https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85046
--- Comment #3 from Nathan Sidwell <nathan at gcc dot gnu.org> --- Created attachment 43758 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43758&action=edit another example I've further altered the testcase. This one ICEs at #3 all the way back to gcc-5 (at least). Notice it gives no error on lines #1 or #2, at least one of those is bogus. Line #1 either implicitly creates union g inside c()::q, or is an error (I need to check the std carefully, but my recollection is that inside a struct definition only 'union g;' forward declares a union. ). If that recollection is wrong, and #1 is ok. then #2 is an error, because we have to tagged types 'g' inside c()::q. What's happening is that the the xref_tag for 'union q' inside c()::q has ts_global for the scope -- so it looks in outer scopes. But this also injects it into an outer scope when we don't find it. That's wrong. Our lazy binding level caching machinery gets confused and scroggs the binding stack, leading to bad reuse. In some cases that bad reuse was hidden by the existence of a namespace-scope binding onto IDENTIFIER_BINDING. That doesn't exist any more. So we ICE in some more cases. I still think this is ICE on invalid, but it's not error-recovery.
