[gcc r11-5059] Delay SLP instance loads gathering

[Richard Biener]

https://gcc.gnu.org/g:c84df34aec3bb845b22384c7e85f0449ca00dd99

commit r11-5059-gc84df34aec3bb845b22384c7e85f0449ca00dd99
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Nov 16 14:26:20 2020 +0100

    Delay SLP instance loads gathering
    
    This delays filling SLP_INSTANCE_LOADS.
    
    2020-11-16  Richard Biener  <rguenther@suse.de>
    
            * tree-vectorizer.h (vect_gather_slp_loads): Declare.
            * tree-vect-loop.c (vect_analyze_loop_2): Call
            vect_gather_slp_loads.
            * tree-vect-slp.c (vect_build_slp_instance): Do not gather
            SLP loads here.
            (vect_gather_slp_loads): Remove wrapper, new function.
            (vect_slp_analyze_bb_1): Call it.

Diff:
---
 gcc/tree-vect-loop.c  |  3 +++
 gcc/tree-vect-slp.c   | 26 ++++++++++++++++++--------
 gcc/tree-vectorizer.h |  1 +
 3 files changed, 22 insertions(+), 8 deletions(-)

diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 4d5532f71d0..ecaaf0116d3 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -2298,6 +2298,9 @@ vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal, unsigned *n_stmts)
 
       /* Optimize the SLP graph with the vectorization factor fixed.  */
       vect_optimize_slp (loop_vinfo);
+
+      /* Gather the loads reachable from the SLP graph entries.  */
+      vect_gather_slp_loads (loop_vinfo);
     }
 
   bool saved_can_use_partial_vectors_p
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index b98d5db9f76..d2f2407ac92 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -2071,13 +2071,6 @@ vect_gather_slp_loads (vec<slp_tree> &loads, slp_tree node,
     }
 }
 
-static void
-vect_gather_slp_loads (slp_instance inst, slp_tree node)
-{
-  hash_set<slp_tree> visited;
-  vect_gather_slp_loads (SLP_INSTANCE_LOADS (inst), node, visited);
-}
-
 
 /* Find the last store in SLP INSTANCE.  */
 
@@ -2252,7 +2245,6 @@ vect_build_slp_instance (vec_info *vinfo,
 	  new_instance->cost_vec = vNULL;
 	  new_instance->subgraph_entries = vNULL;
 
-	  vect_gather_slp_loads (new_instance, node);
 	  if (dump_enabled_p ())
 	    dump_printf_loc (MSG_NOTE, vect_location,
 			     "SLP size %u vs. limit %u.\n",
@@ -3071,6 +3063,21 @@ vect_optimize_slp (vec_info *vinfo)
     }
 }
 
+/* Gather loads reachable from the individual SLP graph entries.  */
+
+void
+vect_gather_slp_loads (vec_info *vinfo)
+{
+  unsigned i;
+  slp_instance instance;
+  FOR_EACH_VEC_ELT (vinfo->slp_instances, i, instance)
+    {
+      hash_set<slp_tree> visited;
+      vect_gather_slp_loads (SLP_INSTANCE_LOADS (instance),
+			     SLP_INSTANCE_TREE (instance), visited);
+    }
+}
+
 
 /* For each possible SLP instance decide whether to SLP it and calculate overall
    unrolling factor needed to SLP the loop.  Return TRUE if decided to SLP at
@@ -4152,6 +4159,9 @@ vect_slp_analyze_bb_1 (bb_vec_info bb_vinfo, int n_stmts, bool &fatal,
   /* Optimize permutations.  */
   vect_optimize_slp (bb_vinfo);
 
+  /* Gather the loads reachable from the SLP graph entries.  */
+  vect_gather_slp_loads (bb_vinfo);
+
   vect_record_base_alignments (bb_vinfo);
 
   /* Analyze and verify the alignment of data references and the
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 3ccd0fd552d..0ee4ef32eb2 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -1974,6 +1974,7 @@ extern opt_result vect_analyze_slp (vec_info *, unsigned);
 extern bool vect_make_slp_decision (loop_vec_info);
 extern void vect_detect_hybrid_slp (loop_vec_info);
 extern void vect_optimize_slp (vec_info *);
+extern void vect_gather_slp_loads (vec_info *);
 extern void vect_get_slp_defs (slp_tree, vec<tree> *);
 extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *,
 			       unsigned n = -1U);