[llvm] [LoopVectorize] Add support for reverse loops in isDereferenceableAndAlignedInLoop (PR #96752)
David Sherwood via llvm-commits
llvm-commits at lists.llvm.org
Mon Aug 19 07:43:40 PDT 2024
================
@@ -293,26 +293,65 @@ bool llvm::isDereferenceableAndAlignedInLoop(LoadInst *LI, Loop *L,
// TODO: Handle overlapping accesses.
// We should be computing AccessSize as (TC - 1) * Step + EltSize.
- if (EltSize.sgt(Step->getAPInt()))
+ bool StepIsNegative = Step->getAPInt().isNegative();
+ APInt AbsStep = Step->getAPInt().abs();
+ if (EltSize.ugt(AbsStep))
+ return false;
+
+ // For the moment, restrict ourselves to the case where the access size is a
+ // multiple of the requested alignment and the base is aligned.
+ // TODO: generalize if a case found which warrants
+ if (EltSize.urem(Alignment.value()) != 0)
return false;
// Compute the total access size for access patterns with unit stride and
// patterns with gaps. For patterns with unit stride, Step and EltSize are the
// same.
// For patterns with gaps (i.e. non unit stride), we are
// accessing EltSize bytes at every Step.
- APInt AccessSize = TC * Step->getAPInt();
+ APInt AccessSize = TC * AbsStep;
assert(SE.isLoopInvariant(AddRec->getStart(), L) &&
"implied by addrec definition");
Value *Base = nullptr;
if (auto *StartS = dyn_cast<SCEVUnknown>(AddRec->getStart())) {
+ if (StepIsNegative)
+ return false;
Base = StartS->getValue();
} else if (auto *StartS = dyn_cast<SCEVAddExpr>(AddRec->getStart())) {
- // Handle (NewBase + offset) as start value.
- const auto *Offset = dyn_cast<SCEVConstant>(StartS->getOperand(0));
- const auto *NewBase = dyn_cast<SCEVUnknown>(StartS->getOperand(1));
- if (StartS->getNumOperands() == 2 && Offset && NewBase) {
+ const SCEV *End = AddRec->evaluateAtIteration(
----------------
david-arm wrote:
I think that's possible, but it's not a drop-in replacement and is quite inefficient in the current context. For example, `getStartAndEndForAccess` recalculates the trip count and element size which we've already calculated, although I could rip out a lot of the code in `isDereferenceableAndAlignedInLoop` so that I avoid doing this. The function also requires passing in a `PointerBounds` DenseMap, which we don't care about. I could probably refactor `getStartAndEndForAccess` by adding a new variant that doesn't require a map. Ultimately I will still need to extract an actual base `Value*` pointer from a SCEV, but perhaps this would reduce the overall complexity.
https://github.com/llvm/llvm-project/pull/96752
More information about the llvm-commits
mailing list