RFA: Define vect_perm for variable-length SVE

Variable-length SVE now supports enough permutes to define vect_perm.

The change to vect_perm_supported is currently a no-op because the
function is only called with a count of 3.

Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf
and x86_64-linux-gnu.  OK for the vect_perm_supported change?
I think the rest is covered by the SVE maintainership.

Richard


2018-08-23  Richard Sandiford  <richard.sandiford@arm.com>

gcc/testsuite/
	* lib/target-supports.exp (vect_perm_supported): Only return
	false for variable-length vectors if the permute size is not
	a power of 2.
	(check_effective_target_vect_perm)
	(check_effective_target_vect_perm_byte)
	(check_effective_target_vect_perm_short): Remove check for
	variable-length vectors.
	* gcc.dg/vect/slp-23.c: Add an XFAIL for variable-length SVE.
	* gcc.dg/vect/slp-perm-10.c: Likewise.
	* gcc.dg/vect/slp-perm-9.c: Add an XFAIL for variable-length vectors.

On Thu, Aug 23, 2018 at 11:15 AM Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Variable-length SVE now supports enough permutes to define vect_perm.
>
> The change to vect_perm_supported is currently a no-op because the
> function is only called with a count of 3.
>
> Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf
> and x86_64-linux-gnu.  OK for the vect_perm_supported change?
> I think the rest is covered by the SVE maintainership.
OK

> Richard
>
>
> 2018-08-23  Richard Sandiford  <richard.sandiford@arm.com>
>
> gcc/testsuite/
>         * lib/target-supports.exp (vect_perm_supported): Only return
>         false for variable-length vectors if the permute size is not
>         a power of 2.
>         (check_effective_target_vect_perm)
>         (check_effective_target_vect_perm_byte)
>         (check_effective_target_vect_perm_short): Remove check for
>         variable-length vectors.
>         * gcc.dg/vect/slp-23.c: Add an XFAIL for variable-length SVE.
>         * gcc.dg/vect/slp-perm-10.c: Likewise.
>         * gcc.dg/vect/slp-perm-9.c: Add an XFAIL for variable-length vectors.
>
> Index: gcc/testsuite/lib/target-supports.exp
> ===================================================================
> --- gcc/testsuite/lib/target-supports.exp       2018-08-21 14:47:06.491178839 +0100
> +++ gcc/testsuite/lib/target-supports.exp       2018-08-23 10:09:45.296442485 +0100
> @@ -5758,8 +5758,7 @@ proc check_effective_target_vect_perm {
>      } else {
>         set et_vect_perm_saved($et_index) 0
>          if { [is-effective-target arm_neon]
> -            || ([istarget aarch64*-*-*]
> -                && ![check_effective_target_vect_variable_length])
> +            || [istarget aarch64*-*-*]
>              || [istarget powerpc*-*-*]
>               || [istarget spu-*-*]
>              || [istarget i?86-*-*] || [istarget x86_64-*-*]
> @@ -5824,7 +5823,9 @@ proc check_effective_target_vect_perm {
>
>  proc vect_perm_supported { count element_bits } {
>      set vector_bits [lindex [available_vector_sizes] 0]
> -    if { $vector_bits <= 0 } {
> +    # The number of vectors has to be a power of 2 when permuting
> +    # variable-length vectors.
> +    if { $vector_bits <= 0 && ($count & -$count) != $count } {
>         return 0
>      }
>      set vf [expr { $vector_bits / $element_bits }]
> @@ -5864,8 +5865,7 @@ proc check_effective_target_vect_perm_by
>          if { ([is-effective-target arm_neon]
>               && [is-effective-target arm_little_endian])
>              || ([istarget aarch64*-*-*]
> -                && [is-effective-target aarch64_little_endian]
> -                && ![check_effective_target_vect_variable_length])
> +                && [is-effective-target aarch64_little_endian])
>              || [istarget powerpc*-*-*]
>              || [istarget spu-*-*]
>              || ([istarget mips-*.*]
> @@ -5904,8 +5904,7 @@ proc check_effective_target_vect_perm_sh
>          if { ([is-effective-target arm_neon]
>               && [is-effective-target arm_little_endian])
>              || ([istarget aarch64*-*-*]
> -                && [is-effective-target aarch64_little_endian]
> -                && ![check_effective_target_vect_variable_length])
> +                && [is-effective-target aarch64_little_endian])
>              || [istarget powerpc*-*-*]
>              || [istarget spu-*-*]
>              || (([istarget i?86-*-*] || [istarget x86_64-*-*])
> Index: gcc/testsuite/gcc.dg/vect/slp-23.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/slp-23.c  2018-05-02 08:37:48.985604715 +0100
> +++ gcc/testsuite/gcc.dg/vect/slp-23.c  2018-08-23 10:09:45.296442485 +0100
> @@ -107,8 +107,8 @@ int main (void)
>
>  /* { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" { target { vect_strided8 && { ! { vect_no_align} } } } } } */
>  /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target { ! { vect_strided8 || vect_no_align } } } } } */
> -/* We fail to vectorize the second loop with variable-length SVE but
> -   fall back to 128-bit vectors, which does use SLP.  */
>  /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" { target { ! vect_perm } } } } */
> -/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 2 "vect" { target vect_perm } } } */
> +/* SLP fails for the second loop with variable-length SVE because
> +   the load size is greater than the minimum vector size.  */
> +/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 2 "vect" { target vect_perm xfail { aarch64_sve && vect_variable_length } } } } */
>
> Index: gcc/testsuite/gcc.dg/vect/slp-perm-10.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/slp-perm-10.c     2016-11-11 17:07:36.516798781 +0000
> +++ gcc/testsuite/gcc.dg/vect/slp-perm-10.c     2018-08-23 10:09:45.296442485 +0100
> @@ -50,4 +50,6 @@ int main ()
>  }
>
>  /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target vect_perm } } } */
> -/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" { target vect_perm } } } */
> +/* SLP fails for variable-length SVE because the load size is greater
> +   than the minimum vector size.  */
> +/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" { target vect_perm xfail { aarch64_sve && vect_variable_length } } } } */
> Index: gcc/testsuite/gcc.dg/vect/slp-perm-9.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/slp-perm-9.c      2018-05-02 08:37:49.013604450 +0100
> +++ gcc/testsuite/gcc.dg/vect/slp-perm-9.c      2018-08-23 10:09:45.296442485 +0100
> @@ -59,7 +59,9 @@ int main (int argc, const char* argv[])
>
>  /* { dg-final { scan-tree-dump-times "vectorized 0 loops" 2 "vect" { target { ! { vect_perm_short || vect_load_lanes } } } } } */
>  /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target { vect_perm_short || vect_load_lanes } } } } */
> -/* { dg-final { scan-tree-dump-times "permutation requires at least three vectors" 1 "vect" { target { vect_perm_short && { ! vect_perm3_short } } } } } */
> +/* We don't try permutes with a group size of 3 for variable-length
> +   vectors.  */
> +/* { dg-final { scan-tree-dump-times "permutation requires at least three vectors" 1 "vect" { target { vect_perm_short && { ! vect_perm3_short } } xfail vect_variable_length } } } */
>  /* { dg-final { scan-tree-dump-not "permutation requires at least three vectors" "vect" { target vect_perm3_short } } } */
>  /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 0 "vect" { target { { ! vect_perm3_short } || vect_load_lanes } } } } */
>  /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" { target { vect_perm3_short && { ! vect_load_lanes } } } } } */