http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52298
--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-20 14:14:48 UTC --- Not sure why in vect_model_reduction_cost we do not consider the reduction at all if nested_in_vect_loop_p (loop, orig_stmt). I think simply dropping that conditional would model the cost more appropriately (though the question is how outer loop vectorization uses the costs).
