GS: Further transfer optimizations

PCSX2 · Oct 3, 2023 · f8c1b84 · f8c1b84
1 parent 2dc3bf1
commit f8c1b84
Show file tree

Hide file tree

Showing 2 changed files with 76 additions and 124 deletions.
diff --git a/pcsx2/GS/GSRegs.h b/pcsx2/GS/GSRegs.h
@@ -1132,8 +1132,6 @@ struct alignas(32) GIFPath
 
 	enum
 	{
-		TYPE_UNKNOWN,
-		TYPE_ADONLY,
 		TYPE_STQRGBAXYZF2,
 		TYPE_STQRGBAXYZ2,
 		TYPE_UVRGBAXYZ2,
@@ -1145,6 +1143,8 @@ struct alignas(32) GIFPath
 		TYPE_STQXYZ2,
 		TYPE_RGBAUVXYZF2,
 		TYPE_UVRGBAXYZF2,
+		TYPE_ADONLY,
+		TYPE_UNKNOWN,
 	};
 
 	__forceinline void SetTag(const void* mem)
@@ -1153,8 +1153,8 @@ struct alignas(32) GIFPath
 
 		// the compiler has a hard time not reloading every time a field of src is accessed
 
-		u32 a = src->U32[0];
-		u32 b = src->U32[1];
+		const u32 a = src->U32[0];
+		const u32 b = src->U32[1];
 
 		tag.U32[0] = a;
 		tag.U32[1] = b;
@@ -1164,7 +1164,7 @@ struct alignas(32) GIFPath
 		if (nloop == 0)
 			return;
 
-		GSVector4i v = GSVector4i::loadl(&src->REGS); // REGS not stored to tag.REGS, only into this->regs, restored before saving the state though
+		const GSVector4i v = GSVector4i::loadl(&src->REGS); // REGS not stored to tag.REGS, only into this->regs, restored before saving the state though
 
 		nreg = (b & 0xf0000000) ? (b >> 28) : 16; // src->NREG
 		regs = v.upl8(v >> 4) & GSVector4i::x0f(nreg);
@@ -1178,49 +1178,49 @@ struct alignas(32) GIFPath
 			{
 				type = TYPE_ADONLY;
 			}
-			else
+			else if(nloop > 4)
 			{
 				switch (nreg)
 				{
 					case 1:
 						break;
 					case 2:
-						if (regs.eq8(GSVector4i::cxpr(0x00000401)).mask() == (1 << 2) - 1)
+						if (regs.U32[0] == 0x00000401)
 							type = TYPE_RGBAXYZF2;
-						else if (regs.eq8(GSVector4i::cxpr(0x0000040)).mask() == (1 << 2) - 1)
+						else if (regs.U32[0] == 0x00000403)
 							type = TYPE_STQXYZF2;
-						else if (regs.eq8(GSVector4i::cxpr(0x00000402)).mask() == (1 << 2) - 13)
+						else if (regs.U32[0] == 0x00000402)
 							type = TYPE_UVXYZF2;
-						else if (regs.eq8(GSVector4i::cxpr(0x00000501)).mask() == (1 << 2) - 1)
+						else if (regs.U32[0] == 0x00000501)
 							type = TYPE_RGBAXYZ2;
-						else if (regs.eq8(GSVector4i::cxpr(0x00000502)).mask() == (1 << 2) - 1)
+						else if (regs.U32[0] == 0x00000502)
 							type = TYPE_STQXYZ2;
-						else if (regs.eq8(GSVector4i::cxpr(0x00000503)).mask() == (1 << 2) - 1)
+						else if (regs.U32[0] == 0x00000503)
 							type = TYPE_UVXYZ2;
 						break;
 					case 3:
 						// many games, TODO: formats mixed with NOPs (xeno2: 040f010f02, 04010f020f, mgs3: 04010f0f02, 0401020f0f, 04010f020f)
-						if (regs.eq8(GSVector4i::cxpr(0x00040102)).mask() == (1 << 3) - 1)
+						if (regs.U32[0] == 0x00040102)
 							type = TYPE_STQRGBAXYZF2;
 						// GoW (has other crazy formats, like ...030503050103)
-						else if (regs.eq8(GSVector4i::cxpr(0x00050102)).mask() == (1 << 3) - 1)
+						else if (regs.U32[0] == 0x00050102)
 							type = TYPE_STQRGBAXYZ2;
 						// TODO: common types with UV instead
-						else if (regs.eq8(GSVector4i::cxpr(0x00050103)).mask() == (1 << 3) - 1)
+						else if (regs.U32[0] == 0x00050103)
 							type = TYPE_UVRGBAXYZ2;
-						else if (regs.eq8(GSVector4i::cxpr(0x00040103)).mask() == (1 << 3) - 1)
+						else if (regs.U32[0] == 0x00040103)
 							type = TYPE_UVRGBAXYZF2;
-						else if (regs.eq8(GSVector4i::cxpr(0x00040301)).mask() == (1 << 3) - 1)
+						else if (regs.U32[0] == 0x00040301)
 							type = TYPE_RGBAUVXYZF2;
 						break;
 					case 4:
-						if (regs.eq8(GSVector4i(0x04030403)).mask() == (1 << 4) - 1)
+						if (regs.U32[0] == 0x04030403)
 						{
 							type = TYPE_UVXYZF2;
 							nreg = 2;
 							nloop *= 2;
 						}
-						else if (regs.eq8(GSVector4i::cxpr(0x05030503)).mask() == (1 << 4) - 1)
+						else if (regs.U32[0] == 0x05030503)
 						{
 							type = TYPE_UVXYZ2;
 							nreg = 2;