x86jit: Add proxy blocks for continuing.

Core/MIPS/JitCommon/JitState.h

@@ -64,6 +64,8 @@ namespace MIPSComp {
 
 		u32 compilerPC;
 		u32 blockStart;
+		u32 lastContinuedPC;
+		u32 initialBlockSize;
 		int nextExit;
 		bool cancel;
 		bool inDelaySlot;

Core/MIPS/x86/CompBranch.cpp

@@ -215,6 +215,7 @@ void Jit::CompBranchExits(CCFlags cc, u32 targetAddr, u32 notTakenAddr, bool del
 			if (likely)
 				CompileDelaySlot(DELAYSLOT_NICE);
 
+			AddContinuedBlock(targetAddr);
 			// Account for the increment in the loop.
 			js.compilerPC = targetAddr - 4;
 			// In case the delay slot was a break or something.
@@ -328,6 +329,7 @@ void Jit::BranchRSRTComp(MIPSOpcode op, Gen::CCFlags cc, bool likely)
 
 		// Branch taken.  Always compile the delay slot, and then go to dest.
 		CompileDelaySlot(DELAYSLOT_NICE);
+		AddContinuedBlock(targetAddr);
 		// Account for the increment in the loop.
 		js.compilerPC = targetAddr - 4;
 		// In case the delay slot was a break or something.
@@ -406,6 +408,7 @@ void Jit::BranchRSZeroComp(MIPSOpcode op, Gen::CCFlags cc, bool andLink, bool li
 		if (andLink)
 			gpr.SetImm(MIPS_REG_RA, js.compilerPC + 8);
 
+		AddContinuedBlock(targetAddr);
 		// Account for the increment in the loop.
 		js.compilerPC = targetAddr - 4;
 		// In case the delay slot was a break or something.
@@ -585,6 +588,7 @@ void Jit::Comp_Jump(MIPSOpcode op) {
 		CompileDelaySlot(DELAYSLOT_NICE);
 		if (jo.continueJumps && js.numInstructions < jo.continueMaxInstructions)
 		{
+			AddContinuedBlock(targetAddr);
 			// Account for the increment in the loop.
 			js.compilerPC = targetAddr - 4;
 			// In case the delay slot was a break or something.
@@ -609,6 +613,7 @@ void Jit::Comp_Jump(MIPSOpcode op) {
 		CompileDelaySlot(DELAYSLOT_NICE);
 		if (jo.continueJumps && js.numInstructions < jo.continueMaxInstructions)
 		{
+			AddContinuedBlock(targetAddr);
 			// Account for the increment in the loop.
 			js.compilerPC = targetAddr - 4;
 			// In case the delay slot was a break or something.
@@ -679,6 +684,7 @@ void Jit::Comp_JumpReg(MIPSOpcode op)
 
 		if (jo.continueJumps && gpr.IsImm(rs) && js.numInstructions < jo.continueMaxInstructions)
 		{
+			AddContinuedBlock(gpr.GetImm(rs));
 			// Account for the increment in the loop.
 			js.compilerPC = gpr.GetImm(rs) - 4;
 			// In case the delay slot was a break or something.

Core/MIPS/x86/Jit.cpp

@@ -116,8 +116,6 @@ static void JitLogMiss(MIPSOpcode op)
 JitOptions::JitOptions()
 {
 	enableBlocklink = true;
-	// WARNING: These options don't work properly with cache clearing.
-	// Need to find a smart way to handle before enabling.
 	immBranches = false;
 	continueBranches = false;
 	continueJumps = false;
@@ -396,6 +394,8 @@ const u8 *Jit::DoJit(u32 em_address, JitBlock *b)
 {
 	js.cancel = false;
 	js.blockStart = js.compilerPC = mips_->pc;
+	js.lastContinuedPC = 0;
+	js.initialBlockSize = 0;
 	js.nextExit = 0;
 	js.downcountAmount = 0;
 	js.curBlock = b;
@@ -466,10 +466,27 @@ const u8 *Jit::DoJit(u32 em_address, JitBlock *b)
 	b->codeSize = (u32)(GetCodePtr() - b->normalEntry);
 	NOP();
 	AlignCode4();
-	b->originalSize = js.numInstructions;
+	if (js.lastContinuedPC == 0)
+		b->originalSize = js.numInstructions;
+	else
+	{
+		// We continued at least once.  Add the last proxy and set the originalSize correctly.
+		blocks.ProxyBlock(js.blockStart, js.lastContinuedPC, (js.compilerPC - js.lastContinuedPC) / sizeof(u32), GetCodePtr());
+		b->originalSize = js.initialBlockSize;
+	}
 	return b->normalEntry;
 }
 
+void Jit::AddContinuedBlock(u32 dest)
+{
+	// The first block is the root block.  When we continue, we create proxy blocks after that.
+	if (js.lastContinuedPC == 0)
+		js.initialBlockSize = js.numInstructions;
+	else
+		blocks.ProxyBlock(js.blockStart, js.lastContinuedPC, (js.compilerPC - js.lastContinuedPC) / sizeof(u32), GetCodePtr());
+	js.lastContinuedPC = dest;
+}
+
 bool Jit::DescribeCodePtr(const u8 *ptr, std::string &name)
 {
 	u32 jitAddr = blocks.GetAddressFromBlockPtr(ptr);

Core/MIPS/x86/Jit.h

@@ -193,6 +193,7 @@ private:
 		CompileDelaySlot(flags, &state);
 	}
 	void EatInstruction(MIPSOpcode op);
+	void AddContinuedBlock(u32 dest);
 
 	void WriteExit(u32 destination, int exit_num);
 	void WriteExitDestInReg(X64Reg reg);