Native code size reduction

oscar64/InterCode.cpp

@@ -8088,6 +8088,7 @@ void InterCodeBasicBlock::PeepholeOptimization(void)
 						mInstructions[i + 2]->mSrc[1].mTemp == mInstructions[i + 1]->mDst.mTemp && mInstructions[i + 2]->mSrc[1].mFinal &&
 						(mInstructions[i + 2]->mSrc[0].mIntConst & 1) == 0)
 					{
+
 						int	shift = mInstructions[i + 0]->mSrc[0].mIntConst;
 						int	mshift = 1;
 						while (!(mInstructions[i + 2]->mSrc[0].mIntConst & (1ULL << mshift)))
@@ -8110,7 +8111,7 @@ void InterCodeBasicBlock::PeepholeOptimization(void)
 						{
 							mInstructions[i + 0]->mCode = IC_LOAD_TEMPORARY;
 							mInstructions[i + 0]->mSrc[0] = mInstructions[i + 0]->mSrc[1];
-							mInstructions[i + 0]->mSrc[0].mTemp = -1;
+							mInstructions[i + 0]->mSrc[1].mTemp = -1;
 
 							mInstructions[i + 1]->mSrc[1].mIntConst = 255ULL >> shift << shift;
 							mInstructions[i + 2]->mSrc[0].mIntConst >>= shift;

oscar64/NativeCodeGenerator.cpp

@@ -9492,7 +9492,64 @@ void NativeCodeBasicBlock::RemEntryBlock(NativeCodeBasicBlock* block)
 		mEntryBlocks.Remove(i);
 }
 
-bool NativeCodeBasicBlock::JoinTailCodeSequences(void)
+NativeCodeBasicBlock * NativeCodeBasicBlock::SplitMatchingTails(NativeCodeProcedure* proc)
+{
+	NativeCodeBasicBlock* nblock = nullptr;
+
+	for (int i = 0; i < mEntryBlocks.Size() - 1; i++)
+	{
+		NativeCodeBasicBlock* bi(mEntryBlocks[i]);
+
+		if (bi->mBranch == ASMIT_JMP && bi->mIns.Size() > 1)
+		{
+			for (int j = i + 1; j < mEntryBlocks.Size(); j++)
+			{
+				NativeCodeBasicBlock* bj(mEntryBlocks[j]);
+
+				if (bj->mBranch == ASMIT_JMP && bj->mIns.Size() > 1)
+				{
+					if (bi->mIns[bi->mIns.Size() - 1].IsSame(bj->mIns[bj->mIns.Size() - 1]) &&
+						bi->mIns[bi->mIns.Size() - 2].IsSame(bj->mIns[bj->mIns.Size() - 2]))
+					{
+						if (!nblock)
+						{
+							nblock = proc->AllocateBlock();
+							nblock->mBranch = ASMIT_JMP;
+							nblock->mVisited = false;
+							nblock->mTrueJump = this;
+
+							nblock->mEntryBlocks.Push(bi);
+							bi->mTrueJump = nblock;
+							mEntryBlocks[i] = nullptr;
+						}
+
+						nblock->mEntryBlocks.Push(bj);
+						bj->mTrueJump = nblock;
+						mEntryBlocks[j] = nullptr;
+					}
+				}
+			}
+
+			if (nblock)
+			{
+				int	i = 0;
+				while (i < mEntryBlocks.Size())
+				{
+					if (mEntryBlocks[i])
+						i++;
+					else
+						mEntryBlocks.Remove(i);
+				}
+
+				return nblock;
+			}
+		}
+	}
+
+	return nullptr;
+}
+
+bool NativeCodeBasicBlock::JoinTailCodeSequences(NativeCodeProcedure* proc)
 {
 	bool	changed = false;
 
@@ -9529,6 +9586,16 @@ bool NativeCodeBasicBlock::JoinTailCodeSequences(void)
 						break;
 				}
 			}
+			
+			if (mEntryBlocks.Size() > 2)
+			{
+				NativeCodeBasicBlock* nblock = SplitMatchingTails(proc);
+				if (nblock)
+				{
+					if (nblock->JoinTailCodeSequences(proc))
+						changed = true;
+				}
+			}
 		}
 #endif
 #if 1
@@ -9630,9 +9697,9 @@ bool NativeCodeBasicBlock::JoinTailCodeSequences(void)
 			}
 		}
 #endif
-		if (mTrueJump && mTrueJump->JoinTailCodeSequences())
+		if (mTrueJump && mTrueJump->JoinTailCodeSequences(proc))
 			changed = true;
-		if (mFalseJump && mFalseJump->JoinTailCodeSequences())
+		if (mFalseJump && mFalseJump->JoinTailCodeSequences(proc))
 			changed = true;
 	}
 
@@ -16168,6 +16235,26 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BypassEmptyBlocks(void)
 	}
 }
 
+int NativeCodeBasicBlock::LeadsInto(NativeCodeBasicBlock* block, int dist)
+{
+	if (mPlaced)
+		return 6;
+	else if (mTrueJump == block || mFalseJump == block)
+		return dist;
+	else if (dist < 5)
+	{
+		int d0 = mTrueJump ? mTrueJump->LeadsInto(block, dist + 1) : 6;
+		int d1 = mFalseJump ? mFalseJump->LeadsInto(block, dist + 1) : 6;
+
+		if (d0 < d1)
+			return d0;
+		else
+			return d1;
+	}
+
+	return 6;
+}
+
 void NativeCodeBasicBlock::BuildPlacement(GrowingArray<NativeCodeBasicBlock*>& placement)
 {
 	if (!mPlaced)
@@ -16190,12 +16277,12 @@ void NativeCodeBasicBlock::BuildPlacement(GrowingArray<NativeCodeBasicBlock*>& p
 
 				mTrueJump->BuildPlacement(placement);
 			}
-			else if (mTrueJump->mFalseJump == mFalseJump || mTrueJump->mTrueJump == mFalseJump)
+			else if (mTrueJump->LeadsInto(mFalseJump, 0) < mFalseJump->LeadsInto(mTrueJump, 0))
 			{
 				mTrueJump->BuildPlacement(placement);
 				mFalseJump->BuildPlacement(placement);
 			}
-			else if (mFalseJump->mFalseJump == mTrueJump || mFalseJump->mTrueJump == mTrueJump)
+			else if (mTrueJump->LeadsInto(mFalseJump, 0) > mFalseJump->LeadsInto(mTrueJump, 0))
 			{
 				mFalseJump->BuildPlacement(placement);
 				mTrueJump->BuildPlacement(placement);
@@ -17136,7 +17223,7 @@ void NativeCodeProcedure::Optimize(void)
 		if (step > 2)
 		{
 			ResetVisited();
-			if (mEntryBlock->JoinTailCodeSequences())
+			if (mEntryBlock->JoinTailCodeSequences(this))
 				changed = true;
 		}
 #endif

oscar64/NativeCodeGenerator.h

@@ -135,6 +135,7 @@ public:
 
 	NativeCodeBasicBlock* BypassEmptyBlocks(void);
 
+	int LeadsInto(NativeCodeBasicBlock* block, int dist);
 	void BuildPlacement(GrowingArray<NativeCodeBasicBlock*>& placement);
 	void InitialOffset(int& total);
 	bool CalculateOffset(int& total);
@@ -246,7 +247,9 @@ public:
 	void AddEntryBlock(NativeCodeBasicBlock* block);
 	void RemEntryBlock(NativeCodeBasicBlock* block);
 
-	bool JoinTailCodeSequences(void);
+	NativeCodeBasicBlock * SplitMatchingTails(NativeCodeProcedure* proc);
+
+	bool JoinTailCodeSequences(NativeCodeProcedure* proc);
 	bool SameTail(const NativeCodeInstruction& ins) const;
 
 	NativeRegisterDataSet	mEntryRegisterDataSet;

oscar64/oscar64.cpp

@@ -73,7 +73,7 @@ int main2(int argc, const char** argv)
 
 #else
 		strcpy(strProductName, "oscar64");
-		strcpy(strProductVersion, "1.4.94");
+		strcpy(strProductVersion, "1.4.95");
 
 #ifdef __APPLE__
 		uint32_t length = sizeof(basePath);

oscar64/oscar64.rc

@@ -25,8 +25,8 @@ LANGUAGE LANG_ENGLISH, SUBLANG_NEUTRAL
 //
 
 VS_VERSION_INFO VERSIONINFO
- FILEVERSION 1,4,94,0
+ FILEVERSION 1,4,95,0
- PRODUCTVERSION 1,4,94,0
+ PRODUCTVERSION 1,4,95,0
  FILEFLAGSMASK 0x3fL
 #ifdef _DEBUG
  FILEFLAGS 0x1L
@@ -43,12 +43,12 @@ BEGIN
         BEGIN
             VALUE "CompanyName", "oscar64"
             VALUE "FileDescription", "oscar64 compiler"
-            VALUE "FileVersion", "1.4.94.0"
+            VALUE "FileVersion", "1.4.95.0"
             VALUE "InternalName", "oscar64.exe"
             VALUE "LegalCopyright", "Copyright (C) 2021"
             VALUE "OriginalFilename", "oscar64.exe"
             VALUE "ProductName", "oscar64"
-            VALUE "ProductVersion", "1.4.94.0"
+            VALUE "ProductVersion", "1.4.95.0"
         END
     END
     BLOCK "VarFileInfo"

oscar64setup/oscar64setup.vdproj

@@ -3752,15 +3752,15 @@
         {
         "Name" = "8:Microsoft Visual Studio"
         "ProductName" = "8:oscar64"
-        "ProductCode" = "8:{DB788758-808B-4A56-B99C-F2BDBCD45FE3}"
+        "ProductCode" = "8:{37C3F00F-2A58-4FAD-B5BA-5574475C7831}"
-        "PackageCode" = "8:{9E6C1E34-6564-490E-B047-ED558001FF36}"
+        "PackageCode" = "8:{0F1773AB-6C73-4E53-B2D9-A794C417CED2}"
         "UpgradeCode" = "8:{9AB61EFF-ACAC-4079-9950-8D96615CD4EF}"
         "AspNetVersion" = "8:2.0.50727.0"
         "RestartWWWService" = "11:FALSE"
         "RemovePreviousVersions" = "11:TRUE"
         "DetectNewerInstalledVersion" = "11:TRUE"
         "InstallAllUsers" = "11:FALSE"
-        "ProductVersion" = "8:1.4.94"
+        "ProductVersion" = "8:1.4.95"
         "Manufacturer" = "8:oscar64"
         "ARPHELPTELEPHONE" = "8:"
         "ARPHELPLINK" = "8:"

samples/games/hscrollshmup.c

@@ -171,14 +171,15 @@ void tiles_draw1(char * dp, char * tm)
 
 struct Shot
 {
-	char	x, y, dx, n;
+	byte	ty, x, ry, n;
-}	shots[5][4];
+	sbyte	dx;
+}	shots[18];
+
+Shot	*	firstShot;
+Shot	*	lastShot;
 
 inline void shot_draw(char * dp, char i, char xp, char yp)
 {
-	__assume(i < 32);
-	__assume(yp < 8);
-
 	char		c = dp[xp];
 	dp[xp] = i | 0xe0;
 
@@ -189,7 +190,42 @@ inline void shot_draw(char * dp, char i, char xp, char yp)
 	fdp[4] = fsp[4]; fdp[5] = fsp[5]; fdp[6] = fsp[6]; fdp[7] = fsp[7];
 
 	fdp[yp] = 0x00;
+}
 
+void shot_add(int dx, int sy)
+{
+	char	py = sy - 6;
+	char	gy = py >> 5;
+	char	ey = (py >> 3) & 3;
+	char	ry = py & 7;
+
+	Shot	*	s = lastShot - 1;
+	while (s->ty > gy)
+	{		
+		s[1] = s[0];
+		s--;
+	}
+	s++;
+
+	lastShot++;
+	lastShot->ty = 6;
+
+	s->ty = gy;
+	s->ry = ry;
+	if (dx < 0)
+	{
+		s->dx = -1;
+		char	x = (148 - 4 * dx) >> 3;
+		s->n = x - 1;
+		s->x = 40 * ey + x;
+	}
+	else
+	{
+		s->dx = 1;
+		char	x = (156 - 4 * dx) >> 3;
+		s->x = 40 * ey + x;
+		s->n = 39 - x;
+	}
 }
 
 void tiles_draw(unsigned x)
@@ -202,6 +238,8 @@ void tiles_draw(unsigned x)
 	char	yl = 0;
 	char	ci = 0;
 
+	Shot	*	ss = firstShot, * ts = firstShot;
+
 	for(int iy=0; iy<5; iy++)
 	{
 		char	*	dp = Screen + 80 + 160 * iy;
@@ -262,22 +300,26 @@ void tiles_draw(unsigned x)
 			dp[k] = 0xf8;			
 		}
 
-
+		while (ss->ty == iy)
-		Shot	*	s = shots[iy];
-		for(char si=0; si<4; si++)
 		{
-			if (s->n)
+			ss->x += ss->dx;
+			ss->n--;
+			shot_draw(dp, ci++, ss->x, ss->ry);
+			if (ss->n)
 			{
-				s->x += s->dx;
+				if (ss != ts)
-				s->n--;
+					*ts = *ss;
-				shot_draw(dp, ci++, s->x, s->y);
+				ts++;
 			}
-			s++;
+			ss++;
 		}
 
 		yl += 4;
 	}
 
+	lastShot = ts;
+	lastShot->ty = 6;
+
 	Font[248 * 8 + 2] = ~(1 << xs);
 
 	vic.ctrl2 = VIC_CTRL2_MCM + xs;	
@@ -321,14 +363,10 @@ int main(void)
 	for(int i=0; i<24; i++)
 		stars[i] = rand() % 40 + 40 * (i & 3);
 
-	for(int i=0; i<5; i++)
+	shots[0].ty = 0;
-	{
+	firstShot = shots + 1;
-		for(int j=0; j<8; j++)
+	lastShot = firstShot;
-		{
+	lastShot->ty = 6;
-			shots[i][j].x = rand() % 160;
-			shots[i][j].y = rand() & 7;
-		}
-	}
 
 	spr_set(0, true, 160, 100, 64,          VCOL_BLUE, true, false, false);
 	spr_set(1, true, 160, 100, 64 + 16, VCOL_MED_GREY, true, false, false);
@@ -394,36 +432,8 @@ int main(void)
 			fdelay--;
 		else if (joyb[0] && vpx != 0)
 		{
-			char	py = spy - 6;
+			shot_add(vpx, spy);
-			char	gy = py >> 5;
+			fdelay = 5;
-			char	ey = (py >> 3) & 3;
-			char	ry = py & 7;
-
-			Shot	*	s = shots[gy];
-
-			char	i = 0;
-			while (i < 4 && s[i].n != 0)
-				i++;
-
-			if (i < 4)
-			{
-				s[i].y = ry;
-				if (vpx < 0)
-				{
-					s[i].dx = -1;
-					char	x = (148 - 4 * vpx) >> 3;
-					s[i].n = x - 1;
-					s[i].x = 40 * ey + x;
-				}
-				else if (vpx > 0)
-				{
-					s[i].dx = 1;
-					char	x = (156 - 4 * vpx) >> 3;
-					s[i].x = 40 * ey + x;
-					s[i].n = 39 - x;
-				}
-				fdelay = 4;
-			}
 		}
 
 		spr_move(0, 172 - 4 * vpx, 50 + spy);