Add interprocedural optimization option (#4457)

Summary:
Pull Request resolved: #4457

X-link: facebookresearch/FBGEMM#1515

Add IPO option to enable link time optimization, which is disabled by default. AVX files use inline asm code and can't be parsed by lto implementations, so they are excluded. CUDA code is also excluded.

This optimization has reduced libfbgemm.a from 80Mb to 64Mb on my host, using GCC 15.

We have attempts to enable link time optimization on PyTorch, fixing FBGEMM builds is the important prerequisite.

Pull Request resolved: #4452

Reviewed By: cthi

Differential Revision: D77912148

Pulled By: q10

fbshipit-source-id: 8ad52ad4c556020e69a39166c867248024a04917

Author: cyyever

CMakeLists.txt

@@ -105,6 +105,7 @@ option(FBGEMM_BUILD_TESTS "Build fbgemm unit tests" ON)
 option(FBGEMM_BUILD_BENCHMARKS "Build fbgemm benchmarks" ON)
 option(FBGEMM_BUILD_DOCS "Build fbgemm documentation" OFF)
 option(FBGEMM_BUILD_FBGEMM_GPU "Build fbgemm_gpu library" OFF)
+option(FBGEMM_USE_IPO "Build fbgemm with interprocedural optimization" OFF)
 option(DISABLE_FBGEMM_AUTOVEC "Disable FBGEMM Autovec" OFF)
 
 if(FBGEMM_BUILD_TESTS)
@@ -420,3 +421,12 @@ endif()
 if(NOT HAVE_GNU_F2H_IEEE)
   add_definitions(-DMISSING_GNU_F2H_IEEE)
 endif()
+
+if(FBGEMM_USE_IPO)
+  include(CheckIPOSupported)
+  check_ipo_supported()
+  set_property(TARGET fbgemm_generic PROPERTY INTERPROCEDURAL_OPTIMIZATION TRUE)
+  set_property(TARGET fbgemm_avx2 PROPERTY INTERPROCEDURAL_OPTIMIZATION FALSE)
+  set_property(TARGET fbgemm_avx512 PROPERTY INTERPROCEDURAL_OPTIMIZATION FALSE)
+  set_property(TARGET fbgemm_autovec PROPERTY INTERPROCEDURAL_OPTIMIZATION TRUE)
+endif()