https://gcc.gnu.org/g:e717aabb6eff417a613a1247124bbd817d1414b3
commit r16-3908-ge717aabb6eff417a613a1247124bbd817d1414b3 Author: Richard Biener <[email protected]> Date: Mon Sep 15 14:03:00 2025 +0200 Unify last two vect_transform_slp_perm_load calls The following unifies the vect_transform_slp_perm_load call done in vectorizable_load with that eventually done in get_load_store_type. On the way it fixes the conditions on which we can allow VMAT_ELEMENTWISE or VMAT_GATHER_SCATTER when there's a SLP permutation (and we arrange to not code generate that). In particular that only works for single-lane SLP of non-grouped loads or groups of size one. VMAT_ELEMENTWISE does not (yet?) materialize a permutation upon vector build but still relies on vect_transform_slp_perm_load. * tree-vect-stmts.cc (get_load_store_type): Get in a flag whether a SLP_TREE_LOAD_PERMUTATION on the node can be code generated and use it. Fix the condition on using strided gather/scatter to avoid dropping a meaningful permutation. (vectorizable_store): Adjust. (vectorizable_load): Analyze the permutation early and pass the result down to get_load_store_type. Fix the condition on when we are allowed to elide a load permutation. Diff: --- gcc/tree-vect-stmts.cc | 39 +++++++++++++++++++++++---------------- 1 file changed, 23 insertions(+), 16 deletions(-) diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc index d46c1e3d56d2..6274956e2a52 100644 --- a/gcc/tree-vect-stmts.cc +++ b/gcc/tree-vect-stmts.cc @@ -1963,13 +1963,16 @@ vector_vector_composition_type (tree vtype, poly_uint64 nelts, tree *ptype) VECTYPE is the vector type that the vectorized statements will use. If ELSVALS is nonzero the supported else values will be stored in the - vector ELSVALS points to. */ + vector ELSVALS points to. + + For loads PERM_OK indicates whether we can code generate a + SLP_TREE_LOAD_PERMUTATION on the node. */ static bool get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, tree vectype, slp_tree slp_node, bool masked_p, vec_load_store_type vls_type, - vect_load_store_data *ls) + bool perm_ok, vect_load_store_data *ls) { vect_memory_access_type *memory_access_type = &ls->memory_access_type; poly_int64 *poffset = &ls->poffset; @@ -2153,17 +2156,13 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, /* For single-element interleaving also fall back to elementwise access in case we did not lower a permutation and cannot code generate it. */ - auto_vec<tree> temv; - unsigned n_perms; if (loop_vinfo && single_element_p && SLP_TREE_LANES (slp_node) == 1 && (*memory_access_type == VMAT_CONTIGUOUS || *memory_access_type == VMAT_CONTIGUOUS_REVERSE) && SLP_TREE_LOAD_PERMUTATION (slp_node).exists () - && !vect_transform_slp_perm_load - (loop_vinfo, slp_node, temv, NULL, - LOOP_VINFO_VECT_FACTOR (loop_vinfo), true, &n_perms)) + && !perm_ok) { *memory_access_type = VMAT_ELEMENTWISE; if (dump_enabled_p ()) @@ -2314,11 +2313,13 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, on nearby locations. Or, even if it's a win over scalar code, it might not be a win over vectorizing at a lower VF, if that allows us to use contiguous accesses. */ - if ((*memory_access_type == VMAT_ELEMENTWISE - || *memory_access_type == VMAT_STRIDED_SLP) + if (loop_vinfo + && (*memory_access_type == VMAT_ELEMENTWISE + || *memory_access_type == VMAT_STRIDED_SLP) && !STMT_VINFO_GATHER_SCATTER_P (stmt_info) && SLP_TREE_LANES (slp_node) == 1 - && loop_vinfo) + && (!SLP_TREE_LOAD_PERMUTATION (slp_node).exists () + || single_element_p)) { gather_scatter_info gs_info; if (vect_use_strided_gather_scatters_p (stmt_info, vectype, loop_vinfo, @@ -7881,7 +7882,7 @@ vectorizable_store (vec_info *vinfo, vect_load_store_data &ls = slp_node->get_data (_ls_data); if (cost_vec && !get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask_node, - vls_type, &_ls_data)) + vls_type, false, &_ls_data)) return false; /* Temporary aliases to analysis data, should not be modified through these. */ @@ -9447,11 +9448,17 @@ vectorizable_load (vec_info *vinfo, else group_size = 1; + bool perm_ok = true; + unsigned n_perms = -1U; + if (cost_vec && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) + perm_ok = vect_transform_slp_perm_load (vinfo, slp_node, vNULL, NULL, vf, + true, &n_perms); + vect_load_store_data _ls_data{}; vect_load_store_data &ls = slp_node->get_data (_ls_data); if (cost_vec && !get_load_store_type (vinfo, stmt_info, vectype, slp_node, mask_node, - VLS_LOAD, &ls)) + VLS_LOAD, perm_ok, &ls)) return false; /* Temporary aliases to analysis data, should not be modified through these. */ @@ -9474,11 +9481,12 @@ vectorizable_load (vec_info *vinfo, /* ??? The following checks should really be part of get_load_store_type. */ - unsigned n_perms = -1U; if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists () && !((memory_access_type == VMAT_ELEMENTWISE || mat_gather_scatter_p (memory_access_type)) - && SLP_TREE_LANES (slp_node) == 1)) + && SLP_TREE_LANES (slp_node) == 1 + && (!grouped_load + || !DR_GROUP_NEXT_ELEMENT (first_stmt_info)))) { slp_perm = true; @@ -9507,8 +9515,7 @@ vectorizable_load (vec_info *vinfo, if (cost_vec) { - if (!vect_transform_slp_perm_load (vinfo, slp_node, vNULL, NULL, vf, - true, &n_perms)) + if (!perm_ok) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION,
