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

Optimize 32bit shifts

This commit is contained in:
drmortalwombat 2024-07-21 21:43:59 +02:00
parent 5f2e05d1dd
commit 5b50389340
3 changed files with 47 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
autotest/bitshifttest.c
View File

@ -373,21 +373,25 @@ int main(void)
shr16n(0xfedc, 0xfedc); shr16n(0xfedc, 0xfedc);
shl32b(0x00000000UL, 0x00000000L); shl32b(0x00000000UL, 0x00000000L);
shl32b(0x00000001UL, 0x00000001L);
shl32b(0xffffffffUL, 0xffffffffL); shl32b(0xffffffffUL, 0xffffffffL);
shl32b(0x12345678UL, 0x12345678L); shl32b(0x12345678UL, 0x12345678L);
shl32b(0xfedcba98UL, 0xfedcba98L); shl32b(0xfedcba98UL, 0xfedcba98L);
shr32b(0x00000000UL, 0x00000000L); shr32b(0x00000000UL, 0x00000000L);
shr32b(0x00000001UL, 0x00000001L);
shr32b(0xffffffffUL, 0xffffffffL); shr32b(0xffffffffUL, 0xffffffffL);
shr32b(0x12345678UL, 0x12345678L); shr32b(0x12345678UL, 0x12345678L);
shr32b(0xfedcba98UL, 0xfedcba98L); shr32b(0xfedcba98UL, 0xfedcba98L);
shl32n(0x00000000UL, 0x00000000L); shl32n(0x00000000UL, 0x00000000L);
shl32n(0x00000001UL, 0x00000001L);
shl32n(0xffffffffUL, 0xffffffffL); shl32n(0xffffffffUL, 0xffffffffL);
shl32n(0x12345678UL, 0x12345678L); shl32n(0x12345678UL, 0x12345678L);
shl32n(0xfedcba98UL, 0xfedcba98L); shl32n(0xfedcba98UL, 0xfedcba98L);
shr32n(0x00000000UL, 0x00000000L); shr32n(0x00000000UL, 0x00000000L);
shr32n(0x00000001UL, 0x00000001L);
shr32n(0xffffffffUL, 0xffffffffL); shr32n(0xffffffffUL, 0xffffffffL);
shr32n(0x12345678UL, 0x12345678L); shr32n(0x12345678UL, 0x12345678L);
shr32n(0xfedcba98UL, 0xfedcba98L); shr32n(0xfedcba98UL, 0xfedcba98L);

2
oscar64/InterCode.cpp
View File

@ -15247,7 +15247,7 @@ void InterCodeBasicBlock::EliminateDoubleLoopCounter(void)
{ {
if (mEntryBlocks[i] != mLoopPrefix) if (mEntryBlocks[i] != mLoopPrefix)
{ {
if (!mEntryBlocks[i]->CollectLoopBody(this, body)) if (!mEntryBlocks[i]->CollectLoopBodyRecursive(this, body))
innerLoop = false; innerLoop = false;
} }
} }

42
oscar64/NativeCodeGenerator.cpp
View File

@ -9823,6 +9823,39 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
return eblock; return eblock;
} }
} }
else if (ins->mSrc[1].mTemp < 0 && ins->mSrc[1].mIntConst == 1)
{
NativeCodeBasicBlock* lblock = nproc->AllocateBlock();
NativeCodeBasicBlock* hblock = nproc->AllocateBlock();
NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
mIns.Push(NativeCodeInstruction(ins, ASMIT_AND, ASMIM_IMMEDIATE, 0x1f));
mIns.Push(NativeCodeInstruction(ins, ASMIT_TAX, ASMIM_IMPLIED));
mIns.Push(NativeCodeInstruction(ins, ASMIT_AND, ASMIM_IMMEDIATE, 0x10));
this->Close(ins, lblock, hblock, ASMIT_BEQ);
NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("bitshift")));
lblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
lblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
lblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ABSOLUTE_X, frt.mOffset + 8, frt.mLinkerObject));
lblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg));
lblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ABSOLUTE_X, frt.mOffset, frt.mLinkerObject));
lblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
lblock->Close(ins, eblock, nullptr, ASMIT_JMP);
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_IMMEDIATE, 0));
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 0));
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 1));
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ABSOLUTE_X, frt.mOffset - 16, frt.mLinkerObject));
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 2));
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ABSOLUTE_X, frt.mOffset - 8, frt.mLinkerObject));
hblock->mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, treg + 3));
hblock->Close(ins, eblock, nullptr, ASMIT_JMP);
return eblock;
}
else else
{ {
NativeCodeBasicBlock* lblock = nproc->AllocateBlock(); NativeCodeBasicBlock* lblock = nproc->AllocateBlock();
@ -44931,6 +44964,15 @@ bool NativeCodeBasicBlock::PeepHoleOptimizerIterate3(int i, int pass)
mIns[i + 2].mLive |= LIVE_CPU_REG_Y; mIns[i + 2].mLive |= LIVE_CPU_REG_Y;
return true; return true;
} }
else if (
mIns[i + 0].mType == ASMIT_STA &&
mIns[i + 1].mType == ASMIT_STA &&
mIns[i + 2].IsShift() && mIns[i + 2].SameEffectiveAddress(mIns[i + 0]) && !(mIns[i + 2].mLive & (LIVE_CPU_REG_A | LIVE_MEM)))
{
mIns[i + 0].mType = ASMIT_NOP; mIns[i + 0].mMode = ASMIM_IMPLIED;
mIns[i + 2].mMode = ASMIM_IMPLIED;
return true;
}
if ( if (
mIns[i + 0].mType == ASMIT_LDY && mIns[i + 0].mMode == ASMIM_IMMEDIATE && mIns[i + 0].mAddress <= 1 && mIns[i + 0].mType == ASMIT_LDY && mIns[i + 0].mMode == ASMIM_IMMEDIATE && mIns[i + 0].mAddress <= 1 &&