[PATCH] i386: vcvtph2ps and vcvtps2ph should be used to convert _Float16 to SFmode with -mf16c [PR 102811]

[Hongtao Liu]

On Tue, Nov 16, 2021 at 4:15 PM Kong, Lingling via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Hi,
>
> vcvtph2ps and vcvtps2ph should be used to convert _Float16 to SFmode with -mf16c. So added define_insn extendhfsf2 and truncsfhf2 for target_f16c.
>
> OK for master?
>
> gcc/ChangeLog:
>
>         PR target/102811
>         * config/i386/i386.md (extendhfsf2): Add extenndhfsf2 for f16c.
>         (extendhfdf2): Split extendhf<mode>2 into separate extendhfsf2, extendhfdf2.
>         (truncsfhf2): Likewise.
>         (truncdfhf2): Likewise.
>
> gcc/testsuite/ChangeLog:
>
>         PR target/102811
>         * gcc.target/i386/avx512vl-vcvtps2ph-pr102811.c: New test.
> ---
>  gcc/config/i386/i386.md                       | 48 +++++++++++++++----
>  .../i386/avx512vl-vcvtps2ph-pr102811.c        | 10 ++++
>  2 files changed, 49 insertions(+), 9 deletions(-)  create mode 100644 gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-pr102811.c
>
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md index 6eb9de81921..c5415475342 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -4574,15 +4574,30 @@
>    emit_move_insn (operands[0], CONST0_RTX (V2DFmode));
>  })
>
> -(define_insn "extendhf<mode>2"
> -  [(set (match_operand:MODEF 0 "nonimm_ssenomem_operand" "=v")
> -        (float_extend:MODEF
> +(define_insn "extendhfsf2"
> +  [(set (match_operand:SF 0 "register_operand" "=v")
> +       (float_extend:SF
> +         (match_operand:HF 1 "nonimmediate_operand" "vm")))]
> +  "TARGET_AVX512FP16 || TARGET_F16C || TARGET_AVX512VL"
> +{
> +  if (TARGET_AVX512FP16)
> +    return "vcvtsh2ss\t{%1, %0, %0|%0, %0, %1}";
> +  else
> +    return "vcvtph2ps\t{%1, %0|%0, %1}"; }
> +  [(set_attr "type" "ssecvt")
> +   (set_attr "prefix" "maybe_evex")
> +   (set_attr "mode" "SF")])
> +
> +(define_insn "extendhfdf2"
> +  [(set (match_operand:DF 0 "nonimm_ssenomem_operand" "=v")
> +       (float_extend:DF
>           (match_operand:HF 1 "nonimmediate_operand" "vm")))]
>    "TARGET_AVX512FP16"
> -  "vcvtsh2<ssemodesuffix>\t{%1, %0, %0|%0, %0, %1}"
> +  "vcvtsh2sd\t{%1, %0, %0|%0, %0, %1}"
>    [(set_attr "type" "ssecvt")
>     (set_attr "prefix" "evex")
> -   (set_attr "mode" "<MODE>")])
> +   (set_attr "mode" "DF")])
>
>
>  (define_expand "extend<mode>xf2"
> @@ -4766,12 +4781,27 @@
>
>  ;; Conversion from {SF,DF}mode to HFmode.
>
> -(define_insn "trunc<mode>hf2"
> +(define_insn "truncsfhf2"
> +  [(set (match_operand:HF 0 "register_operand" "=v")
> +       (float_truncate:HF
> +         (match_operand:SF 1 "nonimmediate_operand" "vm")))]
> +  "TARGET_AVX512FP16 || TARGET_F16C || TARGET_AVX512VL"
> +  {
> +    if (TARGET_AVX512FP16)
> +      return "vcvtss2sh\t{%1, %d0|%d0, %1}";
> +    else
> +      return "vcvtps2ph\t{0, %1, %0|%0, %1, 0}";
> +  }
> +  [(set_attr "type" "ssecvt")
> +   (set_attr "prefix" "evex")
> +   (set_attr "mode" "HF")])
> +
> +(define_insn "truncdfhf2"
>    [(set (match_operand:HF 0 "register_operand" "=v")
> -       (float_truncate:HF
> -         (match_operand:MODEF 1 "nonimmediate_operand" "vm")))]
> +       (float_truncate:HF
> +         (match_operand:DF 1 "nonimmediate_operand" "vm")))]
>    "TARGET_AVX512FP16"
> -  "vcvt<ssemodesuffix>2sh\t{%1, %d0|%d0, %1}"
> +  "vcvtsd2sh\t{%1, %d0|%d0, %1}"
>    [(set_attr "type" "ssecvt")
>     (set_attr "prefix" "evex")
>     (set_attr "mode" "HF")])
> diff --git a/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-pr102811.c b/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-pr102811.c
> new file mode 100644
> index 00000000000..ab44a304a03
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/avx512vl-vcvtps2ph-pr102811.c
> @@ -0,0 +1,10 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mf16c -mno-avx512fp16" } */
> +/* { dg-final { scan-assembler-times "vcvtph2ps\[ \\t\]" 2 } } */
> +/* { dg-final { scan-assembler-times "vcvtps2ph\[ \\t\]" 1 } } */
> +/* { dg-final { scan-assembler-not "__truncsfhf2\[ \\t\]"} } */
> +/* { dg-final { scan-assembler-not "__extendhfsf2\[ \\t\]"} } */
> +_Float16 test (_Float16 a, _Float16 b)
> +{
> +  return a + b;
> +}
> --
> 2.18.1
>


-- 
BR,
Hongtao