Fix crashes in Single Core mode on Dolphin Win64.

Cherry-picked from 6209067daa51.

5 files changed, 28 insertions, 28 deletions

diff --git a/Source/Core/Common/Src/x64ABI.cpp b/Source/Core/Common/Src/x64ABI.cpp
index b5726947fc..1b4fa0b5ff 100644
--- a/Source/Core/Common/Src/x64ABI.cpp
+++ b/Source/Core/Common/Src/x64ABI.cpp
@@ -15,9 +15,9 @@ unsigned int XEmitter::ABI_GetAlignedFrameSize(unsigned int frameSize, bool noPr
 	// the stack pointer is 4/8 bytes less than a multiple of 16; however, the
 	// function prolog immediately subtracts an appropriate amount to align
 	// it, so no alignment is required around a call. 
-	// In the functions generated by ThunkManager::ProtectFunction, we add the
-	// necessary subtraction (and 0x20 bytes shadow space for Win64) into this
-	// rather than having a separate prolog.
+	// In the functions generated by ThunkManager::ProtectFunction and some
+	// others, we add the necessary subtraction (and 0x20 bytes shadow space
+	// for Win64) into this rather than having a separate prolog.
 	// On Windows 32-bit, the required alignment is only 4 bytes, so we just
 	// ensure that the frame size isn't misaligned.
 #ifdef _M_X64
@@ -142,13 +142,13 @@ void XEmitter::ABI_CallFunctionR(void *func, X64Reg reg1) {
 }
 
 // Pass two registers as parameters.
-void XEmitter::ABI_CallFunctionRR(void *func, Gen::X64Reg reg1, Gen::X64Reg reg2)
+void XEmitter::ABI_CallFunctionRR(void *func, Gen::X64Reg reg1, Gen::X64Reg reg2, bool noProlog)
 {
-	ABI_AlignStack(2 * 4);
+	ABI_AlignStack(2 * 4, noProlog);
 	PUSH(32, R(reg2));
 	PUSH(32, R(reg1));
 	CALL(func);
-	ABI_RestoreStack(2 * 4);
+	ABI_RestoreStack(2 * 4, noProlog);
 }
 
 void XEmitter::ABI_CallFunctionAC(void *func, const Gen::OpArg &arg1, u32 param2)
@@ -351,8 +351,8 @@ void XEmitter::ABI_CallFunctionR(void *func, X64Reg reg1) {
 }
 
 // Pass two registers as parameters.
-void XEmitter::ABI_CallFunctionRR(void *func, X64Reg reg1, X64Reg reg2) {
-	ABI_AlignStack(0);
+void XEmitter::ABI_CallFunctionRR(void *func, X64Reg reg1, X64Reg reg2, bool noProlog) {
+	ABI_AlignStack(0, noProlog);
 	if (reg2 != ABI_PARAM1) {
 		if (reg1 != ABI_PARAM1)
 			MOV(64, R(ABI_PARAM1), R(reg1));
@@ -373,7 +373,7 @@ void XEmitter::ABI_CallFunctionRR(void *func, X64Reg reg1, X64Reg reg2) {
 	} else {
 		CALL(func);
 	}
-	ABI_RestoreStack(0);
+	ABI_RestoreStack(0, noProlog);
 }
 
 void XEmitter::ABI_CallFunctionAC(void *func, const Gen::OpArg &arg1, u32 param2)
diff --git a/Source/Core/Common/Src/x64Emitter.h b/Source/Core/Common/Src/x64Emitter.h
index c42c955ec4..27a735ab90 100644
--- a/Source/Core/Common/Src/x64Emitter.h
+++ b/Source/Core/Common/Src/x64Emitter.h
@@ -639,7 +639,7 @@ public:
 
 	// Pass a register as a parameter.
 	void ABI_CallFunctionR(void *func, Gen::X64Reg reg1);
-	void ABI_CallFunctionRR(void *func, Gen::X64Reg reg1, Gen::X64Reg reg2);
+	void ABI_CallFunctionRR(void *func, Gen::X64Reg reg1, Gen::X64Reg reg2, bool noProlog = false);
 
 	// A function that doesn't have any control over what it will do to regs,
 	// such as the dispatcher, should be surrounded by these.
diff --git a/Source/Core/Core/Src/PowerPC/JitCommon/JitAsmCommon.cpp b/Source/Core/Core/Src/PowerPC/JitCommon/JitAsmCommon.cpp
index 5187ef1478..d80b0bb3f8 100644
--- a/Source/Core/Core/Src/PowerPC/JitCommon/JitAsmCommon.cpp
+++ b/Source/Core/Core/Src/PowerPC/JitCommon/JitAsmCommon.cpp
@@ -167,7 +167,7 @@ void CommonAsmRoutines::GenQuantizedStores() {
 	MOV(64, MComplex(RBX, RCX, SCALE_1, 0), R(RAX));
 	FixupBranch skip_complex = J();
 	SetJumpTarget(too_complex);
-	ABI_CallFunctionRR(thunks.ProtectFunction((void *)&WriteDual32, 2), RAX, RCX); 
+	ABI_CallFunctionRR(thunks.ProtectFunction((void *)&WriteDual32, 2), RAX, RCX, /* noProlog = */ true); 
 	SetJumpTarget(skip_complex);
 	RET();
 #else
@@ -184,10 +184,10 @@ void CommonAsmRoutines::GenQuantizedStores() {
 	FixupBranch arg2 = J();
 	SetJumpTarget(argh);
 	MOV(32, R(EAX), M(((char*)&psTemp)));
-	ABI_CallFunctionRR(thunks.ProtectFunction((void *)&Memory::Write_U32, 2), EAX, ECX); 
+	ABI_CallFunctionRR(thunks.ProtectFunction((void *)&Memory::Write_U32, 2), EAX, ECX, /* noProlog = */ true); 
 	MOV(32, R(EAX), M(((char*)&psTemp)+4));
 	ADD(32, R(ECX), Imm32(4));
-	ABI_CallFunctionRR(thunks.ProtectFunction((void *)&Memory::Write_U32, 2), EAX, ECX); 
+	ABI_CallFunctionRR(thunks.ProtectFunction((void *)&Memory::Write_U32, 2), EAX, ECX, /* noProlog = */ true); 
 	SetJumpTarget(arg2);
 	RET();
 #endif
@@ -206,7 +206,7 @@ void CommonAsmRoutines::GenQuantizedStores() {
 	PACKSSDW(XMM0, R(XMM0));
 	PACKUSWB(XMM0, R(XMM0));
 	MOVD_xmm(R(EAX), XMM0);
-	SafeWriteRegToReg(AX, ECX, 16, 0, false);
+	SafeWriteRegToReg(AX, ECX, 16, 0, false, true);
 	
 	RET();
 
@@ -225,7 +225,7 @@ void CommonAsmRoutines::GenQuantizedStores() {
 	PACKSSWB(XMM0, R(XMM0));
 	MOVD_xmm(R(EAX), XMM0);
 
-	SafeWriteRegToReg(AX, ECX, 16, 0, false);
+	SafeWriteRegToReg(AX, ECX, 16, 0, false, true);
 	
 	RET();
 
@@ -251,7 +251,7 @@ void CommonAsmRoutines::GenQuantizedStores() {
 	MOV(16, R(AX), M((char*)psTemp + 4));
 
 	BSWAP(32, EAX);
-	SafeWriteRegToReg(EAX, ECX, 32, 0, false);
+	SafeWriteRegToReg(EAX, ECX, 32, 0, false, true);
 	
 	RET();
 
@@ -271,7 +271,7 @@ void CommonAsmRoutines::GenQuantizedStores() {
 	MOVD_xmm(R(EAX), XMM0);
 	BSWAP(32, EAX);
 	ROL(32, R(EAX), Imm8(16));
-	SafeWriteRegToReg(EAX, ECX, 32, 0, false);
+	SafeWriteRegToReg(EAX, ECX, 32, 0, false, true);
 	
 	RET();
 
@@ -303,11 +303,11 @@ void CommonAsmRoutines::GenQuantizedSingleStores() {
 		// TODO: SafeWriteFloat
 		MOVSS(M(&psTemp[0]), XMM0);
 		MOV(32, R(EAX), M(&psTemp[0]));
-		SafeWriteRegToReg(EAX, ECX, 32, 0, false);
+		SafeWriteRegToReg(EAX, ECX, 32, 0, false, true);
 	} else {
 		MOVSS(M(&psTemp[0]), XMM0);
 		MOV(32, R(EAX), M(&psTemp[0]));
-		SafeWriteRegToReg(EAX, ECX, 32, 0, true);
+		SafeWriteRegToReg(EAX, ECX, 32, 0, true, true);
 	}*/
 
 	const u8* storeSingleU8 = AlignCode4();  // Used by MKWii
@@ -318,7 +318,7 @@ void CommonAsmRoutines::GenQuantizedSingleStores() {
 	MAXSS(XMM0, R(XMM1));
 	MINSS(XMM0, M((void *)&m_255));
 	CVTTSS2SI(EAX, R(XMM0));
-	SafeWriteRegToReg(AL, ECX, 8, 0, true);
+	SafeWriteRegToReg(AL, ECX, 8, 0, true, true);
 	RET();
 
 	const u8* storeSingleS8 = AlignCode4();
@@ -328,7 +328,7 @@ void CommonAsmRoutines::GenQuantizedSingleStores() {
 	MAXSS(XMM0, M((void *)&m_m128));
 	MINSS(XMM0, M((void *)&m_127));
 	CVTTSS2SI(EAX, R(XMM0));
-	SafeWriteRegToReg(AL, ECX, 8, 0, true);
+	SafeWriteRegToReg(AL, ECX, 8, 0, true, true);
 	RET();
 
 	const u8* storeSingleU16 = AlignCode4();  // Used by MKWii
@@ -339,7 +339,7 @@ void CommonAsmRoutines::GenQuantizedSingleStores() {
 	MAXSS(XMM0, R(XMM1));
 	MINSS(XMM0, M((void *)&m_65535));
 	CVTTSS2SI(EAX, R(XMM0));
-	SafeWriteRegToReg(EAX, ECX, 16, 0, true);
+	SafeWriteRegToReg(EAX, ECX, 16, 0, true, true);
 	RET();
 
 	const u8* storeSingleS16 = AlignCode4();
@@ -349,7 +349,7 @@ void CommonAsmRoutines::GenQuantizedSingleStores() {
 	MAXSS(XMM0, M((void *)&m_m32768));
 	MINSS(XMM0, M((void *)&m_32767));
 	CVTTSS2SI(EAX, R(XMM0));
-	SafeWriteRegToReg(EAX, ECX, 16, 0, true);
+	SafeWriteRegToReg(EAX, ECX, 16, 0, true, true);
 	RET();
 
 	singleStoreQuantized = reinterpret_cast<const u8**>(const_cast<u8*>(AlignCode16()));
diff --git a/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.cpp b/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.cpp
index 18c8044385..637dff171b 100644
--- a/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.cpp
+++ b/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.cpp
@@ -223,7 +223,7 @@ void EmuCodeBlock::UnsafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int ac
 }
 
 // Destroys both arg registers
-void EmuCodeBlock::SafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int accessSize, s32 offset, bool swap)
+void EmuCodeBlock::SafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int accessSize, s32 offset, bool swap, bool noProlog)
 {
 	if (offset)
 		ADD(32, R(reg_addr), Imm32((u32)offset));
@@ -247,9 +247,9 @@ void EmuCodeBlock::SafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int acce
 	MOV(32, M(&PC), Imm32(jit->js.compilerPC)); // Helps external systems know which instruction triggered the write
 	switch (accessSize)
 	{
-	case 32: ABI_CallFunctionRR(thunks.ProtectFunction(swap ? ((void *)&Memory::Write_U32) : ((void *)&Memory::Write_U32_Swap), 2), reg_value, reg_addr); break;
-	case 16: ABI_CallFunctionRR(thunks.ProtectFunction(swap ? ((void *)&Memory::Write_U16) : ((void *)&Memory::Write_U16_Swap), 2), reg_value, reg_addr); break;
-	case 8:  ABI_CallFunctionRR(thunks.ProtectFunction((void *)&Memory::Write_U8, 2), reg_value, reg_addr);  break;
+	case 32: ABI_CallFunctionRR(thunks.ProtectFunction(swap ? ((void *)&Memory::Write_U32) : ((void *)&Memory::Write_U32_Swap), 2), reg_value, reg_addr, noProlog); break;
+	case 16: ABI_CallFunctionRR(thunks.ProtectFunction(swap ? ((void *)&Memory::Write_U16) : ((void *)&Memory::Write_U16_Swap), 2), reg_value, reg_addr, noProlog); break;
+	case 8:  ABI_CallFunctionRR(thunks.ProtectFunction((void *)&Memory::Write_U8, 2), reg_value, reg_addr, noProlog);  break;
 	}
 	FixupBranch exit = J();
 	SetJumpTarget(fast);
diff --git a/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.h b/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.h
index 7708ab8598..57dce19790 100644
--- a/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.h
+++ b/Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.h
@@ -16,7 +16,7 @@ public:
 	void UnsafeWriteRegToReg(Gen::X64Reg reg_value, Gen::X64Reg reg_addr, int accessSize, s32 offset = 0, bool swap = true);
 	void UnsafeLoadToEAX(const Gen::OpArg & opAddress, int accessSize, s32 offset, bool signExtend);
 	void SafeLoadToEAX(const Gen::OpArg & opAddress, int accessSize, s32 offset, bool signExtend);
-	void SafeWriteRegToReg(Gen::X64Reg reg_value, Gen::X64Reg reg_addr, int accessSize, s32 offset, bool swap = true);
+	void SafeWriteRegToReg(Gen::X64Reg reg_value, Gen::X64Reg reg_addr, int accessSize, s32 offset, bool swap = true, bool noProlog = false);
 
 	// Trashes both inputs and EAX.
 	void SafeWriteFloatToReg(Gen::X64Reg xmm_value, Gen::X64Reg reg_addr);