> -----Original Message-----
> From: Richard Biener <[email protected]>
> Sent: Tuesday, August 26, 2025 8:59 AM
> To: [email protected]
> Cc: Tamar Christina <[email protected]>
> Subject: [PATCH][aarch64] Compute vect_reduc_type off SLP node instead of
> stmt-info
>
> The following changes the vect_reduc_type API to work on the SLP node.
> The API is only used from the aarch64 backend, so all changes are there.
> In particular I noticed aarch64_force_single_cycle is invoked even
> for scalar costing (where the flag tested isn't computed yet), I
> figured in scalar costing all reductions are a single cycle.
Yeah, this was a fix for PR110625 which changed scalar reductions
to correctly be 1.
>
> Bootstrap and regtest running on x86_64-unknown-linux-gnu.
>
> OK for the aarch64 parts?
OK.
Thanks,
Tamar
>
> Thanks,
> Richard.
>
> * tree-vectorizer.h (vect_reduc_type): Get SLP node as argument.
> * config/aarch64/aarch64.cc (aarch64_sve_in_loop_reduction_latency):
> Take SLO node as argument and adjust.
> (aarch64_in_loop_reduction_latency): Likewise.
> (aarch64_detect_vector_stmt_subtype): Adjust.
> (aarch64_vector_costs::count_ops): Likewise. Treat reductions
> during scalar costing as single-cycle.
> ---
> gcc/config/aarch64/aarch64.cc | 21 ++++++++++++++-------
> gcc/tree-vectorizer.h | 16 ++++++++++------
> 2 files changed, 24 insertions(+), 13 deletions(-)
>
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index fb8311b655d..eb9e2cfaab0 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -17420,10 +17420,11 @@ aarch64_bool_compound_p (vec_info *vinfo,
> stmt_vec_info stmt_info,
> instructions. */
> static unsigned int
> aarch64_sve_in_loop_reduction_latency (vec_info *vinfo,
> + slp_tree node,
> stmt_vec_info stmt_info,
> const sve_vec_cost *sve_costs)
> {
> - switch (vect_reduc_type (vinfo, stmt_info))
> + switch (vect_reduc_type (vinfo, node))
> {
> case EXTRACT_LAST_REDUCTION:
> return sve_costs->clast_cost;
> @@ -17463,7 +17464,9 @@ aarch64_sve_in_loop_reduction_latency (vec_info
> *vinfo,
> - If VEC_FLAGS & VEC_ANY_SVE, return the loop carry latency of the
> SVE implementation. */
> static unsigned int
> -aarch64_in_loop_reduction_latency (vec_info *vinfo, stmt_vec_info stmt_info,
> +aarch64_in_loop_reduction_latency (vec_info *vinfo,
> + slp_tree node,
> + stmt_vec_info stmt_info,
> unsigned int vec_flags)
> {
> const cpu_vector_cost *vec_costs = aarch64_tune_params.vec_costs;
> @@ -17476,7 +17479,8 @@ aarch64_in_loop_reduction_latency (vec_info
> *vinfo, stmt_vec_info stmt_info,
> if (sve_costs)
> {
> unsigned int latency
> - = aarch64_sve_in_loop_reduction_latency (vinfo, stmt_info, sve_costs);
> + = aarch64_sve_in_loop_reduction_latency (vinfo, node,
> + stmt_info, sve_costs);
> if (latency)
> return latency;
> }
> @@ -17575,7 +17579,8 @@ aarch64_detect_vector_stmt_subtype (vec_info
> *vinfo, vect_cost_for_stmt kind,
> && sve_costs)
> {
> unsigned int latency
> - = aarch64_sve_in_loop_reduction_latency (vinfo, stmt_info, sve_costs);
> + = aarch64_sve_in_loop_reduction_latency (vinfo, node,
> + stmt_info, sve_costs);
> if (latency)
> return latency;
> }
> @@ -17787,8 +17792,10 @@ aarch64_vector_costs::count_ops (unsigned int
> count, vect_cost_for_stmt kind,
> && vect_is_reduction (stmt_info))
> {
> unsigned int base
> - = aarch64_in_loop_reduction_latency (m_vinfo, stmt_info, m_vec_flags);
> - if (aarch64_force_single_cycle (m_vinfo, stmt_info))
> + = aarch64_in_loop_reduction_latency (m_vinfo, node,
> + stmt_info, m_vec_flags);
> + if (m_costing_for_scalar
> + || aarch64_force_single_cycle (m_vinfo, stmt_info))
> /* ??? Ideally we'd use a tree to reduce the copies down to 1 vector,
> and then accumulate that, but at the moment the loop-carried
> dependency includes all copies. */
> @@ -17901,7 +17908,7 @@ aarch64_vector_costs::count_ops (unsigned int
> count, vect_cost_for_stmt kind,
> have only accounted for one. */
> if (stmt_info
> && (kind == vector_stmt || kind == vec_to_scalar)
> - && vect_reduc_type (m_vinfo, stmt_info) == COND_REDUCTION)
> + && vect_reduc_type (m_vinfo, node) == COND_REDUCTION)
> ops->general_ops += count;
>
> /* Count the predicate operations needed by an SVE comparison. */
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 2948deffdd6..58311318706 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2877,14 +2877,18 @@ vect_is_reduction (stmt_vec_info stmt_info)
> /* If STMT_INFO describes a reduction, return the vect_reduction_type
> of the reduction it describes, otherwise return -1. */
> inline int
> -vect_reduc_type (vec_info *vinfo, stmt_vec_info stmt_info)
> +vect_reduc_type (vec_info *vinfo, slp_tree node)
> {
> if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
> - if (STMT_VINFO_REDUC_DEF (stmt_info))
> - {
> - stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info);
> - return int (STMT_VINFO_REDUC_TYPE (reduc_info));
> - }
> + {
> + stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
> + if (STMT_VINFO_REDUC_DEF (stmt_info))
> + {
> + stmt_vec_info reduc_info
> + = info_for_reduction (loop_vinfo, stmt_info);
> + return int (STMT_VINFO_REDUC_TYPE (reduc_info));
> + }
> + }
> return -1;
> }
>
> --
> 2.43.0
