This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] If !-prefer-avx128, prefer using even 32-byte integer vector modes
- From: Richard Guenther <richard dot guenther at gmail dot com>
- To: Jakub Jelinek <jakub at redhat dot com>
- Cc: Richard Henderson <rth at redhat dot com>, Uros Bizjak <ubizjak at gmail dot com>, gcc-patches at gcc dot gnu dot org
- Date: Tue, 8 Nov 2011 12:14:21 +0100
- Subject: Re: [PATCH] If !-prefer-avx128, prefer using even 32-byte integer vector modes
- References: <20111108064455.GA17997@tyan-ft48-01.lab.bos.redhat.com>
On Tue, Nov 8, 2011 at 7:44 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> Working virtually out of Pago Pago.
>
> Is there a reason why we don't prefer 32-byte integer vector modes
> even for AVX? ?If a vectorized loop needs some operation that is only
> supported by AVX2 we would retry whenever seeing such stmt, so
> what I think this costs us mainly some small amount of time
> during vectorization analysis. ?Examples of functions that can
> be vectorized with this patch using 32-byte vectors (and the runtime
> of the testcase decreased from ~ 0m2.876s to ~ 0m2.258s).
>
> #define N 1024
> int a[N], b[N];
> float c[N];
>
> __attribute__((noinline, noclone)) void
> f1 ()
> {
> ?int i;
> ?for (i = 0; i < N; i++)
> ? ?a[i] = b[i];
> }
>
> __attribute__((noinline, noclone)) void
> f2 ()
> {
> ?a[0] = b[0];
> ?a[1] = b[1];
> ?a[2] = b[2];
> ?a[3] = b[3];
> ?a[4] = b[4];
> ?a[5] = b[5];
> ?a[6] = b[6];
> ?a[7] = b[7];
> }
Other examples would for example be bitwise operations.
Richard.
> __attribute__((noinline, noclone)) void
> f3 ()
> {
> ?int i;
> ?for (i = 0; i < N; i++)
> ? ?c[i] = a[i];
> }
>
> __attribute__((noinline, noclone)) void
> f4 ()
> {
> ?int i;
> ?for (i = 0; i < N; i++)
> ? ?a[i] = 19;
> }
>
> int
> main ()
> {
> ?int i;
> ?for (i = 0; i < N; i++)
> ? ?b[i] = i;
> ?for (i = 0; i < 10000000; i++)
> ? ?{
> ? ? ?f1 ();
> ? ? ?f2 ();
> ? ? ?f3 ();
> ? ? ?f4 ();
> ? ?}
> ?return 0;
> }
>
>
> 2011-11-08 ?Jakub Jelinek ?<jakub@redhat.com>
>
> ? ? ? ?* config/i386/i386.c (ix86_preferred_simd_mode): Even for TARGET_AVX
> ? ? ? ?if not TARGET_PREFER_AVX128 return 32-byte SI/DI vectors.
>
> --- gcc/config/i386/i386.c.jj ? 2011-11-07 08:32:09.000000000 -1100
> +++ gcc/config/i386/i386.c ? ? ?2011-11-07 19:19:57.000000000 -1100
> @@ -37881,9 +37881,9 @@ ix86_preferred_simd_mode (enum machine_m
> ? ? case HImode:
> ? ? ? return TARGET_AVX2 ? V16HImode : V8HImode;
> ? ? case SImode:
> - ? ? ?return TARGET_AVX2 ? V8SImode : V4SImode;
> + ? ? ?return (TARGET_AVX && !TARGET_PREFER_AVX128) ? V8SImode : V4SImode;
> ? ? case DImode:
> - ? ? ?return TARGET_AVX2 ? V4DImode : V2DImode;
> + ? ? ?return (TARGET_AVX && !TARGET_PREFER_AVX128) ? V4DImode : V2DImode;
>
> ? ? case SFmode:
> ? ? ? if (TARGET_AVX && !TARGET_PREFER_AVX128)
>
> ? ? ? ?Jakub
>