Fix left semi project join when all build side rows have null join ke…

…ys (#3635)

Summary:
When build side is not empty, but all its rows have null join keys, the result
of the semi project join should be NULL, not FALSE.

Fixes #3626.

Pull Request resolved: #3635

Reviewed By: xiaoxmeng

Differential Revision: D42344123

Pulled By: mbasmanova

fbshipit-source-id: 41dc97ba3fb5b0167f6b8e650a1c183891719ccd

velox/exec/HashProbe.cpp

@@ -642,9 +642,12 @@ void HashProbe::prepareOutput(vector_size_t size) {
 
 void HashProbe::fillLeftSemiProjectMatchColumn(vector_size_t size) {
   if (emptyBuildSide()) {
-    // Build side is empty. All rows should return 'match = false', even ones
-    // with a null join key.
-    matchColumn() = BaseVector::createConstant(false, size, pool());
+    // Build side is empty or all rows have null join keys.
+    if (nullAware_ && buildSideHasNullKeys_) {
+      matchColumn() = BaseVector::createNullConstant(BOOLEAN(), size, pool());
+    } else {
+      matchColumn() = BaseVector::createConstant(false, size, pool());
+    }
   } else {
     auto flatMatch = matchColumn()->as<FlatVector<bool>>();
     flatMatch->resize(size);

velox/exec/tests/HashJoinTest.cpp

@@ -3089,6 +3089,39 @@ TEST_F(HashJoinTest, semiProjectWithNullKeys) {
       .referenceQuery(
           "SELECT t0, t1, t0 IN (SELECT u0 FROM u WHERE u0 < 0) FROM t")
       .run();
+
+  // Build side with all rows having null join keys.
+  plan = makePlan(false /*nullAware*/, "", "u0 IS NULL");
+
+  HashJoinBuilder(*pool_, duckDbQueryRunner_, executor_.get())
+      .planNode(plan)
+      .checkSpillStats(false)
+      .referenceQuery(
+          "SELECT t0, t1, EXISTS (SELECT * FROM u WHERE u0 = t0 AND u0 IS NULL) FROM t")
+      .run();
+
+  HashJoinBuilder(*pool_, duckDbQueryRunner_, executor_.get())
+      .planNode(flipJoinSides(plan))
+      .checkSpillStats(false)
+      .referenceQuery(
+          "SELECT t0, t1, EXISTS (SELECT * FROM u WHERE u0 = t0 AND u0 IS NULL) FROM t")
+      .run();
+
+  plan = makePlan(true /*nullAware*/, "", "u0 IS NULL");
+
+  HashJoinBuilder(*pool_, duckDbQueryRunner_, executor_.get())
+      .planNode(plan)
+      .checkSpillStats(false)
+      .referenceQuery(
+          "SELECT t0, t1, t0 IN (SELECT u0 FROM u WHERE u0 IS NULL) FROM t")
+      .run();
+
+  HashJoinBuilder(*pool_, duckDbQueryRunner_, executor_.get())
+      .planNode(flipJoinSides(plan))
+      .checkSpillStats(false)
+      .referenceQuery(
+          "SELECT t0, t1, t0 IN (SELECT u0 FROM u WHERE u0 IS NULL) FROM t")
+      .run();
 }
 
 TEST_F(HashJoinTest, semiProjectOverLazyVectors) {