On Tue, Dec 18, 2007 at 07:46:31AM +0000, Zack Weinberg wrote:
> (apologies for breaking threads, I'm not subscribed and can't hack the
> References line)
>
> Steve Langasek wrote:
> > Jay Berkenbilt wrote:
> > > (Though I'm surprised that going from gcc 4.1 to 4.2 is really an ABI
> > > change.)
> >
> > It's not, except that the boost Debian packages are encoding the gcc version
> > in the soname by hand. This is wrong, but needs to be addressed in order to
> > get icu through in a timely fashion.
>
> Back when monotone used a bunch of boost modules that included
> libraries (as opposed to just headers), the upstream mailing list
> would regularly get reports that the binary segfaulted on most
> operations, and the cause was *invariably* that the user had
> hand-built boost, upgraded gcc, and then hand-built monotone against
> the boost libraries built with the old gcc (using static linkage, so
> the boost soname convention didn't stop them). We saw this a few
> times going between gcc *patch levels*!
>
> Thus I do not think it is safe to remove the gcc version from the
> boost library sonames, despite how nice it would be in terms of
> speeding up testing transitions.
This is completely insane. "In the past" there was the gcc3 to gcc4
C++ ABI transition, and that's it. Debian IS tracking g++/libstdc++ ABI
breakages and does transitions on their own.
Here are 3 things:
* having the gcc version in the soname changes _noting_ to the
cluelessness of users: it doesn't prevent you to link with another
GCC version AFAICT, so this point is moot ;
* if boost wants a delirious soname for custom builds to allow
co-installability to test performance with various g++ versions,
fine by me, it even makes sense after all ;
* boost should come with a decent soname change policy for vendors
that *KNOW* about library transitions and prevent user from shooting
themselves in the foot. The current scheme just manage to make us
want to yell at the upstream, and achieves nothing.
--
·O· Pierre Habouzit
··O madcoder@debian.org
OOO http://www.madism.org
Attachment:
pgpnIsvIjxuu9.pgp
Description: PGP signature