https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79654
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Yeah, I'd prefer both patches (of course just a single testcase). For the decomp, there is really nothing wrong with the artificial var's type, and there could be other places lurking which are upset about error_mark_node type on decls.
