Re: [PATCH, i386] Fix PR target/65671. Generate 32x4 extract even for DF in absence of AVX-512DQ.

[Uros Bizjak <ubizjak@gmail.com>]

On Thu, Apr 9, 2015 at 4:37 PM, Kirill Yukhin <kirill.yukhin@gmail.com> wrote:

> Patch in the bottom fixes PR target/65671.
>
> It simply generates vextract32x4 (float form) for double extract.
>
> Bootstrap & regtesting in progress.
>
> I'll check it in if pass and back port to 4.9.x.
> Feel free comment.
>
> gcc/
>         * config/i386/sse.md: Generate vextract32x4 if AVX-512DQ
>         is disabled.
>
> gcc/testsuite/
>         * gcc.target/i386/pr65671.c: New.
>

OK with fixed ChangeLogs. (Please add PR target/65671 to both ChangeLogs).

Thanks,
Uros.

> commit cb8d5b1c3156d81ae81600217d0861be1aade0ec
> Author: Kirill Yukhin <kirill.yukhin@intel.com>
> Date:   Thu Apr 9 13:05:54 2015 +0300
>
>     Fix PR target/65671. Generate 32x4 extract even for DF in absence of AVX-512DQ.
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index 490fd6b..6d3b54a 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -7015,10 +7015,15 @@
>         (vec_select:<ssehalfvecmode>
>           (match_operand:VI8F_256 1 "register_operand" "v,v")
>           (parallel [(const_int 2) (const_int 3)])))]
> -  "TARGET_AVX"
> +  "TARGET_AVX && <mask_avx512vl_condition> && <mask_avx512dq_condition>"
>  {
> -  if (TARGET_AVX512DQ && TARGET_AVX512VL)
> -    return "vextract<shuffletype>64x2\t{$0x1, %1, %0<mask_operand2>|%0<mask_operand2>, %1, 0x1}";
> +  if (TARGET_AVX512VL)
> +  {
> +    if (TARGET_AVX512DQ)
> +      return "vextract<shuffletype>64x2\t{$0x1, %1, %0<mask_operand2>|%0<mask_operand2>, %1, 0x1}";
> +    else
> +      return "vextract<shuffletype>32x4\t{$0x1, %1, %0|%0, %1, 0x1}";
> +  }
>    else
>      return "vextract<i128>\t{$0x1, %1, %0|%0, %1, 0x1}";
>  }
> diff --git a/gcc/testsuite/gcc.target/i386/pr65671.c b/gcc/testsuite/gcc.target/i386/pr65671.c
> new file mode 100644
> index 0000000..8e5d00d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr65671.c
> @@ -0,0 +1,15 @@
> +/* PR target/65671 */
> +/* { dg-do assemble } */
> +/* { dg-require-effective-target lp64 } */
> +/* { dg-options "-O2 -mavx512vl -ffixed-ymm16" } */
> +
> +#include <x86intrin.h>
> +
> +register __m256d a asm ("ymm16");
> +__m128d b;
> +
> +void
> +foo ()
> +{
> +  b = _mm256_extractf128_pd (a, 1);
> +}