Add optional static memory profile output file

README.md

@@ -124,6 +124,7 @@ The compiler is command line driven, and creates an executable .prg file.
 * -Oz : enable auto placement of global variables in zero page (part of O3)
 * -Op : optimize constant parameters
 * -g : create source level debug info and add source line numbers to asm listing
+* -gp : create source level debug info and add source line numbers to asm listing and static profile data
 * -tf : target format, may be prg, crt or bin
 * -tm : target machine
 * -d64 : create a d64 disk image

oscar64/Array.h

@@ -424,6 +424,28 @@ protected:
 		size = to;
 	}
 
+	template<typename F>
+	void Parition(const F & f, int l, int r)
+	{
+		if (r > l + 1)
+		{
+			int pi = l;
+			T	p(array[pi]);
+
+			for (int i = l + 1; i < r; i++)
+			{
+				if (f(array[i], p))
+				{
+					array[pi++] = array[i];
+					array[i] = array[pi];
+				}
+			}
+			array[pi] = p;
+
+			Parition(f, l, pi);
+			Parition(f, pi + 1, r);
+		}
+	}
 public:
 	ExpandingArray(void)
 	{
@@ -596,6 +618,12 @@ public:
 		return false;
 	}
 
+	template<typename F>
+	void Sort(const F & f)
+	{
+		Parition(f, 0, size);
+	}
+
 	__forceinline T& operator[](int n)
 	{
 		assert(n >= 0 && n < size);

oscar64/Compiler.cpp

@@ -1270,7 +1270,7 @@ bool Compiler::WriteErrorFile(const char* targetPath)
 
 bool Compiler::WriteOutputFile(const char* targetPath, DiskImage * d64)
 {
-	char	prgPath[200], mapPath[200], asmPath[200], lblPath[200], intPath[200], bcsPath[200], dbjPath[200];
+	char	prgPath[200], mapPath[200], asmPath[200], lblPath[200], intPath[200], bcsPath[200], dbjPath[200], cszPath[200];
 	char	basePath[200];
 
 	strcpy_s(basePath, targetPath);
@@ -1293,6 +1293,7 @@ bool Compiler::WriteOutputFile(const char* targetPath, DiskImage * d64)
 	strcpy_s(intPath, prgPath);
 	strcpy_s(bcsPath, prgPath);
 	strcpy_s(dbjPath, prgPath);
+	strcpy_s(cszPath, prgPath);
 
 	strcat_s(mapPath, "map");
 	strcat_s(asmPath, "asm");
@@ -1300,6 +1301,7 @@ bool Compiler::WriteOutputFile(const char* targetPath, DiskImage * d64)
 	strcat_s(intPath, "int");
 	strcat_s(bcsPath, "bcs");
 	strcat_s(dbjPath, "dbj");
+	strcat_s(cszPath, "csz");
 
 	if (mCompilerOptions & COPT_TARGET_PRG)
 	{
@@ -1382,6 +1384,9 @@ bool Compiler::WriteOutputFile(const char* targetPath, DiskImage * d64)
 	if (mCompilerOptions & COPT_DEBUGINFO)
 		WriteDbjFile(dbjPath);
 
+	if (mCompilerOptions & COPT_PROFILEINFO)
+		WriteCszFile(cszPath);
+
 	if (!(mCompilerOptions & COPT_NATIVE))
 	{
 		if (mCompilerOptions & COPT_VERBOSE)
@@ -1448,6 +1453,111 @@ static void DumpReferences(FILE* file, Declaration* dec)
 
 }
 
+bool Compiler::WriteCszFile(const char* filename)
+{
+	FILE* file;
+	fopen_s(&file, filename, "wb");
+	if (file)
+	{
+		for (int i = 0; i < mInterCodeModule->mProcedures.Size(); i++)
+		{
+			InterCodeProcedure* p(mInterCodeModule->mProcedures[i]);
+			if (p->mLinkerObject && p->mIdent && p->mDeclaration)
+			{
+				LinkerObject* lo = p->mLinkerObject;
+
+				struct SourceCount
+				{
+					const char* mFileName;
+					int			mLine, mAddress;
+					int			mBytes;
+				};
+
+				ExpandingArray<SourceCount>	ea;
+
+				for (int j = 0; j < lo->mCodeLocations.Size(); j++)
+				{
+					const CodeLocation& co(lo->mCodeLocations[j]);
+					const Location* ls = &(co.mLocation);
+					while (ls->mFrom)
+						ls = ls->mFrom;
+
+					int k = 0;
+					while (k < ea.Size() && (ea[k].mFileName != ls->mFileName || ea[k].mLine != ls->mLine))
+						k++;
+					if (k == ea.Size())
+					{
+						SourceCount	sc;
+						sc.mFileName = ls->mFileName;
+						sc.mLine = ls->mLine;
+						sc.mBytes = 0;
+						sc.mAddress = co.mStart + lo->mAddress;
+						ea.Push(sc);
+					}
+
+					ea[k].mBytes += co.mEnd - co.mStart;
+				}
+
+				if (ea.Size())
+				{
+					ea.Sort([](const SourceCount& l, const SourceCount& r)->bool {
+						return l.mFileName == r.mFileName ? l.mLine < r.mLine : l.mAddress < r.mAddress; 
+					});
+
+					FILE* fsrc;
+					fopen_s(&fsrc, ea[0].mFileName, "r");
+					if (fsrc)
+					{
+						fprintf(file, "<%s, %s>\n", p->mDeclaration->mQualIdent->mString, ea[0].mFileName);
+#if 0
+						for (int i = 0; i < ea.Size(); i++)
+						{
+							fprintf(file, "%s:%d : %d\n", ea[i].mFileName, ea[i].mLine, ea[i].mBytes);
+						}
+#endif
+						int		k = 0, l = 1;
+						char	line[1024];
+						while (k < ea.Size() && fgets(line, 1024, fsrc))
+						{
+							size_t ll = strlen(line);
+							while (ll > 0 && (line[ll - 1] == '\r' || line[ll - 1] == '\n'))
+								ll--;
+							line[ll] = 0;
+
+							if (l < ea[k].mLine)
+							{
+								if (k > 0)
+									fprintf(file, "%5d,---- (   ) : %s\n", l, line);
+							}
+							else
+							{
+								int ks = k, sum = 0;
+								do {
+									sum += ea[ks].mBytes;
+									ks++;
+								} while (ks < ea.Size() && ea[ks].mFileName != ea[0].mFileName);
+
+								fprintf(file, "%5d,%04x (%3d) : %s\n", l, ea[k].mAddress, ea[k].mBytes, line);
+
+								k = ks;
+							}
+							l++;
+						}	
+
+						fclose(fsrc);
+						fprintf(file, "\n");
+					}
+				}
+			}
+		}
+
+		fclose(file);
+		return true;
+	}
+	else
+		return false;
+}
+
 bool Compiler::WriteDbjFile(const char* filename)
 {
 	FILE* file;

oscar64/Compiler.h

@@ -61,4 +61,5 @@ public:
 	void CompleteTemplateExpansion(void);
 
 	bool WriteDbjFile(const char* filename);
+	bool WriteCszFile(const char* filename);
 };

oscar64/CompilerTypes.h

@@ -41,6 +41,8 @@ static const uint64 COPT_CPLUSPLUS = 1ULL << 52;
 static const uint64 COPT_PETSCII = 1ULL << 53;
 static const uint64 COPT_ERROR_FILES = 1ULL << 54;
 
+static const uint64 COPT_PROFILEINFO = 1ULL << 55;
+
 
 
 static const uint64 COPT_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS;

oscar64/InterCode.cpp

@@ -11472,9 +11472,10 @@ bool InterCodeBasicBlock::SimplifyIntegerNumeric(const GrowingInstructionPtrArra
 
 						if (pins->mSrc[0].mTemp < 0 && ins->mSrc[1].mIntConst + pins->mSrc[0].mIntConst >= 0)
 						{
+							ins->mSrc[1].mMemory = pins->mSrc[1].mMemory;
 							ins->mSrc[1].ForwardTemp(pins->mSrc[1]);
 							pins->mSrc[1].mFinal = false;
-							ins->mSrc[1].mIntConst += pins->mSrc[0].mIntConst;
+							ins->mSrc[1].mIntConst += pins->mSrc[0].mIntConst + pins->mSrc[1].mIntConst;
 							changed = true;
 						}
 #if 1
@@ -17036,6 +17037,7 @@ bool InterCodeBasicBlock::CheapInlining(int & numTemps)
 						{
 							int k = FindStore(this, i, nins->mSrc[0]);
 							InterInstruction* pins = mInstructions[k]->Clone();
+							pins->mLocation.mFrom = &(ins->mLocation);
 							mInstructions[k]->mRemove = true;
 
 							if (pins->mSrc[0].mTemp < 0)
@@ -17058,6 +17060,7 @@ bool InterCodeBasicBlock::CheapInlining(int & numTemps)
 						case IC_STORE:
 						{
 							InterInstruction* pins = nins->Clone();
+							pins->mLocation.mFrom = &(ins->mLocation);
 							if (pins->mSrc[0].mTemp >= 0)
 								pins->mSrc[0].mTemp = tmap[pins->mSrc[0].mTemp];
 							if (pins->mSrc[1].mTemp >= 0)
@@ -17069,6 +17072,7 @@ bool InterCodeBasicBlock::CheapInlining(int & numTemps)
 						case IC_CALL_NATIVE:
 						{
 							InterInstruction* pins = nins->Clone();
+							pins->mLocation.mFrom = &(ins->mLocation);
 							if (pins->mDst.mTemp >= 0)
 								pins->mDst.mTemp = ntemps;
 							mInstructions.Insert(i, pins);
@@ -17081,6 +17085,7 @@ bool InterCodeBasicBlock::CheapInlining(int & numTemps)
 							if (ins->mDst.mTemp >= 0)
 							{
 								InterInstruction* pins = nins->Clone();
+								pins->mLocation.mFrom = &(ins->mLocation);
 								if (pins->mSrc[0].mTemp < 0)
 								{
 									pins->mCode = IC_CONSTANT;
@@ -17103,6 +17108,7 @@ bool InterCodeBasicBlock::CheapInlining(int & numTemps)
 						default:
 						{
 							InterInstruction* pins = nins->Clone();
+							pins->mLocation.mFrom = &(ins->mLocation);
 							for (int k = 0; k < pins->mNumOperands; k++)
 								if (pins->mSrc[k].mTemp >= 0)
 									pins->mSrc[k].mTemp = tmap[pins->mSrc[k].mTemp];
@@ -22737,7 +22743,10 @@ void InterCodeProcedure::SimplifyIntegerNumeric(FastNumberSet& activeSet)
 			mEntryBlock->RemapActiveTemporaries(activeSet);
 		}
 
+		DisassembleDebug("PreSimplifyIntegerNumeric");
+
 		ResetVisited();
+
 	} while (mEntryBlock->SimplifyIntegerNumeric(silvalues, silvused));
 
 	assert(silvused == mTemporaries.Size());
@@ -22904,7 +22913,7 @@ void InterCodeProcedure::Close(void)
 {
 	GrowingTypeArray	tstack(IT_NONE);
 
-	CheckFunc = !strcmp(mIdent->mString, "room_prepare_castle");
+	CheckFunc = !strcmp(mIdent->mString, "room_show");
 	CheckCase = false;
 
 	mEntryBlock = mBlocks[0];

oscar64/InterCodeGenerator.cpp

@@ -4093,7 +4093,7 @@ InterCodeGenerator::ExValue InterCodeGenerator::TranslateExpression(Declaration*
 						}
 
 
-						InterInstruction* wins = new InterInstruction(texp->mLocation, IC_STORE);
+						InterInstruction* wins = new InterInstruction(MapLocation(exp, inlineMapper), IC_STORE);
 						wins->mSrc[1].mMemory = IM_INDIRECT;
 						wins->mSrc[0].mType = InterTypeOf(vr.mType);;
 						wins->mSrc[0].mTemp = vr.mTemp;

oscar64/NativeCodeGenerator.cpp

@@ -5131,9 +5131,8 @@ void NativeCodeBasicBlock::PutLocation(const Location& location, bool weak)
 {
 	int sz = mCodeLocations.Size();
 	if (sz > 0 && 
-		(weak ||
+		mCodeLocations[sz - 1].mLocation.mFileName == location.mFileName &&
-		 (mCodeLocations[sz - 1].mLocation.mFileName == location.mFileName &&
+		mCodeLocations[sz - 1].mLocation.mLine == location.mLine)
-		 mCodeLocations[sz - 1].mLocation.mLine == location.mLine)))
 	{
 		mCodeLocations[sz - 1].mEnd = this->mCode.Size();
 		if (mCodeLocations[sz - 1].mWeak)
@@ -19930,6 +19929,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP;
 					mIns[i + 1].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
@@ -19951,6 +19951,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -19973,6 +19974,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP; mIns[i + 0].mMode = ASMIM_IMPLIED;
 					mIns[i + 1].mType = ASMIT_NOP;
@@ -19990,6 +19992,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -20012,6 +20015,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP; mIns[i + 0].mMode = ASMIM_IMPLIED;
 					mIns[i + 1].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
@@ -20028,6 +20032,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 2].mIns;
 					break;
 				}
 
@@ -20051,6 +20056,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					fblock->mIns.Push(mIns[i + 1]);
 					fblock->mIns[0].mType = ASMIT_DEC;
@@ -20071,6 +20077,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -20092,6 +20099,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
 					mIns[i + 1].mType = ASMIT_NOP;
@@ -20114,6 +20122,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -20135,6 +20144,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
 					mIns[i + 1].mType = ASMIT_NOP;
@@ -20157,6 +20167,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -20177,6 +20188,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP; 
 					mIns[i + 1].mType = ASMIT_NOP;
@@ -20197,6 +20209,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -20218,6 +20231,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					mIns[i + 0].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
 					mIns[i + 1].mType = ASMIT_NOP;
@@ -20240,6 +20254,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BNE;
+					mBranchIns = mIns[i + 3].mIns;
 					break;
 				}
 
@@ -20266,6 +20281,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				rblock->mTrueJump = mTrueJump;
 				rblock->mFalseJump = mFalseJump;
 				rblock->mBranch = mBranch;
+				rblock->mBranchIns = mBranchIns;
 
 				for (int j = i + 6; j < mIns.Size(); j++)
 					rblock->mIns.Push(mIns[j]);
@@ -20275,6 +20291,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				mTrueJump = neblock;
 				mFalseJump = eblock;
 				mBranch = ASMIT_BNE;
+				mBranchIns = mIns[i + 0].mIns;
 
 				if (veq != 0)
 					eblock->mIns.Push(NativeCodeInstruction(mIns[i].mIns, ASMIT_LDA, ASMIM_IMMEDIATE, veq));
@@ -20301,6 +20318,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				rblock->mTrueJump = mTrueJump;
 				rblock->mFalseJump = mFalseJump;
 				rblock->mBranch = mBranch;
+				rblock->mBranchIns = mBranchIns;
 
 				const InterInstruction* iins = mIns[i].mIns;
 
@@ -20312,6 +20330,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				mTrueJump = iblock;
 				mFalseJump = rblock;
 				mBranch = ASMIT_BCS;
+				mBranchIns = iins;
 
 				iblock->mIns.Push(NativeCodeInstruction(iins, ASMIT_INY));
 
@@ -20336,6 +20355,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				rblock->mTrueJump = mTrueJump;
 				rblock->mFalseJump = mFalseJump;
 				rblock->mBranch = mBranch;
+				rblock->mBranchIns = mBranchIns;
 
 				for (int j = i + 4; j < mIns.Size(); j++)
 					rblock->mIns.Push(mIns[j]);
@@ -20345,6 +20365,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				mTrueJump = neblock;
 				mFalseJump = eblock;
 				mBranch = ASMIT_BNE;
+				mBranchIns = mIns[i + 0].mIns;
 
 				neblock->mIns.Push(NativeCodeInstruction(mIns[i].mIns, ASMIT_LDA, ASMIM_IMMEDIATE, 1));
 
@@ -20372,6 +20393,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				rblock->mTrueJump = mTrueJump;
 				rblock->mFalseJump = mFalseJump;
 				rblock->mBranch = mBranch;
+				rblock->mBranchIns = mBranchIns;
 
 				for (int j = i + 5; j < mIns.Size(); j++)
 					rblock->mIns.Push(mIns[j]);
@@ -20386,6 +20408,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				mTrueJump = neblock;
 				mFalseJump = eblock;
 				mBranch = ASMIT_BNE;
+				mBranchIns = mIns[i + 0].mIns;
 
 				if (veq != 0)
 					eblock->mIns.Push(NativeCodeInstruction(mIns[i].mIns, ASMIT_LDA, ASMIM_IMMEDIATE, veq));
@@ -20415,6 +20438,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				rblock->mTrueJump = mTrueJump;
 				rblock->mFalseJump = mFalseJump;
 				rblock->mBranch = mBranch;
+				rblock->mBranchIns = mBranchIns;
 
 				for (int j = i + 5; j < mIns.Size(); j++)
 					rblock->mIns.Push(mIns[j]);
@@ -20424,6 +20448,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				mTrueJump = neblock;
 				mFalseJump = eblock;
 				mBranch = ASMIT_BNE;
+				mBranchIns = mIns[i + 0].mIns;
 
 				if (veq != 0)
 					eblock->mIns.Push(NativeCodeInstruction(mIns[i].mIns, ASMIT_LDA, ASMIM_IMMEDIATE, veq));
@@ -20451,6 +20476,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				rblock->mTrueJump = mTrueJump;
 				rblock->mFalseJump = mFalseJump;
 				rblock->mBranch = mBranch;
+				rblock->mBranchIns = mBranchIns;
 
 				for (int j = i + 5; j < mIns.Size(); j++)
 					rblock->mIns.Push(mIns[j]);
@@ -20460,6 +20486,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 				mTrueJump = neblock;
 				mFalseJump = eblock;
 				mBranch = ASMIT_BNE;
+				mBranchIns = mIns[i + 0].mIns;
 
 				neblock->mIns.Push(NativeCodeInstruction(mIns[i].mIns, ASMIT_LDA, ASMIM_IMMEDIATE, 0xff));
 
@@ -20610,6 +20637,9 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
+
+					const InterInstruction* iins = mIns[0].mIns;
 
 					for (int j = i + 3; j < mIns.Size(); j++)
 						fblock->mIns.Push(mIns[j]);
@@ -20622,6 +20652,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCC;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_LDA && mIns[i + 0].mMode == ASMIM_IMMEDIATE && mIns[i + 0].mAddress == 0 &&
@@ -20638,6 +20669,9 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
+
+					const InterInstruction* iins = mIns[0].mIns;
 
 					for (int j = i + 3; j < mIns.Size(); j++)
 						fblock->mIns.Push(mIns[j]);
@@ -20650,6 +20684,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCC;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_LDA &&
@@ -20666,6 +20701,9 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
+
+					const InterInstruction* iins = mIns[0].mIns;
 
 					for (int j = i + 3; j < mIns.Size(); j++)
 						fblock->mIns.Push(mIns[j]);
@@ -20678,6 +20716,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCS;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_TXA &&
@@ -20693,6 +20732,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					const InterInstruction* iins(mIns[i].mIns);
 
@@ -20707,6 +20747,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCC;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_TYA &&
@@ -20722,6 +20763,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					const InterInstruction* iins(mIns[i].mIns);
 
@@ -20738,6 +20780,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCC;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_TXA &&
@@ -20753,6 +20796,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					const InterInstruction* iins(mIns[i].mIns);
 
@@ -20769,6 +20813,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCC;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_TYA &&
@@ -20784,6 +20829,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
 
 					const InterInstruction* iins(mIns[i].mIns);
 
@@ -20798,6 +20844,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCC;
+					mBranchIns = iins;
 					break;
 				}
 				else if (mIns[i + 0].mType == ASMIT_LDA &&
@@ -20814,6 +20861,9 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					fblock->mTrueJump = mTrueJump;
 					fblock->mFalseJump = mFalseJump;
 					fblock->mBranch = mBranch;
+					fblock->mBranchIns = mBranchIns;
+
+					const InterInstruction* iins(mIns[i].mIns);
 
 					for (int j = i + 3; j < mIns.Size(); j++)
 						fblock->mIns.Push(mIns[j]);
@@ -20826,6 +20876,7 @@ bool NativeCodeBasicBlock::ExpandADCToBranch(NativeCodeProcedure* proc)
 					mTrueJump = fblock;
 					mFalseJump = iblock;
 					mBranch = ASMIT_BCS;
+					mBranchIns = iins;
 					break;
 				}
 
@@ -51575,7 +51626,7 @@ void NativeCodeProcedure::Compile(InterCodeProcedure* proc)
 	mInterProc = proc;
 	mInterProc->mLinkerObject->mNativeProc = this;
 
-	CheckFunc = !strcmp(mInterProc->mIdent->mString, "room_prep_chest");
+	CheckFunc = !strcmp(mInterProc->mIdent->mString, "floor_prepare");
 
 	int	nblocks = proc->mBlocks.Size();
 	tblocks = new NativeCodeBasicBlock * [nblocks];
@@ -52126,7 +52177,7 @@ void NativeCodeProcedure::Compile(InterCodeProcedure* proc)
 
 void NativeCodeProcedure::Assemble(void)
 {
-	CheckFunc = !strcmp(mInterProc->mIdent->mString, "a");
+	CheckFunc = !strcmp(mInterProc->mIdent->mString, "floor_prepare");
 
 	if (mInterProc->mCompilerOptions & COPT_OPTIMIZE_MERGE_CALLS)
 	{
@@ -52189,7 +52240,7 @@ void NativeCodeProcedure::Assemble(void)
 		mInterProc->mLinkerObject->AddReference(rl);
 	}
 
-	if (mGenerator->mCompilerOptions & COPT_DEBUGINFO)
+	if (mGenerator->mCompilerOptions & (COPT_DEBUGINFO | COPT_PROFILEINFO))
 	{
 		if (mCodeLocations.Size() > 0)
 		{

oscar64/Parser.cpp

@@ -9917,6 +9917,7 @@ Expression* Parser::ParseStatement(void)
 						{
 							Declaration* iterVarDec = new Declaration(mScanner->mLocation, DT_VARIABLE);
 							iterVarDec->mBase = containerExp->mDecType->mBase->BuildPointer(mScanner->mLocation);
+							iterVarDec->mBase = containerExp->mDecType->mBase->BuildPointer(mScanner->mLocation);
 							iterVarDec->mVarIndex = mLocalIndex++;
 							iterVarDec->mSize = iterVarDec->mBase->mSize;
 							iterVarDec->mFlags |= DTF_DEFINED;
@@ -10158,6 +10159,15 @@ Expression* Parser::ParseStatement(void)
 								int	stride = (endValue - startValue) / numLoops;
 								int	remain = (endValue - startValue) - numIterations * numLoops;
 
+								// Switch from 3 to 4 to avoid cost of two remains
+								if (numLoops == 3 && remain == 2 && !((endValue - startValue) & 3))
+								{
+									numLoops = 4;
+									numIterations = (endValue - startValue) / numLoops;
+									stride = (endValue - startValue) / numLoops;
+									remain = (endValue - startValue) - numIterations * numLoops;
+								}
+
 								Expression* unrollBody = new Expression(mScanner->mLocation, EX_SEQUENCE);
 								unrollBody->mLeft = bodyExp;
 								Expression* bexp = unrollBody;

oscar64/oscar64.cpp

@@ -297,6 +297,10 @@ int main2(int argc, const char** argv)
 				{
 					compiler->mCompilerOptions |= COPT_DEBUGINFO;
 				}
+				else if (arg[1] == 'g' && arg[2] == 'p' && !arg[3])
+				{
+					compiler->mCompilerOptions |= COPT_DEBUGINFO | COPT_PROFILEINFO;
+				}
 				else if (arg[1] == 'v')
 				{
 					compiler->mCompilerOptions |= COPT_VERBOSE;