The following patch adds a heuristic to prefer store/load-lanes
over SLP when vectorizing. Compared to the variant attached to
the PR I made the STMT_VINFO_STRIDED_P behavior explicit (matching
what you've tested).
It's a heuristic that may end up vectorizing less loops or loops
in a less optimal way.
Thus I wait for your ok (it's essentially ARM specific).
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Ok? It will require a bunch of vectorizer tests to be adjusted for
ARM I think.
Thanks,
Richard.
2015-12-16 Richard Biener <rguent...@suse.de>
PR tree-optimization/68707
PR tree-optimization/67323
* tree-vect-slp.c (vect_analyze_slp_instance): Drop SLP instances
if they can be vectorized using load/store-lane instructions.
Index: gcc/tree-vect-slp.c
===================================================================
*** gcc/tree-vect-slp.c (revision 231673)
--- gcc/tree-vect-slp.c (working copy)
*************** vect_analyze_slp_instance (vec_info *vin
*** 1808,1813 ****
--- 1802,1836 ----
}
}
+ /* If the loads and stores can be handled with load/store-lane
+ instructions do not generate this SLP instance. */
+ if (is_a <loop_vec_info> (vinfo)
+ && loads_permuted
+ && dr && vect_store_lanes_supported (vectype, group_size))
+ {
+ slp_tree load_node;
+ FOR_EACH_VEC_ELT (loads, i, load_node)
+ {
+ gimple *first_stmt = GROUP_FIRST_ELEMENT
+ (vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (load_node)[0]));
+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (first_stmt);
+ if (! STMT_VINFO_STRIDED_P (stmt_vinfo)
+ && ! vect_load_lanes_supported
+ (STMT_VINFO_VECTYPE (stmt_vinfo),
+ GROUP_SIZE (stmt_vinfo)))
+ break;
+ }
+ if (i == loads.length ())
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+ "Built SLP cancelled: can use "
+ "load/store-lanes\n");
+ vect_free_slp_instance (new_instance);
+ return false;
+ }
+ }
+
vinfo->slp_instances.safe_push (new_instance);
if (dump_enabled_p ())
