Fix #9935: Use more selectivity when building SSE specific code

src/CMakeLists.txt

@@ -29,11 +29,6 @@ add_files(
     viewport_sprite_sorter_sse4.cpp
     CONDITION SSE_FOUND
 )
-if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang")
-    set_compile_flags(
-        viewport_sprite_sorter_sse4.cpp
-        COMPILE_FLAGS -msse4.1)
-endif()
 
 add_files(
     aircraft.h

src/blitter/32bpp_anim_sse2.cpp

@@ -19,6 +19,7 @@
 /** Instantiation of the partially SSSE2 32bpp with animation blitter factory. */
 static FBlitter_32bppSSE2_Anim iFBlitter_32bppSSE2_Anim;
 
+GNU_TARGET("sse2")
 void Blitter_32bppSSE2_Anim::PaletteAnimate(const Palette &palette)
 {
 	assert(!_screen_disable_anim);

src/blitter/32bpp_anim_sse2.hpp

@@ -16,6 +16,10 @@
 #define SSE_VERSION 2
 #endif
 
+#ifndef SSE_TARGET
+#define SSE_TARGET "sse2"
+#endif
+
 #ifndef FULL_ANIMATION
 #define FULL_ANIMATION 1
 #endif

src/blitter/32bpp_anim_sse4.cpp

@@ -29,6 +29,7 @@ static FBlitter_32bppSSE4_Anim iFBlitter_32bppSSE4_Anim;
  */
 IGNORE_UNINITIALIZED_WARNING_START
 template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last, bool translucent, bool animated>
+GNU_TARGET("sse4.1")
 inline void Blitter_32bppSSE4_Anim::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
 {
 	const byte * const remap = bp->remap;

src/blitter/32bpp_anim_sse4.hpp

@@ -16,6 +16,10 @@
 #define SSE_VERSION 4
 #endif
 
+#ifndef SSE_TARGET
+#define SSE_TARGET "sse4.1"
+#endif
+
 #ifndef FULL_ANIMATION
 #define FULL_ANIMATION 1
 #endif

src/blitter/32bpp_sse2.hpp

@@ -16,6 +16,10 @@
 #define SSE_VERSION 2
 #endif
 
+#ifndef SSE_TARGET
+#define SSE_TARGET "sse2"
+#endif
+
 #ifndef FULL_ANIMATION
 #define FULL_ANIMATION 0
 #endif

src/blitter/32bpp_sse4.hpp

@@ -16,6 +16,10 @@
 #define SSE_VERSION 4
 #endif
 
+#ifndef SSE_TARGET
+#define SSE_TARGET "sse4.1"
+#endif
+
 #ifndef FULL_ANIMATION
 #define FULL_ANIMATION 0
 #endif

src/blitter/32bpp_sse_func.hpp

@@ -12,6 +12,7 @@
 
 #ifdef WITH_SSE
 
+GNU_TARGET(SSE_TARGET)
 static inline void InsertFirstUint32(const uint32 value, __m128i &into)
 {
 #if (SSE_VERSION >= 4)
@@ -22,6 +23,7 @@ static inline void InsertFirstUint32(const uint32 value, __m128i &into)
 #endif
 }
 
+GNU_TARGET(SSE_TARGET)
 static inline void InsertSecondUint32(const uint32 value, __m128i &into)
 {
 #if (SSE_VERSION >= 4)
@@ -32,6 +34,7 @@ static inline void InsertSecondUint32(const uint32 value, __m128i &into)
 #endif
 }
 
+GNU_TARGET(SSE_TARGET)
 static inline void LoadUint64(const uint64 value, __m128i &into)
 {
 #ifdef POINTER_IS_64BIT
@@ -46,6 +49,7 @@ static inline void LoadUint64(const uint64 value, __m128i &into)
 #endif
 }
 
+GNU_TARGET(SSE_TARGET)
 static inline __m128i PackUnsaturated(__m128i from, const __m128i &mask)
 {
 #if (SSE_VERSION == 2)
@@ -56,6 +60,7 @@ static inline __m128i PackUnsaturated(__m128i from, const __m128i &mask)
 #endif
 }
 
+GNU_TARGET(SSE_TARGET)
 static inline __m128i DistributeAlpha(const __m128i from, const __m128i &mask)
 {
 #if (SSE_VERSION == 2)
@@ -66,6 +71,7 @@ static inline __m128i DistributeAlpha(const __m128i from, const __m128i &mask)
 #endif
 }
 
+GNU_TARGET(SSE_TARGET)
 static inline __m128i AlphaBlendTwoPixels(__m128i src, __m128i dst, const __m128i &distribution_mask, const __m128i &pack_mask)
 {
 	__m128i srcAB = _mm_unpacklo_epi8(src, _mm_setzero_si128());   // PUNPCKLBW, expand each uint8 into uint16
@@ -86,6 +92,7 @@ static inline __m128i AlphaBlendTwoPixels(__m128i src, __m128i dst, const __m128
 /* Darken 2 pixels.
  * rgb = rgb * ((256/4) * 4 - (alpha/4)) / ((256/4) * 4)
  */
+GNU_TARGET(SSE_TARGET)
 static inline __m128i DarkenTwoPixels(__m128i src, __m128i dst, const __m128i &distribution_mask, const __m128i &tr_nom_base)
 {
 	__m128i srcAB = _mm_unpacklo_epi8(src, _mm_setzero_si128());
@@ -99,6 +106,7 @@ static inline __m128i DarkenTwoPixels(__m128i src, __m128i dst, const __m128i &d
 }
 
 IGNORE_UNINITIALIZED_WARNING_START
+GNU_TARGET(SSE_TARGET)
 static Colour ReallyAdjustBrightness(Colour colour, uint8 brightness)
 {
 	uint64 c16 = colour.b | (uint64) colour.g << 16 | (uint64) colour.r << 32;
@@ -141,6 +149,7 @@ static inline Colour AdjustBrightneSSE(Colour colour, uint8 brightness)
 	return ReallyAdjustBrightness(colour, brightness);
 }
 
+GNU_TARGET(SSE_TARGET)
 static inline __m128i AdjustBrightnessOfTwoPixels(__m128i from, uint32 brightness)
 {
 #if (SSE_VERSION < 3)
@@ -192,6 +201,7 @@ static inline __m128i AdjustBrightnessOfTwoPixels(__m128i from, uint32 brightnes
  */
 IGNORE_UNINITIALIZED_WARNING_START
 template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last, bool translucent>
+GNU_TARGET(SSE_TARGET)
 #if (SSE_VERSION == 2)
 inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
 #elif (SSE_VERSION == 3)

src/blitter/32bpp_ssse3.hpp

@@ -16,6 +16,10 @@
 #define SSE_VERSION 3
 #endif
 
+#ifndef SSE_TARGET
+#define SSE_TARGET "ssse3"
+#endif
+
 #ifndef FULL_ANIMATION
 #define FULL_ANIMATION 0
 #endif

src/blitter/CMakeLists.txt

@@ -38,21 +38,6 @@ add_files(
     CONDITION NOT OPTION_DEDICATED AND OPENGL_FOUND
 )
 
-
-if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang")
-    set_compile_flags(
-        32bpp_anim_sse2.cpp
-        32bpp_sse2.cpp
-        COMPILE_FLAGS -msse2)
-    set_compile_flags(
-        32bpp_ssse3.cpp
-        COMPILE_FLAGS -mssse3)
-    set_compile_flags(
-        32bpp_anim_sse4.cpp
-        32bpp_sse4.cpp
-        COMPILE_FLAGS -msse4.1)
-endif()
-
 add_files(
     base.hpp
     common.hpp

src/stdafx.h

@@ -418,11 +418,13 @@ static_assert(SIZE_MAX >= UINT32_MAX);
 #endif /* __APPLE__ */
 
 #if defined(__GNUC__) || defined(__clang__)
-#	define likely(x)   __builtin_expect(!!(x), 1)
+#	define likely(x)     __builtin_expect(!!(x), 1)
-#	define unlikely(x) __builtin_expect(!!(x), 0)
+#	define unlikely(x)   __builtin_expect(!!(x), 0)
+#	define GNU_TARGET(x) [[gnu::target(x)]]
 #else
-#	define likely(x)   (x)
+#	define likely(x)     (x)
-#	define unlikely(x) (x)
+#	define unlikely(x)   (x)
+#	define GNU_TARGET(x)
 #endif /* __GNUC__ || __clang__ */
 
 /* For the FMT library we only want to use the headers, not link to some library. */

src/viewport_sprite_sorter_sse4.cpp

@@ -26,6 +26,7 @@
 #	define LOAD_128 _mm_loadu_si128
 #endif
 
+GNU_TARGET("sse4.1")
 void ViewportSortParentSpritesSSE41(ParentSpriteToSortVector *psdv)
 {
 	if (psdv->size() < 2) return;