Further optimised the JIT cache lookup. Attempted to fix the Wii games in ICC builds.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6136 8ced0084-cf51-0410-be5f-012b33b47a6e

2 files changed, 135 insertions, 147 deletions

diff --git a/Source/Core/Core/Src/PowerPC/Jit64/JitAsm.cpp b/Source/Core/Core/Src/PowerPC/Jit64/JitAsm.cpp
index 511e1b5dbb..cc8b6edfda 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64/JitAsm.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64/JitAsm.cpp
@@ -77,7 +77,7 @@ void Jit64AsmRoutineManager::Generate()
 		dispatcher = GetCodePtr();
 			// The result of slice decrementation should be in flags if somebody jumped here
 			// IMPORTANT - We jump on negative, not carry!!!
-			FixupBranch bail = J_CC(CC_BE);
+			FixupBranch bail = J_CC(CC_BE, true);
 
 			if (Core::g_CoreStartupParameter.bEnableDebugging)
 			{
@@ -95,8 +95,72 @@ void Jit64AsmRoutineManager::Generate()
 			MOV(32, R(EAX), M(&PowerPC::ppcState.pc));
 			dispatcherPcInEAX = GetCodePtr();
 
-			FixupBranch needinst = J(true);
-			const u8* haveinst = GetCodePtr();
+#ifdef JIT_UNLIMITED_ICACHE
+			u32 mask = 0;
+			FixupBranch no_mem;
+			FixupBranch exit_mem;
+			FixupBranch exit_vmem;
+			if (Core::g_CoreStartupParameter.bWii)
+				mask = JIT_ICACHE_EXRAM_BIT;
+			if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+				mask |= JIT_ICACHE_VMEM_BIT;
+			if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+			{
+				TEST(32, R(EAX), Imm32(mask));
+				no_mem = J_CC(CC_NZ);
+			}
+			AND(32, R(EAX), Imm32(JIT_ICACHE_MASK));
+#ifdef _M_IX86
+			MOV(32, R(EAX), MDisp(EAX, (u32)jit->GetBlockCache()->GetICache()));
+#else
+			MOV(64, R(RSI), Imm64((u64)jit->GetBlockCache()->GetICache()));
+			MOV(32, R(EAX), MComplex(RSI, EAX, SCALE_1, 0));
+#endif
+			if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+			{
+				exit_mem = J();
+				SetJumpTarget(no_mem);
+			}
+			if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+			{
+				TEST(32, R(EAX), Imm32(JIT_ICACHE_VMEM_BIT));
+				FixupBranch no_vmem = J_CC(CC_Z);
+				AND(32, R(EAX), Imm32(JIT_ICACHE_MASK));
+#ifdef _M_IX86
+				MOV(32, R(EAX), MDisp(EAX, (u32)jit->GetBlockCache()->GetICacheVMEM()));
+#else
+				MOV(64, R(RSI), Imm64((u64)jit->GetBlockCache()->GetICacheVMEM()));
+				MOV(32, R(EAX), MComplex(RSI, EAX, SCALE_1, 0));
+#endif
+				if (Core::g_CoreStartupParameter.bWii) exit_vmem = J();
+				SetJumpTarget(no_vmem);
+			}
+			if (Core::g_CoreStartupParameter.bWii)
+			{
+				TEST(32, R(EAX), Imm32(JIT_ICACHE_EXRAM_BIT));
+				FixupBranch no_exram = J_CC(CC_Z);
+				AND(32, R(EAX), Imm32(JIT_ICACHEEX_MASK));
+#ifdef _M_IX86
+				MOV(32, R(EAX), MDisp(EAX, (u32)jit->GetBlockCache()->GetICacheEx()));
+#else
+				MOV(64, R(RSI), Imm64((u64)jit->GetBlockCache()->GetICacheEx()));
+				MOV(32, R(EAX), MComplex(RSI, EAX, SCALE_1, 0));
+#endif
+				SetJumpTarget(no_exram);
+			}
+			if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
+				SetJumpTarget(exit_mem);
+			if (Core::g_CoreStartupParameter.bWii && (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack))
+				SetJumpTarget(exit_vmem);
+#else
+#ifdef _M_IX86
+			AND(32, R(EAX), Imm32(Memory::MEMVIEW32_MASK));
+			MOV(32, R(EBX), Imm32((u32)Memory::base));
+			MOV(32, R(EAX), MComplex(EBX, EAX, SCALE_1, 0));
+#else
+			MOV(32, R(EAX), MComplex(RBX, RAX, SCALE_1, 0));
+#endif
+#endif
 
 			TEST(32, R(EAX), Imm32(0xFC));
 			FixupBranch notfound = J_CC(CC_NZ);
@@ -133,7 +197,7 @@ void Jit64AsmRoutineManager::Generate()
 			ABI_CallFunction(reinterpret_cast<void *>(&PowerPC::CheckExceptions));
 			MOV(32, R(EAX), M(&NPC));
 			MOV(32, M(&PC), R(EAX));
-			JMP(dispatcher);
+			JMP(dispatcher, true);
 
 		SetJumpTarget(bail);
 		doTiming = GetCodePtr();
@@ -157,77 +221,6 @@ void Jit64AsmRoutineManager::Generate()
 	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	RET();
 
-
-	SetJumpTarget(needinst);
-#ifdef JIT_UNLIMITED_ICACHE
-	u32 mask = 0;
-	FixupBranch no_mem;
-	FixupBranch exit_mem;
-	FixupBranch exit_vmem;
-	if (Core::g_CoreStartupParameter.bWii)
-		mask = JIT_ICACHE_EXRAM_BIT;
-	if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
-		mask |= JIT_ICACHE_VMEM_BIT;
-	if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
-	{
-		TEST(32, R(EAX), Imm32(mask));
-		no_mem = J_CC(CC_NZ);
-	}
-	AND(32, R(EAX), Imm32(JIT_ICACHE_MASK));
-#ifdef _M_IX86
-	MOV(32, R(EAX), MDisp(EAX, (u32)jit->GetBlockCache()->GetICache()));
-#else
-	MOV(64, R(RSI), Imm64((u64)jit->GetBlockCache()->GetICache()));
-	MOV(32, R(EAX), MComplex(RSI, EAX, SCALE_1, 0));
-#endif
-	if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
-	{
-		exit_mem = J();
-		SetJumpTarget(no_mem);
-	}
-	if (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
-	{
-		TEST(32, R(EAX), Imm32(JIT_ICACHE_VMEM_BIT));
-		FixupBranch no_vmem = J_CC(CC_Z);
-		AND(32, R(EAX), Imm32(JIT_ICACHE_MASK));
-#ifdef _M_IX86
-		MOV(32, R(EAX), MDisp(EAX, (u32)jit->GetBlockCache()->GetICacheVMEM()));
-#else
-		MOV(64, R(RSI), Imm64((u64)jit->GetBlockCache()->GetICacheVMEM()));
-		MOV(32, R(EAX), MComplex(RSI, EAX, SCALE_1, 0));
-#endif
-		if (Core::g_CoreStartupParameter.bWii) exit_vmem = J();
-		SetJumpTarget(no_vmem);
-	}
-	if (Core::g_CoreStartupParameter.bWii)
-	{
-		TEST(32, R(EAX), Imm32(JIT_ICACHE_EXRAM_BIT));
-		FixupBranch no_exram = J_CC(CC_Z);
-		AND(32, R(EAX), Imm32(JIT_ICACHEEX_MASK));
-#ifdef _M_IX86
-		MOV(32, R(EAX), MDisp(EAX, (u32)jit->GetBlockCache()->GetICacheEx()));
-#else
-		MOV(64, R(RSI), Imm64((u64)jit->GetBlockCache()->GetICacheEx()));
-		MOV(32, R(EAX), MComplex(RSI, EAX, SCALE_1, 0));
-#endif
-		SetJumpTarget(no_exram);
-	}
-	if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
-		SetJumpTarget(exit_mem);
-	if (Core::g_CoreStartupParameter.bWii)
-		SetJumpTarget(exit_vmem);
-#else
-#ifdef _M_IX86
-	AND(32, R(EAX), Imm32(Memory::MEMVIEW32_MASK));
-	MOV(32, R(EBX), Imm32((u32)Memory::base));
-	MOV(32, R(EAX), MComplex(EBX, EAX, SCALE_1, 0));
-#else
-	MOV(32, R(EAX), MComplex(RBX, RAX, SCALE_1, 0));
-#endif
-#endif
-	JMP(haveinst, true);
-
-
 	breakpointBailout = GetCodePtr();
 	//Landing pad for drec space
 	ABI_PopAllCalleeSavedRegsAndAdjustStack();
diff --git a/Source/Core/Core/Src/PowerPC/Jit64IL/JitILAsm.cpp b/Source/Core/Core/Src/PowerPC/Jit64IL/JitILAsm.cpp
index 38ba5ec1ca..cd2708512f 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64IL/JitILAsm.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64IL/JitILAsm.cpp
@@ -79,7 +79,7 @@ void JitILAsmRoutineManager::Generate()
 			//This is the place for CPUCompare!
 
 			//The result of slice decrement should be in flags if somebody jumped here
-			FixupBranch bail = J_CC(CC_BE);
+			FixupBranch bail = J_CC(CC_BE, true);
 
 			if (Core::g_CoreStartupParameter.bEnableDebugging)
 			{
@@ -97,75 +97,6 @@ void JitILAsmRoutineManager::Generate()
 			MOV(32, R(EAX), M(&PowerPC::ppcState.pc));
 			dispatcherPcInEAX = GetCodePtr();
 
-			FixupBranch needinst = J(true);
-			const u8* haveinst = GetCodePtr();
-
-			TEST(32, R(EAX), Imm32(0xFC));
-			FixupBranch notfound = J_CC(CC_NZ);
-				BSWAP(32, EAX);
-				//IDEA - we have 26 bits, why not just use offsets from base of code?
-				if (enableDebug)
-				{
-					ADD(32, M(&PowerPC::ppcState.DebugCount), Imm8(1));
-				}
-				//grab from list and jump to it
-#ifdef _M_IX86
-				MOV(32, R(EDX), ImmPtr(jit->GetBlockCache()->GetCodePointers()));
-				JMPptr(MComplex(EDX, EAX, 4, 0));
-#else
-				JMPptr(MComplex(R15, RAX, 8, 0));
-#endif
-			SetJumpTarget(notfound);
-
-			//Ok, no block, let's jit
-#ifdef _M_IX86
-			ABI_AlignStack(4);
-			PUSH(32, M(&PowerPC::ppcState.pc));
-			CALL(reinterpret_cast<void *>(&Jit));
-			ABI_RestoreStack(4);
-#else
-			MOV(32, R(ABI_PARAM1), M(&PowerPC::ppcState.pc));
-			CALL((void *)&Jit);
-#endif
-			JMP(dispatcherNoCheck); // no point in special casing this
-
-			//FP blocks test for FPU available, jump here if false
-			fpException = AlignCode4(); 
-			MOV(32, R(EAX), M(&PC));
-			MOV(32, M(&NPC), R(EAX));
-			OR(32, M(&PowerPC::ppcState.Exceptions), Imm32(EXCEPTION_FPU_UNAVAILABLE));
-			ABI_CallFunction(reinterpret_cast<void *>(&PowerPC::CheckExceptions));
-			MOV(32, R(EAX), M(&NPC));
-			MOV(32, M(&PC), R(EAX));
-			JMP(dispatcher);
-
-		SetJumpTarget(bail);
-		doTiming = GetCodePtr();
-
-		ABI_CallFunction(reinterpret_cast<void *>(&CoreTiming::Advance));
-		
-		testExceptions = GetCodePtr();
-		TEST(32, M(&PowerPC::ppcState.Exceptions), Imm32(0xFFFFFFFF));
-		FixupBranch skipExceptions = J_CC(CC_Z);
-			MOV(32, R(EAX), M(&PC));
-			MOV(32, M(&NPC), R(EAX));
-			ABI_CallFunction(reinterpret_cast<void *>(&PowerPC::CheckExceptions));
-			MOV(32, R(EAX), M(&NPC));
-			MOV(32, M(&PC), R(EAX));
-		SetJumpTarget(skipExceptions);
-		
-		TEST(32, M((void*)PowerPC::GetStatePtr()), Imm32(0xFFFFFFFF));
-		J_CC(CC_Z, outerLoop, true);
-	//Landing pad for drec space
-	ABI_PopAllCalleeSavedRegsAndAdjustStack();
-	RET();
-
-	breakpointBailout = GetCodePtr();
-	//Landing pad for drec space
-	ABI_PopAllCalleeSavedRegsAndAdjustStack();
-	RET();
-
-			SetJumpTarget(needinst);
 #ifdef JIT_UNLIMITED_ICACHE
 			u32 mask = 0;
 			FixupBranch no_mem;
@@ -221,7 +152,7 @@ void JitILAsmRoutineManager::Generate()
 			}
 			if (Core::g_CoreStartupParameter.bWii || Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack)
 				SetJumpTarget(exit_mem);
-			if (Core::g_CoreStartupParameter.bWii)
+			if (Core::g_CoreStartupParameter.bWii && (Core::g_CoreStartupParameter.bMMU || Core::g_CoreStartupParameter.iTLBHack))
 				SetJumpTarget(exit_vmem);
 #else
 #ifdef _M_IX86
@@ -232,7 +163,71 @@ void JitILAsmRoutineManager::Generate()
 			MOV(32, R(EAX), MComplex(RBX, RAX, SCALE_1, 0));
 #endif
 #endif
-			JMP(haveinst, true);
+
+			TEST(32, R(EAX), Imm32(0xFC));
+			FixupBranch notfound = J_CC(CC_NZ);
+				BSWAP(32, EAX);
+				//IDEA - we have 26 bits, why not just use offsets from base of code?
+				if (enableDebug)
+				{
+					ADD(32, M(&PowerPC::ppcState.DebugCount), Imm8(1));
+				}
+				//grab from list and jump to it
+#ifdef _M_IX86
+				MOV(32, R(EDX), ImmPtr(jit->GetBlockCache()->GetCodePointers()));
+				JMPptr(MComplex(EDX, EAX, 4, 0));
+#else
+				JMPptr(MComplex(R15, RAX, 8, 0));
+#endif
+			SetJumpTarget(notfound);
+
+			//Ok, no block, let's jit
+#ifdef _M_IX86
+			ABI_AlignStack(4);
+			PUSH(32, M(&PowerPC::ppcState.pc));
+			CALL(reinterpret_cast<void *>(&Jit));
+			ABI_RestoreStack(4);
+#else
+			MOV(32, R(ABI_PARAM1), M(&PowerPC::ppcState.pc));
+			CALL((void *)&Jit);
+#endif
+			JMP(dispatcherNoCheck); // no point in special casing this
+
+			//FP blocks test for FPU available, jump here if false
+			fpException = AlignCode4(); 
+			MOV(32, R(EAX), M(&PC));
+			MOV(32, M(&NPC), R(EAX));
+			OR(32, M(&PowerPC::ppcState.Exceptions), Imm32(EXCEPTION_FPU_UNAVAILABLE));
+			ABI_CallFunction(reinterpret_cast<void *>(&PowerPC::CheckExceptions));
+			MOV(32, R(EAX), M(&NPC));
+			MOV(32, M(&PC), R(EAX));
+			JMP(dispatcher, true);
+
+		SetJumpTarget(bail);
+		doTiming = GetCodePtr();
+
+		ABI_CallFunction(reinterpret_cast<void *>(&CoreTiming::Advance));
+		
+		testExceptions = GetCodePtr();
+		TEST(32, M(&PowerPC::ppcState.Exceptions), Imm32(0xFFFFFFFF));
+		FixupBranch skipExceptions = J_CC(CC_Z);
+			MOV(32, R(EAX), M(&PC));
+			MOV(32, M(&NPC), R(EAX));
+			ABI_CallFunction(reinterpret_cast<void *>(&PowerPC::CheckExceptions));
+			MOV(32, R(EAX), M(&NPC));
+			MOV(32, M(&PC), R(EAX));
+		SetJumpTarget(skipExceptions);
+		
+		TEST(32, M((void*)PowerPC::GetStatePtr()), Imm32(0xFFFFFFFF));
+		J_CC(CC_Z, outerLoop, true);
+	//Landing pad for drec space
+	ABI_PopAllCalleeSavedRegsAndAdjustStack();
+	RET();
+
+	breakpointBailout = GetCodePtr();
+	//Landing pad for drec space
+	ABI_PopAllCalleeSavedRegsAndAdjustStack();
+	RET();
 
 	GenerateCommon();
 }