Add __forceinline storage specifier

oscar64/ByteCodeGenerator.cpp

@@ -6548,7 +6548,7 @@ void ByteCodeProcedure::Compile(ByteCodeGenerator* generator, InterCodeProcedure
 {
 	mID = proc->mID;
 
-	mNumBlocks = proc->mBlocks.Size();
+	mNumBlocks = proc->mNumBlocks;
 
 	tblocks = new ByteCodeBasicBlock * [mNumBlocks];
 	for (int i = 0; i < mNumBlocks; i++)

oscar64/Declaration.h

@@ -104,6 +104,7 @@ static const uint64	DTF_CONSTEXPR		= (1ULL << 31);
 static const uint64 DTF_AUTO_TEMPLATE	= (1ULL << 32);
 static const uint64 DTF_BANK_INLAY		= (1ULL << 33);
 static const uint64 DTF_PURE_VIRTUAL	= (1ULL << 34);
+static const uint64 DTF_FORCE_INLINE	= (1ULL << 35);
 
 static const uint64 DTF_FUNC_VARIABLE	= (1ULL << 36);
 static const uint64 DTF_FUNC_ASSEMBLER	= (1ULL << 37);
@@ -127,6 +128,7 @@ static const uint64 DTF_DEPRECATED		= (1ULL << 50);
 static const uint64 DTF_FUNC_NO_RETURN	= (1ULL << 51);
 
 
+
 class Declaration;
 
 enum ScopeLevel

oscar64/Errors.h

@@ -43,6 +43,7 @@ enum ErrorID
 	EWARN_INVALID_VALUE_RANGE,
 	EWARN_DEFAULT_COPY_DEPRECATED,
 	EWARN_INSUFFICIENT_MEMORY,
+	EWARN_FUNCTION_NOT_INLINED,
 
 	EERR_GENERIC = 3000,
 	EERR_FILE_NOT_FOUND,

oscar64/GlobalAnalyzer.cpp

@@ -161,7 +161,7 @@ int GlobalAnalyzer::CallerInvokes(Declaration* called)
 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))
+	if (caller->mType == DT_CONST_FUNCTION && (caller->mFlags & (DTF_INLINE | DTF_FORCE_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;
 }
@@ -206,7 +206,7 @@ void GlobalAnalyzer::AutoInline(void)
 //				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))
+				if ((f->mCompilerOptions & COPT_OPTIMIZE_INLINE) && (f->mFlags & DTF_REQUEST_INLINE) || (f->mFlags & DTF_FORCE_INLINE))
 					doinline = true;
 				if (f->mLocalSize < 100)
 				{

oscar64/InterCode.cpp

@@ -5844,7 +5844,7 @@ InterCodeBasicBlock::InterCodeBasicBlock(InterCodeProcedure * proc)
 	mUnreachable = false;
 	mValueRangeValid = false;
 
-	mIndex = proc->mBlocks.Size();
+	mIndex = proc->mNumBlocks++;
 	proc->mBlocks.Push(this);
 }
 
@@ -7441,6 +7441,8 @@ bool InterCodeBasicBlock::EarlyBranchElimination(const GrowingInstructionPtrArra
 			if (cins->mConst.mType == IT_BOOL)
 			{
 				mInstructions[sz - 1]->mCode = IC_JUMP;
+				mInstructions[sz - 1]->mNumOperands = 0;
+
 				if (cins->mConst.mIntConst)
 					mFalseJump->mNumEntries--;
 				else
@@ -21668,7 +21670,7 @@ void InterCodeBasicBlock::WarnUnreachable(void)
 
 
 InterCodeProcedure::InterCodeProcedure(InterCodeModule * mod, const Location & location, const Ident* ident, LinkerObject * linkerObject)
-	: mTemporaries(IT_NONE), mBlocks(nullptr), mLocation(location), mTempOffset(-1), mTempSizes(0), 
+	: mTemporaries(IT_NONE), mBlocks(nullptr), mLocation(location), mTempOffset(-1), mTempSizes(0), mNumBlocks(0),
 	mRenameTable(-1), mRenameUnionTable(-1), mGlobalRenameTable(-1),
 	mValueForwardingTable(nullptr), mLocalVars(nullptr), mParamVars(nullptr), mModule(mod),
 	mIdent(ident), mLinkerObject(linkerObject),
@@ -21809,7 +21811,10 @@ void InterCodeProcedure::EarlyBranchElimination(void)
 
 	ResetVisited();
 	while (mEntryBlock->EarlyBranchElimination(ctemps))
-	{}
+	{
+		BuildTraces(true);
+		TrimBlocks();
+	}
 }
 
 void InterCodeProcedure::BuildTraces(bool expand, bool dominators, bool compact)
@@ -21847,6 +21852,24 @@ void InterCodeProcedure::BuildTraces(bool expand, bool dominators, bool compact)
 		mEntryBlock->BuildDominatorTree(nullptr);
 }
 
+void InterCodeProcedure::TrimBlocks(void)
+{
+	int j = 1;
+	for (int i = 1; i < mBlocks.Size(); i++)
+	{
+		InterCodeBasicBlock* block = mBlocks[i];
+		if (block->mNumEntries > 0)
+		{
+			block->mIndex = j;
+			mBlocks[j++] = block;
+		}
+		else
+			delete block;
+	}
+	mBlocks.SetSize(j, false);
+	mNumBlocks = j;
+}
+
 void InterCodeProcedure::BuildDataFlowSets(void)
 {
 	int	numTemps = mTemporaries.Size();

oscar64/InterCode.h

@@ -690,6 +690,7 @@ protected:
 	void ResetEntryBlocks(void);
 public:
 	InterCodeBasicBlock				*	mEntryBlock;
+	int									mNumBlocks;
 	GrowingInterCodeBasicBlockPtrArray	mBlocks;
 	GrowingTypeArray					mTemporaries;
 	GrowingIntArray						mTempOffset, mTempSizes;
@@ -748,6 +749,7 @@ public:
 protected:
 	void BuildLocalAliasTable(void);
 	void BuildTraces(bool expand, bool dominators = true, bool compact = false);
+	void TrimBlocks(void);
 	void EarlyBranchElimination(void);
 	void BuildDataFlowSets(void);
 	void RenameTemporaries(void);

oscar64/InterCodeGenerator.cpp

@@ -1797,7 +1797,7 @@ void InterCodeGenerator::CopyStruct(InterCodeProcedure* proc, Expression* exp, I
 		if (!ccdec->mLinkerObject)
 			this->TranslateProcedure(proc->mModule, ccdec->mValue, ccdec);
 
-		bool	canInline = (mCompilerOptions & COPT_OPTIMIZE_INLINE) && !(inlineMapper && inlineMapper->mDepth > 10);
+		bool	canInline = ((mCompilerOptions & COPT_OPTIMIZE_INLINE) || (ccdec->mFlags & DTF_FORCE_INLINE)) && !(inlineMapper && inlineMapper->mDepth > 10);
 		bool	doInline = false;
 
 		if (canInline)
@@ -1812,6 +1812,9 @@ void InterCodeGenerator::CopyStruct(InterCodeProcedure* proc, Expression* exp, I
 			}
 		}
 
+		if ((ccdec->mFlags & DTF_FORCE_INLINE) && !doInline)
+			mErrors->Error(exp->mLocation, EWARN_FUNCTION_NOT_INLINED, "Function with __forceinline not inlined", ccdec->mQualIdent);
+
 		if (doInline)
 		{
 			DestructStack* destack = nullptr;
@@ -3781,7 +3784,7 @@ InterCodeGenerator::ExValue InterCodeGenerator::TranslateExpression(Declaration*
 
 			bool	canInline = exp->mLeft->mType == EX_CONSTANT && 
 								exp->mLeft->mDecValue->mType == DT_CONST_FUNCTION && 
-								(mCompilerOptions & COPT_OPTIMIZE_INLINE) &&
+								((mCompilerOptions & COPT_OPTIMIZE_INLINE) || (exp->mLeft->mDecValue->mFlags & DTF_FORCE_INLINE)) &&
 								!(inlineMapper && inlineMapper->mDepth > 10) &&
 								exp->mType != EX_VCALL;
 			bool	doInline = false, inlineConstexpr = false;
@@ -3831,7 +3834,10 @@ InterCodeGenerator::ExValue InterCodeGenerator::TranslateExpression(Declaration*
 			}
 			else
 			{
-				Expression	*	funcexp = exp->mLeft;
+				Expression* funcexp = exp->mLeft;
+
+				if (funcexp->mType == EX_CONSTANT && (funcexp->mDecValue->mFlags & DTF_FORCE_INLINE))
+					mErrors->Error(exp->mLocation, EWARN_FUNCTION_NOT_INLINED, "Function with __forceinline not inlined", funcexp->mDecValue->mQualIdent);
 
 				vl = TranslateExpression(procType, proc, block, funcexp, destack, gotos, breakBlock, continueBlock, inlineMapper);
 

oscar64/Parser.cpp

@@ -4699,6 +4699,11 @@ Declaration* Parser::ParseDeclaration(Declaration * pdec, bool variable, bool ex
 					storageFlags |= DTF_PREVENT_INLINE;
 					mScanner->NextToken();
 				}
+				else if (mScanner->mToken == TK_FORCEINLINE)
+				{
+					storageFlags |= DTF_FORCE_INLINE | DTF_REQUEST_INLINE;
+					mScanner->NextToken();
+				}
 				else if (mScanner->mToken == TK_INLINE)
 				{
 					storageFlags |= DTF_REQUEST_INLINE;
@@ -5341,7 +5346,7 @@ Declaration* Parser::ParseDeclaration(Declaration * pdec, bool variable, bool ex
 									}
 								}
 
-								pdec->mFlags |= ndec->mFlags & (DTF_REQUEST_INLINE | DTF_PREVENT_INLINE);
+								pdec->mFlags |= ndec->mFlags & (DTF_REQUEST_INLINE | DTF_PREVENT_INLINE | DTF_FORCE_INLINE);
 
 								ndec = pdec;
 							}
@@ -5436,7 +5441,7 @@ Declaration* Parser::ParseDeclaration(Declaration * pdec, bool variable, bool ex
 			}
 			else if (pthis)
 			{
-				ndec->mFlags |= storageFlags & (DTF_STATIC | DTF_PREVENT_INLINE);
+				ndec->mFlags |= storageFlags & (DTF_STATIC | DTF_PREVENT_INLINE | DTF_FORCE_INLINE);
 			}
 
 			ndec->mOffset = 0;
@@ -11372,6 +11377,7 @@ void Parser::ParseTemplateDeclarationBody(Declaration * tdec, Declaration * pthi
 
 		ConsumeTokenIf(TK_INLINE);
 		ConsumeTokenIf(TK_NOINLINE);
+		ConsumeTokenIf(TK_FORCEINLINE);
 		ConsumeTokenIf(TK_CONSTEXPR);
 
 		Declaration* bdec = ParseBaseTypeDeclaration(0, true);

oscar64/Scanner.cpp

@@ -60,6 +60,7 @@ const char* TokenNames[] =
 	"__export",
 	"__zeropage",
 	"__noinline",
+	"__forceinline",
 	"__striped",
 	"__dynstack",
 
@@ -1798,6 +1799,8 @@ void Scanner::NextRawToken(void)
 					mToken = TK_ZEROPAGE;
 				else if (!strcmp(tkident, "__noinline"))
 					mToken = TK_NOINLINE;
+				else if (!strcmp(tkident, "__forceinline"))
+					mToken = TK_FORCEINLINE;
 				else if (!strcmp(tkident, "__striped"))
 					mToken = TK_STRIPED;
 				else if (!strcmp(tkident, "__dynstack"))

oscar64/Scanner.h

@@ -58,6 +58,7 @@ enum Token
 	TK_EXPORT,
 	TK_ZEROPAGE,
 	TK_NOINLINE,
+	TK_FORCEINLINE,
 	TK_STRIPED,
 	TK_DYNSTACK,