JIT compiler:

* Improved constants folding in load instructions
* Merged load instructions
* Fixed the register allocator so that it can restore the registers state after jumping to the dispatcher (in case of a conditional jump)

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6076 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Common/Src/ABI.cpp

@@ -127,12 +127,20 @@ void XEmitter::ABI_CallFunctionRR(void *func, Gen::X64Reg reg1, Gen::X64Reg reg2
 void XEmitter::ABI_CallFunctionAC(void *func, const Gen::OpArg &arg1, u32 param2)
 {
 	ABI_AlignStack(2 * 4);
-	PUSH(32, arg1);
 	PUSH(32, Imm32(param2));
+	PUSH(32, arg1);
 	CALL(func);
 	ABI_RestoreStack(2 * 4);
 }
 
+void XEmitter::ABI_CallFunctionA(void *func, const Gen::OpArg &arg1)
+{
+	ABI_AlignStack(1 * 4);
+	PUSH(32, arg1);
+	CALL(func);
+	ABI_RestoreStack(1 * 4);
+}
+
 void XEmitter::ABI_PushAllCalleeSavedRegsAndAdjustStack() {
 	// Note: 4 * 4 = 16 bytes, so alignment is preserved.
 	PUSH(EBP);

Source/Core/Common/Src/x64Emitter.h

@@ -600,6 +600,7 @@ public:
 	void ABI_CallFunctionCCC(void *func, u32 param1, u32 param2, u32 param3);
 	void ABI_CallFunctionCCP(void *func, u32 param1, u32 param2, void *param3);
 	void ABI_CallFunctionAC(void *func, const Gen::OpArg &arg1, u32 param2);
+	void ABI_CallFunctionA(void *func, const Gen::OpArg &arg1);
 
 	// Pass a register as a paremeter.
 	void ABI_CallFunctionR(void *func, Gen::X64Reg reg1);

Source/Core/Core/Src/PowerPC/Jit64/Jit.cpp

@@ -300,6 +300,14 @@ void Jit64::Cleanup()
 
 void Jit64::WriteExit(u32 destination, int exit_num)
 {
+	// We are about to jump to the dispatcher => save and flush regs
+	RegCacheState regCacheStateGPR;
+	RegCacheState regCacheStateFPR;
+	gpr.SaveState(regCacheStateGPR);
+	fpr.SaveState(regCacheStateFPR);
+	gpr.Flush(FLUSH_ALL);
+	fpr.Flush(FLUSH_ALL);
+
 	Cleanup();
 	SUB(32, M(&CoreTiming::downcount), js.downcountAmount > 127 ? Imm32(js.downcountAmount) : Imm8(js.downcountAmount)); 
 
@@ -321,29 +329,69 @@ void Jit64::WriteExit(u32 destination, int exit_num)
 		MOV(32, M(&PC), Imm32(destination));
 		JMP(asm_routines.dispatcher, true);
 	}
+
+	// Restore registers states so that the next instructions could still use the cached values
+	gpr.LoadState(regCacheStateGPR);
+	fpr.LoadState(regCacheStateFPR);
 }
 
 void Jit64::WriteExitDestInEAX(int exit_num) 
 {
+	// We are about to jump to the dispatcher => save and flush regs
+	RegCacheState regCacheStateGPR;
+	RegCacheState regCacheStateFPR;
+	gpr.SaveState(regCacheStateGPR);
+	fpr.SaveState(regCacheStateFPR);
+	gpr.Flush(FLUSH_ALL);
+	fpr.Flush(FLUSH_ALL);
+
 	MOV(32, M(&PC), R(EAX));
 	Cleanup();
 	SUB(32, M(&CoreTiming::downcount), js.downcountAmount > 127 ? Imm32(js.downcountAmount) : Imm8(js.downcountAmount)); 
 	JMP(asm_routines.dispatcher, true);
+
+	// Restore registers states so that the next instructions could still use the cached values
+	gpr.LoadState(regCacheStateGPR);
+	fpr.LoadState(regCacheStateFPR);
 }
 
 void Jit64::WriteRfiExitDestInEAX() 
 {
+	// We are about to jump to the exception handler => save and flush regs
+	RegCacheState regCacheStateGPR;
+	RegCacheState regCacheStateFPR;
+	gpr.SaveState(regCacheStateGPR);
+	fpr.SaveState(regCacheStateFPR);
+	gpr.Flush(FLUSH_ALL);
+	fpr.Flush(FLUSH_ALL);
+	
 	MOV(32, M(&PC), R(EAX));
 	Cleanup();
 	SUB(32, M(&CoreTiming::downcount), js.downcountAmount > 127 ? Imm32(js.downcountAmount) : Imm8(js.downcountAmount)); 
 	JMP(asm_routines.testExceptions, true);
+
+	// Restore registers states so that the next instructions could still use the cached values
+	gpr.LoadState(regCacheStateGPR);
+	fpr.LoadState(regCacheStateFPR);
 }
 
 void Jit64::WriteExceptionExit()
 {
+	// We are about to jump to the exception handler => save and flush regs
+	RegCacheState regCacheStateGPR;
+	RegCacheState regCacheStateFPR;
+	gpr.SaveState(regCacheStateGPR);
+	fpr.SaveState(regCacheStateFPR);
+	gpr.Flush(FLUSH_ALL);
+	fpr.Flush(FLUSH_ALL);
+	
 	Cleanup();
 	SUB(32, M(&CoreTiming::downcount), js.downcountAmount > 127 ? Imm32(js.downcountAmount) : Imm8(js.downcountAmount)); 
 	JMP(asm_routines.testExceptions, true);
+
+	// Restore registers states so that the next instructions could still use the cached values
+	gpr.LoadState(regCacheStateGPR);
+	fpr.LoadState(regCacheStateFPR);
 }
 
 void STACKALIGN Jit64::Run()
@@ -552,7 +600,15 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 			{
 				//This instruction uses FPU - needs to add FP exception bailout
 				TEST(32, M(&PowerPC::ppcState.msr), Imm32(1 << 13)); // Test FP enabled bit
-				FixupBranch b1 = J_CC(CC_NZ);
+				FixupBranch b1 = J_CC(CC_NZ, true);
+
+				// We are about to jump to the exception handler => save and flush regs
+				RegCacheState regCacheStateGPR;
+				RegCacheState regCacheStateFPR;
+				gpr.SaveState(regCacheStateGPR);
+				fpr.SaveState(regCacheStateFPR);
+				gpr.Flush(FLUSH_ALL);
+				fpr.Flush(FLUSH_ALL);
 
 				// If a FPU exception occurs, the exception handler will read
 				// from PC.  Update PC with the latest value in case that happens.
@@ -560,18 +616,19 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 				SUB(32, M(&CoreTiming::downcount), js.downcountAmount > 127 ? Imm32(js.downcountAmount) : Imm8(js.downcountAmount)); 
 				JMP(asm_routines.fpException, true);
 				SetJumpTarget(b1);
+
+				// Restore registers states
+				// (If no memory exception occured, the next instructions could still use the cached values)
+				gpr.LoadState(regCacheStateGPR);
+				fpr.LoadState(regCacheStateFPR);
 			}
 
 			Jit64Tables::CompileInstruction(ops[i]);
 
 			if (js.memcheck && (opinfo->flags & FL_LOADSTORE))
 			{
-				// In case we are about to jump to the dispatcher, flush regs
-				gpr.Flush(FLUSH_ALL);
-				fpr.Flush(FLUSH_ALL);
-
 				TEST(32, M(&PowerPC::ppcState.Exceptions), Imm32(EXCEPTION_DSI));
-				FixupBranch noMemException = J_CC(CC_Z);
+				FixupBranch noMemException = J_CC(CC_Z, true);
 
 				// If a memory exception occurs, the exception handler will read
 				// from PC.  Update PC with the latest value in case that happens.
@@ -607,8 +664,6 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 
 	if (broken_block)
 	{
-		gpr.Flush(FLUSH_ALL);
-		fpr.Flush(FLUSH_ALL);
 		WriteExit(nextPC, 0);
 	}
 

Source/Core/Core/Src/PowerPC/Jit64/Jit.h

@@ -236,8 +236,6 @@ public:
 	void fmaddXX(UGeckoInstruction inst);
 	void fsign(UGeckoInstruction inst);
 	void stX(UGeckoInstruction inst); //stw sth stb
-	void lXz(UGeckoInstruction inst);
-	void lha(UGeckoInstruction inst);
 	void rlwinmx(UGeckoInstruction inst);
 	void rlwimix(UGeckoInstruction inst);
 	void rlwnmx(UGeckoInstruction inst);
@@ -254,11 +252,7 @@ public:
 	void subfmex(UGeckoInstruction inst);
 	void subfzex(UGeckoInstruction inst);
 
-	void lbzx(UGeckoInstruction inst);
+	void lXXx(UGeckoInstruction inst);
-	void lwzx(UGeckoInstruction inst);
-	void lhax(UGeckoInstruction inst);
-	
-	void lwzux(UGeckoInstruction inst);
 	
 	void stXx(UGeckoInstruction inst);
 

Source/Core/Core/Src/PowerPC/Jit64/Jit64_Tables.cpp

@@ -77,14 +77,14 @@ static GekkoOPTemplate primarytable[] =
 	{28, &Jit64::reg_imm}, //"andi_rc",  OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_SET_CR0}},
 	{29, &Jit64::reg_imm}, //"andis_rc", OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_SET_CR0}},
 
-	{32, &Jit64::lXz}, //"lwz",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
+	{32, &Jit64::lXXx}, //"lwz",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
-	{33, &Jit64::Default}, //"lwzu", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
+	{33, &Jit64::lXXx}, //"lwzu", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
-	{34, &Jit64::lXz}, //"lbz",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
+	{34, &Jit64::lXXx}, //"lbz",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
-	{35, &Jit64::Default}, //"lbzu", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
+	{35, &Jit64::lXXx}, //"lbzu", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
-	{40, &Jit64::lXz}, //"lhz",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
+	{40, &Jit64::lXXx}, //"lhz",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
-	{41, &Jit64::Default}, //"lhzu", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
+	{41, &Jit64::lXXx}, //"lhzu", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
-	{42, &Jit64::lha}, //"lha",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
+	{42, &Jit64::lXXx}, //"lha",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A}},
-	{43, &Jit64::Default}, //"lhau", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
+	{43, &Jit64::lXXx}, //"lhau", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A}},
 
 	{44, &Jit64::stX}, //"sth",  OPTYPE_STORE, FL_IN_A | FL_IN_S}},
 	{45, &Jit64::stX}, //"sthu", OPTYPE_STORE, FL_OUT_A | FL_IN_A | FL_IN_S}},
@@ -220,20 +220,20 @@ static GekkoOPTemplate table31[] =
 	{1014, &Jit64::dcbz}, //"dcbz",   OPTYPE_DCACHE, 0, 4}},
 
 	//load word
-	{23,  &Jit64::lwzx}, //"lwzx",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
+	{23,  &Jit64::lXXx}, //"lwzx",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
-	{55,  &Jit64::lwzux}, //"lwzux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
+	{55,  &Jit64::lXXx}, //"lwzux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
 
 	//load halfword
-	{279, &Jit64::Default}, //"lhzx",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
+	{279, &Jit64::lXXx}, //"lhzx",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
-	{311, &Jit64::Default}, //"lhzux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
+	{311, &Jit64::lXXx}, //"lhzux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
 
 	//load halfword signextend
-	{343, &Jit64::lhax}, //"lhax",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
+	{343, &Jit64::lXXx}, //"lhax",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
-	{375, &Jit64::Default}, //"lhaux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
+	{375, &Jit64::lXXx}, //"lhaux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
 
 	//load byte
-	{87,  &Jit64::lbzx}, //"lbzx",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
+	{87,  &Jit64::lXXx}, //"lbzx",  OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},
-	{119, &Jit64::Default}, //"lbzux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
+	{119, &Jit64::lXXx}, //"lbzux", OPTYPE_LOAD, FL_OUT_D | FL_OUT_A | FL_IN_A | FL_IN_B}},
 
 	//load byte reverse
 	{534, &Jit64::Default}, //"lwbrx", OPTYPE_LOAD, FL_OUT_D | FL_IN_A0 | FL_IN_B}},

Source/Core/Core/Src/PowerPC/Jit64/JitRegCache.cpp

@@ -28,18 +28,15 @@ using namespace PowerPC;
 RegCache::RegCache() : emit(0) {
 	memset(locks, 0, sizeof(locks));
 	memset(xlocks, 0, sizeof(xlocks));
-	memset(saved_locks, 0, sizeof(saved_locks));
-	memset(saved_xlocks, 0, sizeof(saved_xlocks));
 	memset(regs, 0, sizeof(regs));
 	memset(xregs, 0, sizeof(xregs));
-	memset(saved_regs, 0, sizeof(saved_regs));
-	memset(saved_xregs, 0, sizeof(saved_xregs));
 }
 
 void RegCache::Start(PPCAnalyst::BlockRegStats &stats)
 {
 	for (int i = 0; i < NUMXREGS; i++)
 	{
+		xregs[i].ppcReg = -1;
 		xregs[i].free = true;
 		xregs[i].dirty = false;
 		xlocks[i] = false;
@@ -48,6 +45,7 @@ void RegCache::Start(PPCAnalyst::BlockRegStats &stats)
 	{
 		regs[i].location = GetDefaultLocation(i);
 		regs[i].away = false;
+		locks[i] = false;
 	}
 	
 	// todo: sort to find the most popular regs
@@ -137,20 +135,20 @@ X64Reg RegCache::GetFreeXReg()
 	return (X64Reg) -1;
 }
 
-void RegCache::SaveState()
+void RegCache::SaveState(RegCacheState & state)
 {
-	memcpy(saved_locks, locks, sizeof(locks));
+	memcpy(state.locks, locks, sizeof(locks));
-	memcpy(saved_xlocks, xlocks, sizeof(xlocks));
+	memcpy(state.xlocks, xlocks, sizeof(xlocks));
-	memcpy(saved_regs, regs, sizeof(regs));
+	memcpy(state.regs, regs, sizeof(regs));
-	memcpy(saved_xregs, xregs, sizeof(xregs));
+	memcpy(state.xregs, xregs, sizeof(xregs));
 }
 
-void RegCache::LoadState()
+void RegCache::LoadState(const RegCacheState & state)
 {
-	memcpy(xlocks, saved_xlocks, sizeof(xlocks));
+	memcpy(xlocks, state.xlocks, sizeof(xlocks));
-	memcpy(locks, saved_locks, sizeof(locks));
+	memcpy(locks, state.locks, sizeof(locks));
-	memcpy(regs, saved_regs, sizeof(regs));
+	memcpy(regs, state.regs, sizeof(regs));
-	memcpy(xregs, saved_xregs, sizeof(xregs));
+	memcpy(xregs, state.xregs, sizeof(xregs));
 }
 
 void RegCache::FlushR(X64Reg reg)

Source/Core/Core/Src/PowerPC/Jit64/JitRegCache.h

@@ -55,21 +55,17 @@ typedef int PReg;
 #define NUMXREGS 8
 #endif
 
-class RegCache
+struct RegCacheState
 {
-private:
 	bool locks[32];
-	bool saved_locks[32];
-	bool saved_xlocks[NUMXREGS];
-
-protected:
 	bool xlocks[NUMXREGS];
 	PPCCachedReg regs[32];
 	X64CachedReg xregs[NUMXREGS];
+};
 
-	PPCCachedReg saved_regs[32];
+class RegCache : protected RegCacheState
-	X64CachedReg saved_xregs[NUMXREGS];
+{
-
+protected:
 	virtual const int *GetAllocationOrder(int &count) = 0;
 	
 	XEmitter *emit;
@@ -123,8 +119,8 @@ public:
 
 	X64Reg GetFreeXReg();
 
-	void SaveState();
+	void SaveState(RegCacheState & state);
-	void LoadState();
+	void LoadState(const RegCacheState & state);
 };
 
 class GPRRegCache : public RegCache

Source/Core/Core/Src/PowerPC/Jit64/Jit_Branch.cpp

@@ -45,8 +45,6 @@ void Jit64::sc(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(Branch)
 
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
 	MOV(32, M(&PC), Imm32(js.compilerPC + 4));
 	OR(32, M(&PowerPC::ppcState.Exceptions), Imm32(EXCEPTION_SYSCALL));
 	WriteExceptionExit();
@@ -57,8 +55,6 @@ void Jit64::rfi(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(Branch)
 
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
 	// See Interpreter rfi for details
 	const u32 mask = 0x87C0FFFF;
 	const u32 clearMSR13 = 0xFFFBFFFF; // Mask used to clear the bit MSR[13]
@@ -79,8 +75,6 @@ void Jit64::bx(UGeckoInstruction inst)
 
 	if (inst.LK)
 		MOV(32, M(&LR), Imm32(js.compilerPC + 4));
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
 
 	if (js.isLastInstruction)
 	{
@@ -120,17 +114,14 @@ void Jit64::bcx(UGeckoInstruction inst)
 	// USES_CR
 	_assert_msg_(DYNA_REC, js.isLastInstruction, "bcx not last instruction of block");
 
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
-
 	FixupBranch pCTRDontBranch;
 	if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0)  // Decrement and test CTR
 	{
 		SUB(32, M(&CTR), Imm8(1));
 		if (inst.BO & BO_BRANCH_IF_CTR_0)
-			pCTRDontBranch = J_CC(CC_NZ);
+			pCTRDontBranch = J_CC(CC_NZ, true);
 		else
-			pCTRDontBranch = J_CC(CC_Z);
+			pCTRDontBranch = J_CC(CC_Z, true);
 	}
 
 	FixupBranch pConditionDontBranch;
@@ -138,9 +129,9 @@ void Jit64::bcx(UGeckoInstruction inst)
 	{
 		TEST(8, M(&PowerPC::ppcState.cr_fast[inst.BI >> 2]), Imm8(8 >> (inst.BI & 3)));
 		if (inst.BO & BO_BRANCH_IF_TRUE)  // Conditional branch 
-			pConditionDontBranch = J_CC(CC_Z);
+			pConditionDontBranch = J_CC(CC_Z, true);
 		else
-			pConditionDontBranch = J_CC(CC_NZ);
+			pConditionDontBranch = J_CC(CC_NZ, true);
 	}
 	
 	if (inst.LK)
@@ -165,9 +156,6 @@ void Jit64::bcctrx(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(Branch)
 
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
-
 	// bcctrx doesn't decrement and/or test CTR
 	_dbg_assert_msg_(POWERPC, inst.BO_2 & BO_DONT_DECREMENT_FLAG, "bcctrx with decrement and test CTR option is invalid!");
 
@@ -196,7 +184,7 @@ void Jit64::bcctrx(UGeckoInstruction inst)
 			branch = CC_Z;
 		else
 			branch = CC_NZ; 
-		FixupBranch b = J_CC(branch, false);
+		FixupBranch b = J_CC(branch, true);
 		MOV(32, R(EAX), M(&CTR));
 		AND(32, R(EAX), Imm32(0xFFFFFFFC));
 		//MOV(32, M(&PC), R(EAX)); => Already done in WriteExitDestInEAX()
@@ -214,17 +202,14 @@ void Jit64::bclrx(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(Branch)
 
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
-
 	FixupBranch pCTRDontBranch;
 	if ((inst.BO & BO_DONT_DECREMENT_FLAG) == 0)  // Decrement and test CTR
 	{
 		SUB(32, M(&CTR), Imm8(1));
 		if (inst.BO & BO_BRANCH_IF_CTR_0)
-			pCTRDontBranch = J_CC(CC_NZ);
+			pCTRDontBranch = J_CC(CC_NZ, true);
 		else
-			pCTRDontBranch = J_CC(CC_Z);
+			pCTRDontBranch = J_CC(CC_Z, true);
 	}
 
 	FixupBranch pConditionDontBranch;
@@ -232,9 +217,9 @@ void Jit64::bclrx(UGeckoInstruction inst)
 	{
 		TEST(8, M(&PowerPC::ppcState.cr_fast[inst.BI >> 2]), Imm8(8 >> (inst.BI & 3)));
 		if (inst.BO & BO_BRANCH_IF_TRUE)  // Conditional branch 
-			pConditionDontBranch = J_CC(CC_Z);
+			pConditionDontBranch = J_CC(CC_Z, true);
 		else
-			pConditionDontBranch = J_CC(CC_NZ);
+			pConditionDontBranch = J_CC(CC_NZ, true);
 	}
 
 		// This below line can be used to prove that blr "eats flags" in practice.

Source/Core/Core/Src/PowerPC/Jit64/Jit_Integer.cpp

@@ -264,9 +264,6 @@ void Jit64::cmpXX(UGeckoInstruction inst)
 		{
 			js.downcountAmount++;
 
-			gpr.Flush(FLUSH_ALL);
-			fpr.Flush(FLUSH_ALL);
-
 			int test_bit = 8 >> (js.next_inst.BI & 3);
 			u8 conditionResult = (js.next_inst.BO & BO_BRANCH_IF_TRUE) ? test_bit : 0;
 			if ((compareResult & test_bit) == conditionResult)
@@ -359,21 +356,19 @@ void Jit64::cmpXX(UGeckoInstruction inst)
 			// if (rand() & 1)
 			//     std::swap(destination1, destination2), condition = !condition;
 
-			gpr.Flush(FLUSH_ALL);
-			fpr.Flush(FLUSH_ALL);
 			FixupBranch pLesser  = J_CC(less_than);
 			FixupBranch pGreater = J_CC(greater_than);
 			MOV(8, M(&PowerPC::ppcState.cr_fast[crf]), Imm8(0x2));  //  == 0
-			FixupBranch continue1 = J();
+			FixupBranch continue1 = J(true);
 
 			SetJumpTarget(pGreater);
 			MOV(8, M(&PowerPC::ppcState.cr_fast[crf]), Imm8(0x4));  //  > 0
-			FixupBranch continue2 = J();
+			FixupBranch continue2 = J(true);
 
 			SetJumpTarget(pLesser);
 			MOV(8, M(&PowerPC::ppcState.cr_fast[crf]), Imm8(0x8));  //  < 0
 			FixupBranch continue3;
-			if (!!(8 & test_bit) == condition) continue3 = J();
+			if (!!(8 & test_bit) == condition) continue3 = J(true);
 			if (!!(4 & test_bit) != condition) SetJumpTarget(continue2);
 			if (!!(2 & test_bit) != condition) SetJumpTarget(continue1);
 			if (js.next_inst.OPCD == 16) // bcx

Source/Core/Core/Src/PowerPC/Jit64/Jit_LoadStore.cpp

@@ -34,96 +34,84 @@
 #include "JitAsm.h"
 #include "JitRegCache.h"
 
-void Jit64::lbzx(UGeckoInstruction inst)
+void Jit64::lXXx(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
 	JITDISABLE(LoadStore)
+	int a = inst.RA, b = inst.RB, d = inst.RD;
 
-	if (Core::g_CoreStartupParameter.bJITLoadStorelbzxOff)
+	// Skip disabled JIT instructions
+	if (Core::g_CoreStartupParameter.bJITLoadStorelbzxOff && (inst.OPCD == 31) && (inst.SUBOP10 == 87))
+	{ Default(inst); return; }
+	if (Core::g_CoreStartupParameter.bJITLoadStorelXzOff && ((inst.OPCD == 34) || (inst.OPCD == 40) || (inst.OPCD == 32)))
+	{ Default(inst); return; }
+	if (Core::g_CoreStartupParameter.bJITLoadStorelwzOff && (inst.OPCD == 32))
 	{ Default(inst); return; }
 
-	int a = inst.RA, b = inst.RB, d = inst.RD;
+	// Determine memory access size and sign extend
-	gpr.FlushLockX(ABI_PARAM1);
+	int accessSize;
-	MOV(32, R(ABI_PARAM1), gpr.R(b));
+	bool signExtend;
-	if (a)
+	switch (inst.OPCD)
 	{
-		ADD(32, R(ABI_PARAM1), gpr.R(a));
+	case 32: /* lwz */
+	case 33: /* lwzu */
+		accessSize = 32;
+		signExtend = false;
+		break;
+		
+	case 34: /* lbz */
+	case 35: /* lbzu */
+		accessSize = 8;
+		signExtend = false;
+		break;
+		
+	case 40: /* lhz */
+	case 41: /* lhzu */
+		accessSize = 16;
+		signExtend = false;
+		break;
+		
+	case 42: /* lha */
+	case 43: /* lhau */
+		accessSize = 16;
+		signExtend = true;
+		break;
+		
+	case 31:
+		switch (inst.SUBOP10)
+		{
+		case 23:  /* lwzx */
+		case 55:  /* lwzux */
+			accessSize = 32;
+			signExtend = false;
+			break;
+			
+		case 87:  /* lbzx */
+		case 119: /* lbzux */
+			accessSize = 8;
+			signExtend = false;
+			break;
+		case 279: /* lhzx */
+		case 311: /* lhzux */
+			accessSize = 16;
+			signExtend = false;
+			break;
+			
+		case 343: /* lhax */
+		case 375: /* lhaux */
+			accessSize = 16;
+			signExtend = true;
+			break;
+			
+		default:
+			PanicAlert("Invalid instruction");
+		}
+		break;
+		
+	default:
+		PanicAlert("Invalid instruction");
 	}
 
-	SafeLoadRegToEAX(ABI_PARAM1, 8, 0);
-
-	MEMCHECK_START
-
-	gpr.KillImmediate(d, false, true);
-	MOV(32, gpr.R(d), R(EAX));
-
-	MEMCHECK_END
-
-	gpr.UnlockAllX();
-}
-
-void Jit64::lhax(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(LoadStore)
-
-	int a = inst.RA, b = inst.RB, d = inst.RD;
-	gpr.FlushLockX(ABI_PARAM1);
-	MOV(32, R(ABI_PARAM1), gpr.R(b));
-	if (a)
-	{
-		ADD(32, R(ABI_PARAM1), gpr.R(a));
-	}
-
-	// Some homebrew actually loads from a hw reg with this instruction
-	SafeLoadRegToEAX(ABI_PARAM1, 16, 0, true);
-
-	MEMCHECK_START
-
-	gpr.KillImmediate(d, false, true);
-	MOV(32, gpr.R(d), R(EAX));
-
-	MEMCHECK_END
-
-	gpr.UnlockAllX();
-}
-
-void Jit64::lwzx(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(LoadStore)
-
-	int a = inst.RA, b = inst.RB, d = inst.RD;
-	gpr.FlushLockX(ABI_PARAM1);
-	MOV(32, R(ABI_PARAM1), gpr.R(b));
-	if (a)
-	{
-		ADD(32, R(ABI_PARAM1), gpr.R(a));
-	}
-
-	SafeLoadRegToEAX(ABI_PARAM1, 32, 0);
-
-	MEMCHECK_START
-
-	gpr.KillImmediate(d, false, true);
-	MOV(32, gpr.R(d), R(EAX));
-
-	MEMCHECK_END
-
-	gpr.UnlockAllX();
-}
-
-void Jit64::lXz(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(LoadStore)
-
-	if (Core::g_CoreStartupParameter.bJITLoadStorelXzOff)
-	{ Default(inst); return; }
-
-	int d = inst.RD;
-	int a = inst.RA;
-
 	// TODO(ector): Make it dynamically enable/disable idle skipping where appropriate
 	// Will give nice boost to dual core mode
 	// (mb2): I agree, 
@@ -144,23 +132,24 @@ void Jit64::lXz(UGeckoInstruction inst)
 
 		// do our job at first
 		s32 offset = (s32)(s16)inst.SIMM_16;
-		gpr.FlushLockX(ABI_PARAM1);
 		gpr.Lock(d);
-		MOV(32, R(ABI_PARAM1), gpr.R(a));
+		SafeLoadToEAX(gpr.R(a), accessSize, offset, signExtend);
-		SafeLoadRegToEAX(ABI_PARAM1, 32, offset);
 		gpr.KillImmediate(d, false, true);
 		MOV(32, gpr.R(d), R(EAX));
 		gpr.UnlockAll();
-		gpr.UnlockAllX();
-
-		gpr.Flush(FLUSH_ALL); 
 		
 		// if it's still 0, we can wait until the next event
 		CMP(32, R(RAX), Imm32(0));
 		FixupBranch noIdle = J_CC(CC_NE);
 
+		// We are about to jump to the exception handler, save and flush regs
+		RegCacheState regCacheStateGPR;
+		RegCacheState regCacheStateFPR;
+		gpr.SaveState(regCacheStateGPR);
+		fpr.SaveState(regCacheStateFPR);
 		gpr.Flush(FLUSH_ALL);
 		fpr.Flush(FLUSH_ALL);
+
 		ABI_CallFunctionC((void *)&PowerPC::OnIdle, PowerPC::ppcState.gpr[a] + (s32)(s16)inst.SIMM_16);
 
 		// ! we must continue executing of the loop after exception handling, maybe there is still 0 in r0
@@ -169,113 +158,100 @@ void Jit64::lXz(UGeckoInstruction inst)
 
 		SetJumpTarget(noIdle);
 
+		// Restore registers states so that the next instructions could still use the cached values
+		gpr.LoadState(regCacheStateGPR);
+		fpr.LoadState(regCacheStateFPR);
+
 		//js.compilerPC += 8;
 		return;
 	}
-
+	else if ((inst.OPCD == 32 /* lwz */) && accessSize == 32 && jo.enableFastMem && !Core::g_CoreStartupParameter.bMMU)
-	// R2 always points to the small read-only data area. We could bake R2-relative loads into immediates.
-	// R13 always points to the small read/write data area. Not so exciting but at least could drop checks in 32-bit safe mode.
-
-	s32 offset = (s32)(s16)inst.SIMM_16;
-	if (!a) 
-	{
-		Default(inst);
-		return;
-	}
-
-	int accessSize;
-	switch (inst.OPCD)
-	{
-	case 32:
-		accessSize = 32;
-		if (Core::g_CoreStartupParameter.bJITLoadStorelwzOff) {Default(inst); return;}
-		break; //lwz	
-	case 40: accessSize = 16; break; //lhz
-	case 34: accessSize = 8;  break; //lbz
-	default:
-		//_assert_msg_(DYNA_REC, 0, "lXz: invalid access size");
-		PanicAlert("lXz: invalid access size");
-		return;
-	}
-
-	if (accessSize == 32 && jo.enableFastMem && !Core::g_CoreStartupParameter.bMMU)
 	{
 		// Fast and daring
+		s32 offset = (s32)(s16)inst.SIMM_16;
 		gpr.Lock(a, d);
 		gpr.BindToRegister(a, true, false);
 		gpr.BindToRegister(d, a == d, true);
 		MOV(accessSize, gpr.R(d), MComplex(RBX, gpr.R(a).GetSimpleReg(), SCALE_1, offset));
 		BSWAP(32, gpr.R(d).GetSimpleReg());
 		gpr.UnlockAll();
+		return;
+	}
+	
+	// Determine whether this instruction updates inst.RA
+	bool update;
+	if (inst.OPCD == 31)
+		update = ((inst.SUBOP10 & 0x20) != 0);
+	else
+		update = ((inst.OPCD & 1) != 0);
+	
+	// Prepare address operand
+	Gen::OpArg opAddress;
+	if (!update && !a)
+	{
+		if (inst.OPCD == 31)
+		{
+			gpr.Lock(b);
+			opAddress = gpr.R(b);
+		}
+		else
+		{
+			opAddress = Imm32((u32)(s32)inst.SIMM_16);
+		}
+	}
+	else if (update && ((a == 0) || (d == a)))
+	{
+		PanicAlert("Invalid instruction");
 	}
 	else
 	{
-		gpr.FlushLockX(ABI_PARAM1);
+		if ((inst.OPCD != 31) && gpr.R(a).IsImm())
-		gpr.Lock(a);
+		{
-		gpr.BindToRegister(a, true, false);
+			opAddress = Imm32((u32)gpr.R(a).offset + (s32)inst.SIMM_16);
-		MOV(32, R(ABI_PARAM1), gpr.R(a));
+		}
-		SafeLoadRegToEAX(ABI_PARAM1, accessSize, offset);
+		else if ((inst.OPCD == 31) && gpr.R(a).IsImm() && gpr.R(b).IsImm())
+		{
+			opAddress = Imm32((u32)gpr.R(a).offset + (u32)gpr.R(b).offset);
+		}
+		else
+		{
+			gpr.FlushLockX(ABI_PARAM1);
+			opAddress = R(ABI_PARAM1);
+			MOV(32, opAddress, gpr.R(a));
 			
-		MEMCHECK_START
+			if (inst.OPCD == 31)
-
+				ADD(32, opAddress, gpr.R(b));
-		gpr.KillImmediate(d, false, true);
+			else
-		MOV(32, gpr.R(d), R(EAX));
+				ADD(32, opAddress, Imm32((u32)(s32)inst.SIMM_16));
-		
+		}
-		MEMCHECK_END
-
-		gpr.UnlockAll();
-		gpr.UnlockAllX();
 	}
-}
 
-void Jit64::lha(UGeckoInstruction inst)
+	SafeLoadToEAX(opAddress, accessSize, 0, signExtend);
-{
-	INSTRUCTION_START
-	JITDISABLE(LoadStore)
 
-	int d = inst.RD;
+	// We must flush immediate values from the following registers because
-	int a = inst.RA;
+	// they may change at runtime if no MMU exception has been raised
-	s32 offset = (s32)(s16)inst.SIMM_16;
+	gpr.KillImmediate(d, true, true);
-	// Safe and boring
+	if (update)
-	gpr.FlushLockX(ABI_PARAM1);
-	MOV(32, R(ABI_PARAM1), gpr.R(a));
-	SafeLoadRegToEAX(ABI_PARAM1, 16, offset, true);
-
-	MEMCHECK_START
-
-	gpr.KillImmediate(d, false, true);
-	MOV(32, gpr.R(d), R(EAX));
-
-	MEMCHECK_END
-
-	gpr.UnlockAllX();
-}
-
-void Jit64::lwzux(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(LoadStore)
-
-	int a = inst.RA, b = inst.RB, d = inst.RD;
-	if (!a || a == d || a == b)
 	{
-		Default(inst);
+		gpr.Lock(a);
-		return;
+		gpr.BindToRegister(a, true, true);
 	}
-	gpr.Lock(a);
-	gpr.BindToRegister(a, true, true);
-	ADD(32, gpr.R(a), gpr.R(b));
-	MOV(32, R(EAX), gpr.R(a));
-	SafeLoadRegToEAX(EAX, 32, 0, false);
 	
 	MEMCHECK_START
 
-	gpr.KillImmediate(d, false, true);
+	if (update)
+	{
+		if (inst.OPCD == 31)
+			ADD(32, gpr.R(a), gpr.R(b));
+		else
+			ADD(32, gpr.R(a), Imm32((u32)(s32)inst.SIMM_16));
+	}
 	MOV(32, gpr.R(d), R(EAX));
 
 	MEMCHECK_END
 	
 	gpr.UnlockAll();
+	gpr.UnlockAll();
 }
 
 // Zero cache line.
@@ -312,7 +288,7 @@ void Jit64::stX(UGeckoInstruction inst)
 	bool update = inst.OPCD & 1;
 
 	s32 offset = (s32)(s16)inst.SIMM_16;
-	if (a || update) 
+	if (a || !update) 
 	{
 		int accessSize;
 		switch (inst.OPCD & ~1)
@@ -323,18 +299,18 @@ void Jit64::stX(UGeckoInstruction inst)
 		default: _assert_msg_(DYNA_REC, 0, "AWETKLJASDLKF"); return;
 		}
 
-		if (gpr.R(a).IsImm())
+		if ((a == 0) || gpr.R(a).IsImm())
 		{
 			// If we already know the address through constant folding, we can do some
 			// fun tricks...
-			u32 addr = (u32)gpr.R(a).offset;
+			u32 addr = ((a == 0) ? 0 : (u32)gpr.R(a).offset);
 			addr += offset;
 			if ((addr & 0xFFFFF000) == 0xCC008000 && jo.optimizeGatherPipe)
 			{
-				if (offset && update)
-					gpr.SetImmediate32(a, addr);
 				gpr.FlushLockX(ABI_PARAM1);
 				MOV(32, R(ABI_PARAM1), gpr.R(s));
+				if (update)
+					gpr.SetImmediate32(a, addr);
 				switch (accessSize)
 				{	
 					// No need to protect these, they don't touch any state
@@ -347,16 +323,27 @@ void Jit64::stX(UGeckoInstruction inst)
 				gpr.UnlockAllX();
 				return;
 			}
-			else if (Memory::IsRAMAddress(addr) && accessSize == 32)
+			else if (Memory::IsRAMAddress(addr))
 			{
-				if (offset && update)
+				MOV(32, R(EAX), gpr.R(s));
-					gpr.SetImmediate32(a, addr);
-				MOV(accessSize, R(EAX), gpr.R(s));
 				BSWAP(accessSize, EAX);
 				WriteToConstRamAddress(accessSize, R(EAX), addr);
+				if (update)
+					gpr.SetImmediate32(a, addr);
+				return;
+			}
+			else
+			{
+				switch (accessSize)
+				{
+				case 32: ABI_CallFunctionAC(thunks.ProtectFunction(true ? ((void *)&Memory::Write_U32) : ((void *)&Memory::Write_U32_Swap), 2), gpr.R(s), addr); break;
+				case 16: ABI_CallFunctionAC(thunks.ProtectFunction(true ? ((void *)&Memory::Write_U16) : ((void *)&Memory::Write_U16_Swap), 2), gpr.R(s), addr); break;
+				case 8:  ABI_CallFunctionAC(thunks.ProtectFunction((void *)&Memory::Write_U8, 2), gpr.R(s), addr);  break;
+				}
+				if (update)
+					gpr.SetImmediate32(a, addr);
 				return;
 			}
-			// Other IO not worth the trouble.
 		}
 		
 		// Optimized stack access?
@@ -372,7 +359,7 @@ void Jit64::stX(UGeckoInstruction inst)
 			AND(32, R(ABI_PARAM1), Imm32(Memory::MEMVIEW32_MASK));
 			MOV(accessSize, MDisp(ABI_PARAM1, (u32)Memory::base + (u32)offset), R(EAX));
 #endif
-			if (update)
+			if (update && offset)
 			{
 				gpr.Lock(a);
 				gpr.KillImmediate(a, true, true);
@@ -406,9 +393,9 @@ void Jit64::stX(UGeckoInstruction inst)
 
 		if (update && offset)
 		{
+			gpr.KillImmediate(a, true, true);
 			MEMCHECK_START
 			
-			gpr.KillImmediate(a, true, true);
 			ADD(32, gpr.R(a), Imm32((u32)offset));
 
 			MEMCHECK_END
@@ -419,7 +406,7 @@ void Jit64::stX(UGeckoInstruction inst)
 	}
 	else
 	{
-		Default(inst);
+		PanicAlert("Invalid stX");
 	}
 }
 

Source/Core/Core/Src/PowerPC/Jit64/Jit_LoadStoreFloating.cpp

@@ -62,15 +62,13 @@ void Jit64::lfs(UGeckoInstruction inst)
 		return;
 	}
 	s32 offset = (s32)(s16)inst.SIMM_16;
-	gpr.FlushLockX(ABI_PARAM1);
-	MOV(32, R(ABI_PARAM1), gpr.R(a));
 	if (jo.assumeFPLoadFromMem)
 	{
-		UnsafeLoadRegToReg(ABI_PARAM1, EAX, 32, offset, false);
+		UnsafeLoadToEAX(gpr.R(a), 32, offset, false);
 	}
 	else
 	{
-		SafeLoadRegToEAX(ABI_PARAM1, 32, offset);
+		SafeLoadToEAX(gpr.R(a), 32, offset, false);
 	}
 
 	MEMCHECK_START
@@ -83,7 +81,6 @@ void Jit64::lfs(UGeckoInstruction inst)
 
 	MEMCHECK_END
 
-	gpr.UnlockAllX();
 	fpr.UnlockAll();
 }
 
@@ -299,9 +296,12 @@ void Jit64::stfs(UGeckoInstruction inst)
 	ADD(32, R(ABI_PARAM2), Imm32(offset));
 	if (update && offset)
 	{
+		// We must flush immediate values from the following register because
+		// it may take another value at runtime if no MMU exception has been raised
+		gpr.KillImmediate(a, true, true);
+
 		MEMCHECK_START
 
-		gpr.KillImmediate(a, false, true);
 		MOV(32, gpr.R(a), R(ABI_PARAM2));
 
 		MEMCHECK_END

Source/Core/Core/Src/PowerPC/Jit64/Jit_SystemRegisters.cpp

@@ -123,8 +123,6 @@ void Jit64::mtmsr(UGeckoInstruction inst)
 	}
 	MOV(32, M(&MSR), gpr.R(inst.RS));
 	gpr.UnlockAll();
-	gpr.Flush(FLUSH_ALL);
-	fpr.Flush(FLUSH_ALL);
 	WriteExit(js.compilerPC + 4, 0);
 }
 // ==============

Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.cpp

@@ -71,6 +71,56 @@ void EmuCodeBlock::UnsafeLoadRegToRegNoSwap(X64Reg reg_addr, X64Reg reg_value, i
 #endif
 }
 
+void EmuCodeBlock::UnsafeLoadToEAX(const Gen::OpArg & opAddress, int accessSize, s32 offset, bool signExtend)
+{
+	if (opAddress.IsImm())
+	{
+	#ifdef _M_IX86
+		MOVZX(32, accessSize, EAX, M(Memory::base + (((u32)opAddress.offset + offset) & Memory::MEMVIEW32_MASK)));
+	#else
+		MOVZX(32, accessSize, EAX, M(Memory::base + (u32)opAddress.offset + offset));
+	#endif
+	}
+	else
+	{
+	#ifdef _M_IX86
+		MOV(32, R(EAX), opAddress);
+		if (offset)
+			ADD(32, R(EAX), Imm32(offset));
+		AND(32, R(EAX), Imm32(Memory::MEMVIEW32_MASK));
+		MOVZX(32, accessSize, EAX, MDisp(EAX, (u32)Memory::base));
+	#else
+		if (opAddress.IsSimpleReg())
+		{
+			MOVZX(32, accessSize, EAX, MComplex(RBX, opAddress.GetSimpleReg(), SCALE_1, offset));
+		}
+		else
+		{
+			MOV(32, R(EAX), opAddress);
+			MOVZX(32, accessSize, EAX, MComplex(RBX, EAX, SCALE_1, offset));
+		}
+	#endif
+	}
+	
+	if (accessSize == 32)
+	{
+		BSWAP(32, EAX);
+	}
+	else if (accessSize == 16)
+	{
+		BSWAP(32, EAX);
+		if (signExtend)
+			SAR(32, R(EAX), Imm8(16));
+		else
+			SHR(32, R(EAX), Imm8(16));
+	}
+	else if (signExtend)
+	{
+		// TODO: bake 8-bit into the original load.
+		MOVSX(32, accessSize, EAX, R(EAX));   
+	}
+}
+
 void EmuCodeBlock::SafeLoadRegToEAX(X64Reg reg_addr, int accessSize, s32 offset, bool signExtend)
 {
 	if (Core::g_CoreStartupParameter.bUseFastMem && (accessSize == 32 || accessSize == 8) && !Core::g_CoreStartupParameter.bMMU)
@@ -112,6 +162,95 @@ void EmuCodeBlock::SafeLoadRegToEAX(X64Reg reg_addr, int accessSize, s32 offset,
 	}
 }
 
+void EmuCodeBlock::SafeLoadToEAX(const Gen::OpArg & opAddress, int accessSize, s32 offset, bool signExtend)
+{
+	if (Core::g_CoreStartupParameter.bUseFastMem && (accessSize == 32 || accessSize == 8) && !Core::g_CoreStartupParameter.bMMU)
+	{
+		// FIXME: accessSize == 16 does not work.  Breaks mkdd
+		UnsafeLoadToEAX(opAddress, accessSize, offset, signExtend);
+	}
+	else
+	{
+		u32 mem_mask = Memory::ADDR_MASK_HW_ACCESS;
+		if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+		{
+			mem_mask |= Memory::ADDR_MASK_MEM1;
+		}
+		
+		if (opAddress.IsImm())
+		{
+			u32 address = (u32)opAddress.offset + offset;
+			if ((address & mem_mask) == 0)
+			{
+				UnsafeLoadToEAX(opAddress, accessSize, offset, signExtend);
+			}
+			else
+			{
+				switch (accessSize)
+				{
+				case 32: ABI_CallFunctionC(thunks.ProtectFunction((void *)&Memory::Read_U32, 1), address); break;
+				case 16: ABI_CallFunctionC(thunks.ProtectFunction((void *)&Memory::Read_U16_ZX, 1), address); break;
+				case 8:  ABI_CallFunctionC(thunks.ProtectFunction((void *)&Memory::Read_U8_ZX, 1), address); break;
+				}
+				if (signExtend && accessSize < 32)
+				{
+					// Need to sign extend values coming from the Read_U* functions.
+					MOVSX(32, accessSize, EAX, R(EAX));
+				}
+			}
+		}
+		else
+		{
+			if (offset)
+			{
+				MOV(32, R(EAX), opAddress);
+				ADD(32, R(EAX), Imm32(offset));
+				TEST(32, R(EAX), Imm32(mem_mask));
+				FixupBranch fast = J_CC(CC_Z);
+
+				switch (accessSize)
+				{
+				case 32: ABI_CallFunctionR(thunks.ProtectFunction((void *)&Memory::Read_U32, 1), EAX); break;
+				case 16: ABI_CallFunctionR(thunks.ProtectFunction((void *)&Memory::Read_U16_ZX, 1), EAX); break;
+				case 8:  ABI_CallFunctionR(thunks.ProtectFunction((void *)&Memory::Read_U8_ZX, 1), EAX);  break;
+				}
+				if (signExtend && accessSize < 32)
+				{
+					// Need to sign extend values coming from the Read_U* functions.
+					MOVSX(32, accessSize, EAX, R(EAX));
+				}
+
+				FixupBranch exit = J();
+				SetJumpTarget(fast);
+				UnsafeLoadToEAX(opAddress, accessSize, offset, signExtend);
+				SetJumpTarget(exit);
+			}
+			else
+			{
+				TEST(32, opAddress, Imm32(mem_mask));
+				FixupBranch fast = J_CC(CC_Z);
+
+				switch (accessSize)
+				{
+				case 32: ABI_CallFunctionA(thunks.ProtectFunction((void *)&Memory::Read_U32, 1), opAddress); break;
+				case 16: ABI_CallFunctionA(thunks.ProtectFunction((void *)&Memory::Read_U16_ZX, 1), opAddress); break;
+				case 8:  ABI_CallFunctionA(thunks.ProtectFunction((void *)&Memory::Read_U8_ZX, 1), opAddress);  break;
+				}
+				if (signExtend && accessSize < 32)
+				{
+					// Need to sign extend values coming from the Read_U* functions.
+					MOVSX(32, accessSize, EAX, R(EAX));
+				}
+
+				FixupBranch exit = J();
+				SetJumpTarget(fast);
+				UnsafeLoadToEAX(opAddress, accessSize, offset, signExtend);
+				SetJumpTarget(exit);
+			}
+		}
+	}
+}
+
 void EmuCodeBlock::UnsafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int accessSize, s32 offset, bool swap)
 {
 	if (accessSize == 8 && reg_value >= 4) {

Source/Core/Core/Src/PowerPC/JitCommon/Jit_Util.h

@@ -27,8 +27,10 @@ public:
 	void UnsafeLoadRegToReg(Gen::X64Reg reg_addr, Gen::X64Reg reg_value, int accessSize, s32 offset = 0, bool signExtend = false);
 	void UnsafeLoadRegToRegNoSwap(Gen::X64Reg reg_addr, Gen::X64Reg reg_value, int accessSize, s32 offset);
 	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 SafeLoadRegToEAX(Gen::X64Reg reg, int accessSize, s32 offset, bool signExtend = false);
 	void SafeWriteRegToReg(Gen::X64Reg reg_value, Gen::X64Reg reg_addr, int accessSize, s32 offset, bool swap = true);
+	void SafeLoadToEAX(const Gen::OpArg & opAddress, int accessSize, s32 offset, bool signExtend);
 
 	// Trashes both inputs and EAX.
 	void SafeWriteFloatToReg(Gen::X64Reg xmm_value, Gen::X64Reg reg_addr);