On Tue, Nov 29, 2016 at 10:42 AM, Andreas Krebbel
<kreb...@linux.vnet.ibm.com> wrote:
> Although the S/390 backend states that the machine supports unaligned
> vector accesses the loop vectorizer still tries to peel loop
> iterations to get higher alignments. Setting
> vect_max_peeling_for_alignment to 0 prevents this.
Hmm, having proper costs should also do this... but I see you neither
have TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST nor
TARGET_VECTORIZE_{INIT,ADD,FINISH}_COST hooks.
The default implementation (default_builtin_vectorization_cost)
makes unaligned loads/stores twice as expensive as the aligned
variants.
So - please instead of setting this param provide
TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST.
Richard.
> gcc/ChangeLog:
>
> 2016-11-29 Andreas Krebbel <kreb...@linux.vnet.ibm.com>
>
> * config/s390/s390.c (s390_option_override_internal): Set
> vect_max_peeling_for_alignment to 0.
>
> gcc/testsuite/ChangeLog:
>
> 2016-11-29 Andreas Krebbel <kreb...@linux.vnet.ibm.com>
>
> * gcc.dg/tree-ssa/gen-vect-26.c: Disable peeling check for s390.
> * gcc.dg/tree-ssa/gen-vect-28.c: Likewise.
> ---
> gcc/config/s390/s390.c | 13 +++++++++++++
> gcc/testsuite/gcc.dg/tree-ssa/gen-vect-26.c | 5 ++---
> gcc/testsuite/gcc.dg/tree-ssa/gen-vect-28.c | 4 ++--
> 3 files changed, 17 insertions(+), 5 deletions(-)
>
> diff --git a/gcc/config/s390/s390.c b/gcc/config/s390/s390.c
> index dab4f43..2e71745 100644
> --- a/gcc/config/s390/s390.c
> +++ b/gcc/config/s390/s390.c
> @@ -14586,6 +14586,19 @@ s390_option_override_internal (bool main_args_p,
> opts->x_param_values,
> opts_set->x_param_values);
>
> + /* S/390 can deal with unaligned accesses without a performance
> + penalty (as long as we do not cross a cache line boundary). This
> + setting prevents the vectorizer from generating expensive extra
> + code emitted to reach a better alignment.
> + Don't do this when vectorize_support_vector_misalignment falls
> + back to the default path in order to avoid effects on software
> + vectorization. */
> + if (TARGET_VX)
> + maybe_set_param_value (PARAM_VECT_MAX_PEELING_FOR_ALIGNMENT,
> + 0,
> + opts->x_param_values,
> + opts_set->x_param_values);
> +
> /* Call target specific restore function to do post-init work. At the
> moment,
> this just sets opts->x_s390_cost_pointer. */
> s390_function_specific_restore (opts, NULL);
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-26.c
> b/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-26.c
> index 8e5f141..461a952 100644
> --- a/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-26.c
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-26.c
> @@ -28,7 +28,6 @@ int main ()
> return 0;
> }
>
> -
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target
> { ! avr-*-* } } } } */
> -/* { dg-final { scan-tree-dump-times "Vectorizing an unaligned access" 0
> "vect" { target { ! avr-*-* } } } } */
> -/* { dg-final { scan-tree-dump-times "Alignment of access forced using
> peeling" 1 "vect" { target { ! avr-*-* } } } } */
> +/* { dg-final { scan-tree-dump-times "Vectorizing an unaligned access" 0
> "vect" { target { ! { avr-*-* s390*-*-* } } } } } */
> +/* { dg-final { scan-tree-dump-times "Alignment of access forced using
> peeling" 1 "vect" { target { ! { avr-*-* s390*-*-* } } } } } */
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-28.c
> b/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-28.c
> index ce97e09..fe44e85 100644
> --- a/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-28.c
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/gen-vect-28.c
> @@ -38,5 +38,5 @@ int main (void)
>
>
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target
> { ! avr-*-* } } } } */
> -/* { dg-final { scan-tree-dump-times "Vectorizing an unaligned access" 0
> "vect" { target { ! avr-*-* } } } } */
> -/* { dg-final { scan-tree-dump-times "Alignment of access forced using
> peeling" 1 "vect" { target { ! avr-*-* } } } } */
> +/* { dg-final { scan-tree-dump-times "Vectorizing an unaligned access" 0
> "vect" { target { ! { avr-*-* s390*-*-* } } } } } */
> +/* { dg-final { scan-tree-dump-times "Alignment of access forced using
> peeling" 1 "vect" { target { ! { avr-*-* s390*-*-* } } } } } */
> --
> 2.9.1
>
