Emulator-Archive/ppsspp
1
0
Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2024-11-23 05:19:56 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Add new IR optimization pass, OptimizeLoadsAfterStores

Browse Source
This commit is contained in:
Henrik Rydgård 2024-06-07 11:07:21 +02:00
parent da88011805
commit bd0beb68a4
4 changed files with 47 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Core/MIPS/IR/IRFrontend.cpp
View File

@ -284,6 +284,7 @@ void IRFrontend::DoJit(u32 em_address, std::vector<IRInst> &instructions, u32 &m
&PropagateConstants,
&PurgeTemps,
&ReduceVec4Flush,
&OptimizeLoadsAfterStores,
// &ReorderLoadStore,
// &MergeLoadStore,
// &ThreeOpToTwoOp,

8
Core/MIPS/IR/IRJit.cpp
View File

@ -255,15 +255,19 @@ void IRJit::RunLoopUntil(u64 globalticks) {
u32 opcode = inst & 0xFF000000;
if (opcode == MIPS_EMUHACK_OPCODE) {
u32 offset = inst & 0x00FFFFFF; // Alternatively, inst - opcode
const IRInst *instPtr = blocks_.GetArenaPtr() + offset;
_dbg_assert_(instPtr->op == IROp::Downcount);
mips->downcount -= instPtr->constant;
instPtr++;
#ifdef IR_PROFILING
IRBlock *block = blocks_.GetBlock(blocks_.GetBlockNumFromOffset(offset));
TimeSpan span;
mips->pc = IRInterpret(mips, blocks_.GetArenaPtr() + offset);
mips->pc = IRInterpret(mips, instPtr);
int64_t elapsedNanos = span.ElapsedNanos();
block->profileStats_.executions += 1;
block->profileStats_.totalNanos += elapsedNanos;
#else
mips->pc = IRInterpret(mips, blocks_.GetArenaPtr() + offset);
mips->pc = IRInterpret(mips, instPtr);
#endif
// Note: this will "jump to zero" on a badly constructed block missing exits.
if (!Memory::IsValid4AlignedAddress(mips->pc)) {

39
Core/MIPS/IR/IRPassSimplify.cpp
View File

@ -2150,6 +2150,45 @@ bool ReduceVec4Flush(const IRWriter &in, IRWriter &out, const IROptions &opts) {
return logBlocks;
}
// This optimizes away redundant loads-after-stores, which are surprisingly not that uncommon.
bool OptimizeLoadsAfterStores(const IRWriter &in, IRWriter &out, const IROptions &opts) {
CONDITIONAL_DISABLE;
// This tells us to skip an AND op that has been optimized out.
// Maybe we could skip multiple, but that'd slow things down and is pretty uncommon.
int nextSkip = -1;
bool logBlocks = false;
for (int i = 0, n = (int)in.GetInstructions().size(); i < n; i++) {
IRInst inst = in.GetInstructions()[i];
// Just copy the last instruction.
if (i == n - 1) {
out.Write(inst);
break;
}
out.Write(inst);
IRInst next = in.GetInstructions()[i + 1];
switch (inst.op) {
case IROp::Store32:
if (next.op == IROp::Load32 &&
next.constant == inst.constant &&
next.dest == inst.src3 &&
next.src1 == inst.src1) {
// The upcoming load is completely redundant.
// Skip it.
i++;
}
break;
default:
break;
}
}
return logBlocks;
}
bool OptimizeForInterpreter(const IRWriter &in, IRWriter &out, const IROptions &opts) {
CONDITIONAL_DISABLE;
// This tells us to skip an AND op that has been optimized out.

1
Core/MIPS/IR/IRPassSimplify.h
View File

@ -17,4 +17,5 @@ bool MergeLoadStore(const IRWriter &in, IRWriter &out, const IROptions &opts);
bool ApplyMemoryValidation(const IRWriter &in, IRWriter &out, const IROptions &opts);
bool ReduceVec4Flush(const IRWriter &in, IRWriter &out, const IROptions &opts);
bool OptimizeLoadsAfterStores(const IRWriter &in, IRWriter &out, const IROptions &opts);
bool OptimizeForInterpreter(const IRWriter &in, IRWriter &out, const IROptions &opts);