Fix misuse of RelOptInfo.unique_for_rels cache by SJE

When SJE uses RelOptInfo.unique_for_rels cache, it passes filtered quals to
innerrel_is_unique_ext().  That might lead to an invalid match to cache entries
made by previous non self-join checking calls.  Add UniqueRelInfo.self_join
flag to prevent such cases.  Also, fix that SJE should require a strict match
of outerrelids to make sure UniqueRelInfo.extra_clauses are valid.

Reported-by: Alexander Lakhin
Discussion: https://postgr.es/m/4788f781-31bd-9796-d7d6-588a751c8787%40gmail.com

src/backend/optimizer/plan/analyzejoins.c

@@ -1247,8 +1247,10 @@ innerrel_is_unique(PlannerInfo *root,
 
 /*
  * innerrel_is_unique_ext
- *	  Do the same as innerrel_is_unique(), but also return additional clauses
- *	  from a baserestrictinfo list that were used to prove uniqueness.
+ *	  Do the same as innerrel_is_unique(), but also set to '*extra_clauses'
+ *	  additional clauses from a baserestrictinfo list that were used to prove
+ *	  uniqueness.  A non NULL 'extra_clauses' indicates that we're checking
+ *	  for self-join and correspondingly dealing with filtered clauses.
  */
 bool
 innerrel_is_unique_ext(PlannerInfo *root,
@@ -1264,6 +1266,7 @@ innerrel_is_unique_ext(PlannerInfo *root,
 	ListCell   *lc;
 	UniqueRelInfo *uniqueRelInfo;
 	List	   *outer_exprs = NIL;
+	bool		self_join = (extra_clauses != NULL);
 
 	/* Certainly can't prove uniqueness when there are no joinclauses */
 	if (restrictlist == NIL)
@@ -1278,16 +1281,23 @@ innerrel_is_unique_ext(PlannerInfo *root,
 
 	/*
 	 * Query the cache to see if we've managed to prove that innerrel is
-	 * unique for any subset of this outerrel.  We don't need an exact match,
-	 * as extra outerrels can't make the innerrel any less unique (or more
-	 * formally, the restrictlist for a join to a superset outerrel must be a
-	 * superset of the conditions we successfully used before).
+	 * unique for any subset of this outerrel.  For non self-join search, we
+	 * don't need an exact match, as extra outerrels can't make the innerrel
+	 * any less unique (or more formally, the restrictlist for a join to a
+	 * superset outerrel must be a superset of the conditions we successfully
+	 * used before). For self-join search, we require an exact match of
+	 * outerrels, because we need extra clauses to be valid for our case.
+	 * Also, for self-join checking we've filtered the clauses list.  Thus,
+	 * for a self-join search, we can match only the result cached for another
+	 * self-join check.
 	 */
 	foreach(lc, innerrel->unique_for_rels)
 	{
 		uniqueRelInfo = (UniqueRelInfo *) lfirst(lc);
 
-		if (bms_is_subset(uniqueRelInfo->outerrelids, outerrelids))
+		if ((!self_join && bms_is_subset(uniqueRelInfo->outerrelids, outerrelids)) ||
+			(self_join && bms_equal(uniqueRelInfo->outerrelids, outerrelids) &&
+			 uniqueRelInfo->self_join))
 		{
 			if (extra_clauses)
 				*extra_clauses = uniqueRelInfo->extra_clauses;
@@ -1309,7 +1319,8 @@ innerrel_is_unique_ext(PlannerInfo *root,
 
 	/* No cached information, so try to make the proof. */
 	if (is_innerrel_unique_for(root, joinrelids, outerrelids, innerrel,
-							   jointype, restrictlist, &outer_exprs))
+							   jointype, restrictlist,
+							   self_join ? &outer_exprs : NULL))
 	{
 		/*
 		 * Cache the positive result for future probes, being sure to keep it
@@ -1323,8 +1334,9 @@ innerrel_is_unique_ext(PlannerInfo *root,
 		 */
 		old_context = MemoryContextSwitchTo(root->planner_cxt);
 		uniqueRelInfo = makeNode(UniqueRelInfo);
-		uniqueRelInfo->extra_clauses = outer_exprs;
 		uniqueRelInfo->outerrelids = bms_copy(outerrelids);
+		uniqueRelInfo->self_join = self_join;
+		uniqueRelInfo->extra_clauses = outer_exprs;
 		innerrel->unique_for_rels = lappend(innerrel->unique_for_rels,
 											uniqueRelInfo);
 		MemoryContextSwitchTo(old_context);

src/include/nodes/pathnodes.h

@@ -3407,6 +3407,12 @@ typedef struct UniqueRelInfo
 	 */
 	Relids		outerrelids;
 
+	/*
+	 * The relation in consideration is unique when considering only clauses
+	 * suitable for self-join (passed split_selfjoin_quals()).
+	 */
+	bool		self_join;
+
 	/*
 	 * Additional clauses from a baserestrictinfo list that were used to prove
 	 * the uniqueness.   We cache it for the self-join checking procedure: a

src/test/regress/expected/join.out

@@ -6905,6 +6905,29 @@ where false;
 ----------
 (0 rows)
 
+-- Check that SJE does not mistakenly re-use knowledge of relation uniqueness
+-- made with different set of quals
+insert into emp1 values (2, 1);
+explain (costs off)
+select * from emp1 t1 where exists (select * from emp1 t2
+                                    where t2.id = t1.code and t2.code > 0);
+                 QUERY PLAN                  
+---------------------------------------------
+ Nested Loop
+   ->  Seq Scan on emp1 t1
+   ->  Index Scan using emp1_pkey on emp1 t2
+         Index Cond: (id = t1.code)
+         Filter: (code > 0)
+(5 rows)
+
+select * from emp1 t1 where exists (select * from emp1 t2
+                                    where t2.id = t1.code and t2.code > 0);
+ id | code 
+----+------
+  1 |    1
+  2 |    1
+(2 rows)
+
 -- We can remove the join even if we find the join can't duplicate rows and
 -- the base quals of each side are different.  In the following case we end up
 -- moving quals over to s1 to make it so it can't match any rows.

src/test/regress/sql/join.sql

@@ -2638,6 +2638,15 @@ select 1 from emp1 full join
     where false) s on true
 where false;
 
+-- Check that SJE does not mistakenly re-use knowledge of relation uniqueness
+-- made with different set of quals
+insert into emp1 values (2, 1);
+explain (costs off)
+select * from emp1 t1 where exists (select * from emp1 t2
+                                    where t2.id = t1.code and t2.code > 0);
+select * from emp1 t1 where exists (select * from emp1 t2
+                                    where t2.id = t1.code and t2.code > 0);
+
 -- We can remove the join even if we find the join can't duplicate rows and
 -- the base quals of each side are different.  In the following case we end up
 -- moving quals over to s1 to make it so it can't match any rows.