Hi Nathan,
> As Rainer pointed out, there were some regressions in the library tests.
> That's because we didn't build the correct ehspec for __cxa_atexit.
> This adds that, but also, I realize we can use the 'hidden' flag on
> pushdecl to make this lazy-builtin not visible to user name lookup
> without them already declaring it. And I realized we should be setting the
> location as BUILTIN_LOCATION because that's what we're doing. Finally the
> duplicate_decl processing needed augmenting to allow such a lazy builtin's
> eh spec to differ from one already known. (The converse is already dealt
> with, as that's the only case we had before.)
>
> As I mentioned in the first patch, a cleanup of the 'lazily declare a
> builtin' API would be a nice stage-1 thing.
>
> I'll hold off applying this until next week. Rainer, I don't yet know if
> this resolves all of the issues you encountered in the PR.
unfortunately, the results are not very encouraging: while the new
libstdc++ failures caused by the v1 patch are gone, the g++.dg/modules
ICEs remain unchanged, and on Solaris 11.3 (or with
-fno-use-cxa-atexit), the new tests ICE, too. Full details in the PR.
Rainer
--
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University
