arcbbb wrote: > Could we detect them differently? I see that plan0 now keeps track of the widen-phi and its backedge. Maybe there's a way to use that information to improve how we handle the IV increment at exit. https://github.com/llvm/llvm-project/pull/147016