Hi Richi,
> -----Original Message-----
> From: rguent...@c653.arch.suse.de <rguent...@c653.arch.suse.de> On
> Behalf Of Richard Biener
> Sent: Wednesday, November 4, 2020 1:36 PM
> To: Tamar Christina <tamar.christ...@arm.com>
> Cc: gcc-patches@gcc.gnu.org; nd <n...@arm.com>; o...@ucw.cz
> Subject: Re: [PATCH v2 10/18]middle-end simplify lane permutes which
> selects from loads from the same DR.
>
> On Tue, 3 Nov 2020, Tamar Christina wrote:
>
> > Hi All,
> >
> > This change allows one to simplify lane permutes that select from
> > multiple load leafs that load from the same DR group by promoting the
> > VEC_PERM node into a load itself and pushing the lane permute into it as a
> load permute.
> >
> > This saves us from having to calculate where to materialize a new load node.
> > If the resulting loads are now unused they are freed and are removed
> > from the graph.
> >
> > This allows us to handle cases where we would have generated:
> >
> > movi v4.4s, 0
> > adrp x3, .LC0
> > ldr q5, [x3, #:lo12:.LC0]
> > mov x3, 0
> > .p2align 3,,7
> > .L2:
> > mov v0.16b, v4.16b
> > mov v3.16b, v4.16b
> > ldr q1, [x1, x3]
> > ldr q2, [x0, x3]
> > fcmla v0.4s, v2.4s, v1.4s, #0
> > fcmla v3.4s, v1.4s, v2.4s, #0
> > fcmla v0.4s, v2.4s, v1.4s, #270
> > fcmla v3.4s, v1.4s, v2.4s, #270
> > mov v1.16b, v3.16b
> > tbl v0.16b, {v0.16b - v1.16b}, v5.16b
> > str q0, [x2, x3]
> > add x3, x3, 16
> > cmp x3, 1600
> > bne .L2
> > ret
> >
> > and instead generate
> >
> > mov x3, 0
> > .p2align 3,,7
> > .L27:
> > ldr q0, [x2, x3]
> > ldr q1, [x0, x3]
> > ldr q2, [x1, x3]
> > fcmla v0.2d, v1.2d, v2.2d, #0
> > fcmla v0.2d, v1.2d, v2.2d, #270
> > str q0, [x2, x3]
> > add x3, x3, 16
> > cmp x3, 512
> > bne .L27
> > ret
> >
> > This runs as a pre step such that permute simplification can still
> > inspect this permute is needed
> >
> > Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> > Tests are included as part of the final patch as they need the SLP
> > pattern matcher to insert permutes in between.
> >
> > Ok for master?
>
> So I think this is too specialized for the general issue that we're doing a
> bad
> job in CSEing the load part of different permutes of the same group. I've
> played with fixing this half a year ago (again) in multiple general ways but
> they all caused some regressions.
>
> So you're now adding some heuristics as to when to anticipate "CSE" (or
> merging with followup permutes).
>
> To quickly recap what I did consider two loads (V2DF) one { a[0], a[1] } and
> the other { a[1], a[0] }. They currently are two SLP nodes and one with a
> load_permutation.
> My original attempts focused on trying to get rid of load_permutation in
> favor of lane_permute nodes and thus during SLP discovery I turned the
> second into { a[0], a[1] } (magically unified with the other load) and a
> followup lane-permute node.
>
> So for your case you have IIUC { a[0], a[0] } and { a[1], a[1] } which
> eventually
> will (due to patterns) be lane-permuted into { a[0], a[1] }, right? So
> generalizing this as a single { a[0], a[1] } plus two lane-permute nodes {
> 0, 0 }
> and { 1, 1 } early would solve the issue as well?
Correct, I did wonder why it was generating two different nodes instead of a
lane
permute but didn't pay much attention that it was just a short coming.
> Now, in general it might be
> more profitable to generate the { a[0], a[0] } and { a[1], a[1] } via
> scalar-load-
> and-splat rather than vector load and permute so we have to be careful to
> not over-optimize here or be prepared to do the reverse transform.
This in principle can be done in optimize_slp then right? Since it would do
a lot of the same work already and find the materialization points.
>
> The patch itself is a bit ugly since it modifies the SLP graph when we already
> produced the graphds graph so I would do any of this before. I did consider
> gathering all loads nodes loading from a group and then trying to apply some
> heuristic to alter the SLP graph so it can be better optimized. In fact when
> we
> want to generate the same code as the non-SLP interleaving scheme does
> we do have to look at those since we have to unify loads there.
>
Yes.. I will concede the patch isn't my finest work.. I also don't like the
fact that I
had to keep leafs in tact less I break things later. But wanted feedback :)
> I'd put this after vect_slp_build_vertices but before the new_graph call -
> altering 'vertices' / 'leafs' should be more easily possible and the 'leafs'
> array
> contains all loads already (vect_slp_build_vertices could be massaged to
> provide a map from DR_GROUP_FIRST_ELEMENT to slp_tree, giving us the
> meta we want).
>
> That said, I'd like to see something more forward-looking rather than the ad-
> hoc special-casing of what you run into with the pattern matching.
>
Yeah, I like your suggestion about doing it at build time and CSEing early, but
don't think I can get that work in a week given that you've already tried
multiple times :)
Happy to give it a go next stage-1 opening though.
> In case we want to still go with the special-casing it should IMHO be done in
> a
> pre-order walk simply looking for lane permute nodes with children that all
> load from the same group performing what you do before any of the
> vertices/graph stuff is built. That's probably easiest at this point and it
> can be
> done when then bst_map is still around so you can properly CSE the new
> load you build.
That's fair enough. I do think I need a temporary (not terrible)
workaround...This would
then need to be somewhere in vect_analyze_slp. Would you prefer I do it during
the
construction of the instance of afterwards?
Regards,
Tamar
>
> Thanks,
> Richard.
>
>
>
> > Thanks,
> > Tamar
> >
> > gcc/ChangeLog:
> >
> > * tree-vect-slp.c (vect_optimize_slp): Promote permutes.
> >
> >
>
> --
> Richard Biener <rguent...@suse.de>
> SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409
> Nuernberg, Germany; GF: Felix Imend
