Merge pull request #19233 from hrydgard/minor-ir-opt

Minor IR Interpreter optimizations, other bugfixes

Core/MIPS/IR/IRInterpreter.cpp

@@ -260,16 +260,24 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 			float temp[4];
 			for (int i = 0; i < 4; i++)
 				temp[i] = mips->f[inst->src1 + ((inst->src2 >> (i * 2)) & 3)];
+			const int dest = inst->dest;
 			for (int i = 0; i < 4; i++)
-				mips->f[inst->dest + i] = temp[i];
+				mips->f[dest + i] = temp[i];
 			break;
 		}
 
 		case IROp::Vec4Blend:
+		{
+			const int dest = inst->dest;
+			const int src1 = inst->src1;
+			const int src2 = inst->src2;
+			const int constant = inst->constant;
+			// 90% of calls to this is inst->constant == 7 or inst->constant == 8. Some are 1 and 4, others very rare.
 			// Could use _mm_blendv_ps (SSE4+BMI), vbslq_f32 (ARM), __riscv_vmerge_vvm (RISC-V)
 			for (int i = 0; i < 4; i++)
-				mips->f[inst->dest + i] = ((inst->constant >> i) & 1) ? mips->f[inst->src2 + i] : mips->f[inst->src1 + i];
+				mips->f[dest + i] = ((constant >> i) & 1) ? mips->f[src2 + i] : mips->f[src1 + i];
 			break;
+		}
 
 		case IROp::Vec4Mov:
 		{
@@ -377,15 +385,19 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 
 		case IROp::Vec2Unpack16To31:
 		{
-			mips->fi[inst->dest] = (mips->fi[inst->src1] << 16) >> 1;
-			mips->fi[inst->dest + 1] = (mips->fi[inst->src1] & 0xFFFF0000) >> 1;
+			const int dest = inst->dest;
+			const int src1 = inst->src1;
+			mips->fi[dest] = (mips->fi[src1] << 16) >> 1;
+			mips->fi[dest + 1] = (mips->fi[src1] & 0xFFFF0000) >> 1;
 			break;
 		}
 
 		case IROp::Vec2Unpack16To32:
 		{
-			mips->fi[inst->dest] = (mips->fi[inst->src1] << 16);
-			mips->fi[inst->dest + 1] = (mips->fi[inst->src1] & 0xFFFF0000);
+			const int dest = inst->dest;
+			const int src1 = inst->src1;
+			mips->fi[dest] = (mips->fi[src1] << 16);
+			mips->fi[dest + 1] = (mips->fi[src1] & 0xFFFF0000);
 			break;
 		}
 
@@ -467,9 +479,11 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 			val = _mm_andnot_si128(mask, val);
 			_mm_store_si128((__m128i *)&mips->fi[inst->dest], val);
 #else
+			const int src1 = inst->src1;
+			const int dest = inst->dest;
 			for (int i = 0; i < 4; i++) {
-				u32 val = mips->fi[inst->src1 + i];
-				mips->fi[inst->dest + i] = (int)val >= 0 ? val : 0;
+				u32 val = mips->fi[src1 + i];
+				mips->fi[dest + i] = (int)val >= 0 ? val : 0;
 			}
 #endif
 			break;
@@ -477,12 +491,14 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 
 		case IROp::Vec4DuplicateUpperBitsAndShift1:  // For vuc2i, the weird one.
 		{
+			const int src1 = inst->src1;
+			const int dest = inst->dest;
 			for (int i = 0; i < 4; i++) {
-				u32 val = mips->fi[inst->src1 + i];
+				u32 val = mips->fi[src1 + i];
 				val = val | (val >> 8);
 				val = val | (val >> 16);
 				val >>= 1;
-				mips->fi[inst->dest + i] = val;
+				mips->fi[dest + i] = val;
 			}
 			break;
 		}
@@ -1111,11 +1127,10 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst) {
 			break;
 
 		case IROp::Nop:
-			_assert_(false);
-			break;
 		default:
-			// Unimplemented IR op. Bad.
 			Crash();
+			break;
+			// Unimplemented IR op. Bad.
 		}
 
 #ifdef _DEBUG

GPU/Common/VertexDecoderCommon.cpp

@@ -1320,6 +1320,10 @@ void VertexDecoder::DecodeVerts(u8 *decodedptr, const void *verts, const UVScale
 	int count = indexUpperBound - indexLowerBound + 1;
 	int stride = decFmt.stride;
 
+#ifdef _DEBUG
+	decodedCount += count;
+#endif
+
 	// Check alignment before running the decoder, as we may crash if it's bad (as should the real PSP but doesn't always)
 	if (((uintptr_t)verts & (biggest - 1)) != 0) {
 		// Bad alignment. Not really sure what to do here... zero the verts to be safe?
@@ -1475,7 +1479,7 @@ static const char * const colnames[8] = { "", "?", "?", "?", "565", "5551", "444
 
 int VertexDecoder::ToString(char *output, bool spaces) const {
 	char *start = output;
-	
+	output += sprintf(output, "[%08x] ", fmt_);
 	output += sprintf(output, "P: %s ", posnames[pos]);
 	if (nrm)
 		output += sprintf(output, "N: %s ", nrmnames[nrm]);
@@ -1502,6 +1506,10 @@ int VertexDecoder::ToString(char *output, bool spaces) const {
 		}
 	}
 
+#ifdef _DEBUG
+	output += sprintf(output, " (%llu)", (long long)decodedCount);
+#endif
+
 	return output - start;
 }
 

GPU/Common/VertexDecoderCommon.h

@@ -475,6 +475,10 @@ public:
 
 	u8 biggest;  // in practice, alignment.
 
+#ifdef _DEBUG
+	mutable u64 decodedCount = 0;
+#endif
+
 	friend class VertexDecoderJitCache;
 
 private:

GPU/Common/VertexDecoderHandwritten.cpp

@@ -16,12 +16,30 @@
 #endif
 #endif
 
-
 // Candidates for hand-writing
 // (found using our custom Very Sleepy).
 // GPU::P:_f_N:_s8_C:_8888_T:_u16__(24b)_040001BE  (5%+ of God of War execution)
 // GPU::P:_f_N:_s8_C:_8888_T:_u16_W:_f_(1x)__(28b)_040007BE (1%+ of God of War execution)
 
+// Tekken 6:
+// (found using the vertex counter that's active in _DEBUG)
+// [04000111] P: s16 C: 565 T: u8  (10b) (736949)    // Also in Midnight Club
+
+// Wipeout Pure:
+// [0400013f] P: s16 N: s8 C: 8888 T: f  (24b) (1495430)
+
+// Flatout:
+// [04000122] P: s16 N: s8 T: u16  (14b) (3901754)
+// [04000116] P: s16 C: 5551 T: u16  (12b) (2225841)
+
+// Test drive:
+// [05000100] P: s16  (6b) (2827872)
+// [050011ff] P: f N: f C: 8888 T: f I: u16  (36b) (3812112)
+
+// Burnout Dominator:
+// [04000122] P: s16 N: s8 T: u16  (14b) (1710813)
+// [04000116] P: s16 C: 5551 T: u16  (12b) (7688298)
+
 // This is the first GoW one.
 void VtxDec_Tu16_C8888_Pfloat(const u8 *srcp, u8 *dstp, int count, const UVScale *uvScaleOffset) {
 	struct GOWVTX {

UI/DevScreens.cpp

@@ -1245,16 +1245,18 @@ void JitCompareScreen::OnRandomBlock(int flag) {
 		int tries = 0;
 		while (!anyWanted && tries < numBlocks) {
 			currentBlock_ = rand() % numBlocks;
-			JitBlockDebugInfo b = blockCache->GetBlockDebugInfo(currentBlock_);
-			u32 mipsBytes = (u32)b.origDisasm.size() * 4;
-			for (u32 addr = b.originalAddress; addr < b.originalAddress + mipsBytes; addr += 4) {
-				MIPSOpcode opcode = Memory::Read_Instruction(addr);
-				if (MIPSGetInfo(opcode) & flag) {
-					char temp[256];
-					MIPSDisAsm(opcode, addr, temp, sizeof(temp));
-					// INFO_LOG(HLE, "Stopping at random instruction: %08x %s", addr, temp);
-					anyWanted = true;
-					break;
+			if (blockCache->IsValidBlock(currentBlock_)) {
+				JitBlockDebugInfo b = blockCache->GetBlockDebugInfo(currentBlock_);
+				u32 mipsBytes = (u32)b.origDisasm.size() * 4;
+				for (u32 addr = b.originalAddress; addr < b.originalAddress + mipsBytes; addr += 4) {
+					MIPSOpcode opcode = Memory::Read_Instruction(addr);
+					if (MIPSGetInfo(opcode) & flag) {
+						char temp[256];
+						MIPSDisAsm(opcode, addr, temp, sizeof(temp));
+						// INFO_LOG(HLE, "Stopping at random instruction: %08x %s", addr, temp);
+						anyWanted = true;
+						break;
+					}
 				}
 			}
 			tries++;
@@ -1360,6 +1362,16 @@ void ShaderViewScreen::CreateViews() {
 	layout->Add(new Button(di->T("Back")))->OnClick.Handle<UIScreen>(this, &UIScreen::OnBack);
 }
 
+bool ShaderViewScreen::key(const KeyInput &ki) {
+	if (ki.flags & KEY_CHAR) {
+		if (ki.unicodeChar == 'C' || ki.unicodeChar == 'c') {
+			System_CopyStringToClipboard(gpu->DebugGetShaderString(id_, type_, SHADER_STRING_SHORT_DESC));
+		}
+	}
+	return UIDialogScreenWithBackground::key(ki);
+}
+
+
 const std::string framedumpsBaseUrl = "http://framedump.ppsspp.org/repro/";
 
 FrameDumpTestScreen::FrameDumpTestScreen() {

UI/DevScreens.h

@@ -205,6 +205,7 @@ public:
 		: id_(id), type_(type) {}
 
 	void CreateViews() override;
+	bool key(const KeyInput &ki) override;
 
 	const char *tag() const override { return "ShaderView"; }