frederik
/
oscar64
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix data dependency lost due to instruction reshuffling

This commit is contained in:
drmortalwombat 2022-06-22 22:09:32 +02:00
parent 002c10ad13
commit 2eeef2d71c
3 changed files with 101 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

113
oscar64/InterCode.cpp
View File

@ -4583,6 +4583,30 @@ bool InterCodeBasicBlock::PropagateNonLocalUsedConstTemps(void)
return changed; return changed;
} }
void InterCodeBasicBlock::CollectAllUsedDefinedTemps(NumberSet& defined, NumberSet& used)
{
if (!mVisited)
{
mVisited = true;
for (int i = 0; i < mInstructions.Size(); i++)
{
InterInstruction* ins(mInstructions[i]);
if (ins->mDst.mTemp >= 0)
defined += ins->mDst.mTemp;
for (int j = 0; j < ins->mNumOperands; j++)
{
if (ins->mSrc[j].mTemp >= 0)
used += ins->mSrc[j].mTemp;
}
}
if (mTrueJump) mTrueJump->CollectAllUsedDefinedTemps(defined, used);
if (mFalseJump) mFalseJump->CollectAllUsedDefinedTemps(defined, used);
}
}
void InterCodeBasicBlock::CollectLocalUsedTemps(int numTemps) void InterCodeBasicBlock::CollectLocalUsedTemps(int numTemps)
{ {
if (!mVisited) if (!mVisited)
@ -9995,6 +10019,24 @@ void InterCodeBasicBlock::CompactInstructions(void)
} }
} }
static void SwapInstructions(InterInstruction* it, InterInstruction* ib)
{
for (int i = 0; i < ib->mNumOperands; i++)
{
if (ib->mSrc[i].mTemp >= 0 && ib->mSrc[i].mFinal)
{
for (int j = 0; j < it->mNumOperands; j++)
{
if (it->mSrc[j].mTemp == ib->mSrc[i].mTemp)
{
it->mSrc[j].mFinal = true;
ib->mSrc[i].mFinal = false;
}
}
}
}
}
void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& staticVars) void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& staticVars)
{ {
int i; int i;
@ -10070,6 +10112,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j < limit && CanBypass(ins, mInstructions[j + 1])) while (j < limit && CanBypass(ins, mInstructions[j + 1]))
{ {
SwapInstructions(ins, mInstructions[j + 1]);
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
j++; j++;
} }
@ -10098,6 +10141,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j < limit && CanBypassLoad(ins, mInstructions[j + 1])) while (j < limit && CanBypassLoad(ins, mInstructions[j + 1]))
{ {
SwapInstructions(ins, mInstructions[j + 1]);
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
j++; j++;
} }
@ -10120,6 +10164,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j < k) while (j < k)
{ {
SwapInstructions(ins, mInstructions[j + 1]);
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
j++; j++;
} }
@ -10133,6 +10178,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j < limit && CanBypass(ins, mInstructions[j + 1])) while (j < limit && CanBypass(ins, mInstructions[j + 1]))
{ {
SwapInstructions(ins, mInstructions[j + 1]);
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
j++; j++;
} }
@ -10187,6 +10233,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j > 0 && CanBypassStore(ins, mInstructions[j - 1])) while (j > 0 && CanBypassStore(ins, mInstructions[j - 1]))
{ {
SwapInstructions(mInstructions[j - 1], ins);
mInstructions[j] = mInstructions[j - 1]; mInstructions[j] = mInstructions[j - 1];
j--; j--;
} }
@ -10199,6 +10246,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j > 0 && CanBypassUp(ins, mInstructions[j - 1])) while (j > 0 && CanBypassUp(ins, mInstructions[j - 1]))
{ {
SwapInstructions(mInstructions[j - 1], ins);
mInstructions[j] = mInstructions[j - 1]; mInstructions[j] = mInstructions[j - 1];
j--; j--;
} }
@ -10211,17 +10259,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j > 0 && CanBypassLoadUp(ins, mInstructions[j - 1])) while (j > 0 && CanBypassLoadUp(ins, mInstructions[j - 1]))
{ {
if (ins->mSrc[0].mFinal) SwapInstructions(mInstructions[j - 1], ins);
{
for (int k = 0; k < mInstructions[j - 1]->mNumOperands; k++)
{
if (mInstructions[j - 1]->mSrc[k].mTemp == ins->mSrc[0].mTemp)
{
mInstructions[j - 1]->mSrc[k].mFinal = true;
ins->mSrc[0].mFinal = false;
}
}
}
mInstructions[j] = mInstructions[j - 1]; mInstructions[j] = mInstructions[j - 1];
j--; j--;
} }
@ -10244,18 +10282,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j < limit && CanBypassLoad(ins, mInstructions[j + 1])) while (j < limit && CanBypassLoad(ins, mInstructions[j + 1]))
{ {
if (!ins->mSrc[0].mFinal) SwapInstructions(ins, mInstructions[j + 1]);
{
for (int k = 0; k < mInstructions[j + 1]->mNumOperands; k++)
{
if (mInstructions[j + 1]->mSrc[k].mTemp == ins->mSrc[0].mTemp && mInstructions[j + 1]->mSrc[k].mFinal)
{
mInstructions[j + 1]->mSrc[k].mFinal = false;
ins->mSrc[0].mFinal = true;
}
}
}
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
j++; j++;
} }
@ -10268,6 +10295,7 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
int j = i; int j = i;
while (j < limit && CanBypass(ins, mInstructions[j + 1])) while (j < limit && CanBypass(ins, mInstructions[j + 1]))
{ {
SwapInstructions(ins, mInstructions[j + 1]);
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
j++; j++;
} }
@ -10323,6 +10351,11 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
mInstructions[i + 1]->mCode == IC_LOAD_TEMPORARY && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i]->mDst.mTemp && mInstructions[i + 1]->mCode == IC_LOAD_TEMPORARY && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i]->mDst.mTemp &&
(mInstructions[i + 2]->mCode == IC_RELATIONAL_OPERATOR || mInstructions[i + 2]->mCode == IC_BINARY_OPERATOR) && mInstructions[i + 2]->mSrc[0].mTemp == mInstructions[i]->mDst.mTemp && mInstructions[i + 2]->mSrc[0].mFinal) (mInstructions[i + 2]->mCode == IC_RELATIONAL_OPERATOR || mInstructions[i + 2]->mCode == IC_BINARY_OPERATOR) && mInstructions[i + 2]->mSrc[0].mTemp == mInstructions[i]->mDst.mTemp && mInstructions[i + 2]->mSrc[0].mFinal)
{ {
#if _DEBUG
for (int j = i + 3; j < mInstructions.Size(); j++)
assert(!mInstructions[j]->ReferencesTemp(mInstructions[i]->mDst.mTemp));
#endif
int t = mInstructions[i + 0]->mDst.mTemp; int t = mInstructions[i + 0]->mDst.mTemp;
mInstructions[i + 0]->mDst.mTemp = mInstructions[i + 1]->mDst.mTemp; mInstructions[i + 0]->mDst.mTemp = mInstructions[i + 1]->mDst.mTemp;
mInstructions[i + 1]->mCode = IC_NONE; mInstructions[i + 1]->mCode = IC_NONE;
@ -10339,6 +10372,11 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
mInstructions[i + 1]->mCode == IC_LOAD_TEMPORARY && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i]->mDst.mTemp && mInstructions[i + 1]->mCode == IC_LOAD_TEMPORARY && mInstructions[i + 1]->mSrc[0].mTemp == mInstructions[i]->mDst.mTemp &&
(mInstructions[i + 2]->mCode == IC_RELATIONAL_OPERATOR || mInstructions[i + 2]->mCode == IC_BINARY_OPERATOR) && mInstructions[i + 2]->mSrc[1].mTemp == mInstructions[i]->mDst.mTemp && mInstructions[i + 2]->mSrc[1].mFinal) (mInstructions[i + 2]->mCode == IC_RELATIONAL_OPERATOR || mInstructions[i + 2]->mCode == IC_BINARY_OPERATOR) && mInstructions[i + 2]->mSrc[1].mTemp == mInstructions[i]->mDst.mTemp && mInstructions[i + 2]->mSrc[1].mFinal)
{ {
#if _DEBUG
for (int j = i + 3; j < mInstructions.Size(); j++)
assert(!mInstructions[j]->ReferencesTemp(mInstructions[i]->mDst.mTemp));
#endif
mInstructions[i + 0]->mDst.mTemp = mInstructions[i + 1]->mDst.mTemp; mInstructions[i + 0]->mDst.mTemp = mInstructions[i + 1]->mDst.mTemp;
mInstructions[i + 1]->mCode = IC_NONE; mInstructions[i + 1]->mCode = IC_NONE;
mInstructions[i + 2]->mSrc[1].mTemp = mInstructions[i + 1]->mDst.mTemp; mInstructions[i + 2]->mSrc[1].mTemp = mInstructions[i + 1]->mDst.mTemp;
@ -11213,10 +11251,30 @@ void InterCodeProcedure::SingleAssignmentForwarding(void)
} }
void InterCodeProcedure::CheckUsedDefinedTemps(void)
{
#if _DEBUG
int numTemps = mTemporaries.Size();
NumberSet defined(numTemps), used(numTemps);
ResetVisited();
mEntryBlock->CollectAllUsedDefinedTemps(defined, used);
for (int i = 0; i < numTemps; i++)
{
assert(!used[i] || defined[i]);
}
#endif
}
void InterCodeProcedure::TempForwarding(void) void InterCodeProcedure::TempForwarding(void)
{ {
int numTemps = mTemporaries.Size(); int numTemps = mTemporaries.Size();
CheckUsedDefinedTemps();
ValueSet valueSet; ValueSet valueSet;
FastNumberSet tvalidSet(numTemps); FastNumberSet tvalidSet(numTemps);
@ -11425,10 +11483,12 @@ void InterCodeProcedure::PromoteSimpleLocalsToTemp(InterMemory paramMemory, int
void InterCodeProcedure::SimplifyIntegerNumeric(FastNumberSet& activeSet) void InterCodeProcedure::SimplifyIntegerNumeric(FastNumberSet& activeSet)
{ {
GrowingInstructionPtrArray silvalues(nullptr); GrowingInstructionPtrArray silvalues(nullptr);
int silvused; int silvused = mTemporaries.Size();
do do
{ {
mTemporaries.SetSize(silvused, true);
BuildDataFlowSets(); BuildDataFlowSets();
TempForwarding(); TempForwarding();
@ -11677,12 +11737,17 @@ void InterCodeProcedure::Close(void)
BuildDataFlowSets(); BuildDataFlowSets();
#endif #endif
CheckUsedDefinedTemps();
SingleAssignmentForwarding(); SingleAssignmentForwarding();
CheckUsedDefinedTemps();
ResetVisited(); ResetVisited();
mEntryBlock->PeepholeOptimization(mModule->mGlobalVars); mEntryBlock->PeepholeOptimization(mModule->mGlobalVars);
DisassembleDebug("Broken Peephole");
TempForwarding(); TempForwarding();
RemoveUnusedInstructions(); RemoveUnusedInstructions();

3
oscar64/InterCode.h
View File

@ -377,6 +377,8 @@ public:
void LocalToTemp(int vindex, int temp); void LocalToTemp(int vindex, int temp);
void CollectAllUsedDefinedTemps(NumberSet& defined, NumberSet& used);
void CollectLocalAddressTemps(GrowingIntArray& localTable, GrowingIntArray& paramTable); void CollectLocalAddressTemps(GrowingIntArray& localTable, GrowingIntArray& paramTable);
void MarkAliasedLocalTemps(const GrowingIntArray& localTable, NumberSet& aliasedLocals, const GrowingIntArray& paramTable, NumberSet& aliasedParams); void MarkAliasedLocalTemps(const GrowingIntArray& localTable, NumberSet& aliasedLocals, const GrowingIntArray& paramTable, NumberSet& aliasedParams);
@ -565,6 +567,7 @@ protected:
void LoadStoreForwarding(InterMemory paramMemory); void LoadStoreForwarding(InterMemory paramMemory);
void MergeBasicBlocks(void); void MergeBasicBlocks(void);
void CheckUsedDefinedTemps(void);
void DisassembleDebug(const char* name); void DisassembleDebug(const char* name);
}; };

9
oscar64/NativeCodeGenerator.cpp
View File

@ -6269,6 +6269,15 @@ void NativeCodeBasicBlock::ShiftRegisterLeftFromByte(InterCodeProcedure* proc, i
if (shift == 0) if (shift == 0)
{ {
}
else if (shift >= 8)
{
mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, reg));
for (int i = 8; i < shift; i++)
mIns.Push(NativeCodeInstruction(ASMIT_ASL, ASMIM_IMPLIED));
mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg + 1));
mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_IMMEDIATE, 0));
mIns.Push(NativeCodeInstruction(ASMIT_STA, ASMIM_ZERO_PAGE, reg));
} }
else if (shift == 1) else if (shift == 1)
{ {