[llvm] [AggressiveInstCombine] Expand strchr/memchr with small constant strings (PR #98501)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Jul 15 01:52:02 PDT 2024
================
@@ -1103,6 +1108,85 @@ void StrNCmpInliner::inlineCompare(Value *LHS, StringRef RHS, uint64_t N,
}
}
+/// Convert strchr/memchr with a small constant string into a switch
+static bool foldStrChr(CallInst *Call, LibFunc Func, DomTreeUpdater *DTU,
+ const DataLayout &DL) {
+ assert((Func == LibFunc_strchr || Func == LibFunc_memchr) &&
+ "Unexpected LibFunc");
+ if (isa<Constant>(Call->getArgOperand(1)))
+ return false;
+
+ StringRef Str;
+ Value *Base = Call->getArgOperand(0);
+ if (!getConstantStringInfo(Base, Str, /*TrimAtNul=*/Func == LibFunc_strchr))
+ return false;
+
+ uint64_t N = Str.size();
+ if (Func == LibFunc_memchr) {
+ if (auto *ConstInt = dyn_cast<ConstantInt>(Call->getArgOperand(2))) {
+ uint64_t Val = ConstInt->getZExtValue();
+ // Ignore the case that n is larger than the size of string.
+ if (Val > N)
+ return false;
+ N = Val;
+ } else
+ return false;
+ }
+
+ if (N > StrChrInlineThreshold)
+ return false;
+
+ BasicBlock *BB = Call->getParent();
+ BasicBlock *BBNext = SplitBlock(BB, Call, DTU);
+ IRBuilder<> IRB(BB);
+ IntegerType *ByteTy = IRB.getInt8Ty();
+ BB->getTerminator()->eraseFromParent();
+ SwitchInst *SI = IRB.CreateSwitch(
+ IRB.CreateTrunc(Call->getArgOperand(1), ByteTy), BBNext, N);
+ Type *IndexTy = DL.getIndexType(Call->getType());
+
+ PHINode *PHI = PHINode::Create(Call->getType(), 2, "", BBNext->begin());
+ PHI->addIncoming(Constant::getNullValue(Call->getType()), BB);
+
+ SmallVector<DominatorTree::UpdateType, 8> Updates;
+
+ BasicBlock *BBSuccess = BasicBlock::Create(
+ Call->getContext(), "strchr.success", BB->getParent(), BBNext);
+ IRB.SetInsertPoint(BBSuccess);
+ PHINode *IndexPHI = IRB.CreatePHI(IndexTy, N);
+ Value *FirstOccursLocation = IRB.CreateInBoundsPtrAdd(Base, IndexPHI);
+ PHI->addIncoming(FirstOccursLocation, BBSuccess);
+ IRB.CreateBr(BBNext);
+ if (DTU)
+ Updates.push_back({DominatorTree::Insert, BBSuccess, BBNext});
+
+ SmallPtrSet<ConstantInt *, 4> Cases;
+ for (uint64_t I = 0; I < N; ++I) {
+ ConstantInt *CaseVal = ConstantInt::get(ByteTy, Str[I]);
+ if (!Cases.insert(CaseVal).second)
+ continue;
+
+ BasicBlock *BBCase = BasicBlock::Create(Call->getContext(), "strchr.case",
+ BB->getParent(), BBNext);
+ SI->addCase(CaseVal, BBCase);
+ IRB.SetInsertPoint(BBCase);
+ IndexPHI->addIncoming(ConstantInt::get(IndexTy, I), BBCase);
+ IRB.CreateBr(BBSuccess);
+ if (DTU) {
+ Updates.push_back({DominatorTree::Insert, BB, BBCase});
+ Updates.push_back({DominatorTree::Insert, BBCase, BBSuccess});
+ }
+ }
----------------
goldsteinn wrote:
Does the new code get properly optimized if the call is only used for equality?
Can you add a few tests / modify a few of your existing tests to test that case?
https://github.com/llvm/llvm-project/pull/98501
More information about the llvm-commits
mailing list