[llvm] [LAA] Honor no-wrap predicates already in PSE in isNoWrap (PR #226430)

[Tim Besard via llvm-commits]

================
@@ -1153,7 +1153,12 @@ isNoWrap(PredicatedScalarEvolution &PSE, const SCEVAddRecExpr *AR, Value *Ptr,
     return true;
   }
 
-  return false;
+  // Without adding a new predicate, AR may still be known not to wrap if the
+  // predicates of PSE already imply it, e.g. because a wrap predicate for AR
+  // was added while analyzing the dependences of the loop.
+  ScalarEvolution &SE = *PSE.getSE();
+  return PSE.getPredicate().implies(
+      SE.getWrapPredicate(AR, SCEVWrapPredicate::IncrementNUSW), SE);
----------------
maleadt wrote:

Thanks, but I don't think this works. Applying your suggestion breaks the case I'm interested in: `isDependenceDistanceOfOne` calls `getPtrStride` without a `Predicates` vector, so `isNoWrap` never enters `if (Ptr && Predicates)` and still returns false for pointers whose NUSW predicate was already added. With the suggestion applied, both new tests fail and LLE doesn't eliminate any loads.
AFAIU skipping the `push_back` also doesn't change anything: every caller passes the collected predicates to `PSE.addPredicates`, which already drops implied ones, so the check would just run the `implies()` scan twice for each new predicate.

I've instead pushed a version that takes what I understand to be the idea of your suggestion, building a single NUSW wrap predicate for the pointer, and uses it for both outcomes. Can you take another look?

https://github.com/llvm/llvm-project/pull/226430