http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46488
--- Comment #33 from rguenther at suse dot de <rguenther at suse dot de> 2010-11-30 15:41:36 UTC --- On Tue, 30 Nov 2010, ebotcazou at gcc dot gnu.org wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46488 > > --- Comment #32 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-30 > 15:10:58 UTC --- > The problem appears to be deeply rooted in the Ring construct, more precisely > in the HEAD trick. IIUC the idea is to "attach" a doubly-linked list to > another structure by means of a "virtual" member overlaid on top of the > structure; the only thing they actually share is a special APR_RING_ENTRY (the > APR_RING_HEAD). But this overlay fundamentally violates the aliasing rules > even if one try to narrow the accesses to just the shared part. > > Richard, is that how the aliasing rules are implemented in the 4.5.x series? Yes, also in 4.4 and 4.3, but maybe you need to be more lucky to trigger the problem there. > Has this been changed in 4.6.0? No. But with 4.6 we can ignore pointer types when doing copy-propagation and thus we probably see that they must-alias (in which case we will not apply TBAA to be more nice to our users). Richard.
