On Wed, Aug 15, 2012 at 2:07 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> On Wed, Aug 15, 2012 at 01:46:05PM +0200, Richard Guenther wrote:
>> Well, we're waiting for someone to break the tie ... I'd go with the original
>> patch, improving the backends where necessary.
>
> E.g. i?86/x86_64 with just plain -msse2 has only very small subset of
> constant shuffles (and no variable shuffle), so by doing the transformation
> you could end up with a scalarized shuffle instead of two constant vector
> shuffles. Expecting the backend to figure it out and doing the two constant
> vector shuffles in every case is not realistic, i386/x86_64 has way too many
> different shuffling instructions (and worse different ISA levels have
> different subsets of them) and while a lot of effort has been spent on
> it already by Richard, me, Marc and others, we are nowhere close to having
> optimal sequence in many cases for various modes and ISA levels. Doing a
> brute force might be too expensive.
Ok. That would still leave us with the issue Ramana brought up - the
target hook returning true unconditionally if a generic permute is implemented.
We just avoid generic expansion by tree-vect-generic.c that way.
(I wonder if there exists some automated machinery that finds a path using
existing instructions to shuffle from {0, 1, ..., n } to the mask and if that
would be a viable way to handle shuffle expansions, or pre-compute them all)
Richard.
> Jakub
