Generate vpblendd instead of vpblendw for V4SI under AVX2.
Commit Message
Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
Ok for GCC14 stage-1(or maybe trunk)?
gcc/ChangeLog:
* config/i386/i386-expand.cc (expand_vec_perm_blend): Generate
vpblendd instead of vpblendw for V4SI under avx2.
gcc/testsuite/ChangeLog:
* gcc.target/i386/pr88828-0.c: Adjust testcase.
---
gcc/config/i386/i386-expand.cc | 18 ++++++++++++++----
gcc/testsuite/gcc.target/i386/pr88828-0.c | 2 +-
2 files changed, 15 insertions(+), 5 deletions(-)
Comments
On Wed, Mar 29, 2023 at 9:21 AM liuhongt <hongtao.liu@intel.com> wrote:
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> Ok for GCC14 stage-1(or maybe trunk)?
>
> gcc/ChangeLog:
>
> * config/i386/i386-expand.cc (expand_vec_perm_blend): Generate
> vpblendd instead of vpblendw for V4SI under avx2.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr88828-0.c: Adjust testcase.
Looks safe, so OK for trunk.
Thanks,
Uros.
> ---
> gcc/config/i386/i386-expand.cc | 18 ++++++++++++++----
> gcc/testsuite/gcc.target/i386/pr88828-0.c | 2 +-
> 2 files changed, 15 insertions(+), 5 deletions(-)
>
> diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> index c1300dc4e26..1c436262ee5 100644
> --- a/gcc/config/i386/i386-expand.cc
> +++ b/gcc/config/i386/i386-expand.cc
> @@ -19069,10 +19069,20 @@ expand_vec_perm_blend (struct expand_vec_perm_d *d)
> goto do_subreg;
>
> case E_V4SImode:
> - for (i = 0; i < 4; ++i)
> - mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2);
> - vmode = V8HImode;
> - goto do_subreg;
> + if (TARGET_AVX2)
> + {
> + /* Use vpblendd instead of vpblendw. */
> + for (i = 0; i < nelt; ++i)
> + mask |= ((unsigned HOST_WIDE_INT) (d->perm[i] >= nelt)) << i;
> + break;
> + }
> + else
> + {
> + for (i = 0; i < 4; ++i)
> + mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2);
> + vmode = V8HImode;
> + goto do_subreg;
> + }
>
> case E_V16QImode:
> /* See if bytes move in pairs so we can use pblendw with
> diff --git a/gcc/testsuite/gcc.target/i386/pr88828-0.c b/gcc/testsuite/gcc.target/i386/pr88828-0.c
> index 3ddb2d13526..441c441b51d 100644
> --- a/gcc/testsuite/gcc.target/i386/pr88828-0.c
> +++ b/gcc/testsuite/gcc.target/i386/pr88828-0.c
> @@ -1,5 +1,5 @@
> /* { dg-do compile } */
> -/* { dg-options "-O2 -msse4.2" } */
> +/* { dg-options "-O2 -msse4.2 -mno-avx2" } */
>
> typedef int v4si __attribute__((vector_size(16)));
> typedef float v4sf __attribute__((vector_size(16)));
> --
> 2.39.1.388.g2fc9e9ca3c
>
@@ -19069,10 +19069,20 @@ expand_vec_perm_blend (struct expand_vec_perm_d *d)
goto do_subreg;
case E_V4SImode:
- for (i = 0; i < 4; ++i)
- mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2);
- vmode = V8HImode;
- goto do_subreg;
+ if (TARGET_AVX2)
+ {
+ /* Use vpblendd instead of vpblendw. */
+ for (i = 0; i < nelt; ++i)
+ mask |= ((unsigned HOST_WIDE_INT) (d->perm[i] >= nelt)) << i;
+ break;
+ }
+ else
+ {
+ for (i = 0; i < 4; ++i)
+ mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2);
+ vmode = V8HImode;
+ goto do_subreg;
+ }
case E_V16QImode:
/* See if bytes move in pairs so we can use pblendw with
@@ -1,5 +1,5 @@
/* { dg-do compile } */
-/* { dg-options "-O2 -msse4.2" } */
+/* { dg-options "-O2 -msse4.2 -mno-avx2" } */
typedef int v4si __attribute__((vector_size(16)));
typedef float v4sf __attribute__((vector_size(16)));