GCC: inline limit (AMReX-Codes#3841)

By default, set GCC inline limit to a huge number (43210). GCC's default value is version dependent. It appears to be smaller than 2000 in recent versions. The motivation for setting this to a huge number is that GCC especially the recent versions (e.g., 12) no longer inline big lambda functions passed to ParallelFor, resulting in serious performance issues.
ajnonaka · Apr 1, 2024 · 728eec4 · 728eec4
1 parent bb7d5cf
commit 728eec4
Show file tree

Hide file tree

Showing 9 changed files with 50 additions and 3 deletions.
diff --git a/Docs/sphinx_documentation/source/BuildingAMReX.rst b/Docs/sphinx_documentation/source/BuildingAMReX.rst
@@ -535,6 +535,12 @@ The list of available options is reported in the :ref:`table <tab:cmakevar>` bel
    | AMReX_FLATTEN_FOR            |  Enable flattening of ParallelFor and similar   | NO                      | YES, NO               |
    |                              |  functions for host code                        |                         |                       |
    +------------------------------+-------------------------------------------------+-------------------------+-----------------------+
+   | AMReX_COMPILER_DEFAULT_INLINE|  Use default inline behavior of compiler,       | NO for GCC              | YES, NO               |
+   |                              |  so far relevant for GCC Only                   | YES otherwise           |                       |
+   +------------------------------+-------------------------------------------------+-------------------------+-----------------------+
+   | AMReX_INLINE_LIMIT           |  Inline limit. Relevant only when               | 43210                   | Non-negative number   |
+   |                              |  AMReX_COMPILER_DEFAULT_INLINE is NO.           |                         |                       |
+   +------------------------------+-------------------------------------------------+-------------------------+-----------------------+
 .. raw:: latex
 
    \end{center}

diff --git a/Src/CMakeLists.txt b/Src/CMakeLists.txt
@@ -86,6 +86,14 @@ foreach(D IN LISTS AMReX_SPACEDIM)
           $<BUILD_INTERFACE:Flags_FPE>
           )
     endif ()
+
+    if (NOT AMReX_COMPILER_DEFULT_INLINE)
+       target_link_libraries(amrex_${D}d
+          PUBLIC
+          $<BUILD_INTERFACE:Flags_INLINE>
+          )
+    endif ()
+
 endforeach()
 
 # General configuration

diff --git a/Tools/CMake/AMReXClangTidy.cmake b/Tools/CMake/AMReXClangTidy.cmake
@@ -17,7 +17,7 @@ macro(setup_clang_tidy)
 
       # Need --extra-arg to suppress warnings like clang-diagnostic-unknown-warning-option
       # when GCC is used.
-      set(AMReX_CLANG_TIDY_COMMAND "${AMReX_CLANG_TIDY_EXE};--extra-arg=-Wno-unknown-warning-option")
+      set(AMReX_CLANG_TIDY_COMMAND "${AMReX_CLANG_TIDY_EXE};--extra-arg=-Wno-unknown-warning-option;--extra-arg=-Wno-ignored-optimization-argument")
       if (AMReX_CLANG_TIDY_CONFIG_FILE_NAME)
          set(AMReX_CLANG_TIDY_COMMAND "${AMReX_CLANG_TIDY_COMMAND}"
              "--config-file=${AMReX_CLANG_TIDY_CONFIG_FILE_NAME}")

diff --git a/Tools/CMake/AMReXConfig.cmake.in b/Tools/CMake/AMReXConfig.cmake.in
@@ -91,6 +91,8 @@ set(AMReX_FPE_FOUND                 @AMReX_FPE@)
 set(AMReX_PIC_FOUND                 @AMReX_PIC@)
 set(AMReX_ASSERTIONS_FOUND          @AMReX_ASSERTIONS@)
 set(AMReX_FLATTEN_FOR_FOUND         @AMReX_FLATTEN_FOR@)
+set(AMReX_COMPILER_DEFAULT_INLINE_FOUND   @AMReX_COMPILER_DEFAULT_INLINE@)
+set(AMReX_INLINE_LIMIT_FOUND              @AMReX_INLINE_LIMIT@)
 
 # Profiling options
 set(AMReX_BASEP_FOUND               @AMReX_BASE_PROFILE@)

diff --git a/Tools/CMake/AMReXFlagsTargets.cmake b/Tools/CMake/AMReXFlagsTargets.cmake
@@ -5,6 +5,7 @@
 #   Flags_CXX                 --> Optional flags for C++ code
 #   Flags_Fortran             --> Optional flags for Fortran code
 #   Flags_FPE                 --> Floating-Point Exception flags for both C++ and Fortran
+#   Flags_INLINE              --> Optional flags for inlining
 #
 # These INTERFACE targets can be added to the AMReX export set.
 #
@@ -93,6 +94,17 @@ target_compile_options( Flags_CXX
    $<${_cxx_intelllvm_rel}:>
    )
 
+add_library(Flags_INLINE INTERFACE)
+add_library(AMReX::Flags_INLINE ALIAS Flags_INLINE)
+
+if (NOT AMReX_COMPILER_DEFAULT_INLINE)
+   target_compile_options( Flags_INLINE
+      INTERFACE
+      $<${_cxx_gnu_rwdbg}:-finline-limit=${AMReX_INLINE_LIMIT}>
+      $<${_cxx_gnu_rel}:-finline-limit=${AMReX_INLINE_LIMIT}>
+      )
+endif ()
+
 #
 # Fortran flags
 #

diff --git a/Tools/CMake/AMReXOptions.cmake b/Tools/CMake/AMReXOptions.cmake
@@ -369,6 +369,20 @@ print_option( AMReX_IPO )
 option(AMReX_FPE "Enable Floating Point Exceptions checks" OFF)
 print_option( AMReX_FPE )
 
+if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+   option(AMReX_COMPILER_DEFAULT_INLINE "Use compiler default inline behavior" OFF)
+   set(AMReX_INLINE_LIMIT 43210 CACHE STRING "Inline limit")
+   if (NOT AMReX_COMPILER_DEFAULT_INLINE)
+      if (AMReX_INLINE_LIMIT LESS 0)
+         message(FATAL_ERROR "AMReX_INLINE_LIMIT, if set, must be non-negative")
+      endif()
+      message(STATUS "   AMReX_INLINE_LIMIT = ${AMReX_INLINE_LIMIT}")
+   endif ()
+else ()
+   set(AMReX_COMPILER_DEFAULT_INLINE ON)
+endif ()
+
+
 if ( "${CMAKE_BUILD_TYPE}" MATCHES "Debug" )
    option( AMReX_ASSERTIONS "Enable assertions" ON)
 else ()

diff --git a/Tools/C_scripts/mmclt.py b/Tools/C_scripts/mmclt.py
@@ -28,7 +28,7 @@ def mmclt(argv):
     fout = open(args.output, "w")
 
     fout.write("CLANG_TIDY ?= clang-tidy\n")
-    fout.write("override CLANG_TIDY_ARGS += --extra-arg=-Wno-unknown-warning-option --extra-arg-before=--driver-mode=g++\n")
+    fout.write("override CLANG_TIDY_ARGS += --extra-arg=-Wno-unknown-warning-option --extra-arg=-Wno-ignored-optimization-argument --extra-arg-before=--driver-mode=g++\n")
     fout.write("\n")
 
     fout.write(".SECONDEXPANSION:\n")

diff --git a/Tools/GNUMake/comps/gnu.mak b/Tools/GNUMake/comps/gnu.mak
@@ -44,6 +44,8 @@ gcc_major_ge_10 = $(shell expr $(gcc_major_version) \>= 10)
 gcc_major_ge_11 = $(shell expr $(gcc_major_version) \>= 11)
 gcc_major_ge_12 = $(shell expr $(gcc_major_version) \>= 12)
 
+INLINE_LIMIT ?= 43210
+
 ifneq ($(NO_CONFIG_CHECKING),TRUE)
 ifneq ($(gcc_major_ge_8),1)
   $(error GCC < 8 not supported)
@@ -96,6 +98,9 @@ else
     CXXFLAGS += -g -O3
     CFLAGS   += -g -O3
   endif
+  ifneq ($(USE_COMPILER_DEFAULT_INLINE),TRUE)
+    CXXFLAGS += -finline-limit=$(INLINE_LIMIT)
+  endif
 endif
 
 ifeq ($(WARN_ALL),TRUE)

diff --git a/Tools/GNUMake/tools/Make.clang-tidy b/Tools/GNUMake/tools/Make.clang-tidy
@@ -1,6 +1,6 @@
 
 CLANG_TIDY = clang-tidy
-CLANG_TIDY_ARGS = --extra-arg=-Wno-unknown-warning-option --extra-arg-before=--driver-mode=g++
+CLANG_TIDY_ARGS = --extra-arg=-Wno-unknown-warning-option --extra-arg=-Wno-ignored-optimization-argument --extra-arg-before=--driver-mode=g++
 # space-separated list of paths to skip, matched against the absolute path of each source file
 CLANG_TIDY_IGNORE_SOURCES ?=