Optimize parameter handling in loops

include/c64/joystick.c

@@ -5,7 +5,7 @@ bool	joyb[2];
 
 void joy_poll(char n)
 {
-	char	b = ((char *)0xdc00)[n];
+	char	b = ((volatile char *)0xdc00)[n];
 
 	if (!(b & 1))
 		joyy[n] = -1;

oscar64/InterCode.cpp

@@ -380,6 +380,8 @@ static bool CollidingMem(const InterOperand& op1, const InterOperand& op2, const
 		{
 			if (op2.mMemory == IM_GLOBAL)
 				return staticVars[op2.mVarIndex]->mAliased;
+			else if (op2.mMemory == IM_FPARAM)
+				return false;
 			else
 				return true;
 		}
@@ -387,6 +389,8 @@ static bool CollidingMem(const InterOperand& op1, const InterOperand& op2, const
 		{
 			if (op1.mMemory == IM_GLOBAL)
 				return staticVars[op1.mVarIndex]->mAliased;
+			else if (op1.mMemory == IM_FPARAM)
+				return false;
 			else
 				return true;
 		}
@@ -1204,6 +1208,11 @@ void TempForwardingTable::Intersect(const TempForwardingTable& table)
 	}
 }
 
+int TempForwardingTable::Size(void) const
+{
+	return mAssoc.Size();
+}
+
 void TempForwardingTable::SetSize(int size)
 {
 	int i;
@@ -2767,6 +2776,15 @@ bool InterInstruction::RemoveUnusedResultInstructions(InterInstruction* pre, Num
 
 		changed = true;
 	}
+	else if (mCode == IC_LOAD_TEMPORARY && mDst.mTemp == mSrc[0].mTemp)
+	{
+		mCode = IC_NONE;
+		mDst.mTemp = -1;
+		for (int i = 0; i < mNumOperands; i++)
+			mSrc[i].mTemp = -1;
+
+		changed = true;
+	}
 	else if (mDst.mTemp != -1)
 	{
 		if (!requiredTemps[mDst.mTemp] && mDst.mTemp >= 0)
@@ -2793,15 +2811,6 @@ bool InterInstruction::RemoveUnusedResultInstructions(InterInstruction* pre, Num
 		else
 			requiredTemps -= mDst.mTemp;
 	}
-	else if (mCode == IC_LOAD_TEMPORARY && mDst.mTemp == mSrc[0].mTemp)
-	{
-		mCode = IC_NONE;
-		mDst.mTemp = -1;
-		for (int i = 0; i < mNumOperands; i++)
-			mSrc[i].mTemp = -1;
-
-		changed = true;
-	}
 
 	for (int i = 0; i < mNumOperands; i++)
 	{
@@ -3710,7 +3719,7 @@ static bool IsInfiniteLoop(InterCodeBasicBlock* head, InterCodeBasicBlock* block
 	return false;
 }
 
-void InterCodeBasicBlock::GenerateTraces(bool expand)
+void InterCodeBasicBlock::GenerateTraces(bool expand, bool compact)
 {
 	int i;
 
@@ -3760,12 +3769,31 @@ void InterCodeBasicBlock::GenerateTraces(bool expand)
 						mFalseJump->mNumEntries++;
 				}
 			}
+#if 1
+			else if (compact && mTrueJump && !mFalseJump && mTrueJump->mInstructions.Size() == 1 && mTrueJump->mInstructions[0]->mCode == IC_BRANCH && mTrueJump->mFalseJump)
+			{
+				InterCodeBasicBlock* tj = mTrueJump;
+
+				int	ns = mInstructions.Size();
+
+				tj->mNumEntries--;
+				tj->mTrueJump->mNumEntries++;
+				tj->mFalseJump->mNumEntries++;
+				
+				mInstructions[ns - 1]->mCode = IC_BRANCH;
+				mInstructions[ns - 1]->mOperator = tj->mInstructions[0]->mOperator;
+				mInstructions[ns - 1]->mSrc[0].Forward(tj->mInstructions[0]->mSrc[0]);
+				
+				mTrueJump = tj->mTrueJump;
+				mFalseJump = tj->mFalseJump;
+			}
+#endif
 			else
 				break;
 		}
 
-		if (mTrueJump) mTrueJump->GenerateTraces(expand);
+		if (mTrueJump) mTrueJump->GenerateTraces(expand, compact);
-		if (mFalseJump) mFalseJump->GenerateTraces(expand);
+		if (mFalseJump) mFalseJump->GenerateTraces(expand, compact);
 
 		mInPath = false;
 	}
@@ -6353,6 +6381,17 @@ void InterCodeBasicBlock::PerformTempForwarding(TempForwardingTable& forwardingT
 
 		if (mLoopHead)
 		{
+			if (mNumEntries == 2 && (mTrueJump == this || mFalseJump == this) && mLocalModifiedTemps.Size())
+			{
+				assert(localForwardingTable.Size() == mLocalModifiedTemps.Size());
+
+				for (int i = 0; i < mLocalModifiedTemps.Size(); i++)
+				{
+					if (mLocalModifiedTemps[i])
+						localForwardingTable.Destroy(i);
+				}
+			}
+			else
 				localForwardingTable.Reset();
 		}
 #if 0
@@ -7550,10 +7589,46 @@ bool InterCodeBasicBlock::LoadStoreForwarding(const GrowingInstructionPtrArray&
 					int i = 0;
 					while (i < mLoadStoreInstructions.Size())
 					{
-						if (tvalue.IndexOf(mLoadStoreInstructions[i]) == -1)
+						InterInstruction* ins(mLoadStoreInstructions[i]);
-							mLoadStoreInstructions.Remove(i);
+						InterInstruction* nins = nullptr;
+
+						int j = tvalue.IndexOf(ins);
+						if (j != -1)
+							nins = ins;
 						else
-							i++;
+						{
+							if (ins->mCode == IC_LOAD)
+							{
+								j = 0;
+								while (j < tvalue.Size() && !SameMem(ins->mSrc[0], tvalue[j]))
+									j++;
+								if (j < tvalue.Size())
+								{
+									if (tvalue[j]->mCode == IC_LOAD && tvalue[j]->mDst.IsEqual(ins->mDst))
+										nins = ins;
+									else if (tvalue[j]->mCode == IC_STORE && tvalue[j]->mSrc[0].IsEqual(ins->mDst))
+										nins = ins;
+								}
+							}
+							else if (ins->mCode == IC_STORE)
+							{
+								j = 0;
+								while (j < tvalue.Size() && !SameMem(ins->mSrc[1], tvalue[j]))
+									j++;
+								if (j < tvalue.Size())
+								{
+									if (tvalue[j]->mCode == IC_LOAD && tvalue[j]->mDst.IsEqual(ins->mSrc[0]))
+										nins = tvalue[j];
+									else if (tvalue[j]->mCode == IC_STORE && tvalue[j]->mSrc[0].IsEqual(ins->mSrc[0]))
+										nins = ins;
+								}
+							}
+						}
+
+						if (nins)
+							mLoadStoreInstructions[i++] = nins;
+						else
+							mLoadStoreInstructions.Remove(i);
 					}
 				}
 
@@ -8008,7 +8083,9 @@ bool InterCodeBasicBlock::MergeCommonPathInstructions(void)
 			while (ti < mTrueJump->mInstructions.Size() && !changed)
 			{
 				InterInstruction* tins = mTrueJump->mInstructions[ti];
-				if (tins->mCode != IC_BRANCH && tins->mCode != IC_JUMP && tins->mCode != IC_RELATIONAL_OPERATOR)
+				InterInstruction* nins = (ti + 1 < mTrueJump->mInstructions.Size()) ? mTrueJump->mInstructions[ti + 1] : nullptr;
+
+				if (tins->mCode != IC_BRANCH && tins->mCode != IC_JUMP && !(nins && nins->mCode == IC_BRANCH && tins->mDst.mTemp == nins->mSrc[0].mTemp))
 				{
 					int	fi = 0;
 					while (fi < mFalseJump->mInstructions.Size() && !tins->IsEqualSource(mFalseJump->mInstructions[fi]))
@@ -8789,6 +8866,50 @@ void InterCodeBasicBlock::FollowJumps(void)
 	}
 }
 
+void InterCodeBasicBlock::BuildLoopSuffix(InterCodeProcedure* proc)
+{
+	if (!mVisited)
+	{
+		mVisited = true;
+
+		if (mLoopHead && mNumEntries == 2 && mFalseJump)
+		{
+			if (mTrueJump == this && mFalseJump != this)
+			{
+				if (mFalseJump->mNumEntries > 1)
+				{
+					InterCodeBasicBlock* suffix = new InterCodeBasicBlock();
+					proc->Append(suffix);
+					InterInstruction* jins = new InterInstruction();
+					jins->mCode = IC_JUMP;
+					suffix->Append(jins);
+					suffix->Close(mFalseJump, nullptr);
+					mFalseJump = suffix;
+					suffix->mNumEntries = 1;
+				}
+			}
+			else if (mFalseJump == this && mTrueJump != this)
+			{
+				if (mTrueJump->mNumEntries > 1)
+				{
+					InterCodeBasicBlock* suffix = new InterCodeBasicBlock();
+					proc->Append(suffix);
+					InterInstruction* jins = new InterInstruction();
+					jins->mCode = IC_JUMP;
+					suffix->Append(jins);
+					suffix->Close(mTrueJump, nullptr);
+					mTrueJump = suffix;
+					suffix->mNumEntries = 1;
+				}
+			}
+		}
+
+		if (mTrueJump)
+			mTrueJump->BuildLoopSuffix(proc);
+		if (mFalseJump)
+			mFalseJump->BuildLoopSuffix(proc);
+	}
+}
 
 InterCodeBasicBlock* InterCodeBasicBlock::BuildLoopPrefix(InterCodeProcedure* proc)
 {
@@ -8809,40 +8930,6 @@ InterCodeBasicBlock* InterCodeBasicBlock::BuildLoopPrefix(InterCodeProcedure* pr
 			jins->mCode = IC_JUMP;
 			mLoopPrefix->Append(jins);
 			mLoopPrefix->Close(this, nullptr);
-
-			if (mNumEntries == 2 && mFalseJump)
-			{
-				if (mTrueJump == this && mFalseJump != this)
-				{
-					if (mFalseJump->mNumEntries > 1)
-					{
-						InterCodeBasicBlock* suffix = new InterCodeBasicBlock();
-						proc->Append(suffix);
-						InterInstruction* jins = new InterInstruction();
-						jins->mCode = IC_JUMP;
-						suffix->Append(jins);
-						suffix->Close(mFalseJump, nullptr);
-						mFalseJump->mNumEntries--;
-						mFalseJump = suffix;
-						suffix->mNumEntries = 1;
-					}
-				}
-				else if (mFalseJump == this && mTrueJump != this)
-				{
-					if (mTrueJump->mNumEntries > 1)
-					{
-						InterCodeBasicBlock* suffix = new InterCodeBasicBlock();
-						proc->Append(suffix);
-						InterInstruction* jins = new InterInstruction();
-						jins->mCode = IC_JUMP;
-						suffix->Append(jins);
-						suffix->Close(mTrueJump, nullptr);
-						mTrueJump->mNumEntries--;
-						mTrueJump = suffix;
-						suffix->mNumEntries = 1;
-					}
-				}
-			}
 		}
 	}
 
@@ -9276,6 +9363,8 @@ void InterCodeBasicBlock::SingleBlockLoopOptimisation(const NumberSet& aliasedPa
 #if 1
 			if (!hasCall)
 			{
+				assert(this == mTrueJump && mFalseJump->mNumEntries == 1 || this == mFalseJump && mTrueJump->mNumEntries == 1);
+
 				// Check forwarding globals
 
 				int i = 0;
@@ -9284,7 +9373,7 @@ void InterCodeBasicBlock::SingleBlockLoopOptimisation(const NumberSet& aliasedPa
 					InterInstruction* ins = mInstructions[i];
 
 					// A global load
-					if (ins->mCode == IC_LOAD && ins->mSrc[0].mTemp < 0 && ins->mSrc[0].mMemory == IM_GLOBAL)
+					if (ins->mCode == IC_LOAD && ins->mSrc[0].mTemp < 0 && (ins->mSrc[0].mMemory == IM_GLOBAL || ins->mSrc[0].mMemory == IM_FPARAM))
 					{
 						// Find the last store that overlaps the load
 						int	j = mInstructions.Size() - 1;
@@ -9312,6 +9401,8 @@ void InterCodeBasicBlock::SingleBlockLoopOptimisation(const NumberSet& aliasedPa
 											k++;
 										if (k == mInstructions.Size())
 										{
+											assert(!mEntryRequiredTemps[sins->mSrc[0].mTemp]);
+
 											// Move load before loop
 											mLoopPrefix->mInstructions.Insert(mLoopPrefix->mInstructions.Size() - 1, ins);
 											InterInstruction* nins = new InterInstruction();
@@ -9760,13 +9851,13 @@ void InterCodeBasicBlock::SingleBlockLoopOptimisation(const NumberSet& aliasedPa
 
 			mInstructions.SetSize(j);
 
-			NumberSet		requiredTemps(mTrueJump == this ? mFalseJump->mExitRequiredTemps : mTrueJump->mExitRequiredTemps);
+			NumberSet		requiredTemps(mTrueJump == this ? mFalseJump->mEntryRequiredTemps : mTrueJump->mEntryRequiredTemps);
 
 			for (int i = 0; i < mInstructions.Size(); i++)
 			{
 				InterInstruction* ins = mInstructions[i];
 				for(int j=0; j<ins->mNumOperands; j++)
-					if (ins->mSrc[j].mTemp >= 0)
+					if (ins->mSrc[j].mTemp >= 0 && ins->mDst.mTemp != ins->mSrc[j].mTemp)
 						requiredTemps += ins->mSrc[j].mTemp;
 			}
 
@@ -9787,6 +9878,16 @@ void InterCodeBasicBlock::SingleBlockLoopOptimisation(const NumberSet& aliasedPa
 					di++;
 				}
 			}
+
+			int	i = 0;
+			while (i < mInstructions.Size())
+			{
+				InterInstruction* ins = mInstructions[i];
+				if (!HasSideEffect(ins->mCode) && !ins->mVolatile && ins->mDst.mTemp >= 0 && !requiredTemps[ins->mDst.mTemp])
+					mInstructions.Remove(i);
+				else
+					i++;
+			}
 		}
 
 		if (mTrueJump)
@@ -10103,7 +10204,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
 		}
 #endif
 
-		bool	changed;
+		bool	changed = false;
 		do
 		{
 			int	j = 0;
@@ -10272,6 +10373,28 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
 						changed = true;
 					}
 #endif
+#if 1
+					else if (
+						mInstructions[i + 0]->mCode == IC_LOAD_TEMPORARY &&
+						mInstructions[i + 1]->mCode == IC_LEA && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i + 0]->mDst.mTemp && mInstructions[i + 1]->mSrc[0].mFinal
+						)
+					{	
+						mInstructions[i + 1]->mSrc[0].Forward(mInstructions[i + 0]->mSrc[0]);
+						mInstructions[i + 0]->mCode = IC_NONE; mInstructions[i + 0]->mNumOperands = 0;
+						changed = true;
+					}
+#endif
+#if 1
+					else if (
+						mInstructions[i + 0]->mCode == IC_LOAD_TEMPORARY &&
+						mInstructions[i + 1]->mCode == IC_LOAD && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i + 0]->mDst.mTemp && mInstructions[i + 1]->mSrc[0].mFinal
+						)
+					{	
+						mInstructions[i + 1]->mSrc[0].Forward(mInstructions[i + 0]->mSrc[0]);
+						mInstructions[i + 0]->mCode = IC_NONE; mInstructions[i + 0]->mNumOperands = 0;
+						changed = true;
+					}
+#endif
 #if 1
 					else if (
 						mInstructions[i + 1]->mCode == IC_LOAD_TEMPORARY && mExitRequiredTemps[mInstructions[i + 1]->mDst.mTemp] &&
@@ -10409,6 +10532,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
 						changed = true;
 					}
 #endif
+
 #if 1
 					// Postincrement artifact
 					if (mInstructions[i + 0]->mCode == IC_LOAD_TEMPORARY && mInstructions[i + 1]->mCode == IC_BINARY_OPERATOR &&
@@ -10435,6 +10559,30 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
 
 #endif
 				}
+
+#if 1
+				if (i + 1 < mInstructions.Size())
+				{
+					if (
+						mInstructions[i + 0]->mCode == IC_LOAD_TEMPORARY &&
+						mInstructions[i + 1]->mCode == IC_RELATIONAL_OPERATOR && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i + 0]->mSrc[0].mTemp && mInstructions[i + 1]->mSrc[0].mFinal
+						)
+					{
+						mInstructions[i + 1]->mSrc[0].mTemp = mInstructions[i + 0]->mDst.mTemp;
+						mInstructions[i + 1]->mSrc[0].mFinal = false;
+						changed = true;
+					}
+					else if (
+						mInstructions[i + 0]->mCode == IC_LOAD_TEMPORARY &&
+						mInstructions[i + 1]->mCode == IC_RELATIONAL_OPERATOR && mInstructions[i + 1]->mSrc[1].mTemp == mInstructions[i + 0]->mSrc[0].mTemp && mInstructions[i + 1]->mSrc[1].mFinal
+						)
+					{
+						mInstructions[i + 1]->mSrc[1].mTemp = mInstructions[i + 0]->mDst.mTemp;
+						mInstructions[i + 1]->mSrc[1].mFinal = false;
+						changed = true;
+					}
+				}
+#endif
 			}
 
 		} while (changed);
@@ -10772,7 +10920,7 @@ void InterCodeProcedure::DisassembleDebug(const char* name)
 	Disassemble(name);
 }
 
-void InterCodeProcedure::BuildTraces(bool expand, bool dominators)
+void InterCodeProcedure::BuildTraces(bool expand, bool dominators, bool compact)
 {
 	// Count number of entries
 //
@@ -10788,7 +10936,7 @@ void InterCodeProcedure::BuildTraces(bool expand, bool dominators)
 	// Build traces
 	//
 	ResetVisited();
-	mEntryBlock->GenerateTraces(expand);
+	mEntryBlock->GenerateTraces(expand, compact);
 
 	ResetVisited();
 	for (int i = 0; i < mBlocks.Size(); i++)
@@ -10827,6 +10975,9 @@ void InterCodeProcedure::BuildDataFlowSets(void)
 	do {
 		ResetVisited();
 	} while (mEntryBlock->BuildGlobalRequiredTempSet(totalRequired));
+
+	ResetVisited();
+	mEntryBlock->CollectLocalUsedTemps(numTemps);
 }
 
 void InterCodeProcedure::RenameTemporaries(void)
@@ -11075,6 +11226,8 @@ void InterCodeProcedure::Close(void)
 
 	RenameTemporaries();
 
+	BuildDataFlowSets();
+
 	TempForwarding();
 
 	int	numTemps = mTemporaries.Size();
@@ -11130,6 +11283,9 @@ void InterCodeProcedure::Close(void)
 		}
 
 		mTemporaries.SetSize(numTemps, true);
+
+		BuildDataFlowSets();
+
 		TempForwarding();
 		retries--;
 
@@ -11225,6 +11381,8 @@ void InterCodeProcedure::Close(void)
 		RenameTemporaries();
 
 		do {
+			BuildDataFlowSets();
+
 			TempForwarding();
 		} while (GlobalConstantPropagation());
 
@@ -11240,9 +11398,6 @@ void InterCodeProcedure::Close(void)
 	}
 
 
-	BuildLoopPrefix();
-	DisassembleDebug("added dominators");
-
 	ResetVisited();
 	mEntryBlock->CompactInstructions();
 
@@ -11266,6 +11421,9 @@ void InterCodeProcedure::Close(void)
 
 	DisassembleDebug("Peephole optimized");
 
+	BuildLoopPrefix();
+	DisassembleDebug("added dominators");
+
 	BuildDataFlowSets();
 
 	TempForwarding();
@@ -11307,9 +11465,6 @@ void InterCodeProcedure::Close(void)
 	{
 		BuildDataFlowSets();
 
-		ResetVisited();
-		mEntryBlock->CollectLocalUsedTemps(mTemporaries.Size());
-
 		ResetVisited();
 		changed = mEntryBlock->PushSinglePathResultInstructions();
 
@@ -11343,6 +11498,8 @@ void InterCodeProcedure::Close(void)
 		ResetVisited();
 		changed = mEntryBlock->MergeCommonPathInstructions();
 
+		DisassembleDebug("Merged common path part");
+
 		if (changed)
 		{
 			TempForwarding();
@@ -11536,9 +11693,6 @@ void InterCodeProcedure::Close(void)
 
 	BuildDataFlowSets();
 
-	ResetVisited();
-	mEntryBlock->CollectLocalUsedTemps(mTemporaries.Size());
-
 	ResetVisited();
 	mEntryBlock->ForwardDiamondMovedTemp();
 	DisassembleDebug("Diamond move forwarding");
@@ -11662,9 +11816,6 @@ void InterCodeProcedure::Close(void)
 #if 1
 	BuildDataFlowSets();
 
-	ResetVisited();
-	mEntryBlock->CollectLocalUsedTemps(mTemporaries.Size());
-
 	ResetVisited();
 	mEntryBlock->ForwardDiamondMovedTemp();
 	DisassembleDebug("Diamond move forwarding 2");
@@ -11711,10 +11862,109 @@ void InterCodeProcedure::Close(void)
 	TempForwarding();
 	RemoveUnusedInstructions();
 
-	DisassembleDebug("Peephole optimized");
+	DisassembleDebug("Global Constant Prop 1");
 
 #endif
 
+#if 1
+	for (int i = 0; i < 4; i++)
+	{
+		ResetVisited();
+		mEntryBlock->PeepholeOptimization(mModule->mGlobalVars);
+
+		DisassembleDebug("Peephole Temp Check");
+
+		ReduceTemporaries();
+
+		MergeBasicBlocks();
+
+		BuildDataFlowSets();
+
+		TempForwarding();
+
+		BuildLoopPrefix();
+
+		BuildDataFlowSets();
+
+		DisassembleDebug("Checking Unused");
+
+		RemoveUnusedInstructions();
+
+		DisassembleDebug("Checked Unused");
+
+		BuildDataFlowSets();
+
+		RenameTemporaries();
+
+		BuildDataFlowSets();
+
+		TempForwarding();
+#if 1
+
+		BuildDataFlowSets();
+		do {
+			TempForwarding();
+		} while (GlobalConstantPropagation());
+
+		do {
+			GrowingInstructionPtrArray	gipa(nullptr);
+			ResetVisited();
+			changed = mEntryBlock->LoadStoreForwarding(gipa, mModule->mGlobalVars);
+
+			DisassembleDebug("Load/Store forwardingX");
+
+			RemoveUnusedStoreInstructions(paramMemory);
+
+			TempForwarding();
+			RemoveUnusedInstructions();
+
+			DisassembleDebug("Load/Store forwarding");
+		} while (changed);
+
+		do
+		{
+			ResetVisited();
+			mEntryBlock->CompactInstructions();
+
+			BuildDataFlowSets();
+
+			ResetVisited();
+			changed = mEntryBlock->MergeCommonPathInstructions();
+
+			DisassembleDebug("Merged common path part");
+
+			if (changed)
+			{
+				TempForwarding();
+				RemoveUnusedInstructions();
+
+			}
+
+		} while (changed);
+
+		DisassembleDebug("Merged common path instructions");
+
+#if 1
+		do
+		{
+			BuildDataFlowSets();
+
+			ResetVisited();
+			changed = mEntryBlock->PushSinglePathResultInstructions();
+
+			DisassembleDebug("Pushed single path result");
+
+		} while (changed);
+#endif
+
+		TempForwarding();
+		RemoveUnusedInstructions();
+#endif
+
+		DisassembleDebug("Global Constant Prop 2");
+	}
+#endif
+
 	MapVariables();
 
 	DisassembleDebug("mapped variabled");
@@ -11726,7 +11976,7 @@ void InterCodeProcedure::Close(void)
 	// Optimize for size
 
 	MergeBasicBlocks();
-	BuildTraces(false, false);
+	BuildTraces(false, false, true);
 	DisassembleDebug("Final Merged basic blocks");
 
 	if (mSaveTempsLinkerObject && mTempSize > 16)
@@ -12007,6 +12257,9 @@ void InterCodeProcedure::BuildLoopPrefix(void)
 	for (int i = 0; i < mBlocks.Size(); i++)
 		mBlocks[i]->mNumEntries = 0;
 	mEntryBlock->CollectEntries();
+
+	ResetVisited();
+	mEntryBlock->BuildLoopSuffix(this);
 }
 
 bool InterCodeProcedure::PropagateNonLocalUsedTemps(void)
@@ -12123,6 +12376,7 @@ void InterCodeProcedure::ReduceTemporaries(void)
 	callerSavedTemps = freeCallerSavedTemps;
 
 	mTempOffset.SetSize(0);
+	mTempSizes.SetSize(0);
 
 	for (int i = 0; i < mTemporaries.Size(); i++)
 	{

oscar64/InterCode.h

@@ -212,6 +212,7 @@ public:
 
 	void Intersect(const TempForwardingTable& table);
 
+	int Size(void) const;
 	void SetSize(int size);
 
 	void Reset(void);
@@ -371,7 +372,7 @@ public:
 
 	void CollectEntries(void);
 	void CollectEntryBlocks(InterCodeBasicBlock* from);
-	void GenerateTraces(bool expand);
+	void GenerateTraces(bool expand, bool compact);
 	void BuildDominatorTree(InterCodeBasicBlock * from);
 
 	void LocalToTemp(int vindex, int temp);
@@ -472,6 +473,7 @@ public:
 	void InnerLoopOptimization(const NumberSet& aliasedParams);
 
 	InterCodeBasicBlock* BuildLoopPrefix(InterCodeProcedure * proc);
+	void BuildLoopSuffix(InterCodeProcedure* proc);
 
 	void SplitBranches(InterCodeProcedure* proc);
 	void FollowJumps(void);
@@ -543,7 +545,7 @@ public:
 	void Disassemble(FILE* file);
 	void Disassemble(const char* name, bool dumpSets = false);
 protected:
-	void BuildTraces(bool expand, bool dominators = true);
+	void BuildTraces(bool expand, bool dominators = true, bool compact = false);
 	void BuildDataFlowSets(void);
 	void RenameTemporaries(void);
 	void TempForwarding(void);

oscar64/NativeCodeGenerator.cpp

@@ -5759,7 +5759,7 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
 				if (InterTypeSize[ins->mDst.mType] == 1)
 				{
 					mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index));
-					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 					if (ainsl)
 					{
 						if (ainsl->mType == ASMIT_ADC)
@@ -5779,7 +5779,7 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
 				else if (InterTypeSize[ins->mDst.mType] == 2)
 				{
 					mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index));
-					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 					if (ainsl)
 					{
 						if (ainsl->mType == ASMIT_ADC)
@@ -5797,7 +5797,7 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
 							mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg));
 
 						mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index + 1));
-						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+						mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 						if (ainsh) mIns.Push(*ainsh);
 						mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg + 1));
 						if (reg == areg)
@@ -5812,17 +5812,17 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
 				else if (InterTypeSize[ins->mDst.mType] == 4)
 				{
 					mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index));
-					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg));
 
 					mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index + 1));
-					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg + 1));
 					mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index + 2));
-					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg + 2));
 					mIns.Push(NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, index + 3));
-					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
+					mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_INDIRECT_Y, areg, nullptr, flags));
 					mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg + 3));
 				}
 			}
@@ -13280,6 +13280,7 @@ bool NativeCodeBasicBlock::JoinTailCodeSequences(NativeCodeProcedure* proc, bool
 
 		}
 #endif
+#if 1
 		if (mEntryBlocks.Size() == 1)
 		{
 			NativeCodeBasicBlock* eblock = mEntryBlocks[0];
@@ -13288,11 +13289,13 @@ bool NativeCodeBasicBlock::JoinTailCodeSequences(NativeCodeProcedure* proc, bool
 				if (mIns[0].mType == ASMIT_ORA && mIns[0].mMode == ASMIM_IMMEDIATE && mIns[0].mAddress == 0 && eblock->mIns.Last().ChangesAccuAndFlag())
 				{
 					eblock->mExitRequiredRegs += CPU_REG_Z;
+					mEntryRequiredRegs += CPU_REG_Z;
 					mIns.Remove(0);
 					changed = true;
 				}
 			}
 		}
+#endif
 #if 1
 		if (loops && mIns.Size() >= 1 && mEntryBlocks.Size() == 2)
 		{
@@ -13837,7 +13840,7 @@ bool NativeCodeBasicBlock::CheckForwardSumYPointer(const NativeCodeBasicBlock* b
 		{
 			NativeCodeInstruction& ins(mIns[at]);
 
-			if (ins.mMode == ASMIM_ZERO_PAGE && (ins.mAddress == reg || ins.mAddress == reg + 1 || ins.mAddress == index))
+			if (ins.mMode == ASMIM_ZERO_PAGE && (ins.mAddress == reg || ins.mAddress == reg + 1))
 				return false;
 			else if (ins.mMode == ASMIM_INDIRECT_Y && ins.mAddress == reg)
 			{
@@ -14786,6 +14789,23 @@ bool NativeCodeBasicBlock::MergeXYSameValue(int from)
 	return false;
 }
 
+int NativeCodeBasicBlock::RetrieveYValue(int at) const
+{
+	while (at > 0 && !mIns[at].ChangesYReg())
+		at--;
+	if (mIns[at].mType == ASMIT_LDY && mIns[at].mMode == ASMIM_IMMEDIATE)
+		return mIns[at].mAddress;
+	else
+		return -1;
+}
+
+void NativeCodeBasicBlock::InsertLoadYImmediate(int at, int val)
+{
+	while (at < mIns.Size() && !mIns[at].ReferencesYReg())
+		at++;
+	mIns.Insert(at, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, val));
+}
+
 bool NativeCodeBasicBlock::PatchGlobalAdressSumYByX(int at, int reg, const NativeCodeInstruction& ains, int addr)
 {
 	int	yindex = 0;
@@ -14867,8 +14887,9 @@ bool NativeCodeBasicBlock::PatchDirectAddressSumY(int at, int reg, int apos, int
 
 	if (mIns[last].mLive & LIVE_CPU_REG_Y)
 	{
-		mIns.Insert(last + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yindex));
+		InsertLoadYImmediate(last + 1, yindex);
-		mIns[last + 1].mLive |= CPU_REG_Y;
+//		mIns.Insert(last + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yindex));
+//		mIns[last + 1].mLive |= CPU_REG_Y;
 	}
 
 	mIns[apos].mType = ASMIT_TAY;
@@ -14934,8 +14955,9 @@ bool NativeCodeBasicBlock::PatchAddressSumY(int at, int reg, int apos, int breg,
 
 	if (mIns[last].mLive & LIVE_CPU_REG_Y)
 	{
-		mIns.Insert(last + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yindex));
+		InsertLoadYImmediate(last + 1, yindex);
-		mIns[last + 1].mLive |= CPU_REG_Y;
+//		mIns.Insert(last + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yindex));
+//		mIns[last + 1].mLive |= CPU_REG_Y;
 	}
 
 	for (int i = 0; i < 5; i++)
@@ -15323,9 +15345,9 @@ bool NativeCodeBasicBlock::MoveLoadIndirectTempStoreUp(int at)
 					mIns[j - 0].mAddress == mIns[at + 1].mAddress + 1 && 
 					mIns[j - 1].mAddress == mIns[j - 3].mAddress + 1)
 				{
+					mIns[at + 0].mLive |= mIns[j].mLive;
 					mIns[at + 1].mLive |= mIns[j].mLive;
 					mIns[at + 2].mLive |= mIns[j].mLive;
-					mIns[at + 3].mLive |= mIns[j].mLive;
 
 					mIns[at + 1].mAddress = mIns[j - 3].mAddress;
 					mIns[at + 1].mLive |= LIVE_MEM;
@@ -15364,9 +15386,9 @@ bool NativeCodeBasicBlock::MoveLoadIndirectTempStoreUp(int at)
 							return false;
 					}
 
+					mIns[at + 0].mLive |= mIns[j].mLive;
 					mIns[at + 1].mLive |= mIns[j].mLive;
 					mIns[at + 2].mLive |= mIns[j].mLive;
-					mIns[at + 3].mLive |= mIns[j].mLive;
 
 					mIns[at + 1].mAddress = addr;
 					mIns[at + 1].mLive |= LIVE_MEM;
@@ -20441,8 +20463,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 							}
 							if (mIns[i + 0].mLive & LIVE_CPU_REG_Y)
 							{
-								mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+								InsertLoadYImmediate(i + 1, 0);
-								mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
+//								mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+//								mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
 							}
 							mIns.Insert(i + 0, NativeCodeInstruction(ASMIT_LDY, ASMIM_ZERO_PAGE, ireg));
 							mIns[i + 0].mLive |= LIVE_CPU_REG_Y | LIVE_MEM;
@@ -20464,13 +20487,17 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 						{
 							if (mIns[i + 0].mLive & LIVE_CPU_REG_Y)
 							{
-								mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
-								mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
 								if (mIns[i + 0].mLive & LIVE_CPU_REG_Z)
 								{
+									mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+									mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
 									mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_ORA, ASMIM_IMMEDIATE, 0));
 									mIns[i + 2].mLive |= LIVE_CPU_REG_Y | LIVE_CPU_REG_Z;
 								}
+								else
+								{
+									InsertLoadYImmediate(i + 1, 0);
+								}
 							}
 
 							if (flags & LIVE_CPU_REG_Y)
@@ -20557,8 +20584,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 #if 1
 						if (mIns[i + 0].mLive & LIVE_CPU_REG_Y)
 						{
-							mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+							InsertLoadYImmediate(i + 1, 0);
-							mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
+//							mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+//							mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
 						}
 						mIns.Insert(i + 0, NativeCodeInstruction(ASMIT_LDY, ASMIM_ZERO_PAGE, ireg));
 						mIns[i + 0].mLive |= LIVE_CPU_REG_Y | LIVE_MEM;
@@ -21012,13 +21040,17 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 							{
 								if (mIns[i + 1].mLive & LIVE_CPU_REG_Y)
 								{
-									mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, mIns[i + 0].mAddress));
-									mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
 									if (mIns[i + 1].mLive & LIVE_CPU_REG_Z)
 									{
+										mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, mIns[i + 0].mAddress));
+										mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
 										mIns.Insert(i + 3, NativeCodeInstruction(ASMIT_ORA, ASMIM_IMMEDIATE, 0));
 										mIns[i + 3].mLive |= LIVE_CPU_REG_Y | LIVE_CPU_REG_Z;
 									}
+									else
+									{
+										InsertLoadYImmediate(i + 2, 0);
+									}
 								}
 
 								if (flags & LIVE_CPU_REG_Y)
@@ -21180,8 +21212,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 								}
 								if (mIns[i + 1].mLive & LIVE_CPU_REG_Y)
 								{
-									mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+									InsertLoadYImmediate(i + 2, 0);
-									mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
+//									mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+//									mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
 								}
 								mIns.Insert(i + 0, NativeCodeInstruction(ASMIT_LDY, ASMIM_ZERO_PAGE, ireg));
 								mIns[i + 0].mLive |= LIVE_CPU_REG_Y | LIVE_MEM;
@@ -21215,8 +21248,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 								}
 								if (mIns[i + 1].mLive & LIVE_CPU_REG_Y)
 								{
-									mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yoffset));
+									InsertLoadYImmediate(i + 2, 0);
-									mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
+//									mIns.Insert(i + 2, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yoffset));
+//									mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
 								}
 
 								mIns[i + 0].mMode = ASMIM_ZERO_PAGE;
@@ -21302,7 +21336,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 					else if (
 						mIns[i + 0].mType == ASMIT_LDA && mIns[i + 0].mMode == ASMIM_ZERO_PAGE &&
 						mIns[i + 2].mType == ASMIT_TAY && !(mIns[i + 2].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_Z)) &&
-						!mIns[i + 1].ChangesAccu() && !mIns[i + 1].RequiresAccu() && !(mIns[i + 1].mLive & LIVE_CPU_REG_Y))
+						!mIns[i + 1].ChangesAccu() && !mIns[i + 1].RequiresAccu() && !mIns[i + 1].RequiresYReg())
 					{
 						mIns[i + 0].mType = ASMIT_LDY;
 						mIns[i + 2].mType = ASMIT_NOP; mIns[i + 2].mMode = ASMIM_IMPLIED;
@@ -22135,8 +22169,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 							}
 							if (mIns[i + 2].mLive & LIVE_CPU_REG_Y)
 							{
-								mIns.Insert(i + 3, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yoffset));
+								InsertLoadYImmediate(i + 3, yoffset);
-								mIns[i + 3].mLive |= LIVE_CPU_REG_Y;
+//								mIns.Insert(i + 3, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, yoffset));
+//								mIns[i + 3].mLive |= LIVE_CPU_REG_Y;
 							}
 							
 							int ypos = i;
@@ -22897,8 +22932,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 							}
 							if (mIns[i + 3].mLive & LIVE_CPU_REG_Y)
 							{
-								mIns.Insert(i + 4, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+								InsertLoadYImmediate(i + 4, 0);
-								mIns[i + 4].mLive |= LIVE_CPU_REG_Y;
+//								mIns.Insert(i + 4, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+//								mIns[i + 4].mLive |= LIVE_CPU_REG_Y;
 							}
 							mIns[i + 0].mMode = ASMIM_ZERO_PAGE;
 							mIns[i + 0].mAddress = ireg;
@@ -23066,8 +23102,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 							}
 							if (mIns[i + 4].mLive & LIVE_CPU_REG_Y)
 							{
-								mIns.Insert(i + 5, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+								InsertLoadYImmediate(i + 5, 0);
-								mIns[i + 5].mLive |= LIVE_CPU_REG_Y;
+//								mIns.Insert(i + 5, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
+//								mIns[i + 5].mLive |= LIVE_CPU_REG_Y;
 							}
 							mIns[i + 0].mMode = ASMIM_ZERO_PAGE;
 							mIns[i + 0].mAddress = ireg;
@@ -23269,9 +23306,13 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 						mIns[i + 1].mType == ASMIT_STA && mIns[i + 1].mMode == ASMIM_ZERO_PAGE)
 					{
 						int	n = 3;
+						if (mIns[i + 0].mFlags & NCIF_VOLATILE)
+							n = 1;
 						if (mIns[i + 1].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_Z))
 							n--;
 
+						if (n > 0)
+						{
 							proc->ResetPatched();
 							if (CheckSingleUseGlobalLoad(this, mIns[i + 1].mAddress, i + 2, mIns[i], n))
 							{
@@ -23293,6 +23334,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 								CheckLive();
 							}
 						}
+					}
 					else if (
 						mIns[i + 0].mType == ASMIT_STA && mIns[i + 0].mMode == ASMIM_ZERO_PAGE &&
 						mIns[i + 1].mType == ASMIT_STA && mIns[i + 1].mMode == ASMIM_ABSOLUTE)
@@ -23554,8 +23596,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 						mIns[i + 6].mType == ASMIT_STA && mIns[i + 6].mMode == ASMIM_ZERO_PAGE && mIns[i + 6].mAddress == mIns[i + 3].mAddress + 1 &&
 						!(mIns[i + 6].mLive & LIVE_CPU_REG_A))
 					{
+						int yval = RetrieveYValue(i);
 						proc->ResetPatched();
-						if (CheckForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 1].mAddress, mIns[i + 2].mAddress, i + 7, -1))
+						if (CheckForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 1].mAddress, mIns[i + 2].mAddress, i + 7, yval))
 						{
 							if (mIns[i + 3].mAddress == mIns[i + 1].mAddress)
 							{
@@ -23566,7 +23609,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 							}
 
 							proc->ResetPatched();
-							if (PatchForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 1].mAddress, mIns[i + 2].mAddress, i + 7, -1))
+							if (PatchForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 1].mAddress, mIns[i + 2].mAddress, i + 7, yval))
 								progress = true;
 						}
 					}
@@ -23582,19 +23625,20 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 						mIns[i + 6].mType == ASMIT_STA && mIns[i + 6].mMode == ASMIM_ZERO_PAGE && mIns[i + 6].mAddress == mIns[i + 3].mAddress + 1 &&
 						!(mIns[i + 6].mLive & LIVE_CPU_REG_A))
 					{
+						int yval = RetrieveYValue(i);
 						proc->ResetPatched();
-						if (CheckForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 2].mAddress, mIns[i + 0].mAddress, i + 7, -1))
+						if (CheckForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 2].mAddress, mIns[i + 0].mAddress, i + 7, yval))
 						{
 							if (mIns[i + 3].mAddress == mIns[i + 2].mAddress)
 							{
-								for (int j = 0; j < 7; j++)
+								for (int j = 1; j < 7; j++)
 								{
 									mIns[i + j].mType = ASMIT_NOP; mIns[i + j].mMode = ASMIM_IMPLIED;
 								}
 							}
 
 							proc->ResetPatched();
-							if (PatchForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 2].mAddress, mIns[i + 0].mAddress, i + 7, -1))
+							if (PatchForwardSumYPointer(this, mIns[i + 3].mAddress, mIns[i + 2].mAddress, mIns[i + 0].mAddress, i + 7, yval))
 								progress = true;
 						}
 					}
@@ -23770,6 +23814,8 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 
 		} while (progress);
 
+		CheckLive();
+
 		int	sz = mIns.Size();
 #if 1
 		if (sz >= 2 &&
@@ -23807,43 +23853,47 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 #if 1
 		else if (sz >= 2 &&
 			mIns[sz - 2].mType == ASMIT_LDA && mIns[sz - 2].mMode == ASMIM_IMMEDIATE && mIns[sz - 2].mAddress == 0 &&
-			mIns[sz - 1].mType == ASMIT_ROL && mIns[sz - 1].mMode == ASMIM_IMPLIED && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_C)))
+			mIns[sz - 1].mType == ASMIT_ROL && mIns[sz - 1].mMode == ASMIM_IMPLIED && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_C)) && !mExitRequiredRegs[CPU_REG_Z])
 		{
 			if (mBranch == ASMIT_BNE)
 			{
 				mBranch = ASMIT_BCS;
 				mIns.SetSize(sz - 2);
+				sz -= 2;
 				changed = true;
 			}
 			else if (mBranch == ASMIT_BEQ)
 			{
 				mBranch = ASMIT_BCC;
 				mIns.SetSize(sz - 2);
+				sz -= 2;
 				changed = true;
 			}
 		}
 		else if (sz >= 3 &&
 			mIns[sz - 3].mType == ASMIT_LDA && mIns[sz - 3].mMode == ASMIM_IMMEDIATE && mIns[sz - 3].mAddress == 0 &&
 			mIns[sz - 2].mType == ASMIT_ROL && mIns[sz - 2].mMode == ASMIM_IMPLIED && 
-			mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_IMMEDIATE && mIns[sz - 1].mAddress == 0x0 && !(mIns[sz - 1].mLive & LIVE_CPU_REG_A))
+			mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_IMMEDIATE && mIns[sz - 1].mAddress == 0x0 && !(mIns[sz - 1].mLive & LIVE_CPU_REG_A) && !mExitRequiredRegs[CPU_REG_Z])
 		{
 			if (mBranch == ASMIT_BNE)
 			{
 				mBranch = ASMIT_BCS;
 				mIns.SetSize(sz - 3);
+				sz -= 3;
 				changed = true;
 			}
 			else if (mBranch == ASMIT_BEQ)
 			{
 				mBranch = ASMIT_BCC;
 				mIns.SetSize(sz - 3);
+				sz -= 3;
 				changed = true;
 			}
 		}
 
 		else if (sz >= 2 &&
 			mIns[sz - 2].mType == ASMIT_EOR && mIns[sz - 2].mMode == ASMIM_IMMEDIATE && mIns[sz - 2].mAddress == 0x80 &&
-			mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_IMMEDIATE && mIns[sz - 1].mAddress == 0x80 && !(mIns[sz - 1].mLive & LIVE_CPU_REG_A))
+			mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_IMMEDIATE && mIns[sz - 1].mAddress == 0x80 && !(mIns[sz - 1].mLive & LIVE_CPU_REG_A) && !mExitRequiredRegs[CPU_REG_Z])
 		{
 			if (mBranch == ASMIT_BNE && mTrueJump->mIns.Size() == 0 && mTrueJump->mBranch == ASMIT_BCC)
 			{
@@ -23865,6 +23915,8 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 
 				mIns[sz - 2].mType = ASMIT_NOP; mIns[sz - 2].mMode = ASMIM_IMPLIED;
 				mIns[sz - 1].mType = ASMIT_ORA; mIns[sz - 1].mAddress = 0; mIns[sz - 1].mLive |= LIVE_CPU_REG_Z;
+
+				CheckLive();
 			}
 		}
 
@@ -23872,7 +23924,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 			mIns[sz - 4].mType == ASMIT_EOR && mIns[sz - 4].mMode == ASMIM_IMMEDIATE && mIns[sz - 4].mAddress == 0x80 &&
 			mIns[sz - 3].mType == ASMIT_STA && mIns[sz - 3].mMode == ASMIM_ZERO_PAGE &&
 			mIns[sz - 2].mType == ASMIT_LDA && mIns[sz - 2].mMode == ASMIM_IMMEDIATE && mIns[sz - 2].mAddress == 0x80 &&
-			mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_ZERO_PAGE && mIns[sz - 1].mAddress == mIns[sz - 3].mAddress && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_MEM)))
+			mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].mMode == ASMIM_ZERO_PAGE && mIns[sz - 1].mAddress == mIns[sz - 3].mAddress && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_MEM)) && !mExitRequiredRegs[CPU_REG_Z])
 		{
 			if (mBranch == ASMIT_BNE && mTrueJump->mIns.Size() == 0 && mTrueJump->mBranch == ASMIT_BCC)
 			{
@@ -23896,12 +23948,14 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 				mIns[sz - 3].mType = ASMIT_NOP; mIns[sz - 3].mMode = ASMIM_IMPLIED;
 				mIns[sz - 2].mType = ASMIT_NOP; mIns[sz - 2].mMode = ASMIM_IMPLIED;
 				mIns[sz - 1].mType = ASMIT_ORA; mIns[sz - 1].mMode = ASMIM_IMMEDIATE;  mIns[sz - 1].mAddress = 0; mIns[sz - 1].mLive |= LIVE_CPU_REG_Z;
+
+				CheckLive();
 			}
 		}
 
 		else if (sz >= 2 &&
 			mIns[sz - 2].mType == ASMIT_LDA && mIns[sz - 2].mMode == ASMIM_IMMEDIATE && mIns[sz - 2].mAddress == 0 &&
-			mIns[sz - 1].mType == ASMIT_SBC && mIns[sz - 1].mMode == ASMIM_IMMEDIATE && mIns[sz - 1].mAddress == 0 && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_C)))
+			mIns[sz - 1].mType == ASMIT_SBC && mIns[sz - 1].mMode == ASMIM_IMMEDIATE && mIns[sz - 1].mAddress == 0 && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_C)) && !mExitRequiredRegs[CPU_REG_Z])
 		{
 			if (mBranch == ASMIT_BNE)
 			{
@@ -23915,6 +23969,9 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 				mIns.SetSize(sz - 2);
 				changed = true;
 			}
+
+			sz -= 2;
+			CheckLive();
 		}
 
 		if (sz >= 1 && 
@@ -23931,16 +23988,20 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 			changed = true;
 		}
 #endif
+
+		CheckLive();
+
 #if 1
 		if (mTrueJump && mFalseJump && !mTrueJump->mFalseJump && !mFalseJump->mFalseJump && mTrueJump->mTrueJump == mFalseJump->mTrueJump &&
 			mTrueJump->mIns.Size() == 1 && mFalseJump->mIns.Size() == 1 && 
 			mTrueJump->mIns[0].mType == ASMIT_LDA && mTrueJump->mIns[0].mMode == ASMIM_IMMEDIATE &&
 			mFalseJump->mIns[0].mType == ASMIT_LDA && mFalseJump->mIns[0].mMode == ASMIM_IMMEDIATE)
 		{
-			if (mBranch == ASMIT_BCS && mTrueJump->mIns[0].mAddress == 1 && mFalseJump->mIns[0].mAddress == 0)
+			if (mBranch == ASMIT_BCS && mTrueJump->mIns[0].mAddress == 1 && mFalseJump->mIns[0].mAddress == 0 && !(mExitRequiredRegs[CPU_REG_C]))
 			{
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
 				mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
+				mExitProvidedRegs += CPU_REG_A;
 				mBranch = ASMIT_JMP;
 				mTrueJump = mTrueJump->mTrueJump;
 				mFalseJump = nullptr;
@@ -23948,10 +24009,11 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 
 				CheckLive();
 			}
-			else if (mBranch == ASMIT_BCC && mTrueJump->mIns[0].mAddress == 0 && mFalseJump->mIns[0].mAddress == 1)
+			else if (mBranch == ASMIT_BCC && mTrueJump->mIns[0].mAddress == 0 && mFalseJump->mIns[0].mAddress == 1 && !(mExitRequiredRegs[CPU_REG_C]))
 			{
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
 				mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
+				mExitProvidedRegs += CPU_REG_A;
 				mBranch = ASMIT_JMP;
 				mTrueJump = mTrueJump->mTrueJump;
 				mFalseJump = nullptr;
@@ -23959,11 +24021,12 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 
 				CheckLive();
 			}
-			else if (mBranch == ASMIT_BCS && mTrueJump->mIns[0].mAddress == 0 && mFalseJump->mIns[0].mAddress == 1)
+			else if (mBranch == ASMIT_BCS && mTrueJump->mIns[0].mAddress == 0 && mFalseJump->mIns[0].mAddress == 1 && !(mExitRequiredRegs[CPU_REG_C]))
 			{
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
 				mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
 				mIns.Push(NativeCodeInstruction(ASMIT_EOR, ASMIM_IMMEDIATE, 1));
+				mExitProvidedRegs += CPU_REG_A;
 				mBranch = ASMIT_JMP;
 				mTrueJump = mTrueJump->mTrueJump;
 				mFalseJump = nullptr;
@@ -23971,11 +24034,12 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 
 				CheckLive();
 			}
-			else if (mBranch == ASMIT_BCC && mTrueJump->mIns[0].mAddress == 1 && mFalseJump->mIns[0].mAddress == 0)
+			else if (mBranch == ASMIT_BCC && mTrueJump->mIns[0].mAddress == 1 && mFalseJump->mIns[0].mAddress == 0 && !(mExitRequiredRegs[CPU_REG_C]))
 			{
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
 				mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
 				mIns.Push(NativeCodeInstruction(ASMIT_EOR, ASMIM_IMMEDIATE, 1));
+				mExitProvidedRegs += CPU_REG_A;
 				mBranch = ASMIT_JMP;
 				mTrueJump = mTrueJump->mTrueJump;
 				mFalseJump = nullptr;
@@ -23990,6 +24054,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 					mIns.Insert(mIns.Size() - 1, NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
 					mIns[mIns.Size() - 1].mType = ASMIT_CMP; mIns[mIns.Size() - 1].mLive |= LIVE_CPU_REG_C;
 					mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
+					mExitProvidedRegs += CPU_REG_A;
 					mBranch = ASMIT_JMP;
 					mTrueJump = mTrueJump->mTrueJump;
 					mFalseJump = nullptr;
@@ -24002,6 +24067,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 					mIns.Insert(mIns.Size() - 1, NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
 					mIns[mIns.Size() - 1].mType = ASMIT_CMP; mIns[mIns.Size() - 1].mLive |= LIVE_CPU_REG_C;
 					mIns.Push(NativeCodeInstruction(ASMIT_ROL, ASMIM_IMPLIED));
+					mExitProvidedRegs += CPU_REG_A;
 					mBranch = ASMIT_JMP;
 					mTrueJump = mTrueJump->mTrueJump;
 					mFalseJump = nullptr;
@@ -24012,6 +24078,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
 			}
 		}
 #endif
+		CheckLive();
 
 #endif
 		assert(mIndex == 1000 || mNumEntries == mEntryBlocks.Size());
@@ -24056,6 +24123,8 @@ void NativeCodeBasicBlock::CheckLive(void)
 
 	for (int j = mIns.Size() - 1; j >= 0; j--)
 	{
+		assert(mIns[j].mType != ASMIT_INV);
+
 		if (mIns[j].mType != ASMIT_NOP)
 		{
 			assert(!(live & ~mIns[j].mLive));
@@ -25037,6 +25106,7 @@ void NativeCodeProcedure::Optimize(void)
 			changed = true;
 #endif
 
+
 #if 1
 		if (step < 6)
 		{
@@ -25046,6 +25116,7 @@ void NativeCodeProcedure::Optimize(void)
 				changed = true;
 			}
 		}
+
 #if 1
 		if (step == 3)
 		{
@@ -25067,10 +25138,11 @@ void NativeCodeProcedure::Optimize(void)
 				changed = true;
 		}
 #endif
+#if 1
 		ResetVisited();
 		if (mEntryBlock->MergeBasicBlocks())
 			changed = true;
-
+#endif
 		ResetEntryBlocks();
 		ResetVisited();
 		mEntryBlock->CollectEntryBlocks(nullptr);
@@ -25082,6 +25154,7 @@ void NativeCodeProcedure::Optimize(void)
 				changed = true;
 		}
 #endif
+
 #if 1
 		if (step > 2 && !changed)
 		{
@@ -25108,9 +25181,9 @@ void NativeCodeProcedure::Optimize(void)
 			if (mEntryBlock->ReduceLocalYPressure())
 				changed = true;
 #endif
-
 		}
 #endif
+
 #if 1
 		if (step == 4 || step == 5)
 		{
@@ -25262,12 +25335,12 @@ void NativeCodeProcedure::Optimize(void)
 #endif
 
 #endif
-
 		if (cnt > 200)
 		{
 			changed = false;
 			mGenerator->mErrors->Error(mInterProc->mLocation, EWARN_OPTIMIZER_LOCKED, "Optimizer locked in infinite loop", mInterProc->mIdent);
 		}
+
 #if 1
 		if (!changed && step < 8)
 		{
@@ -25275,7 +25348,10 @@ void NativeCodeProcedure::Optimize(void)
 			changed = true;
 		}
 #endif
+
+
 		cnt++;
+
 	} while (changed);
 
 #if 1
@@ -25650,7 +25726,7 @@ void NativeCodeProcedure::CompileInterBlock(InterCodeProcedure* iproc, InterCode
 		}	break;
 
 		case IC_RELATIONAL_OPERATOR:
-			if (i + 1 < iblock->mInstructions.Size() && iblock->mInstructions[i + 1]->mCode == IC_BRANCH && iblock->mInstructions[i + 1]->mSrc[0].mFinal)
+			if (i + 1 < iblock->mInstructions.Size() && iblock->mInstructions[i + 1]->mCode == IC_BRANCH && iblock->mInstructions[i + 1]->mSrc[0].mTemp == ins->mDst.mTemp && iblock->mInstructions[i + 1]->mSrc[0].mFinal)
 			{
 				block->RelationalOperator(iproc, ins, this, CompileBlock(iproc, iblock->mTrueJump), CompileBlock(iproc, iblock->mFalseJump));
 				return;

oscar64/NativeCodeGenerator.h

@@ -303,6 +303,8 @@ public:
 	bool JoinTAYARange(int from, int to);
 	bool PatchGlobalAdressSumYByX(int at, int reg, const NativeCodeInstruction& ains, int addr);
 	bool MergeXYSameValue(int from);
+	void InsertLoadYImmediate(int at, int val);
+	int RetrieveYValue(int at) const;
 
 	bool ReverseReplaceTAX(int at);