]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
re PR tree-optimization/87105 (Autovectorization [X86, SSE2, AVX2, DoublePrecision])
authorRichard Biener <rguenther@suse.de>
Fri, 26 Oct 2018 07:38:59 +0000 (07:38 +0000)
committerRichard Biener <rguenth@gcc.gnu.org>
Fri, 26 Oct 2018 07:38:59 +0000 (07:38 +0000)
2018-10-26  Richard Biener  <rguenther@suse.de>

PR tree-optimization/87105
* tree-vectorizer.h (_slp_tree::refcnt): New member.
* tree-vect-slp.c (vect_free_slp_tree): Decrement and honor
refcnt.
(vect_create_new_slp_node): Initialize refcnt to one.
(bst_traits): Move.
(scalar_stmts_set_t, bst_fail): Remove.
(vect_build_slp_tree_2): Add bst_map argument and adjust calls.
(vect_build_slp_tree): Add bst_map argument and lookup
already created SLP nodes.
(vect_print_slp_tree): Handle a SLP graph, print SLP node
addresses.
(vect_slp_rearrange_stmts): Handle a SLP graph.
(vect_analyze_slp_instance): Adjust and free SLP nodes from
the CSE map.  Fix indenting.
(vect_schedule_slp_instance): Add short-cut.

* g++.dg/vect/slp-pr87105.cc: Adjust.
* gcc.dg/torture/20181024-1.c: New testcase.
* g++.dg/opt/20181025-1.C: Likewise.

From-SVN: r265522

gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/g++.dg/opt/20181025-1.C [new file with mode: 0644]
gcc/testsuite/g++.dg/vect/slp-pr87105.cc
gcc/testsuite/gcc.dg/torture/20181024-1.c [new file with mode: 0644]
gcc/tree-vect-slp.c
gcc/tree-vectorizer.h

index b60988c910e45158a6da505d099413887bb7113f..d46a3b47a55718865bbfab3117655f752a02347e 100644 (file)
@@ -1,3 +1,22 @@
+2018-10-26  Richard Biener  <rguenther@suse.de>
+
+       PR tree-optimization/87105
+       * tree-vectorizer.h (_slp_tree::refcnt): New member.
+       * tree-vect-slp.c (vect_free_slp_tree): Decrement and honor
+       refcnt.
+       (vect_create_new_slp_node): Initialize refcnt to one.
+       (bst_traits): Move.
+       (scalar_stmts_set_t, bst_fail): Remove.
+       (vect_build_slp_tree_2): Add bst_map argument and adjust calls.
+       (vect_build_slp_tree): Add bst_map argument and lookup
+       already created SLP nodes.
+       (vect_print_slp_tree): Handle a SLP graph, print SLP node
+       addresses.
+       (vect_slp_rearrange_stmts): Handle a SLP graph.
+       (vect_analyze_slp_instance): Adjust and free SLP nodes from
+       the CSE map.  Fix indenting.
+       (vect_schedule_slp_instance): Add short-cut.
+
 2018-10-26  Martin Liska  <mliska@suse.cz>
 
        PR testsuite/86158
index 73739030993f5e390072f95b7ff2dabd02125bdf..36627afeaa366622f736644bed60ece0ca415e80 100644 (file)
@@ -1,3 +1,10 @@
+2018-10-26  Richard Biener  <rguenther@suse.de>
+
+       PR tree-optimization/87105
+       * g++.dg/vect/slp-pr87105.cc: Adjust.
+       * gcc.dg/torture/20181024-1.c: New testcase.
+       * g++.dg/opt/20181025-1.C: Likewise.
+
 2018-10-26  Richard Biener  <rguenther@suse.de>
 
        PR testsuite/87754
diff --git a/gcc/testsuite/g++.dg/opt/20181025-1.C b/gcc/testsuite/g++.dg/opt/20181025-1.C
new file mode 100644 (file)
index 0000000..43d1614
--- /dev/null
@@ -0,0 +1,31 @@
+// { dg-do compile }
+// { dg-options "-Ofast" }
+
+template <typename Number>
+class Vector {
+    typedef Number value_type;
+    typedef const value_type *const_iterator;
+    Number norm_sqr () const;
+    const_iterator begin () const;
+    unsigned int dim;
+};
+template <typename Number>
+static inline Number
+local_sqr (const Number x)
+{
+  return x*x;
+}
+template <typename Number>
+Number
+Vector<Number>::norm_sqr () const
+{
+  Number sum0 = 0, sum1 = 0, sum2 = 0, sum3 = 0;
+  const_iterator ptr = begin(), eptr = ptr + (dim/4)*4;
+  while (ptr!=eptr) 
+    {
+      sum0 += ::local_sqr(*ptr++);
+      sum1 += ::local_sqr(*ptr++);
+    }
+  return sum0+sum1+sum2+sum3;
+}
+template class Vector<double>;
index 1023d915201cb3e584fec4a4b0dbcd971715447c..949b16c848f6717c04b35beaeafadf7f211f03ce 100644 (file)
@@ -2,7 +2,7 @@
 // { dg-require-effective-target c++11 }
 // { dg-require-effective-target vect_double }
 // For MIN/MAX recognition
-// { dg-additional-options "-ffast-math -fvect-cost-model" }
+// { dg-additional-options "-ffast-math" }
 
 #include <algorithm>
 #include <cmath>
@@ -99,6 +99,7 @@ void quadBoundingBoxA(const Point bez[3], Box& bBox) noexcept {
 
 // We should have if-converted everything down to straight-line code
 // { dg-final { scan-tree-dump-times "<bb \[0-9\]+>" 1 "slp2" } }
-// We fail to elide an earlier store which makes us not handle a later
-// duplicate one for vectorization.
-// { dg-final { scan-tree-dump-times "basic block part vectorized" 1 "slp2" { xfail *-*-* } } }
+// { dg-final { scan-tree-dump-times "basic block part vectorized" 1 "slp2" } }
+// It's a bit awkward to detect that all stores were vectorized but the
+// following more or less does the trick
+// { dg-final { scan-tree-dump "vect_iftmp\[^\r\m\]* = MIN" "slp2" } }
diff --git a/gcc/testsuite/gcc.dg/torture/20181024-1.c b/gcc/testsuite/gcc.dg/torture/20181024-1.c
new file mode 100644 (file)
index 0000000..f2cfe7f
--- /dev/null
@@ -0,0 +1,41 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-march=core-avx2" { target { x86_64-*-* i?86-*-* } } } */
+
+typedef enum {
+ C = 0,               N, S, E, W, T, B,               NE, NW, SE, SW,               NT, NB, ST, SB,               ET, EB, WT, WB,               FLAGS, N_CELL_ENTRIES} CELL_ENTRIES;
+typedef double LBM_Grid[(130)*100*100*N_CELL_ENTRIES];
+void foo( LBM_Grid srcGrid )
+{
+  double ux , uy , uz , rho ,         ux1, uy1, uz1, rho1,         ux2, uy2, uz2, rho2,         u2, px, py;
+  int i;
+  for( i = 0;
+       i < (N_CELL_ENTRIES*( 100*100));
+       i += N_CELL_ENTRIES )
+    {
+      rho1 = + ((srcGrid)[((C)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((N)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((S)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((E)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((W)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((T)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((B)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((NE)+N_CELL_ENTRIES*( 100*100))+(i)]) 
+         + ((srcGrid)[((NW)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((SE)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((SW)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((NT)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((NB)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((ST)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((SB)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((ET)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((EB)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((WT)+N_CELL_ENTRIES*( 100*100))+(i)])
+         + ((srcGrid)[((WB)+N_CELL_ENTRIES*( 100*100))+(i)]);
+      rho = 2.0*rho1 - rho2;
+      px = (((i / N_CELL_ENTRIES) % 100) / (0.5*(100-1))) - 1.0;
+      uz = 0.01 * (1.0-px*px) * (1.0-py*py);
+      u2 = 1.5 * (ux*ux + uy*uy + uz*uz);
+      (((srcGrid)[((C))+(i)])) = (1.0/ 3.0)*rho*(1.0 - u2);
+      (((srcGrid)[((N))+(i)])) = (1.0/18.0)*rho*(1.0 + uy*(4.5*uy + 3.0) - u2);
+    }
+}
index 3aae1776ef91b27ea98d4e68e6e11cf566dfcbd6..ab8504a10bd18a6be4fa0c6741c7c0046c102caf 100644 (file)
@@ -57,6 +57,9 @@ vect_free_slp_tree (slp_tree node, bool final_p)
   int i;
   slp_tree child;
 
+  if (--node->refcnt != 0)
+    return;
+
   FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
     vect_free_slp_tree (child, final_p);
 
@@ -82,7 +85,6 @@ vect_free_slp_tree (slp_tree node, bool final_p)
   free (node);
 }
 
-
 /* Free the memory allocated for the SLP instance.  FINAL_P is true if we
    have vectorized the instance or if we have made a final decision not
    to vectorize the statements in any way.  */
@@ -126,6 +128,7 @@ vect_create_new_slp_node (vec<stmt_vec_info> scalar_stmts)
   SLP_TREE_LOAD_PERMUTATION (node) = vNULL;
   SLP_TREE_TWO_OPERATORS (node) = false;
   SLP_TREE_DEF_TYPE (node) = vect_internal_def;
+  node->refcnt = 1;
 
   unsigned i;
   FOR_EACH_VEC_ELT (scalar_stmts, i, stmt_info)
@@ -1021,9 +1024,6 @@ bst_traits::equal (value_type existing, value_type candidate)
   return true;
 }
 
-typedef hash_set <vec <gimple *>, bst_traits> scalar_stmts_set_t;
-static scalar_stmts_set_t *bst_fail;
-
 typedef hash_map <vec <gimple *>, slp_tree,
                  simple_hashmap_traits <bst_traits, slp_tree> >
   scalar_stmts_to_slp_tree_map_t;
@@ -1034,30 +1034,33 @@ vect_build_slp_tree_2 (vec_info *vinfo,
                       poly_uint64 *max_nunits,
                       vec<slp_tree> *loads,
                       bool *matches, unsigned *npermutes, unsigned *tree_size,
-                      unsigned max_tree_size);
+                      unsigned max_tree_size,
+                      scalar_stmts_to_slp_tree_map_t *bst_map);
 
 static slp_tree
 vect_build_slp_tree (vec_info *vinfo,
                     vec<stmt_vec_info> stmts, unsigned int group_size,
                     poly_uint64 *max_nunits, vec<slp_tree> *loads,
                     bool *matches, unsigned *npermutes, unsigned *tree_size,
-                    unsigned max_tree_size)
+                    unsigned max_tree_size,
+                    scalar_stmts_to_slp_tree_map_t *bst_map)
 {
-  if (bst_fail->contains (stmts))
-    return NULL;
-  slp_tree res = vect_build_slp_tree_2 (vinfo, stmts, group_size, max_nunits,
-                                       loads, matches, npermutes, tree_size,
-                                       max_tree_size);
-  /* When SLP build fails for stmts record this, otherwise SLP build
-     can be exponential in time when we allow to construct parts from
-     scalars, see PR81723.  */
-  if (! res)
+  if (slp_tree *leader = bst_map->get (stmts))
     {
-      vec <stmt_vec_info> x;
-      x.create (stmts.length ());
-      x.splice (stmts);
-      bst_fail->add (x);
+      if (dump_enabled_p ())
+       dump_printf_loc (MSG_NOTE, vect_location, "re-using %sSLP tree %p\n",
+                        *leader ? "" : "failed ", *leader);
+      if (*leader)
+       (*leader)->refcnt++;
+      return *leader;
     }
+  slp_tree res = vect_build_slp_tree_2 (vinfo, stmts, group_size, max_nunits,
+                                       loads, matches, npermutes, tree_size,
+                                       max_tree_size, bst_map);
+  /* Keep a reference for the bst_map use.  */
+  if (res)
+    res->refcnt++;
+  bst_map->put (stmts.copy (), res);
   return res;
 }
 
@@ -1074,7 +1077,8 @@ vect_build_slp_tree_2 (vec_info *vinfo,
                       poly_uint64 *max_nunits,
                       vec<slp_tree> *loads,
                       bool *matches, unsigned *npermutes, unsigned *tree_size,
-                      unsigned max_tree_size)
+                      unsigned max_tree_size,
+                      scalar_stmts_to_slp_tree_map_t *bst_map)
 {
   unsigned nops, i, this_tree_size = 0;
   poly_uint64 this_max_nunits = *max_nunits;
@@ -1205,7 +1209,7 @@ vect_build_slp_tree_2 (vec_info *vinfo,
                                        group_size, &this_max_nunits,
                                        &this_loads, matches, npermutes,
                                        &this_tree_size,
-                                       max_tree_size)) != NULL)
+                                       max_tree_size, bst_map)) != NULL)
        {
          /* If we have all children of child built up from scalars then just
             throw that away and build it up this node from scalars.  */
@@ -1348,7 +1352,7 @@ vect_build_slp_tree_2 (vec_info *vinfo,
                                            group_size, &this_max_nunits,
                                            &this_loads, tem, npermutes,
                                            &this_tree_size,
-                                           max_tree_size)) != NULL)
+                                           max_tree_size, bst_map)) != NULL)
            {
              /* ... so if successful we can apply the operand swapping
                 to the GIMPLE IL.  This is necessary because for example
@@ -1441,21 +1445,37 @@ fail:
 
 static void
 vect_print_slp_tree (dump_flags_t dump_kind, dump_location_t loc,
-                    slp_tree node)
+                    slp_tree node, hash_set<slp_tree> &visited)
 {
   int i;
   stmt_vec_info stmt_info;
   slp_tree child;
 
-  dump_printf_loc (dump_kind, loc, "node%s\n",
+  if (visited.add (node))
+    return;
+
+  dump_printf_loc (dump_kind, loc, "node%s %p\n",
                   SLP_TREE_DEF_TYPE (node) != vect_internal_def
-                  ? " (external)" : "");
+                  ? " (external)" : "", node);
   FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
     dump_printf_loc (dump_kind, loc, "\tstmt %d %G", i, stmt_info->stmt);
+  if (SLP_TREE_CHILDREN (node).is_empty ())
+    return;
+  dump_printf_loc (dump_kind, loc, "\tchildren");
   FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
-    vect_print_slp_tree (dump_kind, loc, child);
+    dump_printf (dump_kind, " %p", (void *)child);
+  dump_printf (dump_kind, "\n");
+  FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
+    vect_print_slp_tree (dump_kind, loc, child, visited);
 }
 
+static void
+vect_print_slp_tree (dump_flags_t dump_kind, dump_location_t loc,
+                    slp_tree node)
+{
+  hash_set<slp_tree> visited;
+  vect_print_slp_tree (dump_kind, loc, node, visited);
+}
 
 /* Mark the tree rooted at NODE with MARK (PURE_SLP or HYBRID).
    If MARK is HYBRID, it refers to a specific stmt in NODE (the stmt at index
@@ -1509,15 +1529,19 @@ vect_mark_slp_stmts_relevant (slp_tree node)
 
 static void
 vect_slp_rearrange_stmts (slp_tree node, unsigned int group_size,
-                          vec<unsigned> permutation)
+                          vec<unsigned> permutation,
+                         hash_set<slp_tree> &visited)
 {
   stmt_vec_info stmt_info;
   vec<stmt_vec_info> tmp_stmts;
   unsigned int i;
   slp_tree child;
 
+  if (visited.add (node))
+    return;
+
   FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
-    vect_slp_rearrange_stmts (child, group_size, permutation);
+    vect_slp_rearrange_stmts (child, group_size, permutation, visited);
 
   gcc_assert (group_size == SLP_TREE_SCALAR_STMTS (node).length ());
   tmp_stmts.create (group_size);
@@ -1578,8 +1602,9 @@ vect_attempt_slp_rearrange_stmts (slp_instance slp_instn)
      statements in the nodes is not important unless they are memory
      accesses, we can rearrange the statements in all the nodes
      according to the order of the loads.  */
+  hash_set<slp_tree> visited;
   vect_slp_rearrange_stmts (SLP_INSTANCE_TREE (slp_instn), group_size,
-                           node->load_permutation);
+                           node->load_permutation, visited);
 
   /* We are done, no actual permutations need to be generated.  */
   poly_uint64 unrolling_factor = SLP_INSTANCE_UNROLLING_FACTOR (slp_instn);
@@ -1889,12 +1914,18 @@ vect_analyze_slp_instance (vec_info *vinfo,
   /* Build the tree for the SLP instance.  */
   bool *matches = XALLOCAVEC (bool, group_size);
   unsigned npermutes = 0;
-  bst_fail = new scalar_stmts_set_t ();
+  scalar_stmts_to_slp_tree_map_t *bst_map
+    = new scalar_stmts_to_slp_tree_map_t ();
   poly_uint64 max_nunits = nunits;
   node = vect_build_slp_tree (vinfo, scalar_stmts, group_size,
                              &max_nunits, &loads, matches, &npermutes,
-                             NULL, max_tree_size);
-  delete bst_fail;
+                             NULL, max_tree_size, bst_map);
+  /* The map keeps a reference on SLP nodes built, release that.  */
+  for (scalar_stmts_to_slp_tree_map_t::iterator it = bst_map->begin ();
+       it != bst_map->end (); ++it)
+    if ((*it).second)
+      vect_free_slp_tree ((*it).second, false);
+  delete bst_map;
   if (node != NULL)
     {
       /* Calculate the unrolling factor based on the smallest type.  */
@@ -1924,109 +1955,109 @@ vect_analyze_slp_instance (vec_info *vinfo,
        }
       else
        {
-      /* Create a new SLP instance.  */
-      new_instance = XNEW (struct _slp_instance);
-      SLP_INSTANCE_TREE (new_instance) = node;
-      SLP_INSTANCE_GROUP_SIZE (new_instance) = group_size;
-      SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor;
-      SLP_INSTANCE_LOADS (new_instance) = loads;
-
-      /* Compute the load permutation.  */
-      slp_tree load_node;
-      bool loads_permuted = false;
-      FOR_EACH_VEC_ELT (loads, i, load_node)
-       {
-         vec<unsigned> load_permutation;
-         int j;
-         stmt_vec_info load_info;
-         bool this_load_permuted = false;
-         load_permutation.create (group_size);
-         stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT
-           (SLP_TREE_SCALAR_STMTS (load_node)[0]);
-         FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node), j, load_info)
+         /* Create a new SLP instance.  */
+         new_instance = XNEW (struct _slp_instance);
+         SLP_INSTANCE_TREE (new_instance) = node;
+         SLP_INSTANCE_GROUP_SIZE (new_instance) = group_size;
+         SLP_INSTANCE_UNROLLING_FACTOR (new_instance) = unrolling_factor;
+         SLP_INSTANCE_LOADS (new_instance) = loads;
+
+         /* Compute the load permutation.  */
+         slp_tree load_node;
+         bool loads_permuted = false;
+         FOR_EACH_VEC_ELT (loads, i, load_node)
            {
-             int load_place = vect_get_place_in_interleaving_chain
-               (load_info, first_stmt_info);
-             gcc_assert (load_place != -1);
-             if (load_place != j)
-               this_load_permuted = true;
-             load_permutation.safe_push (load_place);
+             vec<unsigned> load_permutation;
+             int j;
+             stmt_vec_info load_info;
+             bool this_load_permuted = false;
+             load_permutation.create (group_size);
+             stmt_vec_info first_stmt_info = DR_GROUP_FIRST_ELEMENT
+                 (SLP_TREE_SCALAR_STMTS (load_node)[0]);
+             FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node), j, load_info)
+               {
+                 int load_place = vect_get_place_in_interleaving_chain
+                     (load_info, first_stmt_info);
+                 gcc_assert (load_place != -1);
+                 if (load_place != j)
+                   this_load_permuted = true;
+                 load_permutation.safe_push (load_place);
+               }
+             if (!this_load_permuted
+                 /* The load requires permutation when unrolling exposes
+                    a gap either because the group is larger than the SLP
+                    group-size or because there is a gap between the groups.  */
+                 && (known_eq (unrolling_factor, 1U)
+                     || (group_size == DR_GROUP_SIZE (first_stmt_info)
+                         && DR_GROUP_GAP (first_stmt_info) == 0)))
+               {
+                 load_permutation.release ();
+                 continue;
+               }
+             SLP_TREE_LOAD_PERMUTATION (load_node) = load_permutation;
+             loads_permuted = true;
            }
-         if (!this_load_permuted
-             /* The load requires permutation when unrolling exposes
-                a gap either because the group is larger than the SLP
-                group-size or because there is a gap between the groups.  */
-             && (known_eq (unrolling_factor, 1U)
-                 || (group_size == DR_GROUP_SIZE (first_stmt_info)
-                     && DR_GROUP_GAP (first_stmt_info) == 0)))
+
+         if (loads_permuted)
            {
-             load_permutation.release ();
-             continue;
+             if (!vect_supported_load_permutation_p (new_instance))
+               {
+                 if (dump_enabled_p ())
+                   dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                                    "Build SLP failed: unsupported load "
+                                    "permutation %G", stmt_info->stmt);
+                 vect_free_slp_instance (new_instance, false);
+                 return false;
+               }
            }
-         SLP_TREE_LOAD_PERMUTATION (load_node) = load_permutation;
-         loads_permuted = true;
-       }
-
-      if (loads_permuted)
-        {
-          if (!vect_supported_load_permutation_p (new_instance))
-            {
-              if (dump_enabled_p ())
-               dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
-                                "Build SLP failed: unsupported load "
-                                "permutation %G", stmt_info->stmt);
-             vect_free_slp_instance (new_instance, false);
-              return false;
-            }
-        }
 
          /* If the loads and stores can be handled with load/store-lan
-        instructions do not generate this SLP instance.  */
-      if (is_a <loop_vec_info> (vinfo)
-         && loads_permuted
-         && dr && vect_store_lanes_supported (vectype, group_size, false))
-       {
-         slp_tree load_node;
-         FOR_EACH_VEC_ELT (loads, i, load_node)
+            instructions do not generate this SLP instance.  */
+         if (is_a <loop_vec_info> (vinfo)
+             && loads_permuted
+             && dr && vect_store_lanes_supported (vectype, group_size, false))
            {
-             stmt_vec_info stmt_vinfo = DR_GROUP_FIRST_ELEMENT
-               (SLP_TREE_SCALAR_STMTS (load_node)[0]);
-             /* Use SLP for strided accesses (or if we can't load-lanes).  */
-             if (STMT_VINFO_STRIDED_P (stmt_vinfo)
-                 || ! vect_load_lanes_supported
-                       (STMT_VINFO_VECTYPE (stmt_vinfo),
-                        DR_GROUP_SIZE (stmt_vinfo), false))
-               break;
+             slp_tree load_node;
+             FOR_EACH_VEC_ELT (loads, i, load_node)
+               {
+                 stmt_vec_info stmt_vinfo = DR_GROUP_FIRST_ELEMENT
+                     (SLP_TREE_SCALAR_STMTS (load_node)[0]);
+                 /* Use SLP for strided accesses (or if we can't load-lanes).  */
+                 if (STMT_VINFO_STRIDED_P (stmt_vinfo)
+                     || ! vect_load_lanes_supported
+                     (STMT_VINFO_VECTYPE (stmt_vinfo),
+                      DR_GROUP_SIZE (stmt_vinfo), false))
+                   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, false);
+                 return false;
+               }
            }
-         if (i == loads.length ())
+
+         vinfo->slp_instances.safe_push (new_instance);
+
+         if (dump_enabled_p ())
            {
-             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, false);
-             return false;
+             dump_printf_loc (MSG_NOTE, vect_location,
+                              "Final SLP tree for instance:\n");
+             vect_print_slp_tree (MSG_NOTE, vect_location, node);
            }
-       }
 
-      vinfo->slp_instances.safe_push (new_instance);
-
-      if (dump_enabled_p ())
-       {
-         dump_printf_loc (MSG_NOTE, vect_location,
-                          "Final SLP tree for instance:\n");
-         vect_print_slp_tree (MSG_NOTE, vect_location, node);
+         return true;
        }
-
-      return true;
-    }
     }
   else
     {
-  /* Failed to SLP.  */
-  /* Free the allocated memory.  */
-  scalar_stmts.release ();
-  loads.release ();
+      /* Failed to SLP.  */
+      /* Free the allocated memory.  */
+      scalar_stmts.release ();
+      loads.release ();
     }
 
   /* For basic block SLP, try to break the group up into multiples of the
@@ -3749,8 +3780,13 @@ vect_schedule_slp_instance (slp_tree node, slp_instance instance,
   if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
     return;
 
+  /* See if we have already vectorized the node in the graph of the
+     SLP instance.  */
+  if (SLP_TREE_VEC_STMTS (node).exists ())
+    return;
+
   /* See if we have already vectorized the same set of stmts and reuse their
-     vectorized stmts.  */
+     vectorized stmts across instances.  */
   if (slp_tree *leader = bst_map->get (SLP_TREE_SCALAR_STMTS (node)))
     {
       SLP_TREE_VEC_STMTS (node).safe_splice (SLP_TREE_VEC_STMTS (*leader));
@@ -3778,8 +3814,7 @@ vect_schedule_slp_instance (slp_tree node, slp_instance instance,
   group_size = SLP_INSTANCE_GROUP_SIZE (instance);
 
   gcc_assert (SLP_TREE_NUMBER_OF_VEC_STMTS (node) != 0);
-  if (!SLP_TREE_VEC_STMTS (node).exists ())
-    SLP_TREE_VEC_STMTS (node).create (SLP_TREE_NUMBER_OF_VEC_STMTS (node));
+  SLP_TREE_VEC_STMTS (node).create (SLP_TREE_NUMBER_OF_VEC_STMTS (node));
 
   if (dump_enabled_p ())
     dump_printf_loc (MSG_NOTE, vect_location,
index 08d696a2f7c85c784e8c64f7c157d5fedcb0719d..e1292aa6eb6b743844869c5343b2c462429757f2 100644 (file)
@@ -130,6 +130,8 @@ struct _slp_tree {
      scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
      divided by vector size.  */
   unsigned int vec_stmts_size;
+  /* Reference count in the SLP graph.  */
+  unsigned int refcnt;
   /* Whether the scalar computations use two different operators.  */
   bool two_operators;
   /* The DEF type of this node.  */