Further tweaking of logic that decides when to materialize an uncorrelated
authorTom Lane
Sat, 11 Mar 2000 23:53:41 +0000 (23:53 +0000)
committerTom Lane
Sat, 11 Mar 2000 23:53:41 +0000 (23:53 +0000)
subplan: do it if subplan has subplans itself, and always do it if the
subplan is an indexscan.  (I originally set it to materialize an indexscan
only if the indexqual is fairly selective, but I dunno what I was
thinking ... an unselective indexscan is still expensive ...)

src/backend/optimizer/plan/subselect.c

index eea5b332cfebeef9711fd13faf8f195f0dff971c..a5df56d92046c51e967d2d8838a1b1c69e622897 100644 (file)
@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  * IDENTIFICATION
- *   $Header: /cvsroot/pgsql/src/backend/optimizer/plan/subselect.c,v 1.29 2000/03/02 04:08:16 tgl Exp $
+ *   $Header: /cvsroot/pgsql/src/backend/optimizer/plan/subselect.c,v 1.30 2000/03/11 23:53:41 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -302,8 +302,8 @@ make_subplan(SubLink *slink)
         * the top of an uncorrelated/undirect correlated subplan, which lets
         * us do the work of evaluating the subplan only once.  We do this
         * if the subplan's top plan node is anything more complicated than
-        * a sequential or index scan, and we do it even for those plan types
-        * if the qual appears selective enough to eliminate many tuples.
+        * a plain sequential scan, and we do it even for seqscan if the
+        * qual appears selective enough to eliminate many tuples.
         */
        if (node->parParam == NIL)
        {
@@ -312,28 +312,19 @@ make_subplan(SubLink *slink)
            switch (nodeTag(plan))
            {
                case T_SeqScan:
-               {
-                   Selectivity qualsel;
-
-                   qualsel = clauselist_selectivity(subquery, plan->qual, 0);
-                   /* Is 10% selectivity a good threshold?? */
-                   use_material = qualsel < 0.10;
-                   break;
-               }
-               case T_IndexScan:
-               {
-                   List       *indxqual = ((IndexScan *) plan)->indxqualorig;
-                   Selectivity qualsel;
-
-                   qualsel = clauselist_selectivity(subquery, plan->qual, 0);
-                   qualsel *= clauselist_selectivity(subquery, indxqual, 0);
-                   /* Note: if index is lossy, we just double-counted the
-                    * index selectivity.  Worth fixing?
-                    */
-                   /* Is 10% selectivity a good threshold?? */
-                   use_material = qualsel < 0.10;
+                   if (plan->initPlan || plan->subPlan)
+                       use_material = true;
+                   else
+                   {
+                       Selectivity qualsel;
+
+                       qualsel = clauselist_selectivity(subquery,
+                                                        plan->qual,
+                                                        0);
+                       /* Is 10% selectivity a good threshold?? */
+                       use_material = qualsel < 0.10;
+                   }
                    break;
-               }
                case T_Material:
                case T_Sort:
                    /* Don't add another Material node if there's one already,