http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57584
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> --- niter analysis computes (((unsigned long) sp_35(ab) - (unsigned long) (sp_3(ab) + 8)) + 7) / 8)) which the vectorizer inserts, obviously breaking coalescing of sp_35 and sp_3. It's expand_simple_operations done in niter analysis that breaks this and I have a patch to fix that.