Merge pull request #19124 from hrydgard/minor-ir-optimization

Minor IR interpreter optimizations
hrydgard · May 10, 2024 · e0e9620 · e0e9620
2 parents 9bef910 + 092179c
commit e0e9620
Show file tree

Hide file tree

Showing 4 changed files with 17 additions and 25 deletions.
diff --git a/Core/MIPS/IR/IRInterpreter.cpp b/Core/MIPS/IR/IRInterpreter.cpp
@@ -283,33 +283,20 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst, int count) {
 		case IROp::LoadVec4:
 		{
 			u32 base = mips->r[inst->src1] + inst->constant;
-#if defined(_M_SSE)
-			_mm_store_ps(&mips->f[inst->dest], _mm_load_ps((const float *)Memory::GetPointerUnchecked(base)));
-#else
-			for (int i = 0; i < 4; i++)
-				mips->f[inst->dest + i] = Memory::ReadUnchecked_Float(base + 4 * i);
-#endif
+			// This compiles to a nice SSE load/store on x86, and hopefully similar on ARM.
+			memcpy(&mips->f[inst->dest], Memory::GetPointerUnchecked(base), 4 * 4);
 			break;
 		}
 		case IROp::StoreVec4:
 		{
 			u32 base = mips->r[inst->src1] + inst->constant;
-#if defined(_M_SSE)
-			_mm_store_ps((float *)Memory::GetPointerUnchecked(base), _mm_load_ps(&mips->f[inst->dest]));
-#else
-			for (int i = 0; i < 4; i++)
-				Memory::WriteUnchecked_Float(mips->f[inst->dest + i], base + 4 * i);
-#endif
+			memcpy((float *)Memory::GetPointerUnchecked(base), &mips->f[inst->dest], 4 * 4);
 			break;
 		}
 
 		case IROp::Vec4Init:
 		{
-#if defined(_M_SSE)
-			_mm_store_ps(&mips->f[inst->dest], _mm_load_ps(vec4InitValues[inst->src1]));
-#else
 			memcpy(&mips->f[inst->dest], vec4InitValues[inst->src1], 4 * sizeof(float));
-#endif
 			break;
 		}
 
@@ -398,8 +385,9 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst, int count) {
 #if defined(_M_SSE)
 			_mm_store_ps(&mips->f[inst->dest], _mm_mul_ps(_mm_load_ps(&mips->f[inst->src1]), _mm_set1_ps(mips->f[inst->src2])));
 #else
+			const float factor = mips->f[inst->src2];
 			for (int i = 0; i < 4; i++)
-				mips->f[inst->dest + i] = mips->f[inst->src1 + i] * mips->f[inst->src2];
+				mips->f[inst->dest + i] = mips->f[inst->src1 + i] * factor;
 #endif
 			break;
 		}
@@ -792,7 +780,7 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst, int count) {
 			mips->f[inst->dest] = mips->f[inst->src1] - mips->f[inst->src2];
 			break;
 		case IROp::FMul:
-			if ((my_isinf(mips->f[inst->src1]) && mips->f[inst->src2] == 0.0f) || (my_isinf(mips->f[inst->src2]) && mips->f[inst->src1] == 0.0f)) {
+			if ((mips->f[inst->src2] == 0.0f && my_isinf(mips->f[inst->src1])) || (mips->f[inst->src1] == 0.0f && my_isinf(mips->f[inst->src2]))) {
 				mips->fi[inst->dest] = 0x7fc00000;
 			} else {
 				mips->f[inst->dest] = mips->f[inst->src1] * mips->f[inst->src2];
@@ -1138,13 +1126,14 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst, int count) {
 			// Unimplemented IR op. Bad.
 			Crash();
 		}
-#ifdef _DEBUG
-		if (mips->r[0] != 0)
-			Crash();
-#endif
 		inst++;
 	}
 
+#ifdef _DEBUG
+	if (mips->r[0] != 0)
+		Crash();
+#endif
+
 	// We hit count.  If this is a full block, it was badly constructed.
 	return 0;
 }
diff --git a/Core/MIPS/IR/IRJit.cpp b/Core/MIPS/IR/IRJit.cpp
@@ -250,8 +250,7 @@ void IRJit::RunLoopUntil(u64 globalticks) {
 			u32 inst = Memory::ReadUnchecked_U32(mips_->pc);
 			u32 opcode = inst & 0xFF000000;
 			if (opcode == MIPS_EMUHACK_OPCODE) {
-				u32 data = inst & 0xFFFFFF;
-				IRBlock *block = blocks_.GetBlock(data);
+				IRBlock *block = blocks_.GetBlockUnchecked(inst & 0xFFFFFF);
 				u32 startPC = mips_->pc;
 				mips_->pc = IRInterpret(mips_, block->GetInstructions(), block->GetNumInstructions());
 				// Note: this will "jump to zero" on a badly constructed block missing exits.

diff --git a/Core/MIPS/IR/IRJit.h b/Core/MIPS/IR/IRJit.h
@@ -127,6 +127,9 @@ class IRBlockCache : public JitBlockCacheDebugInterface {
 			return nullptr;
 		}
 	}
+	IRBlock *GetBlockUnchecked(int i) {
+		return &blocks_[i];
+	}
 	const IRBlock *GetBlock(int i) const {
 		if (i >= 0 && i < (int)blocks_.size()) {
 			return &blocks_[i];

diff --git a/Core/MIPS/MIPSIntVFPU.cpp b/Core/MIPS/MIPSIntVFPU.cpp
@@ -621,8 +621,9 @@ namespace MIPSInt
 			break;
 		default:
 			ApplySwizzleS(s, sz);
+			break;
 		}
-		for (int i = 0; i < n; i++) {
+		for (int i = 0; i < (int)n; i++) {
 			switch (optype) {
 			case 0: d[i] = s[i]; break; //vmov
 			case 1: d[i] = s[i]; break; //vabs (prefix)