Back-patch code to deduce implied equalities from transitivity of

tglsfdc · tglsfdc · commit 730e2ffebc98 · 2000-09-23T23:50:47.000Z
mergejoin clauses, and add these equalities to the given WHERE clauses.
This is necessary to ensure that sort keys we think are equivalent
really are equivalent as soon as their rels have been joined.  Without
this, 7.0 may create an incorrect mergejoin plan.
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/pathkeys.c,v 1.21 2000/04/12 17:15:20 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/path/pathkeys.c,v 1.21.2.1 2000/09/23 23:50:47 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -19,6 +19,7 @@
 #include "optimizer/joininfo.h"
 #include "optimizer/pathnode.h"
 #include "optimizer/paths.h"
+#include "optimizer/planmain.h"
 #include "optimizer/tlist.h"
 #include "optimizer/var.h"
 #include "parser/parsetree.h"
@@ -227,35 +228,107 @@ add_equijoined_keys(Query *root, RestrictInfo *restrictinfo)
 	 * into our new set. When done, we add the new set to the front of
 	 * equi_key_list.
 	 *
+	 * It may well be that the two items we're given are already known to
+	 * be equijoin-equivalent, in which case we don't need to change our
+	 * data structure.  If we find both of them in the same equivalence
+	 * set to start with, we can quit immediately.
+	 *
 	 * This is a standard UNION-FIND problem, for which there exist better
 	 * data structures than simple lists.  If this code ever proves to be
 	 * a bottleneck then it could be sped up --- but for now, simple is
 	 * beautiful.
 	 */
-	newset = lcons(item1, lcons(item2, NIL));
+	newset = NIL;
 
 	foreach(cursetlink, root->equi_key_list)
 	{
 		List	   *curset = lfirst(cursetlink);
+		bool		item1here = member(item1, curset);
+		bool		item2here = member(item2, curset);
 
-		if (member(item1, curset) || member(item2, curset))
+		if (item1here || item2here)
 		{
+			/* If find both in same equivalence set, no need to do any more */
+			if (item1here && item2here)
+			{
+				/* Better not have seen only one in an earlier set... */
+				Assert(newset == NIL);
+				return;
+			}
+
+			/* Build the new set only when we know we must */
+			if (newset == NIL)
+				newset = lcons(item1, lcons(item2, NIL));
+
 			/* Found a set to merge into our new set */
 			newset = LispUnion(newset, curset);
 
 			/*
 			 * Remove old set from equi_key_list.  NOTE this does not
-			 * change lnext(cursetlink), so the outer foreach doesn't
-			 * break.
+			 * change lnext(cursetlink), so the foreach loop doesn't break.
 			 */
 			root->equi_key_list = lremove(curset, root->equi_key_list);
 			freeList(curset);	/* might as well recycle old cons cells */
 		}
 	}
 
+	/* Build the new set only when we know we must */
+	if (newset == NIL)
+		newset = lcons(item1, lcons(item2, NIL));
+
 	root->equi_key_list = lcons(newset, root->equi_key_list);
 }
 
+/*
+ * generate_implied_equalities
+ *	  Scan the completed equi_key_list for the query, and generate explicit
+ *	  qualifications (WHERE clauses) for all the pairwise equalities not
+ *	  already mentioned in the quals.  This is useful because the additional
+ *	  clauses help the selectivity-estimation code, and in fact it's
+ *	  *necessary* to ensure that sort keys we think are equivalent really
+ *	  are (see src/backend/optimizer/README for more info).
+ *
+ * This routine just walks the equi_key_list to find all pairwise equalities.
+ * We call process_implied_equality (in plan/initsplan.c) to determine whether
+ * each is already known and add it to the proper restrictinfo list if not.
+ */
+void
+generate_implied_equalities(Query *root)
+{
+	List	   *cursetlink;
+
+	foreach(cursetlink, root->equi_key_list)
+	{
+		List	   *curset = lfirst(cursetlink);
+		List	   *ptr1;
+
+		/*
+		 * A set containing only two items cannot imply any equalities
+		 * beyond the one that created the set, so we can skip it.
+		 */
+		if (length(curset) < 3)
+			continue;
+
+		/*
+		 * Match each item in the set with all that appear after it
+		 * (it's sufficient to generate A=B, need not process B=A too).
+		 */
+		foreach(ptr1, curset)
+		{
+			PathKeyItem *item1 = (PathKeyItem *) lfirst(ptr1);
+			List	   *ptr2;
+
+			foreach(ptr2, lnext(ptr1))
+			{
+				PathKeyItem *item2 = (PathKeyItem *) lfirst(ptr2);
+
+				process_implied_equality(root, item1->key, item2->key,
+										 item1->sortop, item2->sortop);
+			}
+		}
+	}
+}
+
 /*
  * make_canonical_pathkey
  *	  Given a PathKeyItem, find the equi_key_list subset it is a member of,
diff --git a/src/backend/optimizer/plan/initsplan.c b/src/backend/optimizer/plan/initsplan.c
@@ -8,13 +8,14 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/initsplan.c,v 1.46 2000/04/12 17:15:21 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/initsplan.c,v 1.46.2.1 2000/09/23 23:50:47 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
 #include <sys/types.h>
 
 #include "postgres.h"
+#include "catalog/pg_operator.h"
 #include "catalog/pg_type.h"
 #include "nodes/makefuncs.h"
 #include "optimizer/clauses.h"
@@ -25,6 +26,9 @@
 #include "optimizer/planmain.h"
 #include "optimizer/tlist.h"
 #include "optimizer/var.h"
+#include "parser/parse_expr.h"
+#include "parser/parse_oper.h"
+#include "parser/parse_type.h"
 #include "utils/lsyscache.h"
 
 
@@ -280,6 +284,113 @@ add_join_info_to_rels(Query *root, RestrictInfo *restrictinfo,
 	}
 }
 
+/*
+ * process_implied_equality
+ *	  Check to see whether we already have a restrictinfo item that says
+ *	  item1 = item2, and create one if not.  This is a consequence of
+ *	  transitivity of mergejoin equality: if we have mergejoinable
+ *	  clauses A = B and B = C, we can deduce A = C (where = is an
+ *	  appropriate mergejoinable operator).
+ */
+void
+process_implied_equality(Query *root, Node *item1, Node *item2,
+						 Oid sortop1, Oid sortop2)
+{
+	Index		irel1;
+	Index		irel2;
+	RelOptInfo *rel1;
+	List	   *restrictlist;
+	List	   *itm;
+	Oid			ltype,
+				rtype;
+	Operator	eq_operator;
+	Form_pg_operator pgopform;
+	Expr	   *clause;
+
+	/*
+	 * Currently, since check_mergejoinable only accepts Var = Var clauses,
+	 * we should only see Var nodes here.  Would have to work a little
+	 * harder to locate the right rel(s) if more-general mergejoin clauses
+	 * were accepted.
+	 */
+	Assert(IsA(item1, Var));
+	irel1 = ((Var *) item1)->varno;
+	Assert(IsA(item2, Var));
+	irel2 = ((Var *) item2)->varno;
+	/*
+	 * If both vars belong to same rel, we need to look at that rel's
+	 * baserestrictinfo list.  If different rels, each will have a
+	 * joininfo node for the other, and we can scan either list.
+	 */
+	rel1 = get_base_rel(root, irel1);
+	if (irel1 == irel2)
+		restrictlist = rel1->baserestrictinfo;
+	else
+	{
+		JoinInfo   *joininfo = find_joininfo_node(rel1,
+												  lconsi(irel2, NIL));
+
+		restrictlist = joininfo->jinfo_restrictinfo;
+	}
+	/*
+	 * Scan to see if equality is already known.
+	 */
+	foreach(itm, restrictlist)
+	{
+		RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(itm);
+		Node	   *left,
+				   *right;
+
+		if (restrictinfo->mergejoinoperator == InvalidOid)
+			continue;			/* ignore non-mergejoinable clauses */
+		/* We now know the restrictinfo clause is a binary opclause */
+		left = (Node *) get_leftop(restrictinfo->clause);
+		right = (Node *) get_rightop(restrictinfo->clause);
+		if ((equal(item1, left) && equal(item2, right)) ||
+			(equal(item2, left) && equal(item1, right)))
+			return;				/* found a matching clause */
+	}
+	/*
+	 * This equality is new information, so construct a clause
+	 * representing it to add to the query data structures.
+	 */
+	ltype = exprType(item1);
+	rtype = exprType(item2);
+	eq_operator = oper("=", ltype, rtype, true);
+	if (!HeapTupleIsValid(eq_operator))
+	{
+		/*
+		 * Would it be safe to just not add the equality to the query if
+		 * we have no suitable equality operator for the combination of
+		 * datatypes?  NO, because sortkey selection may screw up anyway.
+		 */
+		elog(ERROR, "Unable to identify an equality operator for types '%s' and '%s'",
+			 typeidTypeName(ltype), typeidTypeName(rtype));
+	}
+	pgopform = (Form_pg_operator) GETSTRUCT(eq_operator);
+	/*
+	 * Let's just make sure this appears to be a compatible operator.
+	 */
+	if (pgopform->oprlsortop != sortop1 ||
+		pgopform->oprrsortop != sortop2 ||
+		pgopform->oprresult != BOOLOID)
+		elog(ERROR, "Equality operator for types '%s' and '%s' should be mergejoinable, but isn't",
+			 typeidTypeName(ltype), typeidTypeName(rtype));
+
+	clause = makeNode(Expr);
+	clause->typeOid = BOOLOID;
+	clause->opType = OP_EXPR;
+	clause->oper = (Node *) makeOper(oprid(eq_operator), /* opno */
+									 InvalidOid, /* opid */
+									 BOOLOID, /* operator result type */
+									 0,
+									 NULL);
+	clause->args = lcons(item1, lcons(item2, NIL));
+
+	add_restrict_and_join_to_rel(root, (Node *) clause);
+}
+
+
 /*****************************************************************************
  *
  *	 CHECKS FOR MERGEJOINABLE AND HASHJOINABLE CLAUSES
diff --git a/src/backend/optimizer/plan/planmain.c b/src/backend/optimizer/plan/planmain.c
@@ -14,7 +14,7 @@
  *
  *
  * IDENTIFICATION
- *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planmain.c,v 1.55 2000/04/12 17:15:22 momjian Exp $
+ *	  $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planmain.c,v 1.55.2.1 2000/09/23 23:50:47 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -184,7 +184,7 @@ subplanner(Query *root,
 	 * base_rel_list as relation references are found (e.g., in the
 	 * qualification, the targetlist, etc.).  Restrict and join clauses
 	 * are added to appropriate lists belonging to the mentioned
-	 * relations, and we also build lists of equijoined keys for pathkey
+	 * relations.  We also build lists of equijoined keys for pathkey
 	 * construction.
 	 */
 	root->base_rel_list = NIL;
@@ -193,8 +193,18 @@ subplanner(Query *root,
 
 	make_var_only_tlist(root, flat_tlist);
 	add_restrict_and_join_to_rels(root, qual);
+
+	/*
+	 * Make sure we have RelOptInfo nodes for all relations used.
+	 */
 	add_missing_rels_to_query(root);
 
+	/*
+	 * Use the completed lists of equijoined keys to deduce any implied
+	 * but unstated equalities (for example, A=B and B=C imply A=C).
+	 */
+	generate_implied_equalities(root);
+
 	/*
 	 * We should now have all the pathkey equivalence sets built, so it's
 	 * now possible to convert the requested query_pathkeys to canonical
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
@@ -8,7 +8,7 @@
  * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: paths.h,v 1.44 2000/04/12 17:16:42 momjian Exp $
+ * $Id: paths.h,v 1.44.2.1 2000/09/23 23:50:46 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -90,6 +90,7 @@ typedef enum
 } PathKeysComparison;
 
 extern void add_equijoined_keys(Query *root, RestrictInfo *restrictinfo);
+extern void generate_implied_equalities(Query *root);
 extern List *canonicalize_pathkeys(Query *root, List *pathkeys);
 extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
 extern bool pathkeys_contained_in(List *keys1, List *keys2);
diff --git a/src/include/optimizer/planmain.h b/src/include/optimizer/planmain.h
@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $Id: planmain.h,v 1.39 2000/04/12 17:16:42 momjian Exp $
+ * $Id: planmain.h,v 1.39.2.1 2000/09/23 23:50:46 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -43,6 +43,8 @@ extern Result *make_result(List *tlist, Node *resconstantqual, Plan *subplan);
 extern void make_var_only_tlist(Query *root, List *tlist);
 extern void add_restrict_and_join_to_rels(Query *root, List *clauses);
 extern void add_missing_rels_to_query(Query *root);
+extern void process_implied_equality(Query *root, Node *item1, Node *item2,
+									 Oid sortop1, Oid sortop2);
 
 /*
  * prototypes for plan/setrefs.c
