[Bug middle-end/51766] [4.7 regression] sync_fetch_and_xxx atomicity

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51766

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-01-10 14:48:54 UTC ---
(In reply to comment #5)
> I understand that fixing __sync_* is a hassle.  This is why I opened a separate
> bug for libstdc++.
> 
> While __sync_* is deprecated in favor of __atomic_*, use of __sync_* for
> portability is fairly pervasive in FOSS applications that need it because of
> its implementation in GCC.  Most programmers do not know about memory models
> and do not care about memory models.  And it will take time for programmers to
> switch to __atomic_*, if they even bother to choose a memory model and don't
> introduce a bug.
> 
> The basic problem is MEMMODEL_SEQ_CST only makes a performance difference for
> POWER and developers are going to continue to use __sync_* builtins for a
> while.  This change in default behavior only hurts performance for applications
> on POWER relative to all other architectures, which sucks. :-(

Yes, I see that.  But my question is - did a developer reading the
documentation
get _correct_ code on POWER (which uses a laxer memory model than documented!)
in all cases?  Or can you construct a testcase that works fine on IA64
while surprisingly (after reading docs) does not work on POWER?  Thus, didn't
we simply fix a wrong-code bug (albeit by producing slower code)?