https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82524
--- Comment #13 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Jakub Jelinek from comment #12) > (In reply to Jakub Jelinek from comment #11) > > Created attachment 42346 [details] > > gcc8-pr82524.patch > > Actually, that doesn't work, we ICE on the match_dups. rtx_equal_p in the > conditions works though and could be easily reverted (but the =Q to +Q kept) > if LRA is taught to do this (if possible). Yes, this approach is much less invasive. Perhaps a FIXME comment referring to this PR should be added somewhere. I also expect that gcc.target/i386/pr78904-?.c testcases will need to be XFAILed with your patch.
