class CondEval {
         if (auto ndef = old2new_.lookup(odef))
             return *ndef;
 
-        if (odef->isa_structural()) {
+        if (odef->isa_structural() && !odef->isa<Param>()) {
             Array<const Def*> nops(odef->num_ops());
             for (size_t i = 0; i != odef->num_ops(); ++i)
                 nops[i] = instantiate(odef->op(i));

From d1bce61b74ab92431da545bd96b2c843c11d7e7a Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 30 Jan 2023 19:26:37 +0100
Subject: [PATCH 163/342] Fix issues with Type::stream.

This patch moves Type::stream to rec_stream.cpp and calls Type::stream
in Def::stream.
---
 src/thorin/rec_stream.cpp | 49 ++++++++++++++++++++++++++++++++++
 src/thorin/type.cpp       | 56 ---------------------------------------
 src/thorin/type.h         |  2 +-
 3 files changed, 50 insertions(+), 57 deletions(-)

diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 4980cd91d..ddf2eb067 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -69,6 +69,7 @@ void Def::dump() const { dump(0); }
 void Def::dump(size_t max) const { Stream s(std::cout); stream(s, max).endl(); }
 
 Stream& Def::stream(Stream& s) const {
+    if (isa<Type>()) return ((Type*)this)->stream(s);
     if (isa<Param>() || isa<App>() || no_dep()) return stream1(s);
     return s << unique_name();
 }
@@ -148,4 +149,52 @@ Stream& Scope::stream(Stream& s) const {
     THORIN_UNREACHABLE;
 }
 
+Stream& Type::stream(Stream& s) const {
+    if (false) {}
+    else if (isa<BottomType>()) return s.fmt("!!");
+    else if (isa<   MemType>()) return s.fmt("mem");
+    else if (isa< FrameType>()) return s.fmt("frame");
+    else if (auto t = isa<DefiniteArrayType>()) {
+        return s.fmt("[{} x {}]", t->dim(), t->elem_type());
+    } else if (auto t = isa<FnType>()) {
+        return s.fmt("fn[{, }]", t->ops());
+    } else if (auto t = isa<ClosureType>()) {
+        return s.fmt("closure [{, }]", t->ops());
+    } else if (auto t = isa<IndefiniteArrayType>()) {
+        return s.fmt("[{}]", t->elem_type());
+    } else if (auto t = isa<StructType>()) {
+        return s.fmt("struct {}", t->name());
+    } else if (auto t = isa<VariantType>()) {
+        return s.fmt("variant {}", t->name());
+    } else if (auto t = isa<TupleType>()) {
+        return s.fmt("[{, }]", t->ops());
+    } else if (auto t = isa<PtrType>()) {
+        if (t->is_vector()) s.fmt("<{} x", t->length());
+        s.fmt("{}*", t->pointee());
+        if (t->is_vector()) s.fmt(">");
+        if (t->device() != -1) s.fmt("[{}]", t->device());
+
+        switch (t->addr_space()) {
+            case AddrSpace::Global:   s.fmt("[Global]");   break;
+            case AddrSpace::Texture:  s.fmt("[Tex]");      break;
+            case AddrSpace::Shared:   s.fmt("[Shared]");   break;
+            case AddrSpace::Constant: s.fmt("[Constant]"); break;
+            default: /* ignore unknown address space */    break;
+        }
+        return s;
+    } else if (auto t = isa<PrimType>()) {
+        if (t->is_vector()) s.fmt("<{} x", t->length());
+
+        switch (t->primtype_tag()) {
+#define THORIN_ALL_TYPE(T, M) case Node_PrimType_##T: s.fmt(#T); break;
+#include "thorin/tables/primtypetable.h"
+            default: THORIN_UNREACHABLE;
+        }
+
+        if (t->is_vector()) s.fmt(">");
+        return s;
+    }
+    THORIN_UNREACHABLE;
+}
+
 }
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index 76f393d93..de6bb3bcb 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -130,62 +130,6 @@ bool PtrType::equal(const Def* other) const {
     return ptr->device() == device() && ptr->addr_space() == addr_space();
 }
 
-//------------------------------------------------------------------------------
-
-/*
- * stream
- */
-
-Stream& Type::stream(Stream& s) const {
-    if (false) {}
-    else if (isa<BottomType>()) return s.fmt("!!");
-    else if (isa<   MemType>()) return s.fmt("mem");
-    else if (isa< FrameType>()) return s.fmt("frame");
-    else if (auto t = isa<DefiniteArrayType>()) {
-        return s.fmt("[{} x {}]", t->dim(), t->elem_type());
-    } else if (auto t = isa<FnType>()) {
-        return s.fmt("fn[{, }]", t->ops());
-    } else if (auto t = isa<ClosureType>()) {
-        return s.fmt("closure [{, }]", t->ops());
-    } else if (auto t = isa<IndefiniteArrayType>()) {
-        return s.fmt("[{}]", t->elem_type());
-    } else if (auto t = isa<StructType>()) {
-        return s.fmt("struct {}", t->name());
-    } else if (auto t = isa<VariantType>()) {
-        return s.fmt("variant {}", t->name());
-    } else if (auto t = isa<TupleType>()) {
-        return s.fmt("[{, }]", t->ops());
-    } else if (auto t = isa<PtrType>()) {
-        if (t->is_vector()) s.fmt("<{} x", t->length());
-        s.fmt("{}*", t->pointee());
-        if (t->is_vector()) s.fmt(">");
-        if (t->device() != -1) s.fmt("[{}]", t->device());
-
-        switch (t->addr_space()) {
-            case AddrSpace::Global:   s.fmt("[Global]");   break;
-            case AddrSpace::Texture:  s.fmt("[Tex]");      break;
-            case AddrSpace::Shared:   s.fmt("[Shared]");   break;
-            case AddrSpace::Constant: s.fmt("[Constant]"); break;
-            default: /* ignore unknown address space */    break;
-        }
-        return s;
-    } else if (auto t = isa<PrimType>()) {
-        if (t->is_vector()) s.fmt("<{} x", t->length());
-
-        switch (t->primtype_tag()) {
-#define THORIN_ALL_TYPE(T, M) case Node_PrimType_##T: s.fmt(#T); break;
-#include "thorin/tables/primtypetable.h"
-            default: THORIN_UNREACHABLE;
-        }
-
-        if (t->is_vector()) s.fmt(">");
-        return s;
-    }
-    THORIN_UNREACHABLE;
-}
-
-//------------------------------------------------------------------------------
-
 TypeTable::TypeTable(World& world)
     : world_(world)
     , star_     (world.put<Star>((world)))
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 8af1bfda2..b2e387f78 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -16,7 +16,7 @@ class Type;
 using Types = ArrayRef<const Type*>;
 
 /// Base class for all \p Type%s.
-class Type : public Def, public Streamable<Type> {
+class Type : public Def {
 protected:
     /// Constructor for a @em structural Type.
     Type(World& w, NodeTag tag, const Type* type, Defs args, Debug dbg) : Def(w, tag, type, args, dbg) {}

From 45920464a8cc70423e82c1a1c8dedda59c6ebb2b Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 1 Feb 2023 17:10:32 +0100
Subject: [PATCH 164/342] Revert most changes on norecursion branch.

This reverts commit 9eaba5daceac48e191683268aa3d8a4ed28feeb9.
This reverts commit 72d2deed2b07094d41b01d606ff5dce6450d6ddb.
This reverts commit a9ad15284133a0d5b6a5594f71457c250b539337.
This reverts commit dc54ccaac495d898f0dfb6338953df2aa6d944a6.
This reverts commit 2c0ce6c68db2ebd70ec40c4c3229cd04f2ea7ad1.
---
 src/thorin/analyses/schedule.cpp      | 78 ++-------------------------
 src/thorin/analyses/schedule.h        |  3 --
 src/thorin/be/emitter.h               |  3 +-
 src/thorin/transform/hoist_enters.cpp | 16 ++----
 src/thorin/transform/importer.cpp     | 69 ++----------------------
 src/thorin/transform/importer.h       |  9 ----
 src/thorin/transform/mangle.cpp       | 19 +------
 7 files changed, 17 insertions(+), 180 deletions(-)

diff --git a/src/thorin/analyses/schedule.cpp b/src/thorin/analyses/schedule.cpp
index 38a67a4b6..798de7255 100644
--- a/src/thorin/analyses/schedule.cpp
+++ b/src/thorin/analyses/schedule.cpp
@@ -9,8 +9,6 @@
 #include "thorin/analyses/looptree.h"
 #include "thorin/analyses/scope.h"
 
-#include <stack>
-
 namespace thorin {
 
 Scheduler::Scheduler(const Scope& s)
@@ -46,103 +44,37 @@ Scheduler::Scheduler(const Scope& s)
     }
 }
 
-std::stack<const Def*> early_todo;
-
-Continuation* Scheduler::early(const Def * def) {
+Continuation* Scheduler::early(const Def* def) {
     if (auto cont = early_.lookup(def)) return *cont;
-
-    early_todo.push(def);
-    Continuation *return_cont = nullptr;
-    while (!early_todo.empty()) {
-        return_cont = early_intern();
-    }
-    assert(return_cont);
-    return return_cont;
-}
-
-Continuation* Scheduler::early_intern() {
-    const Def* def = early_todo.top();
-
-    if (auto cont = early_.lookup(def)) {
-        early_todo.pop();
-        return *cont;
-    }
-    if (auto param = def->isa<Param>()) {
-        early_todo.pop();
-        return early_[def] = param->continuation();
-    }
-
-    bool todo_empty = true;
-    for (auto op : def->as_structural()->ops()) {
-        if (!op->isa_nom<Continuation>() && def2uses_.find(op) != def2uses_.end()) {
-            if (!early_.lookup(op)) {
-                early_todo.push(op);
-                todo_empty = false;
-            }
-        }
-    }
-    if (!todo_empty)
-        return nullptr;
+    if (auto param = def->isa<Param>()) return early_[def] = param->continuation();
 
     auto result = scope().entry();
     for (auto op : def->as_structural()->ops()) {
         if (!op->isa_nom<Continuation>() && def2uses_.find(op) != def2uses_.end()) {
-            Continuation *cont = *early_.lookup(op);
-            assert(cont);
+            auto cont = early(op);
             if (domtree().depth(cfg(cont)) > domtree().depth(cfg(result)))
                 result = cont;
         }
     }
 
-    early_todo.pop();
     return early_[def] = result;
 }
 
-std::stack<const Def*> late_todo;
-
-Continuation* Scheduler::late(const Def * def) {
+Continuation* Scheduler::late(const Def* def) {
     if (auto cont = late_.lookup(def)) return *cont;
 
-    late_todo.push(def);
-    Continuation *return_cont = nullptr;
-    while (!late_todo.empty()) {
-        return_cont = late_intern();
-    }
-    assert(return_cont);
-    return return_cont;
-}
-
-Continuation* Scheduler::late_intern() {
-    const Def* def = late_todo.top();
-
-    if (auto cont = late_.lookup(def)) {
-        late_todo.pop();
-        return *cont;
-    }
-
     Continuation* result = nullptr;
     if (auto continuation = def->isa_nom<Continuation>()) {
         result = continuation;
     } else if (auto param = def->isa<Param>()) {
         result = param->continuation();
     } else {
-        bool todo_empty = true;
-        for (auto use : uses(def)) {
-            if (!late_.lookup(use)) {
-                late_todo.push(use);
-                todo_empty = false;
-            }
-        }
-        if (!todo_empty)
-            return nullptr;
         for (auto use : uses(def)) {
-            Continuation* cont = *late_.lookup(use);
-            assert(cont);
+            auto cont = late(use);
             result = result ? domtree().least_common_ancestor(cfg(result), cfg(cont))->continuation() : cont;
         }
     }
 
-    late_todo.pop();
     return late_[def] = result;
 }
 
diff --git a/src/thorin/analyses/schedule.h b/src/thorin/analyses/schedule.h
index 57d874274..b2cfa65aa 100644
--- a/src/thorin/analyses/schedule.h
+++ b/src/thorin/analyses/schedule.h
@@ -48,9 +48,6 @@ class Scheduler {
     DefMap<Continuation*> late_;
     DefMap<Continuation*> smart_;
     DefMap<Uses> def2uses_;
-
-    Continuation* early_intern();
-    Continuation* late_intern();
 };
 
 using Schedule = std::vector<Continuation*>;
diff --git a/src/thorin/be/emitter.h b/src/thorin/be/emitter.h
index f41d4fec8..a13ad8d49 100644
--- a/src/thorin/be/emitter.h
+++ b/src/thorin/be/emitter.h
@@ -35,7 +35,8 @@ class Emitter {
         }
 
         while (!required_defs.empty()) {
-            auto r = pop(required_defs);
+            auto r = required_defs.top();
+            required_defs.pop();
             emit_unsafe(r);
         }
 
diff --git a/src/thorin/transform/hoist_enters.cpp b/src/thorin/transform/hoist_enters.cpp
index 7595ca1fb..a51cf8d64 100644
--- a/src/thorin/transform/hoist_enters.cpp
+++ b/src/thorin/transform/hoist_enters.cpp
@@ -4,12 +4,8 @@
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/verify.h"
 
-#include <stack>
-
 namespace thorin {
 
-static std::stack<const Def*> hoist_enters_todo;
-
 static void find_enters(std::deque<const Enter*>& enters, const Def* def) {
     if (auto enter = def->isa<Enter>())
         enters.push_back(enter);
@@ -19,19 +15,13 @@ static void find_enters(std::deque<const Enter*>& enters, const Def* def) {
 
     for (auto use : def->uses()) {
         if (auto memop = use->isa<MemOp>())
-            hoist_enters_todo.push(memop);
+            find_enters(enters, memop);
     }
 }
 
 static void find_enters(std::deque<const Enter*>& enters, Continuation* continuation) {
-    if (auto mem_param = continuation->mem_param()) {
-        hoist_enters_todo.push(mem_param);
-        while (!hoist_enters_todo.empty()) {
-            auto next_item = pop(hoist_enters_todo);
-
-            find_enters(enters, next_item);
-        }
-    }
+    if (auto mem_param = continuation->mem_param())
+        find_enters(enters, mem_param);
 }
 
 static void hoist_enters(const Scope& scope) {
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index f746c9c46..4691e9795 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -8,70 +8,20 @@ const Def* Importer::import(const Def* odef) {
         return *ndef;
     }
 
-    assert(required_defs.empty());
-    required_defs.push(std::pair(odef, false));
-
-    const Def* return_def = nullptr;
-    while (!required_defs.empty()) {
-        return_def = import_nonrecursive();
-    }
-
-    assert(return_def);
-    assert(return_def == def_old2new_.lookup(odef));
-
-    return return_def;
-}
-
-const Def* Importer::import_nonrecursive() {
-    const Def* odef = required_defs.top().first;
-    bool jump_to_analyze = required_defs.top().second;
-
-    std::optional<const Def*> ndef = std::nullopt;
-    if (ndef = def_old2new_.lookup(odef)) {
-        assert(&(*ndef)->world() == &world());
-        if (!jump_to_analyze) {
-            required_defs.pop();
-            return *ndef;
-        }
-    }
-
-    if (!jump_to_analyze) {
-        required_defs.pop();
-        required_defs.push(std::pair(odef, true));
-    }
-
     if (odef == odef->world().star()) {
         def_old2new_[odef] = world().star();
-        required_defs.pop();
         return world().star();
     }
 
-    if (!def_old2new_.lookup(odef->type())) {
-        required_defs.push(std::pair(odef->type(), false));
-        return nullptr;
-    }
     auto ntype = import(odef->type())->as<Type>();
 
     Def* stub = nullptr;
     if (odef->isa_nom()) {
-        if (ndef) {
-            stub = (*ndef)->as_nom();
-        } else {
-            stub = odef->stub(world(), ntype);
-            def_old2new_[odef] = stub;
-        }
+        stub = odef->stub(world(), ntype);
+        def_old2new_[odef] = stub;
     }
 
     size_t size = odef->num_ops();
-    bool unfinished = false;
-    for (size_t i = 0; i != size; ++i)
-        if (!def_old2new_.lookup(odef->op(i))) {
-            required_defs.push(std::pair(odef->op(i), false));
-            unfinished = true;
-        }
-    if (unfinished)
-        return nullptr;
-
     Array<const Def*> nops(size);
     for (size_t i = 0; i != size; ++i) {
         assert(odef->op(i) != odef);
@@ -80,21 +30,12 @@ const Def* Importer::import_nonrecursive() {
     }
 
     if (odef->isa_structural()) {
-        if (!ndef)
-            ndef = odef->rebuild(world(), ntype, nops);
-
-        if (auto oglobal = odef->isa<Global>()) {
-            if (oglobal->is_external())
-                world().make_external(const_cast<Def*>(*ndef));
-        }
-
-        todo_ |= odef->tag() != (*ndef)->tag();
-        required_defs.pop();
-        return def_old2new_[odef] = *ndef;
+        auto ndef = odef->rebuild(world(), ntype, nops);
+        todo_ |= odef->tag() != ndef->tag();
+        return def_old2new_[odef] = ndef;
     } else {
         assert(odef->isa_nom() && stub);
         stub->rebuild_from(odef, nops);
-        required_defs.pop();
         return stub;
     }
 }
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index 0f3db273c..10941b915 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -4,10 +4,6 @@
 #include "thorin/world.h"
 #include "thorin/config.h"
 
-#include <set>
-#include <stack>
-#include <utility>
-
 namespace thorin {
 
 class Importer {
@@ -28,11 +24,6 @@ class Importer {
     const Def* import(const Def*);
     bool todo() const { return todo_; }
 
-private:
-    const Def* import_nonrecursive();
-
-    std::stack<std::pair<const Def*, bool>> required_defs;
-
 public:
     Def2Def def_old2new_;
     World& src;
diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index 3a417c34d..b184d97c5 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -5,9 +5,6 @@
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
 
-#include<queue>
-#include<utility>
-
 namespace thorin {
 
 const Def* Rewriter::instantiate(const Def* odef) {
@@ -57,8 +54,6 @@ Mangler::Mangler(const Scope& scope, Defs args, Defs lift)
     }
 }
 
-std::queue<std::pair<Continuation*, const Continuation*>> bodies_to_mangle;
-
 Continuation* Mangler::mangle() {
     // create new_entry - but first collect and specialize all param types
     std::vector<const Type*> param_types;
@@ -102,17 +97,7 @@ Continuation* Mangler::mangle() {
         new_entry()->set_filter(world().filter(new_conditions, old_entry()->filter()->debug()));
     }
 
-    bodies_to_mangle.push(std::pair(new_entry(), old_entry()));
-
-    while (!bodies_to_mangle.empty()) {
-        auto task = pop(bodies_to_mangle);
-
-        auto new_cont = task.first;
-        auto old_cont = task.second;
-
-        assert(!new_cont->has_body());
-        new_cont->set_body(mangle_body(old_cont->body()));
-    }
+    new_entry()->set_body(mangle_body(old_entry()->body()));
 
     new_entry()->verify();
 
@@ -171,7 +156,7 @@ const Def* Mangler::mangle(const Def* old_def) {
     else if (auto old_continuation = old_def->isa_nom<Continuation>()) {
         auto new_continuation = mangle_head(old_continuation);
         if (old_continuation->has_body())
-            bodies_to_mangle.push(std::pair(new_continuation, old_continuation));
+            new_continuation->set_body(mangle_body(old_continuation->body()));
         return new_continuation;
     } else if (auto param = old_def->isa<Param>()) {
         assert(within(param->continuation()));

From f5b51a12dbb675d0490bf6b03b5ab435c2c09b9b Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 3 Feb 2023 17:34:23 +0100
Subject: [PATCH 165/342] [JSON]: AddrSpace Private for ptrtpye added.

---
 src/thorin/be/json/json.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 6e8621334..d289603df 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -152,6 +152,9 @@ class TypeTable {
             case AddrSpace::Constant:
                 result["addrspace"] = "constant";
                 break;
+            case AddrSpace::Private:
+                result["addrspace"] = "private";
+                break;
             }
         } else {
             std::cerr << "type cannot be translated\n";

From d3bafe1b8aa5758f6c2b9788adc918362d54e56a Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 28 Feb 2023 14:37:45 +0100
Subject: [PATCH 166/342] Add Star to type dump.

---
 src/thorin/rec_stream.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index f48bf3039..0965e5318 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -217,6 +217,8 @@ Stream& Type::stream(Stream& s) const {
 
         if (t->is_vector()) s.fmt(">");
         return s;
+    } else if (isa<Star>()) {
+        return s.fmt("★");
     }
     THORIN_UNREACHABLE;
 }

From 82a287c0f5f5aec6973c62b91cae3474499b3490 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Fri, 26 May 2023 15:05:07 +0200
Subject: [PATCH 167/342] Add CodeGen for AMDGPU + PAL.

---
 src/thorin/CMakeLists.txt                     |  6 +-
 src/thorin/be/codegen.cpp                     | 21 +++--
 src/thorin/be/codegen.h                       |  2 +-
 .../be/llvm/{amdgpu.cpp => amdgpu_hsa.cpp}    | 14 +--
 src/thorin/be/llvm/{amdgpu.h => amdgpu_hsa.h} |  8 +-
 src/thorin/be/llvm/amdgpu_pal.cpp             | 91 +++++++++++++++++++
 src/thorin/be/llvm/amdgpu_pal.h               | 35 +++++++
 src/thorin/be/llvm/llvm.cpp                   |  3 +-
 src/thorin/be/llvm/runtime.h                  |  3 +-
 src/thorin/continuation.cpp                   |  3 +-
 src/thorin/continuation.h                     |  3 +-
 11 files changed, 162 insertions(+), 27 deletions(-)
 rename src/thorin/be/llvm/{amdgpu.cpp => amdgpu_hsa.cpp} (85%)
 rename src/thorin/be/llvm/{amdgpu.h => amdgpu_hsa.h} (77%)
 create mode 100644 src/thorin/be/llvm/amdgpu_pal.cpp
 create mode 100644 src/thorin/be/llvm/amdgpu_pal.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 944c785d2..440d7cafc 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -95,8 +95,10 @@ if(LLVM_FOUND)
         be/llvm/cpu.h
         be/llvm/llvm.cpp
         be/llvm/llvm.h
-        be/llvm/amdgpu.cpp
-        be/llvm/amdgpu.h
+        be/llvm/amdgpu_hsa.cpp
+        be/llvm/amdgpu_hsa.h
+        be/llvm/amdgpu_pal.cpp
+        be/llvm/amdgpu_pal.h
         be/llvm/nvvm.cpp
         be/llvm/nvvm.h
         be/llvm/parallel.cpp
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index db6f6d802..12247b032 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -6,7 +6,8 @@
 #if THORIN_ENABLE_LLVM
 #include "thorin/be/llvm/cpu.h"
 #include "thorin/be/llvm/nvvm.h"
-#include "thorin/be/llvm/amdgpu.h"
+#include "thorin/be/llvm/amdgpu_hsa.h"
+#include "thorin/be/llvm/amdgpu_pal.h"
 #endif
 #include "thorin/be/c/c.h"
 
@@ -87,11 +88,12 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         Continuation* imported = nullptr;
 
         static const auto backend_intrinsics = std::array {
-            std::pair { CUDA,   Intrinsic::CUDA   },
-            std::pair { NVVM,   Intrinsic::NVVM   },
-            std::pair { OpenCL, Intrinsic::OpenCL },
-            std::pair { AMDGPU, Intrinsic::AMDGPU },
-            std::pair { HLS,    Intrinsic::HLS    }
+            std::pair { CUDA,       Intrinsic::CUDA      },
+            std::pair { NVVM,       Intrinsic::NVVM      },
+            std::pair { OpenCL,     Intrinsic::OpenCL    },
+            std::pair { AMDGPU_HSA, Intrinsic::AMDGPUHSA },
+            std::pair { AMDGPU_PAL, Intrinsic::AMDGPUPAL },
+            std::pair { HLS,        Intrinsic::HLS       }
         };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
@@ -112,7 +114,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         kernels.emplace_back(continuation);
     });
 
-    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU }) {
+    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU_HSA, AMDGPU_PAL }) {
         if (!importers_[backend].world().empty()) {
             get_kernel_configs(importers_[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 auto app = use->body();
@@ -185,8 +187,9 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     hls_kernel_launch(world, hls_host_params);
 
 #if THORIN_ENABLE_LLVM
-    if (!importers_[NVVM  ].world().empty()) cgs[NVVM  ] = std::make_unique<llvm::NVVMCodeGen  >(importers_[NVVM  ].world(), kernel_config,      debug);
-    if (!importers_[AMDGPU].world().empty()) cgs[AMDGPU] = std::make_unique<llvm::AMDGPUCodeGen>(importers_[AMDGPU].world(), kernel_config, opt, debug);
+    if (!importers_[NVVM      ].world().empty()) cgs[NVVM      ] = std::make_unique<llvm::NVVMCodeGen     >(importers_[NVVM      ].world(), kernel_config,      debug);
+    if (!importers_[AMDGPU_HSA].world().empty()) cgs[AMDGPU_HSA] = std::make_unique<llvm::AMDGPUHSACodeGen>(importers_[AMDGPU_HSA].world(), kernel_config, opt, debug);
+    if (!importers_[AMDGPU_PAL].world().empty()) cgs[AMDGPU_PAL] = std::make_unique<llvm::AMDGPUPALCodeGen>(importers_[AMDGPU_PAL].world(), kernel_config, opt, debug);
 #else
     (void)opt;
 #endif
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index cd7c5689f..3a615d4ae 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -44,7 +44,7 @@ struct DeviceBackends {
     Cont2Config kernel_config;
     std::vector<Continuation*> kernels;
 
-    enum { CUDA, NVVM, OpenCL, AMDGPU, HLS, BackendCount };
+    enum { CUDA, NVVM, OpenCL, AMDGPU_HSA, AMDGPU_PAL, HLS, BackendCount };
     std::array<std::unique_ptr<CodeGen>, BackendCount> cgs;
 private:
     std::vector<Importer> importers_;
diff --git a/src/thorin/be/llvm/amdgpu.cpp b/src/thorin/be/llvm/amdgpu_hsa.cpp
similarity index 85%
rename from src/thorin/be/llvm/amdgpu.cpp
rename to src/thorin/be/llvm/amdgpu_hsa.cpp
index 1d19ff7a9..b245010ac 100644
--- a/src/thorin/be/llvm/amdgpu.cpp
+++ b/src/thorin/be/llvm/amdgpu_hsa.cpp
@@ -1,4 +1,4 @@
-#include "thorin/be/llvm/amdgpu.h"
+#include "thorin/be/llvm/amdgpu_hsa.h"
 
 #include <unordered_map> // TODO don't use std::unordered_*
 
@@ -7,7 +7,7 @@
 
 namespace thorin::llvm {
 
-AMDGPUCodeGen::AMDGPUCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
+AMDGPUHSACodeGen::AMDGPUHSACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
     : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
     , kernel_config_(kernel_config)
 {
@@ -19,7 +19,7 @@ AMDGPUCodeGen::AMDGPUCodeGen(World& world, const Cont2Config& kernel_config, int
 // Kernel code
 //------------------------------------------------------------------------------
 
-void AMDGPUCodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Function* f) {
+void AMDGPUHSACodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Function* f) {
     auto config = kernel_config_.find(continuation);
     if (config != kernel_config_.end()) {
         auto block = config->second->as<GPUKernelConfig>()->block_size();
@@ -34,7 +34,7 @@ void AMDGPUCodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Functio
     }
 }
 
-llvm::Function* AMDGPUCodeGen::emit_fun_decl(Continuation* continuation) {
+llvm::Function* AMDGPUHSACodeGen::emit_fun_decl(Continuation* continuation) {
     if (continuation->name() == "llvm.amdgcn.implicitarg.ptr")
         if (auto f = defs_.lookup(entry_); f && llvm::isa<llvm::Function>(*f))
             llvm::cast<llvm::Function>(*f)->addFnAttr("amdgpu-implicitarg-ptr");
@@ -44,13 +44,13 @@ llvm::Function* AMDGPUCodeGen::emit_fun_decl(Continuation* continuation) {
     return CodeGen::emit_fun_decl(continuation);
 }
 
-llvm::Value* AMDGPUCodeGen::emit_global(const Global* global) {
+llvm::Value* AMDGPUHSACodeGen::emit_global(const Global* global) {
     if (global->is_mutable())
         world().wdef(global, "AMDGPU: Global variable '{}' will not be synced with host", global);
     return CodeGen::emit_global(global);
 }
 
-llvm::Value* AMDGPUCodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const MathOp* mathop) {
+llvm::Value* AMDGPUHSACodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const MathOp* mathop) {
     auto make_key = [] (MathOpTag tag, unsigned bitwidth) { return (static_cast<unsigned>(tag) << 16) | bitwidth; };
     static const std::unordered_map<unsigned, std::string> ocml_functions = {
 #define MATH_FUNCTION(name) \
@@ -84,7 +84,7 @@ llvm::Value* AMDGPUCodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const Math
     return call_math_function(irbuilder, mathop, ocml_functions.at(key));
 }
 
-Continuation* AMDGPUCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
+Continuation* AMDGPUHSACodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
     return emit_reserve_shared(irbuilder, continuation, true);
 }
 
diff --git a/src/thorin/be/llvm/amdgpu.h b/src/thorin/be/llvm/amdgpu_hsa.h
similarity index 77%
rename from src/thorin/be/llvm/amdgpu.h
rename to src/thorin/be/llvm/amdgpu_hsa.h
index 640fc5b69..05fb39044 100644
--- a/src/thorin/be/llvm/amdgpu.h
+++ b/src/thorin/be/llvm/amdgpu_hsa.h
@@ -1,5 +1,5 @@
-#ifndef THORIN_BE_LLVM_AMDGPU_H
-#define THORIN_BE_LLVM_AMDGPU_H
+#ifndef THORIN_BE_LLVM_AMDGPU_HSA_H
+#define THORIN_BE_LLVM_AMDGPU_HSA_H
 
 #include "thorin/be/llvm/llvm.h"
 
@@ -11,9 +11,9 @@ namespace llvm {
 
 namespace llvm = ::llvm;
 
-class AMDGPUCodeGen : public CodeGen {
+class AMDGPUHSACodeGen : public CodeGen {
 public:
-    AMDGPUCodeGen(World& world, const Cont2Config&, int opt, bool debug);
+    AMDGPUHSACodeGen(World& world, const Cont2Config&, int opt, bool debug);
 
     const char* file_ext() const override { return ".amdgpu"; }
 
diff --git a/src/thorin/be/llvm/amdgpu_pal.cpp b/src/thorin/be/llvm/amdgpu_pal.cpp
new file mode 100644
index 000000000..77ed8c838
--- /dev/null
+++ b/src/thorin/be/llvm/amdgpu_pal.cpp
@@ -0,0 +1,91 @@
+#include "thorin/be/llvm/amdgpu_pal.h"
+
+#include <unordered_map> // TODO don't use std::unordered_*
+
+#include "thorin/primop.h"
+#include "thorin/world.h"
+
+namespace thorin::llvm {
+
+AMDGPUPALCodeGen::AMDGPUPALCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
+    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
+    , kernel_config_(kernel_config)
+{
+    module().setDataLayout("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-ni:7");
+    module().setTargetTriple("amdgcn-amd-amdpal");
+}
+
+//------------------------------------------------------------------------------
+// Kernel code
+//------------------------------------------------------------------------------
+
+void AMDGPUPALCodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Function* f) {
+    auto config = kernel_config_.find(continuation);
+    if (config != kernel_config_.end()) {
+        auto block = config->second->as<GPUKernelConfig>()->block_size();
+        if (std::get<0>(block) > 0 && std::get<1>(block) > 0 && std::get<2>(block) > 0) {
+            Array<llvm::Metadata*> annotation_values_wgsize(3);
+            auto int32_type = llvm::IntegerType::get(context(), 32);
+            annotation_values_wgsize[0] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<0>(block)));
+            annotation_values_wgsize[1] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<1>(block)));
+            annotation_values_wgsize[2] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<2>(block)));
+            f->setMetadata(llvm::StringRef("reqd_work_group_size"), llvm::MDNode::get(context(), llvm_ref(annotation_values_wgsize)));
+        }
+    }
+}
+
+llvm::Function* AMDGPUPALCodeGen::emit_fun_decl(Continuation* continuation) {
+    if (continuation->name() == "llvm.amdgcn.implicitarg.ptr")
+        if (auto f = defs_.lookup(entry_); f && llvm::isa<llvm::Function>(*f))
+            llvm::cast<llvm::Function>(*f)->addFnAttr("amdgpu-implicitarg-ptr");
+    if (continuation->name() == "__ockl_printf_begin")
+        if (auto f = defs_.lookup(entry_); f && llvm::isa<llvm::Function>(*f))
+            llvm::cast<llvm::Function>(*f)->addFnAttr("amdgpu-implicitarg-num-bytes", "32");
+    return CodeGen::emit_fun_decl(continuation);
+}
+
+llvm::Value* AMDGPUPALCodeGen::emit_global(const Global* global) {
+    if (global->is_mutable())
+        world().wdef(global, "AMDGPU: Global variable '{}' will not be synced with host", global);
+    return CodeGen::emit_global(global);
+}
+
+llvm::Value* AMDGPUPALCodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const MathOp* mathop) {
+    auto make_key = [] (MathOpTag tag, unsigned bitwidth) { return (static_cast<unsigned>(tag) << 16) | bitwidth; };
+    static const std::unordered_map<unsigned, std::string> ocml_functions = {
+#define MATH_FUNCTION(name) \
+        { make_key(MathOp_##name, 32), "__ocml_" #name "_f32" }, \
+        { make_key(MathOp_##name, 64), "__ocml_" #name "_f64" },
+        MATH_FUNCTION(fabs)
+        MATH_FUNCTION(copysign)
+        MATH_FUNCTION(round)
+        MATH_FUNCTION(floor)
+        MATH_FUNCTION(ceil)
+        MATH_FUNCTION(fmin)
+        MATH_FUNCTION(fmax)
+        MATH_FUNCTION(cos)
+        MATH_FUNCTION(sin)
+        MATH_FUNCTION(tan)
+        MATH_FUNCTION(acos)
+        MATH_FUNCTION(asin)
+        MATH_FUNCTION(atan)
+        MATH_FUNCTION(atan2)
+        MATH_FUNCTION(sqrt)
+        MATH_FUNCTION(cbrt)
+        MATH_FUNCTION(pow)
+        MATH_FUNCTION(exp)
+        MATH_FUNCTION(exp2)
+        MATH_FUNCTION(log)
+        MATH_FUNCTION(log2)
+        MATH_FUNCTION(log10)
+#undef MATH_FUNCTION
+    };
+    auto key = make_key(mathop->mathop_tag(), num_bits(mathop->type()->primtype_tag()));
+    return call_math_function(irbuilder, mathop, ocml_functions.at(key));
+}
+
+Continuation* AMDGPUPALCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
+    return emit_reserve_shared(irbuilder, continuation, true);
+}
+
+}
diff --git a/src/thorin/be/llvm/amdgpu_pal.h b/src/thorin/be/llvm/amdgpu_pal.h
new file mode 100644
index 000000000..ab2e7fcdd
--- /dev/null
+++ b/src/thorin/be/llvm/amdgpu_pal.h
@@ -0,0 +1,35 @@
+#ifndef THORIN_BE_LLVM_AMDGPU_PAL_H
+#define THORIN_BE_LLVM_AMDGPU_PAL_H
+
+#include "thorin/be/llvm/llvm.h"
+
+namespace thorin {
+
+class Load;
+
+namespace llvm {
+
+namespace llvm = ::llvm;
+
+class AMDGPUPALCodeGen : public CodeGen {
+public:
+    AMDGPUPALCodeGen(World& world, const Cont2Config&, int opt, bool debug);
+
+    const char* file_ext() const override { return ".amdgpu"; }
+
+protected:
+    void emit_fun_decl_hook(Continuation*, llvm::Function*) override;
+    llvm::Function* emit_fun_decl(Continuation*) override;
+    llvm::Value* emit_global(const Global*) override;
+    llvm::Value* emit_mathop(llvm::IRBuilder<>&, const MathOp*) override;
+    Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
+    std::string get_alloc_name() const override { return "malloc"; }
+
+    const Cont2Config& kernel_config_;
+};
+
+}
+
+}
+
+#endif
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 49d923bf2..d33300f15 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1158,7 +1158,8 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
         case Intrinsic::CUDA:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation);
         case Intrinsic::NVVM:        return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
         case Intrinsic::OpenCL:      return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
-        case Intrinsic::AMDGPU:      return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
+        case Intrinsic::AMDGPUHSA:   return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
+        case Intrinsic::AMDGPUPAL:   return runtime_->emit_host_code(*this, irbuilder, Runtime::PAL_PLATFORM,    ".amdgpu", continuation);
         case Intrinsic::HLS:         return emit_hls(irbuilder, continuation);
         case Intrinsic::Parallel:    return emit_parallel(irbuilder, continuation);
         case Intrinsic::Fibers:      return emit_fibers(irbuilder, continuation);
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index 4ebf5f5ef..aed5736b3 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -22,7 +22,8 @@ class Runtime {
         CPU_PLATFORM,
         CUDA_PLATFORM,
         OPENCL_PLATFORM,
-        HSA_PLATFORM
+        HSA_PLATFORM,
+        PAL_PLATFORM
     };
 
     /// Emits a call to anydsl_launch_kernel.
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index f729937f8..37724ad1a 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -214,7 +214,8 @@ void Continuation::set_intrinsic() {
     if      (name() == "cuda")           attributes().intrinsic = Intrinsic::CUDA;
     else if (name() == "nvvm")           attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")         attributes().intrinsic = Intrinsic::OpenCL;
-    else if (name() == "amdgpu")         attributes().intrinsic = Intrinsic::AMDGPU;
+    else if (name() == "amdgpu_hsa")     attributes().intrinsic = Intrinsic::AMDGPUHSA;
+    else if (name() == "amdgpu_pal")     attributes().intrinsic = Intrinsic::AMDGPUPAL;
     else if (name() == "hls")            attributes().intrinsic = Intrinsic::HLS;
     else if (name() == "parallel")       attributes().intrinsic = Intrinsic::Parallel;
     else if (name() == "fibers")         attributes().intrinsic = Intrinsic::Fibers;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 9f34fe992..6fa0329ed 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -90,7 +90,8 @@ enum class Intrinsic : uint8_t {
     CUDA = AcceleratorBegin,    ///< Internal CUDA-Backend.
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
-    AMDGPU,                     ///< Internal AMDGPU-Backend.
+    AMDGPUHSA,                  ///< Internal AMDGPU-HSA-Backend.
+    AMDGPUPAL,                  ///< Internal AMDGPU-PAL-Backend.
     HLS,                        ///< Internal HLS-Backend.
     Parallel,                   ///< Internal Parallel-CPU-Backend.
     Fibers,                     ///< Internal Parallel-CPU-Backend using resumable fibers.

From 4eefa459a5e90b88ec747419be43c8c96d5e5ad7 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Tue, 13 Jun 2023 11:11:48 +0200
Subject: [PATCH 168/342] PAL: set kernel calling convention to AMDGPU_CS.

---
 src/thorin/be/llvm/amdgpu_pal.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/llvm/amdgpu_pal.cpp b/src/thorin/be/llvm/amdgpu_pal.cpp
index 77ed8c838..3fa0d2a04 100644
--- a/src/thorin/be/llvm/amdgpu_pal.cpp
+++ b/src/thorin/be/llvm/amdgpu_pal.cpp
@@ -8,7 +8,7 @@
 namespace thorin::llvm {
 
 AMDGPUPALCodeGen::AMDGPUPALCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
+    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_CS, opt, debug)
     , kernel_config_(kernel_config)
 {
     module().setDataLayout("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-ni:7");

From ec5691c417523b3802070f0ae6733a9feab38dd3 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Tue, 1 Aug 2023 17:45:29 +0200
Subject: [PATCH 169/342] PAL: set function calling convention to AMDGPU_Gfx.

---
 src/thorin/be/llvm/amdgpu_pal.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/llvm/amdgpu_pal.cpp b/src/thorin/be/llvm/amdgpu_pal.cpp
index 3a7459d55..76a8b909a 100644
--- a/src/thorin/be/llvm/amdgpu_pal.cpp
+++ b/src/thorin/be/llvm/amdgpu_pal.cpp
@@ -8,7 +8,7 @@
 namespace thorin::llvm {
 
 AMDGPUPALCodeGen::AMDGPUPALCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_CS, opt, debug)
+    : CodeGen(world, llvm::CallingConv::AMDGPU_Gfx, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_CS, opt, debug)
     , kernel_config_(kernel_config)
 {
     module().setDataLayout("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7");

From 5098a739efd2e596443710bafaaf3c8eb6b79729 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Tue, 1 Aug 2023 17:53:51 +0200
Subject: [PATCH 170/342] PAL: remove HSA code.

---
 src/thorin/be/llvm/amdgpu_pal.cpp | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/src/thorin/be/llvm/amdgpu_pal.cpp b/src/thorin/be/llvm/amdgpu_pal.cpp
index 76a8b909a..9ea8562c2 100644
--- a/src/thorin/be/llvm/amdgpu_pal.cpp
+++ b/src/thorin/be/llvm/amdgpu_pal.cpp
@@ -35,12 +35,6 @@ void AMDGPUPALCodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Func
 }
 
 llvm::Function* AMDGPUPALCodeGen::emit_fun_decl(Continuation* continuation) {
-    if (continuation->name() == "llvm.amdgcn.implicitarg.ptr")
-        if (auto f = defs_.lookup(entry_); f && llvm::isa<llvm::Function>(*f))
-            llvm::cast<llvm::Function>(*f)->addFnAttr("amdgpu-implicitarg-ptr");
-    if (continuation->name() == "__ockl_printf_begin")
-        if (auto f = defs_.lookup(entry_); f && llvm::isa<llvm::Function>(*f))
-            llvm::cast<llvm::Function>(*f)->addFnAttr("amdgpu-implicitarg-num-bytes", "32");
     return CodeGen::emit_fun_decl(continuation);
 }
 

From 9d212b634a2ed4ee26b6096acaeb3d88b79f1cdb Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 4 Sep 2023 17:06:48 +0200
Subject: [PATCH 171/342] Fix convert_closure_type in llvm be.

---
 src/thorin/be/llvm/llvm.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 7002c0d80..dcbe63a6c 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -250,14 +250,14 @@ llvm::FunctionType* CodeGen::convert_closure_type(const Type* type) {
     auto fn = type->as<FnType>();
     llvm::Type* ret = nullptr;
     std::vector<llvm::Type*> ops;
-    for (auto op : fn->ops()) {
-        if (op->isa<MemType>() || op == world().unit()) continue;
+    for (auto op : fn->types()) {
+        if (op->isa<MemType>() || op == world().unit_type()) continue;
         auto fn = op->isa<FnType>();
         if (fn && !op->isa<ClosureType>()) {
             assert(!ret && "only one 'return' supported");
             std::vector<llvm::Type*> ret_types;
-            for (auto fn_op : fn->ops()) {
-                if (fn_op->isa<MemType>() || fn_op == world().unit()) continue;
+            for (auto fn_op : fn->types()) {
+                if (fn_op->isa<MemType>() || fn_op == world().unit_type()) continue;
                 ret_types.push_back(convert(fn_op));
             }
             if (ret_types.size() == 0)      ret = llvm::Type::getVoidTy(context());

From 135aa4e6a7c651465a110365e5410da6157408e1 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 14 Feb 2023 22:05:22 +0100
Subject: [PATCH 172/342] Fix CMake to allow for cmake based meta repository.

Remove CMAKE_CONFIGURATION_TYPES to ensure LLVM can be build.
---
 CMakeLists.txt               | 10 +++++++---
 cmake/thorin-config.cmake.in |  2 +-
 src/thorin/CMakeLists.txt    |  4 ++--
 3 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3d8f55c1e..616e2877a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -3,7 +3,7 @@ cmake_minimum_required(VERSION 3.13.4 FATAL_ERROR)
 project(Thorin)
 
 set(PACKAGE_VERSION "0.3.9")
-set(CMAKE_CONFIGURATION_TYPES "Debug;Release" CACHE STRING "limited config" FORCE)
+#set(CMAKE_CONFIGURATION_TYPES "Debug;Release" CACHE STRING "limited config" FORCE)
 set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS 1)
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
@@ -26,9 +26,13 @@ find_package(Half REQUIRED)
 message(STATUS "Building with Half library from ${Half_INCLUDE_DIRS}.")
 
 # find json package for json output support.
-find_package(nlohmann_json 3.2.0)
-if(nlohmann_json_FOUND)
+if(TARGET nlohmann_json)
     set(THORIN_ENABLE_JSON TRUE)
+else()
+    find_package(nlohmann_json 3.2.0 QUIET)
+    if(nlohmann_json_FOUND)
+        set(THORIN_ENABLE_JSON TRUE)
+    endif()
 endif()
 
 # check for possible llvm extension
diff --git a/cmake/thorin-config.cmake.in b/cmake/thorin-config.cmake.in
index 121db87e8..60fa5ccc4 100644
--- a/cmake/thorin-config.cmake.in
+++ b/cmake/thorin-config.cmake.in
@@ -28,7 +28,7 @@ list(APPEND CMAKE_MODULE_PATH "${Thorin_ROOT_DIR}/cmake/modules")
 find_path(Half_DIR NAMES half.hpp PATHS ${Half_DIR} $ENV{Half_DIR} "@Half_DIR@" "@Half_INCLUDE_DIR@")
 find_package(Half REQUIRED)
 
-set(Thorin_HAS_JSON_SUPPORT @nlohmann_json_FOUND@)
+set(Thorin_HAS_JSON_SUPPORT @THORIN_ENABLE_JSON@)
 set(Thorin_HAS_LLVM_SUPPORT @LLVM_FOUND@)
 set(Thorin_HAS_RV_SUPPORT @RV_FOUND@)
 set(Thorin_HAS_SHADY_SUPPORT @shady_FOUND@)
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index c1adc632e..07e28605e 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -113,7 +113,7 @@ if (shady_FOUND)
     )
 endif()
 
-if(nlohmann_json_FOUND)
+if(THORIN_ENABLE_JSON)
     list(APPEND THORIN_SOURCES
         be/json/json.cpp
         be/json/json.h
@@ -139,6 +139,6 @@ if (shady_FOUND)
     target_link_libraries(thorin PRIVATE shady::shady)
 endif()
 
-if(nlohmann_json_FOUND)
+if(THORIN_ENABLE_JSON)
     target_link_libraries(thorin PRIVATE nlohmann_json::nlohmann_json)
 endif()

From f8177297eb6fa850e48192e7b7c85f36f107b92e Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 16 Feb 2023 16:20:05 +0100
Subject: [PATCH 173/342] Shady as global subproject needs special support.

---
 CMakeLists.txt               | 11 ++++++++---
 cmake/thorin-config.cmake.in |  2 +-
 src/thorin/CMakeLists.txt    | 10 +++++++---
 3 files changed, 16 insertions(+), 7 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 616e2877a..99f4f2948 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -54,9 +54,14 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
-find_package(shady CONFIG)
-if (shady_FOUND)
-    message(STATUS "Found shady at ${shady_DIR}")
+if (NOT TARGET shady)
+    find_package(shady CONFIG)
+    if (shady_FOUND)
+        message(STATUS "Found shady at ${shady_DIR}")
+        set(THORIN_ENABLE_SHADY TRUE)
+    endif()
+else()
+    export(TARGETS shady api FILE ${CMAKE_BINARY_DIR}/share/anydsl/cmake/shady-exports.cmake)
     set(THORIN_ENABLE_SHADY TRUE)
 endif()
 
diff --git a/cmake/thorin-config.cmake.in b/cmake/thorin-config.cmake.in
index 60fa5ccc4..3a53346dd 100644
--- a/cmake/thorin-config.cmake.in
+++ b/cmake/thorin-config.cmake.in
@@ -31,7 +31,7 @@ find_package(Half REQUIRED)
 set(Thorin_HAS_JSON_SUPPORT @THORIN_ENABLE_JSON@)
 set(Thorin_HAS_LLVM_SUPPORT @LLVM_FOUND@)
 set(Thorin_HAS_RV_SUPPORT @RV_FOUND@)
-set(Thorin_HAS_SHADY_SUPPORT @shady_FOUND@)
+set(Thorin_HAS_SHADY_SUPPORT @THORIN_ENABLE_SHADY@)
 set(AnyDSL_LLVM_LINK_SHARED @AnyDSL_LLVM_LINK_SHARED@)
 
 if(Thorin_HAS_LLVM_SUPPORT)
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 07e28605e..f459e9ca8 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -107,7 +107,7 @@ if(LLVM_FOUND)
     )
 endif()
 
-if (shady_FOUND)
+if (THORIN_ENABLE_SHADY)
     list(APPEND THORIN_SOURCES
         be/shady/shady.cpp
     )
@@ -135,8 +135,12 @@ if(LLVM_FOUND)
     llvm_config(thorin ${AnyDSL_LLVM_LINK_SHARED} ${Thorin_LLVM_COMPONENTS})
 endif()
 
-if (shady_FOUND)
-    target_link_libraries(thorin PRIVATE shady::shady)
+if (THORIN_ENABLE_SHADY)
+    if (shady_FOUND)
+        target_link_libraries(thorin PRIVATE shady::shady)
+    else()
+        target_link_libraries(thorin PRIVATE shady)
+    endif()
 endif()
 
 if(THORIN_ENABLE_JSON)

From 91703e0c10855b2eec93c24b0e747b9df307e4fa Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 27 Jan 2023 18:19:12 +0100
Subject: [PATCH 174/342] Add support for anyopt based linking by adding a CC
 "internal".

Externals with CC == Internal will only be exported as such in the json
backend. The intended use is to be able to combine multiple json files
during compilation.
---
 src/thorin/be/c/c.cpp                       | 2 +-
 src/thorin/be/json/json.cpp                 | 8 ++++++--
 src/thorin/be/llvm/llvm.cpp                 | 7 ++++++-
 src/thorin/be/shady/shady.cpp               | 7 ++++++-
 src/thorin/continuation.h                   | 1 +
 src/thorin/rec_stream.cpp                   | 8 ++++++--
 src/thorin/transform/partial_evaluation.cpp | 2 +-
 7 files changed, 27 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 62131e68a..f46c1f458 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -358,7 +358,7 @@ void CCodeGen::emit_module() {
     Scope::for_each(world(), [&] (const Scope& scope) {
         if (scope.entry()->name() == "hls_top")
             hls_top = scope.entry();
-        else
+        else if (scope.entry()->cc() != CC::Internal)
             emit_scope(scope);
     });
     if (hls_top) {
diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index d289603df..e8ead38c8 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -230,8 +230,12 @@ class DefTable {
                 forward_decl["type"] = "continuation";
                 forward_decl["fn_type"] = type;
                 forward_decl["arg_names"] = arg_names;
-                if (cont->is_external())
-                    forward_decl["external"] = cont->name();
+                if (cont->is_external()) {
+                    if (cont->cc() == CC::Internal)
+                        forward_decl["internal"] = cont->name();
+                    else
+                        forward_decl["external"] = cont->name();
+                }
                 if (cont->cc() == CC::DeviceHostCode)
                     forward_decl["device"] = cont->name();
                 decl_table.push_back(forward_decl);
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index dcbe63a6c..c21a4fdca 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -309,7 +309,12 @@ CodeGen::emit_module() {
         }
     }
 
-    Scope::for_each(world(), [&] (const Scope& scope) { emit_scope(scope); });
+    Scope::for_each(world(), [&] (const Scope& scope) {
+        if(scope.entry()->cc() == CC::Internal) {
+            return;
+        }
+        emit_scope(scope);
+    });
 
     if (debug()) dibuilder_.finalize();
 
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index cd79f70c1..4654d7ac9 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -18,7 +18,12 @@ void CodeGen::emit_stream(std::ostream& out) {
     arena = shady::new_ir_arena(config);
     module = shady::new_module(arena, world().name().c_str());
 
-    Scope::for_each(world(), [&](const Scope& scope) { emit_scope(scope); });
+    Scope::for_each(world(), [&](const Scope& scope) {
+        if(scope.entry()->cc() == CC::Internal) {
+            return;
+        }
+        emit_scope(scope);
+    });
 
     char* bufptr;
     size_t size;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 155e899f0..010d9bd8f 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -87,6 +87,7 @@ enum class CC : uint8_t {
     C,              ///< C calling convention.
     Device,         ///< Device calling convention. These are special functions only available on a particular device.
     DeviceHostCode, ///< Calling convention to denote continuations that are generated as device code.
+    Internal,       ///< External, but only for linking with artic or anyopt.
 };
 
 enum class Intrinsic : uint8_t {
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 0965e5318..343ad4bd6 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -48,8 +48,12 @@ void RecStreamer::run() {
         auto cont = conts.pop();
         s.endl().endl();
 
-        if (cont->world().is_external(cont))
-            s.fmt("extern ");
+        if (cont->world().is_external(cont)) {
+            if (cont->attributes().cc == CC::Internal)
+                s.fmt("intern ");
+            else
+                s.fmt("extern ");
+        }
 
         if (cont->has_body()) {
             std::vector<std::string> param_names;
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 5a18d3495..08c8d07ae 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -81,7 +81,7 @@ class CondEval {
             return true;
         }
 
-        return (!callee_->is_exported() && callee_->can_be_inlined()) || is_one(instantiate(filter(i)));
+        return ((!callee_->is_exported() || callee_->attributes().cc == CC::Internal) && callee_->can_be_inlined()) || is_one(instantiate(filter(i)));
         //return is_one(instantiate(filter(i)));
     }
 

From c292516ed178b9c934dc91e81845aa3652d48d07 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 21 Feb 2023 16:25:29 +0100
Subject: [PATCH 175/342] [JSON] divide emit_stream and emit_json for better
 runtime usability.

---
 src/thorin/be/json/json.cpp | 10 +++++++---
 src/thorin/be/json/json.h   |  1 +
 2 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index e8ead38c8..2271625d0 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -641,9 +641,7 @@ class DefTable {
     }
 };
 
-void CodeGen::emit_stream(std::ostream& stream) {
-    json j;
-
+void CodeGen::emit_json(json& j) {
     j["module"] = world().name();
     if (target_triple != "")
         j["target_triple"] = target_triple;
@@ -666,6 +664,12 @@ void CodeGen::emit_stream(std::ostream& stream) {
     j["defs"] = def_table.decl_table;
     for (auto it : def_table.def_table)
         j["defs"] += it;
+}
+
+void CodeGen::emit_stream(std::ostream& stream) {
+    json j;
+
+    emit_json(j);
     
     Stream s(stream);
     s << j.dump(2) << "\n";
diff --git a/src/thorin/be/json/json.h b/src/thorin/be/json/json.h
index db976094b..ba11199ad 100644
--- a/src/thorin/be/json/json.h
+++ b/src/thorin/be/json/json.h
@@ -24,6 +24,7 @@ class CodeGen : public thorin::CodeGen {
         , target_attr(target_attr)
     {}
     
+    void emit_json(json& j);
     void emit_stream(std::ostream& stream) override;
 
     const char* file_ext() const override {

From 01a4a46e48be38866824fa07b37e2329a0c62a41 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 14 Mar 2023 17:48:24 +0100
Subject: [PATCH 176/342] Shady: Emit additional ops.

---
 src/thorin/be/shady/shady.cpp | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 4654d7ac9..a2d83740f 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -270,6 +270,8 @@ void CodeGen::emit_epilogue(Continuation* cont) {
     NodeVec args;
     for (auto& arg : body->args()) {
         if (convert(arg->type()) == nullptr) {
+            if (is_mem(arg))
+                emit_unsafe(arg);
             args.push_back(nullptr);
         } else if (auto target = arg->isa_nom<Continuation>(); target && target->is_basicblock()) {
             // Emitting basic blocks as values isn't legal - but for convenience we'll put them in our list.
@@ -382,7 +384,8 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             type_arguments.push_back(convert(type_arg));
         payload.operands = vec2nodes(operands);
         payload.type_arguments = vec2nodes(type_arguments);
-        return shady::first(shady::bind_instruction(bb.builder, shady::prim_op(arena, payload)));
+        auto ret = shady::bind_instruction(bb.builder, shady::prim_op(arena, payload));
+        return ret.count ? shady::first(ret) : nullptr;
     };
 
     if (auto prim_lit = def->isa<PrimLit>()) {
@@ -409,9 +412,9 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             contents.push_back(emit(e));
         }
         shady::ArrType payload;
-        const shady::Type* arr_type = shady::arr_type(arena, payload);
         payload.element_type = convert(arr->elem_type());
         payload.size = shady::int32_literal(arena, contents.size());
+        const shady::Type* arr_type = shady::arr_type(arena, payload);
         v = shady::composite(arena, arr_type, vec2nodes(contents));
     } else if (auto cmp = def->isa<Cmp>()) {
         switch (cmp->cmp_tag()) {
@@ -435,10 +438,21 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             case ArithOp_shl: v = mk_primop(shady::Op::lshift_op, { arith->lhs(), arith->rhs() }); break;
             case ArithOp_shr: v = mk_primop(shady::Op::rshift_logical_op, { arith->lhs(), arith->rhs() }); break;
         }
+    } else if (auto store = def->isa<Store>()) {
+        mk_primop(shady::Op::store_op, { store->ptr(), store->val() });
+        defs_[def] = nullptr;
+        return nullptr;
+    } else if (auto lea = def->isa<LEA>()) {
+        v = mk_primop(shady::Op::lea_op, { lea->ptr(), world().zero(lea->index()->type()), lea->index() });
     } else if (auto param = def->isa<Param>()) {
         assert(param->type() == world().mem_type());
         defs_[def] = nullptr;
         return nullptr;
+    } else if (auto bitcast = def->isa<Bitcast>()) {
+        v = emit(bitcast->from());
+    } else {
+        def->dump();
+        THORIN_UNREACHABLE;
     }
     assert(v && shady::is_value(v));
     defs_[def] = v;

From 2d400e8248c38863f91416f7e4a0cbc502e534b6 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 23 Jan 2023 17:15:41 +0100
Subject: [PATCH 177/342] Fixed long memop chains in json emitter.

---
 src/thorin/be/json/json.cpp | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 2271625d0..2bd09fb21 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -183,6 +183,33 @@ class DefTable {
             return it->second;
         }
 
+        if (def->isa<MemOp>()) {
+            std::stack<const Def*> required_defs;
+            std::queue<const Def*> todo;
+            todo.push(def);
+
+            while (!todo.empty()) {
+                auto def = todo.front();
+                todo.pop();
+                if (known_defs.lookup(def)) continue;
+
+                if (auto memop = def->isa<MemOp>()) {
+                    todo.push(memop->mem());
+                    required_defs.push(memop->mem());
+                } else if (auto extract = def->isa<Extract>()) {
+                    if (is_mem(extract)) {
+                        todo.push(extract->agg());
+                        required_defs.push(extract->agg());
+                    }
+                }
+            }
+
+            while (!required_defs.empty()) {
+                auto r = pop(required_defs);
+                translate_def(r);
+            }
+        }
+
         json result;
         if (auto cont = def->isa<Continuation>()) {
             if (cont->is_intrinsic()) {

From 8d4f09b413342f2780adee585a5e100e94cc286c Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 2 Mar 2023 12:05:22 +0100
Subject: [PATCH 178/342] Support stack size changes.

---
 CMakeLists.txt         |  3 +++
 src/thorin/config.h.in |  1 +
 src/thorin/world.cpp   | 21 +++++++++++++++++++++
 src/thorin/world.h     |  3 +++
 4 files changed, 28 insertions(+)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 99f4f2948..7939482bb 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -54,6 +54,9 @@ else()
     message(STATUS "Building without LLVM and RV. Specify LLVM_DIR to compile with LLVM.")
 endif()
 
+include(CheckIncludeFile)
+check_include_file(sys/resource.h THORIN_ENABLE_RLIMITS)
+
 if (NOT TARGET shady)
     find_package(shady CONFIG)
     if (shady_FOUND)
diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index bfbe79d18..e8f2dfc42 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -8,5 +8,6 @@
 #cmakedefine01 THORIN_ENABLE_JSON
 #cmakedefine01 THORIN_ENABLE_RV
 #cmakedefine01 THORIN_ENABLE_SHADY
+#cmakedefine01 THORIN_ENABLE_RLIMITS
 
 #endif
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 33ff8d2fa..8574c8eda 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -12,6 +12,10 @@
 #include <cmath>
 #include <execinfo.h>
 
+#ifdef THORIN_ENABLE_RLIMITS
+#include <sys/resource.h>
+#endif
+
 #include "thorin/def.h"
 #include "thorin/primop.h"
 #include "thorin/continuation.h"
@@ -1310,4 +1314,21 @@ void Thorin::opt() {
     RUN_PASS(codegen_prepare(world()))
 }
 
+bool Thorin::ensure_stack_size(size_t new_size) {
+#ifdef THORIN_ENABLE_RLIMITS
+    struct rlimit rl;
+    int result = getrlimit(RLIMIT_STACK, &rl);
+    if(result != 0) return false;
+
+    rl.rlim_cur = new_size;
+    result = setrlimit(RLIMIT_STACK, &rl);
+    if(result != 0) return false;
+
+    return true;
+#else
+    return false;
+#endif
+}
+
+
 }
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 4dc253ee5..56dda1c30 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -401,6 +401,9 @@ class Thorin {
     /// Performs dead code, unreachable code and unused type elimination.
     void cleanup();
     void opt();
+
+    bool ensure_stack_size(size_t new_size);
+
 private:
     std::unique_ptr<World> world_;
 };

From ed05cc01568b720c19048b09e8a0c2040ba405b4 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 1 Aug 2023 12:24:36 +0200
Subject: [PATCH 179/342] Fix warnings across the compiler.

---
 src/thorin/be/c/c.cpp                      |  2 +-
 src/thorin/be/llvm/llvm.cpp                |  2 +-
 src/thorin/continuation.cpp                |  2 +-
 src/thorin/def.cpp                         |  2 +-
 src/thorin/rec_stream.cpp                  |  2 +-
 src/thorin/transform/hls_kernel_launch.cpp |  2 +-
 src/thorin/transform/split_slots.cpp       |  4 +++-
 src/thorin/type.cpp                        | 22 +++++++++++-----------
 src/thorin/type.h                          |  1 +
 src/thorin/world.cpp                       |  4 +++-
 10 files changed, 24 insertions(+), 19 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index f46c1f458..61e47d3e0 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -976,7 +976,7 @@ std::string CCodeGen::emit_def(BB* bb, const Def* def) {
 
     if (is_unit(def)) return "";
     else if (auto bin = def->isa<BinOp>()) {
-        const char* op;
+        const char* op = "";
         if (auto cmp = bin->isa<Cmp>()) {
             switch (cmp->cmp_tag()) {
                 case Cmp_eq: op = "=="; break;
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index c21a4fdca..045763c3b 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -204,7 +204,7 @@ llvm::Type* CodeGen::convert(const Type* type) {
             size_t max_align = 0, max_size = 0;
 
             auto layout = module().getDataLayout();
-            llvm::Type* max_align_type;
+            llvm::Type* max_align_type = llvm::Type::getVoidTy(context());
             for (auto op : variant_type->types()) {
                 auto op_type = convert(op);
                 size_t size  = layout.getTypeAllocSize(op_type);
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 509cdfcd7..d82041aa8 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -18,7 +18,7 @@ Param::Param(World& world, const Type* type, const Continuation* continuation, s
     //set_op(0, continuation);
 }
 
-const Def* Param::rebuild(World& world, const Type* t, Defs defs) const {
+const Def* Param::rebuild(World&, const Type*, Defs defs) const {
     assert(defs.size() == 1);
     auto cont = defs[0]->as<Continuation>();
     return cont->param(index());
diff --git a/src/thorin/def.cpp b/src/thorin/def.cpp
index eb74c00a4..8915d42ba 100644
--- a/src/thorin/def.cpp
+++ b/src/thorin/def.cpp
@@ -138,7 +138,7 @@ bool is_minus_zero(const Def* def) {
     return false;
 }
 
-void Def::rebuild_from(const Def* old, Defs new_ops) {
+void Def::rebuild_from(const Def*, Defs new_ops) {
     assert(new_ops.size() == num_ops());
     for (size_t i = 0; i < num_ops(); i++)
         set_op(i, new_ops[i]);
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 343ad4bd6..29ae570cc 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -173,7 +173,7 @@ Stream& World::stream(Stream& s) const {
     return s.endl();
 }
 
-Stream& Scope::stream(Stream& s) const {
+Stream& Scope::stream(Stream&) const {
     THORIN_UNREACHABLE;
 }
 
diff --git a/src/thorin/transform/hls_kernel_launch.cpp b/src/thorin/transform/hls_kernel_launch.cpp
index cab3bd7e3..413c24506 100644
--- a/src/thorin/transform/hls_kernel_launch.cpp
+++ b/src/thorin/transform/hls_kernel_launch.cpp
@@ -87,7 +87,7 @@ void hls_kernel_launch(World& world, DeviceParams& device_params) {
             if (auto hls_callee = has_hls_callee(block)) {
                 auto cont_mem_obj = block->mem_param();
                 auto callee_continuation = hls_callee->isa_nom<Continuation>();
-                Continuation* last_hls_cont;
+                Continuation* last_hls_cont = nullptr;
                 if (!last_hls_found) {
                     // TODO I'm at a loss for what is intended here. This is an assignment - not a check, the net result
                     // is the _only the first_ block with an HLS callee will enter this, which means the first block in the schedule
diff --git a/src/thorin/transform/split_slots.cpp b/src/thorin/transform/split_slots.cpp
index 1f81e763d..87f643352 100644
--- a/src/thorin/transform/split_slots.cpp
+++ b/src/thorin/transform/split_slots.cpp
@@ -13,6 +13,7 @@ struct IndexHash {
     static u32 sentinel() { return 0xFFFFFFFF; }
 };
 
+#if 0
 static void split(const Slot* slot) {
     auto array_type = slot->alloced_type()->as<DefiniteArrayType>();
     auto dim = array_type->dim();
@@ -67,8 +68,9 @@ static bool can_split(const Slot* slot) {
 
     return true;
 }
+#endif
 
-static bool split_slots(const Scope& scope) {
+static bool split_slots(const Scope& /* scope */) {
     bool todo = false;
     // TODO
 #if 0
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index de6bb3bcb..b940e5f11 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -46,20 +46,20 @@ Array<const Type*> defs2types(ArrayRef<const Def*> defs) {
  * rebuild
  */
 
-const Type* NominalType::rebuild(World& w, const Type* t, Defs o) const {
+const Type* NominalType::rebuild(World&  , const Type*  , Defs  ) const {
     THORIN_UNREACHABLE;
 }
 
-const Type* BottomType         ::rebuild(World& w, const Type* t, Defs o) const { return w.bottom_type(); }
-const Type* ClosureType        ::rebuild(World& w, const Type* t, Defs o) const { return w.closure_type(defs2types(o)); }
-const Type* DefiniteArrayType  ::rebuild(World& w, const Type* t, Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
-const Type* FnType             ::rebuild(World& w, const Type* t, Defs o) const { return w.fn_type(defs2types(o)); }
-const Type* FrameType          ::rebuild(World& w, const Type* t, Defs o) const { return w.frame_type(); }
-const Type* IndefiniteArrayType::rebuild(World& w, const Type* t, Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
-const Type* MemType            ::rebuild(World& w, const Type* t, Defs o) const { return w.mem_type(); }
-const Type* PrimType           ::rebuild(World& w, const Type* t, Defs o) const { return w.prim_type(primtype_tag(), length()); }
-const Type* PtrType            ::rebuild(World& w, const Type* t, Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
-const Type* TupleType          ::rebuild(World& w, const Type* t, Defs o) const { return w.tuple_type(defs2types(o)); }
+const Type* BottomType         ::rebuild(World& w, const Type*  , Defs  ) const { return w.bottom_type(); }
+const Type* ClosureType        ::rebuild(World& w, const Type*  , Defs o) const { return w.closure_type(defs2types(o)); }
+const Type* DefiniteArrayType  ::rebuild(World& w, const Type*  , Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
+const Type* FnType             ::rebuild(World& w, const Type*  , Defs o) const { return w.fn_type(defs2types(o)); }
+const Type* FrameType          ::rebuild(World& w, const Type*  , Defs  ) const { return w.frame_type(); }
+const Type* IndefiniteArrayType::rebuild(World& w, const Type*  , Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
+const Type* MemType            ::rebuild(World& w, const Type*  , Defs  ) const { return w.mem_type(); }
+const Type* PrimType           ::rebuild(World& w, const Type*  , Defs  ) const { return w.prim_type(primtype_tag(), length()); }
+const Type* PtrType            ::rebuild(World& w, const Type*  , Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
+const Type* TupleType          ::rebuild(World& w, const Type*  , Defs o) const { return w.tuple_type(defs2types(o)); }
 
 /*
  * stub
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 96cdf49fa..1018a95c3 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -97,6 +97,7 @@ class NominalType : public Type {
 
 public:
     Symbol name() const { return name_; }
+    using Type::op_name; //Would be hidden otherwise.
     Symbol op_name(size_t i) const { return op_names_[i]; }
     void set_op_name(size_t i, Symbol name) const {
         const_cast<NominalType*>(this)->op_names_[i] = name;
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 8574c8eda..51de82e52 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -140,7 +140,9 @@ const Def* World::arithop(ArithOpTag tag, const Def* a, const Def* b, Debug dbg)
                     }
                 case ArithOp_mul:
                     switch (type) {
-#define THORIN_ALL_TYPE(T, M) case PrimType_##T: return literal(type, Box(T(l.get_##T() * r.get_##T())), dbg);
+#define THORIN_P_TYPE(T, M) case PrimType_##T: return literal(type, Box(T(l.get_##T() * r.get_##T())), dbg);
+#define THORIN_Q_TYPE(T, M) case PrimType_##T: return literal(type, Box(T(l.get_##T() * r.get_##T())), dbg);
+#define THORIN_BOOL_TYPE(T, M) case PrimType_##T: return literal(type, Box(T(l.get_##T() && r.get_##T())), dbg);
 #include "thorin/tables/primtypetable.h"
                         default: THORIN_UNREACHABLE;
                     }

From e2c74bb3ffc6ac69a568b7a22e38f188344db02f Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 30 Jun 2023 15:37:36 +0200
Subject: [PATCH 180/342] Fix World::extract. Indeces of different types need
 to be taken into account. This fixes an issue where some insertvalue
 instructions could be lost.

---
 src/thorin/world.cpp | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 51de82e52..53e0270da 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -709,9 +709,14 @@ const Def* World::extract(const Def* agg, const Def* index, Debug dbg) {
     if (auto insert = agg->isa<Insert>()) {
         if (index == insert->index())
             return insert->value();
-        else if (index->template isa<PrimLit>()) {
-            if (insert->index()->template isa<PrimLit>())
-                return extract(insert->agg(), index, dbg);
+        else if (auto index_lit = index->isa<PrimLit>()) {
+            if (auto insert_index_lit = insert->index()->isa<PrimLit>()) {
+                if (index_lit->value() == insert_index_lit->value()) {
+                    return insert->value();
+                } else {
+                    return extract(insert->agg(), index, dbg);
+                }
+            }
         }
     }
 

From 07602459177795916dd47f76404fe32ef1a3ed29 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 4 Sep 2023 15:49:30 +0200
Subject: [PATCH 181/342] Rewrote scoping analysis, added new debugging
 facilities

---
 src/thorin/CMakeLists.txt                   |   7 +-
 src/thorin/analyses/cfg.cpp                 |   4 +-
 src/thorin/analyses/cfg.h                   |   2 -
 src/thorin/analyses/domfrontier.cpp         |  25 --
 src/thorin/analyses/domfrontier.h           |  50 ---
 src/thorin/analyses/schedule.cpp            |  36 +-
 src/thorin/analyses/schedule.h              |  10 +-
 src/thorin/analyses/scope.cpp               | 364 ++++++++++++++++----
 src/thorin/analyses/scope.h                 |  61 +++-
 src/thorin/analyses/verify.cpp              |  39 ++-
 src/thorin/be/c/c.cpp                       |   8 +-
 src/thorin/be/codegen.cpp                   |   3 +-
 src/thorin/be/emitter.h                     |  14 +-
 src/thorin/be/llvm/llvm.cpp                 |   5 +-
 src/thorin/be/shady/shady.cpp               |   5 +-
 src/thorin/rec_stream.cpp                   |  13 +-
 src/thorin/transform/cleanup_world.cpp      |   3 +-
 src/thorin/transform/clone_bodies.cpp       |  46 ---
 src/thorin/transform/clone_bodies.h         |  12 -
 src/thorin/transform/codegen_prepare.cpp    |   3 +-
 src/thorin/transform/dead_load_opt.cpp      |   3 +-
 src/thorin/transform/hls_channels.cpp       |   2 +-
 src/thorin/transform/hls_kernel_launch.cpp  |   2 +-
 src/thorin/transform/hoist_enters.cpp       |  32 +-
 src/thorin/transform/inliner.cpp            |   2 +-
 src/thorin/transform/lift_builtins.cpp      |   3 +-
 src/thorin/transform/partial_evaluation.cpp |  60 +---
 src/thorin/transform/split_slots.cpp        |   2 +-
 src/thorin/util/graphviz_dump.cpp           | 315 +++++++++++++++++
 src/thorin/util/scoped_dump.cpp             | 161 +++++++++
 src/thorin/util/scoped_dump.h               |  52 +++
 src/thorin/world.cpp                        |   2 -
 src/thorin/world.h                          |   9 +
 33 files changed, 1018 insertions(+), 337 deletions(-)
 delete mode 100644 src/thorin/analyses/domfrontier.cpp
 delete mode 100644 src/thorin/analyses/domfrontier.h
 delete mode 100644 src/thorin/transform/clone_bodies.cpp
 delete mode 100644 src/thorin/transform/clone_bodies.h
 create mode 100644 src/thorin/util/graphviz_dump.cpp
 create mode 100644 src/thorin/util/scoped_dump.cpp
 create mode 100644 src/thorin/util/scoped_dump.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index f459e9ca8..e3247e5f7 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -16,8 +16,6 @@ set(THORIN_SOURCES
     world.h
     analyses/cfg.cpp
     analyses/cfg.h
-    analyses/domfrontier.cpp
-    analyses/domfrontier.h
     analyses/domtree.cpp
     analyses/domtree.h
     analyses/free_defs.cpp
@@ -44,8 +42,6 @@ set(THORIN_SOURCES
     tables/mathoptable.h
     transform/cleanup_world.cpp
     transform/cleanup_world.h
-    transform/clone_bodies.cpp
-    transform/clone_bodies.h
     transform/closure_conversion.cpp
     transform/closure_conversion.h
     transform/codegen_prepare.h
@@ -87,6 +83,9 @@ set(THORIN_SOURCES
     util/symbol.h
     util/types.h
     util/utility.h
+    util/graphviz_dump.cpp
+    util/scoped_dump.h
+    util/scoped_dump.cpp
     )
 
 if(LLVM_FOUND)
diff --git a/src/thorin/analyses/cfg.cpp b/src/thorin/analyses/cfg.cpp
index e95b261f3..809944b60 100644
--- a/src/thorin/analyses/cfg.cpp
+++ b/src/thorin/analyses/cfg.cpp
@@ -6,7 +6,6 @@
 #include <stack>
 
 #include "thorin/world.h"
-#include "thorin/analyses/domfrontier.h"
 #include "thorin/analyses/domtree.h"
 #include "thorin/analyses/looptree.h"
 #include "thorin/analyses/scope.h"
@@ -30,7 +29,7 @@ Stream& CFNode::stream(Stream& s) const { return s << continuation(); }
 CFA::CFA(const Scope& scope)
     : scope_(scope)
     , entry_(node(scope.entry()))
-    , exit_ (node(scope.exit() ))
+    , exit_ (node(scope.entry()->world().end_scope()))
 {
     std::queue<Continuation*> cfg_queue;
     ContinuationSet cfg_done;
@@ -194,7 +193,6 @@ template<bool forward> const CFNodes& CFG<forward>::preds(const CFNode* n) const
 template<bool forward> const CFNodes& CFG<forward>::succs(const CFNode* n) const { assert(n != nullptr); return forward ? n->succs() : n->preds(); }
 template<bool forward> const DomTreeBase<forward>& CFG<forward>::domtree() const { return lazy_init(this, domtree_); }
 template<bool forward> const LoopTree<forward>& CFG<forward>::looptree() const { return lazy_init(this, looptree_); }
-template<bool forward> const DomFrontierBase<forward>& CFG<forward>::domfrontier() const { return lazy_init(this, domfrontier_); }
 
 template class CFG<true>;
 template class CFG<false>;
diff --git a/src/thorin/analyses/cfg.h b/src/thorin/analyses/cfg.h
index 1d1786d78..c61207553 100644
--- a/src/thorin/analyses/cfg.h
+++ b/src/thorin/analyses/cfg.h
@@ -132,7 +132,6 @@ class CFG {
     const CFNode* operator [] (Continuation* continuation) const { return cfa()[continuation]; }    ///< Maps from @p l to @p CFNode.
     const DomTreeBase<forward>& domtree() const;
     const LoopTree<forward>& looptree() const;
-    const DomFrontierBase<forward>& domfrontier() const;
 
     static size_t index(const CFNode* n) { return forward ? n->f_index_ : n->b_index_; }
 
@@ -143,7 +142,6 @@ class CFG {
     Map<const CFNode*> rpo_;
     mutable std::unique_ptr<const DomTreeBase<forward>> domtree_;
     mutable std::unique_ptr<const LoopTree<forward>> looptree_;
-    mutable std::unique_ptr<const DomFrontierBase<forward>> domfrontier_;
 };
 
 //------------------------------------------------------------------------------
diff --git a/src/thorin/analyses/domfrontier.cpp b/src/thorin/analyses/domfrontier.cpp
deleted file mode 100644
index 98e4d4f1d..000000000
--- a/src/thorin/analyses/domfrontier.cpp
+++ /dev/null
@@ -1,25 +0,0 @@
-#include "thorin/analyses/domfrontier.h"
-
-#include "thorin/analyses/domtree.h"
-
-namespace thorin {
-
-template<bool forward>
-void DomFrontierBase<forward>::create() {
-    const auto& domtree = cfg().domtree();
-    for (auto n : cfg().reverse_post_order().skip_front()) {
-        const auto& preds = cfg().preds(n);
-        if (preds.size() > 1) {
-            auto idom = domtree.idom(n);
-            for (auto pred : preds) {
-                for (auto i = pred; i != idom; i = domtree.idom(i))
-                    link(i, n);
-            }
-        }
-    }
-}
-
-template class DomFrontierBase<true>;
-template class DomFrontierBase<false>;
-
-}
diff --git a/src/thorin/analyses/domfrontier.h b/src/thorin/analyses/domfrontier.h
deleted file mode 100644
index 3014e1be8..000000000
--- a/src/thorin/analyses/domfrontier.h
+++ /dev/null
@@ -1,50 +0,0 @@
-#ifndef THORIN_ANALYSES_DOMFRONTIER_H
-#define THORIN_ANALYSES_DOMFRONTIER_H
-
-#include "thorin/analyses/cfg.h"
-
-namespace thorin {
-
-/**
- * A Dominance Frontier Graph.
- * The template parameter @p forward determines whether to compute
- * regular dominance frontiers or post-dominance frontiers (i.e. control dependence).
- * This template parameter is associated with @p CFG's @c forward parameter.
- * See Cooper et al, 2001. A Simple, Fast Dominance Algorithm: http://www.cs.rice.edu/~keith/EMBED/dom.pdf
- */
-template<bool forward>
-class DomFrontierBase {
-public:
-    DomFrontierBase(const DomFrontierBase &) = delete;
-    DomFrontierBase& operator=(DomFrontierBase) = delete;
-
-    explicit DomFrontierBase(const CFG<forward> &cfg)
-        : cfg_(cfg)
-        , preds_(cfg)
-        , succs_(cfg)
-    {
-        create();
-    }
-
-    const CFG<forward>& cfg() const { return cfg_; }
-    const std::vector<const CFNode*>& preds(const CFNode* n) const { return preds_[n]; }
-    const std::vector<const CFNode*>& succs(const CFNode* n) const { return succs_[n]; }
-
-private:
-    void create();
-    void link(const CFNode* src, const CFNode* dst) {
-        succs_[src].push_back(dst);
-        preds_[dst].push_back(src);
-    }
-
-    const CFG<forward>& cfg_;
-    typename CFG<forward>::template Map<std::vector<const CFNode*>> preds_;
-    typename CFG<forward>::template Map<std::vector<const CFNode*>> succs_;
-};
-
-typedef DomFrontierBase<true>  DomFrontiers;
-typedef DomFrontierBase<false> ControlDeps;
-
-}
-
-#endif
diff --git a/src/thorin/analyses/schedule.cpp b/src/thorin/analyses/schedule.cpp
index 798de7255..999d5da85 100644
--- a/src/thorin/analyses/schedule.cpp
+++ b/src/thorin/analyses/schedule.cpp
@@ -1,18 +1,17 @@
 #include "thorin/analyses/schedule.h"
 
-#include "thorin/config.h"
 #include "thorin/continuation.h"
 #include "thorin/primop.h"
 #include "thorin/world.h"
-#include "thorin/analyses/cfg.h"
 #include "thorin/analyses/domtree.h"
 #include "thorin/analyses/looptree.h"
 #include "thorin/analyses/scope.h"
 
 namespace thorin {
 
-Scheduler::Scheduler(const Scope& s)
-    : scope_(&s)
+Scheduler::Scheduler(const Scope& s, ScopesForest& forest)
+    : forest_(&forest)
+    , scope_(&s)
     , cfg_(&scope().f_cfg())
     , domtree_(&cfg().domtree())
 {
@@ -42,22 +41,26 @@ Scheduler::Scheduler(const Scope& s)
                 enqueue(def, i, def->op(i));
         }
     }
+
+    register_defs(s);
 }
 
-Continuation* Scheduler::early(const Def* def) {
-    if (auto cont = early_.lookup(def)) return *cont;
-    if (auto param = def->isa<Param>()) return early_[def] = param->continuation();
+void Scheduler::register_defs(const Scope& s) {
+    for (auto child : s.children_scopes()) {
+        Scope& cs = forest_->get_scope(child);
+        register_defs(cs);
+    }
 
-    auto result = scope().entry();
-    for (auto op : def->as_structural()->ops()) {
-        if (!op->isa_nom<Continuation>() && def2uses_.find(op) != def2uses_.end()) {
-            auto cont = early(op);
-            if (domtree().depth(cfg(cont)) > domtree().depth(cfg(result)))
-                result = cont;
-        }
+    for (auto def : s.defs()) {
+        if (!early_.lookup(def))
+            early_[def] = s.entry();
     }
+}
 
-    return early_[def] = result;
+Continuation* Scheduler::early(const Def* def, DefSet* seen) {
+    if (auto cont = early_.lookup(def)) return *cont;
+    if (auto param = def->isa<Param>()) return early_[def] = param->continuation();
+    assert(false);
 }
 
 Continuation* Scheduler::late(const Def* def) {
@@ -68,6 +71,9 @@ Continuation* Scheduler::late(const Def* def) {
         result = continuation;
     } else if (auto param = def->isa<Param>()) {
         result = param->continuation();
+    } else if (auto rec = def->isa_nom()) {
+        // don't try to late-schedule recursive nodes for now
+        result = early(def);
     } else {
         for (auto use : uses(def)) {
             auto cont = late(use);
diff --git a/src/thorin/analyses/schedule.h b/src/thorin/analyses/schedule.h
index b2cfa65aa..fd0ea8c96 100644
--- a/src/thorin/analyses/schedule.h
+++ b/src/thorin/analyses/schedule.h
@@ -11,7 +11,7 @@ using DomTree = DomTreeBase<true>;
 class Scheduler {
 public:
     Scheduler() = default;
-    explicit Scheduler(const Scope&);
+    explicit Scheduler(const Scope&, ScopesForest&);
 
     /// @name getters
     //@{
@@ -19,18 +19,21 @@ class Scheduler {
     const F_CFG& cfg() const { return *cfg_; }
     const CFNode* cfg(Continuation* cont) const { return cfg()[cont]; }
     const DomTree& domtree() const { return *domtree_; }
-    const Uses& uses(const Def* def) const { return def2uses_.find(def)->second; }
+    const Uses& uses(const Def* def) const { assert(def2uses_.contains(def)); return def2uses_.find(def)->second; }
     //@}
 
+    void register_defs(const Scope&);
+
     /// @name compute schedules
     //@{
-    Continuation* early(const Def*);
+    Continuation* early(const Def*, DefSet* seen = nullptr);
     Continuation* late (const Def*);
     Continuation* smart(const Def*);
     //@}
 
     friend void swap(Scheduler& s1, Scheduler& s2) {
         using std::swap;
+        swap(s1.forest_,   s2.forest_);
         swap(s1.scope_,    s2.scope_);
         swap(s1.cfg_,      s2.cfg_);
         swap(s1.domtree_,  s2.domtree_);
@@ -41,6 +44,7 @@ class Scheduler {
     }
 
 private:
+    ScopesForest* forest_ = nullptr;
     const Scope* scope_     = nullptr;
     const F_CFG* cfg_       = nullptr;
     const DomTree* domtree_ = nullptr;
diff --git a/src/thorin/analyses/scope.cpp b/src/thorin/analyses/scope.cpp
index 8fd518885..d9e7d6fc7 100644
--- a/src/thorin/analyses/scope.cpp
+++ b/src/thorin/analyses/scope.cpp
@@ -12,134 +12,354 @@
 
 namespace thorin {
 
-Scope::Scope(Continuation* entry)
-    : world_(entry->world())
-    , entry_(entry)
-    , exit_(world().end_scope())
-{
+Scope::Scope(Continuation* entry) : world_(entry->world()), root(std::make_unique<ScopesForest>(world())), forest_(*root), entry_(entry) {
     run();
 }
 
+Scope::Scope(Continuation* entry, ScopesForest& forest)
+    : world_(entry->world())
+    , forest_(forest)
+    , entry_(entry)
+{}
+
 Scope::~Scope() {}
 
 Scope& Scope::update() {
     defs_.clear();
-    free_        = nullptr;
+    free_frontier_.clear();
+    first_free_param_ = nullptr;
     free_params_ = nullptr;
     cfa_         = nullptr;
     run();
     return *this;
 }
 
-void Scope::run() {
+DefSet Scope::potentially_contained() const {
+    DefSet potential_defs;
     std::queue<const Def*> queue;
 
     auto enqueue = [&] (const Def* def) {
-        if (defs_.insert(def).second) {
+        if (potential_defs.insert(def).second) {
             queue.push(def);
-
-            if (auto continuation = def->isa_nom<Continuation>()) {
-                // when a continuation is part of this scope, we also enqueue its params, and we assert those to be unique
-                // TODO most likely redundant once params have the continuation in their ops
-                for (auto param : continuation->params()) {
-                    auto p = defs_.insert(param);
-                    assert_unused(p.second);
-                    queue.push(param);
-                }
-            }
         }
     };
 
-    enqueue(entry_);
+    for (auto param : entry()->params())
+        enqueue(param);
 
     while (!queue.empty()) {
         auto def = pop(queue);
-        if (def != entry_) {
-            for (auto use : def->uses())
+        if (def != entry()) {
+            for (auto use: def->uses())
                 enqueue(use);
         }
     }
 
-    enqueue(exit_);
+    return potential_defs;
+}
+
+void Scope::run() {
+    DefSet potential_defs = potentially_contained();
+
+    unique_queue<DefSet> queue;
+
+    if (entry()->has_body())
+        queue.push(entry()->body());
+    queue.push(entry()->filter());
+
+    defs_.insert(entry());
+    for (auto p : entry()->params())
+        defs_.insert(p);
+
+    while (!queue.empty()) {
+        auto def = queue.pop();
+
+        if (potential_defs.contains(def)) {
+            defs_.insert(def);
+            for (auto op : def->ops())
+                queue.push(op);
+        } else {
+            free_frontier_.insert(def);
+        }
+    }
 }
 
-const DefSet& Scope::free() const {
-    if (!free_) {
-        free_ = std::make_unique<DefSet>();
+void Scope::verify() {
+    for (auto def : defs()) {
+        if (auto cont = def->isa_nom<Continuation>()) {
+            if (cont == entry())
+                continue;
+            Scope& cont_scope = forest_.get_scope(cont);
+            assert(cont_scope.has_free_params());
+            assert(!cont_scope.contains(entry()));
+        }
+    }
+}
 
-        for (auto def : defs_) {
-            for (auto op : def->ops()) {
-                if (!contains(op))
-                    free_->emplace(op);
+static auto first_or_null = [](ParamSet& set) -> const Param* {
+    for (auto p : set) {
+        return p;
+    }
+    return nullptr;
+};
+
+// searches for free variable in this scope, starting at the free frontier and transitively searching the scopes of the free continuations we encounter
+// stop_after_first is used when we only care about whether this is a top-level scope (ie has no free params) or not, we can stop as soon as one param is found
+template<bool stop_after_first>
+std::tuple<ParamSet, bool> Scope::search_free_params() const {
+    if (free_params_) {
+        //world().WLOG("free variables analysis: reusing cached results for {}", entry());
+        return std::make_tuple(*free_params_, true);
+    }
+
+    if (stop_after_first && first_free_param_) {
+        ParamSet one_or_zero_params;
+        if (*first_free_param_ != nullptr)
+            one_or_zero_params.insert(first_free_param());
+        return std::make_tuple(one_or_zero_params, true);
+    }
+
+    ParamSet free_params;
+    /// as much as possible, we'd like to keep the results of those searches, but in the recursive case it's not always possible:
+    /// if there is a cycle, we stop when we encounter a continuation we're already searching the free variables for
+    /// this variable keeps track of whether we did that or not, if we didn't and this is a full search, we can safely save the results
+    bool thorough = true;
+    bool is_root = forest_.stack_.empty();
+    //world().WLOG("free variables analysis: searching transitive ops of: {} (root={}, depth={})", entry(), root, forest_->stack_.size());
+    forest_.stack_.push_back(entry());
+
+    unique_queue<DefSet> queue;
+
+    for (auto def : free_frontier_)
+        queue.push(def);
+
+    while (!queue.empty()) {
+        auto free_def = queue.pop();
+        assert(!contains(free_def));
+
+        if (auto param = free_def->isa<Param>(); param && !param->continuation()->dead_) {
+            free_params.insert(param);
+            if (stop_after_first)
+                break;
+        } else if (auto cont = free_def->isa_nom<Continuation>()) {
+            // the free variables analysis can be recursive, but it's not necessary to inspect our own scope again ...
+            assert(cont != entry());
+
+            // if we hit the recursion wall, the results for this free variable search are only valid for the callee
+            if (std::find(forest_.stack_.begin(), forest_.stack_.end(), cont) != forest_.stack_.end()) {
+                //world().WLOG("free variables analysis: skipping {} to prevent infinite recursion", cont);
+                thorough = false;
+                continue;
             }
+
+            Scope& scope = forest_.get_scope(cont);
+            assert(!scope.defs().empty() || !scope.entry()->has_body());
+
+            // When we have a free continuation in the body of our fn, their free variables are also free in us
+            auto [callee_free_params, callee_results_thorough] = scope.search_free_params<stop_after_first>();
+            if (!is_root)
+                thorough &= callee_results_thorough;
+
+            for (auto p: callee_free_params) {
+                // (those variables have to be free here! otherwise that continuation should be in this scope and not free)
+                if (contains(p)) {
+                    world().WLOG("Potentially broken scoping: free variable {} showed up in the free variables of {} despite that continuation being part of its scope", p, entry());
+                    assert(false);
+                }
+                free_params.insert(p);
+                if (stop_after_first)
+                    break;
+            }
+        }
+        else {
+            for (auto op : free_def->ops()) {
+                assert(op && "scope analysis doesn't work with unfinished nodes...");
+                // the entry might be referenced by the outside world, but that's completely fine
+                if (op == entry())
+                    continue;
+                assert(!contains(op));
+                queue.push(op);
+            }
+        }
+    }
+
+    //world().WLOG("free variables analysis: done with : {} (hit_wall={})", entry(), hit_recursion_wall);
+
+    assert(forest_.stack_.back() == entry());
+    forest_.stack_.pop_back();
+
+    // save the results if we can
+    if (thorough) {
+        if (stop_after_first) {
+            first_free_param_ = std::make_optional<const Param*>(first_or_null(free_params));
+            return std::make_tuple(free_params, true);
+        } else {
+            free_params_ = std::make_unique<ParamSet>(std::move(free_params));
+            return std::make_tuple(*free_params_, true);
         }
     }
 
-    return *free_;
+    for (auto fp : free_params) {
+        assert(fp->continuation() != entry());
+    }
+
+    return std::make_tuple(free_params, thorough);
 }
 
 const ParamSet& Scope::free_params() const {
     if (!free_params_) {
-        free_params_ = std::make_unique<ParamSet>();
-        unique_queue<DefSet> queue;
-
-        auto enqueue = [&](const Def* def) {
-            if (auto param = def->isa<Param>(); param && !param->continuation()->dead_)
-                free_params_->emplace(param);
-            else if (def->isa<Continuation>())
-                return;
-            else
-                queue.push(def);
-        };
-
-        for (auto def : free())
-            enqueue(def);
-        while (!queue.empty()) {
-            for (auto op : queue.pop()->ops())
-                enqueue(op);
-        }
+        auto [set, valid] = search_free_params<false>();
+        assert(valid);
+        free_params_ = std::make_unique<ParamSet>(set);
     }
 
     return *free_params_;
 }
 
+const Param* Scope::first_free_param() const {
+    if (!first_free_param_) {
+        // if we already computed the full free params list, let's reuse that !
+        if (free_params_) {
+            first_free_param_ = std::make_optional<const Param*>(first_or_null(*free_params_));
+        } else {
+            auto [set, valid] = search_free_params<true>();
+            assert(valid);
+            first_free_param_ = std::make_optional<const Param*>(first_or_null(set));
+        }
+    }
+
+    return *first_free_param_;
+}
+
+bool Scope::has_free_params() const {
+    return first_free_param() != nullptr;
+}
+
 const CFA& Scope::cfa() const { return lazy_init(this, cfa_); }
 const F_CFG& Scope::f_cfg() const { return cfa().f_cfg(); }
 const B_CFG& Scope::b_cfg() const { return cfa().b_cfg(); }
 
-template<bool elide_empty>
-void Scope::for_each(const World& world, std::function<void(Scope&)> f) {
-    unique_queue<ContinuationSet> continuation_queue;
+Continuation* Scope::parent_scope() const {
+    if (!parent_scope_) {
+        ContinuationSet candidates_set;
+        std::queue<Continuation*> candidates;
 
-    for (auto&& [_, def] : world.externals()) {
-        if (auto cont = def->template isa<Continuation>())
-            if (cont->has_body()) continuation_queue.push(cont);
-    }
+        for (auto param : free_params()) {
+            if(candidates_set.insert(param->continuation()).second)
+                candidates.push(param->continuation());
+        }
 
-    while (!continuation_queue.empty()) {
-        auto continuation = continuation_queue.pop();
-        if (elide_empty && !continuation->has_body())
-            continue;
-        Scope scope(continuation);
-        f(scope);
+        if (candidates.empty())
+            parent_scope_ = std::make_optional<Continuation*>();
+        else {
+            while (true) {
+                auto candidate = pop(candidates);
+                // when there is only one candidate left, that's our parent
+                if (candidates.empty()) {
+                    parent_scope_ = std::make_optional<Continuation*>(candidate);
+                    break;
+                }
 
-        unique_queue<DefSet> def_queue;
-        for (auto def : scope.free())
-            def_queue.push(def);
+                auto other_candidate = pop(candidates);
+                assert(candidate != other_candidate);
+                if (forest_.get_scope(candidate).contains(other_candidate))
+                    candidates.push(other_candidate);
+                else {
+                    assert(forest_.get_scope(other_candidate).contains(candidate) && "a scope cannot be nested in two unrelated parent scopes");
+                    candidates.push(candidate);
+                }
+            }
+        }
+    }
 
-        while (!def_queue.empty()) {
-            auto def = def_queue.pop();
-            if (auto continuation = def->isa_nom<Continuation>())
-                continuation_queue.push(continuation);
+    return *parent_scope_;
+}
+
+ContinuationSet Scope::children_scopes() const {
+    ContinuationSet set;
+    for (auto def : defs()) {
+        if (auto cont = def->isa_nom<Continuation>()) {
+            auto& scope = forest_.get_scope(cont);
+            if (scope.parent_scope() == entry())
+                set.insert(cont);
             else {
-                for (auto op : def->ops())
-                    def_queue.push(op);
+                //TODO: add check that the parent is eventually us!
             }
         }
     }
+    return set;
+}
+
+template void ScopesForest::for_each<true> ( std::function<void(Scope&)>);
+template void ScopesForest::for_each<false>( std::function<void(Scope&)>);
+
+Scope& ScopesForest::get_scope(Continuation* entry) {
+    if (scopes_.contains(entry)) {
+        auto existing = scopes_.find(entry);
+        assert((size_t)(existing->second.get()) != 0xbebebebe00000000);
+        return *existing->second;
+    }
+    auto scope = std::make_unique<Scope>(entry, *this);
+    Scope* ptr = scope.get();
+    ptr->run();
+    scopes_[entry] = std::move(scope);
+    if (stack_.empty())
+       ptr->verify();
+    return *ptr;
+}
+
+ContinuationSet ScopesForest::top_level_scopes() {
+    ContinuationSet set;
+    for (auto cont : world_.copy_continuations()) {
+        auto& scope = get_scope(cont);
+        assert(stack_.empty());
+        if(!scope.has_free_params()) {
+            set.insert(cont);
+        }
+        assert(stack_.empty());
+    }
+    return set;
 }
 
-template void Scope::for_each<true> (const World&, std::function<void(Scope&)>);
-template void Scope::for_each<false>(const World&, std::function<void(Scope&)>);
+template<bool elide_empty>
+void ScopesForest::for_each(std::function<void(Scope&)> f) {
+    for (auto cont : top_level_scopes()) {
+        if (elide_empty && !cont->has_body())
+            continue;
+        auto& scope = get_scope(cont);
+        f(scope);
+    }
+}
+
+DEBUG_UTIL void dump_scope_tree(Scope& s, ScopesForest* f = nullptr, int depth = 0) {
+    for (int i = 0; i < depth; i++)
+        std::cerr << "  ";
+    std::unique_ptr<ScopesForest> sf;
+    if (!f) {
+        sf = std::make_unique<ScopesForest>(s.world());
+        f = sf.get();
+    }
+
+    std::cerr << s.entry()->unique_name() << std::endl;
+    for (auto c : s.children_scopes()) {
+        auto& cs = f->get_scope(c);
+        dump_scope_tree(cs, f, depth + 1);
+    }
+}
+
+DEBUG_UTIL void dump_scopes(World& w) {
+    ScopesForest forest(w);
+    std::vector<Scope*> q;
+    for (auto cont : w.copy_continuations()) {
+        auto& s = forest.get_scope(cont);
+        if (!s.parent_scope()) {
+            q.push_back(&s);
+        }
+    }
+
+    for (auto root : q) {
+        dump_scope_tree(*root, &forest);
+    }
+}
 
 }
diff --git a/src/thorin/analyses/scope.h b/src/thorin/analyses/scope.h
index fa301a457..c0aa9c7cf 100644
--- a/src/thorin/analyses/scope.h
+++ b/src/thorin/analyses/scope.h
@@ -16,6 +16,8 @@ typedef CFG<false> B_CFG;
 class CFA;
 class CFNode;
 
+class ScopesForest;
+
 /**
  * A @p Scope represents a region of @p Continuation%s which are live from the view of an @p entry @p Continuation.
  * Transitively, all user's of the @p entry's parameters are pooled into this @p Scope.
@@ -28,6 +30,7 @@ class Scope : public Streamable<Scope> {
     Scope& operator=(Scope) = delete;
 
     explicit Scope(Continuation* entry);
+    explicit Scope(Continuation* entry, ScopesForest&);
     ~Scope();
 
     /// Invoke if you have modified sth in this Scope.
@@ -36,19 +39,23 @@ class Scope : public Streamable<Scope> {
     //@{ misc getters
     World& world() const { return world_; }
     Continuation* entry() const { return entry_; }
-    Continuation* exit() const { return exit_; }
+    ScopesForest& forest() const { return forest_; }
     //@}
 
     //@{ get Def%s contained in this Scope
     const DefSet& defs() const { return defs_; }
+    const DefSet& free_frontier() const { return free_frontier_; }
     bool contains(const Def* def) const { return defs_.contains(def); }
-    /// All @p Def%s referenced but @em not contained in this @p Scope.
-    const DefSet& free() const;
+    //@}
+
     /// All @p Param%s that appear free in this @p Scope.
     const ParamSet& free_params() const;
     /// Are there any free @p Param%s within this @p Scope.
-    bool has_free_params() const { return !free_params().empty(); }
-    //@}
+    bool has_free_params() const;
+    const Param* first_free_param() const;
+
+    Continuation* parent_scope() const;
+    ContinuationSet children_scopes() const;
 
     //@{ simple CFA to construct a CFG
     const CFA& cfa() const;
@@ -61,24 +68,50 @@ class Scope : public Streamable<Scope> {
     Stream& stream(Stream&) const;                  ///< Streams thorin to file @p out.
     //@}
 
+    void verify();
+private:
+    void run();
+    DefSet potentially_contained() const;
+
+    template<bool stop_after_first>
+    std::tuple<ParamSet, bool> search_free_params() const;
+
+    World& world_;
+    std::unique_ptr<ScopesForest> root;
+    ScopesForest& forest_;
+    Continuation* entry_ = nullptr;
+    DefSet defs_;
+    DefSet free_frontier_;
+    mutable std::optional<const Param*> first_free_param_;
+    mutable std::unique_ptr<ParamSet> free_params_;
+    mutable std::unique_ptr<const CFA> cfa_;
+    mutable std::optional<Continuation*> parent_scope_;
+
+    friend ScopesForest;
+};
+
+class ScopesForest {
+public:
+    ScopesForest(World& world) : world_(world) {}
+
+    Scope& get_scope(Continuation* entry);
+
+    ContinuationSet top_level_scopes();
+
     /**
      * Transitively visits all @em reachable Scope%s in @p world that do not have free variables.
      * We call these Scope%s @em top-level Scope%s.
      * Select with @p elide_empty whether you want to visit trivial Scope%s of Continuation%s without body.
      */
     template<bool elide_empty = true>
-    static void for_each(const World&, std::function<void(Scope&)>);
+    void for_each(std::function<void(Scope&)>);
 
 private:
-    void run();
-
     World& world_;
-    DefSet defs_;
-    Continuation* entry_ = nullptr;
-    Continuation* exit_ = nullptr;
-    mutable std::unique_ptr<DefSet> free_;
-    mutable std::unique_ptr<ParamSet> free_params_;
-    mutable std::unique_ptr<const CFA> cfa_;
+    std::vector<Continuation*> stack_;
+    ContinuationMap<std::unique_ptr<Scope>> scopes_;
+
+    friend Scope;
 };
 
 }
diff --git a/src/thorin/analyses/verify.cpp b/src/thorin/analyses/verify.cpp
index ba9a5f967..0eb3aee65 100644
--- a/src/thorin/analyses/verify.cpp
+++ b/src/thorin/analyses/verify.cpp
@@ -8,7 +8,7 @@ namespace thorin {
 
 // TODO this needs serious rewriting
 
-static bool verify_calls(World& world) {
+static bool verify_calls(World& world, ScopesForest& forest) {
     bool ok = true;
     for (auto def : world.defs()) {
         if (auto cont = def->isa<Continuation>())
@@ -17,16 +17,32 @@ static bool verify_calls(World& world) {
     return ok;
 }
 
-static bool verify_top_level(World& world) {
+static bool verify_top_level(World& world, ScopesForest& forest) {
     bool ok = true;
-    Scope::for_each(world, [&] (const Scope& scope) {
-        if (scope.has_free_params()) {
-            for (auto param : scope.free_params())
-                world.ELOG("top-level continuation '{}' got free param '{}' belonging to continuation {}", scope.entry(), param, param->continuation());
-            world.ELOG("here: {}", scope.entry());
-            ok = false;
+    unique_queue<DefSet> defs;
+    for (auto& external : world.externals())
+        defs.push(external.second);
+    while (!defs.empty()) {
+        auto def = defs.pop();
+        if (auto cont = def->isa_nom<Continuation>()) {
+            world.VLOG("verifying external continuation '{}'", cont);
+            auto& scope = forest.get_scope(cont);
+            scope.verify();
+            if (scope.has_free_params()) {
+                for (auto param : scope.free_params())
+                    world.ELOG("top-level continuation '{}' got free param '{}' belonging to continuation {}", scope.entry(), param, param->continuation());
+                world.ELOG("here: {}", scope.entry());
+                ok = false;
+            }
+        } else {
+            for (auto op : def->ops())
+                defs.push(op);
         }
-    });
+    }
+    for (auto cont : world.copy_continuations()) {
+        auto& scope = forest.get_scope(cont);
+        scope.verify();
+    }
     return ok;
 }
 
@@ -45,9 +61,10 @@ static bool verify_param(World& world) {
 }
 
 void verify(World& world) {
+    ScopesForest forest(world);
     bool ok = true;
-    ok &= verify_calls(world);
-    ok &= verify_top_level(world);
+    ok &= verify_calls(world, forest);
+    ok &= verify_top_level(world, forest);
     //TODO: This should not fail!
     //ok &= verify_param(world);
     if (!ok)
diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 61e47d3e0..ab6b503ad 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -75,6 +75,7 @@ class CCodeGen : public thorin::Emitter<std::string, std::string, BB, CCodeGen>
 public:
     CCodeGen(Thorin& thorin, const Cont2Config& kernel_config, Stream& stream, Lang lang, bool debug, std::string& flags)
         : thorin_(thorin)
+        , forest_(world())
         , kernel_config_(kernel_config)
         , lang_(lang)
         , fn_mem_(world().fn_type({world().mem_type()}))
@@ -116,6 +117,7 @@ class CCodeGen : public thorin::Emitter<std::string, std::string, BB, CCodeGen>
     std::string tuple_name(const TupleType*);
 
     Thorin& thorin_;
+    ScopesForest forest_;
     const Cont2Config& kernel_config_;
     Lang lang_;
     const FnType* fn_mem_;
@@ -355,15 +357,15 @@ void CCodeGen::emit_module() {
     Continuation* hls_top = nullptr;
     interface_status = get_interface(interface, gmem_config);
 
-    Scope::for_each(world(), [&] (const Scope& scope) {
+    forest_.for_each([&] (const Scope& scope) {
         if (scope.entry()->name() == "hls_top")
             hls_top = scope.entry();
         else if (scope.entry()->cc() != CC::Internal)
-            emit_scope(scope);
+            emit_scope(scope, forest_);
     });
     if (hls_top) {
         hls_top_scope = true;
-        emit_scope(Scope(hls_top));
+        emit_scope(Scope(hls_top), forest_);
     }
 
     if (lang_ == Lang::OpenCL) {
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 56b4190d2..e1a155256 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -93,7 +93,8 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     }
 
     // determine different parts of the world which need to be compiled differently
-    Scope::for_each(world, [&] (const Scope& scope) {
+    ScopesForest forest(world);
+    forest.for_each([&] (const Scope& scope) {
         auto continuation = scope.entry();
         Continuation* imported = nullptr;
 
diff --git a/src/thorin/be/emitter.h b/src/thorin/be/emitter.h
index a13ad8d49..c138ad5e7 100644
--- a/src/thorin/be/emitter.h
+++ b/src/thorin/be/emitter.h
@@ -40,7 +40,8 @@ class Emitter {
             emit_unsafe(r);
         }
 
-        auto place = def->no_dep() ? entry_ : scheduler_.smart(def);
+        //auto place = def->no_dep() ? entry_ : scheduler_.smart(def);
+        auto place = !scheduler_.scope().contains(def) ? entry_ : scheduler_.smart(def);
 
         if (place) {
             auto& bb = cont2bb_[place];
@@ -73,22 +74,23 @@ class Emitter {
         return defs_[def] = val;
     }
 
-    void emit_scope(const Scope& scope) {
+    void emit_scope(const Scope& scope, ScopesForest& forest) {
+        scope_ = &scope;
         auto conts = schedule(scope);
         entry_ = scope.entry();
-        assert(entry_->is_returning());
+        //assert(entry_->is_returning());
 
         auto fct = child().prepare(scope);
         for (auto cont : conts) {
             if (cont->intrinsic() != Intrinsic::EndScope) child().prepare(cont, fct);
         }
 
-        Scheduler new_scheduler(scope);
+        Scheduler new_scheduler(scope, forest);
         swap(scheduler_, new_scheduler);
 
         for (auto cont : conts) {
             if (cont->intrinsic() == Intrinsic::EndScope) continue;
-            assert(cont == entry_ || cont->is_basicblock());
+            //assert(cont == entry_ || cont->is_basicblock());
             child().emit_epilogue(cont);
         }
 
@@ -96,6 +98,7 @@ class Emitter {
             if (cont->intrinsic() != Intrinsic::EndScope) child().finalize(cont);
         }
         child().finalize(scope);
+        scope_ = nullptr;
     }
 
     Scheduler scheduler_;
@@ -103,6 +106,7 @@ class Emitter {
     DefMap<Type> types_;
     ContinuationMap<BB> cont2bb_;
     Continuation* entry_ = nullptr;
+    const Scope* scope_ = nullptr;
 };
 
 }
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 045763c3b..e1b36028a 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -309,11 +309,12 @@ CodeGen::emit_module() {
         }
     }
 
-    Scope::for_each(world(), [&] (const Scope& scope) {
+    ScopesForest forest(world());
+    forest.for_each([&](const Scope& scope) {
         if(scope.entry()->cc() == CC::Internal) {
             return;
         }
-        emit_scope(scope);
+        emit_scope(scope, forest);
     });
 
     if (debug()) dibuilder_.finalize();
diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index a2d83740f..297606171 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -18,11 +18,12 @@ void CodeGen::emit_stream(std::ostream& out) {
     arena = shady::new_ir_arena(config);
     module = shady::new_module(arena, world().name().c_str());
 
-    Scope::for_each(world(), [&](const Scope& scope) {
+    ScopesForest forest(world());
+    forest.for_each([&](const Scope& scope) {
         if(scope.entry()->cc() == CC::Internal) {
             return;
         }
-        emit_scope(scope);
+        emit_scope(scope, forest);
     });
 
     char* bufptr;
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 29ae570cc..037891722 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -55,10 +55,18 @@ void RecStreamer::run() {
                 s.fmt("extern ");
         }
 
+        Scope scope(cont);
+        if (!scope.has_free_params())
+            s.fmt("top_level ");
+        else {
+            s.fmt("// free variables: {, }\n", scope.free_params());
+            s.fmt("// free frontier: {, }\n", scope.free_frontier());
+        }
+
         if (cont->has_body()) {
             std::vector<std::string> param_names;
             for (auto param : cont->params()) param_names.push_back(param->unique_name());
-            s.fmt("{}: {} = ({, }) => {{\t\n", cont->unique_name(), cont->type(), param_names);
+            s.fmt("{}: {} = ({, }) @({}) => {{\t\n", cont->unique_name(), cont->type(), param_names, cont->filter());
             run(cont->filter());
             if (defs.contains(cont->body())) {
                 auto body = cont->body();
@@ -179,6 +187,7 @@ Stream& Scope::stream(Stream&) const {
 
 Stream& Type::stream(Stream& s) const {
     if (false) {}
+    else if (isa<Star>()) return s.fmt("*");
     else if (isa<BottomType>()) return s.fmt("!!");
     else if (isa<   MemType>()) return s.fmt("mem");
     else if (isa< FrameType>()) return s.fmt("frame");
@@ -221,8 +230,6 @@ Stream& Type::stream(Stream& s) const {
 
         if (t->is_vector()) s.fmt(">");
         return s;
-    } else if (isa<Star>()) {
-        return s.fmt("★");
     }
     THORIN_UNREACHABLE;
 }
diff --git a/src/thorin/transform/cleanup_world.cpp b/src/thorin/transform/cleanup_world.cpp
index 635b34a56..9a8eeaa99 100644
--- a/src/thorin/transform/cleanup_world.cpp
+++ b/src/thorin/transform/cleanup_world.cpp
@@ -36,7 +36,8 @@ class Cleaner {
 };
 
 void Cleaner::eliminate_tail_rec() {
-    Scope::for_each(world(), [&](Scope& scope) {
+    ScopesForest forest(world());
+    forest.for_each([&](Scope& scope) {
         auto entry = scope.entry();
 
         bool only_tail_calls = true;
diff --git a/src/thorin/transform/clone_bodies.cpp b/src/thorin/transform/clone_bodies.cpp
deleted file mode 100644
index 6b48fc456..000000000
--- a/src/thorin/transform/clone_bodies.cpp
+++ /dev/null
@@ -1,46 +0,0 @@
-#include "thorin/world.h"
-#include "thorin/analyses/scope.h"
-#include "thorin/transform/mangle.h"
-
-namespace thorin {
-
-/*
- * TODO rewrite
- */
-
-// TODO merge this with lift_builtins
-void clone_bodies(World& /*world*/) {
-#if 0
-    std::vector<Continuation*> todo;
-
-    // TODO this looks broken: I guess we should do that in post-order as in lift_builtins
-    for (auto continuation : world.copy_continuations()) {
-        if (is_passed_to_accelerator(continuation))
-            todo.push_back(continuation);
-    }
-
-    for (auto continuation : todo) {
-        Scope scope(continuation);
-        bool first = true;
-        for (auto use : continuation->copy_uses()) {
-            if (first) {
-                first = false; // re-use the initial continuation as first clone
-            } else {
-                auto ncontinuation = clone(scope);
-                if (auto uapp = use->isa<App>()) {
-                    if (uapp->is_replaced()) continue; // dead app node
-                    auto napp = uapp->with_different_op(use.index(), ncontinuation);
-                    uapp->replace(napp);
-                } else if (auto primop = use->isa<PrimOp>()) {
-                    Array<const Def*> nops(primop->num_ops());
-                    std::copy(primop->ops().begin(), primop->ops().end(), nops.begin());
-                    nops[use.index()] = ncontinuation;
-                    primop->replace(primop->rebuild(world, primop->type(), nops));
-                }
-            }
-        }
-    }
-#endif
-}
-
-}
diff --git a/src/thorin/transform/clone_bodies.h b/src/thorin/transform/clone_bodies.h
deleted file mode 100644
index 063815c66..000000000
--- a/src/thorin/transform/clone_bodies.h
+++ /dev/null
@@ -1,12 +0,0 @@
-#ifndef THORIN_TRANSFORM_CLONE_BODIES_H
-#define THORIN_TRANSFORM_CLONE_BODIES_H
-
-namespace thorin {
-
-class World;
-
-void clone_bodies(World&);
-
-}
-
-#endif
diff --git a/src/thorin/transform/codegen_prepare.cpp b/src/thorin/transform/codegen_prepare.cpp
index 95fbbb972..6ae0401d9 100644
--- a/src/thorin/transform/codegen_prepare.cpp
+++ b/src/thorin/transform/codegen_prepare.cpp
@@ -5,7 +5,8 @@ namespace thorin {
 
 void codegen_prepare(World& world) {
     world.VLOG("start codegen_prepare");
-    Scope::for_each(world, [&](Scope& scope) {
+    ScopesForest forest(world);
+    forest.for_each([&](Scope& scope) {
         world.DLOG("scope: {}", scope.entry());
         bool dirty = false;
         auto ret_param = scope.entry()->ret_param();
diff --git a/src/thorin/transform/dead_load_opt.cpp b/src/thorin/transform/dead_load_opt.cpp
index c88830b79..d3a83e3a8 100644
--- a/src/thorin/transform/dead_load_opt.cpp
+++ b/src/thorin/transform/dead_load_opt.cpp
@@ -37,7 +37,8 @@ static void dead_load_opt(const Scope& scope) {
 }
 
 void dead_load_opt(World& world) {
-    Scope::for_each(world, [&] (const Scope& scope) { dead_load_opt(scope); });
+    ScopesForest forest(world);
+    forest.for_each([&] (const Scope& scope) { dead_load_opt(scope); });
 }
 
 }
diff --git a/src/thorin/transform/hls_channels.cpp b/src/thorin/transform/hls_channels.cpp
index bae0f5291..f4c891133 100644
--- a/src/thorin/transform/hls_channels.cpp
+++ b/src/thorin/transform/hls_channels.cpp
@@ -160,7 +160,7 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
     Def2Def arg2param;
 
 
-    Scope::for_each(world, [&] (Scope& scope) {
+    ScopesForest(world).for_each([&] (Scope& scope) {
             auto old_kernel = scope.entry();
             Def2Mode def2mode;
             extract_kernel_channels(schedule(scope), def2mode);
diff --git a/src/thorin/transform/hls_kernel_launch.cpp b/src/thorin/transform/hls_kernel_launch.cpp
index 413c24506..94c3f32ed 100644
--- a/src/thorin/transform/hls_kernel_launch.cpp
+++ b/src/thorin/transform/hls_kernel_launch.cpp
@@ -76,7 +76,7 @@ void hls_kernel_launch(World& world, DeviceParams& device_params) {
     const size_t base_opencl_param_num = 6;
     Array<const Def*> opencl_args(base_opencl_param_num + device_params.size());
 
-    Scope::for_each(world, [&] (Scope& scope) {
+    ScopesForest(world).for_each([&] (Scope& scope) {
         Schedule scheduled = schedule(scope);
 
         for (auto& block : scheduled) {
diff --git a/src/thorin/transform/hoist_enters.cpp b/src/thorin/transform/hoist_enters.cpp
index a51cf8d64..3f8997915 100644
--- a/src/thorin/transform/hoist_enters.cpp
+++ b/src/thorin/transform/hoist_enters.cpp
@@ -24,41 +24,53 @@ static void find_enters(std::deque<const Enter*>& enters, Continuation* continua
         find_enters(enters, mem_param);
 }
 
-static void hoist_enters(const Scope& scope) {
+static bool hoist_enters(const Scope& scope) {
     World& world = scope.world();
     std::deque<const Enter*> enters;
 
     for (auto n : scope.f_cfg().reverse_post_order())
         find_enters(enters, n->continuation());
 
-
     if (enters.empty() || enters[0]->mem() != scope.entry()->mem_param()) {
         world.VLOG("cannot optimize {} - didn't find entry enter", scope.entry());
-        return;
+        return false;
     }
 
     auto entry_enter = enters[0];
     auto frame = entry_enter->out_frame();
     enters.pop_front();
 
+    bool todo = false;
     for (auto i = enters.rbegin(), e = enters.rend(); i != e; ++i) {
         auto old_enter = *i;
         for (auto use : old_enter->out_frame()->uses()) {
             auto slot = use->as<Slot>();
+            if (slot->uses().size() > 0)
+                todo = true;
             slot->replace_uses(world.slot(slot->alloced_type(), frame, slot->debug()));
             assert(slot->num_uses() == 0);
         }
     }
-
-    for (auto i = enters.rbegin(), e = enters.rend(); i != e; ++i)
-        (*i)->out_mem()->replace_uses((*i)->mem());
-
-    if (frame->num_uses() == 0)
-        entry_enter->out_mem()->replace_uses(entry_enter->mem());
+    return todo;
 }
 
+// TODO: rewrite this and put it out of its misery
 void hoist_enters(Thorin& thorin) {
-    Scope::for_each(thorin.world(), [] (const Scope& scope) { hoist_enters(scope); });
+    bool todo = false;
+    do {
+        todo = false;
+        ScopesForest forest(thorin.world());
+        for (auto cont : thorin.world().copy_continuations()) {
+            if (!cont->has_body())
+                continue;
+            auto& scope = forest.get_scope(cont);
+            if(!scope.has_free_params()) {
+                if (!todo) todo = hoist_enters(scope);
+                if (todo)
+                    break;
+            }
+        }
+    } while (todo);
     thorin.cleanup();
 }
 
diff --git a/src/thorin/transform/inliner.cpp b/src/thorin/transform/inliner.cpp
index d499051e0..ffaa8f309 100644
--- a/src/thorin/transform/inliner.cpp
+++ b/src/thorin/transform/inliner.cpp
@@ -69,7 +69,7 @@ void inliner(Thorin& thorin) {
         return nullptr;
     };
 
-    Scope::for_each(world, [&] (Scope& scope) {
+    ScopesForest(world).for_each([&] (Scope& scope) {
         bool dirty = false;
         for (auto n : scope.f_cfg().post_order()) {
             auto continuation = n->continuation();
diff --git a/src/thorin/transform/lift_builtins.cpp b/src/thorin/transform/lift_builtins.cpp
index 0ececb949..8f21d44ee 100644
--- a/src/thorin/transform/lift_builtins.cpp
+++ b/src/thorin/transform/lift_builtins.cpp
@@ -70,7 +70,8 @@ void lift_builtins(Thorin& thorin) {
     while (true) {
         World& world = thorin.world();
         Continuation* cur = nullptr;
-        Scope::for_each(world, [&] (const Scope& scope) {
+        ScopesForest forest(world);
+        forest.for_each([&] (const Scope& scope) {
             if (cur) return;
             for (auto n : scope.f_cfg().post_order()) {
                 if (n->continuation()->order() <= 1)
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 08c8d07ae..0968e45b1 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -2,6 +2,7 @@
 #include "thorin/world.h"
 #include "thorin/transform/mangle.h"
 #include "thorin/util/hash.h"
+#include "partial_evaluation.h"
 
 namespace thorin {
 
@@ -18,14 +19,12 @@ class PartialEvaluator {
     PartialEvaluator(World& world, bool lower2cff)
         : world_(world)
         , lower2cff_(lower2cff)
-        , boundary_(Def::gid_counter())
     {}
 
     World& world() { return world_; }
     bool run();
     void enqueue(Continuation* continuation) {
-        if (continuation->gid() < 2 * boundary_ && done_.emplace(continuation).second)
-            queue_.push(continuation);
+        queue_.push(continuation);
     }
     void eat_pe_info(Continuation*);
 
@@ -34,16 +33,14 @@ class PartialEvaluator {
     bool lower2cff_;
     HashMap<const App*, Continuation*, HashApp> cache_;
     ContinuationSet done_;
-    std::queue<Continuation*> queue_;
-    ContinuationMap<bool> top_level_;
-    size_t boundary_;
+    unique_queue<ContinuationSet> queue_;
 };
 
 class CondEval {
 public:
-    CondEval(Continuation* callee, Defs args, ContinuationMap<bool>& top_level)
+    CondEval(Continuation* callee, ScopesForest& forest, Defs args)
         : callee_(callee)
-        , top_level_(top_level)
+        , forest_(forest)
     {
         assert(callee->filter()->is_empty() || callee->filter()->size() == args.size());
         assert(callee->num_params() == args.size());
@@ -75,11 +72,11 @@ class CondEval {
         auto order = callee_->param(i)->order();
         if (lower2cff)
             if(order >= 2 || (order == 1
-                        && (!callee_->param(i)->type()->isa<FnType>()
-                            || (!callee_->is_returning() || (!is_top_level(callee_)))))) {
-            world().DLOG("bad param({}) {} of continuation {}", i, callee_->param(i), callee_);
-            return true;
-        }
+                              && (!callee_->param(i)->type()->isa<FnType>()
+                                  || (!callee_->is_returning() || (!is_top_level(callee_)))))) {
+                world().DLOG("bad param({}) {} of continuation {}", i, callee_->param(i), callee_);
+                return true;
+            }
 
         return ((!callee_->is_exported() || callee_->attributes().cc == CC::Internal) && callee_->can_be_inlined()) || is_one(instantiate(filter(i)));
         //return is_one(instantiate(filter(i)));
@@ -90,39 +87,13 @@ class CondEval {
     }
 
     bool is_top_level(Continuation* continuation) {
-        auto p = top_level_.emplace(continuation, true);
-        if (!p.second)
-            return p.first->second;
-
-        Scope scope(continuation);
-        unique_queue<DefSet> queue;
-
-        for (auto def : scope.free())
-            queue.push(def);
-
-        while (!queue.empty()) {
-            auto def = queue.pop();
-
-            if (def->isa<Param>()) // if FV in this scope is a param, this cont can't be top-level
-                return top_level_[continuation] = false;
-            if (auto free_cn = def->isa_nom<Continuation>()) {
-                // if we have a non-top level continuation in scope as a free variable,
-                // then it must be bound by some outer continuation, and so we aren't top-level
-                if (!is_top_level(free_cn))
-                    return top_level_[continuation] = false;
-            } else {
-                for (auto op : def->ops())
-                    queue.push(op);
-            }
-        }
-
-        return top_level_[continuation] = true;
+        return !forest_.get_scope(continuation).has_free_params();
     }
 
 private:
     Continuation* callee_;
     Def2Def old2new_;
-    ContinuationMap<bool>& top_level_;
+    ScopesForest& forest_;
 };
 
 void PartialEvaluator::eat_pe_info(Continuation* cur) {
@@ -151,11 +122,10 @@ bool PartialEvaluator::run() {
         if (!cont) continue;
         if (!cont->has_body()) continue;
         enqueue(cont);
-        top_level_[cont] = true;
     }
 
     while (!queue_.empty()) {
-        auto continuation = pop(queue_);
+        auto continuation = queue_.pop();
 
         bool force_fold = false;
 
@@ -176,7 +146,9 @@ bool PartialEvaluator::run() {
             }
 
             if (callee->has_body()) {
-                CondEval cond_eval(callee, body->args(), top_level_);
+                // TODO cache the forest and only rebuild it when we need to
+                ScopesForest forest(world());
+                CondEval cond_eval(callee, forest, body->args());
 
                 std::vector<const Def*> specialize(body->num_args());
 
diff --git a/src/thorin/transform/split_slots.cpp b/src/thorin/transform/split_slots.cpp
index 87f643352..054587755 100644
--- a/src/thorin/transform/split_slots.cpp
+++ b/src/thorin/transform/split_slots.cpp
@@ -92,7 +92,7 @@ void split_slots(Thorin& thorin) {
     bool todo = true;
     while (todo) {
         todo = false;
-        Scope::for_each(thorin.world(), [&] (const Scope& scope) { todo |= split_slots(scope); });
+        ScopesForest(thorin.world()).for_each([&] (const Scope& scope) { todo |= split_slots(scope); });
         thorin.cleanup();
     }
 }
diff --git a/src/thorin/util/graphviz_dump.cpp b/src/thorin/util/graphviz_dump.cpp
new file mode 100644
index 000000000..d85ea4bb4
--- /dev/null
+++ b/src/thorin/util/graphviz_dump.cpp
@@ -0,0 +1,315 @@
+#ifndef DOT_DUMP_H
+#define DOT_DUMP_H
+
+#include "thorin/world.h"
+#include "thorin/analyses/scope.h"
+
+namespace thorin {
+
+/// Outputs the raw thorin IR as a graph without performing any scope or scheduling analysis
+struct DotPrinter {
+    DotPrinter(World& world, const char* filename = "world.dot") : world_(world), forest_(world) {
+        file = std::ofstream(filename);
+        begin();
+    }
+
+    ~DotPrinter() {
+        end();
+    }
+private:
+    int indent = 0;
+    std::string endl() {
+        std::string s = "\n";
+        for (int i = 0; i < indent; i++)
+            s += " ";
+        return s;
+    }
+    std::string up() {
+        indent++;
+        return "";
+    }
+    std::string down() {
+        indent--;
+        return "";
+    }
+
+    void begin() {
+        file << "digraph " << "world" << " {" << up();
+        file << endl() << "bgcolor=transparent;";
+    }
+
+    void end() {
+        file << arrows.str();
+        file << down() << endl() << "}" << endl();
+    }
+
+    std::string def_id(const Def* def) {
+        return std::string(tag2str(def->tag())) + "_" + std::to_string(def->gid());
+    }
+
+    std::string emit_def(const Def* def);
+    std::string dump_literal(const Literal*);
+    std::string dump_continuation(Continuation* cont);
+
+    std::stringstream arrows;
+
+    void arrow(const std::string& src, const std::string& dst, const std::string& extra) {
+        if (src.empty() || dst.empty())
+            return;
+        arrows << endl() << src << " -> " << dst << " " << extra << ";";
+    }
+public:
+    std::string dump_def(const Def* def);
+
+    void run() {
+        delay_printing_ops = false;
+        while (!todo.empty()) {
+            auto def = todo.pop();
+            dump_def(def);
+        }
+    }
+
+    void print_scope(Scope& scope) {
+        file << "subgraph cluster_" << u_++ << " {" << up() << endl();
+
+        auto cont = scope.entry();
+        emit_def(cont);
+
+        for (auto child : scope.children_scopes())
+            print_scope(scope.forest().get_scope(child));
+
+        for (auto def : scope.defs()) {
+            if (!done.contains(def) && !def->isa<Continuation>()) {
+                emit_def(def);
+            }
+        }
+
+        file << down() << endl() << "}" << endl();
+    };
+
+    bool print_lower_order_args = true;
+    bool print_instanced_filters = false;
+    bool print_literals = true;
+    bool delay_printing_ops = true;
+    Scope* single_scope = nullptr;
+
+private:
+    thorin::World& world_;
+    ScopesForest forest_;
+
+    int u_ = 0;
+
+    unique_queue<DefSet> todo;
+    DefMap<std::string> done;
+    std::ofstream file;
+};
+
+std::string DotPrinter::dump_def(const Def* def) {
+    if (done.contains(def))
+        return done[def];
+
+    if (delay_printing_ops) {
+        todo.push(def);
+        return def_id(def);
+    }
+
+    return emit_def(def);
+}
+
+std::string DotPrinter::emit_def(const Def* def) {
+    assert(!done.contains(def));
+    if (auto cont = def->isa_nom<Continuation>())
+        return dump_continuation(cont);
+    else if (def->isa<Literal>())
+        return dump_literal(def->as<Literal>());
+    else {
+        // default (primops)
+        std::string color = "";
+        std::string fillcolor = "darkseagreen1";
+        std::string style = "filled";
+        std::string shape = "oval";
+        std::string label = std::string(def->op_name()) + " :: " + def->name();
+
+        std::unique_ptr<std::vector<std::tuple<std::string, const Def*>>> filtered_ops;
+
+        if (single_scope && single_scope->free_frontier().contains(def))
+            color = "blue";
+
+        if (def->isa<Param>()) {
+            fillcolor = "grey";
+            shape = "oval";
+            label = def->unique_name();
+        } else if (auto app = def->isa<App>()) {
+            fillcolor = "darkgreen";
+
+            filtered_ops = std::make_unique<std::vector<std::tuple<std::string, const Def*>>>();
+
+            if (auto callee_cont = app->callee()->isa_nom<Continuation>()) {
+                switch (callee_cont->intrinsic()) {
+                    case Intrinsic::Branch: {
+                        fillcolor = "lightblue";
+                        label = "branch";
+                        if (print_lower_order_args) {
+                            (*filtered_ops).emplace_back("mem", app->arg(0));
+                            (*filtered_ops).emplace_back("condition", app->arg(1));
+                        }
+                        (*filtered_ops).emplace_back("true", app->arg(2));
+                        (*filtered_ops).emplace_back("false", app->arg(3));
+
+                        goto print_node;
+                    }
+                    case Intrinsic::Match: {
+                        fillcolor = "lightblue";
+                        label = "match";
+
+                        if (print_lower_order_args) {
+                            (*filtered_ops).emplace_back("mem", app->arg(0));
+                            (*filtered_ops).emplace_back("inspectee", app->arg(1));
+                        }
+
+                        (*filtered_ops).emplace_back("default_case", app->arg(2));
+                        for (size_t i = 3; i < app->num_args(); i+=2) {
+                            if (print_lower_order_args)
+                                (*filtered_ops).emplace_back("case", app->arg(i));
+                            (*filtered_ops).emplace_back("case", app->arg(i + 1));
+                        }
+
+                        goto print_node;
+                    }
+                    default: break;
+                }
+            }
+
+            (*filtered_ops).emplace_back("callee", app->callee());
+            for (size_t i = 0; i < app->num_args(); i++) {
+                if (print_lower_order_args || app->arg(i)->type()->order() >= 1)
+                    (*filtered_ops).emplace_back("arg"+std::to_string(i), app->arg(i));
+            }
+        } else if (auto variant_ctor = def->isa<Variant>()) {
+            label = "variant(" + std::to_string(variant_ctor->index()) + ")";
+        } else if (auto variant_extract = def->isa<VariantExtract>()) {
+            label = "variant_extract(" + std::to_string(variant_extract->index()) + ")";
+        }
+
+        print_node:
+
+        file << endl() << def_id(def) << " [" << up();
+
+        file << endl() << "label = \"";
+        file << label;
+        file << "\";";
+
+        file << endl() << "shape = " << shape << ";";
+        file << endl() << "style = " << style << ";";
+        file << endl() << "fillcolor = " << fillcolor << ";";
+        if (color != "")
+            file << endl() << "color = " << color << ";";
+
+        file << down() << endl() << "]";
+        done.emplace(def, def_id(def));
+
+        if (!filtered_ops) {
+            for (size_t i = 0; i < def->num_ops(); i++) {
+                const auto& op = def->op(i);
+                arrow(def_id(def), dump_def(op), "[arrowhead=vee,label=\"o" + std::to_string(i) + "\",fontsize=8,fontcolor=grey]");
+            }
+        } else {
+            for (auto [edge_label, op] : *filtered_ops) {
+                arrow(def_id(def), dump_def(op), "[arrowhead=vee,label=\"" + edge_label + "\",fontsize=8,fontcolor=grey]");
+            }
+        }
+
+        return def_id(def);
+    }
+}
+
+std::string DotPrinter::dump_literal(const Literal* def) {
+    if (!print_literals)
+        return "";
+    assert(def->num_ops() == 0);
+    file << endl() << def_id(def) << " [" << up();
+
+    file << endl() << "label = \"";
+    file << def->to_string();
+    file << "\";";
+
+    file << endl() << "style = dotted;";
+
+    file << down() << endl() << "]";
+
+    done.emplace(def, def_id(def));
+    return def_id(def);
+}
+
+std::string DotPrinter::dump_continuation(Continuation* cont) {
+    done.emplace(cont, def_id(cont));
+    auto intrinsic = cont->intrinsic();
+    file << endl() << def_id(cont) << " [" << up();
+
+    file << endl() << "label = \"";
+    if (cont->is_external())
+        file << "[extern]\\n";
+    auto name = cont->name();
+    if (!cont->is_external())
+        name = cont->unique_name();
+
+    file << name << "(";
+    for (size_t i = 0; i < cont->num_params(); i++) {
+        file << cont->param(i)->type()->to_string() << (i + 1 == cont->num_params() ? "" : ", ");
+    }
+    file << ")";
+
+    file << "\";";
+
+    file << endl() << "shape = rectangle;";
+    if (intrinsic != Intrinsic::None) {
+        file << endl() << "color = lightblue;";
+        file << endl() << "style = filled;";
+    }
+    if (cont->is_external()) {
+        file << endl() << "color = pink;";
+        file << endl() << "style = filled;";
+    }
+
+    file << down() << endl() << "]";
+
+    if (cont->has_body())
+        arrow(def_id(cont), dump_def(cont->body()), "[arrowhead=normal]");
+
+    return def_id(cont);
+}
+
+DEBUG_UTIL void dump_dot_world(World& world) {
+    DotPrinter printer(world);
+    for (auto& external: world.externals()) {
+        printer.dump_def(external.second);
+    }
+    printer.run();
+}
+
+DEBUG_UTIL void dump_dot_def(const Def* def) {
+    DotPrinter printer(def->world());
+    printer.dump_def(def);
+    printer.run();
+}
+
+DEBUG_UTIL void dump_dot_scopes(World& world) {
+    DotPrinter printer(world);
+    ScopesForest forest(world);
+    for (auto& top_level : forest.top_level_scopes()) {
+        auto& scope = forest.get_scope(top_level);
+        printer.print_scope(scope);
+    }
+    printer.run();
+}
+
+DEBUG_UTIL void dump_dot_scope(Scope& scope) {
+    DotPrinter printer(scope.world());
+    printer.single_scope = &scope;
+    printer.print_scope(scope);
+    printer.run();
+}
+
+}
+
+#endif //DOT_DUMP_H
diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
new file mode 100644
index 000000000..1821c32d3
--- /dev/null
+++ b/src/thorin/util/scoped_dump.cpp
@@ -0,0 +1,161 @@
+#include "scoped_dump.h"
+
+namespace thorin {
+
+void ScopedWorld::stream_cont(thorin::Stream& s, Continuation* cont) const {
+    s.fmt(Magenta);
+    if (cont->is_external())
+        s.fmt("extern ");
+    if (cont->is_intrinsic())
+        s.fmt("intrinsic ");
+
+    s.fmt(Red);
+    s.fmt("{}", cont->unique_name());
+    s.fmt(Reset);
+    s.fmt("(");
+    const FnType* t = cont->type();
+    int ret_pi = -1;
+    for (size_t i = 0; i < cont->num_params(); i++) {
+        s.fmt(Yellow);
+        s.fmt("{}: ", cont->param(i)->unique_name());
+        s.fmt(Blue);
+        s.fmt("{}", t->types()[i]);
+        s.fmt(Reset);
+        if (i + 1 < cont->num_params())
+            s.fmt(", ");
+    }
+    s.fmt(")");
+    if (!cont->has_body()) {
+        s.fmt(";");
+        return;
+    }
+
+    s.fmt(" = {{\t\n");
+
+    for (auto p : cont->params())
+        done_.insert(p);
+
+    Scope& sc = forest_.get_scope(cont);
+    scopes_to_defs_[cont] = std::make_unique<std::vector<const Def*>>();
+    auto children = sc.children_scopes();
+    size_t i = 0;
+    for (auto child : children) {
+        stream_cont(s, child);
+        //if (i + 1 < children.size())
+            s.fmt("\n\n");
+        i++;
+    }
+
+    prepare_def(cont, cont->body());
+
+    auto defs = *scopes_to_defs_[cont];
+    stream_defs(s, defs);
+
+    s.fmt("\b\n}}");
+}
+
+void ScopedWorld::prepare_def(Continuation* in, const thorin::Def* def) const {
+    if (done_.contains(def))
+        return;
+    done_.insert(def);
+    if (def->isa_nom<Continuation>())
+        return;
+
+    while (in) {
+        Scope* scope = &forest_.get_scope(in);
+        if (scope->contains(def))
+            break;
+        in = scope->parent_scope();
+    }
+
+    for (auto op : def->ops())
+        prepare_def(in, op);
+
+    if (!in)
+        top_lvl_.push_back(def);
+    else
+        scopes_to_defs_[in]->push_back(def);
+}
+
+void ScopedWorld::stream_op(thorin::Stream& s, const thorin::Def* op) const {
+    //if (is_mem(op))
+    //    s.fmt(Gray);
+    if (op->isa<PrimLit>())
+        s.fmt(Cyan);
+    if (op->isa<Param>())
+        s.fmt(Yellow);
+    if (op->isa<Continuation>())
+        s.fmt(Red);
+    s.fmt("{}", op);
+    s.fmt(Reset);
+}
+
+void ScopedWorld::stream_ops(thorin::Stream& s, Defs ops) const {
+    s.fmt("(");
+    size_t j = 0;
+    for (auto op : ops) {
+        stream_op(s, op);
+        if (j + 1 < ops.size())
+            s.fmt(", ");
+        j++;
+    }
+    s.fmt(")");
+}
+
+void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
+    if (auto app = def->isa<App>()) {
+        stream_op(s, app->callee());
+        stream_ops(s, app->args());
+        return;
+    }
+
+    s.fmt(Green);
+    s.fmt("{}", def->op_name());
+    s.fmt(Reset);
+    stream_ops(s, def->ops());
+}
+
+void ScopedWorld::stream_defs(thorin::Stream& s, std::vector<const Def*>& defs) const {
+    size_t i = 0;
+    for (auto def : defs) {
+        s.fmt("{}: ", def->unique_name());
+        s.fmt(Blue);
+        s.fmt("{}", def->type());
+        s.fmt(Reset);
+        s.fmt(" = ");
+        stream_def(s, def);
+        if (i + 1 < defs.size())
+            s.fmt("\n");
+        i++;
+    }
+}
+
+Stream& ScopedWorld::stream(thorin::Stream& s) const {
+    auto tl = forest_.top_level_scopes();
+    size_t i = 0;
+    for (auto root : tl) {
+        stream_cont(s, root);
+        //if (i + 1 < tl.size())
+            s.fmt("\n\n");
+        i++;
+    }
+
+    stream_defs(s, top_lvl_);
+
+    return s;
+}
+
+void World::dump_scoped() const {
+    ScopedWorld s(*const_cast<World*>(this));
+    s.dump();
+}
+
+void World::dump_scoped_to_disk() const {
+    ScopedWorld s(*const_cast<World*>(this), (ScopedWorld::Config) { false });
+    auto name = this->name() + ".dump";
+    std::ofstream file(name);
+    Stream st(file);
+    s.stream(st);
+}
+
+}
diff --git a/src/thorin/util/scoped_dump.h b/src/thorin/util/scoped_dump.h
new file mode 100644
index 000000000..388c4f60d
--- /dev/null
+++ b/src/thorin/util/scoped_dump.h
@@ -0,0 +1,52 @@
+#include "thorin/world.h"
+#include "thorin/analyses/scope.h"
+#include "thorin/util/stream.h"
+
+namespace thorin {
+
+#define COLORS(C) \
+C(Black, "\u001b[30m") \
+C(Red, "\u001b[31m") \
+C(Green, "\u001b[32m") \
+C(Yellow, "\u001b[33m") \
+C(Blue, "\u001b[34m") \
+C(Magenta, "\u001b[35m") \
+C(Cyan, "\u001b[36m") \
+C(White, "\u001b[37m") \
+C(Reset, "\u001b[0m")  \
+
+struct ScopedWorld : public Streamable<ScopedWorld> {
+    struct Config {
+        bool use_color;
+    };
+
+    ScopedWorld(World& w, Config cfg = { true }) : world_(w), forest_(w), config_(cfg) {
+#define T(n, c) n = cfg.use_color ? c : "";
+        COLORS(T)
+#undef T
+    }
+
+    World& world_;
+    mutable ScopesForest forest_;
+
+    mutable DefSet done_;
+    mutable ContinuationMap<std::unique_ptr<std::vector<const Def*>>> scopes_to_defs_;
+    mutable std::vector<const Def*> top_lvl_;
+    Config config_;
+
+#define T(n, c) const char* n;
+    COLORS(T)
+#undef T
+
+    Stream& stream(Stream&) const;
+private:
+
+    void stream_cont(thorin::Stream& s, Continuation* cont) const;
+    void prepare_def(Continuation* in, const Def* def) const;
+    void stream_op(thorin::Stream&, const Def* op) const;
+    void stream_ops(thorin::Stream& s, Defs defs) const;
+    void stream_def(thorin::Stream& s, const Def* def) const;
+    void stream_defs(thorin::Stream& s, std::vector<const Def*>& defs) const;
+};
+
+}
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 53e0270da..661bc5e92 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -23,7 +23,6 @@
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/verify.h"
 #include "thorin/transform/cleanup_world.h"
-#include "thorin/transform/clone_bodies.h"
 #include "thorin/transform/closure_conversion.h"
 #include "thorin/transform/codegen_prepare.h"
 #include "thorin/transform/dead_load_opt.h"
@@ -1310,7 +1309,6 @@ void Thorin::opt() {
     RUN_PASS(cleanup())
     RUN_PASS(while (partial_evaluation(world(), true))); // lower2cff
     RUN_PASS(flatten_tuples(*this))
-    RUN_PASS(clone_bodies(world()))
     RUN_PASS(split_slots(*this))
     RUN_PASS(closure_conversion(world()))
     RUN_PASS(lift_builtins(*this))
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 56dda1c30..0e154afb6 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -93,6 +93,13 @@ class World : public Streamable<World> {
     void make_internal(Def* cont) { assert(&cont->world() == this); data_.externals_.erase(cont->unique_name()); }
     bool is_external(const Def* cont) { return data_.externals_.contains(cont->unique_name()); }
     Def* lookup(const std::string& name) { return data_.externals_.lookup(name).value_or(nullptr); }
+    DEBUG_UTIL Continuation* find_cont(const char* name) {
+        for (auto cont : copy_continuations()) {
+            if (cont->unique_name() == name)
+                return cont;
+        }
+        return nullptr;
+    }
     //@}
 
     // types
@@ -293,6 +300,8 @@ class World : public Streamable<World> {
 
     /// @name logging
     //@{
+    void dump_scoped() const;
+    void dump_scoped_to_disk() const;
     Stream& stream(Stream&) const;
     Stream& stream() { return *stream_; }
     /// Writes to a file named @c name().

From 7e090163392c2f9e5582e1527c03dde468cf7188 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 11 Sep 2023 10:01:21 +0200
Subject: [PATCH 182/342] improved diagnostics for PE

---
 src/thorin/continuation.cpp                 | 2 +-
 src/thorin/transform/partial_evaluation.cpp | 1 +
 src/thorin/world.h                          | 1 +
 3 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index d82041aa8..b42263b9b 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -165,7 +165,7 @@ const Param* Continuation::ret_param() const {
 }
 
 void Continuation::destroy(const char* cause) {
-    world().VLOG("{} has been destroyed by {}", this, cause);
+    world().ddef(this, "{} has been destroyed by {}", this, cause);
     destroy_filter();
     unset_op(0);
     set_op(0, world().bottom(world().bottom_type()));
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 0968e45b1..bf5af1f30 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -166,6 +166,7 @@ bool PartialEvaluator::run() {
                     Continuation*& target = p.first->second;
                     // create new specialization if not found in cache
                     if (p.second) {
+                        world_.ddef(continuation, "Specializing call to {}", callee);
                         target = drop(callee, specialize);
                         todo = true;
                     }
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 0e154afb6..ae070ce04 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -328,6 +328,7 @@ class World : public Streamable<World> {
     // Ignore log
     void ignore() {}
 
+    template<class... Args> void ddef(const Def* def, const char* fmt, Args&&... args) { log(LogLevel::Debug, def->loc(), fmt, std::forward<Args&&>(args)...); }
     template<class... Args> void idef(const Def* def, const char* fmt, Args&&... args) { log(LogLevel::Info, def->loc(), fmt, std::forward<Args&&>(args)...); }
     template<class... Args> void wdef(const Def* def, const char* fmt, Args&&... args) { log(LogLevel::Warn, def->loc(), fmt, std::forward<Args&&>(args)...); }
     template<class... Args> void edef(const Def* def, const char* fmt, Args&&... args) { error(def->loc(), fmt, std::forward<Args&&>(args)...); }

From 32ce4d7127245c9d0c489c9f1e3a119c84472ce8 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 11 Sep 2023 10:01:40 +0200
Subject: [PATCH 183/342] reintroduce boundary check in PE

---
 src/thorin/transform/partial_evaluation.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index bf5af1f30..92ab41fa7 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -19,11 +19,14 @@ class PartialEvaluator {
     PartialEvaluator(World& world, bool lower2cff)
         : world_(world)
         , lower2cff_(lower2cff)
+        , boundary_(Def::gid_counter())
     {}
 
     World& world() { return world_; }
     bool run();
     void enqueue(Continuation* continuation) {
+        if (continuation->gid() < 2 * boundary_ && done_.emplace(continuation).second)
+            queue_.push(continuation);
         queue_.push(continuation);
     }
     void eat_pe_info(Continuation*);
@@ -34,6 +37,7 @@ class PartialEvaluator {
     HashMap<const App*, Continuation*, HashApp> cache_;
     ContinuationSet done_;
     unique_queue<ContinuationSet> queue_;
+    size_t boundary_;
 };
 
 class CondEval {

From 09fe3d52792bcd4ccf96018b15f58cbc93910e72 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Sep 2023 16:27:49 +0200
Subject: [PATCH 184/342] LLVM: Make the codegen more value-oriented,
 generalise jumping code

---
 src/thorin/be/llvm/amdgpu.cpp    |   2 +-
 src/thorin/be/llvm/amdgpu.h      |   2 +-
 src/thorin/be/llvm/llvm.cpp      | 274 ++++++++++++++++---------------
 src/thorin/be/llvm/llvm.h        |  33 ++--
 src/thorin/be/llvm/nvvm.cpp      |   2 +-
 src/thorin/be/llvm/nvvm.h        |   2 +-
 src/thorin/be/llvm/parallel.cpp  |  22 +--
 src/thorin/be/llvm/runtime.cpp   |   4 +-
 src/thorin/be/llvm/runtime.h     |   2 +-
 src/thorin/be/llvm/vectorize.cpp |   4 +-
 10 files changed, 176 insertions(+), 171 deletions(-)

diff --git a/src/thorin/be/llvm/amdgpu.cpp b/src/thorin/be/llvm/amdgpu.cpp
index d07f9f389..37468fe72 100644
--- a/src/thorin/be/llvm/amdgpu.cpp
+++ b/src/thorin/be/llvm/amdgpu.cpp
@@ -86,7 +86,7 @@ llvm::Value* AMDGPUCodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const Math
     return call;
 }
 
-Continuation* AMDGPUCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
+llvm::Value* AMDGPUCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
     return emit_reserve_shared(irbuilder, continuation, true);
 }
 
diff --git a/src/thorin/be/llvm/amdgpu.h b/src/thorin/be/llvm/amdgpu.h
index fd5bd5779..999c60e81 100644
--- a/src/thorin/be/llvm/amdgpu.h
+++ b/src/thorin/be/llvm/amdgpu.h
@@ -22,7 +22,7 @@ class AMDGPUCodeGen : public CodeGen {
     llvm::Function* emit_fun_decl(Continuation*) override;
     llvm::Value* emit_global(const Global*) override;
     llvm::Value* emit_mathop(llvm::IRBuilder<>&, const MathOp*) override;
-    Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
+    llvm::Value* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
     std::string get_alloc_name() const override { return "malloc"; }
 
     const Cont2Config& kernel_config_;
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index e1b36028a..93f825fc2 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -449,36 +449,103 @@ void CodeGen::finalize(const Scope&) {
         defs_.erase(def);
 }
 
-void CodeGen::emit_epilogue(Continuation* continuation) {
-    assert(continuation->has_body());
-    auto body = continuation->body();
-
-    auto& [bb, ptr_irbuilder] = cont2bb_[continuation];
-    auto& irbuilder = *ptr_irbuilder;
-
-    if (body->callee() == entry_->ret_param()) { // return
-        std::vector<llvm::Value*> values;
-        std::vector<llvm::Type *> types;
+std::vector<llvm::Value*> CodeGen::split_values(llvm::IRBuilder<>& irbuilder, Types domain, llvm::Value* value) {
+    size_t n = 0;
+    for (auto t : domain) {
+        if (t == world().unit_type() || t->isa<MemType>())
+            continue;
+        n++;
+    }
 
-        for (auto arg : body->args()) {
-            if (auto val = emit_unsafe(arg)) {
-                values.emplace_back(val);
-                types.emplace_back(val->getType());
+    switch (n) {
+        case 0: return {};
+        case 1: return { value };
+        default: {
+            std::vector<llvm::Value*> values;
+            values.resize(n);
+            for (size_t i = 0; i < n; i++) {
+                values[i] = irbuilder.CreateExtractValue(value, i);
             }
+            return values;
         }
+    }
+}
 
-        switch (values.size()) {
+llvm::CallInst* CodeGen::emit_call(llvm::IRBuilder<>& irbuilder, const Def* callee, std::vector<llvm::Value*>& args) {
+    if (callee == entry_->ret_param()) { // normal return
+        std::vector<llvm::Type *> types;
+        for (auto val : args)
+            types.emplace_back(val->getType());
+        switch (args.size()) {
             case 0:  irbuilder.CreateRetVoid();      break;
-            case 1:  irbuilder.CreateRet(values[0]); break;
-            default:
+            case 1:  irbuilder.CreateRet(args[0]); break;
+            default: {
                 llvm::Value* agg = llvm::UndefValue::get(llvm::StructType::get(context(), types));
 
-                for (size_t i = 0, e = values.size(); i != e; ++i)
-                    agg = irbuilder.CreateInsertValue(agg, values[i], { unsigned(i) });
+                for (size_t i = 0, e = args.size(); i != e; ++i)
+                    agg = irbuilder.CreateInsertValue(agg, args[i], { unsigned(i) });
 
                 irbuilder.CreateRet(agg);
+            }
         }
-    } else if (body->callee() == world().branch()) {
+        return nullptr;
+    } else if (callee->isa<Bottom>()) {
+        irbuilder.CreateUnreachable();
+        return nullptr;
+    } else if (auto cont = callee->isa_nom<Continuation>(); cont && scope_->contains(cont) && cont != entry_) {
+        assert(cont->is_basicblock());
+        size_t j = 0, i = 0;
+        for (auto t: cont->type()->types()) {
+            i++;
+            assert(t->order() == 0);
+            if (t->isa<MemType>() || t == world().unit_type())
+                continue;
+            emit_phi_arg(irbuilder, cont->param(i - 1), args[j++]);
+        }
+        irbuilder.CreateBr(cont2bb(cont));
+        return nullptr;
+    } else if (auto closure_t = callee->type()->isa<ClosureType>()) {
+        auto closure = emit(callee);
+        auto fnt = world().fn_type(concat(closure_t->types(), closure_t->as<Type>()));
+        args.push_back(closure);
+        auto func = irbuilder.CreateExtractValue(closure, 0);
+        auto call = irbuilder.CreateCall(llvm::cast<llvm::FunctionType>(convert(fnt)), irbuilder.CreatePointerCast(func, convert(world().ptr_type(fnt))), args);
+        return call;
+    } else if (callee->type()->tag() == Node_FnType) {
+        auto call = irbuilder.CreateCall(llvm::cast<llvm::Function>(emit(callee)), args);
+        if (cont->is_exported())
+            call->setCallingConv(kernel_calling_convention_);
+        else if (cont->cc() == CC::Device)
+            call->setCallingConv(device_calling_convention_);
+        else
+            call->setCallingConv(function_calling_convention_);
+        return call;
+    }
+
+    THORIN_UNREACHABLE;
+}
+
+static const Type* mangle_for_codegen(World& world, ArrayRef<const Type*> ret_types) {
+    // treat non-returning calls as if they return nothing, for now
+    std::vector<const Type*> types;
+    for (auto op: ret_types) {
+        assert(op->order() == 0);
+        if (op->isa<MemType>() || is_type_unit(op)) continue;
+        types.push_back(op);
+    }
+    return world.tuple_type(types);
+}
+
+void CodeGen::emit_epilogue(Continuation* continuation) {
+    assert(continuation->has_body());
+    auto body = continuation->body();
+
+    auto& [bb, ptr_irbuilder] = cont2bb_[continuation];
+    auto& irbuilder = *ptr_irbuilder;
+
+    llvm::CallInst* call_instr = nullptr;
+
+    if (body->callee() == world().branch()) {
         auto mem = body->arg(0);
         emit_unsafe(mem);
 
@@ -499,18 +566,11 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             auto case_bb    = cont2bb(arg->op(1)->as_nom<Continuation>());
             match->addCase(case_const, case_bb);
         }
-    } else if (body->callee()->isa<Bottom>()) {
-        irbuilder.CreateUnreachable();
-    } else if (auto callee = body->callee()->isa_nom<Continuation>(); callee && callee->is_basicblock()) { // ordinary jump
-        for (size_t i = 0, e = body->num_args(); i != e; ++i) {
-            if (auto val = emit_unsafe(body->arg(i))) emit_phi_arg(irbuilder, callee->param(i), val);
-        }
-        irbuilder.CreateBr(cont2bb(callee));
     } else if (auto callee = body->callee()->isa_nom<Continuation>(); callee && callee->is_intrinsic()) {
-        auto ret_continuation = emit_intrinsic(irbuilder, continuation);
-        irbuilder.CreateBr(cont2bb(ret_continuation));
-    } else { // function/closure call
-        // put all first-order args into an array
+        auto args = emit_intrinsic(irbuilder, continuation);
+        call_instr = emit_call(irbuilder, body->arg(callee->ret_param()->index()), args);
+    } else {
+        // plain continuation call: we can just emit all the arguments
         std::vector<llvm::Value*> args;
         const Def* ret_arg = nullptr;
         for (auto arg : body->args()) {
@@ -523,61 +583,26 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             }
         }
 
-        llvm::CallInst* call = nullptr;
-        if (auto callee = body->callee()->isa_nom<Continuation>()) {
-            call = irbuilder.CreateCall(llvm::cast<llvm::Function>(emit(callee)), args);
-            if (callee->is_exported())
-                call->setCallingConv(kernel_calling_convention_);
-            else if (callee->cc() == CC::Device)
-                call->setCallingConv(device_calling_convention_);
-            else
-                call->setCallingConv(function_calling_convention_);
-        } else {
-            // must be a closure
-            auto closure = emit(body->callee());
-            args.push_back(irbuilder.CreateExtractValue(closure, 1));
-            auto func = irbuilder.CreateExtractValue(closure, 0);
-            auto call_type = convert_closure_type(body->callee()->type());
-            call = irbuilder.CreateCall(call_type, func, args);
-        }
-
-        // must be call + continuation --- call + return has been removed by codegen_prepare
-        auto succ = ret_arg->as_nom<Continuation>();
+        call_instr = emit_call(irbuilder, body->callee(), args);
+        if (body->callee()->type()->as<FnType>()->is_returning()) {
+            assert(call_instr && "returning calls always involve one of those");
+            assert(ret_arg && "we need a return argument too!");
 
-        size_t n = 0;
-        const Param* last_param = nullptr;
-        for (auto param : succ->params()) {
-            if (is_mem(param) || is_unit(param))
-                continue;
-            last_param = param;
-            n++;
+            auto ret_args = split_values(irbuilder, ret_arg->type()->as<FnType>()->types(), call_instr);
+            call_instr = emit_call(irbuilder, ret_arg, ret_args);
         }
+    }
 
-        if (n == 0) {
-            irbuilder.CreateBr(cont2bb(succ));
-        } else if (n == 1) {
-            irbuilder.CreateBr(cont2bb(succ));
-            emit_phi_arg(irbuilder, last_param, call);
-        } else {
-            Array<llvm::Value*> extracts(n);
-            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-                auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
-                    continue;
-                extracts[j] = irbuilder.CreateExtractValue(call, unsigned(j));
-                j++;
-            }
-
-            irbuilder.CreateBr(cont2bb(succ));
-
-            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-                auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
-                    continue;
-                emit_phi_arg(irbuilder, param, extracts[j]);
-                j++;
-            }
-        }
+    if (call_instr) {
+        // we need to add a dummy return terminator if the last instruction is a call
+        if (entry_->type()->is_returning()) {
+            auto entry_return_t = mangle_for_codegen(world(), entry_->ret_param()->type()->as<FnType>()->types());
+            if (entry_return_t != world().unit_type()) {
+                irbuilder.CreateRet(llvm::UndefValue::get(convert(entry_return_t)));
+            } else
+                irbuilder.CreateRetVoid();
+        } else
+            irbuilder.CreateRetVoid();
     }
 
     // new insert point is just before the terminator for all other instructions we have to add later on
@@ -586,12 +611,14 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
 
 llvm::Value* CodeGen::emit_constant(const Def* def) {
     auto irbuilder = llvm::IRBuilder(context());
-    return emit_builder(irbuilder, def);
+    auto val = emit_builder(irbuilder, def);
+    return val;
 }
 
 llvm::Value* CodeGen::emit_bb(BB& bb, const Def* def) {
     auto& irbuilder = *bb.second;
-    return emit_builder(irbuilder, def);
+    auto val = emit_builder(irbuilder, def);
+    return val;
 }
 
 llvm::Value* CodeGen::emit_builder(llvm::IRBuilder<>& irbuilder, const Def* def) {
@@ -1221,7 +1248,7 @@ llvm::Value* CodeGen::emit_assembly(llvm::IRBuilder<>& irbuilder, const Assembly
  * emit intrinsic
  */
 
-Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+std::vector<llvm::Value*> CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     auto callee = body->callee()->as_nom<Continuation>();
@@ -1236,33 +1263,35 @@ Continuation* CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder, Continuation
     }
 
     switch (callee->intrinsic()) {
-        case Intrinsic::Atomic:       return emit_atomic(irbuilder, continuation);
-        case Intrinsic::AtomicLoad:   return emit_atomic_load(irbuilder, continuation);
-        case Intrinsic::AtomicStore:  return emit_atomic_store(irbuilder, continuation);
+        case Intrinsic::Atomic:       return { emit_atomic(irbuilder, continuation) };
+        case Intrinsic::AtomicLoad:   return { emit_atomic_load(irbuilder, continuation) };
+        case Intrinsic::AtomicStore:  emit_atomic_store(irbuilder, continuation); break;
         case Intrinsic::CmpXchg:      return emit_cmpxchg(irbuilder, continuation, false);
         case Intrinsic::CmpXchgWeak:  return emit_cmpxchg(irbuilder, continuation, true);
-        case Intrinsic::Fence:        return emit_fence(irbuilder, continuation);
-        case Intrinsic::Reserve:      return emit_reserve(irbuilder, continuation);
-        case Intrinsic::CUDA:         return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation);
-        case Intrinsic::NVVM:         return runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation);
-        case Intrinsic::OpenCL:       return runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation);
-        case Intrinsic::AMDGPU:       return runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation);
-        case Intrinsic::ShadyCompute: return runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation);
-        case Intrinsic::HLS:          return emit_hls(irbuilder, continuation);
-        case Intrinsic::Parallel:     return emit_parallel(irbuilder, continuation);
-        case Intrinsic::Fibers:       return emit_fibers(irbuilder, continuation);
-        case Intrinsic::Spawn:        return emit_spawn(irbuilder, continuation);
-        case Intrinsic::Sync:         return emit_sync(irbuilder, continuation);
+        case Intrinsic::Fence:        emit_fence(irbuilder, continuation); break;
+        case Intrinsic::Reserve:      return { emit_reserve(irbuilder, continuation) };
+        case Intrinsic::CUDA:         runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation); break;
+        case Intrinsic::NVVM:         runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation); break;
+        case Intrinsic::OpenCL:       runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation); break;
+        case Intrinsic::AMDGPU:       runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation); break;
+        case Intrinsic::ShadyCompute: runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation); break;
+        case Intrinsic::HLS:          emit_hls(irbuilder, continuation);      break;
+        case Intrinsic::Parallel:     emit_parallel(irbuilder, continuation); break;
+        case Intrinsic::Fibers:       emit_fibers(irbuilder, continuation);   break;
+        case Intrinsic::Spawn:        return { emit_spawn(irbuilder, continuation) };
+        case Intrinsic::Sync:         emit_sync(irbuilder, continuation);     break;
 #if THORIN_ENABLE_RV
-        case Intrinsic::Vectorize:   return emit_vectorize_continuation(irbuilder, continuation);
+        case Intrinsic::Vectorize:    emit_vectorize_continuation(irbuilder, continuation); break;
 #else
-        case Intrinsic::Vectorize:   throw std::runtime_error("rebuild with RV support");
+        case Intrinsic::Vectorize:    throw std::runtime_error("rebuild with RV support");
 #endif
         default: THORIN_UNREACHABLE;
     }
+
+    return {};
 }
 
-Continuation* CodeGen::emit_atomic(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+llvm::Value* CodeGen::emit_atomic(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() == 7 && "required arguments are missing");
@@ -1282,13 +1311,11 @@ Continuation* CodeGen::emit_atomic(llvm::IRBuilder<>& irbuilder, Continuation* c
     assert(int(llvm::AtomicOrdering::NotAtomic) <= int(order_tag) && int(order_tag) <= int(llvm::AtomicOrdering::SequentiallyConsistent) && "unsupported atomic ordering");
     auto order = (llvm::AtomicOrdering)order_tag;
     auto scope = body->arg(5)->as<ConvOp>()->from()->as<Global>()->init()->as<DefiniteArray>();
-    auto cont = body->arg(6)->as_nom<Continuation>();
     auto call = irbuilder.CreateAtomicRMW(binop, ptr, val, llvm::MaybeAlign(), order, context().getOrInsertSyncScopeID(scope->as_string()));
-    emit_phi_arg(irbuilder, cont->param(1), call);
-    return cont;
+    return call;
 }
 
-Continuation* CodeGen::emit_atomic_load(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+llvm::Value* CodeGen::emit_atomic_load(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() == 5 && "required arguments are missing");
@@ -1303,11 +1330,10 @@ Continuation* CodeGen::emit_atomic_load(llvm::IRBuilder<>& irbuilder, Continuati
     auto align = module().getDataLayout().getABITypeAlign(load_type);
     load->setAlignment(align);
     load->setAtomic(order, context().getOrInsertSyncScopeID(scope->as_string()));
-    emit_phi_arg(irbuilder, cont->param(1), load);
-    return cont;
+    return load;
 }
 
-Continuation* CodeGen::emit_atomic_store(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_atomic_store(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() == 6 && "required arguments are missing");
@@ -1322,10 +1348,9 @@ Continuation* CodeGen::emit_atomic_store(llvm::IRBuilder<>& irbuilder, Continuat
     auto align = module().getDataLayout().getABITypeAlign(convert(body->arg(2)->type()));
     store->setAlignment(align);
     store->setAtomic(order, context().getOrInsertSyncScopeID(scope->as_string()));
-    return cont;
 }
 
-Continuation* CodeGen::emit_cmpxchg(llvm::IRBuilder<>& irbuilder, Continuation* continuation, bool is_weak) {
+std::vector<llvm::Value*> CodeGen::emit_cmpxchg(llvm::IRBuilder<>& irbuilder, Continuation* continuation, bool is_weak) {
     assert(continuation->has_body());
     auto body = continuation->body();
 
@@ -1345,12 +1370,10 @@ Continuation* CodeGen::emit_cmpxchg(llvm::IRBuilder<>& irbuilder, Continuation*
     auto cont = body->arg(7)->as_nom<Continuation>();
     auto call = irbuilder.CreateAtomicCmpXchg(ptr, cmp, val, llvm::MaybeAlign(), success_order, failure_order, context().getOrInsertSyncScopeID(scope->as_string()));
     call->setWeak(is_weak);
-    emit_phi_arg(irbuilder, cont->param(1), irbuilder.CreateExtractValue(call, 0));
-    emit_phi_arg(irbuilder, cont->param(2), irbuilder.CreateExtractValue(call, 1));
-    return cont;
+    return { irbuilder.CreateExtractValue(call, 0), irbuilder.CreateExtractValue(call, 1) };
 }
 
-Continuation* CodeGen::emit_fence(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_fence(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() == 4 && "required arguments are missing");
@@ -1360,15 +1383,14 @@ Continuation* CodeGen::emit_fence(llvm::IRBuilder<>& irbuilder, Continuation* co
     auto scope = body->arg(2)->as<ConvOp>()->from()->as<Global>()->init()->as<DefiniteArray>();
     auto cont = body->arg(3)->as_nom<Continuation>();
     irbuilder.CreateFence(order, context().getOrInsertSyncScopeID(scope->as_string()));
-    return cont;
 }
 
-Continuation* CodeGen::emit_reserve(llvm::IRBuilder<>&, const Continuation* continuation) {
+llvm::Value* CodeGen::emit_reserve(llvm::IRBuilder<>&, const Continuation* continuation) {
     world().edef(continuation, "reserve_shared: only allowed in device code"); // TODO debug
     THORIN_UNREACHABLE;
 }
 
-Continuation* CodeGen::emit_reserve_shared(llvm::IRBuilder<>& irbuilder, const Continuation* continuation, bool init_undef) {
+llvm::Value* CodeGen::emit_reserve_shared(llvm::IRBuilder<>& irbuilder, const Continuation* continuation, bool init_undef) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() == 3 && "required arguments are missing");
@@ -1385,31 +1407,23 @@ Continuation* CodeGen::emit_reserve_shared(llvm::IRBuilder<>& irbuilder, const C
     std::replace(name.begin(), name.end(), '.', '_');
     auto global = emit_global_variable(smem_type, name, 3, init_undef);
     auto call = irbuilder.CreatePointerCast(global, type);
-    emit_phi_arg(irbuilder, cont->param(1), call);
-    return cont;
+    return call;
 }
 
 /*
  * backend-specific stuff
  */
 
-Continuation* CodeGen::emit_hls(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_hls(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     std::vector<llvm::Value*> args(body->num_args()-3);
-    Continuation* ret = nullptr;
     for (size_t i = 2, j = 0; i < body->num_args(); ++i) {
-        if (auto cont = body->arg(i)->isa_nom<Continuation>()) {
-            ret = cont;
-            continue;
-        }
         args[j++] = emit(body->arg(i));
     }
     auto callee = body->arg(1)->as<Global>()->init()->as_nom<Continuation>();
     world().make_external(callee);
     irbuilder.CreateCall(emit_fun_decl(callee), args);
-    assert(ret);
-    return ret;
 }
 
 /*
diff --git a/src/thorin/be/llvm/llvm.h b/src/thorin/be/llvm/llvm.h
index c146f37d1..7d5414331 100644
--- a/src/thorin/be/llvm/llvm.h
+++ b/src/thorin/be/llvm/llvm.h
@@ -83,8 +83,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<llvm::Value*, llv
     virtual llvm::Value* emit_lea     (llvm::IRBuilder<>&, const LEA*);
     virtual llvm::Value* emit_assembly(llvm::IRBuilder<>&, const Assembly* assembly);
 
-    virtual Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*);
-    Continuation* emit_reserve_shared(llvm::IRBuilder<>&, const Continuation*, bool=false);
+    virtual llvm::Value* emit_reserve(llvm::IRBuilder<>&, const Continuation*);
+    llvm::Value* emit_reserve_shared(llvm::IRBuilder<>&, const Continuation*, bool=false);
 
     virtual std::string get_alloc_name() const = 0;
     llvm::BasicBlock* cont2bb(Continuation* cont) { return cont2bb_[cont].first; }
@@ -96,23 +96,26 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<llvm::Value*, llv
     void verify() const;
     void create_loop(llvm::IRBuilder<>&, llvm::Value*, llvm::Value*, llvm::Value*, llvm::Function*, std::function<void(llvm::Value*)>);
     llvm::Value* create_tmp_alloca(llvm::IRBuilder<>&, llvm::Type*, std::function<llvm::Value* (llvm::AllocaInst*)>);
-
     llvm::Value* call_math_function(llvm::IRBuilder<>&, const MathOp*, const std::string&);
 
+    std::vector<llvm::Value*> split_values(llvm::IRBuilder<>&, Types domain, llvm::Value* value);
+    /// Emits a 'call' using already emitted arguments. Returns the call instruction if appropriate
+    llvm::CallInst* emit_call(llvm::IRBuilder<>&, const Def* callee, std::vector<llvm::Value*>& args);
+
 private:
     Continuation* emit_peinfo(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_intrinsic(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_hls(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_parallel(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_fibers(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_spawn(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_sync(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_vectorize_continuation(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_atomic(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_cmpxchg(llvm::IRBuilder<>&, Continuation*, bool);
-    Continuation* emit_fence(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_atomic_load(llvm::IRBuilder<>&, Continuation*);
-    Continuation* emit_atomic_store(llvm::IRBuilder<>&, Continuation*);
+    std::vector<llvm::Value*> emit_intrinsic(llvm::IRBuilder<>&, Continuation*);
+    void emit_hls(llvm::IRBuilder<>&, Continuation*);
+    void emit_parallel(llvm::IRBuilder<>&, Continuation*);
+    void emit_fibers(llvm::IRBuilder<>&, Continuation*);
+    llvm::Value* emit_spawn(llvm::IRBuilder<>&, Continuation*);
+    void emit_sync(llvm::IRBuilder<>&, Continuation*);
+    void emit_vectorize_continuation(llvm::IRBuilder<>&, Continuation*);
+    llvm::Value* emit_atomic(llvm::IRBuilder<>&, Continuation*);
+    std::vector<llvm::Value*> emit_cmpxchg(llvm::IRBuilder<>&, Continuation*, bool);
+    void emit_fence(llvm::IRBuilder<>&, Continuation*);
+    llvm::Value* emit_atomic_load(llvm::IRBuilder<>&, Continuation*);
+    void emit_atomic_store(llvm::IRBuilder<>&, Continuation*);
     llvm::Value* emit_bitcast(llvm::IRBuilder<>&, const Def*, const Type*);
     void emit_vectorize(u32, llvm::Function*, llvm::CallInst*);
     void emit_phi_arg(llvm::IRBuilder<>&, const Param*, llvm::Value*);
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index ec41f6e58..20dd5e2d4 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -242,7 +242,7 @@ llvm::Value* NVVMCodeGen::emit_lea(llvm::IRBuilder<>& irbuilder, const LEA* lea)
     }
 }
 
-Continuation* NVVMCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
+llvm::Value* NVVMCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
     return emit_reserve_shared(irbuilder, continuation);
 }
 
diff --git a/src/thorin/be/llvm/nvvm.h b/src/thorin/be/llvm/nvvm.h
index 204ce6648..742f6127c 100644
--- a/src/thorin/be/llvm/nvvm.h
+++ b/src/thorin/be/llvm/nvvm.h
@@ -28,7 +28,7 @@ class NVVMCodeGen : public CodeGen {
     llvm::Value* emit_lea(llvm::IRBuilder<>&,    const LEA*) override;
     llvm::Value* emit_mathop(llvm::IRBuilder<>&, const MathOp*) override;
 
-    Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
+    llvm::Value* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
 
     llvm::Value* emit_global(const Global*) override;
 
diff --git a/src/thorin/be/llvm/parallel.cpp b/src/thorin/be/llvm/parallel.cpp
index c9012525f..aae4a0208 100644
--- a/src/thorin/be/llvm/parallel.cpp
+++ b/src/thorin/be/llvm/parallel.cpp
@@ -12,7 +12,7 @@ enum {
     PAR_NUM_ARGS
 };
 
-Continuation* CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     // Emit memory dependencies up to this point
@@ -36,7 +36,7 @@ Continuation* CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation*
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(PAR_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->body()->callee()->type()->as<FnType>()->types().skip_front(PAR_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -91,8 +91,6 @@ Continuation* CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation*
 
     // restore old insert point
     irbuilder.SetInsertPoint(old_bb);
-
-    return body->arg(PAR_ARG_RETURN)->as_nom<Continuation>();
 }
 
 enum {
@@ -105,7 +103,7 @@ enum {
     FIB_NUM_ARGS
 };
 
-Continuation* CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     // Emit memory dependencies up to this point
@@ -183,8 +181,6 @@ Continuation* CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* c
 
     // restore old insert point
     irbuilder.SetInsertPoint(old_bb);
-
-    return body->arg(FIB_ARG_RETURN)->as_nom<Continuation>();
 }
 
 enum {
@@ -194,7 +190,7 @@ enum {
     SPAWN_NUM_ARGS
 };
 
-Continuation* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+llvm::Value* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() >= SPAWN_NUM_ARGS && "required arguments are missing");
@@ -213,7 +209,7 @@ Continuation* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* co
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(SPAWN_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->body()->callee()->type()->as<FnType>()->types().skip_front(SPAWN_NUM_ARGS)));
     llvm::Value* closure = nullptr;
     if (closure_type->isStructTy()) {
         closure = llvm::UndefValue::get(closure_type);
@@ -262,10 +258,7 @@ Continuation* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* co
     // restore old insert point
     irbuilder.SetInsertPoint(old_bb);
 
-    // bind parameter of continuation to received handle
-    auto cont = body->arg(SPAWN_ARG_RETURN)->as_nom<Continuation>();
-    emit_phi_arg(irbuilder, cont->param(1), call);
-    return cont;
+    return call;
 }
 
 enum {
@@ -275,7 +268,7 @@ enum {
     SYNC_NUM_ARGS
 };
 
-Continuation* CodeGen::emit_sync(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_sync(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     assert(body->num_args() == SYNC_NUM_ARGS && "wrong number of arguments");
@@ -285,7 +278,6 @@ Continuation* CodeGen::emit_sync(llvm::IRBuilder<>& irbuilder, Continuation* con
 
     auto id = emit(body->arg(SYNC_ARG_ID));
     runtime_->sync_thread(*this, irbuilder, id);
-    return body->arg(SYNC_ARG_RETURN)->as_nom<Continuation>();
 }
 
 }
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index bc0267a73..375713091 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -60,7 +60,7 @@ static bool contains_ptrtype(const Type* type) {
     }
 }
 
-Continuation* Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Platform platform, const std::string& ext, Continuation* continuation) {
+void Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Platform platform, const std::string& ext, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     // to-target is the desired kernel call
@@ -184,8 +184,6 @@ Continuation* Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& buil
                   grid_size, block_size,
                   args, sizes, aligns, allocs, types,
                   builder.getInt32(num_kernel_args));
-
-    return body->arg(LaunchArgs::Return)->as_nom<Continuation>();
 }
 
 llvm::Value* Runtime::launch_kernel(
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index d99de5c27..7c4516e8d 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -51,7 +51,7 @@ class Runtime {
     /// Emits a call to anydsl_sync_thread.
     llvm::Value* sync_thread(CodeGen&, llvm::IRBuilder<>&, llvm::Value* id);
 
-    Continuation* emit_host_code(
+    void emit_host_code(
         CodeGen& code_gen, llvm::IRBuilder<>& builder,
         Platform platform, const std::string& ext, Continuation* continuation);
 
diff --git a/src/thorin/be/llvm/vectorize.cpp b/src/thorin/be/llvm/vectorize.cpp
index 591f518c1..f2c74e619 100644
--- a/src/thorin/be/llvm/vectorize.cpp
+++ b/src/thorin/be/llvm/vectorize.cpp
@@ -48,7 +48,7 @@ struct VectorizeArgs {
     };
 };
 
-Continuation* CodeGen::emit_vectorize_continuation(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
+void CodeGen::emit_vectorize_continuation(llvm::IRBuilder<>& irbuilder, Continuation* continuation) {
     assert(continuation->has_body());
     auto body = continuation->body();
     auto target = body->callee()->as_nom<Continuation>();
@@ -93,8 +93,6 @@ Continuation* CodeGen::emit_vectorize_continuation(llvm::IRBuilder<>& irbuilder,
         world().edef(body->arg(VectorizeArgs::Length), "vector length must be known at compile-time");
     u32 vector_length_constant = body->arg(VectorizeArgs::Length)->as<PrimLit>()->qu32_value();
     vec_todo_.emplace_back(vector_length_constant, emit_fun_decl(kernel), simd_kernel_call);
-
-    return body->arg(VectorizeArgs::Return)->as_nom<Continuation>();
 }
 
 void CodeGen::emit_vectorize(u32 vector_length, llvm::Function* kernel_func, llvm::CallInst* simd_kernel_call) {

From 21422b59704bcd38d285cb45cc1c09349f7b9e32 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 14 Sep 2023 16:28:29 +0200
Subject: [PATCH 185/342] Mark Continuations dropped as device code as being
 external. No special calling convention needed any more.

---
 src/thorin/be/codegen.cpp   | 2 +-
 src/thorin/be/json/json.cpp | 2 --
 src/thorin/continuation.h   | 1 -
 3 files changed, 1 insertion(+), 4 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index e1a155256..8b992d8d3 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -48,7 +48,7 @@ static void get_kernel_configs(
             return false;
         }, true);
 
-        continuation->attributes().cc = CC::DeviceHostCode;
+        continuation->world().make_external(continuation);
         continuation->destroy("codegen");
     }
 }
diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 2bd09fb21..ea0ae03c8 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -263,8 +263,6 @@ class DefTable {
                     else
                         forward_decl["external"] = cont->name();
                 }
-                if (cont->cc() == CC::DeviceHostCode)
-                    forward_decl["device"] = cont->name();
                 decl_table.push_back(forward_decl);
 
                 if(cont->has_body()) {
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 010d9bd8f..3a1f03281 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -86,7 +86,6 @@ class App : public Def {
 enum class CC : uint8_t {
     C,              ///< C calling convention.
     Device,         ///< Device calling convention. These are special functions only available on a particular device.
-    DeviceHostCode, ///< Calling convention to denote continuations that are generated as device code.
     Internal,       ///< External, but only for linking with artic or anyopt.
 };
 

From 8fe13b85428667f475191da586e29339d3bb4914 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 15 Sep 2023 14:02:23 +0200
Subject: [PATCH 186/342] Bug fix: Include config.h in debug.h

---
 src/thorin/debug.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/debug.h b/src/thorin/debug.h
index 05613fccb..ad23841d7 100644
--- a/src/thorin/debug.h
+++ b/src/thorin/debug.h
@@ -4,6 +4,7 @@
 #include <string>
 #include <tuple>
 
+#include "thorin/config.h"
 #include "thorin/util/stream.h"
 
 namespace thorin {

From e8148592473c1fb39ccf718e0c0a708053e4424a Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 15 Sep 2023 14:56:11 +0200
Subject: [PATCH 187/342] Only exported continuations need to be generated on
 top level.

---
 src/thorin/be/c/c.cpp                    | 2 +-
 src/thorin/be/llvm/llvm.cpp              | 2 ++
 src/thorin/transform/codegen_prepare.cpp | 2 ++
 3 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index ab6b503ad..973b84dd2 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -360,7 +360,7 @@ void CCodeGen::emit_module() {
     forest_.for_each([&] (const Scope& scope) {
         if (scope.entry()->name() == "hls_top")
             hls_top = scope.entry();
-        else if (scope.entry()->cc() != CC::Internal)
+        else if (scope.entry()->cc() != CC::Internal && scope.entry()->is_exported())
             emit_scope(scope, forest_);
     });
     if (hls_top) {
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index e1b36028a..b89c66f2a 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -314,6 +314,8 @@ CodeGen::emit_module() {
         if(scope.entry()->cc() == CC::Internal) {
             return;
         }
+        if (!scope.entry()->is_exported())
+            return;
         emit_scope(scope, forest);
     });
 
diff --git a/src/thorin/transform/codegen_prepare.cpp b/src/thorin/transform/codegen_prepare.cpp
index 6ae0401d9..0f0279d56 100644
--- a/src/thorin/transform/codegen_prepare.cpp
+++ b/src/thorin/transform/codegen_prepare.cpp
@@ -28,6 +28,8 @@ void codegen_prepare(World& world) {
 
         if (dirty)
             scope.update();
+        else
+            ret_cont->destroy("codegen_prepare"); //Destroy the ret_cont if it was never used.
     });
     world.VLOG("end codegen_prepare");
 }

From e6fece82e70129915be7fdbcf1eb2b14714b7e0a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Sep 2023 13:59:19 +0200
Subject: [PATCH 188/342] Refactored rewriting and passes

---
 src/thorin/CMakeLists.txt                   |   3 +-
 src/thorin/be/codegen.cpp                   |  10 +-
 src/thorin/be/llvm/parallel.cpp             |   6 +-
 src/thorin/continuation.cpp                 |  90 ++++-------
 src/thorin/continuation.h                   |   9 +-
 src/thorin/def.cpp                          |   7 +-
 src/thorin/def.h                            |   6 +-
 src/thorin/primop.cpp                       |   5 +
 src/thorin/primop.h                         |   4 +-
 src/thorin/transform/cleanup_world.cpp      | 125 +++-------------
 src/thorin/transform/cleanup_world.h        |  12 --
 src/thorin/transform/closure_conversion.cpp |   2 +-
 src/thorin/transform/codegen_prepare.cpp    |  68 +++++----
 src/thorin/transform/codegen_prepare.h      |   2 +-
 src/thorin/transform/hls_channels.cpp       |  18 ++-
 src/thorin/transform/importer.cpp           | 139 +++++++++++++----
 src/thorin/transform/importer.h             |  26 ++--
 src/thorin/transform/lift_builtins.cpp      |   2 +-
 src/thorin/transform/mangle.cpp             | 158 ++++++++------------
 src/thorin/transform/mangle.h               |  28 ++--
 src/thorin/transform/partial_evaluation.cpp |  31 ++--
 src/thorin/transform/partial_evaluation.h   |  14 ++
 src/thorin/transform/rewrite.cpp            |  66 ++++++++
 src/thorin/transform/rewrite.h              |  32 ++++
 src/thorin/type.cpp                         |  31 ++--
 src/thorin/type.h                           |   4 +-
 src/thorin/world.cpp                        |  13 +-
 src/thorin/world.h                          |   3 +-
 28 files changed, 478 insertions(+), 436 deletions(-)
 delete mode 100644 src/thorin/transform/cleanup_world.h
 create mode 100644 src/thorin/transform/rewrite.cpp
 create mode 100644 src/thorin/transform/rewrite.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index e3247e5f7..58c57ce5a 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -41,7 +41,6 @@ set(THORIN_SOURCES
     tables/primtypetable.h
     tables/mathoptable.h
     transform/cleanup_world.cpp
-    transform/cleanup_world.h
     transform/closure_conversion.cpp
     transform/closure_conversion.h
     transform/codegen_prepare.h
@@ -64,6 +63,8 @@ set(THORIN_SOURCES
     transform/resolve_loads.h
     transform/partial_evaluation.cpp
     transform/partial_evaluation.h
+    transform/rewrite.cpp
+    transform/rewrite.h
     transform/split_slots.cpp
     transform/split_slots.h
     transform/hls_channels.cpp
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 8b992d8d3..e53fc2efb 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -93,8 +93,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     }
 
     // determine different parts of the world which need to be compiled differently
-    ScopesForest forest(world);
-    forest.for_each([&] (const Scope& scope) {
+    ScopesForest(world).for_each([&] (const Scope& scope) {
         auto continuation = scope.entry();
         Continuation* imported = nullptr;
 
@@ -142,8 +141,9 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
                     has_restrict &= p.second;
                 }
 
-                auto it_config = app->arg(LaunchArgs::Config)->as<Tuple>();
-                if (it_config->op(0)->isa<PrimLit>() &&
+                auto it_config = app->arg(LaunchArgs::Config)->isa<Tuple>();
+                if (it_config &&
+                    it_config->op(0)->isa<PrimLit>() &&
                     it_config->op(1)->isa<PrimLit>() &&
                     it_config->op(2)->isa<PrimLit>()) {
                     return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{
@@ -194,7 +194,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
             }
             return std::make_unique<HLSKernelConfig>(param_sizes);
         });
-        hls_annotate_top(importers[HLS].world(), top2kernel, kernel_config);
+        hls_annotate_top(accelerator_code[HLS].world(), top2kernel, kernel_config);
     }
     hls_kernel_launch(world, hls_host_params);
 
diff --git a/src/thorin/be/llvm/parallel.cpp b/src/thorin/be/llvm/parallel.cpp
index c9012525f..54e206510 100644
--- a/src/thorin/be/llvm/parallel.cpp
+++ b/src/thorin/be/llvm/parallel.cpp
@@ -36,7 +36,7 @@ Continuation* CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation*
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(PAR_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->body()->callee()->type()->as<FnType>()->types().skip_front(PAR_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -130,7 +130,7 @@ Continuation* CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* c
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(FIB_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->body()->callee()->type()->as<FnType>()->types().skip_front(FIB_NUM_ARGS)));
     llvm::Value* closure = llvm::UndefValue::get(closure_type);
     if (num_kernel_args != 1) {
         for (size_t i = 0; i < num_kernel_args; ++i)
@@ -213,7 +213,7 @@ Continuation* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* co
     }
 
     // fetch values and create a unified struct which contains all values (closure)
-    auto closure_type = convert(world().tuple_type(continuation->arg_fn_type()->types().skip_front(SPAWN_NUM_ARGS)));
+    auto closure_type = convert(world().tuple_type(continuation->body()->callee()->type()->as<FnType>()->types().skip_front(SPAWN_NUM_ARGS)));
     llvm::Value* closure = nullptr;
     if (closure_type->isStructTy()) {
         closure = llvm::UndefValue::get(closure_type);
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index b42263b9b..28f7dad8e 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -2,6 +2,7 @@
 
 #include <iostream>
 
+#include "thorin/transform/rewrite.h"
 #include "thorin/type.h"
 #include "thorin/world.h"
 #include "thorin/transform/mangle.h"
@@ -20,7 +21,10 @@ Param::Param(World& world, const Type* type, const Continuation* continuation, s
 
 const Def* Param::rebuild(World&, const Type*, Defs defs) const {
     assert(defs.size() == 1);
-    auto cont = defs[0]->as<Continuation>();
+    const Def* c = defs[0];
+    if (auto r = c->isa<Run>())
+        c = r->def()->as_nom<Continuation>();
+    auto cont = c->as<Continuation>();
     return cont->param(index());
 }
 
@@ -34,7 +38,7 @@ bool Param::equal(const Def* other) const {
 
 //------------------------------------------------------------------------------
 
-App::App(World& world, const Defs ops, Debug dbg) : Def(world, Node_App, ops[0]->world().bottom_type(), ops, dbg) {
+App::App(World& world, const Defs ops, Debug dbg) : Def(world, Node_App, world.bottom_type(), ops, dbg) {
 #if THORIN_ENABLE_CHECKS
     verify();
     if (auto cont = callee()->isa_nom<Continuation>())
@@ -69,6 +73,7 @@ Continuation::Continuation(World& w, const FnType* pi, const Attributes& attribu
     : Def(w, Node_Continuation, pi, 2, dbg)
     , attributes_(attributes)
 {
+    assert(pi->tag() == Node_FnType && "continuations may not be closures");
     params_.reserve(pi->num_ops());
     set_op(0, world().bottom(world().bottom_type()));
     set_op(1, world().filter({}, dbg));
@@ -80,61 +85,41 @@ Continuation::Continuation(World& w, const FnType* pi, const Attributes& attribu
     }
 }
 
-// TODO: merge with regular stub()
-Continuation* Continuation::mangle_stub() const {
-    Rewriter rewriter;
-    return mangle_stub(rewriter);
-}
-
-Continuation* Continuation::mangle_stub(Rewriter& rewriter) const {
-    auto result = world().continuation(type(), attributes(), debug_history());
-    for (size_t i = 0, e = num_params(); i != e; ++i) {
-        result->param(i)->set_name(debug_history().name);
-        rewriter.old2new[param(i)] = result->param(i);
-    }
-
-    if (!filter()->is_empty()) {
-        Array<const Def*> new_conditions(num_params());
-        for (size_t i = 0, e = num_params(); i != e; ++i)
-            new_conditions[i] = rewriter.instantiate(filter()->condition(i));
-
-        result->set_filter(world().filter(new_conditions, filter()->debug()));
-    }
+Continuation* Continuation::stub(Rewriter& rewriter, const Type* nty) const {
+    assert(!dead_);
+    auto& nworld = rewriter.dst();
 
-    return result;
-}
+    auto npi = nty->isa<FnType>();
+    assert(npi && npi->tag() == Node_FnType);
 
-Continuation* Continuation::stub(World& nworld, const Type* t) const {
-    assert(!dead_);
-    // TODO maybe we want to deal with intrinsics in a more streamlined way
-    if (this == world().branch())
-        return nworld.branch();
-    if (this == world().end_scope())
-        return nworld.end_scope();
-
-    auto npi = t->isa<FnType>();
-    assert(npi);
-    Continuation* ncontinuation = nworld.continuation(npi, attributes(), debug_history());
+    Continuation* ncontinuation = nworld.continuation(npi, attributes(), debug());
     assert(&ncontinuation->world() == &nworld);
     assert(&npi->world() == &nworld);
-    for (size_t i = 0, e = num_params(); i != e; ++i)
-        ncontinuation->param(i)->set_name(param(i)->debug_history().name);
 
-    if (is_external())
-        nworld.make_external(ncontinuation);
+    // TODO: investigate why this hangs
+    // ncontinuation->set_filter(rewriter.instantiate(filter())->as<Filter>());
     return ncontinuation;
 }
 
-void Continuation::rebuild_from(const Def*, Defs nops) {
-    if (this == world().branch())
-        return;
-    if (this == world().end_scope())
-        return;
+void Continuation::rebuild_from(Rewriter& rewriter, const Def* old) {
+    auto ocont = old->as<Continuation>();
+    assert(ocont);
+
+    assert(num_params() >= ocont->num_params());
+    for (size_t i = 0, e = ocont->num_params(); i != e; ++i)
+        param(i)->set_name(ocont->param(i)->debug().name);
+
+    set_filter(rewriter.instantiate(ocont->filter())->as<Filter>());
+
+    if (ocont->is_external())
+        world().make_external(this);
 
-    auto napp = nops[0]->isa<App>();
-    if (napp)
+    if (ocont->has_body()) {
+        auto napp = rewriter.instantiate(ocont->body())->isa<App>();
+        // i feel like this ought to be a warning, but maybe a later pass might legitimately do that
+        assert(napp && "is it legitimate to substitute away an App when rebuilding ?");
         set_body(napp);
-    set_filter(nops[1]->as<Filter>());
+    }
     verify();
 }
 
@@ -172,17 +157,6 @@ void Continuation::destroy(const char* cause) {
     dead_ = true;
 }
 
-const FnType* Continuation::arg_fn_type() const {
-    assert(has_body());
-    Array<const Type*> args(body()->num_args());
-    for (size_t i = 0, e = body()->num_args(); i != e; ++i)
-        args[i] = body()->arg(i)->type();
-
-    return body()->callee()->type()->isa<ClosureType>()
-           ? world().closure_type(args)->as<FnType>()
-           : world().fn_type(args);
-}
-
 const Param* Continuation::append_param(const Type* param_type, Debug dbg) {
     size_t size = type()->num_ops();
     Array<const Type*> ops(size + 1);
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 3a1f03281..ba72dbf91 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -142,10 +142,8 @@ class Continuation : public Def {
 public:
     const FnType* type() const { return Def::type()->as<FnType>(); }
 
-    Continuation* mangle_stub() const;
-    Continuation* mangle_stub(Rewriter& rewriter) const;
-    Continuation* stub(World&, const Type*) const override;
-    void rebuild_from(const Def* old, Defs new_ops) override;
+    Continuation* stub(Rewriter&, const Type*) const override;
+    void rebuild_from(Rewriter&, const Def* old) override;
     const Param* append_param(const Type* type, Debug dbg = {});
     Continuations preds() const;
     Continuations succs() const;
@@ -156,9 +154,6 @@ class Continuation : public Def {
     const Param* ret_param() const;
     size_t num_params() const { return params().size(); }
 
-    // TODO only used in parallel.cpp to create a dummy value, should be refactored in something cleaner
-    const FnType* arg_fn_type() const;
-
     Attributes& attributes() { return attributes_; }
     const Attributes& attributes() const { return attributes_; }
     Intrinsic intrinsic() const { return attributes().intrinsic; }
diff --git a/src/thorin/def.cpp b/src/thorin/def.cpp
index 8915d42ba..988811acf 100644
--- a/src/thorin/def.cpp
+++ b/src/thorin/def.cpp
@@ -3,6 +3,7 @@
 #include <algorithm>
 #include <stack>
 
+#include "thorin/transform/rewrite.h"
 #include "thorin/continuation.h"
 #include "thorin/primop.h"
 #include "thorin/type.h"
@@ -138,10 +139,10 @@ bool is_minus_zero(const Def* def) {
     return false;
 }
 
-void Def::rebuild_from(const Def*, Defs new_ops) {
-    assert(new_ops.size() == num_ops());
+void Def::rebuild_from(Rewriter& rewriter, const Def* old) {
+    assert(old->num_ops() == num_ops());
     for (size_t i = 0; i < num_ops(); i++)
-        set_op(i, new_ops[i]);
+        set_op(i, rewriter.instantiate(old->op(i)));
 }
 
 void Def::replace_uses(const Def* with) const {
diff --git a/src/thorin/def.h b/src/thorin/def.h
index a58b2a216..83419e8dc 100644
--- a/src/thorin/def.h
+++ b/src/thorin/def.h
@@ -14,7 +14,7 @@ namespace thorin {
 
 class Continuation;
 class Def;
-class Tracker;
+class Rewriter;
 class Use;
 class World;
 class Type;
@@ -221,8 +221,8 @@ class Def : public RuntimeCast<Def>, public Streamable<Def> {
     /// @name rebuild/stub
     //@{
     virtual const Def* rebuild(World&, const Type*, Defs) const { THORIN_UNREACHABLE; }
-    virtual       Def* stub(World&, const Type*) const { THORIN_UNREACHABLE; }
-    virtual       void rebuild_from(const Def* old, Defs new_ops);
+    virtual       Def* stub(Rewriter&, const Type*) const { THORIN_UNREACHABLE; }
+    virtual       void rebuild_from(Rewriter&, const Def* old);
     //@}
 
     void replace_uses(const Def*) const;
diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index 49208b173..a1222f4fe 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -1,5 +1,6 @@
 #include "thorin/primop.h"
 #include "thorin/continuation.h"
+#include "thorin/transform/rewrite.h"
 
 #include "thorin/config.h"
 #include "thorin/type.h"
@@ -312,6 +313,10 @@ const PtrType* Closure::environment_ptr_type(World& world) {
     return world.ptr_type(world.type_pu8());
 }
 
+Continuation* Closure::fn() const {
+    return op(0)->as_nom<Continuation>();
+}
+
 //------------------------------------------------------------------------------
 
 }
diff --git a/src/thorin/primop.h b/src/thorin/primop.h
index 6347c85e6..5b2a7eb33 100644
--- a/src/thorin/primop.h
+++ b/src/thorin/primop.h
@@ -343,6 +343,8 @@ class Closure : public Aggregate {
     static const Type*    environment_type(World&);
     static const PtrType* environment_ptr_type(World&);
 
+    Continuation* fn() const;
+
     friend class World;
 };
 
@@ -534,11 +536,11 @@ class Global : public Def {
 
     bool is_external() const;
     void set_init(const Def* new_init) { unset_op(0); set_op(0, new_init); }
+    const Def* rebuild(World&, const Type*, Defs) const override;
 
 private:
     hash_t vhash() const override { return murmur3(gid()); }
     bool equal(const Def* other) const override { return this == other; }
-    const Def* rebuild(World&, const Type*, Defs) const override;
 
     bool is_mutable_;
 
diff --git a/src/thorin/transform/cleanup_world.cpp b/src/thorin/transform/cleanup_world.cpp
index 9a8eeaa99..65a860e48 100644
--- a/src/thorin/transform/cleanup_world.cpp
+++ b/src/thorin/transform/cleanup_world.cpp
@@ -2,8 +2,6 @@
 #include "thorin/world.h"
 #include "thorin/analyses/cfg.h"
 #include "thorin/analyses/scope.h"
-#include "thorin/analyses/domtree.h"
-#include "thorin/analyses/scope.h"
 #include "thorin/analyses/verify.h"
 #include "thorin/transform/importer.h"
 #include "thorin/transform/mangle.h"
@@ -14,14 +12,13 @@ namespace thorin {
 
 class Cleaner {
 public:
-    Cleaner(std::unique_ptr<World>& world)
-        : world_(world)
+    Cleaner(Thorin& thorin)
+        : thorin_(thorin)
     {}
 
-    World& world() { return *world_; }
+    World& world() { return thorin_.world(); }
     void cleanup();
     void eliminate_tail_rec();
-    void eta_conversion();
     void eliminate_params();
     void rebuild();
     void verify_closedness();
@@ -31,13 +28,12 @@ class Cleaner {
 private:
     void cleanup_fix_point();
     void clean_pe_info(std::queue<Continuation*>, Continuation*);
-    std::unique_ptr<World>& world_;
+    Thorin& thorin_;
     bool todo_ = true;
 };
 
 void Cleaner::eliminate_tail_rec() {
-    ScopesForest forest(world());
-    forest.for_each([&](Scope& scope) {
+    ScopesForest(world()).for_each([&](Scope& scope) {
         auto entry = scope.entry();
 
         bool only_tail_calls = true;
@@ -100,86 +96,6 @@ void Cleaner::eliminate_tail_rec() {
     });
 }
 
-void Cleaner::eta_conversion() {
-    for (bool todo = true; todo;) {
-        todo = false;
-        for (auto def : world().copy_defs()) {
-            auto continuation = def->isa_nom<Continuation>();
-            if (!continuation || !continuation->has_body()) continue;
-
-            // eat calls to known continuations that are only used once
-            while (auto callee = continuation->body()->callee()->isa_nom<Continuation>()) {
-                auto body = continuation->body();
-                if (callee == continuation) break;
-
-                if (callee->has_body() && !world().is_external(callee) && callee->can_be_inlined()) {
-                    auto callee_body = callee->body();
-                    for (size_t i = 0, e = body->num_args(); i != e; ++i)
-                        callee->param(i)->replace_uses(body->arg(i));
-
-                    // because App nodes are hash-consed (thus reusable), there is a risk to invalidate their other uses here, if there are indeed any
-                    // can_be_inlined() should account for that by counting reused apps multiple times, but in case it fails we have this pair of asserts as insurance
-                    assert(body->num_uses() == 1);
-                    continuation->jump(callee_body->callee(), callee_body->args(), callee->debug()); // TODO debug
-                    callee->destroy("cleanup: continuation only called once");
-                    assert(body->num_uses() == 0);
-                    todo_ = todo = true;
-                } else
-                    break;
-            }
-
-            auto body = continuation->body();
-            // try to subsume continuations which call a parameter
-            // (that is free within that continuation) with that parameter
-            if (auto param = body->callee()->isa<Param>()) {
-                if (param->continuation() == continuation || world().is_external(continuation))
-                    continue;
-
-                if (body->args() == continuation->params_as_defs()) {
-                    continuation->replace_uses(body->callee());
-                    continuation->destroy("cleanup: calls a parameter (no perm)");
-                    todo_ = todo = true;
-                    continue;
-                }
-
-                // build the permutation of the arguments
-                Array<size_t> perm(body->num_args());
-                bool is_permutation = true;
-                for (size_t i = 0, e = body->num_args(); i != e; ++i)  {
-                    auto param_it = std::find(continuation->params().begin(),
-                                                continuation->params().end(),
-                                                body->arg(i));
-
-                    if (param_it == continuation->params().end()) {
-                        is_permutation = false;
-                        break;
-                    }
-
-                    perm[i] = param_it - continuation->params().begin();
-                }
-
-                if (!is_permutation) continue;
-
-                // for every use of the continuation at a call site,
-                // permute the arguments and call the parameter instead
-                for (auto use : continuation->copy_uses()) {
-                    auto uapp = use->isa<App>();
-                    if (uapp && use.index() == 0) {
-                        for (auto ucontinuation : uapp->using_continuations()) {
-                            Array<const Def*> new_args(perm.size());
-                            for (size_t i = 0, e = perm.size(); i != e; ++i) {
-                                new_args[i] = uapp->arg(perm[i]);
-                            }
-                            ucontinuation->jump(param, new_args, ucontinuation->debug()); // TODO debug
-                            todo_ = todo = true;
-                        }
-                    }
-                }
-            }
-        }
-    }
-}
-
 void Cleaner::eliminate_params() {
     // TODO
     for (auto ocontinuation : world().copy_continuations()) {
@@ -222,10 +138,9 @@ void Cleaner::eliminate_params() {
                     assert(use.index() == 0);
                     for (auto ucontinuation : uapp->using_continuations()) {
                         ucontinuation->jump(ncontinuation, uapp->args().cut(proxy_idx), ucontinuation->debug());
+                        todo_ = true;
                     }
                 }
-
-                todo_ = true;
             }
         }
 next_continuation:;
@@ -235,7 +150,6 @@ next_continuation:;
 void Cleaner::rebuild() {
     auto fresh_world = std::make_unique<World>(world());
     Importer importer(world(), *fresh_world);
-    importer.def_old2new_.rehash(world_->defs().capacity());
 
     for (auto&& [_, def] : world().externals()) {
         if (auto cont = def->isa<Continuation>(); cont && cont->is_exported())
@@ -244,7 +158,7 @@ void Cleaner::rebuild() {
             importer.import(global);
     }
 
-    std::swap(world_, fresh_world);
+    std::swap(thorin_.world_container(), fresh_world);
 
     // verify(world());
 
@@ -281,7 +195,7 @@ void Cleaner::clean_pe_info(std::queue<Continuation*> queue, Continuation* cur)
     auto next = body->arg(3);
     auto msg = body->arg(1)->as<Bitcast>()->from()->as<Global>()->init()->as<DefiniteArray>();
 
-    world_->idef(body->callee(), "pe_info was not constant: {}: {}", msg->as_string(), body->arg(2));
+    world().idef(body->callee(), "pe_info was not constant: {}: {}", msg->as_string(), body->arg(2));
     cur->jump(next, {body->arg(0)}, cur->debug()); // TODO debug
     todo_ = true;
 
@@ -290,7 +204,7 @@ void Cleaner::clean_pe_info(std::queue<Continuation*> queue, Continuation* cur)
 }
 
 void Cleaner::clean_pe_infos() {
-    world_->VLOG("cleaning remaining pe_infos");
+    world().VLOG("cleaning remaining pe_infos");
     std::queue<Continuation*> queue;
     ContinuationSet done;
     auto enqueue = [&](Continuation* continuation) {
@@ -321,24 +235,23 @@ void Cleaner::clean_pe_infos() {
 void Cleaner::cleanup_fix_point() {
     int i = 0;
     for (; todo_; ++i) {
-        world_->VLOG("iteration: {}", i);
+        world().VLOG("iteration: {}", i);
         todo_ = false;
-        if (world_->is_pe_done())
-            eliminate_tail_rec();
-        eta_conversion();
+        rebuild();
+        eliminate_tail_rec();
         eliminate_params();
         rebuild(); // resolve replaced defs before going to resolve_loads
         todo_ |= resolve_loads(world());
         rebuild();
-        if (!world().is_pe_done())
-            todo_ |= partial_evaluation(*world_);
-        else
-            clean_pe_infos();
+        //if (!world().is_pe_done())
+            todo_ |= partial_evaluation(world());
+        //else
+        //    clean_pe_infos();
     }
 }
 
 void Cleaner::cleanup() {
-    world_->VLOG("start cleanup");
+    world().VLOG("start cleanup");
     cleanup_fix_point();
 
     if (!world().is_pe_done()) {
@@ -352,13 +265,13 @@ void Cleaner::cleanup() {
         cleanup_fix_point();
     }
 
-    world_->VLOG("end cleanup");
+    world().VLOG("end cleanup");
 #if THORIN_ENABLE_CHECKS
     verify_closedness();
     debug_verify(world());
 #endif
 }
 
-void cleanup_world(std::unique_ptr<World>& world) { Cleaner(world).cleanup(); }
+void Thorin::cleanup() { Cleaner(*this).cleanup(); }
 
 }
diff --git a/src/thorin/transform/cleanup_world.h b/src/thorin/transform/cleanup_world.h
deleted file mode 100644
index afcb68b75..000000000
--- a/src/thorin/transform/cleanup_world.h
+++ /dev/null
@@ -1,12 +0,0 @@
-#ifndef THORIN_CLEANUP_WORLD_H
-#define THORIN_CLEANUP_WORLD_H
-
-namespace thorin {
-
-class World;
-
-void cleanup_world(std::unique_ptr<World>& world);
-
-}
-
-#endif
diff --git a/src/thorin/transform/closure_conversion.cpp b/src/thorin/transform/closure_conversion.cpp
index a91b55865..9140f4132 100644
--- a/src/thorin/transform/closure_conversion.cpp
+++ b/src/thorin/transform/closure_conversion.cpp
@@ -124,7 +124,7 @@ class ClosureConversion {
                 return continuation && (!continuation->has_body() || continuation->is_intrinsic());
             });
             free_vars.shrink(filtered_out - free_vars.begin());
-            auto lifted = lift(scope, free_vars);
+            auto lifted = lift(scope, scope.entry(), free_vars);
 
             // get the environment type
             const Type* env_type = nullptr;
diff --git a/src/thorin/transform/codegen_prepare.cpp b/src/thorin/transform/codegen_prepare.cpp
index 0f0279d56..e8218a571 100644
--- a/src/thorin/transform/codegen_prepare.cpp
+++ b/src/thorin/transform/codegen_prepare.cpp
@@ -1,37 +1,53 @@
 #include "thorin/world.h"
 #include "thorin/analyses/scope.h"
+#include "thorin/transform/rewrite.h"
 
 namespace thorin {
 
-void codegen_prepare(World& world) {
-    world.VLOG("start codegen_prepare");
-    ScopesForest forest(world);
-    forest.for_each([&](Scope& scope) {
-        world.DLOG("scope: {}", scope.entry());
-        bool dirty = false;
-        auto ret_param = scope.entry()->ret_param();
-        assert(ret_param && "scopes should have a return parameter");
-        auto ret_cont = world.continuation(ret_param->type()->as<FnType>(), ret_param->debug());
-        ret_cont->jump(ret_param, ret_cont->params_as_defs(), ret_param->debug());
-
-        for (auto use : ret_param->copy_uses()) {
-            if (auto uapp = use->isa<App>()) {
-                if (use.index() != 0) {
-                    auto nops = uapp->copy_ops();
-                    nops[use.index()] = ret_cont;
-                    auto napp = uapp->rebuild(world, uapp->type(), nops);
-                    uapp->replace_uses(napp);
-                    dirty = true;
+struct CodegenPrepare : public Rewriter {
+    CodegenPrepare(World& src, World& dst) : Rewriter(src, dst) {}
+
+    Continuation* make_wrapper(const Def* old_return_param) {
+        assert(old_return_param);
+        auto npi = instantiate(old_return_param->type())->as<FnType>();
+        npi = dst().fn_type(npi->types());
+        auto wrapper = dst().continuation(npi, old_return_param->debug());
+        return wrapper;
+    }
+
+    const Def* rewrite(const Def* odef) override {
+        if (auto app = odef->isa<App>()) {
+            auto new_ops = Array<const Def*>(app->num_args(), [&](size_t i) -> const Def* {
+                auto oarg = app->arg(i);
+                if (auto oparam = oarg->isa<Param>()) {
+                    if (oparam == oparam->continuation()->ret_param()) {
+                        auto wrapped = make_wrapper(oarg);
+                        insert(oarg, wrapped);
+                        auto imported_param = instantiate(oparam->continuation())->as_nom<Continuation>()->ret_param();
+                        wrapped->jump(imported_param, wrapped->params_as_defs(), imported_param->debug());
+                        return wrapped;
+                    }
                 }
-            }
+                return instantiate(app->arg(i));
+            });
+            return dst().app(instantiate(app->callee()), new_ops);
         }
+        return Rewriter::rewrite(odef);
+    }
+};
+
+/// this pass makes sure the return param is only called directly, by eta-expanding any uses where it appears in another position
+void codegen_prepare(Thorin& thorin) {
+    thorin.world().VLOG("start codegen_prepare");
+    auto& src = thorin.world();
+    auto destination = std::make_unique<World>(src);
+    CodegenPrepare pass(src, *destination.get());
+
+    for (auto& external : src.externals())
+        pass.instantiate(external.second);
 
-        if (dirty)
-            scope.update();
-        else
-            ret_cont->destroy("codegen_prepare"); //Destroy the ret_cont if it was never used.
-    });
-    world.VLOG("end codegen_prepare");
+    thorin.world_container().swap(destination);
+    thorin.world().VLOG("end codegen_prepare");
 }
 
 }
diff --git a/src/thorin/transform/codegen_prepare.h b/src/thorin/transform/codegen_prepare.h
index fb4c91bd8..919878260 100644
--- a/src/thorin/transform/codegen_prepare.h
+++ b/src/thorin/transform/codegen_prepare.h
@@ -5,7 +5,7 @@ namespace thorin {
 
 class World;
 
-void codegen_prepare(World&);
+void codegen_prepare(Thorin&);
 
 }
 
diff --git a/src/thorin/transform/hls_channels.cpp b/src/thorin/transform/hls_channels.cpp
index f4c891133..a5e968a79 100644
--- a/src/thorin/transform/hls_channels.cpp
+++ b/src/thorin/transform/hls_channels.cpp
@@ -193,12 +193,14 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
 
             world.make_internal(old_kernel);
 
-            Rewriter rewriter;
+            // TODO this is now broken
+            // TODO this should likely use the mangler
+            Rewriter rewriter(world);
             // Map the parameters of the old kernel to the first N parameters of the new one
             // The channels used inside the kernel are mapped to the parameters N + 1, N + 2, ...
-            for (auto pair : index2def) {
+            /*for (auto pair : index2def) {
                 auto param = new_kernel->param(pair.first);
-                rewriter.old2new[pair.second] = param;
+                rewriter.insert(pair.second, param);
                 param2arg[param] = pair.second; // (channel params, globals)
             }
             for (auto def : scope.defs()) {
@@ -206,9 +208,9 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
                     // Copy the basic block by calling stub
                     // Or reuse the newly created kernel copy if def is the old kernel
                     auto new_cont = def == old_kernel ? new_kernel : cont->mangle_stub();
-                    rewriter.old2new[cont] = new_cont;
+                    rewriter.insert(cont, new_cont);
                     for (size_t i = 0; i < cont->num_params(); ++i)
-                        rewriter.old2new[cont->param(i)] = new_cont->param(i);
+                        rewriter.insert(cont->param(i), new_cont->param(i));
                 }
             }
             // Rewriting the basic blocks of the kernel using the map
@@ -216,7 +218,7 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
                 if (auto cont = def->isa_nom<Continuation>()) { // all basic blocks of the scope
                     if (!cont->has_body()) continue;
                     auto body = cont->body();
-                    auto new_cont = rewriter.old2new[cont]->isa_nom<Continuation>();
+                    auto new_cont = rewriter.instantiate(cont)->isa_nom<Continuation>();
                     auto new_callee = rewriter.instantiate(body->callee());
                     Array<const Def*> new_args(body->num_args());
                     for ( size_t i = 0; i < body->num_args(); ++i)
@@ -225,7 +227,7 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
                 }
             }
             if (!is_single_kernel(new_kernel))
-                kernels_ch_modes.emplace_back(def2mode);
+                kernels_ch_modes.emplace_back(def2mode);*/
     });
 
 
@@ -274,7 +276,7 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
         for (auto def : old_world.defs()) {
             if (auto ocontinuation = def->isa_nom<Continuation>()) {
                 auto ncontinuation = elem->as<Param>()->continuation();
-                if (ncontinuation == importer.def_old2new_[ocontinuation]) {
+                if (ncontinuation == importer.import(ocontinuation)) {
                     elem = ocontinuation->param(elem->as<Param>()->index());
                     break;
                 }
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index 4691e9795..9bc97bc9b 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -1,43 +1,128 @@
 #include "thorin/transform/importer.h"
+#include "thorin/transform/mangle.h"
+#include "thorin/primop.h"
 
 namespace thorin {
 
-const Def* Importer::import(const Def* odef) {
-    if (auto ndef = def_old2new_.lookup(odef)) {
-        assert(&(*ndef)->world() == &world());
-        return *ndef;
-    }
+const Def* Importer::rewrite(const Def* const odef) {
+    assert(&odef->world() == &src());
+    if (auto memop = odef->isa<MemOp>()) {
+        // Optimise out dead loads when importing
+        if (memop->isa<Load>() || memop->isa<Enter>()) {
+            if (memop->out(1)->num_uses() == 0) {
+                auto imported_mem = import(memop->mem());
+                auto imported_ty = import(memop->out(1)->type())->as<Type>();
+                todo_ = true;
+                return(dst().tuple({ imported_mem, dst().bottom(imported_ty) }));
+            }
+        }
+    } else if (auto app = odef->isa<App>()) {
+        // eat calls to known continuations that are only used once
+        if (auto callee = app->callee()->isa_nom<Continuation>()) {
+            if (callee->has_body() && !src().is_external(callee) && callee->can_be_inlined()) {
+                todo_ = true;
+                src().VLOG("simplify: inlining continuation {} because it is called exactly once", callee);
+                for (size_t i = 0; i < callee->num_params(); i++)
+                    insert(callee->param(i), import(app->arg(i)));
 
-    if (odef == odef->world().star()) {
-        def_old2new_[odef] = world().star();
-        return world().star();
-    }
+                return instantiate(callee->body());
+            }
+        }
+    } else if (auto closure = odef->isa<Closure>()) {
+        bool only_called = true;
+        for (auto use : closure->uses()) {
+            if (use.def()->isa<App>() && use.index() == 0)
+                continue;
+            only_called = false;
+            break;
+        }
+        if (only_called) {
+            bool self_param_ok = true;
+            for (auto use: closure->fn()->params().back()->uses()) {
+                // the closure argument can be used, but only to extract the environment!
+                if (auto extract = use.def()->isa<Extract>(); extract && is_primlit(extract->index(), 1))
+                    continue;
+                self_param_ok = false;
+                break;
+            }
+            if (self_param_ok) {
+                src().VLOG("simplify: eliminating closure {} as it is never passed as an argument, and is not recursive", closure);
+                Array<const Def*> args(closure->fn()->num_params());
+                args.back() = closure;
+                todo_ = true;
+                return instantiate(drop(closure->fn(), args));
+            }
+        }
+    } else if (auto cont = odef->isa_nom<Continuation>()) {
+        if (cont->has_body()) {
+            auto body = cont->body();
+            // try to subsume continuations which call a def
+            // (that is free within that continuation) with that def
+            auto callee = body->callee();
+            auto& scope = forest_->get_scope(cont);
+            if (!scope.contains(callee)) {
+                if (src().is_external(cont) || callee->type()->tag() != Node_FnType)
+                    goto rebuild;
 
-    auto ntype = import(odef->type())->as<Type>();
+                if (body->args() == cont->params_as_defs()) {
+                    src().VLOG("simplify: continuation {} calls a free def: {}", cont->unique_name(), body->callee());
+                    // We completely replace the original continuation
+                    // If we don't do so, then we miss some simplifications
+                    return instantiate(body->callee());
+                } else {
+                    // build the permutation of the arguments
+                    Array<size_t> perm(body->num_args());
+                    bool is_permutation = true;
+                    for (size_t i = 0, e = body->num_args(); i != e; ++i) {
+                        auto param_it = std::find(cont->params().begin(),
+                                                  cont->params().end(),
+                                                  body->arg(i));
 
-    Def* stub = nullptr;
-    if (odef->isa_nom()) {
-        stub = odef->stub(world(), ntype);
-        def_old2new_[odef] = stub;
-    }
+                        if (param_it == cont->params().end()) {
+                            is_permutation = false;
+                            break;
+                        }
+
+                        perm[i] = param_it - cont->params().begin();
+                    }
+
+                    if (!is_permutation)
+                        goto rebuild;
+                }
+
+                bool has_calls = false;
+                // for every use of the continuation at a call site,
+                // permute the arguments and call the parameter instead
+                for (auto use : cont->copy_uses()) {
+                    auto uapp = use->isa<App>();
+                    if (uapp && use.index() == 0) {
+                        todo_ = true;
+                        has_calls = true;
+                        break;
+                    }
+                }
+
+                if (has_calls) {
+                    auto rebuilt = cont->stub(*this, instantiate(cont->type())->as<Type>());
+                    src().VLOG("simplify: continuation {} calls a free def: {} (with permuted args), introducing a wrapper: {}", cont->unique_name(), body->callee(), rebuilt);
+                    auto wrapped = dst().run(rebuilt);
+                    insert(odef, wrapped);
 
-    size_t size = odef->num_ops();
-    Array<const Def*> nops(size);
-    for (size_t i = 0; i != size; ++i) {
-        assert(odef->op(i) != odef);
-        nops[i] = import(odef->op(i));
-        assert(&nops[i]->world() == &world());
+                    rebuilt->set_body(instantiate(body)->as<App>());
+                    return wrapped;
+                }
+            }
+        }
     }
 
+    rebuild:
+    auto ndef = Rewriter::rewrite(odef);
     if (odef->isa_structural()) {
-        auto ndef = odef->rebuild(world(), ntype, nops);
+        // If some substitution took place
+        // TODO: this might be dead code at the moment
         todo_ |= odef->tag() != ndef->tag();
-        return def_old2new_[odef] = ndef;
-    } else {
-        assert(odef->isa_nom() && stub);
-        stub->rebuild_from(odef, nops);
-        return stub;
     }
+    return ndef;
 }
 
 }
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index 10941b915..91110ade9 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -3,31 +3,29 @@
 
 #include "thorin/world.h"
 #include "thorin/config.h"
+#include "thorin/transform/rewrite.h"
+#include "thorin/analyses/scope.h"
 
 namespace thorin {
 
-class Importer {
+class Importer : Rewriter {
 public:
     explicit Importer(World& src, World& dst)
-        : src(src)
-        , dst(dst)
+        : Rewriter(src, dst), forest_(std::make_unique<ScopesForest>(src))
     {
+        assert(&src != &dst);
         if (src.is_pe_done())
-            world().mark_pe_done();
-#if THORIN_ENABLE_CHECKS
-        if (src.track_history())
-            world().enable_history(true);
-#endif
+            dst.mark_pe_done();
     }
 
-    World& world() { return dst; }
-    const Def* import(const Def*);
+    const Def* import(const Def* odef) { return instantiate(odef); }
     bool todo() const { return todo_; }
 
-public:
-    Def2Def def_old2new_;
-    World& src;
-    World& dst;
+protected:
+    const Def* rewrite(const Def* odef) override;
+
+private:
+    std::unique_ptr<ScopesForest> forest_;
     bool todo_ = false;
 };
 
diff --git a/src/thorin/transform/lift_builtins.cpp b/src/thorin/transform/lift_builtins.cpp
index 8f21d44ee..6bd509da6 100644
--- a/src/thorin/transform/lift_builtins.cpp
+++ b/src/thorin/transform/lift_builtins.cpp
@@ -105,7 +105,7 @@ void lift_builtins(Thorin& thorin) {
             }
         }
 
-        auto lifted = lift(scope, defs);
+        auto lifted = lift(scope, scope.entry(), defs);
         for (auto use : cur->copy_uses()) {
             if (auto uapp = use->isa<App>()) {
                 if (auto callee = uapp->callee()->isa_nom<Continuation>()) {
diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index b184d97c5..77bd201ab 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -7,31 +7,16 @@
 
 namespace thorin {
 
-const Def* Rewriter::instantiate(const Def* odef) {
-    if (auto ndef = old2new.lookup(odef)) return *ndef;
-
-    if (odef->isa_structural() && !odef->isa<Param>()) {
-        Array<const Def*> nops(odef->num_ops());
-        for (size_t i = 0; i != odef->num_ops(); ++i)
-            nops[i] = instantiate(odef->op(i));
-
-        auto nprimop = odef->rebuild(odef->world(), odef->type(), nops);
-        return old2new[odef] = nprimop;
-    }
-
-    return old2new[odef] = odef;
-}
-
 /// Mangles a continuation's scope
 /// @p args has the size of the original continuation, a null entry means the parameter remains, non-null substitutes it in scope and removes it from the signature
 /// @p lift lists defs that should be replaced by a fresh param, to be appended at the end of the signature
-Mangler::Mangler(const Scope& scope, Defs args, Defs lift)
-    : scope_(scope)
+Mangler::Mangler(const Scope& scope, Continuation* entry, Defs args, Defs lift)
+    : Rewriter(scope.world())
+    , scope_(scope)
     , args_(args)
     , lift_(lift)
-    , old_entry_(scope.entry())
+    , old_entry_(entry)
     , defs_(scope.defs().capacity())
-    , def2def_(scope.defs().capacity())
 {
     assert(old_entry()->has_body());
     assert(args.size() == old_entry()->num_params());
@@ -52,6 +37,10 @@ Mangler::Mangler(const Scope& scope, Defs args, Defs lift)
         for (auto use : pop(queue)->uses())
             enqueue(use);
     }
+
+    is_dropping_ = std::any_of(args.begin(), args.end(), [&](const auto& item) {
+        return item != nullptr;
+    });
 }
 
 Continuation* Mangler::mangle() {
@@ -62,122 +51,95 @@ Continuation* Mangler::mangle() {
             param_types.emplace_back(old_entry()->param(i)->type()); // TODO reduce
     }
 
-    auto fn_type = world().fn_type(param_types);
-    new_entry_ = world().continuation(fn_type, old_entry()->debug_history());
+    auto fn_type = dst().fn_type(param_types);
+    new_entry_ = dst().continuation(fn_type, old_entry()->debug());
 
-    // map value params
-    def2def_[old_entry()] = old_entry();
     for (size_t i = 0, j = 0, e = old_entry()->num_params(); i != e; ++i) {
         auto old_param = old_entry()->param(i);
         if (auto def = args_[i])
-            def2def_[old_param] = def;
+            insert(old_param, def);
         else {
             // we recreate params that aren't specialized
             auto new_param = new_entry()->param(j++);
-            def2def_[old_param] = new_param;
+            insert(old_param, new_param);
             new_param->set_name(old_param->name());
         }
     }
 
     for (auto def : lift_)
-        def2def_[def] = new_entry()->append_param(def->type()); // TODO reduce
+        insert(def, new_entry()->append_param(def->type()));
+
+    // if we are dropping parameters, we can't necessarily rewrite the entry, see also note about applications in Mangler::rewrite()
+    if (is_dropping_)
+        insert(old_entry(), old_entry());
+    else {
+        // if we're only adding parameters, we can replace the entry by a small wrapper calling into the lifted entry
+        auto recursion_wrapper = dst().continuation(old_entry()->type());
+        insert(old_entry(), recursion_wrapper);
+        std::vector<const Def*> args;
+        for (auto p : recursion_wrapper->params_as_defs())
+            args.push_back(p);
+        size_t i = 0;
+        for (auto def : lift_)
+            args.push_back(new_entry()->param(recursion_wrapper->num_params() + i++));
+        recursion_wrapper->jump(new_entry(), args);
+    }
 
-    // mangle filter
+    // cut/widen filter
     if (!old_entry()->filter()->is_empty()) {
         Array<const Def*> new_conditions(new_entry()->num_params());
         size_t j = 0;
         for (size_t i = 0, e = old_entry()->num_params(); i != e; ++i) {
             if (args_[i] == nullptr)
-                new_conditions[j++] = mangle(old_entry()->filter()->condition(i));
+                new_conditions[j++] = instantiate(old_entry()->filter()->condition(i));
         }
 
         for (size_t e = new_entry()->num_params(); j != e; ++j)
-            new_conditions[j] = world().literal_bool(false, Debug{});
+            new_conditions[j] = dst().literal_bool(false, Debug{});
 
-        new_entry()->set_filter(world().filter(new_conditions, old_entry()->filter()->debug()));
+        new_entry()->set_filter(dst().filter(new_conditions, old_entry()->filter()->debug()));
     }
 
-    new_entry()->set_body(mangle_body(old_entry()->body()));
-
+    new_entry()->set_body(instantiate(old_entry()->body())->as<App>());
     new_entry()->verify();
 
     return new_entry();
 }
 
-Continuation* Mangler::mangle_head(Continuation* old_continuation) {
-    assert(!def2def_.contains(old_continuation));
-    assert(old_continuation->has_body());
-    Rewriter rewriter{def2def_};
-    Continuation* new_continuation = old_continuation->mangle_stub(rewriter);
-    def2def_[old_continuation] = new_continuation;
-
-    for (size_t i = 0, e = old_continuation->num_params(); i != e; ++i)
-        def2def_[old_continuation->param(i)] = new_continuation->param(i);
-
-    return new_continuation;
-}
-
-const App* Mangler::mangle_body(const App* old_body) {
-    Array<const Def*> nops(old_body->num_ops());
-    for (size_t i = 0, e = nops.size(); i != e; ++i)
-        nops[i] = mangle(old_body->op(i));
-
-    Defs nargs(nops.skip_front()); // new args of body
-    auto ntarget = nops.front();   // new target of body
-
-    // check whether we can optimize tail recursion
-    if (ntarget == old_entry()) {
-        std::vector<size_t> cut;
-        bool substitute = true;
-        for (size_t i = 0, e = args_.size(); i != e && substitute; ++i) {
-            if (auto def = args_[i]) {
-                substitute &= def == nargs[i];
-                cut.push_back(i);
+const Def* Mangler::rewrite(const Def* old_def) {
+    if (!within(old_def))
+        return old_def;  // we leave free variables alone
+    if (auto param = old_def->isa<Param>())
+        assert(within(param->continuation()) && "if the param is not free, the continuation should not be either!");
+    auto ndef = Rewriter::rewrite(old_def);
+    if (auto app = ndef->isa<App>()) {
+        // If you drop a parameter it is replaced by some other def, which will be identical for all recursive calls, because it's now specialised
+        // If there originally was a recursive call that specified the to-be-dropped parameter to something else, we need to call the unmangled original to preserve semantics
+        if (is_dropping_ && app->callee() == old_entry()) {
+            auto oargs = app->args();
+            auto nargs = Array<const Def*>(oargs.size(), [&](size_t i) { return rewrite(oargs[i]); });
+            std::vector<size_t> cut;
+            bool substitute = true;
+            for (size_t i = 0, e = args_.size(); i != e && substitute; ++i) {
+                if (auto def = args_[i]) {
+                    substitute &= def == nargs[i];
+                    cut.push_back(i);
+                }
             }
-        }
 
-        if (substitute) {
-            // Q: why not always change to the mangled continuation ?
-            // A: if you drop a parameter it is replaced by some def (likely a free param), which will be identical for all recursive calls, since they live in the same scope (that's how scopes work)
-            // so if there originally was a recursive call that specified the to-be-dropped parameter to something else, we need to call the unmangled original to preserve semantics
-            const auto& args = concat(nargs.cut(cut), new_entry()->params().get_back(lift_.size()));
-            return world().app(new_entry(), args, old_body->debug()); // TODO debug
+            if (substitute) {
+                const auto& args = concat(nargs.cut(cut), new_entry()->params().get_back(lift_.size()));
+                return dst().app(new_entry(), args, old_def->debug()); // TODO debug
+            }
         }
     }
-
-    return world().app(ntarget, nargs, old_body->debug()); // TODO debug
-}
-
-const Def* Mangler::mangle(const Def* old_def) {
-    if (auto new_def = def2def_.lookup(old_def))
-        return *new_def;
-    else if (!within(old_def))
-        return old_def;  // we leave free variables alone
-    else if (auto old_continuation = old_def->isa_nom<Continuation>()) {
-        auto new_continuation = mangle_head(old_continuation);
-        if (old_continuation->has_body())
-            new_continuation->set_body(mangle_body(old_continuation->body()));
-        return new_continuation;
-    } else if (auto param = old_def->isa<Param>()) {
-        assert(within(param->continuation()));
-        mangle(param->continuation());
-        assert(def2def_.contains(param));
-        return def2def_[param];
-    } else {
-        Array<const Def*> nops(old_def->num_ops());
-        for (size_t i = 0, e = old_def->num_ops(); i != e; ++i)
-            nops[i] = mangle(old_def->op(i));
-
-        auto type = old_def->type(); // TODO reduce
-        assert(!old_def->isa<Type>());
-        return def2def_[old_def] = old_def->rebuild(world(), type, nops);
-    }
+    return ndef;
 }
 
 //------------------------------------------------------------------------------
 
-Continuation* mangle(const Scope& scope, Defs args, Defs lift) {
-    return Mangler(scope, args, lift).mangle();
+Continuation* mangle(const Scope& scope, Continuation* entry, Defs args, Defs lift) {
+    return Mangler(scope, entry, args, lift).mangle();
 }
 
 Continuation* drop(const Def* callee, const Defs specialized_args) {
diff --git a/src/thorin/transform/mangle.h b/src/thorin/transform/mangle.h
index 162da6071..353c69fee 100644
--- a/src/thorin/transform/mangle.h
+++ b/src/thorin/transform/mangle.h
@@ -3,54 +3,48 @@
 
 #include "thorin/type.h"
 #include "thorin/analyses/scope.h"
+#include "thorin/transform/rewrite.h"
 
 namespace thorin {
 
-struct Rewriter {
-    const Def* instantiate(const Def* odef);
-    Def2Def old2new;
-};
-
-class Mangler {
+class Mangler : Rewriter {
 public:
-    Mangler(const Scope& scope, Defs args, Defs lift);
+    Mangler(const Scope& scope, Continuation* entry, Defs args, Defs lift);
 
     const Scope& scope() const { return scope_; }
-    World& world() const { return scope_.world(); }
     Continuation* mangle();
     Continuation* old_entry() const { return old_entry_; }
     Continuation* new_entry() const { return new_entry_; }
 
 private:
-    const App* mangle_body(const App* obody);
-    Continuation* mangle_head(Continuation* ocontinuation);
-    const Def* mangle(const Def* odef);
+    const Def* rewrite(const Def* odef) override;
     bool within(const Def* def) { return scope().contains(def) || defs_.contains(def); }
 
+    bool is_dropping_;
+
     const Scope& scope_;
     Defs args_;
     Defs lift_;
     Continuation* old_entry_;
     Continuation* new_entry_;
     DefSet defs_;
-    Def2Def def2def_;
 };
 
 
-Continuation* mangle(const Scope&, Defs args, Defs lift);
+Continuation* mangle(const Scope&, Continuation* entry, Defs args, Defs lift);
 
 inline Continuation* drop(const Scope& scope, Defs args) {
-    return mangle(scope, args, Array<const Def*>());
+    return mangle(scope, scope.entry(), args, Array<const Def*>());
 }
 
 Continuation* drop(const Def* callee, const Defs specialized_args);
 
-inline Continuation* lift(const Scope& scope, Defs defs) {
-    return mangle(scope, Array<const Def*>(scope.entry()->num_params()), defs);
+inline Continuation* lift(const Scope& scope, Continuation* entry, Defs defs) {
+    return mangle(scope, entry, Array<const Def*>(entry->num_params()), defs);
 }
 
 inline Continuation* clone(const Scope& scope) {
-    return mangle(scope, Array<const Def*>(scope.entry()->num_params()), Defs());
+    return mangle(scope, scope.entry(), Array<const Def*>(scope.entry()->num_params()), Defs());
 }
 
 }
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 92ab41fa7..7117fd15a 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -40,35 +40,28 @@ class PartialEvaluator {
     size_t boundary_;
 };
 
+const Def* BetaReducer::rewrite(const Def* odef) {
+    // leave nominal defs alone
+    if (odef->isa_nom())
+        return odef;
+    return Rewriter::rewrite(odef);
+}
+
 class CondEval {
 public:
     CondEval(Continuation* callee, ScopesForest& forest, Defs args)
-        : callee_(callee)
+        : reducer_(callee->world())
+        , callee_(callee)
         , forest_(forest)
     {
         assert(callee->filter()->is_empty() || callee->filter()->size() == args.size());
         assert(callee->num_params() == args.size());
 
         for (size_t i = 0, e = args.size(); i != e; ++i)
-            old2new_[callee->param(i)] = args[i];
+            reducer_.provide_arg(callee->param(i), args[i]);
     }
 
     World& world() { return callee_->world(); }
-    const Def* instantiate(const Def* odef) {
-        if (auto ndef = old2new_.lookup(odef))
-            return *ndef;
-
-        if (odef->isa_structural() && !odef->isa<Param>()) {
-            Array<const Def*> nops(odef->num_ops());
-            for (size_t i = 0; i != odef->num_ops(); ++i)
-                nops[i] = instantiate(odef->op(i));
-
-            auto nprimop = odef->rebuild(world(), odef->type(), nops);
-            return old2new_[odef] = nprimop;
-        }
-
-        return old2new_[odef] = odef;
-    }
 
     bool eval(size_t i, bool lower2cff) {
         // the only higher order parameter that is allowed is a single 1st-order fn-parameter of a top-level continuation
@@ -82,7 +75,7 @@ class CondEval {
                 return true;
             }
 
-        return ((!callee_->is_exported() || callee_->attributes().cc == CC::Internal) && callee_->can_be_inlined()) || is_one(instantiate(filter(i)));
+        return ((!callee_->is_exported() || callee_->attributes().cc == CC::Internal) && callee_->can_be_inlined()) || is_one(reducer_.instantiate(filter(i)));
         //return is_one(instantiate(filter(i)));
     }
 
@@ -95,8 +88,8 @@ class CondEval {
     }
 
 private:
+    BetaReducer reducer_;
     Continuation* callee_;
-    Def2Def old2new_;
     ScopesForest& forest_;
 };
 
diff --git a/src/thorin/transform/partial_evaluation.h b/src/thorin/transform/partial_evaluation.h
index 4231e9fd6..fd5c2f908 100644
--- a/src/thorin/transform/partial_evaluation.h
+++ b/src/thorin/transform/partial_evaluation.h
@@ -1,10 +1,24 @@
 #ifndef THORIN_TRANSFORM_PARTIAL_EVALUATION_H
 #define THORIN_TRANSFORM_PARTIAL_EVALUATION_H
 
+#include "thorin/transform/rewrite.h"
+
 namespace thorin {
 
 class World;
 
+class BetaReducer : public Rewriter {
+public:
+    BetaReducer(World& w) : Rewriter(w) {}
+
+    void provide_arg(const Param* param, const Def* arg) {
+        insert(param, arg);
+    }
+
+protected:
+    const Def* rewrite(const Def* odef) override;
+};
+
 bool partial_evaluation(World&, bool lower2cff = false);
 
 }
diff --git a/src/thorin/transform/rewrite.cpp b/src/thorin/transform/rewrite.cpp
new file mode 100644
index 000000000..ee4e4bec0
--- /dev/null
+++ b/src/thorin/transform/rewrite.cpp
@@ -0,0 +1,66 @@
+#include "rewrite.h"
+
+namespace thorin {
+
+Rewriter::Rewriter(World& src, World& dst) : src_(src), dst_(dst) {
+    old2new_.rehash(src.defs().capacity());
+}
+
+Rewriter::Rewriter(World& src, World& dst, Rewriter& parent) : Rewriter(src, dst) {
+    old2new_ = parent.old2new_;
+}
+
+const Def* Rewriter::lookup(const thorin::Def* odef) {
+    if (auto ndef = old2new_.lookup(odef)) return *ndef;
+
+    // TODO maybe we want to deal with intrinsics in a more streamlined way
+    if (odef == src().branch())
+        return dst().branch();
+    if (odef == src().end_scope())
+        return dst().end_scope();
+    return nullptr;
+}
+
+const Def* Rewriter::instantiate(const Def* odef) {
+    auto found = lookup(odef);
+    if (found) return found;
+
+    return old2new_[odef] = rewrite(odef);
+}
+
+const Def* Rewriter::insert(const Def* odef, const Def* ndef) {
+    assert(&odef->world() == &src());
+    assert(&ndef->world() == &dst());
+    return old2new_[odef] = ndef;
+}
+
+const Def* Rewriter::rewrite(const Def* odef) {
+    if (odef == odef->world().star())
+        return insert(odef, dst().star());
+
+    auto ntype = instantiate(odef->type())->as<Type>();
+
+    Def* stub = nullptr;
+    if (odef->isa_nom()) {
+        stub = odef->stub(*this, ntype);
+        insert(odef, stub);
+    }
+
+    if (odef->isa_structural()) {
+        size_t size = odef->num_ops();
+        Array<const Def*> nops(size);
+        for (size_t i = 0; i != size; ++i) {
+            assert(odef->op(i) != odef);
+            nops[i] = instantiate(odef->op(i));
+            assert(&nops[i]->world() == &dst());
+        }
+        auto ndef = odef->rebuild(dst(), ntype, nops);
+        return ndef;
+    } else {
+        assert(odef->isa_nom() && stub);
+        stub->rebuild_from(*this, odef);
+        return stub;
+    }
+}
+
+}
\ No newline at end of file
diff --git a/src/thorin/transform/rewrite.h b/src/thorin/transform/rewrite.h
new file mode 100644
index 000000000..60fd1fa3d
--- /dev/null
+++ b/src/thorin/transform/rewrite.h
@@ -0,0 +1,32 @@
+#ifndef THORIN_REWRITE_H
+#define THORIN_REWRITE_H
+
+#include "thorin/world.h"
+
+namespace thorin {
+
+class Rewriter {
+public:
+    explicit Rewriter(World& src, World& dst);
+    explicit Rewriter(World& world) : Rewriter(world, world) {}
+
+    const Def* instantiate(const Def* odef);
+    const Def* insert(const Def* odef, const Def* ndef);
+
+    World& src() { return src_; }
+    World& dst() { return dst_; }
+
+protected:
+    explicit Rewriter(World& src, World& dst, Rewriter& parent);
+    virtual const Def* lookup(const Def* odef);
+    virtual const Def* rewrite(const Def* odef);
+
+private:
+    Def2Def old2new_;
+    World& src_;
+    World& dst_;
+};
+
+}
+
+#endif
\ No newline at end of file
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index b940e5f11..6285890ce 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -5,6 +5,7 @@
 #include <sstream>
 #include <stack>
 
+#include "thorin/transform/rewrite.h"
 #include "thorin/continuation.h"
 #include "thorin/primop.h"
 #include "thorin/world.h"
@@ -46,33 +47,33 @@ Array<const Type*> defs2types(ArrayRef<const Def*> defs) {
  * rebuild
  */
 
-const Type* NominalType::rebuild(World&  , const Type*  , Defs  ) const {
+const Type* NominalType::rebuild(World& w, const Type* t, Defs o) const {
     THORIN_UNREACHABLE;
 }
 
-const Type* BottomType         ::rebuild(World& w, const Type*  , Defs  ) const { return w.bottom_type(); }
-const Type* ClosureType        ::rebuild(World& w, const Type*  , Defs o) const { return w.closure_type(defs2types(o)); }
-const Type* DefiniteArrayType  ::rebuild(World& w, const Type*  , Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
-const Type* FnType             ::rebuild(World& w, const Type*  , Defs o) const { return w.fn_type(defs2types(o)); }
-const Type* FrameType          ::rebuild(World& w, const Type*  , Defs  ) const { return w.frame_type(); }
-const Type* IndefiniteArrayType::rebuild(World& w, const Type*  , Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
-const Type* MemType            ::rebuild(World& w, const Type*  , Defs  ) const { return w.mem_type(); }
-const Type* PrimType           ::rebuild(World& w, const Type*  , Defs  ) const { return w.prim_type(primtype_tag(), length()); }
-const Type* PtrType            ::rebuild(World& w, const Type*  , Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
-const Type* TupleType          ::rebuild(World& w, const Type*  , Defs o) const { return w.tuple_type(defs2types(o)); }
+const Type* BottomType         ::rebuild(World& w, const Type* t, Defs o) const { return w.bottom_type(); }
+const Type* ClosureType        ::rebuild(World& w, const Type* t, Defs o) const { return w.closure_type(defs2types(o)); }
+const Type* DefiniteArrayType  ::rebuild(World& w, const Type* t, Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
+const Type* FnType             ::rebuild(World& w, const Type* t, Defs o) const { return w.fn_type(defs2types(o)); }
+const Type* FrameType          ::rebuild(World& w, const Type* t, Defs o) const { return w.frame_type(); }
+const Type* IndefiniteArrayType::rebuild(World& w, const Type* t, Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
+const Type* MemType            ::rebuild(World& w, const Type* t, Defs o) const { return w.mem_type(); }
+const Type* PrimType           ::rebuild(World& w, const Type* t, Defs o) const { return w.prim_type(primtype_tag(), length()); }
+const Type* PtrType            ::rebuild(World& w, const Type* t, Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
+const Type* TupleType          ::rebuild(World& w, const Type* t, Defs o) const { return w.tuple_type(defs2types(o)); }
 
 /*
  * stub
  */
 
-StructType* StructType::stub(World& world, const Type*) const {
-    auto type = world.struct_type(name(), num_ops());
+StructType* StructType::stub(Rewriter& rewriter, const Type*) const {
+    auto type = rewriter.dst().struct_type(name(), num_ops());
     std::copy(op_names_.begin(), op_names_.end(), type->op_names().begin());
     return type;
 }
 
-VariantType* VariantType::stub(World& world, const Type*) const {
-    auto type = world.variant_type(name(), num_ops());
+VariantType* VariantType::stub(Rewriter& rewriter, const Type*) const {
+    auto type = rewriter.dst().variant_type(name(), num_ops());
     std::copy(op_names_.begin(), op_names_.end(), type->op_names().begin());
     return type;
 }
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 1018a95c3..95719d36c 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -114,7 +114,7 @@ class StructType : public NominalType, public TypeOpsMixin<TupleType> {
     {}
 
 public:
-    virtual StructType* stub(World&, const Type*) const override;
+    virtual StructType* stub(Rewriter&, const Type*) const override;
 
     friend class World;
 };
@@ -126,7 +126,7 @@ class VariantType : public NominalType, public TypeOpsMixin<TupleType> {
     {}
 
 public:
-    virtual VariantType* stub(World&, const Type*) const override;
+    virtual VariantType* stub(Rewriter&, const Type*) const override;
 
     bool has_payload() const;
 
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 661bc5e92..8f242c70c 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -22,7 +22,6 @@
 #include "thorin/type.h"
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/verify.h"
-#include "thorin/transform/cleanup_world.h"
 #include "thorin/transform/closure_conversion.h"
 #include "thorin/transform/codegen_prepare.h"
 #include "thorin/transform/dead_load_opt.h"
@@ -1296,14 +1295,14 @@ Thorin::Thorin(const std::string& name)
     : world_(std::make_unique<World>(name))
 {}
 
-void Thorin::cleanup() { cleanup_world(world_); }
-
 void Thorin::opt() {
+    bool debug_passes = getenv("THORIN_DEBUG_PASSES");
 #define RUN_PASS(pass) \
 { \
-    world().VLOG("running pass {}", #pass); \
-    pass; \
-    debug_verify(world()); \
+    world().VLOG("running pass {}", #pass);  \
+    pass;                                    \
+    debug_verify(world());                   \
+    if (debug_passes) world().dump_scoped(); \
 }
 
     RUN_PASS(cleanup())
@@ -1316,7 +1315,7 @@ void Thorin::opt() {
     RUN_PASS(hoist_enters(*this))
     RUN_PASS(dead_load_opt(world()))
     RUN_PASS(cleanup())
-    RUN_PASS(codegen_prepare(world()))
+    RUN_PASS(codegen_prepare(*this))
 }
 
 bool Thorin::ensure_stack_size(size_t new_size) {
diff --git a/src/thorin/world.h b/src/thorin/world.h
index ae070ce04..547dfdca6 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -272,6 +272,7 @@ class World : public Streamable<World> {
     Continuation* branch() const { return data_.branch_; }
     Continuation* match(const Type* type, size_t num_patterns);
     Continuation* end_scope() const { return data_.end_scope_; }
+    const App* app(const Def* callee, const Defs args, Debug dbg = {});
     const Filter* filter(const Defs, Debug dbg = {});
 
     // getters
@@ -340,7 +341,6 @@ class World : public Streamable<World> {
 
 private:
     const Param* param(const Type* type, const Continuation*, size_t index, Debug dbg);
-    const App* app(const Def* callee, const Defs args, Debug dbg = {});
     const Def* try_fold_aggregate(const Aggregate*);
     template <class F> const Def* transcendental(MathOpTag, const Def*, Debug, F&&);
     template <class F> const Def* transcendental(MathOpTag, const Def*, const Def*, Debug, F&&);
@@ -407,6 +407,7 @@ class Thorin {
     explicit Thorin(const std::string& name);
 
     World& world() { return *world_; };
+    std::unique_ptr<World>& world_container() { return world_; }
 
     /// Performs dead code, unreachable code and unused type elimination.
     void cleanup();

From e92ef28d74fcd4a1f2df680dc8a5489211550190 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 19 Sep 2023 12:00:33 +0200
Subject: [PATCH 189/342] BE: Check for is_returning, not is_exported. Resolves
 isssues where internal Continuations were not emitted at all.

---
 src/thorin/be/c/c.cpp       | 2 +-
 src/thorin/be/llvm/llvm.cpp | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 973b84dd2..c0804238b 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -360,7 +360,7 @@ void CCodeGen::emit_module() {
     forest_.for_each([&] (const Scope& scope) {
         if (scope.entry()->name() == "hls_top")
             hls_top = scope.entry();
-        else if (scope.entry()->cc() != CC::Internal && scope.entry()->is_exported())
+        else if (scope.entry()->cc() != CC::Internal && scope.entry()->is_returning())
             emit_scope(scope, forest_);
     });
     if (hls_top) {
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index b89c66f2a..c846ba293 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -314,7 +314,7 @@ CodeGen::emit_module() {
         if(scope.entry()->cc() == CC::Internal) {
             return;
         }
-        if (!scope.entry()->is_exported())
+        if (!scope.entry()->is_returning())
             return;
         emit_scope(scope, forest);
     });

From 33b54a8d313fe145628cf74fd1fc7fb86c658a56 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Sep 2023 12:05:17 +0200
Subject: [PATCH 190/342] scoped_dump: use stream1 for dumping literals

---
 src/thorin/util/scoped_dump.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index 1821c32d3..6951af638 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -108,6 +108,10 @@ void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
         stream_ops(s, app->args());
         return;
     }
+    if (auto prim_lit = def->isa<PrimLit>()) {
+        def->stream1(s);
+        return;
+    }
 
     s.fmt(Green);
     s.fmt("{}", def->op_name());

From 6e3ce324bf6267e66d0a6afa90edad32e8be067b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 4 Oct 2023 16:33:40 +0200
Subject: [PATCH 191/342] simplified/rewrote lifitng code in hls_channels

---
 src/thorin/transform/hls_channels.cpp | 66 ++++++---------------------
 1 file changed, 15 insertions(+), 51 deletions(-)

diff --git a/src/thorin/transform/hls_channels.cpp b/src/thorin/transform/hls_channels.cpp
index a5e968a79..cafabb63e 100644
--- a/src/thorin/transform/hls_channels.cpp
+++ b/src/thorin/transform/hls_channels.cpp
@@ -165,24 +165,16 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
             Def2Mode def2mode;
             extract_kernel_channels(schedule(scope), def2mode);
 
-            Array<const Type*> new_param_types(def2mode.size() + old_kernel->num_params());
-            std::copy(old_kernel->type()->types().begin(),
-                    old_kernel->type()->types().end(),
-                    new_param_types.begin());
-            size_t i = old_kernel->num_params();
-            // This vector records pairs containing:
-            // - The position of the channel parameter for the new kernel
-            // - The old global definition for the channel
-            std::vector<std::pair<size_t, const Def*>> index2def;
-            for (auto map : def2mode) {
-                index2def.emplace_back(i, map.first);
-                new_param_types[i++] = map.first->type();
-            }
+            std::vector<const Def*> channels;
+            for (auto pair : def2mode)
+                channels.push_back(pair.first);
+
+            // Map the parameters of the old kernel to the first N parameters of the new one
+            // The channels used inside the kernel are mapped to the parameters N + 1, N + 2, ...
+            auto new_kernel = lift(scope, old_kernel, channels);
 
-            // new kernels signature
-            // fn(mem, ret_cnt, ... , /channels/ )
-            auto new_kernel = world.continuation(world.fn_type(new_param_types), old_kernel->debug());
             world.make_external(new_kernel);
+            world.make_internal(old_kernel);
 
             kernel_new2old.emplace(new_kernel, old_kernel);
 
@@ -191,43 +183,14 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
             else
                 new_kernels.emplace_back(new_kernel);
 
-            world.make_internal(old_kernel);
-
-            // TODO this is now broken
-            // TODO this should likely use the mangler
-            Rewriter rewriter(world);
-            // Map the parameters of the old kernel to the first N parameters of the new one
-            // The channels used inside the kernel are mapped to the parameters N + 1, N + 2, ...
-            /*for (auto pair : index2def) {
-                auto param = new_kernel->param(pair.first);
-                rewriter.insert(pair.second, param);
-                param2arg[param] = pair.second; // (channel params, globals)
-            }
-            for (auto def : scope.defs()) {
-                if (auto cont = def->isa_nom<Continuation>()) {
-                    // Copy the basic block by calling stub
-                    // Or reuse the newly created kernel copy if def is the old kernel
-                    auto new_cont = def == old_kernel ? new_kernel : cont->mangle_stub();
-                    rewriter.insert(cont, new_cont);
-                    for (size_t i = 0; i < cont->num_params(); ++i)
-                        rewriter.insert(cont->param(i), new_cont->param(i));
-                }
-            }
-            // Rewriting the basic blocks of the kernel using the map
-            for (auto def : scope.defs()) {
-                if (auto cont = def->isa_nom<Continuation>()) { // all basic blocks of the scope
-                    if (!cont->has_body()) continue;
-                    auto body = cont->body();
-                    auto new_cont = rewriter.instantiate(cont)->isa_nom<Continuation>();
-                    auto new_callee = rewriter.instantiate(body->callee());
-                    Array<const Def*> new_args(body->num_args());
-                    for ( size_t i = 0; i < body->num_args(); ++i)
-                        new_args[i] = rewriter.instantiate(body->arg(i));
-                    new_cont->jump(new_callee, new_args, cont->debug());
-                }
+            for (size_t i = old_kernel->num_params(); i < new_kernel->num_params(); i++) {
+                auto param = new_kernel->param(i);
+                assert(param);
+                param2arg[param] = channels[i - old_kernel->num_params()]; // (channel params -> former globals)
             }
+
             if (!is_single_kernel(new_kernel))
-                kernels_ch_modes.emplace_back(def2mode);*/
+                kernels_ch_modes.emplace_back(def2mode);
     });
 
 
@@ -388,6 +351,7 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
             } else if (param == ret_param) {
                 args[i] = ret;
             } else if (auto arg = param2arg[param]) {
+                assert(arg != nullptr);
                 args[i] = arg->isa<Global>() && is_channel_type(arg->type()) ? global2slot[arg] : arg;
             } else {
                 assert(false);

From a0017801999bd72b7bbd5413d235c1b9595fb8b1 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 10 Oct 2023 14:16:58 +0200
Subject: [PATCH 192/342] fix: intrinsics can have multiple ret-alike params

---
 src/thorin/continuation.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 28f7dad8e..fec796937 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -142,7 +142,7 @@ const Param* Continuation::ret_param() const {
     const Param* result = nullptr;
     for (auto param : params()) {
         if (param->order() >= 1) {
-            assertf(result == nullptr, "only one ret_param allowed");
+            assertf(is_intrinsic() || result == nullptr, "only one ret_param allowed");
             result = param;
         }
     }

From deee19101943ce940cc9d8a4b17b95a84b2c3a3e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 12 Oct 2023 11:43:21 +0200
Subject: [PATCH 193/342] fixed regression with boundary in PE

---
 src/thorin/transform/partial_evaluation.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index 7117fd15a..af2e4cec6 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -27,7 +27,6 @@ class PartialEvaluator {
     void enqueue(Continuation* continuation) {
         if (continuation->gid() < 2 * boundary_ && done_.emplace(continuation).second)
             queue_.push(continuation);
-        queue_.push(continuation);
     }
     void eat_pe_info(Continuation*);
 

From 26eeecd642baefacfc0397ab59ee13c93d2d8e09 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 12 Oct 2023 11:51:15 +0200
Subject: [PATCH 194/342] mangle: don't instantiate dead branches if possible

---
 src/thorin/transform/mangle.cpp | 25 +++++++++++++++++++++++++
 src/thorin/world.cpp            |  1 +
 2 files changed, 26 insertions(+)

diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index 77bd201ab..96e4b8a4b 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -111,6 +111,31 @@ const Def* Mangler::rewrite(const Def* old_def) {
         return old_def;  // we leave free variables alone
     if (auto param = old_def->isa<Param>())
         assert(within(param->continuation()) && "if the param is not free, the continuation should not be either!");
+    if (auto app = old_def->isa<App>()) {
+        // HACK: only rebuild the branch we actually take
+        // this is a hack because the uses can be stale (dead stuff can have transitive uses)
+        if (auto br = app->callee()->isa_nom<Continuation>(); br && br->intrinsic() == Intrinsic::Branch) {
+            auto condition = instantiate(app->arg(1));
+            if (auto lit = condition->isa<PrimLit>()) {
+                auto mem = instantiate(app->arg(0));
+                auto target = lit->value().get_bool() ? instantiate(app->arg(2)) : instantiate(app->arg(3));
+                return dst().app(target, { mem });
+            }
+        }
+        if (auto sw = app->callee()->isa_nom<Continuation>(); sw && sw->intrinsic() == Intrinsic::Match) {
+            auto index = instantiate(app->arg(1));
+            if (auto lit = index->isa<PrimLit>()) {
+                for (size_t i = 3; i < app->num_args(); i++) {
+                    auto opattern = src().extract(app->arg(i), 0_s)->as<PrimLit>();
+                    if (instantiate(opattern) == lit) {
+                        auto mem = instantiate(app->arg(0));
+                        auto target = dst().extract(instantiate(app->arg(i)), 1);
+                        return dst().app(target, { mem }, old_def->debug());
+                    }
+                }
+            }
+        }
+    }
     auto ndef = Rewriter::rewrite(old_def);
     if (auto app = ndef->isa<App>()) {
         // If you drop a parameter it is replaced by some other def, which will be identical for all recursive calls, because it's now specialised
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 8f242c70c..6d83d3570 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -1153,6 +1153,7 @@ const Filter* World::filter(const Defs defs, Debug dbg) {
 const App* World::app(const Def* callee, const Defs args, Debug dbg) {
     if (auto continuation = callee->isa<Continuation>()) {
         switch (continuation->intrinsic()) {
+            // See also mangle::instantiate when modifying this.
             case Intrinsic::Branch: {
                 assert(args.size() == 4);
                 auto mem = args[0], cond = args[1], t = args[2], f = args[3];

From 8ca2c0a1ac71f8c5247d5ff5cd838485f34795f3 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 19 Sep 2023 13:53:35 +0200
Subject: [PATCH 195/342] CC::Thorin for internal continuations.

---
 src/thorin/be/c/c.cpp                       | 2 +-
 src/thorin/be/codegen.cpp                   | 1 +
 src/thorin/be/json/json.cpp                 | 2 +-
 src/thorin/be/llvm/llvm.cpp                 | 3 +--
 src/thorin/continuation.h                   | 6 +++---
 src/thorin/rec_stream.cpp                   | 2 +-
 src/thorin/transform/partial_evaluation.cpp | 2 +-
 7 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index c0804238b..eee2db783 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -360,7 +360,7 @@ void CCodeGen::emit_module() {
     forest_.for_each([&] (const Scope& scope) {
         if (scope.entry()->name() == "hls_top")
             hls_top = scope.entry();
-        else if (scope.entry()->cc() != CC::Internal && scope.entry()->is_returning())
+        else if (scope.entry()->cc() != CC::Thorin && scope.entry()->is_returning())
             emit_scope(scope, forest_);
     });
     if (hls_top) {
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index e53fc2efb..3f8204ef0 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -121,6 +121,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
             imported->param(i)->set_name(continuation->param(i)->name());
         imported->world().make_external(imported);
+        imported->attributes().cc = CC::C;
 
         kernels.emplace_back(continuation);
     });
diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index ea0ae03c8..8188caa3f 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -258,7 +258,7 @@ class DefTable {
                 forward_decl["fn_type"] = type;
                 forward_decl["arg_names"] = arg_names;
                 if (cont->is_external()) {
-                    if (cont->cc() == CC::Internal)
+                    if (cont->cc() == CC::Thorin)
                         forward_decl["internal"] = cont->name();
                     else
                         forward_decl["external"] = cont->name();
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 41af5aece..b4ca93de3 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -311,9 +311,8 @@ CodeGen::emit_module() {
 
     ScopesForest forest(world());
     forest.for_each([&](const Scope& scope) {
-        if(scope.entry()->cc() == CC::Internal) {
+        if(scope.entry()->cc() == CC::Thorin)
             return;
-        }
         if (!scope.entry()->is_returning())
             return;
         emit_scope(scope, forest);
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index ba72dbf91..e1aada870 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -84,9 +84,9 @@ class App : public Def {
 //------------------------------------------------------------------------------
 
 enum class CC : uint8_t {
+    Thorin,         ///< Standard calling convention for everything that solely lives inside thorin.
     C,              ///< C calling convention.
     Device,         ///< Device calling convention. These are special functions only available on a particular device.
-    Internal,       ///< External, but only for linking with artic or anyopt.
 };
 
 enum class Intrinsic : uint8_t {
@@ -129,10 +129,10 @@ class Continuation : public Def {
 public:
     struct Attributes {
         Intrinsic intrinsic = Intrinsic::None;
-        CC cc = CC::C;
+        CC cc = CC::Thorin;
 
         Attributes(Intrinsic intrinsic) : intrinsic(intrinsic) {}
-        Attributes(CC cc = CC::C) : cc(cc) {}
+        Attributes(CC cc = CC::Thorin) : cc(cc) {}
     };
 
 private:
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 037891722..cb7c5ecac 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -49,7 +49,7 @@ void RecStreamer::run() {
         s.endl().endl();
 
         if (cont->world().is_external(cont)) {
-            if (cont->attributes().cc == CC::Internal)
+            if (cont->attributes().cc == CC::Thorin)
                 s.fmt("intern ");
             else
                 s.fmt("extern ");
diff --git a/src/thorin/transform/partial_evaluation.cpp b/src/thorin/transform/partial_evaluation.cpp
index af2e4cec6..f0595640e 100644
--- a/src/thorin/transform/partial_evaluation.cpp
+++ b/src/thorin/transform/partial_evaluation.cpp
@@ -74,7 +74,7 @@ class CondEval {
                 return true;
             }
 
-        return ((!callee_->is_exported() || callee_->attributes().cc == CC::Internal) && callee_->can_be_inlined()) || is_one(reducer_.instantiate(filter(i)));
+        return ((!callee_->is_exported() || callee_->attributes().cc == CC::Thorin) && callee_->can_be_inlined()) || is_one(reducer_.instantiate(filter(i)));
         //return is_one(instantiate(filter(i)));
     }
 

From e369f91db7c6e56efebd1d8d5d2d8e40917710e5 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 19 Sep 2023 15:40:25 +0200
Subject: [PATCH 196/342] BE: Generate code for exported continuations and
 everything reachable.

---
 src/thorin/be/llvm/llvm.cpp | 39 ++++++++++++++++++++++++++++++++-----
 1 file changed, 34 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index b4ca93de3..da245fd6d 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -310,14 +310,43 @@ CodeGen::emit_module() {
     }
 
     ScopesForest forest(world());
+    std::queue<Continuation*> queue;
+    ContinuationSet emitted;
+
+    auto enqueue = [&] (Continuation* cont) {
+        if (emitted.insert(cont).second) {
+            queue.push(cont);
+        }
+    };
+
     forest.for_each([&](const Scope& scope) {
-        if(scope.entry()->cc() == CC::Thorin)
-            return;
-        if (!scope.entry()->is_returning())
-            return;
-        emit_scope(scope, forest);
+        if (scope.entry()->is_exported() && scope.entry()->cc() != CC::Thorin)
+            enqueue(scope.entry());
     });
 
+    while (!queue.empty()) {
+        Continuation* todo = pop(queue);
+        if (!todo->has_body())
+            continue;
+
+        Scope& scope = forest.get_scope(todo);
+
+        emit_scope(scope, forest);
+
+        for(auto free : scope.free_frontier()) {
+            if (const Continuation* cont_const = free->isa<Continuation>()) {
+                Continuation* cont = const_cast<Continuation*>(cont_const);
+                enqueue(cont);
+            }
+            if (const Global* global_const = free->isa<Global>()) {
+                if (auto cont_const = global_const->init()->isa<Continuation>()) {
+                    Continuation* cont = const_cast<Continuation*>(cont_const);
+                    enqueue(cont);
+                }
+            }
+        }
+    }
+
     if (debug()) dibuilder_.finalize();
 
 #if THORIN_ENABLE_RV

From dd2b925c8afd328291adad927580ddab2275cb16 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 1 Aug 2023 12:24:36 +0200
Subject: [PATCH 197/342] Fix warnings across the compiler.

---
 src/thorin/analyses/schedule.cpp |  4 ++--
 src/thorin/analyses/schedule.h   |  2 +-
 src/thorin/analyses/verify.cpp   |  4 +++-
 src/thorin/be/llvm/llvm.cpp      |  4 ----
 src/thorin/transform/mangle.cpp  |  2 +-
 src/thorin/type.cpp              | 22 +++++++++++-----------
 src/thorin/util/scoped_dump.cpp  |  3 +--
 7 files changed, 19 insertions(+), 22 deletions(-)

diff --git a/src/thorin/analyses/schedule.cpp b/src/thorin/analyses/schedule.cpp
index 999d5da85..2f7a6bd64 100644
--- a/src/thorin/analyses/schedule.cpp
+++ b/src/thorin/analyses/schedule.cpp
@@ -57,7 +57,7 @@ void Scheduler::register_defs(const Scope& s) {
     }
 }
 
-Continuation* Scheduler::early(const Def* def, DefSet* seen) {
+Continuation* Scheduler::early(const Def* def) {
     if (auto cont = early_.lookup(def)) return *cont;
     if (auto param = def->isa<Param>()) return early_[def] = param->continuation();
     assert(false);
@@ -71,7 +71,7 @@ Continuation* Scheduler::late(const Def* def) {
         result = continuation;
     } else if (auto param = def->isa<Param>()) {
         result = param->continuation();
-    } else if (auto rec = def->isa_nom()) {
+    } else if (def->isa_nom()) {
         // don't try to late-schedule recursive nodes for now
         result = early(def);
     } else {
diff --git a/src/thorin/analyses/schedule.h b/src/thorin/analyses/schedule.h
index fd0ea8c96..223fec3c1 100644
--- a/src/thorin/analyses/schedule.h
+++ b/src/thorin/analyses/schedule.h
@@ -26,7 +26,7 @@ class Scheduler {
 
     /// @name compute schedules
     //@{
-    Continuation* early(const Def*, DefSet* seen = nullptr);
+    Continuation* early(const Def*);
     Continuation* late (const Def*);
     Continuation* smart(const Def*);
     //@}
diff --git a/src/thorin/analyses/verify.cpp b/src/thorin/analyses/verify.cpp
index 0eb3aee65..614bca846 100644
--- a/src/thorin/analyses/verify.cpp
+++ b/src/thorin/analyses/verify.cpp
@@ -8,7 +8,7 @@ namespace thorin {
 
 // TODO this needs serious rewriting
 
-static bool verify_calls(World& world, ScopesForest& forest) {
+static bool verify_calls(World& world, ScopesForest&) {
     bool ok = true;
     for (auto def : world.defs()) {
         if (auto cont = def->isa<Continuation>())
@@ -46,6 +46,7 @@ static bool verify_top_level(World& world, ScopesForest& forest) {
     return ok;
 }
 
+#if 0
 static bool verify_param(World& world) {
     bool ok = true;
     for (auto def : world.defs()) {
@@ -59,6 +60,7 @@ static bool verify_param(World& world) {
     }
     return ok;
 }
+#endif
 
 void verify(World& world) {
     ScopesForest forest(world);
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index da245fd6d..a54cd2882 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1355,7 +1355,6 @@ llvm::Value* CodeGen::emit_atomic_load(llvm::IRBuilder<>& irbuilder, Continuatio
     assert(int(llvm::AtomicOrdering::NotAtomic) <= int(tag) && int(tag) <= int(llvm::AtomicOrdering::SequentiallyConsistent) && "unsupported atomic ordering");
     auto order = (llvm::AtomicOrdering)tag;
     auto scope = body->arg(3)->as<ConvOp>()->from()->as<Global>()->init()->as<DefiniteArray>();
-    auto cont = body->arg(4)->as_nom<Continuation>();
     auto load = irbuilder.CreateLoad(load_type, ptr);
     auto align = module().getDataLayout().getABITypeAlign(load_type);
     load->setAlignment(align);
@@ -1373,7 +1372,6 @@ void CodeGen::emit_atomic_store(llvm::IRBuilder<>& irbuilder, Continuation* cont
     assert(int(llvm::AtomicOrdering::NotAtomic) <= int(tag) && int(tag) <= int(llvm::AtomicOrdering::SequentiallyConsistent) && "unsupported atomic ordering");
     auto order = (llvm::AtomicOrdering)tag;
     auto scope = body->arg(4)->as<ConvOp>()->from()->as<Global>()->init()->as<DefiniteArray>();
-    auto cont = body->arg(5)->as_nom<Continuation>();
     auto store = irbuilder.CreateStore(val, ptr);
     auto align = module().getDataLayout().getABITypeAlign(convert(body->arg(2)->type()));
     store->setAlignment(align);
@@ -1397,7 +1395,6 @@ std::vector<llvm::Value*> CodeGen::emit_cmpxchg(llvm::IRBuilder<>& irbuilder, Co
     auto success_order = (llvm::AtomicOrdering)success_order_tag;
     auto failure_order = (llvm::AtomicOrdering)failure_order_tag;
     auto scope = body->arg(6)->as<ConvOp>()->from()->as<Global>()->init()->as<DefiniteArray>();
-    auto cont = body->arg(7)->as_nom<Continuation>();
     auto call = irbuilder.CreateAtomicCmpXchg(ptr, cmp, val, llvm::MaybeAlign(), success_order, failure_order, context().getOrInsertSyncScopeID(scope->as_string()));
     call->setWeak(is_weak);
     return { irbuilder.CreateExtractValue(call, 0), irbuilder.CreateExtractValue(call, 1) };
@@ -1411,7 +1408,6 @@ void CodeGen::emit_fence(llvm::IRBuilder<>& irbuilder, Continuation* continuatio
     assert(int(llvm::AtomicOrdering::NotAtomic) <= int(order_tag) && int(order_tag) <= int(llvm::AtomicOrdering::SequentiallyConsistent) && "unsupported atomic ordering");
     auto order = (llvm::AtomicOrdering)order_tag;
     auto scope = body->arg(2)->as<ConvOp>()->from()->as<Global>()->init()->as<DefiniteArray>();
-    auto cont = body->arg(3)->as_nom<Continuation>();
     irbuilder.CreateFence(order, context().getOrInsertSyncScopeID(scope->as_string()));
 }
 
diff --git a/src/thorin/transform/mangle.cpp b/src/thorin/transform/mangle.cpp
index 96e4b8a4b..98b3a559c 100644
--- a/src/thorin/transform/mangle.cpp
+++ b/src/thorin/transform/mangle.cpp
@@ -80,7 +80,7 @@ Continuation* Mangler::mangle() {
         for (auto p : recursion_wrapper->params_as_defs())
             args.push_back(p);
         size_t i = 0;
-        for (auto def : lift_)
+        for ([[maybe_unused]] auto def : lift_)
             args.push_back(new_entry()->param(recursion_wrapper->num_params() + i++));
         recursion_wrapper->jump(new_entry(), args);
     }
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index 6285890ce..d5b84b0dd 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -47,20 +47,20 @@ Array<const Type*> defs2types(ArrayRef<const Def*> defs) {
  * rebuild
  */
 
-const Type* NominalType::rebuild(World& w, const Type* t, Defs o) const {
+const Type* NominalType::rebuild(World&  , const Type*  , Defs  ) const {
     THORIN_UNREACHABLE;
 }
 
-const Type* BottomType         ::rebuild(World& w, const Type* t, Defs o) const { return w.bottom_type(); }
-const Type* ClosureType        ::rebuild(World& w, const Type* t, Defs o) const { return w.closure_type(defs2types(o)); }
-const Type* DefiniteArrayType  ::rebuild(World& w, const Type* t, Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
-const Type* FnType             ::rebuild(World& w, const Type* t, Defs o) const { return w.fn_type(defs2types(o)); }
-const Type* FrameType          ::rebuild(World& w, const Type* t, Defs o) const { return w.frame_type(); }
-const Type* IndefiniteArrayType::rebuild(World& w, const Type* t, Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
-const Type* MemType            ::rebuild(World& w, const Type* t, Defs o) const { return w.mem_type(); }
-const Type* PrimType           ::rebuild(World& w, const Type* t, Defs o) const { return w.prim_type(primtype_tag(), length()); }
-const Type* PtrType            ::rebuild(World& w, const Type* t, Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
-const Type* TupleType          ::rebuild(World& w, const Type* t, Defs o) const { return w.tuple_type(defs2types(o)); }
+const Type* BottomType         ::rebuild(World& w, const Type*  , Defs  ) const { return w.bottom_type(); }
+const Type* ClosureType        ::rebuild(World& w, const Type*  , Defs o) const { return w.closure_type(defs2types(o)); }
+const Type* DefiniteArrayType  ::rebuild(World& w, const Type*  , Defs o) const { return w.definite_array_type(o[0]->as<Type>(), dim()); }
+const Type* FnType             ::rebuild(World& w, const Type*  , Defs o) const { return w.fn_type(defs2types(o)); }
+const Type* FrameType          ::rebuild(World& w, const Type*  , Defs  ) const { return w.frame_type(); }
+const Type* IndefiniteArrayType::rebuild(World& w, const Type*  , Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
+const Type* MemType            ::rebuild(World& w, const Type*  , Defs  ) const { return w.mem_type(); }
+const Type* PrimType           ::rebuild(World& w, const Type*  , Defs  ) const { return w.prim_type(primtype_tag(), length()); }
+const Type* PtrType            ::rebuild(World& w, const Type*  , Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
+const Type* TupleType          ::rebuild(World& w, const Type*  , Defs o) const { return w.tuple_type(defs2types(o)); }
 
 /*
  * stub
diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index 6951af638..36b03c015 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -14,7 +14,6 @@ void ScopedWorld::stream_cont(thorin::Stream& s, Continuation* cont) const {
     s.fmt(Reset);
     s.fmt("(");
     const FnType* t = cont->type();
-    int ret_pi = -1;
     for (size_t i = 0; i < cont->num_params(); i++) {
         s.fmt(Yellow);
         s.fmt("{}: ", cont->param(i)->unique_name());
@@ -108,7 +107,7 @@ void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
         stream_ops(s, app->args());
         return;
     }
-    if (auto prim_lit = def->isa<PrimLit>()) {
+    if (def->isa<PrimLit>()) {
         def->stream1(s);
         return;
     }

From d0af0711750c4c8f84baac11c2921e33f4af7e77 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 16 Oct 2023 16:07:35 +0200
Subject: [PATCH 198/342] Add stdbool.h and stdint.h to generated header files.

---
 src/thorin/be/c/c.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index eee2db783..f8a6c510d 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -1508,6 +1508,9 @@ void CCodeGen::emit_c_int() {
     stream_.fmt("extern \"C\" {{\n");
     stream_.fmt("#endif\n\n");
 
+    stream_.fmt("#include <stdbool.h>\n"    // for the 'bool' type
+                "#include <stdint.h>\n\n");   // for the fixed-width integer types
+
     stream_.fmt("typedef   int8_t  i8;\n"
                 "typedef  uint8_t  u8;\n"
                 "typedef  int16_t i16;\n"

From a947aa84ef546d2b6b3ec0f970b53fbe14a4da45 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 15 Dec 2023 09:21:11 +0100
Subject: [PATCH 199/342] shady: updated API

---
 src/thorin/be/shady/shady.cpp | 50 ++++++++++++++++++++++++-----------
 1 file changed, 34 insertions(+), 16 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index 297606171..b3a6ddf87 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -20,7 +20,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     ScopesForest forest(world());
     forest.for_each([&](const Scope& scope) {
-        if(scope.entry()->cc() == CC::Internal) {
+        if(scope.entry()->cc() == CC::Thorin) {
             return;
         }
         emit_scope(scope, forest);
@@ -74,9 +74,9 @@ const shady::Type* CodeGen::convert(const Type* type) {
             case PrimType_pu32: case PrimType_qu32: t = shady::int32_type(arena); break;
             case PrimType_ps64: case PrimType_qs64:
             case PrimType_pu64: case PrimType_qu64: t = shady::int64_type(arena); break;
-            case PrimType_pf16: case PrimType_qf16: assert(false && "TODO");
-            case PrimType_pf32: case PrimType_qf32: t = shady::float_type(arena); break;
-            case PrimType_pf64: case PrimType_qf64: assert(false && "TODO");
+            case PrimType_pf16: case PrimType_qf16: t = shady::fp16_type(arena); break;
+            case PrimType_pf32: case PrimType_qf32: t = shady::fp32_type(arena); break;
+            case PrimType_pf64: case PrimType_qf64: t = shady::fp64_type(arena); break;
             default: THORIN_UNREACHABLE;
         }
     } else if (auto ptr = type->isa<PtrType>()) {
@@ -251,12 +251,12 @@ void CodeGen::prepare(Continuation* cont, shady::Node*) {
     } else
         assert(bb.head);
 
-    bb.builder = shady::begin_body(module);
+    bb.builder = shady::begin_body(arena);
 }
 
 static std::optional<shady::Op> is_shady_prim_op(const Continuation* cont) {
     for (int i = 0; i < shady::PRIMOPS_COUNT; i++) {
-        if (cont->name() == shady::primop_names[i])
+        if (cont->name() == shady::get_primop_name(static_cast<shady::Op>(i)))
             return std::make_optional((shady::Op) i);
     }
     return std::nullopt;
@@ -293,10 +293,9 @@ void CodeGen::emit_epilogue(Continuation* cont) {
         bb.terminator = shady::fn_ret(arena, payload);
     } else if (body->callee() == world().branch()) {
         shady::Branch payload = {};
-        payload.args = shady::nodes(arena, 0, nullptr);
         payload.branch_condition = args[0];
-        payload.true_target      = args[1];
-        payload.false_target     = args[2];
+        payload.true_jump = shady::jump_helper(arena, args[1], shady::empty(arena));
+        payload.false_jump = shady::jump_helper(arena, args[2], shady::empty(arena));
         bb.terminator = shady::branch(arena, payload);
     } else if (auto match = body->callee()->as_nom<Continuation>(); match && match->intrinsic() == Intrinsic::Match) {
         assert(false);
@@ -332,12 +331,12 @@ void CodeGen::emit_epilogue(Continuation* cont) {
             }
         }
 
-        shady::BodyBuilder* builder = shady::begin_body(module);
+        shady::BodyBuilder* builder = shady::begin_body(arena);
 
-        shady::IndirectCall icall_payload;
+        shady::Call icall_payload;
         icall_payload.args = vec2nodes(args);
         icall_payload.callee = emit(body->callee());
-        shady::Nodes results = shady::bind_instruction(builder, shady::indirect_call(arena, icall_payload));
+        shady::Nodes results = shady::bind_instruction(builder, shady::call(arena, icall_payload));
 
         assert(args[ret_param]->tag == shady::BasicBlock_TAG);
         shady::Jump jump_payload;
@@ -391,6 +390,7 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
 
     if (auto prim_lit = def->isa<PrimLit>()) {
         const auto& box = prim_lit->value();
+        shady::FloatLiteral fl {};
         switch (prim_lit->primtype_tag()) {
             case PrimType_bool:                     v = box.get_bool() ? shady::true_lit(arena) : shady::false_lit(arena); break;
             case PrimType_ps8:  case PrimType_qs8:  v =  shady::int8_literal (arena,  box.get_s8()); break;
@@ -401,9 +401,27 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             case PrimType_pu32: case PrimType_qu32: v = shady::uint32_literal(arena, box.get_u32()); break;
             case PrimType_ps64: case PrimType_qs64: v =  shady::int64_literal(arena, box.get_s64()); break;
             case PrimType_pu64: case PrimType_qu64: v = shady::uint64_literal(arena, box.get_u64()); break;
-            case PrimType_pf16: case PrimType_qf16: assert(false && "TODO");
-            case PrimType_pf32: case PrimType_qf32: v = shady::float_type(arena); break;
-            case PrimType_pf64: case PrimType_qf64: assert(false && "TODO");
+            case PrimType_pf16: case PrimType_qf16: {
+                fl.width = shady::FloatTy16;
+                auto f = box.get_f16();
+                memcpy(&fl.value, &f, sizeof(f));
+                v = shady::float_literal(arena, fl);
+                break;
+            }
+            case PrimType_pf32: case PrimType_qf32: {
+                fl.width = shady::FloatTy32;
+                auto f = box.get_f32();
+                memcpy(&fl.value, &f, sizeof(f));
+                v = shady::float_literal(arena, fl);
+                break;
+            }
+            case PrimType_pf64: case PrimType_qf64:{
+                fl.width = shady::FloatTy64;
+                auto f = box.get_f64();
+                memcpy(&fl.value, &f, sizeof(f));
+                v = shady::float_literal(arena, fl);
+                break;
+            }
             default: THORIN_UNREACHABLE;
         }
     } else if (auto arr = def->isa<DefiniteArray>()) {
@@ -416,7 +434,7 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
         payload.element_type = convert(arr->elem_type());
         payload.size = shady::int32_literal(arena, contents.size());
         const shady::Type* arr_type = shady::arr_type(arena, payload);
-        v = shady::composite(arena, arr_type, vec2nodes(contents));
+        v = shady::composite_helper(arena, arr_type, vec2nodes(contents));
     } else if (auto cmp = def->isa<Cmp>()) {
         switch (cmp->cmp_tag()) {
             case Cmp_eq: v = mk_primop(shady::Op::eq_op,  { cmp->lhs(), cmp->rhs() }); break;

From a3d2c2292cc88eb67814579d61597d079344f170 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 15 Dec 2023 09:24:41 +0100
Subject: [PATCH 200/342] shady: fixed conversion ops

---
 src/thorin/be/shady/shady.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index b3a6ddf87..a5cc92cb3 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -468,7 +468,9 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
         defs_[def] = nullptr;
         return nullptr;
     } else if (auto bitcast = def->isa<Bitcast>()) {
-        v = emit(bitcast->from());
+        v = mk_primop(shady::Op::reinterpret_op, { bitcast->from() }, { bitcast->type() });
+    } else if (auto conversion = def->isa<Cast>()) {
+        v = mk_primop(shady::Op::convert_op, {conversion->from() }, {conversion->type() });
     } else {
         def->dump();
         THORIN_UNREACHABLE;

From 9c15f686c298b2b42a1e8495ae3943af01a0b54e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 15 Dec 2023 09:33:35 +0100
Subject: [PATCH 201/342] shady: implement a bunch of mathops

---
 src/thorin/be/shady/shady.cpp | 27 ++++++++++++++++++++++++---
 src/thorin/be/shady/shady.h   |  3 ++-
 2 files changed, 26 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/shady/shady.cpp b/src/thorin/be/shady/shady.cpp
index a5cc92cb3..3e4a91e7a 100644
--- a/src/thorin/be/shady/shady.cpp
+++ b/src/thorin/be/shady/shady.cpp
@@ -370,10 +370,15 @@ const shady::Node* CodeGen::emit_fun_decl(Continuation* cont) {
     return shady::fn_addr(arena, payload);
 }
 
+void CodeGen::unimplemented(const Def* def) {
+    world().error(def->loc(), "We don't know how to emit {} !", def);
+    abort();
+}
+
 const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
     const shady::Node* v = nullptr;
 
-    auto mk_primop = [&](shady::Op op, std::vector<const Def*> args, std::vector<const Type*> types = {}) -> const shady::Node* {
+    auto mk_primop = [&](shady::Op op, ArrayRef<const Def*> args, ArrayRef<const Type*> types = {}) -> const shady::Node* {
         shady::PrimOp payload = {};
         payload.op = op;
         std::vector<const shady::Node*> operands;
@@ -457,6 +462,23 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
             case ArithOp_shl: v = mk_primop(shady::Op::lshift_op, { arith->lhs(), arith->rhs() }); break;
             case ArithOp_shr: v = mk_primop(shady::Op::rshift_logical_op, { arith->lhs(), arith->rhs() }); break;
         }
+    } else if (auto math = def->isa<MathOp>()) {
+        switch (math->mathop_tag()) {
+            case MathOp_fmin: v = mk_primop(shady::Op::min_op, math->ops()); break;
+            case MathOp_fmax: v = mk_primop(shady::Op::max_op, math->ops()); break;
+            case MathOp_cos: v = mk_primop(shady::Op::cos_op, math->ops()); break;
+            case MathOp_sin: v = mk_primop(shady::Op::sin_op, math->ops()); break;
+            case MathOp_fabs: v = mk_primop(shady::Op::abs_op, math->ops()); break;
+            case MathOp_floor: v = mk_primop(shady::Op::floor_op, math->ops()); break;
+            case MathOp_round: v = mk_primop(shady::Op::round_op, math->ops()); break;
+            case MathOp_pow: v = mk_primop(shady::Op::pow_op, math->ops()); break;
+            case MathOp_exp: v = mk_primop(shady::Op::exp_op, math->ops()); break;
+            case MathOp_sqrt: v = mk_primop(shady::Op::sqrt_op, math->ops()); break;
+            default: {
+                unimplemented(def);
+                THORIN_UNREACHABLE;
+            }
+        }
     } else if (auto store = def->isa<Store>()) {
         mk_primop(shady::Op::store_op, { store->ptr(), store->val() });
         defs_[def] = nullptr;
@@ -472,8 +494,7 @@ const shady::Node* CodeGen::emit_bb(BB& bb, const Def* def) {
     } else if (auto conversion = def->isa<Cast>()) {
         v = mk_primop(shady::Op::convert_op, {conversion->from() }, {conversion->type() });
     } else {
-        def->dump();
-        THORIN_UNREACHABLE;
+        unimplemented(def);
     }
     assert(v && shady::is_value(v));
     defs_[def] = v;
diff --git a/src/thorin/be/shady/shady.h b/src/thorin/be/shady/shady.h
index 75280caff..648cdf0c4 100644
--- a/src/thorin/be/shady/shady.h
+++ b/src/thorin/be/shady/shady.h
@@ -55,7 +55,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<const shady::Node
 
     shady::IrArena* arena = nullptr;
     shady::Module* module = nullptr;
-    //std::vector<shady::Node*> top_level;
+
+    void unimplemented(const Def* def);
 
     shady::Node* curr_fn;
 

From ebed23c6093f35fd6fbbeeec2c3671e8ed817e0d Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 16 Jan 2024 15:15:32 +0100
Subject: [PATCH 202/342] Fixes to compile HLS code.

* CodeGen: fix get_kernel_configs call if multiple backends use the same
  host continuation.
* Keep kernel calling convention consistent in hls_channels.
* Fix Closure Conversion for HLS.
* Add find_origin to importer to simplify hls_channels and stop it from
  importing everything into the new world.
---
 src/thorin/be/codegen.cpp                   | 25 ++++++++++++---------
 src/thorin/transform/closure_conversion.cpp |  2 ++
 src/thorin/transform/hls_channels.cpp       | 15 ++++++-------
 src/thorin/transform/importer.cpp           |  8 +++++++
 src/thorin/transform/importer.h             |  1 +
 5 files changed, 33 insertions(+), 18 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 3f8204ef0..c1e5f6b41 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -92,19 +92,19 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         importers.emplace_back(world, accelerator_code.back().world());
     }
 
+    static const auto backend_intrinsics = std::array {
+        std::pair { CUDA,   Intrinsic::CUDA         },
+        std::pair { NVVM,   Intrinsic::NVVM         },
+        std::pair { OpenCL, Intrinsic::OpenCL       },
+        std::pair { AMDGPU, Intrinsic::AMDGPU       },
+        std::pair { HLS,    Intrinsic::HLS          },
+        std::pair { Shady,  Intrinsic::ShadyCompute }
+    };
+
     // determine different parts of the world which need to be compiled differently
     ScopesForest(world).for_each([&] (const Scope& scope) {
         auto continuation = scope.entry();
         Continuation* imported = nullptr;
-
-        static const auto backend_intrinsics = std::array {
-            std::pair { CUDA,   Intrinsic::CUDA         },
-            std::pair { NVVM,   Intrinsic::NVVM         },
-            std::pair { OpenCL, Intrinsic::OpenCL       },
-            std::pair { AMDGPU, Intrinsic::AMDGPU       },
-            std::pair { HLS,    Intrinsic::HLS          },
-            std::pair { Shady,  Intrinsic::ShadyCompute }
-        };
         for (auto [backend, intrinsic] : backend_intrinsics) {
             if (is_passed_to_intrinsic(continuation, intrinsic)) {
                 imported = importers[backend].import(continuation)->as_nom<Continuation>();
@@ -126,10 +126,15 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
         kernels.emplace_back(continuation);
     });
 
-    for (auto backend : std::array { CUDA, NVVM, OpenCL, AMDGPU, Shady }) {
+    for (auto [backend, intrinsic] : backend_intrinsics) {
+        if (backend == HLS)
+            continue;
+
         if (!accelerator_code[backend].world().empty()) {
             get_kernel_configs(accelerator_code[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
                 auto app = use->body();
+                if (app->callee()->as<Continuation>()->intrinsic() != intrinsic)
+                    return std::unique_ptr<GPUKernelConfig>(nullptr);
                 // determine whether or not this kernel uses restrict pointers
                 bool has_restrict = true;
                 DefSet allocs;
diff --git a/src/thorin/transform/closure_conversion.cpp b/src/thorin/transform/closure_conversion.cpp
index 9140f4132..2048d77a5 100644
--- a/src/thorin/transform/closure_conversion.cpp
+++ b/src/thorin/transform/closure_conversion.cpp
@@ -86,6 +86,8 @@ class ClosureConversion {
                     new_args[i] = body->arg(i);
                 else if (callee->intrinsic() == Intrinsic::Match && i > 2)
                     new_args[i] = body->arg(i);
+                else if (callee->intrinsic() == Intrinsic::HLS && i > 1)
+                    new_args[i] = body->arg(i);
                 else
                     new_args[i] = convert_def(body->arg(i));
             }
diff --git a/src/thorin/transform/hls_channels.cpp b/src/thorin/transform/hls_channels.cpp
index cafabb63e..f515cd436 100644
--- a/src/thorin/transform/hls_channels.cpp
+++ b/src/thorin/transform/hls_channels.cpp
@@ -174,7 +174,9 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
             auto new_kernel = lift(scope, old_kernel, channels);
 
             world.make_external(new_kernel);
+            new_kernel->attributes().cc = CC::C;
             world.make_internal(old_kernel);
+            old_kernel->attributes().cc = CC::Thorin;
 
             kernel_new2old.emplace(new_kernel, old_kernel);
 
@@ -236,14 +238,11 @@ DeviceParams hls_channels(Thorin& thorin, Importer& importer, Top2Kernel& top2ke
 
     // Maping hls world params (from old kernels) to old_world params. Required for host code (runtime) generation
     for (auto& elem : old_kernels_params) {
-        for (auto def : old_world.defs()) {
-            if (auto ocontinuation = def->isa_nom<Continuation>()) {
-                auto ncontinuation = elem->as<Param>()->continuation();
-                if (ncontinuation == importer.import(ocontinuation)) {
-                    elem = ocontinuation->param(elem->as<Param>()->index());
-                    break;
-                }
-            }
+        auto ncontinuation = elem->as<Param>()->continuation();
+        auto odef = importer.find_origin(ncontinuation);
+        if (odef) {
+            auto ocontinuation = odef->as<Continuation>();
+            elem = ocontinuation->param(elem->as<Param>()->index());
         }
     }
 
diff --git a/src/thorin/transform/importer.cpp b/src/thorin/transform/importer.cpp
index 9bc97bc9b..2bebb9d07 100644
--- a/src/thorin/transform/importer.cpp
+++ b/src/thorin/transform/importer.cpp
@@ -125,4 +125,12 @@ const Def* Importer::rewrite(const Def* const odef) {
     return ndef;
 }
 
+const Def* Importer::find_origin(const Def* ndef) {
+    for (auto def : src().defs()) {
+        if (ndef == lookup(def))
+            return def;
+    }
+    return nullptr;
+}
+
 }
diff --git a/src/thorin/transform/importer.h b/src/thorin/transform/importer.h
index 91110ade9..0962a4da2 100644
--- a/src/thorin/transform/importer.h
+++ b/src/thorin/transform/importer.h
@@ -19,6 +19,7 @@ class Importer : Rewriter {
     }
 
     const Def* import(const Def* odef) { return instantiate(odef); }
+    const Def* find_origin(const Def* ndef);
     bool todo() const { return todo_; }
 
 protected:

From f076776570758b2e0d2c68ab8609b5e8f2053d6e Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 1 Feb 2024 13:55:06 +0100
Subject: [PATCH 203/342] Fix GIDHash to not segfault computing a hash for
 "(Def*) nullptr".

---
 src/thorin/type.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/type.h b/src/thorin/type.h
index 2ce899290..c2f9b74ae 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -433,7 +433,7 @@ struct GIDLt {
 
 template<class T>
 struct GIDHash {
-    static hash_t hash(T n) { return thorin::murmur3(n->gid()); }
+    static hash_t hash(T n) { return thorin::murmur3(n ? n->gid() : 0); }
     static bool eq(T a, T b) { return a == b; }
     static T sentinel() { return T(1); }
 };

From 76dabb1baaf938bd19c325c791d600b67785a74f Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Tue, 6 Feb 2024 22:21:12 +0100
Subject: [PATCH 204/342] Abstract class for AMDGPUCodeGen.

---
 src/thorin/CMakeLists.txt         |  2 +
 src/thorin/be/llvm/amdgpu.cpp     | 79 +++++++++++++++++++++++++++++++
 src/thorin/be/llvm/amdgpu.h       | 33 +++++++++++++
 src/thorin/be/llvm/amdgpu_hsa.cpp | 69 +--------------------------
 src/thorin/be/llvm/amdgpu_hsa.h   | 15 +-----
 src/thorin/be/llvm/amdgpu_pal.cpp | 68 +-------------------------
 src/thorin/be/llvm/amdgpu_pal.h   | 15 +-----
 7 files changed, 120 insertions(+), 161 deletions(-)
 create mode 100644 src/thorin/be/llvm/amdgpu.cpp
 create mode 100644 src/thorin/be/llvm/amdgpu.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 440d7cafc..5280c005a 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -95,6 +95,8 @@ if(LLVM_FOUND)
         be/llvm/cpu.h
         be/llvm/llvm.cpp
         be/llvm/llvm.h
+        be/llvm/amdgpu.cpp
+        be/llvm/amdgpu.h
         be/llvm/amdgpu_hsa.cpp
         be/llvm/amdgpu_hsa.h
         be/llvm/amdgpu_pal.cpp
diff --git a/src/thorin/be/llvm/amdgpu.cpp b/src/thorin/be/llvm/amdgpu.cpp
new file mode 100644
index 000000000..cda674b45
--- /dev/null
+++ b/src/thorin/be/llvm/amdgpu.cpp
@@ -0,0 +1,79 @@
+#include "thorin/be/llvm/amdgpu.h"
+
+#include <unordered_map> // TODO don't use std::unordered_*
+
+#include "thorin/primop.h"
+#include "thorin/world.h"
+
+namespace thorin::llvm {
+
+AMDGPUCodeGen::AMDGPUCodeGen(World& world, llvm::CallingConv::ID function_calling_convention, llvm::CallingConv::ID device_calling_convention, llvm::CallingConv::ID kernel_calling_convention, const Cont2Config& kernel_config, int opt, bool debug)
+    : CodeGen(world, function_calling_convention, device_calling_convention, kernel_calling_convention, opt, debug)
+    , kernel_config_(kernel_config) {}
+
+//------------------------------------------------------------------------------
+// Kernel code
+//------------------------------------------------------------------------------
+
+void AMDGPUCodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Function* f) {
+    auto config = kernel_config_.find(continuation);
+    if (config != kernel_config_.end()) {
+        auto block = config->second->as<GPUKernelConfig>()->block_size();
+        if (std::get<0>(block) > 0 && std::get<1>(block) > 0 && std::get<2>(block) > 0) {
+            Array<llvm::Metadata*> annotation_values_wgsize(3);
+            auto int32_type = llvm::IntegerType::get(context(), 32);
+            annotation_values_wgsize[0] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<0>(block)));
+            annotation_values_wgsize[1] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<1>(block)));
+            annotation_values_wgsize[2] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<2>(block)));
+            f->setMetadata(llvm::StringRef("reqd_work_group_size"), llvm::MDNode::get(context(), llvm_ref(annotation_values_wgsize)));
+        }
+    }
+}
+
+llvm::Value* AMDGPUCodeGen::emit_global(const Global* global) {
+    if (global->is_mutable())
+        world().wdef(global, "AMDGPU: Global variable '{}' will not be synced with host", global);
+    return CodeGen::emit_global(global);
+}
+
+llvm::Value* AMDGPUCodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const MathOp* mathop) {
+    auto make_key = [] (MathOpTag tag, unsigned bitwidth) { return (static_cast<unsigned>(tag) << 16) | bitwidth; };
+    static const std::unordered_map<unsigned, std::string> ocml_functions = {
+#define MATH_FUNCTION(name) \
+        { make_key(MathOp_##name, 32), "__ocml_" #name "_f32" }, \
+        { make_key(MathOp_##name, 64), "__ocml_" #name "_f64" },
+        MATH_FUNCTION(fabs)
+        MATH_FUNCTION(copysign)
+        MATH_FUNCTION(round)
+        MATH_FUNCTION(floor)
+        MATH_FUNCTION(ceil)
+        MATH_FUNCTION(fmin)
+        MATH_FUNCTION(fmax)
+        MATH_FUNCTION(cos)
+        MATH_FUNCTION(sin)
+        MATH_FUNCTION(tan)
+        MATH_FUNCTION(acos)
+        MATH_FUNCTION(asin)
+        MATH_FUNCTION(atan)
+        MATH_FUNCTION(atan2)
+        MATH_FUNCTION(sqrt)
+        MATH_FUNCTION(cbrt)
+        MATH_FUNCTION(pow)
+        MATH_FUNCTION(exp)
+        MATH_FUNCTION(exp2)
+        MATH_FUNCTION(log)
+        MATH_FUNCTION(log2)
+        MATH_FUNCTION(log10)
+#undef MATH_FUNCTION
+    };
+    auto key = make_key(mathop->mathop_tag(), num_bits(mathop->type()->primtype_tag()));
+    auto call = call_math_function(irbuilder, mathop, ocml_functions.at(key));
+    llvm::cast<llvm::CallInst>(call)->setCallingConv(function_calling_convention_);
+    return call;
+}
+
+Continuation* AMDGPUCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
+    return emit_reserve_shared(irbuilder, continuation, true);
+}
+
+}
diff --git a/src/thorin/be/llvm/amdgpu.h b/src/thorin/be/llvm/amdgpu.h
new file mode 100644
index 000000000..82e53187d
--- /dev/null
+++ b/src/thorin/be/llvm/amdgpu.h
@@ -0,0 +1,33 @@
+#ifndef THORIN_BE_LLVM_AMDGPU_H
+#define THORIN_BE_LLVM_AMDGPU_H
+
+#include "thorin/be/llvm/llvm.h"
+
+namespace thorin {
+
+namespace llvm {
+
+namespace llvm = ::llvm;
+
+class AMDGPUCodeGen : public CodeGen {
+public:
+    AMDGPUCodeGen(World&, llvm::CallingConv::ID, llvm::CallingConv::ID, llvm::CallingConv::ID, const Cont2Config&, int opt, bool debug);
+
+    const char* file_ext() const override { return ".amdgpu"; }
+
+protected:
+    void emit_fun_decl_hook(Continuation*, llvm::Function*) override;
+    virtual llvm::Function* emit_fun_decl(Continuation*) = 0;
+    llvm::Value* emit_global(const Global*) override;
+    llvm::Value* emit_mathop(llvm::IRBuilder<>&, const MathOp*) override;
+    Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
+    std::string get_alloc_name() const override { return "malloc"; }
+
+    const Cont2Config& kernel_config_;
+};
+
+}
+
+}
+
+#endif
diff --git a/src/thorin/be/llvm/amdgpu_hsa.cpp b/src/thorin/be/llvm/amdgpu_hsa.cpp
index 8b88facd4..4b886159f 100644
--- a/src/thorin/be/llvm/amdgpu_hsa.cpp
+++ b/src/thorin/be/llvm/amdgpu_hsa.cpp
@@ -1,15 +1,9 @@
 #include "thorin/be/llvm/amdgpu_hsa.h"
 
-#include <unordered_map> // TODO don't use std::unordered_*
-
-#include "thorin/primop.h"
-#include "thorin/world.h"
-
 namespace thorin::llvm {
 
 AMDGPUHSACodeGen::AMDGPUHSACodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, opt, debug)
-    , kernel_config_(kernel_config)
+    : AMDGPUCodeGen(world, llvm::CallingConv::C, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_KERNEL, kernel_config, opt, debug)
 {
     module().setDataLayout("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7");
     module().setTargetTriple("amdgcn-amd-amdhsa");
@@ -19,21 +13,6 @@ AMDGPUHSACodeGen::AMDGPUHSACodeGen(World& world, const Cont2Config& kernel_confi
 // Kernel code
 //------------------------------------------------------------------------------
 
-void AMDGPUHSACodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Function* f) {
-    auto config = kernel_config_.find(continuation);
-    if (config != kernel_config_.end()) {
-        auto block = config->second->as<GPUKernelConfig>()->block_size();
-        if (std::get<0>(block) > 0 && std::get<1>(block) > 0 && std::get<2>(block) > 0) {
-            Array<llvm::Metadata*> annotation_values_wgsize(3);
-            auto int32_type = llvm::IntegerType::get(context(), 32);
-            annotation_values_wgsize[0] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<0>(block)));
-            annotation_values_wgsize[1] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<1>(block)));
-            annotation_values_wgsize[2] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<2>(block)));
-            f->setMetadata(llvm::StringRef("reqd_work_group_size"), llvm::MDNode::get(context(), llvm_ref(annotation_values_wgsize)));
-        }
-    }
-}
-
 llvm::Function* AMDGPUHSACodeGen::emit_fun_decl(Continuation* continuation) {
     if (continuation->name() == "llvm.amdgcn.implicitarg.ptr")
         if (auto f = defs_.lookup(entry_); f && llvm::isa<llvm::Function>(*f))
@@ -44,50 +23,4 @@ llvm::Function* AMDGPUHSACodeGen::emit_fun_decl(Continuation* continuation) {
     return CodeGen::emit_fun_decl(continuation);
 }
 
-llvm::Value* AMDGPUHSACodeGen::emit_global(const Global* global) {
-    if (global->is_mutable())
-        world().wdef(global, "AMDGPU: Global variable '{}' will not be synced with host", global);
-    return CodeGen::emit_global(global);
-}
-
-llvm::Value* AMDGPUHSACodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const MathOp* mathop) {
-    auto make_key = [] (MathOpTag tag, unsigned bitwidth) { return (static_cast<unsigned>(tag) << 16) | bitwidth; };
-    static const std::unordered_map<unsigned, std::string> ocml_functions = {
-#define MATH_FUNCTION(name) \
-        { make_key(MathOp_##name, 32), "__ocml_" #name "_f32" }, \
-        { make_key(MathOp_##name, 64), "__ocml_" #name "_f64" },
-        MATH_FUNCTION(fabs)
-        MATH_FUNCTION(copysign)
-        MATH_FUNCTION(round)
-        MATH_FUNCTION(floor)
-        MATH_FUNCTION(ceil)
-        MATH_FUNCTION(fmin)
-        MATH_FUNCTION(fmax)
-        MATH_FUNCTION(cos)
-        MATH_FUNCTION(sin)
-        MATH_FUNCTION(tan)
-        MATH_FUNCTION(acos)
-        MATH_FUNCTION(asin)
-        MATH_FUNCTION(atan)
-        MATH_FUNCTION(atan2)
-        MATH_FUNCTION(sqrt)
-        MATH_FUNCTION(cbrt)
-        MATH_FUNCTION(pow)
-        MATH_FUNCTION(exp)
-        MATH_FUNCTION(exp2)
-        MATH_FUNCTION(log)
-        MATH_FUNCTION(log2)
-        MATH_FUNCTION(log10)
-#undef MATH_FUNCTION
-    };
-    auto key = make_key(mathop->mathop_tag(), num_bits(mathop->type()->primtype_tag()));
-    auto call = call_math_function(irbuilder, mathop, ocml_functions.at(key));
-    llvm::cast<llvm::CallInst>(call)->setCallingConv(function_calling_convention_);
-    return call;
-}
-
-Continuation* AMDGPUHSACodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
-    return emit_reserve_shared(irbuilder, continuation, true);
-}
-
 }
diff --git a/src/thorin/be/llvm/amdgpu_hsa.h b/src/thorin/be/llvm/amdgpu_hsa.h
index 05fb39044..32c4ace94 100644
--- a/src/thorin/be/llvm/amdgpu_hsa.h
+++ b/src/thorin/be/llvm/amdgpu_hsa.h
@@ -1,31 +1,20 @@
 #ifndef THORIN_BE_LLVM_AMDGPU_HSA_H
 #define THORIN_BE_LLVM_AMDGPU_HSA_H
 
-#include "thorin/be/llvm/llvm.h"
+#include "thorin/be/llvm/amdgpu.h"
 
 namespace thorin {
 
-class Load;
-
 namespace llvm {
 
 namespace llvm = ::llvm;
 
-class AMDGPUHSACodeGen : public CodeGen {
+class AMDGPUHSACodeGen : public AMDGPUCodeGen {
 public:
     AMDGPUHSACodeGen(World& world, const Cont2Config&, int opt, bool debug);
 
-    const char* file_ext() const override { return ".amdgpu"; }
-
 protected:
-    void emit_fun_decl_hook(Continuation*, llvm::Function*) override;
     llvm::Function* emit_fun_decl(Continuation*) override;
-    llvm::Value* emit_global(const Global*) override;
-    llvm::Value* emit_mathop(llvm::IRBuilder<>&, const MathOp*) override;
-    Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
-    std::string get_alloc_name() const override { return "malloc"; }
-
-    const Cont2Config& kernel_config_;
 };
 
 }
diff --git a/src/thorin/be/llvm/amdgpu_pal.cpp b/src/thorin/be/llvm/amdgpu_pal.cpp
index 9ea8562c2..8fac12d08 100644
--- a/src/thorin/be/llvm/amdgpu_pal.cpp
+++ b/src/thorin/be/llvm/amdgpu_pal.cpp
@@ -1,15 +1,9 @@
 #include "thorin/be/llvm/amdgpu_pal.h"
 
-#include <unordered_map> // TODO don't use std::unordered_*
-
-#include "thorin/primop.h"
-#include "thorin/world.h"
-
 namespace thorin::llvm {
 
 AMDGPUPALCodeGen::AMDGPUPALCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
-    : CodeGen(world, llvm::CallingConv::AMDGPU_Gfx, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_CS, opt, debug)
-    , kernel_config_(kernel_config)
+    : AMDGPUCodeGen(world, llvm::CallingConv::AMDGPU_Gfx, llvm::CallingConv::C, llvm::CallingConv::AMDGPU_CS, kernel_config, opt, debug)
 {
     module().setDataLayout("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7");
     module().setTargetTriple("amdgcn-amd-amdpal");
@@ -19,69 +13,9 @@ AMDGPUPALCodeGen::AMDGPUPALCodeGen(World& world, const Cont2Config& kernel_confi
 // Kernel code
 //------------------------------------------------------------------------------
 
-void AMDGPUPALCodeGen::emit_fun_decl_hook(Continuation* continuation, llvm::Function* f) {
-    auto config = kernel_config_.find(continuation);
-    if (config != kernel_config_.end()) {
-        auto block = config->second->as<GPUKernelConfig>()->block_size();
-        if (std::get<0>(block) > 0 && std::get<1>(block) > 0 && std::get<2>(block) > 0) {
-            Array<llvm::Metadata*> annotation_values_wgsize(3);
-            auto int32_type = llvm::IntegerType::get(context(), 32);
-            annotation_values_wgsize[0] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<0>(block)));
-            annotation_values_wgsize[1] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<1>(block)));
-            annotation_values_wgsize[2] = llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(int32_type, std::get<2>(block)));
-            f->setMetadata(llvm::StringRef("reqd_work_group_size"), llvm::MDNode::get(context(), llvm_ref(annotation_values_wgsize)));
-        }
-    }
-}
 
 llvm::Function* AMDGPUPALCodeGen::emit_fun_decl(Continuation* continuation) {
     return CodeGen::emit_fun_decl(continuation);
 }
 
-llvm::Value* AMDGPUPALCodeGen::emit_global(const Global* global) {
-    if (global->is_mutable())
-        world().wdef(global, "AMDGPU: Global variable '{}' will not be synced with host", global);
-    return CodeGen::emit_global(global);
-}
-
-llvm::Value* AMDGPUPALCodeGen::emit_mathop(llvm::IRBuilder<>& irbuilder, const MathOp* mathop) {
-    auto make_key = [] (MathOpTag tag, unsigned bitwidth) { return (static_cast<unsigned>(tag) << 16) | bitwidth; };
-    static const std::unordered_map<unsigned, std::string> ocml_functions = {
-#define MATH_FUNCTION(name) \
-        { make_key(MathOp_##name, 32), "__ocml_" #name "_f32" }, \
-        { make_key(MathOp_##name, 64), "__ocml_" #name "_f64" },
-        MATH_FUNCTION(fabs)
-        MATH_FUNCTION(copysign)
-        MATH_FUNCTION(round)
-        MATH_FUNCTION(floor)
-        MATH_FUNCTION(ceil)
-        MATH_FUNCTION(fmin)
-        MATH_FUNCTION(fmax)
-        MATH_FUNCTION(cos)
-        MATH_FUNCTION(sin)
-        MATH_FUNCTION(tan)
-        MATH_FUNCTION(acos)
-        MATH_FUNCTION(asin)
-        MATH_FUNCTION(atan)
-        MATH_FUNCTION(atan2)
-        MATH_FUNCTION(sqrt)
-        MATH_FUNCTION(cbrt)
-        MATH_FUNCTION(pow)
-        MATH_FUNCTION(exp)
-        MATH_FUNCTION(exp2)
-        MATH_FUNCTION(log)
-        MATH_FUNCTION(log2)
-        MATH_FUNCTION(log10)
-#undef MATH_FUNCTION
-    };
-    auto key = make_key(mathop->mathop_tag(), num_bits(mathop->type()->primtype_tag()));
-    auto call = call_math_function(irbuilder, mathop, ocml_functions.at(key));
-    llvm::cast<llvm::CallInst>(call)->setCallingConv(function_calling_convention_);
-    return call;
-}
-
-Continuation* AMDGPUPALCodeGen::emit_reserve(llvm::IRBuilder<>& irbuilder, const Continuation* continuation) {
-    return emit_reserve_shared(irbuilder, continuation, true);
-}
-
 }
diff --git a/src/thorin/be/llvm/amdgpu_pal.h b/src/thorin/be/llvm/amdgpu_pal.h
index ab2e7fcdd..29a792136 100644
--- a/src/thorin/be/llvm/amdgpu_pal.h
+++ b/src/thorin/be/llvm/amdgpu_pal.h
@@ -1,31 +1,20 @@
 #ifndef THORIN_BE_LLVM_AMDGPU_PAL_H
 #define THORIN_BE_LLVM_AMDGPU_PAL_H
 
-#include "thorin/be/llvm/llvm.h"
+#include "thorin/be/llvm/amdgpu.h"
 
 namespace thorin {
 
-class Load;
-
 namespace llvm {
 
 namespace llvm = ::llvm;
 
-class AMDGPUPALCodeGen : public CodeGen {
+class AMDGPUPALCodeGen : public AMDGPUCodeGen {
 public:
     AMDGPUPALCodeGen(World& world, const Cont2Config&, int opt, bool debug);
 
-    const char* file_ext() const override { return ".amdgpu"; }
-
 protected:
-    void emit_fun_decl_hook(Continuation*, llvm::Function*) override;
     llvm::Function* emit_fun_decl(Continuation*) override;
-    llvm::Value* emit_global(const Global*) override;
-    llvm::Value* emit_mathop(llvm::IRBuilder<>&, const MathOp*) override;
-    Continuation* emit_reserve(llvm::IRBuilder<>&, const Continuation*) override;
-    std::string get_alloc_name() const override { return "malloc"; }
-
-    const Cont2Config& kernel_config_;
 };
 
 }

From 47f4b88ce28e83ce8419aa6a91ca3002ee7e000e Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Fri, 15 Mar 2024 12:00:21 +0100
Subject: [PATCH 205/342] NVVM: already optimize at compile time.

---
 src/thorin/be/codegen.cpp   | 2 +-
 src/thorin/be/llvm/nvvm.cpp | 4 ++--
 src/thorin/be/llvm/nvvm.h   | 2 +-
 3 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 12247b032..a690d258a 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -187,7 +187,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
     hls_kernel_launch(world, hls_host_params);
 
 #if THORIN_ENABLE_LLVM
-    if (!importers_[NVVM      ].world().empty()) cgs[NVVM      ] = std::make_unique<llvm::NVVMCodeGen     >(importers_[NVVM      ].world(), kernel_config,      debug);
+    if (!importers_[NVVM      ].world().empty()) cgs[NVVM      ] = std::make_unique<llvm::NVVMCodeGen     >(importers_[NVVM      ].world(), kernel_config, opt, debug);
     if (!importers_[AMDGPU_HSA].world().empty()) cgs[AMDGPU_HSA] = std::make_unique<llvm::AMDGPUHSACodeGen>(importers_[AMDGPU_HSA].world(), kernel_config, opt, debug);
     if (!importers_[AMDGPU_PAL].world().empty()) cgs[AMDGPU_PAL] = std::make_unique<llvm::AMDGPUPALCodeGen>(importers_[AMDGPU_PAL].world(), kernel_config, opt, debug);
 #else
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index ab8d45ecc..ac6058084 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -19,8 +19,8 @@
 
 namespace thorin::llvm {
 
-NVVMCodeGen::NVVMCodeGen(World& world, const Cont2Config& kernel_config, bool debug)
-    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::PTX_Device, llvm::CallingConv::PTX_Kernel, 0, debug)
+NVVMCodeGen::NVVMCodeGen(World& world, const Cont2Config& kernel_config, int opt, bool debug)
+    : CodeGen(world, llvm::CallingConv::C, llvm::CallingConv::PTX_Device, llvm::CallingConv::PTX_Kernel, opt, debug)
     , kernel_config_(kernel_config)
 {
     auto triple = llvm::Triple(llvm::sys::getDefaultTargetTriple());
diff --git a/src/thorin/be/llvm/nvvm.h b/src/thorin/be/llvm/nvvm.h
index ee6f5239f..1cb4c802a 100644
--- a/src/thorin/be/llvm/nvvm.h
+++ b/src/thorin/be/llvm/nvvm.h
@@ -13,7 +13,7 @@ namespace llvm = ::llvm;
 
 class NVVMCodeGen : public CodeGen {
 public:
-    NVVMCodeGen(World& world, const Cont2Config&, bool debug); // NVVM-specific optimizations are run in the runtime
+    NVVMCodeGen(World& world, const Cont2Config&, int opt, bool debug);
 
     const char* file_ext() const override { return ".nvvm"; }
 

From 924feda0bb54c07acefbf4768a2000c0177538d0 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 1 Apr 2024 21:01:50 +0200
Subject: [PATCH 206/342] Fix: Do not include execinfo.h if
 THORIN_ENABLE_CREATION_CONTEXT is not set. It is not required then, and the
 header is Linux only.

---
 src/thorin/world.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 6d83d3570..7fd1b9d0d 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -10,7 +10,10 @@
 #endif
 
 #include <cmath>
+
+#if THORIN_ENABLE_CREATION_CONTEXT
 #include <execinfo.h>
+#endif
 
 #ifdef THORIN_ENABLE_RLIMITS
 #include <sys/resource.h>

From fb0f72189c16826f69ad947fbf197ebac3cc520c Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 9 Apr 2024 18:27:09 +0200
Subject: [PATCH 207/342] Fixes #155, thanks @PearCoding for the patch.

---
 src/thorin/util/scoped_dump.cpp | 2 +-
 src/thorin/world.cpp            | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index 36b03c015..3b76427d7 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -154,7 +154,7 @@ void World::dump_scoped() const {
 }
 
 void World::dump_scoped_to_disk() const {
-    ScopedWorld s(*const_cast<World*>(this), (ScopedWorld::Config) { false });
+    ScopedWorld s(*const_cast<World*>(this), ScopedWorld::Config { false });
     auto name = this->name() + ".dump";
     std::ofstream file(name);
     Stream st(file);
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 7fd1b9d0d..96b3322f5 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -15,7 +15,7 @@
 #include <execinfo.h>
 #endif
 
-#ifdef THORIN_ENABLE_RLIMITS
+#if THORIN_ENABLE_RLIMITS
 #include <sys/resource.h>
 #endif
 
@@ -1323,7 +1323,7 @@ void Thorin::opt() {
 }
 
 bool Thorin::ensure_stack_size(size_t new_size) {
-#ifdef THORIN_ENABLE_RLIMITS
+#if THORIN_ENABLE_RLIMITS
     struct rlimit rl;
     int result = getrlimit(RLIMIT_STACK, &rl);
     if(result != 0) return false;

From 3ff6142cf3d6fa450a5d33159c2af4a8ea477d29 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 23 May 2024 10:11:19 +0200
Subject: [PATCH 208/342] fix incorrect usage of TypeOpsMixin leading to UB
 cast

---
 src/thorin/type.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/src/thorin/type.h b/src/thorin/type.h
index 95719d36c..39c799011 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -107,7 +107,7 @@ class NominalType : public Type {
     }
 };
 
-class StructType : public NominalType, public TypeOpsMixin<TupleType> {
+class StructType : public NominalType, public TypeOpsMixin<StructType> {
 private:
     StructType(World& world, Symbol name, size_t size, Debug dbg)
         : NominalType(world, Node_StructType, name, size, dbg)
@@ -119,7 +119,7 @@ class StructType : public NominalType, public TypeOpsMixin<TupleType> {
     friend class World;
 };
 
-class VariantType : public NominalType, public TypeOpsMixin<TupleType> {
+class VariantType : public NominalType, public TypeOpsMixin<VariantType> {
 private:
     VariantType(World& world, Symbol name, size_t size, Debug dbg)
         : NominalType(world, Node_VariantType, name, size, dbg)
@@ -236,7 +236,7 @@ enum class AddrSpace : uint32_t {
 };
 
 /// Pointer type.
-class PtrType : public VectorType, public TypeOpsMixin<TupleType> {
+class PtrType : public VectorType, public TypeOpsMixin<PtrType> {
 private:
     PtrType(World& world, const Type* pointee, size_t length, int32_t device, AddrSpace addr_space, Debug dbg)
         : VectorType(world, Node_PtrType, {pointee}, length, dbg)
@@ -267,7 +267,7 @@ inline bool is_thin(const Type* type) {
     return type->isa<PrimType>() || type->isa<PtrType>() || is_type_unit(type);
 }
 
-class FnType : public Type, public TypeOpsMixin<TupleType> {
+class FnType : public Type, public TypeOpsMixin<FnType> {
 protected:
     FnType(World& world, Defs ops, NodeTag tag, Debug dbg)
         : Type(world, tag, ops, dbg)
@@ -306,7 +306,7 @@ class ClosureType : public FnType {
 
 //------------------------------------------------------------------------------
 
-class ArrayType : public Type, public TypeOpsMixin<TupleType> {
+class ArrayType : public Type, public TypeOpsMixin<ArrayType> {
 protected:
     ArrayType(World& world, NodeTag tag, const Type* elem_type, Debug dbg)
         : Type(world, tag, {elem_type}, dbg)

From 4515acf6e355ac756fef42e05375e4ecc246df42 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Tue, 28 May 2024 16:31:20 +0200
Subject: [PATCH 209/342] Fix compiler warning.

---
 src/thorin/continuation.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 3d39ae816..73e47e36a 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -12,8 +12,8 @@
 namespace thorin {
 
 class Continuation;
+class Rewriter;
 class Scope;
-struct Rewriter;
 
 typedef std::vector<Continuation*> Continuations;
 

From 93258e5d86739299aaa8abffad296e889f29d319 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Tue, 28 May 2024 16:33:36 +0200
Subject: [PATCH 210/342] Add option to dump thorin IR without color.

---
 src/thorin/util/scoped_dump.cpp | 4 ++--
 src/thorin/world.h              | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index 3b76427d7..b035ac3ce 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -148,8 +148,8 @@ Stream& ScopedWorld::stream(thorin::Stream& s) const {
     return s;
 }
 
-void World::dump_scoped() const {
-    ScopedWorld s(*const_cast<World*>(this));
+void World::dump_scoped(bool use_color) const {
+    ScopedWorld s(*const_cast<World*>(this), ScopedWorld::Config { use_color });
     s.dump();
 }
 
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 547dfdca6..6e62023d3 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -301,7 +301,7 @@ class World : public Streamable<World> {
 
     /// @name logging
     //@{
-    void dump_scoped() const;
+    void dump_scoped(bool=true) const;
     void dump_scoped_to_disk() const;
     Stream& stream(Stream&) const;
     Stream& stream() { return *stream_; }

From 7793f10424502c9e3afe84cef99c3e3c12728764 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Wed, 29 May 2024 16:14:54 +0200
Subject: [PATCH 211/342] Lift builtins: Wrap continuations without bodies into
 globals.

---
 src/thorin/transform/lift_builtins.cpp | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/src/thorin/transform/lift_builtins.cpp b/src/thorin/transform/lift_builtins.cpp
index 6bd509da6..ea4261578 100644
--- a/src/thorin/transform/lift_builtins.cpp
+++ b/src/thorin/transform/lift_builtins.cpp
@@ -71,7 +71,7 @@ void lift_builtins(Thorin& thorin) {
         World& world = thorin.world();
         Continuation* cur = nullptr;
         ScopesForest forest(world);
-        forest.for_each([&] (const Scope& scope) {
+        forest.for_each<false>([&] (const Scope& scope) {
             if (cur) return;
             for (auto n : scope.f_cfg().post_order()) {
                 if (n->continuation()->order() <= 1)
@@ -105,7 +105,13 @@ void lift_builtins(Thorin& thorin) {
             }
         }
 
-        auto lifted = lift(scope, scope.entry(), defs);
+        Continuation * lifted;
+        if (scope.entry()->has_body())
+            lifted = lift(scope, scope.entry(), defs);
+        else {
+            assert(defs.size() == 0 && "Scopes without body cannot have free defs.");
+            lifted = scope.entry();
+        }
         for (auto use : cur->copy_uses()) {
             if (auto uapp = use->isa<App>()) {
                 if (auto callee = uapp->callee()->isa_nom<Continuation>()) {

From 5c2f8af80cc9106d6b5ddb009c56b8772917b347 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Fri, 15 Dec 2023 17:49:06 +0100
Subject: [PATCH 212/342] Fixes to work with a module build of RV and a
 prebuild LLVM.

---
 src/thorin/CMakeLists.txt        |  2 +-
 src/thorin/be/llvm/vectorize.cpp | 18 +-----------------
 2 files changed, 2 insertions(+), 18 deletions(-)

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 15817401d..ce15ca0c2 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -136,7 +136,7 @@ if(LLVM_FOUND)
         target_link_libraries(thorin PRIVATE ${RV_LIBRARIES})
         list(APPEND Thorin_LLVM_COMPONENTS analysis passes transformutils)
     endif()
-    llvm_config(thorin ${AnyDSL_LLVM_LINK_SHARED} ${Thorin_LLVM_COMPONENTS})
+    llvm_config(thorin ${Thorin_LLVM_COMPONENTS})
 endif()
 
 if (THORIN_ENABLE_SHADY)
diff --git a/src/thorin/be/llvm/vectorize.cpp b/src/thorin/be/llvm/vectorize.cpp
index f2c74e619..2df5454f3 100644
--- a/src/thorin/be/llvm/vectorize.cpp
+++ b/src/thorin/be/llvm/vectorize.cpp
@@ -29,7 +29,6 @@
 #include <rv/vectorizationInfo.h>
 #include <rv/resolver/resolvers.h>
 #include <rv/passes/loopExitCanonicalizer.h>
-#include <rv/passes.h>
 #include <rv/region/FunctionRegion.h>
 
 #include "thorin/primop.h"
@@ -164,20 +163,7 @@ void CodeGen::emit_vectorize(u32 vector_length, llvm::Function* kernel_func, llv
     rv::PlatformInfo platform_info(*module_.get(), &tti, &tli);
 
     if (vector_length == 1) {
-        llvm::ValueToValueMapTy argMap;
-        auto itCalleeArgs = simd_kernel_func->args().begin();
-        auto itSourceArgs = kernel_func->args().begin();
-        auto endSourceArgs = kernel_func->args().end();
-
-        for (; itSourceArgs != endSourceArgs; ++itCalleeArgs, ++itSourceArgs) {
-            argMap[&*itSourceArgs] = &*itCalleeArgs;
-        }
-
-        llvm::SmallVector<llvm::ReturnInst*,4> retVec;
-        llvm::CloneFunctionInto(simd_kernel_func, kernel_func, argMap, llvm::CloneFunctionChangeType::LocalChangesOnly, retVec);
-
-        // lower mask intrinsics for scalar code (vector_length == 1)
-        rv::lowerIntrinsics(*simd_kernel_func);
+        rv::cloneFunctionAndLowerIntrinsics(*kernel_func, *simd_kernel_func);
     } else {
         rv::Config config = rv::Config::createForFunction(*kernel_func);
         config.enableIRPolish = config.useAVX2;
@@ -215,8 +201,6 @@ void CodeGen::emit_vectorize(u32 vector_length, llvm::Function* kernel_func, llv
 
         bool vectorize_ok = vectorizer.vectorize(vec_info, FAM, nullptr);
         assert_unused(vectorize_ok);
-
-        vectorizer.finalize();
     }
 
     // inline kernel

From d06da8f462a613f17b4d07dc8ec747b1c94da52e Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 11 Jun 2024 19:29:25 +0200
Subject: [PATCH 213/342] Revert removing ${AnyDSL_LLVM_LINK_SHARED} from
 llvm_config

---
 src/thorin/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index ce15ca0c2..15817401d 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -136,7 +136,7 @@ if(LLVM_FOUND)
         target_link_libraries(thorin PRIVATE ${RV_LIBRARIES})
         list(APPEND Thorin_LLVM_COMPONENTS analysis passes transformutils)
     endif()
-    llvm_config(thorin ${Thorin_LLVM_COMPONENTS})
+    llvm_config(thorin ${AnyDSL_LLVM_LINK_SHARED} ${Thorin_LLVM_COMPONENTS})
 endif()
 
 if (THORIN_ENABLE_SHADY)

From e173b0bf37194316228622e2be7e09948a761c28 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Thu, 13 Jun 2024 15:06:47 +0200
Subject: [PATCH 214/342] Add error message if thorin is build with shared libs
 and LLVM is not.

---
 CMakeLists.txt             |  4 ++++
 cmake/modules/FindRV.cmake | 34 ----------------------------------
 2 files changed, 4 insertions(+), 34 deletions(-)
 delete mode 100644 cmake/modules/FindRV.cmake

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7939482bb..c31f7152a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -42,6 +42,10 @@ if(LLVM_FOUND)
     message(STATUS "Using LLVMConfig.cmake in: ${LLVM_DIR}")
     if(LLVM_LINK_LLVM_DYLIB)
         set(AnyDSL_LLVM_LINK_SHARED "USE_SHARED")
+    else()
+        if (BUILD_SHARED_LIBS)
+            message(SEND_ERROR "Cannot build thorin as a shared library with the current build of LLVM. Build LLVM with LLVM_LINK_LLVM_DYLIB or change BUILD_SHARED_LIBS to off.")
+        endif()
     endif()
     # check for RV
     find_package(RV)
diff --git a/cmake/modules/FindRV.cmake b/cmake/modules/FindRV.cmake
deleted file mode 100644
index 8c852e6b7..000000000
--- a/cmake/modules/FindRV.cmake
+++ /dev/null
@@ -1,34 +0,0 @@
-# Find the RV library
-#
-# Once done this will define
-#  RV_INCLUDE_DIRS  - where to find RV library include file
-#  RV_LIBRARIES     - where to find RV library
-#  RV_FOUND         - True if RV library is found
-
-find_path(RV_INCLUDE_DIR rv/rv.h
-    PATHS
-        ${LLVM_INCLUDE_DIRS}
-        ${LLVM_EXTERNAL_RV_SOURCE_DIR}/include
-        ${LLVM_BUILD_MAIN_SRC_DIR}/../rv/include
-        ${LLVM_BUILD_MAIN_SRC_DIR}/tools/rv/include)
-if(TARGET RV)
-    set(RV_LIBRARY RV)
-else()
-    find_library(RV_LIBRARY       RV       PATHS ${LLVM_LIBRARY_DIRS})
-endif()
-if(TARGET gensleef)
-    set(RV_SLEEF_LIBRARY gensleef)
-else()
-    find_library(RV_SLEEF_LIBRARY gensleef PATHS ${LLVM_LIBRARY_DIRS})
-endif()
-
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(RV DEFAULT_MSG RV_INCLUDE_DIR RV_LIBRARY)
-
-set(RV_INCLUDE_DIRS ${RV_INCLUDE_DIR})
-set(RV_LIBRARIES    ${RV_LIBRARY})
-if(RV_SLEEF_LIBRARY)
-    list(APPEND RV_LIBRARIES ${RV_SLEEF_LIBRARY})
-endif()
-
-mark_as_advanced(RV_INCLUDE_DIR RV_LIBRARY RV_SLEEF_LIBRARY)

From 2400342888fe9e04111ed6c74003c7901717e221 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 17 Jun 2024 21:35:07 +0200
Subject: [PATCH 215/342] llvm: tolerate empty call instr in case of Bottom
 calls

---
 src/thorin/be/llvm/llvm.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 8e7b5d8ff..ecea16edd 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -614,7 +614,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         }
 
         call_instr = emit_call(irbuilder, body->callee(), args);
-        if (body->callee()->type()->as<FnType>()->is_returning()) {
+        if (body->callee()->type()->as<FnType>()->is_returning() && !body->callee()->isa<Bottom>()) {
             assert(call_instr && "returning calls always involve one of those");
             assert(ret_arg && "we need a return argument too!");
 

From 0b7694d94f46bec59f079d668d3c629d70883c1b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 19 Jun 2024 10:11:33 +0200
Subject: [PATCH 216/342] hotfix: don't pre-size hashmap in Rewriter

---
 src/thorin/transform/rewrite.cpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/thorin/transform/rewrite.cpp b/src/thorin/transform/rewrite.cpp
index ee4e4bec0..ac352568f 100644
--- a/src/thorin/transform/rewrite.cpp
+++ b/src/thorin/transform/rewrite.cpp
@@ -3,7 +3,11 @@
 namespace thorin {
 
 Rewriter::Rewriter(World& src, World& dst) : src_(src), dst_(dst) {
-    old2new_.rehash(src.defs().capacity());
+    // TODO: rehash is slow-ish, especially in Debug mode
+    // Many short-lived Rewriters are created that only rebuild a tiny portion of the
+    // world, such as in CondEval. For these, we end up paying a significant amount
+    // of time just running this, leading to bad performance. Be smarter or don't do this.
+    //old2new_.rehash(src.defs().capacity());
 }
 
 Rewriter::Rewriter(World& src, World& dst, Rewriter& parent) : Rewriter(src, dst) {

From 1d5aa13217ee9681047dac206e73bdf60ed64b35 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 19 Jun 2024 10:15:56 +0200
Subject: [PATCH 217/342] scope-analysis: gate verify() behind
 THORIN_ENABLE_CHECKS

---
 src/thorin/analyses/scope.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/thorin/analyses/scope.cpp b/src/thorin/analyses/scope.cpp
index d9e7d6fc7..4688c8a4a 100644
--- a/src/thorin/analyses/scope.cpp
+++ b/src/thorin/analyses/scope.cpp
@@ -303,8 +303,10 @@ Scope& ScopesForest::get_scope(Continuation* entry) {
     Scope* ptr = scope.get();
     ptr->run();
     scopes_[entry] = std::move(scope);
+#if THORIN_ENABLE_CHECKS
     if (stack_.empty())
        ptr->verify();
+#endif
     return *ptr;
 }
 

From 5b1b623eced1d3e7ba38f68a4047c5021b24e31e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 20 Jun 2024 18:29:55 +0200
Subject: [PATCH 218/342] fix_this())

---
 src/thorin/rec_stream.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index cb7c5ecac..1dfde9ef3 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -154,10 +154,10 @@ Stream& Def::stream1(Stream& s) const {
         if (global->is_external())
             return s.fmt("{}", unique_name());
         else
-            return s.fmt("{}({, }))", op_name(), ops());
+            return s.fmt("{}({, })", op_name(), ops());
     }
 
-    return s.fmt("{}({, }))", op_name(), ops());
+    return s.fmt("{}({, })", op_name(), ops());
 }
 
 Stream& Def::stream_let(Stream& s) const {

From 73f3c2168325c78a683ab3adb9e753269dccb61c Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 24 Jun 2024 10:24:03 +0200
Subject: [PATCH 219/342] dump: don't include cont. name in param names anymore

---
 src/thorin/rec_stream.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 1dfde9ef3..1ddf66f4a 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -117,7 +117,7 @@ Stream& Def::stream(Stream& s, size_t max) const {
 
 Stream& Def::stream1(Stream& s) const {
     if (auto param = isa<Param>()) {
-        return s.fmt("{}.{}", param->continuation(), param->unique_name());
+        return s.fmt("{}", param->unique_name());
     } else if (isa<Continuation>()) {
 #if THORIN_ENABLE_CREATION_CONTEXT
         if (debug().creation_context != "")

From 607e79f87706c9da008e05d307245f78664adea1 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 24 Jun 2024 17:09:57 +0200
Subject: [PATCH 220/342] First quickfix for some of the problems mentioned in
 issue #163.

* C backend: do not expect body->arg(0) to always exist in emit_epilogue.
* LLVM Backend: Do not expect ret to exist in CodeGen::convert(Type*).
---
 src/thorin/be/c/c.cpp       | 4 +++-
 src/thorin/be/llvm/llvm.cpp | 5 ++++-
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index f8a6c510d..636eecfac 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -686,7 +686,9 @@ void CCodeGen::emit_epilogue(Continuation* cont) {
     auto&& bb = cont2bb_[cont];
     assert(cont->has_body());
     auto body = cont->body();
-    emit_debug_info(bb.tail, body->arg(0));
+    if (body->num_args() > 0) {
+        emit_debug_info(bb.tail, body->arg(0));
+    }
 
     if ((lang_ == Lang::OpenCL || (lang_ == Lang::HLS && hls_top_scope)) && (cont->is_exported()))
         emit_fun_decl(cont);
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index ecea16edd..afe2a95af 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -160,7 +160,10 @@ llvm::Type* CodeGen::convert(const Type* type) {
                 } else
                     ops.push_back(convert(op));
             }
-            assert(ret);
+
+            if (!ret) {
+                ret = llvm::Type::getVoidTy(context());
+            }
 
             if (type->tag() == Node_FnType) {
                 auto llvm_type = llvm::FunctionType::get(ret, ops, false);

From 396e97401dc7b55de7ac648c406bd481835aa6e0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 14:02:36 +0200
Subject: [PATCH 221/342] scoped_dump: fix broken output when app node is
 shared

---
 src/thorin/util/scoped_dump.cpp | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index b035ac3ce..d370eeb3a 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -49,7 +49,7 @@ void ScopedWorld::stream_cont(thorin::Stream& s, Continuation* cont) const {
 
     auto defs = *scopes_to_defs_[cont];
     stream_defs(s, defs);
-
+    s.fmt("{}", cont->body()->unique_name());
     s.fmt("\b\n}}");
 }
 
@@ -119,7 +119,6 @@ void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
 }
 
 void ScopedWorld::stream_defs(thorin::Stream& s, std::vector<const Def*>& defs) const {
-    size_t i = 0;
     for (auto def : defs) {
         s.fmt("{}: ", def->unique_name());
         s.fmt(Blue);
@@ -127,9 +126,7 @@ void ScopedWorld::stream_defs(thorin::Stream& s, std::vector<const Def*>& defs)
         s.fmt(Reset);
         s.fmt(" = ");
         stream_def(s, def);
-        if (i + 1 < defs.size())
-            s.fmt("\n");
-        i++;
+        s.fmt("\n");
     }
 }
 

From f62781b4755ee4ab7decf7be91482fcf930f5344 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 14:02:59 +0200
Subject: [PATCH 222/342] scoped_dump: print CC as part of continuations if
 non-default

---
 src/thorin/util/scoped_dump.cpp | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index d370eeb3a..7cb4102c1 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -9,6 +9,17 @@ void ScopedWorld::stream_cont(thorin::Stream& s, Continuation* cont) const {
     if (cont->is_intrinsic())
         s.fmt("intrinsic ");
 
+    switch (cont->cc()) {
+        case CC::Thorin: break;
+        case CC::C: s.fmt("cc(C) "); break;
+        case CC::Device: s.fmt("cc(Device) "); break;
+        default: s.fmt("cc(?)"); break;
+    }
+
+    s.fmt(Green);
+    s.fmt("cont ");
+    s.fmt(Reset);
+
     s.fmt(Red);
     s.fmt("{}", cont->unique_name());
     s.fmt(Reset);

From cf797d7bb8744e7408c6a42f4e60be835cfe5ca2 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 14:03:27 +0200
Subject: [PATCH 223/342] scoped_dump: ignore literals at the top-level

---
 src/thorin/util/scoped_dump.cpp | 12 ++++++++++++
 src/thorin/util/scoped_dump.h   |  1 +
 2 files changed, 13 insertions(+)

diff --git a/src/thorin/util/scoped_dump.cpp b/src/thorin/util/scoped_dump.cpp
index 7cb4102c1..d471a4320 100644
--- a/src/thorin/util/scoped_dump.cpp
+++ b/src/thorin/util/scoped_dump.cpp
@@ -112,6 +112,12 @@ void ScopedWorld::stream_ops(thorin::Stream& s, Defs ops) const {
     s.fmt(")");
 }
 
+bool ScopedWorld::print_inline(const thorin::Def* def) const {
+    if (def->isa<PrimLit>())
+        return true;
+    return false;
+}
+
 void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
     if (auto app = def->isa<App>()) {
         stream_op(s, app->callee());
@@ -122,6 +128,10 @@ void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
         def->stream1(s);
         return;
     }
+    if (def->isa_nom()) {
+        s.fmt("{}", def->unique_name());
+        return;
+    }
 
     s.fmt(Green);
     s.fmt("{}", def->op_name());
@@ -131,6 +141,8 @@ void ScopedWorld::stream_def(thorin::Stream& s, const thorin::Def* def) const {
 
 void ScopedWorld::stream_defs(thorin::Stream& s, std::vector<const Def*>& defs) const {
     for (auto def : defs) {
+        if (print_inline(def))
+            continue;
         s.fmt("{}: ", def->unique_name());
         s.fmt(Blue);
         s.fmt("{}", def->type());
diff --git a/src/thorin/util/scoped_dump.h b/src/thorin/util/scoped_dump.h
index 388c4f60d..09aff4df5 100644
--- a/src/thorin/util/scoped_dump.h
+++ b/src/thorin/util/scoped_dump.h
@@ -41,6 +41,7 @@ struct ScopedWorld : public Streamable<ScopedWorld> {
     Stream& stream(Stream&) const;
 private:
 
+    bool print_inline(const Def* def) const;
     void stream_cont(thorin::Stream& s, Continuation* cont) const;
     void prepare_def(Continuation* in, const Def* def) const;
     void stream_op(thorin::Stream&, const Def* op) const;

From 671515c1eb627d2afa346f39559149ca1b6b8a19 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 16:30:08 +0200
Subject: [PATCH 224/342] remove 'device' from PtrType

---
 src/thorin/be/json/json.cpp |  3 ---
 src/thorin/primop.cpp       |  8 ++++----
 src/thorin/rec_stream.cpp   |  1 -
 src/thorin/type.cpp         | 10 +++++-----
 src/thorin/type.h           |  6 +-----
 src/thorin/world.h          |  2 +-
 6 files changed, 11 insertions(+), 19 deletions(-)

diff --git a/src/thorin/be/json/json.cpp b/src/thorin/be/json/json.cpp
index 8188caa3f..cee970f22 100644
--- a/src/thorin/be/json/json.cpp
+++ b/src/thorin/be/json/json.cpp
@@ -128,14 +128,11 @@ class TypeTable {
             }
         } else if (auto ptrtype = type->isa<PtrType>()) {
             auto pointee_type = translate_type(ptrtype->pointee());
-            auto device = ptrtype->device();
 
             result["type"] = "ptr";
             result["args"] = { pointee_type };
             result["name"] = pointee_type + "_p_" + std::to_string(type_table.size());
             result["length"] = ptrtype->length();
-            if (device != -1)
-                result["device"] = device;
             switch (ptrtype->addr_space()) {
             case AddrSpace::Generic:
                 //result["addrspace"] = "generic"; //Default
diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index a1222f4fe..786a834d0 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -68,14 +68,14 @@ LEA::LEA(World& world, const Def* ptr, const Def* index, Debug dbg)
 {
     auto type = ptr_type();
     if (auto tuple = ptr_pointee()->isa<TupleType>()) {
-        set_type(world.ptr_type(get(tuple->types(), index), type->length(), type->device(), type->addr_space()));
+        set_type(world.ptr_type(get(tuple->types(), index), type->length(), type->addr_space()));
     } else if (auto array = ptr_pointee()->isa<ArrayType>()) {
-        set_type(world.ptr_type(array->elem_type(), type->length(), type->device(), type->addr_space()));
+        set_type(world.ptr_type(array->elem_type(), type->length(), type->addr_space()));
     } else if (auto struct_type = ptr_pointee()->isa<StructType>()) {
-        set_type(world.ptr_type(get(struct_type->types(), index), type->length(), type->device(), type->addr_space()));
+        set_type(world.ptr_type(get(struct_type->types(), index), type->length(), type->addr_space()));
     } else if (auto prim_type = ptr_pointee()->isa<PrimType>()) {
         assert(prim_type->length() > 1);
-        set_type(world.ptr_type(world.prim_type(prim_type->primtype_tag()), type->length(), type->device(), type->addr_space()));
+        set_type(world.ptr_type(world.prim_type(prim_type->primtype_tag()), type->length(), type->addr_space()));
     } else {
         THORIN_UNREACHABLE;
     }
diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 1ddf66f4a..0eb0fe449 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -209,7 +209,6 @@ Stream& Type::stream(Stream& s) const {
         if (t->is_vector()) s.fmt("<{} x", t->length());
         s.fmt("{}*", t->pointee());
         if (t->is_vector()) s.fmt(">");
-        if (t->device() != -1) s.fmt("[{}]", t->device());
 
         switch (t->addr_space()) {
             case AddrSpace::Global:   s.fmt("[Global]");   break;
diff --git a/src/thorin/type.cpp b/src/thorin/type.cpp
index d5b84b0dd..837d6ff2d 100644
--- a/src/thorin/type.cpp
+++ b/src/thorin/type.cpp
@@ -59,7 +59,7 @@ const Type* FrameType          ::rebuild(World& w, const Type*  , Defs  ) const
 const Type* IndefiniteArrayType::rebuild(World& w, const Type*  , Defs o) const { return w.indefinite_array_type(o[0]->as<Type>()); }
 const Type* MemType            ::rebuild(World& w, const Type*  , Defs  ) const { return w.mem_type(); }
 const Type* PrimType           ::rebuild(World& w, const Type*  , Defs  ) const { return w.prim_type(primtype_tag(), length()); }
-const Type* PtrType            ::rebuild(World& w, const Type*  , Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), device(), addr_space()); }
+const Type* PtrType            ::rebuild(World& w, const Type*  , Defs o) const { return w.ptr_type(o[0]->as<Type>(), length(), addr_space()); }
 const Type* TupleType          ::rebuild(World& w, const Type*  , Defs o) const { return w.tuple_type(defs2types(o)); }
 
 /*
@@ -115,7 +115,7 @@ bool use_lea(const Type* type) { return type->isa<StructType>() || type->isa<Arr
  */
 
 hash_t PtrType::vhash() const {
-    return hash_combine(VectorType::vhash(), (hash_t)device(), (hash_t)addr_space());
+    return hash_combine(VectorType::vhash(), (hash_t)addr_space());
 }
 
 //------------------------------------------------------------------------------
@@ -128,7 +128,7 @@ bool PtrType::equal(const Def* other) const {
     if (!VectorType::equal(other))
         return false;
     auto ptr = other->as<PtrType>();
-    return ptr->device() == device() && ptr->addr_space() == addr_space();
+    return ptr->addr_space() == addr_space();
 }
 
 TypeTable::TypeTable(World& world)
@@ -163,8 +163,8 @@ const PrimType* World::prim_type(PrimTypeTag tag, size_t length) {
     return length == 1 ? types_.primtypes_[i] : make<PrimType>(*this, tag, length, Debug());
 }
 
-const PtrType* World::ptr_type(const Type* pointee, size_t length, int32_t device, AddrSpace addr_space) {
-    return make<PtrType>(*this, pointee, length, device, addr_space, Debug());
+const PtrType* World::ptr_type(const Type* pointee, size_t length, AddrSpace addr_space) {
+    return make<PtrType>(*this, pointee, length, addr_space, Debug());
 }
 
 const FnType*              World::fn_type(Types args) { return make<FnType>(*this, types2defs(args), Node_FnType, Debug()); }
diff --git a/src/thorin/type.h b/src/thorin/type.h
index 39c799011..e73ad5b6f 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -238,17 +238,14 @@ enum class AddrSpace : uint32_t {
 /// Pointer type.
 class PtrType : public VectorType, public TypeOpsMixin<PtrType> {
 private:
-    PtrType(World& world, const Type* pointee, size_t length, int32_t device, AddrSpace addr_space, Debug dbg)
+    PtrType(World& world, const Type* pointee, size_t length, AddrSpace addr_space, Debug dbg)
         : VectorType(world, Node_PtrType, {pointee}, length, dbg)
         , addr_space_(addr_space)
-        , device_(device)
     {}
 
 public:
     const Type* pointee() const { return op(0)->as<Type>(); }
     AddrSpace addr_space() const { return addr_space_; }
-    int32_t device() const { return device_; }
-    bool is_host_device() const { return device_ == -1; }
 
     hash_t vhash() const override;
     bool equal(const Def* other) const override;
@@ -257,7 +254,6 @@ class PtrType : public VectorType, public TypeOpsMixin<PtrType> {
     const Type* rebuild(World&, const Type*, Defs) const override;
 
     AddrSpace addr_space_;
-    int32_t device_;
 
     friend class World;
 };
diff --git a/src/thorin/world.h b/src/thorin/world.h
index 6e62023d3..a35c95ad3 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -118,7 +118,7 @@ class World : public Streamable<World> {
     const BottomType* bottom_type() { return make<BottomType>(*this, Debug()); }
     const MemType* mem_type() { return make<MemType>(*this, Debug()); }
     const FrameType* frame_type() { return make<FrameType>(*this, Debug()); }
-    const PtrType* ptr_type(const Type* pointee, size_t length = 1, int32_t device = -1, AddrSpace addr_space = AddrSpace::Generic);
+    const PtrType* ptr_type(const Type* pointee, size_t length = 1, AddrSpace addr_space = AddrSpace::Generic);
     const FnType* fn_type() { return fn_type({}); } ///< Returns an empty @p FnType.
     const FnType* fn_type(Types args);
     const ClosureType* closure_type(Types args);

From 955acc202a0578d7ec8df13783c5e8c19d4e25a8 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 14:45:37 +0200
Subject: [PATCH 225/342] revive spir-v backend from the grave

---
 CMakeLists.txt                        |   6 +
 src/thorin/CMakeLists.txt             |   7 +
 src/thorin/be/spirv/spirv.cpp         | 975 ++++++++++++++++++++++++++
 src/thorin/be/spirv/spirv.h           | 186 +++++
 src/thorin/be/spirv/spirv_builder.hpp | 669 ++++++++++++++++++
 5 files changed, 1843 insertions(+)
 create mode 100644 src/thorin/be/spirv/spirv.cpp
 create mode 100644 src/thorin/be/spirv/spirv.h
 create mode 100644 src/thorin/be/spirv/spirv_builder.hpp

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c31f7152a..17d22bfe6 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -72,6 +72,12 @@ else()
     set(THORIN_ENABLE_SHADY TRUE)
 endif()
 
+find_package(SPIRV-Headers)
+if (SPIRV-Headers_FOUND)
+    message(STATUS "Found SPIRV-Headers at ${SPIRV-Headers_DIR}")
+    set(THORIN_ENABLE_SPIRV TRUE)
+endif()
+
 message(STATUS "Using Debug flags: ${CMAKE_CXX_FLAGS_DEBUG}")
 message(STATUS "Using Release flags: ${CMAKE_CXX_FLAGS_RELEASE}")
 if(DEFINED CMAKE_BUILD_TYPE)
diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 15817401d..ec3bc9ef1 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -124,6 +124,13 @@ if(THORIN_ENABLE_JSON)
     )
 endif()
 
+if(THORIN_ENABLE_SPIRV)
+    list(APPEND THORIN_SOURCES
+        be/spirv/spirv.cpp
+        be/spirv/spirv.h
+    )
+endif()
+
 add_library(thorin ${THORIN_SOURCES})
 target_include_directories(thorin PUBLIC ${Half_INCLUDE_DIRS} ${Thorin_ROOT_DIR}/src ${CMAKE_BINARY_DIR}/include)
 
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
new file mode 100644
index 000000000..3b2d0e146
--- /dev/null
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -0,0 +1,975 @@
+#include "thorin/be/spirv/spirv.h"
+
+#include "thorin/analyses/scope.h"
+#include "thorin/analyses/schedule.h"
+#include "thorin/analyses/domtree.h"
+
+#include <iostream>
+
+namespace thorin::spirv {
+
+/// Used as a dummy SSA value for emitting things like mem/unit
+/// Should never make it in the binary files !
+constexpr SpvId spv_none { 0 };
+
+// SPIR-V has 3 "kinds" of primitives, and the user may declare arbitrary bitwidths, the following helps in translation:
+enum class PrimTypeKind {
+    Signed, Unsigned, Float
+};
+inline PrimTypeKind classify_primtype(const PrimType* type) {
+    switch (type->tag()) {
+#define THORIN_QS_TYPE(T, M) THORIN_PS_TYPE(T, M)
+#define THORIN_PS_TYPE(T, M) \
+case PrimType_##T: \
+    return PrimTypeKind::Signed; \
+    break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_QS_TYPE
+#undef THORIN_PS_TYPE
+
+#define THORIN_QU_TYPE(T, M) THORIN_PU_TYPE(T, M)
+#define THORIN_PU_TYPE(T, M) \
+case PrimType_##T: \
+    return PrimTypeKind::Unsigned; \
+    break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_QU_TYPE
+#undef THORIN_PU_TYPE
+
+#define THORIN_QF_TYPE(T, M) THORIN_PF_TYPE(T, M)
+#define THORIN_PF_TYPE(T, M) \
+case PrimType_##T: \
+    return PrimTypeKind::Float; \
+    break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_QF_TYPE
+#undef THORIN_PF_TYPE
+        default: THORIN_UNREACHABLE;
+    }
+}
+inline const PrimType* get_primtype(World& world, PrimTypeKind kind, int bitwidth, int length) {
+#define GET_PRIMTYPE_WITH_KIND(kind) \
+switch (bitwidth) { \
+    case 8:  return world.type_p##kind##8 (length); \
+    case 16: return world.type_p##kind##16(length); \
+    case 32: return world.type_p##kind##32(length); \
+    case 64: return world.type_p##kind##64(length); \
+}
+
+#define GET_PRIMTYPE_WITH_KIND_F(kind) \
+switch (bitwidth) { \
+    case 8: world.ELOG("8-bit floats do not exist"); \
+    case 16: return world.type_p##kind##16(length); \
+    case 32: return world.type_p##kind##32(length); \
+    case 64: return world.type_p##kind##64(length); \
+}
+
+    switch (kind) {
+        case PrimTypeKind::Signed:   GET_PRIMTYPE_WITH_KIND(s);   THORIN_UNREACHABLE;
+        case PrimTypeKind::Unsigned: GET_PRIMTYPE_WITH_KIND(u);   THORIN_UNREACHABLE;
+        case PrimTypeKind::Float:    GET_PRIMTYPE_WITH_KIND_F(f); THORIN_UNREACHABLE;
+        default: THORIN_UNREACHABLE;
+    }
+
+#undef GET_PRIMTYPE_WITH_KIND
+#undef GET_PRIMTYPE_WITH_KIND_F
+}
+
+BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
+        : builder::SpvBasicBlockBuilder(fn_builder.file_builder), fn_builder(fn_builder), file_builder(fn_builder.file_builder) {
+    label = file_builder.generate_fresh_id();
+}
+
+FnBuilder::FnBuilder(CodeGen* cg, FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), cg(cg), file_builder(file_builder) {}
+
+FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
+    capability(spv::Capability::CapabilityShader);
+    capability(spv::Capability::CapabilityVariablePointers);
+    capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
+    // capability(spv::Capability::CapabilityInt16);
+    capability(spv::Capability::CapabilityInt64);
+
+    addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
+    memory_model = spv::MemoryModel::MemoryModelGLSL450;
+}
+
+SpvId FileBuilder::u32_t() {
+    if (u32_t_.id == 0)
+        u32_t_ = cg->convert(cg->world().type_pu32())->type_id;
+    return u32_t_;
+}
+
+SpvId FileBuilder::u32_constant(uint32_t pattern) {
+    return constant(u32_t(), { pattern });
+}
+
+Builtins::Builtins(FileBuilder& builder) {
+    auto& world = builder.cg->world();
+    auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
+    auto spv_uint_t = builder.cg->convert(world.type_pu32());
+    auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t->type_id);
+    auto spv_uvec3_ptp = builder.declare_ptr_type(spv::StorageClassPrivate, spv_uvec3_t->type_id);
+    auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t->type_id);
+
+    // Because we technically can have multiple entry points, we take the easy way out and make each entry point
+    // write to a private variable the actual workgroup size for that specific kernel. Dirty, but simple.
+    workgroup_size = builder.variable(spv_uvec3_ptp, spv::StorageClassPrivate);
+    builder.name(workgroup_size, "BuiltInWorkgroupSize");
+
+    num_workgroups = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(num_workgroups, spv::DecorationBuiltIn, { spv::BuiltInNumWorkgroups });
+    builder.name(num_workgroups, "BuiltInNumWorkgroups");
+
+    workgroup_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(workgroup_id, spv::DecorationBuiltIn, { spv::BuiltInWorkgroupId });
+    builder.name(workgroup_id, "BuiltInWorkgroupId");
+
+    local_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(local_id, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationId });
+    builder.name(local_id, "BuiltInLocalInvocationId");
+
+    global_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
+    builder.decorate(global_id, spv::DecorationBuiltIn, { spv::BuiltInGlobalInvocationId });
+    builder.name(global_id, "BuiltInGlobalInvocationId");
+
+    local_invocation_index = builder.variable(spv_uint_pt, spv::StorageClassInput);
+    builder.decorate(local_invocation_index, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationIndex });
+    builder.name(local_invocation_index, "BuiltInLocalInvocationIndex");
+}
+
+ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
+    builder.extension("SPV_KHR_non_semantic_info");
+    shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
+}
+
+CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
+        : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
+{}
+
+void CodeGen::emit_stream(std::ostream& out) {
+    builder_ = std::make_unique<FileBuilder>(this);
+
+    builder_->builtins = std::make_unique<Builtins>(*builder_);
+    builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
+
+    structure_loops();
+    structure_flow();
+    // cleanup_world(world());
+
+    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
+
+    auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
+    auto push_constant_struct_type = builder_->declare_struct_type({ push_constant_arr_type });
+    auto push_constant_struct_ptr_type = builder_->declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
+    builder_->name(push_constant_struct_type, "ThorinPushConstant");
+    builder_->decorate(push_constant_struct_type, spv::DecorationBlock);
+    builder_->decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
+    builder_->decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
+    auto push_constant_struct_ptr = builder_->variable(push_constant_struct_ptr_type, spv::StorageClassPushConstant);
+    builder_->name(push_constant_struct_ptr, "thorin_push_constant_data");
+
+    auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
+    for (auto& cont : world().continuations()) {
+        if (cont->is_exported()) {
+            assert(defs_.contains(cont) && kernel_config_.contains(cont));
+            auto config = kernel_config_.find(cont);
+
+            SpvId callee = defs_[cont];
+
+            FnBuilder fn_builder(this, *builder_.get());
+            fn_builder.fn_type = entry_pt_signature;
+            fn_builder.fn_ret_type = builder_->void_type;
+
+            BasicBlockBuilder* bb = fn_builder.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn_builder)).get();
+            fn_builder.bbs_to_emit.push_back(bb);
+
+            auto block = config->second->as<GPUKernelConfig>()->block_size();
+            std::vector<uint32_t> local_size = {
+                    (uint32_t) std::get<0>(block),
+                    (uint32_t) std::get<1>(block),
+                    (uint32_t) std::get<2>(block),
+            };
+
+            auto spv_uvec3_t = convert(world().type_pu32(3));
+            SpvId wg_size_constant = builder_->constant_composite(spv_uvec3_t->type_id, {
+                    builder_->u32_constant(local_size[0]),
+                    builder_->u32_constant(local_size[1]),
+                    builder_->u32_constant(local_size[2]),
+            });
+            bb->store(wg_size_constant, builder_->builtins->workgroup_size);
+
+            // iterate on cont type and extract the arguments
+            auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push))->type_id;
+            auto zero = bb->file_builder.u32_constant(0);
+            auto arr_ref = bb->access_chain(ptr_type, push_constant_struct_ptr, { zero });
+            uint32_t offset = 0;
+            std::vector<SpvId> args;
+            for (size_t i = 0; i < cont->num_params(); i++) {
+                auto param = cont->param(i);
+                auto param_type = param->type();
+                if (param_type == world().unit() || param_type == world().mem_type() || param_type->isa<FnType>()) continue;
+                assert(param_type->order() == 0);
+                auto converted = convert(param_type);
+                assert(converted->datatype != nullptr);
+                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.u32_constant(offset));
+                args.push_back(arg);
+                offset += converted->datatype->serialized_size();
+            }
+
+            bb->call(builder_->void_type, callee, args);
+            bb->return_void();
+
+            builder_->define_function(fn_builder);
+            builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
+
+            std::vector<SpvId> interface = {
+                    push_constant_struct_ptr,
+                    builder_->builtins->workgroup_size,
+                    builder_->builtins->num_workgroups,
+                    builder_->builtins->workgroup_id,
+                    builder_->builtins->local_id,
+                    builder_->builtins->global_id,
+                    builder_->builtins->local_invocation_index,
+            };
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", interface);
+
+            builder_->execution_mode(fn_builder.function_id, spv::ExecutionModeLocalSize, local_size);
+        }
+    }
+
+    builder_->finish(out);
+    builder_ = nullptr;
+}
+
+void CodeGen::emit(const thorin::Scope& scope) {
+    entry_ = scope.entry();
+    assert(entry_->is_returning());
+
+    FnBuilder fn(this, *builder_.get());
+    fn.scope = &scope;
+    fn.fn_type = convert(entry_->type())->type_id;
+    fn.fn_ret_type = get_codom_type(entry_);
+    defs_.emplace(scope.entry(), fn.function_id);
+
+    current_fn_ = &fn;
+
+    auto conts = schedule(scope);
+
+    fn.bbs_to_emit.reserve(conts.size());
+    fn.bbs.reserve(conts.size());
+    auto& bbs = fn.bbs;
+
+    for (auto cont : conts) {
+        if (cont->intrinsic() == Intrinsic::EndScope) continue;
+
+        BasicBlockBuilder* bb = bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn)).get();
+        fn.bbs_to_emit.emplace_back(bb);
+        auto [i, b] = fn.bbs_map.emplace(cont, bb);
+        assert(b);
+
+        if (debug())
+            builder_->name(bb->label, cont->name().c_str());
+        fn.labels.emplace(cont, bb->label);
+
+        if (entry_ == cont) {
+            for (auto param : entry_->params()) {
+                if (is_mem(param) || is_unit(param)) {
+                    // Nothing
+                } else if (param->order() == 0) {
+                    auto param_t = convert(param->type());
+                    auto id = fn.parameter(param_t->type_id);
+                    fn.params[param] = id;
+                    if (param->type()->isa<PtrType>()) {
+                        builder_->decorate(id, spv::DecorationAliased);
+                    }
+                }
+            }
+        } else {
+            for (auto param : cont->params()) {
+                if (is_mem(param) || is_unit(param)) {
+                    // Nothing
+                } else {
+                    // OpPhi requires the full list of predecessors (values, labels)
+                    // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
+                    // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
+                    auto type = convert(param->type())->type_id;
+                    assert(type.id != 0);
+                    bb->phis_map[param] = { type, builder_->generate_fresh_id(), {} };
+                }
+            }
+        }
+    }
+
+    for (auto cont : conts) {
+        if (cont->intrinsic() == Intrinsic::EndScope) continue;
+        assert(cont == entry_ || cont->is_basicblock());
+        emit_epilogue(cont, fn.bbs_map[cont]);
+    }
+
+    for(auto& bb : fn.bbs) {
+        for (auto& [param, phi] : bb->phis_map) {
+            bb->phis.emplace_back(&phi);
+        }
+    }
+
+    builder_->define_function(fn);
+    builder_->name(fn.function_id, scope.entry()->name());
+}
+
+SpvId CodeGen::get_codom_type(const Continuation* fn) {
+    auto ret_cont_type = fn->ret_param()->type();
+    std::vector<SpvId> types;
+    for (auto& op : ret_cont_type->ops()) {
+        if (op->isa<MemType>() || is_type_unit(op))
+            continue;
+        assert(op->order() == 0);
+        types.push_back(convert(op)->type_id);
+    }
+    if (types.empty())
+        return builder_->void_type;
+    if (types.size() == 1)
+        return types[0];
+    return builder_->declare_struct_type(types);
+}
+
+void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
+    // Handles the potential nuances of jumping to another continuation
+    auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
+        bb->branch(current_fn_->labels[succ]);
+        for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+            auto param = succ->param(i);
+            if (is_mem(param) || is_unit(param))
+                continue;
+            auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
+            phi.preds.emplace_back(args[j], current_fn_->labels[continuation]);
+            j++;
+        }
+    };
+
+    if (continuation->callee() == entry_->ret_param()) {
+        std::vector<SpvId> values;
+
+        for (auto arg : continuation->args()) {
+            assert(arg->order() == 0);
+            auto val = emit(arg, bb);
+            if (is_mem(arg) || is_unit(arg))
+                continue;
+            values.emplace_back(val);
+        }
+
+        switch (values.size()) {
+            case 0:  bb->return_void();      break;
+            case 1:  bb->return_value(values[0]); break;
+            default: bb->return_value(bb->composite(current_fn_->fn_ret_type, values));
+        }
+    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+        int index = -1;
+        for (auto& arg : continuation->args()) {
+            index++;
+            auto val = emit(arg, bb);
+            if (is_mem(arg) || is_unit(arg)) continue;
+            bb->args[arg] = val;
+            auto* param = callee->param(index);
+            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
+            phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
+        }
+        bb->branch(current_fn_->labels[callee]);
+    } else if (continuation->callee() == world().branch()) {
+        auto& domtree = current_fn_->scope->b_cfg().domtree();
+        auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
+        SpvId merge_bb;
+        if (merge_cont == current_fn_->scope->exit()) {
+            BasicBlockBuilder* unreachable_merge_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
+            current_fn_->bbs_to_emit.emplace_back(unreachable_merge_bb);
+            builder_->name(unreachable_merge_bb->label, "merge_unreachable" + continuation->name());
+            unreachable_merge_bb->unreachable();
+            merge_bb = unreachable_merge_bb->label;
+        } else {
+            // TODO create a dedicated merge bb if this one is the merge blocks for more than 1 selection construct
+            merge_bb = current_fn_->labels[merge_cont];
+        }
+
+        auto cond = emit(continuation->arg(0), bb);
+        bb->args.emplace(continuation->arg(0), cond);
+        auto tbb = current_fn_->labels[continuation->arg(1)->as_continuation()];
+        auto fbb = current_fn_->labels[continuation->arg(2)->as_continuation()];
+        bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
+        bb->branch_conditional(cond, tbb, fbb);
+    } else if (continuation->callee()->isa<Continuation>() && continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
+        /*auto val = emit(continuation->arg(0));
+        auto otherwise_bb = cont2bb(continuation->arg(1)->as_continuation());
+        auto match = irbuilder.CreateSwitch(val, otherwise_bb, continuation->num_args() - 2);
+        for (size_t i = 2; i < continuation->num_args(); i++) {
+            auto arg = continuation->arg(i)->as<Tuple>();
+            auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
+            auto case_bb    = cont2bb(arg->op(1)->as_continuation());
+            match->addCase(case_const, case_bb);
+        }*/
+        THORIN_UNREACHABLE;
+    } else if (continuation->callee()->isa<Bottom>()) {
+        bb->unreachable();
+    } else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
+        auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
+        auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
+        bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
+
+        BasicBlockBuilder* dispatch_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
+
+        auto header_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), bb);
+
+        current_fn_->bbs_to_emit.emplace(header_bb_location + 1, dispatch_bb);
+        builder_->name(dispatch_bb->label, "dispatch_" + continuation->name());
+        bb->branch(dispatch_bb->label);
+        int targets = continuation->num_ops();
+        assert(targets > 0);
+
+        // TODO handle dispatching to multiple targets
+        assert(targets == 1);
+        auto dispatch_target = continuation->op(0)->as_continuation();
+        // Extract the relevant variant & expand the tuple if necessary
+        auto arg = world().variant_extract(continuation->param(0), 0);
+        auto extracted = emit(arg, dispatch_bb);
+
+        if (dispatch_target->param(0)->type()->equal(arg->type())) {
+            auto* param = dispatch_target->param(0);
+            auto& phi = current_fn_->bbs_map[dispatch_target]->phis_map[param];
+            phi.preds.emplace_back(extracted, dispatch_bb->label);
+        } else {
+            assert(false && "TODO destructure argument");
+        }
+
+        dispatch_bb->branch(current_fn_->bbs_map[dispatch_target]->label);
+
+    } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
+        auto loop_header = continuation->op(0)->as_continuation();
+        auto header_label = current_fn_->bbs_map[loop_header]->label;
+
+        auto arg = continuation->param(0);
+        bb->args[arg] = emit(arg, bb);
+        auto* param = loop_header->param(0);
+        auto& phi = current_fn_->bbs_map[loop_header]->phis_map[param];
+        phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
+
+        bb->branch(header_label);
+    } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
+
+        int targets = continuation->num_ops();
+        assert(targets > 0);
+
+        // TODO handle dispatching to multiple targets
+        assert(targets == 1);
+        auto callee = continuation->op(0)->as_continuation();
+        // TODO phis
+        bb->branch(current_fn_->bbs_map[callee]->label);
+    } else if (auto builtin = continuation->callee()->isa_continuation(); builtin->is_imported()) {
+        // Ensure we emit previous memory operations
+        assert(is_mem(continuation->arg(0)));
+        emit(continuation->arg(0), bb);
+
+        auto productions = emit_builtin(continuation, builtin, bb);
+        auto succ = continuation->args().back()->as_continuation();
+        jump_to_next_cont_with_args(succ, productions);
+    } else if (auto intrinsic = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+        THORIN_UNREACHABLE;
+    } else { // function/closure call
+        // put all first-order args into an array
+        std::vector<SpvId> call_args;
+        const Def* ret_arg = nullptr;
+        for (auto arg : continuation->args()) {
+            if (arg->order() == 0) {
+                auto arg_type = arg->type();
+                auto arg_val = emit(arg, bb);
+                if (arg_type == world().unit() || arg_type == world().mem_type()) continue;
+                call_args.push_back(arg_val);
+            } else {
+                assert(!ret_arg);
+                ret_arg = arg;
+            }
+        }
+
+        auto ret_type = get_codom_type(continuation);
+
+        SpvId call_result;
+        if (auto called_continuation = continuation->callee()->isa_continuation()) {
+            call_result = bb->call(ret_type, emit(called_continuation, bb), call_args);
+        } else {
+            // must be a closure
+            THORIN_UNREACHABLE;
+
+            // auto closure = emit(callee);
+            // args.push_back(irbuilder.CreateExtractValue(closure, 1));
+            // call = irbuilder.CreateCall(irbuilder.CreateExtractValue(closure, 0), args);
+        }
+
+        // must be call + continuation --- call + return has been removed by codegen_prepare
+        auto succ = ret_arg->as_continuation();
+
+        size_t n = 0;
+        const Param* last_param = nullptr;
+        for (auto param : succ->params()) {
+            if (is_mem(param) || is_unit(param))
+                continue;
+            last_param = param;
+            n++;
+        }
+
+        if (n == 0) {
+            bb->branch(current_fn_->labels[succ]);
+        } else if (n == 1) {
+            bb->branch(current_fn_->labels[succ]);
+
+            auto& phi = current_fn_->bbs_map[succ]->phis_map[last_param];
+            phi.preds.emplace_back(call_result, current_fn_->labels[continuation]);
+        } else {
+            Array<SpvId> extracts(n);
+            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+                auto param = succ->param(i);
+                if (is_mem(param) || is_unit(param))
+                    continue;
+                extracts[j] = bb->extract(convert(param->type())->type_id, call_result, { (uint32_t) j });
+                j++;
+            }
+
+            bb->branch(current_fn_->labels[succ]);
+
+            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+                auto param = succ->param(i);
+                if (is_mem(param) || is_unit(param))
+                    continue;
+
+                auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
+                phi.preds.emplace_back(extracts[j], current_fn_->labels[continuation]);
+
+                j++;
+            }
+        }
+    }
+}
+
+SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
+    if (auto bin = def->isa<BinOp>()) {
+        SpvId lhs = emit(bin->lhs(), bb);
+        SpvId rhs = emit(bin->rhs(), bb);
+        ConvertedType* result_types = convert(def->type());
+        SpvId result_type = result_types->type_id;
+
+        if (auto cmp = bin->isa<Cmp>()) {
+            auto type = cmp->lhs()->type();
+            if (is_type_s(type)) {
+                switch (cmp->cmp_tag()) {
+                    case Cmp_eq: return bb->binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb->binop(spv::Op::OpSGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb->binop(spv::Op::OpSGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb->binop(spv::Op::OpSLessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb->binop(spv::Op::OpSLessThanEqual    , result_type, lhs, rhs);
+                }
+            } else if (is_type_u(type)) {
+                switch (cmp->cmp_tag()) {
+                    case Cmp_eq: return bb->binop(spv::Op::OpIEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpINotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb->binop(spv::Op::OpUGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb->binop(spv::Op::OpUGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb->binop(spv::Op::OpULessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb->binop(spv::Op::OpULessThanEqual    , result_type, lhs, rhs);
+                }
+            } else if (is_type_f(type)) {
+                switch (cmp->cmp_tag()) {
+                    // TODO look into the NaN story
+                    case Cmp_eq: return bb->binop(spv::Op::OpFOrdEqual            , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpFOrdNotEqual         , result_type, lhs, rhs);
+                    case Cmp_gt: return bb->binop(spv::Op::OpFOrdGreaterThan      , result_type, lhs, rhs);
+                    case Cmp_ge: return bb->binop(spv::Op::OpFOrdGreaterThanEqual , result_type, lhs, rhs);
+                    case Cmp_lt: return bb->binop(spv::Op::OpFOrdLessThan         , result_type, lhs, rhs);
+                    case Cmp_le: return bb->binop(spv::Op::OpFOrdLessThanEqual    , result_type, lhs, rhs);
+                }
+            } else if (type->isa<PtrType>()) {
+                assertf(false, "Physical pointers are unsupported");
+            } else if(is_type_bool(type)) {
+                switch (cmp->cmp_tag()) {
+                    // TODO look into the NaN story
+                    case Cmp_eq: return bb->binop(spv::Op::OpLogicalEqual    , result_type, lhs, rhs);
+                    case Cmp_ne: return bb->binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
+                    default: THORIN_UNREACHABLE;
+                }
+                assertf(false, "TODO: should we emulate the other comparison ops ?");
+            }
+        }
+
+        if (auto arithop = bin->isa<ArithOp>()) {
+            auto type = arithop->type();
+
+            if (is_type_f(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_add: return bb->binop(spv::Op::OpFAdd, result_type, lhs, rhs);
+                    case ArithOp_sub: return bb->binop(spv::Op::OpFSub, result_type, lhs, rhs);
+                    case ArithOp_mul: return bb->binop(spv::Op::OpFMul, result_type, lhs, rhs);
+                    case ArithOp_div: return bb->binop(spv::Op::OpFDiv, result_type, lhs, rhs);
+                    case ArithOp_rem: return bb->binop(spv::Op::OpFRem, result_type, lhs, rhs);
+                    case ArithOp_and:
+                    case ArithOp_or:
+                    case ArithOp_xor:
+                    case ArithOp_shl:
+                    case ArithOp_shr: THORIN_UNREACHABLE;
+                }
+            }
+
+            if (is_type_s(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_add: return bb->binop(spv::Op::OpIAdd                 , result_type, lhs, rhs);
+                    case ArithOp_sub: return bb->binop(spv::Op::OpISub                 , result_type, lhs, rhs);
+                    case ArithOp_mul: return bb->binop(spv::Op::OpIMul                 , result_type, lhs, rhs);
+                    case ArithOp_div: return bb->binop(spv::Op::OpSDiv                 , result_type, lhs, rhs);
+                    case ArithOp_rem: return bb->binop(spv::Op::OpSRem                 , result_type, lhs, rhs);
+                    case ArithOp_and: return bb->binop(spv::Op::OpBitwiseAnd           , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb->binop(spv::Op::OpBitwiseOr            , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb->binop(spv::Op::OpBitwiseXor           , result_type, lhs, rhs);
+                    case ArithOp_shl: return bb->binop(spv::Op::OpShiftLeftLogical     , result_type, lhs, rhs);
+                    case ArithOp_shr: return bb->binop(spv::Op::OpShiftRightArithmetic , result_type, lhs, rhs);
+                }
+            } else if (is_type_u(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_add: return bb->binop(spv::Op::OpIAdd              , result_type, lhs, rhs);
+                    case ArithOp_sub: return bb->binop(spv::Op::OpISub              , result_type, lhs, rhs);
+                    case ArithOp_mul: return bb->binop(spv::Op::OpIMul              , result_type, lhs, rhs);
+                    case ArithOp_div: return bb->binop(spv::Op::OpUDiv              , result_type, lhs, rhs);
+                    case ArithOp_rem: return bb->binop(spv::Op::OpUMod              , result_type, lhs, rhs);
+                    case ArithOp_and: return bb->binop(spv::Op::OpBitwiseAnd        , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb->binop(spv::Op::OpBitwiseOr         , result_type, lhs, rhs);
+                    case ArithOp_xor: return bb->binop(spv::Op::OpBitwiseXor        , result_type, lhs, rhs);
+                    case ArithOp_shl: return bb->binop(spv::Op::OpShiftLeftLogical  , result_type, lhs, rhs);
+                    case ArithOp_shr: return bb->binop(spv::Op::OpShiftRightLogical , result_type, lhs, rhs);
+                }
+            } else if(is_type_bool(type)) {
+                switch (arithop->arithop_tag()) {
+                    case ArithOp_and: return bb->binop(spv::Op::OpLogicalAnd      , result_type, lhs, rhs);
+                    case ArithOp_or:  return bb->binop(spv::Op::OpLogicalOr       , result_type, lhs, rhs);
+                        // Note: there is no OpLogicalXor
+                    case ArithOp_xor: return bb->binop(spv::Op::OpLogicalNotEqual , result_type, lhs, rhs);
+                    default: THORIN_UNREACHABLE;
+                }
+            }
+            THORIN_UNREACHABLE;
+        }
+    } else if (auto primlit = def->isa<PrimLit>()) {
+        Box box = primlit->value();
+        auto type = convert(def->type())->type_id;
+        SpvId constant;
+        switch (primlit->primtype_tag()) {
+            case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
+            case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
+            case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
+            case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
+            case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
+            case PrimType_ps32: case PrimType_qs32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
+            case PrimType_pu32: case PrimType_qu32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
+            case PrimType_ps64: case PrimType_qs64:
+            case PrimType_pu64: case PrimType_qu64: {
+                uint64_t value = static_cast<uint64_t>(box.get_u64());
+                uint64_t upper = value >> 32U;
+                uint64_t lower = value & 0xFFFFFFFFU;
+                constant = bb->file_builder.constant(type, { (uint32_t) lower, (uint32_t) upper });
+                break;
+            }
+            case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
+            case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
+            case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
+        }
+        return constant;
+    } else if (auto param = def->isa<Param>()) {
+        if (is_mem(param)) return spv_none;
+        if (auto param_id = current_fn_->params.lookup(param)) {
+            assert((*param_id).id != 0);
+            return *param_id;
+        } else {
+            auto val = (*current_fn_->bbs_map[param->continuation()]).phis_map[param].value;
+            assert(val.id != 0);
+            return val;
+        }
+    } else if (auto variant = def->isa<Variant>()) {
+        auto variant_type = def->type()->as<VariantType>();
+        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
+        auto tag = builder_->u32_constant(variant->index());
+
+        if (variant_datatype->elements_types.size() > 1) {
+            auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+            auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
+            auto converted_payload_type = convert(variant_type->op(variant->index()));
+
+            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0), emit(variant->value(), bb));
+            auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
+
+            std::vector<SpvId> with_tag = {tag, payload};
+            return bb->composite(convert(variant->type())->type_id, with_tag);
+        } else {
+            // Zero-sized payload case
+            std::vector<SpvId> with_tag = { tag };
+            return bb->composite(convert(variant->type())->type_id, with_tag);
+        }
+    } else if (auto vextract = def->isa<VariantExtract>()) {
+        auto variant_type = vextract->value()->type()->as<VariantType>();
+        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
+
+        auto target_type = convert(def->type());
+
+        assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
+        auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+        auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
+        auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
+        bb->store(payload, payload_arr);
+
+        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));
+    } else if (auto vindex = def->isa<VariantIndex>()) {
+        auto value = emit(vindex->op(0), bb);
+        return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
+    } else if (auto tuple = def->isa<Tuple>()) {
+        std::vector<SpvId> elements;
+        elements.resize(tuple->num_ops());
+        size_t x = 0;
+        for (auto& e : tuple->ops()) {
+            elements[x++] = emit(e, bb);
+        }
+        return bb->composite(convert(tuple->type())->type_id, elements);
+    } else if (auto structagg = def->isa<StructAgg>()) {
+        std::vector<SpvId> elements;
+        elements.resize(structagg->num_ops());
+        size_t x = 0;
+        for (auto& e : structagg->ops()) {
+            elements[x++] = emit(e, bb);
+        }
+        return bb->composite(convert(structagg->type())->type_id, elements);
+    } else if (auto access = def->isa<Access>()) {
+        // emit dependent operations first
+        emit(access->mem(), bb);
+
+        std::vector<uint32_t> operands;
+        auto ptr_type = access->ptr()->type()->as<PtrType>();
+        if (ptr_type->addr_space() == AddrSpace::Global) {
+            operands.push_back(spv::MemoryAccessAlignedMask);
+            operands.push_back( 4 ); // TODO: SPIR-V docs say to consult client API for valid values.
+        }
+        if (auto load = def->isa<Load>()) {
+            return bb->load(convert(load->out_val_type())->type_id, emit(load->ptr(), bb), operands);
+        } else if (auto store = def->isa<Store>()) {
+            bb->store(emit(store->val(), bb), emit(store->ptr(), bb), operands);
+            return spv_none;
+        } else THORIN_UNREACHABLE;
+    } else if (auto lea = def->isa<LEA>()) {
+        switch (lea->ptr_type()->addr_space()) {
+            case AddrSpace::Global:
+            case AddrSpace::Shared:
+                break;
+            default:
+                world().ELOG("LEA is only allowed in global & shared address spaces");
+                break;
+        }
+        auto type = convert(lea->ptr_type());
+        auto offset = emit(lea->index(), bb);
+        return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
+    } else if (auto aggop = def->isa<AggOp>()) {
+        auto spv_agg = emit(aggop->agg(), bb);
+        auto agg_type = convert(aggop->agg()->type())->type_id;
+
+        bool mem = false;
+        if (auto tt = aggop->agg()->type()->isa<TupleType>(); tt && tt->op(0)->isa<MemType>()) mem = true;
+
+        auto copy_to_alloca = [&] (SpvId target_type) {
+            world().wdef(def, "slow: alloca and loads/stores needed for aggregate '{}'", def);
+            auto agg_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, agg_type);
+
+            auto variable = bb->fn_builder.variable(agg_ptr_type, spv::StorageClassFunction);
+            bb->store(spv_agg, variable);
+
+            auto cell_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, target_type);
+            auto cell = bb->access_chain(cell_ptr_type, variable, { emit(aggop->index(), bb)} );
+            return std::make_pair(variable, cell);
+        };
+
+        if (auto extract = aggop->isa<Extract>()) {
+            if (is_mem(extract)) return spv_none;
+
+            auto target_type = convert(extract->type())->type_id;
+            auto constant_index = aggop->index()->isa<PrimLit>();
+
+            // We have a fast-path: if the index is constant, we can simply use OpCompositeExtract
+            if (aggop->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
+                assert(aggop->agg()->type()->isa<DefiniteArrayType>());
+                assert(!is_mem(extract));
+                return bb->load(target_type, copy_to_alloca(target_type).second);
+            }
+
+            if (extract->agg()->type()->isa<VectorType>())
+                return bb->vector_extract_dynamic(target_type, spv_agg, emit(extract->index(), bb));
+
+            // index *must* be constant for the remaining possible cases
+            assert(constant_index != nullptr);
+            uint32_t index = constant_index->value().get_u32();
+
+            unsigned offset = 0;
+            if (mem) {
+                if (aggop->agg()->type()->num_ops() == 2) return spv_agg;
+                offset = 1;
+            }
+
+            return bb->extract(target_type, spv_agg, { index - offset });
+        } else if (auto insert = def->isa<Insert>()) {
+            auto value = emit(insert->value(), bb);
+            auto constant_index = aggop->index()->isa<PrimLit>();
+
+            // TODO deal with mem - but I think for now this case shouldn't happen
+
+            if (insert->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
+                assert(aggop->agg()->type()->isa<DefiniteArrayType>());
+                auto [variable, cell] = copy_to_alloca(agg_type);
+                bb->store(value, cell);
+                return bb->load(agg_type, variable);
+            }
+
+            if (insert->agg()->type()->isa<VectorType>())
+                return bb->vector_insert_dynamic(agg_type, spv_agg, value, emit(insert->index(), bb));
+
+            // index *must* be constant for the remaining possible cases
+            assert(constant_index != nullptr);
+            uint32_t index = constant_index->value().get_u32();
+
+            return bb->insert(agg_type, value, spv_agg, { index });
+        } else THORIN_UNREACHABLE;
+    } else if (auto conv = def->isa<ConvOp>()) {
+        auto src_type = conv->from()->type();
+        auto dst_type = conv->type();
+
+        auto conv_src_type = convert(src_type);
+        auto conv_dst_type = convert(dst_type);
+
+        if (auto bitcast = def->isa<Bitcast>()) {
+            if (conv_src_type->datatype->serialized_size() != conv_dst_type->datatype->serialized_size())
+                world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type->datatype->serialized_size(), conv_dst_type->datatype->serialized_size());
+
+            return bb->convert(spv::OpBitcast, convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
+        } else if (auto cast = def->isa<Cast>()) {
+            // NB: all ops used here are scalar/vector agnostic
+            auto src_prim = src_type->isa<PrimType>();
+            auto dst_prim = dst_type->isa<PrimType>();
+            if (!src_prim || !dst_prim || src_prim->length() != dst_prim->length())
+                world().ELOG("Illegal cast: % to %, casts are only supported between primitives with identical vector length", src_type->to_string(), dst_type->to_string());
+
+            auto length = src_prim->length();
+
+            auto src_kind = classify_primtype(src_prim);
+            auto dst_kind = classify_primtype(dst_prim);
+            size_t src_bitwidth = conv_src_type->datatype->serialized_size();
+            size_t dst_bitwidth = conv_src_type->datatype->serialized_size();
+
+            SpvId data = emit(cast->from(), bb);
+
+            // If floating point is involved (src or dst), OpConvert*ToF and OpConvertFTo* can take care of the bit width transformation so no need for any chopping/expanding
+            if (src_kind == PrimTypeKind::Float || dst_kind == PrimTypeKind::Float) {
+                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
+                switch (src_kind) {
+                    case PrimTypeKind::Signed:    data = bb->convert(spv::OpConvertSToF, target_type, data); break;
+                    case PrimTypeKind::Unsigned:  data = bb->convert(spv::OpConvertUToF, target_type, data); break;
+                    case PrimTypeKind::Float:
+                        switch (dst_kind) {
+                            case PrimTypeKind::Signed:   data = bb->convert(spv::OpConvertFToS, target_type, data); break;
+                            case PrimTypeKind::Unsigned: data = bb->convert(spv::OpConvertFToU, target_type, data); break;
+                            default: THORIN_UNREACHABLE;
+                        }
+                        break;
+                }
+            } else {
+                // we expand first and shrink last to minimize precision losses, with bitcast in the middle
+                bool needs_chopping = src_bitwidth > dst_bitwidth;
+                bool needs_expanding = src_bitwidth < dst_bitwidth;
+
+                if (needs_expanding) {
+                    auto target_type = convert(get_primtype(world(), src_kind, src_bitwidth, length))->type_id;
+                    switch (src_kind) {
+                        case PrimTypeKind::Signed:
+                            data = bb->convert(spv::OpSConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Unsigned:
+                            data = bb->convert(spv::OpUConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Float:
+                            data = bb->convert(spv::OpFConvert, target_type, data);
+                            break;
+                    }
+                }
+
+                auto expanded_bitwidth = needs_expanding ? dst_bitwidth : src_bitwidth;
+                auto bitcast_target_type = convert(get_primtype(world(), dst_kind, expanded_bitwidth, length))->type_id;
+                data = bb->convert(spv::OpBitcast, bitcast_target_type, data);
+
+                if (needs_chopping) {
+                    auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
+                    switch (dst_kind) {
+                        case PrimTypeKind::Signed:
+                            data = bb->convert(spv::OpSConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Unsigned:
+                            data = bb->convert(spv::OpUConvert, target_type, data);
+                            break;
+                        case PrimTypeKind::Float:
+                            data = bb->convert(spv::OpFConvert, target_type, data);
+                            break;
+                    }
+                }
+            }
+        } else THORIN_UNREACHABLE;
+    } else if (def->isa<Bottom>()) {
+        return bb->undef(convert(def->type())->type_id);
+    }
+    assertf(false, "Incomplete emit(def) definition");
+}
+
+std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const Continuation* builtin, BasicBlockBuilder* bb) {
+    std::vector<SpvId> productions;
+    auto uvec3_t = convert(world().type_pu32(3));
+    auto u32_t = convert(world().type_pu32());
+    auto i32_t = convert(world().type_ps32());
+    if (builtin->name() == "spirv.nonsemantic.printf") {
+        std::vector<SpvId> args;
+        auto string = source_cont->arg(1);
+        if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
+            auto arr = string->as<DefiniteArray>();
+            std::vector<char> the_string;
+            for (size_t i = 0; i < arr_type->dim(); i++)
+                the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
+            the_string.push_back('\0');
+            args.push_back(builder_->debug_string(the_string.data()));
+        } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
+
+        for (size_t i = 2; i < source_cont->num_args() - 1; i++) {
+            args.push_back(emit(source_cont->arg(i), bb));
+        }
+
+        bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
+    } else if (builtin->name() == "get_work_dim") {
+        THORIN_UNREACHABLE;
+    } else if (builtin->name() == "get_global_id") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->global_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_local_size") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_size);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_local_id") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_num_groups") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->num_workgroups);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else if (builtin->name() == "get_group_id") {
+        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+    } else {
+        world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
+    }
+    return productions;
+}
+
+}
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
new file mode 100644
index 000000000..75ed0bb73
--- /dev/null
+++ b/src/thorin/be/spirv/spirv.h
@@ -0,0 +1,186 @@
+#ifndef THORIN_SPIRV_H
+#define THORIN_SPIRV_H
+
+#include "thorin/be/spirv/spirv_builder.hpp"
+#include "thorin/be/codegen.h"
+
+namespace thorin::spirv {
+
+using SpvId = builder::SpvId;
+
+class CodeGen;
+struct Datatype;
+struct PtrDatatype;
+
+struct FileBuilder;
+struct FnBuilder;
+
+struct ConvertedType {
+    ConvertedType(CodeGen* cg) : code_gen(cg) {}
+    ConvertedType(const ConvertedType&) = delete;
+
+    spirv::CodeGen* code_gen;
+    const thorin::Type* src_type;
+    SpvId type_id { 0 };
+    std::unique_ptr<Datatype> datatype;
+
+    bool is_known_size() { return datatype != nullptr; }
+};
+
+struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
+    explicit BasicBlockBuilder(FnBuilder& fn_builder);
+    BasicBlockBuilder(const BasicBlockBuilder&) = delete;
+
+    FnBuilder& fn_builder;
+    FileBuilder& file_builder;
+    std::unordered_map<const Param*, Phi> phis_map;
+    DefMap<SpvId> args;
+};
+
+struct FnBuilder : public builder::SpvFnBuilder {
+    explicit FnBuilder(CodeGen* cg, FileBuilder& file_builder);
+    FnBuilder(const FnBuilder&) = delete;
+
+    CodeGen* cg;
+    FileBuilder& file_builder;
+
+    const Scope* scope = nullptr;
+    std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
+    std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
+    ContinuationMap<SpvId> labels;
+    DefMap<SpvId> params;
+};
+
+struct Builtins {
+    SpvId workgroup_size;
+    SpvId num_workgroups;
+    SpvId workgroup_id;
+    SpvId local_id;
+    SpvId global_id;
+    SpvId local_invocation_index;
+
+    explicit Builtins(FileBuilder&);
+};
+
+struct ImportedInstructions {
+    SpvId shader_printf;
+
+    explicit ImportedInstructions(FileBuilder&);
+};
+
+struct FileBuilder : public builder::SpvFileBuilder {
+    explicit FileBuilder(CodeGen* cg);
+    FileBuilder(const FileBuilder&) = delete;
+
+    CodeGen* cg;
+
+    std::unique_ptr<Builtins> builtins;
+    std::unique_ptr<ImportedInstructions> imported_instrs;
+
+    SpvId u32_t();
+    SpvId u32_constant(uint32_t);
+
+private:
+    SpvId u32_t_ { 0 };
+    /*SpvId i32_t;
+    SpvId u32_t;
+    SpvId i64_t;
+    SpvId u64_t;
+    SpvId i32_constant(int32_t);
+    SpvId i64_constant(int64_t);
+    SpvId u64_constant(uint64_t);*/
+};
+
+class CodeGen : public thorin::CodeGen {
+public:
+    CodeGen(World&, Cont2Config&, bool debug);
+
+    void emit_stream(std::ostream& stream) override;
+    const char* file_ext() const override { return ".spv"; }
+
+    ConvertedType* convert(const Type*);
+protected:
+    void structure_loops();
+    void structure_flow();
+
+    void emit(const Scope& scope);
+    void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
+    SpvId emit(const Def* def, BasicBlockBuilder* bb);
+    std::vector<SpvId> emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
+
+    SpvId get_codom_type(const Continuation* fn);
+
+    std::unique_ptr<FileBuilder> builder_;
+    Continuation* entry_ = nullptr;
+    FnBuilder* current_fn_ = nullptr;
+    DefMap<std::unique_ptr<ConvertedType>> types_;
+    DefMap<SpvId> defs_;
+    const Cont2Config& kernel_config_;
+
+    friend PtrDatatype;
+};
+
+/// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
+/// potentially different mappings, depending on the capabilities of the target platform. The serdes code deals with pointers
+/// in arrays of unsigned 32 bit words, and is there to get around the limitation of not being able to bitcast pointers in the
+/// logical addressing mode.
+struct Datatype {
+public:
+    ConvertedType* type;
+    Datatype(ConvertedType* type) : type(type) {}
+
+    // Datatypes are serialized using a base element, for now it is hardcoded to use 32-bit scalar unsigned integers
+    static constexpr size_t base_element_bitwidth = 32;
+    static constexpr size_t base_element_bytes = base_element_bitwidth / 8;
+
+    virtual size_t serialized_size() = 0;
+    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) = 0;
+    virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) = 0;
+};
+
+/// For scalar datatypes
+struct ScalarDatatype : public Datatype {
+    int type_tag;
+    size_t size_in_bytes;
+    size_t alignment;
+    ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
+
+    size_t serialized_size() override { return (size_in_bytes + 3) / 4; };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
+};
+
+struct PtrDatatype : public Datatype {
+    static constexpr size_t bitwidth = 64;
+    PtrDatatype(ConvertedType* type) : Datatype(type) {}
+
+    size_t serialized_size() override { return bitwidth / 32; };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
+};
+
+struct DefiniteArrayDatatype : public Datatype {
+    ConvertedType* element_type;
+    size_t length;
+
+    DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length);
+
+    size_t serialized_size() override { return element_type->datatype->serialized_size(); };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
+};
+
+struct ProductDatatype : public Datatype {
+    std::vector<ConvertedType*> elements_types;
+    size_t total_size = 0;
+
+    ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types);
+
+    size_t serialized_size() override { return total_size; };
+    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
+    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
+};
+
+}
+
+#endif //THORIN_SPIRV_H
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
new file mode 100644
index 000000000..238343e79
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -0,0 +1,669 @@
+#include <spirv/unified1/spirv.hpp>
+
+#include <string>
+#include <cassert>
+#include <vector>
+#include <unordered_map>
+#include <ostream>
+#include <memory>
+
+namespace thorin::spirv::builder {
+
+struct SpvId { uint32_t id; };
+
+struct SpvSectionBuilder;
+struct SpvBasicBlockBuilder;
+struct SpvFnBuilder;
+struct SpvFileBuilder;
+
+inline int div_roundup(int a, int b) {
+    if (a % b == 0)
+        return a / b;
+    else
+        return (a / b) + 1;
+}
+
+struct SpvSectionBuilder {
+    std::vector<uint32_t> data_;
+
+private:
+    void output_word(uint32_t word) {
+        data_.push_back(word);
+    }
+public:
+    void op(spv::Op op, int ops_size) {
+        uint32_t lower = op & 0xFFFFu;
+        uint32_t upper = (ops_size << 16) & 0xFFFF0000u;
+        output_word(lower | upper);
+    }
+
+    void ref_id(SpvId id) {
+        assert(id.id != 0);
+        output_word(id.id);
+    }
+
+    void literal_name(std::string_view str) {
+        int i = 0;
+        uint32_t cword = 0;
+        for (char c : str) {
+            cword = cword | (c & 0xFF) << (i * 8);
+            i++;
+            if (i == 4) {
+                output_word(cword);
+                cword = 0;
+                i = 0;
+            }
+        }
+        output_word(cword);
+    }
+
+    void literal_int(uint32_t i) {
+        output_word(i);
+    }
+};
+
+struct SpvBasicBlockBuilder : public SpvSectionBuilder {
+    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
+            : file_builder(file_builder)
+    {}
+
+    SpvFileBuilder& file_builder;
+
+    struct Phi {
+        SpvId type;
+        SpvId value;
+        std::vector<std::pair<SpvId, SpvId>> preds;
+    };
+    std::vector<Phi*> phis;
+    SpvId label;
+
+    SpvId undef(SpvId type) {
+        op(spv::Op::OpUndef, 3);
+        ref_id(type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        return id;
+    }
+
+    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
+        op(spv::Op::OpCompositeConstruct, 3 + elements.size());
+        ref_id(aggregate_t);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        for (auto e : elements)
+            ref_id(e);
+        return id;
+    }
+
+    SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
+        op(spv::Op::OpCompositeExtract, 4 + indices.size());
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(composite);
+        for (auto i : indices)
+            literal_int(i);
+        return id;
+    }
+
+    SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
+        op(spv::Op::OpCompositeInsert, 5 + indices.size());
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(object);
+        ref_id(composite);
+        for (auto i : indices)
+            literal_int(i);
+        return id;
+    }
+
+    SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
+        op(spv::Op::OpVectorExtractDynamic, 5);
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(vector);
+        ref_id(index);
+        return id;
+    }
+
+    SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
+        op(spv::Op::OpVectorInsertDynamic, 6);
+        ref_id(target_type);
+        auto id = generate_fresh_id();
+        ref_id(id);
+        ref_id(vector);
+        ref_id(component);
+        ref_id(index);
+        return id;
+    }
+
+    // Used for almost all conversion operations
+    SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
+        op(op_, 4);
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(value);
+        return id;
+    }
+
+    SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
+        op(spv::Op::OpAccessChain, 4 + indexes.size());
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(element);
+        for (auto index : indexes)
+            ref_id(index);
+        return id;
+    }
+
+    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
+        op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(base);
+        ref_id(element);
+        for (auto index : indexes)
+            ref_id(index);
+        return id;
+    }
+
+    SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
+        op(spv::Op::OpLoad, 4 + operands.size());
+        auto id = generate_fresh_id();
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(pointer);
+        for (auto op : operands)
+            literal_int(op);
+        return id;
+    }
+
+    void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
+        op(spv::Op::OpStore, 3 + operands.size());
+        ref_id(pointer);
+        ref_id(value);
+        for (auto op : operands)
+            literal_int(op);
+    }
+
+    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
+        op(op_, 5);
+        auto id = generate_fresh_id();
+        ref_id(result_type);
+        ref_id(id);
+        ref_id(lhs);
+        ref_id(rhs);
+        return id;
+    }
+
+    void branch(SpvId target) {
+        op(spv::Op::OpBranch, 2);
+        ref_id(target);
+    }
+
+    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
+        op(spv::Op::OpBranchConditional, 4);
+        ref_id(condition);
+        ref_id(true_target);
+        ref_id(false_target);
+    }
+
+    void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
+        op(spv::Op::OpSelectionMerge, 3);
+        ref_id(merge_bb);
+        literal_int(selection_control);
+    }
+
+    void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
+        op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
+        ref_id(merge_bb);
+        ref_id(continue_bb);
+        literal_int(loop_control);
+
+        for (auto e : loop_control_ops)
+            literal_int(e);
+    }
+
+    SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
+        op(spv::Op::OpFunctionCall, 4 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(callee);
+
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
+    }
+
+    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
+        op(spv::Op::OpExtInst, 5 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(set);
+        literal_int(instruction);
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
+    }
+
+    void return_void() {
+        op(spv::Op::OpReturn, 1);
+    }
+
+    void return_value(SpvId value) {
+        op(spv::Op::OpReturnValue, 2);
+        ref_id(value);
+    }
+
+    void unreachable() {
+        op(spv::Op::OpUnreachable, 1);
+    }
+
+private:
+    SpvId generate_fresh_id();
+};
+
+struct SpvFnBuilder {
+    explicit SpvFnBuilder(SpvFileBuilder* file_builder)
+            : file_builder(file_builder)
+    {
+        function_id = generate_fresh_id();
+    }
+
+    SpvFileBuilder* file_builder;
+    SpvId function_id;
+
+    SpvId fn_type;
+    SpvId fn_ret_type;
+    std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
+
+    // Contains OpFunctionParams
+    SpvSectionBuilder header;
+
+    SpvSectionBuilder variables;
+
+    SpvId parameter(SpvId param_type) {
+        header.op(spv::Op::OpFunctionParameter, 3);
+        auto id = generate_fresh_id();
+        header.ref_id(param_type);
+        header.ref_id(id);
+        return id;
+    }
+
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        variables.op(spv::Op::OpVariable, 4);
+        variables.ref_id(type);
+        auto id = generate_fresh_id();
+        variables.ref_id(id);
+        variables.literal_int(storage_class);
+        return id;
+    }
+
+private:
+    SpvId generate_fresh_id();
+};
+
+struct SpvFileBuilder {
+
+    enum UniqueDeclTag {
+        NONE,
+        FN_TYPE,
+        PTR_TYPE,
+        DEF_ARR_TYPE,
+        CONSTANT,
+        CONSTANT_COMPOSITE,
+    };
+
+    /// Prevents duplicate declarations
+    struct UniqueDeclKey {
+        UniqueDeclTag tag;
+        std::vector<uint32_t> members;
+
+        bool operator==(const UniqueDeclKey &b) const {
+            return tag == b.tag && members == b.members;
+        }
+    };
+
+    struct UniqueDeclKeyHasher {
+        size_t operator() (const UniqueDeclKey& key) const {
+            size_t acc = 0;
+            for (auto id : key.members)
+                acc ^= std::hash<uint32_t>{}(id);
+            return std::hash<size_t>{}(key.tag) ^ acc;
+        }
+    };
+
+    SpvFileBuilder()
+            : void_type(declare_void_type())
+    {}
+    SpvFileBuilder(const SpvFileBuilder&) = delete;
+
+    SpvId generate_fresh_id() { return { bound++ }; }
+
+    void name(SpvId id, std::string_view str) {
+        assert(id.id < bound);
+        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
+        debug_names.ref_id(id);
+        debug_names.literal_name(str);
+    }
+
+    SpvId declare_bool_type() {
+        types_constants.op(spv::Op::OpTypeBool, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    SpvId declare_int_type(int width, bool signed_) {
+        types_constants.op(spv::Op::OpTypeInt, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        types_constants.literal_int(signed_ ? 1 : 0);
+        return id;
+    }
+
+    SpvId declare_float_type(int width) {
+        types_constants.op(spv::Op::OpTypeFloat, 3);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        return id;
+    }
+
+    SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
+        auto key = UniqueDeclKey { PTR_TYPE, { element_type.id, (uint32_t) storage_class } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpTypePointer, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
+        types_constants.ref_id(element_type);
+        unique_decls[key] = id;
+        return id;
+    }
+
+    SpvId declare_array_type(SpvId element_type, SpvId dim) {
+        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type.id, dim.id } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpTypeArray, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(element_type);
+        types_constants.ref_id(dim);
+        unique_decls[key] = id;
+        return id;
+    }
+
+    SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
+        auto key = UniqueDeclKey { FN_TYPE, {} };
+        for (auto d : dom) key.members.push_back(d.id);
+        key.members.push_back(codom.id);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+
+        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(codom);
+        for (auto arg : dom)
+            types_constants.ref_id(arg);
+        unique_decls[key] = id;
+        return id;
+    }
+
+    SpvId declare_struct_type(std::vector<SpvId> elements) {
+        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        for (auto arg : elements)
+            types_constants.ref_id(arg);
+        return id;
+    }
+
+    SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
+        types_constants.op(spv::Op::OpTypeVector, 4);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.ref_id(component_type);
+        types_constants.literal_int(dim);
+        return id;
+    }
+
+    void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+        annotations.op(spv::Op::OpDecorate, 3 + extra.size());
+        annotations.ref_id(target);
+        annotations.literal_int(decoration);
+        for (auto e : extra)
+            annotations.literal_int(e);
+    }
+
+    void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+        annotations.op(spv::Op::OpMemberDecorate, 4 + extra.size());
+        annotations.ref_id(target);
+        annotations.literal_int(member);
+        annotations.literal_int(decoration);
+        for (auto e : extra)
+            annotations.literal_int(e);
+    }
+
+    SpvId debug_string(std::string string) {
+        debug_string_source.op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
+        auto id = generate_fresh_id();
+        debug_string_source.ref_id(id);
+        debug_string_source.literal_name(string);
+        return id;
+    }
+
+    SpvId bool_constant(SpvId type, bool value) {
+        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
+        auto key = UniqueDeclKey { CONSTANT, bit_pattern };
+        key.members.push_back(type.id);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto arg : bit_pattern)
+            types_constants.literal_int(arg);
+        unique_decls[key] = id;
+        return id;
+    }
+
+    SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
+        auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
+        key.members.push_back(type.id);
+        for (auto op : ops) key.members.push_back(op.id);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto op : ops)
+            types_constants.ref_id(op);
+        unique_decls[key] = id;
+        return id;
+    }
+
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        types_constants.op(spv::Op::OpVariable, 4);
+        types_constants.ref_id(type);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
+        return id;
+    }
+
+    SpvId define_function(SpvFnBuilder& fn_builder) {
+        fn_defs.op(spv::Op::OpFunction, 5);
+        fn_defs.ref_id(fn_builder.fn_ret_type);
+        fn_defs.ref_id(fn_builder.function_id);
+        fn_defs.data_.push_back(spv::FunctionControlMaskNone);
+        fn_defs.ref_id(fn_builder.fn_type);
+
+        // Includes stuff like OpFunctionParameters
+        for (auto w : fn_builder.header.data_)
+            fn_defs.data_.push_back(w);
+
+        bool first = true;
+        for (auto& bb : fn_builder.bbs_to_emit) {
+            fn_defs.op(spv::Op::OpLabel, 2);
+            fn_defs.ref_id(bb->label);
+
+            if (first) {
+                for (auto w : fn_builder.variables.data_)
+                    fn_defs.data_.push_back(w);
+                first = false;
+            }
+
+            for (auto& phi : bb->phis) {
+                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
+                fn_defs.ref_id(phi->type);
+                fn_defs.ref_id(phi->value);
+                assert(!phi->preds.empty());
+                for (auto& [pred_value, pred_label] : phi->preds) {
+                    fn_defs.ref_id(pred_value);
+                    fn_defs.ref_id(pred_label);
+                }
+            }
+
+            for (auto w : bb->data_)
+                fn_defs.data_.push_back(w);
+        }
+
+        fn_defs.op(spv::Op::OpFunctionEnd, 1);
+        return fn_builder.function_id;
+    }
+
+    void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
+        entry_points.op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
+        entry_points.literal_int(execution_model);
+        entry_points.ref_id(entry_point);
+        entry_points.literal_name(name);
+        for (auto i : interface)
+            entry_points.ref_id(i);
+    }
+
+    void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
+        entry_points.op(spv::Op::OpExecutionMode, 3 + payloads.size());
+        entry_points.ref_id(entry_point);
+        entry_points.literal_int(execution_mode);
+        for (auto d : payloads)
+            entry_points.literal_int(d);
+    }
+
+    void capability(spv::Capability cap) {
+        capabilities.op(spv::Op::OpCapability, 2);
+        capabilities.data_.push_back(cap);
+    }
+
+    void extension(std::string name) {
+        extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
+        extensions.literal_name(name);
+    }
+
+    SpvId extended_import(std::string name) {
+        ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
+        auto id = generate_fresh_id();
+        ext_inst_import.ref_id(id);
+        ext_inst_import.literal_name(name);
+        return id;
+    }
+
+    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
+
+private:
+    std::ostream* output_ = nullptr;
+    uint32_t bound = 1;
+
+    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
+    SpvSectionBuilder capabilities;
+    SpvSectionBuilder extensions;
+    SpvSectionBuilder ext_inst_import;
+    SpvSectionBuilder entry_points;
+    SpvSectionBuilder execution_modes;
+    SpvSectionBuilder debug_string_source;
+    SpvSectionBuilder debug_names;
+    SpvSectionBuilder debug_module_processed;
+    SpvSectionBuilder annotations;
+    SpvSectionBuilder types_constants;
+    SpvSectionBuilder fn_decls;
+    SpvSectionBuilder fn_defs;
+
+    // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
+    std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
+
+    SpvId declare_void_type() {
+        types_constants.op(spv::Op::OpTypeVoid, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
+
+    void output_word_le(uint32_t word) {
+        output_->put((word >> 0) & 0xFFu);
+        output_->put((word >> 8) & 0xFFu);
+        output_->put((word >> 16) & 0xFFu);
+        output_->put((word >> 24) & 0xFFu);
+    }
+
+    void output_section(SpvSectionBuilder& section) {
+        for (auto& word : section.data_) {
+            output_word_le(word);
+        }
+    }
+public:
+    const SpvId void_type;
+
+    void finish(std::ostream& output) {
+        output_ = &output;
+        SpvSectionBuilder memory_model_section;
+        memory_model_section.op(spv::Op::OpMemoryModel, 3);
+        memory_model_section.data_.push_back(addressing_model);
+        memory_model_section.data_.push_back(memory_model);
+
+        output_word_le(spv::MagicNumber);
+        output_word_le(spv::Version); // TODO: target a specific spirv version
+        output_word_le(uint32_t(0)); // TODO get a magic number ?
+        output_word_le(bound);
+        output_word_le(uint32_t(0)); // instruction schema padding
+
+        output_section(capabilities);
+        output_section(extensions);
+        output_section(ext_inst_import);
+        output_section(memory_model_section);
+        output_section(entry_points);
+        output_section(execution_modes);
+        output_section(debug_string_source);
+        output_section(debug_names);
+        output_section(debug_module_processed);
+        output_section(annotations);
+        output_section(types_constants);
+        output_section(fn_decls);
+        output_section(fn_defs);
+    }
+};
+
+inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
+    return file_builder.generate_fresh_id();
+}
+
+inline SpvId SpvFnBuilder::generate_fresh_id() {
+    return file_builder->generate_fresh_id();
+}
+
+}

From cfeb4d5e8e235fea57bfceeba8163c7ce526965a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 15:01:26 +0200
Subject: [PATCH 226/342] spirv: API updates

---
 src/thorin/be/spirv/spirv.cpp | 90 ++++++++++++++++++-----------------
 src/thorin/be/spirv/spirv.h   |  6 +--
 2 files changed, 49 insertions(+), 47 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3b2d0e146..0f0ffb9c6 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -142,8 +142,8 @@ ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
     shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
 }
 
-CodeGen::CodeGen(thorin::World& world, Cont2Config& kernel_config, bool debug)
-        : thorin::CodeGen(world, debug), kernel_config_(kernel_config)
+CodeGen::CodeGen(Thorin& thorin, Cont2Config& kernel_config, bool debug)
+        : thorin::CodeGen(thorin, debug), kernel_config_(kernel_config)
 {}
 
 void CodeGen::emit_stream(std::ostream& out) {
@@ -156,7 +156,7 @@ void CodeGen::emit_stream(std::ostream& out) {
     structure_flow();
     // cleanup_world(world());
 
-    Scope::for_each(world(), [&](const Scope& scope) { emit(scope); });
+    ScopesForest(world()).for_each([&](const Scope& scope) { emit(scope); });
 
     auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
     auto push_constant_struct_type = builder_->declare_struct_type({ push_constant_arr_type });
@@ -169,7 +169,7 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_->name(push_constant_struct_ptr, "thorin_push_constant_data");
 
     auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
-    for (auto& cont : world().continuations()) {
+    for (auto& cont : world().copy_continuations()) {
         if (cont->is_exported()) {
             assert(defs_.contains(cont) && kernel_config_.contains(cont));
             auto config = kernel_config_.find(cont);
@@ -207,7 +207,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             for (size_t i = 0; i < cont->num_params(); i++) {
                 auto param = cont->param(i);
                 auto param_type = param->type();
-                if (param_type == world().unit() || param_type == world().mem_type() || param_type->isa<FnType>()) continue;
+                if (param_type == world().unit_type() || param_type == world().mem_type() || param_type->isa<FnType>()) continue;
                 assert(param_type->order() == 0);
                 auto converted = convert(param_type);
                 assert(converted->datatype != nullptr);
@@ -320,7 +320,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
     auto ret_cont_type = fn->ret_param()->type();
     std::vector<SpvId> types;
     for (auto& op : ret_cont_type->ops()) {
-        if (op->isa<MemType>() || is_type_unit(op))
+        if (op->isa<MemType>() || is_type_unit(op->type()))
             continue;
         assert(op->order() == 0);
         types.push_back(convert(op)->type_id);
@@ -346,10 +346,12 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
     };
 
-    if (continuation->callee() == entry_->ret_param()) {
+    auto& app = *continuation->body();
+
+    if (app.callee() == entry_->ret_param()) {
         std::vector<SpvId> values;
 
-        for (auto arg : continuation->args()) {
+        for (auto arg : app.args()) {
             assert(arg->order() == 0);
             auto val = emit(arg, bb);
             if (is_mem(arg) || is_unit(arg))
@@ -362,19 +364,19 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             case 1:  bb->return_value(values[0]); break;
             default: bb->return_value(bb->composite(current_fn_->fn_ret_type, values));
         }
-    } else if (auto callee = continuation->callee()->isa_continuation(); callee && callee->is_basicblock()) { // ordinary jump
+    } else if (auto dst_cont = app.callee()->isa_nom<Continuation>(); dst_cont && dst_cont->is_basicblock()) { // ordinary jump
         int index = -1;
-        for (auto& arg : continuation->args()) {
+        for (auto& arg : app.args()) {
             index++;
             auto val = emit(arg, bb);
             if (is_mem(arg) || is_unit(arg)) continue;
             bb->args[arg] = val;
-            auto* param = callee->param(index);
-            auto& phi = current_fn_->bbs_map[callee]->phis_map[param];
+            auto* param = dst_cont->param(index);
+            auto& phi = current_fn_->bbs_map[dst_cont]->phis_map[param];
             phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
         }
-        bb->branch(current_fn_->labels[callee]);
-    } else if (continuation->callee() == world().branch()) {
+        bb->branch(current_fn_->labels[dst_cont]);
+    } else if (app.callee() == world().branch()) {
         auto& domtree = current_fn_->scope->b_cfg().domtree();
         auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
         SpvId merge_bb;
@@ -389,24 +391,24 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
             merge_bb = current_fn_->labels[merge_cont];
         }
 
-        auto cond = emit(continuation->arg(0), bb);
-        bb->args.emplace(continuation->arg(0), cond);
-        auto tbb = current_fn_->labels[continuation->arg(1)->as_continuation()];
-        auto fbb = current_fn_->labels[continuation->arg(2)->as_continuation()];
+        auto cond = emit(app.arg(0), bb);
+        bb->args.emplace(app.arg(0), cond);
+        auto tbb = current_fn_->labels[app.arg(1)->isa_nom<Continuation>()];
+        auto fbb = current_fn_->labels[app.arg(2)->isa_nom<Continuation>()];
         bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
         bb->branch_conditional(cond, tbb, fbb);
-    } else if (continuation->callee()->isa<Continuation>() && continuation->callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
+    } else if (app.callee()->isa<Continuation>() && app.callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         /*auto val = emit(continuation->arg(0));
-        auto otherwise_bb = cont2bb(continuation->arg(1)->as_continuation());
+        auto otherwise_bb = cont2bb(continuation->arg(1)->isa_nom<Continuation>());
         auto match = irbuilder.CreateSwitch(val, otherwise_bb, continuation->num_args() - 2);
         for (size_t i = 2; i < continuation->num_args(); i++) {
             auto arg = continuation->arg(i)->as<Tuple>();
             auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
-            auto case_bb    = cont2bb(arg->op(1)->as_continuation());
+            auto case_bb    = cont2bb(arg->op(1)->isa_nom<Continuation>());
             match->addCase(case_const, case_bb);
         }*/
         THORIN_UNREACHABLE;
-    } else if (continuation->callee()->isa<Bottom>()) {
+    } else if (app.callee()->isa<Bottom>()) {
         bb->unreachable();
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
         auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
@@ -425,7 +427,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         // TODO handle dispatching to multiple targets
         assert(targets == 1);
-        auto dispatch_target = continuation->op(0)->as_continuation();
+        auto dispatch_target = continuation->op(0)->isa_nom<Continuation>();
         // Extract the relevant variant & expand the tuple if necessary
         auto arg = world().variant_extract(continuation->param(0), 0);
         auto extracted = emit(arg, dispatch_bb);
@@ -441,7 +443,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         dispatch_bb->branch(current_fn_->bbs_map[dispatch_target]->label);
 
     } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
-        auto loop_header = continuation->op(0)->as_continuation();
+        auto loop_header = continuation->op(0)->isa_nom<Continuation>();
         auto header_label = current_fn_->bbs_map[loop_header]->label;
 
         auto arg = continuation->param(0);
@@ -458,28 +460,28 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         // TODO handle dispatching to multiple targets
         assert(targets == 1);
-        auto callee = continuation->op(0)->as_continuation();
+        auto callee = continuation->op(0)->isa_nom<Continuation>();
         // TODO phis
         bb->branch(current_fn_->bbs_map[callee]->label);
-    } else if (auto builtin = continuation->callee()->isa_continuation(); builtin->is_imported()) {
+    } else if (auto builtin = app.callee()->isa_nom<Continuation>(); builtin->is_imported()) {
         // Ensure we emit previous memory operations
-        assert(is_mem(continuation->arg(0)));
-        emit(continuation->arg(0), bb);
+        assert(is_mem(app.arg(0)));
+        emit(app.arg(0), bb);
 
-        auto productions = emit_builtin(continuation, builtin, bb);
-        auto succ = continuation->args().back()->as_continuation();
+        auto productions = emit_builtin(app, builtin, bb);
+        auto succ = app.args().back()->isa_nom<Continuation>();
         jump_to_next_cont_with_args(succ, productions);
-    } else if (auto intrinsic = continuation->callee()->isa_continuation(); callee && callee->is_intrinsic()) {
+    } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
         THORIN_UNREACHABLE;
     } else { // function/closure call
         // put all first-order args into an array
         std::vector<SpvId> call_args;
         const Def* ret_arg = nullptr;
-        for (auto arg : continuation->args()) {
+        for (auto arg : app.args()) {
             if (arg->order() == 0) {
                 auto arg_type = arg->type();
                 auto arg_val = emit(arg, bb);
-                if (arg_type == world().unit() || arg_type == world().mem_type()) continue;
+                if (arg_type == world().unit_type() || arg_type == world().mem_type()) continue;
                 call_args.push_back(arg_val);
             } else {
                 assert(!ret_arg);
@@ -490,7 +492,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         auto ret_type = get_codom_type(continuation);
 
         SpvId call_result;
-        if (auto called_continuation = continuation->callee()->isa_continuation()) {
+        if (auto called_continuation = app.callee()->isa_nom<Continuation>()) {
             call_result = bb->call(ret_type, emit(called_continuation, bb), call_args);
         } else {
             // must be a closure
@@ -502,7 +504,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
 
         // must be call + continuation --- call + return has been removed by codegen_prepare
-        auto succ = ret_arg->as_continuation();
+        auto succ = ret_arg->isa_nom<Continuation>();
 
         size_t n = 0;
         const Param* last_param = nullptr;
@@ -922,14 +924,14 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const Continuation* builtin, BasicBlockBuilder* bb) {
+std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* builtin, BasicBlockBuilder* bb) {
     std::vector<SpvId> productions;
     auto uvec3_t = convert(world().type_pu32(3));
     auto u32_t = convert(world().type_pu32());
     auto i32_t = convert(world().type_ps32());
     if (builtin->name() == "spirv.nonsemantic.printf") {
         std::vector<SpvId> args;
-        auto string = source_cont->arg(1);
+        auto string = app.arg(1);
         if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
             auto arr = string->as<DefiniteArray>();
             std::vector<char> the_string;
@@ -939,8 +941,8 @@ std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const
             args.push_back(builder_->debug_string(the_string.data()));
         } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
 
-        for (size_t i = 2; i < source_cont->num_args() - 1; i++) {
-            args.push_back(emit(source_cont->arg(i), bb));
+        for (size_t i = 2; i < app.num_args() - 1; i++) {
+            args.push_back(emit(app.arg(i), bb));
         }
 
         bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
@@ -948,23 +950,23 @@ std::vector<SpvId> CodeGen::emit_builtin(const Continuation* source_cont, const
         THORIN_UNREACHABLE;
     } else if (builtin->name() == "get_global_id") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->global_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else if (builtin->name() == "get_local_size") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_size);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else if (builtin->name() == "get_local_id") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else if (builtin->name() == "get_num_groups") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->num_workgroups);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else if (builtin->name() == "get_group_id") {
         auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(source_cont->arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 75ed0bb73..d834f7fec 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -93,12 +93,12 @@ struct FileBuilder : public builder::SpvFileBuilder {
 
 class CodeGen : public thorin::CodeGen {
 public:
-    CodeGen(World&, Cont2Config&, bool debug);
+    CodeGen(Thorin& thorin, Cont2Config&, bool debug);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
 
-    ConvertedType* convert(const Type*);
+    ConvertedType* convert(const Def*);
 protected:
     void structure_loops();
     void structure_flow();
@@ -106,7 +106,7 @@ class CodeGen : public thorin::CodeGen {
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     SpvId emit(const Def* def, BasicBlockBuilder* bb);
-    std::vector<SpvId> emit_builtin(const Continuation*, const Continuation*, BasicBlockBuilder*);
+    std::vector<SpvId> emit_builtin(const App&, const Continuation*, BasicBlockBuilder*);
 
     SpvId get_codom_type(const Continuation* fn);
 

From ffae8bd81f0d261bd688336f37e2f36919e7d359 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 15:03:15 +0200
Subject: [PATCH 227/342] spirv: remove all mentions of structured control flow

---
 src/thorin/be/spirv/spirv.cpp | 72 -----------------------------------
 src/thorin/be/spirv/spirv.h   |  3 --
 2 files changed, 75 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 0f0ffb9c6..d533aabb4 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -152,10 +152,6 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_->builtins = std::make_unique<Builtins>(*builder_);
     builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
 
-    structure_loops();
-    structure_flow();
-    // cleanup_world(world());
-
     ScopesForest(world()).for_each([&](const Scope& scope) { emit(scope); });
 
     auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
@@ -377,25 +373,10 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
         bb->branch(current_fn_->labels[dst_cont]);
     } else if (app.callee() == world().branch()) {
-        auto& domtree = current_fn_->scope->b_cfg().domtree();
-        auto merge_cont = domtree.idom(current_fn_->scope->f_cfg().operator[](continuation))->continuation();
-        SpvId merge_bb;
-        if (merge_cont == current_fn_->scope->exit()) {
-            BasicBlockBuilder* unreachable_merge_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
-            current_fn_->bbs_to_emit.emplace_back(unreachable_merge_bb);
-            builder_->name(unreachable_merge_bb->label, "merge_unreachable" + continuation->name());
-            unreachable_merge_bb->unreachable();
-            merge_bb = unreachable_merge_bb->label;
-        } else {
-            // TODO create a dedicated merge bb if this one is the merge blocks for more than 1 selection construct
-            merge_bb = current_fn_->labels[merge_cont];
-        }
-
         auto cond = emit(app.arg(0), bb);
         bb->args.emplace(app.arg(0), cond);
         auto tbb = current_fn_->labels[app.arg(1)->isa_nom<Continuation>()];
         auto fbb = current_fn_->labels[app.arg(2)->isa_nom<Continuation>()];
-        bb->selection_merge(merge_bb,spv::SelectionControlMaskNone);
         bb->branch_conditional(cond, tbb, fbb);
     } else if (app.callee()->isa<Continuation>() && app.callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         /*auto val = emit(continuation->arg(0));
@@ -410,59 +391,6 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         THORIN_UNREACHABLE;
     } else if (app.callee()->isa<Bottom>()) {
         bb->unreachable();
-    } else if (continuation->intrinsic() == Intrinsic::SCFLoopHeader) {
-        auto merge_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.merge_target)]->label;
-        auto continue_label = current_fn_->bbs_map[const_cast<Continuation*>(continuation->attributes_.scf_metadata.loop_header.continue_target)]->label;
-        bb->loop_merge(merge_label, continue_label, spv::LoopControlMaskNone, {});
-
-        BasicBlockBuilder* dispatch_bb = current_fn_->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*current_fn_)).get();
-
-        auto header_bb_location = std::find(current_fn_->bbs_to_emit.begin(), current_fn_->bbs_to_emit.end(), bb);
-
-        current_fn_->bbs_to_emit.emplace(header_bb_location + 1, dispatch_bb);
-        builder_->name(dispatch_bb->label, "dispatch_" + continuation->name());
-        bb->branch(dispatch_bb->label);
-        int targets = continuation->num_ops();
-        assert(targets > 0);
-
-        // TODO handle dispatching to multiple targets
-        assert(targets == 1);
-        auto dispatch_target = continuation->op(0)->isa_nom<Continuation>();
-        // Extract the relevant variant & expand the tuple if necessary
-        auto arg = world().variant_extract(continuation->param(0), 0);
-        auto extracted = emit(arg, dispatch_bb);
-
-        if (dispatch_target->param(0)->type()->equal(arg->type())) {
-            auto* param = dispatch_target->param(0);
-            auto& phi = current_fn_->bbs_map[dispatch_target]->phis_map[param];
-            phi.preds.emplace_back(extracted, dispatch_bb->label);
-        } else {
-            assert(false && "TODO destructure argument");
-        }
-
-        dispatch_bb->branch(current_fn_->bbs_map[dispatch_target]->label);
-
-    } else if (continuation->intrinsic() == Intrinsic::SCFLoopContinue) {
-        auto loop_header = continuation->op(0)->isa_nom<Continuation>();
-        auto header_label = current_fn_->bbs_map[loop_header]->label;
-
-        auto arg = continuation->param(0);
-        bb->args[arg] = emit(arg, bb);
-        auto* param = loop_header->param(0);
-        auto& phi = current_fn_->bbs_map[loop_header]->phis_map[param];
-        phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
-
-        bb->branch(header_label);
-    } else if (continuation->intrinsic() == Intrinsic::SCFLoopMerge) {
-
-        int targets = continuation->num_ops();
-        assert(targets > 0);
-
-        // TODO handle dispatching to multiple targets
-        assert(targets == 1);
-        auto callee = continuation->op(0)->isa_nom<Continuation>();
-        // TODO phis
-        bb->branch(current_fn_->bbs_map[callee]->label);
     } else if (auto builtin = app.callee()->isa_nom<Continuation>(); builtin->is_imported()) {
         // Ensure we emit previous memory operations
         assert(is_mem(app.arg(0)));
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index d834f7fec..ae6ea7a05 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -100,9 +100,6 @@ class CodeGen : public thorin::CodeGen {
 
     ConvertedType* convert(const Def*);
 protected:
-    void structure_loops();
-    void structure_flow();
-
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
     SpvId emit(const Def* def, BasicBlockBuilder* bb);

From 6f0d6f6fbe6444f71340277a3f6d69bd0c066def Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 15:04:39 +0200
Subject: [PATCH 228/342] spirv: update branch() to support mem arguments

---
 src/thorin/be/spirv/spirv.cpp | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index d533aabb4..24939162c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -373,10 +373,13 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
         bb->branch(current_fn_->labels[dst_cont]);
     } else if (app.callee() == world().branch()) {
-        auto cond = emit(app.arg(0), bb);
-        bb->args.emplace(app.arg(0), cond);
-        auto tbb = current_fn_->labels[app.arg(1)->isa_nom<Continuation>()];
-        auto fbb = current_fn_->labels[app.arg(2)->isa_nom<Continuation>()];
+        auto mem = app.arg(0);
+        emit_unsafe(mem);
+
+        auto cond = emit(app.arg(1), bb);
+        bb->args.emplace(app.arg(2), cond);
+        auto tbb = current_fn_->labels[app.arg(2)->isa_nom<Continuation>()];
+        auto fbb = current_fn_->labels[app.arg(3)->isa_nom<Continuation>()];
         bb->branch_conditional(cond, tbb, fbb);
     } else if (app.callee()->isa<Continuation>() && app.callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         /*auto val = emit(continuation->arg(0));

From 9a4644aac49f256384091aa6bfa95d54000fe51a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 16:18:45 +0200
Subject: [PATCH 229/342] re-introduce select parts of spirv type codegen

---
 src/thorin/CMakeLists.txt             |   1 +
 src/thorin/be/spirv/spirv.cpp         | 144 +++++++++++++-------------
 src/thorin/be/spirv/spirv.h           | 103 ++++--------------
 src/thorin/be/spirv/spirv_builder.hpp |  43 ++++----
 src/thorin/type.h                     |   2 +
 5 files changed, 116 insertions(+), 177 deletions(-)

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index ec3bc9ef1..201d73665 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -127,6 +127,7 @@ endif()
 if(THORIN_ENABLE_SPIRV)
     list(APPEND THORIN_SOURCES
         be/spirv/spirv.cpp
+        be/spirv/spirv_types.cpp
         be/spirv/spirv.h
     )
 endif()
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 24939162c..651d44e71 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -94,8 +94,8 @@ FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
 }
 
 SpvId FileBuilder::u32_t() {
-    if (u32_t_.id == 0)
-        u32_t_ = cg->convert(cg->world().type_pu32())->type_id;
+    if (u32_t_ == 0)
+        u32_t_ = cg->convert(cg->world().type_pu32()).id;
     return u32_t_;
 }
 
@@ -107,9 +107,9 @@ Builtins::Builtins(FileBuilder& builder) {
     auto& world = builder.cg->world();
     auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
     auto spv_uint_t = builder.cg->convert(world.type_pu32());
-    auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t->type_id);
-    auto spv_uvec3_ptp = builder.declare_ptr_type(spv::StorageClassPrivate, spv_uvec3_t->type_id);
-    auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t->type_id);
+    auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t.id);
+    auto spv_uvec3_ptp = builder.declare_ptr_type(spv::StorageClassPrivate, spv_uvec3_t.id);
+    auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t.id);
 
     // Because we technically can have multiple entry points, we take the easy way out and make each entry point
     // write to a private variable the actual workgroup size for that specific kernel. Dirty, but simple.
@@ -142,8 +142,8 @@ ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
     shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
 }
 
-CodeGen::CodeGen(Thorin& thorin, Cont2Config& kernel_config, bool debug)
-        : thorin::CodeGen(thorin, debug), kernel_config_(kernel_config)
+CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, Cont2Config& kernel_config, bool debug)
+        : thorin::CodeGen(thorin, debug), target_info_(target_info), kernel_config_(kernel_config)
 {}
 
 void CodeGen::emit_stream(std::ostream& out) {
@@ -154,7 +154,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     ScopesForest(world()).for_each([&](const Scope& scope) { emit(scope); });
 
-    auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128))->type_id;
+    auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128)).id;
     auto push_constant_struct_type = builder_->declare_struct_type({ push_constant_arr_type });
     auto push_constant_struct_ptr_type = builder_->declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
     builder_->name(push_constant_struct_type, "ThorinPushConstant");
@@ -187,7 +187,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             };
 
             auto spv_uvec3_t = convert(world().type_pu32(3));
-            SpvId wg_size_constant = builder_->constant_composite(spv_uvec3_t->type_id, {
+            SpvId wg_size_constant = builder_->constant_composite(spv_uvec3_t, {
                     builder_->u32_constant(local_size[0]),
                     builder_->u32_constant(local_size[1]),
                     builder_->u32_constant(local_size[2]),
@@ -195,7 +195,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             bb->store(wg_size_constant, builder_->builtins->workgroup_size);
 
             // iterate on cont type and extract the arguments
-            auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push))->type_id;
+            auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push));
             auto zero = bb->file_builder.u32_constant(0);
             auto arr_ref = bb->access_chain(ptr_type, push_constant_struct_ptr, { zero });
             uint32_t offset = 0;
@@ -243,7 +243,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
 
     FnBuilder fn(this, *builder_.get());
     fn.scope = &scope;
-    fn.fn_type = convert(entry_->type())->type_id;
+    fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
     defs_.emplace(scope.entry(), fn.function_id);
 
@@ -273,7 +273,7 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // Nothing
                 } else if (param->order() == 0) {
                     auto param_t = convert(param->type());
-                    auto id = fn.parameter(param_t->type_id);
+                    auto id = fn.parameter(param_t.id);
                     fn.params[param] = id;
                     if (param->type()->isa<PtrType>()) {
                         builder_->decorate(id, spv::DecorationAliased);
@@ -288,8 +288,8 @@ void CodeGen::emit(const thorin::Scope& scope) {
                     // OpPhi requires the full list of predecessors (values, labels)
                     // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                     // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    auto type = convert(param->type())->type_id;
-                    assert(type.id != 0);
+                    auto type = convert(param->type()).id;
+                    assert(type != 0);
                     bb->phis_map[param] = { type, builder_->generate_fresh_id(), {} };
                 }
             }
@@ -313,16 +313,16 @@ void CodeGen::emit(const thorin::Scope& scope) {
 }
 
 SpvId CodeGen::get_codom_type(const Continuation* fn) {
-    auto ret_cont_type = fn->ret_param()->type();
+    auto ret_cont_type = fn->ret_param()->type()->as<FnType>();
     std::vector<SpvId> types;
-    for (auto& op : ret_cont_type->ops()) {
+    for (auto& op : ret_cont_type->types()) {
         if (op->isa<MemType>() || is_type_unit(op->type()))
             continue;
         assert(op->order() == 0);
-        types.push_back(convert(op)->type_id);
+        types.push_back(convert(op).id);
     }
     if (types.empty())
-        return builder_->void_type;
+        return convert(world().unit_type()).id;
     if (types.size() == 1)
         return types[0];
     return builder_->declare_struct_type(types);
@@ -459,7 +459,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
                 auto param = succ->param(i);
                 if (is_mem(param) || is_unit(param))
                     continue;
-                extracts[j] = bb->extract(convert(param->type())->type_id, call_result, { (uint32_t) j });
+                extracts[j] = bb->extract(convert(param->type()).id, call_result, { (uint32_t) j });
                 j++;
             }
 
@@ -483,8 +483,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs(), bb);
         SpvId rhs = emit(bin->rhs(), bb);
-        ConvertedType* result_types = convert(def->type());
-        SpvId result_type = result_types->type_id;
+        SpvId result_type = convert(def->type()).id;
 
         if (auto cmp = bin->isa<Cmp>()) {
             auto type = cmp->lhs()->type();
@@ -586,7 +585,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         }
     } else if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
-        auto type = convert(def->type())->type_id;
+        auto type = convert(def->type()).id;
         SpvId constant;
         switch (primlit->primtype_tag()) {
             case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
@@ -612,49 +611,51 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
     } else if (auto param = def->isa<Param>()) {
         if (is_mem(param)) return spv_none;
         if (auto param_id = current_fn_->params.lookup(param)) {
-            assert((*param_id).id != 0);
+            assert((*param_id) != 0);
             return *param_id;
         } else {
             auto val = (*current_fn_->bbs_map[param->continuation()]).phis_map[param].value;
-            assert(val.id != 0);
+            assert(val != 0);
             return val;
         }
     } else if (auto variant = def->isa<Variant>()) {
-        auto variant_type = def->type()->as<VariantType>();
+        assert(false && "TODO: rewrite");
+        /*auto variant_type = def->type()->as<VariantType>();
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
         auto tag = builder_->u32_constant(variant->index());
 
         if (variant_datatype->elements_types.size() > 1) {
-            auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+            auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function));
             auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
             auto converted_payload_type = convert(variant_type->op(variant->index()));
 
             converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0), emit(variant->value(), bb));
-            auto payload = bb->load(variant_datatype->elements_types[1]->type_id, payload_arr);
+            auto payload = bb->load(variant_datatype->elements_types[1], payload_arr);
 
             std::vector<SpvId> with_tag = {tag, payload};
-            return bb->composite(convert(variant->type())->type_id, with_tag);
+            return bb->composite(convert(variant->type()), with_tag);
         } else {
             // Zero-sized payload case
             std::vector<SpvId> with_tag = { tag };
-            return bb->composite(convert(variant->type())->type_id, with_tag);
-        }
+            return bb->composite(convert(variant->type()), with_tag);
+        }*/
     } else if (auto vextract = def->isa<VariantExtract>()) {
-        auto variant_type = vextract->value()->type()->as<VariantType>();
+        assert(false && "TODO: rewrite");
+        /*auto variant_type = vextract->value()->type()->as<VariantType>();
         auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
 
         auto target_type = convert(def->type());
 
         assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
-        auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function))->type_id;
+        auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function));
         auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
-        auto payload = bb->extract(variant_datatype->elements_types[1]->type_id, emit(vextract->value(), bb), {1});
+        auto payload = bb->extract(variant_datatype->elements_types[1], emit(vextract->value(), bb), {1});
         bb->store(payload, payload_arr);
 
-        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));
+        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));*/
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0), bb);
-        return bb->extract(convert(world().type_pu32())->type_id, value, { 0 });
+        return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
         std::vector<SpvId> elements;
         elements.resize(tuple->num_ops());
@@ -662,7 +663,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         for (auto& e : tuple->ops()) {
             elements[x++] = emit(e, bb);
         }
-        return bb->composite(convert(tuple->type())->type_id, elements);
+        return bb->composite(convert(tuple->type()).id, elements);
     } else if (auto structagg = def->isa<StructAgg>()) {
         std::vector<SpvId> elements;
         elements.resize(structagg->num_ops());
@@ -670,7 +671,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         for (auto& e : structagg->ops()) {
             elements[x++] = emit(e, bb);
         }
-        return bb->composite(convert(structagg->type())->type_id, elements);
+        return bb->composite(convert(structagg->type()).id, elements);
     } else if (auto access = def->isa<Access>()) {
         // emit dependent operations first
         emit(access->mem(), bb);
@@ -682,7 +683,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             operands.push_back( 4 ); // TODO: SPIR-V docs say to consult client API for valid values.
         }
         if (auto load = def->isa<Load>()) {
-            return bb->load(convert(load->out_val_type())->type_id, emit(load->ptr(), bb), operands);
+            return bb->load(convert(load->out_val_type()).id, emit(load->ptr(), bb), operands);
         } else if (auto store = def->isa<Store>()) {
             bb->store(emit(store->val(), bb), emit(store->ptr(), bb), operands);
             return spv_none;
@@ -696,12 +697,12 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
                 world().ELOG("LEA is only allowed in global & shared address spaces");
                 break;
         }
-        auto type = convert(lea->ptr_type());
+        auto type = convert(lea->ptr_type()).id;
         auto offset = emit(lea->index(), bb);
-        return bb->ptr_access_chain(type->type_id, emit(lea->ptr(), bb), offset, {});
+        return bb->ptr_access_chain(type, emit(lea->ptr(), bb), offset, {});
     } else if (auto aggop = def->isa<AggOp>()) {
         auto spv_agg = emit(aggop->agg(), bb);
-        auto agg_type = convert(aggop->agg()->type())->type_id;
+        auto agg_type = convert(aggop->agg()->type()).id;
 
         bool mem = false;
         if (auto tt = aggop->agg()->type()->isa<TupleType>(); tt && tt->op(0)->isa<MemType>()) mem = true;
@@ -721,7 +722,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         if (auto extract = aggop->isa<Extract>()) {
             if (is_mem(extract)) return spv_none;
 
-            auto target_type = convert(extract->type())->type_id;
+            auto target_type = convert(extract->type()).id;
             auto constant_index = aggop->index()->isa<PrimLit>();
 
             // We have a fast-path: if the index is constant, we can simply use OpCompositeExtract
@@ -775,10 +776,11 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         auto conv_dst_type = convert(dst_type);
 
         if (auto bitcast = def->isa<Bitcast>()) {
-            if (conv_src_type->datatype->serialized_size() != conv_dst_type->datatype->serialized_size())
-                world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type->datatype->serialized_size(), conv_dst_type->datatype->serialized_size());
+            assert(conv_src_type.layout && conv_dst_type.layout);
+            if (conv_src_type.layout->size != conv_dst_type.layout->size)
+                world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type.layout->size, conv_dst_type.layout->size);
 
-            return bb->convert(spv::OpBitcast, convert(bitcast->type())->type_id, emit(bitcast->from(), bb));
+            return bb->convert(spv::OpBitcast, convert(bitcast->type()).id, emit(bitcast->from(), bb));
         } else if (auto cast = def->isa<Cast>()) {
             // NB: all ops used here are scalar/vector agnostic
             auto src_prim = src_type->isa<PrimType>();
@@ -790,14 +792,14 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
 
             auto src_kind = classify_primtype(src_prim);
             auto dst_kind = classify_primtype(dst_prim);
-            size_t src_bitwidth = conv_src_type->datatype->serialized_size();
-            size_t dst_bitwidth = conv_src_type->datatype->serialized_size();
+            size_t src_bitwidth = conv_src_type.layout->size;
+            size_t dst_bitwidth = conv_src_type.layout->size;
 
             SpvId data = emit(cast->from(), bb);
 
             // If floating point is involved (src or dst), OpConvert*ToF and OpConvertFTo* can take care of the bit width transformation so no need for any chopping/expanding
             if (src_kind == PrimTypeKind::Float || dst_kind == PrimTypeKind::Float) {
-                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
+                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length)).id;
                 switch (src_kind) {
                     case PrimTypeKind::Signed:    data = bb->convert(spv::OpConvertSToF, target_type, data); break;
                     case PrimTypeKind::Unsigned:  data = bb->convert(spv::OpConvertUToF, target_type, data); break;
@@ -815,7 +817,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
                 bool needs_expanding = src_bitwidth < dst_bitwidth;
 
                 if (needs_expanding) {
-                    auto target_type = convert(get_primtype(world(), src_kind, src_bitwidth, length))->type_id;
+                    auto target_type = convert(get_primtype(world(), src_kind, src_bitwidth, length)).id;
                     switch (src_kind) {
                         case PrimTypeKind::Signed:
                             data = bb->convert(spv::OpSConvert, target_type, data);
@@ -830,11 +832,11 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
                 }
 
                 auto expanded_bitwidth = needs_expanding ? dst_bitwidth : src_bitwidth;
-                auto bitcast_target_type = convert(get_primtype(world(), dst_kind, expanded_bitwidth, length))->type_id;
+                auto bitcast_target_type = convert(get_primtype(world(), dst_kind, expanded_bitwidth, length)).id;
                 data = bb->convert(spv::OpBitcast, bitcast_target_type, data);
 
                 if (needs_chopping) {
-                    auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length))->type_id;
+                    auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length)).id;
                     switch (dst_kind) {
                         case PrimTypeKind::Signed:
                             data = bb->convert(spv::OpSConvert, target_type, data);
@@ -850,16 +852,16 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             }
         } else THORIN_UNREACHABLE;
     } else if (def->isa<Bottom>()) {
-        return bb->undef(convert(def->type())->type_id);
+        return bb->undef(convert(def->type()).id);
     }
     assertf(false, "Incomplete emit(def) definition");
 }
 
 std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* builtin, BasicBlockBuilder* bb) {
     std::vector<SpvId> productions;
-    auto uvec3_t = convert(world().type_pu32(3));
-    auto u32_t = convert(world().type_pu32());
-    auto i32_t = convert(world().type_ps32());
+    auto uvec3_t = convert(world().type_pu32(3)).id;
+    auto u32_t = convert(world().type_pu32()).id;
+    auto i32_t = convert(world().type_ps32()).id;
     if (builtin->name() == "spirv.nonsemantic.printf") {
         std::vector<SpvId> args;
         auto string = app.arg(1);
@@ -876,29 +878,29 @@ std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* bui
             args.push_back(emit(app.arg(i), bb));
         }
 
-        bb->ext_instruction(bb->file_builder.void_type, builder_->imported_instrs->shader_printf, 1, args);
+        bb->ext_instruction(convert(world().unit_type()).id, builder_->imported_instrs->shader_printf, 1, args);
     } else if (builtin->name() == "get_work_dim") {
         THORIN_UNREACHABLE;
     } else if (builtin->name() == "get_global_id") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->global_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+        auto vector = bb->load(uvec3_t, builder_->builtins->global_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_local_size") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_size);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+        auto vector = bb->load(uvec3_t, builder_->builtins->workgroup_size);
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_local_id") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->local_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+        auto vector = bb->load(uvec3_t, builder_->builtins->local_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_num_groups") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->num_workgroups);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+        auto vector = bb->load(uvec3_t, builder_->builtins->num_workgroups);
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_group_id") {
-        auto vector = bb->load(uvec3_t->type_id, builder_->builtins->workgroup_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t->type_id, vector, emit(app.arg(1), bb));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t->type_id, extracted));
+        auto vector = bb->load(uvec3_t, builder_->builtins->workgroup_id);
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
     }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index ae6ea7a05..2239b677d 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -9,22 +9,28 @@ namespace thorin::spirv {
 using SpvId = builder::SpvId;
 
 class CodeGen;
-struct Datatype;
-struct PtrDatatype;
 
 struct FileBuilder;
 struct FnBuilder;
 
-struct ConvertedType {
-    ConvertedType(CodeGen* cg) : code_gen(cg) {}
-    ConvertedType(const ConvertedType&) = delete;
+struct SpvTargetInfo {
+    struct {
+        // Either '4' or '8'
+        size_t pointer_size;
+    } mem_layout;
 
-    spirv::CodeGen* code_gen;
-    const thorin::Type* src_type;
-    SpvId type_id { 0 };
-    std::unique_ptr<Datatype> datatype;
+    enum Dialect{
+        OpenCL,
+        Shady
+    };
+};
 
-    bool is_known_size() { return datatype != nullptr; }
+struct ConvertedType {
+    SpvId id;
+    struct Layout {
+        size_t size, alignment;
+    };
+    std::optional<Layout> layout;
 };
 
 struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
@@ -82,23 +88,16 @@ struct FileBuilder : public builder::SpvFileBuilder {
 
 private:
     SpvId u32_t_ { 0 };
-    /*SpvId i32_t;
-    SpvId u32_t;
-    SpvId i64_t;
-    SpvId u64_t;
-    SpvId i32_constant(int32_t);
-    SpvId i64_constant(int64_t);
-    SpvId u64_constant(uint64_t);*/
 };
 
 class CodeGen : public thorin::CodeGen {
 public:
-    CodeGen(Thorin& thorin, Cont2Config&, bool debug);
+    CodeGen(Thorin& thorin, SpvTargetInfo, Cont2Config&, bool debug);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
 
-    ConvertedType* convert(const Def*);
+    ConvertedType convert(const Type*);
 protected:
     void emit(const Scope& scope);
     void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
@@ -107,75 +106,13 @@ class CodeGen : public thorin::CodeGen {
 
     SpvId get_codom_type(const Continuation* fn);
 
+    SpvTargetInfo target_info_;
     std::unique_ptr<FileBuilder> builder_;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
-    DefMap<std::unique_ptr<ConvertedType>> types_;
+    DefMap<ConvertedType> types_;
     DefMap<SpvId> defs_;
     const Cont2Config& kernel_config_;
-
-    friend PtrDatatype;
-};
-
-/// Thorin data types are mapped to SPIR-V in non-trivial ways, this interface is used by the emission code to abstract over
-/// potentially different mappings, depending on the capabilities of the target platform. The serdes code deals with pointers
-/// in arrays of unsigned 32 bit words, and is there to get around the limitation of not being able to bitcast pointers in the
-/// logical addressing mode.
-struct Datatype {
-public:
-    ConvertedType* type;
-    Datatype(ConvertedType* type) : type(type) {}
-
-    // Datatypes are serialized using a base element, for now it is hardcoded to use 32-bit scalar unsigned integers
-    static constexpr size_t base_element_bitwidth = 32;
-    static constexpr size_t base_element_bytes = base_element_bitwidth / 8;
-
-    virtual size_t serialized_size() = 0;
-    virtual SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) = 0;
-    virtual void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) = 0;
-};
-
-/// For scalar datatypes
-struct ScalarDatatype : public Datatype {
-    int type_tag;
-    size_t size_in_bytes;
-    size_t alignment;
-    ScalarDatatype(ConvertedType* type, int type_tag, size_t size_in_bytes, size_t alignment_in_bytes);
-
-    size_t serialized_size() override { return (size_in_bytes + 3) / 4; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-struct PtrDatatype : public Datatype {
-    static constexpr size_t bitwidth = 64;
-    PtrDatatype(ConvertedType* type) : Datatype(type) {}
-
-    size_t serialized_size() override { return bitwidth / 32; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-struct DefiniteArrayDatatype : public Datatype {
-    ConvertedType* element_type;
-    size_t length;
-
-    DefiniteArrayDatatype(ConvertedType* type, ConvertedType* element_type, size_t length);
-
-    size_t serialized_size() override { return element_type->datatype->serialized_size(); };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
-};
-
-struct ProductDatatype : public Datatype {
-    std::vector<ConvertedType*> elements_types;
-    size_t total_size = 0;
-
-    ProductDatatype(ConvertedType* type, const std::vector<ConvertedType*>&& elements_types);
-
-    size_t serialized_size() override { return total_size; };
-    SpvId emit_deserialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset) override;
-    void emit_serialization(BasicBlockBuilder& bb, spv::StorageClass storage_class, SpvId array, SpvId base_offset, SpvId data) override;
 };
 
 }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 238343e79..6a600b7a6 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -9,7 +9,8 @@
 
 namespace thorin::spirv::builder {
 
-struct SpvId { uint32_t id; };
+//struct SpvId { uint32_t id; };
+using SpvId = uint32_t;
 
 struct SpvSectionBuilder;
 struct SpvBasicBlockBuilder;
@@ -38,8 +39,8 @@ struct SpvSectionBuilder {
     }
 
     void ref_id(SpvId id) {
-        assert(id.id != 0);
-        output_word(id.id);
+        assert(id != 0);
+        output_word(id);
     }
 
     void literal_name(std::string_view str) {
@@ -340,15 +341,13 @@ struct SpvFileBuilder {
         }
     };
 
-    SpvFileBuilder()
-            : void_type(declare_void_type())
-    {}
+    SpvFileBuilder() {}
     SpvFileBuilder(const SpvFileBuilder&) = delete;
 
     SpvId generate_fresh_id() { return { bound++ }; }
 
     void name(SpvId id, std::string_view str) {
-        assert(id.id < bound);
+        assert(id < bound);
         debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
         debug_names.ref_id(id);
         debug_names.literal_name(str);
@@ -379,7 +378,7 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
-        auto key = UniqueDeclKey { PTR_TYPE, { element_type.id, (uint32_t) storage_class } };
+        auto key = UniqueDeclKey { PTR_TYPE, { element_type, (uint32_t) storage_class } };
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpTypePointer, 4);
         auto id = generate_fresh_id();
@@ -391,7 +390,7 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_array_type(SpvId element_type, SpvId dim) {
-        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type.id, dim.id } };
+        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type, dim } };
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpTypeArray, 4);
         auto id = generate_fresh_id();
@@ -404,8 +403,8 @@ struct SpvFileBuilder {
 
     SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
         auto key = UniqueDeclKey { FN_TYPE, {} };
-        for (auto d : dom) key.members.push_back(d.id);
-        key.members.push_back(codom.id);
+        for (auto d : dom) key.members.push_back(d);
+        key.members.push_back(codom);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
 
         types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
@@ -471,7 +470,7 @@ struct SpvFileBuilder {
 
     SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
         auto key = UniqueDeclKey { CONSTANT, bit_pattern };
-        key.members.push_back(type.id);
+        key.members.push_back(type);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
         auto id = generate_fresh_id();
@@ -485,8 +484,8 @@ struct SpvFileBuilder {
 
     SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
         auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
-        key.members.push_back(type.id);
-        for (auto op : ops) key.members.push_back(op.id);
+        key.members.push_back(type);
+        for (auto op : ops) key.members.push_back(op);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
         auto id = generate_fresh_id();
@@ -507,6 +506,13 @@ struct SpvFileBuilder {
         return id;
     }
 
+    SpvId declare_void_type() {
+        types_constants.op(spv::Op::OpTypeVoid, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
+    }
+
     SpvId define_function(SpvFnBuilder& fn_builder) {
         fn_defs.op(spv::Op::OpFunction, 5);
         fn_defs.ref_id(fn_builder.fn_ret_type);
@@ -607,13 +613,6 @@ struct SpvFileBuilder {
     // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
     std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
 
-    SpvId declare_void_type() {
-        types_constants.op(spv::Op::OpTypeVoid, 2);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        return id;
-    }
-
     void output_word_le(uint32_t word) {
         output_->put((word >> 0) & 0xFFu);
         output_->put((word >> 8) & 0xFFu);
@@ -627,8 +626,6 @@ struct SpvFileBuilder {
         }
     }
 public:
-    const SpvId void_type;
-
     void finish(std::ostream& output) {
         output_ = &output;
         SpvSectionBuilder memory_model_section;
diff --git a/src/thorin/type.h b/src/thorin/type.h
index e73ad5b6f..bf7cb26df 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -233,6 +233,8 @@ enum class AddrSpace : uint32_t {
     Shared   = 3,
     Constant = 4,
     Private =  5, // Corresponds to the 'private' storage class in compute kernels/shaders, as in thread-private
+    Function = 6, // Corresponds to the 'function' storage class in SPIR-V
+    Push     = 7, // Corresponds to the 'push constant' storage class in SPIR-V
 };
 
 /// Pointer type.

From 2fbfc56475615fbcdee0faa36f8514ff92993260 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 16:25:58 +0200
Subject: [PATCH 230/342] spirv: just emit entry points naively

---
 src/thorin/be/spirv/spirv.cpp | 70 ++++-------------------------------
 1 file changed, 8 insertions(+), 62 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 651d44e71..7d1f38727 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -154,17 +154,12 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     ScopesForest(world()).for_each([&](const Scope& scope) { emit(scope); });
 
-    auto push_constant_arr_type = convert(world().definite_array_type(world().type_pu32(), 128)).id;
-    auto push_constant_struct_type = builder_->declare_struct_type({ push_constant_arr_type });
-    auto push_constant_struct_ptr_type = builder_->declare_ptr_type(spv::StorageClassPushConstant, push_constant_struct_type);
-    builder_->name(push_constant_struct_type, "ThorinPushConstant");
-    builder_->decorate(push_constant_struct_type, spv::DecorationBlock);
-    builder_->decorate_member(push_constant_struct_type, 0, spv::DecorationOffset, { 0 });
-    builder_->decorate(push_constant_arr_type, spv::DecorationArrayStride, { 4 });
-    auto push_constant_struct_ptr = builder_->variable(push_constant_struct_ptr_type, spv::StorageClassPushConstant);
-    builder_->name(push_constant_struct_ptr, "thorin_push_constant_data");
-
-    auto entry_pt_signature = builder_->declare_fn_type({}, builder_->void_type);
+    std::vector<SpvId> interface;
+    for (auto def : world().defs()) {
+        if (auto global = def->isa<Global>())
+            interface.push_back(emit(global));
+    }
+
     for (auto& cont : world().copy_continuations()) {
         if (cont->is_exported()) {
             assert(defs_.contains(cont) && kernel_config_.contains(cont));
@@ -172,13 +167,6 @@ void CodeGen::emit_stream(std::ostream& out) {
 
             SpvId callee = defs_[cont];
 
-            FnBuilder fn_builder(this, *builder_.get());
-            fn_builder.fn_type = entry_pt_signature;
-            fn_builder.fn_ret_type = builder_->void_type;
-
-            BasicBlockBuilder* bb = fn_builder.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn_builder)).get();
-            fn_builder.bbs_to_emit.push_back(bb);
-
             auto block = config->second->as<GPUKernelConfig>()->block_size();
             std::vector<uint32_t> local_size = {
                     (uint32_t) std::get<0>(block),
@@ -186,50 +174,8 @@ void CodeGen::emit_stream(std::ostream& out) {
                     (uint32_t) std::get<2>(block),
             };
 
-            auto spv_uvec3_t = convert(world().type_pu32(3));
-            SpvId wg_size_constant = builder_->constant_composite(spv_uvec3_t, {
-                    builder_->u32_constant(local_size[0]),
-                    builder_->u32_constant(local_size[1]),
-                    builder_->u32_constant(local_size[2]),
-            });
-            bb->store(wg_size_constant, builder_->builtins->workgroup_size);
-
-            // iterate on cont type and extract the arguments
-            auto ptr_type = convert(world().ptr_type(world().definite_array_type(world().type_pu32(), 128), 1, 4, AddrSpace::Push));
-            auto zero = bb->file_builder.u32_constant(0);
-            auto arr_ref = bb->access_chain(ptr_type, push_constant_struct_ptr, { zero });
-            uint32_t offset = 0;
-            std::vector<SpvId> args;
-            for (size_t i = 0; i < cont->num_params(); i++) {
-                auto param = cont->param(i);
-                auto param_type = param->type();
-                if (param_type == world().unit_type() || param_type == world().mem_type() || param_type->isa<FnType>()) continue;
-                assert(param_type->order() == 0);
-                auto converted = convert(param_type);
-                assert(converted->datatype != nullptr);
-                SpvId arg = converted->datatype->emit_deserialization(*bb, spv::StorageClassPushConstant, arr_ref, bb->file_builder.u32_constant(offset));
-                args.push_back(arg);
-                offset += converted->datatype->serialized_size();
-            }
-
-            bb->call(builder_->void_type, callee, args);
-            bb->return_void();
-
-            builder_->define_function(fn_builder);
-            builder_->name(fn_builder.function_id, "entry_point_" + cont->name());
-
-            std::vector<SpvId> interface = {
-                    push_constant_struct_ptr,
-                    builder_->builtins->workgroup_size,
-                    builder_->builtins->num_workgroups,
-                    builder_->builtins->workgroup_id,
-                    builder_->builtins->local_id,
-                    builder_->builtins->global_id,
-                    builder_->builtins->local_invocation_index,
-            };
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, fn_builder.function_id, "kernel_main", interface);
-
-            builder_->execution_mode(fn_builder.function_id, spv::ExecutionModeLocalSize, local_size);
+            builder_->declare_entry_point(spv::ExecutionModelGLCompute, callee, cont->name().c_str(), interface);
+            builder_->execution_mode(callee, spv::ExecutionModeLocalSize, local_size);
         }
     }
 

From 8d29a53f813e3196958eca55e389879759438c61 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 17:46:08 +0200
Subject: [PATCH 231/342] spirv: base off Emitter

---
 src/thorin/be/spirv/spirv.cpp | 217 +++++++++++++++++-----------------
 src/thorin/be/spirv/spirv.h   |  28 +++--
 2 files changed, 125 insertions(+), 120 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 7d1f38727..ee732ff4b 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -80,7 +80,7 @@ BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
     label = file_builder.generate_fresh_id();
 }
 
-FnBuilder::FnBuilder(CodeGen* cg, FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), cg(cg), file_builder(file_builder) {}
+FnBuilder::FnBuilder(FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), file_builder(file_builder) {}
 
 FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
     capability(spv::Capability::CapabilityShader);
@@ -152,7 +152,8 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_->builtins = std::make_unique<Builtins>(*builder_);
     builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
 
-    ScopesForest(world()).for_each([&](const Scope& scope) { emit(scope); });
+    ScopesForest forest(world());
+    forest.for_each([&](const Scope& scope) { emit_scope(scope, forest); });
 
     std::vector<SpvId> interface;
     for (auto def : world().defs()) {
@@ -183,79 +184,73 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_ = nullptr;
 }
 
-void CodeGen::emit(const thorin::Scope& scope) {
-    entry_ = scope.entry();
-    assert(entry_->is_returning());
+SpvId CodeGen::emit_fun_decl(thorin::Continuation* continuation) {
+    return get_fn_builder(continuation).function_id;
+}
+
+FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
+    if (auto found = fn_builders_.find(continuation); found != fn_builders_.end()) {
+        return *found->second;
+    }
 
-    FnBuilder fn(this, *builder_.get());
-    fn.scope = &scope;
+    auto& fn = *(fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
     fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
-    defs_.emplace(scope.entry(), fn.function_id);
-
-    current_fn_ = &fn;
-
-    auto conts = schedule(scope);
-
-    fn.bbs_to_emit.reserve(conts.size());
-    fn.bbs.reserve(conts.size());
-    auto& bbs = fn.bbs;
-
-    for (auto cont : conts) {
-        if (cont->intrinsic() == Intrinsic::EndScope) continue;
-
-        BasicBlockBuilder* bb = bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn)).get();
-        fn.bbs_to_emit.emplace_back(bb);
-        auto [i, b] = fn.bbs_map.emplace(cont, bb);
-        assert(b);
-
-        if (debug())
-            builder_->name(bb->label, cont->name().c_str());
-        fn.labels.emplace(cont, bb->label);
-
-        if (entry_ == cont) {
-            for (auto param : entry_->params()) {
-                if (is_mem(param) || is_unit(param)) {
-                    // Nothing
-                } else if (param->order() == 0) {
-                    auto param_t = convert(param->type());
-                    auto id = fn.parameter(param_t.id);
-                    fn.params[param] = id;
-                    if (param->type()->isa<PtrType>()) {
-                        builder_->decorate(id, spv::DecorationAliased);
-                    }
+    return fn;
+}
+
+FnBuilder& CodeGen::prepare(const thorin::Scope& scope) {
+    auto& fn = get_fn_builder(scope.entry());
+    builder_->name(fn.function_id, scope.entry()->name());
+    return fn;
+}
+
+void CodeGen::prepare(thorin::Continuation* cont, FnBuilder& fn) {
+    auto& bb = *fn.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn));
+    cont2bb_[cont] = &bb;
+    fn.bbs_to_emit.emplace_back(&bb);
+
+    builder_->name(bb.label, cont->name().c_str());
+
+    if (entry_ == cont) {
+        for (auto param : cont->params()) {
+            if (is_mem(param) || is_unit(param)) {
+                // Nothing
+            } else if (param->order() == 0) {
+                auto param_t = convert(param->type());
+                auto id = fn.parameter(param_t.id);
+                fn.params[param] = id;
+                if (param->type()->isa<PtrType>()) {
+                    builder_->decorate(id, spv::DecorationAliased);
                 }
             }
-        } else {
-            for (auto param : cont->params()) {
-                if (is_mem(param) || is_unit(param)) {
-                    // Nothing
-                } else {
-                    // OpPhi requires the full list of predecessors (values, labels)
-                    // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
-                    // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
-                    auto type = convert(param->type()).id;
-                    assert(type != 0);
-                    bb->phis_map[param] = { type, builder_->generate_fresh_id(), {} };
-                }
+        }
+    } else {
+        for (auto param : cont->params()) {
+            if (is_mem(param) || is_unit(param)) {
+                // Nothing
+            } else {
+                // OpPhi requires the full list of predecessors (values, labels)
+                // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
+                // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
+                auto type = convert(param->type()).id;
+                assert(type != 0);
+                bb.phis_map[param] = { type, builder_->generate_fresh_id(), {} };
             }
         }
     }
 
-    for (auto cont : conts) {
-        if (cont->intrinsic() == Intrinsic::EndScope) continue;
-        assert(cont == entry_ || cont->is_basicblock());
-        emit_epilogue(cont, fn.bbs_map[cont]);
-    }
+}
 
-    for(auto& bb : fn.bbs) {
-        for (auto& [param, phi] : bb->phis_map) {
-            bb->phis.emplace_back(&phi);
-        }
+void CodeGen::finalize(thorin::Continuation* cont) {
+    auto& bb = *cont2bb_[cont];
+    for (auto& [param, phi] : bb.phis_map) {
+        bb.phis.emplace_back(&phi);
     }
+}
 
-    builder_->define_function(fn);
-    builder_->name(fn.function_id, scope.entry()->name());
+void CodeGen::finalize(const thorin::Scope&) {
+    builder_->define_function(*current_fn_);
 }
 
 SpvId CodeGen::get_codom_type(const Continuation* fn) {
@@ -274,16 +269,18 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
     return builder_->declare_struct_type(types);
 }
 
-void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
+void CodeGen::emit_epilogue(Continuation* continuation) {
+    auto& bb = cont2bb_[continuation];
     // Handles the potential nuances of jumping to another continuation
     auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
-        bb->branch(current_fn_->labels[succ]);
+        assert(succ->is_basicblock());
+        bb->branch(emit(succ));
         for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
             auto param = succ->param(i);
             if (is_mem(param) || is_unit(param))
                 continue;
-            auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
-            phi.preds.emplace_back(args[j], current_fn_->labels[continuation]);
+            auto& phi = cont2bb_[succ]->phis_map[param];
+            phi.preds.emplace_back(args[j], emit(continuation));
             j++;
         }
     };
@@ -295,7 +292,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         for (auto arg : app.args()) {
             assert(arg->order() == 0);
-            auto val = emit(arg, bb);
+            auto val = emit(arg);
             if (is_mem(arg) || is_unit(arg))
                 continue;
             values.emplace_back(val);
@@ -310,22 +307,22 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         int index = -1;
         for (auto& arg : app.args()) {
             index++;
-            auto val = emit(arg, bb);
+            auto val = emit(arg);
             if (is_mem(arg) || is_unit(arg)) continue;
             bb->args[arg] = val;
             auto* param = dst_cont->param(index);
-            auto& phi = current_fn_->bbs_map[dst_cont]->phis_map[param];
-            phi.preds.emplace_back(bb->args[arg], current_fn_->labels[continuation]);
+            auto& phi = cont2bb_[dst_cont]->phis_map[param];
+            phi.preds.emplace_back(bb->args[arg], emit(continuation));
         }
-        bb->branch(current_fn_->labels[dst_cont]);
+        bb->branch(emit(dst_cont));
     } else if (app.callee() == world().branch()) {
         auto mem = app.arg(0);
         emit_unsafe(mem);
 
-        auto cond = emit(app.arg(1), bb);
+        auto cond = emit(app.arg(1));
         bb->args.emplace(app.arg(2), cond);
-        auto tbb = current_fn_->labels[app.arg(2)->isa_nom<Continuation>()];
-        auto fbb = current_fn_->labels[app.arg(3)->isa_nom<Continuation>()];
+        auto tbb = emit(app.arg(2));
+        auto fbb = emit(app.arg(3));
         bb->branch_conditional(cond, tbb, fbb);
     } else if (app.callee()->isa<Continuation>() && app.callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         /*auto val = emit(continuation->arg(0));
@@ -343,7 +340,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     } else if (auto builtin = app.callee()->isa_nom<Continuation>(); builtin->is_imported()) {
         // Ensure we emit previous memory operations
         assert(is_mem(app.arg(0)));
-        emit(app.arg(0), bb);
+        emit(app.arg(0));
 
         auto productions = emit_builtin(app, builtin, bb);
         auto succ = app.args().back()->isa_nom<Continuation>();
@@ -357,7 +354,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         for (auto arg : app.args()) {
             if (arg->order() == 0) {
                 auto arg_type = arg->type();
-                auto arg_val = emit(arg, bb);
+                auto arg_val = emit(arg);
                 if (arg_type == world().unit_type() || arg_type == world().mem_type()) continue;
                 call_args.push_back(arg_val);
             } else {
@@ -370,7 +367,7 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
 
         SpvId call_result;
         if (auto called_continuation = app.callee()->isa_nom<Continuation>()) {
-            call_result = bb->call(ret_type, emit(called_continuation, bb), call_args);
+            call_result = bb->call(ret_type, emit(called_continuation), call_args);
         } else {
             // must be a closure
             THORIN_UNREACHABLE;
@@ -393,12 +390,12 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
         }
 
         if (n == 0) {
-            bb->branch(current_fn_->labels[succ]);
+            bb->branch(emit(succ));
         } else if (n == 1) {
-            bb->branch(current_fn_->labels[succ]);
+            bb->branch(emit(succ));
 
-            auto& phi = current_fn_->bbs_map[succ]->phis_map[last_param];
-            phi.preds.emplace_back(call_result, current_fn_->labels[continuation]);
+            auto& phi = cont2bb_[succ]->phis_map[last_param];
+            phi.preds.emplace_back(call_result, emit(continuation));
         } else {
             Array<SpvId> extracts(n);
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
@@ -409,15 +406,15 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
                 j++;
             }
 
-            bb->branch(current_fn_->labels[succ]);
+            bb->branch(emit(succ));
 
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
                 auto param = succ->param(i);
                 if (is_mem(param) || is_unit(param))
                     continue;
 
-                auto& phi = current_fn_->bbs_map[succ]->phis_map[param];
-                phi.preds.emplace_back(extracts[j], current_fn_->labels[continuation]);
+                auto& phi = cont2bb_[succ]->phis_map[last_param];
+                phi.preds.emplace_back(extracts[j], emit(continuation));
 
                 j++;
             }
@@ -425,10 +422,10 @@ void CodeGen::emit_epilogue(Continuation* continuation, BasicBlockBuilder* bb) {
     }
 }
 
-SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
+SpvId CodeGen::emit_bb(const Def* def, BasicBlockBuilder* bb) {
     if (auto bin = def->isa<BinOp>()) {
-        SpvId lhs = emit(bin->lhs(), bb);
-        SpvId rhs = emit(bin->rhs(), bb);
+        SpvId lhs = emit(bin->lhs());
+        SpvId rhs = emit(bin->rhs());
         SpvId result_type = convert(def->type()).id;
 
         if (auto cmp = bin->isa<Cmp>()) {
@@ -560,7 +557,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             assert((*param_id) != 0);
             return *param_id;
         } else {
-            auto val = (*current_fn_->bbs_map[param->continuation()]).phis_map[param].value;
+            auto val = cont2bb_[param->continuation()]->phis_map[param].value;
             assert(val != 0);
             return val;
         }
@@ -600,14 +597,14 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
 
         return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));*/
     } else if (auto vindex = def->isa<VariantIndex>()) {
-        auto value = emit(vindex->op(0), bb);
+        auto value = emit(vindex->op(0));
         return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
         std::vector<SpvId> elements;
         elements.resize(tuple->num_ops());
         size_t x = 0;
         for (auto& e : tuple->ops()) {
-            elements[x++] = emit(e, bb);
+            elements[x++] = emit(e);
         }
         return bb->composite(convert(tuple->type()).id, elements);
     } else if (auto structagg = def->isa<StructAgg>()) {
@@ -615,12 +612,12 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
         elements.resize(structagg->num_ops());
         size_t x = 0;
         for (auto& e : structagg->ops()) {
-            elements[x++] = emit(e, bb);
+            elements[x++] = emit(e);
         }
         return bb->composite(convert(structagg->type()).id, elements);
     } else if (auto access = def->isa<Access>()) {
         // emit dependent operations first
-        emit(access->mem(), bb);
+        emit(access->mem());
 
         std::vector<uint32_t> operands;
         auto ptr_type = access->ptr()->type()->as<PtrType>();
@@ -629,9 +626,9 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             operands.push_back( 4 ); // TODO: SPIR-V docs say to consult client API for valid values.
         }
         if (auto load = def->isa<Load>()) {
-            return bb->load(convert(load->out_val_type()).id, emit(load->ptr(), bb), operands);
+            return bb->load(convert(load->out_val_type()).id, emit(load->ptr()), operands);
         } else if (auto store = def->isa<Store>()) {
-            bb->store(emit(store->val(), bb), emit(store->ptr(), bb), operands);
+            bb->store(emit(store->val()), emit(store->ptr()), operands);
             return spv_none;
         } else THORIN_UNREACHABLE;
     } else if (auto lea = def->isa<LEA>()) {
@@ -644,10 +641,10 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
                 break;
         }
         auto type = convert(lea->ptr_type()).id;
-        auto offset = emit(lea->index(), bb);
-        return bb->ptr_access_chain(type, emit(lea->ptr(), bb), offset, {});
+        auto offset = emit(lea->index());
+        return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {});
     } else if (auto aggop = def->isa<AggOp>()) {
-        auto spv_agg = emit(aggop->agg(), bb);
+        auto spv_agg = emit(aggop->agg());
         auto agg_type = convert(aggop->agg()->type()).id;
 
         bool mem = false;
@@ -661,7 +658,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             bb->store(spv_agg, variable);
 
             auto cell_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, target_type);
-            auto cell = bb->access_chain(cell_ptr_type, variable, { emit(aggop->index(), bb)} );
+            auto cell = bb->access_chain(cell_ptr_type, variable, { emit(aggop->index())} );
             return std::make_pair(variable, cell);
         };
 
@@ -679,7 +676,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             }
 
             if (extract->agg()->type()->isa<VectorType>())
-                return bb->vector_extract_dynamic(target_type, spv_agg, emit(extract->index(), bb));
+                return bb->vector_extract_dynamic(target_type, spv_agg, emit(extract->index()));
 
             // index *must* be constant for the remaining possible cases
             assert(constant_index != nullptr);
@@ -693,7 +690,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
 
             return bb->extract(target_type, spv_agg, { index - offset });
         } else if (auto insert = def->isa<Insert>()) {
-            auto value = emit(insert->value(), bb);
+            auto value = emit(insert->value());
             auto constant_index = aggop->index()->isa<PrimLit>();
 
             // TODO deal with mem - but I think for now this case shouldn't happen
@@ -706,7 +703,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             }
 
             if (insert->agg()->type()->isa<VectorType>())
-                return bb->vector_insert_dynamic(agg_type, spv_agg, value, emit(insert->index(), bb));
+                return bb->vector_insert_dynamic(agg_type, spv_agg, value, emit(insert->index()));
 
             // index *must* be constant for the remaining possible cases
             assert(constant_index != nullptr);
@@ -726,7 +723,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             if (conv_src_type.layout->size != conv_dst_type.layout->size)
                 world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type.layout->size, conv_dst_type.layout->size);
 
-            return bb->convert(spv::OpBitcast, convert(bitcast->type()).id, emit(bitcast->from(), bb));
+            return bb->convert(spv::OpBitcast, convert(bitcast->type()).id, emit(bitcast->from()));
         } else if (auto cast = def->isa<Cast>()) {
             // NB: all ops used here are scalar/vector agnostic
             auto src_prim = src_type->isa<PrimType>();
@@ -741,7 +738,7 @@ SpvId CodeGen::emit(const Def* def, BasicBlockBuilder* bb) {
             size_t src_bitwidth = conv_src_type.layout->size;
             size_t dst_bitwidth = conv_src_type.layout->size;
 
-            SpvId data = emit(cast->from(), bb);
+            SpvId data = emit(cast->from());
 
             // If floating point is involved (src or dst), OpConvert*ToF and OpConvertFTo* can take care of the bit width transformation so no need for any chopping/expanding
             if (src_kind == PrimTypeKind::Float || dst_kind == PrimTypeKind::Float) {
@@ -821,7 +818,7 @@ std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* bui
         } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
 
         for (size_t i = 2; i < app.num_args() - 1; i++) {
-            args.push_back(emit(app.arg(i), bb));
+            args.push_back(emit(app.arg(i)));
         }
 
         bb->ext_instruction(convert(world().unit_type()).id, builder_->imported_instrs->shader_printf, 1, args);
@@ -829,23 +826,23 @@ std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* bui
         THORIN_UNREACHABLE;
     } else if (builtin->name() == "get_global_id") {
         auto vector = bb->load(uvec3_t, builder_->builtins->global_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_local_size") {
         auto vector = bb->load(uvec3_t, builder_->builtins->workgroup_size);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_local_id") {
         auto vector = bb->load(uvec3_t, builder_->builtins->local_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_num_groups") {
         auto vector = bb->load(uvec3_t, builder_->builtins->num_workgroups);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else if (builtin->name() == "get_group_id") {
         auto vector = bb->load(uvec3_t, builder_->builtins->workgroup_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1), bb));
+        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
         productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 2239b677d..e3e880c7e 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -3,6 +3,7 @@
 
 #include "thorin/be/spirv/spirv_builder.hpp"
 #include "thorin/be/codegen.h"
+#include "thorin/be/emitter.h"
 
 namespace thorin::spirv {
 
@@ -44,16 +45,11 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {
-    explicit FnBuilder(CodeGen* cg, FileBuilder& file_builder);
+    explicit FnBuilder(FileBuilder& file_builder);
     FnBuilder(const FnBuilder&) = delete;
 
-    CodeGen* cg;
     FileBuilder& file_builder;
-
-    const Scope* scope = nullptr;
     std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
-    std::unordered_map<Continuation*, BasicBlockBuilder*> bbs_map;
-    ContinuationMap<SpvId> labels;
     DefMap<SpvId> params;
 };
 
@@ -90,18 +86,29 @@ struct FileBuilder : public builder::SpvFileBuilder {
     SpvId u32_t_ { 0 };
 };
 
-class CodeGen : public thorin::CodeGen {
+class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, SpvId, BasicBlockBuilder*, CodeGen> {
 public:
     CodeGen(Thorin& thorin, SpvTargetInfo, Cont2Config&, bool debug);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
 
+    bool is_valid(SpvId id) {
+        return id > 0;
+    }
+
     ConvertedType convert(const Type*);
+
+    SpvId emit_fun_decl(Continuation*);
+
+    FnBuilder& prepare(const Scope&);
+    void prepare(Continuation*, FnBuilder&);
+    void finalize(const Scope&);
+    void finalize(Continuation*);
 protected:
-    void emit(const Scope& scope);
-    void emit_epilogue(Continuation*, BasicBlockBuilder* bb);
-    SpvId emit(const Def* def, BasicBlockBuilder* bb);
+    FnBuilder& get_fn_builder(Continuation*);
+    void emit_epilogue(Continuation*);
+    SpvId emit_bb(const Def* def, BasicBlockBuilder* bb);
     std::vector<SpvId> emit_builtin(const App&, const Continuation*, BasicBlockBuilder*);
 
     SpvId get_codom_type(const Continuation* fn);
@@ -110,6 +117,7 @@ class CodeGen : public thorin::CodeGen {
     std::unique_ptr<FileBuilder> builder_;
     Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
+    ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
     DefMap<ConvertedType> types_;
     DefMap<SpvId> defs_;
     const Cont2Config& kernel_config_;

From 01855fd2044ad1db0ed8780c16c1b624f9092bdc Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 17:54:44 +0200
Subject: [PATCH 232/342] builds

---
 src/thorin/be/spirv/spirv.cpp | 90 ++++++++++++++++++-----------------
 src/thorin/be/spirv/spirv.h   | 11 +++--
 2 files changed, 54 insertions(+), 47 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index ee732ff4b..f16b7b340 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -199,16 +199,16 @@ FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
     return fn;
 }
 
-FnBuilder& CodeGen::prepare(const thorin::Scope& scope) {
+FnBuilder* CodeGen::prepare(const thorin::Scope& scope) {
     auto& fn = get_fn_builder(scope.entry());
     builder_->name(fn.function_id, scope.entry()->name());
-    return fn;
+    return &fn;
 }
 
-void CodeGen::prepare(thorin::Continuation* cont, FnBuilder& fn) {
-    auto& bb = *fn.bbs.emplace_back(std::make_unique<BasicBlockBuilder>(fn));
+void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
+    auto& bb = *fn->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*fn));
     cont2bb_[cont] = &bb;
-    fn.bbs_to_emit.emplace_back(&bb);
+    fn->bbs_to_emit.emplace_back(&bb);
 
     builder_->name(bb.label, cont->name().c_str());
 
@@ -218,8 +218,8 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder& fn) {
                 // Nothing
             } else if (param->order() == 0) {
                 auto param_t = convert(param->type());
-                auto id = fn.parameter(param_t.id);
-                fn.params[param] = id;
+                auto id = fn->parameter(param_t.id);
+                fn->params[param] = id;
                 if (param->type()->isa<PtrType>()) {
                     builder_->decorate(id, spv::DecorationAliased);
                 }
@@ -422,7 +422,46 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
     }
 }
 
-SpvId CodeGen::emit_bb(const Def* def, BasicBlockBuilder* bb) {
+SpvId CodeGen::emit_constant(const thorin::Def* def) {
+    if (auto primlit = def->isa<PrimLit>()) {
+        Box box = primlit->value();
+        auto type = convert(def->type()).id;
+        SpvId constant;
+        switch (primlit->primtype_tag()) {
+            case PrimType_bool:                     constant = builder_->bool_constant(type, box.get_bool()); break;
+            case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
+            case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
+            case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
+            case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
+            case PrimType_ps32: case PrimType_qs32: constant = builder_->constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
+            case PrimType_pu32: case PrimType_qu32: constant = builder_->constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
+            case PrimType_ps64: case PrimType_qs64:
+            case PrimType_pu64: case PrimType_qu64: {
+                uint64_t value = static_cast<uint64_t>(box.get_u64());
+                uint64_t upper = value >> 32U;
+                uint64_t lower = value & 0xFFFFFFFFU;
+                constant = builder_->constant(type, { (uint32_t) lower, (uint32_t) upper });
+                break;
+            }
+            case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
+            case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
+            case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
+        }
+        return constant;
+    } else if (auto param = def->isa<Param>()) {
+        if (is_mem(param)) return spv_none;
+        if (auto param_id = current_fn_->params.lookup(param)) {
+            assert((*param_id) != 0);
+            return *param_id;
+        } else {
+            auto val = cont2bb_[param->continuation()]->phis_map[param].value;
+            assert(val != 0);
+            return val;
+        }
+    } else return emit_bb(nullptr, def);
+}
+
+SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs());
         SpvId rhs = emit(bin->rhs());
@@ -526,41 +565,6 @@ SpvId CodeGen::emit_bb(const Def* def, BasicBlockBuilder* bb) {
             }
             THORIN_UNREACHABLE;
         }
-    } else if (auto primlit = def->isa<PrimLit>()) {
-        Box box = primlit->value();
-        auto type = convert(def->type()).id;
-        SpvId constant;
-        switch (primlit->primtype_tag()) {
-            case PrimType_bool:                     constant = bb->file_builder.bool_constant(type, box.get_bool()); break;
-            case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
-            case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
-            case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
-            case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
-            case PrimType_ps32: case PrimType_qs32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
-            case PrimType_pu32: case PrimType_qu32: constant = bb->file_builder.constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
-            case PrimType_ps64: case PrimType_qs64:
-            case PrimType_pu64: case PrimType_qu64: {
-                uint64_t value = static_cast<uint64_t>(box.get_u64());
-                uint64_t upper = value >> 32U;
-                uint64_t lower = value & 0xFFFFFFFFU;
-                constant = bb->file_builder.constant(type, { (uint32_t) lower, (uint32_t) upper });
-                break;
-            }
-            case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
-            case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
-            case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
-        }
-        return constant;
-    } else if (auto param = def->isa<Param>()) {
-        if (is_mem(param)) return spv_none;
-        if (auto param_id = current_fn_->params.lookup(param)) {
-            assert((*param_id) != 0);
-            return *param_id;
-        } else {
-            auto val = cont2bb_[param->continuation()]->phis_map[param].value;
-            assert(val != 0);
-            return val;
-        }
     } else if (auto variant = def->isa<Variant>()) {
         assert(false && "TODO: rewrite");
         /*auto variant_type = def->type()->as<VariantType>();
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index e3e880c7e..c474b92ae 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -2,6 +2,7 @@
 #define THORIN_SPIRV_H
 
 #include "thorin/be/spirv/spirv_builder.hpp"
+#include "thorin/analyses/schedule.h"
 #include "thorin/be/codegen.h"
 #include "thorin/be/emitter.h"
 
@@ -101,14 +102,16 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, SpvId, Bas
 
     SpvId emit_fun_decl(Continuation*);
 
-    FnBuilder& prepare(const Scope&);
-    void prepare(Continuation*, FnBuilder&);
+    FnBuilder* prepare(const Scope&);
+    void prepare(Continuation*, FnBuilder*);
+    void emit_epilogue(Continuation*);
     void finalize(const Scope&);
     void finalize(Continuation*);
+
+    SpvId emit_constant(const Def*);
+    SpvId emit_bb(BasicBlockBuilder* bb, const Def* def);
 protected:
     FnBuilder& get_fn_builder(Continuation*);
-    void emit_epilogue(Continuation*);
-    SpvId emit_bb(const Def* def, BasicBlockBuilder* bb);
     std::vector<SpvId> emit_builtin(const App&, const Continuation*, BasicBlockBuilder*);
 
     SpvId get_codom_type(const Continuation* fn);

From 3d03c0d94694eae28b597f835573792f113dc50b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 27 Jun 2024 18:07:37 +0200
Subject: [PATCH 233/342] cleanup

---
 src/thorin/be/spirv/spirv.h | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index c474b92ae..61b4728b9 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -87,7 +87,7 @@ struct FileBuilder : public builder::SpvFileBuilder {
     SpvId u32_t_ { 0 };
 };
 
-class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, SpvId, BasicBlockBuilder*, CodeGen> {
+class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
 public:
     CodeGen(Thorin& thorin, SpvTargetInfo, Cont2Config&, bool debug);
 
@@ -118,11 +118,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, SpvId, Bas
 
     SpvTargetInfo target_info_;
     std::unique_ptr<FileBuilder> builder_;
-    Continuation* entry_ = nullptr;
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
-    DefMap<ConvertedType> types_;
-    DefMap<SpvId> defs_;
     const Cont2Config& kernel_config_;
 };
 

From 912ab06701bfd855fc9109d0fb27a3a02c860b1e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 14:15:07 +0200
Subject: [PATCH 234/342] spirv: don't leak spirv headers downstream

---
 src/thorin/be/spirv/spirv.cpp       |  2 +-
 src/thorin/be/spirv/spirv.h         | 38 ++--------------------
 src/thorin/be/spirv/spirv_private.h | 49 +++++++++++++++++++++++++++++
 3 files changed, 52 insertions(+), 37 deletions(-)
 create mode 100644 src/thorin/be/spirv/spirv_private.h

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index f16b7b340..f2106cd96 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -1,4 +1,4 @@
-#include "thorin/be/spirv/spirv.h"
+#include "spirv_private.h"
 
 #include "thorin/analyses/scope.h"
 #include "thorin/analyses/schedule.h"
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 61b4728b9..955e4df15 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -1,14 +1,13 @@
 #ifndef THORIN_SPIRV_H
 #define THORIN_SPIRV_H
 
-#include "thorin/be/spirv/spirv_builder.hpp"
 #include "thorin/analyses/schedule.h"
 #include "thorin/be/codegen.h"
 #include "thorin/be/emitter.h"
 
 namespace thorin::spirv {
 
-using SpvId = builder::SpvId;
+using SpvId = uint32_t;
 
 class CodeGen;
 
@@ -35,25 +34,6 @@ struct ConvertedType {
     std::optional<Layout> layout;
 };
 
-struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
-    explicit BasicBlockBuilder(FnBuilder& fn_builder);
-    BasicBlockBuilder(const BasicBlockBuilder&) = delete;
-
-    FnBuilder& fn_builder;
-    FileBuilder& file_builder;
-    std::unordered_map<const Param*, Phi> phis_map;
-    DefMap<SpvId> args;
-};
-
-struct FnBuilder : public builder::SpvFnBuilder {
-    explicit FnBuilder(FileBuilder& file_builder);
-    FnBuilder(const FnBuilder&) = delete;
-
-    FileBuilder& file_builder;
-    std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
-    DefMap<SpvId> params;
-};
-
 struct Builtins {
     SpvId workgroup_size;
     SpvId num_workgroups;
@@ -71,21 +51,7 @@ struct ImportedInstructions {
     explicit ImportedInstructions(FileBuilder&);
 };
 
-struct FileBuilder : public builder::SpvFileBuilder {
-    explicit FileBuilder(CodeGen* cg);
-    FileBuilder(const FileBuilder&) = delete;
-
-    CodeGen* cg;
-
-    std::unique_ptr<Builtins> builtins;
-    std::unique_ptr<ImportedInstructions> imported_instrs;
-
-    SpvId u32_t();
-    SpvId u32_constant(uint32_t);
-
-private:
-    SpvId u32_t_ { 0 };
-};
+struct BasicBlockBuilder;
 
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
 public:
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
new file mode 100644
index 000000000..228f73b2d
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -0,0 +1,49 @@
+#ifndef THORIN_SPIRV_PRIVATE_H
+#define THORIN_SPIRV_PRIVATE_H
+
+#include "spirv.h"
+
+#include "thorin/be/spirv/spirv_builder.hpp"
+
+namespace thorin::spirv {
+
+struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
+    explicit BasicBlockBuilder(FnBuilder& fn_builder);
+
+    BasicBlockBuilder(const BasicBlockBuilder&) = delete;
+
+    FnBuilder& fn_builder;
+    FileBuilder& file_builder;
+    std::unordered_map<const Param*, Phi> phis_map;
+    DefMap<SpvId> args;
+};
+
+struct FnBuilder : public builder::SpvFnBuilder {
+    explicit FnBuilder(FileBuilder& file_builder);
+
+    FnBuilder(const FnBuilder&) = delete;
+
+    FileBuilder& file_builder;
+    std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
+    DefMap<SpvId> params;
+};
+
+struct FileBuilder : public builder::SpvFileBuilder {
+    explicit FileBuilder(CodeGen* cg);
+    FileBuilder(const FileBuilder&) = delete;
+
+    CodeGen* cg;
+
+    std::unique_ptr<Builtins> builtins;
+    std::unique_ptr<ImportedInstructions> imported_instrs;
+
+    SpvId u32_t();
+    SpvId u32_constant(uint32_t);
+
+private:
+    SpvId u32_t_ { 0 };
+};
+
+}
+
+#endif // THORIN_SPIRV_PRIVATE_H

From 6ebf90e7a03eb6a0412f02c4c6278e0be9c0eab7 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 14:15:33 +0200
Subject: [PATCH 235/342] spirv: advertise in thorin-config.cmake.in

---
 cmake/thorin-config.cmake.in | 1 +
 1 file changed, 1 insertion(+)

diff --git a/cmake/thorin-config.cmake.in b/cmake/thorin-config.cmake.in
index 3a53346dd..641f33e0c 100644
--- a/cmake/thorin-config.cmake.in
+++ b/cmake/thorin-config.cmake.in
@@ -32,6 +32,7 @@ set(Thorin_HAS_JSON_SUPPORT @THORIN_ENABLE_JSON@)
 set(Thorin_HAS_LLVM_SUPPORT @LLVM_FOUND@)
 set(Thorin_HAS_RV_SUPPORT @RV_FOUND@)
 set(Thorin_HAS_SHADY_SUPPORT @THORIN_ENABLE_SHADY@)
+set(Thorin_HAS_SPIRV_SUPPORT @THORIN_ENABLE_SPIRV@)
 set(AnyDSL_LLVM_LINK_SHARED @AnyDSL_LLVM_LINK_SHARED@)
 
 if(Thorin_HAS_LLVM_SUPPORT)

From e371c8d4cf9480227c2e15d27a69c5728b66658a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 14:43:04 +0200
Subject: [PATCH 236/342] spirv: make kernel_config optional

---
 src/thorin/be/spirv/spirv.cpp | 10 ++++++----
 src/thorin/be/spirv/spirv.h   |  4 ++--
 2 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index f2106cd96..add9e2314 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -142,7 +142,7 @@ ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
     shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
 }
 
-CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, Cont2Config& kernel_config, bool debug)
+CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, bool debug, const Cont2Config* kernel_config)
         : thorin::CodeGen(thorin, debug), target_info_(target_info), kernel_config_(kernel_config)
 {}
 
@@ -162,9 +162,11 @@ void CodeGen::emit_stream(std::ostream& out) {
     }
 
     for (auto& cont : world().copy_continuations()) {
-        if (cont->is_exported()) {
-            assert(defs_.contains(cont) && kernel_config_.contains(cont));
-            auto config = kernel_config_.find(cont);
+        assert(defs_.contains(cont));
+        if (cont->is_exported() && kernel_config_) {
+            auto config = kernel_config_->find(cont);
+            if (config == kernel_config_->end())
+                continue;
 
             SpvId callee = defs_[cont];
 
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 955e4df15..714497a79 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -55,7 +55,7 @@ struct BasicBlockBuilder;
 
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
 public:
-    CodeGen(Thorin& thorin, SpvTargetInfo, Cont2Config&, bool debug);
+    CodeGen(Thorin& thorin, SpvTargetInfo, bool debug, const Cont2Config* = nullptr);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
@@ -86,7 +86,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     std::unique_ptr<FileBuilder> builder_;
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
-    const Cont2Config& kernel_config_;
+    const Cont2Config* kernel_config_;
 };
 
 }

From c4969b3511ab23fb8f2b11ea622ef66007d55c6d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 14:48:53 +0200
Subject: [PATCH 237/342] add spirv_types to the index

---
 src/thorin/be/spirv/spirv_types.cpp | 225 ++++++++++++++++++++++++++++
 1 file changed, 225 insertions(+)
 create mode 100644 src/thorin/be/spirv/spirv_types.cpp

diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
new file mode 100644
index 000000000..51bb71e8c
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -0,0 +1,225 @@
+#include "spirv_private.h"
+#include "thorin/util/stream.h"
+#include "thorin/util/utility.h"
+
+namespace thorin::spirv {
+
+ConvertedType CodeGen::convert(const Type* type) {
+    // Spir-V requires each primitive type to be "unique", it doesn't allow for example two 32-bit signed integer types.
+    // Therefore we must enforce that precise/quick types map to the same thing.
+    switch (type->tag()) {
+#define THORIN_Q_TYPE(T, M) \
+    case PrimType_##T: \
+        type = world().prim_type(PrimType_p##M, type->as<VectorType>()->length()); \
+        break;
+#include "thorin/tables/primtypetable.h"
+#undef THORIN_Q_TYPE
+        default: break;
+    }
+
+    if (auto iter = types_.find(type); iter != types_.end())
+        return iter->second;
+
+    ConvertedType converted = { 0, std::nullopt };
+
+    switch (type->tag()) {
+        // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
+        // Using extensions, we could use 16 or 8-bit ints instead
+        // We can also pack them inside structures using bit-twiddling tricks, if the need arises
+        // Note: this only affects storing booleans inside structures, for regular variables the actual spir-v bool type is used.
+        case Node_PrimType_bool:
+            converted.id = builder_->declare_bool_type();
+            converted.layout = { 1, 1 };
+            break;
+        case Node_PrimType_ps8:
+            converted.id = builder_->declare_int_type(8, true);
+            converted.layout = { 1, 1 };
+            break;
+        case Node_PrimType_pu8:
+            converted.id = builder_->declare_int_type(8, false);
+            converted.layout = { 1, 1 };
+            break;
+        case Node_PrimType_ps16:
+            converted.id = builder_->declare_int_type(16, true);
+            converted.layout = { 2, 2 };
+            break;
+        case Node_PrimType_pu16:
+            converted.id = builder_->declare_int_type(16, false);
+            converted.layout = { 2, 2 };
+            break;
+        case Node_PrimType_ps32:
+            converted.id = builder_->declare_int_type(32, true );
+            converted.layout = { 4, 4 };
+            break;
+        case Node_PrimType_pu32:
+            converted.id = builder_->declare_int_type(32, false);
+            converted.layout = { 4, 4 };
+            break;
+        case Node_PrimType_ps64:
+            converted.id = builder_->declare_int_type(64, true);
+            converted.layout = { 8, 8 };
+            break;
+        case Node_PrimType_pu64:
+            converted.id = builder_->declare_int_type(64, false);
+            converted.layout = { 8, 8 };
+            break;
+        case Node_PrimType_pf16:
+            converted.id = builder_->declare_float_type(16);
+            converted.layout = { 2, 2 };
+            break;
+        case Node_PrimType_pf32:
+            converted.id = builder_->declare_float_type(32);
+            converted.layout = { 4, 4 };
+            break;
+        case Node_PrimType_pf64:
+            converted.id = builder_->declare_float_type(64);
+            converted.layout = { 8, 8 };
+            break;
+        case Node_PtrType: {
+            auto ptr = type->as<PtrType>();
+            spv::StorageClass storage_class;
+            switch (ptr->addr_space()) {
+                case AddrSpace::Function: storage_class = spv::StorageClassFunction;              break;
+                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;               break;
+                case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
+                case AddrSpace::Global:   storage_class = spv::StorageClassCrossWorkgroup;        break;
+                case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
+                default:
+                    assert(false && "This address space is not supported");
+                    break;
+            }
+
+            const Type* pointee = ptr->pointee();
+            while (auto arr = pointee->isa<IndefiniteArrayType>())
+                pointee = arr->elem_type();
+            converted.id = builder_->declare_ptr_type(storage_class, convert(pointee).id);
+            converted.layout = { target_info_.mem_layout.pointer_size, target_info_.mem_layout.pointer_size };
+            break;
+        }
+        case Node_IndefiniteArrayType: {
+            world().ELOG("Using indefinite types directly is not permitted - they may only be pointed to");
+            std::abort();
+        }
+        case Node_DefiniteArrayType: {
+            auto array = type->as<DefiniteArrayType>();
+            auto element = convert(array->elem_type());
+            converted.id = builder_->declare_array_type(element.id, builder_->u32_constant(array->dim()));
+            converted.layout = { element.layout->size * array->dim(), element.layout->alignment };
+            break;
+        }
+
+        case Node_ClosureType:
+        case Node_FnType: {
+            // extract "return" type, collect all other types
+            auto fn = type->as<FnType>();
+            SpvId ret = 0;
+            std::vector<SpvId> ops;
+            for (auto op : fn->types()) {
+                if (op->isa<MemType>() || op == world().unit_type())
+                    continue;
+                auto fn_type = op->isa<FnType>();
+                if (fn_type && !op->isa<ClosureType>()) {
+                    assert(!ret && "only one 'return' supported");
+                    std::vector<SpvId> ret_types;
+                    for (auto fn_op : fn_type->types()) {
+                        if (fn_op->isa<MemType>() || fn_op == world().unit_type())
+                            continue;
+                        ret_types.push_back(convert(fn_op).id);
+                    }
+                    if (ret_types.empty())          ret = convert(world().tuple_type({})).id;
+                    else if (ret_types.size() == 1) ret = ret_types.back();
+                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
+                } else
+                    ops.push_back(convert(op).id);
+            }
+            assert(ret);
+
+            if (type->tag() == Node_FnType) {
+                converted.id = builder_->declare_fn_type(ops, ret);
+            } else {
+                assert(false && "TODO: handle closure mess");
+                THORIN_UNREACHABLE;
+            }
+            break;
+        }
+
+        case Node_Vector: {
+            auto vec = type->as<VectorType>();
+            assert(vec->length() > 1);
+            auto component = vec->scalarize();
+            auto conv_comp = convert(component);
+            converted.id = builder_->declare_vector_type(conv_comp.id, (uint32_t)vec->length());
+            converted.layout = conv_comp.layout;
+            converted.layout->size *= vec->length();
+            break;
+        }
+
+        case Node_StructType:
+        case Node_TupleType: {
+            std::vector<SpvId> spv_types;
+            size_t total_serialized_size = 0;
+            converted.layout = { 0, 0 };
+            for (auto member : type->ops()) {
+                auto member_type = member->as<Type>();
+                if (member_type == world().unit_type() || member_type == world().mem_type()) continue;
+                auto converted_member_type = convert(member_type);
+                assert(converted_member_type.layout);
+                spv_types.push_back(converted_member_type.id);
+                converted.layout->alignment = std::max(converted.layout->alignment, converted_member_type.layout->alignment);
+                converted.layout->size = pad(converted.layout->size + converted_member_type.layout->size, converted.layout->alignment);
+            }
+            if (total_serialized_size == 0) {
+                outf("this one is void");
+                converted.id = builder_->declare_void_type();
+                converted.layout = std::nullopt;
+                break;
+            }
+
+            converted.id = builder_->declare_struct_type(spv_types);
+            builder_->name(converted.id, type->to_string());
+            break;
+        }
+
+        case Node_VariantType: {
+            assert(type->num_ops() > 0 && "empty variants not supported");
+            auto tag_type = world().type_pu32();
+            SpvId converted_tag_type = convert(tag_type).id;
+
+            size_t max_serialized_size = 0;
+            for (auto member : type->as<VariantType>()->types()) {
+                auto member_type = member->as<Type>();
+                if (member_type == world().unit_type() || member_type == world().mem_type()) continue;
+                auto converted_member_type = convert(member_type);
+                assert(converted_member_type.layout);
+                if (converted_member_type.layout->size > max_serialized_size)
+                    max_serialized_size = converted_member_type.layout->size;
+            }
+
+            if (max_serialized_size > 0) {
+                auto payload_type = world().definite_array_type(world().type_pu8(), max_serialized_size);
+                auto struct_t = world().struct_type(type->name(), 2);
+                struct_t->set_op(0, tag_type);
+                struct_t->set_op(1, payload_type);
+                return convert(struct_t);
+            } else {
+                // We keep this useless level of struct so the rest of the code doesn't need a special path to extract the tag
+                auto struct_t = world().struct_type(type->name(), 1);
+                struct_t->set_op(0, tag_type);
+                return convert(struct_t);
+            }
+            THORIN_UNREACHABLE;
+        }
+
+        case Node_MemType: {
+            assert(false && "MemType cannot be converted to SPIR-V");
+        }
+
+        default:
+            THORIN_UNREACHABLE;
+    }
+
+    types_[type] = converted;
+    return converted;
+}
+
+}

From e77abf90438564ac9b1090cfb1e3fa9228836b9f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 14:57:11 +0200
Subject: [PATCH 238/342] spirv: adjust header to not rely on unavailable
 definitions

---
 src/thorin/be/spirv/spirv.cpp       | 13 +++++++------
 src/thorin/be/spirv/spirv.h         |  4 +---
 src/thorin/be/spirv/spirv_private.h |  3 +++
 3 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index add9e2314..daa23f920 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -147,7 +147,8 @@ CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, bool debug, const Co
 {}
 
 void CodeGen::emit_stream(std::ostream& out) {
-    builder_ = std::make_unique<FileBuilder>(this);
+    FileBuilder builder(this);
+    builder_ = &builder;
 
     builder_->builtins = std::make_unique<Builtins>(*builder_);
     builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
@@ -191,11 +192,11 @@ SpvId CodeGen::emit_fun_decl(thorin::Continuation* continuation) {
 }
 
 FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
-    if (auto found = fn_builders_.find(continuation); found != fn_builders_.end()) {
+    if (auto found = builder_->fn_builders_.find(continuation); found != builder_->fn_builders_.end()) {
         return *found->second;
     }
 
-    auto& fn = *(fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
+    auto& fn = *(builder_->fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
     fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
     return fn;
@@ -252,7 +253,7 @@ void CodeGen::finalize(thorin::Continuation* cont) {
 }
 
 void CodeGen::finalize(const thorin::Scope&) {
-    builder_->define_function(*current_fn_);
+    builder_->define_function(*builder_->current_fn_);
 }
 
 SpvId CodeGen::get_codom_type(const Continuation* fn) {
@@ -303,7 +304,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         switch (values.size()) {
             case 0:  bb->return_void();      break;
             case 1:  bb->return_value(values[0]); break;
-            default: bb->return_value(bb->composite(current_fn_->fn_ret_type, values));
+            default: bb->return_value(bb->composite(builder_->current_fn_->fn_ret_type, values));
         }
     } else if (auto dst_cont = app.callee()->isa_nom<Continuation>(); dst_cont && dst_cont->is_basicblock()) { // ordinary jump
         int index = -1;
@@ -452,7 +453,7 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
         return constant;
     } else if (auto param = def->isa<Param>()) {
         if (is_mem(param)) return spv_none;
-        if (auto param_id = current_fn_->params.lookup(param)) {
+        if (auto param_id = builder_->current_fn_->params.lookup(param)) {
             assert((*param_id) != 0);
             return *param_id;
         } else {
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 714497a79..05d27f628 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -83,9 +83,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     SpvId get_codom_type(const Continuation* fn);
 
     SpvTargetInfo target_info_;
-    std::unique_ptr<FileBuilder> builder_;
-    FnBuilder* current_fn_ = nullptr;
-    ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
+    FileBuilder* builder_;
     const Cont2Config* kernel_config_;
 };
 
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index 228f73b2d..745b1ea0d 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -37,6 +37,9 @@ struct FileBuilder : public builder::SpvFileBuilder {
     std::unique_ptr<Builtins> builtins;
     std::unique_ptr<ImportedInstructions> imported_instrs;
 
+    FnBuilder* current_fn_ = nullptr;
+    ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
+
     SpvId u32_t();
     SpvId u32_constant(uint32_t);
 

From 559ef0c31946f048b355add77956db2dbfe9416f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 15:10:56 +0200
Subject: [PATCH 239/342] spirv: emit parameters/phis properly

---
 src/thorin/be/spirv/spirv.cpp | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index daa23f920..d0ac89908 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -219,9 +219,11 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
         for (auto param : cont->params()) {
             if (is_mem(param) || is_unit(param)) {
                 // Nothing
+                defs_[param] = 0;
             } else if (param->order() == 0) {
                 auto param_t = convert(param->type());
                 auto id = fn->parameter(param_t.id);
+                defs_[param] = id;
                 fn->params[param] = id;
                 if (param->type()->isa<PtrType>()) {
                     builder_->decorate(id, spv::DecorationAliased);
@@ -232,13 +234,16 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
         for (auto param : cont->params()) {
             if (is_mem(param) || is_unit(param)) {
                 // Nothing
+                defs_[param] = 0;
             } else {
                 // OpPhi requires the full list of predecessors (values, labels)
                 // We don't have that yet! But we will need the Phi node identifier to build the basic blocks ...
                 // To solve this we generate an id for the phi node now, but defer emission of it to a later stage
                 auto type = convert(param->type()).id;
                 assert(type != 0);
-                bb.phis_map[param] = { type, builder_->generate_fresh_id(), {} };
+                auto id = builder_->generate_fresh_id();
+                defs_[param] = id;
+                bb.phis_map[param] = { type, id, {} };
             }
         }
     }

From aad02da1cc962ad4ce77aeb4c42361fed1790355 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 15:13:51 +0200
Subject: [PATCH 240/342] spirv: deal with untangible args properly

---
 src/thorin/be/spirv/spirv.cpp | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index d0ac89908..74ca3eadd 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -315,8 +315,11 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         int index = -1;
         for (auto& arg : app.args()) {
             index++;
+            if (is_mem(arg) || is_unit(arg)) {
+                emit_unsafe(arg);
+                continue;
+            }
             auto val = emit(arg);
-            if (is_mem(arg) || is_unit(arg)) continue;
             bb->args[arg] = val;
             auto* param = dst_cont->param(index);
             auto& phi = cont2bb_[dst_cont]->phis_map[param];
@@ -362,8 +365,11 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         for (auto arg : app.args()) {
             if (arg->order() == 0) {
                 auto arg_type = arg->type();
+                if (arg_type == world().unit_type() || arg_type == world().mem_type()) {
+                    emit_unsafe(arg);
+                    continue;
+                }
                 auto arg_val = emit(arg);
-                if (arg_type == world().unit_type() || arg_type == world().mem_type()) continue;
                 call_args.push_back(arg_val);
             } else {
                 assert(!ret_arg);

From abf4b6a84ed3e2006114340d97ca34b838a3bce0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 15:14:08 +0200
Subject: [PATCH 241/342] spirv: emit_bb redirects to emit_constant

---
 src/thorin/be/spirv/spirv.cpp | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 74ca3eadd..51f41c2b5 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -300,9 +300,11 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
 
         for (auto arg : app.args()) {
             assert(arg->order() == 0);
-            auto val = emit(arg);
-            if (is_mem(arg) || is_unit(arg))
+            if (is_mem(arg) || is_unit(arg)) {
+                emit_unsafe(arg);
                 continue;
+            }
+            auto val = emit(arg);
             values.emplace_back(val);
         }
 
@@ -472,7 +474,9 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
             assert(val != 0);
             return val;
         }
-    } else return emit_bb(nullptr, def);
+    }
+
+    assertf(false, "Incomplete emit(def) definition");
 }
 
 SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
@@ -815,6 +819,10 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     } else if (def->isa<Bottom>()) {
         return bb->undef(convert(def->type()).id);
     }
+
+    if (!def->has_dep(Dep::Param))
+        return emit_constant(def);
+
     assertf(false, "Incomplete emit(def) definition");
 }
 

From ae4207520fb0e19dfb457a887435a93fc46ed546 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 15:18:35 +0200
Subject: [PATCH 242/342] spirv: deal with constants of all primtypes

---
 src/thorin/be/spirv/spirv.cpp | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 51f41c2b5..f06639b3f 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -445,11 +445,13 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
         SpvId constant;
         switch (primlit->primtype_tag()) {
             case PrimType_bool:                     constant = builder_->bool_constant(type, box.get_bool()); break;
-            case PrimType_ps8:  case PrimType_qs8:  assertf(false, "not implemented yet");
-            case PrimType_pu8:  case PrimType_qu8:  assertf(false, "not implemented yet");
-            case PrimType_ps16: case PrimType_qs16: assertf(false, "not implemented yet");
-            case PrimType_pu16: case PrimType_qu16: assertf(false, "not implemented yet");
-            case PrimType_ps32: case PrimType_qs32: constant = builder_->constant(type, { static_cast<unsigned int>(box.get_s32()) }); break;
+            case PrimType_ps8:  case PrimType_qs8:
+            case PrimType_pu8:  case PrimType_qu8:  constant = builder_->constant(type, { static_cast<unsigned int>(box.get_u8()) }); break;
+            case PrimType_ps16: case PrimType_qs16:
+            case PrimType_pu16: case PrimType_qu16:
+            case PrimType_pf16: case PrimType_qf16: constant = builder_->constant(type, { static_cast<unsigned int>(box.get_u16()) }); break;
+            case PrimType_pf32: case PrimType_qf32:
+            case PrimType_ps32: case PrimType_qs32:
             case PrimType_pu32: case PrimType_qu32: constant = builder_->constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
             case PrimType_ps64: case PrimType_qs64:
             case PrimType_pu64: case PrimType_qu64: {
@@ -459,8 +461,6 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
                 constant = builder_->constant(type, { (uint32_t) lower, (uint32_t) upper });
                 break;
             }
-            case PrimType_pf16: case PrimType_qf16: assertf(false, "not implemented yet");
-            case PrimType_pf32: case PrimType_qf32: assertf(false, "not implemented yet");
             case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
         }
         return constant;

From 6daf1eac7483e5d348558f9e66c6b03e83838f88 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 16:35:26 +0200
Subject: [PATCH 243/342] spirv_builder: improved extended instructions
 handling

---
 src/thorin/be/spirv/spirv.cpp         |  3 +-
 src/thorin/be/spirv/spirv_builder.hpp | 48 +++++++++++++++++++--------
 2 files changed, 36 insertions(+), 15 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index f06639b3f..3ec8b4139 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -847,7 +847,8 @@ std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* bui
             args.push_back(emit(app.arg(i)));
         }
 
-        bb->ext_instruction(convert(world().unit_type()).id, builder_->imported_instrs->shader_printf, 1, args);
+        builder_->extension("SPV_KHR_non_semantic_info");
+        bb->ext_instruction(convert(world().unit_type()).id, { "NonSemantic.DebugPrintf", 1}, args);
     } else if (builtin->name() == "get_work_dim") {
         THORIN_UNREACHABLE;
     } else if (builtin->name() == "get_global_id") {
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 6a600b7a6..f110cc47f 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -17,6 +17,11 @@ struct SpvBasicBlockBuilder;
 struct SpvFnBuilder;
 struct SpvFileBuilder;
 
+struct ExtendedInstruction {
+    const char* set_name;
+    uint32_t id;
+};
+
 inline int div_roundup(int a, int b) {
     if (a % b == 0)
         return a / b;
@@ -242,17 +247,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
-        op(spv::Op::OpExtInst, 5 + arguments.size());
-        auto id = generate_fresh_id();
-        ref_id(return_type);
-        ref_id(id);
-        ref_id(set);
-        literal_int(instruction);
-        for (auto a : arguments)
-            ref_id(a);
-        return id;
-    }
+    SpvId ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments);
 
     void return_void() {
         op(spv::Op::OpReturn, 1);
@@ -269,6 +264,19 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
 private:
     SpvId generate_fresh_id();
+
+protected:
+    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
+        op(spv::Op::OpExtInst, 5 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(set);
+        literal_int(instruction);
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
+    }
 };
 
 struct SpvFnBuilder {
@@ -581,17 +589,22 @@ struct SpvFileBuilder {
         extensions.literal_name(name);
     }
 
+    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
+
+protected:
     SpvId extended_import(std::string name) {
+        auto found = extended_instruction_sets.find(name);
+        if (found != extended_instruction_sets.end())
+            return found->second;
         ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
         auto id = generate_fresh_id();
         ext_inst_import.ref_id(id);
         ext_inst_import.literal_name(name);
+        extended_instruction_sets[name] = id;
         return id;
     }
 
-    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
-    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
-
 private:
     std::ostream* output_ = nullptr;
     uint32_t bound = 1;
@@ -612,6 +625,7 @@ struct SpvFileBuilder {
 
     // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
     std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
+    std::unordered_map<std::string, SpvId> extended_instruction_sets;
 
     void output_word_le(uint32_t word) {
         output_->put((word >> 0) & 0xFFu);
@@ -653,6 +667,8 @@ struct SpvFileBuilder {
         output_section(fn_decls);
         output_section(fn_defs);
     }
+
+    friend SpvBasicBlockBuilder;
 };
 
 inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
@@ -663,4 +679,8 @@ inline SpvId SpvFnBuilder::generate_fresh_id() {
     return file_builder->generate_fresh_id();
 }
 
+SpvId SpvBasicBlockBuilder::ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments) {
+    return ext_instruction(return_type, file_builder.extended_import(instr.set_name), instr.id, arguments);
+}
+
 }

From 19c0f2f77f1f59b70af6e158c10876cfff38504f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 16:38:32 +0200
Subject: [PATCH 244/342] spirv: cache declared capabilities and extensions

---
 src/thorin/be/spirv/spirv_builder.hpp | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index f110cc47f..c8886979e 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -4,6 +4,7 @@
 #include <cassert>
 #include <vector>
 #include <unordered_map>
+#include <unordered_set>
 #include <ostream>
 #include <memory>
 
@@ -580,13 +581,21 @@ struct SpvFileBuilder {
     }
 
     void capability(spv::Capability cap) {
+        auto found = capabilities_set.find(cap);
+        if (found != capabilities_set.end())
+            return;
         capabilities.op(spv::Op::OpCapability, 2);
         capabilities.data_.push_back(cap);
+        capabilities_set.insert(cap);
     }
 
     void extension(std::string name) {
+        auto found = extensions_set.find(name);
+        if (found != extensions_set.end())
+            return;
         extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
         extensions.literal_name(name);
+        extensions_set.insert(name);
     }
 
     spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
@@ -626,6 +635,8 @@ struct SpvFileBuilder {
     // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
     std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
     std::unordered_map<std::string, SpvId> extended_instruction_sets;
+    std::unordered_set<spv::Capability> capabilities_set;
+    std::unordered_set<std::string> extensions_set;
 
     void output_word_le(uint32_t word) {
         output_->put((word >> 0) & 0xFFu);

From 9ada962929fb9bc07ff65773cd7d1dc1e6c2d475 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:05:39 +0200
Subject: [PATCH 245/342] spirv: fix builder multiple definition errors

---
 src/thorin/be/spirv/spirv_builder.hpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index c8886979e..2e741f2fe 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -690,7 +690,7 @@ inline SpvId SpvFnBuilder::generate_fresh_id() {
     return file_builder->generate_fresh_id();
 }
 
-SpvId SpvBasicBlockBuilder::ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments) {
+inline SpvId SpvBasicBlockBuilder::ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments) {
     return ext_instruction(return_type, file_builder.extended_import(instr.set_name), instr.id, arguments);
 }
 

From a1509ab36c21f757d9ec3218517217b56493ae7e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:06:05 +0200
Subject: [PATCH 246/342] spirv: remove ImportedInstructions

---
 src/thorin/be/spirv/spirv.cpp       | 6 ------
 src/thorin/be/spirv/spirv.h         | 6 ------
 src/thorin/be/spirv/spirv_private.h | 1 -
 3 files changed, 13 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3ec8b4139..ed5531bee 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -137,11 +137,6 @@ Builtins::Builtins(FileBuilder& builder) {
     builder.name(local_invocation_index, "BuiltInLocalInvocationIndex");
 }
 
-ImportedInstructions::ImportedInstructions(FileBuilder& builder) {
-    builder.extension("SPV_KHR_non_semantic_info");
-    shader_printf = builder.extended_import("NonSemantic.DebugPrintf");
-}
-
 CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, bool debug, const Cont2Config* kernel_config)
         : thorin::CodeGen(thorin, debug), target_info_(target_info), kernel_config_(kernel_config)
 {}
@@ -151,7 +146,6 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_ = &builder;
 
     builder_->builtins = std::make_unique<Builtins>(*builder_);
-    builder_->imported_instrs = std::make_unique<ImportedInstructions>(*builder_);
 
     ScopesForest forest(world());
     forest.for_each([&](const Scope& scope) { emit_scope(scope, forest); });
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 05d27f628..bf93c1c0c 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -45,12 +45,6 @@ struct Builtins {
     explicit Builtins(FileBuilder&);
 };
 
-struct ImportedInstructions {
-    SpvId shader_printf;
-
-    explicit ImportedInstructions(FileBuilder&);
-};
-
 struct BasicBlockBuilder;
 
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index 745b1ea0d..9af46295b 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -35,7 +35,6 @@ struct FileBuilder : public builder::SpvFileBuilder {
     CodeGen* cg;
 
     std::unique_ptr<Builtins> builtins;
-    std::unique_ptr<ImportedInstructions> imported_instrs;
 
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;

From 2566aecc3c1f105001a69c2cbf396f78d746dd08 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:08:27 +0200
Subject: [PATCH 247/342] spirv: initial support for MathOps

---
 src/thorin/CMakeLists.txt                  |  1 +
 src/thorin/be/spirv/spirv.cpp              |  3 ++
 src/thorin/be/spirv/spirv.h                |  2 +
 src/thorin/be/spirv/spirv_instructions.cpp | 55 ++++++++++++++++++++++
 4 files changed, 61 insertions(+)
 create mode 100644 src/thorin/be/spirv/spirv_instructions.cpp

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 201d73665..01044d98f 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -128,6 +128,7 @@ if(THORIN_ENABLE_SPIRV)
     list(APPEND THORIN_SOURCES
         be/spirv/spirv.cpp
         be/spirv/spirv_types.cpp
+        be/spirv/spirv_instructions.cpp
         be/spirv/spirv.h
     )
 endif()
diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index ed5531bee..17c7d414f 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -474,6 +474,9 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
 }
 
 SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
+    if (auto mathop = def->isa<MathOp>())
+        return emit_mathop(bb, *mathop);
+
     if (auto bin = def->isa<BinOp>()) {
         SpvId lhs = emit(bin->lhs());
         SpvId rhs = emit(bin->rhs());
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index bf93c1c0c..de7026bd8 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -74,6 +74,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     FnBuilder& get_fn_builder(Continuation*);
     std::vector<SpvId> emit_builtin(const App&, const Continuation*, BasicBlockBuilder*);
 
+    SpvId emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
+
     SpvId get_codom_type(const Continuation* fn);
 
     SpvTargetInfo target_info_;
diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
new file mode 100644
index 000000000..509179e7b
--- /dev/null
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -0,0 +1,55 @@
+#include "spirv_private.h"
+
+namespace thorin::spirv {
+
+struct SpirMathOps {
+#define THORIN_MATHOP(mathop_name) builder::ExtendedInstruction mathop_name;
+#include "thorin/tables/mathoptable.h"
+#undef THORIN_MATHOP
+};
+
+#include <spirv/unified1/OpenCL.std.h>
+
+SpirMathOps opencl_std = {
+    .fabs =     { "OpenCL.std", OpenCLLIB::Fabs },
+    .copysign = { "OpenCL.std", OpenCLLIB::Sign },
+    .round =    { "OpenCL.std", OpenCLLIB::Round },
+    .floor =    { "OpenCL.std", OpenCLLIB::Floor },
+    .ceil =     { "OpenCL.std", OpenCLLIB::Ceil },
+    .fmin =     { "OpenCL.std", OpenCLLIB::Fmin },
+    .fmax =     { "OpenCL.std", OpenCLLIB::Fmax },
+    .cos =      { "OpenCL.std", OpenCLLIB::Cos },
+    .sin =      { "OpenCL.std", OpenCLLIB::Sin },
+    .tan =      { "OpenCL.std", OpenCLLIB::Tan },
+    .acos =     { "OpenCL.std", OpenCLLIB::Acos },
+    .asin =     { "OpenCL.std", OpenCLLIB::Asin },
+    .atan =     { "OpenCL.std", OpenCLLIB::Atan },
+    .atan2 =    { "OpenCL.std", OpenCLLIB::Atan2 },
+    .sqrt =     { "OpenCL.std", OpenCLLIB::Sqrt },
+    .cbrt =     { "OpenCL.std", OpenCLLIB::Cbrt },
+    .pow =      { "OpenCL.std", OpenCLLIB::Pow },
+    .exp =      { "OpenCL.std", OpenCLLIB::Exp },
+    .exp2 =     { "OpenCL.std", OpenCLLIB::Exp2 },
+    .log =      { "OpenCL.std", OpenCLLIB::Log },
+    .log2 =     { "OpenCL.std", OpenCLLIB::Log2 },
+    .log10 =    { "OpenCL.std", OpenCLLIB::Log10 },
+};
+
+SpvId CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
+    auto type = mathop.type();
+
+    SpirMathOps& impl = opencl_std;
+    std::vector<SpvId> ops;
+    for (auto& op : mathop.ops()) {
+        ops.push_back(emit(op));
+    }
+
+    if (is_type_f(type)) {
+        switch (mathop.mathop_tag()) {
+#define THORIN_MATHOP(mathop_name) case MathOp_##mathop_name: return bb->ext_instruction(convert(type).id, impl.mathop_name, ops);
+#include "thorin/tables/mathoptable.h"
+        }
+    }
+}
+
+}
\ No newline at end of file

From df2db758ed5934478f8071beff183f7a0e164d05 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:28:01 +0200
Subject: [PATCH 248/342] spirv: fix OpenCL copysign

---
 src/thorin/be/spirv/spirv_instructions.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 509179e7b..4a9b53104 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -12,7 +12,7 @@ struct SpirMathOps {
 
 SpirMathOps opencl_std = {
     .fabs =     { "OpenCL.std", OpenCLLIB::Fabs },
-    .copysign = { "OpenCL.std", OpenCLLIB::Sign },
+    .copysign = { "OpenCL.std", OpenCLLIB::Copysign },
     .round =    { "OpenCL.std", OpenCLLIB::Round },
     .floor =    { "OpenCL.std", OpenCLLIB::Floor },
     .ceil =     { "OpenCL.std", OpenCLLIB::Ceil },

From 7028f0aab7622e91547536edf9ec6c3f60c239af Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:28:18 +0200
Subject: [PATCH 249/342] spirv: fix some more issues

---
 src/thorin/be/spirv/spirv.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 17c7d414f..65afc7ace 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -157,12 +157,12 @@ void CodeGen::emit_stream(std::ostream& out) {
     }
 
     for (auto& cont : world().copy_continuations()) {
-        assert(defs_.contains(cont));
         if (cont->is_exported() && kernel_config_) {
             auto config = kernel_config_->find(cont);
             if (config == kernel_config_->end())
                 continue;
 
+            assert(defs_.contains(cont));
             SpvId callee = defs_[cont];
 
             auto block = config->second->as<GPUKernelConfig>()->block_size();
@@ -193,18 +193,21 @@ FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
     auto& fn = *(builder_->fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
     fn.fn_type = convert(entry_->type()).id;
     fn.fn_ret_type = get_codom_type(entry_);
+    defs_[continuation] = fn.function_id;
     return fn;
 }
 
 FnBuilder* CodeGen::prepare(const thorin::Scope& scope) {
     auto& fn = get_fn_builder(scope.entry());
     builder_->name(fn.function_id, scope.entry()->name());
+    builder_->current_fn_ = &fn;
     return &fn;
 }
 
 void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
     auto& bb = *fn->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*fn));
     cont2bb_[cont] = &bb;
+    defs_[cont] = bb.label;
     fn->bbs_to_emit.emplace_back(&bb);
 
     builder_->name(bb.label, cont->name().c_str());
@@ -373,10 +376,9 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             }
         }
 
-        auto ret_type = get_codom_type(continuation);
-
         SpvId call_result;
         if (auto called_continuation = app.callee()->isa_nom<Continuation>()) {
+            auto ret_type = get_codom_type(called_continuation);
             call_result = bb->call(ret_type, emit(called_continuation), call_args);
         } else {
             // must be a closure

From 78ea186c708d8094a66892ac39a3240983211ac6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:28:54 +0200
Subject: [PATCH 250/342] spirv: cleanup

---
 src/thorin/be/spirv/spirv.cpp | 10 ----------
 1 file changed, 10 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 65afc7ace..9c090429f 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -460,16 +460,6 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
             case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
         }
         return constant;
-    } else if (auto param = def->isa<Param>()) {
-        if (is_mem(param)) return spv_none;
-        if (auto param_id = builder_->current_fn_->params.lookup(param)) {
-            assert((*param_id) != 0);
-            return *param_id;
-        } else {
-            auto val = cont2bb_[param->continuation()]->phis_map[param].value;
-            assert(val != 0);
-            return val;
-        }
     }
 
     assertf(false, "Incomplete emit(def) definition");

From d72883099ca5534f372f8c20cc9383f1864f3181 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:33:32 +0200
Subject: [PATCH 251/342] spirv: don't use labels as values.

---
 src/thorin/be/spirv/spirv.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 9c090429f..8374c570c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -207,7 +207,6 @@ FnBuilder* CodeGen::prepare(const thorin::Scope& scope) {
 void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
     auto& bb = *fn->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*fn));
     cont2bb_[cont] = &bb;
-    defs_[cont] = bb.label;
     fn->bbs_to_emit.emplace_back(&bb);
 
     builder_->name(bb.label, cont->name().c_str());

From 74ad6e49a93a0fa808ef037b471f402011b35230 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 17:41:39 +0200
Subject: [PATCH 252/342] spirv: fix labels

---
 src/thorin/be/spirv/spirv.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 8374c570c..f88a396f7 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -227,6 +227,7 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
             }
         }
     } else {
+        defs_[cont] = bb.label;
         for (auto param : cont->params()) {
             if (is_mem(param) || is_unit(param)) {
                 // Nothing

From 0cb78ef8cd7918af3f377a7038b5fed0c7ef1ffe Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 18:09:08 +0200
Subject: [PATCH 253/342] spirv: always use labels for OpPhi

---
 src/thorin/be/spirv/spirv.cpp | 13 +++++++++----
 src/thorin/be/spirv/spirv.h   |  1 +
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index f88a396f7..542a7a2d5 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -274,6 +274,11 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
     return builder_->declare_struct_type(types);
 }
 
+SpvId CodeGen::emit_as_bb(thorin::Continuation* cont) {
+    emit(cont);
+    return cont2bb_[cont]->label;
+}
+
 void CodeGen::emit_epilogue(Continuation* continuation) {
     auto& bb = cont2bb_[continuation];
     // Handles the potential nuances of jumping to another continuation
@@ -285,7 +290,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             if (is_mem(param) || is_unit(param))
                 continue;
             auto& phi = cont2bb_[succ]->phis_map[param];
-            phi.preds.emplace_back(args[j], emit(continuation));
+            phi.preds.emplace_back(args[j], emit_as_bb(continuation));
             j++;
         }
     };
@@ -322,7 +327,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             bb->args[arg] = val;
             auto* param = dst_cont->param(index);
             auto& phi = cont2bb_[dst_cont]->phis_map[param];
-            phi.preds.emplace_back(bb->args[arg], emit(continuation));
+            phi.preds.emplace_back(bb->args[arg], emit_as_bb(continuation));
         }
         bb->branch(emit(dst_cont));
     } else if (app.callee() == world().branch()) {
@@ -407,7 +412,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             bb->branch(emit(succ));
 
             auto& phi = cont2bb_[succ]->phis_map[last_param];
-            phi.preds.emplace_back(call_result, emit(continuation));
+            phi.preds.emplace_back(call_result, emit_as_bb(continuation));
         } else {
             Array<SpvId> extracts(n);
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
@@ -426,7 +431,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
                     continue;
 
                 auto& phi = cont2bb_[succ]->phis_map[last_param];
-                phi.preds.emplace_back(extracts[j], emit(continuation));
+                phi.preds.emplace_back(extracts[j], emit_as_bb(continuation));
 
                 j++;
             }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index de7026bd8..d3121dda2 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -74,6 +74,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     FnBuilder& get_fn_builder(Continuation*);
     std::vector<SpvId> emit_builtin(const App&, const Continuation*, BasicBlockBuilder*);
 
+    SpvId emit_as_bb(Continuation*);
     SpvId emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
 
     SpvId get_codom_type(const Continuation* fn);

From 93da33228f77f518a4a84e1023c8fdbec9df9820 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 18:09:32 +0200
Subject: [PATCH 254/342] spirv: deal with vectors properly

---
 src/thorin/be/spirv/spirv_types.cpp | 21 +++++++++------------
 1 file changed, 9 insertions(+), 12 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 51bb71e8c..e73fd83e4 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -20,9 +20,17 @@ ConvertedType CodeGen::convert(const Type* type) {
     if (auto iter = types_.find(type); iter != types_.end())
         return iter->second;
 
+    // Vector types are stupid and dangerous!
+
     ConvertedType converted = { 0, std::nullopt };
 
-    switch (type->tag()) {
+    if (auto vec = type->isa<VectorType>(); vec && vec->length() > 1) {
+        auto component = vec->scalarize();
+        auto conv_comp = convert(component);
+        converted.id = builder_->declare_vector_type(conv_comp.id, (uint32_t) vec->length());
+        converted.layout = conv_comp.layout;
+        converted.layout->size *= vec->length();
+    } else switch (type->tag()) {
         // Boolean types are typically packed intelligently when declaring in local variables, however with vanilla Vulkan 1.0 they can only be represented via 32-bit integers
         // Using extensions, we could use 16 or 8-bit ints instead
         // We can also pack them inside structures using bit-twiddling tricks, if the need arises
@@ -143,17 +151,6 @@ ConvertedType CodeGen::convert(const Type* type) {
             break;
         }
 
-        case Node_Vector: {
-            auto vec = type->as<VectorType>();
-            assert(vec->length() > 1);
-            auto component = vec->scalarize();
-            auto conv_comp = convert(component);
-            converted.id = builder_->declare_vector_type(conv_comp.id, (uint32_t)vec->length());
-            converted.layout = conv_comp.layout;
-            converted.layout->size *= vec->length();
-            break;
-        }
-
         case Node_StructType:
         case Node_TupleType: {
             std::vector<SpvId> spv_types;

From b56a3da7646b21732f0dd698f150f6dd07e86e3c Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 18:12:42 +0200
Subject: [PATCH 255/342] spirv: declare Linkage capability when without entry
 points

---
 src/thorin/be/spirv/spirv.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 542a7a2d5..50e816a15 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -156,6 +156,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             interface.push_back(emit(global));
     }
 
+    int entry_points_count = 0;
     for (auto& cont : world().copy_continuations()) {
         if (cont->is_exported() && kernel_config_) {
             auto config = kernel_config_->find(cont);
@@ -174,9 +175,14 @@ void CodeGen::emit_stream(std::ostream& out) {
 
             builder_->declare_entry_point(spv::ExecutionModelGLCompute, callee, cont->name().c_str(), interface);
             builder_->execution_mode(callee, spv::ExecutionModeLocalSize, local_size);
+            entry_points_count++;
         }
     }
 
+    if (entry_points_count == 0) {
+        builder_->capability(spv::Capability::CapabilityLinkage);
+    }
+
     builder_->finish(out);
     builder_ = nullptr;
 }

From 99415933971e50c04391a0fc56ae7d72eaa802d3 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 18:30:39 +0200
Subject: [PATCH 256/342] declare capabilities lazily, based on dialect

---
 src/thorin/be/spirv/spirv.cpp       | 22 ++++++++++++++--------
 src/thorin/be/spirv/spirv.h         |  6 ++++--
 src/thorin/be/spirv/spirv_types.cpp | 13 ++++++++++++-
 3 files changed, 30 insertions(+), 11 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 50e816a15..86688633b 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -83,14 +83,6 @@ BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
 FnBuilder::FnBuilder(FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), file_builder(file_builder) {}
 
 FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
-    capability(spv::Capability::CapabilityShader);
-    capability(spv::Capability::CapabilityVariablePointers);
-    capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
-    // capability(spv::Capability::CapabilityInt16);
-    capability(spv::Capability::CapabilityInt64);
-
-    addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
-    memory_model = spv::MemoryModel::MemoryModelGLSL450;
 }
 
 SpvId FileBuilder::u32_t() {
@@ -145,6 +137,20 @@ void CodeGen::emit_stream(std::ostream& out) {
     FileBuilder builder(this);
     builder_ = &builder;
 
+    switch (target_info_.dialect) {
+        case SpvTargetInfo::OpenCL:
+            builder_->capability(spv::Capability::CapabilityKernel);
+            builder_->capability(spv::Capability::CapabilityAddresses);
+            builder_->addressing_model = target_info_.mem_layout.pointer_size == 4 ? spv::AddressingModelPhysical32 : spv::AddressingModelPhysical64;
+            builder_->memory_model = spv::MemoryModel::MemoryModelOpenCL;
+            break;
+        case SpvTargetInfo::Vulkan:
+            builder_->capability(spv::Capability::CapabilityShader);
+            builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
+            builder_->memory_model = spv::MemoryModel::MemoryModelGLSL450;
+            break;
+    }
+
     builder_->builtins = std::make_unique<Builtins>(*builder_);
 
     ScopesForest forest(world());
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index d3121dda2..e12abe270 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -17,13 +17,15 @@ struct FnBuilder;
 struct SpvTargetInfo {
     struct {
         // Either '4' or '8'
-        size_t pointer_size;
+        size_t pointer_size = 8;
     } mem_layout;
 
     enum Dialect{
         OpenCL,
-        Shady
+        Vulkan
     };
+
+    Dialect dialect = OpenCL;
 };
 
 struct ConvertedType {
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index e73fd83e4..c75adf9ee 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -48,10 +48,12 @@ ConvertedType CodeGen::convert(const Type* type) {
             converted.layout = { 1, 1 };
             break;
         case Node_PrimType_ps16:
+            builder_->capability(spv::Capability::CapabilityInt16);
             converted.id = builder_->declare_int_type(16, true);
             converted.layout = { 2, 2 };
             break;
         case Node_PrimType_pu16:
+            builder_->capability(spv::Capability::CapabilityInt16);
             converted.id = builder_->declare_int_type(16, false);
             converted.layout = { 2, 2 };
             break;
@@ -64,10 +66,12 @@ ConvertedType CodeGen::convert(const Type* type) {
             converted.layout = { 4, 4 };
             break;
         case Node_PrimType_ps64:
+            builder_->capability(spv::Capability::CapabilityInt64);
             converted.id = builder_->declare_int_type(64, true);
             converted.layout = { 8, 8 };
             break;
         case Node_PrimType_pu64:
+            builder_->capability(spv::Capability::CapabilityInt64);
             converted.id = builder_->declare_int_type(64, false);
             converted.layout = { 8, 8 };
             break;
@@ -90,8 +94,15 @@ ConvertedType CodeGen::convert(const Type* type) {
                 case AddrSpace::Function: storage_class = spv::StorageClassFunction;              break;
                 case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;               break;
                 case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
-                case AddrSpace::Global:   storage_class = spv::StorageClassCrossWorkgroup;        break;
                 case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
+                case AddrSpace::Global: {
+                    if (target_info_.dialect == SpvTargetInfo::Vulkan) {
+                        builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
+                        storage_class = spv::StorageClassPhysicalStorageBuffer;
+                    } else
+                        storage_class = spv::StorageClassCrossWorkgroup;
+                    break;
+                }
                 default:
                     assert(false && "This address space is not supported");
                     break;

From 6e6777988b1d8b71a6bdddc445655e810a13b192 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 19:16:52 +0200
Subject: [PATCH 257/342] spirv: remove old Builtins code

---
 src/thorin/be/spirv/spirv.cpp       | 69 ++---------------------------
 src/thorin/be/spirv/spirv.h         | 13 +-----
 src/thorin/be/spirv/spirv_private.h |  2 -
 3 files changed, 4 insertions(+), 80 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 86688633b..236fb915b 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -95,40 +95,6 @@ SpvId FileBuilder::u32_constant(uint32_t pattern) {
     return constant(u32_t(), { pattern });
 }
 
-Builtins::Builtins(FileBuilder& builder) {
-    auto& world = builder.cg->world();
-    auto spv_uvec3_t = builder.cg->convert(world.type_pu32(3));
-    auto spv_uint_t = builder.cg->convert(world.type_pu32());
-    auto spv_uvec3_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uvec3_t.id);
-    auto spv_uvec3_ptp = builder.declare_ptr_type(spv::StorageClassPrivate, spv_uvec3_t.id);
-    auto spv_uint_pt = builder.declare_ptr_type(spv::StorageClassInput, spv_uint_t.id);
-
-    // Because we technically can have multiple entry points, we take the easy way out and make each entry point
-    // write to a private variable the actual workgroup size for that specific kernel. Dirty, but simple.
-    workgroup_size = builder.variable(spv_uvec3_ptp, spv::StorageClassPrivate);
-    builder.name(workgroup_size, "BuiltInWorkgroupSize");
-
-    num_workgroups = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(num_workgroups, spv::DecorationBuiltIn, { spv::BuiltInNumWorkgroups });
-    builder.name(num_workgroups, "BuiltInNumWorkgroups");
-
-    workgroup_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(workgroup_id, spv::DecorationBuiltIn, { spv::BuiltInWorkgroupId });
-    builder.name(workgroup_id, "BuiltInWorkgroupId");
-
-    local_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(local_id, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationId });
-    builder.name(local_id, "BuiltInLocalInvocationId");
-
-    global_id = builder.variable(spv_uvec3_pt, spv::StorageClassInput);
-    builder.decorate(global_id, spv::DecorationBuiltIn, { spv::BuiltInGlobalInvocationId });
-    builder.name(global_id, "BuiltInGlobalInvocationId");
-
-    local_invocation_index = builder.variable(spv_uint_pt, spv::StorageClassInput);
-    builder.decorate(local_invocation_index, spv::DecorationBuiltIn, { spv::BuiltInLocalInvocationIndex });
-    builder.name(local_invocation_index, "BuiltInLocalInvocationIndex");
-}
-
 CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, bool debug, const Cont2Config* kernel_config)
         : thorin::CodeGen(thorin, debug), target_info_(target_info), kernel_config_(kernel_config)
 {}
@@ -151,8 +117,6 @@ void CodeGen::emit_stream(std::ostream& out) {
             break;
     }
 
-    builder_->builtins = std::make_unique<Builtins>(*builder_);
-
     ScopesForest forest(world());
     forest.for_each([&](const Scope& scope) { emit_scope(scope, forest); });
 
@@ -364,16 +328,14 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         THORIN_UNREACHABLE;
     } else if (app.callee()->isa<Bottom>()) {
         bb->unreachable();
-    } else if (auto builtin = app.callee()->isa_nom<Continuation>(); builtin->is_imported()) {
+    } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
         // Ensure we emit previous memory operations
         assert(is_mem(app.arg(0)));
         emit(app.arg(0));
 
-        auto productions = emit_builtin(app, builtin, bb);
+        auto productions = emit_intrinsic(app, intrinsic, bb);
         auto succ = app.args().back()->isa_nom<Continuation>();
         jump_to_next_cont_with_args(succ, productions);
-    } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
-        THORIN_UNREACHABLE;
     } else { // function/closure call
         // put all first-order args into an array
         std::vector<SpvId> call_args;
@@ -832,11 +794,8 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* builtin, BasicBlockBuilder* bb) {
+std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* builtin, BasicBlockBuilder* bb) {
     std::vector<SpvId> productions;
-    auto uvec3_t = convert(world().type_pu32(3)).id;
-    auto u32_t = convert(world().type_pu32()).id;
-    auto i32_t = convert(world().type_ps32()).id;
     if (builtin->name() == "spirv.nonsemantic.printf") {
         std::vector<SpvId> args;
         auto string = app.arg(1);
@@ -855,28 +814,6 @@ std::vector<SpvId> CodeGen::emit_builtin(const App& app, const Continuation* bui
 
         builder_->extension("SPV_KHR_non_semantic_info");
         bb->ext_instruction(convert(world().unit_type()).id, { "NonSemantic.DebugPrintf", 1}, args);
-    } else if (builtin->name() == "get_work_dim") {
-        THORIN_UNREACHABLE;
-    } else if (builtin->name() == "get_global_id") {
-        auto vector = bb->load(uvec3_t, builder_->builtins->global_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
-    } else if (builtin->name() == "get_local_size") {
-        auto vector = bb->load(uvec3_t, builder_->builtins->workgroup_size);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
-    } else if (builtin->name() == "get_local_id") {
-        auto vector = bb->load(uvec3_t, builder_->builtins->local_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
-    } else if (builtin->name() == "get_num_groups") {
-        auto vector = bb->load(uvec3_t, builder_->builtins->num_workgroups);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
-    } else if (builtin->name() == "get_group_id") {
-        auto vector = bb->load(uvec3_t, builder_->builtins->workgroup_id);
-        auto extracted = bb->vector_extract_dynamic(u32_t, vector, emit(app.arg(1)));
-        productions.push_back(bb->convert(spv::OpBitcast, i32_t, extracted));
     } else {
         world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
     }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index e12abe270..9154761a3 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -36,17 +36,6 @@ struct ConvertedType {
     std::optional<Layout> layout;
 };
 
-struct Builtins {
-    SpvId workgroup_size;
-    SpvId num_workgroups;
-    SpvId workgroup_id;
-    SpvId local_id;
-    SpvId global_id;
-    SpvId local_invocation_index;
-
-    explicit Builtins(FileBuilder&);
-};
-
 struct BasicBlockBuilder;
 
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
@@ -74,7 +63,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     SpvId emit_bb(BasicBlockBuilder* bb, const Def* def);
 protected:
     FnBuilder& get_fn_builder(Continuation*);
-    std::vector<SpvId> emit_builtin(const App&, const Continuation*, BasicBlockBuilder*);
+    std::vector<SpvId> emit_intrinsic(const App& app, const Continuation* builtin, BasicBlockBuilder* bb);
 
     SpvId emit_as_bb(Continuation*);
     SpvId emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index 9af46295b..62bf73368 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -34,8 +34,6 @@ struct FileBuilder : public builder::SpvFileBuilder {
 
     CodeGen* cg;
 
-    std::unique_ptr<Builtins> builtins;
-
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
 

From 26669681a83708ac624f17d7938b64177ecdda43 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 1 Jul 2024 19:17:05 +0200
Subject: [PATCH 258/342] spirv: use VectorComputeINTEL for Private

---
 src/thorin/be/spirv/spirv_types.cpp | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index c75adf9ee..4e7e755ba 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -92,7 +92,12 @@ ConvertedType CodeGen::convert(const Type* type) {
             spv::StorageClass storage_class;
             switch (ptr->addr_space()) {
                 case AddrSpace::Function: storage_class = spv::StorageClassFunction;              break;
-                case AddrSpace::Private:  storage_class = spv::StorageClassPrivate;               break;
+                case AddrSpace::Private: {
+                    storage_class = spv::StorageClassPrivate;
+                    if (target_info_.dialect != SpvTargetInfo::Dialect::Vulkan)
+                        builder_->capability(spv::CapabilityVectorComputeINTEL);
+                    break;
+                }
                 case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
                 case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
                 case AddrSpace::Global: {

From a548bccb9d09aadff2906224741bd773defc5698 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 09:17:15 +0200
Subject: [PATCH 259/342] spirv: renamed Target

---
 src/thorin/be/spirv/spirv.cpp       |  8 +++++---
 src/thorin/be/spirv/spirv.h         | 10 ++++++----
 src/thorin/be/spirv/spirv_types.cpp |  4 ++--
 3 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 236fb915b..4d3d00414 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -95,7 +95,9 @@ SpvId FileBuilder::u32_constant(uint32_t pattern) {
     return constant(u32_t(), { pattern });
 }
 
-CodeGen::CodeGen(Thorin& thorin, SpvTargetInfo target_info, bool debug, const Cont2Config* kernel_config)
+
+
+CodeGen::CodeGen(Thorin& thorin, Target& target_info, bool debug, const Cont2Config* kernel_config)
         : thorin::CodeGen(thorin, debug), target_info_(target_info), kernel_config_(kernel_config)
 {}
 
@@ -104,13 +106,13 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_ = &builder;
 
     switch (target_info_.dialect) {
-        case SpvTargetInfo::OpenCL:
+        case Target::OpenCL:
             builder_->capability(spv::Capability::CapabilityKernel);
             builder_->capability(spv::Capability::CapabilityAddresses);
             builder_->addressing_model = target_info_.mem_layout.pointer_size == 4 ? spv::AddressingModelPhysical32 : spv::AddressingModelPhysical64;
             builder_->memory_model = spv::MemoryModel::MemoryModelOpenCL;
             break;
-        case SpvTargetInfo::Vulkan:
+        case Target::Vulkan:
             builder_->capability(spv::Capability::CapabilityShader);
             builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
             builder_->memory_model = spv::MemoryModel::MemoryModelGLSL450;
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 9154761a3..94530e475 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -14,13 +14,13 @@ class CodeGen;
 struct FileBuilder;
 struct FnBuilder;
 
-struct SpvTargetInfo {
+struct Target {
     struct {
         // Either '4' or '8'
         size_t pointer_size = 8;
     } mem_layout;
 
-    enum Dialect{
+    enum Dialect {
         OpenCL,
         Vulkan
     };
@@ -40,7 +40,7 @@ struct BasicBlockBuilder;
 
 class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
 public:
-    CodeGen(Thorin& thorin, SpvTargetInfo, bool debug, const Cont2Config* = nullptr);
+    CodeGen(Thorin& thorin, Target&, bool debug, const Cont2Config* = nullptr);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
@@ -70,9 +70,11 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
 
     SpvId get_codom_type(const Continuation* fn);
 
-    SpvTargetInfo target_info_;
+    Target& target_info_;
     FileBuilder* builder_;
     const Cont2Config* kernel_config_;
+
+    friend Target;
 };
 
 }
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 4e7e755ba..312727a5c 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -94,14 +94,14 @@ ConvertedType CodeGen::convert(const Type* type) {
                 case AddrSpace::Function: storage_class = spv::StorageClassFunction;              break;
                 case AddrSpace::Private: {
                     storage_class = spv::StorageClassPrivate;
-                    if (target_info_.dialect != SpvTargetInfo::Dialect::Vulkan)
+                    if (target_info_.dialect != Target::Dialect::Vulkan)
                         builder_->capability(spv::CapabilityVectorComputeINTEL);
                     break;
                 }
                 case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
                 case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
                 case AddrSpace::Global: {
-                    if (target_info_.dialect == SpvTargetInfo::Vulkan) {
+                    if (target_info_.dialect == Target::Dialect::Vulkan) {
                         builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
                         storage_class = spv::StorageClassPhysicalStorageBuffer;
                     } else

From bb32c8226639ebdc9105bbe4f183b323833a9907 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 10:17:25 +0200
Subject: [PATCH 260/342] spirv: opencl doesn't want signed integers

---
 src/thorin/be/spirv/spirv_types.cpp | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 312727a5c..797b20d04 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -17,6 +17,25 @@ ConvertedType CodeGen::convert(const Type* type) {
         default: break;
     }
 
+    // OpenCL has no signed integer types
+    if (target_info_.dialect == Target::OpenCL) {
+        switch (type->tag()) {
+            case Node_PrimType_ps8:
+                type = world().prim_type(PrimType_pu8, type->as<VectorType>()->length()); \
+                break;
+            case Node_PrimType_ps16:
+                type = world().prim_type(PrimType_pu16, type->as<VectorType>()->length()); \
+                break;
+            case Node_PrimType_ps32:
+                type = world().prim_type(PrimType_pu32, type->as<VectorType>()->length()); \
+                break;
+            case Node_PrimType_ps64:
+                type = world().prim_type(PrimType_pu64, type->as<VectorType>()->length()); \
+                break;
+            default: break;
+        }
+    }
+
     if (auto iter = types_.find(type); iter != types_.end())
         return iter->second;
 

From 97fecff4fbf7206f8820fa4478bb0a4269d9ddc5 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 13:27:25 +0200
Subject: [PATCH 261/342] spirv: fix some emission issues

---
 src/thorin/be/spirv/spirv.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 4d3d00414..ca4868451 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -330,10 +330,10 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         THORIN_UNREACHABLE;
     } else if (app.callee()->isa<Bottom>()) {
         bb->unreachable();
-    } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && intrinsic->is_intrinsic()) {
+    } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && (intrinsic->is_intrinsic() || intrinsic->cc() == CC::Device)) {
         // Ensure we emit previous memory operations
         assert(is_mem(app.arg(0)));
-        emit(app.arg(0));
+        emit_unsafe(app.arg(0));
 
         auto productions = emit_intrinsic(app, intrinsic, bb);
         auto succ = app.args().back()->isa_nom<Continuation>();
@@ -609,7 +609,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         return bb->composite(convert(structagg->type()).id, elements);
     } else if (auto access = def->isa<Access>()) {
         // emit dependent operations first
-        emit(access->mem());
+        emit_unsafe(access->mem());
 
         std::vector<uint32_t> operands;
         auto ptr_type = access->ptr()->type()->as<PtrType>();

From 14584d37866cd5fd84b0517d1d90b7d69ded577f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 13:27:52 +0200
Subject: [PATCH 262/342] spirv: add spirv.builtin intrinsic

---
 src/thorin/be/spirv/spirv.cpp       | 22 ++++++++++--
 src/thorin/be/spirv/spirv.h         |  3 +-
 src/thorin/be/spirv/spirv_private.h |  1 +
 src/thorin/be/spirv/spirv_types.cpp | 55 ++++++++++++++++-------------
 4 files changed, 52 insertions(+), 29 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index ca4868451..e926308e3 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -796,9 +796,9 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* builtin, BasicBlockBuilder* bb) {
+std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
     std::vector<SpvId> productions;
-    if (builtin->name() == "spirv.nonsemantic.printf") {
+    if (intrinsic->name() == "spirv.nonsemantic.printf") {
         std::vector<SpvId> args;
         auto string = app.arg(1);
         if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
@@ -816,8 +816,24 @@ std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* b
 
         builder_->extension("SPV_KHR_non_semantic_info");
         bb->ext_instruction(convert(world().unit_type()).id, { "NonSemantic.DebugPrintf", 1}, args);
+    } else if (intrinsic->name() == "spirv.builtin") {
+        if (auto spv_builtin_lit = app.arg(1)->isa<PrimLit>()) {
+            auto spv_builtin = spv_builtin_lit->value().get_u32();
+            auto found = builder_->builtins_.find(spv_builtin);
+            if (found != builder_->builtins_.end()) {
+                productions.push_back(found->second);
+            } else {
+                auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
+                auto desired_type = ret_type->types()[1]->as<PtrType>();
+                auto id = builder_->variable(convert(desired_type).id, static_cast<spv::StorageClass>(convert(desired_type->addr_space())));
+                builder_->decorate(id, spv::Decoration::DecorationBuiltIn, { spv_builtin });
+                builder_->builtins_[spv_builtin] = id;
+                productions.push_back(id);
+            }
+        } else
+            world().ELOG("spirv.builtin requires an integer literal as the argument");
     } else {
-        world().ELOG("This spir-v builtin isn't recognised: %s", builtin->name());
+        world().ELOG("This spir-v builtin isn't recognised: %s", intrinsic->name());
     }
     return productions;
 }
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 94530e475..22886eab8 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -49,6 +49,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
         return id > 0;
     }
 
+    uint32_t convert(AddrSpace);
     ConvertedType convert(const Type*);
 
     SpvId emit_fun_decl(Continuation*);
@@ -63,7 +64,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     SpvId emit_bb(BasicBlockBuilder* bb, const Def* def);
 protected:
     FnBuilder& get_fn_builder(Continuation*);
-    std::vector<SpvId> emit_intrinsic(const App& app, const Continuation* builtin, BasicBlockBuilder* bb);
+    std::vector<SpvId> emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb);
 
     SpvId emit_as_bb(Continuation*);
     SpvId emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index 62bf73368..ef1d72979 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -36,6 +36,7 @@ struct FileBuilder : public builder::SpvFileBuilder {
 
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
+    std::unordered_map<uint32_t, SpvId> builtins_;
 
     SpvId u32_t();
     SpvId u32_constant(uint32_t);
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 797b20d04..e29812081 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -4,6 +4,35 @@
 
 namespace thorin::spirv {
 
+uint32_t CodeGen::convert(AddrSpace as) {
+    spv::StorageClass storage_class;
+    switch (as) {
+        case AddrSpace::Function: storage_class = spv::StorageClassFunction;              break;
+        case AddrSpace::Private: {
+            storage_class = spv::StorageClassPrivate;
+            if (target_info_.dialect != Target::Dialect::Vulkan)
+                builder_->capability(spv::CapabilityVectorComputeINTEL);
+            break;
+        }
+        case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
+        case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
+        case AddrSpace::Input:    storage_class = spv::StorageClassInput;                 break;
+        case AddrSpace::Output:   storage_class = spv::StorageClassOutput;                break;
+        case AddrSpace::Global: {
+            if (target_info_.dialect == Target::Dialect::Vulkan) {
+                builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
+                storage_class = spv::StorageClassPhysicalStorageBuffer;
+            } else
+                storage_class = spv::StorageClassCrossWorkgroup;
+            break;
+        }
+        default:
+            assert(false && "This address space is not supported");
+            break;
+    }
+    return storage_class;
+}
+
 ConvertedType CodeGen::convert(const Type* type) {
     // Spir-V requires each primitive type to be "unique", it doesn't allow for example two 32-bit signed integer types.
     // Therefore we must enforce that precise/quick types map to the same thing.
@@ -108,34 +137,10 @@ ConvertedType CodeGen::convert(const Type* type) {
             break;
         case Node_PtrType: {
             auto ptr = type->as<PtrType>();
-            spv::StorageClass storage_class;
-            switch (ptr->addr_space()) {
-                case AddrSpace::Function: storage_class = spv::StorageClassFunction;              break;
-                case AddrSpace::Private: {
-                    storage_class = spv::StorageClassPrivate;
-                    if (target_info_.dialect != Target::Dialect::Vulkan)
-                        builder_->capability(spv::CapabilityVectorComputeINTEL);
-                    break;
-                }
-                case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
-                case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
-                case AddrSpace::Global: {
-                    if (target_info_.dialect == Target::Dialect::Vulkan) {
-                        builder_->capability(spv::Capability::CapabilityPhysicalStorageBufferAddresses);
-                        storage_class = spv::StorageClassPhysicalStorageBuffer;
-                    } else
-                        storage_class = spv::StorageClassCrossWorkgroup;
-                    break;
-                }
-                default:
-                    assert(false && "This address space is not supported");
-                    break;
-            }
-
             const Type* pointee = ptr->pointee();
             while (auto arr = pointee->isa<IndefiniteArrayType>())
                 pointee = arr->elem_type();
-            converted.id = builder_->declare_ptr_type(storage_class, convert(pointee).id);
+            converted.id = builder_->declare_ptr_type(static_cast<spv::StorageClass>(convert(ptr->addr_space())), convert(pointee).id);
             converted.layout = { target_info_.mem_layout.pointer_size, target_info_.mem_layout.pointer_size };
             break;
         }

From e5f9031ff8933b51076424775244a6367e3600bc Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 13:28:07 +0200
Subject: [PATCH 263/342] spirv: fix tuple type codegen

---
 src/thorin/be/spirv/spirv_types.cpp | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index e29812081..27f2ca413 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -205,8 +205,7 @@ ConvertedType CodeGen::convert(const Type* type) {
                 converted.layout->alignment = std::max(converted.layout->alignment, converted_member_type.layout->alignment);
                 converted.layout->size = pad(converted.layout->size + converted_member_type.layout->size, converted.layout->alignment);
             }
-            if (total_serialized_size == 0) {
-                outf("this one is void");
+            if (converted.layout->size == 0) {
                 converted.id = builder_->declare_void_type();
                 converted.layout = std::nullopt;
                 break;

From 9f22df5f1c975260b4029b64105f1a3613393612 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 13:28:20 +0200
Subject: [PATCH 264/342] added Input and Output address spaces

---
 src/thorin/rec_stream.cpp | 7 ++++++-
 src/thorin/type.h         | 2 ++
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/src/thorin/rec_stream.cpp b/src/thorin/rec_stream.cpp
index 0eb0fe449..590693e31 100644
--- a/src/thorin/rec_stream.cpp
+++ b/src/thorin/rec_stream.cpp
@@ -211,11 +211,16 @@ Stream& Type::stream(Stream& s) const {
         if (t->is_vector()) s.fmt(">");
 
         switch (t->addr_space()) {
+            case AddrSpace::Generic: break;
             case AddrSpace::Global:   s.fmt("[Global]");   break;
             case AddrSpace::Texture:  s.fmt("[Tex]");      break;
             case AddrSpace::Shared:   s.fmt("[Shared]");   break;
             case AddrSpace::Constant: s.fmt("[Constant]"); break;
-            default: /* ignore unknown address space */    break;
+            case AddrSpace::Private:  s.fmt("[Private]");  break;
+            case AddrSpace::Function: s.fmt("[Function]"); break;
+            case AddrSpace::Input:    s.fmt("[Input]");    break;
+            case AddrSpace::Output:   s.fmt("[Output]");   break;
+            default: s.fmt("[{}]", (int) t->addr_space()); break;
         }
         return s;
     } else if (auto t = isa<PrimType>()) {
diff --git a/src/thorin/type.h b/src/thorin/type.h
index bf7cb26df..8d69b4aa2 100644
--- a/src/thorin/type.h
+++ b/src/thorin/type.h
@@ -235,6 +235,8 @@ enum class AddrSpace : uint32_t {
     Private =  5, // Corresponds to the 'private' storage class in compute kernels/shaders, as in thread-private
     Function = 6, // Corresponds to the 'function' storage class in SPIR-V
     Push     = 7, // Corresponds to the 'push constant' storage class in SPIR-V
+    Input    = 8,
+    Output   = 9,
 };
 
 /// Pointer type.

From cec8afef60d74f664cdefce14451efb2963ac28b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 13:28:51 +0200
Subject: [PATCH 265/342] whitespace

---
 src/thorin/be/spirv/spirv.cpp | 2 --
 1 file changed, 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index e926308e3..c6f8fc669 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -95,8 +95,6 @@ SpvId FileBuilder::u32_constant(uint32_t pattern) {
     return constant(u32_t(), { pattern });
 }
 
-
-
 CodeGen::CodeGen(Thorin& thorin, Target& target_info, bool debug, const Cont2Config* kernel_config)
         : thorin::CodeGen(thorin, debug), target_info_(target_info), kernel_config_(kernel_config)
 {}

From 4ffffed1dc9a4acb7fd021d0478ca397bf5f3c9a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 13:57:22 +0200
Subject: [PATCH 266/342] spirv: implement Match

---
 src/thorin/be/spirv/spirv.cpp         | 21 +++++++++++----------
 src/thorin/be/spirv/spirv_builder.hpp | 11 +++++++++++
 2 files changed, 22 insertions(+), 10 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c6f8fc669..b530ee5db 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -316,16 +316,17 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         auto fbb = emit(app.arg(3));
         bb->branch_conditional(cond, tbb, fbb);
     } else if (app.callee()->isa<Continuation>() && app.callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
-        /*auto val = emit(continuation->arg(0));
-        auto otherwise_bb = cont2bb(continuation->arg(1)->isa_nom<Continuation>());
-        auto match = irbuilder.CreateSwitch(val, otherwise_bb, continuation->num_args() - 2);
-        for (size_t i = 2; i < continuation->num_args(); i++) {
-            auto arg = continuation->arg(i)->as<Tuple>();
-            auto case_const = llvm::cast<llvm::ConstantInt>(emit(arg->op(0)));
-            auto case_bb    = cont2bb(arg->op(1)->isa_nom<Continuation>());
-            match->addCase(case_const, case_bb);
-        }*/
-        THORIN_UNREACHABLE;
+        emit_unsafe(app.arg(0));
+        auto val = emit(app.arg(1));
+        auto otherwise_bb = emit_as_bb(app.arg(2)->isa_nom<Continuation>());
+        std::vector<SpvId> literals;
+        std::vector<SpvId> cases;
+        for (size_t i = 3; i < app.num_args(); i++) {
+            auto arg = app.arg(i)->as<Tuple>();
+            literals.push_back(emit(arg->op(0)));
+            cases.push_back(emit_as_bb(arg->op(1)->as_nom<Continuation>()));
+        }
+        bb->branch_switch(val, otherwise_bb, literals, cases);
     } else if (app.callee()->isa<Bottom>()) {
         bb->unreachable();
     } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && (intrinsic->is_intrinsic() || intrinsic->cc() == CC::Device)) {
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 2e741f2fe..b94f11eb1 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -220,6 +220,17 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         ref_id(false_target);
     }
 
+    void branch_switch(SpvId selector, SpvId default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
+        assert(literals.size() == cases.size());
+        op(spv::Op::OpSwitch, 3 + literals.size() * 2);
+        ref_id(selector);
+        ref_id(default_case);
+        for (size_t i = 0; i < literals.size(); i++) {
+            ref_id(literals[i]);
+            ref_id(cases[i]);
+        }
+    }
+
     void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
         op(spv::Op::OpSelectionMerge, 3);
         ref_id(merge_bb);

From 4f6563b77be5bdb0622aaf8f47e015fdf8fb8d57 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 2 Jul 2024 16:14:54 +0200
Subject: [PATCH 267/342] spirv: Implement Slot, Enter

---
 src/thorin/be/spirv/spirv.cpp       | 22 +++++++++++++++++-----
 src/thorin/be/spirv/spirv_types.cpp |  8 ++++++--
 2 files changed, 23 insertions(+), 7 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index b530ee5db..057436611 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -622,6 +622,14 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             bb->store(emit(store->val()), emit(store->ptr()), operands);
             return spv_none;
         } else THORIN_UNREACHABLE;
+    } else if (auto slot = def->isa<Slot>()) {
+        emit_unsafe(slot->frame());
+        auto type = slot->type();
+        auto id = bb->fn_builder.variable(convert(world().ptr_type(type->pointee(), 1, AddrSpace::Function)).id, spv::StorageClass::StorageClassFunction);
+        id = bb->convert(spv::Op::OpBitcast, convert(type).id, id);
+        return id;
+    } else if (auto enter = def->isa<Enter>()) {
+        return emit_unsafe(enter->mem());
     } else if (auto lea = def->isa<LEA>()) {
         switch (lea->ptr_type()->addr_space()) {
             case AddrSpace::Global:
@@ -635,13 +643,12 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto offset = emit(lea->index());
         return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {});
     } else if (auto aggop = def->isa<AggOp>()) {
-        auto spv_agg = emit(aggop->agg());
         auto agg_type = convert(aggop->agg()->type()).id;
 
         bool mem = false;
         if (auto tt = aggop->agg()->type()->isa<TupleType>(); tt && tt->op(0)->isa<MemType>()) mem = true;
 
-        auto copy_to_alloca = [&] (SpvId target_type) {
+        auto copy_to_alloca = [&] (SpvId spv_agg, SpvId target_type) {
             world().wdef(def, "slow: alloca and loads/stores needed for aggregate '{}'", def);
             auto agg_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, agg_type);
 
@@ -654,7 +661,11 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         };
 
         if (auto extract = aggop->isa<Extract>()) {
-            if (is_mem(extract)) return spv_none;
+            if (is_mem(extract) || extract->type()->isa<FrameType>()) {
+                emit_unsafe(extract->agg());
+                return spv_none;
+            }
+            auto spv_agg = emit(aggop->agg());
 
             auto target_type = convert(extract->type()).id;
             auto constant_index = aggop->index()->isa<PrimLit>();
@@ -663,7 +674,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             if (aggop->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
                 assert(aggop->agg()->type()->isa<DefiniteArrayType>());
                 assert(!is_mem(extract));
-                return bb->load(target_type, copy_to_alloca(target_type).second);
+                return bb->load(target_type, copy_to_alloca(spv_agg, target_type).second);
             }
 
             if (extract->agg()->type()->isa<VectorType>())
@@ -681,6 +692,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
 
             return bb->extract(target_type, spv_agg, { index - offset });
         } else if (auto insert = def->isa<Insert>()) {
+            auto spv_agg = emit(aggop->agg());
             auto value = emit(insert->value());
             auto constant_index = aggop->index()->isa<PrimLit>();
 
@@ -688,7 +700,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
 
             if (insert->agg()->type()->isa<ArrayType>() && constant_index == nullptr) {
                 assert(aggop->agg()->type()->isa<DefiniteArrayType>());
-                auto [variable, cell] = copy_to_alloca(agg_type);
+                auto [variable, cell] = copy_to_alloca(spv_agg, agg_type);
                 bb->store(value, cell);
                 return bb->load(agg_type, variable);
             }
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 27f2ca413..90a718697 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -14,8 +14,12 @@ uint32_t CodeGen::convert(AddrSpace as) {
                 builder_->capability(spv::CapabilityVectorComputeINTEL);
             break;
         }
+        case AddrSpace::Generic: {
+            storage_class = spv::StorageClassGeneric;
+            builder_->capability(spv::Capability::CapabilityGenericPointer);
+            break;
+        }
         case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
-        case AddrSpace::Generic:  storage_class = spv::StorageClassGeneric;               break;
         case AddrSpace::Input:    storage_class = spv::StorageClassInput;                 break;
         case AddrSpace::Output:   storage_class = spv::StorageClassOutput;                break;
         case AddrSpace::Global: {
@@ -198,7 +202,7 @@ ConvertedType CodeGen::convert(const Type* type) {
             converted.layout = { 0, 0 };
             for (auto member : type->ops()) {
                 auto member_type = member->as<Type>();
-                if (member_type == world().unit_type() || member_type == world().mem_type()) continue;
+                if (member_type == world().unit_type() || member_type == world().mem_type() || member_type->isa<FrameType>()) continue;
                 auto converted_member_type = convert(member_type);
                 assert(converted_member_type.layout);
                 spv_types.push_back(converted_member_type.id);

From ec932facb60a88f503b572e982fbfd042acd1c9c Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 4 Jul 2024 14:25:09 +0200
Subject: [PATCH 268/342] rewrote DeviceBackends entirely

---
 src/thorin/be/codegen.cpp | 346 +++++++++++++++++++++++---------------
 src/thorin/be/codegen.h   |  42 ++++-
 src/thorin/world.cpp      |   2 +
 src/thorin/world.h        |   1 +
 4 files changed, 253 insertions(+), 138 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 83b5a0bf0..a01279af5 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -1,58 +1,23 @@
 #include "thorin/be/codegen.h"
-#include "thorin/analyses/scope.h"
-#include "thorin/transform/hls_channels.h"
-#include "thorin/transform/hls_kernel_launch.h"
+
+#include "thorin/be/c/c.h"
 
 #if THORIN_ENABLE_LLVM
-#include "thorin/be/llvm/cpu.h"
 #include "thorin/be/llvm/nvvm.h"
 #include "thorin/be/llvm/amdgpu_hsa.h"
 #include "thorin/be/llvm/amdgpu_pal.h"
 #endif
+
 #if THORIN_ENABLE_SHADY
 #include "thorin/be/shady/shady.h"
 #undef empty
 #undef nodes
 #endif
-#include "thorin/be/c/c.h"
-
-namespace thorin {
 
-static void get_kernel_configs(
-    Thorin& thorin,
-    const std::vector<Continuation*>& kernels,
-    Cont2Config& kernel_configs,
-    std::function<std::unique_ptr<KernelConfig> (Continuation*, Continuation*)> use_callback)
-{
-    thorin.opt();
-
-    auto externals = thorin.world().externals();
-    for (auto continuation : kernels) {
-        // recover the imported continuation (lost after the call to opt)
-        Continuation* imported = nullptr;
-        for (auto [_, def] : externals) {
-            auto exported = def->isa<Continuation>();
-            if (!exported) continue;
-            if (!exported->has_body()) continue;
-            if (exported->name() == continuation->name())
-                imported = exported;
-        }
-        if (!imported) continue;
-
-        visit_uses(continuation, [&] (Continuation* use) {
-            assert(use->has_body());
-            auto config = use_callback(use, imported);
-            if (config) {
-                auto p = kernel_configs.emplace(imported, std::move(config));
-                assert_unused(p.second && "single kernel config entry expected");
-            }
-            return false;
-        }, true);
+#include "thorin/transform/hls_channels.h"
+#include "thorin/transform/hls_kernel_launch.h"
 
-        continuation->world().make_external(continuation);
-        continuation->destroy("codegen");
-    }
-}
+namespace thorin {
 
 static const App* get_alloc_call(const Def* def) {
     // look through casts
@@ -84,95 +49,104 @@ static uint64_t get_alloc_size(const Def* def) {
     return size ? static_cast<uint64_t>(size->value().get_qu64()) : 0_u64;
 }
 
-DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& flags)
-    : cgs {}
-{
-    std::vector<Importer> importers;
-    for (auto& name : backend_names) {
-        accelerator_code.emplace_back(name);
-        importers.emplace_back(world, accelerator_code.back().world());
-    }
-
-    static const auto backend_intrinsics = std::array {
-        std::pair { CUDA,       Intrinsic::CUDA         },
-        std::pair { NVVM,       Intrinsic::NVVM         },
-        std::pair { OpenCL,     Intrinsic::OpenCL       },
-        std::pair { AMDGPU_HSA, Intrinsic::AMDGPUHSA    },
-        std::pair { AMDGPU_PAL, Intrinsic::AMDGPUPAL    },
-        std::pair { HLS,        Intrinsic::HLS          },
-        std::pair { Shady,      Intrinsic::ShadyCompute }
-    };
+static std::unique_ptr<GPUKernelConfig> get_gpu_kernel_config(const App* app, Continuation* imported) {
+    // determine whether or not this kernel uses restrict pointers
+    bool has_restrict = true;
+    DefSet allocs;
+    for (size_t i = LaunchArgs::Num, e = app->num_args(); has_restrict && i != e; ++i) {
+        auto arg = app->arg(i);
+        if (!arg->type()->isa<PtrType>()) continue;
+        auto alloc = get_alloc_call(arg);
+        if (!alloc) has_restrict = false;
+        auto p = allocs.insert(alloc);
+        has_restrict &= p.second;
+    }
 
-    // determine different parts of the world which need to be compiled differently
-    ScopesForest(world).for_each([&] (const Scope& scope) {
-        auto continuation = scope.entry();
-        Continuation* imported = nullptr;
-        for (auto [backend, intrinsic] : backend_intrinsics) {
-            if (is_passed_to_intrinsic(continuation, intrinsic)) {
-                imported = importers[backend].import(continuation)->as_nom<Continuation>();
-                break;
-            }
-        }
+    auto it_config = app->arg(LaunchArgs::Config)->isa<Tuple>();
+    if (it_config &&
+        it_config->op(0)->isa<PrimLit>() &&
+        it_config->op(1)->isa<PrimLit>() &&
+        it_config->op(2)->isa<PrimLit>()) {
+        return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{
+                it_config->op(0)->as<PrimLit>()->qu32_value().data(),
+                it_config->op(1)->as<PrimLit>()->qu32_value().data(),
+                it_config->op(2)->as<PrimLit>()->qu32_value().data()
+        }, has_restrict);
+    }
+    return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{-1, -1, -1}, has_restrict);
+}
 
-        if (imported == nullptr)
-            return;
+Backend::Backend(thorin::DeviceBackends& backends, World& src) : backends_(backends), device_code_(src), importer_(std::make_unique<Importer>(src, device_code_.world())) {}
 
-        // Necessary so that the names match in the original and imported worlds
-        imported->set_name(continuation->unique_name());
-        continuation->set_name(continuation->unique_name());
-        for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
-            imported->param(i)->set_name(continuation->param(i)->name());
-        imported->world().make_external(imported);
-        imported->attributes().cc = CC::C;
+struct CudaBackend : public Backend {
+    explicit CudaBackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::CUDA, *this, get_gpu_kernel_config);
+    }
 
-        kernels.emplace_back(continuation);
-    });
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        std::string empty;
+        return std::make_unique<c::CodeGen>(device_code_, config, c::Lang::CUDA, backends_.debug(), empty);
+    }
+};
 
-    for (auto [backend, intrinsic] : backend_intrinsics) {
-        if (backend == HLS)
-            continue;
+struct OpenCLBackend : public Backend {
+    explicit OpenCLBackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::OpenCL, *this, get_gpu_kernel_config);
+    }
 
-        if (!accelerator_code[backend].world().empty()) {
-            get_kernel_configs(accelerator_code[backend], kernels, kernel_config, [&](Continuation *use, Continuation * /* imported */) {
-                auto app = use->body();
-                if (app->callee()->as<Continuation>()->intrinsic() != intrinsic)
-                    return std::unique_ptr<GPUKernelConfig>(nullptr);
-                // determine whether or not this kernel uses restrict pointers
-                bool has_restrict = true;
-                DefSet allocs;
-                for (size_t i = LaunchArgs::Num, e = app->num_args(); has_restrict && i != e; ++i) {
-                    auto arg = app->arg(i);
-                    if (!arg->type()->isa<PtrType>()) continue;
-                    auto alloc = get_alloc_call(arg);
-                    if (!alloc) has_restrict = false;
-                    auto p = allocs.insert(alloc);
-                    has_restrict &= p.second;
-                }
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        std::string empty;
+        return std::make_unique<c::CodeGen>(device_code_, config, c::Lang::OpenCL, backends_.debug(), empty);
+    }
+};
 
-                auto it_config = app->arg(LaunchArgs::Config)->isa<Tuple>();
-                if (it_config &&
-                    it_config->op(0)->isa<PrimLit>() &&
-                    it_config->op(1)->isa<PrimLit>() &&
-                    it_config->op(2)->isa<PrimLit>()) {
-                    return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{
-                        it_config->op(0)->as<PrimLit>()->qu32_value().data(),
-                        it_config->op(1)->as<PrimLit>()->qu32_value().data(),
-                        it_config->op(2)->as<PrimLit>()->qu32_value().data()
-                    }, has_restrict);
-                }
-                return std::make_unique<GPUKernelConfig>(std::tuple<int, int, int>{-1, -1, -1}, has_restrict);
-            });
-        }
+#if THORIN_ENABLE_LLVM
+struct AMDHSABackend : public Backend {
+    explicit AMDHSABackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::AMDGPUHSA, *this, get_gpu_kernel_config);
+    }
+
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        return std::make_unique<llvm::AMDGPUHSACodeGen>(device_code_, config, backends_.opt(), backends_.debug());
+    }
+};
+
+struct AMDPALBackend : public Backend {
+    explicit AMDPALBackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::AMDGPUPAL, *this, get_gpu_kernel_config);
+    }
+
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        return std::make_unique<llvm::AMDGPUPALCodeGen>(device_code_, config, backends_.opt(), backends_.debug());
+    }
+};
+
+struct NVVMBackend : public Backend {
+    explicit NVVMBackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::NVVM, *this, get_gpu_kernel_config);
+    }
+
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        return std::make_unique<llvm::NVVMCodeGen>(device_code_, config, backends_.opt(), backends_.debug());
+    }
+};
+#endif
+
+#if THORIN_ENABLE_SHADY
+struct ShadyBackend : public Backend {
+    explicit ShadyBackend(DeviceBackends2& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::ShadyCompute, get_gpu_kernel_config);
     }
 
-    // get the HLS kernel configurations
-    Top2Kernel top2kernel;
-    DeviceParams hls_host_params;
-    if (!accelerator_code[HLS].world().empty()) {
-        hls_host_params = hls_channels(accelerator_code[HLS], importers[HLS], top2kernel, world);
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        return std::make_unique<shady_be::CodeGen>(device_code_, config, backends_.debug());
+    }
+};
+#endif
 
-        get_kernel_configs(accelerator_code[HLS], kernels, kernel_config, [&] (Continuation* use, Continuation* imported) {
-            auto app = use->body();
+struct HLSBackend : public Backend {
+    explicit HLSBackend(DeviceBackends& b, World& src, std::string& hls_flags) : Backend(b, src), hls_flags_(hls_flags) {
+        b.register_intrinsic(Intrinsic::NVVM, *this, [&](const App* app, Continuation* imported) {
             HLSKernelConfig::Param2Size param_sizes;
             for (size_t i = hls_free_vars_offset, e = app->num_args(); i != e; ++i) {
                 auto arg = app->arg(i);
@@ -180,7 +154,7 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
                 if (!ptr_type) continue;
                 auto size = get_alloc_size(arg);
                 if (size == 0)
-                    world.edef(arg, "array size is not known at compile time");
+                    b.world().edef(arg, "array size is not known at compile time");
                 auto elem_type = ptr_type->pointee();
                 size_t multiplier = 1;
                 if (!elem_type->isa<PrimType>()) {
@@ -195,29 +169,137 @@ DeviceBackends::DeviceBackends(World& world, int opt, bool debug, std::string& f
                 }
                 auto prim_type = elem_type->isa<PrimType>();
                 if (!prim_type)
-                    world.edef(arg, "only pointers to arrays of primitive types are supported");
+                    b.world().edef(arg, "only pointers to arrays of primitive types are supported");
                 auto num_elems = size / (multiplier * num_bits(prim_type->primtype_tag()) / 8);
                 // imported has type: fn (mem, fn (mem), ...)
                 param_sizes.emplace(imported->param(i - hls_free_vars_offset + 2), num_elems);
             }
             return std::make_unique<HLSKernelConfig>(param_sizes);
         });
-        hls_annotate_top(accelerator_code[HLS].world(), top2kernel, kernel_config);
     }
-    hls_kernel_launch(world, hls_host_params);
 
+    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+        Top2Kernel top2kernel;
+        DeviceParams hls_host_params;
+
+        hls_host_params = hls_channels(device_code_, *importer_, top2kernel, backends_.world());
+        hls_annotate_top(device_code_.world(), top2kernel, const_cast<Cont2Config&>(config));
+        hls_kernel_launch(device_code_.world(), hls_host_params);
+
+        return std::make_unique<c::CodeGen>(device_code_, config, c::Lang::HLS, backends_.debug(), hls_flags_);
+    }
+
+    std::string& hls_flags_;
+};
+
+DeviceBackends::DeviceBackends(thorin::World& world, int opt, bool debug, std::string& hls_flags) : world_(world), opt_(opt), debug_(debug) {
+    register_backend(std::make_unique<CudaBackend>(*this, world));
+    register_backend(std::make_unique<OpenCLBackend>(*this, world));
 #if THORIN_ENABLE_LLVM
-    if (!accelerator_code[NVVM      ].world().empty()) cgs[NVVM      ] = std::make_unique<llvm::NVVMCodeGen     >(accelerator_code[NVVM      ], kernel_config, opt, debug);
-    if (!accelerator_code[AMDGPU_HSA].world().empty()) cgs[AMDGPU_HSA] = std::make_unique<llvm::AMDGPUHSACodeGen>(accelerator_code[AMDGPU_HSA], kernel_config, opt, debug);
-    if (!accelerator_code[AMDGPU_PAL].world().empty()) cgs[AMDGPU_PAL] = std::make_unique<llvm::AMDGPUPALCodeGen>(accelerator_code[AMDGPU_PAL], kernel_config, opt, debug);
-#else
-    (void)opt;
+    register_backend(std::make_unique<AMDHSABackend>(*this, world));
+    register_backend(std::make_unique<AMDPALBackend>(*this, world));
+    register_backend(std::make_unique<NVVMBackend>(*this, world));
 #endif
 #if THORIN_ENABLE_SHADY
-    if (!accelerator_code[Shady].world().empty()) cgs[Shady] = std::make_unique<shady_be::CodeGen>(accelerator_code[Shady], kernel_config, debug);
+    register_backend(std::make_unique<ShadyBackend>(*this, world))
 #endif
-    for (auto [backend, lang] : std::array { std::pair { CUDA, c::Lang::CUDA }, std::pair { OpenCL, c::Lang::OpenCL }, std::pair { HLS, c::Lang::HLS } })
-        if (!accelerator_code[backend].world().empty()) cgs[backend] = std::make_unique<c::CodeGen>(accelerator_code[backend], kernel_config, lang, debug, flags);
+    register_backend(std::make_unique<HLSBackend>(*this, world, hls_flags));
+
+    search_for_device_code();
+}
+
+void DeviceBackends::register_backend(std::unique_ptr<Backend> backend) {
+    backends_.push_back(std::move(backend));
+}
+
+World& DeviceBackends::world() { return world_; }
+bool DeviceBackends::debug() { return debug_; }
+int DeviceBackends::opt() { return opt_; }
+
+void DeviceBackends::register_intrinsic(thorin::Intrinsic intrinsic, Backend& backend, GetKernelConfigFn f) {
+    intrinsics_[intrinsic] = std::make_pair(&backend, f);
+}
+
+void DeviceBackends::search_for_device_code() {
+    // determine different parts of the world which need to be compiled differently
+    ScopesForest(world_).for_each([&] (const Scope& scope) {
+        auto continuation = scope.entry();
+        Continuation* imported = nullptr;
+
+        Intrinsic intrinsic = Intrinsic::None;
+        visit_capturing_intrinsics(continuation, [&] (Continuation* continuation) {
+            if (continuation->is_accelerator()) {
+                intrinsic = continuation->intrinsic();
+                return true;
+            }
+            return false;
+        });
+
+        if (intrinsic == Intrinsic::None)
+            return;
+
+        auto handler = intrinsics_.find(intrinsic);
+        assert(handler != intrinsics_.end());
+        auto [backend, get_config] = handler->second;
+
+        imported = backend->importer_->import(continuation)->as_nom<Continuation>();
+        if (imported == nullptr)
+            return;
+
+        // Necessary so that the names match in the original and imported worlds
+        imported->set_name(continuation->unique_name());
+        continuation->set_name(continuation->unique_name());
+        for (size_t i = 0, e = continuation->num_params(); i != e; ++i)
+            imported->param(i)->set_name(continuation->param(i)->name());
+        imported->world().make_external(imported);
+        imported->attributes().cc = CC::C;
+
+        backend->kernels_.emplace_back(continuation);
+    });
+
+    for (auto& backend : backends_) {
+        if (backend->thorin().world().empty())
+            continue;
+
+        backend->thorin().opt();
+
+        Cont2Config kernel_configs;
+
+        auto externals = world_.externals();
+        for (auto continuation : backend->kernels_) {
+            // recover the imported continuation (lost after the call to opt)
+            Continuation* imported = nullptr;
+            for (auto [_, def] : externals) {
+                auto exported = def->isa<Continuation>();
+                if (!exported) continue;
+                if (!exported->has_body()) continue;
+                if (exported->name() == continuation->name())
+                    imported = exported;
+            }
+            if (!imported) continue;
+
+            visit_uses(continuation, [&] (Continuation* use) {
+                assert(use->has_body());
+
+                auto handler = intrinsics_.find(use->body()->callee()->as<Continuation>()->intrinsic());
+                assert(handler != intrinsics_.end());
+                auto [backend2, get_config] = handler->second;
+                assert(backend2 == &*backend);
+
+                auto config = get_config(use->body(), imported);
+                if (config) {
+                    auto p = kernel_configs.emplace(imported, std::move(config));
+                    assert_unused(p.second && "single kernel config entry expected");
+                }
+                return false;
+            }, true);
+
+            continuation->world().make_external(continuation);
+            continuation->destroy("codegen");
+        }
+
+        cgs.emplace_back(backend->create_cg(kernel_configs));
+    }
 }
 
 CodeGen::CodeGen(Thorin& thorin, bool debug)
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index c46cee7c0..47a68ac45 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -39,17 +39,47 @@ struct LaunchArgs {
     };
 };
 
+struct DeviceBackends;
+
+struct Backend {
+    Backend(DeviceBackends& backends, World& src);
+
+    virtual std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) = 0;
+
+    Thorin& thorin() { return device_code_; }
+    Importer& importer() { return *importer_; }
+
+protected:
+    DeviceBackends& backends_;
+    Thorin device_code_;
+    std::unique_ptr<Importer> importer_;
+
+    std::vector<Continuation*> kernels_;
+    friend DeviceBackends;
+};
+
 struct DeviceBackends {
     DeviceBackends(World& world, int opt, bool debug, std::string& hls_flags);
 
-    Cont2Config kernel_config;
-    std::vector<Continuation*> kernels;
+    World& world();
+    std::vector<std::unique_ptr<CodeGen>> cgs;
+
+    int opt();
+    bool debug();
+
+    void register_backend(std::unique_ptr<Backend>);
+    using GetKernelConfigFn = std::function<std::unique_ptr<KernelConfig>(const App*, Continuation*)>;
+    void register_intrinsic(Intrinsic, Backend&, GetKernelConfigFn);
 
-    enum { CUDA, NVVM, OpenCL, AMDGPU_HSA, AMDGPU_PAL, HLS, Shady, BackendCount };
-    std::array<std::unique_ptr<CodeGen>, BackendCount> cgs;
 private:
-    std::array<const char*, BackendCount> backend_names = { "CUDA", "NVVM", "OpenCL", "AMDGPU_HSA", "AMDGPU_PAL", "HLS", "Shady" };
-    std::vector<Thorin> accelerator_code;
+    World& world_;
+    std::vector<std::unique_ptr<Backend>> backends_;
+    std::unordered_map<Intrinsic, std::pair<Backend*, GetKernelConfigFn>> intrinsics_;
+
+    int opt_;
+    bool debug_;
+
+    void search_for_device_code();
 };
 
 }
diff --git a/src/thorin/world.cpp b/src/thorin/world.cpp
index 96b3322f5..5c3339685 100644
--- a/src/thorin/world.cpp
+++ b/src/thorin/world.cpp
@@ -1299,6 +1299,8 @@ Thorin::Thorin(const std::string& name)
     : world_(std::make_unique<World>(name))
 {}
 
+Thorin::Thorin(thorin::World& src) : world_(std::make_unique<World>(src)) {}
+
 void Thorin::opt() {
     bool debug_passes = getenv("THORIN_DEBUG_PASSES");
 #define RUN_PASS(pass) \
diff --git a/src/thorin/world.h b/src/thorin/world.h
index a35c95ad3..4b77d8072 100644
--- a/src/thorin/world.h
+++ b/src/thorin/world.h
@@ -405,6 +405,7 @@ class Thorin {
 public:
     /// Initial world constructor
     explicit Thorin(const std::string& name);
+    explicit Thorin(World& src);
 
     World& world() { return *world_; };
     std::unique_ptr<World>& world_container() { return world_; }

From 6dc95f7283c134c6d83e4145ab97c948435b6fd4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 4 Jul 2024 14:29:49 +0200
Subject: [PATCH 269/342] move and rename KernelLaunchArgs

---
 src/thorin/be/codegen.cpp      |  4 ++--
 src/thorin/be/codegen.h        | 12 ------------
 src/thorin/be/kernel_config.h  | 12 ++++++++++++
 src/thorin/be/llvm/runtime.cpp | 14 +++++++-------
 4 files changed, 21 insertions(+), 21 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index a01279af5..fecc60720 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -53,7 +53,7 @@ static std::unique_ptr<GPUKernelConfig> get_gpu_kernel_config(const App* app, Co
     // determine whether or not this kernel uses restrict pointers
     bool has_restrict = true;
     DefSet allocs;
-    for (size_t i = LaunchArgs::Num, e = app->num_args(); has_restrict && i != e; ++i) {
+    for (size_t i = KernelLaunchArgs::Num, e = app->num_args(); has_restrict && i != e; ++i) {
         auto arg = app->arg(i);
         if (!arg->type()->isa<PtrType>()) continue;
         auto alloc = get_alloc_call(arg);
@@ -62,7 +62,7 @@ static std::unique_ptr<GPUKernelConfig> get_gpu_kernel_config(const App* app, Co
         has_restrict &= p.second;
     }
 
-    auto it_config = app->arg(LaunchArgs::Config)->isa<Tuple>();
+    auto it_config = app->arg(KernelLaunchArgs::Config)->isa<Tuple>();
     if (it_config &&
         it_config->op(0)->isa<PrimLit>() &&
         it_config->op(1)->isa<PrimLit>() &&
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index 47a68ac45..3bb5d5636 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -27,18 +27,6 @@ class CodeGen {
     bool debug_;
 };
 
-struct LaunchArgs {
-    enum {
-        Mem = 0,
-        Device,
-        Space,
-        Config,
-        Body,
-        Return,
-        Num
-    };
-};
-
 struct DeviceBackends;
 
 struct Backend {
diff --git a/src/thorin/be/kernel_config.h b/src/thorin/be/kernel_config.h
index c1d36aeb3..787b579a6 100644
--- a/src/thorin/be/kernel_config.h
+++ b/src/thorin/be/kernel_config.h
@@ -6,6 +6,18 @@
 
 namespace thorin {
 
+struct KernelLaunchArgs {
+    enum {
+        Mem = 0,
+        Device,
+        Space,
+        Config,
+        Body,
+        Return,
+        Num
+    };
+};
+
 class KernelConfig : public RuntimeCast<KernelConfig> {
 public:
     virtual ~KernelConfig() {}
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index 375713091..0d8ff9092 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -67,22 +67,22 @@ void Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Plat
     // target(mem, device, (dim.x, dim.y, dim.z), (block.x, block.y, block.z), body, return, free_vars)
     auto target = body->callee()->as_nom<Continuation>();
     assert_unused(target->is_intrinsic());
-    assert(body->num_args() >= LaunchArgs::Num && "required arguments are missing");
+    assert(body->num_args() >= KernelLaunchArgs::Num && "required arguments are missing");
 
     // arguments
-    auto target_device_id = code_gen.emit(body->arg(LaunchArgs::Device));
+    auto target_device_id = code_gen.emit(body->arg(KernelLaunchArgs::Device));
     auto target_platform = builder.getInt32(platform);
     auto target_device = builder.CreateOr(target_platform, builder.CreateShl(target_device_id, builder.getInt32(4)));
 
-    auto it_space = body->arg(LaunchArgs::Space);
-    auto it_config = body->arg(LaunchArgs::Config);
-    auto kernel = body->arg(LaunchArgs::Body)->as<Global>()->init()->as<Continuation>();
+    auto it_space = body->arg(KernelLaunchArgs::Space);
+    auto it_config = body->arg(KernelLaunchArgs::Config);
+    auto kernel = body->arg(KernelLaunchArgs::Body)->as<Global>()->init()->as<Continuation>();
 
     auto& world = continuation->world();
     //auto kernel_name = builder.CreateGlobalStringPtr(kernel->name() == "hls_top" ? kernel->name() : kernel->name());
     auto kernel_name = builder.CreateGlobalStringPtr(kernel->name());
     auto file_name = builder.CreateGlobalStringPtr(world.name() + ext);
-    const size_t num_kernel_args = body->num_args() - LaunchArgs::Num;
+    const size_t num_kernel_args = body->num_args() - KernelLaunchArgs::Num;
 
     // allocate argument pointers, sizes, and types
     llvm::Value* args   = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt8PtrTy(), num_kernel_args), "args");
@@ -93,7 +93,7 @@ void Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Plat
 
     // fill array of arguments
     for (size_t i = 0; i < num_kernel_args; ++i) {
-        auto target_arg = body->arg(i + LaunchArgs::Num);
+        auto target_arg = body->arg(i + KernelLaunchArgs::Num);
         const auto target_val = code_gen.emit(target_arg);
 
         KernelArgType arg_type;

From 735d8aa05974e8bd80d5e8d62bae58fc96dfa266 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 4 Jul 2024 15:08:19 +0200
Subject: [PATCH 270/342] move get_kernel_configs into method

---
 src/thorin/be/codegen.cpp | 81 +++++++++++++++++++++------------------
 src/thorin/be/codegen.h   |  2 +
 2 files changed, 45 insertions(+), 38 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index fecc60720..e9df0a49d 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -19,6 +19,47 @@
 
 namespace thorin {
 
+std::unique_ptr<Cont2Config> Backend::get_kernel_configs() {
+    device_code_.opt();
+
+    auto kernel_configs = std::make_unique<Cont2Config>();
+
+    auto& externals = backends_.world().externals();
+    for (auto continuation : kernels_) {
+        // recover the imported continuation (lost after the call to opt)
+        Continuation* imported = nullptr;
+        for (auto [_, def] : externals) {
+            auto exported = def->isa<Continuation>();
+            if (!exported) continue;
+            if (!exported->has_body()) continue;
+            if (exported->name() == continuation->name())
+                imported = exported;
+        }
+        if (!imported) continue;
+
+        visit_uses(continuation, [&] (Continuation* use) {
+            assert(use->has_body());
+
+            auto handler = backends_.intrinsics_.find(use->body()->callee()->as<Continuation>()->intrinsic());
+            assert(handler != backends_.intrinsics_.end());
+            auto [backend2, get_config] = handler->second;
+            assert(backend2 == this);
+
+            auto config = get_config(use->body(), imported);
+            if (config) {
+                auto p = kernel_configs->emplace(imported, std::move(config));
+                assert_unused(p.second && "single kernel config entry expected");
+            }
+            return false;
+        }, true);
+
+        continuation->world().make_external(continuation);
+        continuation->destroy("codegen");
+    }
+
+    return kernel_configs;
+}
+
 static const App* get_alloc_call(const Def* def) {
     // look through casts
     while (auto conv_op = def->isa<ConvOp>())
@@ -261,44 +302,8 @@ void DeviceBackends::search_for_device_code() {
         if (backend->thorin().world().empty())
             continue;
 
-        backend->thorin().opt();
-
-        Cont2Config kernel_configs;
-
-        auto externals = world_.externals();
-        for (auto continuation : backend->kernels_) {
-            // recover the imported continuation (lost after the call to opt)
-            Continuation* imported = nullptr;
-            for (auto [_, def] : externals) {
-                auto exported = def->isa<Continuation>();
-                if (!exported) continue;
-                if (!exported->has_body()) continue;
-                if (exported->name() == continuation->name())
-                    imported = exported;
-            }
-            if (!imported) continue;
-
-            visit_uses(continuation, [&] (Continuation* use) {
-                assert(use->has_body());
-
-                auto handler = intrinsics_.find(use->body()->callee()->as<Continuation>()->intrinsic());
-                assert(handler != intrinsics_.end());
-                auto [backend2, get_config] = handler->second;
-                assert(backend2 == &*backend);
-
-                auto config = get_config(use->body(), imported);
-                if (config) {
-                    auto p = kernel_configs.emplace(imported, std::move(config));
-                    assert_unused(p.second && "single kernel config entry expected");
-                }
-                return false;
-            }, true);
-
-            continuation->world().make_external(continuation);
-            continuation->destroy("codegen");
-        }
-
-        cgs.emplace_back(backend->create_cg(kernel_configs));
+        auto kernel_configs = backend->get_kernel_configs();
+        cgs.emplace_back(backend->create_cg(*kernel_configs));
     }
 }
 
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index 3bb5d5636..c413e1121 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -32,6 +32,7 @@ struct DeviceBackends;
 struct Backend {
     Backend(DeviceBackends& backends, World& src);
 
+    std::unique_ptr<Cont2Config> get_kernel_configs();
     virtual std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) = 0;
 
     Thorin& thorin() { return device_code_; }
@@ -68,6 +69,7 @@ struct DeviceBackends {
     bool debug_;
 
     void search_for_device_code();
+friend Backend;
 };
 
 }

From 410c03a83ff02aae27ec853adc4fb3b80eaeac62 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 4 Jul 2024 15:47:13 +0200
Subject: [PATCH 271/342] backends: fix HLS using NVVM intrinsic

---
 src/thorin/be/codegen.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index e9df0a49d..1c5d4f35e 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -187,7 +187,7 @@ struct ShadyBackend : public Backend {
 
 struct HLSBackend : public Backend {
     explicit HLSBackend(DeviceBackends& b, World& src, std::string& hls_flags) : Backend(b, src), hls_flags_(hls_flags) {
-        b.register_intrinsic(Intrinsic::NVVM, *this, [&](const App* app, Continuation* imported) {
+        b.register_intrinsic(Intrinsic::HLS, *this, [&](const App* app, Continuation* imported) {
             HLSKernelConfig::Param2Size param_sizes;
             for (size_t i = hls_free_vars_offset, e = app->num_args(); i != e; ++i) {
                 auto arg = app->arg(i);

From 5fea491c201f6d6b747e006d4de8913661068f7e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 5 Jul 2024 11:21:58 +0200
Subject: [PATCH 272/342] spirv: fixed LEA using src pointer type instead of
 dst

---
 src/thorin/be/spirv/spirv.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 057436611..1e4e66532 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -631,7 +631,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     } else if (auto enter = def->isa<Enter>()) {
         return emit_unsafe(enter->mem());
     } else if (auto lea = def->isa<LEA>()) {
-        switch (lea->ptr_type()->addr_space()) {
+        switch (lea->type()->addr_space()) {
             case AddrSpace::Global:
             case AddrSpace::Shared:
                 break;
@@ -639,7 +639,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
                 world().ELOG("LEA is only allowed in global & shared address spaces");
                 break;
         }
-        auto type = convert(lea->ptr_type()).id;
+        auto type = convert(lea->type()).id;
         auto offset = emit(lea->index());
         return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {});
     } else if (auto aggop = def->isa<AggOp>()) {

From 0f2bfce8aa2c0a0e467fbd3cae5301b0feb7641b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 5 Jul 2024 12:23:51 +0200
Subject: [PATCH 273/342] fix issues with kernel config ownership

---
 src/thorin/be/codegen.cpp | 12 ++++--------
 src/thorin/be/codegen.h   |  5 ++++-
 2 files changed, 8 insertions(+), 9 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 1c5d4f35e..159bde465 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -19,11 +19,9 @@
 
 namespace thorin {
 
-std::unique_ptr<Cont2Config> Backend::get_kernel_configs() {
+void Backend::prepare_kernel_configs() {
     device_code_.opt();
 
-    auto kernel_configs = std::make_unique<Cont2Config>();
-
     auto& externals = backends_.world().externals();
     for (auto continuation : kernels_) {
         // recover the imported continuation (lost after the call to opt)
@@ -47,7 +45,7 @@ std::unique_ptr<Cont2Config> Backend::get_kernel_configs() {
 
             auto config = get_config(use->body(), imported);
             if (config) {
-                auto p = kernel_configs->emplace(imported, std::move(config));
+                auto p = kernel_configs_.emplace(imported, std::move(config));
                 assert_unused(p.second && "single kernel config entry expected");
             }
             return false;
@@ -56,8 +54,6 @@ std::unique_ptr<Cont2Config> Backend::get_kernel_configs() {
         continuation->world().make_external(continuation);
         continuation->destroy("codegen");
     }
-
-    return kernel_configs;
 }
 
 static const App* get_alloc_call(const Def* def) {
@@ -302,8 +298,8 @@ void DeviceBackends::search_for_device_code() {
         if (backend->thorin().world().empty())
             continue;
 
-        auto kernel_configs = backend->get_kernel_configs();
-        cgs.emplace_back(backend->create_cg(*kernel_configs));
+        backend->prepare_kernel_configs();
+        cgs.emplace_back(backend->create_cg(backend->kernel_configs()));
     }
 }
 
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index c413e1121..1e85b597d 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -32,7 +32,7 @@ struct DeviceBackends;
 struct Backend {
     Backend(DeviceBackends& backends, World& src);
 
-    std::unique_ptr<Cont2Config> get_kernel_configs();
+    Cont2Config& kernel_configs() { return kernel_configs_; };
     virtual std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) = 0;
 
     Thorin& thorin() { return device_code_; }
@@ -44,6 +44,9 @@ struct Backend {
     std::unique_ptr<Importer> importer_;
 
     std::vector<Continuation*> kernels_;
+    Cont2Config kernel_configs_;
+
+    void prepare_kernel_configs();
     friend DeviceBackends;
 };
 

From d9f5726a20708c3cd148d82300788ba30c735457 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 5 Jul 2024 12:30:16 +0200
Subject: [PATCH 274/342] remove kernel_config from backend API

---
 src/thorin/be/codegen.cpp | 28 ++++++++++++++--------------
 src/thorin/be/codegen.h   |  3 +--
 2 files changed, 15 insertions(+), 16 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 159bde465..5953283d4 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -120,9 +120,9 @@ struct CudaBackend : public Backend {
         b.register_intrinsic(Intrinsic::CUDA, *this, get_gpu_kernel_config);
     }
 
-    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+    std::unique_ptr<CodeGen> create_cg() override {
         std::string empty;
-        return std::make_unique<c::CodeGen>(device_code_, config, c::Lang::CUDA, backends_.debug(), empty);
+        return std::make_unique<c::CodeGen>(device_code_, kernel_configs_, c::Lang::CUDA, backends_.debug(), empty);
     }
 };
 
@@ -131,9 +131,9 @@ struct OpenCLBackend : public Backend {
         b.register_intrinsic(Intrinsic::OpenCL, *this, get_gpu_kernel_config);
     }
 
-    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+    std::unique_ptr<CodeGen> create_cg() override {
         std::string empty;
-        return std::make_unique<c::CodeGen>(device_code_, config, c::Lang::OpenCL, backends_.debug(), empty);
+        return std::make_unique<c::CodeGen>(device_code_, kernel_configs_, c::Lang::OpenCL, backends_.debug(), empty);
     }
 };
 
@@ -143,8 +143,8 @@ struct AMDHSABackend : public Backend {
         b.register_intrinsic(Intrinsic::AMDGPUHSA, *this, get_gpu_kernel_config);
     }
 
-    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
-        return std::make_unique<llvm::AMDGPUHSACodeGen>(device_code_, config, backends_.opt(), backends_.debug());
+    std::unique_ptr<CodeGen> create_cg() override {
+        return std::make_unique<llvm::AMDGPUHSACodeGen>(device_code_, kernel_configs_, backends_.opt(), backends_.debug());
     }
 };
 
@@ -153,8 +153,8 @@ struct AMDPALBackend : public Backend {
         b.register_intrinsic(Intrinsic::AMDGPUPAL, *this, get_gpu_kernel_config);
     }
 
-    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
-        return std::make_unique<llvm::AMDGPUPALCodeGen>(device_code_, config, backends_.opt(), backends_.debug());
+    std::unique_ptr<CodeGen> create_cg() override {
+        return std::make_unique<llvm::AMDGPUPALCodeGen>(device_code_, kernel_configs_, backends_.opt(), backends_.debug());
     }
 };
 
@@ -163,8 +163,8 @@ struct NVVMBackend : public Backend {
         b.register_intrinsic(Intrinsic::NVVM, *this, get_gpu_kernel_config);
     }
 
-    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
-        return std::make_unique<llvm::NVVMCodeGen>(device_code_, config, backends_.opt(), backends_.debug());
+    std::unique_ptr<CodeGen> create_cg() override {
+        return std::make_unique<llvm::NVVMCodeGen>(device_code_, kernel_configs_, backends_.opt(), backends_.debug());
     }
 };
 #endif
@@ -215,15 +215,15 @@ struct HLSBackend : public Backend {
         });
     }
 
-    std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) override {
+    std::unique_ptr<CodeGen> create_cg() override {
         Top2Kernel top2kernel;
         DeviceParams hls_host_params;
 
         hls_host_params = hls_channels(device_code_, *importer_, top2kernel, backends_.world());
-        hls_annotate_top(device_code_.world(), top2kernel, const_cast<Cont2Config&>(config));
+        hls_annotate_top(device_code_.world(), top2kernel, kernel_configs_);
         hls_kernel_launch(device_code_.world(), hls_host_params);
 
-        return std::make_unique<c::CodeGen>(device_code_, config, c::Lang::HLS, backends_.debug(), hls_flags_);
+        return std::make_unique<c::CodeGen>(device_code_, kernel_configs_, c::Lang::HLS, backends_.debug(), hls_flags_);
     }
 
     std::string& hls_flags_;
@@ -299,7 +299,7 @@ void DeviceBackends::search_for_device_code() {
             continue;
 
         backend->prepare_kernel_configs();
-        cgs.emplace_back(backend->create_cg(backend->kernel_configs()));
+        cgs.emplace_back(backend->create_cg());
     }
 }
 
diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index 1e85b597d..215a54b59 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -32,8 +32,7 @@ struct DeviceBackends;
 struct Backend {
     Backend(DeviceBackends& backends, World& src);
 
-    Cont2Config& kernel_configs() { return kernel_configs_; };
-    virtual std::unique_ptr<CodeGen> create_cg(const Cont2Config& config) = 0;
+    virtual std::unique_ptr<CodeGen> create_cg() = 0;
 
     Thorin& thorin() { return device_code_; }
     Importer& importer() { return *importer_; }

From a256ca6d7cbb9f70bad77c99c64debc804df7785 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 22 Jul 2024 18:54:35 +0200
Subject: [PATCH 275/342] c: fix broken syntax when accessing tuple components

---
 src/thorin/be/c/c.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 636eecfac..947551dc9 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -943,7 +943,7 @@ void CCodeGen::emit_access(Stream& s, const Type* agg_type, const Def* index, co
     } else if (agg_type->isa<IndefiniteArrayType>()) {
         s.fmt("[{}]", emit(index));
     } else if (agg_type->isa<TupleType>()) {
-        s.fmt("{}e{}", prefix, emit_constant(index));
+        s.fmt("{}e{}", prefix, primlit_value<size_t>(index));
     } else if (agg_type->isa<StructType>()) {
         s.fmt("{}{}", prefix, agg_type->as<StructType>()->op_name(primlit_value<size_t>(index)));
     } else if (agg_type->isa<VectorType>()) {

From c4b5b8de68a6727942d74b9a09927e64c11c2352 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 22 Jul 2024 18:55:19 +0200
Subject: [PATCH 276/342] backends: fix broken imported cont recovery logic

---
 src/thorin/be/codegen.cpp | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 5953283d4..9125830e7 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -22,16 +22,15 @@ namespace thorin {
 void Backend::prepare_kernel_configs() {
     device_code_.opt();
 
-    auto& externals = backends_.world().externals();
+    auto conts = device_code_.world().copy_continuations();
     for (auto continuation : kernels_) {
         // recover the imported continuation (lost after the call to opt)
         Continuation* imported = nullptr;
-        for (auto [_, def] : externals) {
-            auto exported = def->isa<Continuation>();
-            if (!exported) continue;
-            if (!exported->has_body()) continue;
-            if (exported->name() == continuation->name())
-                imported = exported;
+        for (auto original_cont : conts) {
+            if (!original_cont) continue;
+            if (!original_cont->has_body()) continue;
+            if (original_cont->name() == continuation->name())
+                imported = original_cont;
         }
         if (!imported) continue;
 

From 129895a425b17b305b52294b2cf12fb84dc2f7e4 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 22 Jul 2024 19:00:03 +0200
Subject: [PATCH 277/342] c: get rid of typedefs and move primtype emission
 logic into a fn

---
 src/thorin/be/c/c.cpp | 159 +++++++++++++++++++++++++-----------------
 1 file changed, 95 insertions(+), 64 deletions(-)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index 947551dc9..b945bace8 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -103,6 +103,7 @@ class CCodeGen : public thorin::Emitter<std::string, std::string, BB, CCodeGen>
     void finalize(Continuation*);
 
 private:
+    void convert_primtype(StringStream&s, PrimTypeTag tag, int len);
     std::string convert(const Type*);
     std::string addr_space_prefix(AddrSpace);
     std::string constructor_prefix(const Type*);
@@ -210,6 +211,97 @@ bool CCodeGen::get_interface(HlsInterface &interface, HlsInterface &gmem) {
     return false;
 }
 
+inline const char* stddef_primtype_name(PrimTypeTag tag) {
+    switch (tag) {
+        case PrimType_ps8:  case PrimType_qs8:  return "int8_t";
+        case PrimType_pu8:  case PrimType_qu8:  return "uint8_t";
+        case PrimType_ps16: case PrimType_qs16: return "int16_t";
+        case PrimType_pu16: case PrimType_qu16: return "uint16_t";
+        case PrimType_ps32: case PrimType_qs32: return "int32_t";
+        case PrimType_pu32: case PrimType_qu32: return "uint32_t";
+        case PrimType_ps64: case PrimType_qs64: return "int64_t";
+        case PrimType_pu64: case PrimType_qu64: return "uint64_t";
+        case PrimType_pf16: case PrimType_qf16: return "half";
+        default: THORIN_UNREACHABLE;
+    }
+}
+
+inline const char* cuda_scalar_primtype(PrimTypeTag tag) {
+    switch (tag) {
+        case PrimType_ps8:  case PrimType_qs8:  return "char";
+        case PrimType_pu8:  case PrimType_qu8:  return "unsigned char";
+        case PrimType_ps16: case PrimType_qs16: return "short";
+        case PrimType_pu16: case PrimType_qu16: return "unsigned short";
+        case PrimType_ps32: case PrimType_qs32: return "int";
+        case PrimType_pu32: case PrimType_qu32: return "unsigned int";
+        case PrimType_ps64: case PrimType_qs64: return "long long";
+        case PrimType_pu64: case PrimType_qu64: return "unsigned long long";
+        case PrimType_pf16: case PrimType_qf16: return "f16"; // typedef'd with macro magic
+        default: THORIN_UNREACHABLE;
+    }
+}
+
+/// OpenCL uses these for scalar and vectors.
+/// CUDA actually uses the same prefixes for its vectors
+/// See
+/// https://registry.khronos.org/OpenCL/sdk/3.0/docs/man/html/vectorDataTypes.html
+/// https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#built-in-vector-types
+inline const char* opencl_cuda_vectorbase(PrimTypeTag tag) {
+    switch (tag) {
+        case PrimType_ps8:  case PrimType_qs8:  return "char";
+        case PrimType_pu8:  case PrimType_qu8:  return "uchar";
+        case PrimType_ps16: case PrimType_qs16: return "short";
+        case PrimType_pu16: case PrimType_qu16: return "ushort";
+        case PrimType_ps32: case PrimType_qs32: return "int";
+        case PrimType_pu32: case PrimType_qu32: return "uint";
+        case PrimType_ps64: case PrimType_qs64: return "long";
+        case PrimType_pu64: case PrimType_qu64: return "ulong";
+        case PrimType_pf16: case PrimType_qf16: return "half"; // NB: cuda has no vectors of half.
+        default: THORIN_UNREACHABLE;
+    }
+}
+
+void CCodeGen::convert_primtype(StringStream& s, PrimTypeTag tag, int len) {
+    assert(len > 0);
+
+    // Enable special code paths for f16 and f64
+    switch (tag) {
+        case PrimType_pf16: case PrimType_qf16: use_fp_16_ = true; break;
+        case PrimType_pf64: case PrimType_qf64: use_fp_64_ = true; break;
+        default: break;
+    }
+
+    // 'bool', 'float' and 'double' are identical everywhere
+    switch (tag) {
+        case PrimType_bool: s << "bool"; break;
+        case PrimType_pf32: case PrimType_qf32: s << "float"; break;
+        case PrimType_pf64: case PrimType_qf64: s << "double"; break;
+        default: {
+            if (lang_ == Lang::CUDA && len == 1)
+                s << cuda_scalar_primtype(tag);
+            else if (lang_ == Lang::CUDA || lang_ == Lang::OpenCL)
+                s << opencl_cuda_vectorbase(tag);
+            else
+                s << stddef_primtype_name(tag);
+            break;
+        }
+    }
+
+    // length suffixes
+    if (len == 1)
+        return;
+    switch (lang_) {
+        case Lang::CUDA:
+        case Lang::OpenCL:
+            s << len;
+            break;
+        case Lang::HLS:
+        case Lang::C99:
+            s.fmt(" __attribute__ ((ext_vector_size ({})))", len);
+            break;
+    }
+}
+
 /*
  * convert
  */
@@ -223,23 +315,7 @@ std::string CCodeGen::convert(const Type* type) {
     if (type == world().unit_type() || type->isa<MemType>() || type->isa<FrameType>())
         s << "void";
     else if (auto primtype = type->isa<PrimType>()) {
-        switch (primtype->primtype_tag()) {
-            case PrimType_bool:                     s << "bool";                     break;
-            case PrimType_ps8:  case PrimType_qs8:  s <<   "i8";                     break;
-            case PrimType_pu8:  case PrimType_qu8:  s <<   "u8";                     break;
-            case PrimType_ps16: case PrimType_qs16: s <<  "i16";                     break;
-            case PrimType_pu16: case PrimType_qu16: s <<  "u16";                     break;
-            case PrimType_ps32: case PrimType_qs32: s <<  "i32";                     break;
-            case PrimType_pu32: case PrimType_qu32: s <<  "u32";                     break;
-            case PrimType_ps64: case PrimType_qs64: s <<  "i64";                     break;
-            case PrimType_pu64: case PrimType_qu64: s <<  "u64";                     break;
-            case PrimType_pf16: case PrimType_qf16: s <<  "f16";  use_fp_16_ = true; break;
-            case PrimType_pf32: case PrimType_qf32: s <<  "f32";                     break;
-            case PrimType_pf64: case PrimType_qf64: s <<  "f64";  use_fp_64_ = true; break;
-            default: THORIN_UNREACHABLE;
-        }
-        if (primtype->is_vector())
-            s << primtype->length();
+        convert_primtype(s, primtype->primtype_tag(), vector_length(primtype));
     } else if (auto array = type->isa<IndefiniteArrayType>()) {
         return types_[type] = convert(array->elem_type()); // IndefiniteArrayType always occurs within a pointer
     } else if (type->isa<FnType>()) {
@@ -402,21 +478,6 @@ void CCodeGen::emit_module() {
             stream_ << "#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n";
         if (use_fp_64_)
             stream_ << "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n";
-
-        stream_.fmt(    "\n"
-                        "typedef   char  i8;\n"
-                        "typedef  uchar  u8;\n"
-                        "typedef  short i16;\n"
-                        "typedef ushort u16;\n"
-                        "typedef    int i32;\n"
-                        "typedef   uint u32;\n"
-                        "typedef   long i64;\n"
-                        "typedef  ulong u64;\n");
-        if (use_fp_16_)
-            stream_.fmt("typedef   half f16;\n");
-        stream_.fmt(    "typedef  float f32;\n");
-        if (use_fp_64_)
-            stream_.fmt("typedef double f64;\n");
     }
 
     stream_.endl();
@@ -441,45 +502,15 @@ void CCodeGen::emit_module() {
             stream_.fmt("#include <hls_half.h>\n");
     }
 
-    if (lang_ == Lang::C99 || lang_ == Lang::HLS) {
-        stream_.fmt(    "\n"
-                        "typedef   int8_t  i8;\n"
-                        "typedef  uint8_t  u8;\n"
-                        "typedef  int16_t i16;\n"
-                        "typedef uint16_t u16;\n"
-                        "typedef  int32_t i32;\n"
-                        "typedef uint32_t u32;\n"
-                        "typedef  int64_t i64;\n"
-                        "typedef uint64_t u64;\n"
-                        "typedef    float f32;\n"
-                        "typedef   double f64;\n"
-                        "\n");
-
-         if (use_fp_16_ && lang_ == Lang::HLS)
-            stream_.fmt("typedef     half f16;\n");
-    }
-
     if (lang_ == Lang::CUDA) {
-        if (use_fp_16_)
+        if (use_fp_16_) {
             stream_.fmt("#include <cuda_fp16.h>\n\n");
-        stream_.fmt(    "typedef               char  i8;\n"
-                        "typedef      unsigned char  u8;\n"
-                        "typedef              short i16;\n"
-                        "typedef     unsigned short u16;\n"
-                        "typedef                int i32;\n"
-                        "typedef       unsigned int u32;\n"
-                        "typedef          long long i64;\n"
-                        "typedef unsigned long long u64;\n"
-                        "\n");
-        if (use_fp_16_)
             stream_.fmt("#if __CUDACC_VER_MAJOR__ <= 8\n"
                         "typedef               half f16;\n"
                         "#else\n"
                         "typedef         __half_raw f16;\n"
                         "#endif\n");
-        stream_.fmt(    "typedef              float f32;\n"
-                        "typedef             double f64;\n"
-                        "\n");
+        }
     }
 
     if (lang_ == Lang::CUDA || lang_ == Lang::HLS) {

From ff8060fe34b7a4abc144575a52aae2adb8afcfb6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 25 Jul 2024 14:59:38 +0200
Subject: [PATCH 278/342] added OpenCL_SPIRV backend

---
 src/thorin/be/codegen.cpp   | 20 ++++++++++++++++++++
 src/thorin/be/llvm/llvm.cpp |  1 +
 src/thorin/config.h.in      |  1 +
 src/thorin/continuation.cpp |  1 +
 src/thorin/continuation.h   |  1 +
 5 files changed, 24 insertions(+)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 9125830e7..da96a3b19 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -14,6 +14,10 @@
 #undef nodes
 #endif
 
+#if THORIN_ENABLE_SPIRV
+#include "thorin/be/spirv/spirv.h"
+#endif
+
 #include "thorin/transform/hls_channels.h"
 #include "thorin/transform/hls_kernel_launch.h"
 
@@ -136,6 +140,19 @@ struct OpenCLBackend : public Backend {
     }
 };
 
+#if THORIN_ENABLE_SPIRV
+struct OpenCLSPIRVBackend : public Backend {
+    explicit OpenCLSPIRVBackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::OpenCL_SPIRV, *this, get_gpu_kernel_config);
+    }
+
+    std::unique_ptr<CodeGen> create_cg() override {
+        spirv::Target target;
+        return std::make_unique<spirv::CodeGen>(device_code_, target, backends_.debug(), &kernel_configs_);
+    }
+};
+#endif
+
 #if THORIN_ENABLE_LLVM
 struct AMDHSABackend : public Backend {
     explicit AMDHSABackend(DeviceBackends& b, World& src) : Backend(b, src) {
@@ -238,6 +255,9 @@ DeviceBackends::DeviceBackends(thorin::World& world, int opt, bool debug, std::s
 #endif
 #if THORIN_ENABLE_SHADY
     register_backend(std::make_unique<ShadyBackend>(*this, world))
+#endif
+#if THORIN_ENABLE_SPIRV
+    register_backend(std::make_unique<OpenCLSPIRVBackend>(*this, world));
 #endif
     register_backend(std::make_unique<HLSBackend>(*this, world, hls_flags));
 
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index afe2a95af..89032b405 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1306,6 +1306,7 @@ std::vector<llvm::Value*> CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder,
         case Intrinsic::CUDA:         runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation); break;
         case Intrinsic::NVVM:         runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation); break;
         case Intrinsic::OpenCL:       runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation); break;
+        case Intrinsic::OpenCL_SPIRV: runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv",    continuation); break;
         case Intrinsic::AMDGPUHSA:    runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation); break;
         case Intrinsic::AMDGPUPAL:    runtime_->emit_host_code(*this, irbuilder, Runtime::PAL_PLATFORM,    ".amdgpu", continuation); break;
         case Intrinsic::ShadyCompute: runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation); break;
diff --git a/src/thorin/config.h.in b/src/thorin/config.h.in
index e8f2dfc42..3a6b28644 100644
--- a/src/thorin/config.h.in
+++ b/src/thorin/config.h.in
@@ -8,6 +8,7 @@
 #cmakedefine01 THORIN_ENABLE_JSON
 #cmakedefine01 THORIN_ENABLE_RV
 #cmakedefine01 THORIN_ENABLE_SHADY
+#cmakedefine01 THORIN_ENABLE_SPIRV
 #cmakedefine01 THORIN_ENABLE_RLIMITS
 
 #endif
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 9001206b4..1d97faa9a 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -248,6 +248,7 @@ void Continuation::set_intrinsic() {
     if      (name() == "cuda")           attributes().intrinsic = Intrinsic::CUDA;
     else if (name() == "nvvm")           attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")         attributes().intrinsic = Intrinsic::OpenCL;
+    else if (name() == "opencl_spirv")   attributes().intrinsic = Intrinsic::OpenCL_SPIRV;
     else if (name() == "amdgpu_hsa")     attributes().intrinsic = Intrinsic::AMDGPUHSA;
     else if (name() == "amdgpu_pal")     attributes().intrinsic = Intrinsic::AMDGPUPAL;
     else if (name() == "shady_compute")  attributes().intrinsic = Intrinsic::ShadyCompute;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 73e47e36a..d537134fe 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -95,6 +95,7 @@ enum class Intrinsic : uint8_t {
     CUDA = AcceleratorBegin,    ///< Internal CUDA-Backend.
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
+    OpenCL_SPIRV,               ///< Internal OpenCL-Backend.
     AMDGPUHSA,                  ///< Internal AMDGPU-HSA-Backend.
     AMDGPUPAL,                  ///< Internal AMDGPU-PAL-Backend.
     ShadyCompute,               ///< Internal Shady Compute Backend.

From cfa14c5eaa53d3fcef76c2037d20c6c1b3f70a79 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 25 Jul 2024 15:00:13 +0200
Subject: [PATCH 279/342] commented out weird LEA address space stuff

---
 src/thorin/be/spirv/spirv.cpp | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 1e4e66532..af485c169 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -631,14 +631,14 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     } else if (auto enter = def->isa<Enter>()) {
         return emit_unsafe(enter->mem());
     } else if (auto lea = def->isa<LEA>()) {
-        switch (lea->type()->addr_space()) {
-            case AddrSpace::Global:
-            case AddrSpace::Shared:
-                break;
-            default:
-                world().ELOG("LEA is only allowed in global & shared address spaces");
-                break;
-        }
+        //switch (lea->type()->addr_space()) {
+        //    case AddrSpace::Global:
+        //    case AddrSpace::Shared:
+        //        break;
+        //    default:
+        //        world().ELOG("LEA is only allowed in global & shared address spaces");
+        //        break;
+        //}
         auto type = convert(lea->type()).id;
         auto offset = emit(lea->index());
         return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {});

From ca9b0d60f4f636cfdb4c1a5b2a7d9a6ecc1c8901 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 25 Jul 2024 15:00:29 +0200
Subject: [PATCH 280/342] c/opencl: handle ptr as casts

---
 src/thorin/be/c/c.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/thorin/be/c/c.cpp b/src/thorin/be/c/c.cpp
index b945bace8..398e6dac4 100644
--- a/src/thorin/be/c/c.cpp
+++ b/src/thorin/be/c/c.cpp
@@ -1088,6 +1088,8 @@ std::string CCodeGen::emit_def(BB* bb, const Def* def) {
             s.fmt("(({})->e)", src);
         } else if (s_ptr && d_ptr && s_ptr->addr_space() == d_ptr->addr_space()) {
             s.fmt("(({}) {})", d_t, src);
+        } else if (s_ptr && d_ptr && s_ptr->addr_space() != d_ptr->addr_space() && lang_ == Lang::OpenCL) {
+            s.fmt("(({}) ((size_t) {}))", d_t, src);
         } else if (conv->isa<Cast>()) {
             auto s_prim = s_type->isa<PrimType>();
             auto d_prim = d_type->isa<PrimType>();

From 0cb3002cba991713df64cd5062621d7feffae226 Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Tue, 1 Oct 2024 12:37:21 +0200
Subject: [PATCH 281/342] Slight updates to switch to llvm 18.

---
 src/thorin/be/llvm/cpu.cpp      |  2 +-
 src/thorin/be/llvm/llvm.cpp     |  4 ++--
 src/thorin/be/llvm/nvvm.cpp     |  4 ++--
 src/thorin/be/llvm/parallel.cpp |  6 +++---
 src/thorin/be/llvm/runtime.cpp  | 32 ++++++++++++++++----------------
 5 files changed, 24 insertions(+), 24 deletions(-)

diff --git a/src/thorin/be/llvm/cpu.cpp b/src/thorin/be/llvm/cpu.cpp
index 4a2d6dfbc..133d5ddfe 100644
--- a/src/thorin/be/llvm/cpu.cpp
+++ b/src/thorin/be/llvm/cpu.cpp
@@ -1,7 +1,7 @@
 #include "thorin/be/llvm/cpu.h"
 
 #include <llvm/MC/TargetRegistry.h>
-#include <llvm/Support/Host.h>
+#include <llvm/TargetParser/Host.h>
 #include <llvm/Support/TargetSelect.h>
 #include <llvm/Target/TargetMachine.h>
 #include <llvm/Target/TargetOptions.h>
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index ecea16edd..31f4440c9 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -4,7 +4,7 @@
 #include <stdexcept>
 #include <unordered_map> // TODO don't used std::unordered_*
 
-#include <llvm/ADT/Triple.h>
+#include <llvm/TargetParser/Triple.h>
 #include <llvm/IR/Constant.h>
 #include <llvm/IR/Constants.h>
 #include <llvm/IR/Function.h>
@@ -17,7 +17,7 @@
 #include <llvm/IR/Type.h>
 #include <llvm/IR/Verifier.h>
 #include <llvm/Passes/PassBuilder.h>
-#include <llvm/Support/Host.h>
+#include <llvm/TargetParser/Host.h>
 #include <llvm/Support/Path.h>
 #include <llvm/Support/raw_os_ostream.h>
 #include <llvm/Support/FileSystem.h>
diff --git a/src/thorin/be/llvm/nvvm.cpp b/src/thorin/be/llvm/nvvm.cpp
index 88c488da8..7c39c160a 100644
--- a/src/thorin/be/llvm/nvvm.cpp
+++ b/src/thorin/be/llvm/nvvm.cpp
@@ -3,7 +3,7 @@
 #include <sstream>
 #include <unordered_map> // TODO don't used std::unordered_*
 
-#include <llvm/ADT/Triple.h>
+#include <llvm/TargetParser/Triple.h>
 #include <llvm/IR/Function.h>
 #include <llvm/IR/Metadata.h>
 #include <llvm/IR/Module.h>
@@ -11,7 +11,7 @@
 #include <llvm/IR/GlobalVariable.h>
 #include <llvm/IR/InlineAsm.h>
 #include <llvm/IRReader/IRReader.h>
-#include <llvm/Support/Host.h>
+#include <llvm/TargetParser/Host.h>
 #include <llvm/Support/SourceMgr.h>
 
 #include "thorin/primop.h"
diff --git a/src/thorin/be/llvm/parallel.cpp b/src/thorin/be/llvm/parallel.cpp
index 5b83bdcd2..96ab7142d 100644
--- a/src/thorin/be/llvm/parallel.cpp
+++ b/src/thorin/be/llvm/parallel.cpp
@@ -51,7 +51,7 @@ void CodeGen::emit_parallel(llvm::IRBuilder<>& irbuilder, Continuation* continua
 
     // create wrapper function and call the runtime
     // wrapper(void* closure, int lower, int upper)
-    llvm::Type* wrapper_arg_types[] = { irbuilder.getInt8PtrTy(0), irbuilder.getInt32Ty(), irbuilder.getInt32Ty() };
+    llvm::Type* wrapper_arg_types[] = { irbuilder.getPtrTy(), irbuilder.getInt32Ty(), irbuilder.getInt32Ty() };
     auto wrapper_ft = llvm::FunctionType::get(irbuilder.getVoidTy(), wrapper_arg_types, false);
     auto wrapper_name = kernel->unique_name() + "_parallel_for";
     auto wrapper = (llvm::Function*)module_->getOrInsertFunction(wrapper_name, wrapper_ft).getCallee()->stripPointerCasts();
@@ -143,7 +143,7 @@ void CodeGen::emit_fibers(llvm::IRBuilder<>& irbuilder, Continuation* continuati
 
     // create wrapper function and call the runtime
     // wrapper(void* closure, int lower, int upper)
-    llvm::Type* wrapper_arg_types[] = { irbuilder.getInt8PtrTy(0), irbuilder.getInt32Ty(), irbuilder.getInt32Ty() };
+    llvm::Type* wrapper_arg_types[] = { irbuilder.getPtrTy(), irbuilder.getInt32Ty(), irbuilder.getInt32Ty() };
     auto wrapper_ft = llvm::FunctionType::get(irbuilder.getVoidTy(), wrapper_arg_types, false);
     auto wrapper_name = kernel->unique_name() + "_fibers";
     auto wrapper = (llvm::Function*)module_->getOrInsertFunction(wrapper_name, wrapper_ft).getCallee()->stripPointerCasts();
@@ -225,7 +225,7 @@ llvm::Value* CodeGen::emit_spawn(llvm::IRBuilder<>& irbuilder, Continuation* con
 
     // create wrapper function and call the runtime
     // wrapper(void* closure)
-    llvm::Type* wrapper_arg_types[] = { irbuilder.getInt8PtrTy(0) };
+    llvm::Type* wrapper_arg_types[] = { irbuilder.getPtrTy() };
     auto wrapper_ft = llvm::FunctionType::get(irbuilder.getVoidTy(), wrapper_arg_types, false);
     auto wrapper_name = kernel->unique_name() + "_spawn_thread";
     auto wrapper = (llvm::Function*)module_->getOrInsertFunction(wrapper_name, wrapper_ft).getCallee()->stripPointerCasts();
diff --git a/src/thorin/be/llvm/runtime.cpp b/src/thorin/be/llvm/runtime.cpp
index 375713091..d4467edf8 100644
--- a/src/thorin/be/llvm/runtime.cpp
+++ b/src/thorin/be/llvm/runtime.cpp
@@ -85,11 +85,11 @@ void Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Plat
     const size_t num_kernel_args = body->num_args() - LaunchArgs::Num;
 
     // allocate argument pointers, sizes, and types
-    llvm::Value* args   = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt8PtrTy(), num_kernel_args), "args");
-    llvm::Value* sizes  = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt32Ty(),   num_kernel_args), "sizes");
-    llvm::Value* aligns = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt32Ty(),   num_kernel_args), "aligns");
-    llvm::Value* allocs = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt32Ty(),   num_kernel_args), "allocs");
-    llvm::Value* types  = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt8Ty(),    num_kernel_args), "types");
+    llvm::Value* args   = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getPtrTy(),   num_kernel_args), "args");
+    llvm::Value* sizes  = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt32Ty(), num_kernel_args), "sizes");
+    llvm::Value* aligns = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt32Ty(), num_kernel_args), "aligns");
+    llvm::Value* allocs = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt32Ty(), num_kernel_args), "allocs");
+    llvm::Value* types  = code_gen.emit_alloca(builder, llvm::ArrayType::get(builder.getInt8Ty(),  num_kernel_args), "types");
 
     // fill array of arguments
     for (size_t i = 0; i < num_kernel_args; ++i) {
@@ -109,7 +109,7 @@ void Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Plat
             if (!contains_ptrtype(target_arg->type()))
                 world.wdef(target_arg, "argument '{}' of aggregate type '{}' contains pointer (not supported in OpenCL 1.2)", target_arg, target_arg->type());
 
-            void_ptr = builder.CreatePointerCast(alloca, builder.getInt8PtrTy());
+            void_ptr = builder.CreatePointerCast(alloca, builder.getPtrTy());
             arg_type = KernelArgType::Struct;
         } else if (target_arg->type()->isa<PtrType>()) {
             auto ptr = target_arg->type()->as<PtrType>();
@@ -118,17 +118,17 @@ void Runtime::emit_host_code(CodeGen& code_gen, llvm::IRBuilder<>& builder, Plat
             if (!rtype->isa<ArrayType>())
                 world.edef(target_arg, "currently only pointers to arrays supported as kernel argument; argument has different type: {}", ptr);
 
-            auto alloca = code_gen.emit_alloca(builder, builder.getInt8PtrTy(), target_arg->name());
-            auto target_ptr = builder.CreatePointerCast(target_val, builder.getInt8PtrTy());
+            auto alloca = code_gen.emit_alloca(builder, builder.getPtrTy(), target_arg->name());
+            auto target_ptr = builder.CreatePointerCast(target_val, builder.getPtrTy());
             builder.CreateStore(target_ptr, alloca);
-            void_ptr = builder.CreatePointerCast(alloca, builder.getInt8PtrTy());
+            void_ptr = builder.CreatePointerCast(alloca, builder.getPtrTy());
             arg_type = KernelArgType::Ptr;
         } else {
             // normal variable
             auto alloca = code_gen.emit_alloca(builder, target_val->getType(), target_arg->name());
             builder.CreateStore(target_val, alloca);
 
-            void_ptr = builder.CreatePointerCast(alloca, builder.getInt8PtrTy());
+            void_ptr = builder.CreatePointerCast(alloca, builder.getPtrTy());
             arg_type = KernelArgType::Val;
         }
 
@@ -203,8 +203,8 @@ llvm::Value* Runtime::parallel_for(
 {
     llvm::Value* parallel_args[] = {
         num_threads, lower, upper,
-        builder.CreatePointerCast(closure_ptr, builder.getInt8PtrTy()),
-        builder.CreatePointerCast(fun_ptr, builder.getInt8PtrTy())
+        builder.CreatePointerCast(closure_ptr, builder.getPtrTy()),
+        builder.CreatePointerCast(fun_ptr, builder.getPtrTy())
     };
     return builder.CreateCall(get(code_gen, "anydsl_parallel_for"), parallel_args);
 }
@@ -215,16 +215,16 @@ llvm::Value* Runtime::spawn_fibers(
 {
     llvm::Value* fibers_args[] = {
         num_threads, num_blocks, num_warps,
-        builder.CreatePointerCast(closure_ptr, builder.getInt8PtrTy()),
-        builder.CreatePointerCast(fun_ptr, builder.getInt8PtrTy())
+        builder.CreatePointerCast(closure_ptr, builder.getPtrTy()),
+        builder.CreatePointerCast(fun_ptr, builder.getPtrTy())
     };
     return builder.CreateCall(get(code_gen, "anydsl_fibers_spawn"), fibers_args);
 }
 
 llvm::Value* Runtime::spawn_thread(CodeGen& code_gen, llvm::IRBuilder<>& builder, llvm::Value* closure_ptr, llvm::Value* fun_ptr) {
     llvm::Value* spawn_args[] = {
-        builder.CreatePointerCast(closure_ptr, builder.getInt8PtrTy()),
-        builder.CreatePointerCast(fun_ptr, builder.getInt8PtrTy())
+        builder.CreatePointerCast(closure_ptr, builder.getPtrTy()),
+        builder.CreatePointerCast(fun_ptr, builder.getPtrTy())
     };
     return builder.CreateCall(get(code_gen, "anydsl_spawn_thread"), spawn_args);
 }

From 11ab8dfe1188a49924b0195d52f0f8c6a49be2f3 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 16 Oct 2024 14:11:14 +0200
Subject: [PATCH 282/342] spirv: fix Cast def emission

---
 src/thorin/be/spirv/spirv.cpp | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index af485c169..be1b22532 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -738,8 +738,8 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
 
             auto src_kind = classify_primtype(src_prim);
             auto dst_kind = classify_primtype(dst_prim);
-            size_t src_bitwidth = conv_src_type.layout->size;
-            size_t dst_bitwidth = conv_src_type.layout->size;
+            size_t src_bitwidth = conv_src_type.layout->size * 8;
+            size_t dst_bitwidth = conv_src_type.layout->size * 8;
 
             SpvId data = emit(cast->from());
 
@@ -796,6 +796,8 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
                     }
                 }
             }
+
+            return data;
         } else THORIN_UNREACHABLE;
     } else if (def->isa<Bottom>()) {
         return bb->undef(convert(def->type()).id);

From 80291f022c5e67749dab93bc346a4313d4776720 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 16 Oct 2024 14:11:30 +0200
Subject: [PATCH 283/342] added LevelZero runtime platform

---
 src/thorin/be/codegen.cpp    | 11 ++++++++++
 src/thorin/be/llvm/llvm.cpp  | 39 ++++++++++++++++++------------------
 src/thorin/be/llvm/runtime.h |  1 +
 src/thorin/continuation.cpp  |  1 +
 src/thorin/continuation.h    |  1 +
 5 files changed, 34 insertions(+), 19 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index da96a3b19..82213c87e 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -151,6 +151,17 @@ struct OpenCLSPIRVBackend : public Backend {
         return std::make_unique<spirv::CodeGen>(device_code_, target, backends_.debug(), &kernel_configs_);
     }
 };
+
+struct LevelZeroSPIRVBackend : public Backend {
+    explicit LevelZeroSPIRVBackend(DeviceBackends& b, World& src) : Backend(b, src) {
+        b.register_intrinsic(Intrinsic::LevelZero_SPIRV, *this, get_gpu_kernel_config);
+    }
+
+    std::unique_ptr<CodeGen> create_cg() override {
+        spirv::Target target;
+        return std::make_unique<spirv::CodeGen>(device_code_, target, backends_.debug(), &kernel_configs_);
+    }
+};
 #endif
 
 #if THORIN_ENABLE_LLVM
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 89032b405..7b66725d1 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1296,25 +1296,26 @@ std::vector<llvm::Value*> CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder,
     }
 
     switch (callee->intrinsic()) {
-        case Intrinsic::Atomic:       return { emit_atomic(irbuilder, continuation) };
-        case Intrinsic::AtomicLoad:   return { emit_atomic_load(irbuilder, continuation) };
-        case Intrinsic::AtomicStore:  emit_atomic_store(irbuilder, continuation); break;
-        case Intrinsic::CmpXchg:      return emit_cmpxchg(irbuilder, continuation, false);
-        case Intrinsic::CmpXchgWeak:  return emit_cmpxchg(irbuilder, continuation, true);
-        case Intrinsic::Fence:        emit_fence(irbuilder, continuation); break;
-        case Intrinsic::Reserve:      return { emit_reserve(irbuilder, continuation) };
-        case Intrinsic::CUDA:         runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".cu",     continuation); break;
-        case Intrinsic::NVVM:         runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,   ".nvvm",   continuation); break;
-        case Intrinsic::OpenCL:       runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".cl",     continuation); break;
-        case Intrinsic::OpenCL_SPIRV: runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM, ".spv",    continuation); break;
-        case Intrinsic::AMDGPUHSA:    runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,    ".amdgpu", continuation); break;
-        case Intrinsic::AMDGPUPAL:    runtime_->emit_host_code(*this, irbuilder, Runtime::PAL_PLATFORM,    ".amdgpu", continuation); break;
-        case Intrinsic::ShadyCompute: runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,  ".shady",  continuation); break;
-        case Intrinsic::HLS:          emit_hls(irbuilder, continuation);      break;
-        case Intrinsic::Parallel:     emit_parallel(irbuilder, continuation); break;
-        case Intrinsic::Fibers:       emit_fibers(irbuilder, continuation);   break;
-        case Intrinsic::Spawn:        return { emit_spawn(irbuilder, continuation) };
-        case Intrinsic::Sync:         emit_sync(irbuilder, continuation);     break;
+        case Intrinsic::Atomic:          return { emit_atomic(irbuilder, continuation) };
+        case Intrinsic::AtomicLoad:      return { emit_atomic_load(irbuilder, continuation) };
+        case Intrinsic::AtomicStore:     emit_atomic_store(irbuilder, continuation); break;
+        case Intrinsic::CmpXchg:         return emit_cmpxchg(irbuilder, continuation, false);
+        case Intrinsic::CmpXchgWeak:     return emit_cmpxchg(irbuilder, continuation, true);
+        case Intrinsic::Fence:           emit_fence(irbuilder, continuation); break;
+        case Intrinsic::Reserve:         return { emit_reserve(irbuilder, continuation) };
+        case Intrinsic::CUDA:            runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,       ".cu",     continuation); break;
+        case Intrinsic::NVVM:            runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,       ".nvvm",   continuation); break;
+        case Intrinsic::OpenCL:          runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM,     ".cl",     continuation); break;
+        case Intrinsic::OpenCL_SPIRV:    runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM,     ".spv",    continuation); break;
+        case Intrinsic::LevelZero_SPIRV: runtime_->emit_host_code(*this, irbuilder, Runtime::LEVEL_ZERO_PLATFORM, ".spv",    continuation); break;
+        case Intrinsic::AMDGPUHSA:       runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,        ".amdgpu", continuation); break;
+        case Intrinsic::AMDGPUPAL:       runtime_->emit_host_code(*this, irbuilder, Runtime::PAL_PLATFORM,        ".amdgpu", continuation); break;
+        case Intrinsic::ShadyCompute:    runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,      ".shady",  continuation); break;
+        case Intrinsic::HLS:             emit_hls(irbuilder, continuation);      break;
+        case Intrinsic::Parallel:        emit_parallel(irbuilder, continuation); break;
+        case Intrinsic::Fibers:          emit_fibers(irbuilder, continuation);   break;
+        case Intrinsic::Spawn:           return { emit_spawn(irbuilder, continuation) };
+        case Intrinsic::Sync:            emit_sync(irbuilder, continuation);     break;
 #if THORIN_ENABLE_RV
         case Intrinsic::Vectorize:    emit_vectorize_continuation(irbuilder, continuation); break;
 #else
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index 9038040a0..551d06797 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -24,6 +24,7 @@ class Runtime {
         OPENCL_PLATFORM,
         HSA_PLATFORM,
         PAL_PLATFORM,
+        LEVEL_ZERO_PLATFORM,
         SHADY_PLATFORM,
     };
 
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 1d97faa9a..d99dd0ae9 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -249,6 +249,7 @@ void Continuation::set_intrinsic() {
     else if (name() == "nvvm")           attributes().intrinsic = Intrinsic::NVVM;
     else if (name() == "opencl")         attributes().intrinsic = Intrinsic::OpenCL;
     else if (name() == "opencl_spirv")   attributes().intrinsic = Intrinsic::OpenCL_SPIRV;
+    else if (name() == "levelzero")      attributes().intrinsic = Intrinsic::LevelZero_SPIRV;
     else if (name() == "amdgpu_hsa")     attributes().intrinsic = Intrinsic::AMDGPUHSA;
     else if (name() == "amdgpu_pal")     attributes().intrinsic = Intrinsic::AMDGPUPAL;
     else if (name() == "shady_compute")  attributes().intrinsic = Intrinsic::ShadyCompute;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index d537134fe..fb747c5b7 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -96,6 +96,7 @@ enum class Intrinsic : uint8_t {
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
     OpenCL_SPIRV,               ///< Internal OpenCL-Backend.
+    LevelZero_SPIRV,            ///< Internal SPIRV for Level0-Backend.
     AMDGPUHSA,                  ///< Internal AMDGPU-HSA-Backend.
     AMDGPUPAL,                  ///< Internal AMDGPU-PAL-Backend.
     ShadyCompute,               ///< Internal Shady Compute Backend.

From a9757904024addf77f8b6db7ddfcd9fd8d4e3698 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 16 Oct 2024 14:33:37 +0200
Subject: [PATCH 284/342] move runtime enums outside of llvm backend

---
 src/thorin/CMakeLists.txt     |  1 +
 src/thorin/be/codegen.cpp     |  1 +
 src/thorin/be/kernel_config.h | 12 ------------
 src/thorin/be/llvm/llvm.cpp   | 16 ++++++++--------
 src/thorin/be/llvm/runtime.h  | 11 +----------
 src/thorin/be/runtime.h       | 29 +++++++++++++++++++++++++++++
 6 files changed, 40 insertions(+), 30 deletions(-)
 create mode 100644 src/thorin/be/runtime.h

diff --git a/src/thorin/CMakeLists.txt b/src/thorin/CMakeLists.txt
index 01044d98f..7910f0dfa 100644
--- a/src/thorin/CMakeLists.txt
+++ b/src/thorin/CMakeLists.txt
@@ -33,6 +33,7 @@ set(THORIN_SOURCES
     be/emitter.h
     be/c/c.cpp
     be/c/c.h
+    be/runtime.h
     be/kernel_config.h
     tables/allnodes.h
     tables/arithoptable.h
diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 82213c87e..4d9ec7c83 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -1,6 +1,7 @@
 #include "thorin/be/codegen.h"
 
 #include "thorin/be/c/c.h"
+#include "thorin/be/runtime.h"
 
 #if THORIN_ENABLE_LLVM
 #include "thorin/be/llvm/nvvm.h"
diff --git a/src/thorin/be/kernel_config.h b/src/thorin/be/kernel_config.h
index 787b579a6..c1d36aeb3 100644
--- a/src/thorin/be/kernel_config.h
+++ b/src/thorin/be/kernel_config.h
@@ -6,18 +6,6 @@
 
 namespace thorin {
 
-struct KernelLaunchArgs {
-    enum {
-        Mem = 0,
-        Device,
-        Space,
-        Config,
-        Body,
-        Return,
-        Num
-    };
-};
-
 class KernelConfig : public RuntimeCast<KernelConfig> {
 public:
     virtual ~KernelConfig() {}
diff --git a/src/thorin/be/llvm/llvm.cpp b/src/thorin/be/llvm/llvm.cpp
index 7b66725d1..362703d21 100644
--- a/src/thorin/be/llvm/llvm.cpp
+++ b/src/thorin/be/llvm/llvm.cpp
@@ -1303,14 +1303,14 @@ std::vector<llvm::Value*> CodeGen::emit_intrinsic(llvm::IRBuilder<>& irbuilder,
         case Intrinsic::CmpXchgWeak:     return emit_cmpxchg(irbuilder, continuation, true);
         case Intrinsic::Fence:           emit_fence(irbuilder, continuation); break;
         case Intrinsic::Reserve:         return { emit_reserve(irbuilder, continuation) };
-        case Intrinsic::CUDA:            runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,       ".cu",     continuation); break;
-        case Intrinsic::NVVM:            runtime_->emit_host_code(*this, irbuilder, Runtime::CUDA_PLATFORM,       ".nvvm",   continuation); break;
-        case Intrinsic::OpenCL:          runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM,     ".cl",     continuation); break;
-        case Intrinsic::OpenCL_SPIRV:    runtime_->emit_host_code(*this, irbuilder, Runtime::OPENCL_PLATFORM,     ".spv",    continuation); break;
-        case Intrinsic::LevelZero_SPIRV: runtime_->emit_host_code(*this, irbuilder, Runtime::LEVEL_ZERO_PLATFORM, ".spv",    continuation); break;
-        case Intrinsic::AMDGPUHSA:       runtime_->emit_host_code(*this, irbuilder, Runtime::HSA_PLATFORM,        ".amdgpu", continuation); break;
-        case Intrinsic::AMDGPUPAL:       runtime_->emit_host_code(*this, irbuilder, Runtime::PAL_PLATFORM,        ".amdgpu", continuation); break;
-        case Intrinsic::ShadyCompute:    runtime_->emit_host_code(*this, irbuilder, Runtime::SHADY_PLATFORM,      ".shady",  continuation); break;
+        case Intrinsic::CUDA:            runtime_->emit_host_code(*this, irbuilder, Platform::CUDA_PLATFORM,       ".cu",     continuation); break;
+        case Intrinsic::NVVM:            runtime_->emit_host_code(*this, irbuilder, Platform::CUDA_PLATFORM,       ".nvvm",   continuation); break;
+        case Intrinsic::OpenCL:          runtime_->emit_host_code(*this, irbuilder, Platform::OPENCL_PLATFORM,     ".cl",     continuation); break;
+        case Intrinsic::OpenCL_SPIRV:    runtime_->emit_host_code(*this, irbuilder, Platform::OPENCL_PLATFORM,     ".spv",    continuation); break;
+        case Intrinsic::LevelZero_SPIRV: runtime_->emit_host_code(*this, irbuilder, Platform::LEVEL_ZERO_PLATFORM, ".spv",    continuation); break;
+        case Intrinsic::AMDGPUHSA:       runtime_->emit_host_code(*this, irbuilder, Platform::HSA_PLATFORM,        ".amdgpu", continuation); break;
+        case Intrinsic::AMDGPUPAL:       runtime_->emit_host_code(*this, irbuilder, Platform::PAL_PLATFORM,        ".amdgpu", continuation); break;
+        case Intrinsic::ShadyCompute:    runtime_->emit_host_code(*this, irbuilder, Platform::SHADY_PLATFORM,      ".shady",  continuation); break;
         case Intrinsic::HLS:             emit_hls(irbuilder, continuation);      break;
         case Intrinsic::Parallel:        emit_parallel(irbuilder, continuation); break;
         case Intrinsic::Fibers:          emit_fibers(irbuilder, continuation);   break;
diff --git a/src/thorin/be/llvm/runtime.h b/src/thorin/be/llvm/runtime.h
index 551d06797..fca64aa6f 100644
--- a/src/thorin/be/llvm/runtime.h
+++ b/src/thorin/be/llvm/runtime.h
@@ -7,6 +7,7 @@
 #include <llvm/IR/Module.h>
 
 #include "thorin/world.h"
+#include "thorin/be/runtime.h"
 
 namespace thorin::llvm {
 
@@ -18,16 +19,6 @@ class Runtime {
 public:
     Runtime(llvm::LLVMContext&, llvm::Module&);
 
-    enum Platform {
-        CPU_PLATFORM,
-        CUDA_PLATFORM,
-        OPENCL_PLATFORM,
-        HSA_PLATFORM,
-        PAL_PLATFORM,
-        LEVEL_ZERO_PLATFORM,
-        SHADY_PLATFORM,
-    };
-
     /// Emits a call to anydsl_launch_kernel.
     llvm::Value* launch_kernel(
         CodeGen&, llvm::IRBuilder<>&, llvm::Value* device,
diff --git a/src/thorin/be/runtime.h b/src/thorin/be/runtime.h
new file mode 100644
index 000000000..68cdd9b97
--- /dev/null
+++ b/src/thorin/be/runtime.h
@@ -0,0 +1,29 @@
+#ifndef THORIN_RUNTIME_H
+#define THORIN_RUNTIME_H
+
+/// Backend-agnostic information to interface with the runtime component
+namespace thorin {
+
+enum Platform {
+    CPU_PLATFORM,
+    CUDA_PLATFORM,
+    OPENCL_PLATFORM,
+    HSA_PLATFORM,
+    PAL_PLATFORM,
+    LEVEL_ZERO_PLATFORM,
+    SHADY_PLATFORM,
+};
+
+enum KernelLaunchArgs {
+    Mem = 0,
+    Device,
+    Space,
+    Config,
+    Body,
+    Return,
+    Num
+};
+
+}
+
+#endif

From 6b41a16524e3ec5a28663e9920d511e992cde7b2 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 16 Oct 2024 14:56:52 +0200
Subject: [PATCH 285/342] spirv: use Ptr variant of OpAccessChain only where
 appropriate

---
 src/thorin/be/spirv/spirv.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index be1b22532..edae2cf59 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -641,7 +641,9 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         //}
         auto type = convert(lea->type()).id;
         auto offset = emit(lea->index());
-        return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {});
+        if (lea->ptr_pointee()->isa<ArrayType>())
+            return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {  });
+        return bb->access_chain(type, emit(lea->ptr()), { offset });
     } else if (auto aggop = def->isa<AggOp>()) {
         auto agg_type = convert(aggop->agg()->type()).id;
 

From 2c214ec1d448796f5a56e6cd4adf0db59e0faa7a Mon Sep 17 00:00:00 2001
From: Matthias Kurtenacker <matthias.kurtenacker@dfki.de>
Date: Mon, 24 Jun 2024 19:40:06 +0200
Subject: [PATCH 286/342] Do not warn users if RV is not found. We have info
 messages for that.

---
 CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c31f7152a..8bfbeecd2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -48,7 +48,7 @@ if(LLVM_FOUND)
         endif()
     endif()
     # check for RV
-    find_package(RV)
+    find_package(RV QUIET CONFIG)
     if(RV_FOUND)
         message(STATUS "Building with RV from LLVM installation.")
     else()

From e99f9b8bf8e388d5d4e7bcd190782bd8720206d6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 18 Oct 2024 14:26:43 +0200
Subject: [PATCH 287/342] remove useless args field in spirv::BB

---
 src/thorin/be/spirv/spirv.cpp       | 4 +---
 src/thorin/be/spirv/spirv_private.h | 1 -
 2 files changed, 1 insertion(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index edae2cf59..25f2e7b3d 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -300,10 +300,9 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
                 continue;
             }
             auto val = emit(arg);
-            bb->args[arg] = val;
             auto* param = dst_cont->param(index);
             auto& phi = cont2bb_[dst_cont]->phis_map[param];
-            phi.preds.emplace_back(bb->args[arg], emit_as_bb(continuation));
+            phi.preds.emplace_back(val, emit_as_bb(continuation));
         }
         bb->branch(emit(dst_cont));
     } else if (app.callee() == world().branch()) {
@@ -311,7 +310,6 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         emit_unsafe(mem);
 
         auto cond = emit(app.arg(1));
-        bb->args.emplace(app.arg(2), cond);
         auto tbb = emit(app.arg(2));
         auto fbb = emit(app.arg(3));
         bb->branch_conditional(cond, tbb, fbb);
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index ef1d72979..3d6701673 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -15,7 +15,6 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     FnBuilder& fn_builder;
     FileBuilder& file_builder;
     std::unordered_map<const Param*, Phi> phis_map;
-    DefMap<SpvId> args;
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {

From 64454fc4cb19d30a2ced4e5ac5983ea256d28a95 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 18 Oct 2024 14:42:32 +0200
Subject: [PATCH 288/342] spirv: share logic between call & intrinsic paths

---
 src/thorin/be/spirv/spirv.cpp | 32 ++++++++++----------------------
 1 file changed, 10 insertions(+), 22 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 25f2e7b3d..098ed9204 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -262,6 +262,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         assert(succ->is_basicblock());
         bb->branch(emit(succ));
         for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
+            assert(j < args.size());
             auto param = succ->param(i);
             if (is_mem(param) || is_unit(param))
                 continue;
@@ -370,45 +371,32 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         // must be call + continuation --- call + return has been removed by codegen_prepare
         auto succ = ret_arg->isa_nom<Continuation>();
 
-        size_t n = 0;
+        size_t real_params_count = 0;
         const Param* last_param = nullptr;
         for (auto param : succ->params()) {
             if (is_mem(param) || is_unit(param))
                 continue;
             last_param = param;
-            n++;
+            real_params_count++;
         }
 
-        if (n == 0) {
-            bb->branch(emit(succ));
-        } else if (n == 1) {
-            bb->branch(emit(succ));
+        std::vector<SpvId> args(real_params_count);
 
-            auto& phi = cont2bb_[succ]->phis_map[last_param];
-            phi.preds.emplace_back(call_result, emit_as_bb(continuation));
-        } else {
-            Array<SpvId> extracts(n);
+        if (real_params_count == 1) {
+            args[0] = call_result;
+        } else if (real_params_count > 1) {
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
                 auto param = succ->param(i);
                 if (is_mem(param) || is_unit(param))
                     continue;
-                extracts[j] = bb->extract(convert(param->type()).id, call_result, { (uint32_t) j });
+                args[j] = bb->extract(convert(param->type()).id, call_result, { (uint32_t) j });
                 j++;
             }
 
             bb->branch(emit(succ));
-
-            for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-                auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
-                    continue;
-
-                auto& phi = cont2bb_[succ]->phis_map[last_param];
-                phi.preds.emplace_back(extracts[j], emit_as_bb(continuation));
-
-                j++;
-            }
         }
+
+        jump_to_next_cont_with_args(succ, args);
     }
 }
 

From 2eba44d2c197f3e4d2de7bb28d41c370fc2bbf88 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 18 Oct 2024 15:40:54 +0200
Subject: [PATCH 289/342] fix: Branch and Match don't have ret_param s

---
 src/thorin/continuation.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index d99dd0ae9..61d53d2cf 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -139,6 +139,12 @@ const Param* Continuation::mem_param() const {
 }
 
 const Param* Continuation::ret_param() const {
+    switch (intrinsic()) {
+        case Intrinsic::Branch:
+        case Intrinsic::Match:
+            return nullptr;
+        default: break;
+    }
     const Param* result = nullptr;
     for (auto param : params()) {
         if (param->order() >= 1) {

From 8191629094927f1fac18face91be82fe0d11378d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 18 Oct 2024 15:41:31 +0200
Subject: [PATCH 290/342] added enum for ground truth of App operands

---
 src/thorin/continuation.h | 13 +++++++++----
 src/thorin/primop.cpp     |  2 +-
 2 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index fb747c5b7..11c87ee25 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -60,10 +60,15 @@ class App : public Def {
     App(World&, const Defs ops, Debug dbg);
 
 public:
-    const Def* callee() const { return op(0); }
-    const Def* arg(size_t i) const { return op(1 + i); }
-    size_t num_args() const { return num_ops() - 1; }
-    const Defs args() const { return ops().skip_front(); }
+    enum Ops {
+        Callee = 0,
+        FirstArg = 1,
+    };
+
+    const Def* callee() const { return op(Ops::Callee); }
+    const Def* arg(size_t i) const { return op(Ops::FirstArg + i); }
+    size_t num_args() const { return num_ops() - Ops::FirstArg; }
+    const Defs args() const { return ops().skip_front(Ops::FirstArg); }
     const Def* rebuild(World&, const Type*, Defs) const override;
 
     Continuations using_continuations() const {
diff --git a/src/thorin/primop.cpp b/src/thorin/primop.cpp
index 786a834d0..8b6fa09d7 100644
--- a/src/thorin/primop.cpp
+++ b/src/thorin/primop.cpp
@@ -188,7 +188,7 @@ bool Slot::equal(const Def* other) const { return this == other; }
  * rebuild
  */
 
-const Def* App           ::rebuild(World& w, const Type*  , Defs o) const { return w.app(o[0], o.skip_front(), debug()); }
+const Def* App           ::rebuild(World& w, const Type*  , Defs o) const { return w.app(o[App::Ops::Callee], o.skip_front(App::Ops::FirstArg), debug()); }
 const Def* ArithOp       ::rebuild(World& w, const Type*  , Defs o) const { return w.arithop(arithop_tag(), o[0], o[1], debug()); }
 const Def* Bitcast       ::rebuild(World& w, const Type* t, Defs o) const { return w.bitcast(t, o[0], debug()); }
 const Def* Bottom        ::rebuild(World& w, const Type* t, Defs  ) const { return w.bottom(t, debug()); }

From 12e48c53ae3bac2b4e7717f15e8bd17fc1bbe32e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 18 Oct 2024 15:42:20 +0200
Subject: [PATCH 291/342] spirv: added a hack to avoid phi nodes with return
 blocks

---
 src/thorin/be/spirv/spirv.cpp       | 48 ++++++++++++++++++++++++++---
 src/thorin/be/spirv/spirv_private.h |  2 ++
 2 files changed, 45 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 098ed9204..55e4e8234 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -180,6 +180,28 @@ FnBuilder* CodeGen::prepare(const thorin::Scope& scope) {
     return &fn;
 }
 
+static bool is_return_block(thorin::Continuation* cont) {
+    if (!cont->is_basicblock())
+        return false;
+    int uses_as_ret_param = 0;
+    for (auto use : cont->copy_uses()) {
+        if (use.def()->isa<Param>())
+            continue; // the block can have params
+        else if (auto app = use.def()->isa<App>()) {
+            if (auto callee = app->callee()->isa_nom<Continuation>()) {
+                auto arg_index = use.index() - App::FirstArg;
+                auto ret_param = callee->ret_param();
+                if (ret_param && arg_index == ret_param->index()) {
+                    uses_as_ret_param++;
+                    continue;
+                }
+            }
+        }
+        return false; // any other use disqualifies the block
+    }
+    return uses_as_ret_param == 1;
+}
+
 void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
     auto& bb = *fn->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*fn));
     cont2bb_[cont] = &bb;
@@ -187,6 +209,10 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
 
     builder_->name(bb.label, cont->name().c_str());
 
+    bb.semi_inline = is_return_block(cont);
+    if (bb.semi_inline)
+        world().ddef(cont, "Emitting {} as return block", cont);
+
     if (entry_ == cont) {
         for (auto param : cont->params()) {
             if (is_mem(param) || is_unit(param)) {
@@ -204,6 +230,8 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
         }
     } else {
         defs_[cont] = bb.label;
+        if (bb.semi_inline)
+            return;
         for (auto param : cont->params()) {
             if (is_mem(param) || is_unit(param)) {
                 // Nothing
@@ -256,20 +284,30 @@ SpvId CodeGen::emit_as_bb(thorin::Continuation* cont) {
 }
 
 void CodeGen::emit_epilogue(Continuation* continuation) {
-    auto& bb = cont2bb_[continuation];
+    BasicBlockBuilder* bb = cont2bb_[continuation];
+
     // Handles the potential nuances of jumping to another continuation
     auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
         assert(succ->is_basicblock());
-        bb->branch(emit(succ));
+        BasicBlockBuilder* dstbb = cont2bb_[succ];
+
         for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
             assert(j < args.size());
             auto param = succ->param(i);
-            if (is_mem(param) || is_unit(param))
+            if (is_mem(param) || is_unit(param)) {
+                if (dstbb->semi_inline)
+                    defs_[param] = 0;
                 continue;
-            auto& phi = cont2bb_[succ]->phis_map[param];
-            phi.preds.emplace_back(args[j], emit_as_bb(continuation));
+            }
+            if (dstbb->semi_inline) {
+                defs_[param] = args[j];
+            } else {
+                auto& phi = cont2bb_[succ]->phis_map[param];
+                phi.preds.emplace_back(args[j], emit_as_bb(continuation));
+            }
             j++;
         }
+        bb->branch(emit(succ));
     };
 
     auto& app = *continuation->body();
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index 3d6701673..9f1546682 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -15,6 +15,8 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     FnBuilder& fn_builder;
     FileBuilder& file_builder;
     std::unordered_map<const Param*, Phi> phis_map;
+
+    bool semi_inline;
 };
 
 struct FnBuilder : public builder::SpvFnBuilder {

From 63da5f84a21ab9531566a4778acd6431060a69c6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 17:22:15 +0100
Subject: [PATCH 292/342] fix levelzero intrinsic not being registered in
 codegen.cpp

---
 src/thorin/be/codegen.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 4d9ec7c83..6b50fa273 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -270,6 +270,7 @@ DeviceBackends::DeviceBackends(thorin::World& world, int opt, bool debug, std::s
 #endif
 #if THORIN_ENABLE_SPIRV
     register_backend(std::make_unique<OpenCLSPIRVBackend>(*this, world));
+    register_backend(std::make_unique<LevelZeroSPIRVBackend>(*this, world));
 #endif
     register_backend(std::make_unique<HLSBackend>(*this, world, hls_flags));
 

From 5074b03f0aedc091e5fbb0d958cbc5c8070094f9 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 17:22:34 +0100
Subject: [PATCH 293/342] spirv: fix dangling reference to target_info_

---
 src/thorin/be/spirv/spirv.cpp | 1 +
 src/thorin/be/spirv/spirv.h   | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 55e4e8234..78d299e0b 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -115,6 +115,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             builder_->addressing_model = spv::AddressingModelPhysicalStorageBuffer64;
             builder_->memory_model = spv::MemoryModel::MemoryModelGLSL450;
             break;
+        default: assert(false && "unknown spirv dialect");
     }
 
     ScopesForest forest(world());
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 22886eab8..44dfa9685 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -71,7 +71,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
 
     SpvId get_codom_type(const Continuation* fn);
 
-    Target& target_info_;
+    Target target_info_;
     FileBuilder* builder_;
     const Cont2Config* kernel_config_;
 

From 52c208b99fe6a9d3d5ed0fe4f9dae6cf673ffefe Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 17:22:58 +0100
Subject: [PATCH 294/342] spirv: fix typo in error msg

---
 src/thorin/be/spirv/spirv.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 78d299e0b..14c0e42c4 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -873,7 +873,7 @@ std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* i
         } else
             world().ELOG("spirv.builtin requires an integer literal as the argument");
     } else {
-        world().ELOG("This spir-v builtin isn't recognised: %s", intrinsic->name());
+        world().ELOG("This spir-v builtin isn't recognised: {}", intrinsic->name());
     }
     return productions;
 }

From 702cf3c34872af12e5453466cb1f65ea9bc8082a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 17:23:06 +0100
Subject: [PATCH 295/342] spirv: use GLCompute or Kernel appropriately

---
 src/thorin/be/spirv/spirv.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 14c0e42c4..3fa8a0111 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -144,7 +144,7 @@ void CodeGen::emit_stream(std::ostream& out) {
                     (uint32_t) std::get<2>(block),
             };
 
-            builder_->declare_entry_point(spv::ExecutionModelGLCompute, callee, cont->name().c_str(), interface);
+            builder_->declare_entry_point(target_info_.dialect == Target::Vulkan ? spv::ExecutionModelGLCompute : spv::ExecutionModelKernel, callee, cont->name().c_str(), interface);
             builder_->execution_mode(callee, spv::ExecutionModeLocalSize, local_size);
             entry_points_count++;
         }

From 5e5fc8ab9ff7e992ec824ee8b10f920d41b76f14 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 17:48:53 +0100
Subject: [PATCH 296/342] spv: fix typo in Cast

---
 src/thorin/be/spirv/spirv.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 3fa8a0111..144e8a450 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -766,7 +766,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             auto src_kind = classify_primtype(src_prim);
             auto dst_kind = classify_primtype(dst_prim);
             size_t src_bitwidth = conv_src_type.layout->size * 8;
-            size_t dst_bitwidth = conv_src_type.layout->size * 8;
+            size_t dst_bitwidth = conv_dst_type.layout->size * 8;
 
             SpvId data = emit(cast->from());
 
@@ -801,6 +801,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
                         case PrimTypeKind::Float:
                             data = bb->convert(spv::OpFConvert, target_type, data);
                             break;
+                        default: assert(false);
                     }
                 }
 
@@ -820,6 +821,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
                         case PrimTypeKind::Float:
                             data = bb->convert(spv::OpFConvert, target_type, data);
                             break;
+                        default: assert(false);
                     }
                 }
             }

From 8e8c309a4139cc62c55fee603d0e9d2edb251565 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 17:58:48 +0100
Subject: [PATCH 297/342] fix interface of generated spv kernel to include
 builtins

---
 src/thorin/be/spirv/spirv.cpp       | 6 +++---
 src/thorin/be/spirv/spirv_private.h | 1 +
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 144e8a450..ad7b79d6c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -121,10 +121,9 @@ void CodeGen::emit_stream(std::ostream& out) {
     ScopesForest forest(world());
     forest.for_each([&](const Scope& scope) { emit_scope(scope, forest); });
 
-    std::vector<SpvId> interface;
     for (auto def : world().defs()) {
         if (auto global = def->isa<Global>())
-            interface.push_back(emit(global));
+            builder.interface.push_back(emit(global));
     }
 
     int entry_points_count = 0;
@@ -144,7 +143,7 @@ void CodeGen::emit_stream(std::ostream& out) {
                     (uint32_t) std::get<2>(block),
             };
 
-            builder_->declare_entry_point(target_info_.dialect == Target::Vulkan ? spv::ExecutionModelGLCompute : spv::ExecutionModelKernel, callee, cont->name().c_str(), interface);
+            builder_->declare_entry_point(target_info_.dialect == Target::Vulkan ? spv::ExecutionModelGLCompute : spv::ExecutionModelKernel, callee, cont->name().c_str(), builder.interface);
             builder_->execution_mode(callee, spv::ExecutionModeLocalSize, local_size);
             entry_points_count++;
         }
@@ -868,6 +867,7 @@ std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* i
                 auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
                 auto desired_type = ret_type->types()[1]->as<PtrType>();
                 auto id = builder_->variable(convert(desired_type).id, static_cast<spv::StorageClass>(convert(desired_type->addr_space())));
+                builder_->interface.push_back(id);
                 builder_->decorate(id, spv::Decoration::DecorationBuiltIn, { spv_builtin });
                 builder_->builtins_[spv_builtin] = id;
                 productions.push_back(id);
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index 9f1546682..d36d4e81a 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -38,6 +38,7 @@ struct FileBuilder : public builder::SpvFileBuilder {
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
     std::unordered_map<uint32_t, SpvId> builtins_;
+    std::vector<SpvId> interface;
 
     SpvId u32_t();
     SpvId u32_constant(uint32_t);

From cbd75e59caed9ede0893b069f11a471a01e01576 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 30 Oct 2024 18:04:01 +0100
Subject: [PATCH 298/342] spv: emit spv 1.2 for OpenCL

---
 src/thorin/be/spirv/spirv.cpp         | 1 +
 src/thorin/be/spirv/spirv_builder.hpp | 4 +++-
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index ad7b79d6c..d59a489e2 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -109,6 +109,7 @@ void CodeGen::emit_stream(std::ostream& out) {
             builder_->capability(spv::Capability::CapabilityAddresses);
             builder_->addressing_model = target_info_.mem_layout.pointer_size == 4 ? spv::AddressingModelPhysical32 : spv::AddressingModelPhysical64;
             builder_->memory_model = spv::MemoryModel::MemoryModelOpenCL;
+            builder_->version = 0x10200;
             break;
         case Target::Vulkan:
             builder_->capability(spv::Capability::CapabilityShader);
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index b94f11eb1..876dab6e0 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -609,6 +609,8 @@ struct SpvFileBuilder {
         extensions_set.insert(name);
     }
 
+    uint32_t version = spv::Version;
+
     spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
     spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
 
@@ -670,7 +672,7 @@ struct SpvFileBuilder {
         memory_model_section.data_.push_back(memory_model);
 
         output_word_le(spv::MagicNumber);
-        output_word_le(spv::Version); // TODO: target a specific spirv version
+        output_word_le(version); // TODO: target a specific spirv version
         output_word_le(uint32_t(0)); // TODO get a magic number ?
         output_word_le(bound);
         output_word_le(uint32_t(0)); // instruction schema padding

From 336e2628feaf3abb5f77641401474e39589b80d6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 4 Nov 2024 15:33:20 +0100
Subject: [PATCH 299/342] codegen: add seperate notion of 'Offload' intrinsics

---
 src/thorin/be/codegen.cpp   | 2 +-
 src/thorin/continuation.cpp | 3 +++
 src/thorin/continuation.h   | 5 ++++-
 3 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/codegen.cpp b/src/thorin/be/codegen.cpp
index 6b50fa273..554a4c023 100644
--- a/src/thorin/be/codegen.cpp
+++ b/src/thorin/be/codegen.cpp
@@ -297,7 +297,7 @@ void DeviceBackends::search_for_device_code() {
 
         Intrinsic intrinsic = Intrinsic::None;
         visit_capturing_intrinsics(continuation, [&] (Continuation* continuation) {
-            if (continuation->is_accelerator()) {
+            if (continuation->is_offload_intrinsic()) {
                 intrinsic = continuation->intrinsic();
                 return true;
             }
diff --git a/src/thorin/continuation.cpp b/src/thorin/continuation.cpp
index 61d53d2cf..1632708b1 100644
--- a/src/thorin/continuation.cpp
+++ b/src/thorin/continuation.cpp
@@ -250,6 +250,9 @@ const Filter* Continuation::all_true_filter() const {
 }
 
 bool Continuation::is_accelerator() const { return Intrinsic::AcceleratorBegin <= intrinsic() && intrinsic() < Intrinsic::AcceleratorEnd; }
+
+bool Continuation::is_offload_intrinsic() const { return Intrinsic::OffloadBegin <= intrinsic() && intrinsic() < Intrinsic::OffloadEnd; }
+
 void Continuation::set_intrinsic() {
     if      (name() == "cuda")           attributes().intrinsic = Intrinsic::CUDA;
     else if (name() == "nvvm")           attributes().intrinsic = Intrinsic::NVVM;
diff --git a/src/thorin/continuation.h b/src/thorin/continuation.h
index 11c87ee25..bdedec64f 100644
--- a/src/thorin/continuation.h
+++ b/src/thorin/continuation.h
@@ -97,7 +97,8 @@ enum class CC : uint8_t {
 enum class Intrinsic : uint8_t {
     None,
     AcceleratorBegin,
-    CUDA = AcceleratorBegin,    ///< Internal CUDA-Backend.
+    OffloadBegin = AcceleratorBegin,
+    CUDA = OffloadBegin,    ///< Internal CUDA-Backend.
     NVVM,                       ///< Internal NNVM-Backend.
     OpenCL,                     ///< Internal OpenCL-Backend.
     OpenCL_SPIRV,               ///< Internal OpenCL-Backend.
@@ -106,6 +107,7 @@ enum class Intrinsic : uint8_t {
     AMDGPUPAL,                  ///< Internal AMDGPU-PAL-Backend.
     ShadyCompute,               ///< Internal Shady Compute Backend.
     HLS,                        ///< Internal HLS-Backend.
+    OffloadEnd = HLS,
     Parallel,                   ///< Internal Parallel-CPU-Backend.
     Fibers,                     ///< Internal Parallel-CPU-Backend using resumable fibers.
     Spawn,                      ///< Internal Parallel-CPU-Backend.
@@ -186,6 +188,7 @@ class Continuation : public Def {
     bool is_channel() const { return name().find("channel") != std::string::npos; }
     bool is_pipe() const { return name().find("pipe") != std::string::npos; }
     bool is_accelerator() const;
+    bool is_offload_intrinsic() const;
 
     const App* body() const { return op(0)->as<App>(); }
     bool has_body() const { return !op(0)->isa<Bottom>(); }

From 30df11e0a1455467820ad6f44d878b1c6a997a90 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Mon, 4 Nov 2024 16:47:32 +0100
Subject: [PATCH 300/342] fix fp64 constant emission

---
 src/thorin/be/spirv/spirv.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index d59a489e2..c0be0e581 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -439,6 +439,8 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
     }
 }
 
+static_assert(sizeof(double) == sizeof(uint64_t), "This code assumes 64-bit double");
+
 SpvId CodeGen::emit_constant(const thorin::Def* def) {
     if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
@@ -454,6 +456,7 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
             case PrimType_pf32: case PrimType_qf32:
             case PrimType_ps32: case PrimType_qs32:
             case PrimType_pu32: case PrimType_qu32: constant = builder_->constant(type, { static_cast<unsigned int>(box.get_u32()) }); break;
+            case PrimType_pf64: case PrimType_qf64:
             case PrimType_ps64: case PrimType_qs64:
             case PrimType_pu64: case PrimType_qu64: {
                 uint64_t value = static_cast<uint64_t>(box.get_u64());
@@ -462,7 +465,6 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
                 constant = builder_->constant(type, { (uint32_t) lower, (uint32_t) upper });
                 break;
             }
-            case PrimType_pf64: case PrimType_qf64: assertf(false, "not implemented yet");
         }
         return constant;
     }

From 2b1b4fc092259bce54fb676a83f7db78f9164a2a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:27:09 +0100
Subject: [PATCH 301/342] spirv_builder.hpp: move FileBuilder first

---
 src/thorin/be/spirv/spirv_builder.hpp | 935 +++++++++++++-------------
 1 file changed, 470 insertions(+), 465 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 876dab6e0..64efb134d 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -69,631 +69,636 @@ struct SpvSectionBuilder {
     }
 };
 
-struct SpvBasicBlockBuilder : public SpvSectionBuilder {
-    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
-            : file_builder(file_builder)
-    {}
 
-    SpvFileBuilder& file_builder;
 
-    struct Phi {
-        SpvId type;
-        SpvId value;
-        std::vector<std::pair<SpvId, SpvId>> preds;
+struct SpvFileBuilder {
+
+    enum UniqueDeclTag {
+        NONE,
+        FN_TYPE,
+        PTR_TYPE,
+        DEF_ARR_TYPE,
+        CONSTANT,
+        CONSTANT_COMPOSITE,
     };
-    std::vector<Phi*> phis;
-    SpvId label;
 
-    SpvId undef(SpvId type) {
-        op(spv::Op::OpUndef, 3);
-        ref_id(type);
+    /// Prevents duplicate declarations
+    struct UniqueDeclKey {
+        UniqueDeclTag tag;
+        std::vector<uint32_t> members;
+
+        bool operator==(const UniqueDeclKey &b) const {
+            return tag == b.tag && members == b.members;
+        }
+    };
+
+    struct UniqueDeclKeyHasher {
+        size_t operator() (const UniqueDeclKey& key) const {
+            size_t acc = 0;
+            for (auto id : key.members)
+                acc ^= std::hash<uint32_t>{}(id);
+            return std::hash<size_t>{}(key.tag) ^ acc;
+        }
+    };
+
+    SpvFileBuilder() {}
+    SpvFileBuilder(const SpvFileBuilder&) = delete;
+
+    SpvId generate_fresh_id() { return { bound++ }; }
+
+    void name(SpvId id, std::string_view str) {
+        assert(id < bound);
+        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
+        debug_names.ref_id(id);
+        debug_names.literal_name(str);
+    }
+
+    SpvId declare_bool_type() {
+        types_constants.op(spv::Op::OpTypeBool, 2);
         auto id = generate_fresh_id();
-        ref_id(id);
+        types_constants.ref_id(id);
         return id;
     }
 
-    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
-        op(spv::Op::OpCompositeConstruct, 3 + elements.size());
-        ref_id(aggregate_t);
+    SpvId declare_int_type(int width, bool signed_) {
+        types_constants.op(spv::Op::OpTypeInt, 4);
         auto id = generate_fresh_id();
-        ref_id(id);
-        for (auto e : elements)
-            ref_id(e);
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
+        types_constants.literal_int(signed_ ? 1 : 0);
         return id;
     }
 
-    SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
-        op(spv::Op::OpCompositeExtract, 4 + indices.size());
-        ref_id(target_type);
+    SpvId declare_float_type(int width) {
+        types_constants.op(spv::Op::OpTypeFloat, 3);
         auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(composite);
-        for (auto i : indices)
-            literal_int(i);
+        types_constants.ref_id(id);
+        types_constants.literal_int(width);
         return id;
     }
 
-    SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
-        op(spv::Op::OpCompositeInsert, 5 + indices.size());
-        ref_id(target_type);
+    SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
+        auto key = UniqueDeclKey { PTR_TYPE, { element_type, (uint32_t) storage_class } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpTypePointer, 4);
         auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(object);
-        ref_id(composite);
-        for (auto i : indices)
-            literal_int(i);
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
+        types_constants.ref_id(element_type);
+        unique_decls[key] = id;
         return id;
     }
 
-    SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
-        op(spv::Op::OpVectorExtractDynamic, 5);
-        ref_id(target_type);
+    SpvId declare_array_type(SpvId element_type, SpvId dim) {
+        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type, dim } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpTypeArray, 4);
         auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(vector);
-        ref_id(index);
+        types_constants.ref_id(id);
+        types_constants.ref_id(element_type);
+        types_constants.ref_id(dim);
+        unique_decls[key] = id;
         return id;
     }
 
-    SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
-        op(spv::Op::OpVectorInsertDynamic, 6);
-        ref_id(target_type);
+    SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
+        auto key = UniqueDeclKey { FN_TYPE, {} };
+        for (auto d : dom) key.members.push_back(d);
+        key.members.push_back(codom);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+
+        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
         auto id = generate_fresh_id();
-        ref_id(id);
-        ref_id(vector);
-        ref_id(component);
-        ref_id(index);
+        types_constants.ref_id(id);
+        types_constants.ref_id(codom);
+        for (auto arg : dom)
+            types_constants.ref_id(arg);
+        unique_decls[key] = id;
         return id;
     }
 
-    // Used for almost all conversion operations
-    SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
-        op(op_, 4);
+    SpvId declare_struct_type(std::vector<SpvId> elements) {
+        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
         auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(value);
+        types_constants.ref_id(id);
+        for (auto arg : elements)
+            types_constants.ref_id(arg);
         return id;
     }
 
-    SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
-        op(spv::Op::OpAccessChain, 4 + indexes.size());
+    SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
+        types_constants.op(spv::Op::OpTypeVector, 4);
         auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(element);
-        for (auto index : indexes)
-            ref_id(index);
+        types_constants.ref_id(id);
+        types_constants.ref_id(component_type);
+        types_constants.literal_int(dim);
         return id;
     }
 
-    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
-        op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
+    void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+        annotations.op(spv::Op::OpDecorate, 3 + extra.size());
+        annotations.ref_id(target);
+        annotations.literal_int(decoration);
+        for (auto e : extra)
+            annotations.literal_int(e);
+    }
+
+    void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+        annotations.op(spv::Op::OpMemberDecorate, 4 + extra.size());
+        annotations.ref_id(target);
+        annotations.literal_int(member);
+        annotations.literal_int(decoration);
+        for (auto e : extra)
+            annotations.literal_int(e);
+    }
+
+    SpvId debug_string(std::string string) {
+        debug_string_source.op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
         auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(base);
-        ref_id(element);
-        for (auto index : indexes)
-            ref_id(index);
+        debug_string_source.ref_id(id);
+        debug_string_source.literal_name(string);
         return id;
     }
 
-    SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
-        op(spv::Op::OpLoad, 4 + operands.size());
+    SpvId bool_constant(SpvId type, bool value) {
+        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
         auto id = generate_fresh_id();
-        ref_id(target_type);
-        ref_id(id);
-        ref_id(pointer);
-        for (auto op : operands)
-            literal_int(op);
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
         return id;
     }
 
-    void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
-        op(spv::Op::OpStore, 3 + operands.size());
-        ref_id(pointer);
-        ref_id(value);
-        for (auto op : operands)
-            literal_int(op);
+    SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
+        auto key = UniqueDeclKey { CONSTANT, bit_pattern };
+        key.members.push_back(type);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
+        auto id = generate_fresh_id();
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto arg : bit_pattern)
+            types_constants.literal_int(arg);
+        unique_decls[key] = id;
+        return id;
     }
 
-    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
-        op(op_, 5);
+    SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
+        auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
+        key.members.push_back(type);
+        for (auto op : ops) key.members.push_back(op);
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
+        types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
         auto id = generate_fresh_id();
-        ref_id(result_type);
-        ref_id(id);
-        ref_id(lhs);
-        ref_id(rhs);
+        types_constants.ref_id(type);
+        types_constants.ref_id(id);
+        for (auto op : ops)
+            types_constants.ref_id(op);
+        unique_decls[key] = id;
         return id;
     }
 
-    void branch(SpvId target) {
-        op(spv::Op::OpBranch, 2);
-        ref_id(target);
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        types_constants.op(spv::Op::OpVariable, 4);
+        types_constants.ref_id(type);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        types_constants.literal_int(storage_class);
+        return id;
     }
 
-    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
-        op(spv::Op::OpBranchConditional, 4);
-        ref_id(condition);
-        ref_id(true_target);
-        ref_id(false_target);
+    SpvId declare_void_type() {
+        types_constants.op(spv::Op::OpTypeVoid, 2);
+        auto id = generate_fresh_id();
+        types_constants.ref_id(id);
+        return id;
     }
 
-    void branch_switch(SpvId selector, SpvId default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
-        assert(literals.size() == cases.size());
-        op(spv::Op::OpSwitch, 3 + literals.size() * 2);
-        ref_id(selector);
-        ref_id(default_case);
-        for (size_t i = 0; i < literals.size(); i++) {
-            ref_id(literals[i]);
-            ref_id(cases[i]);
-        }
-    }
-
-    void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
-        op(spv::Op::OpSelectionMerge, 3);
-        ref_id(merge_bb);
-        literal_int(selection_control);
-    }
+    SpvId define_function(SpvFnBuilder& fn_builder);
 
-    void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
-        op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
-        ref_id(merge_bb);
-        ref_id(continue_bb);
-        literal_int(loop_control);
-
-        for (auto e : loop_control_ops)
-            literal_int(e);
+    void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
+        entry_points.op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
+        entry_points.literal_int(execution_model);
+        entry_points.ref_id(entry_point);
+        entry_points.literal_name(name);
+        for (auto i : interface)
+            entry_points.ref_id(i);
     }
 
-    SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
-        op(spv::Op::OpFunctionCall, 4 + arguments.size());
-        auto id = generate_fresh_id();
-        ref_id(return_type);
-        ref_id(id);
-        ref_id(callee);
-
-        for (auto a : arguments)
-            ref_id(a);
-        return id;
+    void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
+        entry_points.op(spv::Op::OpExecutionMode, 3 + payloads.size());
+        entry_points.ref_id(entry_point);
+        entry_points.literal_int(execution_mode);
+        for (auto d : payloads)
+            entry_points.literal_int(d);
     }
 
-    SpvId ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments);
-
-    void return_void() {
-        op(spv::Op::OpReturn, 1);
+    void capability(spv::Capability cap) {
+        auto found = capabilities_set.find(cap);
+        if (found != capabilities_set.end())
+            return;
+        capabilities.op(spv::Op::OpCapability, 2);
+        capabilities.data_.push_back(cap);
+        capabilities_set.insert(cap);
     }
 
-    void return_value(SpvId value) {
-        op(spv::Op::OpReturnValue, 2);
-        ref_id(value);
+    void extension(std::string name) {
+        auto found = extensions_set.find(name);
+        if (found != extensions_set.end())
+            return;
+        extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
+        extensions.literal_name(name);
+        extensions_set.insert(name);
     }
 
-    void unreachable() {
-        op(spv::Op::OpUnreachable, 1);
-    }
+    uint32_t version = spv::Version;
 
-private:
-    SpvId generate_fresh_id();
+    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
+    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
 
 protected:
-    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
-        op(spv::Op::OpExtInst, 5 + arguments.size());
+    SpvId extended_import(std::string name) {
+        auto found = extended_instruction_sets.find(name);
+        if (found != extended_instruction_sets.end())
+            return found->second;
+        ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
         auto id = generate_fresh_id();
-        ref_id(return_type);
-        ref_id(id);
-        ref_id(set);
-        literal_int(instruction);
-        for (auto a : arguments)
-            ref_id(a);
+        ext_inst_import.ref_id(id);
+        ext_inst_import.literal_name(name);
+        extended_instruction_sets[name] = id;
         return id;
     }
-};
-
-struct SpvFnBuilder {
-    explicit SpvFnBuilder(SpvFileBuilder* file_builder)
-            : file_builder(file_builder)
-    {
-        function_id = generate_fresh_id();
-    }
-
-    SpvFileBuilder* file_builder;
-    SpvId function_id;
 
-    SpvId fn_type;
-    SpvId fn_ret_type;
-    std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
+private:
+    std::ostream* output_ = nullptr;
+    uint32_t bound = 1;
 
-    // Contains OpFunctionParams
-    SpvSectionBuilder header;
+    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
+    SpvSectionBuilder capabilities;
+    SpvSectionBuilder extensions;
+    SpvSectionBuilder ext_inst_import;
+    SpvSectionBuilder entry_points;
+    SpvSectionBuilder execution_modes;
+    SpvSectionBuilder debug_string_source;
+    SpvSectionBuilder debug_names;
+    SpvSectionBuilder debug_module_processed;
+    SpvSectionBuilder annotations;
+    SpvSectionBuilder types_constants;
+    SpvSectionBuilder fn_decls;
+    SpvSectionBuilder fn_defs;
 
-    SpvSectionBuilder variables;
+    // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
+    std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
+    std::unordered_map<std::string, SpvId> extended_instruction_sets;
+    std::unordered_set<spv::Capability> capabilities_set;
+    std::unordered_set<std::string> extensions_set;
 
-    SpvId parameter(SpvId param_type) {
-        header.op(spv::Op::OpFunctionParameter, 3);
-        auto id = generate_fresh_id();
-        header.ref_id(param_type);
-        header.ref_id(id);
-        return id;
+    void output_word_le(uint32_t word) {
+        output_->put((word >> 0) & 0xFFu);
+        output_->put((word >> 8) & 0xFFu);
+        output_->put((word >> 16) & 0xFFu);
+        output_->put((word >> 24) & 0xFFu);
     }
 
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        variables.op(spv::Op::OpVariable, 4);
-        variables.ref_id(type);
-        auto id = generate_fresh_id();
-        variables.ref_id(id);
-        variables.literal_int(storage_class);
-        return id;
+    void output_section(SpvSectionBuilder& section) {
+        for (auto& word : section.data_) {
+            output_word_le(word);
+        }
     }
+public:
+    void finish(std::ostream& output) {
+        output_ = &output;
+        SpvSectionBuilder memory_model_section;
+        memory_model_section.op(spv::Op::OpMemoryModel, 3);
+        memory_model_section.data_.push_back(addressing_model);
+        memory_model_section.data_.push_back(memory_model);
 
-private:
-    SpvId generate_fresh_id();
-};
+        output_word_le(spv::MagicNumber);
+        output_word_le(version); // TODO: target a specific spirv version
+        output_word_le(uint32_t(0)); // TODO get a magic number ?
+        output_word_le(bound);
+        output_word_le(uint32_t(0)); // instruction schema padding
 
-struct SpvFileBuilder {
+        output_section(capabilities);
+        output_section(extensions);
+        output_section(ext_inst_import);
+        output_section(memory_model_section);
+        output_section(entry_points);
+        output_section(execution_modes);
+        output_section(debug_string_source);
+        output_section(debug_names);
+        output_section(debug_module_processed);
+        output_section(annotations);
+        output_section(types_constants);
+        output_section(fn_decls);
+        output_section(fn_defs);
+    }
 
-    enum UniqueDeclTag {
-        NONE,
-        FN_TYPE,
-        PTR_TYPE,
-        DEF_ARR_TYPE,
-        CONSTANT,
-        CONSTANT_COMPOSITE,
-    };
+    friend SpvBasicBlockBuilder;
+};
 
-    /// Prevents duplicate declarations
-    struct UniqueDeclKey {
-        UniqueDeclTag tag;
-        std::vector<uint32_t> members;
+struct SpvBasicBlockBuilder : public SpvSectionBuilder {
+    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
+            : file_builder(file_builder)
+    {}
 
-        bool operator==(const UniqueDeclKey &b) const {
-            return tag == b.tag && members == b.members;
-        }
-    };
+    SpvFileBuilder& file_builder;
 
-    struct UniqueDeclKeyHasher {
-        size_t operator() (const UniqueDeclKey& key) const {
-            size_t acc = 0;
-            for (auto id : key.members)
-                acc ^= std::hash<uint32_t>{}(id);
-            return std::hash<size_t>{}(key.tag) ^ acc;
-        }
+    struct Phi {
+        SpvId type;
+        SpvId value;
+        std::vector<std::pair<SpvId, SpvId>> preds;
     };
+    std::vector<Phi*> phis;
+    SpvId label;
 
-    SpvFileBuilder() {}
-    SpvFileBuilder(const SpvFileBuilder&) = delete;
-
-    SpvId generate_fresh_id() { return { bound++ }; }
-
-    void name(SpvId id, std::string_view str) {
-        assert(id < bound);
-        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
-        debug_names.ref_id(id);
-        debug_names.literal_name(str);
-    }
-
-    SpvId declare_bool_type() {
-        types_constants.op(spv::Op::OpTypeBool, 2);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        return id;
-    }
-
-    SpvId declare_int_type(int width, bool signed_) {
-        types_constants.op(spv::Op::OpTypeInt, 4);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(width);
-        types_constants.literal_int(signed_ ? 1 : 0);
-        return id;
-    }
-
-    SpvId declare_float_type(int width) {
-        types_constants.op(spv::Op::OpTypeFloat, 3);
-        auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(width);
-        return id;
-    }
-
-    SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
-        auto key = UniqueDeclKey { PTR_TYPE, { element_type, (uint32_t) storage_class } };
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpTypePointer, 4);
+    SpvId undef(SpvId type) {
+        op(spv::Op::OpUndef, 3);
+        ref_id(type);
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(storage_class);
-        types_constants.ref_id(element_type);
-        unique_decls[key] = id;
+        ref_id(id);
         return id;
     }
 
-    SpvId declare_array_type(SpvId element_type, SpvId dim) {
-        auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type, dim } };
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpTypeArray, 4);
+    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
+        op(spv::Op::OpCompositeConstruct, 3 + elements.size());
+        ref_id(aggregate_t);
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(element_type);
-        types_constants.ref_id(dim);
-        unique_decls[key] = id;
+        ref_id(id);
+        for (auto e : elements)
+            ref_id(e);
         return id;
     }
 
-    SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
-        auto key = UniqueDeclKey { FN_TYPE, {} };
-        for (auto d : dom) key.members.push_back(d);
-        key.members.push_back(codom);
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-
-        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
+    SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
+        op(spv::Op::OpCompositeExtract, 4 + indices.size());
+        ref_id(target_type);
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(codom);
-        for (auto arg : dom)
-            types_constants.ref_id(arg);
-        unique_decls[key] = id;
+        ref_id(id);
+        ref_id(composite);
+        for (auto i : indices)
+            literal_int(i);
         return id;
     }
 
-    SpvId declare_struct_type(std::vector<SpvId> elements) {
-        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
+    SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
+        op(spv::Op::OpCompositeInsert, 5 + indices.size());
+        ref_id(target_type);
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        for (auto arg : elements)
-            types_constants.ref_id(arg);
+        ref_id(id);
+        ref_id(object);
+        ref_id(composite);
+        for (auto i : indices)
+            literal_int(i);
         return id;
     }
 
-    SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
-        types_constants.op(spv::Op::OpTypeVector, 4);
+    SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
+        op(spv::Op::OpVectorExtractDynamic, 5);
+        ref_id(target_type);
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.ref_id(component_type);
-        types_constants.literal_int(dim);
+        ref_id(id);
+        ref_id(vector);
+        ref_id(index);
         return id;
     }
 
-    void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
-        annotations.op(spv::Op::OpDecorate, 3 + extra.size());
-        annotations.ref_id(target);
-        annotations.literal_int(decoration);
-        for (auto e : extra)
-            annotations.literal_int(e);
-    }
-
-    void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
-        annotations.op(spv::Op::OpMemberDecorate, 4 + extra.size());
-        annotations.ref_id(target);
-        annotations.literal_int(member);
-        annotations.literal_int(decoration);
-        for (auto e : extra)
-            annotations.literal_int(e);
-    }
-
-    SpvId debug_string(std::string string) {
-        debug_string_source.op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
+    SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
+        op(spv::Op::OpVectorInsertDynamic, 6);
+        ref_id(target_type);
         auto id = generate_fresh_id();
-        debug_string_source.ref_id(id);
-        debug_string_source.literal_name(string);
+        ref_id(id);
+        ref_id(vector);
+        ref_id(component);
+        ref_id(index);
         return id;
     }
 
-    SpvId bool_constant(SpvId type, bool value) {
-        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
+    // Used for almost all conversion operations
+    SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
+        op(op_, 4);
         auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(value);
         return id;
     }
 
-    SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
-        auto key = UniqueDeclKey { CONSTANT, bit_pattern };
-        key.members.push_back(type);
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
+    SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
+        op(spv::Op::OpAccessChain, 4 + indexes.size());
         auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        for (auto arg : bit_pattern)
-            types_constants.literal_int(arg);
-        unique_decls[key] = id;
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(element);
+        for (auto index : indexes)
+            ref_id(index);
         return id;
     }
 
-    SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
-        auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
-        key.members.push_back(type);
-        for (auto op : ops) key.members.push_back(op);
-        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
+    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
+        op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
         auto id = generate_fresh_id();
-        types_constants.ref_id(type);
-        types_constants.ref_id(id);
-        for (auto op : ops)
-            types_constants.ref_id(op);
-        unique_decls[key] = id;
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(base);
+        ref_id(element);
+        for (auto index : indexes)
+            ref_id(index);
         return id;
     }
 
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        types_constants.op(spv::Op::OpVariable, 4);
-        types_constants.ref_id(type);
+    SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
+        op(spv::Op::OpLoad, 4 + operands.size());
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
-        types_constants.literal_int(storage_class);
+        ref_id(target_type);
+        ref_id(id);
+        ref_id(pointer);
+        for (auto op : operands)
+            literal_int(op);
         return id;
     }
 
-    SpvId declare_void_type() {
-        types_constants.op(spv::Op::OpTypeVoid, 2);
+    void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
+        op(spv::Op::OpStore, 3 + operands.size());
+        ref_id(pointer);
+        ref_id(value);
+        for (auto op : operands)
+            literal_int(op);
+    }
+
+    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
+        op(op_, 5);
         auto id = generate_fresh_id();
-        types_constants.ref_id(id);
+        ref_id(result_type);
+        ref_id(id);
+        ref_id(lhs);
+        ref_id(rhs);
         return id;
     }
 
-    SpvId define_function(SpvFnBuilder& fn_builder) {
-        fn_defs.op(spv::Op::OpFunction, 5);
-        fn_defs.ref_id(fn_builder.fn_ret_type);
-        fn_defs.ref_id(fn_builder.function_id);
-        fn_defs.data_.push_back(spv::FunctionControlMaskNone);
-        fn_defs.ref_id(fn_builder.fn_type);
+    void branch(SpvId target) {
+        op(spv::Op::OpBranch, 2);
+        ref_id(target);
+    }
 
-        // Includes stuff like OpFunctionParameters
-        for (auto w : fn_builder.header.data_)
-            fn_defs.data_.push_back(w);
+    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
+        op(spv::Op::OpBranchConditional, 4);
+        ref_id(condition);
+        ref_id(true_target);
+        ref_id(false_target);
+    }
+
+    void branch_switch(SpvId selector, SpvId default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
+        assert(literals.size() == cases.size());
+        op(spv::Op::OpSwitch, 3 + literals.size() * 2);
+        ref_id(selector);
+        ref_id(default_case);
+        for (size_t i = 0; i < literals.size(); i++) {
+            ref_id(literals[i]);
+            ref_id(cases[i]);
+        }
+    }
 
-        bool first = true;
-        for (auto& bb : fn_builder.bbs_to_emit) {
-            fn_defs.op(spv::Op::OpLabel, 2);
-            fn_defs.ref_id(bb->label);
+    void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
+        op(spv::Op::OpSelectionMerge, 3);
+        ref_id(merge_bb);
+        literal_int(selection_control);
+    }
 
-            if (first) {
-                for (auto w : fn_builder.variables.data_)
-                    fn_defs.data_.push_back(w);
-                first = false;
-            }
+    void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
+        op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
+        ref_id(merge_bb);
+        ref_id(continue_bb);
+        literal_int(loop_control);
 
-            for (auto& phi : bb->phis) {
-                fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
-                fn_defs.ref_id(phi->type);
-                fn_defs.ref_id(phi->value);
-                assert(!phi->preds.empty());
-                for (auto& [pred_value, pred_label] : phi->preds) {
-                    fn_defs.ref_id(pred_value);
-                    fn_defs.ref_id(pred_label);
-                }
-            }
+        for (auto e : loop_control_ops)
+            literal_int(e);
+    }
 
-            for (auto w : bb->data_)
-                fn_defs.data_.push_back(w);
-        }
+    SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
+        op(spv::Op::OpFunctionCall, 4 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(callee);
 
-        fn_defs.op(spv::Op::OpFunctionEnd, 1);
-        return fn_builder.function_id;
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
     }
 
-    void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
-        entry_points.op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
-        entry_points.literal_int(execution_model);
-        entry_points.ref_id(entry_point);
-        entry_points.literal_name(name);
-        for (auto i : interface)
-            entry_points.ref_id(i);
+    SpvId ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments);
+
+    void return_void() {
+        op(spv::Op::OpReturn, 1);
     }
 
-    void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
-        entry_points.op(spv::Op::OpExecutionMode, 3 + payloads.size());
-        entry_points.ref_id(entry_point);
-        entry_points.literal_int(execution_mode);
-        for (auto d : payloads)
-            entry_points.literal_int(d);
+    void return_value(SpvId value) {
+        op(spv::Op::OpReturnValue, 2);
+        ref_id(value);
     }
 
-    void capability(spv::Capability cap) {
-        auto found = capabilities_set.find(cap);
-        if (found != capabilities_set.end())
-            return;
-        capabilities.op(spv::Op::OpCapability, 2);
-        capabilities.data_.push_back(cap);
-        capabilities_set.insert(cap);
+    void unreachable() {
+        op(spv::Op::OpUnreachable, 1);
     }
 
-    void extension(std::string name) {
-        auto found = extensions_set.find(name);
-        if (found != extensions_set.end())
-            return;
-        extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
-        extensions.literal_name(name);
-        extensions_set.insert(name);
+private:
+    SpvId generate_fresh_id();
+
+protected:
+    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
+        op(spv::Op::OpExtInst, 5 + arguments.size());
+        auto id = generate_fresh_id();
+        ref_id(return_type);
+        ref_id(id);
+        ref_id(set);
+        literal_int(instruction);
+        for (auto a : arguments)
+            ref_id(a);
+        return id;
     }
+};
 
-    uint32_t version = spv::Version;
+struct SpvFnBuilder {
+    explicit SpvFnBuilder(SpvFileBuilder* file_builder)
+            : file_builder(file_builder)
+    {
+        function_id = generate_fresh_id();
+    }
 
-    spv::AddressingModel addressing_model = spv::AddressingModel::AddressingModelLogical;
-    spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
+    SpvFileBuilder* file_builder;
+    SpvId function_id;
 
-protected:
-    SpvId extended_import(std::string name) {
-        auto found = extended_instruction_sets.find(name);
-        if (found != extended_instruction_sets.end())
-            return found->second;
-        ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
+    SpvId fn_type;
+    SpvId fn_ret_type;
+    std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
+
+    // Contains OpFunctionParams
+    SpvSectionBuilder header;
+
+    SpvSectionBuilder variables;
+
+    SpvId parameter(SpvId param_type) {
+        header.op(spv::Op::OpFunctionParameter, 3);
         auto id = generate_fresh_id();
-        ext_inst_import.ref_id(id);
-        ext_inst_import.literal_name(name);
-        extended_instruction_sets[name] = id;
+        header.ref_id(param_type);
+        header.ref_id(id);
+        return id;
+    }
+
+    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+        variables.op(spv::Op::OpVariable, 4);
+        variables.ref_id(type);
+        auto id = generate_fresh_id();
+        variables.ref_id(id);
+        variables.literal_int(storage_class);
         return id;
     }
 
 private:
-    std::ostream* output_ = nullptr;
-    uint32_t bound = 1;
+    SpvId generate_fresh_id();
+};
 
-    // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
-    SpvSectionBuilder capabilities;
-    SpvSectionBuilder extensions;
-    SpvSectionBuilder ext_inst_import;
-    SpvSectionBuilder entry_points;
-    SpvSectionBuilder execution_modes;
-    SpvSectionBuilder debug_string_source;
-    SpvSectionBuilder debug_names;
-    SpvSectionBuilder debug_module_processed;
-    SpvSectionBuilder annotations;
-    SpvSectionBuilder types_constants;
-    SpvSectionBuilder fn_decls;
-    SpvSectionBuilder fn_defs;
+inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
+    fn_defs.op(spv::Op::OpFunction, 5);
+    fn_defs.ref_id(fn_builder.fn_ret_type);
+    fn_defs.ref_id(fn_builder.function_id);
+    fn_defs.data_.push_back(spv::FunctionControlMaskNone);
+    fn_defs.ref_id(fn_builder.fn_type);
 
-    // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
-    std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
-    std::unordered_map<std::string, SpvId> extended_instruction_sets;
-    std::unordered_set<spv::Capability> capabilities_set;
-    std::unordered_set<std::string> extensions_set;
+    // Includes stuff like OpFunctionParameters
+    for (auto w : fn_builder.header.data_)
+        fn_defs.data_.push_back(w);
 
-    void output_word_le(uint32_t word) {
-        output_->put((word >> 0) & 0xFFu);
-        output_->put((word >> 8) & 0xFFu);
-        output_->put((word >> 16) & 0xFFu);
-        output_->put((word >> 24) & 0xFFu);
-    }
+    bool first = true;
+    for (auto& bb : fn_builder.bbs_to_emit) {
+        fn_defs.op(spv::Op::OpLabel, 2);
+        fn_defs.ref_id(bb->label);
 
-    void output_section(SpvSectionBuilder& section) {
-        for (auto& word : section.data_) {
-            output_word_le(word);
+        if (first) {
+            for (auto w : fn_builder.variables.data_)
+                fn_defs.data_.push_back(w);
+            first = false;
         }
-    }
-public:
-    void finish(std::ostream& output) {
-        output_ = &output;
-        SpvSectionBuilder memory_model_section;
-        memory_model_section.op(spv::Op::OpMemoryModel, 3);
-        memory_model_section.data_.push_back(addressing_model);
-        memory_model_section.data_.push_back(memory_model);
 
-        output_word_le(spv::MagicNumber);
-        output_word_le(version); // TODO: target a specific spirv version
-        output_word_le(uint32_t(0)); // TODO get a magic number ?
-        output_word_le(bound);
-        output_word_le(uint32_t(0)); // instruction schema padding
+        for (auto& phi : bb->phis) {
+            fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
+            fn_defs.ref_id(phi->type);
+            fn_defs.ref_id(phi->value);
+            assert(!phi->preds.empty());
+            for (auto& [pred_value, pred_label] : phi->preds) {
+                fn_defs.ref_id(pred_value);
+                fn_defs.ref_id(pred_label);
+            }
+        }
 
-        output_section(capabilities);
-        output_section(extensions);
-        output_section(ext_inst_import);
-        output_section(memory_model_section);
-        output_section(entry_points);
-        output_section(execution_modes);
-        output_section(debug_string_source);
-        output_section(debug_names);
-        output_section(debug_module_processed);
-        output_section(annotations);
-        output_section(types_constants);
-        output_section(fn_decls);
-        output_section(fn_defs);
+        for (auto w : bb->data_)
+            fn_defs.data_.push_back(w);
     }
 
-    friend SpvBasicBlockBuilder;
-};
+    fn_defs.op(spv::Op::OpFunctionEnd, 1);
+    return fn_builder.function_id;
+}
+
 
 inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
     return file_builder.generate_fresh_id();

From bc35e0fc7318adfc16c5184dbb2e74864fac3e36 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:28:59 +0100
Subject: [PATCH 302/342] spirv_builder.hpp: remove now-useless
 generate_fresh_id wrappers

---
 src/thorin/be/spirv/spirv_builder.hpp | 47 +++++++++------------------
 1 file changed, 16 insertions(+), 31 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 64efb134d..2f7cdf88c 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -414,7 +414,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId undef(SpvId type) {
         op(spv::Op::OpUndef, 3);
         ref_id(type);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(id);
         return id;
     }
@@ -422,7 +422,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
         op(spv::Op::OpCompositeConstruct, 3 + elements.size());
         ref_id(aggregate_t);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(id);
         for (auto e : elements)
             ref_id(e);
@@ -432,7 +432,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
         op(spv::Op::OpCompositeExtract, 4 + indices.size());
         ref_id(target_type);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(id);
         ref_id(composite);
         for (auto i : indices)
@@ -443,7 +443,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
         op(spv::Op::OpCompositeInsert, 5 + indices.size());
         ref_id(target_type);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(id);
         ref_id(object);
         ref_id(composite);
@@ -455,7 +455,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
         op(spv::Op::OpVectorExtractDynamic, 5);
         ref_id(target_type);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(id);
         ref_id(vector);
         ref_id(index);
@@ -465,7 +465,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
         op(spv::Op::OpVectorInsertDynamic, 6);
         ref_id(target_type);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(id);
         ref_id(vector);
         ref_id(component);
@@ -476,7 +476,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     // Used for almost all conversion operations
     SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
         op(op_, 4);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(value);
@@ -485,7 +485,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
         op(spv::Op::OpAccessChain, 4 + indexes.size());
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(element);
@@ -496,7 +496,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
         op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(base);
@@ -508,7 +508,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
         op(spv::Op::OpLoad, 4 + operands.size());
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(pointer);
@@ -527,7 +527,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
         op(op_, 5);
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(result_type);
         ref_id(id);
         ref_id(lhs);
@@ -576,7 +576,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
         op(spv::Op::OpFunctionCall, 4 + arguments.size());
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(return_type);
         ref_id(id);
         ref_id(callee);
@@ -601,13 +601,10 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         op(spv::Op::OpUnreachable, 1);
     }
 
-private:
-    SpvId generate_fresh_id();
-
 protected:
     SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
         op(spv::Op::OpExtInst, 5 + arguments.size());
-        auto id = generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         ref_id(return_type);
         ref_id(id);
         ref_id(set);
@@ -622,7 +619,7 @@ struct SpvFnBuilder {
     explicit SpvFnBuilder(SpvFileBuilder* file_builder)
             : file_builder(file_builder)
     {
-        function_id = generate_fresh_id();
+        function_id = file_builder->generate_fresh_id();
     }
 
     SpvFileBuilder* file_builder;
@@ -639,7 +636,7 @@ struct SpvFnBuilder {
 
     SpvId parameter(SpvId param_type) {
         header.op(spv::Op::OpFunctionParameter, 3);
-        auto id = generate_fresh_id();
+        auto id = file_builder->generate_fresh_id();
         header.ref_id(param_type);
         header.ref_id(id);
         return id;
@@ -648,14 +645,11 @@ struct SpvFnBuilder {
     SpvId variable(SpvId type, spv::StorageClass storage_class) {
         variables.op(spv::Op::OpVariable, 4);
         variables.ref_id(type);
-        auto id = generate_fresh_id();
+        auto id = file_builder->generate_fresh_id();
         variables.ref_id(id);
         variables.literal_int(storage_class);
         return id;
     }
-
-private:
-    SpvId generate_fresh_id();
 };
 
 inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
@@ -699,15 +693,6 @@ inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
     return fn_builder.function_id;
 }
 
-
-inline SpvId SpvBasicBlockBuilder::generate_fresh_id() {
-    return file_builder.generate_fresh_id();
-}
-
-inline SpvId SpvFnBuilder::generate_fresh_id() {
-    return file_builder->generate_fresh_id();
-}
-
 inline SpvId SpvBasicBlockBuilder::ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments) {
     return ext_instruction(return_type, file_builder.extended_import(instr.set_name), instr.id, arguments);
 }

From b87c8631fb9b087dcedc25e116eb7c4a50237e7e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:30:05 +0100
Subject: [PATCH 303/342] spirv_builder.hpp: made file_builder field in
 FnBuilder a reference

---
 src/thorin/be/spirv/spirv.cpp         |  2 +-
 src/thorin/be/spirv/spirv_builder.hpp | 10 +++++-----
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c0be0e581..c879a4f37 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -80,7 +80,7 @@ BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
     label = file_builder.generate_fresh_id();
 }
 
-FnBuilder::FnBuilder(FileBuilder& file_builder) : builder::SpvFnBuilder(&file_builder), file_builder(file_builder) {}
+FnBuilder::FnBuilder(FileBuilder& file_builder) : builder::SpvFnBuilder(file_builder), file_builder(file_builder) {}
 
 FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
 }
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 2f7cdf88c..3c57da195 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -616,13 +616,13 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 };
 
 struct SpvFnBuilder {
-    explicit SpvFnBuilder(SpvFileBuilder* file_builder)
+    explicit SpvFnBuilder(SpvFileBuilder& file_builder)
             : file_builder(file_builder)
     {
-        function_id = file_builder->generate_fresh_id();
+        function_id = file_builder.generate_fresh_id();
     }
 
-    SpvFileBuilder* file_builder;
+    SpvFileBuilder& file_builder;
     SpvId function_id;
 
     SpvId fn_type;
@@ -636,7 +636,7 @@ struct SpvFnBuilder {
 
     SpvId parameter(SpvId param_type) {
         header.op(spv::Op::OpFunctionParameter, 3);
-        auto id = file_builder->generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         header.ref_id(param_type);
         header.ref_id(id);
         return id;
@@ -645,7 +645,7 @@ struct SpvFnBuilder {
     SpvId variable(SpvId type, spv::StorageClass storage_class) {
         variables.op(spv::Op::OpVariable, 4);
         variables.ref_id(type);
-        auto id = file_builder->generate_fresh_id();
+        auto id = file_builder.generate_fresh_id();
         variables.ref_id(id);
         variables.literal_int(storage_class);
         return id;

From e20add369de11cf8cde07f4b2664903a3efc79dc Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:32:18 +0100
Subject: [PATCH 304/342] spirv_builder.hpp: rename op() into begin_op()

---
 src/thorin/be/spirv/spirv_builder.hpp | 107 +++++++++++++-------------
 1 file changed, 53 insertions(+), 54 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 3c57da195..f1fc3e82e 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -38,9 +38,9 @@ struct SpvSectionBuilder {
         data_.push_back(word);
     }
 public:
-    void op(spv::Op op, int ops_size) {
+    void begin_op(spv::Op op, int size_in_words) {
         uint32_t lower = op & 0xFFFFu;
-        uint32_t upper = (ops_size << 16) & 0xFFFF0000u;
+        uint32_t upper = (size_in_words << 16) & 0xFFFF0000u;
         output_word(lower | upper);
     }
 
@@ -72,7 +72,6 @@ struct SpvSectionBuilder {
 
 
 struct SpvFileBuilder {
-
     enum UniqueDeclTag {
         NONE,
         FN_TYPE,
@@ -108,20 +107,20 @@ struct SpvFileBuilder {
 
     void name(SpvId id, std::string_view str) {
         assert(id < bound);
-        debug_names.op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
+        debug_names.begin_op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
         debug_names.ref_id(id);
         debug_names.literal_name(str);
     }
 
     SpvId declare_bool_type() {
-        types_constants.op(spv::Op::OpTypeBool, 2);
+        types_constants.begin_op(spv::Op::OpTypeBool, 2);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         return id;
     }
 
     SpvId declare_int_type(int width, bool signed_) {
-        types_constants.op(spv::Op::OpTypeInt, 4);
+        types_constants.begin_op(spv::Op::OpTypeInt, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.literal_int(width);
@@ -130,7 +129,7 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_float_type(int width) {
-        types_constants.op(spv::Op::OpTypeFloat, 3);
+        types_constants.begin_op(spv::Op::OpTypeFloat, 3);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.literal_int(width);
@@ -140,7 +139,7 @@ struct SpvFileBuilder {
     SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
         auto key = UniqueDeclKey { PTR_TYPE, { element_type, (uint32_t) storage_class } };
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpTypePointer, 4);
+        types_constants.begin_op(spv::Op::OpTypePointer, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.literal_int(storage_class);
@@ -152,7 +151,7 @@ struct SpvFileBuilder {
     SpvId declare_array_type(SpvId element_type, SpvId dim) {
         auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type, dim } };
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpTypeArray, 4);
+        types_constants.begin_op(spv::Op::OpTypeArray, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.ref_id(element_type);
@@ -167,7 +166,7 @@ struct SpvFileBuilder {
         key.members.push_back(codom);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
 
-        types_constants.op(spv::Op::OpTypeFunction, 3 + dom.size());
+        types_constants.begin_op(spv::Op::OpTypeFunction, 3 + dom.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.ref_id(codom);
@@ -178,7 +177,7 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_struct_type(std::vector<SpvId> elements) {
-        types_constants.op(spv::Op::OpTypeStruct, 2 + elements.size());
+        types_constants.begin_op(spv::Op::OpTypeStruct, 2 + elements.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         for (auto arg : elements)
@@ -187,7 +186,7 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
-        types_constants.op(spv::Op::OpTypeVector, 4);
+        types_constants.begin_op(spv::Op::OpTypeVector, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         types_constants.ref_id(component_type);
@@ -196,7 +195,7 @@ struct SpvFileBuilder {
     }
 
     void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
-        annotations.op(spv::Op::OpDecorate, 3 + extra.size());
+        annotations.begin_op(spv::Op::OpDecorate, 3 + extra.size());
         annotations.ref_id(target);
         annotations.literal_int(decoration);
         for (auto e : extra)
@@ -204,7 +203,7 @@ struct SpvFileBuilder {
     }
 
     void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
-        annotations.op(spv::Op::OpMemberDecorate, 4 + extra.size());
+        annotations.begin_op(spv::Op::OpMemberDecorate, 4 + extra.size());
         annotations.ref_id(target);
         annotations.literal_int(member);
         annotations.literal_int(decoration);
@@ -213,7 +212,7 @@ struct SpvFileBuilder {
     }
 
     SpvId debug_string(std::string string) {
-        debug_string_source.op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
+        debug_string_source.begin_op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
         auto id = generate_fresh_id();
         debug_string_source.ref_id(id);
         debug_string_source.literal_name(string);
@@ -221,7 +220,7 @@ struct SpvFileBuilder {
     }
 
     SpvId bool_constant(SpvId type, bool value) {
-        types_constants.op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
+        types_constants.begin_op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
         auto id = generate_fresh_id();
         types_constants.ref_id(type);
         types_constants.ref_id(id);
@@ -232,7 +231,7 @@ struct SpvFileBuilder {
         auto key = UniqueDeclKey { CONSTANT, bit_pattern };
         key.members.push_back(type);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpConstant, 3 + bit_pattern.size());
+        types_constants.begin_op(spv::Op::OpConstant, 3 + bit_pattern.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(type);
         types_constants.ref_id(id);
@@ -247,7 +246,7 @@ struct SpvFileBuilder {
         key.members.push_back(type);
         for (auto op : ops) key.members.push_back(op);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
-        types_constants.op(spv::Op::OpConstantComposite, 3 + ops.size());
+        types_constants.begin_op(spv::Op::OpConstantComposite, 3 + ops.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(type);
         types_constants.ref_id(id);
@@ -258,7 +257,7 @@ struct SpvFileBuilder {
     }
 
     SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        types_constants.op(spv::Op::OpVariable, 4);
+        types_constants.begin_op(spv::Op::OpVariable, 4);
         types_constants.ref_id(type);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -267,7 +266,7 @@ struct SpvFileBuilder {
     }
 
     SpvId declare_void_type() {
-        types_constants.op(spv::Op::OpTypeVoid, 2);
+        types_constants.begin_op(spv::Op::OpTypeVoid, 2);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         return id;
@@ -276,7 +275,7 @@ struct SpvFileBuilder {
     SpvId define_function(SpvFnBuilder& fn_builder);
 
     void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
-        entry_points.op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
+        entry_points.begin_op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
         entry_points.literal_int(execution_model);
         entry_points.ref_id(entry_point);
         entry_points.literal_name(name);
@@ -285,7 +284,7 @@ struct SpvFileBuilder {
     }
 
     void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
-        entry_points.op(spv::Op::OpExecutionMode, 3 + payloads.size());
+        entry_points.begin_op(spv::Op::OpExecutionMode, 3 + payloads.size());
         entry_points.ref_id(entry_point);
         entry_points.literal_int(execution_mode);
         for (auto d : payloads)
@@ -296,7 +295,7 @@ struct SpvFileBuilder {
         auto found = capabilities_set.find(cap);
         if (found != capabilities_set.end())
             return;
-        capabilities.op(spv::Op::OpCapability, 2);
+        capabilities.begin_op(spv::Op::OpCapability, 2);
         capabilities.data_.push_back(cap);
         capabilities_set.insert(cap);
     }
@@ -305,7 +304,7 @@ struct SpvFileBuilder {
         auto found = extensions_set.find(name);
         if (found != extensions_set.end())
             return;
-        extensions.op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
+        extensions.begin_op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
         extensions.literal_name(name);
         extensions_set.insert(name);
     }
@@ -320,7 +319,7 @@ struct SpvFileBuilder {
         auto found = extended_instruction_sets.find(name);
         if (found != extended_instruction_sets.end())
             return found->second;
-        ext_inst_import.op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
+        ext_inst_import.begin_op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
         auto id = generate_fresh_id();
         ext_inst_import.ref_id(id);
         ext_inst_import.literal_name(name);
@@ -368,7 +367,7 @@ struct SpvFileBuilder {
     void finish(std::ostream& output) {
         output_ = &output;
         SpvSectionBuilder memory_model_section;
-        memory_model_section.op(spv::Op::OpMemoryModel, 3);
+        memory_model_section.begin_op(spv::Op::OpMemoryModel, 3);
         memory_model_section.data_.push_back(addressing_model);
         memory_model_section.data_.push_back(memory_model);
 
@@ -412,7 +411,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId label;
 
     SpvId undef(SpvId type) {
-        op(spv::Op::OpUndef, 3);
+        begin_op(spv::Op::OpUndef, 3);
         ref_id(type);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
@@ -420,7 +419,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
-        op(spv::Op::OpCompositeConstruct, 3 + elements.size());
+        begin_op(spv::Op::OpCompositeConstruct, 3 + elements.size());
         ref_id(aggregate_t);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
@@ -430,7 +429,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
-        op(spv::Op::OpCompositeExtract, 4 + indices.size());
+        begin_op(spv::Op::OpCompositeExtract, 4 + indices.size());
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
@@ -441,7 +440,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
-        op(spv::Op::OpCompositeInsert, 5 + indices.size());
+        begin_op(spv::Op::OpCompositeInsert, 5 + indices.size());
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
@@ -453,7 +452,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
-        op(spv::Op::OpVectorExtractDynamic, 5);
+        begin_op(spv::Op::OpVectorExtractDynamic, 5);
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
@@ -463,7 +462,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
-        op(spv::Op::OpVectorInsertDynamic, 6);
+        begin_op(spv::Op::OpVectorInsertDynamic, 6);
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
@@ -475,7 +474,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     // Used for almost all conversion operations
     SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
-        op(op_, 4);
+        begin_op(op_, 4);
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
@@ -484,7 +483,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
-        op(spv::Op::OpAccessChain, 4 + indexes.size());
+        begin_op(spv::Op::OpAccessChain, 4 + indexes.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
@@ -495,7 +494,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
-        op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
+        begin_op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
@@ -507,7 +506,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
-        op(spv::Op::OpLoad, 4 + operands.size());
+        begin_op(spv::Op::OpLoad, 4 + operands.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
         ref_id(id);
@@ -518,7 +517,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
-        op(spv::Op::OpStore, 3 + operands.size());
+        begin_op(spv::Op::OpStore, 3 + operands.size());
         ref_id(pointer);
         ref_id(value);
         for (auto op : operands)
@@ -526,7 +525,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
-        op(op_, 5);
+        begin_op(op_, 5);
         auto id = file_builder.generate_fresh_id();
         ref_id(result_type);
         ref_id(id);
@@ -536,12 +535,12 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     void branch(SpvId target) {
-        op(spv::Op::OpBranch, 2);
+        begin_op(spv::Op::OpBranch, 2);
         ref_id(target);
     }
 
     void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
-        op(spv::Op::OpBranchConditional, 4);
+        begin_op(spv::Op::OpBranchConditional, 4);
         ref_id(condition);
         ref_id(true_target);
         ref_id(false_target);
@@ -549,7 +548,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
 
     void branch_switch(SpvId selector, SpvId default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
         assert(literals.size() == cases.size());
-        op(spv::Op::OpSwitch, 3 + literals.size() * 2);
+        begin_op(spv::Op::OpSwitch, 3 + literals.size() * 2);
         ref_id(selector);
         ref_id(default_case);
         for (size_t i = 0; i < literals.size(); i++) {
@@ -559,13 +558,13 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
-        op(spv::Op::OpSelectionMerge, 3);
+        begin_op(spv::Op::OpSelectionMerge, 3);
         ref_id(merge_bb);
         literal_int(selection_control);
     }
 
     void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
-        op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
+        begin_op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
         ref_id(merge_bb);
         ref_id(continue_bb);
         literal_int(loop_control);
@@ -575,7 +574,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
-        op(spv::Op::OpFunctionCall, 4 + arguments.size());
+        begin_op(spv::Op::OpFunctionCall, 4 + arguments.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(return_type);
         ref_id(id);
@@ -589,21 +588,21 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     SpvId ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments);
 
     void return_void() {
-        op(spv::Op::OpReturn, 1);
+        begin_op(spv::Op::OpReturn, 1);
     }
 
     void return_value(SpvId value) {
-        op(spv::Op::OpReturnValue, 2);
+        begin_op(spv::Op::OpReturnValue, 2);
         ref_id(value);
     }
 
     void unreachable() {
-        op(spv::Op::OpUnreachable, 1);
+        begin_op(spv::Op::OpUnreachable, 1);
     }
 
 protected:
     SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
-        op(spv::Op::OpExtInst, 5 + arguments.size());
+        begin_op(spv::Op::OpExtInst, 5 + arguments.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(return_type);
         ref_id(id);
@@ -635,7 +634,7 @@ struct SpvFnBuilder {
     SpvSectionBuilder variables;
 
     SpvId parameter(SpvId param_type) {
-        header.op(spv::Op::OpFunctionParameter, 3);
+        header.begin_op(spv::Op::OpFunctionParameter, 3);
         auto id = file_builder.generate_fresh_id();
         header.ref_id(param_type);
         header.ref_id(id);
@@ -643,7 +642,7 @@ struct SpvFnBuilder {
     }
 
     SpvId variable(SpvId type, spv::StorageClass storage_class) {
-        variables.op(spv::Op::OpVariable, 4);
+        variables.begin_op(spv::Op::OpVariable, 4);
         variables.ref_id(type);
         auto id = file_builder.generate_fresh_id();
         variables.ref_id(id);
@@ -653,7 +652,7 @@ struct SpvFnBuilder {
 };
 
 inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
-    fn_defs.op(spv::Op::OpFunction, 5);
+    fn_defs.begin_op(spv::Op::OpFunction, 5);
     fn_defs.ref_id(fn_builder.fn_ret_type);
     fn_defs.ref_id(fn_builder.function_id);
     fn_defs.data_.push_back(spv::FunctionControlMaskNone);
@@ -665,7 +664,7 @@ inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
 
     bool first = true;
     for (auto& bb : fn_builder.bbs_to_emit) {
-        fn_defs.op(spv::Op::OpLabel, 2);
+        fn_defs.begin_op(spv::Op::OpLabel, 2);
         fn_defs.ref_id(bb->label);
 
         if (first) {
@@ -675,7 +674,7 @@ inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
         }
 
         for (auto& phi : bb->phis) {
-            fn_defs.op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
+            fn_defs.begin_op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
             fn_defs.ref_id(phi->type);
             fn_defs.ref_id(phi->value);
             assert(!phi->preds.empty());
@@ -689,7 +688,7 @@ inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
             fn_defs.data_.push_back(w);
     }
 
-    fn_defs.op(spv::Op::OpFunctionEnd, 1);
+    fn_defs.begin_op(spv::Op::OpFunctionEnd, 1);
     return fn_builder.function_id;
 }
 

From 5a60b5030ef13c9a88841ee613e042a33cc42688 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:38:05 +0100
Subject: [PATCH 305/342] spirv: renamed a bunch of types to drop the 'Spv'
 prefix

---
 src/thorin/be/spirv/spirv.cpp              |  62 +++----
 src/thorin/be/spirv/spirv.h                |  22 +--
 src/thorin/be/spirv/spirv_builder.hpp      | 182 ++++++++++-----------
 src/thorin/be/spirv/spirv_instructions.cpp |   4 +-
 src/thorin/be/spirv/spirv_private.h        |  18 +-
 src/thorin/be/spirv/spirv_types.cpp        |  10 +-
 6 files changed, 148 insertions(+), 150 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c879a4f37..b8877a567 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -10,7 +10,7 @@ namespace thorin::spirv {
 
 /// Used as a dummy SSA value for emitting things like mem/unit
 /// Should never make it in the binary files !
-constexpr SpvId spv_none { 0 };
+constexpr Id spv_none { 0 };
 
 // SPIR-V has 3 "kinds" of primitives, and the user may declare arbitrary bitwidths, the following helps in translation:
 enum class PrimTypeKind {
@@ -76,22 +76,22 @@ switch (bitwidth) { \
 }
 
 BasicBlockBuilder::BasicBlockBuilder(FnBuilder& fn_builder)
-        : builder::SpvBasicBlockBuilder(fn_builder.file_builder), fn_builder(fn_builder), file_builder(fn_builder.file_builder) {
+        : builder::BasicBlockBuilder(fn_builder.file_builder), fn_builder(fn_builder), file_builder(fn_builder.file_builder) {
     label = file_builder.generate_fresh_id();
 }
 
-FnBuilder::FnBuilder(FileBuilder& file_builder) : builder::SpvFnBuilder(file_builder), file_builder(file_builder) {}
+FnBuilder::FnBuilder(FileBuilder& file_builder) : builder::FnBuilder(file_builder), file_builder(file_builder) {}
 
-FileBuilder::FileBuilder(CodeGen* cg) : builder::SpvFileBuilder(), cg(cg) {
+FileBuilder::FileBuilder(CodeGen* cg) : builder::FileBuilder(), cg(cg) {
 }
 
-SpvId FileBuilder::u32_t() {
+Id FileBuilder::u32_t() {
     if (u32_t_ == 0)
         u32_t_ = cg->convert(cg->world().type_pu32()).id;
     return u32_t_;
 }
 
-SpvId FileBuilder::u32_constant(uint32_t pattern) {
+Id FileBuilder::u32_constant(uint32_t pattern) {
     return constant(u32_t(), { pattern });
 }
 
@@ -135,7 +135,7 @@ void CodeGen::emit_stream(std::ostream& out) {
                 continue;
 
             assert(defs_.contains(cont));
-            SpvId callee = defs_[cont];
+            Id callee = defs_[cont];
 
             auto block = config->second->as<GPUKernelConfig>()->block_size();
             std::vector<uint32_t> local_size = {
@@ -158,7 +158,7 @@ void CodeGen::emit_stream(std::ostream& out) {
     builder_ = nullptr;
 }
 
-SpvId CodeGen::emit_fun_decl(thorin::Continuation* continuation) {
+Id CodeGen::emit_fun_decl(thorin::Continuation* continuation) {
     return get_fn_builder(continuation).function_id;
 }
 
@@ -263,9 +263,9 @@ void CodeGen::finalize(const thorin::Scope&) {
     builder_->define_function(*builder_->current_fn_);
 }
 
-SpvId CodeGen::get_codom_type(const Continuation* fn) {
+Id CodeGen::get_codom_type(const Continuation* fn) {
     auto ret_cont_type = fn->ret_param()->type()->as<FnType>();
-    std::vector<SpvId> types;
+    std::vector<Id> types;
     for (auto& op : ret_cont_type->types()) {
         if (op->isa<MemType>() || is_type_unit(op->type()))
             continue;
@@ -279,7 +279,7 @@ SpvId CodeGen::get_codom_type(const Continuation* fn) {
     return builder_->declare_struct_type(types);
 }
 
-SpvId CodeGen::emit_as_bb(thorin::Continuation* cont) {
+Id CodeGen::emit_as_bb(thorin::Continuation* cont) {
     emit(cont);
     return cont2bb_[cont]->label;
 }
@@ -288,7 +288,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
     BasicBlockBuilder* bb = cont2bb_[continuation];
 
     // Handles the potential nuances of jumping to another continuation
-    auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<SpvId> args) {
+    auto jump_to_next_cont_with_args = [&](Continuation* succ, std::vector<Id> args) {
         assert(succ->is_basicblock());
         BasicBlockBuilder* dstbb = cont2bb_[succ];
 
@@ -314,7 +314,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
     auto& app = *continuation->body();
 
     if (app.callee() == entry_->ret_param()) {
-        std::vector<SpvId> values;
+        std::vector<Id> values;
 
         for (auto arg : app.args()) {
             assert(arg->order() == 0);
@@ -357,8 +357,8 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         emit_unsafe(app.arg(0));
         auto val = emit(app.arg(1));
         auto otherwise_bb = emit_as_bb(app.arg(2)->isa_nom<Continuation>());
-        std::vector<SpvId> literals;
-        std::vector<SpvId> cases;
+        std::vector<Id> literals;
+        std::vector<Id> cases;
         for (size_t i = 3; i < app.num_args(); i++) {
             auto arg = app.arg(i)->as<Tuple>();
             literals.push_back(emit(arg->op(0)));
@@ -377,7 +377,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         jump_to_next_cont_with_args(succ, productions);
     } else { // function/closure call
         // put all first-order args into an array
-        std::vector<SpvId> call_args;
+        std::vector<Id> call_args;
         const Def* ret_arg = nullptr;
         for (auto arg : app.args()) {
             if (arg->order() == 0) {
@@ -394,7 +394,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             }
         }
 
-        SpvId call_result;
+        Id call_result;
         if (auto called_continuation = app.callee()->isa_nom<Continuation>()) {
             auto ret_type = get_codom_type(called_continuation);
             call_result = bb->call(ret_type, emit(called_continuation), call_args);
@@ -419,7 +419,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             real_params_count++;
         }
 
-        std::vector<SpvId> args(real_params_count);
+        std::vector<Id> args(real_params_count);
 
         if (real_params_count == 1) {
             args[0] = call_result;
@@ -441,11 +441,11 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
 
 static_assert(sizeof(double) == sizeof(uint64_t), "This code assumes 64-bit double");
 
-SpvId CodeGen::emit_constant(const thorin::Def* def) {
+Id CodeGen::emit_constant(const thorin::Def* def) {
     if (auto primlit = def->isa<PrimLit>()) {
         Box box = primlit->value();
         auto type = convert(def->type()).id;
-        SpvId constant;
+        Id constant;
         switch (primlit->primtype_tag()) {
             case PrimType_bool:                     constant = builder_->bool_constant(type, box.get_bool()); break;
             case PrimType_ps8:  case PrimType_qs8:
@@ -472,14 +472,14 @@ SpvId CodeGen::emit_constant(const thorin::Def* def) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
+Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     if (auto mathop = def->isa<MathOp>())
         return emit_mathop(bb, *mathop);
 
     if (auto bin = def->isa<BinOp>()) {
-        SpvId lhs = emit(bin->lhs());
-        SpvId rhs = emit(bin->rhs());
-        SpvId result_type = convert(def->type()).id;
+        Id lhs = emit(bin->lhs());
+        Id rhs = emit(bin->rhs());
+        Id result_type = convert(def->type()).id;
 
         if (auto cmp = bin->isa<Cmp>()) {
             auto type = cmp->lhs()->type();
@@ -618,7 +618,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto value = emit(vindex->op(0));
         return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
-        std::vector<SpvId> elements;
+        std::vector<Id> elements;
         elements.resize(tuple->num_ops());
         size_t x = 0;
         for (auto& e : tuple->ops()) {
@@ -626,7 +626,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         }
         return bb->composite(convert(tuple->type()).id, elements);
     } else if (auto structagg = def->isa<StructAgg>()) {
-        std::vector<SpvId> elements;
+        std::vector<Id> elements;
         elements.resize(structagg->num_ops());
         size_t x = 0;
         for (auto& e : structagg->ops()) {
@@ -677,7 +677,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         bool mem = false;
         if (auto tt = aggop->agg()->type()->isa<TupleType>(); tt && tt->op(0)->isa<MemType>()) mem = true;
 
-        auto copy_to_alloca = [&] (SpvId spv_agg, SpvId target_type) {
+        auto copy_to_alloca = [&] (Id spv_agg, Id target_type) {
             world().wdef(def, "slow: alloca and loads/stores needed for aggregate '{}'", def);
             auto agg_ptr_type = builder_->declare_ptr_type(spv::StorageClassFunction, agg_type);
 
@@ -770,7 +770,7 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             size_t src_bitwidth = conv_src_type.layout->size * 8;
             size_t dst_bitwidth = conv_dst_type.layout->size * 8;
 
-            SpvId data = emit(cast->from());
+            Id data = emit(cast->from());
 
             // If floating point is involved (src or dst), OpConvert*ToF and OpConvertFTo* can take care of the bit width transformation so no need for any chopping/expanding
             if (src_kind == PrimTypeKind::Float || dst_kind == PrimTypeKind::Float) {
@@ -840,10 +840,10 @@ SpvId CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-std::vector<SpvId> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
-    std::vector<SpvId> productions;
+std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
+    std::vector<Id> productions;
     if (intrinsic->name() == "spirv.nonsemantic.printf") {
-        std::vector<SpvId> args;
+        std::vector<Id> args;
         auto string = app.arg(1);
         if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
             auto arr = string->as<DefiniteArray>();
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 44dfa9685..fe0c70c78 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -7,7 +7,7 @@
 
 namespace thorin::spirv {
 
-using SpvId = uint32_t;
+using Id = uint32_t;
 
 class CodeGen;
 
@@ -29,7 +29,7 @@ struct Target {
 };
 
 struct ConvertedType {
-    SpvId id;
+    Id id;
     struct Layout {
         size_t size, alignment;
     };
@@ -38,21 +38,21 @@ struct ConvertedType {
 
 struct BasicBlockBuilder;
 
-class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedType, BasicBlockBuilder*, CodeGen> {
+class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType, BasicBlockBuilder*, CodeGen> {
 public:
     CodeGen(Thorin& thorin, Target&, bool debug, const Cont2Config* = nullptr);
 
     void emit_stream(std::ostream& stream) override;
     const char* file_ext() const override { return ".spv"; }
 
-    bool is_valid(SpvId id) {
+    bool is_valid(Id id) {
         return id > 0;
     }
 
     uint32_t convert(AddrSpace);
     ConvertedType convert(const Type*);
 
-    SpvId emit_fun_decl(Continuation*);
+    Id emit_fun_decl(Continuation*);
 
     FnBuilder* prepare(const Scope&);
     void prepare(Continuation*, FnBuilder*);
@@ -60,16 +60,16 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<SpvId, ConvertedT
     void finalize(const Scope&);
     void finalize(Continuation*);
 
-    SpvId emit_constant(const Def*);
-    SpvId emit_bb(BasicBlockBuilder* bb, const Def* def);
+    Id emit_constant(const Def*);
+    Id emit_bb(BasicBlockBuilder* bb, const Def* def);
 protected:
     FnBuilder& get_fn_builder(Continuation*);
-    std::vector<SpvId> emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb);
+    std::vector<Id> emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb);
 
-    SpvId emit_as_bb(Continuation*);
-    SpvId emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
+    Id emit_as_bb(Continuation*);
+    Id emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
 
-    SpvId get_codom_type(const Continuation* fn);
+    Id get_codom_type(const Continuation* fn);
 
     Target target_info_;
     FileBuilder* builder_;
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index f1fc3e82e..aca443356 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -11,12 +11,12 @@
 namespace thorin::spirv::builder {
 
 //struct SpvId { uint32_t id; };
-using SpvId = uint32_t;
+using Id = uint32_t;
 
-struct SpvSectionBuilder;
-struct SpvBasicBlockBuilder;
-struct SpvFnBuilder;
-struct SpvFileBuilder;
+struct SectionBuilder;
+struct BasicBlockBuilder;
+struct FnBuilder;
+struct FileBuilder;
 
 struct ExtendedInstruction {
     const char* set_name;
@@ -30,7 +30,7 @@ inline int div_roundup(int a, int b) {
         return (a / b) + 1;
 }
 
-struct SpvSectionBuilder {
+struct SectionBuilder {
     std::vector<uint32_t> data_;
 
 private:
@@ -44,7 +44,7 @@ struct SpvSectionBuilder {
         output_word(lower | upper);
     }
 
-    void ref_id(SpvId id) {
+    void ref_id(Id id) {
         assert(id != 0);
         output_word(id);
     }
@@ -69,9 +69,7 @@ struct SpvSectionBuilder {
     }
 };
 
-
-
-struct SpvFileBuilder {
+struct FileBuilder {
     enum UniqueDeclTag {
         NONE,
         FN_TYPE,
@@ -100,26 +98,26 @@ struct SpvFileBuilder {
         }
     };
 
-    SpvFileBuilder() {}
-    SpvFileBuilder(const SpvFileBuilder&) = delete;
+    FileBuilder() {}
+    FileBuilder(const FileBuilder&) = delete;
 
-    SpvId generate_fresh_id() { return { bound++ }; }
+    Id generate_fresh_id() { return { bound++ }; }
 
-    void name(SpvId id, std::string_view str) {
+    void name(Id id, std::string_view str) {
         assert(id < bound);
         debug_names.begin_op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
         debug_names.ref_id(id);
         debug_names.literal_name(str);
     }
 
-    SpvId declare_bool_type() {
+    Id declare_bool_type() {
         types_constants.begin_op(spv::Op::OpTypeBool, 2);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         return id;
     }
 
-    SpvId declare_int_type(int width, bool signed_) {
+    Id declare_int_type(int width, bool signed_) {
         types_constants.begin_op(spv::Op::OpTypeInt, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -128,7 +126,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_float_type(int width) {
+    Id declare_float_type(int width) {
         types_constants.begin_op(spv::Op::OpTypeFloat, 3);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -136,7 +134,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_ptr_type(spv::StorageClass storage_class, SpvId element_type) {
+    Id declare_ptr_type(spv::StorageClass storage_class, Id element_type) {
         auto key = UniqueDeclKey { PTR_TYPE, { element_type, (uint32_t) storage_class } };
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.begin_op(spv::Op::OpTypePointer, 4);
@@ -148,7 +146,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_array_type(SpvId element_type, SpvId dim) {
+    Id declare_array_type(Id element_type, Id dim) {
         auto key = UniqueDeclKey { DEF_ARR_TYPE, { element_type, dim } };
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.begin_op(spv::Op::OpTypeArray, 4);
@@ -160,7 +158,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_fn_type(std::vector<SpvId> dom, SpvId codom) {
+    Id declare_fn_type(std::vector<Id> dom, Id codom) {
         auto key = UniqueDeclKey { FN_TYPE, {} };
         for (auto d : dom) key.members.push_back(d);
         key.members.push_back(codom);
@@ -176,7 +174,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_struct_type(std::vector<SpvId> elements) {
+    Id declare_struct_type(std::vector<Id> elements) {
         types_constants.begin_op(spv::Op::OpTypeStruct, 2 + elements.size());
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -185,7 +183,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_vector_type(SpvId component_type, uint32_t dim) {
+    Id declare_vector_type(Id component_type, uint32_t dim) {
         types_constants.begin_op(spv::Op::OpTypeVector, 4);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
@@ -194,7 +192,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    void decorate(SpvId target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+    void decorate(Id target, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
         annotations.begin_op(spv::Op::OpDecorate, 3 + extra.size());
         annotations.ref_id(target);
         annotations.literal_int(decoration);
@@ -202,7 +200,7 @@ struct SpvFileBuilder {
             annotations.literal_int(e);
     }
 
-    void decorate_member(SpvId target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
+    void decorate_member(Id target, uint32_t member, spv::Decoration decoration, std::vector<uint32_t> extra = {}) {
         annotations.begin_op(spv::Op::OpMemberDecorate, 4 + extra.size());
         annotations.ref_id(target);
         annotations.literal_int(member);
@@ -211,7 +209,7 @@ struct SpvFileBuilder {
             annotations.literal_int(e);
     }
 
-    SpvId debug_string(std::string string) {
+    Id debug_string(std::string string) {
         debug_string_source.begin_op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
         auto id = generate_fresh_id();
         debug_string_source.ref_id(id);
@@ -219,7 +217,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId bool_constant(SpvId type, bool value) {
+    Id bool_constant(Id type, bool value) {
         types_constants.begin_op(value ? spv::Op::OpConstantTrue : spv::Op::OpConstantFalse, 3);
         auto id = generate_fresh_id();
         types_constants.ref_id(type);
@@ -227,7 +225,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId constant(SpvId type, std::vector<uint32_t> bit_pattern) {
+    Id constant(Id type, std::vector<uint32_t> bit_pattern) {
         auto key = UniqueDeclKey { CONSTANT, bit_pattern };
         key.members.push_back(type);
         if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
@@ -241,7 +239,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId constant_composite(SpvId type, std::vector<SpvId> ops) {
+    Id constant_composite(Id type, std::vector<Id> ops) {
         auto key = UniqueDeclKey { CONSTANT_COMPOSITE, {} };
         key.members.push_back(type);
         for (auto op : ops) key.members.push_back(op);
@@ -256,7 +254,7 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+    Id variable(Id type, spv::StorageClass storage_class) {
         types_constants.begin_op(spv::Op::OpVariable, 4);
         types_constants.ref_id(type);
         auto id = generate_fresh_id();
@@ -265,16 +263,16 @@ struct SpvFileBuilder {
         return id;
     }
 
-    SpvId declare_void_type() {
+    Id declare_void_type() {
         types_constants.begin_op(spv::Op::OpTypeVoid, 2);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
         return id;
     }
 
-    SpvId define_function(SpvFnBuilder& fn_builder);
+    Id define_function(FnBuilder& fn_builder);
 
-    void declare_entry_point(spv::ExecutionModel execution_model, SpvId entry_point, std::string name, std::vector<SpvId> interface) {
+    void declare_entry_point(spv::ExecutionModel execution_model, Id entry_point, std::string name, std::vector<Id> interface) {
         entry_points.begin_op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
         entry_points.literal_int(execution_model);
         entry_points.ref_id(entry_point);
@@ -283,7 +281,7 @@ struct SpvFileBuilder {
             entry_points.ref_id(i);
     }
 
-    void execution_mode(SpvId entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
+    void execution_mode(Id entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
         entry_points.begin_op(spv::Op::OpExecutionMode, 3 + payloads.size());
         entry_points.ref_id(entry_point);
         entry_points.literal_int(execution_mode);
@@ -315,7 +313,7 @@ struct SpvFileBuilder {
     spv::MemoryModel memory_model = spv::MemoryModel::MemoryModelSimple;
 
 protected:
-    SpvId extended_import(std::string name) {
+    Id extended_import(std::string name) {
         auto found = extended_instruction_sets.find(name);
         if (found != extended_instruction_sets.end())
             return found->second;
@@ -332,22 +330,22 @@ struct SpvFileBuilder {
     uint32_t bound = 1;
 
     // Ordered as per https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.pdf#subsection.2.4
-    SpvSectionBuilder capabilities;
-    SpvSectionBuilder extensions;
-    SpvSectionBuilder ext_inst_import;
-    SpvSectionBuilder entry_points;
-    SpvSectionBuilder execution_modes;
-    SpvSectionBuilder debug_string_source;
-    SpvSectionBuilder debug_names;
-    SpvSectionBuilder debug_module_processed;
-    SpvSectionBuilder annotations;
-    SpvSectionBuilder types_constants;
-    SpvSectionBuilder fn_decls;
-    SpvSectionBuilder fn_defs;
+    SectionBuilder capabilities;
+    SectionBuilder extensions;
+    SectionBuilder ext_inst_import;
+    SectionBuilder entry_points;
+    SectionBuilder execution_modes;
+    SectionBuilder debug_string_source;
+    SectionBuilder debug_names;
+    SectionBuilder debug_module_processed;
+    SectionBuilder annotations;
+    SectionBuilder types_constants;
+    SectionBuilder fn_decls;
+    SectionBuilder fn_defs;
 
     // SPIR-V disallows duplicate non-aggregate type declarations, we protect against these with this
-    std::unordered_map<UniqueDeclKey, SpvId, UniqueDeclKeyHasher> unique_decls;
-    std::unordered_map<std::string, SpvId> extended_instruction_sets;
+    std::unordered_map<UniqueDeclKey, Id, UniqueDeclKeyHasher> unique_decls;
+    std::unordered_map<std::string, Id> extended_instruction_sets;
     std::unordered_set<spv::Capability> capabilities_set;
     std::unordered_set<std::string> extensions_set;
 
@@ -358,7 +356,7 @@ struct SpvFileBuilder {
         output_->put((word >> 24) & 0xFFu);
     }
 
-    void output_section(SpvSectionBuilder& section) {
+    void output_section(SectionBuilder& section) {
         for (auto& word : section.data_) {
             output_word_le(word);
         }
@@ -366,7 +364,7 @@ struct SpvFileBuilder {
 public:
     void finish(std::ostream& output) {
         output_ = &output;
-        SpvSectionBuilder memory_model_section;
+        SectionBuilder memory_model_section;
         memory_model_section.begin_op(spv::Op::OpMemoryModel, 3);
         memory_model_section.data_.push_back(addressing_model);
         memory_model_section.data_.push_back(memory_model);
@@ -392,25 +390,25 @@ struct SpvFileBuilder {
         output_section(fn_defs);
     }
 
-    friend SpvBasicBlockBuilder;
+    friend BasicBlockBuilder;
 };
 
-struct SpvBasicBlockBuilder : public SpvSectionBuilder {
-    explicit SpvBasicBlockBuilder(SpvFileBuilder& file_builder)
+struct BasicBlockBuilder : public SectionBuilder {
+    explicit BasicBlockBuilder(FileBuilder& file_builder)
             : file_builder(file_builder)
     {}
 
-    SpvFileBuilder& file_builder;
+    FileBuilder& file_builder;
 
     struct Phi {
-        SpvId type;
-        SpvId value;
-        std::vector<std::pair<SpvId, SpvId>> preds;
+        Id type;
+        Id value;
+        std::vector<std::pair<Id, Id>> preds;
     };
     std::vector<Phi*> phis;
-    SpvId label;
+    Id label;
 
-    SpvId undef(SpvId type) {
+    Id undef(Id type) {
         begin_op(spv::Op::OpUndef, 3);
         ref_id(type);
         auto id = file_builder.generate_fresh_id();
@@ -418,7 +416,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId composite(SpvId aggregate_t, std::vector<SpvId>& elements) {
+    Id composite(Id aggregate_t, std::vector<Id>& elements) {
         begin_op(spv::Op::OpCompositeConstruct, 3 + elements.size());
         ref_id(aggregate_t);
         auto id = file_builder.generate_fresh_id();
@@ -428,7 +426,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId extract(SpvId target_type, SpvId composite, std::vector<uint32_t> indices) {
+    Id extract(Id target_type, Id composite, std::vector<uint32_t> indices) {
         begin_op(spv::Op::OpCompositeExtract, 4 + indices.size());
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
@@ -439,7 +437,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId insert(SpvId target_type, SpvId object, SpvId composite, std::vector<uint32_t> indices) {
+    Id insert(Id target_type, Id object, Id composite, std::vector<uint32_t> indices) {
         begin_op(spv::Op::OpCompositeInsert, 5 + indices.size());
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
@@ -451,7 +449,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId vector_extract_dynamic(SpvId target_type, SpvId vector, SpvId index) {
+    Id vector_extract_dynamic(Id target_type, Id vector, Id index) {
         begin_op(spv::Op::OpVectorExtractDynamic, 5);
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
@@ -461,7 +459,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId vector_insert_dynamic(SpvId target_type, SpvId vector, SpvId component, SpvId index) {
+    Id vector_insert_dynamic(Id target_type, Id vector, Id component, Id index) {
         begin_op(spv::Op::OpVectorInsertDynamic, 6);
         ref_id(target_type);
         auto id = file_builder.generate_fresh_id();
@@ -473,7 +471,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
     // Used for almost all conversion operations
-    SpvId convert(spv::Op op_, SpvId target_type, SpvId value) {
+    Id convert(spv::Op op_, Id target_type, Id value) {
         begin_op(op_, 4);
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
@@ -482,7 +480,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId access_chain(SpvId target_type, SpvId element, std::vector<SpvId> indexes) {
+    Id access_chain(Id target_type, Id element, std::vector<Id> indexes) {
         begin_op(spv::Op::OpAccessChain, 4 + indexes.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
@@ -493,7 +491,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId ptr_access_chain(SpvId target_type, SpvId base, SpvId element, std::vector<SpvId> indexes) {
+    Id ptr_access_chain(Id target_type, Id base, Id element, std::vector<Id> indexes) {
         begin_op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
@@ -505,7 +503,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId load(SpvId target_type, SpvId pointer, std::vector<uint32_t> operands = {}) {
+    Id load(Id target_type, Id pointer, std::vector<uint32_t> operands = {}) {
         begin_op(spv::Op::OpLoad, 4 + operands.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(target_type);
@@ -516,7 +514,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    void store(SpvId value, SpvId pointer, std::vector<uint32_t> operands = {}) {
+    void store(Id value, Id pointer, std::vector<uint32_t> operands = {}) {
         begin_op(spv::Op::OpStore, 3 + operands.size());
         ref_id(pointer);
         ref_id(value);
@@ -524,7 +522,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
             literal_int(op);
     }
 
-    SpvId binop(spv::Op op_, SpvId result_type, SpvId lhs, SpvId rhs) {
+    Id binop(spv::Op op_, Id result_type, Id lhs, Id rhs) {
         begin_op(op_, 5);
         auto id = file_builder.generate_fresh_id();
         ref_id(result_type);
@@ -534,19 +532,19 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    void branch(SpvId target) {
+    void branch(Id target) {
         begin_op(spv::Op::OpBranch, 2);
         ref_id(target);
     }
 
-    void branch_conditional(SpvId condition, SpvId true_target, SpvId false_target) {
+    void branch_conditional(Id condition, Id true_target, Id false_target) {
         begin_op(spv::Op::OpBranchConditional, 4);
         ref_id(condition);
         ref_id(true_target);
         ref_id(false_target);
     }
 
-    void branch_switch(SpvId selector, SpvId default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
+    void branch_switch(Id selector, Id default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
         assert(literals.size() == cases.size());
         begin_op(spv::Op::OpSwitch, 3 + literals.size() * 2);
         ref_id(selector);
@@ -557,13 +555,13 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         }
     }
 
-    void selection_merge(SpvId merge_bb, spv::SelectionControlMask selection_control) {
+    void selection_merge(Id merge_bb, spv::SelectionControlMask selection_control) {
         begin_op(spv::Op::OpSelectionMerge, 3);
         ref_id(merge_bb);
         literal_int(selection_control);
     }
 
-    void loop_merge(SpvId merge_bb, SpvId continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
+    void loop_merge(Id merge_bb, Id continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
         begin_op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
         ref_id(merge_bb);
         ref_id(continue_bb);
@@ -573,7 +571,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
             literal_int(e);
     }
 
-    SpvId call(SpvId return_type, SpvId callee, std::vector<SpvId> arguments) {
+    Id call(Id return_type, Id callee, std::vector<Id> arguments) {
         begin_op(spv::Op::OpFunctionCall, 4 + arguments.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(return_type);
@@ -585,13 +583,13 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
         return id;
     }
 
-    SpvId ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments);
+    Id ext_instruction(Id return_type, ExtendedInstruction instr, std::vector<Id> arguments);
 
     void return_void() {
         begin_op(spv::Op::OpReturn, 1);
     }
 
-    void return_value(SpvId value) {
+    void return_value(Id value) {
         begin_op(spv::Op::OpReturnValue, 2);
         ref_id(value);
     }
@@ -601,7 +599,7 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 
 protected:
-    SpvId ext_instruction(SpvId return_type, SpvId set, uint32_t instruction, std::vector<SpvId> arguments) {
+    Id ext_instruction(Id return_type, Id set, uint32_t instruction, std::vector<Id> arguments) {
         begin_op(spv::Op::OpExtInst, 5 + arguments.size());
         auto id = file_builder.generate_fresh_id();
         ref_id(return_type);
@@ -614,26 +612,26 @@ struct SpvBasicBlockBuilder : public SpvSectionBuilder {
     }
 };
 
-struct SpvFnBuilder {
-    explicit SpvFnBuilder(SpvFileBuilder& file_builder)
+struct FnBuilder {
+    explicit FnBuilder(FileBuilder& file_builder)
             : file_builder(file_builder)
     {
         function_id = file_builder.generate_fresh_id();
     }
 
-    SpvFileBuilder& file_builder;
-    SpvId function_id;
+    FileBuilder& file_builder;
+    Id function_id;
 
-    SpvId fn_type;
-    SpvId fn_ret_type;
-    std::vector<SpvBasicBlockBuilder*> bbs_to_emit;
+    Id fn_type;
+    Id fn_ret_type;
+    std::vector<BasicBlockBuilder*> bbs_to_emit;
 
     // Contains OpFunctionParams
-    SpvSectionBuilder header;
+    SectionBuilder header;
 
-    SpvSectionBuilder variables;
+    SectionBuilder variables;
 
-    SpvId parameter(SpvId param_type) {
+    Id parameter(Id param_type) {
         header.begin_op(spv::Op::OpFunctionParameter, 3);
         auto id = file_builder.generate_fresh_id();
         header.ref_id(param_type);
@@ -641,7 +639,7 @@ struct SpvFnBuilder {
         return id;
     }
 
-    SpvId variable(SpvId type, spv::StorageClass storage_class) {
+    Id variable(Id type, spv::StorageClass storage_class) {
         variables.begin_op(spv::Op::OpVariable, 4);
         variables.ref_id(type);
         auto id = file_builder.generate_fresh_id();
@@ -651,7 +649,7 @@ struct SpvFnBuilder {
     }
 };
 
-inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
+inline Id FileBuilder::define_function(FnBuilder &fn_builder) {
     fn_defs.begin_op(spv::Op::OpFunction, 5);
     fn_defs.ref_id(fn_builder.fn_ret_type);
     fn_defs.ref_id(fn_builder.function_id);
@@ -692,7 +690,7 @@ inline SpvId SpvFileBuilder::define_function(SpvFnBuilder &fn_builder) {
     return fn_builder.function_id;
 }
 
-inline SpvId SpvBasicBlockBuilder::ext_instruction(SpvId return_type, ExtendedInstruction instr, std::vector<SpvId> arguments) {
+inline Id BasicBlockBuilder::ext_instruction(Id return_type, ExtendedInstruction instr, std::vector<Id> arguments) {
     return ext_instruction(return_type, file_builder.extended_import(instr.set_name), instr.id, arguments);
 }
 
diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 4a9b53104..484e94a73 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -35,11 +35,11 @@ SpirMathOps opencl_std = {
     .log10 =    { "OpenCL.std", OpenCLLIB::Log10 },
 };
 
-SpvId CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
+Id CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
     auto type = mathop.type();
 
     SpirMathOps& impl = opencl_std;
-    std::vector<SpvId> ops;
+    std::vector<Id> ops;
     for (auto& op : mathop.ops()) {
         ops.push_back(emit(op));
     }
diff --git a/src/thorin/be/spirv/spirv_private.h b/src/thorin/be/spirv/spirv_private.h
index d36d4e81a..ae8f03d51 100644
--- a/src/thorin/be/spirv/spirv_private.h
+++ b/src/thorin/be/spirv/spirv_private.h
@@ -7,7 +7,7 @@
 
 namespace thorin::spirv {
 
-struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
+struct BasicBlockBuilder : public builder::BasicBlockBuilder {
     explicit BasicBlockBuilder(FnBuilder& fn_builder);
 
     BasicBlockBuilder(const BasicBlockBuilder&) = delete;
@@ -19,17 +19,17 @@ struct BasicBlockBuilder : public builder::SpvBasicBlockBuilder {
     bool semi_inline;
 };
 
-struct FnBuilder : public builder::SpvFnBuilder {
+struct FnBuilder : public builder::FnBuilder {
     explicit FnBuilder(FileBuilder& file_builder);
 
     FnBuilder(const FnBuilder&) = delete;
 
     FileBuilder& file_builder;
     std::vector<std::unique_ptr<BasicBlockBuilder>> bbs;
-    DefMap<SpvId> params;
+    DefMap<Id> params;
 };
 
-struct FileBuilder : public builder::SpvFileBuilder {
+struct FileBuilder : public builder::FileBuilder {
     explicit FileBuilder(CodeGen* cg);
     FileBuilder(const FileBuilder&) = delete;
 
@@ -37,14 +37,14 @@ struct FileBuilder : public builder::SpvFileBuilder {
 
     FnBuilder* current_fn_ = nullptr;
     ContinuationMap<std::unique_ptr<FnBuilder>> fn_builders_;
-    std::unordered_map<uint32_t, SpvId> builtins_;
-    std::vector<SpvId> interface;
+    std::unordered_map<uint32_t, Id> builtins_;
+    std::vector<Id> interface;
 
-    SpvId u32_t();
-    SpvId u32_constant(uint32_t);
+    Id u32_t();
+    Id u32_constant(uint32_t);
 
 private:
-    SpvId u32_t_ { 0 };
+    Id u32_t_ { 0 };
 };
 
 }
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 90a718697..f3d767a89 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -164,15 +164,15 @@ ConvertedType CodeGen::convert(const Type* type) {
         case Node_FnType: {
             // extract "return" type, collect all other types
             auto fn = type->as<FnType>();
-            SpvId ret = 0;
-            std::vector<SpvId> ops;
+            Id ret = 0;
+            std::vector<Id> ops;
             for (auto op : fn->types()) {
                 if (op->isa<MemType>() || op == world().unit_type())
                     continue;
                 auto fn_type = op->isa<FnType>();
                 if (fn_type && !op->isa<ClosureType>()) {
                     assert(!ret && "only one 'return' supported");
-                    std::vector<SpvId> ret_types;
+                    std::vector<Id> ret_types;
                     for (auto fn_op : fn_type->types()) {
                         if (fn_op->isa<MemType>() || fn_op == world().unit_type())
                             continue;
@@ -197,7 +197,7 @@ ConvertedType CodeGen::convert(const Type* type) {
 
         case Node_StructType:
         case Node_TupleType: {
-            std::vector<SpvId> spv_types;
+            std::vector<Id> spv_types;
             size_t total_serialized_size = 0;
             converted.layout = { 0, 0 };
             for (auto member : type->ops()) {
@@ -223,7 +223,7 @@ ConvertedType CodeGen::convert(const Type* type) {
         case Node_VariantType: {
             assert(type->num_ops() > 0 && "empty variants not supported");
             auto tag_type = world().type_pu32();
-            SpvId converted_tag_type = convert(tag_type).id;
+            Id converted_tag_type = convert(tag_type).id;
 
             size_t max_serialized_size = 0;
             for (auto member : type->as<VariantType>()->types()) {

From 90261bbaf645a6cc1365a2698df492466af55458 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:47:24 +0100
Subject: [PATCH 306/342] spirv_builder.hpp: move terminators into their own
 section

---
 src/thorin/be/spirv/spirv.cpp         |  18 ++--
 src/thorin/be/spirv/spirv_builder.hpp | 113 ++++++++++++++------------
 2 files changed, 71 insertions(+), 60 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index b8877a567..63a60211c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -308,7 +308,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             }
             j++;
         }
-        bb->branch(emit(succ));
+        bb->terminator.branch(emit(succ));
     };
 
     auto& app = *continuation->body();
@@ -327,9 +327,9 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         }
 
         switch (values.size()) {
-            case 0:  bb->return_void();      break;
-            case 1:  bb->return_value(values[0]); break;
-            default: bb->return_value(bb->composite(builder_->current_fn_->fn_ret_type, values));
+            case 0:  bb->terminator.return_void();      break;
+            case 1:  bb->terminator.return_value(values[0]); break;
+            default: bb->terminator.return_value(bb->composite(builder_->current_fn_->fn_ret_type, values));
         }
     } else if (auto dst_cont = app.callee()->isa_nom<Continuation>(); dst_cont && dst_cont->is_basicblock()) { // ordinary jump
         int index = -1;
@@ -344,7 +344,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             auto& phi = cont2bb_[dst_cont]->phis_map[param];
             phi.preds.emplace_back(val, emit_as_bb(continuation));
         }
-        bb->branch(emit(dst_cont));
+        bb->terminator.branch(emit(dst_cont));
     } else if (app.callee() == world().branch()) {
         auto mem = app.arg(0);
         emit_unsafe(mem);
@@ -352,7 +352,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         auto cond = emit(app.arg(1));
         auto tbb = emit(app.arg(2));
         auto fbb = emit(app.arg(3));
-        bb->branch_conditional(cond, tbb, fbb);
+        bb->terminator.branch_conditional(cond, tbb, fbb);
     } else if (app.callee()->isa<Continuation>() && app.callee()->as<Continuation>()->intrinsic() == Intrinsic::Match) {
         emit_unsafe(app.arg(0));
         auto val = emit(app.arg(1));
@@ -364,9 +364,9 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
             literals.push_back(emit(arg->op(0)));
             cases.push_back(emit_as_bb(arg->op(1)->as_nom<Continuation>()));
         }
-        bb->branch_switch(val, otherwise_bb, literals, cases);
+        bb->terminator.branch_switch(val, otherwise_bb, literals, cases);
     } else if (app.callee()->isa<Bottom>()) {
-        bb->unreachable();
+        bb->terminator.unreachable();
     } else if (auto intrinsic = app.callee()->isa_nom<Continuation>(); intrinsic && (intrinsic->is_intrinsic() || intrinsic->cc() == CC::Device)) {
         // Ensure we emit previous memory operations
         assert(is_mem(app.arg(0)));
@@ -432,7 +432,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
                 j++;
             }
 
-            bb->branch(emit(succ));
+            bb->terminator.branch(emit(succ));
         }
 
         jump_to_next_cont_with_args(succ, args);
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index aca443356..0ee6e8773 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -395,8 +395,7 @@ struct FileBuilder {
 
 struct BasicBlockBuilder : public SectionBuilder {
     explicit BasicBlockBuilder(FileBuilder& file_builder)
-            : file_builder(file_builder)
-    {}
+            : file_builder(file_builder), terminator(*this) {}
 
     FileBuilder& file_builder;
 
@@ -532,45 +531,6 @@ struct BasicBlockBuilder : public SectionBuilder {
         return id;
     }
 
-    void branch(Id target) {
-        begin_op(spv::Op::OpBranch, 2);
-        ref_id(target);
-    }
-
-    void branch_conditional(Id condition, Id true_target, Id false_target) {
-        begin_op(spv::Op::OpBranchConditional, 4);
-        ref_id(condition);
-        ref_id(true_target);
-        ref_id(false_target);
-    }
-
-    void branch_switch(Id selector, Id default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
-        assert(literals.size() == cases.size());
-        begin_op(spv::Op::OpSwitch, 3 + literals.size() * 2);
-        ref_id(selector);
-        ref_id(default_case);
-        for (size_t i = 0; i < literals.size(); i++) {
-            ref_id(literals[i]);
-            ref_id(cases[i]);
-        }
-    }
-
-    void selection_merge(Id merge_bb, spv::SelectionControlMask selection_control) {
-        begin_op(spv::Op::OpSelectionMerge, 3);
-        ref_id(merge_bb);
-        literal_int(selection_control);
-    }
-
-    void loop_merge(Id merge_bb, Id continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
-        begin_op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
-        ref_id(merge_bb);
-        ref_id(continue_bb);
-        literal_int(loop_control);
-
-        for (auto e : loop_control_ops)
-            literal_int(e);
-    }
-
     Id call(Id return_type, Id callee, std::vector<Id> arguments) {
         begin_op(spv::Op::OpFunctionCall, 4 + arguments.size());
         auto id = file_builder.generate_fresh_id();
@@ -585,18 +545,66 @@ struct BasicBlockBuilder : public SectionBuilder {
 
     Id ext_instruction(Id return_type, ExtendedInstruction instr, std::vector<Id> arguments);
 
-    void return_void() {
-        begin_op(spv::Op::OpReturn, 1);
-    }
+    struct TerminatorBuilder : public SectionBuilder {
+        TerminatorBuilder(BasicBlockBuilder& bb) : bb(bb) {}
 
-    void return_value(Id value) {
-        begin_op(spv::Op::OpReturnValue, 2);
-        ref_id(value);
-    }
+        void branch(Id target) {
+            begin_op(spv::Op::OpBranch, 2);
+            ref_id(target);
+        }
 
-    void unreachable() {
-        begin_op(spv::Op::OpUnreachable, 1);
-    }
+        void branch_conditional(Id condition, Id true_target, Id false_target) {
+            begin_op(spv::Op::OpBranchConditional, 4);
+            ref_id(condition);
+            ref_id(true_target);
+            ref_id(false_target);
+        }
+
+        void branch_switch(Id selector, Id default_case, std::vector<uint32_t> literals, std::vector<uint32_t> cases) {
+            assert(literals.size() == cases.size());
+            begin_op(spv::Op::OpSwitch, 3 + literals.size() * 2);
+            ref_id(selector);
+            ref_id(default_case);
+            for (size_t i = 0; i < literals.size(); i++) {
+                ref_id(literals[i]);
+                ref_id(cases[i]);
+            }
+        }
+
+        void selection_merge(Id merge_bb, spv::SelectionControlMask selection_control) {
+            begin_op(spv::Op::OpSelectionMerge, 3);
+            ref_id(merge_bb);
+            literal_int(selection_control);
+        }
+
+        void loop_merge(Id merge_bb, Id continue_bb, spv::LoopControlMask loop_control, std::vector<uint32_t> loop_control_ops) {
+            begin_op(spv::Op::OpLoopMerge, 4 + loop_control_ops.size());
+            ref_id(merge_bb);
+            ref_id(continue_bb);
+            literal_int(loop_control);
+
+            for (auto e : loop_control_ops)
+                literal_int(e);
+        }
+
+        void return_void() {
+            begin_op(spv::Op::OpReturn, 1);
+        }
+
+        void return_value(Id value) {
+            begin_op(spv::Op::OpReturnValue, 2);
+            ref_id(value);
+        }
+
+        void unreachable() {
+            begin_op(spv::Op::OpUnreachable, 1);
+        }
+
+    private:
+        BasicBlockBuilder& bb;
+    };
+
+    TerminatorBuilder terminator;
 
 protected:
     Id ext_instruction(Id return_type, Id set, uint32_t instruction, std::vector<Id> arguments) {
@@ -684,6 +692,9 @@ inline Id FileBuilder::define_function(FnBuilder &fn_builder) {
 
         for (auto w : bb->data_)
             fn_defs.data_.push_back(w);
+
+        for (auto w : bb->terminator.data_)
+            fn_defs.data_.push_back(w);
     }
 
     fn_defs.begin_op(spv::Op::OpFunctionEnd, 1);

From 7dd78b9731f9d5ee51206f6f29538d4768c8ec84 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:55:15 +0100
Subject: [PATCH 307/342] spirv: fix int8 and fp16, 64 not declaring relevant
 capabilities

---
 src/thorin/be/spirv/spirv_types.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index f3d767a89..823ad2faf 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -92,10 +92,12 @@ ConvertedType CodeGen::convert(const Type* type) {
             converted.layout = { 1, 1 };
             break;
         case Node_PrimType_ps8:
+            builder_->capability(spv::Capability::CapabilityInt8);
             converted.id = builder_->declare_int_type(8, true);
             converted.layout = { 1, 1 };
             break;
         case Node_PrimType_pu8:
+            builder_->capability(spv::Capability::CapabilityInt8);
             converted.id = builder_->declare_int_type(8, false);
             converted.layout = { 1, 1 };
             break;
@@ -128,6 +130,7 @@ ConvertedType CodeGen::convert(const Type* type) {
             converted.layout = { 8, 8 };
             break;
         case Node_PrimType_pf16:
+            builder_->capability(spv::Capability::CapabilityFloat16);
             converted.id = builder_->declare_float_type(16);
             converted.layout = { 2, 2 };
             break;
@@ -136,6 +139,7 @@ ConvertedType CodeGen::convert(const Type* type) {
             converted.layout = { 4, 4 };
             break;
         case Node_PrimType_pf64:
+            builder_->capability(spv::Capability::CapabilityFloat64);
             converted.id = builder_->declare_float_type(64);
             converted.layout = { 8, 8 };
             break;

From f26d0fdafe108cfad7ec4a2af0831e884abb28d3 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 19:59:45 +0100
Subject: [PATCH 308/342] spirv_builder.hpp: ensure only one TypeVoid exists

---
 src/thorin/be/spirv/spirv_builder.hpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 0ee6e8773..ea0456f36 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -72,6 +72,7 @@ struct SectionBuilder {
 struct FileBuilder {
     enum UniqueDeclTag {
         NONE,
+        VOID_TYPE,
         FN_TYPE,
         PTR_TYPE,
         DEF_ARR_TYPE,
@@ -264,9 +265,12 @@ struct FileBuilder {
     }
 
     Id declare_void_type() {
+        auto key = UniqueDeclKey { VOID_TYPE, { } };
+        if (auto iter = unique_decls.find(key); iter != unique_decls.end()) return iter->second;
         types_constants.begin_op(spv::Op::OpTypeVoid, 2);
         auto id = generate_fresh_id();
         types_constants.ref_id(id);
+        unique_decls[key] = id;
         return id;
     }
 

From b5d6d85e0e1b2b389d885819b3e003c14db75aca Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 20:14:03 +0100
Subject: [PATCH 309/342] spirv: fix LEA on array types

---
 src/thorin/be/spirv/spirv.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 63a60211c..97e2239a2 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -641,7 +641,7 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto ptr_type = access->ptr()->type()->as<PtrType>();
         if (ptr_type->addr_space() == AddrSpace::Global) {
             operands.push_back(spv::MemoryAccessAlignedMask);
-            operands.push_back( 4 ); // TODO: SPIR-V docs say to consult client API for valid values.
+            operands.push_back(4); // TODO: SPIR-V docs say to consult client API for valid values.
         }
         if (auto load = def->isa<Load>()) {
             return bb->load(convert(load->out_val_type()).id, emit(load->ptr()), operands);
@@ -668,8 +668,10 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         //}
         auto type = convert(lea->type()).id;
         auto offset = emit(lea->index());
-        if (lea->ptr_pointee()->isa<ArrayType>())
-            return bb->ptr_access_chain(type, emit(lea->ptr()), offset, {  });
+        if (auto arr_type = lea->ptr_pointee()->isa<ArrayType>()) {
+            auto base = bb->convert(spv::OpBitcast, type, emit(lea->ptr()));
+            return bb->ptr_access_chain(type, base, offset, {  });
+        }
         return bb->access_chain(type, emit(lea->ptr()), { offset });
     } else if (auto aggop = def->isa<AggOp>()) {
         auto agg_type = convert(aggop->agg()->type()).id;

From 19108af9817ca0d2b7069a585ea5f1474131edae Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 20:14:25 +0100
Subject: [PATCH 310/342] spirv: use OpPtrCastToGeneric when allocating on the
 stack

---
 src/thorin/be/spirv/spirv.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 97e2239a2..15ef71773 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -653,7 +653,7 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         emit_unsafe(slot->frame());
         auto type = slot->type();
         auto id = bb->fn_builder.variable(convert(world().ptr_type(type->pointee(), 1, AddrSpace::Function)).id, spv::StorageClass::StorageClassFunction);
-        id = bb->convert(spv::Op::OpBitcast, convert(type).id, id);
+        id = bb->convert(spv::Op::OpPtrCastToGeneric, convert(type).id, id);
         return id;
     } else if (auto enter = def->isa<Enter>()) {
         return emit_unsafe(enter->mem());

From f7819d15d84bf1b08301c18596d0054752b15a6d Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 22:28:36 +0100
Subject: [PATCH 311/342] spirv: workarround for broken OpCompositeConstruct

---
 src/thorin/be/spirv/spirv.cpp | 44 +++++++++++++++++++++++------------
 src/thorin/be/spirv/spirv.h   |  6 +++++
 2 files changed, 35 insertions(+), 15 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 15ef71773..65aa43fb3 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -329,7 +329,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         switch (values.size()) {
             case 0:  bb->terminator.return_void();      break;
             case 1:  bb->terminator.return_value(values[0]); break;
-            default: bb->terminator.return_value(bb->composite(builder_->current_fn_->fn_ret_type, values));
+            default: bb->terminator.return_value(emit_composite(bb, builder_->current_fn_->fn_ret_type, values));
         }
     } else if (auto dst_cont = app.callee()->isa_nom<Continuation>(); dst_cont && dst_cont->is_basicblock()) { // ordinary jump
         int index = -1;
@@ -471,6 +471,32 @@ Id CodeGen::emit_constant(const thorin::Def* def) {
 
     assertf(false, "Incomplete emit(def) definition");
 }
+Id CodeGen::emit_composite(BasicBlockBuilder* bb, Id t, Defs defs) {
+    std::vector<Id> ids;
+    for (auto& def : defs) {
+        ids.push_back(emit(def));
+    }
+    return emit_composite(bb, t, ids);
+}
+
+Id CodeGen::emit_composite(BasicBlockBuilder* bb, Id t, ArrayRef<Id> ids) {
+    if (target_info_.bugs.broken_op_construct_composite) {
+        Id c = bb->undef(t);
+        uint32_t x = 0;
+        for (auto& e : ids) {
+            c = bb->insert(t, e, c, { x++ });
+        }
+        return c;
+    } else {
+        std::vector<Id> elements;
+        elements.resize(ids.size());
+        size_t x = 0;
+        for (auto& e : ids) {
+            elements[x++] = e;
+        }
+        return bb->composite(t, elements);
+    }
+}
 
 Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     if (auto mathop = def->isa<MathOp>())
@@ -618,21 +644,9 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto value = emit(vindex->op(0));
         return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
-        std::vector<Id> elements;
-        elements.resize(tuple->num_ops());
-        size_t x = 0;
-        for (auto& e : tuple->ops()) {
-            elements[x++] = emit(e);
-        }
-        return bb->composite(convert(tuple->type()).id, elements);
+        return emit_composite(bb, convert(tuple->type()).id, tuple->ops());
     } else if (auto structagg = def->isa<StructAgg>()) {
-        std::vector<Id> elements;
-        elements.resize(structagg->num_ops());
-        size_t x = 0;
-        for (auto& e : structagg->ops()) {
-            elements[x++] = emit(e);
-        }
-        return bb->composite(convert(structagg->type()).id, elements);
+        return emit_composite(bb, convert(structagg->type()).id, structagg->ops());
     } else if (auto access = def->isa<Access>()) {
         // emit dependent operations first
         emit_unsafe(access->mem());
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index fe0c70c78..1eb733a13 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -20,6 +20,10 @@ struct Target {
         size_t pointer_size = 8;
     } mem_layout;
 
+    struct {
+        bool broken_op_construct_composite = true;
+    } bugs;
+
     enum Dialect {
         OpenCL,
         Vulkan
@@ -68,6 +72,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
 
     Id emit_as_bb(Continuation*);
     Id emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
+    Id emit_composite(BasicBlockBuilder* bb, Id, Defs);
+    Id emit_composite(BasicBlockBuilder* bb, Id, ArrayRef<Id>);
 
     Id get_codom_type(const Continuation* fn);
 

From d4786b05f82a70338edf9efb8903f652d3f8a1bd Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Mon, 4 Nov 2024 22:49:24 +0100
Subject: [PATCH 312/342] account for another bug in the SPIRV-LLVM translator

---
 src/thorin/be/spirv/spirv.cpp | 2 ++
 src/thorin/be/spirv/spirv.h   | 1 +
 2 files changed, 3 insertions(+)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 65aa43fb3..9f872c522 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -686,6 +686,8 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             auto base = bb->convert(spv::OpBitcast, type, emit(lea->ptr()));
             return bb->ptr_access_chain(type, base, offset, {  });
         }
+        if (target_info_.bugs.static_ac_indices_must_be_i32)
+            offset = emit(world().cast(world().type_pu32(), lea->index()));
         return bb->access_chain(type, emit(lea->ptr()), { offset });
     } else if (auto aggop = def->isa<AggOp>()) {
         auto agg_type = convert(aggop->agg()->type()).id;
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 1eb733a13..ec5645879 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -22,6 +22,7 @@ struct Target {
 
     struct {
         bool broken_op_construct_composite = true;
+        bool static_ac_indices_must_be_i32 = true;
     } bugs;
 
     enum Dialect {

From 17991eb96d2c5111828bfa9b84da825c59244c31 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 8 Nov 2024 15:41:47 +0100
Subject: [PATCH 313/342] move emit_intrinsic to spirv_instructions.cpp

---
 src/thorin/be/spirv/spirv.cpp              | 43 --------------------
 src/thorin/be/spirv/spirv_instructions.cpp | 46 +++++++++++++++++++++-
 2 files changed, 45 insertions(+), 44 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 9f872c522..e0a2e5746 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -858,47 +858,4 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
-std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
-    std::vector<Id> productions;
-    if (intrinsic->name() == "spirv.nonsemantic.printf") {
-        std::vector<Id> args;
-        auto string = app.arg(1);
-        if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
-            auto arr = string->as<DefiniteArray>();
-            std::vector<char> the_string;
-            for (size_t i = 0; i < arr_type->dim(); i++)
-                the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
-            the_string.push_back('\0');
-            args.push_back(builder_->debug_string(the_string.data()));
-        } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
-
-        for (size_t i = 2; i < app.num_args() - 1; i++) {
-            args.push_back(emit(app.arg(i)));
-        }
-
-        builder_->extension("SPV_KHR_non_semantic_info");
-        bb->ext_instruction(convert(world().unit_type()).id, { "NonSemantic.DebugPrintf", 1}, args);
-    } else if (intrinsic->name() == "spirv.builtin") {
-        if (auto spv_builtin_lit = app.arg(1)->isa<PrimLit>()) {
-            auto spv_builtin = spv_builtin_lit->value().get_u32();
-            auto found = builder_->builtins_.find(spv_builtin);
-            if (found != builder_->builtins_.end()) {
-                productions.push_back(found->second);
-            } else {
-                auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
-                auto desired_type = ret_type->types()[1]->as<PtrType>();
-                auto id = builder_->variable(convert(desired_type).id, static_cast<spv::StorageClass>(convert(desired_type->addr_space())));
-                builder_->interface.push_back(id);
-                builder_->decorate(id, spv::Decoration::DecorationBuiltIn, { spv_builtin });
-                builder_->builtins_[spv_builtin] = id;
-                productions.push_back(id);
-            }
-        } else
-            world().ELOG("spirv.builtin requires an integer literal as the argument");
-    } else {
-        world().ELOG("This spir-v builtin isn't recognised: {}", intrinsic->name());
-    }
-    return productions;
-}
-
 }
diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 484e94a73..3a4f3e0ca 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -52,4 +52,48 @@ Id CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
     }
 }
 
-}
\ No newline at end of file
+std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
+    std::vector<Id> productions;
+
+    if (intrinsic->name() == "spirv.nonsemantic.printf") {
+        std::vector<Id> args;
+        auto string = app.arg(1);
+        if (auto arr_type = string->type()->isa<DefiniteArrayType>(); arr_type->elem_type() == world().type_pu8()) {
+            auto arr = string->as<DefiniteArray>();
+            std::vector<char> the_string;
+            for (size_t i = 0; i < arr_type->dim(); i++)
+                the_string.push_back(arr->op(i)->as<PrimLit>()->value().get_u8());
+            the_string.push_back('\0');
+            args.push_back(builder_->debug_string(the_string.data()));
+        } else world().ELOG("spirv.nonsemantic.printf takes a string literal");
+
+        for (size_t i = 2; i < app.num_args() - 1; i++) {
+            args.push_back(emit(app.arg(i)));
+        }
+
+        builder_->extension("SPV_KHR_non_semantic_info");
+        bb->ext_instruction(convert(world().unit_type()).id, { "NonSemantic.DebugPrintf", 1}, args);
+    } else if (intrinsic->name() == "spirv.builtin") {
+        if (auto spv_builtin_lit = app.arg(1)->isa<PrimLit>()) {
+            auto spv_builtin = spv_builtin_lit->value().get_u32();
+            auto found = builder_->builtins_.find(spv_builtin);
+            if (found != builder_->builtins_.end()) {
+                productions.push_back(found->second);
+            } else {
+                auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
+                auto desired_type = ret_type->types()[1]->as<PtrType>();
+                auto id = builder_->variable(convert(desired_type).id, static_cast<spv::StorageClass>(convert(desired_type->addr_space())));
+                builder_->interface.push_back(id);
+                builder_->decorate(id, spv::Decoration::DecorationBuiltIn, { spv_builtin });
+                builder_->builtins_[spv_builtin] = id;
+                productions.push_back(id);
+            }
+        } else
+            world().ELOG("spirv.builtin requires an integer literal as the argument");
+    } else {
+        world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());
+    }
+    return productions;
+}
+
+}

From 720c161d38418323c7876b6678cc232edccce31c Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 8 Nov 2024 15:53:48 +0100
Subject: [PATCH 314/342] spirv: use the mathop path for intrinsics named the
 same

---
 src/thorin/be/spirv/spirv.cpp              | 15 +++++++++++++++
 src/thorin/be/spirv/spirv.h                |  1 +
 src/thorin/be/spirv/spirv_instructions.cpp | 16 ++++++++++++++--
 3 files changed, 30 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index e0a2e5746..a640f0662 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -471,6 +471,21 @@ Id CodeGen::emit_constant(const thorin::Def* def) {
 
     assertf(false, "Incomplete emit(def) definition");
 }
+
+std::vector<Id> CodeGen::emit_args(Defs defs) {
+    std::vector<Id> emitted;
+    for (auto arg : defs) {
+        auto arg_type = arg->type();
+        if (arg_type == world().unit_type() || arg_type == world().mem_type()) {
+            emit_unsafe(arg);
+            continue;
+        } else {
+            emitted.push_back(emit(arg));
+        }
+    }
+    return emitted;
+}
+
 Id CodeGen::emit_composite(BasicBlockBuilder* bb, Id t, Defs defs) {
     std::vector<Id> ids;
     for (auto& def : defs) {
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index ec5645879..0e8dc4355 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -70,6 +70,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
 protected:
     FnBuilder& get_fn_builder(Continuation*);
     std::vector<Id> emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb);
+    std::vector<Id> emit_args(Defs);
 
     Id emit_as_bb(Continuation*);
     Id emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 3a4f3e0ca..a7eb19983 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -55,6 +55,19 @@ Id CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
 std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
     std::vector<Id> productions;
 
+    auto get_produced_type = [&]() {
+        auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
+        return ret_type->types()[1];
+    };
+
+    SpirMathOps& impl = opencl_std;
+    auto intrinsic_name = intrinsic->name();
+#define THORIN_MATHOP(mathop_name) \
+    if ((#mathop_name) == intrinsic_name) { \
+        return { bb->ext_instruction(convert(get_produced_type()).id, impl.mathop_name, emit_args(app.args().skip_back())) }; \
+    }
+#include "thorin/tables/mathoptable.h"
+
     if (intrinsic->name() == "spirv.nonsemantic.printf") {
         std::vector<Id> args;
         auto string = app.arg(1);
@@ -80,8 +93,7 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             if (found != builder_->builtins_.end()) {
                 productions.push_back(found->second);
             } else {
-                auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
-                auto desired_type = ret_type->types()[1]->as<PtrType>();
+                auto desired_type = get_produced_type()->as<PtrType>();
                 auto id = builder_->variable(convert(desired_type).id, static_cast<spv::StorageClass>(convert(desired_type->addr_space())));
                 builder_->interface.push_back(id);
                 builder_->decorate(id, spv::Decoration::DecorationBuiltIn, { spv_builtin });

From 4c703c2a9a1c81dd734981b2320402d316fe1eb2 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 8 Nov 2024 16:42:39 +0100
Subject: [PATCH 315/342] spirv: added support for reserve_shared

---
 src/thorin/be/spirv/spirv_instructions.cpp | 21 ++++++++++++++++-----
 src/thorin/be/spirv/spirv_types.cpp        |  4 ++++
 2 files changed, 20 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index a7eb19983..0552cbcc1 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -53,8 +53,6 @@ Id CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
 }
 
 std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
-    std::vector<Id> productions;
-
     auto get_produced_type = [&]() {
         auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
         return ret_type->types()[1];
@@ -86,7 +84,9 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
 
         builder_->extension("SPV_KHR_non_semantic_info");
         bb->ext_instruction(convert(world().unit_type()).id, { "NonSemantic.DebugPrintf", 1}, args);
+        return {};
     } else if (intrinsic->name() == "spirv.builtin") {
+        std::vector<Id> productions;
         if (auto spv_builtin_lit = app.arg(1)->isa<PrimLit>()) {
             auto spv_builtin = spv_builtin_lit->value().get_u32();
             auto found = builder_->builtins_.find(spv_builtin);
@@ -100,12 +100,23 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
                 builder_->builtins_[spv_builtin] = id;
                 productions.push_back(id);
             }
+            return productions;
         } else
             world().ELOG("spirv.builtin requires an integer literal as the argument");
-    } else {
-        world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());
+    } else if (intrinsic->name() == "reserve_shared") {
+        auto size = app.arg(1)->isa<PrimLit>();
+        if (!size)
+            world().error(app.loc(), "reserve_shared called with non-constant size");
+        else {
+            auto in_bytes = size->value().get_u64();
+            auto type = world().definite_array_type(world().type_pu8(), in_bytes);
+            Id id = builder_->variable(convert(type).id, spv::StorageClass::StorageClassCrossWorkgroup);
+            id = bb->convert(spv::Op::OpBitcast, convert(get_produced_type()).id, id);
+            return { id };
+        }
     }
-    return productions;
+    world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());
+    exit(-1);
 }
 
 }
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 823ad2faf..8f72cb3a5 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -19,6 +19,10 @@ uint32_t CodeGen::convert(AddrSpace as) {
             builder_->capability(spv::Capability::CapabilityGenericPointer);
             break;
         }
+        case AddrSpace::Shared: {
+            storage_class = spv::StorageClassWorkgroup;
+            break;
+        }
         case AddrSpace::Push:     storage_class = spv::StorageClassPushConstant;          break;
         case AddrSpace::Input:    storage_class = spv::StorageClassInput;                 break;
         case AddrSpace::Output:   storage_class = spv::StorageClassOutput;                break;

From e6b9ad2b06becabcdff6be1c255a33f0977c1e67 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 8 Nov 2024 16:42:55 +0100
Subject: [PATCH 316/342] spirv: added support for 'min', 'max' intrinsics

---
 src/thorin/be/spirv/spirv_instructions.cpp | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 0552cbcc1..dd95e76b5 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -114,6 +114,22 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             id = bb->convert(spv::Op::OpBitcast, convert(get_produced_type()).id, id);
             return { id };
         }
+    } else if (intrinsic->name() == "min") {
+        auto type = get_produced_type();
+        if (is_type_f(type))
+            return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::Fmin }, emit_args(app.args().skip_back())) };
+        if (is_type_u(type))
+            return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::UMin }, emit_args(app.args().skip_back())) };
+        if (is_type_i(type))
+            return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::SMin }, emit_args(app.args().skip_back())) };
+    } else if (intrinsic->name() == "max") {
+        auto type = get_produced_type();
+        if (is_type_f(type))
+            return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::Fmax }, emit_args(app.args().skip_back())) };
+        if (is_type_u(type))
+            return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::UMax }, emit_args(app.args().skip_back())) };
+        if (is_type_i(type))
+            return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::SMax }, emit_args(app.args().skip_back())) };
     }
     world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());
     exit(-1);

From 5dd14ae3d47762024f8a100203df47ac45700c8f Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 8 Nov 2024 21:51:49 +0100
Subject: [PATCH 317/342] spirv: implement more ops

---
 src/thorin/be/spirv/spirv.cpp              | 21 ++++++++++++++++---
 src/thorin/be/spirv/spirv_builder.hpp      | 24 ++++++++++++----------
 src/thorin/be/spirv/spirv_instructions.cpp |  9 ++++++++
 3 files changed, 40 insertions(+), 14 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index a640f0662..cf64e2126 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -293,13 +293,13 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         BasicBlockBuilder* dstbb = cont2bb_[succ];
 
         for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
-            assert(j < args.size());
             auto param = succ->param(i);
             if (is_mem(param) || is_unit(param)) {
                 if (dstbb->semi_inline)
                     defs_[param] = 0;
                 continue;
             }
+            assert(j < args.size());
             if (dstbb->semi_inline) {
                 defs_[param] = args[j];
             } else {
@@ -786,15 +786,28 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         if (auto bitcast = def->isa<Bitcast>()) {
             assert(conv_src_type.layout && conv_dst_type.layout);
             if (conv_src_type.layout->size != conv_dst_type.layout->size)
-                world().ELOG("Source (%) and destination (%) datatypes sizes do not match (% vs % bytes)", src_type->to_string(), dst_type->to_string(), conv_src_type.layout->size, conv_dst_type.layout->size);
+                world().ELOG("Source ({}) and destination ({}) datatypes sizes do not match ({} vs {} bytes)", src_type, dst_type, conv_src_type.layout->size, conv_dst_type.layout->size);
 
             return bb->convert(spv::OpBitcast, convert(bitcast->type()).id, emit(bitcast->from()));
         } else if (auto cast = def->isa<Cast>()) {
+            if (auto src_ptr_type = src_type->isa<PtrType>()) {
+                if (auto dst_ptr_type = dst_type->isa<PtrType>()) {
+                    if (src_ptr_type->addr_space() == AddrSpace::Generic) {
+                        return bb->op_with_result(spv::Op::OpGenericCastToPtr, convert(cast->type()).id, { emit(cast->from()) });
+                    } else if (dst_ptr_type->addr_space() == AddrSpace::Generic) {
+                        return bb->op_with_result(spv::Op::OpPtrCastToGeneric, convert(cast->type()).id, { emit(cast->from()) });
+                    } else {
+                        world().WLOG("Abnormal ptr-ptr cast: {} to {}, should be a bitcast", src_ptr_type, dst_ptr_type);
+                        return bb->convert(spv::OpBitcast, convert(cast->type()).id, emit(cast->from()));
+                    }
+                }
+            }
+
             // NB: all ops used here are scalar/vector agnostic
             auto src_prim = src_type->isa<PrimType>();
             auto dst_prim = dst_type->isa<PrimType>();
             if (!src_prim || !dst_prim || src_prim->length() != dst_prim->length())
-                world().ELOG("Illegal cast: % to %, casts are only supported between primitives with identical vector length", src_type->to_string(), dst_type->to_string());
+                world().ELOG("Illegal cast: {} to {}, casts are only supported between primitives with identical vector length", src_type, dst_type);
 
             auto length = src_prim->length();
 
@@ -865,6 +878,8 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         } else THORIN_UNREACHABLE;
     } else if (def->isa<Bottom>()) {
         return bb->undef(convert(def->type()).id);
+    } else if (auto select = def->isa<Select>()) {
+        return bb->op_with_result(spv::Op::OpSelect, convert(def->type()).id, emit_args(select->ops()));
     }
 
     if (!def->has_dep(Dep::Param))
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index ea0456f36..0a2eed189 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -411,24 +411,26 @@ struct BasicBlockBuilder : public SectionBuilder {
     std::vector<Phi*> phis;
     Id label;
 
-    Id undef(Id type) {
-        begin_op(spv::Op::OpUndef, 3);
-        ref_id(type);
-        auto id = file_builder.generate_fresh_id();
-        ref_id(id);
-        return id;
+    void op(spv::Op op, std::vector<uint32_t> operands) {
+        begin_op(op, operands.size() + 1);
+        for (auto e : operands)
+            literal_int(e);
     }
 
-    Id composite(Id aggregate_t, std::vector<Id>& elements) {
-        begin_op(spv::Op::OpCompositeConstruct, 3 + elements.size());
-        ref_id(aggregate_t);
+    Id op_with_result(spv::Op op, Id type, std::vector<uint32_t> operands) {
+        begin_op(op, operands.size() + 3);
+        ref_id(type);
         auto id = file_builder.generate_fresh_id();
         ref_id(id);
-        for (auto e : elements)
-            ref_id(e);
+        for (auto e : operands)
+            literal_int(e);
         return id;
     }
 
+    Id undef(Id type) { return op_with_result(spv::Op::OpUndef, type, {}); }
+
+    Id composite(Id aggregate_t, std::vector<Id> elements) { return op_with_result(spv::Op::OpCompositeConstruct, aggregate_t, elements); }
+
     Id extract(Id target_type, Id composite, std::vector<uint32_t> indices) {
         begin_op(spv::Op::OpCompositeExtract, 4 + indices.size());
         ref_id(target_type);
diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index dd95e76b5..44a41ebfd 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -130,6 +130,15 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::UMax }, emit_args(app.args().skip_back())) };
         if (is_type_i(type))
             return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::SMax }, emit_args(app.args().skip_back())) };
+    } else if (intrinsic->name() == "barrier") {
+        emit_args(app.args().skip_back());
+        bb->op(spv::Op::OpMemoryBarrier, { spv::Scope::ScopeInvocation, spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask });
+        return { };
+    } else if (intrinsic->name() == "atomic_add") {
+        auto args = emit_args(app.args().skip_back());
+        auto [ptr, value] = *(std::array<Id, 2>*)args.data();
+        auto result = bb->op_with_result(spv::Op::OpAtomicIAdd, convert(get_produced_type()).id,  { ptr, spv::Scope::ScopeInvocation, spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask, value });
+        return { result };
     }
     world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());
     exit(-1);

From ce056c16016012308d6845472fba451214dc2f85 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 13:44:04 +0100
Subject: [PATCH 318/342] spirv: fix OpBarrier and OpIAtomicAdd emission

---
 src/thorin/be/spirv/spirv.cpp              | 4 ++++
 src/thorin/be/spirv/spirv.h                | 1 +
 src/thorin/be/spirv/spirv_instructions.cpp | 4 ++--
 3 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index cf64e2126..b1b3a0ee1 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -486,6 +486,10 @@ std::vector<Id> CodeGen::emit_args(Defs defs) {
     return emitted;
 }
 
+Id CodeGen::literal(uint32_t value) {
+    return emit(world().literal_pu32(value, {}));
+}
+
 Id CodeGen::emit_composite(BasicBlockBuilder* bb, Id t, Defs defs) {
     std::vector<Id> ids;
     for (auto& def : defs) {
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 0e8dc4355..4414e917d 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -71,6 +71,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     FnBuilder& get_fn_builder(Continuation*);
     std::vector<Id> emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb);
     std::vector<Id> emit_args(Defs);
+    Id literal(uint32_t);
 
     Id emit_as_bb(Continuation*);
     Id emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 44a41ebfd..89a37e20d 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -132,12 +132,12 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::SMax }, emit_args(app.args().skip_back())) };
     } else if (intrinsic->name() == "barrier") {
         emit_args(app.args().skip_back());
-        bb->op(spv::Op::OpMemoryBarrier, { spv::Scope::ScopeInvocation, spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask });
+        bb->op(spv::Op::OpMemoryBarrier, { literal(spv::Scope::ScopeInvocation), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask) });
         return { };
     } else if (intrinsic->name() == "atomic_add") {
         auto args = emit_args(app.args().skip_back());
         auto [ptr, value] = *(std::array<Id, 2>*)args.data();
-        auto result = bb->op_with_result(spv::Op::OpAtomicIAdd, convert(get_produced_type()).id,  { ptr, spv::Scope::ScopeInvocation, spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask, value });
+        auto result = bb->op_with_result(spv::Op::OpAtomicIAdd, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeInvocation), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask), value });
         return { result };
     }
     world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());

From ffd197fd499b1699d74ee7d5f5f592ad5b08cc3e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 13:44:20 +0100
Subject: [PATCH 319/342] spirv: support for emitting variants

---
 src/thorin/be/spirv/spirv.cpp       | 52 ++++++++++++++++-------------
 src/thorin/be/spirv/spirv.h         |  3 ++
 src/thorin/be/spirv/spirv_types.cpp | 11 +++---
 3 files changed, 37 insertions(+), 29 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index b1b3a0ee1..6066df824 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -625,40 +625,44 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             THORIN_UNREACHABLE;
         }
     } else if (auto variant = def->isa<Variant>()) {
-        assert(false && "TODO: rewrite");
-        /*auto variant_type = def->type()->as<VariantType>();
-        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
+        auto variant_type = def->type()->as<VariantType>();
+        auto converted_type = convert(variant_type);
         auto tag = builder_->u32_constant(variant->index());
 
-        if (variant_datatype->elements_types.size() > 1) {
-            auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function));
-            auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
-            auto converted_payload_type = convert(variant_type->op(variant->index()));
+        if (auto payload_t = converted_type.variant.payload_t) {
+            auto scratch_type = convert(world().ptr_type(payload_t.value(), 1, AddrSpace::Function));
+            auto scratch = bb->fn_builder.variable(scratch_type.id, spv::StorageClassFunction);
 
-            converted_payload_type->datatype->emit_serialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0), emit(variant->value(), bb));
-            auto payload = bb->load(variant_datatype->elements_types[1], payload_arr);
+            auto cast_type = convert(world().ptr_type(variant->value()->type(), 1, AddrSpace::Function));
+            auto scratch_cast = bb->op_with_result(spv::Op::OpBitcast, cast_type.id, { scratch });
+            bb->store(scratch_cast, emit(variant->value()));
 
-            std::vector<SpvId> with_tag = {tag, payload};
-            return bb->composite(convert(variant->type()), with_tag);
+            auto payload = bb->load(convert(payload_t.value()).id, scratch);
+
+            std::vector<Id> with_tag = {tag, payload};
+            return bb->composite(converted_type.id, with_tag);
         } else {
             // Zero-sized payload case
-            std::vector<SpvId> with_tag = { tag };
-            return bb->composite(convert(variant->type()), with_tag);
-        }*/
+            std::vector<Id> with_tag = { tag };
+            return bb->composite(converted_type.id, with_tag);
+        }
     } else if (auto vextract = def->isa<VariantExtract>()) {
-        assert(false && "TODO: rewrite");
-        /*auto variant_type = vextract->value()->type()->as<VariantType>();
-        auto variant_datatype = (ProductDatatype*) convert(variant_type)->datatype.get();
+        auto variant_type = vextract->value()->type()->as<VariantType>();
+        auto converted_type = convert(variant_type);
+        auto target_type = def->type();
+
+        assert(converted_type.variant.payload_t.has_value() && "Can't extract zero-sized datatypes");
+        auto payload_t = converted_type.variant.payload_t.value();
 
-        auto target_type = convert(def->type());
+        auto scratch_type = convert(world().ptr_type(target_type, 1, AddrSpace::Function));
+        auto scratch = bb->fn_builder.variable(scratch_type.id, spv::StorageClassFunction);
 
-        assert(variant_datatype->elements_types.size() > 1 && "Can't extract zero-sized datatypes");
-        auto alloc_type = convert(world().ptr_type(variant_datatype->elements_types[1]->src_type, 1, 4, AddrSpace::Function));
-        auto payload_arr = current_fn_->variable(alloc_type, spv::StorageClassFunction);
-        auto payload = bb->extract(variant_datatype->elements_types[1], emit(vextract->value(), bb), {1});
-        bb->store(payload, payload_arr);
+        auto cast_type = convert(world().ptr_type(payload_t, 1, AddrSpace::Function));
+        auto scratch_cast = bb->op_with_result(spv::Op::OpBitcast, cast_type.id, { scratch });
+        auto payload = bb->extract(convert(payload_t).id, emit(vextract->value()), { 1 });
+        bb->store(payload, scratch_cast);
 
-        return target_type->datatype->emit_deserialization(*bb, spv::StorageClassFunction, payload_arr, bb->file_builder.u32_constant(0));*/
+        return bb->load(convert(target_type).id, scratch);
     } else if (auto vindex = def->isa<VariantIndex>()) {
         auto value = emit(vindex->op(0));
         return bb->extract(convert(world().type_pu32()).id, value, { 0 });
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 4414e917d..a1fe867cc 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -39,6 +39,9 @@ struct ConvertedType {
         size_t size, alignment;
     };
     std::optional<Layout> layout;
+    struct {
+        std::optional<const thorin::Type*> payload_t;
+    } variant;
 };
 
 struct BasicBlockBuilder;
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 8f72cb3a5..b4dfb3796 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -231,14 +231,14 @@ ConvertedType CodeGen::convert(const Type* type) {
         case Node_VariantType: {
             assert(type->num_ops() > 0 && "empty variants not supported");
             auto tag_type = world().type_pu32();
-            Id converted_tag_type = convert(tag_type).id;
 
             size_t max_serialized_size = 0;
             for (auto member : type->as<VariantType>()->types()) {
                 auto member_type = member->as<Type>();
                 if (member_type == world().unit_type() || member_type == world().mem_type()) continue;
                 auto converted_member_type = convert(member_type);
-                assert(converted_member_type.layout);
+                if(!converted_member_type.layout)
+                    continue;
                 if (converted_member_type.layout->size > max_serialized_size)
                     max_serialized_size = converted_member_type.layout->size;
             }
@@ -248,14 +248,15 @@ ConvertedType CodeGen::convert(const Type* type) {
                 auto struct_t = world().struct_type(type->name(), 2);
                 struct_t->set_op(0, tag_type);
                 struct_t->set_op(1, payload_type);
-                return convert(struct_t);
+                converted = convert(struct_t);
+                converted.variant.payload_t = std::make_optional(payload_type);
             } else {
                 // We keep this useless level of struct so the rest of the code doesn't need a special path to extract the tag
                 auto struct_t = world().struct_type(type->name(), 1);
                 struct_t->set_op(0, tag_type);
-                return convert(struct_t);
+                converted = convert(struct_t);
             }
-            THORIN_UNREACHABLE;
+            break;
         }
 
         case Node_MemType: {

From feae51ad3b84615194af711c9812b73b8c76d71e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 15:14:40 +0100
Subject: [PATCH 320/342] spirv: fix emission of external continuations

---
 src/thorin/be/spirv/spirv.cpp         | 11 +++++++-
 src/thorin/be/spirv/spirv_builder.hpp | 37 ++++++++++++++-------------
 2 files changed, 29 insertions(+), 19 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 6066df824..6f29fcf61 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -120,7 +120,11 @@ void CodeGen::emit_stream(std::ostream& out) {
     }
 
     ScopesForest forest(world());
-    forest.for_each([&](const Scope& scope) { emit_scope(scope, forest); });
+    forest.for_each<false>([&](const Scope& scope) {
+        if (scope.entry()->is_intrinsic())
+            return;
+        emit_scope(scope, forest);
+    });
 
     for (auto def : world().defs()) {
         if (auto global = def->isa<Global>())
@@ -204,6 +208,8 @@ static bool is_return_block(thorin::Continuation* cont) {
 }
 
 void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
+    if (!cont->has_body())
+        return;
     auto& bb = *fn->bbs.emplace_back(std::make_unique<BasicBlockBuilder>(*fn));
     cont2bb_[cont] = &bb;
     fn->bbs_to_emit.emplace_back(&bb);
@@ -285,6 +291,9 @@ Id CodeGen::emit_as_bb(thorin::Continuation* cont) {
 }
 
 void CodeGen::emit_epilogue(Continuation* continuation) {
+    if (!continuation->has_body())
+        return;
+
     BasicBlockBuilder* bb = cont2bb_[continuation];
 
     // Handles the potential nuances of jumping to another continuation
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 0a2eed189..954b8d3a1 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -663,47 +663,48 @@ struct FnBuilder {
     }
 };
 
-inline Id FileBuilder::define_function(FnBuilder &fn_builder) {
-    fn_defs.begin_op(spv::Op::OpFunction, 5);
-    fn_defs.ref_id(fn_builder.fn_ret_type);
-    fn_defs.ref_id(fn_builder.function_id);
-    fn_defs.data_.push_back(spv::FunctionControlMaskNone);
-    fn_defs.ref_id(fn_builder.fn_type);
+inline Id FileBuilder::define_function(FnBuilder &fn_builder, bool definition) {
+    auto& tgt = definition ? fn_defs : fn_decls;
+    tgt.begin_op(spv::Op::OpFunction, 5);
+    tgt.ref_id(fn_builder.fn_ret_type);
+    tgt.ref_id(fn_builder.function_id);
+    tgt.data_.push_back(spv::FunctionControlMaskNone);
+    tgt.ref_id(fn_builder.fn_type);
 
     // Includes stuff like OpFunctionParameters
     for (auto w : fn_builder.header.data_)
-        fn_defs.data_.push_back(w);
+        tgt.data_.push_back(w);
 
     bool first = true;
     for (auto& bb : fn_builder.bbs_to_emit) {
-        fn_defs.begin_op(spv::Op::OpLabel, 2);
-        fn_defs.ref_id(bb->label);
+        tgt.begin_op(spv::Op::OpLabel, 2);
+        tgt.ref_id(bb->label);
 
         if (first) {
             for (auto w : fn_builder.variables.data_)
-                fn_defs.data_.push_back(w);
+                tgt.data_.push_back(w);
             first = false;
         }
 
         for (auto& phi : bb->phis) {
-            fn_defs.begin_op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
-            fn_defs.ref_id(phi->type);
-            fn_defs.ref_id(phi->value);
+            tgt.begin_op(spv::Op::OpPhi, 3 + 2 * phi->preds.size());
+            tgt.ref_id(phi->type);
+            tgt.ref_id(phi->value);
             assert(!phi->preds.empty());
             for (auto& [pred_value, pred_label] : phi->preds) {
-                fn_defs.ref_id(pred_value);
-                fn_defs.ref_id(pred_label);
+                tgt.ref_id(pred_value);
+                tgt.ref_id(pred_label);
             }
         }
 
         for (auto w : bb->data_)
-            fn_defs.data_.push_back(w);
+            tgt.data_.push_back(w);
 
         for (auto w : bb->terminator.data_)
-            fn_defs.data_.push_back(w);
+            tgt.data_.push_back(w);
     }
 
-    fn_defs.begin_op(spv::Op::OpFunctionEnd, 1);
+    tgt.begin_op(spv::Op::OpFunctionEnd, 1);
     return fn_builder.function_id;
 }
 

From 4e7393c23e0715720ffb9961dde4ffe8137b3a6a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 15:16:11 +0100
Subject: [PATCH 321/342] spirv: refactor dom/codom code

---
 src/thorin/be/spirv/spirv.cpp       | 35 +++++++--------
 src/thorin/be/spirv/spirv.h         |  5 ++-
 src/thorin/be/spirv/spirv_types.cpp | 66 ++++++++++++++++-------------
 3 files changed, 55 insertions(+), 51 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 6f29fcf61..c6fedabdf 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -173,7 +173,7 @@ FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
 
     auto& fn = *(builder_->fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
     fn.fn_type = convert(entry_->type()).id;
-    fn.fn_ret_type = get_codom_type(entry_);
+    fn.fn_ret_type = get_codom_type(entry_->type());
     defs_[continuation] = fn.function_id;
     return fn;
 }
@@ -259,30 +259,25 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
 }
 
 void CodeGen::finalize(thorin::Continuation* cont) {
+    if (!cont->has_body())
+        return;
     auto& bb = *cont2bb_[cont];
     for (auto& [param, phi] : bb.phis_map) {
         bb.phis.emplace_back(&phi);
     }
 }
 
-void CodeGen::finalize(const thorin::Scope&) {
-    builder_->define_function(*builder_->current_fn_);
-}
-
-Id CodeGen::get_codom_type(const Continuation* fn) {
-    auto ret_cont_type = fn->ret_param()->type()->as<FnType>();
-    std::vector<Id> types;
-    for (auto& op : ret_cont_type->types()) {
-        if (op->isa<MemType>() || is_type_unit(op->type()))
-            continue;
-        assert(op->order() == 0);
-        types.push_back(convert(op).id);
+void CodeGen::finalize(const thorin::Scope& scope) {
+    builder_->define_function(*builder_->current_fn_, scope.entry()->has_body());
+    assert(!scope.entry()->is_intrinsic());
+    if (scope.entry()->is_external()) {
+        if (!scope.entry()->has_body()) {
+            auto v = builder::make_literal_string(scope.entry()->name());
+            v.push_back(spv::LinkageType::LinkageTypeImport);
+            builder_->decorate(builder_->current_fn_->function_id, spv::Decoration::DecorationLinkageAttributes, v);
+            builder_->capability(spv::CapabilityLinkage);
+        }
     }
-    if (types.empty())
-        return convert(world().unit_type()).id;
-    if (types.size() == 1)
-        return types[0];
-    return builder_->declare_struct_type(types);
 }
 
 Id CodeGen::emit_as_bb(thorin::Continuation* cont) {
@@ -336,7 +331,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         }
 
         switch (values.size()) {
-            case 0:  bb->terminator.return_void();      break;
+            case 0:  bb->terminator.return_void(); break;
             case 1:  bb->terminator.return_value(values[0]); break;
             default: bb->terminator.return_value(emit_composite(bb, builder_->current_fn_->fn_ret_type, values));
         }
@@ -405,7 +400,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
 
         Id call_result;
         if (auto called_continuation = app.callee()->isa_nom<Continuation>()) {
-            auto ret_type = get_codom_type(called_continuation);
+            auto ret_type = get_codom_type(called_continuation->type());
             call_result = bb->call(ret_type, emit(called_continuation), call_args);
         } else {
             // must be a closure
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index a1fe867cc..f28be9705 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -58,7 +58,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     }
 
     uint32_t convert(AddrSpace);
-    ConvertedType convert(const Type*);
+    ConvertedType convert(const Type*, bool allow_void = false);
 
     Id emit_fun_decl(Continuation*);
 
@@ -81,7 +81,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     Id emit_composite(BasicBlockBuilder* bb, Id, Defs);
     Id emit_composite(BasicBlockBuilder* bb, Id, ArrayRef<Id>);
 
-    Id get_codom_type(const Continuation* fn);
+    std::tuple<std::vector<Id>, Id> get_dom_codom(const FnType* fn);
+    Id get_codom_type(const FnType*);
 
     Target target_info_;
     FileBuilder* builder_;
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index b4dfb3796..9fe6f3eb0 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -41,7 +41,38 @@ uint32_t CodeGen::convert(AddrSpace as) {
     return storage_class;
 }
 
-ConvertedType CodeGen::convert(const Type* type) {
+Id CodeGen::get_codom_type(const FnType* fn) {
+    auto [dom, codom] = get_dom_codom(fn);
+    assert(codom);
+    return codom;
+}
+
+std::tuple<std::vector<Id>, Id> CodeGen::get_dom_codom(const FnType* fn) {
+    Id ret = 0;
+    std::vector<Id> ops;
+    for (auto op : fn->types()) {
+        if (op->isa<MemType>() || op == world().unit_type())
+            continue;
+        auto fn_type = op->isa<FnType>();
+        if (fn_type && !op->isa<ClosureType>()) {
+            assert(!ret && "only one 'return' supported");
+            std::vector<const Type*> ret_types;
+            for (auto fn_op : fn_type->types()) {
+                if (fn_op->isa<MemType>() || fn_op == world().unit_type())
+                    continue;
+                ret_types.push_back(fn_op);
+            }
+            if (ret_types.size() == 1)
+                ret = convert(ret_types.back()).id;
+            else
+                ret = convert(world().tuple_type(ret_types), true).id;
+        } else
+            ops.push_back(convert(op).id);
+    }
+    return std::make_tuple(ops, ret);
+}
+
+ConvertedType CodeGen::convert(const Type* type, bool allow_void) {
     // Spir-V requires each primitive type to be "unique", it doesn't allow for example two 32-bit signed integer types.
     // Therefore we must enforce that precise/quick types map to the same thing.
     switch (type->tag()) {
@@ -170,32 +201,10 @@ ConvertedType CodeGen::convert(const Type* type) {
 
         case Node_ClosureType:
         case Node_FnType: {
-            // extract "return" type, collect all other types
-            auto fn = type->as<FnType>();
-            Id ret = 0;
-            std::vector<Id> ops;
-            for (auto op : fn->types()) {
-                if (op->isa<MemType>() || op == world().unit_type())
-                    continue;
-                auto fn_type = op->isa<FnType>();
-                if (fn_type && !op->isa<ClosureType>()) {
-                    assert(!ret && "only one 'return' supported");
-                    std::vector<Id> ret_types;
-                    for (auto fn_op : fn_type->types()) {
-                        if (fn_op->isa<MemType>() || fn_op == world().unit_type())
-                            continue;
-                        ret_types.push_back(convert(fn_op).id);
-                    }
-                    if (ret_types.empty())          ret = convert(world().tuple_type({})).id;
-                    else if (ret_types.size() == 1) ret = ret_types.back();
-                    else                            assert(false && "Didn't we refactor this out yet by making functions single-argument ?");
-                } else
-                    ops.push_back(convert(op).id);
-            }
-            assert(ret);
+            auto [dom, codom] = get_dom_codom(type->as<FnType>());
 
             if (type->tag() == Node_FnType) {
-                converted.id = builder_->declare_fn_type(ops, ret);
+                converted.id = builder_->declare_fn_type(dom, codom);
             } else {
                 assert(false && "TODO: handle closure mess");
                 THORIN_UNREACHABLE;
@@ -217,10 +226,10 @@ ConvertedType CodeGen::convert(const Type* type) {
                 converted.layout->alignment = std::max(converted.layout->alignment, converted_member_type.layout->alignment);
                 converted.layout->size = pad(converted.layout->size + converted_member_type.layout->size, converted.layout->alignment);
             }
-            if (converted.layout->size == 0) {
+            if (allow_void && converted.layout->size == 0) {
                 converted.id = builder_->declare_void_type();
                 converted.layout = std::nullopt;
-                break;
+                return converted;
             }
 
             converted.id = builder_->declare_struct_type(spv_types);
@@ -237,8 +246,7 @@ ConvertedType CodeGen::convert(const Type* type) {
                 auto member_type = member->as<Type>();
                 if (member_type == world().unit_type() || member_type == world().mem_type()) continue;
                 auto converted_member_type = convert(member_type);
-                if(!converted_member_type.layout)
-                    continue;
+                assert(converted_member_type.layout);
                 if (converted_member_type.layout->size > max_serialized_size)
                     max_serialized_size = converted_member_type.layout->size;
             }

From 085216677bd4681184a8c1e700da6a1308a1a5bc Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 15:16:30 +0100
Subject: [PATCH 322/342] spirv: fix broken store() when emitting variants

---
 src/thorin/be/spirv/spirv.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c6fedabdf..31e1a20b6 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -639,7 +639,7 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
 
             auto cast_type = convert(world().ptr_type(variant->value()->type(), 1, AddrSpace::Function));
             auto scratch_cast = bb->op_with_result(spv::Op::OpBitcast, cast_type.id, { scratch });
-            bb->store(scratch_cast, emit(variant->value()));
+            bb->store(emit(variant->value()), scratch_cast);
 
             auto payload = bb->load(convert(payload_t.value()).id, scratch);
 

From fdb006d6ae9fe3ac4cde9f586e04349a0d388966 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 15:16:47 +0100
Subject: [PATCH 323/342] spirv_builder: add literal_string infra

---
 src/thorin/be/spirv/spirv_builder.hpp | 31 +++++++++++++++++++++------
 1 file changed, 24 insertions(+), 7 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 954b8d3a1..4d56997ea 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -30,6 +30,23 @@ inline int div_roundup(int a, int b) {
         return (a / b) + 1;
 }
 
+static inline std::vector<uint32_t> make_literal_string(std::string_view str) {
+    std::vector<uint32_t> v;
+    int i = 0;
+    uint32_t cword = 0;
+    for (char c : str) {
+        cword = cword | (c & 0xFF) << (i * 8);
+        i++;
+        if (i == 4) {
+            v.push_back(cword);
+            cword = 0;
+            i = 0;
+        }
+    }
+    v.push_back(cword);
+    return v;
+}
+
 struct SectionBuilder {
     std::vector<uint32_t> data_;
 
@@ -49,7 +66,7 @@ struct SectionBuilder {
         output_word(id);
     }
 
-    void literal_name(std::string_view str) {
+    void literal_string(std::string_view str) {
         int i = 0;
         uint32_t cword = 0;
         for (char c : str) {
@@ -108,7 +125,7 @@ struct FileBuilder {
         assert(id < bound);
         debug_names.begin_op(spv::Op::OpName, 2 + div_roundup(str.size() + 1, 4));
         debug_names.ref_id(id);
-        debug_names.literal_name(str);
+        debug_names.literal_string(str);
     }
 
     Id declare_bool_type() {
@@ -214,7 +231,7 @@ struct FileBuilder {
         debug_string_source.begin_op(spv::Op::OpString, 2 + div_roundup(string.size() + 1, 4));
         auto id = generate_fresh_id();
         debug_string_source.ref_id(id);
-        debug_string_source.literal_name(string);
+        debug_string_source.literal_string(string);
         return id;
     }
 
@@ -274,13 +291,13 @@ struct FileBuilder {
         return id;
     }
 
-    Id define_function(FnBuilder& fn_builder);
+    Id define_function(FnBuilder& fn_builder, bool define);
 
     void declare_entry_point(spv::ExecutionModel execution_model, Id entry_point, std::string name, std::vector<Id> interface) {
         entry_points.begin_op(spv::Op::OpEntryPoint, 3 + div_roundup(name.size() + 1, 4) + interface.size());
         entry_points.literal_int(execution_model);
         entry_points.ref_id(entry_point);
-        entry_points.literal_name(name);
+        entry_points.literal_string(name);
         for (auto i : interface)
             entry_points.ref_id(i);
     }
@@ -307,7 +324,7 @@ struct FileBuilder {
         if (found != extensions_set.end())
             return;
         extensions.begin_op(spv::Op::OpExtension, 1 + div_roundup(name.size() + 1, 4));
-        extensions.literal_name(name);
+        extensions.literal_string(name);
         extensions_set.insert(name);
     }
 
@@ -324,7 +341,7 @@ struct FileBuilder {
         ext_inst_import.begin_op(spv::Op::OpExtInstImport, 2 + div_roundup(name.size() + 1, 4));
         auto id = generate_fresh_id();
         ext_inst_import.ref_id(id);
-        ext_inst_import.literal_name(name);
+        ext_inst_import.literal_string(name);
         extended_instruction_sets[name] = id;
         return id;
     }

From c7469638457a07db79f6a0a68c6d75b286a24be3 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 15:17:06 +0100
Subject: [PATCH 324/342] spirv: fix reserve_shared

---
 src/thorin/be/spirv/spirv_instructions.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 89a37e20d..ee946e1ea 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -110,7 +110,7 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
         else {
             auto in_bytes = size->value().get_u64();
             auto type = world().definite_array_type(world().type_pu8(), in_bytes);
-            Id id = builder_->variable(convert(type).id, spv::StorageClass::StorageClassCrossWorkgroup);
+            Id id = builder_->variable(convert(world().ptr_type(type, 1, AddrSpace::Shared)).id, static_cast<spv::StorageClass>(convert(AddrSpace::Shared)));
             id = bb->convert(spv::Op::OpBitcast, convert(get_produced_type()).id, id);
             return { id };
         }

From dbbd974ed7893906ddcc1aa2da46bbf598403fd0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 17:38:22 +0100
Subject: [PATCH 325/342] spirv: do not emit device functions

---
 src/thorin/be/spirv/spirv.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 31e1a20b6..97a5a1707 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -121,7 +121,7 @@ void CodeGen::emit_stream(std::ostream& out) {
 
     ScopesForest forest(world());
     forest.for_each<false>([&](const Scope& scope) {
-        if (scope.entry()->is_intrinsic())
+        if (scope.entry()->is_intrinsic() || scope.entry()->cc() == CC::Device)
             return;
         emit_scope(scope, forest);
     });
@@ -269,7 +269,7 @@ void CodeGen::finalize(thorin::Continuation* cont) {
 
 void CodeGen::finalize(const thorin::Scope& scope) {
     builder_->define_function(*builder_->current_fn_, scope.entry()->has_body());
-    assert(!scope.entry()->is_intrinsic());
+    assert(!(scope.entry()->is_intrinsic() || scope.entry()->cc() == CC::Device));
     if (scope.entry()->is_external()) {
         if (!scope.entry()->has_body()) {
             auto v = builder::make_literal_string(scope.entry()->name());

From adfd62d67cbb9397f701d70e0bd97cadaaaf7d35 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 17:38:36 +0100
Subject: [PATCH 326/342] spirv: use emit_composite for variants

---
 src/thorin/be/spirv/spirv.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 97a5a1707..4533ee96c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -644,11 +644,11 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             auto payload = bb->load(convert(payload_t.value()).id, scratch);
 
             std::vector<Id> with_tag = {tag, payload};
-            return bb->composite(converted_type.id, with_tag);
+            return emit_composite(bb, converted_type.id, with_tag);
         } else {
             // Zero-sized payload case
             std::vector<Id> with_tag = { tag };
-            return bb->composite(converted_type.id, with_tag);
+            return emit_composite(bb, converted_type.id, with_tag);
         }
     } else if (auto vextract = def->isa<VariantExtract>()) {
         auto variant_type = vextract->value()->type()->as<VariantType>();

From 3638b5953dbb7c70e59895ca13b801a78c9703b6 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 17:38:47 +0100
Subject: [PATCH 327/342] spirv: fix some casting behaviours

---
 src/thorin/be/spirv/spirv.cpp | 68 +++++++++++++----------------------
 1 file changed, 24 insertions(+), 44 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 4533ee96c..ec497cb82 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -804,10 +804,18 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         } else if (auto cast = def->isa<Cast>()) {
             if (auto src_ptr_type = src_type->isa<PtrType>()) {
                 if (auto dst_ptr_type = dst_type->isa<PtrType>()) {
-                    if (src_ptr_type->addr_space() == AddrSpace::Generic) {
-                        return bb->op_with_result(spv::Op::OpGenericCastToPtr, convert(cast->type()).id, { emit(cast->from()) });
-                    } else if (dst_ptr_type->addr_space() == AddrSpace::Generic) {
-                        return bb->op_with_result(spv::Op::OpPtrCastToGeneric, convert(cast->type()).id, { emit(cast->from()) });
+                    if (src_ptr_type->addr_space() == AddrSpace::Generic && dst_ptr_type->addr_space() != AddrSpace::Generic) {
+                        auto casted = emit(cast->from());
+                        if (src_ptr_type->pointee() != dst_ptr_type->pointee())
+                            casted = bb->convert(spv::OpBitcast, convert(world().ptr_type(dst_ptr_type->pointee(), 1, src_ptr_type->addr_space())).id, casted);
+                        casted = bb->op_with_result(spv::Op::OpGenericCastToPtr, convert(cast->type()).id, { casted });
+                        return casted;
+                    } else if (src_ptr_type->addr_space() != AddrSpace::Generic && dst_ptr_type->addr_space() == AddrSpace::Generic) {
+                        auto casted = emit(cast->from());
+                        if (src_ptr_type->pointee() != dst_ptr_type->pointee())
+                            casted = bb->convert(spv::OpBitcast, convert(world().ptr_type(dst_ptr_type->pointee(), 1, src_ptr_type->addr_space())).id, casted);
+                        casted = bb->op_with_result(spv::Op::OpPtrCastToGeneric, convert(cast->type()).id, { casted });
+                        return casted;
                     } else {
                         world().WLOG("Abnormal ptr-ptr cast: {} to {}, should be a bitcast", src_ptr_type, dst_ptr_type);
                         return bb->convert(spv::OpBitcast, convert(cast->type()).id, emit(cast->from()));
@@ -840,50 +848,22 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
                         switch (dst_kind) {
                             case PrimTypeKind::Signed:   data = bb->convert(spv::OpConvertFToS, target_type, data); break;
                             case PrimTypeKind::Unsigned: data = bb->convert(spv::OpConvertFToU, target_type, data); break;
-                            default: THORIN_UNREACHABLE;
+                            case PrimTypeKind::Float:    data = bb->convert(spv::OpFConvert, target_type, data); break;
                         }
                         break;
                 }
+            } else if (src_bitwidth == dst_bitwidth) {
+                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length)).id;
+                data = bb->convert(spv::OpBitcast, target_type, data);
+            }
+            else if (src_kind == PrimTypeKind::Unsigned) {
+                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length)).id;
+                data = bb->convert(spv::OpUConvert, target_type, data);
+            } else if (src_kind == PrimTypeKind::Signed) {
+                auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length)).id;
+                data = bb->convert(spv::OpSConvert, target_type, data);
             } else {
-                // we expand first and shrink last to minimize precision losses, with bitcast in the middle
-                bool needs_chopping = src_bitwidth > dst_bitwidth;
-                bool needs_expanding = src_bitwidth < dst_bitwidth;
-
-                if (needs_expanding) {
-                    auto target_type = convert(get_primtype(world(), src_kind, src_bitwidth, length)).id;
-                    switch (src_kind) {
-                        case PrimTypeKind::Signed:
-                            data = bb->convert(spv::OpSConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Unsigned:
-                            data = bb->convert(spv::OpUConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Float:
-                            data = bb->convert(spv::OpFConvert, target_type, data);
-                            break;
-                        default: assert(false);
-                    }
-                }
-
-                auto expanded_bitwidth = needs_expanding ? dst_bitwidth : src_bitwidth;
-                auto bitcast_target_type = convert(get_primtype(world(), dst_kind, expanded_bitwidth, length)).id;
-                data = bb->convert(spv::OpBitcast, bitcast_target_type, data);
-
-                if (needs_chopping) {
-                    auto target_type = convert(get_primtype(world(), dst_kind, dst_bitwidth, length)).id;
-                    switch (dst_kind) {
-                        case PrimTypeKind::Signed:
-                            data = bb->convert(spv::OpSConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Unsigned:
-                            data = bb->convert(spv::OpUConvert, target_type, data);
-                            break;
-                        case PrimTypeKind::Float:
-                            data = bb->convert(spv::OpFConvert, target_type, data);
-                            break;
-                        default: assert(false);
-                    }
-                }
+                assert(false && "unsupported cast");
             }
 
             return data;

From 1bf66921df05a5b2aba3e467ea9f05429b7f522e Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 12 Nov 2024 17:38:58 +0100
Subject: [PATCH 328/342] spirv: add rv_all and atomic_add for fp types

---
 src/thorin/be/spirv/spirv_instructions.cpp | 21 ++++++++++++++++++++-
 1 file changed, 20 insertions(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index ee946e1ea..5c901827b 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -137,7 +137,26 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
     } else if (intrinsic->name() == "atomic_add") {
         auto args = emit_args(app.args().skip_back());
         auto [ptr, value] = *(std::array<Id, 2>*)args.data();
-        auto result = bb->op_with_result(spv::Op::OpAtomicIAdd, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeInvocation), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask), value });
+        auto produced = get_produced_type();
+        spv::Op op;
+        if (is_type_f(produced)) {
+            op = spv::OpAtomicFAddEXT;
+            auto ct = convert(produced);
+            switch (ct.layout->size) {
+                case 2: builder_->capability(spv::Capability::CapabilityAtomicFloat16AddEXT); break;
+                case 4: builder_->capability(spv::Capability::CapabilityAtomicFloat32AddEXT); break;
+                case 8: builder_->capability(spv::Capability::CapabilityAtomicFloat64AddEXT); break;
+            }
+            builder_->extension("SPV_EXT_shader_atomic_float_add");
+        } else if (is_type_i(produced))
+            op = spv::OpAtomicIAdd;
+        else
+            assert(false && "unknown primitive type for atomic_add");
+        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeInvocation), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask), value });
+        return { result };
+    } else if (intrinsic->name() == "rv_all") {
+        auto args = emit_args(app.args().skip_back());
+        auto result = bb->op_with_result(spv::Op::OpGroupAll, convert(get_produced_type()).id,  { literal(spv::Scope::ScopeInvocation), emit(app.arg(1)) });
         return { result };
     }
     world().ELOG("thorin/spirv: Intrinsic '{}' isn't recognised", intrinsic->name());

From 51cc4a34c024af9cd74a187356adb054b7881d1a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 13 Nov 2024 15:48:47 +0100
Subject: [PATCH 329/342] spirv: hack: avoid emitting plain structs in kernel
 params

---
 src/thorin/be/spirv/spirv.cpp | 49 ++++++++++++++++++++++++++++++++---
 1 file changed, 45 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index ec497cb82..0679327ca 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -166,14 +166,40 @@ Id CodeGen::emit_fun_decl(thorin::Continuation* continuation) {
     return get_fn_builder(continuation).function_id;
 }
 
+static bool prefers_passing_by_ptr(const Type* t) {
+    return t->isa<VariantType>() || t->isa<TupleType>() || t->isa<StructType>();
+}
+
+static const FnType* patch_entry_point_signature(const FnType* type) {
+    assert(type->tag() == Node_FnType && "only plain functions expected");
+    auto& world = type->world();
+    std::vector<const Type*> new_ops;
+    for (auto t : type->types()) {
+        if (prefers_passing_by_ptr(t))
+            t = world.ptr_type(t, 1, AddrSpace::Global);
+        else if (auto fn = t->isa<FnType>())
+            t = patch_entry_point_signature(fn);
+        else if (auto ptr = t->isa<PtrType>()) {
+            if (ptr->addr_space() == AddrSpace::Generic)
+                t = world.ptr_type(ptr->pointee(), 1, AddrSpace::Global);
+        }
+        new_ops.push_back(t);
+    }
+    return world.fn_type(new_ops);
+}
+
 FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
     if (auto found = builder_->fn_builders_.find(continuation); found != builder_->fn_builders_.end()) {
         return *found->second;
     }
 
     auto& fn = *(builder_->fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
-    fn.fn_type = convert(entry_->type()).id;
-    fn.fn_ret_type = get_codom_type(entry_->type());
+    auto fn_type = entry_->type();
+    if (kernel_config_->contains(continuation)) {
+        fn_type = patch_entry_point_signature(fn_type);
+    }
+    fn.fn_type = convert(fn_type).id;
+    fn.fn_ret_type = get_codom_type(fn_type);
     defs_[continuation] = fn.function_id;
     return fn;
 }
@@ -221,13 +247,28 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
         world().ddef(cont, "Emitting {} as return block", cont);
 
     if (entry_ == cont) {
+        bool entry_point = kernel_config_->contains(entry_);
         for (auto param : cont->params()) {
             if (is_mem(param) || is_unit(param)) {
                 // Nothing
                 defs_[param] = 0;
             } else if (param->order() == 0) {
-                auto param_t = convert(param->type());
-                auto id = fn->parameter(param_t.id);
+                Id id;
+                if (entry_point && prefers_passing_by_ptr(param->type())) {
+                    auto param_t = convert(world().ptr_type(param->type(), 1, AddrSpace::Global));
+                    id = fn->parameter(param_t.id);
+                    builder_->decorate(id, spv::DecorationFuncParamAttr, { spv::FunctionParameterAttribute::FunctionParameterAttributeByVal });
+                    builder_->decorate(id, spv::DecorationFuncParamAttr, { spv::FunctionParameterAttribute::FunctionParameterAttributeNoCapture });
+                    builder_->decorate(id, spv::DecorationFuncParamAttr, { spv::FunctionParameterAttribute::FunctionParameterAttributeNoWrite });
+                    id = bb.load(convert(param->type()).id, id);
+                } else {
+                    auto param_t = param->type();
+                    if (auto ptr = param_t->isa<PtrType>()) {
+                        if (ptr->addr_space() == AddrSpace::Generic)
+                            param_t = world().ptr_type(ptr->pointee(), 1, AddrSpace::Global);
+                    }
+                    id = fn->parameter(convert(param_t).id);
+                }
                 defs_[param] = id;
                 fn->params[param] = id;
                 if (param->type()->isa<PtrType>()) {

From 6621979ce401c8360980974ae66cc05977507c48 Mon Sep 17 00:00:00 2001
From: Richard Membarth <richard.membarth@uni-saarland.de>
Date: Fri, 15 Nov 2024 11:17:44 +0100
Subject: [PATCH 330/342] Match LLVM's required CMake version.

---
 CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c31f7152a..40fc8be1d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.13.4 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.20.0 FATAL_ERROR)
 
 project(Thorin)
 

From 537c0fd3b69891f9bf9e9b553d7dfb237cc97a29 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Nov 2024 13:25:47 +0100
Subject: [PATCH 331/342] spirv: improve how bitcasts are handled

---
 src/thorin/be/spirv/spirv.cpp | 55 ++++++++++++++++++++---------------
 src/thorin/be/spirv/spirv.h   |  1 +
 2 files changed, 32 insertions(+), 24 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 0679327ca..c4517245c 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -263,17 +263,22 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
                     id = bb.load(convert(param->type()).id, id);
                 } else {
                     auto param_t = param->type();
+                    bool needs_cast_back_to_generic = false;
                     if (auto ptr = param_t->isa<PtrType>()) {
-                        if (ptr->addr_space() == AddrSpace::Generic)
+                        if (ptr->addr_space() == AddrSpace::Generic) {
                             param_t = world().ptr_type(ptr->pointee(), 1, AddrSpace::Global);
+                            needs_cast_back_to_generic = true;
+                        }
                     }
                     id = fn->parameter(convert(param_t).id);
+                    if (param->type()->isa<PtrType>()) {
+                        builder_->decorate(id, spv::DecorationAliased);
+                    }
+                    if (needs_cast_back_to_generic)
+                        id = emit_ptr_bitcast(&bb, param_t->as<PtrType>(), param->type()->as<PtrType>(), id);
                 }
                 defs_[param] = id;
                 fn->params[param] = id;
-                if (param->type()->isa<PtrType>()) {
-                    builder_->decorate(id, spv::DecorationAliased);
-                }
             }
         }
     } else {
@@ -562,6 +567,22 @@ Id CodeGen::emit_composite(BasicBlockBuilder* bb, Id t, ArrayRef<Id> ids) {
     }
 }
 
+Id CodeGen::emit_ptr_bitcast(BasicBlockBuilder* bb, const thorin::PtrType* from, const thorin::PtrType* to, Id id) {
+    if (from->pointee() != to->pointee()) {
+        auto ptr_t = world().ptr_type(to->pointee(), 1, from->addr_space());
+        id = bb->convert(spv::OpBitcast, convert(ptr_t).id, id);
+        from = ptr_t;
+    }
+
+    if (from->addr_space() == AddrSpace::Generic && to->addr_space() != AddrSpace::Generic) {
+        id = bb->op_with_result(spv::Op::OpGenericCastToPtr, convert(to).id, { id });
+    } else if (from->addr_space() != AddrSpace::Generic && to->addr_space() == AddrSpace::Generic) {
+        id = bb->op_with_result(spv::Op::OpPtrCastToGeneric, convert(to).id, { id });
+    }
+
+    return id;
+}
+
 Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
     if (auto mathop = def->isa<MathOp>())
         return emit_mathop(bb, *mathop);
@@ -836,6 +857,12 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto conv_src_type = convert(src_type);
         auto conv_dst_type = convert(dst_type);
 
+        if (auto src_ptr_type = src_type->isa<PtrType>()) {
+            if (auto dst_ptr_type = dst_type->isa<PtrType>()) {
+                return emit_ptr_bitcast(bb, src_ptr_type, dst_ptr_type, emit(conv->from()));
+            }
+        }
+
         if (auto bitcast = def->isa<Bitcast>()) {
             assert(conv_src_type.layout && conv_dst_type.layout);
             if (conv_src_type.layout->size != conv_dst_type.layout->size)
@@ -843,26 +870,6 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
 
             return bb->convert(spv::OpBitcast, convert(bitcast->type()).id, emit(bitcast->from()));
         } else if (auto cast = def->isa<Cast>()) {
-            if (auto src_ptr_type = src_type->isa<PtrType>()) {
-                if (auto dst_ptr_type = dst_type->isa<PtrType>()) {
-                    if (src_ptr_type->addr_space() == AddrSpace::Generic && dst_ptr_type->addr_space() != AddrSpace::Generic) {
-                        auto casted = emit(cast->from());
-                        if (src_ptr_type->pointee() != dst_ptr_type->pointee())
-                            casted = bb->convert(spv::OpBitcast, convert(world().ptr_type(dst_ptr_type->pointee(), 1, src_ptr_type->addr_space())).id, casted);
-                        casted = bb->op_with_result(spv::Op::OpGenericCastToPtr, convert(cast->type()).id, { casted });
-                        return casted;
-                    } else if (src_ptr_type->addr_space() != AddrSpace::Generic && dst_ptr_type->addr_space() == AddrSpace::Generic) {
-                        auto casted = emit(cast->from());
-                        if (src_ptr_type->pointee() != dst_ptr_type->pointee())
-                            casted = bb->convert(spv::OpBitcast, convert(world().ptr_type(dst_ptr_type->pointee(), 1, src_ptr_type->addr_space())).id, casted);
-                        casted = bb->op_with_result(spv::Op::OpPtrCastToGeneric, convert(cast->type()).id, { casted });
-                        return casted;
-                    } else {
-                        world().WLOG("Abnormal ptr-ptr cast: {} to {}, should be a bitcast", src_ptr_type, dst_ptr_type);
-                        return bb->convert(spv::OpBitcast, convert(cast->type()).id, emit(cast->from()));
-                    }
-                }
-            }
 
             // NB: all ops used here are scalar/vector agnostic
             auto src_prim = src_type->isa<PrimType>();
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index f28be9705..310613494 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -80,6 +80,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     Id emit_mathop(BasicBlockBuilder* bb, const MathOp& op);
     Id emit_composite(BasicBlockBuilder* bb, Id, Defs);
     Id emit_composite(BasicBlockBuilder* bb, Id, ArrayRef<Id>);
+    Id emit_ptr_bitcast(BasicBlockBuilder* bb, const PtrType* from, const PtrType* to, Id);
 
     std::tuple<std::vector<Id>, Id> get_dom_codom(const FnType* fn);
     Id get_codom_type(const FnType*);

From 5abfbc7042de75ed1d19363c6bd670b568126abc Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Nov 2024 13:26:30 +0100
Subject: [PATCH 332/342] spv: place structs in Function arguments

---
 src/thorin/be/spirv/spirv.cpp | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index c4517245c..b66567a7b 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -166,7 +166,7 @@ Id CodeGen::emit_fun_decl(thorin::Continuation* continuation) {
     return get_fn_builder(continuation).function_id;
 }
 
-static bool prefers_passing_by_ptr(const Type* t) {
+static bool cl_demands_passed_by_reference(const Type* t) {
     return t->isa<VariantType>() || t->isa<TupleType>() || t->isa<StructType>();
 }
 
@@ -175,8 +175,8 @@ static const FnType* patch_entry_point_signature(const FnType* type) {
     auto& world = type->world();
     std::vector<const Type*> new_ops;
     for (auto t : type->types()) {
-        if (prefers_passing_by_ptr(t))
-            t = world.ptr_type(t, 1, AddrSpace::Global);
+        if (cl_demands_passed_by_reference(t))
+            t = world.ptr_type(t, 1, AddrSpace::Function);
         else if (auto fn = t->isa<FnType>())
             t = patch_entry_point_signature(fn);
         else if (auto ptr = t->isa<PtrType>()) {
@@ -254,8 +254,8 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
                 defs_[param] = 0;
             } else if (param->order() == 0) {
                 Id id;
-                if (entry_point && prefers_passing_by_ptr(param->type())) {
-                    auto param_t = convert(world().ptr_type(param->type(), 1, AddrSpace::Global));
+                if (entry_point && cl_demands_passed_by_reference(param->type())) {
+                    auto param_t = convert(world().ptr_type(param->type(), 1, AddrSpace::Function));
                     id = fn->parameter(param_t.id);
                     builder_->decorate(id, spv::DecorationFuncParamAttr, { spv::FunctionParameterAttribute::FunctionParameterAttributeByVal });
                     builder_->decorate(id, spv::DecorationFuncParamAttr, { spv::FunctionParameterAttribute::FunctionParameterAttributeNoCapture });

From 14c9793de828d7f31261d6eebbe8e845ac3d5858 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Nov 2024 13:58:07 +0100
Subject: [PATCH 333/342] spirv: fix wrong layout for execution_modes

---
 src/thorin/be/spirv/spirv_builder.hpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 4d56997ea..6c3f5b862 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -303,11 +303,11 @@ struct FileBuilder {
     }
 
     void execution_mode(Id entry_point, spv::ExecutionMode execution_mode, std::vector<uint32_t> payloads) {
-        entry_points.begin_op(spv::Op::OpExecutionMode, 3 + payloads.size());
-        entry_points.ref_id(entry_point);
-        entry_points.literal_int(execution_mode);
+        execution_modes.begin_op(spv::Op::OpExecutionMode, 3 + payloads.size());
+        execution_modes.ref_id(entry_point);
+        execution_modes.literal_int(execution_mode);
         for (auto d : payloads)
-            entry_points.literal_int(d);
+            execution_modes.literal_int(d);
     }
 
     void capability(spv::Capability cap) {

From 719570d2f4b1db2b00ba41de044e3f0e201439ef Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Tue, 19 Nov 2024 14:47:14 +0100
Subject: [PATCH 334/342] spirv: better filtering for unit types

---
 src/thorin/be/spirv/spirv.cpp       | 27 +++++++++++++++++--------
 src/thorin/be/spirv/spirv.h         |  4 +++-
 src/thorin/be/spirv/spirv_types.cpp | 31 ++++++++++++++++++-----------
 3 files changed, 41 insertions(+), 21 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index b66567a7b..024152442 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -249,7 +249,7 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
     if (entry_ == cont) {
         bool entry_point = kernel_config_->contains(entry_);
         for (auto param : cont->params()) {
-            if (is_mem(param) || is_unit(param)) {
+            if (!should_emit(param->type())) {
                 // Nothing
                 defs_[param] = 0;
             } else if (param->order() == 0) {
@@ -286,7 +286,7 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
         if (bb.semi_inline)
             return;
         for (auto param : cont->params()) {
-            if (is_mem(param) || is_unit(param)) {
+            if (!should_emit(param->type())) {
                 // Nothing
                 defs_[param] = 0;
             } else {
@@ -344,7 +344,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
 
         for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
             auto param = succ->param(i);
-            if (is_mem(param) || is_unit(param)) {
+            if (!should_emit(param->type())) {
                 if (dstbb->semi_inline)
                     defs_[param] = 0;
                 continue;
@@ -368,7 +368,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
 
         for (auto arg : app.args()) {
             assert(arg->order() == 0);
-            if (is_mem(arg) || is_unit(arg)) {
+            if (!should_emit(arg->type())) {
                 emit_unsafe(arg);
                 continue;
             }
@@ -385,7 +385,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         int index = -1;
         for (auto& arg : app.args()) {
             index++;
-            if (is_mem(arg) || is_unit(arg)) {
+            if (!should_emit(arg->type())) {
                 emit_unsafe(arg);
                 continue;
             }
@@ -463,7 +463,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         size_t real_params_count = 0;
         const Param* last_param = nullptr;
         for (auto param : succ->params()) {
-            if (is_mem(param) || is_unit(param))
+            if (!should_emit(param->type()))
                 continue;
             last_param = param;
             real_params_count++;
@@ -476,7 +476,7 @@ void CodeGen::emit_epilogue(Continuation* continuation) {
         } else if (real_params_count > 1) {
             for (size_t i = 0, j = 0; i != succ->num_params(); ++i) {
                 auto param = succ->param(i);
-                if (is_mem(param) || is_unit(param))
+                if (!should_emit(param->type()))
                     continue;
                 args[j] = bb->extract(convert(param->type()).id, call_result, { (uint32_t) j });
                 j++;
@@ -522,11 +522,22 @@ Id CodeGen::emit_constant(const thorin::Def* def) {
     assertf(false, "Incomplete emit(def) definition");
 }
 
+bool CodeGen::should_emit(const thorin::Type* type) {
+    if (type == world().mem_type())
+        return false;
+    if (auto fn_t = type->isa<FnType>())
+        return fn_t->is_returning();
+    auto converted = convert_maybe_void(type);
+    if (converted.id == builder_->declare_void_type())
+        return false;
+    return true;
+}
+
 std::vector<Id> CodeGen::emit_args(Defs defs) {
     std::vector<Id> emitted;
     for (auto arg : defs) {
         auto arg_type = arg->type();
-        if (arg_type == world().unit_type() || arg_type == world().mem_type()) {
+        if (!should_emit(arg_type)) {
             emit_unsafe(arg);
             continue;
         } else {
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 310613494..64452513f 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -58,7 +58,8 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     }
 
     uint32_t convert(AddrSpace);
-    ConvertedType convert(const Type*, bool allow_void = false);
+    ConvertedType convert_maybe_void(const Type*);
+    ConvertedType convert(const Type*);
 
     Id emit_fun_decl(Continuation*);
 
@@ -74,6 +75,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     FnBuilder& get_fn_builder(Continuation*);
     std::vector<Id> emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb);
     std::vector<Id> emit_args(Defs);
+    bool should_emit(const Type*);
     Id literal(uint32_t);
 
     Id emit_as_bb(Continuation*);
diff --git a/src/thorin/be/spirv/spirv_types.cpp b/src/thorin/be/spirv/spirv_types.cpp
index 9fe6f3eb0..c40f94406 100644
--- a/src/thorin/be/spirv/spirv_types.cpp
+++ b/src/thorin/be/spirv/spirv_types.cpp
@@ -51,28 +51,40 @@ std::tuple<std::vector<Id>, Id> CodeGen::get_dom_codom(const FnType* fn) {
     Id ret = 0;
     std::vector<Id> ops;
     for (auto op : fn->types()) {
-        if (op->isa<MemType>() || op == world().unit_type())
-            continue;
         auto fn_type = op->isa<FnType>();
         if (fn_type && !op->isa<ClosureType>()) {
             assert(!ret && "only one 'return' supported");
             std::vector<const Type*> ret_types;
             for (auto fn_op : fn_type->types()) {
-                if (fn_op->isa<MemType>() || fn_op == world().unit_type())
+                if (!should_emit(fn_op))
                     continue;
                 ret_types.push_back(fn_op);
             }
             if (ret_types.size() == 1)
-                ret = convert(ret_types.back()).id;
+                ret = convert_maybe_void(ret_types.back()).id;
             else
-                ret = convert(world().tuple_type(ret_types), true).id;
-        } else
+                ret = convert_maybe_void(world().tuple_type(ret_types)).id;
+        } else if (!should_emit(op))
+            continue;
+        else
             ops.push_back(convert(op).id);
     }
     return std::make_tuple(ops, ret);
 }
 
-ConvertedType CodeGen::convert(const Type* type, bool allow_void) {
+ConvertedType CodeGen::convert_maybe_void(const thorin::Type* type) {
+    auto converted = convert(type);
+
+    if ((type->isa<StructType>() || type->isa<TupleType>()) && converted.layout->size == 0) {
+        converted.id = builder_->declare_void_type();
+        converted.layout = std::nullopt;
+        return converted;
+    }
+
+    return converted;
+}
+
+ConvertedType CodeGen::convert(const Type* type) {
     // Spir-V requires each primitive type to be "unique", it doesn't allow for example two 32-bit signed integer types.
     // Therefore we must enforce that precise/quick types map to the same thing.
     switch (type->tag()) {
@@ -226,11 +238,6 @@ ConvertedType CodeGen::convert(const Type* type, bool allow_void) {
                 converted.layout->alignment = std::max(converted.layout->alignment, converted_member_type.layout->alignment);
                 converted.layout->size = pad(converted.layout->size + converted_member_type.layout->size, converted.layout->alignment);
             }
-            if (allow_void && converted.layout->size == 0) {
-                converted.id = builder_->declare_void_type();
-                converted.layout = std::nullopt;
-                return converted;
-            }
 
             converted.id = builder_->declare_struct_type(spv_types);
             builder_->name(converted.id, type->to_string());

From cf42088cc46a155a79fede3467fb3cc2a2497e47 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 21 Nov 2024 14:22:58 +0100
Subject: [PATCH 335/342] spirv: updated spirv_builder.hpp

---
 src/thorin/be/spirv/spirv.cpp         |  9 ++++
 src/thorin/be/spirv/spirv.h           |  1 +
 src/thorin/be/spirv/spirv_builder.hpp | 77 +++++++++++++++------------
 3 files changed, 52 insertions(+), 35 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 024152442..89664230f 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -324,6 +324,11 @@ void CodeGen::finalize(const thorin::Scope& scope) {
             builder_->capability(spv::CapabilityLinkage);
         }
     }
+
+    for (auto def : scope_local_defs_) {
+        defs_.erase(def);
+    }
+    scope_local_defs_.clear();
 }
 
 Id CodeGen::emit_as_bb(thorin::Continuation* cont) {
@@ -706,6 +711,7 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto converted_type = convert(variant_type);
         auto tag = builder_->u32_constant(variant->index());
 
+        scope_local_defs_.insert(def);
         if (auto payload_t = converted_type.variant.payload_t) {
             auto scratch_type = convert(world().ptr_type(payload_t.value(), 1, AddrSpace::Function));
             auto scratch = bb->fn_builder.variable(scratch_type.id, spv::StorageClassFunction);
@@ -744,8 +750,10 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
         auto value = emit(vindex->op(0));
         return bb->extract(convert(world().type_pu32()).id, value, { 0 });
     } else if (auto tuple = def->isa<Tuple>()) {
+        scope_local_defs_.insert(def);
         return emit_composite(bb, convert(tuple->type()).id, tuple->ops());
     } else if (auto structagg = def->isa<StructAgg>()) {
+        scope_local_defs_.insert(def);
         return emit_composite(bb, convert(structagg->type()).id, structagg->ops());
     } else if (auto access = def->isa<Access>()) {
         // emit dependent operations first
@@ -928,6 +936,7 @@ Id CodeGen::emit_bb(BasicBlockBuilder* bb, const Def* def) {
             return data;
         } else THORIN_UNREACHABLE;
     } else if (def->isa<Bottom>()) {
+        scope_local_defs_.insert(def);
         return bb->undef(convert(def->type()).id);
     } else if (auto select = def->isa<Select>()) {
         return bb->op_with_result(spv::Op::OpSelect, convert(def->type()).id, emit_args(select->ops()));
diff --git a/src/thorin/be/spirv/spirv.h b/src/thorin/be/spirv/spirv.h
index 64452513f..cae51c7c6 100644
--- a/src/thorin/be/spirv/spirv.h
+++ b/src/thorin/be/spirv/spirv.h
@@ -90,6 +90,7 @@ class CodeGen : public thorin::CodeGen, public thorin::Emitter<Id, ConvertedType
     Target target_info_;
     FileBuilder* builder_;
     const Cont2Config* kernel_config_;
+    DefSet scope_local_defs_;
 
     friend Target;
 };
diff --git a/src/thorin/be/spirv/spirv_builder.hpp b/src/thorin/be/spirv/spirv_builder.hpp
index 6c3f5b862..81dc3fdc5 100644
--- a/src/thorin/be/spirv/spirv_builder.hpp
+++ b/src/thorin/be/spirv/spirv_builder.hpp
@@ -48,8 +48,13 @@ static inline std::vector<uint32_t> make_literal_string(std::string_view str) {
 }
 
 struct SectionBuilder {
+    explicit SectionBuilder(FileBuilder& file_builder) : file_builder_(file_builder) {}
+
     std::vector<uint32_t> data_;
+    FileBuilder& file_builder_;
 
+protected:
+    inline Id fresh_id();
 private:
     void output_word(uint32_t word) {
         data_.push_back(word);
@@ -84,6 +89,22 @@ struct SectionBuilder {
     void literal_int(uint32_t i) {
         output_word(i);
     }
+
+    void op(spv::Op op, std::vector<uint32_t> operands) {
+        begin_op(op, operands.size() + 1);
+        for (auto e : operands)
+            literal_int(e);
+    }
+
+    Id op_with_result(spv::Op op, Id type, std::vector<uint32_t> operands) {
+        begin_op(op, operands.size() + 3);
+        ref_id(type);
+        auto id = fresh_id();
+        ref_id(id);
+        for (auto e : operands)
+            literal_int(e);
+        return id;
+    }
 };
 
 struct FileBuilder {
@@ -116,7 +137,7 @@ struct FileBuilder {
         }
     };
 
-    FileBuilder() {}
+    FileBuilder() : capabilities(*this), extensions(*this), ext_inst_import(*this), entry_points(*this), execution_modes(*this), debug_string_source(*this), debug_names(*this), debug_module_processed(*this), annotations(*this), types_constants(*this), fn_decls(*this), fn_defs(*this) {}
     FileBuilder(const FileBuilder&) = delete;
 
     Id generate_fresh_id() { return { bound++ }; }
@@ -385,7 +406,7 @@ struct FileBuilder {
 public:
     void finish(std::ostream& output) {
         output_ = &output;
-        SectionBuilder memory_model_section;
+        SectionBuilder memory_model_section(*this);
         memory_model_section.begin_op(spv::Op::OpMemoryModel, 3);
         memory_model_section.data_.push_back(addressing_model);
         memory_model_section.data_.push_back(memory_model);
@@ -414,11 +435,13 @@ struct FileBuilder {
     friend BasicBlockBuilder;
 };
 
+inline Id SectionBuilder::fresh_id() {
+    return file_builder_.generate_fresh_id();
+}
+
 struct BasicBlockBuilder : public SectionBuilder {
     explicit BasicBlockBuilder(FileBuilder& file_builder)
-            : file_builder(file_builder), terminator(*this) {}
-
-    FileBuilder& file_builder;
+            : SectionBuilder(file_builder), terminator(*this) {}
 
     struct Phi {
         Id type;
@@ -428,22 +451,6 @@ struct BasicBlockBuilder : public SectionBuilder {
     std::vector<Phi*> phis;
     Id label;
 
-    void op(spv::Op op, std::vector<uint32_t> operands) {
-        begin_op(op, operands.size() + 1);
-        for (auto e : operands)
-            literal_int(e);
-    }
-
-    Id op_with_result(spv::Op op, Id type, std::vector<uint32_t> operands) {
-        begin_op(op, operands.size() + 3);
-        ref_id(type);
-        auto id = file_builder.generate_fresh_id();
-        ref_id(id);
-        for (auto e : operands)
-            literal_int(e);
-        return id;
-    }
-
     Id undef(Id type) { return op_with_result(spv::Op::OpUndef, type, {}); }
 
     Id composite(Id aggregate_t, std::vector<Id> elements) { return op_with_result(spv::Op::OpCompositeConstruct, aggregate_t, elements); }
@@ -451,7 +458,7 @@ struct BasicBlockBuilder : public SectionBuilder {
     Id extract(Id target_type, Id composite, std::vector<uint32_t> indices) {
         begin_op(spv::Op::OpCompositeExtract, 4 + indices.size());
         ref_id(target_type);
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(id);
         ref_id(composite);
         for (auto i : indices)
@@ -462,7 +469,7 @@ struct BasicBlockBuilder : public SectionBuilder {
     Id insert(Id target_type, Id object, Id composite, std::vector<uint32_t> indices) {
         begin_op(spv::Op::OpCompositeInsert, 5 + indices.size());
         ref_id(target_type);
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(id);
         ref_id(object);
         ref_id(composite);
@@ -474,7 +481,7 @@ struct BasicBlockBuilder : public SectionBuilder {
     Id vector_extract_dynamic(Id target_type, Id vector, Id index) {
         begin_op(spv::Op::OpVectorExtractDynamic, 5);
         ref_id(target_type);
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(id);
         ref_id(vector);
         ref_id(index);
@@ -484,7 +491,7 @@ struct BasicBlockBuilder : public SectionBuilder {
     Id vector_insert_dynamic(Id target_type, Id vector, Id component, Id index) {
         begin_op(spv::Op::OpVectorInsertDynamic, 6);
         ref_id(target_type);
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(id);
         ref_id(vector);
         ref_id(component);
@@ -495,7 +502,7 @@ struct BasicBlockBuilder : public SectionBuilder {
     // Used for almost all conversion operations
     Id convert(spv::Op op_, Id target_type, Id value) {
         begin_op(op_, 4);
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(value);
@@ -504,7 +511,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 
     Id access_chain(Id target_type, Id element, std::vector<Id> indexes) {
         begin_op(spv::Op::OpAccessChain, 4 + indexes.size());
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(element);
@@ -515,7 +522,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 
     Id ptr_access_chain(Id target_type, Id base, Id element, std::vector<Id> indexes) {
         begin_op(spv::Op::OpPtrAccessChain, 5 + indexes.size());
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(base);
@@ -527,7 +534,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 
     Id load(Id target_type, Id pointer, std::vector<uint32_t> operands = {}) {
         begin_op(spv::Op::OpLoad, 4 + operands.size());
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(target_type);
         ref_id(id);
         ref_id(pointer);
@@ -546,7 +553,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 
     Id binop(spv::Op op_, Id result_type, Id lhs, Id rhs) {
         begin_op(op_, 5);
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(result_type);
         ref_id(id);
         ref_id(lhs);
@@ -556,7 +563,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 
     Id call(Id return_type, Id callee, std::vector<Id> arguments) {
         begin_op(spv::Op::OpFunctionCall, 4 + arguments.size());
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(return_type);
         ref_id(id);
         ref_id(callee);
@@ -569,7 +576,7 @@ struct BasicBlockBuilder : public SectionBuilder {
     Id ext_instruction(Id return_type, ExtendedInstruction instr, std::vector<Id> arguments);
 
     struct TerminatorBuilder : public SectionBuilder {
-        TerminatorBuilder(BasicBlockBuilder& bb) : bb(bb) {}
+        TerminatorBuilder(BasicBlockBuilder& bb) : SectionBuilder(bb.file_builder_), bb(bb) {}
 
         void branch(Id target) {
             begin_op(spv::Op::OpBranch, 2);
@@ -632,7 +639,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 protected:
     Id ext_instruction(Id return_type, Id set, uint32_t instruction, std::vector<Id> arguments) {
         begin_op(spv::Op::OpExtInst, 5 + arguments.size());
-        auto id = file_builder.generate_fresh_id();
+        auto id = fresh_id();
         ref_id(return_type);
         ref_id(id);
         ref_id(set);
@@ -645,7 +652,7 @@ struct BasicBlockBuilder : public SectionBuilder {
 
 struct FnBuilder {
     explicit FnBuilder(FileBuilder& file_builder)
-            : file_builder(file_builder)
+            : file_builder(file_builder), header(file_builder), variables(file_builder)
     {
         function_id = file_builder.generate_fresh_id();
     }
@@ -726,7 +733,7 @@ inline Id FileBuilder::define_function(FnBuilder &fn_builder, bool definition) {
 }
 
 inline Id BasicBlockBuilder::ext_instruction(Id return_type, ExtendedInstruction instr, std::vector<Id> arguments) {
-    return ext_instruction(return_type, file_builder.extended_import(instr.set_name), instr.id, arguments);
+    return ext_instruction(return_type, file_builder_.extended_import(instr.set_name), instr.id, arguments);
 }
 
 }

From f7bdd76fa06ffa0951de2f0d4dbd0f0a5fa4b438 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 21 Nov 2024 14:23:08 +0100
Subject: [PATCH 336/342] spirv: emit global atomic instead

---
 src/thorin/be/spirv/spirv_instructions.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 5c901827b..c66d5ab5a 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -152,7 +152,7 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             op = spv::OpAtomicIAdd;
         else
             assert(false && "unknown primitive type for atomic_add");
-        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeInvocation), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask), value });
+        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeDevice), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsCrossWorkgroupMemoryMask), value });
         return { result };
     } else if (intrinsic->name() == "rv_all") {
         auto args = emit_args(app.args().skip_back());

From ca68fb92dede62833fd064dac2496e7790726466 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Fri, 22 Nov 2024 15:24:11 +0100
Subject: [PATCH 337/342] spirv: fix regression

---
 src/thorin/be/spirv/spirv.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv.cpp b/src/thorin/be/spirv/spirv.cpp
index 89664230f..931546ea0 100644
--- a/src/thorin/be/spirv/spirv.cpp
+++ b/src/thorin/be/spirv/spirv.cpp
@@ -195,7 +195,7 @@ FnBuilder& CodeGen::get_fn_builder(thorin::Continuation* continuation) {
 
     auto& fn = *(builder_->fn_builders_[continuation] = std::make_unique<FnBuilder>(*builder_));
     auto fn_type = entry_->type();
-    if (kernel_config_->contains(continuation)) {
+    if (kernel_config_ && kernel_config_->contains(continuation)) {
         fn_type = patch_entry_point_signature(fn_type);
     }
     fn.fn_type = convert(fn_type).id;
@@ -247,7 +247,7 @@ void CodeGen::prepare(thorin::Continuation* cont, FnBuilder* fn) {
         world().ddef(cont, "Emitting {} as return block", cont);
 
     if (entry_ == cont) {
-        bool entry_point = kernel_config_->contains(entry_);
+        bool entry_point = kernel_config_ && kernel_config_->contains(entry_);
         for (auto param : cont->params()) {
             if (!should_emit(param->type())) {
                 // Nothing

From 9af4dcceb3bf7397273b6bab9d4f067943c38309 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <hugo@xol.io>
Date: Fri, 22 Nov 2024 15:50:27 +0100
Subject: [PATCH 338/342] added default virtual destructor to Backend

---
 src/thorin/be/codegen.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/thorin/be/codegen.h b/src/thorin/be/codegen.h
index 215a54b59..d4d804e94 100644
--- a/src/thorin/be/codegen.h
+++ b/src/thorin/be/codegen.h
@@ -31,6 +31,7 @@ struct DeviceBackends;
 
 struct Backend {
     Backend(DeviceBackends& backends, World& src);
+    virtual ~Backend() = default;
 
     virtual std::unique_ptr<CodeGen> create_cg() = 0;
 

From 9acff437a910a7e5fd394c343dc115083795b3e0 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 27 Nov 2024 14:08:21 +0100
Subject: [PATCH 339/342] spirv: implement atomic_min

---
 src/thorin/be/spirv/spirv_instructions.cpp | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index c66d5ab5a..ddac8ed2e 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -154,6 +154,28 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             assert(false && "unknown primitive type for atomic_add");
         auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeDevice), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsCrossWorkgroupMemoryMask), value });
         return { result };
+    } else if (intrinsic->name() == "atomic_min") {
+        auto args = emit_args(app.args().skip_back());
+        auto [ptr, value] = *(std::array<Id, 2>*)args.data();
+        auto produced = get_produced_type();
+        spv::Op op;
+        if (is_type_f(produced)) {
+            op = spv::OpAtomicFMinEXT;
+            auto ct = convert(produced);
+            switch (ct.layout->size) {
+                case 2: builder_->capability(spv::Capability::CapabilityAtomicFloat16MinMaxEXT); break;
+                case 4: builder_->capability(spv::Capability::CapabilityAtomicFloat32MinMaxEXT); break;
+                case 8: builder_->capability(spv::Capability::CapabilityAtomicFloat64MinMaxEXT); break;
+            }
+            builder_->extension("SPV_EXT_shader_atomic_float_min_max");
+        } else if (is_type_s(produced))
+            op = spv::OpAtomicSMin;
+        else if (is_type_u(produced))
+            op = spv::OpAtomicUMin;
+        else
+            assert(false && "unknown primitive type for atomic_add");
+        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeDevice), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsCrossWorkgroupMemoryMask), value });
+        return { result };
     } else if (intrinsic->name() == "rv_all") {
         auto args = emit_args(app.args().skip_back());
         auto result = bb->op_with_result(spv::Op::OpGroupAll, convert(get_produced_type()).id,  { literal(spv::Scope::ScopeInvocation), emit(app.arg(1)) });

From b19176c9495317da14d53e434efe3d08a0a4837b Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 27 Nov 2024 14:18:39 +0100
Subject: [PATCH 340/342] spirv: use appropriate scope/semantic mask depending
 on atomic ptr param

---
 src/thorin/be/spirv/spirv_instructions.cpp | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index ddac8ed2e..fd1a5fce7 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -52,6 +52,18 @@ Id CodeGen::emit_mathop(BasicBlockBuilder* bb, const thorin::MathOp& mathop) {
     }
 }
 
+std::tuple<spv::Scope, spv::MemorySemanticsMask> addrspace_atomics_params(World& world, AddrSpace as) {
+    switch (as) {
+        case AddrSpace::Global:
+            return std::make_tuple(spv::ScopeDevice, spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsCrossWorkgroupMemoryMask);
+        case AddrSpace::Shared:
+            return std::make_tuple(spv::ScopeWorkgroup, spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsWorkgroupMemoryMask);
+        default:
+            world.ELOG("Unsupported address space for atomics: {}", (int) as);
+            THORIN_UNREACHABLE;
+    }
+}
+
 std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intrinsic, BasicBlockBuilder* bb) {
     auto get_produced_type = [&]() {
         auto ret_type = (*intrinsic->params().back()).type()->as<FnType>();
@@ -152,7 +164,8 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             op = spv::OpAtomicIAdd;
         else
             assert(false && "unknown primitive type for atomic_add");
-        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeDevice), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsCrossWorkgroupMemoryMask), value });
+        auto [scope, semantics] = addrspace_atomics_params(world(), app.arg(1)->type()->as<PtrType>()->addr_space());
+        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(scope), literal(semantics), value });
         return { result };
     } else if (intrinsic->name() == "atomic_min") {
         auto args = emit_args(app.args().skip_back());
@@ -174,7 +187,8 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             op = spv::OpAtomicUMin;
         else
             assert(false && "unknown primitive type for atomic_add");
-        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(spv::Scope::ScopeDevice), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsMask::MemorySemanticsCrossWorkgroupMemoryMask), value });
+        auto [scope, semantics] = addrspace_atomics_params(world(), app.arg(1)->type()->as<PtrType>()->addr_space());
+        auto result = bb->op_with_result(op, convert(get_produced_type()).id,  { ptr, literal(scope), literal(semantics), value });
         return { result };
     } else if (intrinsic->name() == "rv_all") {
         auto args = emit_args(app.args().skip_back());

From 13db35de0dda2a47a32636b9bb3a6a68c53e8620 Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Wed, 27 Nov 2024 17:25:16 +0100
Subject: [PATCH 341/342] spirv: fix reserve_shared

---
 src/thorin/be/spirv/spirv_instructions.cpp | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index fd1a5fce7..2bfbaccb6 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -116,14 +116,18 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
         } else
             world().ELOG("spirv.builtin requires an integer literal as the argument");
     } else if (intrinsic->name() == "reserve_shared") {
+        auto produced_t = get_produced_type()->as<PtrType>();
+        auto pointee = produced_t->pointee();
+        if (auto indef = pointee->isa<IndefiniteArrayType>())
+            pointee = indef->elem_type();
         auto size = app.arg(1)->isa<PrimLit>();
         if (!size)
             world().error(app.loc(), "reserve_shared called with non-constant size");
         else {
-            auto in_bytes = size->value().get_u64();
+            auto in_bytes = size->value().get_u64() * convert(pointee).layout->size;
             auto type = world().definite_array_type(world().type_pu8(), in_bytes);
             Id id = builder_->variable(convert(world().ptr_type(type, 1, AddrSpace::Shared)).id, static_cast<spv::StorageClass>(convert(AddrSpace::Shared)));
-            id = bb->convert(spv::Op::OpBitcast, convert(get_produced_type()).id, id);
+            id = bb->convert(spv::Op::OpBitcast, convert(produced_t).id, id);
             return { id };
         }
     } else if (intrinsic->name() == "min") {

From 174499ee63e07d1d7c883c6331c0e85e23faaf8a Mon Sep 17 00:00:00 2001
From: Hugo Devillers <devillers@cg.uni-saarland.de>
Date: Thu, 28 Nov 2024 15:30:35 +0100
Subject: [PATCH 342/342] spirv: emit barrier() as a Control barrier instead

---
 src/thorin/be/spirv/spirv_instructions.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/thorin/be/spirv/spirv_instructions.cpp b/src/thorin/be/spirv/spirv_instructions.cpp
index 2bfbaccb6..1e011542c 100644
--- a/src/thorin/be/spirv/spirv_instructions.cpp
+++ b/src/thorin/be/spirv/spirv_instructions.cpp
@@ -148,7 +148,7 @@ std::vector<Id> CodeGen::emit_intrinsic(const App& app, const Continuation* intr
             return { bb->ext_instruction(convert(get_produced_type()).id, { .set_name = "OpenCL.std", .id = OpenCLLIB::SMax }, emit_args(app.args().skip_back())) };
     } else if (intrinsic->name() == "barrier") {
         emit_args(app.args().skip_back());
-        bb->op(spv::Op::OpMemoryBarrier, { literal(spv::Scope::ScopeInvocation), literal(spv::MemorySemanticsMask::MemorySemanticsAcquireReleaseMask) });
+        bb->op(spv::Op::OpControlBarrier, { literal(spv::Scope::ScopeWorkgroup), literal(spv::Scope::ScopeWorkgroup), literal(spv::MemorySemanticsMask::MemorySemanticsWorkgroupMemoryMask | spv::MemorySemanticsMask::MemorySemanticsSequentiallyConsistentMask) });
         return { };
     } else if (intrinsic->name() == "atomic_add") {
         auto args = emit_args(app.args().skip_back());