VU: optimise entering VU JITs

Keeps note on how many cycles it needs for the next block to save exiting the EE JIT and entering the microVU JIT for no reason

pcsx2/VU.h

@@ -155,6 +155,8 @@ struct __aligned16 VURegs {
 	u32 statusflag;
 	u32 clipflag;
 
+	s32 nextBlockCycles;
+
 	u8 *Mem;
 	u8 *Micro;
 

pcsx2/VUmicro.cpp

@@ -34,25 +34,17 @@ void BaseVUmicroCPU::ExecuteBlock(bool startUp) {
 
 	if (startUp && s) {  // Start Executing a microprogram
 		Execute(s); // Kick start VU
-
-		if (stat & test) {
-			cpuSetNextEventDelta(s);
-
-			if (m_Idx)
-				VU1.cycle = cpuRegs.cycle;
-			else
-				VU0.cycle = cpuRegs.cycle;
-		}
 	}
 	else { // Continue Executing
 		u32 cycle = m_Idx ? VU1.cycle : VU0.cycle;
 		s32 delta = (s32)(u32)(cpuRegs.cycle - cycle);
-		if (delta > 0) {	// Enough time has passed
+		s32 nextblockcycles = m_Idx ? VU1.nextBlockCycles : VU0.nextBlockCycles;
+
+		if (delta < nextblockcycles)
+			return;
+
+		if (delta > 0)		// Enough time has passed
 			Execute(delta);	// Execute the time since the last call
-			if (stat & test) 
-				cpuSetNextEventDelta(delta);
-		}
-		else cpuSetNextEventDelta(-delta); // Haven't caught-up from kick start
 	}
 }
 
@@ -62,11 +54,11 @@ void BaseVUmicroCPU::ExecuteBlock(bool startUp) {
 // This fixes spinning/hanging in some games like Ratchet and Clank's Intro.
 void BaseVUmicroCPU::ExecuteBlockJIT(BaseVUmicroCPU* cpu) {
 	const u32& stat	= VU0.VI[REG_VPU_STAT].UL;
-	const int  test = cpu->m_Idx ? 0x100 : 1;
+	const int  test = 1;
 
 	if (stat & test) {		// VU is running
-		u32 cycle = cpu->m_Idx ? VU1.cycle : VU0.cycle;
+		s32 delta = (s32)(u32)(cpuRegs.cycle - VU0.cycle);
-		s32 delta = (s32)(u32)(cpuRegs.cycle - cycle);
+
 		if (delta > 0) {			// Enough time has passed
 			cpu->Execute(delta);	// Execute the time since the last call
 		}

pcsx2/x86/ix86-32/iR5900LoadStore.cpp

@@ -581,6 +581,7 @@ void recLQC2()
 	xADD(eax, scaleblockcycles_clear());
 	xMOV(ptr32[&cpuRegs.cycle], eax); // update cycles
 	xSUB(eax, ptr32[&VU0.cycle]);
+	xSUB(eax, ptr32[&VU0.nextBlockCycles]);
 	xCMP(eax, 8);
 	xForwardJL32 skip;
 	xLoadFarAddr(arg1reg, CpuVU0);
@@ -625,6 +626,7 @@ void recSQC2()
 	xADD(eax, scaleblockcycles_clear());
 	xMOV(ptr32[&cpuRegs.cycle], eax); // update cycles
 	xSUB(eax, ptr32[&VU0.cycle]);
+	xSUB(eax, ptr32[&VU0.nextBlockCycles]);
 	xCMP(eax, 8);
 	xForwardJL32 skip;
 	xLoadFarAddr(arg1reg, CpuVU0);

pcsx2/x86/microVU.cpp

@@ -95,6 +95,7 @@ void mVUreset(microVU& mVU, bool resetReserve) {
 	mVUdispatcherCD(mVU);
 	mVUemitSearch();
 
+	mVU.regs().nextBlockCycles = 0;
 	// Clear All Program Data
 	//memset(&mVU.prog, 0, sizeof(mVU.prog));
 	memset(&mVU.prog.lpState, 0, sizeof(mVU.prog.lpState));

pcsx2/x86/microVU_Branch.inl

@@ -108,12 +108,15 @@ void mVUDTendProgram(mV, microFlagCycles* mFC, int isEbit) {
 	}
 
 	if (isEbit) { // Clear 'is busy' Flags
+		xMOV(ptr32[&mVU.regs().nextBlockCycles], 0);
 		if (!mVU.index || !THREAD_VU1) {
 			xAND(ptr32[&VU0.VI[REG_VPU_STAT].UL], (isVU1 ? ~0x100 : ~0x001)); // VBS0/VBS1 flag
 		}
 		else
 			xFastCall((void*)mVUTBit);
 	}
+	else
+		xMOV(ptr32[&mVU.regs().nextBlockCycles], mVUcycles);
 	
 	if (isEbit != 2) { // Save PC, and Jump to Exit Point
 		xMOV(ptr32[&mVU.regs().VI[REG_TPC].UL], xPC);
@@ -208,12 +211,15 @@ void mVUendProgram(mV, microFlagCycles* mFC, int isEbit) {
 
 
 	if ((isEbit && isEbit != 3)) { // Clear 'is busy' Flags
+		xMOV(ptr32[&mVU.regs().nextBlockCycles], 0);
 		if (!mVU.index || !THREAD_VU1) {
 			xAND(ptr32[&VU0.VI[REG_VPU_STAT].UL], (isVU1 ? ~0x100 : ~0x001)); // VBS0/VBS1 flag
 		}
 		else
 			xFastCall((void*)mVUEBit);
 	}
+	else
+		xMOV(ptr32[&mVU.regs().nextBlockCycles], mVUcycles);
 
 	if (isEbit != 2 && isEbit != 3) { // Save PC, and Jump to Exit Point
 		xMOV(ptr32[&mVU.regs().VI[REG_TPC].UL], xPC);

pcsx2/x86/microVU_Macro.inl

@@ -302,6 +302,7 @@ static void recCFC2() {
 		xADD(eax, scaleblockcycles_clear());
 		xMOV(ptr32[&cpuRegs.cycle], eax); // update cycles
 		xSUB(eax, ptr32[&vu0Regs.cycle]);
+		xSUB(eax, ptr32[&vu0Regs.nextBlockCycles]);
 		xCMP(eax, 8);
 		xForwardJL32 skip;
 		xLoadFarAddr(arg1reg, CpuVU0);
@@ -380,6 +381,7 @@ static void recCTC2() {
 		xADD(eax, scaleblockcycles_clear());
 		xMOV(ptr32[&cpuRegs.cycle], eax); // update cycles
 		xSUB(eax, ptr32[&vu0Regs.cycle]);
+		xSUB(eax, ptr32[&vu0Regs.nextBlockCycles]);
 		xCMP(eax, 8);
 		xForwardJL32 skip;
 		xLoadFarAddr(arg1reg, CpuVU0);
@@ -460,6 +462,7 @@ static void recQMFC2() {
 		xADD(eax, scaleblockcycles_clear());
 		xMOV(ptr32[&cpuRegs.cycle], eax); // update cycles
 		xSUB(eax, ptr32[&vu0Regs.cycle]);
+		xSUB(eax, ptr32[&vu0Regs.nextBlockCycles]);
 		xCMP(eax, 8);
 		xForwardJL32 skip;
 		xLoadFarAddr(arg1reg, CpuVU0);