optimized compiler flags, spreader is now fast as GCC-13 with other compilers

Author: DiamonDinoia

CMakeLists.txt

@@ -4,10 +4,27 @@ project(finufft VERSION 2.2.0 LANGUAGES C CXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
+
 set(GNU_LIKE_FRONTENDS AppleClang Clang GNU)
 if (CMAKE_CXX_COMPILER_ID IN_LIST GNU_LIKE_FRONTENDS)
     # Set custom compiler flags for gcc-compatible compilers
-    set(FINUFFT_CXX_FLAGS_RELEASE -O3 -funroll-loops -ffp-contract=fast)
+    set(FINUFFT_CXX_FLAGS_RELEASE -O3
+            -funroll-loops
+            -ffp-contract=fast
+            -fexcess-precision=fast
+            -fno-math-errno
+            -fno-signed-zeros
+            -fno-trapping-math
+            -fassociative-math
+            -freciprocal-math
+            # These flags make the spreader 20% faster in 2D and up to 100% faster in 3D with no loss of accuracy.
+            # Why no -fast-math or -funsafe-math-optimizations ?
+            # It breaks the code, and influences other software that depend on finufft bu changing the rounding mode.
+            # GCC-13 recently fixed this issue:
+            # https://github.com/llvm/llvm-project/issues/57589
+            # https://gcc.gnu.org/gcc-13/changes.html
+            # https://trofi.github.io/posts/302-Ofast-and-ffast-math-non-local-effects.html
+    )
     set(FINUFFT_CXX_FLAGS_RELWITHDEBINFO -g ${FINUFFT_CXX_FLAGS_RELEASE})
 endif ()