Avoid unhelpful internal error for incorrect recursive-WITH queries.

checkWellFormedRecursion would issue "missing recursive reference"
if a WITH RECURSIVE query contained a single self-reference but
that self-reference was inside a top-level WITH, ORDER BY, LIMIT,
etc, rather than inside the second arm of the UNION as expected.
We already intended to throw more-on-point errors for such cases,
but those error checks must be done before examining the UNION arm
in order to have the desired results.  So this patch need only
move some code (and improve the comments).

Per bug #18536 from Alexander Lakhin.  Back-patch to all supported
branches.

Discussion: https://postgr.es/m/18536-0a342ec07901203e@postgresql.org

src/backend/parser/parse_cte.c

@@ -883,25 +883,14 @@ checkWellFormedRecursion(CteState *cstate)
 							cte->ctename),
 					 parser_errposition(cstate->pstate, cte->location)));
 
-		/* The left-hand operand mustn't contain self-reference at all */
-		cstate->curitem = i;
-		cstate->innerwiths = NIL;
-		cstate->selfrefcount = 0;
-		cstate->context = RECURSION_NONRECURSIVETERM;
-		checkWellFormedRecursionWalker((Node *) stmt->larg, cstate);
-		Assert(cstate->innerwiths == NIL);
-
-		/* Right-hand operand should contain one reference in a valid place */
-		cstate->curitem = i;
-		cstate->innerwiths = NIL;
-		cstate->selfrefcount = 0;
-		cstate->context = RECURSION_OK;
-		checkWellFormedRecursionWalker((Node *) stmt->rarg, cstate);
-		Assert(cstate->innerwiths == NIL);
-		if (cstate->selfrefcount != 1)	/* shouldn't happen */
-			elog(ERROR, "missing recursive reference");
-
-		/* WITH mustn't contain self-reference, either */
+		/*
+		 * Really, we should insist that there not be a top-level WITH, since
+		 * syntactically that would enclose the UNION.  However, we've not
+		 * done so in the past and it's probably too late to change.  Settle
+		 * for insisting that WITH not contain a self-reference.  Test this
+		 * before examining the UNION arms, to avoid issuing confusing errors
+		 * in such cases.
+		 */
 		if (stmt->withClause)
 		{
 			cstate->curitem = i;
@@ -918,7 +907,9 @@ checkWellFormedRecursion(CteState *cstate)
 		 * don't make sense because it's impossible to figure out what they
 		 * mean when we have only part of the recursive query's results. (If
 		 * we did allow them, we'd have to check for recursive references
-		 * inside these subtrees.)
+		 * inside these subtrees.  As for WITH, we have to do this before
+		 * examining the UNION arms, to avoid issuing confusing errors if
+		 * there is a recursive reference here.)
 		 */
 		if (stmt->sortClause)
 			ereport(ERROR,
@@ -944,6 +935,28 @@ checkWellFormedRecursion(CteState *cstate)
 					 errmsg("FOR UPDATE/SHARE in a recursive query is not implemented"),
 					 parser_errposition(cstate->pstate,
 										exprLocation((Node *) stmt->lockingClause))));
+
+		/*
+		 * Now we can get on with checking the UNION operands themselves.
+		 *
+		 * The left-hand operand mustn't contain a self-reference at all.
+		 */
+		cstate->curitem = i;
+		cstate->innerwiths = NIL;
+		cstate->selfrefcount = 0;
+		cstate->context = RECURSION_NONRECURSIVETERM;
+		checkWellFormedRecursionWalker((Node *) stmt->larg, cstate);
+		Assert(cstate->innerwiths == NIL);
+
+		/* Right-hand operand should contain one reference in a valid place */
+		cstate->curitem = i;
+		cstate->innerwiths = NIL;
+		cstate->selfrefcount = 0;
+		cstate->context = RECURSION_OK;
+		checkWellFormedRecursionWalker((Node *) stmt->rarg, cstate);
+		Assert(cstate->innerwiths == NIL);
+		if (cstate->selfrefcount != 1)	/* shouldn't happen */
+			elog(ERROR, "missing recursive reference");
 	}
 }
 

src/test/regress/expected/with.out

@@ -1976,6 +1976,46 @@ WITH RECURSIVE x(n) AS (SELECT n FROM x UNION ALL SELECT 1)
 ERROR:  recursive reference to query "x" must not appear within its non-recursive term
 LINE 1: WITH RECURSIVE x(n) AS (SELECT n FROM x UNION ALL SELECT 1)
                                               ^
+-- allow this, because we historically have
+WITH RECURSIVE x(n) AS (
+  WITH x1 AS (SELECT 1 AS n)
+    SELECT 0
+    UNION
+    SELECT * FROM x1)
+	SELECT * FROM x;
+ n 
+---
+ 0
+ 1
+(2 rows)
+
+-- but this should be rejected
+WITH RECURSIVE x(n) AS (
+  WITH x1 AS (SELECT 1 FROM x)
+    SELECT 0
+    UNION
+    SELECT * FROM x1)
+	SELECT * FROM x;
+ERROR:  recursive reference to query "x" must not appear within a subquery
+LINE 2:   WITH x1 AS (SELECT 1 FROM x)
+                                    ^
+-- and this too
+WITH RECURSIVE x(n) AS (
+  (WITH x1 AS (SELECT 1 FROM x) SELECT * FROM x1)
+  UNION
+  SELECT 0)
+	SELECT * FROM x;
+ERROR:  recursive reference to query "x" must not appear within its non-recursive term
+LINE 2:   (WITH x1 AS (SELECT 1 FROM x) SELECT * FROM x1)
+                                     ^
+-- and this
+WITH RECURSIVE x(n) AS (
+  SELECT 0 UNION SELECT 1
+  ORDER BY (SELECT n FROM x))
+	SELECT * FROM x;
+ERROR:  ORDER BY in a recursive query is not implemented
+LINE 3:   ORDER BY (SELECT n FROM x))
+                   ^
 CREATE TEMPORARY TABLE y (a INTEGER);
 INSERT INTO y SELECT generate_series(1, 10);
 -- LEFT JOIN

src/test/regress/sql/with.sql

@@ -889,6 +889,35 @@ WITH RECURSIVE x(n) AS (SELECT n FROM x)
 WITH RECURSIVE x(n) AS (SELECT n FROM x UNION ALL SELECT 1)
 	SELECT * FROM x;
 
+-- allow this, because we historically have
+WITH RECURSIVE x(n) AS (
+  WITH x1 AS (SELECT 1 AS n)
+    SELECT 0
+    UNION
+    SELECT * FROM x1)
+	SELECT * FROM x;
+
+-- but this should be rejected
+WITH RECURSIVE x(n) AS (
+  WITH x1 AS (SELECT 1 FROM x)
+    SELECT 0
+    UNION
+    SELECT * FROM x1)
+	SELECT * FROM x;
+
+-- and this too
+WITH RECURSIVE x(n) AS (
+  (WITH x1 AS (SELECT 1 FROM x) SELECT * FROM x1)
+  UNION
+  SELECT 0)
+	SELECT * FROM x;
+
+-- and this
+WITH RECURSIVE x(n) AS (
+  SELECT 0 UNION SELECT 1
+  ORDER BY (SELECT n FROM x))
+	SELECT * FROM x;
+
 CREATE TEMPORARY TABLE y (a INTEGER);
 INSERT INTO y SELECT generate_series(1, 10);