Optimize 32bit shifts

include/c64/asm6502.c

@@ -0,0 +1,94 @@
+#include "asm6502.h"
+
+inline byte asm_np(byte * ip, AsmIns ins)
+{
+	ip[0] = ins & 0xff;
+	return 1;
+}
+
+inline byte asm_ac(byte * ip, AsmIns ins)
+{
+	ip[0] = (ins & 0xff) | 0x08;
+	return 1;
+}
+
+inline byte asm_zp(byte * ip, AsmIns ins, byte addr)
+{
+	ip[0] = (ins & 0xff) | 0x04;
+	ip[1] = addr;
+	return 2;
+}
+
+inline byte asm_rl(byte * ip, AsmIns ins, byte addr)
+{
+	ip[0] = ins & 0xff;
+	ip[1] = addr;
+	return 2;
+}
+
+inline byte asm_im(byte * ip, AsmIns ins, byte value)
+{
+	ip[0] = (ins & 0xff) | ((ins & 0x01) << 3);
+	ip[1] = value;
+	return 2;
+}
+
+inline byte asm_zx(byte * ip, AsmIns ins, byte addr)
+{
+	ip[0] = (ins & 0xff) | 0x05;
+	ip[1] = addr;
+	return 2;
+}
+
+inline byte asm_zy(byte * ip, AsmIns ins, byte addr)
+{
+	ip[0] = (ins & 0xff) | 0x05;
+	ip[1] = addr;
+	return 2;
+}
+
+inline byte asm_ab(byte * ip, AsmIns ins, unsigned addr)
+{
+	ip[0] = (ins & 0xff) ^ 0x0c;
+	ip[1] = addr & 0xff;
+	ip[2] = addr >> 8;
+	return 3;
+}
+
+inline byte asm_in(byte * ip, AsmIns ins, unsigned addr)
+{
+	ip[0] = (ins & 0xff) ^ 0x2c;
+	ip[1] = addr & 0xff;
+	ip[2] = addr >> 8;
+	return 3;
+}
+
+inline byte asm_ax(byte * ip, AsmIns ins, unsigned addr)
+{
+	ip[0] = (ins & 0xff) | 0x1c;
+	ip[1] = addr & 0xff;
+	ip[2] = addr >> 8;
+	return 3;
+}
+
+inline byte asm_ay(byte * ip, AsmIns ins, unsigned addr)
+{
+	ip[0] = (ins & 0xff) | 0x18;
+	ip[1] = addr & 0xff;
+	ip[2] = addr >> 8;
+	return 3;
+}
+
+inline byte asm_ix(byte * ip, AsmIns ins, byte addr)
+{
+	ip[0] = (ins & 0xff) | 0x00;
+	ip[1] = addr;
+	return 2;
+}
+
+inline byte asm_iy(byte * ip, AsmIns ins, byte addr)
+{
+	ip[0] = (ins & 0xff) | 0x10;
+	ip[1] = addr;
+	return 2;
+}

include/c64/asm6502.h

@@ -82,98 +82,33 @@ enum AsmIns
 	ASM_JSR = 0x2c
 };
 
-inline byte asm_np(byte * ip, AsmIns ins)
+inline byte asm_np(byte * ip, AsmIns ins);
-{
-	ip[0] = ins & 0xff;
-	return 1;
-}
 
-inline byte asm_ac(byte * ip, AsmIns ins)
+inline byte asm_ac(byte * ip, AsmIns ins);
-{
-	ip[0] = (ins & 0xff) | 0x08;
-	return 1;
-}
 
-inline byte asm_zp(byte * ip, AsmIns ins, byte addr)
+inline byte asm_zp(byte * ip, AsmIns ins, byte addr);
-{
-	ip[0] = (ins & 0xff) | 0x04;
-	ip[1] = addr;
-	return 2;
-}
 
-inline byte asm_rl(byte * ip, AsmIns ins, byte addr)
+inline byte asm_rl(byte * ip, AsmIns ins, byte addr);
-{
-	ip[0] = ins & 0xff;
-	ip[1] = addr;
-	return 2;
-}
 
-inline byte asm_im(byte * ip, AsmIns ins, byte value)
+inline byte asm_im(byte * ip, AsmIns ins, byte value);
-{
-	ip[0] = (ins & 0xff) | ((ins & 0x01) << 3);
-	ip[1] = value;
-	return 2;
-}
 
-inline byte asm_zx(byte * ip, AsmIns ins, byte addr)
+inline byte asm_zx(byte * ip, AsmIns ins, byte addr);
-{
-	ip[0] = (ins & 0xff) | 0x05;
-	ip[1] = addr;
-	return 2;
-}
 
-inline byte asm_zy(byte * ip, AsmIns ins, byte addr)
+inline byte asm_zy(byte * ip, AsmIns ins, byte addr);
-{
-	ip[0] = (ins & 0xff) | 0x05;
-	ip[1] = addr;
-	return 2;
-}
 
-inline byte asm_ab(byte * ip, AsmIns ins, unsigned addr)
+inline byte asm_ab(byte * ip, AsmIns ins, unsigned addr);
-{
-	ip[0] = (ins & 0xff) ^ 0x0c;
-	ip[1] = addr & 0xff;
-	ip[2] = addr >> 8;
-	return 3;
-}
 
-inline byte asm_in(byte * ip, AsmIns ins, unsigned addr)
+inline byte asm_in(byte * ip, AsmIns ins, unsigned addr);
-{
-	ip[0] = (ins & 0xff) ^ 0x2c;
-	ip[1] = addr & 0xff;
-	ip[2] = addr >> 8;
-	return 3;
-}
 
-inline byte asm_ax(byte * ip, AsmIns ins, unsigned addr)
+inline byte asm_ax(byte * ip, AsmIns ins, unsigned addr);
-{
-	ip[0] = (ins & 0xff) | 0x1c;
-	ip[1] = addr & 0xff;
-	ip[2] = addr >> 8;
-	return 3;
-}
 
-inline byte asm_ay(byte * ip, AsmIns ins, unsigned addr)
+inline byte asm_ay(byte * ip, AsmIns ins, unsigned addr);
-{
-	ip[0] = (ins & 0xff) | 0x18;
-	ip[1] = addr & 0xff;
-	ip[2] = addr >> 8;
-	return 3;
-}
 
-inline byte asm_ix(byte * ip, AsmIns ins, byte addr)
+inline byte asm_ix(byte * ip, AsmIns ins, byte addr);
-{
-	ip[0] = (ins & 0xff) | 0x00;
-	ip[1] = addr;
-	return 2;
-}
 
-inline byte asm_iy(byte * ip, AsmIns ins, byte addr)
+inline byte asm_iy(byte * ip, AsmIns ins, byte addr);
-{
+
-	ip[0] = (ins & 0xff) | 0x10;
+#pragma compile("asm6502.c")
-	ip[1] = addr;
-	return 2;
-}
 
 #endif
 

include/c64/memmap.c

@@ -0,0 +1,61 @@
+#include "memmap.h"
+
+char PLAShadow;
+
+__asm DoneTrampoline
+{
+	lda PLAShadow
+	sta $01
+	pla
+	tax
+	pla
+	rti
+}
+
+__asm IRQTrampoline
+{
+	pha
+	txa
+	pha
+	lda #$36
+	sta $01
+
+	lda #>DoneTrampoline
+	pha
+	lda #<DoneTrampoline
+	pha
+	tsx
+	lda $0105, x
+	pha
+	jmp ($fffa)
+}
+
+__asm NMITrampoline
+{
+	pha
+	txa
+	pha
+	lda #$36
+	sta $01
+
+	lda #>DoneTrampoline
+	pha
+	lda #<DoneTrampoline
+	pha
+	tsx
+	lda $0105, x
+	pha
+	jmp ($fffe)
+}
+
+void mmap_trampoline(void)
+{
+	*((void **)0xfffa) = IRQTrampoline;
+	*((void **)0xfffe) = NMITrampoline;	
+}
+
+void mmap_set(char pla)
+{
+	PLAShadow = pla;
+	*((char *)0x01) = pla;
+}

include/c64/memmap.h

@@ -0,0 +1,25 @@
+#ifndef MEMMAP_H
+#define MEMMAP_H
+
+#include "types.h"
+
+#define MMAP_ROM		0x37
+#define MMAP_NO_BASIC	0x36
+#define MMAP_NO_ROM		0x35
+#define MMAP_RAM		0x30
+
+// Install an IRQ an NMI trampoline, that routes the kernal interrupts
+// through an intermediate trampoline when the kernal ROM is not paged
+// in.  The trampoline enables the ROM, executes the interrupt and
+// restores the memory map setting before returning.
+
+void mmap_trampoline(void);
+
+// Set the memory map in a way that is compatible with the IRQ
+// trampoline
+
+inline void mmap_set(char pla);
+
+#pragma compile("memmap.c")
+
+#endif

include/c64/vic.c

@@ -15,3 +15,33 @@ void vic_sprxy(byte s, int x, int y)
 	else
 		vic.spr_msbx &= ~(1 << s);
 }
+
+void vic_setmode(VicMode mode, char * text, char * font)
+{
+	switch (mode)
+	{
+		case VICM_TEXT:
+			vic.ctrl1 = VIC_CTRL1_DEN | VIC_CTRL1_RSEL | 3;
+			vic.ctrl2 = VIC_CTRL2_CSEL;
+			break;
+		case VICM_TEXT_MC:
+			vic.ctrl1 = VIC_CTRL1_DEN | VIC_CTRL1_RSEL | 3;
+			vic.ctrl2 = VIC_CTRL2_CSEL | VIC_CTRL2_MCM;
+			break;
+		case VICM_TEXT_ECM:
+			vic.ctrl1 = VIC_CTRL1_DEN | VIC_CTRL1_ECM | VIC_CTRL1_RSEL | 3;
+			vic.ctrl2 = VIC_CTRL2_CSEL;
+			break;
+		case VICM_HIRES:
+			vic.ctrl1 = VIC_CTRL1_BMM | VIC_CTRL1_DEN | VIC_CTRL1_RSEL | 3;
+			vic.ctrl2 = VIC_CTRL2_CSEL;
+			break;
+		case VICM_HIRES_MC:
+			vic.ctrl1 = VIC_CTRL1_BMM | VIC_CTRL1_DEN | VIC_CTRL1_RSEL | 3;
+			vic.ctrl2 = VIC_CTRL2_CSEL | VIC_CTRL2_MCM;
+			break;
+	}
+
+	cia2.pra = (cia2.pra & 0xfc) | (((unsigned)text >> 14) ^ 0x03);	
+	vic.memptr = (((unsigned)text >> 6) & 0xf0) | (((unsigned)font >> 10) & 0x0e);
+}

include/c64/vic.h

@@ -75,6 +75,17 @@ struct VIC
 
 void vic_setbank(char bank);
 
+enum VicMode
+{
+	VICM_TEXT,
+	VICM_TEXT_MC,
+	VICM_TEXT_ECM,
+	VICM_HIRES,
+	VICM_HIRES_MC
+};
+
+void vic_setmode(VicMode mode, char * text, char * font);
+
 inline void vic_sprxy(byte s, int x, int y);
 
 

oscar64/Declaration.cpp

@@ -368,7 +368,7 @@ Expression* Expression::ConstantFold(Errors * errors)
 }
 
 Declaration::Declaration(const Location& loc, DecType type)
-	: mLocation(loc), mType(type), mScope(nullptr), mData(nullptr), mIdent(nullptr), mSize(0), mOffset(0), mFlags(0), mComplexity(0), mLocalSize(0), mBase(nullptr), mParams(nullptr), mValue(nullptr), mNext(nullptr), mVarIndex(-1), mLinkerObject(nullptr), mCallers(nullptr), mCalled(nullptr)
+	: mLocation(loc), mType(type), mScope(nullptr), mData(nullptr), mIdent(nullptr), mSize(0), mOffset(0), mFlags(0), mComplexity(0), mLocalSize(0), mBase(nullptr), mParams(nullptr), mValue(nullptr), mNext(nullptr), mVarIndex(-1), mLinkerObject(nullptr), mCallers(nullptr), mCalled(nullptr), mAlignment(1)
 {}
 
 Declaration::~Declaration(void)

oscar64/Declaration.h

@@ -167,7 +167,7 @@ public:
 	Declaration*		mBase, *mParams, * mNext;
 	Expression*			mValue;
 	DeclarationScope*	mScope;
-	int					mOffset, mSize, mVarIndex, mNumVars, mComplexity, mLocalSize;
+	int					mOffset, mSize, mVarIndex, mNumVars, mComplexity, mLocalSize, mAlignment;
 	int64				mInteger;
 	double				mNumber;
 	uint32				mFlags;

oscar64/GlobalAnalyzer.cpp

@@ -59,7 +59,7 @@ void GlobalAnalyzer::AutoInline(void)
 		for (int i = 0; i < mFunctions.Size(); i++)
 		{
 			Declaration* f = mFunctions[i];
-			if (!(f->mFlags & DTF_INLINE) && !(f->mBase->mFlags & DTF_VARIADIC) && !(f->mFlags & DTF_FUNC_VARIABLE) && !(f->mFlags & DTF_FUNC_ASSEMBLER) && !(f->mFlags & DTF_INTRINSIC) && !(f->mFlags & DTF_FUNC_RECURSIVE) && f->mLocalSize < 100)
+			if (!(f->mFlags & DTF_INLINE) && !(f->mBase->mFlags & DTF_VARIADIC) && !(f->mFlags & DTF_FUNC_VARIABLE) && !((f->mFlags & DTF_FUNC_ASSEMBLER) && !(f->mFlags & DTF_REQUEST_INLINE)) && !(f->mFlags & DTF_INTRINSIC) && !(f->mFlags & DTF_FUNC_RECURSIVE) && f->mLocalSize < 100)
 			{
 				int		nparams = 0;
 				Declaration* dec = f->mBase->mParams;

oscar64/InterCodeGenerator.cpp

@@ -214,7 +214,7 @@ void InterCodeGenerator::InitGlobalVariable(InterCodeModule * mod, Declaration*
 		InterVariable	*	var = new InterVariable();
 		var->mOffset = 0;
 		var->mSize = dec->mSize;
-		var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA);
+		var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA, dec->mAlignment);
 		var->mIdent = dec->mIdent;
 
 		Declaration* type = dec->mBase;
@@ -830,7 +830,7 @@ InterCodeGenerator::ExValue InterCodeGenerator::TranslateExpression(Declaration*
 					var->mSize = dec->mSize;
 					if ((dec->mFlags & DTF_VAR_ALIASING) || dec->mBase->mType == DT_TYPE_ARRAY)
 						var->mAliased = true;
-					var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA);
+					var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA, dec->mAlignment);
 					dec->mLinkerObject = var->mLinkerObject;
 					var->mLinkerObject->AddData(dec->mData, dec->mSize);
 					proc->mModule->mGlobalVars.Push(var);
@@ -855,7 +855,7 @@ InterCodeGenerator::ExValue InterCodeGenerator::TranslateExpression(Declaration*
 					InterVariable	* var = new InterVariable();
 					var->mOffset = 0;
 					var->mSize = dec->mSize;
-					var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA);
+					var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA, dec->mAlignment);
 					dec->mLinkerObject = var->mLinkerObject;
 					var->mIdent = dec->mIdent;
 					dec->mVarIndex = proc->mModule->mGlobalVars.Size();
@@ -3222,7 +3222,7 @@ void InterCodeGenerator::BuildInitializer(InterCodeModule * mod, uint8* dp, int
 				var->mIndex = dec->mVarIndex;
 				var->mOffset = 0;
 				var->mSize = dec->mSize;
-				var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA);
+				var->mLinkerObject = mLinker->AddObject(dec->mLocation, dec->mIdent, dec->mSection, LOT_DATA, dec->mAlignment);
 				dec->mLinkerObject = var->mLinkerObject;
 				var->mLinkerObject->AddData(dec->mData, dec->mSize);
 				mod->mGlobalVars.Push(var);
@@ -3233,6 +3233,21 @@ void InterCodeGenerator::BuildInitializer(InterCodeModule * mod, uint8* dp, int
 			variable->mLinkerObject->AddReference(ref);
 			break;
 		}
+		case DT_VARIABLE:
+		{
+			if (!dec->mLinkerObject)
+			{
+				InitGlobalVariable(mod, dec);
+			}
+
+			LinkerReference	ref;
+			ref.mObject = variable->mLinkerObject;
+			ref.mOffset = offset;
+			ref.mFlags = LREF_LOWBYTE | LREF_HIGHBYTE;
+			ref.mRefObject = dec->mLinkerObject;
+			ref.mRefOffset = 0;
+			variable->mLinkerObject->AddReference(ref);
+		}	break;
 		}
 	}
 }

oscar64/Linker.cpp

@@ -12,7 +12,7 @@ LinkerSection::LinkerSection(void)
 {}
 
 LinkerObject::LinkerObject(void)
-	: mReferences(nullptr), mNumTemporaries(0)
+	: mReferences(nullptr), mNumTemporaries(0), mSize(0), mAlignment(1)
 {}
 
 LinkerObject::~LinkerObject(void)
@@ -133,7 +133,7 @@ LinkerObject* Linker::FindObjectByAddr(int addr)
 	return nullptr;
 }
 
-LinkerObject * Linker::AddObject(const Location& location, const Ident* ident, LinkerSection * section, LinkerObjectType type)
+LinkerObject * Linker::AddObject(const Location& location, const Ident* ident, LinkerSection * section, LinkerObjectType type, int alignment)
 {
 	LinkerObject* obj = new LinkerObject;
 	obj->mLocation = location;
@@ -146,6 +146,7 @@ LinkerObject * Linker::AddObject(const Location& location, const Ident* ident, L
 	obj->mRegion = nullptr;
 	obj->mProc = nullptr;
 	obj->mFlags = 0;
+	obj->mAlignment = alignment;
 	section->mObjects.Push(obj);
 	mObjects.Push(obj);
 	return obj;
@@ -201,8 +202,9 @@ void Linker::Link(void)
 					if ((lobj->mFlags & LOBJF_REFERENCED) && !(lobj->mFlags & LOBJF_PLACED) && lrgn->mStart + lrgn->mUsed + lobj->mSize <= lrgn->mEnd)
 					{
 						lobj->mFlags |= LOBJF_PLACED;
-						lobj->mAddress = lrgn->mStart + lrgn->mUsed;
+						lobj->mAddress = (lrgn->mStart + lrgn->mUsed + lobj->mAlignment - 1) & ~(lobj->mAlignment - 1);
-						lrgn->mUsed += lobj->mSize;
+						lrgn->mUsed = lobj->mAddress + lobj->mSize - lrgn->mStart;
+
 						lobj->mRegion = lrgn;
 
 						if (lsec->mType == LST_DATA)

oscar64/Linker.h

@@ -93,7 +93,7 @@ public:
 	LinkerObjectType	mType;
 	int					mID;
 	int					mAddress;
-	int					mSize;
+	int					mSize, mAlignment;
 	LinkerSection	*	mSection;
 	LinkerRegion	*	mRegion;
 	uint8			*	mData;
@@ -129,7 +129,7 @@ public:
 
 	bool IsSectionPlaced(LinkerSection* section);
 
-	LinkerObject * AddObject(const Location & location, const Ident* ident, LinkerSection * section, LinkerObjectType type);
+	LinkerObject * AddObject(const Location & location, const Ident* ident, LinkerSection * section, LinkerObjectType type, int alignment = 1);
 
 //	void AddReference(const LinkerReference& ref);
 

oscar64/NativeCodeGenerator.cpp

@@ -822,6 +822,12 @@ void NativeCodeInstruction::Simulate(NativeRegisterDataSet& data)
 			data.mRegs[BC_REG_ADDR + i].Reset();
 		}
 		data.mRegs[BC_REG_WORK_Y].Reset();
+
+		if (!(mFlags & NCIF_RUNTIME))
+		{
+			for (int i = BC_REG_TMP; i < BC_REG_TMP_SAVED; i++)
+				data.mRegs[i].Reset();
+		}
 		break;
 
 	case ASMIT_ROL:
@@ -1455,6 +1461,12 @@ bool NativeCodeInstruction::ValueForwarding(NativeRegisterDataSet& data, AsmInsT
 		}
 		data.ResetZeroPage(BC_REG_WORK_Y);
 
+		if (!(mFlags & NCIF_RUNTIME))
+		{
+			for (int i = BC_REG_TMP; i < BC_REG_TMP_SAVED; i++)
+				data.ResetZeroPage(i);
+		}
+
 		return false;
 	}
 
@@ -4972,34 +4984,75 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 			if (ins->mSrc[0].mTemp < 0)
 			{
 				int	shift = ins->mSrc[0].mIntConst & 31;
+
+				int	sreg = BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp];
+
+				if (shift >= 24)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					sreg = treg;
+					shift -= 24;
+				}
+				else if (shift >= 16)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					sreg = treg;
+					shift -= 16;
+				}
+				else if (shift >= 8)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					sreg = treg;
+					shift -= 8;
+				}
+
 				if (shift == 0)
 				{
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
 					}
 				}
 				else if (shift == 1)
 				{
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_ASL, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
 					}
@@ -5016,15 +5069,15 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 					NativeCodeBasicBlock* lblock = nproc->AllocateBlock();
 					NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
 
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 					}
 					else
 						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, treg + 3));
@@ -5100,34 +5153,75 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 			if (ins->mSrc[0].mTemp < 0)
 			{
 				int	shift = ins->mSrc[0].mIntConst & 31;
+
+				int	sreg = BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp];
+
+				if (shift >= 24)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					sreg = treg;
+					shift -= 24;
+				}
+				else if (shift >= 16)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					sreg = treg;
+					shift -= 16;
+				}
+				else if (shift >= 8)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					sreg = treg;
+					shift -= 8;
+				}
+
 				if (shift == 0)
 				{
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
 					}
 				}
 				else if (shift == 1)
 				{
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 						mIns.Push(NativeCodeInstruction(ASMIT_LSR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 0));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 0));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 0));
 					}
@@ -5144,15 +5238,15 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 					NativeCodeBasicBlock* lblock = nproc->AllocateBlock();
 					NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
 
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 					}
 					else
 						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, treg + 3));
@@ -5228,35 +5322,84 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 			if (ins->mSrc[0].mTemp < 0)
 			{
 				int	shift = ins->mSrc[0].mIntConst & 31;
+
+				int	sreg = BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp];
+
+				if (shift >= 24)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_ASL, ASMIM_IMPLIED));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0x00));
+					mIns.Push(NativeCodeInstruction(ASMIT_ADC, ASMIM_IMMEDIATE, 0xff));
+					mIns.Push(NativeCodeInstruction(ASMIT_EOR, ASMIM_IMMEDIATE, 0xff));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					sreg = treg;
+					shift -= 24;
+				}
+				else if (shift >= 16)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_ASL, ASMIM_IMPLIED));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0x00));
+					mIns.Push(NativeCodeInstruction(ASMIT_ADC, ASMIM_IMMEDIATE, 0xff));
+					mIns.Push(NativeCodeInstruction(ASMIT_EOR, ASMIM_IMMEDIATE, 0xff));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					sreg = treg;
+					shift -= 16;
+				}
+				else if (shift >= 8)
+				{
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
+					mIns.Push(NativeCodeInstruction(ASMIT_ASL, ASMIM_IMPLIED));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0x00));
+					mIns.Push(NativeCodeInstruction(ASMIT_ADC, ASMIM_IMMEDIATE, 0xff));
+					mIns.Push(NativeCodeInstruction(ASMIT_EOR, ASMIM_IMMEDIATE, 0xff));
+					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
+					sreg = treg;
+					shift -= 8;
+				}
+
 				if (shift == 0)
 				{
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
 					}
 				}
 				else if (shift == 1)
 				{
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 						mIns.Push(NativeCodeInstruction(ASMIT_CMP, ASMIM_IMMEDIATE, 0x80));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 0));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 0));
 						mIns.Push(NativeCodeInstruction(ASMIT_ROR, ASMIM_IMPLIED));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 0));
 					}
@@ -5275,15 +5418,15 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 					NativeCodeBasicBlock* lblock = nproc->AllocateBlock();
 					NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
 
-					if (ins->mSrc[1].mTemp != ins->mDst.mTemp)
+					if (sreg != treg)
 					{
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp]));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 1));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 1));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 2));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 2));
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp] + 3));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, sreg + 3));
 					}
 					else
 						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, treg + 3));
@@ -8087,6 +8230,34 @@ bool NativeCodeBasicBlock::JoinTailCodeSequences(void)
 			}
 		}
 
+		if (mTrueJump && mFalseJump && !mTrueJump->mFalseJump && !mFalseJump->mFalseJump)
+		{
+			if (mTrueJump->mIns.Size() > mFalseJump->mIns.Size())
+			{
+				int	i = 0, offset = mTrueJump->mIns.Size() - mFalseJump->mIns.Size();
+				while (i < mFalseJump->mIns.Size() && mFalseJump->mIns[i].IsSame(mTrueJump->mIns[i + offset]))
+					i++;
+				if (i == mFalseJump->mIns.Size())
+				{
+					mTrueJump->mTrueJump = mFalseJump;
+					mTrueJump->mIns.SetSize(offset);
+					changed = true;
+				}
+			}
+			else
+			{
+				int	i = 0, offset = mFalseJump->mIns.Size() - mTrueJump->mIns.Size();
+				while (i < mTrueJump->mIns.Size() && mTrueJump->mIns[i].IsSame(mFalseJump->mIns[i + offset]))
+					i++;
+				if (i == mTrueJump->mIns.Size())
+				{
+					mFalseJump->mTrueJump = mTrueJump;
+					mFalseJump->mIns.SetSize(offset);
+					changed = true;
+				}
+			}
+		}
+
 		if (mTrueJump && mTrueJump->JoinTailCodeSequences())
 			changed = true;
 		if (mFalseJump && mFalseJump->JoinTailCodeSequences())
@@ -8096,6 +8267,28 @@ bool NativeCodeBasicBlock::JoinTailCodeSequences(void)
 	return changed;
 }
 
+bool NativeCodeBasicBlock::FindPageStartAddress(int at, int reg, int& addr)
+{
+	int	j = at - 2;
+	while (j >= 0)
+	{
+		if (mIns[j + 0].mType == ASMIT_LDA && mIns[j + 0].mMode == ASMIM_IMMEDIATE && 
+			mIns[j + 1].mType == ASMIT_STA && mIns[j + 1].mMode == ASMIM_ZERO_PAGE && mIns[j + 1].mAddress == reg + 1)
+		{
+			addr = mIns[j + 0].mAddress << 8;
+			return true;
+		}
+		if (mIns[j + 1].mMode == ASMIM_ZERO_PAGE && (mIns[j + 1].mAddress == reg || mIns[j + 1].mAddress == reg + 1) && mIns[j + 1].ChangesAddress())
+			return false;
+
+		j--;
+	}
+
+	if (mFromJump)
+		return mFromJump->FindPageStartAddress(mFromJump->mIns.Size(), reg, addr);
+	else
+		return false;
+}
 
 bool NativeCodeBasicBlock::FindGlobalAddress(int at, int reg, int& apos)
 {
@@ -8727,6 +8920,9 @@ bool NativeCodeBasicBlock::MoveLoadAddImmStoreUp(int at)
 	{
 		if (mIns[j].mType == ASMIT_STA && mIns[j].mMode == ASMIM_ZERO_PAGE && mIns[j].mAddress == mIns[at + 1].mAddress)
 		{
+			if (mIns[j].mLive & LIVE_CPU_REG_A)
+				return false;
+
 			mIns.Insert(j + 1, mIns[at + 3]);	// STA
 			mIns.Insert(j + 1, mIns[at + 3]);	// ADC
 			mIns.Insert(j + 1, mIns[at + 2]);	// CLC
@@ -9964,6 +10160,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 		}
 #endif
 
+
 #if 1
 		// move load - store (),y up to initial store
 		// 
@@ -9992,6 +10189,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 		}
 #endif
 
+
 #if 1
 		// move load - add # - store up to initial store
 		// 
@@ -10010,6 +10208,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 		}
 #endif
 
+
 #if 1
 		// move load - add ZP - store up to initial store
 		// 
@@ -10168,6 +10367,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 			}
 		}
 #endif
+
 		bool	progress = false;
 		do {
 			progress = false;
@@ -10274,6 +10474,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 					progress = true;
 				}
 #endif
+
 #if 1
 				if (mIns[i + 0].mMode == ASMIM_INDIRECT_Y && (mIns[i + 0].mFlags & NCIF_YZERO))
 				{
@@ -10752,6 +10953,31 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 						}
 					}
 #endif
+#if 1
+					if (mIns[i + 0].mType == ASMIT_LDY && mIns[i + 0].mMode == ASMIM_IMMEDIATE &&
+						mIns[i + 1].mMode == ASMIM_INDIRECT_Y)
+					{
+						int	addr;
+						if (FindPageStartAddress(i, mIns[i + 1].mAddress, addr))
+						{
+							if (mIns[i + 1].mLive & LIVE_CPU_REG_Y)
+								mIns.Insert(i + 2, mIns[i + 0]);
+
+							int	absaddr = addr + mIns[i + 0].mAddress;
+
+							mIns[i + 0].mMode = ASMIM_ZERO_PAGE;
+							mIns[i + 0].mAddress = mIns[i + 1].mAddress;
+
+							mIns[i + 1].mMode = ASMIM_ABSOLUTE_Y;
+							mIns[i + 1].mAddress = absaddr;
+							mIns[i + 1].mLinkerObject = nullptr;
+
+							progress = true;
+						}
+					}
+#endif
+
+
 #if 1
 					if (
 						mIns[i + 0].mMode == ASMIM_INDIRECT_Y && (mIns[i + 0].mFlags & NCIF_YZERO) &&
@@ -11452,6 +11678,21 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 						progress = true;
 					}
 #endif
+					else if (
+						mIns[i + 0].mType == ASMIT_LDA && mIns[i + 0].mMode == ASMIM_ZERO_PAGE &&
+						mIns[i + 1].IsShift() && mIns[i + 1].mMode == ASMIM_IMPLIED &&
+						mIns[i + 2].mType == ASMIT_STA && mIns[i + 2].mMode == ASMIM_ZERO_PAGE && mIns[i + 2].mAddress == mIns[i + 0].mAddress &&
+						mIns[i + 3].mType == ASMIT_ORA && mIns[i + 3].mMode == ASMIM_IMMEDIATE && mIns[i + 3].mAddress == 0x00 && !(mIns[i + 3].mLive & LIVE_CPU_REG_A))
+					{
+						mIns[i + 0].mType = mIns[i + 1].mType;
+						mIns[i + 0].mLive |= LIVE_MEM | LIVE_CPU_REG_C | LIVE_CPU_REG_Z;
+
+						mIns[i + 1].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
+						mIns[i + 2].mType = ASMIT_NOP; mIns[i + 2].mMode = ASMIM_IMPLIED;
+						mIns[i + 3].mType = ASMIT_NOP; mIns[i + 3].mMode = ASMIM_IMPLIED;
+
+						progress = true;
+					}
 
 #if 1
 					if (
@@ -11560,6 +11801,38 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 				}
 #endif
 
+#if 1
+				if (i + 7 < mIns.Size())
+				{
+					if (
+						mIns[i + 0].mType == ASMIT_CLC &&
+						mIns[i + 1].mType == ASMIT_ADC && mIns[i + 1].mMode == ASMIM_ZERO_PAGE &&
+						mIns[i + 2].mType == ASMIT_STA && mIns[i + 2].mMode == ASMIM_ZERO_PAGE &&
+						mIns[i + 3].mType == ASMIT_LDA && mIns[i + 3].mMode == ASMIM_ZERO_PAGE && mIns[i + 3].mAddress == mIns[i + 1].mAddress + 1 &&
+						mIns[i + 4].mType == ASMIT_ADC && mIns[i + 4].mMode == ASMIM_IMMEDIATE && mIns[i + 4].mAddress == 0 &&
+						mIns[i + 5].mType == ASMIT_STA && mIns[i + 5].mMode == ASMIM_ZERO_PAGE && mIns[i + 5].mAddress == mIns[i + 2].mAddress + 1 &&
+
+						mIns[i + 6].mType == ASMIT_LDY && mIns[i + 6].mMode == ASMIM_IMMEDIATE && mIns[i + 6].mAddress == 0 &&
+						mIns[i + 7].mMode == ASMIM_INDIRECT_Y && mIns[i + 7].mAddress == mIns[i + 2].mAddress && !(mIns[i + 7].mLive & LIVE_MEM))
+					{
+						for (int j = 0; j < 6; j++)
+						{
+							mIns[i + j].mType = ASMIT_NOP; mIns[i + j].mMode = ASMIM_IMPLIED;
+						}
+
+						if (mIns[i + 7].mLive & LIVE_CPU_REG_Y)
+						{
+							mIns.Insert(i + 8, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+							mIns[i + 8].mLive |= LIVE_CPU_REG_Y;
+						}
+
+						mIns[i + 6].mType = ASMIT_TAY;
+						mIns[i + 6].mMode = ASMIM_IMPLIED;
+						mIns[i + 7].mAddress = mIns[i + 1].mAddress;
+						progress = true;
+					}
+				}
+#endif					
 #if 1
 				if (i + 9 < mIns.Size())
 				{
@@ -12051,8 +12324,6 @@ NativeCodeProcedure::~NativeCodeProcedure(void)
 
 void NativeCodeProcedure::CompressTemporaries(void)
 {
-//	return;
-
 	if (mInterProc->mTempSize > 16)
 	{
 		ResetVisited();
@@ -12516,12 +12787,13 @@ void NativeCodeProcedure::Optimize(void)
 			mBlocks[i]->mEntryBlocks.SetSize(0);
 		mEntryBlock->CollectEntryBlocks(nullptr);
 
+#if 1
 		if (step == 2)
 		{
 			if (MapFastParamsToTemps())
 				changed = true;
 		}
-
+#endif
 		if (step > 2)
 		{
 			ResetVisited();
@@ -12919,19 +13191,25 @@ void NativeCodeProcedure::CompileInterBlock(InterCodeProcedure* iproc, InterCode
 				{
 					block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, ins->mSrc[0].mIntConst & 0xff));
 					block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 0));
-					block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, (ins->mSrc[0].mIntConst >> 8) & 0xff));
-					block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
 
-					mExitBlock->mIns[0].mFlags |= NCIF_LOWER | NCIF_UPPER;
+					mExitBlock->mIns[0].mFlags |= NCIF_LOWER;
 
-					if (ins->mSrc[0].mType == IT_INT32)
+					if (InterTypeSize[ins->mSrc[0].mType] > 1)
 					{
-						block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, (ins->mSrc[0].mIntConst >> 16) & 0xff));
+						block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, (ins->mSrc[0].mIntConst >> 8) & 0xff));
-						block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
+						block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
-						block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, (ins->mSrc[0].mIntConst >> 24) & 0xff));
-						block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
 
-						mExitBlock->mIns[0].mFlags |= NCIF_LONG;
+						mExitBlock->mIns[0].mFlags |= NCIF_UPPER;
+
+						if (InterTypeSize[ins->mSrc[0].mType] > 2)
+						{
+							block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, (ins->mSrc[0].mIntConst >> 16) & 0xff));
+							block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
+							block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, (ins->mSrc[0].mIntConst >> 24) & 0xff));
+							block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
+
+							mExitBlock->mIns[0].mFlags |= NCIF_LONG;
+						}
 					}
 				}
 			}				
@@ -12939,19 +13217,25 @@ void NativeCodeProcedure::CompileInterBlock(InterCodeProcedure* iproc, InterCode
 			{
 				block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp]));
 				block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU));
-				block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp] + 1));
-				block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
 
-				mExitBlock->mIns[0].mFlags |= NCIF_LOWER | NCIF_UPPER;
+				mExitBlock->mIns[0].mFlags |= NCIF_LOWER;
 
-				if (ins->mSrc[0].mType == IT_FLOAT || ins->mSrc[0].mType == IT_INT32)
+				if (InterTypeSize[ins->mSrc[0].mType] > 1)
 				{
-					block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp] + 2));
+					block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp] + 1));
-					block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
+					block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
-					block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp] + 3));
-					block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
 
-					mExitBlock->mIns[0].mFlags |= NCIF_LONG;
+					mExitBlock->mIns[0].mFlags |= NCIF_UPPER;
+
+					if (InterTypeSize[ins->mSrc[0].mType] > 2)
+					{
+						block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp] + 2));
+						block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
+						block->mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + iproc->mTempOffset[ins->mSrc[0].mTemp] + 3));
+						block->mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
+
+						mExitBlock->mIns[0].mFlags |= NCIF_LONG;
+					}
 				}
 			}
 

oscar64/NativeCodeGenerator.h

@@ -193,6 +193,7 @@ public:
 	bool FindAddressSumY(int at, int reg, int & apos, int& breg, int& ireg);
 	bool FindGlobalAddress(int at, int reg, int& apos);
 	bool FindGlobalAddressSumY(int at, int reg, bool direct, int& apos, const NativeCodeInstruction * & ains, const NativeCodeInstruction*& iins, uint32 & flags);
+	bool FindPageStartAddress(int at, int reg, int& addr);
 	bool MoveStoreXUp(int at);
 	bool MoveStoreHighByteDown(int at);
 	bool MoveAddHighByteDown(int at);

oscar64/Parser.cpp

@@ -563,6 +563,18 @@ Declaration * Parser::CopyConstantInitializer(int offset, Declaration* dtype, Ex
 			dec->mOffset = offset;
 		}
 	}
+	else if (dtype->mType == DT_TYPE_POINTER && dec->mType == DT_VARIABLE && dec->mBase->mType == DT_TYPE_ARRAY && (dec->mFlags & DTF_STATIC))
+	{
+		if (dtype->CanAssign(exp->mDecType))
+		{
+			Declaration	*	ndec = new Declaration(dec->mLocation, DT_CONST_POINTER);
+			ndec->mValue = exp;
+			ndec->mBase = dtype;
+			dec = ndec;
+		}
+		else
+			mErrors->Error(exp->mLocation, EERR_CONSTANT_INITIALIZER, "Incompatible constant initializer");
+	}
 	else
 		mErrors->Error(exp->mLocation, EERR_CONSTANT_INITIALIZER, "Constant initializer expected");
 
@@ -3106,6 +3118,36 @@ void Parser::ParsePragma(void)
 
 			ConsumeToken(TK_CLOSE_PARENTHESIS);
 		}
+		else if (!strcmp(mScanner->mTokenIdent->mString, "align"))
+		{
+			mScanner->NextToken();
+			ConsumeToken(TK_OPEN_PARENTHESIS);
+
+			if (mScanner->mToken == TK_IDENT)
+			{
+				Declaration* dec = mGlobals->Lookup(mScanner->mTokenIdent);
+				if (dec && dec->mType == DT_VARIABLE && (dec->mFlags & DTF_STATIC))
+				{
+					mScanner->NextToken();
+					ConsumeToken(TK_COMMA);
+
+					Expression * exp = ParseRExpression();
+					if (exp->mType == EX_CONSTANT && exp->mDecValue->mType == DT_CONST_INTEGER)
+						dec->mAlignment = exp->mDecValue->mInteger;
+					else
+						mErrors->Error(mScanner->mLocation, EERR_PRAGMA_PARAMETER, "Integer number for alignment expected");
+				}
+				else
+				{
+					mErrors->Error(mScanner->mLocation, EERR_OBJECT_NOT_FOUND, "Variable not found");
+					mScanner->NextToken();
+				}
+			}
+			else
+				mErrors->Error(mScanner->mLocation, EERR_PRAGMA_PARAMETER, "Variable name expected");
+
+			ConsumeToken(TK_CLOSE_PARENTHESIS);
+		}
 		else
 		{
 			mScanner->NextToken();