[llvm] [LoopVectorize] Add support for reverse loops in isDereferenceableAndAlignedInLoop (PR #96752)

[Florian Hahn via llvm-commits]

================
@@ -276,84 +277,80 @@ static bool AreEquivalentAddressValues(const Value *A, const Value *B) {
 bool llvm::isDereferenceableAndAlignedInLoop(
     LoadInst *LI, Loop *L, ScalarEvolution &SE, DominatorTree &DT,
     AssumptionCache *AC, SmallVectorImpl<const SCEVPredicate *> *Predicates) {
-  auto &DL = LI->getDataLayout();
-  Value *Ptr = LI->getPointerOperand();
-
-  APInt EltSize(DL.getIndexTypeSizeInBits(Ptr->getType()),
-                DL.getTypeStoreSize(LI->getType()).getFixedValue());
-  const Align Alignment = LI->getAlign();
+  const SCEV *Ptr = SE.getSCEV(LI->getPointerOperand());
+  auto *AddRec = dyn_cast<SCEVAddRecExpr>(Ptr);
 
-  Instruction *HeaderFirstNonPHI = L->getHeader()->getFirstNonPHI();
-
-  // If given a uniform (i.e. non-varying) address, see if we can prove the
-  // access is safe within the loop w/o needing predication.
-  if (L->isLoopInvariant(Ptr))
----------------
fhahn wrote:

It looks like this code path was previously not tested and dropping it may cause a regression. I added 747f7f38bddaa4ae60c8d1380c9b28d81fc13e8a which should cover this codepath.

https://github.com/llvm/llvm-project/pull/96752