> On 23 Jul 2026, at 13:14, Pengfei Li <[email protected]> wrote:
>
> The current Neoverse V2 cost model underestimates the costs of SVE
> structure loads and stores. This can cause the vectorizer to make a
> suboptimal choice between AdvSIMD and SVE.
>
> This patch brings the cost model in line with the latest Neoverse V2
> Software Optimization Guide, which indicates that SVE structure loads
> and stores are generally more expensive than the AdvSIMD versions.
>
> Performance testing shows this change improves SPEC CPU 2026 735.ocio_r
> by ~10.8% on Neoverse V2.
>
Ok, I see that the values in the SWOG indeed have slightly worse latencies and
throughputs for some forms of these instructions.
Thanks,
Kyrill
> Bootstrapped and tested on aarch64-linux-gnu.
>
> gcc/ChangeLog:
>
> * config/aarch64/tuning_models/neoversev2.h: Update the costs
> of SVE structure loads and stores.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/aarch64/sve/cost_model_20.c: New test.
> ---
> gcc/config/aarch64/tuning_models/neoversev2.h | 6 ++---
> .../gcc.target/aarch64/sve/cost_model_20.c | 22 +++++++++++++++++++
> 2 files changed, 25 insertions(+), 3 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/cost_model_20.c
>
> diff --git a/gcc/config/aarch64/tuning_models/neoversev2.h
> b/gcc/config/aarch64/tuning_models/neoversev2.h
> index 15eff2807f7..6df0cc444b8 100644
> --- a/gcc/config/aarch64/tuning_models/neoversev2.h
> +++ b/gcc/config/aarch64/tuning_models/neoversev2.h
> @@ -70,9 +70,9 @@ static const sve_vec_cost neoversev2_sve_vector_cost =
> {
> 2, /* int_stmt_cost */
> 2, /* fp_stmt_cost */
> - 2, /* ld2_st2_permute_cost */
> - 3, /* ld3_st3_permute_cost */
> - 3, /* ld4_st4_permute_cost */
> + 3, /* ld2_st2_permute_cost */
> + 4, /* ld3_st3_permute_cost */
> + 4, /* ld4_st4_permute_cost */
> 2, /* permute_cost */
> /* Theoretically, a reduction involving 15 scalar ADDs could
> complete in ~5 cycles and would have a cost of 15. [SU]ADDV
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cost_model_20.c
> b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_20.c
> new file mode 100644
> index 00000000000..77cb00cf1d6
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_20.c
> @@ -0,0 +1,22 @@
> +/* { dg-options "-O3 -mtune=neoverse-v2" } */
> +
> +void
> +f (float *restrict in, float *restrict out, int num)
> +{
> + for (int i = 0; i < num; ++i)
> + {
> + float r = in[4 * i];
> + float g = in[4 * i + 1];
> + float b = in[4 * i + 2];
> + float a = in[4 * i + 3];
> + out[4 * i] = r * 0.9f + g * 1.0f + b * 0.7f + a * 1.0f;
> + out[4 * i + 1] = r * 1.1f + g * 0.8f + b * 1.2f + a * 1.0f;
> + out[4 * i + 2] = r * 0.7f + g * 1.2f + b * 0.8f + a * 1.0f;
> + out[4 * i + 3] = r * 1.2f + g * 0.7f + b * 1.1f + a * 1.0f;
> + }
> +}
> +
> +/* We should use AdvSIMD ld4/st4 rather than SVE for the vectorized main loop
> + on Neoverse-V2. */
> +/* { dg-final { scan-assembler {\tld4\t} } } */
> +/* { dg-final { scan-assembler {\tst4\t} } } */
> --
> 2.43.0
>
