IMPALA-14714: Fix incorrect results for anti-join with constant ON cl…

fe/src/main/java/org/apache/impala/analysis/Analyzer.java

@@ -2128,13 +2128,20 @@ public void registerConjuncts(Expr e, boolean fromHavingClause)
    * evaluates the conjunct. If the conjunct evaluates to false, marks this query
    * block as having an empty result set or as having an empty select-project-join
    * portion, if fromHavingClause is true or false, respectively.
-   * No-op if the conjunct is not constant or is outer joined.
+   * No-op if the conjunct is not constant, is outer joined, or is anti-joined.
+   * For anti-joins, a constant FALSE in the ON clause means no rows from the right
+   * table will match, so all rows from the left table should be returned (for LEFT
+   * ANTI JOIN) or no rows should be returned (for RIGHT ANTI JOIN). We should not
+   * mark the entire query as having an empty result set.
    * Return true, if conjunct is constant FALSE.
    * Throws an AnalysisException if there is an error evaluating `conjunct`
    */
   private boolean markConstantConjunct(Expr conjunct, boolean fromHavingClause)
       throws AnalysisException {
-    if (!conjunct.isConstant() || isOjConjunct(conjunct)) return false;
+    if (!conjunct.isConstant() || isOjConjunct(conjunct)
+        || isAntiJoinedConjunct(conjunct)) {
+      return false;
+    }
     markConjunctAssigned(conjunct);
     if ((!fromHavingClause && !hasEmptySpjResultSet_)
         || (fromHavingClause && !hasEmptyResultSet_)) {
@@ -2495,13 +2502,15 @@ public boolean canEvalOuterJoinedConjunct(Expr e, List<TupleId> tids) {
    */
   public boolean canEvalPredicate(List<TupleId> tupleIds, Expr e) {
     if (!e.isBoundByTupleIds(tupleIds)) return false;
-    List<TupleId> tids = new ArrayList<>();
-    e.getIds(tids, null);
-    if (tids.isEmpty()) return true;
-
+    // Check On-clause conjuncts before the tids.isEmpty() shortcut so that constant
+    // On-clause predicates (e.g. ON false) are routed to canEvalOnClauseConjunct()
+    // instead of being incorrectly assigned to a scan node.
     if (e.isOnClauseConjunct()) {
       return canEvalOnClauseConjunct(tupleIds, e);
     }
+    List<TupleId> tids = new ArrayList<>();
+    e.getIds(tids, null);
+    if (tids.isEmpty()) return true;
     return isLastOjMaterializedByTupleIds(tupleIds, e);
   }
 
@@ -2587,6 +2596,11 @@ public boolean canEvalAntiJoinedConjunct(Expr e, List<TupleId> nodeTupleIds) {
     if (antiJoinRef == null) return true;
     List<TupleId> tids = new ArrayList<>();
     e.getIds(tids, null);
+    if (tids.isEmpty()) {
+      // Constant anti-join On-clause conjuncts (e.g. ON false) must be evaluated
+      // at the anti-join node, not at a scan node below it.
+      return nodeTupleIds.containsAll(antiJoinRef.getAllTableRefIds());
+    }
     if (tids.size() > 1) {
       return nodeTupleIds.containsAll(antiJoinRef.getAllTableRefIds())
           && antiJoinRef.getAllTableRefIds().containsAll(nodeTupleIds);

fe/src/main/java/org/apache/impala/planner/SingleNodePlanner.java

@@ -1317,6 +1317,15 @@ private void getConjunctsToInlineView(final Analyzer analyzer, final String alia
       List<TupleId> tids = new ArrayList<>();
       e.getIds(tids, null);
       if (tids.isEmpty()) {
+        // Do not migrate anti-join conjuncts (e.g., ON FALSE) into the inline view.
+        // For anti-joins, a constant FALSE in the ON clause means no rows from the
+        // right table will match, so the correct behavior depends on the join
+        // type (LEFT ANTI returns all left rows, RIGHT ANTI returns all right rows).
+        // Migrating such conjuncts would incorrectly mark the inline view as having an
+        // empty result set.
+        if (analyzer.isAntiJoinedConjunct(e)) {
+          continue;
+        }
         evalInInlineViewPreds.add(e);
       } else if (e.isOnClauseConjunct()) {
         if (!analyzer.canEvalOnClauseConjunct(tupleIds, e)) continue;
@@ -2083,8 +2092,8 @@ private PlanNode createJoinNode(PlanNode outer, PlanNode inner,
       // Unassigned conjuncts bound by the invisible tuple id of a semi join must have
       // come from the join's On-clause, and therefore, must be added to the other join
       // conjuncts to produce correct results.
-      // TODO This doesn't handle predicates specified in the On clause which are not
-      // bound by any tuple id (e.g. ON (true))
+      // Note: Constant predicates in the On clause (e.g. ON TRUE, ON FALSE) are now
+      // handled correctly through canEvalPredicate() and canEvalAntiJoinedConjunct().
       List<TupleId> tblRefIds = Lists.newArrayList(outer.getTblRefIds());
       tblRefIds.addAll(inner.getTblRefIds());
       otherJoinConjuncts = analyzer.getUnassignedConjuncts(tblRefIds, false);

testdata/workloads/functional-query/queries/QueryTest/semi-joins.test

@@ -346,3 +346,190 @@ RIGHT ANTI JOIN SemiJoinTblB b on v.b = b.b
 ---- TYPES
 INT, INT, INT
 ====
+---- QUERY
+# Testing LEFT ANTI JOIN with ON FALSE should return all rows from left table
+# because no rows from right table can match.
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON false
+---- RESULTS
+1,1,1
+1,1,10
+1,2,10
+1,3,10
+NULL,NULL,30
+2,4,30
+2,NULL,20
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing RIGHT ANTI JOIN with ON FALSE should return all rows from right table
+# because no rows from left table can match.
+SELECT b.* FROM SemiJoinTblA a RIGHT ANTI JOIN SemiJoinTblB b ON false
+---- RESULTS
+1,1,1
+1,1,10
+1,2,5
+1,NULL,10
+2,10,NULL
+3,NULL,NULL
+3,NULL,50
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT ANTI JOIN with ON TRUE should return empty result set
+# because all rows from right table match.
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON true
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing RIGHT ANTI JOIN with ON TRUE should return empty result set
+# because all rows from left table match.
+SELECT b.* FROM SemiJoinTblA a RIGHT ANTI JOIN SemiJoinTblB b ON true
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT ANTI JOIN with constant expression 1=0 (equivalent to FALSE)
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON 1=0
+---- RESULTS
+1,1,1
+1,1,10
+1,2,10
+1,3,10
+NULL,NULL,30
+2,4,30
+2,NULL,20
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT ANTI JOIN with constant expression 1=1 (equivalent to TRUE)
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON 1=1
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT ANTI JOIN with ON FALSE and WHERE clause
+# Should return rows from left table that satisfy WHERE condition
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON false
+WHERE a.a = 1
+---- RESULTS
+1,1,1
+1,1,10
+1,2,10
+1,3,10
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT ANTI JOIN with ON TRUE and WHERE clause
+# Should return empty result (anti-join filters all, then WHERE is applied)
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON true
+WHERE a.a = 1
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT SEMI JOIN with ON FALSE should return empty result
+# because no rows from right table can match.
+SELECT a.* FROM SemiJoinTblA a LEFT SEMI JOIN SemiJoinTblB b ON false
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing LEFT SEMI JOIN with ON TRUE should return all rows from left table
+SELECT a.* FROM SemiJoinTblA a LEFT SEMI JOIN SemiJoinTblB b ON true
+---- RESULTS
+1,1,1
+1,1,10
+1,2,10
+1,3,10
+NULL,NULL,30
+2,4,30
+2,NULL,20
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing nested query with LEFT ANTI JOIN ON FALSE
+SELECT * FROM (
+  SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON false
+) v WHERE v.a = 2
+---- RESULTS
+2,4,30
+2,NULL,20
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing multiple anti-joins with constant predicates
+SELECT a.* FROM SemiJoinTblA a
+LEFT ANTI JOIN SemiJoinTblB b ON false
+LEFT ANTI JOIN SemiJoinTblB c ON true
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing anti-join with mixed constant and regular predicates (ON FALSE OR a.a = b.a)
+# The OR with FALSE should be optimized away, leaving just a.a = b.a
+# Only rows with NULL in column a are returned because NULL doesn't match anything
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON (false OR a.a = b.a)
+---- RESULTS
+NULL,NULL,30
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing anti-join with mixed constant and regular predicates (ON TRUE AND a.a = b.a)
+# The AND with TRUE should be optimized away, leaving just a.a = b.a
+# Only rows with NULL in column a are returned because NULL doesn't match anything
+SELECT a.* FROM SemiJoinTblA a LEFT ANTI JOIN SemiJoinTblB b ON (true AND a.a = b.a)
+---- RESULTS
+NULL,NULL,30
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing anti-join with inline view on the left side and ON FALSE
+# The inline view should not be marked as having an empty result set
+SELECT v.* FROM (SELECT * FROM SemiJoinTblA WHERE a = 1) v
+LEFT ANTI JOIN SemiJoinTblB b ON false
+---- RESULTS
+1,1,1
+1,1,10
+1,2,10
+1,3,10
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing nested inline views with anti-join and ON FALSE
+# Verifies that constant predicates don't propagate incorrectly through nested views
+SELECT * FROM (
+  SELECT v.* FROM (SELECT * FROM SemiJoinTblA WHERE a = 2) v
+  LEFT ANTI JOIN SemiJoinTblB b ON false
+) v2 WHERE v2.a = 2
+---- RESULTS
+2,4,30
+2,NULL,20
+---- TYPES
+INT, INT, INT
+====
+---- QUERY
+# Testing anti-join with inline view and mixed predicates
+# Constant FALSE should not be migrated into the inline view
+SELECT v.* FROM (SELECT * FROM SemiJoinTblA WHERE a = 1) v
+LEFT ANTI JOIN SemiJoinTblB b ON (false OR v.a = b.a)
+---- RESULTS
+---- TYPES
+INT, INT, INT
+====
+