Re: [PATCH][RFC] tree-optimization/92335 - Improve sinking heuristics for vectorization

[Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>]

On Mon, 21 Aug 2023 at 12:27, Richard Biener <rguenther@suse.de> wrote:
>
> On Sat, 19 Aug 2023, Prathamesh Kulkarni wrote:
>
> > On Fri, 18 Aug 2023 at 17:11, Richard Biener <rguenther@suse.de> wrote:
> > >
> > > On Fri, 18 Aug 2023, Richard Biener wrote:
> > >
> > > > On Thu, 17 Aug 2023, Prathamesh Kulkarni wrote:
> > > >
> > > > > On Tue, 15 Aug 2023 at 14:28, Richard Sandiford
> > > > > <richard.sandiford@arm.com> wrote:
> > > > > >
> > > > > > Richard Biener <rguenther@suse.de> writes:
> > > > > > > On Mon, 14 Aug 2023, Prathamesh Kulkarni wrote:
> > > > > > >> On Mon, 7 Aug 2023 at 13:19, Richard Biener <richard.guenther@gmail.com> wrote:
> > > > > > >> > It doesn't seem to make a difference for x86.  That said, the "fix" is
> > > > > > >> > probably sticking the correct target on the dump-check, it seems
> > > > > > >> > that vect_fold_extract_last is no longer correct here.
> > > > > > >> Um sorry, I did go thru various checks in target-supports.exp, but not
> > > > > > >> sure which one will be appropriate for this case,
> > > > > > >> and am stuck here :/ Could you please suggest how to proceed ?
> > > > > > >
> > > > > > > Maybe Richard S. knows the magic thing to test, he originally
> > > > > > > implemented the direct conversion support.  I suggest to implement
> > > > > > > such dg-checks if they are not present (I can't find them),
> > > > > > > possibly quite specific to the modes involved (like we have
> > > > > > > other checks with _qi_to_hi suffixes, for float modes maybe
> > > > > > > just _float).
> > > > > >
> > > > > > Yeah, can't remember specific selectors for that feature.  TBH I think
> > > > > > most (all?) of the tests were AArch64-specific.
> > > > > Hi,
> > > > > As Richi mentioned above, the test now vectorizes on AArch64 because
> > > > > it has support for direct conversion
> > > > > between vectors while x86 doesn't. IIUC this is because
> > > > > supportable_convert_operation returns true
> > > > > for V4HI -> V4SI on Aarch64 since it can use extend_v4hiv4si2 for
> > > > > doing the conversion ?
> > > > >
> > > > > In the attached patch, I added a new target check vect_extend which
> > > > > (currently) returns 1 only for aarch64*-*-*,
> > > > > which makes the test PASS on both the targets, altho I am not sure if
> > > > > this is entirely correct.
> > > > > Does the patch look OK ?
> > > >
> > > > Can you make vect_extend more specific, say vect_extend_hi_si or
> > > > what is specifically needed here?  Note I'll have to investigate
> > > > why x86 cannot vectorize here since in fact it does have
> > > > the extend operation ... it might be also worth splitting the
> > > > sign/zero extend case, so - vect_sign_extend_hi_si or
> > > > vect_extend_short_int?
> > >
> > > And now having anaylzed _why_ x86 doesn't vectorize it's rather
> > > why we get this vectorized with NEON which is because
> > >
> > > static opt_machine_mode
> > > aarch64_vectorize_related_mode (machine_mode vector_mode,
> > >                                 scalar_mode element_mode,
> > >                                 poly_uint64 nunits)
> > > {
> > > ...
> > >   /* Prefer to use 1 128-bit vector instead of 2 64-bit vectors.  */
> > >   if (TARGET_SIMD
> > >       && (vec_flags & VEC_ADVSIMD)
> > >       && known_eq (nunits, 0U)
> > >       && known_eq (GET_MODE_BITSIZE (vector_mode), 64U)
> > >       && maybe_ge (GET_MODE_BITSIZE (element_mode)
> > >                    * GET_MODE_NUNITS (vector_mode), 128U))
> > >     {
> > >       machine_mode res = aarch64_simd_container_mode (element_mode, 128);
> > >       if (VECTOR_MODE_P (res))
> > >         return res;
> > >
> > > which makes us get a V4SImode vector for a V4HImode loop vector_mode.
> > Thanks for the explanation!
> > >
> > > So I think the appropriate effective dejagnu target is
> > > aarch64-*-* (there's none specifically to advsimd, not sure if one
> > > can disable that?)
> > The attached patch uses aarch64*-*-* target check, and additionally
> > for SVE (and other targets supporting vect_fold_extract_last) it
> > checks
> > if the condition reduction was carried out using FOLD_EXTRACT_LAST.
> > Does that look OK ?
>
> Works for me.
Thanks, committed to trunk in dd606dc7c7e49feb7a900902ec6d35b421789173

Thanks,
Prathamesh
>
> Richard.
>
> > Thanks,
> > Prathamesh
> > >
> >
> > > Richard.
> > >
> > > > > Thanks,
> > > > > Prathamesh
> > > > > >
> > > > > > Thanks,
> > > > > > Richard
> > > > >
> > > >
> > > >
> > >
> > > --
> > > Richard Biener <rguenther@suse.de>
> > > SUSE Software Solutions Germany GmbH,
> > > Frankenstrasse 146, 90461 Nuernberg, Germany;
> > > GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
> >
>
> --
> Richard Biener <rguenther@suse.de>
> SUSE Software Solutions Germany GmbH,
> Frankenstrasse 146, 90461 Nuernberg, Germany;
> GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)