Fix same XY optimization

oscar64/InterCode.cpp

@@ -5523,6 +5523,74 @@ bool InterCodeBasicBlock::PropagateConstTemps(const GrowingInstructionPtrArray&
 	return changed;
 }
 
+bool InterCodeBasicBlock::CombineIndirectAddressing(void)
+{
+	bool	changed = false;
+
+	if (!mVisited)
+	{
+		mVisited = true;
+
+		GrowingInstructionPtrArray	tvalue(nullptr);
+
+		for (int i = 0; i < mInstructions.Size(); i++)
+		{
+			InterInstruction* lins = mInstructions[i];
+			InterInstruction* tins = nullptr;
+
+			if (lins->mCode == IC_LEA && lins->mSrc[0].mTemp >= 0 && lins->mSrc[1].mTemp < 0 &&
+				!lins->mSrc[0].IsUByte() &&
+				(lins->mSrc[1].mMemory == IM_ABSOLUTE || lins->mSrc[1].mMemory == IM_GLOBAL || lins->mSrc[1].mMemory == IM_LOCAL))
+			{
+				int j = 0;
+				while (j < tvalue.Size() &&
+					!(tvalue[j]->mSrc[0].mTemp == lins->mSrc[0].mTemp &&
+						tvalue[j]->mSrc[1].mTemp < 0 &&
+						tvalue[j]->mSrc[1].mMemory == lins->mSrc[1].mMemory &&
+						tvalue[j]->mSrc[1].mVarIndex == lins->mSrc[1].mVarIndex &&
+						tvalue[j]->mSrc[1].mIntConst <= lins->mSrc[1].mIntConst &&
+						tvalue[j]->mSrc[1].mIntConst + 256 > lins->mSrc[1].mIntConst))
+					j++;
+
+				if (j < tvalue.Size())
+				{
+					int	offset = lins->mSrc[1].mIntConst - tvalue[j]->mSrc[1].mIntConst;
+					lins->mSrc[1] = tvalue[j]->mDst;
+					lins->mSrc[0].mTemp = -1;
+					lins->mSrc[0].mIntConst = offset;
+					changed = true;
+				}
+				else
+					tins = lins;
+			}
+
+			if (HasSideEffect(lins->mCode))
+				tvalue.SetSize(0);
+			else if (lins->mDst.mTemp >= 0)
+			{
+				int j = 0;
+				while (j < tvalue.Size())
+				{
+					if (tvalue[j]->ReferencesTemp(lins->mDst.mTemp))
+						tvalue.Remove(j);
+					else
+						j++;
+				}
+			}
+
+			if (tins)
+				tvalue.Push(tins);
+		}
+
+		if (mTrueJump && mTrueJump->CombineIndirectAddressing())
+			changed = true;
+		if (mFalseJump && mFalseJump->CombineIndirectAddressing())
+			changed = true;	
+	}
+
+	return changed;
+}
+
 void InterCodeBasicBlock::SimplifyIntegerRangeRelops(void)
 {
 	if (!mVisited)
@@ -13505,6 +13573,17 @@ bool InterCodeBasicBlock::PeepholeReplaceOptimization(const GrowingVariableArray
 				mInstructions[i + 1]->mSrc[0] = mInstructions[i + 0]->mDst;
 				changed = true;
 			}
+			else if (
+				mInstructions[i + 0]->mCode == IC_LEA && mInstructions[i + 0]->mSrc[0].mTemp < 0 && mInstructions[i + 0]->mSrc[1].mTemp >= 0 &&
+				mInstructions[i + 1]->mCode == IC_LOAD && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i + 0]->mDst.mTemp && mInstructions[i + 1]->mSrc[0].mFinal &&
+				mInstructions[i + 1]->mSrc[0].mIntConst + mInstructions[i + 0]->mSrc[0].mIntConst < 256)
+			{
+				mInstructions[i + 1]->mSrc[0].mTemp = mInstructions[i + 0]->mSrc[1].mTemp;
+				mInstructions[i + 1]->mSrc[0].mIntConst += mInstructions[i + 0]->mSrc[0].mIntConst;
+				mInstructions[i + 0]->mCode = IC_NONE;
+				mInstructions[i + 0]->mNumOperands = 0;
+				changed = true;
+			}
 			else if (
 				mInstructions[i + 0]->mCode == IC_LEA && mInstructions[i + 0]->mSrc[1].mTemp < 0 &&
 				mInstructions[i + 1]->mCode == IC_STORE && mInstructions[i + 1]->mSrc[1].mTemp == mInstructions[i + 0]->mDst.mTemp && mInstructions[i + 1]->mSrc[1].mFinal &&
@@ -15302,6 +15381,14 @@ void InterCodeProcedure::LoadStoreForwarding(InterMemory paramMemory)
 	} while (changed);
 }
 
+void InterCodeProcedure::CombineIndirectAddressing(void)
+{
+	ResetVisited();
+
+	mEntryBlock->CombineIndirectAddressing();
+	BuildDataFlowSets();
+}
+
 void InterCodeProcedure::PropagateConstOperationsUp(void)
 {
 #if 1
@@ -16022,6 +16109,8 @@ void InterCodeProcedure::Close(void)
 
 	PropagateConstOperationsUp();
 
+	CombineIndirectAddressing();
+
 #if 1
 	for (int i = 0; i < 4; i++)
 	{

oscar64/InterCode.h

@@ -434,6 +434,8 @@ public:
 	bool BuildGlobalIntegerRangeSets(bool initial, const GrowingVariableArray& localVars, const GrowingVariableArray& paramVars);
 	void SimplifyIntegerRangeRelops(void);
 
+	bool CombineIndirectAddressing(void);
+
 	GrowingIntArray			mEntryRenameTable;
 	GrowingIntArray			mExitRenameTable;
 
@@ -630,6 +632,7 @@ protected:
 	void ExpandSelect(void);
 	void PropagateConstOperationsUp(void);
 	void RebuildIntegerRangeSet(void);
+	void CombineIndirectAddressing(void);
 
 	void MergeBasicBlocks(void);
 	void CheckUsedDefinedTemps(void);

oscar64/NativeCodeGenerator.cpp

@@ -12831,7 +12831,7 @@ bool NativeCodeBasicBlock::ReplaceYRegWithXReg(int start, int end)
 {
 	bool	changed = false;
 
-	CheckLive();
+//	CheckLive();
 
 	for (int i = start; i < end; i++)
 	{
@@ -12840,7 +12840,7 @@ bool NativeCodeBasicBlock::ReplaceYRegWithXReg(int start, int end)
 			changed = true;
 	}
 
-	CheckLive();
+//	CheckLive();
 
 	return changed;
 }
@@ -13061,6 +13061,89 @@ bool NativeCodeBasicBlock::ForwardAccuAddSub(void)
 	return changed;
 }
 
+bool NativeCodeBasicBlock::ForwardAXYReg(void)
+{
+	bool	changed = false;
+
+	if (!mVisited)
+	{
+		mVisited = true;
+
+		bool	xisa = false, yisa = false;
+		int		xoffset = -1, yoffset = -1;
+
+		for (int i = 0; i < mIns.Size(); i++)
+		{
+			if (mIns[i].mType == ASMIT_TAX)
+			{
+				xisa = true;
+				xoffset = i;
+			}
+			else if (mIns[i].mType == ASMIT_TXA)
+			{
+				xisa = true;
+				yisa = false;
+				xoffset = i;
+			}
+			else if (mIns[i].mType == ASMIT_TAY)
+			{
+				yisa = true;
+				yoffset = i;
+			}
+			else if (mIns[i].mType == ASMIT_TYA)
+			{
+				yisa = true;
+				xisa = false;
+				yoffset = i;
+			}
+			else if (mIns[i].ChangesXReg())
+				xisa = false;
+			else if (mIns[i].ChangesYReg())
+				xisa = false;
+			else if (mIns[i].ChangesAccu())
+			{
+				xisa = false;
+				yisa = false;
+			}
+			else if (i + 1 < mIns.Size() && mIns[i].mType == ASMIT_CLC &&
+				mIns[i + 1].mType == ASMIT_ADC && mIns[i + 1].mMode == ASMIM_IMMEDIATE && !(mIns[i + 1].mLive & LIVE_CPU_REG_C))
+			{
+				if (xisa && !(mIns[i + 1].mLive & LIVE_CPU_REG_X))
+				{
+					if (mIns[i + 1].mAddress == 1)
+					{
+						mIns[i + 0].mType = ASMIT_INX;
+						mIns[i + 1].mType = ASMIT_TXA; mIns[i + 1].mMode = ASMIM_IMPLIED;
+						for (int j = xoffset; j < i + 1; j++)
+							mIns[j].mLive |= LIVE_CPU_REG_X;
+						xisa = false;
+						changed = true;
+					}
+				}
+				else if (yisa && !(mIns[i + 1].mLive & LIVE_CPU_REG_Y))
+				{
+					if (mIns[i + 1].mAddress == 1)
+					{
+						mIns[i + 0].mType = ASMIT_INY;
+						mIns[i + 1].mType = ASMIT_TYA; mIns[i + 1].mMode = ASMIM_IMPLIED;
+						for (int j = yoffset; j < i + 1; j++)
+							mIns[j].mLive |= LIVE_CPU_REG_Y;
+						yisa = false;
+						changed = true;
+					}
+				}
+			}
+		}
+
+		if (mTrueJump && mTrueJump->ForwardAXYReg())
+			changed = true;
+		if (mFalseJump && mFalseJump->ForwardAXYReg())
+			changed = true;
+	}
+
+	return changed;
+}
+
 bool NativeCodeBasicBlock::ForwardZpYIndex(bool full)
 {
 	CheckLive();
@@ -13487,21 +13570,41 @@ bool NativeCodeBasicBlock::CombineSameXY(void)
 		int	xpos, ypos;
 		bool	samexy = false;
 
+		CheckLive();
+
 		for (int i = 0; i < mIns.Size(); i++)
 		{
 			NativeCodeInstruction& ins(mIns[i]);
 
 			if (ins.ChangesXReg())
 			{
-				if (samexy && CombineSameXtoY(xpos, ypos, i))
+				if (samexy)
+				{
+					if (!ins.RequiresXReg() && CombineSameXtoY(xpos, ypos, i))
 						changed = true;
+					else if (!ins.RequiresYReg() && !(ins.mLive & LIVE_CPU_REG_Y) && CombineSameYtoX(xpos, ypos, i))
+					{
+						changed = true;
+						yreg = -1;
+					}
+				}
+
 				xreg = -1;
 				samexy = false;
 			}
 			if (ins.ChangesYReg())
 			{
-				if (samexy && CombineSameYtoX(ypos, xpos, i))
+				if (samexy)
+				{
+					if (!ins.RequiresYReg() && CombineSameYtoX(xpos, ypos, i))
 						changed = true;
+					else if (!ins.RequiresXReg() && !(ins.mLive & LIVE_CPU_REG_X) && CombineSameXtoY(xpos, ypos, i))
+					{
+						changed = true;
+						xreg = -1;
+					}
+				}
+
 				yreg = -1;
 				samexy = false;
 			}
@@ -13565,6 +13668,8 @@ bool NativeCodeBasicBlock::CombineSameXY(void)
 				changed = true;
 		}
 
+		CheckLive();
+
 		if (mTrueJump && mTrueJump->CombineSameXY())
 			changed = true;
 		if (mFalseJump && mFalseJump->CombineSameXY())
@@ -13854,14 +13959,14 @@ bool NativeCodeBasicBlock::ReplaceXRegWithYReg(int start, int end)
 {
 	bool	changed = false;
 
-	CheckLive();
+	//CheckLive();
 	for (int i = start; i < end; i++)
 	{
 		NativeCodeInstruction& ins(mIns[i]);
 		if (ins.ReplaceXRegWithYReg())
 			changed = true;
 	}
-	CheckLive();
+	//CheckLive();
 
 	return changed;
 }
@@ -39562,6 +39667,12 @@ void NativeCodeProcedure::Optimize(void)
 
 #endif
 
+		if (step == 8)
+		{
+			ResetVisited();
+			if (mEntryBlock->ForwardAXYReg())
+				changed = true;			
+		}
 #if 1
 		if (step == 10)
 		{

oscar64/NativeCodeGenerator.h

@@ -409,6 +409,7 @@ public:
 	bool ForwardAccuAddSub(void);
 	bool ForwardZpYIndex(bool full);
 	bool ForwardZpXIndex(bool full);
+	bool ForwardAXYReg(void);
 
 	bool RegisterValueForwarding(void);
 	bool CanCombineSameXtoY(int start, int end);

oscar64setup/oscar64setup.vdproj

@@ -268,6 +268,12 @@
         }
         "Entry"
         {
+        "MsmKey" = "8:_414E74B8FB564A7EBE074E9464FAB0F5"
+        "OwnerKey" = "8:_UNDEFINED"
+        "MsmSig" = "8:_UNDEFINED"
+        }
+        "Entry"
+        {
         "MsmKey" = "8:_41A6A127243E4EC2A1E1E2171993C5C1"
         "OwnerKey" = "8:_UNDEFINED"
         "MsmSig" = "8:_UNDEFINED"
@@ -2033,6 +2039,26 @@
             "IsDependency" = "11:FALSE"
             "IsolateTo" = "8:"
             }
+            "{1FB2D0AE-D3B9-43D4-B9DD-F88EC61E35DE}:_414E74B8FB564A7EBE074E9464FAB0F5"
+            {
+            "SourcePath" = "8:..\\samples\\hires\\qsort.prg"
+            "TargetName" = "8:qsort.prg"
+            "Tag" = "8:"
+            "Folder" = "8:_82FDD682B0334DE1B89A843D32862B85"
+            "Condition" = "8:"
+            "Transitive" = "11:FALSE"
+            "Vital" = "11:TRUE"
+            "ReadOnly" = "11:FALSE"
+            "Hidden" = "11:FALSE"
+            "System" = "11:FALSE"
+            "Permanent" = "11:FALSE"
+            "SharedLegacy" = "11:FALSE"
+            "PackageAs" = "3:1"
+            "Register" = "3:1"
+            "Exclude" = "11:FALSE"
+            "IsDependency" = "11:FALSE"
+            "IsolateTo" = "8:"
+            }
             "{1FB2D0AE-D3B9-43D4-B9DD-F88EC61E35DE}:_41A6A127243E4EC2A1E1E2171993C5C1"
             {
             "SourcePath" = "8:..\\samples\\scrolling\\grid2d.c"

samples/hires/build.sh

@@ -4,3 +4,5 @@
 ../../bin/oscar64 lines.c -n
 ../../bin/oscar64 polygon.c -n
 ../../bin/oscar64 bitblit.c -n
+../../bin/oscar64 fractaltree.c -n
+../../bin/oscar64 qsort.c -n

samples/hires/make.bat

@@ -5,3 +5,4 @@ call ..\..\bin\oscar64 polygon.c -n
 call ..\..\bin\oscar64 bitblit.c -n
 call ..\..\bin\oscar64 cube3d.c -n
 call ..\..\bin\oscar64 fractaltree.c -n
+call ..\..\bin\oscar64 qsort.c -n

samples/hires/qsort.c

@@ -0,0 +1,117 @@
+#include <c64/memmap.h>
+#include <c64/vic.h>
+#include <gfx/bitmap.h>
+#include <string.h>
+#include <conio.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <time.h>
+
+#define Color	((char *)0xd000)
+#define Hires	((char *)0xe000)
+
+Bitmap		Screen;
+
+void init(void)
+{
+	mmap_trampoline();
+	mmap_set(MMAP_RAM);
+
+	memset(Color, 0x01, 1000);
+	memset(Hires, 0x00, 8000);
+
+	mmap_set(MMAP_NO_ROM);
+
+	vic_setmode(VICM_HIRES, Color, Hires);
+
+	vic.color_border = VCOL_WHITE;
+
+	bm_init(&Screen, Hires, 40, 25);	
+}
+
+void done(void)
+{
+	mmap_set(MMAP_ROM);
+
+	getch();
+
+	vic_setmode(VICM_TEXT, (char *)0x0400, (char *)0x1000);
+}
+
+char field[160];
+
+void fill(void)
+{
+	for(int i=0; i<160; i++)
+		field[i] = i;
+}
+
+void shuffle(void)
+{
+	for(int i=0; i<160; i++)
+	{
+		int j = rand() % 160;
+		char t = field[i];
+		field[i] = field[j];
+		field[j] = t;
+	}	
+}
+
+void draw(unsigned i)
+{
+	bmu_line(&Screen, 2 * i, 0, 2 * i, field[i], 0x00, LINOP_SET);
+	bmu_line(&Screen, 2 * i, field[i], 2 * i, 160, 0xff, LINOP_SET);
+}
+
+void partition(int l, int r)
+{
+	while (l < r)
+	{
+		int i = l;
+		int j = r;		
+		char pi = field[(r + l) >> 1];
+		while (i <= j)
+		{
+			while (field[i] > pi)
+				i++;
+			while (field[j] < pi)
+				j--;
+			if (i <= j)
+			{
+				char t = field[i];
+				field[i] = field[j];
+				field[j] = t;
+				draw(i);
+				draw(j);
+				i++;
+				j--;
+			}
+		}
+
+		partition(l, j);
+		l = i;
+	}
+}
+
+int main(void)
+{
+	init();
+
+	fill();
+	shuffle();
+
+	for(int i=0; i<160; i++)
+		draw(i);
+
+	clock_t	t0 = clock();
+	partition(0, 159);
+	clock_t	t1 = clock();
+
+	char	t[20];
+	sprintf(t, "TIME : %.1f SECS.", (float)(t1 - t0) / 60);
+	bmu_put_chars(&Screen, 4, 170, t, strlen(t), BLTOP_COPY);
+
+	done();
+	
+	return 0;
+}