[clang-tools-extra] [clang] [llvm] [AArch64] Add an AArch64 pass for loop idiom transformations (PR #72273)
Antonio Frighetto via cfe-commits
cfe-commits at lists.llvm.org
Wed Dec 13 06:44:19 PST 2023
================
@@ -0,0 +1,726 @@
+
+//===- AArch64LoopIdiomTransform.cpp - Loop idiom recognition -------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "AArch64LoopIdiomTransform.h"
+#include "llvm/Analysis/DomTreeUpdater.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/MDBuilder.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "aarch64-lit"
+
+static cl::opt<bool>
+ DisableAll("disable-aarch64-lit-all", cl::Hidden, cl::init(false),
+ cl::desc("Disable AArch64 Loop Idiom Transform Pass."));
+
+static cl::opt<bool> DisableByteCmp(
+ "disable-aarch64-lit-bytecmp", cl::Hidden, cl::init(false),
+ cl::desc("Proceed with AArch64 Loop Idiom Transform Pass, but do "
+ "not convert byte-compare loop(s)."));
+
+namespace llvm {
+
+void initializeAArch64LoopIdiomTransformLegacyPassPass(PassRegistry &);
+Pass *createAArch64LoopIdiomTransformPass();
+
+} // end namespace llvm
+
+namespace {
+
+class AArch64LoopIdiomTransform {
+ Loop *CurLoop = nullptr;
+ DominatorTree *DT;
+ LoopInfo *LI;
+ const TargetTransformInfo *TTI;
+ const DataLayout *DL;
+
+public:
+ explicit AArch64LoopIdiomTransform(DominatorTree *DT, LoopInfo *LI,
+ const TargetTransformInfo *TTI,
+ const DataLayout *DL)
+ : DT(DT), LI(LI), TTI(TTI), DL(DL) {}
+
+ bool run(Loop *L);
+
+private:
+ /// \name Countable Loop Idiom Handling
+ /// @{
+
+ bool runOnCountableLoop();
+ bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
+ SmallVectorImpl<BasicBlock *> &ExitBlocks);
+
+ bool recognizeByteCompare();
+ Value *expandFindMismatch(IRBuilder<> &Builder, GetElementPtrInst *GEPA,
+ GetElementPtrInst *GEPB, Value *Start,
+ Value *MaxLen);
+ void transformByteCompare(GetElementPtrInst *GEPA, GetElementPtrInst *GEPB,
+ Value *MaxLen, Value *Index, Value *Start,
+ bool IncIdx, BasicBlock *FoundBB,
+ BasicBlock *EndBB);
+ /// @}
+};
+
+class AArch64LoopIdiomTransformLegacyPass : public LoopPass {
+public:
+ static char ID;
+
+ explicit AArch64LoopIdiomTransformLegacyPass() : LoopPass(ID) {
+ initializeAArch64LoopIdiomTransformLegacyPassPass(
+ *PassRegistry::getPassRegistry());
+ }
+
+ StringRef getPassName() const override {
+ return "Recognize AArch64-specific loop idioms";
+ }
+
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.addRequired<LoopInfoWrapperPass>();
+ AU.addRequired<DominatorTreeWrapperPass>();
+ AU.addRequired<TargetTransformInfoWrapperPass>();
+ }
+
+ bool runOnLoop(Loop *L, LPPassManager &LPM) override;
+};
+
+bool AArch64LoopIdiomTransformLegacyPass::runOnLoop(Loop *L,
+ LPPassManager &LPM) {
+
+ if (skipLoop(L))
+ return false;
+
+ auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+ auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
+ auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
+ *L->getHeader()->getParent());
+ return AArch64LoopIdiomTransform(
+ DT, LI, &TTI, &L->getHeader()->getModule()->getDataLayout())
+ .run(L);
+}
+
+} // end anonymous namespace
+
+char AArch64LoopIdiomTransformLegacyPass::ID = 0;
+
+INITIALIZE_PASS_BEGIN(
+ AArch64LoopIdiomTransformLegacyPass, "aarch64-lit",
+ "Transform specific loop idioms into optimised vector forms", false, false)
+INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
+INITIALIZE_PASS_END(
+ AArch64LoopIdiomTransformLegacyPass, "aarch64-lit",
+ "Transform specific loop idioms into optimised vector forms", false, false)
+
+Pass *llvm::createAArch64LoopIdiomTransformPass() {
+ return new AArch64LoopIdiomTransformLegacyPass();
+}
+
+PreservedAnalyses
+AArch64LoopIdiomTransformPass::run(Loop &L, LoopAnalysisManager &AM,
+ LoopStandardAnalysisResults &AR,
+ LPMUpdater &) {
+ if (DisableAll)
+ return PreservedAnalyses::all();
+
+ const auto *DL = &L.getHeader()->getModule()->getDataLayout();
+
+ AArch64LoopIdiomTransform LIT(&AR.DT, &AR.LI, &AR.TTI, DL);
+ if (!LIT.run(&L))
+ return PreservedAnalyses::all();
+
+ return PreservedAnalyses::none();
+}
+
+//===----------------------------------------------------------------------===//
+//
+// Implementation of AArch64LoopIdiomTransform
+//
+//===----------------------------------------------------------------------===//
+
+bool AArch64LoopIdiomTransform::run(Loop *L) {
+ CurLoop = L;
+
+ if (DisableAll)
+ return false;
+
+ // If the loop could not be converted to canonical form, it must have an
+ // indirectbr in it, just give up.
+ if (!L->getLoopPreheader())
+ return false;
+
+ LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
+ << CurLoop->getHeader()->getParent()->getName()
+ << "] Loop %" << CurLoop->getHeader()->getName() << "\n");
+
+ return recognizeByteCompare();
+}
+
+/// Match loop-invariant value.
+template <typename SubPattern_t> struct match_LoopInvariant {
+ SubPattern_t SubPattern;
+ const Loop *L;
+
+ match_LoopInvariant(const SubPattern_t &SP, const Loop *L)
+ : SubPattern(SP), L(L) {}
+
+ template <typename ITy> bool match(ITy *V) {
+ return L->isLoopInvariant(V) && SubPattern.match(V);
+ }
+};
+
+/// Matches if the value is loop-invariant.
+template <typename Ty>
+inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) {
+ return match_LoopInvariant<Ty>(M, L);
+}
+
+bool AArch64LoopIdiomTransform::recognizeByteCompare() {
+ if (!TTI->supportsScalableVectors() || !TTI->getMinPageSize().has_value() ||
+ DisableByteCmp)
+ return false;
+
+ BasicBlock *Header = CurLoop->getHeader();
+ BasicBlock *PH = CurLoop->getLoopPreheader();
+
+ // In AArch64LoopIdiomTransform::run we have already checked that the loop
+ // has a preheader so we can assume it's in a canonical form.
+ auto *EntryBI = cast<BranchInst>(PH->getTerminator());
+
+ if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 2)
+ return false;
+
+ PHINode *PN = dyn_cast<PHINode>(&Header->front());
+ if (!PN || PN->getNumIncomingValues() != 2)
+ return false;
+
+ auto LoopBlocks = CurLoop->getBlocks();
+ // The first block in the loop should contain only 4 instructions, e.g.
+ //
+ // while.cond:
+ // %res.phi = phi i32 [ %start, %ph ], [ %inc, %while.body ]
+ // %inc = add i32 %res.phi, 1
+ // %cmp.not = icmp eq i32 %inc, %n
+ // br i1 %cmp.not, label %while.end, label %while.body
+ //
+ auto CondBBInsts = LoopBlocks[0]->instructionsWithoutDebug();
+ if (std::distance(CondBBInsts.begin(), CondBBInsts.end()) > 4)
----------------
antoniofrighetto wrote:
Sorry for replying only now, thanks for the explanation. If I got this correctly, the snippet you provided is not meant to be transformed as the indexes of the two GEPs do not match, thus we would be bailing out. I'm confused though, why the following snippet shouldn't be supported (as per comment [here](https://github.com/llvm/llvm-project/blob/5f09a37fc4d808310b8f0ae23f10f953d05c8749/llvm/lib/Target/AArch64/AArch64LoopIdiomTransform.cpp#L359-L376))? Doesn't it match the specifics of the pass (the index is simply incremented before loading the values), or am I missing something?
```c
int test(unsigned char *p1, unsigned char *p2, unsigned start, unsigned end) {
while (++start != end) {
if (p1[start] != p2[start])
break;
}
return start;
}
```
Also, if we were to support 1), 2), 3) we would still not be catching samples like https://godbolt.org/z/vPGqde4zj, due to the `and` on the induction variable (not sure why it has been converted to an i64 here, but I guess it's an orthogonal problem – provided we want to catch this).
https://github.com/llvm/llvm-project/pull/72273
More information about the cfe-commits
mailing list