Be more realistic about plans involving Materialize nodes: take their

cost into account while planning.

src/backend/executor/execAmi.c

@@ -6,7 +6,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- *	$Id: execAmi.c,v 1.64 2002/06/20 20:29:27 momjian Exp $
+ *	$Id: execAmi.c,v 1.65 2002/11/30 05:21:01 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -170,14 +170,10 @@ ExecReScan(Plan *node, ExprContext *exprCtxt, Plan *parent)
 	}
 }
 
-/* ----------------------------------------------------------------
+/*
- *		ExecMarkPos
+ * ExecMarkPos
  *
- *		Marks the current scan position.
+ * Marks the current scan position.
- *
- *		XXX Needs to be extended to include all the node types,
- *		or at least all the ones that can be directly below a mergejoin.
- * ----------------------------------------------------------------
  */
 void
 ExecMarkPos(Plan *node)
@@ -192,6 +188,10 @@ ExecMarkPos(Plan *node)
 			ExecIndexMarkPos((IndexScan *) node);
 			break;
 
+		case T_TidScan:
+			ExecTidMarkPos((TidScan *) node);
+			break;
+
 		case T_FunctionScan:
 			ExecFunctionMarkPos((FunctionScan *) node);
 			break;
@@ -204,10 +204,6 @@ ExecMarkPos(Plan *node)
 			ExecSortMarkPos((Sort *) node);
 			break;
 
-		case T_TidScan:
-			ExecTidMarkPos((TidScan *) node);
-			break;
-
 		default:
 			/* don't make hard error unless caller asks to restore... */
 			elog(LOG, "ExecMarkPos: node type %d not supported",
@@ -216,14 +212,10 @@ ExecMarkPos(Plan *node)
 	}
 }
 
-/* ----------------------------------------------------------------
+/*
- *		ExecRestrPos
+ * ExecRestrPos
  *
- *		restores the scan position previously saved with ExecMarkPos()
+ * restores the scan position previously saved with ExecMarkPos()
- *
- *		XXX Needs to be extended to include all the node types,
- *		or at least all the ones that can be directly below a mergejoin.
- * ----------------------------------------------------------------
  */
 void
 ExecRestrPos(Plan *node)
@@ -238,6 +230,10 @@ ExecRestrPos(Plan *node)
 			ExecIndexRestrPos((IndexScan *) node);
 			break;
 
+		case T_TidScan:
+			ExecTidRestrPos((TidScan *) node);
+			break;
+
 		case T_FunctionScan:
 			ExecFunctionRestrPos((FunctionScan *) node);
 			break;
@@ -256,3 +252,29 @@ ExecRestrPos(Plan *node)
 			break;
 	}
 }
+
+/*
+ * ExecSupportsMarkRestore - does a plan type support mark/restore?
+ *
+ * XXX Ideally, all plan node types would support mark/restore, and this
+ * wouldn't be needed.  For now, this had better match the routines above.
+ */
+bool
+ExecSupportsMarkRestore(NodeTag plantype)
+{
+	switch (plantype)
+	{
+		case T_SeqScan:
+		case T_IndexScan:
+		case T_TidScan:
+		case T_FunctionScan:
+		case T_Material:
+		case T_Sort:
+			return true;
+
+		default:
+			break;
+	}
+
+	return false;
+}

src/backend/executor/nodeSeqscan.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/executor/nodeSeqscan.c,v 1.37 2002/09/04 20:31:18 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/executor/nodeSeqscan.c,v 1.38 2002/11/30 05:21:01 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -19,9 +19,8 @@
  *		ExecInitSeqScan			creates and initializes a seqscan node.
  *		ExecEndSeqScan			releases any storage allocated.
  *		ExecSeqReScan			rescans the relation
- *		ExecMarkPos				marks scan position
+ *		ExecSeqMarkPos			marks scan position
- *		ExecRestrPos			restores scan position
+ *		ExecSeqRestrPos			restores scan position
- *
  */
 #include "postgres.h"
 

src/backend/executor/nodeTidscan.c

@@ -8,19 +8,19 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/executor/nodeTidscan.c,v 1.26 2002/09/04 20:31:18 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/executor/nodeTidscan.c,v 1.27 2002/11/30 05:21:01 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
 /*
  * INTERFACE ROUTINES
  *
- *		ExecTidScan		scans a relation using tids
+ *		ExecTidScan			scans a relation using tids
  *		ExecInitTidScan		creates and initializes state info.
  *		ExecTidReScan		rescans the tid relation.
  *		ExecEndTidScan		releases all storage.
  *		ExecTidMarkPos		marks scan position.
- *
+ *		ExecTidRestrPos		restores scan position.
  */
 #include "postgres.h"
 
@@ -345,7 +345,6 @@ ExecTidMarkPos(TidScan *node)
 	tidstate->tss_MarkTidPtr = tidstate->tss_TidPtr;
 }
 
-#ifdef NOT_USED
 /* ----------------------------------------------------------------
  *		ExecTidRestrPos
  *
@@ -363,7 +362,6 @@ ExecTidRestrPos(TidScan *node)
 	tidstate = node->tidstate;
 	tidstate->tss_TidPtr = tidstate->tss_MarkTidPtr;
 }
-#endif
 
 /* ----------------------------------------------------------------
  *		ExecInitTidScan

src/backend/nodes/copyfuncs.c

@@ -15,7 +15,7 @@
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/nodes/copyfuncs.c,v 1.224 2002/11/30 00:08:16 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/nodes/copyfuncs.c,v 1.225 2002/11/30 05:21:01 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -1142,6 +1142,27 @@ _copyResultPath(ResultPath *from)
 	return newnode;
 }
 
+/*
+ * _copyMaterialPath
+ */
+static MaterialPath *
+_copyMaterialPath(MaterialPath *from)
+{
+	MaterialPath    *newnode = makeNode(MaterialPath);
+
+	/*
+	 * copy node superclass fields
+	 */
+	CopyPathFields((Path *) from, (Path *) newnode);
+
+	/*
+	 * copy remainder of node
+	 */
+	COPY_NODE_FIELD(subpath);
+
+	return newnode;
+}
+
 /*
  * CopyJoinPathFields
  *
@@ -2739,6 +2760,9 @@ copyObject(void *from)
 		case T_RelOptInfo:
 			retval = _copyRelOptInfo(from);
 			break;
+		case T_IndexOptInfo:
+			retval = _copyIndexOptInfo(from);
+			break;
 		case T_Path:
 			retval = _copyPath(from);
 			break;
@@ -2754,6 +2778,9 @@ copyObject(void *from)
 		case T_ResultPath:
 			retval = _copyResultPath(from);
 			break;
+		case T_MaterialPath:
+			retval = _copyMaterialPath(from);
+			break;
 		case T_NestPath:
 			retval = _copyNestPath(from);
 			break;
@@ -2772,9 +2799,6 @@ copyObject(void *from)
 		case T_JoinInfo:
 			retval = _copyJoinInfo(from);
 			break;
-		case T_IndexOptInfo:
-			retval = _copyIndexOptInfo(from);
-			break;
 		case T_InnerIndexscanInfo:
 			retval = _copyInnerIndexscanInfo(from);
 			break;

src/backend/nodes/equalfuncs.c

@@ -20,7 +20,7 @@
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/nodes/equalfuncs.c,v 1.169 2002/11/25 21:29:36 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/nodes/equalfuncs.c,v 1.170 2002/11/30 05:21:01 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -456,6 +456,16 @@ _equalResultPath(ResultPath *a, ResultPath *b)
 	return true;
 }
 
+static bool
+_equalMaterialPath(MaterialPath *a, MaterialPath *b)
+{
+	if (!_equalPath((Path *) a, (Path *) b))
+		return false;
+	COMPARE_NODE_FIELD(subpath);
+
+	return true;
+}
+
 static bool
 _equalJoinPath(JoinPath *a, JoinPath *b)
 {
@@ -1704,12 +1714,27 @@ equal(void *a, void *b)
 		case T_RelOptInfo:
 			retval = _equalRelOptInfo(a, b);
 			break;
+		case T_IndexOptInfo:
+			retval = _equalIndexOptInfo(a, b);
+			break;
 		case T_Path:
 			retval = _equalPath(a, b);
 			break;
 		case T_IndexPath:
 			retval = _equalIndexPath(a, b);
 			break;
+		case T_TidPath:
+			retval = _equalTidPath(a, b);
+			break;
+		case T_AppendPath:
+			retval = _equalAppendPath(a, b);
+			break;
+		case T_ResultPath:
+			retval = _equalResultPath(a, b);
+			break;
+		case T_MaterialPath:
+			retval = _equalMaterialPath(a, b);
+			break;
 		case T_NestPath:
 			retval = _equalNestPath(a, b);
 			break;
@@ -1731,18 +1756,6 @@ equal(void *a, void *b)
 		case T_InnerIndexscanInfo:
 			retval = _equalInnerIndexscanInfo(a, b);
 			break;
-		case T_TidPath:
-			retval = _equalTidPath(a, b);
-			break;
-		case T_AppendPath:
-			retval = _equalAppendPath(a, b);
-			break;
-		case T_ResultPath:
-			retval = _equalResultPath(a, b);
-			break;
-		case T_IndexOptInfo:
-			retval = _equalIndexOptInfo(a, b);
-			break;
 
 		case T_List:
 			{

src/backend/nodes/outfuncs.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/nodes/outfuncs.c,v 1.184 2002/11/30 00:08:16 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/nodes/outfuncs.c,v 1.185 2002/11/30 05:21:02 tgl Exp $
  *
  * NOTES
  *	  Every node type that can appear in stored rules' parsetrees *must*
@@ -1010,6 +1010,16 @@ _outResultPath(StringInfo str, ResultPath *node)
 	WRITE_NODE_FIELD(constantqual);
 }
 
+static void
+_outMaterialPath(StringInfo str, MaterialPath *node)
+{
+	WRITE_NODE_TYPE("MATERIALPATH");
+
+	_outPathInfo(str, (Path *) node);
+
+	WRITE_NODE_FIELD(subpath);
+}
+
 static void
 _outNestPath(StringInfo str, NestPath *node)
 {
@@ -1557,6 +1567,9 @@ _outNode(StringInfo str, void *obj)
 			case T_ResultPath:
 				_outResultPath(str, obj);
 				break;
+			case T_MaterialPath:
+				_outMaterialPath(str, obj);
+				break;
 			case T_NestPath:
 				_outNestPath(str, obj);
 				break;

src/backend/optimizer/README

@@ -259,7 +259,8 @@ RelOptInfo      - a relation or joined relations
   IndexPath     - index scans
   TidPath       - scan by CTID
   AppendPath    - append multiple subpaths together
-  ResultPath    - a Result plan (used for variable-free tlist or qual)
+  ResultPath    - a Result plan node (used for variable-free tlist or qual)
+  MaterialPath  - a Material plan node
   NestPath      - nested-loop joins
   MergePath     - merge joins
   HashPath      - hash joins

src/backend/optimizer/path/allpaths.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/allpaths.c,v 1.92 2002/11/13 00:39:47 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/allpaths.c,v 1.93 2002/11/30 05:21:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -724,33 +724,34 @@ static void
 print_path(Query *root, Path *path, int indent)
 {
 	const char *ptype;
-	bool		join;
+	bool		join = false;
+	Path	   *subpath = NULL;
 	int			i;
 
 	switch (nodeTag(path))
 	{
 		case T_Path:
 			ptype = "SeqScan";
-			join = false;
 			break;
 		case T_IndexPath:
 			ptype = "IdxScan";
-			join = false;
 			break;
 		case T_TidPath:
 			ptype = "TidScan";
-			join = false;
 			break;
 		case T_AppendPath:
 			ptype = "Append";
-			join = false;
 			break;
 		case T_ResultPath:
 			ptype = "Result";
-			join = false;
+			subpath = ((ResultPath *) path)->subpath;
+			break;
+		case T_MaterialPath:
+			ptype = "Material";
+			subpath = ((MaterialPath *) path)->subpath;
 			break;
 		case T_NestPath:
-			ptype = "Nestloop";
+			ptype = "NestLoop";
 			join = true;
 			break;
 		case T_MergePath:
@@ -763,7 +764,6 @@ print_path(Query *root, Path *path, int indent)
 			break;
 		default:
 			ptype = "???Path";
-			join = false;
 			break;
 	}
 
@@ -814,6 +814,9 @@ print_path(Query *root, Path *path, int indent)
 		print_path(root, jp->outerjoinpath, indent + 1);
 		print_path(root, jp->innerjoinpath, indent + 1);
 	}
+
+	if (subpath)
+		print_path(root, subpath, indent + 1);
 }
 
 void

src/backend/optimizer/path/costsize.c

@@ -42,7 +42,7 @@
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/costsize.c,v 1.92 2002/11/30 00:08:16 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/costsize.c,v 1.93 2002/11/30 05:21:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -230,7 +230,7 @@ cost_index(Path *path, Query *root,
 	Assert(length(baserel->relids) == 1);
 	Assert(baserel->rtekind == RTE_RELATION);
 
-	if (!enable_indexscan && !is_injoin)
+	if (!enable_indexscan)
 		startup_cost += disable_cost;
 
 	/*
@@ -513,6 +513,43 @@ cost_sort(Path *path, Query *root,
 	path->total_cost = startup_cost + run_cost;
 }
 
+/*
+ * cost_material
+ *	  Determines and returns the cost of materializing a relation, including
+ *	  the cost of reading the input data.
+ *
+ * If the total volume of data to materialize exceeds SortMem, we will need
+ * to write it to disk, so the cost is much higher in that case.
+ */
+void
+cost_material(Path *path,
+			  Cost input_cost, double tuples, int width)
+{
+	Cost		startup_cost = input_cost;
+	Cost		run_cost = 0;
+	double		nbytes = relation_byte_size(tuples, width);
+	long		sortmembytes = SortMem * 1024L;
+
+	/* disk costs */
+	if (nbytes > sortmembytes)
+	{
+		double		npages = ceil(nbytes / BLCKSZ);
+
+		/* We'll write during startup and read during retrieval */
+		startup_cost += npages;
+		run_cost += npages;
+	}
+
+	/*
+	 * Also charge a small amount per extracted tuple.  We use cpu_tuple_cost
+	 * so that it doesn't appear worthwhile to materialize a bare seqscan.
+	 */
+	run_cost += cpu_tuple_cost * tuples;
+
+	path->startup_cost = startup_cost;
+	path->total_cost = startup_cost + run_cost;
+}
+
 /*
  * cost_agg
  *		Determines and returns the cost of performing an Agg plan node,
@@ -630,19 +667,17 @@ cost_nestloop(Path *path, Query *root,
 	 * before we can start returning tuples, so the join's startup cost is
 	 * their sum.  What's not so clear is whether the inner path's
 	 * startup_cost must be paid again on each rescan of the inner path.
-	 * This is not true if the inner path is materialized, but probably is
+	 * This is not true if the inner path is materialized or is a hashjoin,
-	 * true otherwise.	Since we don't yet have clean handling of the
+	 * but probably is true otherwise.
-	 * decision whether to materialize a path, we can't tell here which
-	 * will happen.  As a compromise, charge 50% of the inner startup cost
-	 * for each restart.
 	 */
 	startup_cost += outer_path->startup_cost + inner_path->startup_cost;
 	run_cost += outer_path->total_cost - outer_path->startup_cost;
 	run_cost += outer_path->parent->rows *
 		(inner_path->total_cost - inner_path->startup_cost);
-	if (outer_path->parent->rows > 1)
+	if (!(IsA(inner_path, MaterialPath) ||
-		run_cost += (outer_path->parent->rows - 1) *
+		  IsA(inner_path, HashPath)) &&
-			inner_path->startup_cost * 0.5;
+		outer_path->parent->rows > 1)
+		run_cost += (outer_path->parent->rows - 1) * inner_path->startup_cost;
 
 	/*
 	 * Number of tuples processed (not number emitted!).  If inner path is
@@ -1544,7 +1579,7 @@ set_rel_width(Query *root, RelOptInfo *rel)
 static double
 relation_byte_size(double tuples, int width)
 {
-	return tuples * ((double) MAXALIGN(width + sizeof(HeapTupleData)));
+	return tuples * (MAXALIGN(width) + MAXALIGN(sizeof(HeapTupleData)));
 }
 
 /*

src/backend/optimizer/path/joinpath.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/joinpath.c,v 1.73 2002/11/30 00:08:16 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/joinpath.c,v 1.74 2002/11/30 05:21:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -281,8 +281,9 @@ sort_inner_and_outer(Query *root,
  *	  only outer paths that are already ordered well enough for merging).
  *
  * We always generate a nestloop path for each available outer path.
- * In fact we may generate as many as three: one on the cheapest-total-cost
+ * In fact we may generate as many as four: one on the cheapest-total-cost
- * inner path, one on the cheapest-startup-cost inner path (if different),
+ * inner path, one on the same with materialization, one on the
+ * cheapest-startup-cost inner path (if different), 
  * and one on the best inner-indexscan path (if any).
  *
  * We also consider mergejoins if mergejoin clauses are available.	We have
@@ -315,7 +316,8 @@ match_unsorted_outer(Query *root,
 {
 	bool		nestjoinOK;
 	bool		useallclauses;
-	Path	   *bestinnerjoin;
+	Path	   *matpath = NULL;
+	Path	   *bestinnerjoin = NULL;
 	List	   *i;
 
 	/*
@@ -345,12 +347,26 @@ match_unsorted_outer(Query *root,
 			break;
 	}
 
-	/*
+	if (nestjoinOK)
-	 * Get the best innerjoin indexpath (if any) for this outer rel. It's
+	{
-	 * the same for all outer paths.
+		/*
-	 */
+		 * If the cheapest inner path is a join or seqscan, we should consider
-	bestinnerjoin = best_inner_indexscan(root, innerrel,
+		 * materializing it.  (This is a heuristic: we could consider it
-										 outerrel->relids, jointype);
+		 * always, but for inner indexscans it's probably a waste of time.)
+		 */
+		if (!(IsA(innerrel->cheapest_total_path, IndexPath) ||
+			  IsA(innerrel->cheapest_total_path, TidPath)))
+			matpath = (Path *)
+				create_material_path(innerrel, 
+									 innerrel->cheapest_total_path);
+
+		/*
+		 * Get the best innerjoin indexpath (if any) for this outer rel. It's
+		 * the same for all outer paths.
+		 */
+		bestinnerjoin = best_inner_indexscan(root, innerrel,
+											 outerrel->relids, jointype);
+	}
 
 	foreach(i, outerrel->pathlist)
 	{
@@ -376,8 +392,9 @@ match_unsorted_outer(Query *root,
 		{
 			/*
 			 * Always consider a nestloop join with this outer and
-			 * cheapest-total-cost inner.	Consider nestloops using the
+			 * cheapest-total-cost inner.  When appropriate, also consider
-			 * cheapest-startup-cost inner as well, and the best innerjoin
+			 * using the materialized form of the cheapest inner, the
+			 * cheapest-startup-cost inner path, and the best innerjoin
 			 * indexpath.
 			 */
 			add_path(joinrel, (Path *)
@@ -388,6 +405,15 @@ match_unsorted_outer(Query *root,
 										  innerrel->cheapest_total_path,
 										  restrictlist,
 										  merge_pathkeys));
+			if (matpath != NULL)
+				add_path(joinrel, (Path *)
+						 create_nestloop_path(root,
+											  joinrel,
+											  jointype,
+											  outerpath,
+											  matpath,
+											  restrictlist,
+											  merge_pathkeys));
 			if (innerrel->cheapest_startup_path !=
 				innerrel->cheapest_total_path)
 				add_path(joinrel, (Path *)

src/backend/optimizer/plan/createplan.c

@@ -10,7 +10,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/createplan.c,v 1.125 2002/11/30 00:08:17 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/createplan.c,v 1.126 2002/11/30 05:21:02 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -35,6 +35,7 @@ static Scan *create_scan_plan(Query *root, Path *best_path);
 static Join *create_join_plan(Query *root, JoinPath *best_path);
 static Append *create_append_plan(Query *root, AppendPath *best_path);
 static Result *create_result_plan(Query *root, ResultPath *best_path);
+static Material *create_material_plan(Query *root, MaterialPath *best_path);
 static SeqScan *create_seqscan_plan(Path *best_path, List *tlist,
 					List *scan_clauses);
 static IndexScan *create_indexscan_plan(Query *root, IndexPath *best_path,
@@ -141,6 +142,10 @@ create_plan(Query *root, Path *best_path)
 			plan = (Plan *) create_result_plan(root,
 											   (ResultPath *) best_path);
 			break;
+		case T_Material:
+			plan = (Plan *) create_material_plan(root,
+												 (MaterialPath *) best_path);
+			break;
 		default:
 			elog(ERROR, "create_plan: unknown pathtype %d",
 				 best_path->pathtype);
@@ -383,6 +388,28 @@ create_result_plan(Query *root, ResultPath *best_path)
 	return plan;
 }
 
+/*
+ * create_material_plan
+ *	  Create a Material plan for 'best_path' and (recursively) plans
+ *	  for its subpaths.
+ *
+ *	  Returns a Plan node.
+ */
+static Material *
+create_material_plan(Query *root, MaterialPath *best_path)
+{
+	Material   *plan;
+	Plan	   *subplan;
+
+	subplan = create_plan(root, best_path->subpath);
+
+	plan = make_material(best_path->path.parent->targetlist, subplan);
+
+	copy_path_costsize(&plan->plan, (Path *) best_path);
+
+	return plan;
+}
+
 
 /*****************************************************************************
  *
@@ -739,18 +766,6 @@ create_nestloop_plan(Query *root,
 											 inner_tlist,
 											 innerscan->scan.scanrelid);
 	}
-	else if (IsA_Join(inner_plan))
-	{
-		/*
-		 * Materialize the inner join for speed reasons.
-		 *
-		 * XXX It is probably *not* always fastest to materialize an inner
-		 * join --- how can we estimate whether this is a good thing to
-		 * do?
-		 */
-		inner_plan = (Plan *) make_material(inner_tlist,
-											inner_plan);
-	}
 
 	/*
 	 * Set quals to contain INNER/OUTER var references.
@@ -842,44 +857,6 @@ create_mergejoin_plan(Query *root,
 									inner_plan,
 									best_path->innersortkeys);
 
-	/*
-	 * The executor requires the inner side of a mergejoin to support
-	 * "mark" and "restore" operations.  Not all plan types do, so we must
-	 * be careful not to generate an invalid plan.	If necessary, an
-	 * invalid inner plan can be handled by inserting a Materialize node.
-	 *
-	 * Since the inner side must be ordered, and only Sorts and IndexScans
-	 * can create order to begin with, you might think there's no problem
-	 * --- but you'd be wrong.  Nestloop and merge joins can *preserve*
-	 * the order of their inputs, so they can be selected as the input of
-	 * a mergejoin, and that won't work in the present executor.
-	 *
-	 * Doing this here is a bit of a kluge since the cost of the Materialize
-	 * wasn't taken into account in our earlier decisions.  But
-	 * Materialize is hard to estimate a cost for, and the above
-	 * consideration shows that this is a rare case anyway, so this seems
-	 * an acceptable way to proceed.
-	 *
-	 * This check must agree with ExecMarkPos/ExecRestrPos in
-	 * executor/execAmi.c!
-	 */
-	switch (nodeTag(inner_plan))
-	{
-		case T_SeqScan:
-		case T_IndexScan:
-		case T_FunctionScan:
-		case T_Material:
-		case T_Sort:
-			/* OK, these inner plans support mark/restore */
-			break;
-
-		default:
-			/* Ooops, need to materialize the inner plan */
-			inner_plan = (Plan *) make_material(inner_tlist,
-												inner_plan);
-			break;
-	}
-
 	/*
 	 * Now we can build the mergejoin node.
 	 */
@@ -1668,15 +1645,7 @@ make_material(List *tlist, Plan *lefttree)
 	Material   *node = makeNode(Material);
 	Plan	   *plan = &node->plan;
 
-	copy_plan_costsize(plan, lefttree);
+	/* cost should be inserted by caller */
-
-	/*
-	 * For plausibility, make startup & total costs equal total cost of
-	 * input plan; this only affects EXPLAIN display not decisions.
-	 *
-	 * XXX shouldn't we charge some additional cost for materialization?
-	 */
-	plan->startup_cost = plan->total_cost;
 	plan->state = (EState *) NULL;
 	plan->targetlist = tlist;
 	plan->qual = NIL;

src/backend/optimizer/plan/subselect.c

@@ -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 1.57 2002/11/30 00:08:18 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/subselect.c,v 1.58 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -328,9 +328,17 @@ make_subplan(SubLink *slink)
 			if (use_material)
 			{
 				Plan	   *matplan;
+				Path		matpath; /* dummy for result of cost_material */
 
 				matplan = (Plan *) make_material(plan->targetlist, plan);
-				/* kluge --- see comments above */
+				/* need to calculate costs */
+				cost_material(&matpath,
+							  plan->total_cost,
+							  plan->plan_rows,
+							  plan->plan_width);
+				matplan->startup_cost = matpath.startup_cost;
+				matplan->total_cost = matpath.total_cost;
+				/* parameter kluge --- see comments above */
 				matplan->extParam = listCopy(plan->extParam);
 				matplan->locParam = listCopy(plan->locParam);
 				node->plan = plan = matplan;

src/backend/optimizer/util/pathnode.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/util/pathnode.c,v 1.81 2002/11/30 00:08:20 tgl Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/util/pathnode.c,v 1.82 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -16,6 +16,7 @@
 
 #include <math.h>
 
+#include "executor/executor.h"
 #include "nodes/plannodes.h"
 #include "optimizer/cost.h"
 #include "optimizer/pathnode.h"
@@ -450,6 +451,7 @@ create_result_path(RelOptInfo *rel, Path *subpath, List *constantqual)
 	pathnode->subpath = subpath;
 	pathnode->constantqual = constantqual;
 
+	/* Ideally should define cost_result(), but I'm too lazy */
 	if (subpath)
 	{
 		pathnode->path.startup_cost = subpath->startup_cost;
@@ -464,6 +466,31 @@ create_result_path(RelOptInfo *rel, Path *subpath, List *constantqual)
 	return pathnode;
 }
 
+/*
+ * create_material_path
+ *	  Creates a path corresponding to a Material plan, returning the
+ *	  pathnode.
+ */
+MaterialPath *
+create_material_path(RelOptInfo *rel, Path *subpath)
+{
+	MaterialPath *pathnode = makeNode(MaterialPath);
+
+	pathnode->path.pathtype = T_Material;
+	pathnode->path.parent = rel;
+
+	pathnode->path.pathkeys = subpath->pathkeys;
+
+	pathnode->subpath = subpath;
+
+	cost_material(&pathnode->path,
+				  subpath->total_cost,
+				  rel->rows,
+				  rel->width);
+
+	return pathnode;
+}
+
 /*
  * create_subqueryscan_path
  *	  Creates a path corresponding to a sequential scan of a subquery,
@@ -583,6 +610,21 @@ create_mergejoin_path(Query *root,
 	if (innersortkeys &&
 		pathkeys_contained_in(innersortkeys, inner_path->pathkeys))
 		innersortkeys = NIL;
+	/*
+	 * If we are not sorting the inner path, we may need a materialize
+	 * node to ensure it can be marked/restored.  (Sort does support
+	 * mark/restore, so no materialize is needed in that case.)
+	 *
+	 * Since the inner side must be ordered, and only Sorts and IndexScans
+	 * can create order to begin with, you might think there's no problem
+	 * --- but you'd be wrong.  Nestloop and merge joins can *preserve*
+	 * the order of their inputs, so they can be selected as the input of
+	 * a mergejoin, and they don't support mark/restore at present.
+	 */
+	if (innersortkeys == NIL &&
+		!ExecSupportsMarkRestore(inner_path->pathtype))
+		inner_path = (Path *)
+			create_material_path(inner_path->parent, inner_path);
 
 	pathnode->jpath.path.pathtype = T_MergeJoin;
 	pathnode->jpath.path.parent = joinrel;

src/include/executor/executor.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: executor.h,v 1.78 2002/09/04 20:31:42 momjian Exp $
+ * $Id: executor.h,v 1.79 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -33,6 +33,7 @@
 extern void ExecReScan(Plan *node, ExprContext *exprCtxt, Plan *parent);
 extern void ExecMarkPos(Plan *node);
 extern void ExecRestrPos(Plan *node);
+extern bool ExecSupportsMarkRestore(NodeTag plantype);
 
 /*
  * prototypes from functions in execJunk.c

src/include/executor/nodeTidscan.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: nodeTidscan.h,v 1.9 2002/06/20 20:29:49 momjian Exp $
+ * $Id: nodeTidscan.h,v 1.10 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -20,6 +20,7 @@ extern TupleTableSlot *ExecTidScan(TidScan *node);
 extern void ExecTidReScan(TidScan *node, ExprContext *exprCtxt, Plan *parent);
 extern void ExecEndTidScan(TidScan *node);
 extern void ExecTidMarkPos(TidScan *node);
+extern void ExecTidRestrPos(TidScan *node);
 extern bool ExecInitTidScan(TidScan *node, EState *estate, Plan *parent);
 extern int	ExecCountSlotsTidScan(TidScan *node);
 extern void ExecTidReScan(TidScan *node, ExprContext *exprCtxt, Plan *parent);

src/include/nodes/nodes.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: nodes.h,v 1.124 2002/11/24 21:52:14 tgl Exp $
+ * $Id: nodes.h,v 1.125 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -75,6 +75,7 @@ typedef enum NodeTag
 	 * TAGS FOR PLANNER NODES (relation.h)
 	 */
 	T_RelOptInfo = 200,
+	T_IndexOptInfo,
 	T_Path,
 	T_IndexPath,
 	T_NestPath,
@@ -83,10 +84,10 @@ typedef enum NodeTag
 	T_TidPath,
 	T_AppendPath,
 	T_ResultPath,
+	T_MaterialPath,
 	T_PathKeyItem,
 	T_RestrictInfo,
 	T_JoinInfo,
-	T_IndexOptInfo,
 	T_InnerIndexscanInfo,
 
 	/*
@@ -288,17 +289,6 @@ extern Node *newNodeMacroHolder;
 
 #define IsA(nodeptr,_type_)		(nodeTag(nodeptr) == T_##_type_)
 
-/* ----------------------------------------------------------------
- *					  IsA functions (no inheritance any more)
- * ----------------------------------------------------------------
- */
-#define IsA_JoinPath(jp) \
-	(IsA(jp, NestPath) || IsA(jp, MergePath) || IsA(jp, HashPath))
-
-#define IsA_Join(jp) \
-	(IsA(jp, Join) || IsA(jp, NestLoop) || \
-	 IsA(jp, MergeJoin) || IsA(jp, HashJoin))
-
 /* ----------------------------------------------------------------
  *					  extern declarations follow
  * ----------------------------------------------------------------

src/include/nodes/relation.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: relation.h,v 1.71 2002/11/30 00:08:22 tgl Exp $
+ * $Id: relation.h,v 1.72 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -405,6 +405,18 @@ typedef struct ResultPath
 	List	   *constantqual;
 } ResultPath;
 
+/*
+ * MaterialPath represents use of a Material plan node, i.e., caching of
+ * the output of its subpath.  This is used when the subpath is expensive
+ * and needs to be scanned repeatedly, or when we need mark/restore ability
+ * and the subpath doesn't have it.
+ */
+typedef struct MaterialPath
+{
+	Path		path;
+	Path	   *subpath;
+} MaterialPath;
+
 /*
  * All join-type paths share these fields.
  */

src/include/optimizer/cost.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: cost.h,v 1.48 2002/11/21 00:42:19 tgl Exp $
+ * $Id: cost.h,v 1.49 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -60,6 +60,8 @@ extern void cost_functionscan(Path *path, Query *root,
 				  RelOptInfo *baserel);
 extern void cost_sort(Path *path, Query *root,
 		  List *pathkeys, Cost input_cost, double tuples, int width);
+extern void cost_material(Path *path,
+						  Cost input_cost, double tuples, int width);
 extern void cost_agg(Path *path, Query *root,
 					 AggStrategy aggstrategy, int numAggs,
 					 int numGroupCols, double numGroups,

src/include/optimizer/pathnode.h

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: pathnode.h,v 1.45 2002/11/06 00:00:45 tgl Exp $
+ * $Id: pathnode.h,v 1.46 2002/11/30 05:21:03 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -37,6 +37,7 @@ extern TidPath *create_tidscan_path(Query *root, RelOptInfo *rel,
 extern AppendPath *create_append_path(RelOptInfo *rel, List *subpaths);
 extern ResultPath *create_result_path(RelOptInfo *rel, Path *subpath,
 									  List *constantqual);
+extern MaterialPath *create_material_path(RelOptInfo *rel, Path *subpath);
 extern Path *create_subqueryscan_path(RelOptInfo *rel);
 extern Path *create_functionscan_path(Query *root, RelOptInfo *rel);