Commit r5970 part #4: Merge some loops in PPCAnalyst::Flatten()

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5976 8ced0084-cf51-0410-be5f-012b33b47a6e

2 files changed, 72 insertions, 90 deletions

diff --git a/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp b/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp
index 2a00a18cf9..d9fc6f7a0f 100644
--- a/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp
+++ b/Source/Core/Core/Src/PowerPC/PPCAnalyst.cpp
@@ -296,6 +296,13 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa, Bloc
 
 	gpa->any = true;
 	fpa->any = false;
+	for (int i = 0; i < 32; i++)
+	{
+		gpa->firstRead[i]  = -1;
+		gpa->firstWrite[i] = -1;
+		gpa->numReads[i] = 0;
+		gpa->numWrites[i] = 0;
+	}
 
 	u32 blockstart = address;
 	int maxsize = blockSize;
@@ -307,7 +314,8 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa, Bloc
 	CodeOp *code = buffer->codebuffer;
 	bool foundExit = false;
 
-	// Flatten! (Currently just copies, following branches is disabled)
+	// Do analysis of the code, look for dependencies etc
+	int numSystemInstructions = 0;
 	for (int i = 0; i < maxsize; i++)
 	{
 		num_inst++;
@@ -326,63 +334,11 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa, Bloc
 		if (opinfo)
 			numCycles += opinfo->numCyclesMinusOne + 1;
 		_assert_msg_(POWERPC, opinfo != 0, "Invalid Op - Error flattening %08x op %08x", address + i*4, inst.hex);
-		bool follow = false;
-		u32 destination;
-		if (inst.OPCD == 18 && blockSize > 1)
-		{
-			//Is bx - should we inline? yes!
-			if (inst.AA)
-				destination = SignExt26(inst.LI << 2);
-			else
-				destination = address + SignExt26(inst.LI << 2);
-			if (destination != blockstart)
-				follow = true;
-		}
-		if (follow)
-			numFollows++;
-		if (numFollows > 1)
-			follow = false;
-		follow = false;
-		if (!follow)
-		{
-			if (opinfo->flags & FL_ENDBLOCK) //right now we stop early
-			{
-				foundExit = true;
-				break;
-			}
-			address += 4;
-		}
-		else
-		{
-			code[i].skip = true;
-			address = destination;
-		}
-	}	
-	if (!foundExit && blockSize > 1)
-		NOTICE_LOG(POWERPC, "Analyzer ERROR - Function %08x too big, size is 0x%08x", blockstart, address-blockstart);
-	st->numCycles = numCycles;
-
-	// Do analysis of the code, look for dependencies etc
-	int numSystemInstructions = 0;
-	for (int i = 0; i < 32; i++)
-	{
-		gpa->firstRead[i]  = -1;
-		gpa->firstWrite[i] = -1;
-		gpa->numReads[i] = 0;
-		gpa->numWrites[i] = 0;
-	}
 
-	gpa->any = true;
-	for (int i = 0; i < num_inst; i++)
-	{
-		UGeckoInstruction inst = code[i].inst;
-		
 		code[i].wantsCR0 = false;
 		code[i].wantsCR1 = false;
 		code[i].wantsPS1 = false;
 
-		const GekkoOPInfo *opinfo = code[i].opinfo;
-		_assert_msg_(POWERPC, opinfo != 0, "Invalid Op - Error scanning %08x op %08x",address+i*4,inst.hex);
 		int flags = opinfo->flags;
 
 		if (flags & FL_USE_FPU)
@@ -407,54 +363,53 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa, Bloc
 		else
 			code[i].outputCR1 = (flags & FL_SET_CR1) ? true : false;
 
-		for (int j = 0; j < 3; j++)
-		{
-			code[i].fregsIn[j] = -1;
-			code[i].regsIn[j] = -1;
-		}
-		for (int j = 0; j < 2; j++)
-			code[i].regsOut[j] = -1;
-
-		code[i].fregOut = -1;
-
 		int numOut = 0;
 		int numIn = 0;
 		if (flags & FL_OUT_A)
 		{
 			code[i].regsOut[numOut++] = inst.RA;
-			gpa->numWrites[inst.RA]++;
+			gpa->SetOutputRegister(inst.RA, i);
 		}
 		if (flags & FL_OUT_D)
 		{
 			code[i].regsOut[numOut++] = inst.RD;
-			gpa->numWrites[inst.RD]++;
+			gpa->SetOutputRegister(inst.RD, i);
 		}
 		if (flags & FL_OUT_S)
 		{
 			code[i].regsOut[numOut++] = inst.RS;
-			gpa->numWrites[inst.RS]++;
+			gpa->SetOutputRegister(inst.RS, i);
 		}
 		if ((flags & FL_IN_A) || ((flags & FL_IN_A0) && inst.RA != 0))
 		{
 			code[i].regsIn[numIn++] = inst.RA;
-			gpa->numReads[inst.RA]++;
+			gpa->SetInputRegister(inst.RA, i);
 		}
 		if (flags & FL_IN_B)
 		{
 			code[i].regsIn[numIn++] = inst.RB;
-			gpa->numReads[inst.RB]++;
+			gpa->SetInputRegister(inst.RB, i);
 		}
 		if (flags & FL_IN_C)
 		{
 			code[i].regsIn[numIn++] = inst.RC;
-			gpa->numReads[inst.RC]++;
+			gpa->SetInputRegister(inst.RC, i);
 		}
 		if (flags & FL_IN_S)
 		{
 			code[i].regsIn[numIn++] = inst.RS;
-			gpa->numReads[inst.RS]++;
+			gpa->SetInputRegister(inst.RS, i);
 		}
 
+		// Set remaining register slots as unused (-1)
+		for (int j = numIn; j < 3; j++)
+			code[i].regsIn[j] = -1;
+		for (int j = numOut; j < 2; j++)
+			code[i].regsOut[j] = -1;
+		for (int j = 0; j < 3; j++)
+			code[i].fregsIn[j] = -1;
+		code[i].fregOut = -1;
+
 		switch (opinfo->type) 
 		{
 		case OPTYPE_INTEGER:
@@ -479,28 +434,41 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa, Bloc
 			break;
 		}
 
-		for (int j = 0; j < numIn; j++)
+		bool follow = false;
+		u32 destination;
+		if (inst.OPCD == 18 && blockSize > 1)
 		{
-			int r = code[i].regsIn[j];
-			if (r < 0 || r > 31)
-				PanicAlert("wtf");
-			if (gpa->firstRead[r] == -1)
-				gpa->firstRead[r] = (short)(i);
-			gpa->lastRead[r] = (short)(i);
-			gpa->numReads[r]++;
+			//Is bx - should we inline? yes!
+			if (inst.AA)
+				destination = SignExt26(inst.LI << 2);
+			else
+				destination = address + SignExt26(inst.LI << 2);
+			if (destination != blockstart)
+				follow = true;
 		}
-
-		for (int j = 0; j < numOut; j++)
-		{ 
-			int r = code[i].regsOut[j];
-			if (r < 0 || r > 31)
-				PanicAlert("wtf");
-			if (gpa->firstWrite[r] == -1)
-				gpa->firstWrite[r] = (short)(i);
-			gpa->lastWrite[r] = (short)(i);
-			gpa->numWrites[r]++;
+		if (follow)
+			numFollows++;
+		if (numFollows > 1)
+			follow = false;
+		follow = false;
+		if (!follow)
+		{
+			if (opinfo->flags & FL_ENDBLOCK) //right now we stop early
+			{
+				foundExit = true;
+				break;
+			}
+			address += 4;
 		}
-	}
+		else
+		{
+			code[i].skip = true;
+			address = destination;
+		}
+	}	
+	if (!foundExit && blockSize > 1)
+		NOTICE_LOG(POWERPC, "Analyzer ERROR - Function %08x too big, size is 0x%08x", blockstart, address-blockstart);
+	st->numCycles = numCycles;
 
 	// Instruction Reordering Pass
 	if (blockSize > 1)
@@ -530,7 +498,7 @@ u32 Flatten(u32 address, int *realsize, BlockStats *st, BlockRegStats *gpa, Bloc
 	bool wantsCR0 = true;
 	bool wantsCR1 = true;
 	bool wantsPS1 = true;
-	for (int i = num_inst; i; i--)
+	for (int i = num_inst - 1; i >= 0; i--)
 	{
 		if (code[i].outputCR0)
 			wantsCR0 = false;
diff --git a/Source/Core/Core/Src/PowerPC/PPCAnalyst.h b/Source/Core/Core/Src/PowerPC/PPCAnalyst.h
index 5a41f6fda7..d19774e3e8 100644
--- a/Source/Core/Core/Src/PowerPC/PPCAnalyst.h
+++ b/Source/Core/Core/Src/PowerPC/PPCAnalyst.h
@@ -77,6 +77,20 @@ struct BlockRegStats
 	int GetUseRange(int reg) {
 		return max(lastRead[reg], lastWrite[reg]) - 
 			   min(firstRead[reg], firstWrite[reg]);}
+
+	inline void SetInputRegister(int reg, short opindex) {
+		if (firstRead[reg] == -1)
+			firstRead[reg] = (short)(opindex);
+		lastRead[reg] = (short)(opindex);
+		numReads[reg]++;
+	}
+
+	inline void SetOutputRegister(int reg, short opindex) {
+		if (firstWrite[reg] == -1)
+			firstWrite[reg] = (short)(opindex);
+		lastWrite[reg] = (short)(opindex);
+		numWrites[reg]++;
+	}
 };