Re: [committed 0/6] amdgcn: Add V32, V16, V8, V4, and V2 vectors

[Richard Biener <richard.guenther@gmail.com>]

On Tue, Oct 11, 2022 at 1:53 PM Andrew Stubbs <ams@codesourcery.com> wrote:
>
> On 11/10/2022 12:29, Richard Biener wrote:
> > On Tue, Oct 11, 2022 at 1:03 PM Andrew Stubbs <ams@codesourcery.com> wrote:
> >>
> >> This patch series adds additional vector sizes for the amdgcn backend.
> >>
> >> The hardware supports any arbitrary vector length up to 64-lanes via
> >> masking, but GCC cannot (yet) make full use of them due to middle-end
> >> limitations.  Adding smaller "virtual" vector sizes increases the
> >> complexity of the backend a little, but opens up optimization
> >> opportunities for the current middle-end implementation somewhat. In
> >> particular, it enables many more cases of SLP optimization.
> >>
> >> The patchset gives aproximately 100 addtional test PASS and a few extra
> >> FAIL.  However, the failures are not new issues, but rather existing
> >> problems that did not show up because the code did not previously
> >> vectorize.  Expanding the testcase to allow 64-lane vectors shows the
> >> same problems there.
> >>
> >> I shall backport these patches to the OG12 branch shortly.
> >
> > I suppose until you change the related_vector_mode hook the PR107096 issue
> > will not hit you but at least it's then latent ...
>
> How do you mean, change it?
>
> static opt_machine_mode
> gcn_related_vector_mode (machine_mode vector_mode,
>                           scalar_mode element_mode, poly_uint64 nunits)
> {
>    int n = nunits.to_constant ();
>
>    if (n == 0)
>      n = GET_MODE_NUNITS (vector_mode);
>
>    return VnMODE (n, element_mode);
> }
>
>
> It returns what it's asked for, always matching the number of lanes (not
> the bitsize), which is most likely the most natural for GCN.

Yes, change it in any way no longer honoring that.  Or discover the
case (not sure if it actually exists) where the vectorizer itself tricks
you into this by passing down nunits !=0 when vectorizing a loop
(I _think_ that's only done for basic-block vectorization currently).

Richard.

>
> Andrew