Re: libstdc++ ABI break

[Daniel Jacobowitz <drow at false dot org>]

On Fri, Oct 05, 2007 at 05:49:03PM -0700, Mark Mitchell wrote:
> Other than this, is the 4.2.x libstdc++ actually ABI-compatible with
> 4.1.x?  Is there any evidence that people are actually mixing and
> matching in that way?  I'm not asking if the libstdc++ ABI tests pass;
> I'm asking if we actually know that people are actually mixing things.
> Like, are people building Qt applications with 4.2 and linking them to
> 4.1 libstdc++?

The tests that caught this are not the C++ ABI tests.  They're the LSB
(Linux Standard Base) library ABI tests, which are designed to let a
large number of vendors find a common ABI to work on a large number of
distributions.  So I think you can take it as certain that there will
be people running LSB-compliant applications on both GCC 4.2 and GCC
4.2.1.

There will also be other vendors running the LSB compliance testsuite
in the months to come.  They're all going to face commercial pressure
to apply something like Ben's patch; as long as libstdc++ keeps the
libstdc++.so.6 SONAME, its ABI is frozen as far as Linux vendors are
concerned.

No one's noticed it yet, but it would be a pretty nasty bug to track
down, and the affected release was only made ten weeks ago; it hasn't
been widely picked up yet.  I'm sure it will be.  The people who have
already picked it up are hopefully still using it for test releases.

I recommend Ben's patch.

-- 
Daniel Jacobowitz
CodeSourcery