https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89399
--- Comment #5 from Jeffrey A. Law <law at redhat dot com> --- So I think in the immediate term any time we're using PATTERN (cand->insn) we really should be using single_set (cand->insn). We already know we've passed the single_set test when we put the insn onto the candidate list. I haven't audited all the uses of get_sub_rtx, but the ones I have looked at look like it's used on the insn we're modifying rather than the candidate insn (which is the extension we're trying to completely eliminate). I'm going to spin the change to avoid incorrect uses of PATTERN (cand->insn) in my tester.
