RFC: Modeling horizontal vector reductions

[Renato Golin]

On 12 September 2013 15:00, Arnold Schwaighofer <aschwaighofer at apple.com>wrote:

> The shuffles for the first form (splitting of vectors) is:
>
> (0,1) <- This one is free.
> (2,3)
>
> while for the second form (pairwise adds) is
> (0,2)
> (1,3)
>

Oh, my fault, I read it wrongly. I thought both forms were the same, but
the first IR was producing nop-shuffles.

I agree we need this kind of knowledge in the cost-model, and that it'd
benefit more than one arch. I'll check the patch later on...

I'm also not sure about unsafe-math in the isel, but it strikes me as an
important feature to have in. Even if one flag is deprecated, we'll need
some form of knowing that during selection, right?

cheers,
--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130912/888d7e23/attachment.html>