https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69155
--- Comment #13 from rguenther at suse dot de <rguenther at suse dot de> --- On January 13, 2016 7:08:28 PM GMT+01:00, "jakub at gcc dot gnu.org" <gcc-bugzi...@gcc.gnu.org> wrote: >https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69155 > >--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> --- >Created attachment 37335 > --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37335&action=edit >gcc6-pr69155-wip.patch > >Partial patch I've bootstrapped/regtested on x86_64-linux and >i686-linux. >Either our testsuite coverage is insufficient, or it might be safe to >only deal >with PHI arguments and nothing else if we walk in rpo order. >In that case, as the renamer is not prepared to handle PHIs with >non-is_gimple_val arguments and renaming them, it might be much easier >to just >not use the renamer and simply record the PHIs we still need to update, >push >the underlying VAR_DECLs (or extra temporaries for anon SSA_NAMEs) to >the PHI >arguments first and get back to them at the end (or immediately once >the >SSA_NAME is defined?). I think it's enough to deal with PHIs and keeping a worklist of unhandled edge/PHI pairs would indeed work (and use decls as placeholder there). Richard.
