Fix Clang + libc++ compilation and portability issues

include/onnxruntime/core/graph/graph.h

@@ -73,7 +73,10 @@ class Node {
     Fused = 1,      ///< The node refers to a function.
   };
 
-  explicit Node() = default;
+  // Constructor and destructor defined out-of-line in graph.cc so that Graph
+  // is a complete type when std::unique_ptr<Graph> in subgraphs_ is destroyed
+  // (required by libc++).
+  explicit Node();
 
 #if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
   Node(std::string_view name,
@@ -82,15 +85,10 @@ class Node {
        gsl::span<NodeArg* const> input_args,
        gsl::span<NodeArg* const> output_args,
        const NodeAttributes* attributes,
-       std::string_view domain) {
-    Init(name, op_type, description,
-         input_args,
-         output_args,
-         attributes, domain);
-  }
+       std::string_view domain);
 #endif
 
-  ~Node() = default;
+  ~Node();
 
   /**
   @class EdgeEnd
@@ -588,7 +586,7 @@ class Node {
   // NOTE: This friendship relationship should ONLY be used for calling methods of the Node class and not accessing
   // the data members directly, so that the Node can maintain its internal invariants.
   friend class Graph;
-  Node(NodeIndex index, Graph& graph) : index_(index), graph_(&graph), can_be_saved_(true) {}
+  Node(NodeIndex index, Graph& graph);
 
  protected:
 #if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)

include/onnxruntime/core/session/onnxruntime_c_api.h

@@ -91,8 +91,14 @@ extern "C" {
 #define ORT_MUST_USE_RESULT
 #define ORTCHAR_T wchar_t
 #else
-// To make symbols visible on macOS/iOS
-#ifdef __APPLE__
+// Make symbols visible on non-Windows platforms. The visibility attribute is
+// needed when ORT is built as a shared library without a version script
+// (e.g. when compiled within another project's build system). On Linux, the
+// default ORT build uses a generated version script (tools/ci_build/gen_def.py)
+// that exports the needed symbols, so this was previously only enabled for
+// __APPLE__. Expanding to __GNUC__ (GCC/Clang) covers additional embedding
+// scenarios while remaining harmless when a version script is also in use.
+#if defined(__GNUC__)
 #define ORT_EXPORT __attribute__((visibility("default")))
 #else
 #define ORT_EXPORT

onnxruntime/contrib_ops/cpu/bert/bias_gelu.cc

@@ -11,6 +11,7 @@
 #include "core/providers/common.h"
 #include "core/util/math_cpuonly.h"
 #include "core/mlas/inc/mlas.h"
+#include "core/common/math_constants.h"
 using onnxruntime::narrow;
 namespace onnxruntime {
 namespace contrib {

onnxruntime/core/common/math_constants.h

@@ -0,0 +1,24 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+// Portable math constants using C++20 <numbers>.
+// POSIX constants like M_PI and M_SQRT2 are not part of the C++ standard and
+// are not provided by all standard library implementations (e.g. libc++).
+// This header defines them in terms of std::numbers so they are available
+// on any conforming C++20 standard library.
+
+#pragma once
+
+#include <numbers>
+
+#ifndef M_PI
+#define M_PI std::numbers::pi
+#endif
+
+#ifndef M_SQRT1_2
+#define M_SQRT1_2 (1.0 / std::numbers::sqrt2)
+#endif
+
+#ifndef M_SQRT2
+#define M_SQRT2 std::numbers::sqrt2
+#endif

onnxruntime/core/graph/ep_api_types.cc

@@ -110,6 +110,8 @@ static bool IsOptionalAttribute(const Node& node, const std::string& attr_name)
 // EpNode
 //
 
+EpNode::SubgraphState::~SubgraphState() = default;
+
 EpNode::EpNode(const EpGraph* ep_graph, const Node& node, PrivateTag)
     : OrtNode(OrtGraphIrApi::kEpApi), ep_graph_(ep_graph), node_(node) {}
 

onnxruntime/core/graph/ep_api_types.h

@@ -115,6 +115,9 @@ struct EpNode : public OrtNode {
   struct SubgraphState {
     SubgraphState() = default;
     SubgraphState(SubgraphState&& other) = default;
+    // Destructor defined out-of-line so EpGraph is complete when
+    // unique_ptr<EpGraph> is destroyed (required by libc++).
+    ~SubgraphState();
     std::string attribute_name;
     std::unique_ptr<GraphViewer> subgraph_viewer;  // The graph_viewer wrapped by EpGraph below.
     std::unique_ptr<EpGraph> ep_subgraph;

onnxruntime/core/graph/graph.cc

@@ -527,6 +527,28 @@ bool NodeArg::Exists() const noexcept {
   return exists_;
 }
 
+// Out-of-line constructor and destructor so Graph is complete when
+// unique_ptr<Graph> in subgraphs_ is destroyed (required by libc++).
+Node::Node() = default;
+Node::~Node() = default;
+
+Node::Node(NodeIndex index, Graph& graph) : index_(index), graph_(&graph), can_be_saved_(true) {}
+
+#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
+Node::Node(std::string_view name,
+           std::string_view op_type,
+           std::string_view description,
+           gsl::span<NodeArg* const> input_args,
+           gsl::span<NodeArg* const> output_args,
+           const NodeAttributes* attributes,
+           std::string_view domain) {
+  Init(name, op_type, description,
+       input_args,
+       output_args,
+       attributes, domain);
+}
+#endif
+
 Node::EdgeEnd::EdgeEnd(const Node& node, int src_arg_index, int dst_arg_index) noexcept
     : node_(&node),
       src_arg_index_(src_arg_index),

onnxruntime/core/mlas/lib/aarch64/asmmacro.h

@@ -32,6 +32,9 @@ Macro Description:
 #if defined(__APPLE__)
         .globl  _\FunctionName\()
 _\FunctionName\():
+#elif defined(_WIN32)
+        .globl  \FunctionName\()
+\FunctionName\():
 #else
         .globl  \FunctionName\()
         .type   \FunctionName\(),%function

onnxruntime/core/mlas/lib/activate.cpp

@@ -131,7 +131,7 @@ struct MLAS_ACTIVATION_FUNCTION<MlasLeakyReluActivation>
         MLAS_FLOAT32X4 ValueTimesAlpha = MlasMultiplyFloat32x4(Value, AlphaBroadcast);
 
 #if defined(MLAS_NEON_INTRINSICS)
-#if defined(_WIN32)
+#if defined(_WIN32) && !defined(__clang__)
         return vbslq_f32(vcleq_z_f32_ex(Value), ValueTimesAlpha, Value);
 #else
         // N.B. Standard NEON headers lack an intrinsic for the "vcle #0" form.

onnxruntime/core/mlas/lib/mlasi.h

@@ -1803,7 +1803,7 @@ MlasConvDepthwiseMultiplier2CHWKernel7x7S2Avx512F(
 // Also define additional standard NEON intrinsics using the MSVC aliases.
 //
 
-#if defined(_M_ARM64)
+#if defined(_M_ARM64) && !defined(__clang__)
 #ifndef vmaxvq_f32
 #define vmaxvq_f32(src) neon_fmaxv(src)
 #endif

onnxruntime/core/mlas/lib/qladd.h

@@ -50,7 +50,7 @@ MlasCalcQLinearAddParameters(
 
 #if defined(MLAS_NEON_INTRINSICS)
 
-#if ! defined(_MSC_VER)
+#if !defined(_MSC_VER) || defined(__clang__)
 
 #define vld1q_s8_ex(pD, align) vld1q_s8((int8_t*)__builtin_assume_aligned(pD, ((align)/8)))
 #define vst1_s8_ex(pD, D, align) vst1_s8((int8_t*)__builtin_assume_aligned(pD, ((align)/8)), D)

onnxruntime/core/optimizer/gelu_fusion.cc

@@ -7,6 +7,7 @@
 #include "core/graph/graph_utils.h"
 #include "float.h"
 #include <deque>
+#include "core/common/math_constants.h"
 
 using namespace ONNX_NAMESPACE;
 using namespace onnxruntime::common;

onnxruntime/core/optimizer/optimizer_execution_frame.cc

@@ -27,6 +27,8 @@ static size_t EstimateInputsOutputs(gsl::span<const Node* const> nodes) {
   return num;
 }
 
+OptimizerExecutionFrame::Info::~Info() = default;
+
 OptimizerExecutionFrame::Info::Info(const std::vector<const Node*>& nodes,
                                     const InitializedTensorSet& initialized_tensor_set,
                                     const std::filesystem::path& model_path,

onnxruntime/core/optimizer/optimizer_execution_frame.h

@@ -36,7 +36,9 @@ class OptimizerExecutionFrame final : public IExecutionFrame {
          const std::function<bool(const std::string&)>& is_sparse_initializer_func,
          const logging::Logger& logger);
 
-    ~Info() = default;
+    // Destructor defined out-of-line so NodeIndexInfo is complete when
+    // unique_ptr<NodeIndexInfo> is destroyed (required by libc++).
+    ~Info();
 
     const AllocatorPtr& GetAllocator() const {
       return allocator_ptr_;

onnxruntime/core/optimizer/qdq_transformer/selectors_actions/shared/utils.cc

@@ -18,6 +18,20 @@
 namespace onnxruntime {
 namespace QDQ {
 
+// Out-of-line constructor/destructor definitions so NodeGroupSelector is
+// complete when unique_ptr<NodeGroupSelector> is destroyed (required by libc++).
+OpVersionsAndSelector::OpVersionsAndSelector(const OpVersionsMap& ops_and_versions_in,
+                                             std::unique_ptr<NodeGroupSelector> selector_in)
+    : op_versions_map{ops_and_versions_in},
+      selector{std::move(selector_in)} {}
+
+OpVersionsAndSelector::~OpVersionsAndSelector() = default;
+
+Selectors::Selectors() = default;
+Selectors::~Selectors() = default;
+
+SelectorManager::~SelectorManager() = default;
+
 void Selectors::RegisterSelector(const OpVersionsAndSelector::OpVersionsMap& ops_and_versions_in,
                                  std::unique_ptr<NodeGroupSelector> selector_in) {
   auto entry = std::make_unique<OpVersionsAndSelector>(

onnxruntime/core/optimizer/qdq_transformer/selectors_actions/shared/utils.h

@@ -31,9 +31,11 @@ struct OpVersionsAndSelector {
   using OpVersionsMap = std::unordered_map<std::string, std::vector<ONNX_NAMESPACE::OperatorSetVersion>>;
 
   OpVersionsAndSelector(const OpVersionsMap& ops_and_versions_in,
-                        std::unique_ptr<NodeGroupSelector> selector_in)
-      : op_versions_map{ops_and_versions_in},
-        selector{std::move(selector_in)} {}
+                        std::unique_ptr<NodeGroupSelector> selector_in);
+
+  // Destructor defined out-of-line so NodeGroupSelector is complete when
+  // unique_ptr<NodeGroupSelector> is destroyed (required by libc++).
+  ~OpVersionsAndSelector();
 
   OpVersionsMap op_versions_map;
   std::unique_ptr<NodeGroupSelector> selector;
@@ -44,7 +46,11 @@ struct OpVersionsAndSelector {
 // class that manages a set of node group selectors
 class Selectors {
  public:
-  Selectors() = default;
+  // Constructor/destructor defined out-of-line so NodeGroupSelector is
+  // complete when unique_ptr<NodeGroupSelector> is destroyed (required
+  // by libc++, which checks completeness at the point of the destructor).
+  Selectors();
+  ~Selectors();
 
   // register a selector for the specified ops.
   void RegisterSelector(const OpVersionsAndSelector::OpVersionsMap& ops_and_versions_in,
@@ -64,6 +70,7 @@ class Selectors {
 class SelectorManager {
  public:
   SelectorManager();
+  ~SelectorManager();
 
   // Methods that finds and returns a vector of QDQ::NodeGroup in a given graph
   // Can be used in QDQ support in different EPs

onnxruntime/core/optimizer/stft_decomposition.cc

@@ -11,6 +11,7 @@
 #include "core/optimizer/utils.h"
 #include "core/framework/op_kernel.h"
 #include "core/framework/tensorprotoutils.h"
+#include "core/common/math_constants.h"
 
 using namespace onnxruntime::common;
 

onnxruntime/core/platform/windows/env.h

@@ -47,7 +47,7 @@ class WindowsEnv : public Env {
 #endif
   EnvThread* CreateThread(_In_opt_z_ const ORTCHAR_T* name_prefix, int index,
                           unsigned (*start_address)(int id, Eigen::ThreadPoolInterface* param),
-                          Eigen::ThreadPoolInterface* param, const ThreadOptions& thread_options);
+                          Eigen::ThreadPoolInterface* param, const ThreadOptions& thread_options) override;
 #if defined(_MSC_VER) && !defined(__clang__)
 #pragma warning(pop)
 #endif

onnxruntime/core/platform/windows/stacktrace.cc

@@ -6,7 +6,11 @@
 #include <mutex>
 #include <sstream>
 #ifdef __has_include
-#if __has_include(<stacktrace>)
+// libc++ provides a <stacktrace> header that passes __has_include and defines
+// __cpp_lib_stacktrace, but it is a stub that does not contain a working
+// std::stacktrace implementation (as of libc++ 18). Including it leads to
+// compile errors, so we exclude libc++ entirely.
+#if __has_include(<stacktrace>) && !defined(_LIBCPP_VERSION)
 #include <stacktrace>
 #endif
 #endif
@@ -30,7 +34,7 @@ class CaptureStackTrace {
 // Get the stack trace. Currently only enabled for a DEBUG build as we require the DbgHelp library.
 std::vector<std::string> GetStackTrace() {
 #ifndef NDEBUG
-#if (defined __cpp_lib_stacktrace) && !(defined _OPSCHEMA_LIB_) && !(defined _GAMING_XBOX) && !(defined ONNXRUNTIME_ENABLE_MEMLEAK_CHECK)
+#if (defined __cpp_lib_stacktrace) && !defined(_LIBCPP_VERSION) && !(defined _OPSCHEMA_LIB_) && !(defined _GAMING_XBOX) && !(defined ONNXRUNTIME_ENABLE_MEMLEAK_CHECK)
   return detail::CaptureStackTrace().Trace();
 #else
   return {};
@@ -42,7 +46,7 @@ std::vector<std::string> GetStackTrace() {
 
 namespace detail {
 #ifndef NDEBUG
-#if (defined __cpp_lib_stacktrace) && !(defined _OPSCHEMA_LIB_) && !(defined _GAMING_XBOX) && !(defined ONNXRUNTIME_ENABLE_MEMLEAK_CHECK)
+#if (defined __cpp_lib_stacktrace) && !defined(_LIBCPP_VERSION) && !(defined _OPSCHEMA_LIB_) && !(defined _GAMING_XBOX) && !(defined ONNXRUNTIME_ENABLE_MEMLEAK_CHECK)
 
 std::vector<std::string> CaptureStackTrace::Trace() const {
   std::vector<std::string> stacktrace;

onnxruntime/core/providers/cpu/math/cumsum.cc

@@ -106,7 +106,7 @@ CumSum<T>::CumSum(const OpKernelInfo& info) : OpKernel(info), exclusive_(), reve
     if (exclusive == 1 || exclusive == 0) {
       exclusive_ = exclusive;
     } else {
-      ORT_ENFORCE("attribute exclusive can only be 0 or 1");
+      ORT_THROW("attribute exclusive can only be 0 or 1");
     }
   }
   int64_t reverse = 0;
@@ -115,7 +115,7 @@ CumSum<T>::CumSum(const OpKernelInfo& info) : OpKernel(info), exclusive_(), reve
     if (reverse == 1 || reverse == 0) {
       reverse_ = reverse;
     } else {
-      ORT_ENFORCE("attribute reverse can only be 0 or 1");
+      ORT_THROW("attribute reverse can only be 0 or 1");
     }
   }
 }

onnxruntime/core/providers/cpu/ml/ml_common.h

@@ -10,6 +10,7 @@
 #include "core/mlas/inc/mlas.h"
 #include "core/platform/threadpool.h"
 #include "core/common/inlined_containers.h"
+#include "core/common/math_constants.h"
 
 namespace onnxruntime {
 namespace ml {  // name space for onnx.ml operators

onnxruntime/core/providers/cpu/signal/dft.cc

@@ -16,6 +16,7 @@
 #include "core/providers/cpu/signal/utils.h"
 #include "core/util/math_cpuonly.h"
 #include "Eigen/src/Core/Map.h"
+#include "core/common/math_constants.h"
 
 namespace onnxruntime {
 
@@ -221,7 +222,7 @@ static Status dft_bluestein_z_chirp(
     OpKernelContext* ctx, const Tensor* X, Tensor* Y, Tensor& b_fft, Tensor& chirp, size_t X_offset, size_t X_stride, size_t Y_offset, size_t Y_stride,
     int64_t axis, size_t dft_length, const Tensor* window, bool is_onesided, bool inverse, InlinedVector<std::complex<T>>& V,
     InlinedVector<std::complex<T>>& temp_output) {
-  static constexpr T pi = static_cast<T>(M_PI);
+  static const T pi = static_cast<T>(3.14159265358979323846);
 
   AllocatorPtr alloc;
   ORT_RETURN_IF_ERROR(ctx->GetTempSpaceAllocator(&alloc));

onnxruntime/core/providers/cpu/signal/window_functions.cc

@@ -8,6 +8,7 @@
 
 #include "core/providers/common.h"
 #include "core/providers/cpu/signal/utils.h"
+#include "core/common/math_constants.h"
 
 namespace onnxruntime {
 ONNX_CPU_OPERATOR_KERNEL(HannWindow, 17,

onnxruntime/core/session/inference_session.cc

@@ -1733,7 +1733,7 @@ static Status LoadOrtModelBytes(const PathString& model_uri,
 
   bytes_data_holder.resize(num_bytes);
 
-  std::ifstream bytes_stream(model_uri, std::ifstream::in | std::ifstream::binary);
+  std::ifstream bytes_stream(model_uri.c_str(), std::ifstream::in | std::ifstream::binary);
   bytes_stream.read(reinterpret_cast<char*>(bytes_data_holder.data()), num_bytes);
 
   if (!bytes_stream) {

onnxruntime/core/session/plugin_ep/ep_plugin_provider_interfaces.cc

@@ -468,6 +468,10 @@ PluginExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph_vie
   return result;
 }
 
+// Out-of-line destructor: EpNode and EpValueInfo must be complete types
+// when unique_ptr members are destroyed (required by libc++).
+PluginExecutionProvider::FusedNodeState::~FusedNodeState() = default;
+
 Status PluginExecutionProvider::FusedNodeState::AddFusedNode(const Node& fused_node, /*out*/ EpNode*& added_ep_node) {
   std::unique_ptr<EpNode> unique_ep_fused_node = nullptr;
   ORT_RETURN_IF_ERROR(EpNode::Create(fused_node, /*parent graph*/ nullptr, this->value_infos, unique_ep_fused_node));

onnxruntime/core/session/plugin_ep/ep_plugin_provider_interfaces.h

@@ -155,7 +155,11 @@ class PluginExecutionProvider : public IExecutionProvider {
   struct FusedNodeState {
     FusedNodeState() = default;
     FusedNodeState(FusedNodeState&& other) = default;
+    FusedNodeState& operator=(FusedNodeState&& other) = default;
     FusedNodeState(const FusedNodeState& other) = delete;
+    // Destructor defined out-of-line so EpNode/EpValueInfo are complete when
+    // unique_ptr<EpNode>/unique_ptr<EpValueInfo> are destroyed (required by libc++).
+    ~FusedNodeState();
     Status AddFusedNode(const Node& fused_node, /*out*/ EpNode*& added_ep_node);
 
     std::vector<std::unique_ptr<EpNode>> nodes;