Sorry for the wrong email earlier.
│
│
Any opinion on this, Richard S?
BRs,
Lin
-----Original Message-----
From: Richard Biener <[email protected]>
Sent: Wednesday, June 17, 2026 5:11 PM
To: Hu, Lin1 <[email protected]>
Cc: [email protected]; [email protected]
Subject: RE: [PATCH 2/3] dojump: use simplify_expand_binop for vector boolean
mask
On Wed, 17 Jun 2026, Hu, Lin1 wrote:
> Thanks for the review, I hadn't considered this initially.
>
> I think we don't need to worry the question. On this path the mode is
> SCALAR_INT_MODE_P. So emit_cmp_and_jump_insns goes through
> prepare_cmp_insn (optabs.cc:4998). emit_cmp_and_jump_insns force_reg
> op0 when it's a constant (optabs.cc:4989), and prepare_cmp_insn calls
> prepare_operand on both op0 and op1 (optabs.cc:4726-4727);
> prepare_operand copies the operand into a new reg via copy_to_mode_reg
> when it doesn't match the target's cbranch predicate (optabs.cc:4834).
> So if a target requires both operands to be REG, the downstream path
> already handles it.
Thanks for double-checking.
Richard.
> BRs,
> Lin
> -----Original Message-----
> From: Richard Biener <[email protected]>
> Sent: Wednesday, June 17, 2026 3:56 PM
> To: Hu, Lin1 <[email protected]>
> Cc: [email protected]; [email protected]
> Subject: RE: [PATCH 2/3] dojump: use simplify_expand_binop for vector
> boolean mask
>
> On Wed, 17 Jun 2026, Hu, Lin1 wrote:
>
> > Due to the patch modify middle-end.
> > + Richard Biener and Richard Sandiford
>
> LGTM, but please leave others the chance to chime in. I think the current
> code ensures op0 and op1 are a REG_P and emit_cmp_and_jump_insns might not be
> able to force_reg if the target requires?
>
> > Thanks.
> >
> > BRs,
> > Lin
> >
> > -----Original Message-----
> > From: Hu, Lin1 <[email protected]>
> > Sent: Friday, June 5, 2026 2:26 PM
> > To: [email protected]
> > Cc: Liu, Hongtao <[email protected]>; [email protected]
> > Subject: [PATCH 2/3] dojump: use simplify_expand_binop for vector
> > boolean mask
> >
> > When either op0 or op1 is a constant, expand_binop can emit an extra AND
> > during expansion. Use simplify_expand_binop here to avoid creating it in
> > the first place.
> >
> > gcc/ChangeLog:
> >
> > * dojump.cc (do_compare_rtx_and_jump): Use simplify_expand_binop
> > instead of expand_binop.
> > ---
> > gcc/dojump.cc | 11 +++++------
> > 1 file changed, 5 insertions(+), 6 deletions(-)
> >
> > diff --git a/gcc/dojump.cc b/gcc/dojump.cc index
> > 599a546b0de..1ed22116a27 100644
> > --- a/gcc/dojump.cc
> > +++ b/gcc/dojump.cc
> > @@ -1244,12 +1244,11 @@ do_compare_rtx_and_jump (rtx op0, rtx op1, enum
> > rtx_code code, int unsignedp,
> > auto nunits = TYPE_VECTOR_SUBPARTS (TREE_TYPE (val)).to_constant ();
> > if (maybe_ne (GET_MODE_PRECISION (mode), nunits))
> > {
> > - op0 = expand_binop (mode, and_optab, op0,
> > - GEN_INT ((HOST_WIDE_INT_1U << nunits) - 1),
> > - NULL_RTX, true, OPTAB_WIDEN);
> > - op1 = expand_binop (mode, and_optab, op1,
> > - GEN_INT ((HOST_WIDE_INT_1U << nunits) - 1),
> > - NULL_RTX, true, OPTAB_WIDEN);
> > + rtx mask_imm = GEN_INT ((HOST_WIDE_INT_1U << nunits) - 1);
> > + op0 = simplify_expand_binop (mode, and_optab, op0, mask_imm,
> > + NULL_RTX, true, OPTAB_WIDEN);
> > + op1 = simplify_expand_binop (mode, and_optab, op1, mask_imm,
> > + NULL_RTX, true, OPTAB_WIDEN);
> > }
> > }
> >
> > --
> > 2.31.1
> >
> >
>
> --
> Richard Biener <[email protected]>
> SUSE Software Solutions Germany GmbH,
> Frankenstrasse 146, 90461 Nuernberg, Germany;
> GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG
> Nuernberg)
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
