Optimize some common prefixes

Author: hrydgard

Core/MIPS/IR/IRCompVFPU.cpp

@@ -108,6 +108,13 @@ namespace MIPSComp {
 		}
 	}
 
+	static void InitRegs(u8 *vregs, int reg) {
+		vregs[0] = reg;
+		vregs[1] = reg + 1;
+		vregs[2] = reg + 2;
+		vregs[3] = reg + 3;
+	}
+
 	void IRFrontend::ApplyPrefixST(u8 *vregs, u32 prefix, VectorSize sz, int tempReg) {
 		if (prefix == 0xE4)
 			return;
@@ -119,6 +126,27 @@ namespace MIPSComp {
 		for (int i = 0; i < n; i++)
 			origV[i] = vregs[i];
 
+		// Some common vector prefixes
+		if (sz == V_Quad && IsConsecutive4(vregs)) {
+			if (prefix == 0xF00E4 && IsConsecutive4(vregs)) {
+				InitRegs(vregs, tempReg);
+				ir.Write(IROp::Vec4Neg, vregs[0], origV[0]);
+				return;
+			}
+			if (prefix == 0x00FE4 && IsConsecutive4(vregs)) {
+				InitRegs(vregs, tempReg);
+				ir.Write(IROp::Vec4Abs, vregs[0], origV[0]);
+				return;
+			}
+			// Pure shuffle
+			if (prefix == (prefix & 0xFF)) {
+				InitRegs(vregs, tempReg);
+				ir.Write(IROp::Vec4Shuffle, vregs[0], origV[0], prefix);
+				return;
+			}
+		}
+
+		// Alright, fall back to the generic approach.
 		for (int i = 0; i < n; i++) {
 			int regnum = (prefix >> (i * 2)) & 3;
 			int abs = (prefix >> (8 + i)) & 1;
@@ -395,7 +423,6 @@ namespace MIPSComp {
 		GetVectorRegsPrefixT(tregs, sz, vt);
 		GetVectorRegsPrefixD(dregs, V_Single, vd);
 
-		// TODO: applyprefixST here somehow (shuffle, etc...)
 		ir.Write(IROp::FMul, IRVTEMP_0, sregs[0], tregs[0]);
 
 		int n = GetNumVectorElements(sz);
@@ -1050,7 +1077,7 @@ namespace MIPSComp {
 			}
 		} else if (sz == M_4x4) {
 			// Tekken 6 has a case here: MEE
-			logBlocks = 1;
+			// logBlocks = 1;
 		}
 
 		// Fallback. Expands a LOT
@@ -1141,8 +1168,8 @@ namespace MIPSComp {
 			tempregs[i] = temp;
 		}
 		for (int i = 0; i < n; i++) {
-			u8 temp = tempregs[i];
-			ir.Write(IROp::FMov, dregs[i], temp);
+			if (tempregs[i] != dregs[i])
+				ir.Write(IROp::FMov, dregs[i], tempregs[i]);
 		}
 	}
 

Core/MIPS/IR/IRInst.cpp

@@ -109,6 +109,8 @@ static const IRMeta irMeta[] = {
 	{ IROp::Vec4Mul, "Vec4Mul", "FFF" },
 	{ IROp::Vec4Scale, "Vec4Scale", "FFF" },
 	{ IROp::Vec4Dot, "Vec4Dot", "FFF" },
+	{ IROp::Vec4Neg, "Vec4Neg", "FF" },
+	{ IROp::Vec4Abs, "Vec4Abs", "FF" },
 
 	{ IROp::Interpret, "Interpret", "_C" },
 	{ IROp::Downcount, "Downcount", "_II" },

Core/MIPS/IR/IRInst.h

@@ -159,6 +159,8 @@ enum class IROp : u8 {
 	Vec4Div,
 	Vec4Scale,
 	Vec4Dot,
+	Vec4Neg,
+	Vec4Abs,
 
 	// vx2i
 	Vec4ExpandU16ToU32Hi,

Core/MIPS/IR/IRInterpreter.cpp

@@ -212,6 +212,16 @@ u32 IRInterpret(MIPSState *mips, const IRInst *inst, const u32 *constPool, int c
 #endif
 			break;
 
+		case IROp::Vec4Neg:
+			for (int i = 0; i < 4; i++)
+				mips->f[inst->dest + i] = -mips->f[inst->src1 + i];
+			break;
+
+		case IROp::Vec4Abs:
+			for (int i = 0; i < 4; i++)
+				mips->f[inst->dest + i] = fabsf(mips->f[inst->src1 + i]);
+			break;
+
 		case IROp::FCmpVfpuBit:
 		{
 			int op = inst->dest & 0xF;

Core/MIPS/IR/IRPassSimplify.cpp

@@ -114,11 +114,11 @@ bool OptimizeFPMoves(const IRWriter &in, IRWriter &out) {
 				inst.op = IROp::FMov;
 				inst.src1 = prev.src1;
 				out.Write(inst);
-				logBlocks = true;
 			} else {
 				out.Write(inst);
 			}
 			break;
+
 		default:
 			// Remap constants to the new reality
 			const IRMeta *m = GetIRMeta(inst.op);
@@ -487,6 +487,8 @@ bool PropagateConstants(const IRWriter &in, IRWriter &out) {
 		case IROp::Vec4Dot:
 		case IROp::Vec4Scale:
 		case IROp::Vec4Shuffle:
+		case IROp::Vec4Neg:
+		case IROp::Vec4Abs:
 			out.Write(inst);
 			break;