Merge pull request #13421 from hrydgard/darkstalkers-opt

Darkstalkers minor optimizations

Common/ColorConv.h

@@ -66,25 +66,18 @@ inline u32 RGB565ToRGBA8888(u16 src) {
 }
 
 inline u16 RGBA8888ToRGB565(u32 value) {
-	u8 r = value & 0xFF;
-	u8 g = (value >> 8) & 0xFF;
-	u8 b = (value >> 16) & 0xFF;
-	r >>= 3;
-	g >>= 2;
-	b >>= 3;
-	return (u16)r | ((u16)g << 5) | ((u16)b << 11);
+	u32 r = (value >> 3) & 0x1F;
+	u32 g = (value >> 5) & (0x3F << 5);
+	u32 b = (value >> 8) & (0x1F << 11);
+	return (u16)(r | g | b);
 }
 
 inline u16 RGBA8888ToRGBA5551(u32 value) {
-	u8 r = value & 0xFF;
-	u8 g = (value >> 8) & 0xFF;
-	u8 b = (value >> 16) & 0xFF;
-	u8 a = (value >> 24) & 0xFF;
-	r >>= 3;
-	g >>= 3;
-	b >>= 3;
-	a >>= 7;
-	return (u16)r | ((u16)g << 5) | ((u16)b << 10) | ((u16)a << 15);
+	u32 r = (value >> 3) & 0x1F;
+	u32 g = (value >> 6) & (0x1F << 5);
+	u32 b = (value >> 9) & (0x1F << 10);
+	u32 a = (value >> 16) & 0x8000;
+	return (u16)(r | g | b | a);
 }
 
 inline u16 RGBA8888ToRGBA4444(u32 value) {

GPU/Common/VertexDecoderX86.cpp

@@ -167,7 +167,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 	BeginWrite();
 	const u8 *start = this->AlignCode16();
 
-#ifdef _M_IX86
+#if PPSSPP_ARCH(X86)
 	// Store register values
 	PUSH(ESI);
 	PUSH(EDI);
@@ -179,12 +179,20 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 	MOV(32, R(srcReg), MDisp(ESP, 16 + offset + 0));
 	MOV(32, R(dstReg), MDisp(ESP, 16 + offset + 4));
 	MOV(32, R(counterReg), MDisp(ESP, 16 + offset + 8));
+
+	const uint8_t STACK_FIXED_ALLOC = 64;
+#else
+	// Parameters automatically fall into place.
+
+	// This will align the stack properly to 16 bytes (the call of this function pushed RIP, which is 8 bytes).
+	const uint8_t STACK_FIXED_ALLOC = 64 + 8;
 #endif
 
+	// Allocate temporary storage on the stack.
+	SUB(PTRBITS, R(ESP), Imm8(STACK_FIXED_ALLOC));
 	// Save XMM4/XMM5 which apparently can be problematic?
 	// Actually, if they are, it must be a compiler bug because they SHOULD be ok.
 	// So I won't bother.
-	SUB(PTRBITS, R(ESP), Imm8(64));
 	MOVUPS(MDisp(ESP, 0), XMM4);
 	MOVUPS(MDisp(ESP, 16), XMM5);
 	MOVUPS(MDisp(ESP, 32), XMM6);
@@ -265,7 +273,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 	MOVUPS(XMM5, MDisp(ESP, 16));
 	MOVUPS(XMM6, MDisp(ESP, 32));
 	MOVUPS(XMM7, MDisp(ESP, 48));
-	ADD(PTRBITS, R(ESP), Imm8(64));
+	ADD(PTRBITS, R(ESP), Imm8(STACK_FIXED_ALLOC));
 
 #ifdef _M_IX86
 	// Restore register values

GPU/Software/RasterizerRectangle.cpp

@@ -31,18 +31,17 @@ extern bool currentDialogActive;
 namespace Rasterizer {
 
 // Through mode, with the specific Darkstalker settings.
-inline void DrawSinglePixel5551(u16 *pixel, const Vec4<int> &color_in) {
+inline void DrawSinglePixel5551(u16 *pixel, const u32 color_in) {
 	u32 new_color;
-	if (color_in.a() == 255) {
-		new_color = color_in.ToRGBA() & 0xFFFFFF;
+	if ((color_in >> 24) == 255) {
+		new_color = color_in & 0xFFFFFF;
 	} else {
 		const u32 old_color = RGBA5551ToRGBA8888(*pixel);
 		const Vec4<int> dst = Vec4<int>::FromRGBA(old_color);
-		Vec3<int> blended = AlphaBlendingResult(color_in, dst);
+		Vec3<int> blended = AlphaBlendingResult(Vec4<int>::FromRGBA(color_in), dst);
 		// ToRGB() always automatically clamps.
 		new_color = blended.ToRGB();
 	}
-
 	new_color |= (*pixel & 0x8000) ? 0xff000000 : 0x00000000;
 	*pixel = RGBA8888ToRGBA5551(new_color);
 }
@@ -148,7 +147,7 @@ void DrawSprite(const VertexData& v0, const VertexData& v1) {
 					for (int x = pos0.x; x < pos1.x; x++) {
 						u32 tex_color = nearestFunc(s, t, texptr, texbufw, 0);
 						if (tex_color & 0xFF000000) {
-							DrawSinglePixel5551(pixel, Vec4<int>::FromRGBA(tex_color));
+							DrawSinglePixel5551(pixel, tex_color);
 						}
 						s += ds;
 						pixel++;
@@ -159,7 +158,7 @@ void DrawSprite(const VertexData& v0, const VertexData& v1) {
 						Vec4<int> tex_color = Vec4<int>::FromRGBA(nearestFunc(s, t, texptr, texbufw, 0));
 						prim_color = ModulateRGBA(prim_color, tex_color);
 						if (prim_color.a() > 0) {
-							DrawSinglePixel5551(pixel, prim_color);
+							DrawSinglePixel5551(pixel, prim_color.ToRGBA());
 						}
 						s += ds;
 						pixel++;
@@ -208,7 +207,7 @@ void DrawSprite(const VertexData& v0, const VertexData& v1) {
 				u16 *pixel = fb.Get16Ptr(pos0.x, y, gstate.FrameBufStride());
 				for (int x = pos0.x; x < pos1.x; x++) {
 					Vec4<int> prim_color = v0.color0;
-					DrawSinglePixel5551(pixel, prim_color);
+					DrawSinglePixel5551(pixel, prim_color.ToRGBA());
 					pixel++;
 				}
 			}