Teach planner about the idea that a mergejoin won't necessarily read

both input streams to the end.  If one variable's range is much less
than the other, an indexscan-based merge can win by not scanning all
of the other table.  Per example from Reinhard Max.

src/backend/executor/nodeMergejoin.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/executor/nodeMergejoin.c,v 1.47 2001/10/28 06:25:43 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/executor/nodeMergejoin.c,v 1.48 2002/03/01 04:09:22 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -88,97 +88,62 @@ static bool MergeCompare(List *eqQual, List *compareQual, ExprContext *econtext)
 
 
 /* ----------------------------------------------------------------
- *		MJFormSkipQual
+ *		MJFormSkipQuals
  *
  *		This takes the mergeclause which is a qualification of the
- *		form ((= expr expr) (= expr expr) ...) and forms a new
- *		qualification like ((> expr expr) (> expr expr) ...) which
- *		is used by ExecMergeJoin() in order to determine if we should
- *		skip tuples.  The replacement operators are named either ">"
- *		or "<" according to the replaceopname parameter, and have the
- *		same operand data types as the "=" operators they replace.
- *		(We expect there to be such operators because the "=" operators
+ *		form ((= expr expr) (= expr expr) ...) and forms new lists
+ *		of the forms ((< expr expr) (< expr expr) ...) and
+ *		((> expr expr) (> expr expr) ...).  These lists will be used
+ *		by ExecMergeJoin() to determine if we should skip tuples.
+ *		(We expect there to be suitable operators because the "=" operators
  *		were marked mergejoinable; however, there might be a different
  *		one needed in each qual clause.)
  * ----------------------------------------------------------------
  */
-static List *
-MJFormSkipQual(List *qualList, char *replaceopname)
+static void
+MJFormSkipQuals(List *qualList, List **ltQuals, List **gtQuals)
 {
-	List	   *qualCopy;
-	List	   *qualcdr;
-	Expr	   *qual;
-	Oper	   *op;
-	HeapTuple	optup;
-	Form_pg_operator opform;
-	Oid			oprleft,
-				oprright;
+	List	   *ltcdr,
+			   *gtcdr;
 
 	/*
-	 * qualList is a list: ((op .. ..) ...)
-	 *
-	 * first we make a copy of it.	copyObject() makes a deep copy so let's
-	 * use it instead of the old fashoned lispCopy()...
+	 * Make modifiable copies of the qualList.
 	 */
-	qualCopy = (List *) copyObject((Node *) qualList);
+	*ltQuals = (List *) copyObject((Node *) qualList);
+	*gtQuals = (List *) copyObject((Node *) qualList);
 
-	foreach(qualcdr, qualCopy)
+	/*
+	 * Scan both lists in parallel, so that we can update the operators
+	 * with the minimum number of syscache searches.
+	 */
+	ltcdr = *ltQuals;
+	foreach(gtcdr, *gtQuals)
 	{
-		/*
-		 * first get the current (op .. ..) list
-		 */
-		qual = lfirst(qualcdr);
+		Expr	   *ltqual = (Expr *) lfirst(ltcdr);
+		Expr	   *gtqual = (Expr *) lfirst(gtcdr);
+		Oper	   *ltop = (Oper *) ltqual->oper;
+		Oper	   *gtop = (Oper *) gtqual->oper;
 
 		/*
-		 * now get at the op
+		 * The two ops should be identical, so use either one for lookup.
 		 */
-		op = (Oper *) qual->oper;
-		if (!IsA(op, Oper))
-			elog(ERROR, "MJFormSkipQual: op not an Oper!");
+		if (!IsA(ltop, Oper))
+			elog(ERROR, "MJFormSkipQuals: op not an Oper!");
 
 		/*
-		 * Get the declared left and right operand types of the operator.
-		 * Note we do *not* use the actual operand types, since those
-		 * might be different in scenarios with binary-compatible data
-		 * types. There should be "<" and ">" operators matching a
-		 * mergejoinable "=" operator's declared operand types, but we
-		 * might not find them if we search with the actual operand types.
+		 * Lookup the operators, and replace the data in the copied
+		 * operator nodes.
 		 */
-		optup = SearchSysCache(OPEROID,
-							   ObjectIdGetDatum(op->opno),
-							   0, 0, 0);
-		if (!HeapTupleIsValid(optup))	/* shouldn't happen */
-			elog(ERROR, "MJFormSkipQual: operator %u not found", op->opno);
-		opform = (Form_pg_operator) GETSTRUCT(optup);
-		oprleft = opform->oprleft;
-		oprright = opform->oprright;
-		ReleaseSysCache(optup);
-
-		/*
-		 * Now look up the matching "<" or ">" operator.  If there isn't
-		 * one, whoever marked the "=" operator mergejoinable was a loser.
-		 */
-		optup = SearchSysCache(OPERNAME,
-							   PointerGetDatum(replaceopname),
-							   ObjectIdGetDatum(oprleft),
-							   ObjectIdGetDatum(oprright),
-							   CharGetDatum('b'));
-		if (!HeapTupleIsValid(optup))
-			elog(ERROR,
-			"MJFormSkipQual: mergejoin operator %u has no matching %s op",
-				 op->opno, replaceopname);
-		opform = (Form_pg_operator) GETSTRUCT(optup);
-
-		/*
-		 * And replace the data in the copied operator node.
-		 */
-		op->opno = optup->t_data->t_oid;
-		op->opid = opform->oprcode;
-		op->op_fcache = NULL;
-		ReleaseSysCache(optup);
+		op_mergejoin_crossops(ltop->opno,
+							  &ltop->opno,
+							  &gtop->opno,
+							  &ltop->opid,
+							  &gtop->opid);
+		ltop->op_fcache = NULL;
+		gtop->op_fcache = NULL;
+
+		ltcdr = lnext(ltcdr);
 	}
-
-	return qualCopy;
 }
 
 /* ----------------------------------------------------------------
@@ -1430,7 +1395,6 @@ bool
 ExecInitMergeJoin(MergeJoin *node, EState *estate, Plan *parent)
 {
 	MergeJoinState *mergestate;
-	List	   *joinclauses;
 
 	MJ1_printf("ExecInitMergeJoin: %s\n",
 			   "initializing node");
@@ -1522,9 +1486,9 @@ ExecInitMergeJoin(MergeJoin *node, EState *estate, Plan *parent)
 	/*
 	 * form merge skip qualifications
 	 */
-	joinclauses = node->mergeclauses;
-	mergestate->mj_OuterSkipQual = MJFormSkipQual(joinclauses, "<");
-	mergestate->mj_InnerSkipQual = MJFormSkipQual(joinclauses, ">");
+	MJFormSkipQuals(node->mergeclauses,
+					&mergestate->mj_OuterSkipQual,
+					&mergestate->mj_InnerSkipQual);
 
 	MJ_printf("\nExecInitMergeJoin: OuterSkipQual is ");
 	MJ_nodeDisplay(mergestate->mj_OuterSkipQual);

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 1.79 2001/10/25 05:49:32 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/costsize.c,v 1.80 2002/03/01 04:09:24 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -58,6 +58,7 @@
 #include "optimizer/cost.h"
 #include "optimizer/pathnode.h"
 #include "parser/parsetree.h"
+#include "utils/selfuncs.h"
 #include "utils/lsyscache.h"
 #include "utils/syscache.h"
 
@@ -565,12 +566,29 @@ cost_mergejoin(Path *path, Query *root,
 	Cost		startup_cost = 0;
 	Cost		run_cost = 0;
 	Cost		cpu_per_tuple;
+	double		outer_rows,
+				inner_rows;
 	double		ntuples;
+	Selectivity	leftscan,
+				rightscan;
 	Path		sort_path;		/* dummy for result of cost_sort */
 
 	if (!enable_mergejoin)
 		startup_cost += disable_cost;
 
+	/*
+	 * A merge join will stop as soon as it exhausts either input stream.
+	 * Estimate fraction of the left and right inputs that will actually
+	 * need to be scanned.  We use only the first (most significant)
+	 * merge clause for this purpose.
+	 */
+	mergejoinscansel(root,
+					 (Node *) ((RestrictInfo *) lfirst(mergeclauses))->clause,
+					 &leftscan, &rightscan);
+
+	outer_rows = outer_path->parent->rows * leftscan;
+	inner_rows = inner_path->parent->rows * rightscan;
+
 	/* cost of source data */
 
 	/*
@@ -588,12 +606,14 @@ cost_mergejoin(Path *path, Query *root,
 				  outer_path->parent->rows,
 				  outer_path->parent->width);
 		startup_cost += sort_path.startup_cost;
-		run_cost += sort_path.total_cost - sort_path.startup_cost;
+		run_cost += (sort_path.total_cost - sort_path.startup_cost)
+			* leftscan;
 	}
 	else
 	{
 		startup_cost += outer_path->startup_cost;
-		run_cost += outer_path->total_cost - outer_path->startup_cost;
+		run_cost += (outer_path->total_cost - outer_path->startup_cost)
+			* leftscan;
 	}
 
 	if (innersortkeys)			/* do we need to sort inner? */
@@ -605,30 +625,33 @@ cost_mergejoin(Path *path, Query *root,
 				  inner_path->parent->rows,
 				  inner_path->parent->width);
 		startup_cost += sort_path.startup_cost;
-		run_cost += sort_path.total_cost - sort_path.startup_cost;
+		run_cost += (sort_path.total_cost - sort_path.startup_cost)
+			* rightscan;
 	}
 	else
 	{
 		startup_cost += inner_path->startup_cost;
-		run_cost += inner_path->total_cost - inner_path->startup_cost;
+		run_cost += (inner_path->total_cost - inner_path->startup_cost)
+			* rightscan;
 	}
 
 	/*
 	 * The number of tuple comparisons needed depends drastically on the
 	 * number of equal keys in the two source relations, which we have no
-	 * good way of estimating.	Somewhat arbitrarily, we charge one tuple
+	 * good way of estimating.  (XXX could the MCV statistics help?)
+	 * Somewhat arbitrarily, we charge one tuple
 	 * comparison (one cpu_operator_cost) for each tuple in the two source
 	 * relations.  This is probably a lower bound.
 	 */
-	run_cost += cpu_operator_cost *
-		(outer_path->parent->rows + inner_path->parent->rows);
+	run_cost += cpu_operator_cost * (outer_rows + inner_rows);
 
 	/*
 	 * For each tuple that gets through the mergejoin proper, we charge
 	 * cpu_tuple_cost plus the cost of evaluating additional restriction
 	 * clauses that are to be applied at the join.	It's OK to use an
 	 * approximate selectivity here, since in most cases this is a minor
-	 * component of the cost.
+	 * component of the cost.  NOTE: it's correct to use the unscaled rows
+	 * counts here, not the scaled-down counts we obtained above.
 	 */
 	ntuples = approx_selectivity(root, mergeclauses) *
 		outer_path->parent->rows * inner_path->parent->rows;

src/backend/utils/cache/lsyscache.c

@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/utils/cache/lsyscache.c,v 1.59 2001/10/25 05:49:46 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/utils/cache/lsyscache.c,v 1.60 2002/03/01 04:09:26 tgl Exp $
  *
  * NOTES
  *	  Eventually, the index information should go through here, too.
@@ -369,6 +369,76 @@ op_mergejoinable(Oid opno, Oid ltype, Oid rtype, Oid *leftOp, Oid *rightOp)
 	return result;
 }
 
+/*
+ * op_mergejoin_crossops
+ *
+ *		Returns the cross-type comparison operators (ltype "<" rtype and
+ *		ltype ">" rtype) for an operator previously determined to be
+ *		mergejoinable.  Optionally, fetches the regproc ids of these
+ *		operators, as well as their operator OIDs.
+ *
+ * Raises error if operators cannot be found.  Assuming that the operator
+ * had indeed been marked mergejoinable, this indicates that whoever marked
+ * it so was mistaken.
+ */
+void
+op_mergejoin_crossops(Oid opno, Oid *ltop, Oid *gtop,
+					  RegProcedure *ltproc, RegProcedure *gtproc)
+{
+	HeapTuple	tp;
+	Form_pg_operator optup;
+	Oid			oprleft,
+				oprright;
+
+	/*
+	 * Get the declared left and right operand types of the operator.
+	 */
+	tp = SearchSysCache(OPEROID,
+						ObjectIdGetDatum(opno),
+						0, 0, 0);
+	if (!HeapTupleIsValid(tp))	/* shouldn't happen */
+		elog(ERROR, "op_mergejoin_crossops: operator %u not found", opno);
+	optup = (Form_pg_operator) GETSTRUCT(tp);
+	oprleft = optup->oprleft;
+	oprright = optup->oprright;
+	ReleaseSysCache(tp);
+
+	/*
+	 * Look up the "<" operator with the same input types.  If there isn't
+	 * one, whoever marked the "=" operator mergejoinable was a loser.
+	 */
+	tp = SearchSysCache(OPERNAME,
+						PointerGetDatum("<"),
+						ObjectIdGetDatum(oprleft),
+						ObjectIdGetDatum(oprright),
+						CharGetDatum('b'));
+	if (!HeapTupleIsValid(tp))
+		elog(ERROR, "op_mergejoin_crossops: mergejoin operator %u has no matching < operator",
+			 opno);
+	optup = (Form_pg_operator) GETSTRUCT(tp);
+	*ltop = tp->t_data->t_oid;
+	if (ltproc)
+		*ltproc = optup->oprcode;
+	ReleaseSysCache(tp);
+
+	/*
+	 * And the same for the ">" operator.
+	 */
+	tp = SearchSysCache(OPERNAME,
+						PointerGetDatum(">"),
+						ObjectIdGetDatum(oprleft),
+						ObjectIdGetDatum(oprright),
+						CharGetDatum('b'));
+	if (!HeapTupleIsValid(tp))
+		elog(ERROR, "op_mergejoin_crossops: mergejoin operator %u has no matching > operator",
+			 opno);
+	optup = (Form_pg_operator) GETSTRUCT(tp);
+	*gtop = tp->t_data->t_oid;
+	if (gtproc)
+		*gtproc = optup->oprcode;
+	ReleaseSysCache(tp);
+}
+
 /*
  * op_hashjoinable
  *

src/include/utils/lsyscache.h

@@ -6,7 +6,7 @@
  * Portions Copyright (c) 1996-2001, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: lsyscache.h,v 1.39 2001/11/05 17:46:36 momjian Exp $
+ * $Id: lsyscache.h,v 1.40 2002/03/01 04:09:28 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -29,6 +29,8 @@ extern RegProcedure get_opcode(Oid opno);
 extern char *get_opname(Oid opno);
 extern bool op_mergejoinable(Oid opno, Oid ltype, Oid rtype,
 				 Oid *leftOp, Oid *rightOp);
+extern void op_mergejoin_crossops(Oid opno, Oid *ltop, Oid *gtop,
+				 RegProcedure *ltproc, RegProcedure *gtproc);
 extern Oid	op_hashjoinable(Oid opno, Oid ltype, Oid rtype);
 extern bool op_iscachable(Oid opno);
 extern Oid	get_commutator(Oid opno);

src/include/utils/selfuncs.h

@@ -8,7 +8,7 @@
  * Portions Copyright (c) 1996-2001, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: selfuncs.h,v 1.4 2001/11/05 17:46:36 momjian Exp $
+ * $Id: selfuncs.h,v 1.5 2002/03/01 04:09:28 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -66,8 +66,12 @@ extern Datum icregexnejoinsel(PG_FUNCTION_ARGS);
 extern Datum nlikejoinsel(PG_FUNCTION_ARGS);
 extern Datum icnlikejoinsel(PG_FUNCTION_ARGS);
 
-Selectivity booltestsel(Query *root, BooleanTest *clause, int varRelid);
-Selectivity nulltestsel(Query *root, NullTest *clause, int varRelid);
+extern Selectivity booltestsel(Query *root, BooleanTest *clause, int varRelid);
+extern Selectivity nulltestsel(Query *root, NullTest *clause, int varRelid);
+
+extern void mergejoinscansel(Query *root, Node *clause,
+							 Selectivity *leftscan,
+							 Selectivity *rightscan);
 
 extern Datum btcostestimate(PG_FUNCTION_ARGS);
 extern Datum rtcostestimate(PG_FUNCTION_ARGS);