x86 JIT: Join adjacent vrot calls together to avoid redundant sin/cos calls. Add a prototype, fix minor issues.

Core/MIPS/ARM/ArmCompVFPU.cpp

@@ -2111,4 +2111,8 @@ namespace MIPSComp
 		fpr.ReleaseSpillLocksAndDiscardTemps();
 	}
 
+	void Jit::Comp_ColorConv(MIPSOpcode op) {
+		DISABLE;
+	}
+
 }

Core/MIPS/ARM/ArmCompVFPUNEON.cpp

@@ -177,5 +177,9 @@ void Jit::CompNEON_Vocp(MIPSOpcode op) {
 	DISABLE;
 }
 
+void Jit::CompNEON_ColorConv(MIPSOpcode op) {
+	DISABLE;
+}
+
 }
 // namespace MIPSComp

Core/MIPS/ARM/ArmJit.h

@@ -138,6 +138,7 @@ public:
 	void Comp_VCrossQuat(MIPSOpcode op);
 	void Comp_Vsgn(MIPSOpcode op);
 	void Comp_Vocp(MIPSOpcode op);
+	void Comp_ColorConv(MIPSOpcode op);
 
 	// Non-NEON: VPFX
 
@@ -176,6 +177,7 @@ public:
 	void CompNEON_VCrossQuat(MIPSOpcode op);
 	void CompNEON_Vsgn(MIPSOpcode op);
 	void CompNEON_Vocp(MIPSOpcode op);
+	void CompNEON_ColorConv(MIPSOpcode op);
 
 	int Replace_fabsf();
 

Core/MIPS/MIPS/MipsJit.h

@@ -131,7 +131,7 @@ public:
 	void Comp_VCrossQuat(MIPSOpcode op) {}
 	void Comp_Vsgn(MIPSOpcode op) {}
 	void Comp_Vocp(MIPSOpcode op) {}
-
+	void Comp_ColorConv(MIPSOpcode op) {}
 	int Replace_fabsf() { return 0; }
 
 	JitBlockCache *GetBlockCache() { return &blocks; }

Core/MIPS/MIPSCodeUtils.h

@@ -46,6 +46,10 @@
 #define MIPS_GET_FT(op) ((op>>16) & 0x1F)
 #define MIPS_GET_FD(op) ((op>>6 ) & 0x1F)
 
+#define MIPS_GET_VD(op) (op & 0x7F)
+#define MIPS_GET_VS(op) ((op>>8) & 0x7F)
+#define MIPS_GET_VT(op) ((op>>16) & 0x7F)
+
 
 namespace MIPSCodeUtils
 {

Core/MIPS/MIPSTables.cpp

@@ -771,9 +771,9 @@ const MIPSInstruction tableVFPU9[32] = // 110100 00010 xxxxx . ....... . .......
 	INVALID, INVALID, INVALID, INVALID,
 	//24
 	INVALID,
-	INSTR("vt4444", &Jit::Comp_Generic, Dis_ColorConv, Int_ColorConv, IN_OTHER|OUT_OTHER|IS_VFPU|OUT_EAT_PREFIX),
-	INSTR("vt5551", &Jit::Comp_Generic, Dis_ColorConv, Int_ColorConv, IN_OTHER|OUT_OTHER|IS_VFPU|OUT_EAT_PREFIX),
-	INSTR("vt5650", &Jit::Comp_Generic, Dis_ColorConv, Int_ColorConv, IN_OTHER|OUT_OTHER|IS_VFPU|OUT_EAT_PREFIX),
+	INSTR("vt4444", &Jit::Comp_ColorConv, Dis_ColorConv, Int_ColorConv, IN_OTHER|OUT_OTHER|IS_VFPU|OUT_EAT_PREFIX),
+	INSTR("vt5551", &Jit::Comp_ColorConv, Dis_ColorConv, Int_ColorConv, IN_OTHER|OUT_OTHER|IS_VFPU|OUT_EAT_PREFIX),
+	INSTR("vt5650", &Jit::Comp_ColorConv, Dis_ColorConv, Int_ColorConv, IN_OTHER|OUT_OTHER|IS_VFPU|OUT_EAT_PREFIX),
 
 	//28
 	INVALID, INVALID, INVALID, INVALID,

Core/MIPS/PPC/PpcCompVFPU.cpp

@@ -1209,5 +1209,8 @@ namespace MIPSComp
 	void Jit::Comp_Vocp(MIPSOpcode op) {
 		DISABLE;
 	}
+	void Jit::Comp_ColorConv(MIPSOpcode op) {
+		DISABLE;
+	}
 }
 

Core/MIPS/fake/FakeJit.h

@@ -131,6 +131,7 @@ public:
 	void Comp_VCrossQuat(MIPSOpcode op) {}
 	void Comp_Vsgn(MIPSOpcode op) {}
 	void Comp_Vocp(MIPSOpcode op) {}
+	void Comp_ColorConv(MIPSOpcode op) {}
 
 	int Replace_fabsf() { return 0; }
 

Core/MIPS/x86/CompVFPU.cpp

@@ -2986,6 +2986,39 @@ void Jit::Comp_Vfim(MIPSOpcode op) {
 	fpr.ReleaseSpillLocks();
 }
 
+void Jit::CompVrotShuffle(u8 *dregs, int imm, int n, bool negSin) {
+	char what[4] = { '0', '0', '0', '0' };
+	if (((imm >> 2) & 3) == (imm & 3)) {
+		for (int i = 0; i < 4; i++)
+			what[i] = 'S';
+	}
+	what[(imm >> 2) & 3] = 'S';
+	what[imm & 3] = 'C';
+
+	// TODO: shufps SIMD version
+
+	for (int i = 0; i < n; i++) {
+		fpr.MapRegV(dregs[i], MAP_DIRTY | MAP_NOINIT);
+		switch (what[i]) {
+		case 'C': MOVSS(fpr.V(dregs[i]), XMM1); break;
+		case 'S':
+			MOVSS(fpr.V(dregs[i]), XMM0);
+			if (negSin) {
+				XORPS(fpr.VX(dregs[i]), M(&signBitLower));
+			}
+			break;
+		case '0':
+		{
+			XORPS(fpr.VX(dregs[i]), fpr.V(dregs[i]));
+			break;
+		}
+		default:
+			ERROR_LOG(JIT, "Bad what in vrot");
+			break;
+		}
+	}
+}
+
 // Very heavily used by FF:CC
 void Jit::Comp_VRot(MIPSOpcode op) {
 	CONDITIONAL_DISABLE;
@@ -2997,8 +3030,22 @@ void Jit::Comp_VRot(MIPSOpcode op) {
 	int n = GetNumVectorElements(sz);
 
 	u8 dregs[4];
+	u8 dregs2[4];
+
+	u32 nextOp = Memory::Read_Opcode_JIT(js.compilerPC + 4).encoding;
+	int vd2 = -1;
+	int imm2 = -1;
+	if (false && (nextOp >> 26) == 60 && ((nextOp >> 21) & 0x1F) == 29 && _VS == MIPS_GET_VS(nextOp)) {
+		// Pair of vrot. Let's join them.
+		vd2 = MIPS_GET_VD(nextOp);
+		imm2 = (nextOp >> 16) & 0x1f;
+		NOTICE_LOG(JIT, "Joint VFPU at %08x", js.blockStart);
+	}
+
 	u8 sreg;
 	GetVectorRegs(dregs, sz, vd);
+	if (vd2 >= 0)
+		GetVectorRegs(dregs2, sz, vd2);
 	GetVectorRegs(&sreg, V_Single, vs);
 
 	// Flush SIMD.
@@ -3009,44 +3056,31 @@ void Jit::Comp_VRot(MIPSOpcode op) {
 	gpr.FlushBeforeCall();
 	fpr.Flush();
 
-	bool negSin = (imm & 0x10) ? true : false;
+	bool negSin1 = (imm & 0x10) ? true : false;
 
 #ifdef _M_X64
 	MOVSS(XMM0, fpr.V(sreg));
-	ABI_CallFunction(negSin ? (const void *)&SinCosNegSin : (const void *)&SinCos);
+	ABI_CallFunction(negSin1 ? (const void *)&SinCosNegSin : (const void *)&SinCos);
 #else
 	// Sigh, passing floats with cdecl isn't pretty, ends up on the stack.
-	ABI_CallFunctionA(negSin ? (const void *)&SinCosNegSin : (const void *)&SinCos, fpr.V(sreg));
+	ABI_CallFunctionA(negSin2 ? (const void *)&SinCosNegSin : (const void *)&SinCos, fpr.V(sreg));
 #endif
 
 	MOVSS(XMM0, M(&sincostemp[0]));
 	MOVSS(XMM1, M(&sincostemp[1]));
-	
-	char what[4] = {'0', '0', '0', '0'};
-	if (((imm >> 2) & 3) == (imm & 3)) {
-		for (int i = 0; i < 4; i++)
-			what[i] = 'S';
-	}
-	what[(imm >> 2) & 3] = 'S';
-	what[imm & 3] = 'C';
 
-	for (int i = 0; i < n; i++) {
-		fpr.MapRegV(dregs[i], MAP_DIRTY | MAP_NOINIT);
-		switch (what[i]) {
-		case 'C': MOVSS(fpr.V(dregs[i]), XMM1); break;
-		case 'S': MOVSS(fpr.V(dregs[i]), XMM0); break;
-		case '0':
-			{
-				XORPS(fpr.VX(dregs[i]), fpr.V(dregs[i]));
-				break;
-			}
-		default:
-			ERROR_LOG(JIT, "Bad what in vrot");
-			break;
-		}
+	CompVrotShuffle(dregs, imm, n, false);
+	if (vd2 != -1) {
+		// If the negsin setting differs between the two joint invocations, we need to flip the second one.
+		bool negSin2 = (imm2 & 0x10) ? true : false;
+		CompVrotShuffle(dregs2, imm2, n, negSin1 != negSin2);
+		js.compilerPC += 4;
 	}
-
 	fpr.ReleaseSpillLocks();
 }
 
+void Jit::Comp_ColorConv(MIPSOpcode op) {
+	DISABLE;
+}
+
 }

Core/MIPS/x86/Jit.h

@@ -145,6 +145,7 @@ public:
 	void Comp_VCrossQuat(MIPSOpcode op);
 	void Comp_Vsgn(MIPSOpcode op);
 	void Comp_Vocp(MIPSOpcode op);
+	void Comp_ColorConv(MIPSOpcode op);
 
 	void Comp_DoNothing(MIPSOpcode op);
 
@@ -236,6 +237,7 @@ private:
 
 	void CompFPTriArith(MIPSOpcode op, void (XEmitter::*arith)(X64Reg reg, OpArg), bool orderMatters);
 	void CompFPComp(int lhs, int rhs, u8 compare, bool allowNaN = false);
+	void CompVrotShuffle(u8 *dregs, int imm, int n, bool negSin);
 
 	void CallProtectedFunction(const void *func, const OpArg &arg1);
 	void CallProtectedFunction(const void *func, const OpArg &arg1, const OpArg &arg2);

Core/MIPS/x86/RegCacheFPU.cpp

@@ -758,6 +758,8 @@ int FPURegCache::GetTempVS(u8 *v, VectorSize vsz) {
 	for (int i = 0; i < n; ++i) {
 		regs[v[i] + 32].tempLocked = true;
 	}
+
+	return 0;  // ??
 }
 
 void FPURegCache::Flush() {