On Thu, Sep 13, 2018 at 5:22 PM Nathan Sidwell <nat...@acm.org> wrote: > > Richard, Jan, as early debug experts perhaps you can clue me in? > > 87295 is an ICE in lookup_external_ref, we strcmp a NULL > die->die_id.die_symbol. > > The die is being cloned in clone_as_declaration during a > copy_ancestor_tree walk ultimately coming from > copy_decls_for_unworthy_types. > > The incoming die is DW_TAG_structure_type with comdat_type_p set. We > create a new_die_raw, which has NULL die_id.die_symbol and then never > set that field (or another piece of the die_id union). > > How is this supposed to work?
I think we're forgetting to "undo" the -fdebug-types-section part of the early LTO debug output before trying to do it again for the fat part and there we get confused. Test coverage is non-existant for -fdebug-types-section (mostly), so this has gone unnoticed. In dwarf2out_finish you'll find a if (flag_generate_lto || ...) section done early that undoes quite some stuff but appearantly that's not enough ... Richard. > > nathan > -- > Nathan Sidwell
