http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48702
--- Comment #6 from Zdenek Dvorak <rakdver at gcc dot gnu.org> 2011-04-21 11:49:36 UTC --- I think we rely on the assumption that pointer arithmetics satisfies the restrictions of C standard (i.e., that the pointer is within a single memory object) on more places than just this one. So, if the code produced by ivopts is indeed for (p = &a; p != &a - 3; --p) *(p + 3) = ... this could potentially cause other problems other than with alias analysis. But, I would expect the code from ivopts to be rather for (p = &a; ...; p = (pointer) ((unsigned) p - 1) *((pointer) ((unsigned) p + 3)) = ... Which should be a valid code. Is that the case? If so, fixing alias analysis would be the proper solution.
