On 05/20/2013 04:14 PM, Jakub Jelinek wrote:
Isn't this ABI incompatible change? See http://gcc.gnu.org/PR56564 for more info (yeah, different target, but looks exactly the same issue). If what this new DATA_ALIGNMENT value is optimization rather than an ABI requirement, then you'll hit the same issue.
Hmmmm. The originally-posted version of the patch from 2011 (which is the one we've been using locally since then) only changed LOCAL_ALIGNMENT, which is presumably safe. I only extended it to DATA_ALIGNMENT because of the previous review comment that it ought to apply to all arrays and not just local ones. :-S I did wonder about ABI issues when I saw the extra stuff the generic vector increase_alignment pass does (in vect_can_force_dr_alignment_p), but the documentation for DATA_ALIGNMENT in tm.texi seems to suggest it's fine to use it for this kind of optimization purposes.
To tell the truth, we don't care enough about this patch to be motivated to spend a lot of time on a general solution to PR56564, but I could go back to hacking LOCAL_ALIGNMENT only.... OTOH, if the consensus is that the whole idea is bad or that this should be done in the generic vectorizer instead of target code, we'll drop the patch.
-Sandra
