microVU:

- Added partial program searching support (helps to recompile less in games that send a lot of small programs like Tekken 5)
- Optimized logical min/max code some more (knocked off 2 instructions)
- Fixed some cases where 'last_used' should have been set.

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1209 96395faa-99c1-11dd-bbfe-3dabce05a288

pcsx2/x86/microVU.cpp

@@ -77,15 +77,8 @@ microVUt(void) mVUreset() {
 	// Clear All Program Data
 	memset(&mVU->prog, 0, sizeof(mVU->prog));
 
-	// Create Block Managers
-	// Block managers are now allocated "on-demand" by the recompiler -- air
-	/*for (int i = 0; i <= mVU->prog.max; i++) {
-		for (u32 j = 0; j < (mVU->progSize / 2); j++) {
-			mVU->prog.prog[i].block[j] = new microBlockManager();
-		}
-	}*/
-
 	// Program Variables
+	mVU->prog.isSame = -1;
 	mVU->prog.cleared = 1;
 	mVU->prog.cur = -1;
 	mVU->prog.total = -1;
@@ -99,6 +92,8 @@ microVUt(void) mVUreset() {
 		mVU->prog.prog[i].x86ptr = z;
 		z += (mVU->cacheSize / (mVU->prog.max + 1));
 		mVU->prog.prog[i].x86end = z;
+		mVU->prog.prog[i].range[0] = -1; // Set range to 
+		mVU->prog.prog[i].range[1] = -1; // indeterminable status
 	}
 }
 
@@ -138,7 +133,10 @@ microVUt(void) mVUclear(u32 addr, u32 size) {
 microVUt(void) mVUclearProg(int progIndex) {
 	microVU* mVU = mVUx;
 	mVU->prog.prog[progIndex].used = 1;
+	mVU->prog.prog[progIndex].last_used = 3;
 	mVU->prog.prog[progIndex].sFlagHack = 0;
+	mVU->prog.prog[progIndex].range[0] = -1;
+	mVU->prog.prog[progIndex].range[1] = -1;
 	mVU->prog.prog[progIndex].x86ptr = mVU->prog.prog[progIndex].x86start;
 	for (u32 i = 0; i < (mVU->progSize / 2); i++) {
 		if( mVU->prog.prog[progIndex].block[i] )
@@ -160,6 +158,8 @@ microVUt(int) mVUfindLeastUsedProg() {
 	if (mVU->prog.total < mVU->prog.max) {
 		mVU->prog.total++;
 		mVUcacheProg<vuIndex>(mVU->prog.total); // Cache Micro Program
+		mVU->prog.prog[mVU->prog.total].used = 1;
+		mVU->prog.prog[mVU->prog.total].last_used = 3;
 		Console::Notice("microVU%d: Cached MicroPrograms = %d", params vuIndex, mVU->prog.total+1);
 		return mVU->prog.total;
 	}
@@ -214,13 +214,16 @@ microVUt(void) __mVUvsyncUpdate() {
 	}
 }
 
-microVUt(int) mVUcmpProg(int progIndex, bool progUsed, bool needOverflowCheck) {
+// Compare Cached microProgram to mVU->regs->Micro
+microVUt(int) mVUcmpProg(int progIndex, bool progUsed, bool needOverflowCheck, bool cmpWholeProg) {
 	microVU* mVU = mVUx;
-
+	
 	if (progUsed) {
-		if (!memcmp_mmx(mVU->prog.prog[progIndex].data, mVU->regs->Micro, mVU->microSize)) {
+		if (cmpWholeProg && (!memcmp_mmx((u8*)mVUprogI.data, mVU->regs->Micro, mVU->microSize)) ||
+		  (!cmpWholeProg && (!memcmp_mmx((u8*)mVUprogI.data + mVUprogI.range[0], (u8*)mVU->regs->Micro + mVUprogI.range[0], ((mVUprogI.range[1] + 8) - mVUprogI.range[0]))))) {
 			mVU->prog.cur = progIndex;
 			mVU->prog.cleared = 0;
+			mVU->prog.isSame = cmpWholeProg ? 1 : -1;
 			mVU->prog.prog[progIndex].last_used = 3;
 			if (!needOverflowCheck || mVU->prog.prog[progIndex].used < 0x7fffffff) {
 				mVU->prog.prog[progIndex].used++; // increment 'used' (avoiding overflows if necessary)
@@ -237,18 +240,24 @@ microVUt(int) mVUsearchProg() {
 	
 	if (mVU->prog.cleared) { // If cleared, we need to search for new program
 		for (int i = 0; i <= mVU->prog.total; i++) {
-			if (mVUcmpProg<vuIndex>(i, !!mVU->prog.prog[i].used, 1))
+			if (mVUcmpProg<vuIndex>(i, !!mVU->prog.prog[i].used, 1, 0))
 				return 1; // Check Recently Used Programs
 		}
 		for (int i = 0; i <= mVU->prog.total; i++) {
-			if (mVUcmpProg<vuIndex>(i,  !mVU->prog.prog[i].used, 0))
+			if (mVUcmpProg<vuIndex>(i,  !mVU->prog.prog[i].used, 0, 0))
 				return 1; // Check Older Programs
-		}
+		}                                                  
+		/*for (int i = 0; i <= mVU->prog.total; i++) {
+			if (mVUcmpProg<vuIndex>(i, 1, 1, 0))
+				return 1; // Check Partial Program
+		}*/
 		mVU->prog.cur = mVUfindLeastUsedProg<vuIndex>(); // If cleared and program not found, make a new program instance
 		mVU->prog.cleared = 0;
+		mVU->prog.isSame  = 1;
 		return 0;
 	}
 	mVU->prog.prog[mVU->prog.cur].used++;
+	mVU->prog.prog[mVU->prog.cur].last_used = 3;
 	return 1; // If !cleared, then we're still on the same program as last-time ;)
 }
 /*

pcsx2/x86/microVU.h

@@ -88,7 +88,7 @@ struct microProgManager {
 	static const int	max = mMaxProg - 1; 
 	int					cur;			// Index to Current MicroProgram thats running (-1 = uncached)
 	int					total;			// Total Number of valid MicroPrograms minus 1
-	int					isSame;			// Current cached microProgram is Exact Same program as mVU->regs->Micro
+	int					isSame;			// Current cached microProgram is Exact Same program as mVU->regs->Micro (-1 = unknown, 0 = No, 1 = Yes)
 	int					cleared;		// Micro Program is Indeterminate so must be searched for (and if no matches are found then recompile a new one)
 	microRegInfo		lpState;		// Pipeline state from where program left off (useful for continuing execution)
 };

pcsx2/x86/microVU_Compile.inl

@@ -150,6 +150,34 @@ microVUt(void) mVUtestCycles() {
 	SUB32ItoM((uptr)&mVU->cycles, mVUcycles);
 }
 
+microVUt(void) mVUcheckIsSame() {
+	microVU* mVU = mVUx;
+
+	if (mVU->prog.isSame == -1) {
+		mVU->prog.isSame = !!memcmp_mmx(mVU->prog.prog[mVU->prog.cur].data, mVU->regs->Micro, mVU->microSize);
+	}
+	if (mVU->prog.isSame == 0) {
+		mVUcacheProg<vuIndex>(mVU->prog.cur);
+	}
+}
+
+microVUt(void) mVUsetupRange(u32 pc) {
+	microVU* mVU = mVUx;
+
+	if (mVUcurProg.range[0] == -1) { 
+		mVUcurProg.range[0] = (s32)pc;
+		mVUcurProg.range[1] = (s32)pc;
+	}
+	else if (mVUcurProg.range[0] > (s32)pc) {
+		mVUcurProg.range[0] = (s32)pc;
+		mVUcheckIsSame<vuIndex>();
+	}
+	else if (mVUcurProg.range[1] < (s32)pc) {
+		mVUcurProg.range[1] = (s32)pc;
+		mVUcheckIsSame<vuIndex>();
+	}
+}
+
 //------------------------------------------------------------------
 // Recompiler
 //------------------------------------------------------------------
@@ -169,6 +197,9 @@ microVUt(void*) __fastcall mVUcompile(u32 startPC, uptr pState) {
 	microBlock* pBlock = mVUblocks[startPC/8]->search((microRegInfo*)pState);
 	if (pBlock) { return pBlock->x86ptrStart; }
 	
+	// Setup Program Bounds/Range
+	mVUsetupRange<vuIndex>(startPC);
+
 	// First Pass
 	iPC = startPC / 4;
 	setCode();
@@ -230,6 +261,7 @@ microVUt(void*) __fastcall mVUcompile(u32 startPC, uptr pState) {
 		else {
 			microBlock* bBlock = NULL;
 			u32* ajmp = 0;
+			mVUsetupRange<vuIndex>(xPC);
 
 			switch (mVUbranch) {
 				case 3: branchCase(JE32,  JNE32);	// IBEQ
@@ -308,6 +340,7 @@ microVUt(void*) __fastcall mVUcompile(u32 startPC, uptr pState) {
 	if (x == (vuIndex?(0x3fff/8):(0xfff/8))) { Console::Error("microVU%d: Possible infinite compiling loop!", params vuIndex); }
 
 	// Do E-bit end stuff here
+	mVUsetupRange<vuIndex>(xPC - 8);
 	mVUendProgram<vuIndex>(findFlagInst(xStatus, 0x7fffffff), findFlagInst(xMac, 0x7fffffff), findFlagInst(xClip, 0x7fffffff));
 
 	return thisPtr; //ToDo: Save pipeline state?

pcsx2/x86/microVU_Misc.h

@@ -150,6 +150,7 @@ declareAllVariables
 #define pass4 if (recPass == 3)
 
 // Misc Macros...
+#define mVUprogI	 mVU->prog.prog[progIndex]
 #define mVUcurProg	 mVU->prog.prog[mVU->prog.cur]
 #define mVUblocks	 mVU->prog.prog[mVU->prog.cur].block
 #define mVUallocInfo mVU->prog.prog[mVU->prog.cur].allocInfo
@@ -161,13 +162,14 @@ declareAllVariables
 #define mVUregs		 mVUallocInfo.block.pState
 #define mVUregsTemp	 mVUallocInfo.regsTemp
 #define iPC			 mVUallocInfo.curPC
-#define mVUflagInfo	 mVUregs.needExactMatch
 #define mVUsFlagHack mVUallocInfo.sFlagHack
 #define mVUinfo		 mVUallocInfo.info[iPC / 2]
 #define mVUstall	 mVUallocInfo.stall[iPC / 2]
 #define mVUstartPC	 mVUallocInfo.startPC
+#define mVUflagInfo	 mVUregs.needExactMatch
+#define mVUflagHack  (mVUcurProg.sFlagHack)
 #define xPC			 ((iPC / 2) * 8)
-#define curI		 mVUcurProg.data[iPC]
+#define curI		 ((u32*)mVU->regs->Micro)[iPC]//mVUcurProg.data[iPC]
 #define setCode()	 { mVU->code = curI; }
 #define incPC(x)	 { iPC = ((iPC + x) & (mVU->progSize-1)); setCode(); }
 #define incPC2(x)	 { iPC = ((iPC + x) & (mVU->progSize-1)); }
@@ -175,7 +177,11 @@ declareAllVariables
 #define bSaveAddr	 (((xPC + (2 * 8)) & ((vuIndex) ? 0x3ff8:0xff8)) / 8)
 #define branchAddr	 ((xPC + 8 + (_Imm11_ * 8)) & ((vuIndex) ? 0x3ff8:0xff8))
 #define shufflePQ	 (((mVU->p) ? 0xb0 : 0xe0) | ((mVU->q) ? 0x01 : 0x04))
-#define mVUflagHack  (mVUcurProg.sFlagHack)
+
+// Flag Info
+#define __Status	 (mVUflagInfo & (0xf<<0))
+#define __Mac		 (mVUflagInfo & (0xf<<4))
+#define __Clip		 (mVUflagInfo & (0xf<<8))
 
 // Pass 1 uses these to set mVUinfo
 #define _isNOP		 (1<<0) // Skip Lower Instruction
@@ -254,13 +260,6 @@ declareAllVariables
 #define mVUlogQ()	 { mVUlog(", Q"); }
 #define mVUlogCLIP() { mVUlog("w.xyz vf%02d, vf%02dw", _Fs_, _Ft_); }
 
-// Flag Info
-//#define __NeedExact	 (1<<12)
-//#define __ExactMatch (mVUregs.needExactMatch & (1<<12))
-#define __Status	 (mVUflagInfo & (0xf<<0))
-#define __Mac		 (mVUflagInfo & (0xf<<4))
-#define __Clip		 (mVUflagInfo & (0xf<<8))
-
 // Store VI regs in mmx regs?
 #define isMMX(_VIreg_)	0 //(_VIreg_ >= 1 && _VIreg_ <=8)
 #define mmVI(_VIreg_)	(_VIreg_ - 1)

pcsx2/x86/microVU_Misc.inl

@@ -348,12 +348,10 @@ void MIN_MAX_(x86SSERegType to, x86SSERegType from, bool min) {
 
 // Warning: Modifies from and to's upper 3 vectors
 void MIN_MAX_SS(x86SSERegType to, x86SSERegType from, bool min) {
-	SSE2_PSHUFD_XMM_to_XMM(to, to, 0x50);
+	SSE_SHUFPS_XMM_to_XMM (to, from, 0);
 	SSE2_PAND_M128_to_XMM (to,	 (uptr)MIN_MAX_MASK1);
 	SSE2_POR_M128_to_XMM  (to,	 (uptr)MIN_MAX_MASK2);
-	SSE2_PSHUFD_XMM_to_XMM(from, from, 0x50);
+	SSE2_PSHUFD_XMM_to_XMM(from, to, 0xee);
-	SSE2_PAND_M128_to_XMM (from, (uptr)MIN_MAX_MASK1);
-	SSE2_POR_M128_to_XMM  (from, (uptr)MIN_MAX_MASK2);
 	if (min) SSE2_MINPD_XMM_to_XMM(to, from);
 	else	 SSE2_MAXPD_XMM_to_XMM(to, from);
 }