PR12760 plugin vs. warning syms

[Alan Modra]

On Mon, May 16, 2011 at 07:42:44AM -0700, H.J. Lu wrote:
> On Mon, May 16, 2011 at 7:12 AM, Alan Modra <amodra@gmail.com> wrote:
> > On Mon, May 16, 2011 at 05:48:42AM -0700, H.J. Lu wrote:
> >> On Mon, May 16, 2011 at 1:29 AM, Alan Modra <amodra@gmail.com> wrote:
> >> > On Sun, May 15, 2011 at 06:37:58PM -0700, H.J. Lu wrote:
> >> >> We must call bfd_link_repair_undef_list when we change symbol
> >> >> type to bfd_link_hash_undefweak. Â Otherwise, assert may fail in:
> >> >>
> >> >> void
> >> >> bfd_link_add_undef (struct bfd_link_hash_table *table,
> >> >> Â  Â  Â  Â  Â  Â  Â  Â  Â  Â  struct bfd_link_hash_entry *h)
> >> >> {
> >> >> Â  BFD_ASSERT (h->u.undef.next == NULL);
> >> >
> >> > How could that happen? Â We know the new symbol is a definition.
> >> >
> >>
> >> After we turned a definition into an undefweak, we may see a reference in
> >> LTO output, which will call bfd_link_add_undef.
> >
> > No, because the undefweak doesn't last. Â When plugin_notice returns to
> > _bfd_generic_link_add_one_symbol we'll turn it into a defined,
> > defweak, or common.
> >
> 
> If linker sees undefined reference from an LTO output first, linker
> will turn undefweak into undefined, which will trigger the assert.

plugin_notice won't be changing syms (very temporarily) to undefweak
in that case.

-- 
Alan Modra
Australia Development Lab, IBM