x86jit: Special case broadcast shuffles.

Core/MIPS/IR/IRRegCache.cpp

@@ -159,6 +159,25 @@ bool IRNativeRegCacheBase::IsFPRMapped(IRReg fpr) {
 	return mr[fpr + 32].loc == MIPSLoc::FREG || mr[fpr + 32].loc == MIPSLoc::VREG;
 }
 
+int IRNativeRegCacheBase::GetFPRLaneCount(IRReg fpr) {
+	if (!IsFPRMapped(fpr) || mr[fpr].lane > 0)
+		return 0;
+	if (mr[fpr].lane == -1)
+		return 1;
+
+	int c = 1;
+	for (int i = 1; i < 4; ++i) {
+		if (mr[fpr + i].nReg != mr[fpr].nReg || mr[fpr + i].loc != mr[fpr].loc)
+			return c;
+		if (mr[fpr + i].lane != i)
+			return c;
+
+		c++;
+	}
+
+	return c;
+}
+
 int IRNativeRegCacheBase::GetFPRLane(IRReg fpr) {
 	_dbg_assert_(IsValidFPR(fpr));
 	if (mr[fpr + 32].loc == MIPSLoc::FREG || mr[fpr + 32].loc == MIPSLoc::VREG) {

Core/MIPS/IR/IRRegCache.h

@@ -164,6 +164,7 @@ public:
 	bool IsGPRMappedAsPointer(IRReg gpr);
 	bool IsGPRMappedAsStaticPointer(IRReg gpr);
 	int GetFPRLane(IRReg fpr);
+	int GetFPRLaneCount(IRReg fpr);
 
 	bool IsGPRImm(IRReg gpr);
 	bool IsGPR2Imm(IRReg base);

Core/MIPS/x86/X64IRCompVec.cpp

@@ -212,7 +212,13 @@ void X64JitBackend::CompIR_VecAssign(IRInst inst) {
 		break;
 
 	case IROp::Vec4Shuffle:
-		regs_.Map(inst);
+		if (regs_.GetFPRLaneCount(inst.src1) == 1 && (inst.src1 & 3) == 0 && inst.src2 == 0) {
+			// This is a broadcast.  If dest == src1, this won't clear it.
+			regs_.SpillLockFPR(inst.src1);
+			regs_.MapVec4(inst.dest, MIPSMap::NOINIT);
+		} else {
+			regs_.Map(inst);
+		}
 		if (cpu_info.bAVX) {
 			VPERMILPS(128, regs_.FX(inst.dest), regs_.F(inst.src1), inst.src2);
 		} else {