Experimental: Implement skid's MMU Demand Paging in JitIL.

This one needs a lot of testing, since I don't have any games that need it, except Rogue Leader (GSWP64) - and I didn't do more "testing" than watching the stormtroopers dance due to speed reasons (altho it seems twice as fast as JIT for me; it does spit out a lot more warnings too)

btw, ITS OVER 6000!!111

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6006 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Core/Src/PowerPC/Jit64IL/IR.h

@@ -163,6 +163,10 @@ enum Opcode {
 	ShortIdleLoop, // Idle loop seen in homebrew like wii mahjong,
 		       // just a branch
 
+	// used for MMU, at least until someone 
+	// has a better idea of integrating it
+	FPExceptionCheckStart, FPExceptionCheckEnd,
+	ISIException,
 	// "Opcode" representing a register too far away to
 	// reference directly; this is a size optimization
 	Tramp,
@@ -389,6 +393,15 @@ public:
 	InstLoc EmitSystemCall(InstLoc pc) {
 		return FoldUOp(SystemCall, pc);
 	}
+	InstLoc EmitFPExceptionCheckStart(InstLoc pc) {
+		return EmitUOp(FPExceptionCheckStart, pc);
+	}
+	InstLoc EmitFPExceptionCheckEnd(InstLoc pc) {
+		return EmitUOp(FPExceptionCheckEnd, pc);
+	}
+	InstLoc EmitISIException(InstLoc dest) {
+		return EmitUOp(ISIException, dest);
+	}
 	InstLoc EmitRFIExit() {
 		return FoldZeroOp(RFIExit, 0);
 	}		

Source/Core/Core/Src/PowerPC/Jit64IL/IR_X86.cpp

@@ -705,6 +705,9 @@ static void DoWriteCode(IRBuilder* ibuild, JitIL* Jit, bool UseProfile, bool Mak
 		case RFIExit:
 		case InterpreterBranch:		
 		case ShortIdleLoop:
+		case FPExceptionCheckStart:
+		case FPExceptionCheckEnd:
+		case ISIException:
 		case Int3:
 		case Tramp:
 			// No liveness effects
@@ -1607,10 +1610,9 @@ static void DoWriteCode(IRBuilder* ibuild, JitIL* Jit, bool UseProfile, bool Mak
 		}
 		case SystemCall: {
 			unsigned InstLoc = ibuild->GetImmValue(getOp1(I));
-			Jit->Cleanup();
 			Jit->OR(32, M(&PowerPC::ppcState.Exceptions), Imm32(EXCEPTION_SYSCALL));
 			Jit->MOV(32, M(&PC), Imm32(InstLoc + 4));
-			Jit->JMP(((JitIL *)jit)->asm_routines.testExceptions, true);
+			Jit->WriteExceptionExit();
 			break;
 		}
 		case InterpreterBranch: {
@@ -1627,14 +1629,46 @@ static void DoWriteCode(IRBuilder* ibuild, JitIL* Jit, bool UseProfile, bool Mak
 			Jit->AND(32, R(EAX), Imm32(~mask));
 			Jit->AND(32, R(ECX), Imm32(mask));
 			Jit->OR(32, R(EAX), R(ECX));       
-			// MSR &= 0xFFFDFFFF; //TODO: VERIFY
+			// MSR &= 0xFFFBFFFF; // Mask used to clear the bit MSR[13]
-			Jit->AND(32, R(EAX), Imm32(0xFFFDFFFF));
+			Jit->AND(32, R(EAX), Imm32(0xFFFBFFFF));
 			Jit->MOV(32, M(&MSR), R(EAX));
 			// NPC = SRR0; 
 			Jit->MOV(32, R(EAX), M(&SRR0));
 			Jit->WriteRfiExitDestInEAX();
 			break;
 		}
+		case FPExceptionCheckStart: {
+			unsigned InstLoc = ibuild->GetImmValue(getOp1(I));
+			//This instruction uses FPU - needs to add FP exception bailout
+			Jit->TEST(32, M(&PowerPC::ppcState.msr), Imm32(1 << 13)); // Test FP enabled bit
+			FixupBranch b1 = Jit->J_CC(CC_NZ);
+
+			// If a FPU exception occurs, the exception handler will read
+			// from PC.  Update PC with the latest value in case that happens.
+			Jit->MOV(32, M(&PC), Imm32(InstLoc));
+			Jit->SUB(32, M(&CoreTiming::downcount), Jit->js.downcountAmount > 127 ? Imm32(Jit->js.downcountAmount) : Imm8(Jit->js.downcountAmount)); 
+			Jit->JMP(Jit->asm_routines.fpException, true);
+			Jit->SetJumpTarget(b1);
+			break;
+		}
+		case FPExceptionCheckEnd: {
+			unsigned InstLoc = ibuild->GetImmValue(getOp1(I));
+			Jit->TEST(32, M(&PowerPC::ppcState.Exceptions), Imm32(EXCEPTION_DSI));
+			FixupBranch noMemException = Jit->J_CC(CC_Z);
+
+			// If a memory exception occurs, the exception handler will read
+			// from PC.  Update PC with the latest value in case that happens.
+			Jit->MOV(32, M(&PC), Imm32(InstLoc));
+			Jit->WriteExceptionExit();
+			Jit->SetJumpTarget(noMemException);
+			break;
+		}
+		case ISIException: {
+			unsigned InstLoc = ibuild->GetImmValue(getOp1(I));
+			Jit->ABI_CallFunctionC(reinterpret_cast<void *>(&Memory::GenerateISIException_JIT), InstLoc);
+			Jit->WriteExceptionExit();
+			break;
+		}
 		case Int3: {
 			Jit->INT3();
 			break;

Source/Core/Core/Src/PowerPC/Jit64IL/JitIL.cpp

@@ -172,7 +172,11 @@ void JitIL::Init()
 	}
 	else
 	{
-		jo.enableBlocklink = true;  // Speed boost, but not 100% safe
+		if (!Core::g_CoreStartupParameter.bJITBlockLinking)
+			jo.enableBlocklink = false;
+		else
+			// Speed boost, but not 100% safe
+			jo.enableBlocklink = !Core::g_CoreStartupParameter.bMMU;
 	}
 
 #ifdef _M_X64
@@ -185,6 +189,7 @@ void JitIL::Init()
 	jo.optimizeGatherPipe = true;
 	jo.fastInterrupts = false;
 	jo.accurateSinglePrecision = false;
+	js.memcheck = Core::g_CoreStartupParameter.bMMU;
 
 	trampolines.Init();
 	AllocCodeSpace(CODE_SIZE);
@@ -336,11 +341,10 @@ void JitIL::WriteRfiExitDestInEAX()
 	JMP(asm_routines.testExceptions, true);
 }
 
-void JitIL::WriteExceptionExit(u32 exception)
+void JitIL::WriteExceptionExit()
 {
 	Cleanup();
-	OR(32, M(&PowerPC::ppcState.Exceptions), Imm32(exception));
+	SUB(32, M(&CoreTiming::downcount), js.downcountAmount > 127 ? Imm32(js.downcountAmount) : Imm8(js.downcountAmount)); 
-	MOV(32, M(&PC), Imm32(js.compilerPC + 4));
 	JMP(asm_routines.testExceptions, true);
 }
 
@@ -401,6 +405,9 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 {
 	int blockSize = code_buf->GetSize();
 
+	// Memory exception on instruction fetch
+	bool memory_exception = false;
+
 	// A broken block is a block that does not end in a branch
 	bool broken_block = false;
 
@@ -414,7 +421,21 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 	if (em_address == 0)
 		PanicAlert("ERROR : Trying to compile at 0. LR=%08x", LR);
 
-	int size;
+	if (Core::g_CoreStartupParameter.bMMU &&
+				(em_address >> 28) != 0x0 &&
+				(em_address >> 28) != 0x8 &&
+				(em_address >> 28) != 0x9 &&
+				(em_address >> 28) != 0xC &&
+				(em_address >> 28) != 0xD)	
+	{
+		if (!Memory::TranslateAddress(em_address, Memory::FLAG_OPCODE))
+		{
+			// Memory exception occurred during instruction fetch
+			memory_exception = true;
+		}
+	}
+
+	int size = 0;
 	js.isLastInstruction = false;
 	js.blockStart = em_address;
 	js.fifoBytesThisBlock = 0;
@@ -423,7 +444,12 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 
 	//Analyze the block, collect all instructions it is made of (including inlining,
 	//if that is enabled), reorder instructions for optimal performance, and join joinable instructions.
+	b->exitAddress[0] = em_address;
+	if (!memory_exception)
+	{
+		// If there is a memory exception inside a block (broken_block==true), compile up to that instruction.
 		b->exitAddress[0] = PPCAnalyst::Flatten(em_address, &size, &js.st, &js.gpa, &js.fpa, broken_block, code_buf, blockSize);
+	}
 	PPCAnalyst::CodeOp *ops = code_buf->codebuffer;
 
 	const u8 *start = AlignCode4(); //TODO: Test if this or AlignCode16 make a difference from GetCodePtr
@@ -459,8 +485,7 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 	ibuild.Reset();
 
 	
-	js.downcountAmount = js.st.numCycles;
+	js.downcountAmount = 0;
-
 	if (!Core::g_CoreStartupParameter.bEnableDebugging)
 		js.downcountAmount += PatchEngine::GetSpeedhackCycles(em_address);
 
@@ -470,6 +495,9 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 		js.compilerPC = ops[i].address;
 		js.op = &ops[i];
 		js.instructionNumber = i;
+		const GekkoOPInfo *opinfo = GetOpInfo(ops[i].inst);
+		js.downcountAmount += (opinfo->numCyclesMinusOne + 1);
+
 		if (i == (int)size - 1)
 		{
 			js.isLastInstruction = true;
@@ -484,8 +512,23 @@ const u8* JitIL::DoJit(u32 em_address, PPCAnalyst::CodeBuffer *code_buf, JitBloc
 
 		if (!ops[i].skip)
 		{
-			JitILTables::CompileInstruction(ops[i]);
+			if (js.memcheck && (opinfo->flags & FL_USE_FPU))
+			{
+				ibuild.EmitFPExceptionCheckStart(ibuild.EmitIntConst(ops[i].address));
 			}
+			
+			JitILTables::CompileInstruction(ops[i]);
+
+			if (js.memcheck && (opinfo->flags & FL_LOADSTORE))
+			{
+				ibuild.EmitFPExceptionCheckEnd(ibuild.EmitIntConst(ops[i].address));
+			}
+		}
+	}
+
+	if (memory_exception)
+	{
+		ibuild.EmitISIException(ibuild.EmitIntConst(em_address));
 	}
 
 	// Perform actual code generation

Source/Core/Core/Src/PowerPC/Jit64IL/JitIL.h

@@ -112,7 +112,7 @@ public:
 
 	void WriteExit(u32 destination, int exit_num);
 	void WriteExitDestInEAX(int exit_num);
-	void WriteExceptionExit(u32 exception);
+	void WriteExceptionExit();
 	void WriteRfiExitDestInEAX();
 	void WriteCallInterpreter(UGeckoInstruction _inst);
 	void Cleanup();

Source/Core/Core/Src/PowerPC/Jit64IL/JitIL_Integer.cpp

@@ -125,6 +125,7 @@ void JitIL::cmpXX(UGeckoInstruction inst)
 		rhs = ibuild.EmitIntConst(inst.SIMM_16);
 		res = ibuild.EmitICmpCRSigned(lhs, rhs);
 	}
+	js.downcountAmount++; //TODO: should this be somewhere else?
 	
 	ibuild.EmitStoreCR(res, inst.CRFD);
 }