On Thu, 14 Jan 2021, Richard Sandiford wrote:
> At the moment, if we use only one vector of an LD4 result,
> we'll treat the LD4 as having the cost of a single load.
> But all 4 loads and any associated permutes take place
> regardless of which results are actually used.
>
> This patch therefore counts the cost of unused LOAD_LANES
> results against the first statement in a group. An alternative
> would be to multiply the ncopies of the first stmt by the group
> size and treat other stmts in the group as having zero cost,
> but I thought that might be more surprising when reading dumps.
>
> Tested on aarch64-linux-gnu, aarch64_be-elf and x86_64-linux-gnu.
> OK to install?
OK.
Richard.
> Richard
>
>
> gcc/
> * tree-vect-stmts.c (vect_model_load_cost): Account for unused
> IFN_LOAD_LANES results.
>
> gcc/testsuite/
> * gcc.target/aarch64/sve/cost_model_11.c: New test.
> * gcc.target/aarch64/sve/mask_struct_load_5.c: Use
> -fno-vect-cost-model.
> ---
> .../gcc.target/aarch64/sve/cost_model_11.c | 12 ++++++++++
> .../aarch64/sve/mask_struct_load_5.c | 2 +-
> gcc/tree-vect-stmts.c | 24 +++++++++++++++++++
> 3 files changed, 37 insertions(+), 1 deletion(-)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/cost_model_11.c
>
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cost_model_11.c
> b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_11.c
> new file mode 100644
> index 00000000000..d9f4ccc76de
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_11.c
> @@ -0,0 +1,12 @@
> +/* { dg-options "-O2 -ftree-vectorize -msve-vector-bits=128" } */
> +
> +long
> +f (long *x, long *y, long *z, long n)
> +{
> + long res = 0;
> + for (long i = 0; i < n; ++i)
> + z[i] = x[i * 4] + y[i * 4];
> + return res;
> +}
> +
> +/* { dg-final { scan-assembler-not {\tld4d\t} } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_5.c
> b/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_5.c
> index da367e4fd79..2a33ee81d1a 100644
> --- a/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_5.c
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_5.c
> @@ -1,5 +1,5 @@
> /* { dg-do compile } */
> -/* { dg-options "-O2 -ftree-vectorize -ffast-math --param
> aarch64-sve-compare-costs=0" } */
> +/* { dg-options "-O2 -ftree-vectorize -ffast-math -fno-vect-cost-model" } */
>
> #include <stdint.h>
>
> diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
> index 068e4982303..4d72c4db2f7 100644
> --- a/gcc/tree-vect-stmts.c
> +++ b/gcc/tree-vect-stmts.c
> @@ -1120,6 +1120,30 @@ vect_model_load_cost (vec_info *vinfo,
> once per group anyhow. */
> bool first_stmt_p = (first_stmt_info == stmt_info);
>
> + /* An IFN_LOAD_LANES will load all its vector results, regardless of which
> + ones we actually need. Account for the cost of unused results. */
> + if (first_stmt_p && !slp_node && memory_access_type ==
> VMAT_LOAD_STORE_LANES)
> + {
> + unsigned int gaps = DR_GROUP_SIZE (first_stmt_info);
> + stmt_vec_info next_stmt_info = first_stmt_info;
> + do
> + {
> + gaps -= 1;
> + next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
> + }
> + while (next_stmt_info);
> + if (gaps)
> + {
> + if (dump_enabled_p ())
> + dump_printf_loc (MSG_NOTE, vect_location,
> + "vect_model_load_cost: %d unused vectors.\n",
> + gaps);
> + vect_get_load_cost (vinfo, stmt_info, ncopies * gaps, false,
> + &inside_cost, &prologue_cost,
> + cost_vec, cost_vec, true);
> + }
> + }
> +
> /* We assume that the cost of a single load-lanes instruction is
> equivalent to the cost of DR_GROUP_SIZE separate loads. If a grouped
> access is instead being provided by a load-and-permute operation,
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)
