Migrating to LLVM18

.github/workflows/test.yml

@@ -18,7 +18,7 @@ jobs:
     - uses: actions/checkout@v4
 
     - name: install dependencies
-      run: sudo apt-get update && sudo apt-get install -y llvm-14 clang-14 libc++-14-dev libc++abi-14-dev python3-minimal libgoogle-perftools-dev libboost-container-dev python3-dev libbsd-dev
+      run: sudo apt-get update && sudo apt-get install -y llvm-18 clang-18 libc++-18-dev libc++abi-18-dev libunwind-18-dev python3-minimal libgoogle-perftools-dev libboost-container-dev python3-dev libbsd-dev
 
     - name: Cache Z3
       id: cache-z3
@@ -51,13 +51,13 @@ jobs:
         path: ${{ github.workspace }}/aflpp
 
     - name: configure
-      run: CC=clang-14 CXX=clang++-14 cmake -B ${{ github.workspace }}/build -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=${{ github.workspace }}/install -DAFLPP_PATH=${{ github.workspace }}/aflpp
+      run: CC=clang-18 CXX=clang++-18 cmake -B ${{ github.workspace }}/build -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=${{ github.workspace }}/install -DAFLPP_PATH=${{ github.workspace }}/aflpp -DLLVM_DIR=/usr/lib/llvm-18/lib/cmake/llvm
 
     - name: build
-      run: CC=clang-14 CXX=clang++-14 cmake --build ${{ github.workspace }}/build
+      run: CC=clang-18 CXX=clang++-18 cmake --build ${{ github.workspace }}/build
 
     - name: install
-      run: CC=clang-14 CXX=clang++-14 cmake --install ${{ github.workspace }}/build
+      run: CC=clang-18 CXX=clang++-18 cmake --install ${{ github.workspace }}/build
 
     - name: install lit
       run: pip install lit

CMakeLists.txt

@@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.13)
 
 project(symsan VERSION 1.2.2 LANGUAGES C CXX ASM)
 
-find_package(LLVM 14 REQUIRED CONFIG)
+find_package(LLVM REQUIRED CONFIG)
 
 # Find Z3 (minimum version 4.8.15 required for string theory APIs)
 # Prefer /usr/local over system
@@ -44,6 +44,13 @@ if (LLVM_FOUND)
     message(STATUS "LLVM_VERSION_MAJOR: ${LLVM_VERSION_MAJOR}")
     message(STATUS "LLVM_VERSION_MINOR: ${LLVM_VERSION_MINOR}")
     message(STATUS "LLVM_VERSION_PATCH: ${LLVM_VERSION_PATCH}")
+    # SymSan requires LLVM >= 18 (the default on Ubuntu 24.04). Older versions
+    # predate opaque pointers and the DFSan trampoline removal and are no longer
+    # supported. (find_package(LLVM 18) can't be used: LLVM's version config
+    # requires an exact major.minor match.)
+    if (LLVM_VERSION_MAJOR LESS 18)
+        message(FATAL_ERROR "SymSan requires LLVM >= 18, found ${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR}")
+    endif()
 else()
     message(FATAL_ERROR "You haven't install LLVM !")
 endif()

README.md

@@ -23,14 +23,14 @@ it has more strict dependency on the LLVM version. Right now only LLVM 12 is tes
 ### Build Requirements
 
 - Linux-amd64 (Tested on Ubuntu 24.04)
-- [LLVM 14.0.6](http://llvm.org/docs/index.html): clang, libc++, libc++abi
+- [LLVM 18.1.18](http://llvm.org/docs/index.html): clang, libc++, libc++abi
 
 ### Compilation
 
 Create a `build` directory and execute the following commands in it:
 
 ```shell
-$ CC=clang-14 CXX=clang++-14 cmake -DCMAKE_INSTALL_PREFIX=/path/to/install -DCMAKE_BUILD_TYPE=Release /path/to/symsan/source
+$ CC=clang-18 CXX=clang++-18 cmake -DCMAKE_INSTALL_PREFIX=/path/to/install -DCMAKE_BUILD_TYPE=Release /path/to/symsan/source
 $ make
 $ make install
 ```

compiler/CMakeLists.txt

@@ -1,5 +1,9 @@
 add_executable(KOClang ko_clang.c)
 set_target_properties(KOClang PROPERTIES OUTPUT_NAME "ko-clang")
+# Let the driver version-gate the clang flags it forwards.
+target_compile_definitions(KOClang PRIVATE
+    LLVM_VERSION_MAJOR=${LLVM_VERSION_MAJOR}
+    LLVM_VERSION_MINOR=${LLVM_VERSION_MINOR})
 
 add_custom_command(TARGET KOClang POST_BUILD
     COMMAND ln -sf "ko-clang" "ko-clang++")

compiler/ko_clang.c

@@ -444,7 +444,11 @@ static void edit_params(u32 argc, char **argv) {
   }
 
   if (!skip_instrumentation) {
+    // The new pass manager is the default since clang 13 and the flag that
+    // used to request it was removed in clang 16.
+#if LLVM_VERSION_CODE < LLVM_VERSION(16, 0)
     cc_params[cc_par_cnt++] = "-fexperimental-new-pass-manager";
+#endif
     // add UCSanPass first, if specified
     if (use_ucsan) {
       add_ucsan_pass();

compiler/ucsan_opt

@@ -13,15 +13,13 @@ else:
     print(f"Cannot find {ucsan_pass_path}, please make sure you have run this script in the right directory(the directory ko-clang installed)")
     os._exit(1)
 
-KO_FLAGS = ["KO_CC=clang-14"]
+KO_FLAGS = ["KO_CC=clang-18"]
 ko_flags = ' '.join(KO_FLAGS)
 
-# Base opt-14 command: load UCSan pass
-# Load library twice for LLVM-14:
-# -load: registers cl::opt definitions (command-line options)
-# -load-pass-plugin: registers the new pass manager plugin
-cc = f"opt-14 -load {path_to_install}/lib/symsan/UCSanPass.so" \
-     f" -load-pass-plugin={path_to_install}/lib/symsan/UCSanPass.so" \
+# Base opt command: load the UCSan pass plugin. On LLVM 18 -load-pass-plugin
+# alone registers both the new-PM pipeline and the pass's cl::opt options (the
+# separate -load needed on LLVM 14 is no longer required).
+cc = f"opt-18 -load-pass-plugin={path_to_install}/lib/symsan/UCSanPass.so" \
      f" -ucsan-abilist={path_to_install}/lib/symsan/ucsan_abilist.txt"
 
 ko_cc = f"{path_to_install}/bin/ko-clang"
@@ -66,8 +64,8 @@ if not files:
     argparser.print_usage()
     os._exit(1)
 
-# Note: With LLVM-14 new pass manager, we don't add -O# flags to opt
-# Optimization is controlled via KO_DONT_OPTIMIZE and handled by ko-clang
+# Note: with the new pass manager we don't add -O# flags to opt;
+# optimization is controlled via KO_DONT_OPTIMIZE and handled by ko-clang
 if not args.optimization:
     KO_FLAGS.append("KO_DONT_OPTIMIZE=1")
 
@@ -98,9 +96,8 @@ if args.taint:
         os._exit(1)
     # Tell UCSanPass that TaintPass will run after it
     cc += f" -ucsan-with-taint=true"
-    # Load TaintPass library for cl::opt and pass registration
-    cc += f" -load {taint_pass_path}" \
-          f" -load-pass-plugin={taint_pass_path}" \
+    # Load TaintPass plugin (registers its pipeline and cl::opt options)
+    cc += f" -load-pass-plugin={taint_pass_path}" \
           f" -taint-abilist={path_to_install}/lib/symsan/dfsan_abilist.txt" \
           f" -taint-with-ucsan=true" \
           f" -taint-solve-ub=true"
@@ -138,7 +135,7 @@ for file in files:
         cleanup()
         os._exit(1)
     tmp_files.append(bc_file)
-    ret = command(f"llc-14 -filetype=obj --relocation-model=pic -o {obj_file} {bc_file}")
+    ret = command(f"llc-18 -filetype=obj --relocation-model=pic -o {obj_file} {bc_file}")
     if ret != 0:
         print(f"Failed to compile {file}, check the error message above")
         cleanup()

include/version.h

@@ -5,43 +5,14 @@
 #define LLVM_VERSION(major, minor) ((major)*100 + (minor))
 #define LLVM_VERSION_CODE LLVM_VERSION(LLVM_VERSION_MAJOR, LLVM_VERSION_MINOR)
 
-#if LLVM_VERSION_CODE >= LLVM_VERSION(5, 0)
-#define LLVM_ATTRIBUTE_LIST AttributeList
-
-#define LLVM_NEW_ALLOCINST(ty, name, insertp)                                  \
-  (new AllocaInst(ty, getDataLayout().getAllocaAddrSpace(), name, insertp))
-
-#define LLVM_REMOVE_ATTRIBUTE(func, attr, attrbuilder)                         \
-  func->removeAttributes(attr, attrbuilder)
-
-#else
-
-#define LLVM_ATTRIBUTE_LIST AttributeSet
-
-#define LLVM_NEW_ALLOCINST(ty, name, insertp)                                  \
-  (new AllocaInst(ty, name, insertp))
-
-#define LLVM_REMOVE_ATTRIBUTE(func, attr, attrbuilder)                         \
-  func->removeAttributes(                                                      \
-      attr, LLVM_ATTRIBUTE_LIST::get(func->getContext(), attr, attrbuilder))
-
-#endif
-
-#if LLVM_VERSION_CODE >= LLVM_VERSION(6, 0)
-
-#define SCL_INSECTION(scl, section, prefix, query, category)                   \
-  scl->inSection(section, prefix, query, category)
-
-#define LLVM_ADD_PARAM_ATTR(func, argno, attr) func->addParamAttr(argno, attr)
-
-#else
-
-#define SCL_INSECTION(scl, section, prefix, query, category)                   \
-  scl->inSection(prefix, query, category)
-
-#define LLVM_ADD_PARAM_ATTR(func, argno, attr)                                 \
-  func->addAttribute(argno + 1, attr)
-
+// SymSan targets LLVM 18 (the default on Ubuntu 24.04) as its minimum.  These
+// headers are only needed by C++ translation units built against LLVM; the
+// compiler-driver C sources include this file only for the LLVM_VERSION macros
+// above.
+#if defined(__cplusplus) && defined(LLVM_VERSION_MAJOR)
+#include <optional>
+#include "llvm/TargetParser/Triple.h"
+#include "llvm/IR/AttributeMask.h"
 #endif
 
 #endif

instrumentation/CMakeLists.txt

@@ -1,4 +1,4 @@
-set (CMAKE_CXX_STANDARD 14)
+set (CMAKE_CXX_STANDARD 17)
 set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g")
 # fix pass bug: https://github.com/sampsyo/llvm-pass-skeleton/issues/7#issuecomment-401834287
 set (CMAKE_MODULE_LINKER_FLAGS "${CMAKE_CXX_LINK_FLAGS} -Wl,-znodelete")

instrumentation/LoopOutlinePass.cpp

@@ -26,6 +26,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/ADT/SetVector.h"
+#include "llvm/BinaryFormat/Dwarf.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/IR/BasicBlock.h"
@@ -207,7 +208,7 @@ collectGlobals(Function *Out) {
     for (Instruction &I : BB)
       for (Value *Op : I.operands())
         if (auto *gv = dyn_cast<GlobalVariable>(Op->stripPointerCasts()))
-          if (!gv->getName().startswith("llvm.") && !gv->getName().empty())
+          if (!(gv->getName()).starts_with("llvm.") && !gv->getName().empty())
             gvs.insert(gv);
   std::vector<std::pair<std::string, std::string>> out;
   for (GlobalVariable *gv : gvs) {
@@ -240,10 +241,12 @@ static Function *extractTarget(const Target &T, unsigned Idx) {
   DT.recalculate(*T.F);
 
   CodeExtractorAnalysisCache CEAC(*T.F);
+  // NOTE: the trailing Suffix argument is intentionally left at its default;
+  // LLVM 18 inserted an AllocationBlock parameter before it, so passing it
+  // positionally is not source-compatible across versions.
   CodeExtractor CE(T.Blocks, &DT, /*AggregateArgs=*/false,
                    /*BFI=*/nullptr, /*BPI=*/nullptr, /*AC=*/nullptr,
-                   /*AllowVarArgs=*/false, /*AllowAlloca=*/false,
-                   /*Suffix=*/"");
+                   /*AllowVarArgs=*/false, /*AllowAlloca=*/false);
   if (!CE.isEligible()) {
     if (ClVerbose)
       errs() << "[loop-outline] skip " << T.F->getName()
@@ -274,7 +277,7 @@ struct LoopOutlinePass : public PassInfoMixin<LoopOutlinePass> {
     // Phase 1 (read-only): gather innermost-loop body regions.
     std::vector<Target> Targets;
     for (Function &F : M) {
-      if (F.isDeclaration() || F.getName().startswith("__ucsan_loopbody_"))
+      if (F.isDeclaration() || (F.getName()).starts_with("__ucsan_loopbody_"))
         continue;
       DominatorTree DT;
       DT.recalculate(F);