Re: string/list ABI changes

[Jason Merrill <jason at redhat dot com>]

On 03/02/2013 05:09 AM, Paolo Carlini wrote:
> Then what about the rest, all the other things we want to change in the
> new ABI? There are C++11 related changes (eg, that additional virtual in
> time) + old issues we finally would be able to fix + things that in a
> new ABI we would have the leeway to do better.

OK, maybe there are other types that we want to change as well.

> I'm not sure to
> understand whether, as I thought, for 4.9 we really want to globally
> break the ABI by default.

I don't think so.  I've seen a lot of push-back against the idea of 
globally breaking the ABI, and in any case I don't see how we could do 
that thoroughly enough to prevent symbol collisions between old and new 
code, especially if we want to maintain compatibility with the 
cross-vendor ABI.  Changing the soname doesn't help at all with that issue.

It seems to me that Microsoft can get away with breaking their ABI all 
the time because they don't have the global symbol space that we have in 
ELF.  Given that, all ABI changes need to either change affected mangled 
names or have the user decide that they don't care because they're going 
to rebuild the world anyway.  But some of our users care very much.

> On top of that providing some header-only
> older versions (of eg, std::list) seem easy, otherwise we would have to
> maintain forever two different ABIs at the same time, thus two different
> versions of a large portion (consider the interactions with locale too)
> of the library at the same time, something we never even attempted
> (personally I'm not sure I want to be a pioneer in this area ;)

I don't see the interactions with locale as a big deal; we just need to 
compile affected things in the library in both modes.  The additional 
maintenance burden is a hassle, but I think manageable.

Jason