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

Optimize div/mod unsigned pairs

This commit is contained in:
drmortalwombat 2023-11-03 20:16:23 +01:00
parent d9c84e3a15
commit 8ea991db81
4 changed files with 415 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
include/stdio.c
View File

@ -279,10 +279,7 @@ int nformi(const sinfo * si, char * str, int v, bool s)
while (u > 0) while (u > 0)
{ {
int c = u % si->base; int c = u % si->base;
if (c >= 10) c += c >= 10 ? 'A' - 10 : '0';
c += 'A' - 10;
else
c += '0';
buffer[--i] = c; buffer[--i] = c;
u /= si->base; u /= si->base;
} }
@ -340,10 +337,7 @@ int nforml(const sinfo * si, char * str, long v, bool s)
while (u > 0) while (u > 0)
{ {
int c = u % si->base; int c = u % si->base;
if (c >= 10) c += c >= 10 ? 'A' - 10 : '0';
c += 'A' - 10;
else
c += '0';
buffer[--i] = c; buffer[--i] = c;
u /= si->base; u /= si->base;
} }

82
oscar64/InterCode.cpp
View File

@ -8022,6 +8022,11 @@ void InterCodeBasicBlock::UpdateLocalIntegerRangeSets(const GrowingVariableArray
} }
} }
} }
else if (s0 >= 0)
{
mTrueValueRange[s0].LimitMin(mInstructions[sz - 2]->mSrc[1].mIntConst + 1);
mFalseValueRange[s0].LimitMax(mInstructions[sz - 2]->mSrc[1].mIntConst);
}
break; break;
case IA_CMPLEU: case IA_CMPLEU:
if (s0 < 0) if (s0 < 0)
@ -10827,6 +10832,10 @@ bool InterCodeBasicBlock::LoadStoreForwarding(const GrowingInstructionPtrArray&
} }
#if 1 #if 1
int sz = mInstructions.Size() - 1;
if (sz > 2 && mInstructions[sz]->mCode == IC_BRANCH && mInstructions[sz - 1]->mCode == IC_RELATIONAL_OPERATOR)
sz--;
// move loads down as far as possible to avoid false aliasing // move loads down as far as possible to avoid false aliasing
for (int i = mInstructions.Size() - 2; i >= 0; i--) for (int i = mInstructions.Size() - 2; i >= 0; i--)
{ {
@ -10834,7 +10843,7 @@ bool InterCodeBasicBlock::LoadStoreForwarding(const GrowingInstructionPtrArray&
if (ins->mCode == IC_LOAD) if (ins->mCode == IC_LOAD)
{ {
int j = i; int j = i;
while (j + 1 < mInstructions.Size() && CanSwapInstructions(ins, mInstructions[j + 1])) while (j + 1 < sz && CanSwapInstructions(ins, mInstructions[j + 1]))
{ {
SwapInstructions(ins, mInstructions[j + 1]); SwapInstructions(ins, mInstructions[j + 1]);
mInstructions[j] = mInstructions[j + 1]; mInstructions[j] = mInstructions[j + 1];
@ -11904,9 +11913,30 @@ bool InterCodeBasicBlock::PushSinglePathResultInstructions(void)
InterInstruction* ins(mInstructions[i]); InterInstruction* ins(mInstructions[i]);
int dtemp = ins->mDst.mTemp; int dtemp = ins->mDst.mTemp;
bool moved = false; bool moved = false, pair = false;
if (dtemp >= 0 && !providedTemps[dtemp] && !requiredTemps[dtemp]) if (i > 0 && ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_DIVU && i > 0 &&
mInstructions[i - 1]->mCode == IC_BINARY_OPERATOR && mInstructions[i - 1]->mOperator == IA_MODU &&
ins->mSrc[0].IsEqual(mInstructions[i - 1]->mSrc[0]) &&
ins->mSrc[1].IsEqual(mInstructions[i - 1]->mSrc[1]))
pair = true;
else if (i > 0 && ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_MODU && i > 0 &&
mInstructions[i - 1]->mCode == IC_BINARY_OPERATOR && mInstructions[i - 1]->mOperator == IA_DIVU &&
ins->mSrc[0].IsEqual(mInstructions[i - 1]->mSrc[0]) &&
ins->mSrc[1].IsEqual(mInstructions[i - 1]->mSrc[1]))
pair = true;
else if (i + 1 < mInstructions.Size() && ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_DIVU &&
mInstructions[i + 1]->mCode == IC_BINARY_OPERATOR && mInstructions[i + 1]->mOperator == IA_MODU &&
ins->mSrc[0].IsEqual(mInstructions[i + 1]->mSrc[0]) &&
ins->mSrc[1].IsEqual(mInstructions[i + 1]->mSrc[1]))
pair = true;
else if (i + 1 < mInstructions.Size() && ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_MODU &&
mInstructions[i + 1]->mCode == IC_BINARY_OPERATOR && mInstructions[i + 1]->mOperator == IA_DIVU &&
ins->mSrc[0].IsEqual(mInstructions[i + 1]->mSrc[0]) &&
ins->mSrc[1].IsEqual(mInstructions[i + 1]->mSrc[1]))
pair = true;
if (!pair && dtemp >= 0 && !providedTemps[dtemp] && !requiredTemps[dtemp])
{ {
int j = 0; int j = 0;
while (j < ins->mNumOperands && (ins->mSrc[j].mTemp < 0 || !(providedTemps[ins->mSrc[j].mTemp] || IsTempModifiedOnPath(ins->mSrc[j].mTemp, i + 1)))) while (j < ins->mNumOperands && (ins->mSrc[j].mTemp < 0 || !(providedTemps[ins->mSrc[j].mTemp] || IsTempModifiedOnPath(ins->mSrc[j].mTemp, i + 1))))
@ -13000,7 +13030,7 @@ bool InterCodeBasicBlock::SingleTailLoopOptimization(const NumberSet& aliasedPar
} }
} }
else if (ai->mCode == IC_BINARY_OPERATOR && ai->mOperator == IA_ADD && ai->mSrc[0].mTemp < 0 && ai->mDst.mTemp == ai->mSrc[1].mTemp && ai->mSrc[0].mIntConst == 1 && IsIntegerType(ai->mDst.mType) && else if (ai->mCode == IC_BINARY_OPERATOR && ai->mOperator == IA_ADD && ai->mSrc[0].mTemp < 0 && ai->mDst.mTemp == ai->mSrc[1].mTemp && ai->mSrc[0].mIntConst == 1 && IsIntegerType(ai->mDst.mType) &&
ci->mCode == IC_RELATIONAL_OPERATOR && ci->mOperator == IA_CMPLU && ci->mSrc[0].mTemp >= 0 && ci->mSrc[0].IsUnsigned() && ci->mSrc[1].mTemp == ai->mDst.mTemp && ci->mCode == IC_RELATIONAL_OPERATOR && ci->mOperator == IA_CMPLU && ci->mSrc[0].mTemp >= 0 && ci->mSrc[0].IsPositive() && ci->mSrc[1].mTemp == ai->mDst.mTemp &&
bi->mCode == IC_BRANCH && bi->mSrc[0].mTemp == ci->mDst.mTemp && !post->mEntryRequiredTemps[ai->mDst.mTemp] && bi->mCode == IC_BRANCH && bi->mSrc[0].mTemp == ci->mDst.mTemp && !post->mEntryRequiredTemps[ai->mDst.mTemp] &&
!tail->IsTempReferencedInRange(0, tz - 3, ai->mDst.mTemp) && !tail->IsTempModifiedInRange(0, tz - 3, ci->mSrc[0].mTemp)) !tail->IsTempReferencedInRange(0, tz - 3, ai->mDst.mTemp) && !tail->IsTempModifiedInRange(0, tz - 3, ci->mSrc[0].mTemp))
{ {
@ -16950,6 +16980,48 @@ void InterCodeBasicBlock::PeepholeOptimization(const GrowingVariableArray& stati
} while (changed); } while (changed);
// move div up to mod
int imod = -1, idiv = -1;
for (int i = 0; i < mInstructions.Size(); i++)
{
InterInstruction* ins = mInstructions[i];
if (ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_MODU)
{
imod = -1;
if (idiv >= 0 && ins->mSrc[0].IsEqual(mInstructions[idiv]->mSrc[0]) && ins->mSrc[1].IsEqual(mInstructions[idiv]->mSrc[1]))
{
int j = i - 1;
while (j > idiv && CanSwapInstructions(mInstructions[j], ins))
{
SwapInstructions(mInstructions[j], ins);
mInstructions[j + 1] = mInstructions[j];
j--;
}
mInstructions[j + 1] = ins;
}
else
imod = i;
}
else if (ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_DIVU)
{
idiv = -1;
if (imod >= 0 && ins->mSrc[0].IsEqual(mInstructions[imod]->mSrc[0]) && ins->mSrc[1].IsEqual(mInstructions[imod]->mSrc[1]))
{
int j = i - 1;
while (j > imod && CanSwapInstructions(mInstructions[j], ins))
{
SwapInstructions(mInstructions[j + 0], ins);
mInstructions[j + 1] = mInstructions[j];
j--;
}
mInstructions[j + 1] = ins;
}
else
idiv = i;
}
}
// Check case of cmp signed immediate // Check case of cmp signed immediate
if (mFalseJump && mInstructions.Size() > 3) if (mFalseJump && mInstructions.Size() > 3)
{ {
@ -18436,7 +18508,7 @@ void InterCodeProcedure::Close(void)
{ {
GrowingTypeArray tstack(IT_NONE); GrowingTypeArray tstack(IT_NONE);
CheckFunc = !strcmp(mIdent->mString, "game_menu"); CheckFunc = !strcmp(mIdent->mString, "nformi");
CheckCase = false; CheckCase = false;
mEntryBlock = mBlocks[0]; mEntryBlock = mBlocks[0];

334
oscar64/NativeCodeGenerator.cpp
View File

@ -8352,6 +8352,109 @@ void NativeCodeBasicBlock::AddAsrSignedByte(InterCodeProcedure* proc, const Inte
} }
void NativeCodeBasicBlock::BinaryDivModPair(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction* ins1, const InterInstruction* ins2)
{
if (ins1->mSrc[1].mTemp < 0)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, ins1->mSrc[1].mIntConst & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 0));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, (ins1->mSrc[1].mIntConst >> 8) & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
if (ins1->mDst.mType == IT_INT32)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, (ins1->mSrc[1].mIntConst >> 16) & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, (ins1->mSrc[1].mIntConst >> 24) & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
}
}
else
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[1].mTemp]));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 0));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[1].mTemp] + 1));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
if (ins1->mDst.mType == IT_INT32)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[1].mTemp] + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[1].mTemp] + 3));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
}
}
if (ins1->mSrc[0].mTemp < 0)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, ins1->mSrc[0].mIntConst & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 0));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, (ins1->mSrc[0].mIntConst >> 8) & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 1));
if (ins1->mDst.mType == IT_INT32)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, (ins1->mSrc[0].mIntConst >> 16) & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_IMMEDIATE, (ins1->mSrc[0].mIntConst >> 24) & 0xff));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 3));
}
}
else
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[0].mTemp]));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 0));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[0].mTemp] + 1));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 1));
if (ins1->mDst.mType == IT_INT32)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[0].mTemp] + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mSrc[0].mTemp] + 3));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 3));
}
}
if (ins1->mDst.mType == IT_INT32)
{
NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("divu32")));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_JSR, ASMIM_ABSOLUTE, frt.mOffset, frt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER));
}
else
{
NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("divu16")));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_JSR, ASMIM_ABSOLUTE, frt.mOffset, frt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER));
}
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 0));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mDst.mTemp]));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 1));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mDst.mTemp] + 1));
if (ins1->mDst.mType == IT_INT32)
{
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mDst.mTemp] + 2));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
mIns.Push(NativeCodeInstruction(ins1, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins1->mDst.mTemp] + 3));
}
if (ins2->mDst.mType == IT_INT32)
{
mIns.Push(NativeCodeInstruction(ins2, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 4));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins2->mDst.mTemp]));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 5));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins2->mDst.mTemp] + 1));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 6));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins2->mDst.mTemp] + 2));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 7));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins2->mDst.mTemp] + 3));
}
else
{
mIns.Push(NativeCodeInstruction(ins2, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 2));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins2->mDst.mTemp]));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 3));
mIns.Push(NativeCodeInstruction(ins2, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins2->mDst.mTemp] + 1));
}
}
NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins, const InterInstruction * sins1, const InterInstruction * sins0) NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins, const InterInstruction * sins1, const InterInstruction * sins0)
{ {
int treg = BC_REG_TMP + proc->mTempOffset[ins->mDst.mTemp]; int treg = BC_REG_TMP + proc->mTempOffset[ins->mDst.mTemp];
@ -16107,6 +16210,116 @@ bool NativeCodeBasicBlock::Split16BitLoopCount(NativeCodeProcedure* proc)
return changed; return changed;
} }
bool NativeCodeBasicBlock::LoopRegisterWrapAround(void)
{
bool changed = false;
if (!mVisited)
{
mVisited = true;
if (mLoopHead && mNumEntries == 2)
{
NativeCodeBasicBlock* eblock = nullptr, * cblock = nullptr, * bblock = nullptr;
if (mEntryBlocks[0]->mFalseJump)
{
eblock = mEntryBlocks[0];
bblock = mEntryBlocks[1];
}
else if (mEntryBlocks[1]->mFalseJump)
{
eblock = mEntryBlocks[1];
bblock = mEntryBlocks[0];
}
if (eblock && eblock != this && !bblock->mFalseJump)
{
if (eblock->mFalseJump == this)
cblock = eblock->mTrueJump;
else
cblock = eblock->mFalseJump;
int i = eblock->mIns.Size() - 1;
while (i >= 0 && !eblock->mIns[i].ChangesXReg())
i--;
if (i >= 0 && eblock->mIns[i].mType == ASMIT_LDX && eblock->mIns[i].mMode == ASMIM_ZERO_PAGE)
{
int addr = eblock->mIns[i].mAddress;
int j = 0;
while (j < mIns.Size() && !mIns[j].ReferencesXReg() && !mIns[j].ChangesZeroPage(addr))
j++;
if (j < mIns.Size())
{
if (mIns[j].mType == ASMIT_STX && mIns[j].mMode == ASMIM_ZERO_PAGE && mIns[j].mAddress == addr)
{
bblock->mIns.Push(mIns[j]);
bblock->mExitRequiredRegs += addr;
mEntryRequiredRegs += addr;
mIns.Remove(j);
changed = true;
}
}
}
i = eblock->mIns.Size() - 1;
while (i >= 0 && !eblock->mIns[i].ChangesYReg())
i--;
if (i >= 0 && eblock->mIns[i].mType == ASMIT_LDY && eblock->mIns[i].mMode == ASMIM_ZERO_PAGE)
{
int addr = eblock->mIns[i].mAddress;
int j = 0;
while (j < mIns.Size() && !mIns[j].ReferencesYReg() && !mIns[j].ChangesZeroPage(addr))
j++;
if (j < mIns.Size())
{
if (mIns[j].mType == ASMIT_STY && mIns[j].mMode == ASMIM_ZERO_PAGE && mIns[j].mAddress == addr)
{
bblock->mIns.Push(mIns[j]);
bblock->mExitRequiredRegs += addr;
mEntryRequiredRegs += addr;
mIns.Remove(j);
changed = true;
}
}
}
i = eblock->mIns.Size() - 1;
while (i >= 0 && !eblock->mIns[i].ChangesAccu())
i--;
if (i >= 0 && eblock->mIns[i].mType == ASMIT_LDA && eblock->mIns[i].mMode == ASMIM_ZERO_PAGE)
{
int addr = eblock->mIns[i].mAddress;
int j = 0;
while (j < mIns.Size() && !mIns[j].ReferencesAccu() && !mIns[j].ChangesZeroPage(addr))
j++;
if (j < mIns.Size())
{
if (mIns[j].mType == ASMIT_STA && mIns[j].mMode == ASMIM_ZERO_PAGE && mIns[j].mAddress == addr)
{
bblock->mIns.Push(mIns[j]);
bblock->mExitRequiredRegs += addr;
mEntryRequiredRegs += addr;
mIns.Remove(j);
changed = true;
}
}
}
}
}
if (mTrueJump && mTrueJump->LoopRegisterWrapAround())
changed = true;
if (mFalseJump && mFalseJump->LoopRegisterWrapAround())
changed = true;
}
return changed;
}
bool NativeCodeBasicBlock::EliminateDeadLoops(void) bool NativeCodeBasicBlock::EliminateDeadLoops(void)
{ {
bool changed = false; bool changed = false;
@ -24541,6 +24754,69 @@ bool NativeCodeBasicBlock::FindGlobalAddressSumY(int at, int reg, bool direct, i
return false; return false;
} }
bool NativeCodeBasicBlock::FindSharedGlobalAddressSumY(int at, int reg, const NativeCodeInstruction*& ains, const NativeCodeInstruction*& iins)
{
int j = at - 7;
while (j >= 0)
{
if (mIns[j + 0].mType == ASMIT_CLC &&
mIns[j + 1].mType == ASMIT_LDA && mIns[j + 1].mMode == ASMIM_IMMEDIATE_ADDRESS && (mIns[j + 1].mFlags & NCIF_LOWER) && mIns[j + 1].mLinkerObject &&
mIns[j + 2].mType == ASMIT_ADC && mIns[j + 2].mMode == ASMIM_ZERO_PAGE &&
mIns[j + 3].mType == ASMIT_STA && mIns[j + 3].mMode == ASMIM_ZERO_PAGE && mIns[j + 3].mAddress == reg &&
mIns[j + 4].mType == ASMIT_LDA && mIns[j + 4].mMode == ASMIM_IMMEDIATE_ADDRESS && (mIns[j + 4].mFlags & NCIF_UPPER) && mIns[j + 4].mLinkerObject == mIns[j + 1].mLinkerObject &&
mIns[j + 5].mType == ASMIT_ADC && mIns[j + 5].mMode == ASMIM_IMMEDIATE && mIns[j + 5].mAddress == 0 &&
mIns[j + 6].mType == ASMIT_STA && mIns[j + 6].mMode == ASMIM_ZERO_PAGE && mIns[j + 6].mAddress == reg + 1)
{
ains = &(mIns[j + 1]);
iins = &(mIns[j + 2]);
int ireg = iins->mAddress;
if (reg == ireg)
return false;
if (ChangesZeroPage(ireg, j + 7, at))
return false;
return true;
}
if (mIns[j + 6].ChangesZeroPage(reg) || mIns[j + 6].ChangesZeroPage(reg + 1))
return false;
j--;
}
if (mLoopHead)
return false;
if (mEntryBlocks.Size() > 0)
{
while (j >= -6)
{
if (mIns[j + 6].ChangesZeroPage(reg) || mIns[j + 6].ChangesZeroPage(reg + 1))
return false;
j--;
}
if (mEntryBlocks[0]->FindSharedGlobalAddressSumY(mEntryBlocks[0]->mIns.Size(), reg, ains, iins))
{
if (ChangesZeroPage(iins->mAddress, 0, at))
return false;
const NativeCodeInstruction* ains0;
const NativeCodeInstruction* iins0;
int i = 1;
while (i < mEntryBlocks.Size() && mEntryBlocks[i]->FindSharedGlobalAddressSumY(mEntryBlocks[i]->mIns.Size(), reg, ains0, iins0) && ains0->IsSame(*ains) && iins0->IsSame(*iins))
i++;
if (i == mEntryBlocks.Size())
return true;
}
}
return false;
}
bool NativeCodeBasicBlock::JoinTAXARange(int from, int to) bool NativeCodeBasicBlock::JoinTAXARange(int from, int to)
{ {
int start = from; int start = from;
@ -30656,7 +30932,7 @@ bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc
exitBlock->mIns[1].mType = ASMIT_STX; exitBlock->mIns[1].mType = ASMIT_STX;
} }
mIns.Remove(i); mIns.Remove(i); mIns.Remove(i + 1);
mIns.Remove(0); mIns.Remove(0); mIns.Remove(0); mIns.Remove(0);
CheckLive(); CheckLive();
@ -33336,6 +33612,7 @@ bool NativeCodeBasicBlock::OptimizeFindLoop(NativeCodeProcedure* proc)
body->mExitRequiredRegs += CPU_REG_Y; body->mExitRequiredRegs += CPU_REG_Y;
mEntryRequiredRegs += CPU_REG_Y; mEntryRequiredRegs += CPU_REG_Y;
mExitRequiredRegs += CPU_REG_Y; mExitRequiredRegs += CPU_REG_Y;
succ->mEntryRequiredRegs += CPU_REG_Y;
for (int i = 0; i < mIns.Size(); i++) for (int i = 0; i < mIns.Size(); i++)
mIns[i].mLive |= LIVE_CPU_REG_Y; mIns[i].mLive |= LIVE_CPU_REG_Y;
@ -36003,6 +36280,28 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
#endif #endif
} }
else if (FindSharedGlobalAddressSumY(i, sreg, ains, iins))
{
#if 1
if (mIns[i + 0].mLive & LIVE_CPU_REG_Y)
{
InsertLoadYImmediate(mIns[i + 0].mIns, i + 1, 0);
// mIns.Insert(i + 1, NativeCodeInstruction(ASMIT_LDY, ASMIM_IMMEDIATE, 0));
// mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
}
mIns.Insert(i + 0, NativeCodeInstruction(mIns[i + 0].mIns, ASMIT_LDY, *iins));
mIns[i + 0].mLive |= LIVE_CPU_REG_Y | LIVE_MEM;
mIns[i + 1].mMode = ASMIM_ABSOLUTE_Y;
mIns[i + 1].mLinkerObject = ains->mLinkerObject;
mIns[i + 1].mAddress = ains->mAddress;
mIns[i + 1].mFlags &= ~NCIF_YZERO;
progress = true;
CheckLive();
#endif
}
#if 1 #if 1
if (mIns[i + 0].mMode == ASMIM_INDIRECT_Y && (mIns[i + 0].mFlags & NCIF_YZERO) && !(mIns[i + 0].mLive & LIVE_CPU_REG_X)) if (mIns[i + 0].mMode == ASMIM_INDIRECT_Y && (mIns[i + 0].mFlags & NCIF_YZERO) && !(mIns[i + 0].mLive & LIVE_CPU_REG_X))
@ -43024,6 +43323,11 @@ void NativeCodeProcedure::Optimize(void)
changed = true; changed = true;
#endif #endif
#if _DEBUG
ResetVisited();
mEntryBlock->CheckBlocks(true);
#endif
#if 1 #if 1
ResetVisited(); ResetVisited();
if (mEntryBlock->OptimizeFindLoop(this)) if (mEntryBlock->OptimizeFindLoop(this))
@ -43286,6 +43590,14 @@ void NativeCodeProcedure::Optimize(void)
changed = true; changed = true;
} }
if (step == 8)
{
ResetVisited();
if (mEntryBlock->LoopRegisterWrapAround())
changed = true;
}
if (step >= 9) if (step >= 9)
{ {
ResetVisited(); ResetVisited();
@ -43820,6 +44132,26 @@ void NativeCodeProcedure::CompileInterBlock(InterCodeProcedure* iproc, InterCode
block->AddAsrSignedByte(iproc, ins, iblock->mInstructions[i + 1]); block->AddAsrSignedByte(iproc, ins, iblock->mInstructions[i + 1]);
i ++; i ++;
} }
else if (i + 1 < iblock->mInstructions.Size() &&
ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_DIVU &&
iblock->mInstructions[i + 1]->mOperator == IA_MODU &&
ins->mSrc[0].IsEqual(iblock->mInstructions[i + 1]->mSrc[0]) &&
ins->mSrc[1].IsEqual(iblock->mInstructions[i + 1]->mSrc[1]) &&
ins->mSrc[0].mTemp != ins->mDst.mTemp && ins->mSrc[1].mTemp != ins->mDst.mTemp)
{
block->BinaryDivModPair(iproc, this, ins, iblock->mInstructions[i + 1]);
i++;
}
else if (i + 1 < iblock->mInstructions.Size() &&
ins->mCode == IC_BINARY_OPERATOR && ins->mOperator == IA_MODU &&
iblock->mInstructions[i + 1]->mOperator == IA_DIVU &&
ins->mSrc[0].IsEqual(iblock->mInstructions[i + 1]->mSrc[0]) &&
ins->mSrc[1].IsEqual(iblock->mInstructions[i + 1]->mSrc[1]) &&
ins->mSrc[0].mTemp != ins->mDst.mTemp && ins->mSrc[1].mTemp != ins->mDst.mTemp)
{
block->BinaryDivModPair(iproc, this, iblock->mInstructions[i + 1], ins);
i++;
}
else else
block = block->BinaryOperator(iproc, this, ins, nullptr, nullptr); block = block->BinaryOperator(iproc, this, ins, nullptr, nullptr);
break; break;

3
oscar64/NativeCodeGenerator.h
View File

@ -343,6 +343,7 @@ public:
void LoadEffectiveAddress(InterCodeProcedure* proc, const InterInstruction * ins, const InterInstruction* sins1, const InterInstruction* sins0, bool addrvalid); void LoadEffectiveAddress(InterCodeProcedure* proc, const InterInstruction * ins, const InterInstruction* sins1, const InterInstruction* sins0, bool addrvalid);
void LoadStoreOpAbsolute2D(InterCodeProcedure* proc, const InterInstruction* lins1, const InterInstruction* lins2, const InterInstruction* mins); void LoadStoreOpAbsolute2D(InterCodeProcedure* proc, const InterInstruction* lins1, const InterInstruction* lins2, const InterInstruction* mins);
void SignExtendAddImmediate(InterCodeProcedure* proc, const InterInstruction* xins, const InterInstruction* ains); void SignExtendAddImmediate(InterCodeProcedure* proc, const InterInstruction* xins, const InterInstruction* ains);
void BinaryDivModPair(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction* ins1, const InterInstruction* ins2);
void NumericConversion(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins); void NumericConversion(InterCodeProcedure* proc, NativeCodeProcedure* nproc, const InterInstruction * ins);
NativeCodeBasicBlock * CopyValue(InterCodeProcedure* proc, const InterInstruction * ins, NativeCodeProcedure* nproc); NativeCodeBasicBlock * CopyValue(InterCodeProcedure* proc, const InterInstruction * ins, NativeCodeProcedure* nproc);
@ -429,6 +430,7 @@ public:
bool FindGlobalAddress(int at, int reg, int& apos); bool FindGlobalAddress(int at, int reg, int& apos);
bool FindGlobalAddressSumY(int at, int reg, bool direct, int& apos, const NativeCodeInstruction * & ains, const NativeCodeInstruction*& iins, uint32 & flags, int & addr); bool FindGlobalAddressSumY(int at, int reg, bool direct, int& apos, const NativeCodeInstruction * & ains, const NativeCodeInstruction*& iins, uint32 & flags, int & addr);
bool FindExternAddressSumY(int at, int reg, int& breg, int& ireg); bool FindExternAddressSumY(int at, int reg, int& breg, int& ireg);
bool FindSharedGlobalAddressSumY(int at, int reg, const NativeCodeInstruction*& ains, const NativeCodeInstruction*& iins);
bool FindPageStartAddress(int at, int reg, int& addr); bool FindPageStartAddress(int at, int reg, int& addr);
bool FindBypassAddressSumY(int at, int reg, int& apos, int& breg); bool FindBypassAddressSumY(int at, int reg, int& apos, int& breg);
bool PatchBypassAddressSumY(int at, int reg, int apos, int breg); bool PatchBypassAddressSumY(int at, int reg, int apos, int breg);
@ -588,6 +590,7 @@ public:
bool SimplifyLoopEnd(NativeCodeProcedure* proc); bool SimplifyLoopEnd(NativeCodeProcedure* proc);
bool CrossBlockStoreLoadBypass(NativeCodeProcedure* proc); bool CrossBlockStoreLoadBypass(NativeCodeProcedure* proc);
bool EliminateDeadLoops(void); bool EliminateDeadLoops(void);
bool LoopRegisterWrapAround(void);
bool CanBytepassLoad(const NativeCodeInstruction& ains, int from = 0) const; bool CanBytepassLoad(const NativeCodeInstruction& ains, int from = 0) const;
bool CanHoistStore(const NativeCodeInstruction& ains) const; bool CanHoistStore(const NativeCodeInstruction& ains) const;