diff options
| author | Richard Biener <rguenther@suse.de> | 2024-11-12 10:31:30 +0100 |
|---|---|---|
| committer | Richard Biener <rguenth@gcc.gnu.org> | 2024-11-12 14:57:32 +0100 |
| commit | 0d4b254b20a9193ab261d02f8a063e21816f85e4 (patch) | |
| tree | 1b88def31c54824e092b9a3468294399c7959c56 /gcc | |
| parent | 1b35b929354c41f3e2682aa7a30013e1bfe31bd4 (diff) | |
| download | gcc-0d4b254b20a9193ab261d02f8a063e21816f85e4.zip gcc-0d4b254b20a9193ab261d02f8a063e21816f85e4.tar.gz gcc-0d4b254b20a9193ab261d02f8a063e21816f85e4.tar.bz2 | |
tree-optimization/116973 - SLP permute lower heuristic and single-lane SLP
When forcing single-lane SLP to emulate non-SLP behavior we need to
disable heuristics designed to optimize SLP loads and instead in
all cases resort to an interleaving scheme as requested by forcefully
doing single-lane SLP.
This fixes the remaining fallout for --param vect-force-slp=1 on x86.
PR tree-optimization/116973
* tree-vect-slp.cc (vect_lower_load_permutations): Add
force_single_lane parameter. Disable heuristic that keeps
some load-permutations.
(vect_analyze_slp): Pass force_single_lane to
vect_lower_load_permutations.
Diffstat (limited to 'gcc')
| -rw-r--r-- | gcc/tree-vect-slp.cc | 17 |
1 files changed, 11 insertions, 6 deletions
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc index eebac19..d69fdc0 100644 --- a/gcc/tree-vect-slp.cc +++ b/gcc/tree-vect-slp.cc @@ -4402,7 +4402,8 @@ vllp_cmp (const void *a_, const void *b_) static void vect_lower_load_permutations (loop_vec_info loop_vinfo, scalar_stmts_to_slp_tree_map_t *bst_map, - const array_slice<slp_tree> &loads) + const array_slice<slp_tree> &loads, + bool force_single_lane) { /* We at this point want to lower without a fixed VF or vector size in mind which means we cannot actually compute whether we @@ -4494,7 +4495,8 @@ vect_lower_load_permutations (loop_vec_info loop_vinfo, extracting it from the larger load. ??? Long-term some of the lowering should move to where the vector types involved are fixed. */ - if (ld_lanes_lanes == 0 + if (!force_single_lane + && ld_lanes_lanes == 0 && contiguous && (SLP_TREE_LANES (load) > 1 || loads.size () == 1) && pow2p_hwi (SLP_TREE_LANES (load)) @@ -4668,7 +4670,8 @@ vect_lower_load_permutations (loop_vec_info loop_vinfo, static void vect_lower_load_permutations (loop_vec_info loop_vinfo, - scalar_stmts_to_slp_tree_map_t *bst_map) + scalar_stmts_to_slp_tree_map_t *bst_map, + bool force_single_lane) { /* Gather and sort loads across all instances. */ hash_set<slp_tree> visited; @@ -4696,14 +4699,16 @@ vect_lower_load_permutations (loop_vec_info loop_vinfo, if (STMT_VINFO_GROUPED_ACCESS (a0)) vect_lower_load_permutations (loop_vinfo, bst_map, make_array_slice (&loads[firsti], - i - firsti)); + i - firsti), + force_single_lane); firsti = i; } if (firsti < loads.length () && STMT_VINFO_GROUPED_ACCESS (SLP_TREE_SCALAR_STMTS (loads[firsti])[0])) vect_lower_load_permutations (loop_vinfo, bst_map, make_array_slice (&loads[firsti], - loads.length () - firsti)); + loads.length () - firsti), + force_single_lane); } /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP @@ -5097,7 +5102,7 @@ vect_analyze_slp (vec_info *vinfo, unsigned max_tree_size, like schemes. */ if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo)) { - vect_lower_load_permutations (loop_vinfo, bst_map); + vect_lower_load_permutations (loop_vinfo, bst_map, force_single_lane); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, |