cmake is woking nicely but the dispatch plugins fail due to llvm bug

Author: at0m741

CMakeLists.txt

@@ -9,33 +9,30 @@ option(VERBOSE "Verbose runtime logs" OFF)
 option(DEBUG "Build with debug symbols" OFF)
 option(BUILD_PLUGINS "Build Clang/LLVM plugins" ON)
 option(BUILD_TESTS "Build test executables" ON)
-option(BUILD_PYBIND "Build Python bindings" ON)
+option(BUILD_PYBIND "Build Python bindings" OFF)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 message(STATUS "Detecting architecture...")
 
+# ─────────────────────────────── Architecture detection ───────────────────────────────
 if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64")
     message(STATUS "→ Building for x86_64 (AVX/AVX2/AVX512 enabled)")
     add_compile_definitions(TENSORIUM_X86)
     add_compile_options(-mavx2 -mfma)
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "arm64|aarch64")
     message(STATUS "→ Building for ARM64 / Apple Silicon (NEON enabled)")
     add_compile_definitions(TENSORIUM_ARM)
-    find_package(OpenMP)
-    if (OpenMP_CXX_FOUND)
-        target_link_libraries(${PROJECT_NAME} PUBLIC OpenMP::OpenMP_CXX)
-    endif()
 else()
     message(WARNING "→ Unknown architecture: ${CMAKE_SYSTEM_PROCESSOR}, using scalar fallback.")
     add_compile_definitions(TENSORIUM_FALLBACK)
 endif()
-# ─────────────────────────────── Includes ──────────────────────────────
+
 include_directories(${CMAKE_SOURCE_DIR}/Includes)
 
-# ─────────────────────────────── CPU architecture flags ───────────────────────────────
+# ─────────────────────────────── CPU optimization flags ───────────────────────────────
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64" OR CMAKE_SYSTEM_PROCESSOR MATCHES "amd64")
   message(STATUS "Configuring for x86_64: enabling AVX2/FMA or AVX512")
   set(BASE_FLAGS "-O3 -mtune=native -Wno-ignored-attributes -Rpass-analysis=tensorium-align")
@@ -46,7 +43,6 @@ if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64" OR CMAKE_SYSTEM_PROCESSOR MATCHES "am
   else()
     set(CMAKE_CXX_FLAGS "${BASE_FLAGS} ${AVX2_FLAGS}")
   endif()
-
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "arm64" OR CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
   message(STATUS "Configuring for Apple Silicon ARM64: disabling AVX flags")
   set(CMAKE_CXX_FLAGS "-O3 -mcpu=apple-m1 -Wno-ignored-attributes")
@@ -62,9 +58,6 @@ if(VERBOSE)
   add_compile_definitions(VERBOSE)
 endif()
 
-
-
-# ─────────────────────────────── Architecture detection ───────────────────────────────
 message(STATUS "Detected architecture: ${CMAKE_SYSTEM_PROCESSOR}")
 
 # ─────────────────────────────── OpenMP handling ───────────────────────────────
@@ -88,15 +81,12 @@ else()
   endif()
 endif()
 
-# ─────────────────────────────── Subdirectories ───────────────────────────────
+# ─────────────────────────────── Plugins ───────────────────────────────
 if(BUILD_PLUGINS)
   add_subdirectory(Plugins)
 endif()
 
+# ─────────────────────────────── Tests ───────────────────────────────
 if(BUILD_TESTS)
   add_subdirectory(Tests)
 endif()
-
-if(BUILD_PYBIND)
-  add_subdirectory(Pybind)
-endif()

Makefile

@@ -2,12 +2,22 @@ find_package(LLVM REQUIRED CONFIG)
 include_directories(${LLVM_INCLUDE_DIRS})
 add_definitions(${LLVM_DEFINITIONS})
 
-# AST plugin
+# List of components to map to libraries
+set(LLVM_COMPONENTS
+    Core
+    Support
+    IRReader
+    Analysis
+    Passes
+    TransformUtils
+)
+
+llvm_map_components_to_libnames(LLVM_LIBS ${LLVM_COMPONENTS})
+
+# Tensorium Dispatch Plugin (Clang AST)
 add_library(TensoriumDispatchPlugin SHARED TensoriumDispatchPlugin.cpp)
-target_link_libraries(TensoriumDispatchPlugin PRIVATE clang-cpp LLVM)
-set_target_properties(TensoriumDispatchPlugin PROPERTIES COMPILE_FLAGS "")
+target_link_libraries(TensoriumDispatchPlugin PRIVATE clang-cpp ${LLVM_LIBS})
 
-# LLVM IR plugin
+# Tensorium LLVM IR Plugin
 add_library(TensoriumLLVM_IRCheck SHARED TensoriumLLVM_IRCheck.cpp)
-target_link_libraries(TensoriumLLVM_IRCheck PRIVATE LLVM)
-set_target_properties(TensoriumLLVM_IRCheck PROPERTIES COMPILE_FLAGS "")
+target_link_libraries(TensoriumLLVM_IRCheck PRIVATE ${LLVM_LIBS})

Plugins/CMakeLists.txt

@@ -6,6 +6,12 @@
 #include "clang/Lex/Preprocessor.h"
 #include "clang/Tooling/Tooling.h"
 #include "llvm/Support/MemoryBuffer.h"
+#include <iostream>
+//
+// static int _ = []() {
+//     std::cerr << "[TensoriumDispatchPlugin] Plugin loaded into Clang.\n";
+//     return 0;
+// }();
 // #include "LLVM_Handler.hpp"
 /**
  * @file TensoriumPlugin.cpp
@@ -59,7 +65,8 @@ class TensoriumASTConsumer : public ASTConsumer {
         // 		SourceLocation loc = FD->getBeginLoc();
         // 		for (const auto &entry : TensoriumTargetTable) {
         // 			if (Context.getSourceManager().isBeforeInTranslationUnit(entry.loc,
-        // loc)) { 				std::string fname = FD->getNameAsString(); 				llvm::errs() << "[tensorium] Target("
+        // loc)) { 				std::string fname = FD->getNameAsString();
+        // llvm::errs() << "[tensorium] Target("
         // << entry.platform << ", " << entry.isa
         // 					<< ") applies to function " << fname << "\n";
         //
@@ -300,6 +307,16 @@ class TensoriumPluginAction : public PluginASTAction {
 
 } // namespace
 
-/// @brief Register the plugin under the name "tensorium-dispatch"
-static FrontendPluginRegistry::Add<TensoriumPluginAction> X("tensorium-dispatch",
-                                                            "Handle #pragma tensorium directives");
+// @brief Register the plugin under the name "tensorium-dispatch"
+// Important: we avoid double-free issues by wrapping in a ManagedStatic,
+
+#include "clang/Frontend/FrontendPluginRegistry.h"
+
+// …
+
+using clang::FrontendPluginRegistry;
+
+static FrontendPluginRegistry::Add<TensoriumPluginAction>
+    X("tensorium-dispatch", "Handle #pragma tensorium directives");
+
+

Plugins/TensoriumDispatchPlugin.cpp

@@ -1,6 +1,6 @@
-#include "Tensorium/Tensorium.hpp"
+#include "../includes/Tensorium/Tensorium.hpp"
 #include <iostream>
-#include <omp.h>
+// #include <omp.h>
 #include <pybind11/numpy.h>
 #include <pybind11/pybind11.h>
 #include <pybind11/stl.h>

Pybind/bindings.cpp

@@ -1,4 +1,3 @@
-
 file(GLOB TEST_SOURCES
     ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/Matrix/*.cpp
@@ -10,7 +9,7 @@ file(GLOB TEST_SOURCES
 
 add_executable(TensoriumTests ${TEST_SOURCES})
 
-# --- OpenMP flags selon architecture ---
+# ─────────────────────────────── OpenMP flags ───────────────────────────────
 if(APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
   include_directories(${OPENMP_INCLUDE_PATH})
   link_directories(${OPENMP_LIB_PATH})
@@ -20,7 +19,7 @@ elseif(DEFINED OPENMP_LIB)
   target_link_libraries(TensoriumTests PRIVATE ${OPENMP_LIB})
 endif()
 
-# --- MPI / KNL ---
+# ─────────────────────────────── MPI / KNL options ───────────────────────────────
 if(USE_MPI)
   find_package(MPI REQUIRED)
   target_link_libraries(TensoriumTests PRIVATE MPI::MPI_CXX)
@@ -33,3 +32,12 @@ if(USE_KNL)
   target_link_libraries(TensoriumTests PRIVATE memkind)
 endif()
 
+# # ─────────────────────────────── Inject the plugins ───────────────────────────────
+# set(DISPATCH_PLUGIN_PATH "${CMAKE_BINARY_DIR}/Plugins/libTensoriumDispatchPlugin.so")
+#
+# add_dependencies(TensoriumTests TensoriumDispatchPlugin)
+#
+# # inject the clang plugin (frontend)
+# set_property(TARGET TensoriumTests APPEND_STRING PROPERTY COMPILE_FLAGS
+#     " -Xclang -load -Xclang ${DISPATCH_PLUGIN_PATH} -Xclang -add-plugin -Xclang tensorium-dispatch"
+# )

Tests/CMakeLists.txt

@@ -31,7 +31,7 @@ tensorium::Matrix<K> mul_mat_reference(const tensorium::Matrix<K>& A, const tens
 }
 int matrix_bench() {
 	using namespace tensorium;
-	std::vector<std::size_t> sizes = {256};
+	std::vector<std::size_t> sizes = {8192};
 
 	for (std::size_t N : sizes) {
 		Matrix<double> A(N, N);

Tests/Matrix/Matrix.cpp

@@ -9,7 +9,7 @@
 #include <cstdlib>
 #include <cstring>
 #include <algorithm>
-#include <omp.h>
+// #include <omp.h>
 #include <iomanip>
 #include <random>