MMU Speed Optimisations:

* Un-cleaned the Load/Store code.  I think this will fix the stability issues from r6032
* Added memory exception checking to a couple more JIT FPU instructions
* Optimised the 64bit fast memory access slightly
* Optimised the MMU speed hack.

The rest of the speed optimisations from r6032 have been retained.


git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6034 8ced0084-cf51-0410-be5f-012b33b47a6e

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

@@ -44,9 +44,11 @@ void Jit64::lbzx(UGeckoInstruction inst)
 
 	int a = inst.RA, b = inst.RB, d = inst.RD;
 	gpr.FlushLockX(ABI_PARAM1);
+	gpr.Lock(b);
 	MOV(32, R(ABI_PARAM1), gpr.R(b));
 	if (a)
 	{
+		gpr.Lock(a);
 		ADD(32, R(ABI_PARAM1), gpr.R(a));
 	}
 
@@ -54,11 +56,13 @@ void Jit64::lbzx(UGeckoInstruction inst)
 
 	MEMCHECK_START
 
-	gpr.KillImmediate(d);
+	gpr.Lock(d);
+	gpr.LoadToX64(d, (b == d || a == d), true);
 	MOV(32, gpr.R(d), R(EAX));
 
 	MEMCHECK_END
 
+	gpr.UnlockAll();
 	gpr.UnlockAllX();
 }
 
@@ -69,9 +73,11 @@ void Jit64::lhax(UGeckoInstruction inst)
 
 	int a = inst.RA, b = inst.RB, d = inst.RD;
 	gpr.FlushLockX(ABI_PARAM1);
+	gpr.Lock(b);
 	MOV(32, R(ABI_PARAM1), gpr.R(b));
 	if (a)
 	{
+		gpr.Lock(a);
 		ADD(32, R(ABI_PARAM1), gpr.R(a));
 	}
 
@@ -80,11 +86,13 @@ void Jit64::lhax(UGeckoInstruction inst)
 
 	MEMCHECK_START
 
-	gpr.KillImmediate(d);
+	gpr.Lock(d);
+	gpr.LoadToX64(d, (b == d || a == d), true);
 	MOV(32, gpr.R(d), R(EAX));
 
 	MEMCHECK_END
 
+	gpr.UnlockAll();
 	gpr.UnlockAllX();
 }
 
@@ -95,9 +103,11 @@ void Jit64::lwzx(UGeckoInstruction inst)
 
 	int a = inst.RA, b = inst.RB, d = inst.RD;
 	gpr.FlushLockX(ABI_PARAM1);
+	gpr.Lock(b);
 	MOV(32, R(ABI_PARAM1), gpr.R(b));
 	if (a)
 	{
+		gpr.Lock(a);
 		ADD(32, R(ABI_PARAM1), gpr.R(a));
 	}
 
@@ -105,11 +115,13 @@ void Jit64::lwzx(UGeckoInstruction inst)
 
 	MEMCHECK_START
 
-	gpr.KillImmediate(d);
+	gpr.Lock(d);
+	gpr.LoadToX64(d, (b == d || a == d), true);
 	MOV(32, gpr.R(d), R(EAX));
 
 	MEMCHECK_END
 
+	gpr.UnlockAll();
 	gpr.UnlockAllX();
 }
 
@@ -207,21 +219,25 @@ void Jit64::lXz(UGeckoInstruction inst)
 		MOV(accessSize, gpr.R(d), MComplex(RBX, gpr.R(a).GetSimpleReg(), SCALE_1, offset));
 		BSWAP(32, gpr.R(d).GetSimpleReg());
 		gpr.UnlockAll();
-		gpr.Flush(FLUSH_ALL);
 	}
 	else
 	{
 		gpr.FlushLockX(ABI_PARAM1);
+		gpr.Lock(a);
+		gpr.LoadToX64(a, true, false);
+
 		MOV(32, R(ABI_PARAM1), gpr.R(a));
 		SafeLoadRegToEAX(ABI_PARAM1, accessSize, offset);
 		
 		MEMCHECK_START
 
-		gpr.KillImmediate(d);
+		gpr.Lock(d);
+		gpr.LoadToX64(d, a == d, true);
 		MOV(32, gpr.R(d), R(EAX));
 		
 		MEMCHECK_END
 
+		gpr.UnlockAll();
 		gpr.UnlockAllX();
 	}
 }
@@ -236,16 +252,19 @@ void Jit64::lha(UGeckoInstruction inst)
 	s32 offset = (s32)(s16)inst.SIMM_16;
 	// Safe and boring
 	gpr.FlushLockX(ABI_PARAM1);
+	gpr.Lock(a);
 	MOV(32, R(ABI_PARAM1), gpr.R(a));
 	SafeLoadRegToEAX(ABI_PARAM1, 16, offset, true);
 
 	MEMCHECK_START
 
-	gpr.KillImmediate(d);
+	gpr.Lock(d);
+	gpr.LoadToX64(d, d == a, true);
 	MOV(32, gpr.R(d), R(EAX));
 
 	MEMCHECK_END
 
+	gpr.UnlockAll();
 	gpr.UnlockAllX();
 }
 
@@ -268,12 +287,14 @@ void Jit64::lwzux(UGeckoInstruction inst)
 
 	MEMCHECK_START
 
-	gpr.KillImmediate(d);
+	gpr.Lock(d);
+	gpr.LoadToX64(d, b == d, true);
 	MOV(32, gpr.R(d), R(EAX));
 
 	MEMCHECK_END
 
 	gpr.UnlockAll();
+	gpr.UnlockAllX();
 }
 
 // Zero cache line.
@@ -392,7 +413,9 @@ void Jit64::stX(UGeckoInstruction inst)
 		//Still here? Do regular path.
 
 		gpr.FlushLockX(ABI_PARAM1, ABI_PARAM2);
-		gpr.Lock(a);
+		gpr.Lock(s, a);
+		if (update && offset)
+			gpr.LoadToX64(a, true, true);
 		MOV(32, R(ABI_PARAM2), gpr.R(a));
 		MOV(32, R(ABI_PARAM1), gpr.R(s));
 		SafeWriteRegToReg(ABI_PARAM1, ABI_PARAM2, accessSize, offset);
@@ -400,9 +423,8 @@ void Jit64::stX(UGeckoInstruction inst)
 		if (update && offset)
 		{
 			MEMCHECK_START
-
+			
-			gpr.KillImmediate(a);
+			ADD(32, gpr.R(a), Imm32(offset));
-			MOV(32, gpr.R(a), R(ABI_PARAM2));
 
 			MEMCHECK_END
 		}

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

@@ -260,8 +260,6 @@ void Jit64::stfs(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(LoadStoreFloating)
 
-	if (js.memcheck) { Default(inst); return; }
-
 	bool update = inst.OPCD & 1;
 	int s = inst.RS;
 	int a = inst.RA;
@@ -301,7 +299,11 @@ void Jit64::stfs(UGeckoInstruction inst)
 	ADD(32, R(ABI_PARAM2), Imm32(offset));
 	if (update && offset)
 	{
+		MEMCHECK_START
+
 		MOV(32, gpr.R(a), R(ABI_PARAM2));
+
+		MEMCHECK_END
 	}
 	CVTSD2SS(XMM0, fpr.R(s));
 	SafeWriteFloatToReg(XMM0, ABI_PARAM2);
@@ -316,8 +318,6 @@ void Jit64::stfsx(UGeckoInstruction inst)
 	INSTRUCTION_START
 	JITDISABLE(LoadStoreFloating)
 
-	if (js.memcheck) { Default(inst); return; }
-
 	// We can take a shortcut here - it's not likely that a hardware access would use this instruction.
 	gpr.FlushLockX(ABI_PARAM1);
 	fpr.Lock(inst.RS);

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

@@ -74,8 +74,9 @@ void EmuCodeBlock::UnsafeLoadRegToRegNoSwap(X64Reg reg_addr, X64Reg reg_value, i
 
 void EmuCodeBlock::SafeLoadRegToEAX(X64Reg reg_addr, int accessSize, s32 offset, bool signExtend)
 {
-	if (Core::g_CoreStartupParameter.bUseFastMem && accessSize == 32 && !Core::g_CoreStartupParameter.bMMU)
+	if (Core::g_CoreStartupParameter.bUseFastMem && (accessSize == 32 || accessSize == 8) && !Core::g_CoreStartupParameter.bMMU)
 	{
+		// FIXME: accessSize == 16 does not work.  Breaks mkdd
 		UnsafeLoadRegToReg(reg_addr, EAX, accessSize, offset, signExtend);
 	}
 	else
@@ -87,8 +88,11 @@ void EmuCodeBlock::SafeLoadRegToEAX(X64Reg reg_addr, int accessSize, s32 offset,
 		FixupBranch addr20;
 		if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
 		{
-			CMP(32, R(reg_addr), Imm32(0xf0000000));
+			if (Core::g_CoreStartupParameter.bMMU)
-			addrf0 = J_CC(CC_GE);
+			{
+				CMP(32, R(reg_addr), Imm32(0xf0000000));
+				addrf0 = J_CC(CC_GE);
+			}
 			TEST(32, R(reg_addr), Imm32(0x20000000));
 			addr20 = J_CC(CC_NZ);
 		}
@@ -98,8 +102,11 @@ void EmuCodeBlock::SafeLoadRegToEAX(X64Reg reg_addr, int accessSize, s32 offset,
 
 		if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
 		{
+			if (Core::g_CoreStartupParameter.bMMU)
+			{
+				SetJumpTarget(addrf0);				
+			}			
 			SetJumpTarget(addr20);
-			SetJumpTarget(addrf0);
 		}
 
 		switch (accessSize)
@@ -144,8 +151,7 @@ void EmuCodeBlock::SafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int acce
 	// TODO: Figure out a cleaner way to check memory bounds
 	FixupBranch addrf0;
 	FixupBranch addr20;
-	FixupBranch fast;
+	if (Core::g_CoreStartupParameter.bMMU)
-	if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
 	{
 		CMP(32, R(reg_addr), Imm32(0xf0000000));
 		addrf0 = J_CC(CC_GE);
@@ -153,16 +159,13 @@ void EmuCodeBlock::SafeWriteRegToReg(X64Reg reg_value, X64Reg reg_addr, int acce
 		addr20 = J_CC(CC_NZ);
 	}
 
-	if (!Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.bUseFastMem)
+	TEST(32, R(reg_addr), Imm32(0x0C000000));
-	{
+	FixupBranch fast = J_CC(CC_Z);
-		TEST(32, R(reg_addr), Imm32(0x0C000000));
-		fast = J_CC(CC_Z);
-	}
 
-	if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+	if (Core::g_CoreStartupParameter.bMMU)
 	{
-		SetJumpTarget(addr20);
 		SetJumpTarget(addrf0);
+		SetJumpTarget(addr20);
 	}
 
 	switch (accessSize)