On Thu, 25 Mar 2021, Richard Sandiford wrote: > Richard Biener <rguent...@suse.de> writes: > > Whether a vector DR is aligned or not depends on the vectorization > > scheme but at the point of computing peeling for alignment > > that has not been determined so we try to work with scalar DRs. > > > > For the uses by vectorizable_load/store I pondered to pass > > the memory_access_type down to vect_supportable_dr_alignment but > > that obviously doesn't work for peeling. > > > > I suppose we should delay peeling/versioning for alignment until > > after vectorizable_load/store computes the actual vector DRs > > and align/version those (won't versioning for alias have the > > same issues?). > > Yeah, I guess it's a bit chicken-and-egg: vectorizable_load/store > need to know whether the operation can be aligned by peeling before > deciding how to implement it, whereas peeling needs to know how > vectorizable_load/store would implement something in order to > know what alignment is needed.
Indeed. I think it's easier to implement peeling/versioning in a correct way if we work on vector DRs which means doing some optimistic guessing when vectorizing loads/stores (which means _not_ going elementwise loads/stores on targets that cannot do unaligned accesses, but assuming we can somehow re-align them). After peeling we then can have a meaningful validity check of the vector DRs, aka whether they are supported by the target. That of couse heavily sides with "modern" targets that can do unaligned accesses. I suppose we don't have many left that can't (and those do not get much vectorization anyway). Richard.
