Re: Missed lowering to ld1rq from svld1rq for memory operand

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

On Thu, 12 Jan 2023 at 21:02, Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> > On Fri, 5 Aug 2022 at 17:49, Richard Sandiford
> > <richard.sandiford@arm.com> wrote:
> >>
> >> Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> >> > Hi Richard,
> >> > Following from off-list discussion, in the attached patch, I wrote pattern
> >> > similar to vec_duplicate<mode>_reg, which seems to work for the svld1rq tests.
> >> > Does it look OK ?
> >> >
> >> > Sorry, I didn't fully understand your suggestion on integrating with
> >> > vec_duplicate<mode>_reg
> >> > pattern. For vec_duplicate<mode>_reg, the operand to vec_duplicate expects
> >> > mode to be <VEL>, while the pattern in patch expects operand of
> >> > vec_duplicate to have mode <V128>.
> >> > How do we write a pattern so an operand can accept either of the 2 modes ?
> >>
> >> I quoted the wrong one, sorry, should have been
> >> aarch64_vec_duplicate_vq<mode>_le.
> >>
> >> > Also it seems <V128> cannot be used with SVE_ALL ?
> >>
> >> Yeah, these would be SVE_FULL only.
> > Hi Richard,
> > Sorry for the very late reply. I have attached patch, to integrate
> > with vec_duplicate_vq<mode>_le.
> > Bootstrapped+tested on aarch64-linux-gnu.
> > OK to commit ?
> >
> > Thanks,
> > Prathamesh
> >>
> >> Richard
> >>
> >
> > gcc/
> >       * config/aarch64/aarch64-sve.md (aarch64_vec_duplicate_vq<mode>_le):
> >       Change to define_insn_and_split to fold ldr+dup to ld1rq.
> >       * config/aarch64/predicates.md (aarch64_sve_dup_ld1rq_operand): New.
> >
> > testsuite/
> >       * gcc.target/aarch64/sve/acle/general/pr96463-2.c: Adjust.
> >
> > diff --git a/gcc/config/aarch64/aarch64-sve.md b/gcc/config/aarch64/aarch64-sve.md
> > index b8cc47ef5fc..4548375b8d6 100644
> > --- a/gcc/config/aarch64/aarch64-sve.md
> > +++ b/gcc/config/aarch64/aarch64-sve.md
> > @@ -2533,14 +2533,34 @@
> >  )
> >
> >  ;; Duplicate an Advanced SIMD vector to fill an SVE vector (LE version).
> > -(define_insn "@aarch64_vec_duplicate_vq<mode>_le"
> > -  [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
> > +
> > +(define_insn_and_split "@aarch64_vec_duplicate_vq<mode>_le"
> > +  [(set (match_operand:SVE_FULL 0 "register_operand" "=w, w")
> >       (vec_duplicate:SVE_FULL
> > -       (match_operand:<V128> 1 "register_operand" "w")))]
> > +       (match_operand:<V128> 1 "aarch64_sve_dup_ld1rq_operand" "w, UtQ")))
> > +   (clobber (match_scratch:VNx16BI 2 "=X, Upl"))]
> >    "TARGET_SVE && !BYTES_BIG_ENDIAN"
> >    {
> > -    operands[1] = gen_rtx_REG (<MODE>mode, REGNO (operands[1]));
> > -    return "dup\t%0.q, %1.q[0]";
> > +    switch (which_alternative)
> > +      {
> > +     case 0:
> > +       operands[1] = gen_rtx_REG (<MODE>mode, REGNO (operands[1]));
> > +       return "dup\t%0.q, %1.q[0]";
> > +     case 1:
> > +       return "#";
> > +     default:
> > +       gcc_unreachable ();
> > +      }
> > +  }
> > +  "&& MEM_P (operands[1])"
> > +  [(const_int 0)]
> > +  {
> > +    if (GET_CODE (operands[2]) == SCRATCH)
> > +      operands[2] = gen_reg_rtx (VNx16BImode);
> > +    emit_move_insn (operands[2], CONSTM1_RTX (VNx16BImode));
> > +    rtx gp = gen_lowpart (<VPRED>mode, operands[2]);
> > +    emit_insn (gen_aarch64_sve_ld1rq<mode> (operands[0], operands[1], gp));
> > +    DONE;
> >    }
> >  )
> >
> > diff --git a/gcc/config/aarch64/predicates.md b/gcc/config/aarch64/predicates.md
> > index ff7f73d3f30..6062f37025e 100644
> > --- a/gcc/config/aarch64/predicates.md
> > +++ b/gcc/config/aarch64/predicates.md
> > @@ -676,6 +676,10 @@
> >    (ior (match_operand 0 "register_operand")
> >         (match_operand 0 "aarch64_sve_ld1r_operand")))
> >
> > +(define_predicate "aarch64_sve_dup_ld1rq_operand"
> > +  (ior (match_operand 0 "register_operand")
> > +       (match_operand 0 "aarch64_sve_ld1rq_operand")))
> > +
> >  (define_predicate "aarch64_sve_ptrue_svpattern_immediate"
> >    (and (match_code "const")
> >         (match_test "aarch64_sve_ptrue_svpattern_p (op, NULL)")))
> > diff --git a/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr96463-2.c b/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr96463-2.c
> > index 196de3f5e0a..c38204e6874 100644
> > --- a/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr96463-2.c
> > +++ b/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr96463-2.c
> > @@ -26,4 +26,4 @@ TEST(svfloat64_t, float64_t, f64)
> >
> >  TEST(svbfloat16_t, bfloat16_t, bf16)
> >
> > -/* { dg-final { scan-assembler-times {\tdup\tz[0-9]+\.q, z[0-9]+\.q\[0\]} 12 { target aarch64_little_endian } } } */
> > +/* { dg-final { scan-assembler-not {\tdup\t} } } */
>
> It would be good to add something like:
>
> /* { dg-final { scan-assembler-times {\tld1rq\t} 12 } } */
>
> (I assume it'll pass for both endiannesses, but please check!),
> in addition to the scan-assembler-not.
>
> OK with that change, thanks.
Thanks, committed the patch in
a3b99b84609af310c72b4d6221621f5b63a3c169 after adjusting the
test-case,
and verifying that we generate ld1rq for big endian targets, and
bootstrap+test on aarch64-linux-gnu.

Thanks,
Prathamesh
>
> Richard