https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70615
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- cp_gimplify_expr has code to handle the PTRMEM_CST case, but that is unfortunately too late, if it appears in a CONSTRUCTOR, the gimplifier can try to hash the CONSTRUCTOR which obviously can't handle the FE specific trees. So, shall the FE somewhere early remove the PTRMEM_CSTs, or shall cp_genericize_r remove them (well, replace with cplus_expand_constant result), or shall cp_gimplify_expr have a case for CONSTRUCTOR where it would walk_tree the ctor and use cplus_expand_constant in there?