LLE JIT: Changed the ASM dispatcher so that it jumps to the blocks instead of calling them. This removes the need to push and pop all of the registers at each block, speeding up the JIT. Changed the cycle counting to work off memory accesses instead of a register. Removed the C++ JIT dispatcher because it will no longer work with this new format.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6858 8ced0084-cf51-0410-be5f-012b33b47a6e

4 files changed, 21 insertions, 79 deletions

diff --git a/Source/Core/DSPCore/Src/DSPCore.cpp b/Source/Core/DSPCore/Src/DSPCore.cpp
index 42da6410f9..46d9ee1f38 100644
--- a/Source/Core/DSPCore/Src/DSPCore.cpp
+++ b/Source/Core/DSPCore/Src/DSPCore.cpp
@@ -241,19 +241,14 @@ int DSPCore_RunCycles(int cycles)
 {
 	if (jit)
 	{
-		// DSPCore_CheckExceptions();
-		// DSPCore_CheckExternalInterrupt();
 		cyclesLeft = cycles;
-
 		CompiledCode pExecAddr = (CompiledCode)jit->enterDispatcher;
 		pExecAddr();
-
-		// To use the C++ dispatcher, uncomment the line below and comment out the two lines above
-		//jit->RunForCycles(cyclesLeft);
 		return cyclesLeft;
 	}
 
-	while (cycles > 0) {
+	while (cycles > 0)
+	{
 	reswitch:
 		switch (core_state)
 		{
diff --git a/Source/Core/DSPCore/Src/DSPEmitter.cpp b/Source/Core/DSPCore/Src/DSPEmitter.cpp
index 2ec62b1cd4..e68fea7e18 100644
--- a/Source/Core/DSPCore/Src/DSPEmitter.cpp
+++ b/Source/Core/DSPCore/Src/DSPEmitter.cpp
@@ -96,15 +96,10 @@ void DSPEmitter::checkExceptions(u32 retval)
 #endif
 
 	DSPJitRegCache c(gpr);
-
 	SaveDSPRegs();
 	ABI_CallFunction((void *)&DSPCore_CheckExceptions);
-
-	//	ABI_RestoreStack(0);
-	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	MOV(32, R(EAX), Imm32(retval));
-	RET();
-
+	JMP(returnDispatcher, true);
 	gpr.flushRegs(c,false);
 
 	SetJumpTarget(skipCheck);
@@ -201,8 +196,6 @@ void DSPEmitter::Compile(u16 start_addr)
 	unresolvedJumps[start_addr].clear();
 
 	const u8 *entryPoint = AlignCode16();
-	ABI_PushAllCalleeSavedRegsAndAdjustStack();
-	//	ABI_AlignStack(0);
 
 	/*
 	// Check for other exceptions
@@ -276,7 +269,6 @@ void DSPEmitter::Compile(u16 start_addr)
 			DSPJitRegCache c(gpr);
 			HandleLoop();
 			SaveDSPRegs();
-			ABI_PopAllCalleeSavedRegsAndAdjustStack();
 			if (DSPAnalyzer::code_flags[start_addr] & DSPAnalyzer::CODE_IDLE_SKIP)
 			{
 				MOV(16, R(EAX), Imm16(DSP_IDLE_SKIP_CYCLES));
@@ -285,7 +277,7 @@ void DSPEmitter::Compile(u16 start_addr)
 			{
 				MOV(16, R(EAX), Imm16(blockSize[start_addr]));
 			}
-			RET();
+			JMP(returnDispatcher, true);
 			gpr.flushRegs(c,false);
 
 			SetJumpTarget(rLoopAddressExit);
@@ -315,7 +307,6 @@ void DSPEmitter::Compile(u16 start_addr)
 				DSPJitRegCache c(gpr);
 				//don't update g_dsp.pc -- the branch insn already did
 				SaveDSPRegs();
-				ABI_PopAllCalleeSavedRegsAndAdjustStack();
 				if (DSPAnalyzer::code_flags[start_addr] & DSPAnalyzer::CODE_IDLE_SKIP)
 				{
 					MOV(16, R(EAX), Imm16(DSP_IDLE_SKIP_CYCLES));
@@ -324,7 +315,7 @@ void DSPEmitter::Compile(u16 start_addr)
 				{
 					MOV(16, R(EAX), Imm16(blockSize[start_addr]));
 				}
-				RET();
+				JMP(returnDispatcher, true);
 				gpr.flushRegs(c,false);
 
 				SetJumpTarget(rNoBranch);
@@ -382,8 +373,6 @@ void DSPEmitter::Compile(u16 start_addr)
 	}
 
 	SaveDSPRegs();
-
-	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	if (DSPAnalyzer::code_flags[start_addr] & DSPAnalyzer::CODE_IDLE_SKIP)
 	{
 		MOV(16, R(EAX), Imm16(DSP_IDLE_SKIP_CYCLES));
@@ -392,18 +381,16 @@ void DSPEmitter::Compile(u16 start_addr)
 	{
 		MOV(16, R(EAX), Imm16(blockSize[start_addr]));
 	}	
-	RET();
+	JMP(returnDispatcher, true);
 }
 
 const u8 *DSPEmitter::CompileStub()
 {
 	const u8 *entryPoint = AlignCode16();
-	ABI_PushAllCalleeSavedRegsAndAdjustStack();
 	ABI_CallFunction((void *)&CompileCurrent);
-	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	//MOVZX(32, 16, ECX, M(&g_dsp.pc));
 	XOR(32, R(EAX), R(EAX)); // Return 0 cycles executed
-	RET();
+	JMP(returnDispatcher);
 	return entryPoint;
 }
 
@@ -412,18 +399,6 @@ void DSPEmitter::CompileDispatcher()
 	enterDispatcher = AlignCode16();
 	ABI_PushAllCalleeSavedRegsAndAdjustStack();
 
-	// Cache pointers into registers
-#ifdef _M_IX86
-	MOV(16, R(ESI), M(&cyclesLeft));
-	MOV(32, R(EBX), ImmPtr(blocks));
-#else
-	// Using R12 here since it is callee save register on both
-	// linux and windows 64.
-	MOV(64, R(R12), ImmPtr(&cyclesLeft));
-	MOV(16, R(R12), MatR(R12));
-	MOV(64, R(RBX), ImmPtr(blocks));
-#endif
-
 	const u8 *dispatcherLoop = GetCodePtr();
 
 	// Check for DSP halt
@@ -444,16 +419,21 @@ void DSPEmitter::CompileDispatcher()
 
 	// Execute block. Cycles executed returned in EAX.
 #ifdef _M_IX86
-	CALLptr(MComplex(EBX, ECX, SCALE_4, 0));
+	MOV(32, R(EBX), ImmPtr(blocks));
+	JMPptr(MComplex(EBX, ECX, SCALE_4, 0));
 #else
-	CALLptr(MComplex(RBX, RCX, SCALE_8, 0));
+	MOV(64, R(RBX), ImmPtr(blocks));
+	JMPptr(MComplex(RBX, RCX, SCALE_8, 0));
 #endif
 
+	returnDispatcher = GetCodePtr();
+
 	// Decrement cyclesLeft
 #ifdef _M_IX86
-	SUB(16, R(ESI), R(EAX));
+	SUB(16, M(&cyclesLeft), R(EAX));
 #else
-	SUB(16, R(R12), R(EAX));
+	MOV(64, R(R12), ImmPtr(&cyclesLeft));
+	SUB(16, MatR(R12), R(EAX));
 #endif
 
 	J_CC(CC_A, dispatcherLoop);
@@ -464,35 +444,3 @@ void DSPEmitter::CompileDispatcher()
 	ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	RET();
 }
-
-// Don't use the % operator in the inner loop. It's slow.
-int STACKALIGN DSPEmitter::RunForCycles(int cycles)
-{
-	while (!(g_dsp.cr & CR_HALT))
-	{
-		// Compile the block if needed
-		u16 block_addr = g_dsp.pc;
-		int block_size = blockSize[block_addr];
-		if (!block_size)
-		{
-			CompileCurrent();
-			block_size = blockSize[block_addr];
-		}
-		
-		// Execute the block if we have enough cycles
-		if (cycles > block_size)
-		{
-			cycles -= blocks[block_addr]();
-		}
-		else
-		{
-			break;
-		}
-	}
-
-	// DSP gave up the remaining cycles.
-	if (g_dsp.cr & CR_HALT || cycles < 0)
-		return 0;
-
-	return cycles;
-}
diff --git a/Source/Core/DSPCore/Src/DSPEmitter.h b/Source/Core/DSPCore/Src/DSPEmitter.h
index 401c89ae03..17f1850001 100644
--- a/Source/Core/DSPCore/Src/DSPEmitter.h
+++ b/Source/Core/DSPCore/Src/DSPEmitter.h
@@ -50,8 +50,6 @@ public:
 
 	void Default(UDSPInstruction inst);
 
-	int STACKALIGN RunForCycles(int cycles);
-
 	// CC Util
 	void Update_SR_Register64(Gen::X64Reg val = Gen::EAX);
 	void Update_SR_Register64_Carry(Gen::X64Reg val = Gen::EAX);
@@ -253,6 +251,7 @@ public:
 	// CALL this to start the dispatcher
 	const u8 *enterDispatcher;
 	const u8 *stubEntryPoint;
+	const u8 *returnDispatcher;
 	u16 compilePC;
 	u16 startAddr;
 	Block *blockLinks;
diff --git a/Source/Core/DSPCore/Src/Jit/DSPJitBranch.cpp b/Source/Core/DSPCore/Src/Jit/DSPJitBranch.cpp
index 56f5edcd9a..4dd2476e70 100644
--- a/Source/Core/DSPCore/Src/Jit/DSPJitBranch.cpp
+++ b/Source/Core/DSPCore/Src/Jit/DSPJitBranch.cpp
@@ -122,7 +122,6 @@ static void ReJitConditional(const UDSPInstruction opc, DSPEmitter& emitter)
 static void WriteBranchExit(DSPEmitter& emitter)
 {
 	emitter.SaveDSPRegs();
-	emitter.ABI_PopAllCalleeSavedRegsAndAdjustStack();
 	if (DSPAnalyzer::code_flags[emitter.startAddr] & DSPAnalyzer::CODE_IDLE_SKIP)
 	{
 		emitter.MOV(16, R(EAX), Imm16(0x1000));
@@ -131,7 +130,7 @@ static void WriteBranchExit(DSPEmitter& emitter)
 	{
 		emitter.MOV(16, R(EAX), Imm16(emitter.blockSize[emitter.startAddr]));
 	}
-	emitter.RET();
+	emitter.JMP(emitter.returnDispatcher, true);
 }
 
 static void WriteBlockLink(DSPEmitter& emitter, u16 dest)
@@ -152,10 +151,11 @@ static void WriteBlockLink(DSPEmitter& emitter, u16 dest)
 			emitter.MOV(16, M(&cyclesLeft), R(ESI));
 #else
 			// Check if we have enough cycles to execute the next block
-			emitter.CMP(16, R(R12), Imm16(emitter.blockSize[emitter.startAddr] + emitter.blockSize[dest]));
+			emitter.MOV(64, R(R12), ImmPtr(&cyclesLeft));
+			emitter.CMP(16, MatR(R12), Imm16(emitter.blockSize[emitter.startAddr] + emitter.blockSize[dest]));
 			FixupBranch notEnoughCycles = emitter.J_CC(CC_BE);
 
-			emitter.SUB(16, R(R12), Imm16(emitter.blockSize[emitter.startAddr]));
+			emitter.SUB(16, MatR(R12), Imm16(emitter.blockSize[emitter.startAddr]));
 #endif
 			emitter.JMP(emitter.blockLinks[dest], true);
 			emitter.SetJumpTarget(notEnoughCycles);