https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63593
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> --- Ok, one issue is that predictive commoning removes statements and releases SSA names while they are still in use, and then allocates new SSA names before eventually releasing the using stmts. Delaying that keeps released SSA names in the IL (better than filled with invalid re-used ones) after the transform callback. Now we enter gimple_duplicate_loop_to_header_edge with that "invalid" IL... But at least it fixes this testcase and the related one.
