Reduce generated code size with -Os

include/crt.c

@@ -2858,6 +2858,11 @@ __asm freg
 
 __asm faddsub
 {
+fsub:
+		lda tmp + 3
+		eor #$80
+		sta tmp + 3
+fadd:
 		lda #$ff
 		cmp tmp + 4
 		beq INF
@@ -3008,7 +3013,7 @@ L2:		lsr
 __asm inp_binop_add_f32
 {
 		jsr	freg.split_exp
-		jsr faddsub
+		jsr faddsub.fadd
 		jmp	startup.exec
 }
 
@@ -3017,10 +3022,7 @@ __asm inp_binop_add_f32
 __asm inp_binop_sub_f32
 {
 		jsr	freg.split_exp
-		lda	tmp + 3
-		eor	#$80
-		sta	tmp + 3
-		jsr	faddsub
+		jsr	faddsub.fsub
 		jmp	startup.exec
 }
 
@@ -3983,11 +3985,38 @@ W3:		sta	accu + 2
 		jmp	freg.merge_aexp
 }
 
+__asm store32
+{
+		lda	accu + 0
+		sta $00, x
+		lda	accu + 1
+		sta $01, x
+		lda	accu + 2
+		sta $02, x
+		lda	accu + 3
+		sta $03, x
+		rts
+}
+
+__asm load32
+{
+		lda	$00, x
+		sta accu + 0
+		lda	$01, x
+		sta accu + 1
+		lda	$02, x
+		sta accu + 2
+		lda	$03, x
+		sta accu + 3
+		rts
+}
+
 #pragma runtime(fsplita, freg.split_aexp)
 #pragma runtime(fsplitt, freg.split_texp)
 #pragma runtime(fsplitx, freg.split_xexp)
 #pragma runtime(fmergea, freg.merge_aexp)
-#pragma runtime(faddsub, faddsub)
+#pragma runtime(fadd, faddsub.fadd)
+#pragma runtime(fsub, faddsub.fsub)
 #pragma runtime(fmul, fmul)
 #pragma runtime(fdiv, fdiv)
 #pragma runtime(fcmp, fcmp)
@@ -4001,6 +4030,8 @@ W3:		sta	accu + 2
 #pragma runtime(ffromlu, uint32_to_float)
 #pragma runtime(ftoli, f32_to_i32)
 #pragma runtime(ftolu, f32_to_u32)
+#pragma runtime(store32, store32)
+#pragma runtime(load32, load32)
 
 __asm inp_op_floor_f32
 {

oscar64/Compiler.cpp

@@ -943,7 +943,8 @@ bool Compiler::GenerateCode(void)
 		RegisterRuntime(loc, Ident::Unique("fsplitt"));
 		RegisterRuntime(loc, Ident::Unique("fsplitx"));
 		RegisterRuntime(loc, Ident::Unique("fsplita"));
-		RegisterRuntime(loc, Ident::Unique("faddsub"));
+		RegisterRuntime(loc, Ident::Unique("fadd"));
+		RegisterRuntime(loc, Ident::Unique("fsub"));
 		RegisterRuntime(loc, Ident::Unique("fmul"));
 		RegisterRuntime(loc, Ident::Unique("fdiv"));
 		RegisterRuntime(loc, Ident::Unique("mul16"));
@@ -971,6 +972,9 @@ bool Compiler::GenerateCode(void)
 		RegisterRuntime(loc, Ident::Unique("divu32"));
 		RegisterRuntime(loc, Ident::Unique("modu32"));
 
+		RegisterRuntime(loc, Ident::Unique("store32"));
+		RegisterRuntime(loc, Ident::Unique("load32"));
+
 		RegisterRuntime(loc, Ident::Unique("malloc"));
 		RegisterRuntime(loc, Ident::Unique("free"));
 		RegisterRuntime(loc, Ident::Unique("breakpoint"));

oscar64/CompilerTypes.h

@@ -47,7 +47,7 @@ static const uint64 COPT_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE |
 
 static const uint64 COPT_OPTIMIZE_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS;
 
-static const uint64 COPT_OPTIMIZE_SIZE = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_CODE_SIZE | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;
+static const uint64 COPT_OPTIMIZE_SIZE = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_CODE_SIZE | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;
 
 static const uint64 COPT_OPTIMIZE_SPEED = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_AUTO_UNROLL | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_ASSEMBLER | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;
 

oscar64/Errors.h

@@ -42,6 +42,7 @@ enum ErrorID
 	EWARN_UNDEFINED_POINTER_ARITHMETIC,
 	EWARN_INVALID_VALUE_RANGE,
 	EWARN_DEFAULT_COPY_DEPRECATED,
+	EWARN_INSUFFICIENT_MEMORY,
 
 	EERR_GENERIC = 3000,
 	EERR_FILE_NOT_FOUND,

oscar64/GlobalAnalyzer.cpp

@@ -147,6 +147,25 @@ void GlobalAnalyzer::TopoSort(Declaration* procDec)
 	}
 }
 
+int GlobalAnalyzer::CallerInvokes(Declaration* called)
+{
+	int n = 0;
+	for (int i = 0; i < called->mCallers.Size(); i++)
+	{
+		Declaration* f = called->mCallers[i];
+		n += CallerInvokes(f, called);
+	}
+	return n;
+}
+
+int GlobalAnalyzer::CallerInvokes(Declaration* caller, Declaration* called)
+{
+	int n = 1;
+	if (caller->mType == DT_CONST_FUNCTION && (caller->mFlags & (DTF_INLINE | DTF_REQUEST_INLINE)) && !(caller->mFlags & DTF_PREVENT_INLINE) && !(caller->mFlags & DTF_FUNC_RECURSIVE) && !(caller->mFlags & DTF_FUNC_VARIABLE) && !(caller->mFlags & DTF_EXPORT))
+		n = CallerInvokes(caller);
+	return n > 1 ? n : 1;
+}
+
 void GlobalAnalyzer::AutoInline(void)
 {
 	for (int i = 0; i < mFunctions.Size(); i++)
@@ -181,33 +200,39 @@ void GlobalAnalyzer::AutoInline(void)
 					dec = dec->mNext;
 				}
 
+				int invokes = CallerInvokes(f);
 				int	cost = (f->mComplexity - 20 * nparams - 10);
 
-//				printf("CHECK INLINING %s (%d) %d * (%d - 1)\n", f->mIdent->mString, f->mComplexity, cost, f->mCallers.Size());
+//				printf("CHECK INLINING %s (%d) %d * (%d - 1)\n", f->mIdent->mString, f->mComplexity, cost, invokes);
 
 				bool	doinline = false;
 				if ((f->mCompilerOptions & COPT_OPTIMIZE_INLINE) && (f->mFlags & DTF_REQUEST_INLINE))
 					doinline = true;
 				if (f->mLocalSize < 100)
 				{
-					if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE) && ((cost - 20) * (f->mCallers.Size() - 1) <= 20))
+					if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE) && ((cost - 20) * (invokes - 1) <= 20))
 					{
-						if (f->mCallers.Size() == 1 && f->mComplexity > 100)
+						if (f->mCompilerOptions & COPT_OPTIMIZE_CODE_SIZE)
+						{
+							if (invokes == 1 && f->mSection == f->mCallers[0]->mSection || cost < 0)
+								doinline = true;
+						}
+						else if (invokes == 1 && f->mComplexity > 100)
 						{
 //							printf("CHECK INLINING2 %s <- %s %d\n", f->mIdent->mString, f->mCallers[0]->mIdent->mString, f->mCallers[0]->mCalled.Size());
 							if (cost < 0 || f->mCallers[0]->mComplexity + cost  < 1000 || f->mCallers[0]->mCalled.Size() == 1)
 								doinline = true;
 						}
-						else
+						else 
 							doinline = true;
 					}
-					if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE_ALL) && (cost * (f->mCallers.Size() - 1) <= 10000))
+					if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE_ALL) && (cost * (invokes - 1) <= 10000))
 						doinline = true;
 				}
 
 				if (doinline)
 				{
-//					printf("INLINING %s %d * (%d - 1)\n", f->mIdent->mString, cost, f->mCallers.Size());
+//					printf("INLINING %s %d * (%d - 1)\n", f->mIdent->mString, cost, invokes);
 
 					f->mFlags |= DTF_INLINE;
 					for (int j = 0; j < f->mCallers.Size(); j++)

oscar64/GlobalAnalyzer.h

@@ -32,6 +32,8 @@ class GlobalAnalyzer
 	GrowingArray<Declaration*>		mGlobalVariables;
 
 	void AnalyzeInit(Declaration* mdec);
+	int CallerInvokes(Declaration* called);
+	int CallerInvokes(Declaration* caller, Declaration* called);
 
 	Declaration* Analyze(Expression* exp, Declaration* procDec, bool lhs, bool aliasing);
 

oscar64/InterCode.cpp

@@ -23479,7 +23479,7 @@ void InterCodeProcedure::Close(void)
 
 	DisassembleDebug("mapped variabled");
 
-	ReduceTemporaries();
+	ReduceTemporaries(true);
 
 	DisassembleDebug("Reduced Temporaries");
 
@@ -24224,12 +24224,20 @@ void InterCodeProcedure::HoistCommonConditionalPath(void)
 }
 
 
-void InterCodeProcedure::ReduceTemporaries(void)
+void InterCodeProcedure::ReduceTemporaries(bool final)
 {
 	NumberSet* collisionSet;
 	int i, j, numRenamedTemps;
 	int numTemps = mTemporaries.Size();
 
+	NumberSet	callerSaved(numTemps);
+
+	if (final)
+	{
+		ResetVisited();
+		mEntryBlock->BuildCallerSaveTempSet(callerSaved);
+	}
+
 	ResetVisited();
 	mEntryBlock->BuildLocalTempSets(numTemps);
 
@@ -24261,24 +24269,73 @@ void InterCodeProcedure::ReduceTemporaries(void)
 
 	NumberSet	usedTemps(numTemps);
 
-	for (i = 0; i < numTemps; i++)
+	if (final)
 	{
-		usedTemps.Clear();
-
-		for (j = 0; j < numTemps; j++)
+		for (int sz = 4; sz > 0; sz >>= 1)
 		{
-			if (mRenameTable[j] >= 0 && (collisionSet[i][j] || InterTypeSize[mTemporaries[j]] != InterTypeSize[mTemporaries[i]]))
+			for (i = 0; i < numTemps; i++)
 			{
-				usedTemps += mRenameTable[j];
+				if (InterTypeSize[mTemporaries[i]] == sz && !callerSaved[i])
+				{
+					usedTemps.Clear();
+
+					for (j = 0; j < numTemps; j++)
+					{
+						if (mRenameTable[j] >= 0 && collisionSet[i][j])
+							usedTemps += mRenameTable[j];
+					}
+
+					j = 0;
+					while (usedTemps[j])
+						j++;
+
+					mRenameTable[i] = j;
+					if (j >= numRenamedTemps) numRenamedTemps = j + 1;
+				}
+			}
+			for (i = 0; i < numTemps; i++)
+			{
+				if (InterTypeSize[mTemporaries[i]] == sz && callerSaved[i])
+				{
+					usedTemps.Clear();
+
+					for (j = 0; j < numTemps; j++)
+					{
+						if (mRenameTable[j] >= 0 && collisionSet[i][j])
+							usedTemps += mRenameTable[j];
+					}
+
+					j = 0;
+					while (usedTemps[j])
+						j++;
+
+					mRenameTable[i] = j;
+					if (j >= numRenamedTemps) numRenamedTemps = j + 1;
+				}
 			}
 		}
+	}
+	else
+	{
+		for (i = 0; i < numTemps; i++)
+		{
+			usedTemps.Clear();
 
-		j = 0;
-		while (usedTemps[j])
-			j++;
+			for (j = 0; j < numTemps; j++)
+			{
+				if (mRenameTable[j] >= 0 && (collisionSet[i][j] || InterTypeSize[mTemporaries[j]] != InterTypeSize[mTemporaries[i]]))
+				{
+					usedTemps += mRenameTable[j];
+				}
+			}
+
+			j = 0;
+			while (usedTemps[j])
+				j++;
 
-		mRenameTable[i] = j;
-		if (j >= numRenamedTemps) numRenamedTemps = j + 1;
+			mRenameTable[i] = j;
+			if (j >= numRenamedTemps) numRenamedTemps = j + 1;
+		}
 	}
 
 	mTemporaries.SetSize(numRenamedTemps, true);

oscar64/InterCode.h

@@ -739,7 +739,7 @@ friend class InterCodeBasicBlock;
 	bool ModifiesGlobal(int varindex);
 
 	void MapVariables(void);
-	void ReduceTemporaries(void);
+	void ReduceTemporaries(bool final = false);
 	void Disassemble(FILE* file);
 	void Disassemble(const char* name, bool dumpSets = false);
 protected:

oscar64/Linker.cpp

@@ -772,6 +772,8 @@ void LinkerRegion::PlaceStackSection(LinkerSection* stackSection, LinkerSection*
 
 void Linker::CopyObjects(bool inlays)
 {
+	bool	errors = false;
+
 	for (int i = 0; i < mObjects.Size(); i++)
 	{
 		LinkerObject* obj = mObjects[i];
@@ -809,7 +811,9 @@ void Linker::CopyObjects(bool inlays)
 			{
 				if (!obj->mRegion)
 				{
-					mErrors->Error(obj->mLocation, ERRR_INSUFFICIENT_MEMORY, "Could not place object", obj->mIdent);
+					mErrors->Error(obj->mLocation, errors ? EWARN_INSUFFICIENT_MEMORY : ERRR_INSUFFICIENT_MEMORY, "Could not place object", obj->mIdent);
+					if (mCompilerOptions & COPT_ERROR_FILES)
+						errors = true;
 
 					int avail = 0;
 					for (int i = 0; i < mRegions.Size(); i++)
@@ -1645,7 +1649,7 @@ bool Linker::WriteCrtFile(const char* filename, uint16 id)
 
 bool Linker::WriteMapFile(const char* filename)
 {
-	bool	banked = mCartridgeBankUsed[0];
+	bool	banked = mCartridgeBankUsed[0] || mCartridgeBankUsed[1];
 
 	FILE* file;
 	fopen_s(&file, filename, "wb");