This patch adds tuning fields for the total cost of a gather load instruction. Until now, we've costed them as one scalar load per element instead. Those scalar_load-based values are also what the patch uses to fill in the new fields for existing cost structures.
gcc/ * config/aarch64/aarch64-protos.h (sve_vec_cost): Add gather_load_x32_cost and gather_load_x64_cost. * config/aarch64/aarch64.c (generic_sve_vector_cost) (a64fx_sve_vector_cost, neoversev1_sve_vector_cost): Update accordingly, using the values given by the scalar_load * number of elements calculation that we used previously. (aarch64_detect_vector_stmt_subtype): Use the new fields. --- gcc/config/aarch64/aarch64-protos.h | 9 +++++++++ gcc/config/aarch64/aarch64.c | 19 +++++++++++++++++++ 2 files changed, 28 insertions(+) diff --git a/gcc/config/aarch64/aarch64-protos.h b/gcc/config/aarch64/aarch64-protos.h index fb4ce8e9f84..b91eeeba101 100644 --- a/gcc/config/aarch64/aarch64-protos.h +++ b/gcc/config/aarch64/aarch64-protos.h @@ -259,12 +259,16 @@ struct sve_vec_cost : simd_vec_cost unsigned int fadda_f16_cost, unsigned int fadda_f32_cost, unsigned int fadda_f64_cost, + unsigned int gather_load_x32_cost, + unsigned int gather_load_x64_cost, unsigned int scatter_store_elt_cost) : simd_vec_cost (base), clast_cost (clast_cost), fadda_f16_cost (fadda_f16_cost), fadda_f32_cost (fadda_f32_cost), fadda_f64_cost (fadda_f64_cost), + gather_load_x32_cost (gather_load_x32_cost), + gather_load_x64_cost (gather_load_x64_cost), scatter_store_elt_cost (scatter_store_elt_cost) {} @@ -279,6 +283,11 @@ struct sve_vec_cost : simd_vec_cost const int fadda_f32_cost; const int fadda_f64_cost; + /* The cost of a gather load instruction. The x32 value is for loads + of 32-bit elements and the x64 value is for loads of 64-bit elements. */ + const int gather_load_x32_cost; + const int gather_load_x64_cost; + /* The per-element cost of a scatter store. */ const int scatter_store_elt_cost; }; diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index b14b6f22aec..36f11808916 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -675,6 +675,8 @@ static const sve_vec_cost generic_sve_vector_cost = 2, /* fadda_f16_cost */ 2, /* fadda_f32_cost */ 2, /* fadda_f64_cost */ + 4, /* gather_load_x32_cost */ + 2, /* gather_load_x64_cost */ 1 /* scatter_store_elt_cost */ }; @@ -744,6 +746,8 @@ static const sve_vec_cost a64fx_sve_vector_cost = 13, /* fadda_f16_cost */ 13, /* fadda_f32_cost */ 13, /* fadda_f64_cost */ + 64, /* gather_load_x32_cost */ + 32, /* gather_load_x64_cost */ 1 /* scatter_store_elt_cost */ }; @@ -1739,6 +1743,8 @@ static const sve_vec_cost neoversev1_sve_vector_cost = 19, /* fadda_f16_cost */ 11, /* fadda_f32_cost */ 8, /* fadda_f64_cost */ + 32, /* gather_load_x32_cost */ + 16, /* gather_load_x64_cost */ 3 /* scatter_store_elt_cost */ }; @@ -14958,6 +14964,19 @@ aarch64_detect_vector_stmt_subtype (vec_info *vinfo, vect_cost_for_stmt kind, && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info))) return simd_costs->store_elt_extra_cost; + /* Detect SVE gather loads, which are costed as a single scalar_load + for each element. We therefore need to divide the full-instruction + cost by the number of elements in the vector. */ + if (kind == scalar_load + && sve_costs + && STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) == VMAT_GATHER_SCATTER) + { + unsigned int nunits = vect_nunits_for_cost (vectype); + if (GET_MODE_UNIT_BITSIZE (TYPE_MODE (vectype)) == 64) + return { sve_costs->gather_load_x64_cost, nunits }; + return { sve_costs->gather_load_x32_cost, nunits }; + } + /* Detect cases in which a scalar_store is really storing one element in a scatter operation. */ if (kind == scalar_store