On Wed, Aug 15, 2012 at 2:53 PM, Jakub Jelinek <ja...@redhat.com> wrote: > On Wed, Aug 15, 2012 at 02:36:54PM +0200, Richard Guenther wrote: >> On Wed, Aug 15, 2012 at 2:29 PM, Jakub Jelinek <ja...@redhat.com> wrote: >> > On Wed, Aug 15, 2012 at 02:15:03PM +0200, Richard Guenther wrote: >> >> 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. >> > >> > Yeah, if there is a generic permute, can_vec_perm_p will return true always >> > for the modes for which it is available. Which is why I've been suggesting >> > also the target hook which should return false if only generic permute is >> > going to be used. >> >> Well. What about returning a cost instead? We don't want to transform >> two insns to four either. > > That could work, it would just be a lot of work to change it (and more > costly). E.g. on i386 for 16-byte vectors with certain ISAs we return true > right away, when returning cost we'd need to go the slow path always > even when the caller is just interested in a boolean flag whether it is > possible or not. > > CCing rth as the author of the hooks.
We'd want more accurate costs for the vectorizer cost model anyways. Richard. > Jakub
