Fix possible usage of incorrect UPPERREL_SETOP RelOptInfo

03d40e4 allowed dummy UNION [ALL] children to be removed from the plan
by checking for is_dummy_rel().  That commit neglected to still account
for the relids from the dummy rel so that the correct UPPERREL_SETOP
RelOptInfo could be found and used for adding the Paths to.

Not doing this could result in processing of subsequent UNIONs using the
same RelOptInfo as a previously processed UNION, which could result in
add_path() freeing old Paths that are needed by the previous UNION.

The same fix was independently submitted (2 mins later) by Richard Guo.

Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/bee34aec-659c-46f1-9ab7-7bbae0b7616c@gmail.com

Author: david-rowley

src/backend/optimizer/prep/prepunion.c

@@ -773,6 +773,12 @@ generate_union_paths(SetOperationStmt *op, PlannerInfo *root,
 		RelOptInfo *rel = lfirst(lc);
 		Path	   *ordered_path;
 
+		/*
+		 * Record the relids so that we can identify the correct
+		 * UPPERREL_SETOP RelOptInfo below.
+		 */
+		relids = bms_add_members(relids, rel->relids);
+
 		/* Skip any UNION children that are proven not to yield any rows */
 		if (is_dummy_rel(rel))
 			continue;
@@ -815,8 +821,6 @@ generate_union_paths(SetOperationStmt *op, PlannerInfo *root,
 				partial_pathlist = lappend(partial_pathlist,
 										   linitial(rel->partial_pathlist));
 		}
-
-		relids = bms_add_members(relids, rel->relids);
 	}
 
 	/* Build result relation. */

src/test/regress/expected/union.out

@@ -1260,14 +1260,14 @@ SELECT four FROM tenk1 WHERE 1=2
 UNION
 SELECT ten FROM tenk1 WHERE 1=2
 ORDER BY 1;
-              QUERY PLAN              
---------------------------------------
+                                       QUERY PLAN                                        
+-----------------------------------------------------------------------------------------
  Sort
    Output: unnamed_subquery.two
    Sort Key: unnamed_subquery.two
    ->  Result
          Output: unnamed_subquery.two
-         Replaces: Aggregate
+         Replaces: Aggregate on unnamed_subquery, unnamed_subquery_1, unnamed_subquery_2
          One-Time Filter: false
 (7 rows)