On Wed, 27 Aug 2025, Jakub Jelinek wrote:
> On Wed, Aug 27, 2025 at 03:50:26PM +0200, Richard Biener wrote:
> > Ah, thanks for the write-up how to compute the number of uses.
> >
> > I'll test the following.
> > --- a/gcc/tree-vect-patterns.cc
> > +++ b/gcc/tree-vect-patterns.cc
> > @@ -4303,6 +4303,8 @@ vect_synth_mult_by_constant (vec_info *vinfo, tree
> > op, tree val,
> > /* Targets that don't support vector shifts but support vector additions
> > can synthesize shifts that way. */
> > bool synth_shift_p = !vect_supportable_shift (vinfo, LSHIFT_EXPR,
> > multtype);
> > + if (synth_shift_p)
> > + return NULL;
>
> This is wrong, that will punt no matter what, for reductions/non-reductions,
> and even when not really used. Just the other hunk should do it.
Oops, that was supposed to be reverted.
> Also, I'd probably do it only after the target_supports_mult_synth_alg
> check, not before it.
I guess that doesn't really matter, I expect the reduction case to be
a) uncommon and b) the check to be faster?
I'm testing with the above hunk reverted now (I'll note no FAILs with
it on x86, so I guess either we have shifts or mults for all component
modes or test coverage is weak).
Richard.
> > HOST_WIDE_INT hwval = tree_to_shwi (val);
> > /* Use MAX_COST here as we don't want to limit the sequence on rtx costs.
> > @@ -4314,6 +4316,35 @@ vect_synth_mult_by_constant (vec_info *vinfo, tree
> > op, tree val,
> > if (!possible)
> > return NULL;
> >
> > + if (vect_is_reduction (stmt_vinfo))
> > + {
> > + int op_uses = alg.op[0] != alg_zero;
> > + for (int i = 1; i < alg.ops; i++)
> > + switch (alg.op[i])
> > + {
> > + case alg_add_t_m2:
> > + case alg_sub_t_m2:
> > + if (synth_shift_p && alg.log[i])
> > + return NULL;
> > + else
> > + op_uses++;
> > + break;
> > + case alg_add_t2_m:
> > + case alg_sub_t2_m:
> > + op_uses++;
> > + break;
> > + default:
> > + break;
> > + }
> > + if (variant == add_variant)
> > + op_uses++;
> > + /* When we'll synthesize more than a single use of the reduction
> > + operand the reduction constraints are violated. Avoid this
> > + situation. */
> > + if (op_uses > 1)
> > + return NULL;
> > + }
> > +
> > if (!target_supports_mult_synth_alg (&alg, variant, vectype,
> > synth_shift_p))
> > return NULL;
> >
> > --
> > 2.43.0
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
