Re: [SVE] PR86753

[Prathamesh Kulkarni <prathamesh dot kulkarni at linaro dot org>]

On Wed, 16 Oct 2019 at 04:19, Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Richard Biener <richard.guenther@gmail.com> writes:
> > On Tue, Oct 15, 2019 at 8:07 AM Prathamesh Kulkarni
> > <prathamesh.kulkarni@linaro.org> wrote:
> >>
> >> On Wed, 9 Oct 2019 at 08:14, Prathamesh Kulkarni
> >> <prathamesh.kulkarni@linaro.org> wrote:
> >> >
> >> > On Tue, 8 Oct 2019 at 13:21, Richard Sandiford
> >> > <richard.sandiford@arm.com> wrote:
> >> > >
> >> > > Leaving the main review to Richard, just some comments...
> >> > >
> >> > > Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> >> > > > @@ -9774,6 +9777,10 @@ vect_is_simple_cond (tree cond, vec_info *vinfo,
> >> > > >
> >> > > >     When STMT_INFO is vectorized as a nested cycle, for_reduction is true.
> >> > > >
> >> > > > +   For COND_EXPR<C, T, E> if T comes from masked load, and is conditional
> >> > > > +   on C, we apply loop mask to result of vector comparison, if it's present.
> >> > > > +   Similarly for E, if it is conditional on !C.
> >> > > > +
> >> > > >     Return true if STMT_INFO is vectorizable in this way.  */
> >> > > >
> >> > > >  bool
> >> > >
> >> > > I think this is a bit misleading.  But IMO it'd be better not to have
> >> > > a comment here and just rely on the one in the main function body.
> >> > > This optimisation isn't really changing the vectorisation strategy,
> >> > > and the comment could easily get forgotten if things change in future.
> >> > >
> >> > > > [...]
> >> > > > @@ -9999,6 +10006,35 @@ vectorizable_condition (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
> >> > > >    /* Handle cond expr.  */
> >> > > >    for (j = 0; j < ncopies; j++)
> >> > > >      {
> >> > > > +      tree loop_mask = NULL_TREE;
> >> > > > +      bool swap_cond_operands = false;
> >> > > > +
> >> > > > +      /* Look up if there is a loop mask associated with the
> >> > > > +      scalar cond, or it's inverse.  */
> >> > >
> >> > > Maybe:
> >> > >
> >> > >    See whether another part of the vectorized code applies a loop
> >> > >    mask to the condition, or to its inverse.
> >> > >
> >> > > > +
> >> > > > +      if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
> >> > > > +     {
> >> > > > +       scalar_cond_masked_key cond (cond_expr, ncopies);
> >> > > > +       if (loop_vinfo->scalar_cond_masked_set.contains (cond))
> >> > > > +         {
> >> > > > +           vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
> >> > > > +           loop_mask = vect_get_loop_mask (gsi, masks, ncopies, vectype, j);
> >> > > > +         }
> >> > > > +       else
> >> > > > +         {
> >> > > > +           bool honor_nans = HONOR_NANS (TREE_TYPE (cond.op0));
> >> > > > +           cond.code = invert_tree_comparison (cond.code, honor_nans);
> >> > > > +           if (loop_vinfo->scalar_cond_masked_set.contains (cond))
> >> > > > +             {
> >> > > > +               vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
> >> > > > +               loop_mask = vect_get_loop_mask (gsi, masks, ncopies,
> >> > > > +                                               vectype, j);
> >> > > > +               cond_code = cond.code;
> >> > > > +               swap_cond_operands = true;
> >> > > > +             }
> >> > > > +         }
> >> > > > +     }
> >> > > > +
> >> > > >        stmt_vec_info new_stmt_info = NULL;
> >> > > >        if (j == 0)
> >> > > >       {
> >> > > > @@ -10114,6 +10153,47 @@ vectorizable_condition (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
> >> > > >                   }
> >> > > >               }
> >> > > >           }
> >> > > > +
> >> > > > +       /* If loop mask is present, then AND it with
> >> > >
> >> > > Maybe "If we decided to apply a loop mask, ..."
> >> > >
> >> > > > +          result of vec comparison, so later passes (fre4)
> >> > >
> >> > > Probably better not to name the pass -- could easily change in future.
> >> > >
> >> > > > +          will reuse the same condition used in masked load.
> >> > >
> >> > > Could be a masked store, or potentially other things too.
> >> > > So maybe just "will reuse the masked condition"?
> >> > >
> >> > > > +
> >> > > > +          For example:
> >> > > > +          for (int i = 0; i < 100; ++i)
> >> > > > +            x[i] = y[i] ? z[i] : 10;
> >> > > > +
> >> > > > +          results in following optimized GIMPLE:
> >> > > > +
> >> > > > +          mask__35.8_43 = vect__4.7_41 != { 0, ... };
> >> > > > +          vec_mask_and_46 = loop_mask_40 & mask__35.8_43;
> >> > > > +          _19 = &MEM[base: z_12(D), index: ivtmp_56, step: 4, offset: 0B];
> >> > > > +          vect_iftmp.11_47 = .MASK_LOAD (_19, 4B, vec_mask_and_46);
> >> > > > +          vect_iftmp.12_52 = VEC_COND_EXPR <vec_mask_and_46,
> >> > > > +                                            vect_iftmp.11_47, { 10, ... }>;
> >> > > > +
> >> > > > +          instead of recomputing vec != { 0, ... } in vec_cond_expr  */
> >> > >
> >> > > That's true, but gives the impression that avoiding the vec != { 0, ... }
> >> > > is the main goal, whereas we could do that just by forcing a three-operand
> >> > > COND_EXPR.  It's really more about making sure that vec != { 0, ... }
> >> > > and its masked form aren't both live at the same time.  So maybe:
> >> > >
> >> > >              instead of using a masked and unmasked forms of
> >> > >              vect__4.7_41 != { 0, ... } (masked in the MASK_LOAD,
> >> > >              unmasked in the VEC_COND_EXPR).  */
> >> > >
> >> > Hi Richard,
> >> > Thanks for the suggestions, I have updated comments in the attached patch.
> >> Hi,
> >> The attached patch is rebased on trunk, and after PR91532 fix, the
> >> hunk for fmla_2.c is no
> >> longer required.
> >
> > Hmm.  So we already record some mask info - you just add in addition
> > to that the scalar predicate representing the mask.  I wonder if you can
> > integrate that into the existing vec_loop_masks vector instead of
> > adding another data structure on the side?  Not that I am understanding
> > the existing fully masked code at all (or specifically what it computes
> > as nscalars_per_iter, etc. ... :/).
>
> We can AND several different scalar conditions with the same loop
> mask (that's relatively common), and could even AND the same scalar
> condition with different loop masks (although that's less likely).
> So I think having separate info makes sense.
>
> > At least add the new vinfo member right to the other masks related
> > field.
>
> Agree that would be better.
>
> > @@ -10122,6 +10157,48 @@ vectorizable_condition (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
> >                   }
> >               }
> >           }
> > +
> > +       /* If we decided to apply a loop mask to result of vec
> > +          comparison, so later passes will reuse the same condition.
>
> Maybe:
>
>           /* If we decided to apply a loop mask to the result of the vector
>              comparison, AND the comparison with the mask now.  Later passes
>              should then be able to reuse the AND results between mulitple
>              vector statements.
>
> > +          For example:
> > +          for (int i = 0; i < 100; ++i)
> > +            x[i] = y[i] ? z[i] : 10;
> > +
> > +          results in following optimized GIMPLE:
> > +
> > +          mask__35.8_43 = vect__4.7_41 != { 0, ... };
> > +          vec_mask_and_46 = loop_mask_40 & mask__35.8_43;
> > +          _19 = &MEM[base: z_12(D), index: ivtmp_56, step: 4, offset: 0B];
> > +          vect_iftmp.11_47 = .MASK_LOAD (_19, 4B, vec_mask_and_46);
> > +          vect_iftmp.12_52 = VEC_COND_EXPR <vec_mask_and_46,
> > +                                            vect_iftmp.11_47, { 10, ... }>;
> > +
> > +          instead of using a masked and unmasked forms of
> > +             vec != { 0, ... } (masked in the MASK_LOAD,
> > +             unmasked in the VEC_COND_EXPR).  */
>
> The last paragraph uses some space rather than tab indentation.
>
> > +/* If code(T) is comparison op or def of comparison stmt,
> > +   extract it's operands.
> > +   Else return <NE_EXPR, T, 0>.  */
> > +
> > +void
> > +scalar_cond_masked_key::get_cond_ops_from_tree (tree t)
> > +{
>
> Maybe:
>
> /* If the condition represented by T is a comparison or the SSA name
>    result of a comparison, extract the comparison's operands.  Represent
>    T as NE_EXPR <T, 0> otherwise.  */
>
> OK with those changes and the one Richard asked for.
Thanks! Committed in r277141.

Thanks,
Prathamesh
>
> Thanks,
> Richard