Disable vectorization on pre SSE2
chandlerc at gmail.com
Wed Mar 5 13:28:18 PST 2014
The bugs issue actually bothers me here.
I'm really alarmed that we're papering over an inability to scalarize
vector code. I think that's a bad sign.
Should the vectorizer produce it? No, because its slow.
Should the backend be able to handle it? *Yes*, and it is worth some amount
of pain to try to ensure this is the case.
On Wed, Mar 5, 2014 at 11:45 AM, Benjamin Kramer <benny.kra at gmail.com>wrote:
> On 05.03.2014, at 18:02, Arnold Schwaighofer <aschwaighofer at apple.com>
> > Hi all,
> > does anyone have objections against disabling vectorization on pre SSE2
> architectures. I don’t think we have bots testing this and so bugs are
> likely (see PR18036). We fail in isel for a simple vector select.
> This is the right thing to do, getting the bugs fixed would also be a good
> thing :)
> - Ben
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-commits