Fix inter instruction cross block move

include/string.c

@@ -227,7 +227,7 @@ void * memmove(void * dst, const void * src, int size)
 		if (d < s)
 		{
 			do {
-				*d++ = *s++;				
+				*d++ = *s++;
 			} while (--sz);
 		}
 		else if (d > s)

oscar64/InterCode.cpp

@@ -10004,7 +10004,15 @@ int InterCodeBasicBlock::FindSameInstruction(const InterInstruction* ins) const
 
 bool InterCodeBasicBlock::CanMoveInstructionBehindBlock(int ii) const
 {
-	return CanMoveInstructionDown(ii, mInstructions.Size());
+	if (CanMoveInstructionDown(ii, mInstructions.Size() - 1))
+	{
+		InterInstruction* ins = mInstructions.Last();
+		if (ins->mCode == IC_BRANCH && mInstructions[ii]->mDst.mTemp == ins->mSrc[0].mTemp)
+			return false;
+		return true;
+	}
+	else
+		return false;
 }
 
 bool InterCodeBasicBlock::CanMoveInstructionBeforeBlock(int ii, const InterInstruction* ins) const
@@ -12163,6 +12171,39 @@ bool  InterCodeBasicBlock::CheckSingleBlockLimitedLoop(InterCodeBasicBlock*& pbl
 				}
 			}
 		}
+		else if (
+			mInstructions[nins - 1]->mCode == IC_BRANCH &&
+			mInstructions[nins - 2]->mCode == IC_BINARY_OPERATOR && mInstructions[nins - 2]->mOperator == IA_ADD)
+		{
+			InterInstruction* ains = mInstructions[nins - 2];
+			InterInstruction* bins = mInstructions[nins - 1];
+
+			if (bins->mSrc[0].mTemp == ains->mDst.mTemp &&
+				ains->mSrc[1].mTemp == ains->mDst.mTemp &&
+				ains->mSrc[0].mTemp < 0 &&
+				ains->mSrc[0].mIntConst == -1)
+			{
+				int pi = pblock->mInstructions.Size() - 1;
+				while (pi >= 0 && pblock->mInstructions[pi]->mDst.mTemp != ains->mDst.mTemp)
+					pi--;
+
+				if (pi >= 0 && pblock->mInstructions[pi]->mCode == IC_CONSTANT)
+				{
+					int i = 0;
+					while (i < nins - 2 && mInstructions[i]->mDst.mTemp != ains->mDst.mTemp)
+						i++;
+					if (i == nins - 2)
+					{
+						nloop = pblock->mInstructions[pi]->mConst.mIntConst;
+
+						mLocalValueRange[ains->mDst.mTemp].LimitMin(1);
+						mLocalValueRange[ains->mDst.mTemp].LimitMax(pblock->mInstructions[pi]->mConst.mIntConst);
+
+						return true;
+					}
+				}
+			}
+		}
 	}
 	
 	return false;
@@ -15791,7 +15832,7 @@ void InterCodeProcedure::Close(void)
 {
 	GrowingTypeArray	tstack(IT_NONE);
 
-	CheckFunc = !strcmp(mIdent->mString, "strtof");
+	CheckFunc = !strcmp(mIdent->mString, "main");
 
 	mEntryBlock = mBlocks[0];
 

oscar64/NativeCodeGenerator.cpp

@@ -16634,26 +16634,30 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 			int	sz = mIns.Size();
 
 			if (mIns[sz - 3].mType == ASMIT_STA && mIns[sz - 3].mMode == ASMIM_ZERO_PAGE &&
-				mIns[sz - 2].mType == ASMIT_LDA && mIns[sz - 2].mMode == ASMIM_ZERO_PAGE &&
+				mIns[sz - 2].mType == ASMIT_LDA && HasAsmInstructionMode(ASMIT_CMP, mIns[sz - 2].mMode) &&
-				mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_ZERO_PAGE && mIns[sz - 1].mAddress == mIns[sz - 3].mAddress && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_MEM)))
+				mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_ZERO_PAGE && mIns[sz - 1].mAddress == mIns[sz - 3].mAddress && !(mIns[sz - 1].mLive & LIVE_CPU_REG_A))
 			{
 				changed = true;
 
 				NativeCodeBasicBlock* cblock = proc->AllocateBlock();
-				cblock->Close(mBranchIns, mFalseJump, mTrueJump, mBranch);
+				cblock->Close(mBranchIns, mTrueJump, mFalseJump, mBranch);
 
-				mIns[sz - 1].mAddress = mIns[sz - 2].mAddress; mIns[sz - 1].mLive |= LIVE_CPU_REG_Z;
+				mIns[sz - 3].mLive |= LIVE_CPU_REG_A;
-				mIns[sz - 2].mType = ASMIT_NOP; mIns[sz - 2].mMode = ASMIM_IMPLIED;
+				mIns[sz - 2].mType = ASMIT_CMP; mIns[sz - 2].mLive |= LIVE_CPU_REG_Z | LIVE_CPU_REG_C;
+				mIns[sz - 1].mType = ASMIT_NOP; mIns[sz - 1].mMode = ASMIM_IMPLIED;
+				mExitRequiredRegs += CPU_REG_Z;
 
 				if (mBranch == ASMIT_BCC)
 				{
-					mBranch = ASMIT_BNE;
+					mTrueJump = mFalseJump;
-					mTrueJump = cblock;
+					mFalseJump = cblock;
+					cblock->mBranch = ASMIT_BNE;
 				}
 				else
 				{
-					mBranch = ASMIT_BEQ;
+					mFalseJump = mTrueJump;
-					mFalseJump = cblock;
+					mTrueJump = cblock;
+					cblock->mBranch = ASMIT_BEQ;					
 				}
 			}
 		}
@@ -27532,6 +27536,8 @@ bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc
 	NativeCodeBasicBlock* lblock = proc->AllocateBlock();
 	NativeCodeBasicBlock* eblock = proc->AllocateBlock();
 
+	NativeCodeBasicBlock* xblock = mTrueJump == this ? mFalseJump : mTrueJump;
+
 	eblock->mBranch = ASMIT_JMP;
 	eblock->mTrueJump = mFalseJump;
 	eblock->mFalseJump = nullptr;
@@ -27543,8 +27549,8 @@ bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc
 	lblock->mEntryRequiredRegs = mEntryRequiredRegs;
 	lblock->mExitRequiredRegs = mExitRequiredRegs;
 
-	eblock->mEntryRequiredRegs = mExitRequiredRegs;
+	eblock->mEntryRequiredRegs = xblock->mEntryRequiredRegs;
-	eblock->mExitRequiredRegs = mExitRequiredRegs;
+	eblock->mExitRequiredRegs = xblock->mEntryRequiredRegs;
 
 	mExitRequiredRegs = mEntryRequiredRegs;
 
@@ -27920,6 +27926,27 @@ static bool IsPointerIncAddrs(NativeCodeBasicBlock* block, int reg)
 	return changed;
 }
 
+bool NativeCodeBasicBlock::Is16BitImmSum(int at, int& val, int& reg) const
+{
+	if (at + 6 < mIns.Size())
+	{
+		if (mIns[at + 0].mType == ASMIT_CLC &&
+			mIns[at + 1].mType == ASMIT_LDA && mIns[at + 1].mMode == ASMIM_ZERO_PAGE &&
+			mIns[at + 2].mType == ASMIT_ADC && mIns[at + 2].mMode == ASMIM_IMMEDIATE &&
+			mIns[at + 3].mType == ASMIT_STA && mIns[at + 3].mMode == ASMIM_ZERO_PAGE && mIns[at + 3].mAddress == mIns[at + 1].mAddress &&
+			mIns[at + 4].mType == ASMIT_LDA && mIns[at + 4].mMode == ASMIM_ZERO_PAGE && mIns[at + 4].mAddress == mIns[at + 1].mAddress + 1 &&
+			mIns[at + 5].mType == ASMIT_ADC && mIns[at + 5].mMode == ASMIM_IMMEDIATE &&
+			mIns[at + 6].mType == ASMIT_STA && mIns[at + 6].mMode == ASMIM_ZERO_PAGE && mIns[at + 6].mAddress == mIns[at + 4].mAddress)
+		{
+			reg = mIns[at + 1].mAddress;
+			val = mIns[at + 2].mAddress + 256 * mIns[at + 5].mAddress;
+			return true;
+		}
+	}
+
+	return false;
+}
+
 
 bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc, NativeCodeBasicBlock* prevBlock, NativeCodeBasicBlock* exitBlock, bool full)
 {
@@ -27932,6 +27959,14 @@ bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc
 
 	CheckLive();
 
+	if (!exitBlock)
+	{
+		if (mTrueJump == this)
+			exitBlock = mFalseJump;
+		else
+			exitBlock = mTrueJump;
+	}
+
 	int	sz = mIns.Size();
 
 	if (sz == 2 && (mBranch == ASMIT_BEQ || mBranch == ASMIT_BNE) && mIns[0].mType == ASMIT_LDA && mIns[1].mType == ASMIT_CMP && !(mIns[1].mFlags & NCIF_VOLATILE) && !(mIns[1].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_C)))
@@ -29004,6 +29039,116 @@ bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc
 	}
 #endif
 
+
+	if (mIns.Size() > 0 && (mIns.Last().mType == ASMIT_DEX || mIns.Last().mType == ASMIT_DEC || mIns.Last().mType == ASMIT_CPX))
+	{
+		bool	yzero = false;
+
+		if (prevBlock)
+			yzero = prevBlock->RetrieveYValue(prevBlock->mIns.Size() - 1) == 0;
+
+		int		nregs = 0;
+		int		iregs[32];
+
+		for (int i = 0; i + 3 < mIns.Size(); i++)
+		{
+			if (mIns[i].mType == ASMIT_LDY && mIns[i].mMode == ASMIM_IMMEDIATE && mIns[i].mAddress == 0)
+				yzero = true;
+			else if (mIns[i].ChangesYReg())
+			{
+				nregs = 0;
+				break;
+			}
+
+			if (mIns[i].mMode == ASMIM_INDIRECT_Y)
+			{
+				if (!yzero)
+				{
+					nregs = 0;
+					break;
+				}
+
+				int reg = mIns[i].mAddress;
+
+				if (ReferencesZeroPage(mIns[i].mAddress, 0, i) ||
+					ReferencesZeroPage(mIns[i].mAddress + 1, 0, i))
+				{
+					nregs = 0;
+					break;
+				}
+				if (exitBlock->mEntryRequiredRegs[reg] || exitBlock->mEntryRequiredRegs[reg + 1] || exitBlock->mEntryRequiredRegs[CPU_REG_Y])
+				{
+					nregs = 0;
+					break;
+				}
+
+				int j = i + 1;
+				while (j < mIns.Size() && (
+					mIns[j].mMode == ASMIM_INDIRECT ||
+					!(mIns[j].ReferencesZeroPage(reg) || mIns[j].ReferencesZeroPage(reg + 1))))
+					j++;
+
+				if (j == mIns.Size())
+				{
+					nregs = 0;
+					break;
+				}
+
+				int sreg, dreg, offset;
+				if (Is16BitAddSubImmediate(j - 1, sreg, dreg, offset) && sreg == reg && dreg == sreg && offset == 1)
+				{
+					if (ReferencesZeroPage(reg, j + 6) || ReferencesZeroPage(reg + 1, j + 6))
+					{
+						nregs = 0;
+						break;
+					}
+					iregs[nregs++] = sreg;
+				}
+				else
+				{
+					nregs = 0;
+					break;
+				}
+			}
+		}
+
+		if (nregs)
+		{
+			if (!prevBlock)
+				return OptimizeSimpleLoopInvariant(proc, full);
+
+			for (int i = 0; i < mIns.Size(); i++)
+			{
+				if (mIns[i].mMode == ASMIM_ZERO_PAGE)
+				{
+					int j = 0;
+					while (j < nregs && iregs[j] != mIns[i].mAddress && iregs[j] + 1 != mIns[i].mAddress)
+						j++;
+					if (j < nregs)
+					{
+						mIns[i].mType = ASMIT_NOP;
+						mIns[i].mMode = ASMIM_IMPLIED;
+					}
+				}
+				else if (mIns[i].mType == ASMIT_LDY)
+				{
+					mIns[i].mType = ASMIT_NOP;
+					mIns[i].mMode = ASMIM_IMPLIED;
+				}
+
+				mIns[i].mLive |= LIVE_CPU_REG_Y;
+			}
+			int j = mIns.Size();
+			while (j > 0 && (mIns[j - 1].mLive & LIVE_CPU_REG_Z))
+				j--;
+			mIns.Insert(j, NativeCodeInstruction(mIns[0].mIns, ASMIT_INY));
+			prevBlock->mIns.Push(NativeCodeInstruction(mIns[0].mIns, ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+
+			changed = true;
+		}
+	}
+
+
 	CheckLive();
 
 	return changed;

oscar64/NativeCodeGenerator.h

@@ -478,6 +478,8 @@ public:
 	bool BypassRegisterConditionBlock(void);
 	bool FoldLoopEntry(void);
 
+	bool Is16BitImmSum(int at, int & val, int& reg) const;
+
 	bool Check16BitSum(int at, NativeRegisterSum16Info& info);
 	bool Propagate16BitSum(const ExpandingArray<NativeRegisterSum16Info>& cinfo);
 	

oscar64/oscar64.cpp

@@ -74,7 +74,7 @@ int main2(int argc, const char** argv)
 
 #else
 		strcpy(strProductName, "oscar64");
-		strcpy(strProductVersion, "1.19.201");
+		strcpy(strProductVersion, "1.19.202");
 
 #ifdef __APPLE__
 		uint32_t length = sizeof(basePath);

oscar64/oscar64.rc

@@ -25,8 +25,8 @@ LANGUAGE LANG_ENGLISH, SUBLANG_NEUTRAL
 //
 
 VS_VERSION_INFO VERSIONINFO
- FILEVERSION 1,19,201,0
+ FILEVERSION 1,19,202,0
- PRODUCTVERSION 1,19,201,0
+ PRODUCTVERSION 1,19,202,0
  FILEFLAGSMASK 0x3fL
 #ifdef _DEBUG
  FILEFLAGS 0x1L
@@ -43,12 +43,12 @@ BEGIN
         BEGIN
             VALUE "CompanyName", "oscar64"
             VALUE "FileDescription", "oscar64 compiler"
-            VALUE "FileVersion", "1.19.201.0"
+            VALUE "FileVersion", "1.19.202.0"
             VALUE "InternalName", "oscar64.exe"
             VALUE "LegalCopyright", "Copyright (C) 2021"
             VALUE "OriginalFilename", "oscar64.exe"
             VALUE "ProductName", "oscar64"
-            VALUE "ProductVersion", "1.19.201.0"
+            VALUE "ProductVersion", "1.19.202.0"
         END
     END
     BLOCK "VarFileInfo"

oscar64setup/oscar64setup.vdproj

@@ -5233,15 +5233,15 @@
         {
         "Name" = "8:Microsoft Visual Studio"
         "ProductName" = "8:oscar64"
-        "ProductCode" = "8:{AED61249-4A7B-4472-9783-1F5014AC87EF}"
+        "ProductCode" = "8:{121C9EAF-FA66-4A4C-9CBB-B9CAC3B869C7}"
-        "PackageCode" = "8:{99957648-E168-4853-BA75-AA7B61A0AE71}"
+        "PackageCode" = "8:{249C96D2-72FE-4889-8F29-99C7507EC96D}"
         "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" = "8:1.19.201"
+        "ProductVersion" = "8:1.19.202"
         "Manufacturer" = "8:oscar64"
         "ARPHELPTELEPHONE" = "8:"
         "ARPHELPLINK" = "8:"
@@ -5755,7 +5755,7 @@
         {
             "{5259A561-127C-4D43-A0A1-72F10C7B3BF8}:_FB2E467BC172457785F4279BB0BFE8B6"
             {
-            "SourcePath" = "8:..\\Debug\\oscar64.exe"
+            "SourcePath" = "8:..\\Release\\oscar64.exe"
             "TargetName" = "8:"
             "Tag" = "8:"
             "Folder" = "8:_607E75AF0E2A4CB9908C4C39DF8FE6E4"