https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120745
--- Comment #2 from Richard Sandiford <rsandifo at gcc dot gnu.org> --- Created attachment 61697 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61697&action=edit Candidate patch Hmm, yeah. Could you try the attached patch? It seems to work for this testcase and passes an aarch64 bootstrap & regression-test on trunk. (That's not saying much, though, since current trunk AFAIK doesn't try to delete cycles.)
