Build: Turn off floating-point contraction for C++

Clang and GCC fuse a * b + c into one rounding on CPUs that have fused
multiply-add, which on ARM64 made the interpreter's dot products, vavg,
vdet and vqmul round differently from x86-64. Fused operations now only
happen where the code asks for them. MSVC doesn't contract by default.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

CMakeLists.txt

@@ -496,6 +496,9 @@ if(NOT MSVC)
 
 	add_compile_options(-fno-math-errno)
 
+	# No fused multiply-adds unless written explicitly, so float results are the same on every CPU.
+	add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:-ffp-contract=off>")
+
 	if(X86)
 		# enable sse2 code generation (enabled by default on X86_64)
 		if(NOT MACOSX)

android/jni/Locals.mk

@@ -4,7 +4,7 @@
 LOCAL_CFLAGS := -DUSE_FFMPEG -DWITH_UPNP -DUSING_GLES2 -DZ7_CRC_NUM_TABLES=1 -DMOBILE_DEVICE -O3 -fsigned-char -Wall -Wno-multichar -Wno-unused-variable -fno-strict-aliasing -D__STDC_CONSTANT_MACROS -DSPIRV_CROSS_EXCEPTIONS_TO_ASSERTIONS
 # yes, it's really CPPFLAGS for C++
 # deprecated-register is generated by Android default code and causes noise.
-LOCAL_CPPFLAGS := -fexceptions -std=c++17 -frtti -Wno-reorder -Wno-deprecated-register -Wno-nullability-completeness
+LOCAL_CPPFLAGS := -fexceptions -std=c++17 -ffp-contract=off -frtti -Wno-reorder -Wno-deprecated-register -Wno-nullability-completeness
 LOCAL_C_INCLUDES := \
   $(LOCAL_PATH)/../../Common \
   $(LOCAL_PATH)/../.. \

libretro/Makefile

@@ -448,7 +448,8 @@ else
 endif
 
 ifeq (,$(findstring msvc,$(platform)))
-   CXXFLAGS += -std=c++17
+   # No fused multiply-adds unless written explicitly, so float results are the same on every CPU.
+   CXXFLAGS += -std=c++17 -ffp-contract=off
 else
    CXXFLAGS += -std:c++17
 endif