I second @nikolaypanchenko and @efriedma-quic that “correct” long term solution would be to support mentioned pattern by the vectorizer. I believe there shouldn’t be any technically unsolvable problems emitting runtime boundary checks (even target specific) as part of vectoirzation.
Saying the above, it still may make sense to add a separate “pattern matching” pass as a short term solution because the “correct” solution will take ?significant? time to implement. On the other hand, if we align our efforts and work together we can finish implementation much faster. @kmclaughlin-arm do you or anybody else have interest\opportunity to join “[RFC] Vectorization of loops with Data Dependent Exits” initiative?