Set LTO option only if it's supported and the build is not a debug build (#1002)

Previously, the CMakeLists.txt files added `-flto` to some of the
compilation targets unconditionally. The addition should only be done if
the compiler supports LTO, and it shouldn't be done if the cmake
configuration is the debug configuration.

This PR moves the detection of availability of LTO to the top-level
CMakeLists.txt file, and defines a CMake macro to set it in the
sub-CMakeLists.txt files. The macro avoids repeating some gnarly CMake
syntax all over the place.

Author: mhucka

CMakeLists.txt

@@ -70,6 +70,21 @@ else()
     add_compile_options($<$<NOT:$<CONFIG:Debug>>:-O3>)
 endif()
 
+find_package(OpenMP REQUIRED)
+
+# Check for LTO support (which in CMake is lumped in with IPO).
+include(CheckIPOSupported)
+check_ipo_supported(RESULT HAVE_IPO)
+
+# Helper function for setting LTO flag (but only if the config is not debug).
+function(enable_lto target_name)
+    if(HAVE_IPO)
+        set_property(TARGET ${target_name} PROPERTY
+            INTERPROCEDURAL_OPTIMIZATION $<NOT:$<CONFIG:Debug>>
+        )
+    endif()
+endfunction()
+
 # Always build the basic part.
 add_subdirectory(pybind_interface/basic)
 add_subdirectory(pybind_interface/decide)

Makefile

@@ -48,18 +48,10 @@ HIPCCFLAGS := $(BASE_HIPCCFLAGS) $(HIPCCFLAGS)
 
 LTO_FLAGS := -flto=auto
 USING_CLANG := $(shell $(CXX) --version | grep -isq clang && echo "true")
-ifeq ($(USING_CLANG),"true")
+ifeq ($(USING_CLANG),true)
     LTO_FLAGS := -flto
 endif
 
-# Test if OpenMP header files are available and we can link with the library.
-OMP_CHECK_CMD := echo "int main() { return 0; }" | \
-    $(CXX) -fopenmp -x c++ - -o /dev/null 2>/dev/null
-HAVE_OPENMP := $(shell $(OMP_CHECK_CMD) && echo "true")
-ifeq ($(HAVE_OPENMP),true)
-    OPENMP_FLAGS := -fopenmp
-endif
-
 ifdef DEBUG
     DEBUG_FLAGS := -g -O0
     CXXFLAGS += $(DEBUG_FLAGS)

pybind_interface/avx2/CMakeLists.txt

@@ -17,8 +17,11 @@ project(qsim)
 
 if(MSVC)
     add_compile_options(/arch:AVX2 /openmp)
+    # Add /O2 to any configuration that is NOT Debug.
+    # This prevents a conflict with /RTC1 in DEBUG builds.
+    add_compile_options($<$<NOT:$<CONFIG:Debug>>:/O2>)
 else()
-    add_compile_options(-mavx2 -mfma -flto=auto)
+    add_compile_options(-mavx2 -mfma -O3)
     execute_process(
         COMMAND bash --noprofile -c "grep -qs bmi2 /proc/cpuinfo"
         RESULT_VARIABLE _EXIT_CODE
@@ -46,4 +49,6 @@ endif()
 include(../GetPybind11.cmake)
 pybind11_add_module(qsim_avx2 pybind_main_avx2.cpp)
 
-target_link_libraries(qsim_avx2 PRIVATE qsim_openmp_config)
+target_link_libraries(qsim_avx2 PUBLIC OpenMP::OpenMP_CXX)
+
+enable_lto(qsim_avx2)

pybind_interface/avx512/CMakeLists.txt

@@ -18,7 +18,7 @@ project(qsim)
 if(MSVC)
     add_compile_options(/arch:AVX512 /openmp)
 else()
-    add_compile_options(-mavx512f -mbmi2 -flto=auto)
+    add_compile_options(-mavx512f -mbmi2 -O3)
 endif()
 
 if(APPLE)
@@ -39,4 +39,6 @@ endif()
 include(../GetPybind11.cmake)
 pybind11_add_module(qsim_avx512 pybind_main_avx512.cpp)
 
-target_link_libraries(qsim_avx512 PRIVATE qsim_openmp_config)
+target_link_libraries(qsim_avx512 PUBLIC OpenMP::OpenMP_CXX)
+
+enable_lto(qsim_avx512)

pybind_interface/basic/CMakeLists.txt

@@ -18,7 +18,7 @@ project(qsim)
 if(MSVC)
     add_compile_options(/openmp)
 else()
-    add_compile_options(-flto=auto)
+    add_compile_options(-O3)
 endif()
 
 if(APPLE)
@@ -39,4 +39,6 @@ endif()
 include(../GetPybind11.cmake)
 pybind11_add_module(qsim_basic pybind_main_basic.cpp)
 
-target_link_libraries(qsim_basic PRIVATE qsim_openmp_config)
+target_link_libraries(qsim_basic PUBLIC OpenMP::OpenMP_CXX)
+
+enable_lto(qsim_basic)

pybind_interface/sse/CMakeLists.txt

@@ -18,7 +18,7 @@ project(qsim)
 if(MSVC)
     add_compile_options(/openmp)
 else()
-    add_compile_options(-msse4 -flto=auto)
+    add_compile_options(-msse4 -O3)
 endif()
 
 if(APPLE)
@@ -39,4 +39,6 @@ endif()
 include(../GetPybind11.cmake)
 pybind11_add_module(qsim_sse pybind_main_sse.cpp)
 
-target_link_libraries(qsim_sse PRIVATE qsim_openmp_config)
+target_link_libraries(qsim_sse PUBLIC OpenMP::OpenMP_CXX)
+
+enable_lto(qsim_sse)