This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Improve <avx512>_fmadd_<mode>_mask3<round_name>
- From: Kirill Yukhin <kirill dot yukhin at gmail dot com>
- To: Jakub Jelinek <jakub at redhat dot com>
- Cc: Uros Bizjak <ubizjak at gmail dot com>, gcc-patches at gcc dot gnu dot org
- Date: Thu, 5 May 2016 13:01:39 +0300
- Subject: Re: [PATCH] Improve <avx512>_fmadd_<mode>_mask3<round_name>
- Authentication-results: sourceware.org; auth=none
- References: <20160504193142 dot GJ26501 at tucnak dot zalov dot cz>
Hello Jakub,
On 04 May 21:31, Jakub Jelinek wrote:
> Hi!
>
> As the testcase can show, we should be using v constraint and generate
> better code that way.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2016-05-04 Jakub Jelinek <jakub@redhat.com>
>
> * config/i386/sse.md (<avx512>_fmadd_<mode>_mask3<round_name>): Use
> v constraint instead of x.
>
> * gcc.target/i386/avx512f-vfmadd-1.c: New test.
Didn't get what the test checks?
It works fine w/o patch (generating extra moves though)
Maybe scan-asm that xmm{16,17,18} actually hit FMA?
--
Thanks, K
>
> --- gcc/config/i386/sse.md.jj 2016-05-04 14:36:08.000000000 +0200
> +++ gcc/config/i386/sse.md 2016-05-04 15:16:44.180894303 +0200
> @@ -3327,10 +3327,10 @@ (define_insn "<avx512>_fmadd_<mode>_mask
> (set_attr "mode" "<MODE>")])
>
> (define_insn "<avx512>_fmadd_<mode>_mask3<round_name>"
> - [(set (match_operand:VF_AVX512VL 0 "register_operand" "=x")
> + [(set (match_operand:VF_AVX512VL 0 "register_operand" "=v")
> (vec_merge:VF_AVX512VL
> (fma:VF_AVX512VL
> - (match_operand:VF_AVX512VL 1 "register_operand" "x")
> + (match_operand:VF_AVX512VL 1 "register_operand" "v")
> (match_operand:VF_AVX512VL 2 "nonimmediate_operand" "<round_constraint>")
> (match_operand:VF_AVX512VL 3 "register_operand" "0"))
> (match_dup 3)
> --- gcc/testsuite/gcc.target/i386/avx512f-vfmadd-1.c.jj 2016-05-04 15:35:54.919506742 +0200
> +++ gcc/testsuite/gcc.target/i386/avx512f-vfmadd-1.c 2016-05-04 15:36:08.648326113 +0200
> @@ -0,0 +1,24 @@
> +/* { dg-do assemble { target { avx512f && { ! ia32 } } } } */
> +/* { dg-options "-O2 -mavx512f" } */
> +
> +#include <x86intrin.h>
> +
> +void
> +f1 (__m512d x, __m512d y, __m512d z, __mmask8 m)
> +{
> + register __m512d a __asm ("xmm16"), b __asm ("xmm17"), c __asm ("xmm18");
> + a = x; b = y; c = z;
> + asm volatile ("" : "+v" (a), "+v" (b), "+v" (c));
> + a = _mm512_mask3_fmadd_round_pd (c, b, a, m, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC);
> + asm volatile ("" : "+v" (a));
> +}
> +
> +void
> +f2 (__m512 x, __m512 y, __m512 z, __mmask8 m)
> +{
> + register __m512 a __asm ("xmm16"), b __asm ("xmm17"), c __asm ("xmm18");
> + a = x; b = y; c = z;
> + asm volatile ("" : "+v" (a), "+v" (b), "+v" (c));
> + a = _mm512_mask3_fmadd_round_ps (c, b, a, m, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC);
> + asm volatile ("" : "+v" (a));
> +}
>
> Jakub