[Bug tree-optimization/56878] Issue with candidate choice in vect_gen_niters_for_prolog_loop.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56878



--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-10 
14:58:00 UTC ---

Hmm, it seems to be a rather unfortunate choice of aligning x vs. v1 which

seems to only depend on the order in which the data-refs appear in the

data-ref array.  That is, vect_enhance_data_refs_alignment should

prefer amongst several choices that are otherwise equal the data-ref

that has the most invariant base address.