re PR rtl-optimization/40761 (IRA memory hog for insanely nested loops)

2012-01-19  Vladimir Makarov  <vmakarov@redhat.com>

	PR rtl-optimization/40761
	* ira-int.h (struct ira_loop_tree_node): Add comment for member
	loop.  Add new member loop_num.
	(IRA_LOOP_NODE_BY_INDEX): Modify the check.
	(ira_build): Remove the parameter.

	* ira.c (ira_print_disposition): Use loop_num instead of
	loop->num.
	(ira.c): Do not build CFG loops for one region allocation.  Remove
	argument from ira_build call.

	* ira-build.c (init_loop_tree_node): New function.
	(create_loop_tree_nodes): Use it.  Separate the case when CFG
	loops are not built.
	(more_one_region_p): Check current_loops.
	(finish_loop_tree_nodes): Separate the case when CFG loops are not
	built.
	(add_loop_to_tree): Process loop equal to NULL too.
	(form_loop_tree): Separate the case when CFG loops are not built.
	Use explicitly number for the root.
	(rebuild_regno_allocno_maps, create_loop_tree_node_allocnos): Add
	an assertion.
	(ira_print_expanded_allocno, loop_compare_func): Use loop_num
	instead of loop->num.
	(mark_loops_for_removal): Ditto.  Use loop_num instead of
	loop->num.
	(mark_all_loops_for_removal): Ditto.
	(remove_unnecessary_regions): Separate the case when CFG loops
	are not built.
	(ira_build): Remove the parameter.  Use explicit number of regions
	when CFG loops are not built.

	* ira-color.c (print_loop_title): Separate the case for the root
	node.  Use loop_num instead of loop->num.
	(move_spill_restore): Use loop_num instead of loop->num.

	* ira-emit.c (setup_entered_from_non_parent_p): Add an assertion.
	(change_loop): Ditto.
	(change_loop): Use loop_num instead of loop->num.

	* ira-lives.c (process_bb_node_lives): Ditto.

	* ira-costs.c (print_allocno_costs, find_costs_and_classes):
	Ditto.

	* ira-conflicts.c (print_allocno_conflicts): Ditto.

From-SVN: r183312

1 files changed, 95 insertions, 67 deletions

diff --git a/gcc/ira-build.c b/gcc/ira-build.c
index 2d8708a..78d0b36 100644
--- a/gcc/ira-build.c
+++ b/gcc/ira-build.c
@@ -93,15 +93,35 @@ int ira_copies_num;
    basic block.  */
 static int last_basic_block_before_change;
 
-/* The following function allocates the loop tree nodes.  If LOOPS_P
-   is FALSE, the nodes corresponding to the loops (except the root
-   which corresponds the all function) will be not allocated but nodes
-   will still be allocated for basic blocks.  */
+/* Initialize some members in loop tree node NODE.  Use LOOP_NUM for
+   the member loop_num.  */
 static void
-create_loop_tree_nodes (bool loops_p)
+init_loop_tree_node (struct ira_loop_tree_node *node, int loop_num)
+{
+  int max_regno = max_reg_num ();
+
+  node->regno_allocno_map
+    = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t) * max_regno);
+  memset (node->regno_allocno_map, 0, sizeof (ira_allocno_t) * max_regno);
+  memset (node->reg_pressure, 0, sizeof (node->reg_pressure));
+  node->all_allocnos = ira_allocate_bitmap ();
+  node->modified_regnos = ira_allocate_bitmap ();
+  node->border_allocnos = ira_allocate_bitmap ();
+  node->local_copies = ira_allocate_bitmap ();
+  node->loop_num = loop_num;
+  node->children = NULL;
+  node->subloops = NULL;
+}
+
+
+/* The following function allocates the loop tree nodes.  If
+   CURRENT_LOOPS is NULL, the nodes corresponding to the loops (except
+   the root which corresponds the all function) will be not allocated
+   but nodes will still be allocated for basic blocks.  */
+static void
+create_loop_tree_nodes (void)
 {
   unsigned int i, j;
-  int max_regno;
   bool skip_p;
   edge_iterator ei;
   edge e;
@@ -122,17 +142,21 @@ create_loop_tree_nodes (bool loops_p)
       ira_bb_nodes[i].border_allocnos = NULL;
       ira_bb_nodes[i].local_copies = NULL;
     }
+  if (current_loops == NULL)
+    {
+      ira_loop_nodes = ((struct ira_loop_tree_node *)
+			ira_allocate (sizeof (struct ira_loop_tree_node)));
+      init_loop_tree_node (ira_loop_nodes, 0);
+      return;
+    }
   ira_loop_nodes = ((struct ira_loop_tree_node *)
 		    ira_allocate (sizeof (struct ira_loop_tree_node)
 				  * VEC_length (loop_p, ira_loops.larray)));
-  max_regno = max_reg_num ();
   FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
     {
       if (loop != ira_loops.tree_root)
 	{
 	  ira_loop_nodes[i].regno_allocno_map = NULL;
-	  if (! loops_p)
-	    continue;
 	  skip_p = false;
 	  FOR_EACH_EDGE (e, ei, loop->header->preds)
 	    if (e->src != loop->latch
@@ -154,16 +178,7 @@ create_loop_tree_nodes (bool loops_p)
 	  if (skip_p)
 	    continue;
 	}
-      ira_loop_nodes[i].regno_allocno_map
-	= (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t) * max_regno);
-      memset (ira_loop_nodes[i].regno_allocno_map, 0,
-	      sizeof (ira_allocno_t) * max_regno);
-      memset (ira_loop_nodes[i].reg_pressure, 0,
-	      sizeof (ira_loop_nodes[i].reg_pressure));
-      ira_loop_nodes[i].all_allocnos = ira_allocate_bitmap ();
-      ira_loop_nodes[i].modified_regnos = ira_allocate_bitmap ();
-      ira_loop_nodes[i].border_allocnos = ira_allocate_bitmap ();
-      ira_loop_nodes[i].local_copies = ira_allocate_bitmap ();
+      init_loop_tree_node (&ira_loop_nodes[i], loop->num);
     }
 }
 
@@ -175,10 +190,11 @@ more_one_region_p (void)
   unsigned int i;
   loop_p loop;
 
-  FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
-    if (ira_loop_nodes[i].regno_allocno_map != NULL
-	&& ira_loop_tree_root != &ira_loop_nodes[i])
-      return true;
+  if (current_loops != NULL)
+    FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
+      if (ira_loop_nodes[i].regno_allocno_map != NULL
+	  && ira_loop_tree_root != &ira_loop_nodes[i])
+	return true;
   return false;
 }
 
@@ -205,8 +221,11 @@ finish_loop_tree_nodes (void)
   unsigned int i;
   loop_p loop;
 
-  FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
-    finish_loop_tree_node (&ira_loop_nodes[i]);
+  if (current_loops == NULL)
+    finish_loop_tree_node (&ira_loop_nodes[0]);
+  else
+    FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
+      finish_loop_tree_node (&ira_loop_nodes[i]);
   ira_free (ira_loop_nodes);
   for (i = 0; i < (unsigned int) last_basic_block_before_change; i++)
     {
@@ -227,30 +246,39 @@ finish_loop_tree_nodes (void)
 
 
 /* The following recursive function adds LOOP to the loop tree
-   hierarchy.  LOOP is added only once.  */
+   hierarchy.  LOOP is added only once.  If LOOP is NULL we adding
+   loop designating the whole function when CFG loops are not
+   built.  */
 static void
 add_loop_to_tree (struct loop *loop)
 {
+  int loop_num;
   struct loop *parent;
   ira_loop_tree_node_t loop_node, parent_node;
 
   /* We can not use loop node access macros here because of potential
      checking and because the nodes are not initialized enough
      yet.  */
-  if (loop_outer (loop) != NULL)
+  if (loop != NULL && loop_outer (loop) != NULL)
     add_loop_to_tree (loop_outer (loop));
-  if (ira_loop_nodes[loop->num].regno_allocno_map != NULL
-      && ira_loop_nodes[loop->num].children == NULL)
+  loop_num = loop != NULL ? loop->num : 0;
+  if (ira_loop_nodes[loop_num].regno_allocno_map != NULL
+      && ira_loop_nodes[loop_num].children == NULL)
     {
       /* We have not added loop node to the tree yet.  */
-      loop_node = &ira_loop_nodes[loop->num];
+      loop_node = &ira_loop_nodes[loop_num];
       loop_node->loop = loop;
       loop_node->bb = NULL;
-      for (parent = loop_outer (loop);
-	   parent != NULL;
-	   parent = loop_outer (parent))
-	if (ira_loop_nodes[parent->num].regno_allocno_map != NULL)
-	  break;
+      if (loop == NULL)
+	parent = NULL;
+      else
+	{
+	  for (parent = loop_outer (loop);
+	       parent != NULL;
+	       parent = loop_outer (parent))
+	    if (ira_loop_nodes[parent->num].regno_allocno_map != NULL)
+	      break;
+	}
       if (parent == NULL)
 	{
 	  loop_node->next = NULL;
@@ -299,21 +327,13 @@ setup_loop_tree_level (ira_loop_tree_node_t loop_node, int level)
 static void
 form_loop_tree (void)
 {
-  unsigned int i;
   basic_block bb;
   struct loop *parent;
   ira_loop_tree_node_t bb_node, loop_node;
-  loop_p loop;
 
   /* We can not use loop/bb node access macros because of potential
      checking and because the nodes are not initialized enough
      yet.  */
-  FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
-     if (ira_loop_nodes[i].regno_allocno_map != NULL)
-       {
-	 ira_loop_nodes[i].children = NULL;
-	 ira_loop_nodes[i].subloops = NULL;
-       }
   FOR_EACH_BB (bb)
     {
       bb_node = &ira_bb_nodes[bb->index];
@@ -323,18 +343,23 @@ form_loop_tree (void)
       bb_node->children = NULL;
       bb_node->subloop_next = NULL;
       bb_node->next = NULL;
-      for (parent = bb->loop_father;
-	   parent != NULL;
-	   parent = loop_outer (parent))
-	if (ira_loop_nodes[parent->num].regno_allocno_map != NULL)
-	  break;
+      if (current_loops == NULL)
+	parent = NULL;
+      else
+	{
+	  for (parent = bb->loop_father;
+	       parent != NULL;
+	       parent = loop_outer (parent))
+	    if (ira_loop_nodes[parent->num].regno_allocno_map != NULL)
+	      break;
+	}
       add_loop_to_tree (parent);
-      loop_node = &ira_loop_nodes[parent->num];
+      loop_node = &ira_loop_nodes[parent == NULL ? 0 : parent->num];
       bb_node->next = loop_node->children;
       bb_node->parent = loop_node;
       loop_node->children = bb_node;
     }
-  ira_loop_tree_root = IRA_LOOP_NODE_BY_INDEX (ira_loops.tree_root->num);
+  ira_loop_tree_root = IRA_LOOP_NODE_BY_INDEX (0);
   ira_loop_tree_height = setup_loop_tree_level (ira_loop_tree_root, 0);
   ira_assert (ira_loop_tree_root->regno_allocno_map != NULL);
 }
@@ -353,6 +378,7 @@ rebuild_regno_allocno_maps (void)
   loop_p loop;
   ira_allocno_iterator ai;
 
+  ira_assert (current_loops != NULL);
   max_regno = max_reg_num ();
   FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, l, loop)
     if (ira_loop_nodes[l].regno_allocno_map != NULL)
@@ -837,7 +863,7 @@ ira_print_expanded_allocno (ira_allocno_t a)
   if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL)
     fprintf (ira_dump_file, ",b%d", bb->index);
   else
-    fprintf (ira_dump_file, ",l%d", ALLOCNO_LOOP_TREE_NODE (a)->loop->num);
+    fprintf (ira_dump_file, ",l%d", ALLOCNO_LOOP_TREE_NODE (a)->loop_num);
   if (ALLOCNO_CAP_MEMBER (a) != NULL)
     {
       fprintf (ira_dump_file, ":");
@@ -1621,6 +1647,7 @@ create_loop_tree_node_allocnos (ira_loop_tree_node_t loop_node)
       edge e;
       VEC (edge, heap) *edges;
 
+      ira_assert (current_loops != NULL);
       FOR_EACH_EDGE (e, ei, loop_node->loop->header->preds)
 	if (e->src != loop_node->loop->latch)
 	  create_loop_allocnos (e);
@@ -1848,7 +1875,7 @@ loop_compare_func (const void *v1p, const void *v2p)
   if ((diff = (int) loop_depth (l1->loop) - (int) loop_depth (l2->loop)) != 0)
     return diff;
   /* Make sorting stable.  */
-  return l1->loop->num - l2->loop->num;
+  return l1->loop_num - l2->loop_num;
 }
 
 /* Mark loops which should be removed from regional allocation.  We
@@ -1870,6 +1897,7 @@ mark_loops_for_removal (void)
   ira_loop_tree_node_t *sorted_loops;
   loop_p loop;
 
+  ira_assert (current_loops != NULL);
   sorted_loops
     = (ira_loop_tree_node_t *) ira_allocate (sizeof (ira_loop_tree_node_t)
 					     * VEC_length (loop_p,
@@ -1900,7 +1928,7 @@ mark_loops_for_removal (void)
 	fprintf
 	  (ira_dump_file,
 	   "  Mark loop %d (header %d, freq %d, depth %d) for removal (%s)\n",
-	   sorted_loops[i]->loop->num, sorted_loops[i]->loop->header->index,
+	   sorted_loops[i]->loop_num, sorted_loops[i]->loop->header->index,
 	   sorted_loops[i]->loop->header->frequency,
 	   loop_depth (sorted_loops[i]->loop),
 	   low_pressure_loop_node_p (sorted_loops[i]->parent)
@@ -1917,6 +1945,7 @@ mark_all_loops_for_removal (void)
   int i;
   loop_p loop;
 
+  ira_assert (current_loops != NULL);
   FOR_EACH_VEC_ELT (loop_p, ira_loops.larray, i, loop)
     if (ira_loop_nodes[i].regno_allocno_map != NULL)
       {
@@ -1931,7 +1960,7 @@ mark_all_loops_for_removal (void)
 	  fprintf
 	    (ira_dump_file,
 	     "  Mark loop %d (header %d, freq %d, depth %d) for removal\n",
-	     ira_loop_nodes[i].loop->num,
+	     ira_loop_nodes[i].loop_num,
 	     ira_loop_nodes[i].loop->header->index,
 	     ira_loop_nodes[i].loop->header->frequency,
 	     loop_depth (ira_loop_nodes[i].loop));
@@ -2221,6 +2250,8 @@ remove_low_level_allocnos (void)
 static void
 remove_unnecessary_regions (bool all_p)
 {
+  if (current_loops == NULL)
+    return;
   if (all_p)
     mark_all_loops_for_removal ();
   else
@@ -3026,23 +3057,20 @@ update_conflict_hard_reg_costs (void)
 }
 
 /* Create a internal representation (IR) for IRA (allocnos, copies,
-   loop tree nodes).  If LOOPS_P is FALSE the nodes corresponding to
-   the loops (except the root which corresponds the all function) and
-   correspondingly allocnos for the loops will be not created.  Such
-   parameter value is used for Chaitin-Briggs coloring.  The function
-   returns TRUE if we generate loop structure (besides nodes
-   representing all function and the basic blocks) for regional
-   allocation.  A true return means that we really need to flatten IR
-   before the reload.  */
+   loop tree nodes).  The function returns TRUE if we generate loop
+   structure (besides nodes representing all function and the basic
+   blocks) for regional allocation.  A true return means that we
+   really need to flatten IR before the reload.  */
 bool
-ira_build (bool loops_p)
+ira_build (void)
 {
-  df_analyze ();
+  bool loops_p;
 
+  df_analyze ();
   initiate_cost_vectors ();
   initiate_allocnos ();
   initiate_copies ();
-  create_loop_tree_nodes (loops_p);
+  create_loop_tree_nodes ();
   form_loop_tree ();
   create_allocnos ();
   ira_costs ();
@@ -3111,8 +3139,8 @@ ira_build (bool loops_p)
 	    }
 	}
       fprintf (ira_dump_file, "  regions=%d, blocks=%d, points=%d\n",
-	       VEC_length (loop_p, ira_loops.larray), n_basic_blocks,
-	       ira_max_point);
+	       current_loops == NULL ? 1 : VEC_length (loop_p, ira_loops.larray),
+	       n_basic_blocks, ira_max_point);
       fprintf (ira_dump_file,
 	       "    allocnos=%d (big %d), copies=%d, conflicts=%d, ranges=%d\n",
 	       ira_allocnos_num, nr_big, ira_copies_num, n, nr);