Optimized long multiply

autotest/autotest.bat

@@ -232,115 +232,115 @@ echo Failed with error #%errorlevel%.
 exit /b %errorlevel%
 
 :testh
-..\release\oscar64 -e -bc %~1
+..\bin\oscar64 -e -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -n %~1
+..\bin\oscar64 -e -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -bc %~1
+..\bin\oscar64 -e -O2 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -n %~1
+..\bin\oscar64 -e -O2 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -n -dHEAPCHECK %~1
+..\bin\oscar64 -e -O2 -n -dHEAPCHECK %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O0 -bc %~1
+..\bin\oscar64 -e -O0 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O0 -n %~1
+..\bin\oscar64 -e -O0 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -Os -bc %~1
+..\bin\oscar64 -e -Os -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -Os -n %~1
+..\bin\oscar64 -e -Os -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O3 -bc %~1
+..\bin\oscar64 -e -O3 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O3 -n %~1
+..\bin\oscar64 -e -O3 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O3 -n -dHEAPCHECK %~1
+..\bin\oscar64 -e -O3 -n -dHEAPCHECK %~1
 @if %errorlevel% neq 0 goto :error
 
 @exit /b 0
 
 :test
-..\release\oscar64 -e -bc %~1
+..\bin\oscar64 -e -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -n %~1
+..\bin\oscar64 -e -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -bc %~1
+..\bin\oscar64 -e -O2 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -n %~1
+..\bin\oscar64 -e -O2 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O0 -bc %~1
+..\bin\oscar64 -e -O0 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O0 -n %~1
+..\bin\oscar64 -e -O0 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -Os -bc %~1
+..\bin\oscar64 -e -Os -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -Os -n %~1
+..\bin\oscar64 -e -Os -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O3 -bc %~1
+..\bin\oscar64 -e -O3 -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O3 -n %~1
+..\bin\oscar64 -e -O3 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -xz -Oz -n %~1
+..\bin\oscar64 -e -O2 -xz -Oz -n %~1
 @if %errorlevel% neq 0 goto :error
 
 @exit /b 0
 
 :testb
-..\release\oscar64 -e -bc %~1
+..\bin\oscar64 -e -bc %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -bc -O2 %~1
+..\bin\oscar64 -e -bc -O2 %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -bc -O0 %~1
+..\bin\oscar64 -e -bc -O0 %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -bc -Os %~1
+..\bin\oscar64 -e -bc -Os %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -bc -O3 %~1
+..\bin\oscar64 -e -bc -O3 %~1
 @if %errorlevel% neq 0 goto :error
 
 @exit /b 0
 
 :testn
-..\release\oscar64 -e -n %~1
+..\bin\oscar64 -e -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -n %~1
+..\bin\oscar64 -e -O2 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O0 -n %~1
+..\bin\oscar64 -e -O0 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -Os -n %~1
+..\bin\oscar64 -e -Os -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O3 -n %~1
+..\bin\oscar64 -e -O3 -n %~1
 @if %errorlevel% neq 0 goto :error
 
-..\release\oscar64 -e -O2 -xz -Oz -n %~1
+..\bin\oscar64 -e -O2 -xz -Oz -n %~1
 @if %errorlevel% neq 0 goto :error
 
 @exit /b 0

autotest/testint32.c

@@ -5,6 +5,11 @@ void testmuli(long a, long b, long ab)
 	assert (a * b == ab);
 }
 
+void testmulu(unsigned long a, unsigned long b, unsigned long ab)
+{
+	assert (a * b == ab);
+}
+
 void testdivi(long a, long b, long ab)
 {
 	assert (a / b == ab);
@@ -80,6 +85,26 @@ int main(void)
 	testmuli( -1024, 1237, -1266688l);
 	testmuli( -1024,-1237, 1266688l);
 
+	testmulu(0x00000001, 0x0000003c, 0x0000003c);
+	testmulu(0x00000100, 0x0000003c, 0x00003c00);
+	testmulu(0x00010000, 0x0000003c, 0x003c0000);
+	testmulu(0x01000000, 0x0000003c, 0x3c000000);
+
+	testmulu(0x0000003c, 0x00000001, 0x0000003c);
+	testmulu(0x0000003c, 0x00000100, 0x00003c00);
+	testmulu(0x0000003c, 0x00010000, 0x003c0000);
+	testmulu(0x0000003c, 0x01000000, 0x3c000000);
+
+	testmulu(0x0000004b, 0x0000003c, 0x00001194);
+	testmulu(0x00004b00, 0x0000003c, 0x00119400);
+	testmulu(0x004b0000, 0x0000003c, 0x11940000);
+	testmulu(0x4b000000, 0x0000003c, 0x94000000);
+
+	testmulu(0x0000003c, 0x0000004b, 0x00001194);
+	testmulu(0x0000003c, 0x00004b00, 0x00119400);
+	testmulu(0x0000003c, 0x004b0000, 0x11940000);
+	testmulu(0x0000003c, 0x4b000000, 0x94000000);
+
 	testdivi( 1,  1,  1);
 	testdivi(-1,  1, -1);
 	testdivi( 1, -1, -1);

include/crt.c

@@ -695,6 +695,11 @@ L4:
 		sty tmp + 2
 }
 
+__asm mul32by8
+{
+
+}
+
 __asm mul32
 {
 		lda	#0
@@ -703,6 +708,52 @@ __asm mul32
 		sta	tmp + 6
 		sta	tmp + 7
 
+		lda	tmp + 0
+		jsr WM
+		lda	tmp + 1
+		jsr WM
+		lda	tmp + 2
+		jsr WM
+		lda	tmp + 3
+WM:
+		bne W0
+		ldx accu + 2
+		stx accu + 3
+		ldx accu + 1
+		stx accu + 2
+		ldx accu
+		stx accu + 1
+		sta accu
+		rts
+W0:
+		sec
+		ror
+		bcc W1
+L1:		tax
+		clc
+		lda	tmp + 4
+		adc	accu
+		sta	tmp + 4
+		lda	tmp + 5
+		adc	accu + 1
+		sta	tmp + 5
+		lda	tmp + 6
+		adc	accu + 2
+		sta	tmp + 6
+		lda	tmp + 7
+		adc	accu + 3
+		sta	tmp + 7
+		txa
+W1:		asl	accu
+		rol	accu + 1
+		rol	accu + 2
+		rol	accu + 3
+		lsr
+		bcc W1
+		bne L1
+		rts
+
+#if 0
 		ldx	#32
 L1:		lsr	tmp + 3
 		ror	tmp + 2
@@ -729,6 +780,7 @@ W1:		asl	accu
 		dex
 		bne	L1
 		rts
+#endif
 }
 
 __asm mul16by8

oscar64/InterCode.cpp

@@ -967,8 +967,11 @@ static int64 ConstantFolding(InterOperator oper, InterType type, int64 val1, int
 		return val1 - val2;
 		break;
 	case IA_MUL:
+		if (type == IT_INT32 && val1 >= 0 && val2 >= 0)
+			return val1 * val2 & 0xffffffff;
+		else
 			return val1 * val2;
-		break;
+
 	case IA_DIVU:
 		if (val2)
 			return (uint64)val1 / (uint64)val2;
@@ -11042,6 +11045,123 @@ void InterCodeBasicBlock::LinkerObjectForwarding(const GrowingInstructionPtrArra
 	}
 }
 
+void InterCodeBasicBlock::ReduceRecursionTempSpilling(InterMemory paramMemory, const GrowingInstructionPtrArray& tvalue)
+{
+	if (!mVisited)
+	{
+		if (!mLoopHead)
+		{
+			if (mNumEntries > 0)
+			{
+				if (mNumEntered == 0)
+					mLoadStoreInstructions = tvalue;
+				else
+				{
+					int i = 0;
+					while (i < mLoadStoreInstructions.Size())
+					{
+						InterInstruction* ins(mLoadStoreInstructions[i]);
+						InterInstruction* nins = nullptr;
+
+						int j = tvalue.IndexOf(ins);
+						if (j != -1)
+							nins = ins;
+
+						if (nins)
+							mLoadStoreInstructions[i++] = nins;
+						else
+							mLoadStoreInstructions.Remove(i);
+					}
+				}
+
+				mNumEntered++;
+
+				if (mNumEntered < mNumEntries)
+					return;
+			}
+		}
+#if 1
+		else if (mNumEntries == 2 && (mTrueJump == this || mFalseJump == this))
+		{
+			mLoadStoreInstructions = tvalue;
+			for (int i = 0; i < mInstructions.Size(); i++)
+			{
+				InterInstruction* ins(mInstructions[i]);
+				if (ins->mDst.mTemp >= 0)
+				{
+					int j = 0;
+					while (j < mLoadStoreInstructions.Size())
+					{
+						if (mLoadStoreInstructions[j]->ReferencesTemp(ins->mDst.mTemp) || CollidingMem(ins, mLoadStoreInstructions[j]))
+							mLoadStoreInstructions.Remove(j);
+						else
+							j++;
+					}
+				}
+			}
+		}
+#endif
+		else
+			mLoadStoreInstructions.SetSize(0);
+
+		mVisited = true;
+
+		NumberSet	rtemps(mEntryRequiredTemps);
+
+		for (int i = 0; i < mInstructions.Size(); i++)
+		{
+			InterInstruction* ins(mInstructions[i]);
+			InterInstruction* lins = nullptr;
+			bool			flushMem = false;
+
+			if (ins->mCode == IC_CALL || ins->mCode == IC_CALL_NATIVE)
+			{
+				if (ins->mSrc[0].mLinkerObject == mProc->mLinkerObject)
+				{
+					for (int j = 0; j < mLoadStoreInstructions.Size(); j++)
+					{
+						if (rtemps[mLoadStoreInstructions[j]->mDst.mTemp])
+							mInstructions.Insert(i + 1, mLoadStoreInstructions[j]->Clone());
+					}
+				}
+			}
+			else if (ins->mCode == IC_LOAD && ins->mSrc[0].mTemp < 0 && ins->mSrc[0].mMemory == paramMemory)
+			{
+				if (InterTypeSize[ins->mDst.mType] == ins->mSrc[0].mOperandSize)
+					lins = ins;
+			}
+
+			for (int j = 0; j < ins->mNumOperands; j++)
+			{
+				if (ins->mSrc[j].mTemp >= 0 && ins->mSrc[j].mFinal)
+					rtemps -= ins->mSrc[j].mTemp;
+			}
+
+			int	j = 0, k = 0, t = ins->mDst.mTemp;
+			if (t >= 0 || IsObservable(ins->mCode))
+			{
+				while (j < mLoadStoreInstructions.Size())
+				{
+					if (DestroyingMem(mLoadStoreInstructions[j], ins))
+						;
+					else if (t != mLoadStoreInstructions[j]->mDst.mTemp)
+						mLoadStoreInstructions[k++] = mLoadStoreInstructions[j];
+
+					j++;
+				}
+				mLoadStoreInstructions.SetSize(k);
+			}
+
+			if (lins)
+				mLoadStoreInstructions.Push(lins);
+		}
+
+		if (mTrueJump) mTrueJump->ReduceRecursionTempSpilling(paramMemory, mLoadStoreInstructions);
+		if (mFalseJump) mFalseJump->ReduceRecursionTempSpilling(paramMemory, mLoadStoreInstructions);
+	}
+}
+
+
 bool InterCodeBasicBlock::LoadStoreForwarding(const GrowingInstructionPtrArray& tvalue, const GrowingVariableArray& staticVars)
 {
 	bool	changed = false;
@@ -19820,6 +19940,15 @@ void InterCodeProcedure::EliminateAliasValues()
 	DisassembleDebug("EliminateAliasValues");
 }
 
+void InterCodeProcedure::ReduceRecursionTempSpilling(InterMemory paramMemory)
+{
+	GrowingInstructionPtrArray	gipa(nullptr);
+	ResetVisited();
+	mEntryBlock->ReduceRecursionTempSpilling(paramMemory, gipa);
+
+	DisassembleDebug("ReduceRecursionTempSpilling");
+}
+
 void InterCodeProcedure::LoadStoreForwarding(InterMemory paramMemory)
 {
 	DisassembleDebug("Load/Store forwardingY");
@@ -19931,7 +20060,7 @@ void InterCodeProcedure::Close(void)
 {
 	GrowingTypeArray	tstack(IT_NONE);
 
-	CheckFunc = !strcmp(mIdent->mString, "KeyExpansion");
+	CheckFunc = !strcmp(mIdent->mString, "main");
 	CheckCase = false;
 
 	mEntryBlock = mBlocks[0];
@@ -20822,6 +20951,9 @@ void InterCodeProcedure::Close(void)
 
 	DisassembleDebug("Reduced Temporaries");
 
+	if (!mFastCallProcedure)
+		ReduceRecursionTempSpilling(paramMemory);
+
 	// Optimize for size
 
 	MergeBasicBlocks();

oscar64/InterCode.h

@@ -472,6 +472,7 @@ public:
 
 	void LinkerObjectForwarding(const GrowingInstructionPtrArray& tvalue);
 	bool LoadStoreForwarding(const GrowingInstructionPtrArray& tvalue, const GrowingVariableArray& staticVars);
+	void ReduceRecursionTempSpilling(InterMemory paramMemory, const GrowingInstructionPtrArray& tvalue);
 
 	void LocalRenameRegister(const GrowingIntArray& renameTable, int& num);
 	void BuildGlobalRenameRegisterTable(const GrowingIntArray& renameTable, GrowingIntArray& globalRenameTable);
@@ -715,6 +716,7 @@ protected:
 	void MergeIndexedLoadStore(void);
 	void EliminateAliasValues();
 	void LoadStoreForwarding(InterMemory paramMemory);
+	void ReduceRecursionTempSpilling(InterMemory paramMemory);
 	void ExpandSelect(void);
 	void PropagateConstOperationsUp(void);
 	void RebuildIntegerRangeSet(void);
@@ -731,6 +733,7 @@ protected:
 	void PropagateMemoryAliasingInfo(void);
 	void MoveConditionsOutOfLoop(void);
 
+
 	void PeepholeOptimization(void);
 
 	void CheckFinal(void);