On Wed, 22 Jul 2026, Liu, Hongtao wrote:
>
>
> > -----Original Message-----
> > From: Richard Biener <[email protected]>
> > Sent: Tuesday, July 21, 2026 9:52 PM
> > To: [email protected]
> > Cc: Liu, Hongtao <[email protected]>; [email protected]
> > Subject: [PATCH] target/126328 - [x86] insn_cost of
> > *sse3_h{add,sub}v2df3[_low]
> >
> > I run into issues with gcc.target/i386/pr54400.c when SLP vectorizing a
> > horizontal reduction of a V2DF vector. The following plugs the remaining
> > hole
> > in costing which prevents combine from doing its work to recover the single
> > hadd instruction for { p[0] + p[1], q[0] + q[1] }
> >
> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
> >
> > OK for trunk if testing succeeds?
> >
> > To have an effect it depends on the prerequesite [v2] Simplify
> > (vec_select:<scalar> (vec_concat ..)) and unless Support two-lane vector BB
> > reductions without target support is installed gcc.target/i386/pr54400.c is
> > passing anyway.
> >
> > PR target/126328
> > * config/i386/i386.cc (ix86_insn_cost): Match
> > *sse3_h{add,sub}v2df3[_low] and cost it like an add.
> > ---
> > gcc/config/i386/i386.cc | 21 +++++++++++++++++++++
> > 1 file changed, 21 insertions(+)
> >
> > diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc index
> > 56a90333cfc..ab04290bf7c 100644
> > --- a/gcc/config/i386/i386.cc
> > +++ b/gcc/config/i386/i386.cc
> > @@ -22714,6 +22714,27 @@ ix86_insn_cost (rtx_insn *insn, bool speed)
> > : COSTS_N_INSNS (3) + 1;
> > }
> > }
> > + if (rtx set = single_set (insn))
> > + {
> > + rtx src = SET_SRC (set);
> > + /* ix86_rtx_cost makes VEC_CONCAT artificially cheap by not recursing
> > + into operands. That makes *sse3_haddv2df3 win over
> > + *sse3_haddv2df3_low, preventing combine from doing its work for
> > + gcc.target/i386/pr54400.c. Cost both the same explicitly. */
> > + if ((GET_MODE (src) == V2DFmode
> > + && GET_CODE (src) == VEC_CONCAT
> > + && (GET_CODE (XEXP (src, 0)) == PLUS
> > + || GET_CODE (XEXP (src, 0)) == MINUS)
> > + && GET_CODE (XEXP (XEXP (src, 0), 0)) == VEC_SELECT)
> > + || (GET_MODE (src) == DFmode
> > + && (GET_CODE (src) == PLUS || GET_CODE (src) == MINUS)
> > + && GET_CODE (XEXP (src, 0)) == VEC_SELECT
> > + && GET_CODE (XEXP (src, 1)) == VEC_SELECT))
> > + /* Assume this is only recognized as *sse3_h{add,sub}v2df3[_low].
> > */
> > + return insn_cost + ix86_vec_cost (GET_MODE (pat),
> > + (speed ? ix86_tune_cost
> > + : &ix86_size_cost)->addss);
>
> GET_MODE (pat) returns VOIDmode since pat is a SET - you likely intended to
> use V2DFmode here. Alternatively, we could simply drop ix86_vec_cost since it
> is only used in the vectorization cost model and not in rtx_cost.
> In that case, we could just return:
> insn_cost + (speed ? ix86_tune_cost : &ix86_size_cost)->addss;
Ah yes, I missed that GET_MODE, I'll adjust as you suggest in case the
combine change is rejected.
> Also, regarding the fix you proposed for combine [1] - with that fix in
> place, would it be possible to use insn_attr instead of structural pattern
> matching in this patch?
Yes. We'd need to add some appropriate attribute to the
*sse3_h{add,sub}v2df3 patterns though. I do prefer that route, let's
see if I get any feedback on that.
I'll probably push the vectorizer change today, so the testcase with
fail intermittently as we sort this out (so it's easier for folks to
see what happens themselves if they want).
Richard.
> [1] https://gcc.gnu.org/pipermail/gcc-patches/2026-July/724837.html
>
> > + }
> >
> > return insn_cost + pattern_cost (pat, speed); }
> > --
> > 2.51.0
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Abhinav Puri; (HRB 36809, AG Nuernberg)
