svn commit: r233391 - head/contrib/libstdc++/libsupc++

[Konstantin Belousov]

On Sat, Mar 24, 2012 at 11:54:21AM -0400, Alexander Kabaev wrote:
> On Fri, 23 Mar 2012 23:39:44 +0000
> David Chisnall <theraven at FreeBSD.org> wrote:
> 
> > On 23 Mar 2012, at 21:10, Konstantin Belousov wrote:
> > 
> > > The patch just committed made the base-shipped library incompatible
> > > with the ports and manual libstdc++ builds.
> > 
> > To clarify - the one shipped in the base system has been incompatible
> > with the one in ports since late 2007...
> >
> That says something about the extend of the breakage, does it not?
>  
> > In future, however, we will want libstdc++ in ports to be using the
> > system ABI library (e.g. libcxxrt) so that it can be linked with
> > libraries using the system C++ stack, so the layout of std::type_info
> > in its headers doesn't matter too much because it won't be used.
> > 
> > > I think it is wiser to
> > > not 'undo the undo', and instead start living with the
> > > upstream-approved ABI. For symvered library, it does not make any
> > > difference if you break it in 10.0 or 9.1.
> > 
> > Not true.  The vtable layout can not be symbol versioned
> > effectively.  The std::type_info class is required by the ABI to
> > exist (and is in libsupc++ / libcxxrt, so is independent of our STL
> > implementation, either libstdc++ or libc++).  Users expect to be able
> > to cast 
> > 
> > > Also, I think that consumers that depend of the ABI of
> > > std::typeinfo can be hacked to understand the layout of vtable.
> > 
> > Supporting both layouts is not really tenable.  We need to do one of
> > the following:
> > 
> > - Say 'we are going to break this now!' and force people to recompile
> > libsupc++, libcxxrt, and libobjc2.
> > - Keep the layout that we currently ship and patch <typeinfo> in
> > ports.
> > 
> > I don't mind which of these we do since I'm not the one that has to
> > do the work in ensuring compatibility with ports but it will need to
> > pick one and then stick to it...
> > 
> > > I doubt that there are
> > > many users that utilize typeinfo.
> > 
> > Yes, to my knowledge there are three, and two of them are me :-(
> > 
> 
> I will support the switch to 'default' ABI used upstream, provided
> David is onboard. The breakage is very limited at the moment and
> can only grow over time, so will grow the pain, while we insist on being
> diverged from GCC. We might as well bite the bullet and live through
> this at the cost of 9.1 release errata that only affects libobjc2.
Right, this was my point, and you seems to agree with me:
better to make a switch back to the upstream ABI sooner (9.1) rather then
later (10.0).
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/svn-src-all/attachments/20120324/540d5340/attachment.pgp