On September 8, 2018 10:49:02 AM GMT+01:00, Jan Hubicka <hubi...@ucw.cz> wrote:
>Hi
>while working on path to replace type variant by first compatible one I
>run
>into issue that the first vairant was not seen by free_lang_data. I
>think this
>is a bug becuase nothing prevents middle-end from looking up a variant
>there
>and using it when it needs to do so.
>
>I tried to walk the variant lists and free lang data on them but that
>crahses
>because not all types in variant list passes verify_type (C++ puts
>there
>incomplete type variants whose canonical type is complete and thus they
>are
>considered bogus). So pruning those lists seems to be better variant.
>
>Bootstrapped/regtested x86_64-linux, OK?
But if they are still reachable by GC they are now bogously unlinked. We
already do not stream unused variants, so what does this achieve? Maybe we
should make the variant list GCable somehow?
>Honza
>
> * tree.c (free_lang_data_in_cgraph): Prune TYPE_NEXT_VARIANT lists.
>Index: tree.c
>===================================================================
>--- tree.c (revision 263989)
>+++ tree.c (working copy)
>@@ -5845,7 +5845,12 @@ free_lang_data_in_cgraph (void)
>
> /* Traverse every type found freeing its language data. */
> FOR_EACH_VEC_ELT (fld.types, i, t)
>- free_lang_data_in_type (t);
>+ {
>+ free_lang_data_in_type (t);
>+ while (TYPE_NEXT_VARIANT (t)
>+ && !fld.pset.contains (TYPE_NEXT_VARIANT (t)))
>+ TYPE_NEXT_VARIANT (t) = TYPE_NEXT_VARIANT (TYPE_NEXT_VARIANT (t));
>+ }
> if (flag_checking)
> {
> FOR_EACH_VEC_ELT (fld.types, i, t)
