Merge branch 'main' into lluo/wheel_release

Author: lanluo-nvidia

mlir-tensorrt/.clang-format-ignore

@@ -0,0 +1 @@
+**/*.td

mlir-tensorrt/CMakeLists.txt

@@ -7,6 +7,7 @@ include(MTRTCMakePolicy NO_POLICY_SCOPE)
 include(CMakeDependentOption)
 include(CMakePrintHelpers)
 include(MTRTCPM)
+include(MTRTCMakeExtras)
 include(MTRTDependencies)
 include(MTRTFeatures)
 include(AddMLIRTensorRT)

mlir-tensorrt/CMakeOptions.cmake

@@ -39,9 +39,11 @@ mtrt_option(MLIR_TRT_RELATIVE_DEBUG_PATHS
 # whereas the definition of the source groups belonging to those packages
 # are distributed across multiple sub-directories.
 set(MLIR_TRT_ENABLE_PROJECTS_DEFAULT
-  executor tensorrt compiler
+  executor tensorrt
 )
 
+list(APPEND MLIR_TRT_ENABLE_PROJECTS_DEFAULT "compiler")
+
 if(MLIR_TRT_ENABLE_PYTHON)
   list(APPEND MLIR_TRT_ENABLE_PROJECTS_DEFAULT "integrations/python")
 endif()
@@ -74,3 +76,4 @@ if(MLIR_TRT_ENABLE_NCCL)
 else()
   set(MLIR_TRT_NCCL_TARGET "" CACHE INTERNAL "")
 endif()
+

mlir-tensorrt/Version.cmake

@@ -3,3 +3,11 @@ set(MLIR_TENSORRT_VERSION_MINOR "4")
 set(MLIR_TENSORRT_VERSION_PATCH "3")
 set(MLIR_TENSORRT_VERSION
   "${MLIR_TENSORRT_VERSION_MAJOR}.${MLIR_TENSORRT_VERSION_MINOR}.${MLIR_TENSORRT_VERSION_PATCH}")
+
+# The source release script replaces the empty string with a actual
+# hash, so don't change the default here.
+set(MLIR_TENSORRT_GIT_HASH "")
+
+set(CPACK_PACKAGE_VERSION_MAJOR ${MLIR_TENSORRT_VERSION_MAJOR})
+set(CPACK_PACKAGE_VERSION_MINOR ${MLIR_TENSORRT_VERSION_MINOR})
+set(CPACK_PACKAGE_VERSION_PATCH ${MLIR_TENSORRT_VERSION_PATCH})

mlir-tensorrt/build_tools/cmake/MTRTCommonFunctions.cmake

@@ -332,6 +332,8 @@ function(mtrt_add_project_library name)
     add_mlir_library(${name} ${lib_type_args} DISABLE_INSTALL EXCLUDE_FROM_LIBMLIR ${ARG_UNPARSED_ARGUMENTS})
   endif()
 
+  mtrt_apply_extra_check_options("${name}")
+
   if(ARG_MLIR_LIBS)
     list(POP_FRONT ARG_MLIR_LIBS VISIBILITY)
     mtrt_target_link_mlir_libraries(${name} ${VISIBILITY} ${ARG_MLIR_LIBS})

mlir-tensorrt/build_tools/cmake/MTRTCompilationOptions.cmake

@@ -1,42 +1,74 @@
+
+# Force-set internal cache variable "MLIR_TRT_APPLY_EXTRA_CHECKS" to
+# the result of the expression
+# (MLIR_TRT_ENABLE_EXTRA_CHECKS AND (CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")).
+# This is done to avoid repeated checks when we populate target compilation options.
+set(MLIR_TRT_APPLY_EXTRA_CHECKS FALSE CACHE INTERNAL "")
+if(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang" AND MLIR_TRT_ENABLE_EXTRA_CHECKS)
+  set(MLIR_TRT_APPLY_EXTRA_CHECKS TRUE CACHE INTERNAL "")
+endif()
+
 if(MLIR_TRT_ENABLE_WERROR)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror")
   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror")
   set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} -Wno-error")
 endif()
 
-function(append_c_cxx_flags)
+# ------------------------------------------------------------------------------
+# Applies additional compilation options to a C/CXX library or executable
+# target. This only has an effect if MLIR_TRT_ENABLE_EXTRA_CHECKS is ON.
+# ------------------------------------------------------------------------------
+
+macro(mtrt_apply_extra_checks_to_target_helper target)
+  target_compile_options(${target} PRIVATE
+    -Wmissing-declarations
+    -Wmissing-prototypes
+    -Wunused
+    -fstrict-flex-arrays=3
+    )
+endmacro()
+
+macro(mtrt_apply_extra_check_options target)
+  if(MLIR_TRT_APPLY_EXTRA_CHECKS)
+    if(TARGET "obj.${target}")
+      mtrt_apply_extra_checks_to_target_helper("obj.${target}")
+    else()
+      mtrt_apply_extra_checks_to_target_helper("${target}")
+    endif()
+  endif()
+endmacro()
+
+# ------------------------------------------------------------------------------
+# Updates global C/CXX library or executable compilation flags.
+# This only has an effect if MLIR_TRT_ENABLE_EXTRA_CHECKS is ON.
+#
+# Some flags enable additional warnings, others may enable additional runtime
+# safety or preconditions checks.
+#
+# For a description of each flag, see the below reference:
+# https://best.openssf.org/Compiler-Hardening-Guides/Compiler-Options-Hardening-Guide-for-C-and-C++.html
+# ------------------------------------------------------------------------------
+function(mtrt_update_global_c_cxx_flags)
   string(JOIN " " flags ${ARGN})
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${flags}" PARENT_SCOPE)
   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${flags}" PARENT_SCOPE)
 endfunction()
 
-# Enable additional warnings and runtime checks recommended
-# by https://best.openssf.org/Compiler-Hardening-Guides/Compiler-Options-Hardening-Guide-for-C-and-C++.html
-if(MLIR_TRT_ENABLE_EXTRA_CHECKS AND
-   (CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang"))
+if(MLIR_TRT_APPLY_EXTRA_CHECKS)
   # FORTIFY_SOURCE=3 requires O1 or higher.
-  # GCC has default FORTIFY_SOURCE=2.
-  if(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND
-     (NOT CMAKE_BUILD_TYPE STREQUAL "Debug"))
+  # GCC has default FORTIFY_SOURCE=2 and would require `-U_FORTIFY_SOURCE`
+  # to change the value, so we just set it to 3 when using clang.
+  if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug"
+     AND CMAKE_CXX_COMPILER_ID MATCHES "Clang")
     add_compile_definitions(
       _FORTIFY_SOURCE=3
     )
   endif()
 
-  # Note: '-D_GLIBCXX_ASSERTIONS' already added by
-  # the 'HandleLLVMOptions'.
-
-  append_c_cxx_flags(
-    # Enforce that all global functions have declarations. This catches
-    # missing `static` qualifiers on functions which are intended to be TU-local.
-    -Wmissing-declarations
-    -Wmissing-prototypes
-    -Wunused
+  mtrt_update_global_c_cxx_flags(
     # Make template instantiation errors more readable.
     -fdiagnostics-show-template-tree
     # Enable better bounds checking for trailing array members.
-    # Value is 0 to 3, with 3 being the most strict. LLVM has some code that
-    # only allows us to use 1 here.
     -fstrict-flex-arrays=1
     # Enable runtime checks for variable-size stack allocation validity.
     -fstack-clash-protection
@@ -48,25 +80,61 @@ if(MLIR_TRT_ENABLE_EXTRA_CHECKS AND
     -Werror=int-conversion
   )
 
-  if(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR
-     CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
-    append_c_cxx_flags(
+  if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64")
+    mtrt_update_global_c_cxx_flags(
       -fcf-protection=full
-    )
-  endif()
-
-  if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-    append_c_cxx_flags(
       -Wself-assign
     )
   endif()
 endif()
 
-# The CUDA headers will trigger excessive "C++20 extensions warning" with
-# clang19+ in macro usage.
-include(CheckCXXCompilerFlag)
-check_cxx_compiler_flag("-Wc++20-extensions"
-  CXX_SUPPORTS_C20_EXTENSIONS_FLAG)
-if(CXX_SUPPORTS_C20_EXTENSIONS_FLAG)
-  add_compile_options(-Wno-c++20-extensions)
+# This ensures that `__FILE__` in log printing expands to a relative path and file paths
+# in debug info will expand relative to the source or binary directories.
+#
+#
+# It adds the following global compilation flag:
+#   '-fno-canonical-prefixes'
+#   '-ffile-prefix-map=[project_src_root]=.'
+#   '-fdebug-prefix-map=[project_src_root]=.'
+# In addition, this flag is added if the build directory is not a subdirectory of the source directory:
+#   '-ffile-prefix-map=[project_binary_root]=build'.
+#   '-fdebug-prefix-map=[project_binary_root]=build'.
+#
+# This is only applied if the project is the top-level project and MLIR_TRT_RELATIVE_DEBUG_PATHS is ON.
+function(mtrt_apply_relative_path_options)
+  if(NOT COMMAND mtrt_append_compiler_flag_if_supported)
+    message(FATAL_ERROR "MTRTCompilationOptions must be included after MTRTCMakeExtras")
+  endif()
+
+  # In CMake, you can't know where the "cmake" command is invoked from, so we
+  # just assume that we want make paths relative to the source directory.
+  set(relative_src ".")
+
+  # The following checks if the source directory is a prefix of the binary directory.
+  cmake_path(IS_PREFIX CMAKE_SOURCE_DIR "${CMAKE_BINARY_DIR}" src_dir_is_prefix_of_bin_dir)
+
+  set(debug_remaps "-fdebug-prefix-map=${CMAKE_SOURCE_DIR}=.")
+  set(file_remaps "-ffile-prefix-map=${CMAKE_SOURCE_DIR}=.")
+  if(src_dir_is_prefix_of_bin_dir)
+    set(debug_remaps "${debug_remaps} -fdebug-prefix-map=${CMAKE_BINARY_DIR}=build")
+    set(file_remaps "${file_remaps} -ffile-prefix-map=${CMAKE_BINARY_DIR}=build")
+  endif()
+
+  mtrt_append_compiler_flag_if_supported(
+    CHECK "-fdebug-prefix-map=foo=bar"
+    APPEND "${debug_remaps}"
+  )
+
+  mtrt_append_compiler_flag_if_supported(
+    CHECK "-ffile-prefix-map=foo=bar"
+    APPEND "${file_remaps}"
+  )
+  mtrt_append_compiler_flag_if_supported(
+    CHECK "-no-canonical-prefixes"
+    APPEND "-no-canonical-prefixes"
+  )
+endfunction()
+
+if(PROJECT_IS_TOP_LEVEL AND MLIR_TRT_RELATIVE_DEBUG_PATHS)
+  mtrt_apply_relative_path_options()
 endif()

mlir-tensorrt/build_tools/cmake/MTRTRelativeDebugPaths.cmake

@@ -50,12 +50,10 @@ mtrt_add_project_targets(MLIRTensorRT
   MLIRTransformDialectTransforms
   )
 
-
 add_subdirectory(include)
 add_subdirectory(lib)
 add_subdirectory(test)
 
-
 # Tools must be sequenced after tests.
 add_subdirectory(tools)
 
@@ -69,7 +67,6 @@ set(MTRT_COMPILER_HEADER_DIRS
   "${PROJECT_BINARY_DIR}/include/mlir-tensorrt"
 )
 
-
 mtrt_add_header_installation_components(mtrt-headers
   DIRECTORIES
    ${MTRT_COMPILER_HEADER_DIRS}

mlir-tensorrt/compiler/CMakeLists.txt

@@ -41,6 +41,7 @@ def ConvertStablehloToPlanPass : Pass<"convert-stablehlo-to-plan", "::mlir::Modu
       calls.
   }];
   let dependentDialects = [
+    "::mlir::cf::ControlFlowDialect",
     "::mlir::plan::PlanDialect",
     "::mlir::executor::ExecutorDialect",
     "::mlir::tensor::TensorDialect"

mlir-tensorrt/compiler/include/mlir-tensorrt/Conversion/Passes.td

@@ -1,6 +1,5 @@
 add_subdirectory(Analysis)
 
-
 add_subdirectory(Utils)
 add_subdirectory(Interfaces)
 add_subdirectory(Dialect)