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

Loop MSB optimizations

This commit is contained in:
drmortalwombat 2022-03-22 22:21:59 +01:00
parent 8c77a5d256
commit 5f0e0225e1
6 changed files with 254 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
oscar64/InterCode.cpp
View File

@ -8379,6 +8379,7 @@ void InterCodeBasicBlock::PeepholeOptimization(void)
}
mInstructions.SetSize(j);
// shorten lifespan
int loopTmp = -1;
@ -8387,6 +8388,33 @@ void InterCodeBasicBlock::PeepholeOptimization(void)
if (limit >= 0 && mInstructions[limit]->mCode == IC_BRANCH)
{
limit--;
#if 1
// try to move conditional source down
int i = limit;
while (i >= 0 && mInstructions[i]->mDst.mTemp != mInstructions[limit + 1]->mSrc[0].mTemp)
i--;
if (i >= 0 && i != limit)
{
InterInstruction* ins(mInstructions[i]);
if (ins->mCode == IC_BINARY_OPERATOR || ins->mCode == IC_RELATIONAL_OPERATOR)
{
if (ins->mSrc[0].mTemp < 0)
{
int k = i;
while (k < limit && CanBypass(ins, mInstructions[k + 1]))
k++;
if (k == limit)
{
mInstructions.Remove(i);
mInstructions.Insert(limit, ins);
}
}
}
}
#endif
if (limit > 0 && mInstructions[limit]->mCode == IC_RELATIONAL_OPERATOR)
{
if (mInstructions[limit]->mSrc[1].mTemp)
@ -8417,6 +8445,8 @@ void InterCodeBasicBlock::PeepholeOptimization(void)
limit--;
}
}
else if (limit > 0 && mInstructions[limit]->mDst.mTemp == mInstructions[limit + 1]->mSrc[0].mTemp)
limit--;
}
else if (limit >= 0 && mInstructions[limit]->mCode == IC_JUMP)
limit --;

230
oscar64/NativeCodeGenerator.cpp
View File

@ -129,7 +129,7 @@ void NativeRegisterDataSet::Intersect(const NativeRegisterDataSet& set)
}
else if (mRegs[i].mMode == NRDM_ABSOLUTE)
{
if (set.mRegs[i].mMode != NRDM_ABSOLUTE || mRegs[i].mValue != set.mRegs[i].mValue)
if (set.mRegs[i].mMode != NRDM_ABSOLUTE || mRegs[i].mValue != set.mRegs[i].mValue || mRegs[i].mLinkerObject != set.mRegs[i].mLinkerObject)
{
mRegs[i].Reset();
changed = true;
@ -9485,7 +9485,7 @@ bool NativeCodeBasicBlock::ReduceLocalYPressure(void)
return changed;
}
bool NativeCodeBasicBlock::ForwardZpYIndex(void)
bool NativeCodeBasicBlock::ForwardZpYIndex(bool full)
{
bool changed = false;
@ -9588,6 +9588,18 @@ bool NativeCodeBasicBlock::ForwardZpYIndex(void)
mIns[i + 1].mType = ASMIT_STY;
changed = true;
}
else if (i + 2 < mIns.Size() && full &&
mIns[i + 0].mType == ASMIT_INC && mIns[i + 0].mMode == ASMIM_ZERO_PAGE && mIns[i + 0].mAddress == yreg &&
mIns[i + 1].mType == ASMIT_LDA && mIns[i + 1].mMode == ASMIM_ZERO_PAGE && mIns[i + 1].mAddress == yreg && yoffset == 0 && !(mIns[i + 1].mLive & LIVE_CPU_REG_Y))
{
for (int j = ypred; j < i; j++)
mIns[j].mLive |= LIVE_CPU_REG_Y;
mIns.Insert(i, NativeCodeInstruction(ASMIT_INY, ASMIM_IMPLIED));
mIns[i + 1].mType = ASMIT_STY; mIns[i + 1].mLive |= LIVE_CPU_REG_Y;
mIns[i + 2].mType = ASMIT_TYA; mIns[i + 2].mMode = ASMIM_IMPLIED;
yreg = -1;
changed = true;
}
else if (mIns[i].mType == ASMIT_INY)
{
yoffset = (yoffset + 1) & 255;
@ -9621,17 +9633,17 @@ bool NativeCodeBasicBlock::ForwardZpYIndex(void)
}
}
if (mTrueJump && mTrueJump->ForwardZpYIndex())
if (mTrueJump && mTrueJump->ForwardZpYIndex(full))
changed = true;
if (mFalseJump && mFalseJump->ForwardZpYIndex())
if (mFalseJump && mFalseJump->ForwardZpYIndex(full))
changed = true;
}
return changed;
}
bool NativeCodeBasicBlock::ForwardZpXIndex(void)
bool NativeCodeBasicBlock::ForwardZpXIndex(bool full)
{
bool changed = false;
@ -9752,10 +9764,10 @@ bool NativeCodeBasicBlock::ForwardZpXIndex(void)
}
}
if (mTrueJump && mTrueJump->ForwardZpXIndex())
if (mTrueJump && mTrueJump->ForwardZpXIndex(full))
changed = true;
if (mFalseJump && mFalseJump->ForwardZpXIndex())
if (mFalseJump && mFalseJump->ForwardZpXIndex(full))
changed = true;
}
@ -12815,6 +12827,65 @@ bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc
return lblock->OptimizeSimpleLoopInvariant(proc, this, eblock);
}
bool NativeCodeBasicBlock::RemoveSimpleLoopUnusedIndex(void)
{
bool changed = false;
if (!mVisited)
{
mVisited = true;
if (mFalseJump && mTrueJump == this)
{
NumberSet required(mFalseJump->mEntryRequiredRegs);
bool complex = false;
int k = mIns.Size() - 1;
while (k >= 0)
{
if (mIns[k].mType == ASMIT_STA && mIns[k].mMode == ASMIM_ZERO_PAGE && !(mIns[k].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_C | LIVE_CPU_REG_Z)) && !required[mIns[k].mAddress])
{
if (k > 2 && mIns[k - 1].mMode == ASMIM_IMMEDIATE && mIns[k - 2].mType == ASMIT_LDA && mIns[k - 2].mMode == ASMIM_ZERO_PAGE && mIns[k - 2].mAddress == mIns[k].mAddress)
k -= 2;
}
else if (mIns[k].mMode == ASMIM_ZERO_PAGE)
required += mIns[k].mAddress;
else if (mIns[k].mMode == ASMIM_INDIRECT_Y)
{
required += mIns[k].mAddress;
required += mIns[k].mAddress + 1;
}
else if (mIns[k].mType == ASMIT_JSR)
complex = true;
k--;
}
if (!complex)
{
for (int i = 0; i < mIns.Size(); i++)
{
if (mIns[i].mType == ASMIT_STA && mIns[i].mMode == ASMIM_ZERO_PAGE && !required[mIns[i].mAddress])
{
mIns[i].mType = ASMIT_NOP;
mIns[i].mMode = ASMIM_IMPLIED;
changed = true;
}
}
}
}
if (mTrueJump && mTrueJump->RemoveSimpleLoopUnusedIndex())
changed = true;
if (mFalseJump && mFalseJump->RemoveSimpleLoopUnusedIndex())
changed = true;
}
return changed;
}
bool NativeCodeBasicBlock::OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc, NativeCodeBasicBlock* prevBlock, NativeCodeBasicBlock* exitBlock)
{
bool changed = false;
@ -14747,11 +14818,55 @@ void NativeCodeBasicBlock::BlockSizeReduction(NativeCodeProcedure* proc)
}
mIns.SetSize(j);
#if 1
bool yimm = false, ximm = false;
int yval = 0, xval = 0;
bool yimm = false, ximm = false, aimm = false, afail = false;
int yval = 0, xval = 0, aval = 0;
for(int i = 0; i < mIns.Size(); i++)
{
if (mIns[i].mType == ASMIT_LDY)
if (afail)
aimm = false;
if (mIns[i].mType == ASMIT_LDA)
{
if (mIns[i].mMode == ASMIM_IMMEDIATE)
{
if (yimm && mIns[i].mAddress == yval)
{
mIns[i].mType = ASMIT_TYA;
mIns[i].mMode = ASMIM_IMPLIED;
}
else if (ximm && mIns[i].mAddress == xval)
{
mIns[i].mType = ASMIT_TXA;
mIns[i].mMode = ASMIM_IMPLIED;
}
aimm = true;
aval = mIns[i].mAddress;
}
else
aimm = false;
}
else if (mIns[i].mType == ASMIT_TYA)
{
if (yimm)
{
aimm = true;
aval = yval;
}
else
aimm = false;
}
else if (mIns[i].mType == ASMIT_TXA)
{
if (ximm)
{
aimm = true;
aval = xval;
}
else
aimm = false;
}
else if (mIns[i].mType == ASMIT_LDY)
{
if (mIns[i].mMode == ASMIM_IMMEDIATE)
{
@ -14767,6 +14882,13 @@ void NativeCodeBasicBlock::BlockSizeReduction(NativeCodeProcedure* proc)
mIns[i].mType = ASMIT_DEY;
mIns[i].mMode = ASMIM_IMPLIED;
}
else if (aimm && mIns[i].mAddress == aval)
{
mIns[i].mType = ASMIT_TAY;
mIns[i].mMode = ASMIM_IMPLIED;
yimm = true;
yval = aval;
}
else
{
yimm = true;
@ -14786,7 +14908,15 @@ void NativeCodeBasicBlock::BlockSizeReduction(NativeCodeProcedure* proc)
yval = (yval - 1) & 0xff;
}
else if (mIns[i].mType == ASMIT_TAY)
{
if (aimm)
{
yimm = true;
yval = aval;
}
else
yimm = false;
}
else if (mIns[i].mType == ASMIT_LDX)
{
if (mIns[i].mMode == ASMIM_IMMEDIATE)
@ -14803,6 +14933,13 @@ void NativeCodeBasicBlock::BlockSizeReduction(NativeCodeProcedure* proc)
mIns[i].mType = ASMIT_DEX;
mIns[i].mMode = ASMIM_IMPLIED;
}
else if (aimm && mIns[i].mAddress == aval)
{
mIns[i].mType = ASMIT_TAX;
mIns[i].mMode = ASMIM_IMPLIED;
ximm = true;
xval = aval;
}
else
{
ximm = true;
@ -14822,11 +14959,26 @@ void NativeCodeBasicBlock::BlockSizeReduction(NativeCodeProcedure* proc)
xval = (xval - 1) & 0xff;
}
else if (mIns[i].mType == ASMIT_TAX)
{
if (aimm)
{
ximm = true;
xval = aval;
}
else
ximm = false;
}
else if (mIns[i].mType == ASMIT_JSR)
yimm = ximm = false;
else if (mIns[i].mMode == ASMIM_RELATIVE && mIns[i].mAddress < 0)
yimm = ximm = false;
yimm = ximm = aimm = false;
else if (mIns[i].mMode == ASMIM_RELATIVE)
{
if (mIns[i].mAddress < 0)
yimm = ximm = aimm = false;
else
afail = true;
}
else if (mIns[i].ChangesAccu())
aimm = false;
}
#endif
@ -15617,6 +15769,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
}
#endif
#if 1
bool progress = false;
do {
progress = false;
@ -15674,7 +15827,32 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
changed = true;
}
}
#if 1
if (sz > 2 &&
mIns[sz - 3].mType == ASMIT_STA &&
mIns[sz - 2].mType == ASMIT_LDA && mIns[sz - 2].mMode == ASMIM_IMMEDIATE &&
mIns[sz - 1].mType == ASMIT_CMP && mIns[sz - 1].SameEffectiveAddress(mIns[sz - 3]) && !(mIns[sz - 1].mLive & (LIVE_CPU_REG_Z | LIVE_CPU_REG_A)))
{
if (mBranch == ASMIT_BCC && mIns[sz - 2].mAddress < 0xff)
{
mIns[sz - 3].mLive |= LIVE_CPU_REG_A;
mIns[sz - 1].mMode = ASMIM_IMMEDIATE;
mIns[sz - 1].mAddress = mIns[sz - 2].mAddress + 1;
mIns[sz - 2].mType = ASMIT_NOP; mIns[sz - 2].mMode = ASMIM_IMPLIED;
mBranch = ASMIT_BCS;
changed = true;
}
else if (mBranch == ASMIT_BCS && mIns[sz - 2].mAddress < 0xff)
{
mIns[sz - 3].mLive |= LIVE_CPU_REG_A;
mIns[sz - 1].mMode = ASMIM_IMMEDIATE;
mIns[sz - 1].mAddress = mIns[sz - 2].mAddress + 1;
mIns[sz - 2].mType = ASMIT_NOP; mIns[sz - 2].mMode = ASMIM_IMPLIED;
mBranch = ASMIT_BCC;
changed = true;
}
}
#endif
for (int i = 0; i < mIns.Size(); i++)
{
#if 1
@ -16537,15 +16715,18 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
mIns[i + 2].mType = ASMIT_CLC; mIns[i + 2].mMode = ASMIM_IMPLIED;
progress = true;
}
#if 1
else if (
mIns[i + 0].ChangesAccuAndFlag() &&
mIns[i + 1].mType == ASMIT_STA &&
mIns[i + 2].mType == ASMIT_ORA && mIns[i + 2].mMode == ASMIM_IMMEDIATE && mIns[i + 2].mAddress == 0)
{
mIns[i + 2].mType = ASMIT_NOP; mIns[i + 2].mMode = ASMIM_IMPLIED;
mIns[i + 0].mLive |= mIns[i + 2].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_Z);
mIns[i + 1].mLive |= mIns[i + 2].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_Z);
progress = true;
}
#endif
else if (
mIns[i + 0].ChangesAccuAndFlag() &&
mIns[i + 1].mType == ASMIT_STA &&
@ -17987,6 +18168,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
}
}
#endif
#if 1
if (i + 1 < mIns.Size())
{
if (
@ -18014,7 +18196,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
}
}
}
#endif
if (i + 5 < mIns.Size())
{
if (
@ -18404,6 +18586,7 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(NativeCodeProcedure* proc, int pass
} while (progress);
#endif
if (this->mTrueJump && this->mTrueJump->PeepHoleOptimizer(proc, pass))
changed = true;
@ -19431,7 +19614,14 @@ void NativeCodeProcedure::Optimize(void)
{
changed = true;
}
#if 1
if (step == 3)
{
ResetVisited();
if (mEntryBlock->RemoveSimpleLoopUnusedIndex())
changed = true;
}
#endif
if (step > 0)
{
ResetVisited();
@ -19471,11 +19661,13 @@ void NativeCodeProcedure::Optimize(void)
if (step == 3)
{
ResetVisited();
changed = mEntryBlock->OptimizeInnerLoops(this);
if (mEntryBlock->OptimizeInnerLoops(this))
changed = true;
ResetVisited();
if (mEntryBlock->ReduceLocalYPressure())
changed = true;
}
#endif
#if 1
@ -19577,11 +19769,11 @@ void NativeCodeProcedure::Optimize(void)
#if 1
ResetVisited();
if (mEntryBlock->ForwardZpYIndex())
if (mEntryBlock->ForwardZpYIndex(step >= 4))
changed = true;
ResetVisited();
if (mEntryBlock->ForwardZpXIndex())
if (mEntryBlock->ForwardZpXIndex(step >= 4))
changed = true;
#endif
if (!changed && step < 6)

5
oscar64/NativeCodeGenerator.h
View File

@ -164,6 +164,7 @@ public:
bool OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc);
bool OptimizeSimpleLoopInvariant(NativeCodeProcedure* proc, NativeCodeBasicBlock * prevBlock, NativeCodeBasicBlock* exitBlock);
bool RemoveSimpleLoopUnusedIndex(void);
bool OptimizeSimpleLoop(NativeCodeProcedure* proc);
bool SimpleLoopReversal(NativeCodeProcedure* proc);
@ -257,8 +258,8 @@ public:
bool ReplaceYRegWithXReg(int start, int end);
bool ReplaceXRegWithYReg(int start, int end);
bool ForwardZpYIndex(void);
bool ForwardZpXIndex(void);
bool ForwardZpYIndex(bool full);
bool ForwardZpXIndex(bool full);
bool FindImmediateStore(int at, int reg, const NativeCodeInstruction*& ains);

2
oscar64/oscar64.cpp
View File

@ -73,7 +73,7 @@ int main2(int argc, const char** argv)
#else
strcpy(strProductName, "oscar64");
strcpy(strProductVersion, "1.5.104");
strcpy(strProductVersion, "1.5.105");
#ifdef __APPLE__
uint32_t length = sizeof(basePath);

8
oscar64/oscar64.rc
View File

@ -25,8 +25,8 @@ LANGUAGE LANG_ENGLISH, SUBLANG_NEUTRAL
//
VS_VERSION_INFO VERSIONINFO
FILEVERSION 1,5,104,0
PRODUCTVERSION 1,5,104,0
FILEVERSION 1,5,105,0
PRODUCTVERSION 1,5,105,0
FILEFLAGSMASK 0x3fL
#ifdef _DEBUG
FILEFLAGS 0x1L
@ -43,12 +43,12 @@ BEGIN
BEGIN
VALUE "CompanyName", "oscar64"
VALUE "FileDescription", "oscar64 compiler"
VALUE "FileVersion", "1.5.104.0"
VALUE "FileVersion", "1.5.105.0"
VALUE "InternalName", "oscar64.exe"
VALUE "LegalCopyright", "Copyright (C) 2021"
VALUE "OriginalFilename", "oscar64.exe"
VALUE "ProductName", "oscar64"
VALUE "ProductVersion", "1.5.104.0"
VALUE "ProductVersion", "1.5.105.0"
END
END
BLOCK "VarFileInfo"

6
oscar64setup/oscar64setup.vdproj
View File

@ -3945,15 +3945,15 @@
{
"Name" = "8:Microsoft Visual Studio"
"ProductName" = "8:oscar64"
"ProductCode" = "8:{459712F2-F0A9-4C4E-94E9-87EDFDD5FBA4}"
"PackageCode" = "8:{36A3B67B-6D78-4E56-A84E-8E9BD9280818}"
"ProductCode" = "8:{406BC6F9-178C-4F55-8841-568E49B13E6A}"
"PackageCode" = "8:{37C93EF5-4005-4CD6-B3E1-7A1383214775}"
"UpgradeCode" = "8:{9AB61EFF-ACAC-4079-9950-8D96615CD4EF}"
"AspNetVersion" = "8:2.0.50727.0"
"RestartWWWService" = "11:FALSE"
"RemovePreviousVersions" = "11:TRUE"
"DetectNewerInstalledVersion" = "11:TRUE"
"InstallAllUsers" = "11:FALSE"
"ProductVersion" = "8:1.5.104"
"ProductVersion" = "8:1.5.105"
"Manufacturer" = "8:oscar64"
"ARPHELPTELEPHONE" = "8:"
"ARPHELPLINK" = "8:"