https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90231
--- Comment #17 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Note, for the ignoring of not useful candidates we could during the cand_pref computation not just check for step being equal, but also if one step is multiple of the other one or vice versa, of course prefer same step first, then perhaps the * case where info->iv->step is a constant multiple of cand->iv->step and finally the division case, perhaps that one with the requirement that the type of cand->iv->base has then at least ceil_log2 (cand->iv->step) bits larger TYPE_PRECISION.
