Copilot commented on code in PR #12756:
URL: https://github.com/apache/gluten/pull/12756#discussion_r3871264179


##########
backends-velox/src/main/scala/org/apache/gluten/extension/RewriteSelfJoinInequalityToAggregate.scala:
##########
@@ -0,0 +1,674 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.gluten.extension
+
+import org.apache.gluten.config.VeloxConfig
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.RowOrdering
+import org.apache.spark.sql.catalyst.expressions.aggregate._
+import org.apache.spark.sql.catalyst.plans._
+import org.apache.spark.sql.catalyst.plans.logical._
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, 
LogicalRelation}
+import org.apache.spark.sql.execution.datasources.parquet.ParquetFileFormat
+import org.apache.spark.sql.types.LongType
+
+/**
+ * Rewrites self-join with inequality into GROUP BY + HAVING COUNT(DISTINCT) > 
1.
+ *
+ * Targets the two uncorrelated InSubquery shapes exercised by TPC-DS Q95:
+ *
+ *   - Pattern A': the subquery top-level InnerJoin is a direct self-join.
+ *   - Pattern A2: the subquery contains an outer InnerJoin with a self-join 
child; only the
+ *     self-join child is replaced with Aggregate and the outer join is 
preserved.
+ *
+ * Both patterns require an existence-only membership context so row-count 
multiplicity from the
+ * original self-join cross-product does not affect semantics. Correlated 
InSubquery expressions are
+ * intentionally fail-closed because the ExprId remapping performed here does 
not rewrite correlated
+ * predicates.
+ *
+ * Both patterns share:
+ *   - [[buildAggregateHavingDistinctGt1]] to construct `Filter(cnt > 1, 
Aggregate)`
+ *   - [[canonicalizeWrapper]] to rebuild a wrapping Project so every equi-key 
reference points to
+ *     the sjLeft-side attribute, with **fresh exprIds** (Spark's SPARK-21835 
style -- no reuse of
+ *     original exprIds), returning an old->new attribute remap for downstream 
rewrite.
+ *
+ * Controlled by `spark.gluten.sql.rewrite.selfJoinInequality` (default false, 
opt-in).
+ */
+case class RewriteSelfJoinInequalityToAggregate(spark: SparkSession)
+  extends Rule[LogicalPlan]
+  with PredicateHelper
+  with Logging {
+
+  private val CountDistinctAliasName = "_gluten_rw_selfjoin_cnt_distinct"
+
+  override def apply(plan: LogicalPlan): LogicalPlan = {
+    if (!VeloxConfig.get.enableRewriteSelfJoinInequality) {
+      logDebug("RewriteSelfJoinInequalityToAggregate: disabled via config, 
skipping")
+      return plan
+    }
+
+    // Pattern A' / A2: rewrite uncorrelated InSubquery plans.
+    // Correlated subqueries carry outer references / correlated join 
conditions in
+    // `SubqueryExpression.children`; fail closed because this rule does not 
remap them.
+    val rewritten = plan.transformAllExpressions {
+      case in @ InSubquery(_, lq: ListQuery) if lq.children.isEmpty =>
+        rewriteSubqueryPlan(lq.plan) match {
+          case Some(newSub) => in.copy(query = lq.copy(plan = newSub))
+          case None => in
+        }
+    }
+    if (!(rewritten eq plan)) {
+      logDebug(
+        "RewriteSelfJoinInequalityToAggregate: rewrote self-join to " +
+          "GROUP BY + HAVING COUNT(DISTINCT) > 1")
+    }
+    rewritten
+  }
+  // 
============================================================================
+  //  Shared helpers
+  // 
============================================================================
+
+  /**
+   * Build `Filter(cnt > 1, Aggregate(equiKeys, [equiKeys, cnt_alias], 
Filter(IsNotNull(equiKeys),
+   * child)))`. Returns the Filter node whose output is `equiKeys ++ 
[count_alias_attr]`.
+   *
+   * The extra `IsNotNull(equiKeys)` filter is essential to preserve the 
original equi-join's NULL
+   * semantics. Under SQL 3VL, `left.k = right.k` never matches when either 
side is NULL, so the
+   * original self-join drops rows with NULL equi-keys. Aggregate, in 
contrast, groups NULL keys
+   * together into a single "NULL group" -- if that group has >= 2 distinct 
non-null neq values,
+   * COUNT(DISTINCT) > 1 fires and injects NULL into the subquery output. That 
leaked NULL then
+   * turns `NOT IN` into a spurious empty result (Spark's null-aware anti-join 
uses
+   * `Or(equi, IsNull(equi))` which any NULL sub-row satisfies) and can flip 
IN/NOT IN outcomes. The
+   * neq column needs no such filter: `COUNT(DISTINCT col)` already ignores 
NULL.
+   */
+  private def buildAggregateHavingDistinctGt1(
+      equiKeys: Seq[Attribute],
+      neqCol: Attribute,
+      child: LogicalPlan): LogicalPlan = {
+    val countExpr = AggregateExpression(
+      Count(Seq(neqCol)),
+      mode = Complete,
+      isDistinct = true,
+      filter = None,
+      NamedExpression.newExprId)
+    val countAlias = Alias(countExpr, CountDistinctAliasName)()
+    // Seq[Attribute] is a Seq[NamedExpression] via covariance; no cast needed.
+    val aggExprs: Seq[NamedExpression] = equiKeys :+ countAlias
+    val nonNullChild = equiKeys
+      .map(a => IsNotNull(a): Expression)
+      .reduceOption(And)
+      .map(Filter(_, child))
+      .getOrElse(child)
+    val agg = Aggregate(equiKeys, aggExprs, nonNullChild)
+    Filter(GreaterThan(countAlias.toAttribute, Literal(1L, LongType)), agg)
+  }
+
+  /**
+   * Canonicalize a Project so every equi-key reference points at the 
sjLeft-side attribute.
+   * [[parseSelfJoinCondition]] has already verified that each pair refers to 
the same output
+   * position on the two structurally identical self-join sides. Uses **fresh 
exprIds** (no reuse of
+   * original wrapper output exprIds) -- the same technique Spark's own 
`dedupSubqueryOnSelfJoin`
+   * uses when it needs to change subquery output.
+   *
+   * Returns the rebuilt Project and a map `oldWrapperOutputExprId -> 
newWrapperOutputAttr`, so
+   * downstream references (outer join condition, top-level Project) can be 
updated consistently.
+   *
+   * `equiPairs` provides the definitive ExprId-based lookup: `equiPair (l, 
r)` binds
+   * `l.exprId -> l` (identity) and `r.exprId -> l` (sjRight -> sjLeft). 
Attribute identity in
+   * Catalyst is ExprId, not name; two columns can share a name with distinct 
ExprIds. Name-based
+   * lookup would silently drop such entries via `.toMap`.
+   *
+   * Fails (returns None) when a projectList entry is neither an equi-key 
Attribute (by ExprId) nor
+   * `Alias(equi-key Attribute, _)`. Fail-closed.
+   */
+  private def canonicalizeWrapper(
+      projectList: Seq[NamedExpression],
+      equiPairs: Seq[(Attribute, Attribute)],
+      newChild: LogicalPlan): Option[(Project, Map[ExprId, Attribute])] = {
+    // ExprId-based canonical map: any equi-key attribute (either side) -> 
sjLeft attribute.
+    val exprIdToLeft: Map[ExprId, Attribute] =
+      equiPairs.flatMap { case (l, r) => Seq(l.exprId -> l, r.exprId -> l) 
}.toMap
+    val oldOutput: Seq[Attribute] = projectList.map(_.toAttribute)
+    val mapped: Seq[Option[NamedExpression]] = projectList.map {
+      case a: Attribute if exprIdToLeft.contains(a.exprId) =>
+        // Wrap every rewritten output slot in a fresh Alias.
+        //
+        // When a wrapper reprojects BOTH sides of the same equi pair (e.g.
+        // `SELECT s1.k, s2.k FROM T s1 JOIN T s2 ON s1.k = s2.k AND s1.v <> 
s2.v`),
+        // both entries collapse to the same sjLeft Attribute after the 
self-join is
+        // rewritten. Duplicate output ExprIds are not illegal in Spark 
(`SELECT a, a`
+        // is a valid Project), but fresh Aliases give each output slot an 
independent
+        // identity, which keeps the `oldOutput -> newOutput` remap 1-to-1 and 
lets
+        // downstream references (outer join condition, top-level Project) be 
updated
+        // unambiguously via ExprId.
+        //
+        // The fresh ExprId is on the Alias ITSELF; the referenced child keeps 
its
+        // original ExprId. Spark's logical-plan integrity checks reject 
reusing a
+        // referenced ExprId as the Alias's own ExprId, not duplication across 
slots.
+        Some(Alias(exprIdToLeft(a.exprId), a.name)(): NamedExpression)
+      case al @ Alias(a: Attribute, _) if exprIdToLeft.contains(a.exprId) =>
+        // Fresh exprId; do NOT reuse `al.exprId`. Reusing another 
expression's exprId
+        // is the pattern that Spark 3.3 flags via structural-integrity checks.
+        Some(Alias(exprIdToLeft(a.exprId), al.name)(): NamedExpression)

Review Comment:
   When rebuilding wrapper projections, the new `Alias(...)()` drops qualifier 
and explicit metadata from the original `Attribute`/`Alias` outputs. Even if 
semantics are unchanged, losing column metadata can affect downstream consumers 
(e.g., metadata-driven features, column comments, or analyzer/debugging 
output). Consider preserving `qualifier` and `explicitMetadata` from the 
original `Alias`, and preserving attribute metadata where applicable.



##########
backends-velox/src/main/scala/org/apache/gluten/extension/RewriteSelfJoinInequalityToAggregate.scala:
##########
@@ -0,0 +1,674 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.gluten.extension
+
+import org.apache.gluten.config.VeloxConfig
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.RowOrdering
+import org.apache.spark.sql.catalyst.expressions.aggregate._
+import org.apache.spark.sql.catalyst.plans._
+import org.apache.spark.sql.catalyst.plans.logical._
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, 
LogicalRelation}
+import org.apache.spark.sql.execution.datasources.parquet.ParquetFileFormat
+import org.apache.spark.sql.types.LongType
+
+/**
+ * Rewrites self-join with inequality into GROUP BY + HAVING COUNT(DISTINCT) > 
1.
+ *
+ * Targets the two uncorrelated InSubquery shapes exercised by TPC-DS Q95:
+ *
+ *   - Pattern A': the subquery top-level InnerJoin is a direct self-join.
+ *   - Pattern A2: the subquery contains an outer InnerJoin with a self-join 
child; only the
+ *     self-join child is replaced with Aggregate and the outer join is 
preserved.
+ *
+ * Both patterns require an existence-only membership context so row-count 
multiplicity from the
+ * original self-join cross-product does not affect semantics. Correlated 
InSubquery expressions are
+ * intentionally fail-closed because the ExprId remapping performed here does 
not rewrite correlated
+ * predicates.
+ *
+ * Both patterns share:
+ *   - [[buildAggregateHavingDistinctGt1]] to construct `Filter(cnt > 1, 
Aggregate)`
+ *   - [[canonicalizeWrapper]] to rebuild a wrapping Project so every equi-key 
reference points to
+ *     the sjLeft-side attribute, with **fresh exprIds** (Spark's SPARK-21835 
style -- no reuse of
+ *     original exprIds), returning an old->new attribute remap for downstream 
rewrite.
+ *
+ * Controlled by `spark.gluten.sql.rewrite.selfJoinInequality` (default false, 
opt-in).
+ */
+case class RewriteSelfJoinInequalityToAggregate(spark: SparkSession)
+  extends Rule[LogicalPlan]
+  with PredicateHelper
+  with Logging {
+
+  private val CountDistinctAliasName = "_gluten_rw_selfjoin_cnt_distinct"
+
+  override def apply(plan: LogicalPlan): LogicalPlan = {
+    if (!VeloxConfig.get.enableRewriteSelfJoinInequality) {
+      logDebug("RewriteSelfJoinInequalityToAggregate: disabled via config, 
skipping")
+      return plan
+    }
+
+    // Pattern A' / A2: rewrite uncorrelated InSubquery plans.
+    // Correlated subqueries carry outer references / correlated join 
conditions in
+    // `SubqueryExpression.children`; fail closed because this rule does not 
remap them.
+    val rewritten = plan.transformAllExpressions {
+      case in @ InSubquery(_, lq: ListQuery) if lq.children.isEmpty =>
+        rewriteSubqueryPlan(lq.plan) match {
+          case Some(newSub) => in.copy(query = lq.copy(plan = newSub))
+          case None => in
+        }
+    }
+    if (!(rewritten eq plan)) {
+      logDebug(
+        "RewriteSelfJoinInequalityToAggregate: rewrote self-join to " +
+          "GROUP BY + HAVING COUNT(DISTINCT) > 1")
+    }
+    rewritten
+  }
+  // 
============================================================================
+  //  Shared helpers
+  // 
============================================================================
+
+  /**
+   * Build `Filter(cnt > 1, Aggregate(equiKeys, [equiKeys, cnt_alias], 
Filter(IsNotNull(equiKeys),
+   * child)))`. Returns the Filter node whose output is `equiKeys ++ 
[count_alias_attr]`.
+   *
+   * The extra `IsNotNull(equiKeys)` filter is essential to preserve the 
original equi-join's NULL
+   * semantics. Under SQL 3VL, `left.k = right.k` never matches when either 
side is NULL, so the
+   * original self-join drops rows with NULL equi-keys. Aggregate, in 
contrast, groups NULL keys
+   * together into a single "NULL group" -- if that group has >= 2 distinct 
non-null neq values,
+   * COUNT(DISTINCT) > 1 fires and injects NULL into the subquery output. That 
leaked NULL then
+   * turns `NOT IN` into a spurious empty result (Spark's null-aware anti-join 
uses
+   * `Or(equi, IsNull(equi))` which any NULL sub-row satisfies) and can flip 
IN/NOT IN outcomes. The
+   * neq column needs no such filter: `COUNT(DISTINCT col)` already ignores 
NULL.
+   */
+  private def buildAggregateHavingDistinctGt1(
+      equiKeys: Seq[Attribute],
+      neqCol: Attribute,
+      child: LogicalPlan): LogicalPlan = {
+    val countExpr = AggregateExpression(
+      Count(Seq(neqCol)),
+      mode = Complete,
+      isDistinct = true,
+      filter = None,
+      NamedExpression.newExprId)
+    val countAlias = Alias(countExpr, CountDistinctAliasName)()
+    // Seq[Attribute] is a Seq[NamedExpression] via covariance; no cast needed.
+    val aggExprs: Seq[NamedExpression] = equiKeys :+ countAlias
+    val nonNullChild = equiKeys
+      .map(a => IsNotNull(a): Expression)
+      .reduceOption(And)
+      .map(Filter(_, child))
+      .getOrElse(child)
+    val agg = Aggregate(equiKeys, aggExprs, nonNullChild)
+    Filter(GreaterThan(countAlias.toAttribute, Literal(1L, LongType)), agg)
+  }
+
+  /**
+   * Canonicalize a Project so every equi-key reference points at the 
sjLeft-side attribute.
+   * [[parseSelfJoinCondition]] has already verified that each pair refers to 
the same output
+   * position on the two structurally identical self-join sides. Uses **fresh 
exprIds** (no reuse of
+   * original wrapper output exprIds) -- the same technique Spark's own 
`dedupSubqueryOnSelfJoin`
+   * uses when it needs to change subquery output.
+   *
+   * Returns the rebuilt Project and a map `oldWrapperOutputExprId -> 
newWrapperOutputAttr`, so
+   * downstream references (outer join condition, top-level Project) can be 
updated consistently.
+   *
+   * `equiPairs` provides the definitive ExprId-based lookup: `equiPair (l, 
r)` binds
+   * `l.exprId -> l` (identity) and `r.exprId -> l` (sjRight -> sjLeft). 
Attribute identity in
+   * Catalyst is ExprId, not name; two columns can share a name with distinct 
ExprIds. Name-based
+   * lookup would silently drop such entries via `.toMap`.
+   *
+   * Fails (returns None) when a projectList entry is neither an equi-key 
Attribute (by ExprId) nor
+   * `Alias(equi-key Attribute, _)`. Fail-closed.
+   */
+  private def canonicalizeWrapper(
+      projectList: Seq[NamedExpression],
+      equiPairs: Seq[(Attribute, Attribute)],
+      newChild: LogicalPlan): Option[(Project, Map[ExprId, Attribute])] = {
+    // ExprId-based canonical map: any equi-key attribute (either side) -> 
sjLeft attribute.
+    val exprIdToLeft: Map[ExprId, Attribute] =
+      equiPairs.flatMap { case (l, r) => Seq(l.exprId -> l, r.exprId -> l) 
}.toMap
+    val oldOutput: Seq[Attribute] = projectList.map(_.toAttribute)
+    val mapped: Seq[Option[NamedExpression]] = projectList.map {
+      case a: Attribute if exprIdToLeft.contains(a.exprId) =>
+        // Wrap every rewritten output slot in a fresh Alias.
+        //
+        // When a wrapper reprojects BOTH sides of the same equi pair (e.g.
+        // `SELECT s1.k, s2.k FROM T s1 JOIN T s2 ON s1.k = s2.k AND s1.v <> 
s2.v`),
+        // both entries collapse to the same sjLeft Attribute after the 
self-join is
+        // rewritten. Duplicate output ExprIds are not illegal in Spark 
(`SELECT a, a`
+        // is a valid Project), but fresh Aliases give each output slot an 
independent
+        // identity, which keeps the `oldOutput -> newOutput` remap 1-to-1 and 
lets
+        // downstream references (outer join condition, top-level Project) be 
updated
+        // unambiguously via ExprId.
+        //
+        // The fresh ExprId is on the Alias ITSELF; the referenced child keeps 
its
+        // original ExprId. Spark's logical-plan integrity checks reject 
reusing a
+        // referenced ExprId as the Alias's own ExprId, not duplication across 
slots.
+        Some(Alias(exprIdToLeft(a.exprId), a.name)(): NamedExpression)
+      case al @ Alias(a: Attribute, _) if exprIdToLeft.contains(a.exprId) =>
+        // Fresh exprId; do NOT reuse `al.exprId`. Reusing another 
expression's exprId
+        // is the pattern that Spark 3.3 flags via structural-integrity checks.
+        Some(Alias(exprIdToLeft(a.exprId), al.name)(): NamedExpression)
+      case _ => None
+    }
+    if (mapped.exists(_.isEmpty)) {
+      None
+    } else {
+      val newProjectList = mapped.flatten
+      val newWrapper = Project(newProjectList, newChild)
+      val newOutput = newWrapper.output
+      val remap: Map[ExprId, Attribute] =
+        oldOutput.zip(newOutput).map { case (o, n) => o.exprId -> n }.toMap
+      Some((newWrapper, remap))
+    }
+  }
+
+  /**
+   * Replace equi-key attribute references inside a NamedExpression according 
to `remap`, while
+   * preserving the NamedExpression shape.
+   *
+   * `Expression.transformUp` returns `Expression`, not `NamedExpression`. We 
avoid a blanket
+   * `asInstanceOf[NamedExpression]` by handling the two shapes that can 
appear in a Project's
+   * `projectList` explicitly: a bare Attribute (whose top-level may itself be 
replaced) and an
+   * Alias (which stays an Alias while its child is transformed). Anything 
else in a projectList --
+   * e.g. computed expressions we don't own -- is passed through unchanged.
+   */
+  private def remapNamedExpressionAttributes(
+      ne: NamedExpression,
+      remap: Map[ExprId, Attribute]): NamedExpression = ne match {
+    case a: Attribute if remap.contains(a.exprId) => remap(a.exprId)
+    case a: Attribute => a
+    case al: Alias =>
+      val newChild = al.child.transformUp {
+        case a: Attribute if remap.contains(a.exprId) => remap(a.exprId)
+      }
+      if (newChild eq al.child) al
+      else Alias(newChild, al.name)(al.exprId, al.qualifier, 
al.explicitMetadata)
+    case other => other
+  }
+
+  // 
============================================================================
+  //  Pattern A' / A2 dispatch (subquery plans of InSubquery)
+  // 
============================================================================
+
+  private def rewriteSubqueryPlan(plan: LogicalPlan): Option[LogicalPlan] = {
+    // Candidate-level nondeterminism guard: reject if ANY node in the whole 
subquery plan
+    // is non-repeatable (Rand, LIMIT-without-ORDER-BY, Sample, Offset, 
streaming). This
+    // catches nondeterminism that has been hoisted above the self-join by an 
earlier
+    // optimizer rule -- the per-side `isSameBaseRelation` check alone would 
miss it because
+    // both innerLeft/innerRight can look deterministic after such a hoist.
+    if (!isRepeatablePlan(plan)) return None
+
+    val (projectListOpt, innerJoin): (Option[Seq[NamedExpression]], Join) = 
plan match {
+      case Project(pl, j: Join) if j.joinType == Inner && 
j.condition.isDefined =>
+        (Some(pl), j)
+      case j: Join if j.joinType == Inner && j.condition.isDefined =>
+        (None, j)
+      case _ => return None
+    }
+
+    if (isSameBaseRelation(innerJoin.left, innerJoin.right)) {
+      rewriteDirectSelfJoin(projectListOpt, innerJoin)
+    } else {
+      rewriteNestedSelfJoin(projectListOpt, innerJoin)
+    }
+  }
+
+  // 
============================================================================
+  //  Pattern A' : direct self-join at subquery top level
+  // 
============================================================================
+
+  private def rewriteDirectSelfJoin(
+      projectListOpt: Option[Seq[NamedExpression]],
+      innerJoin: Join): Option[LogicalPlan] = {
+    val innerLeft = innerJoin.left
+    val innerRight = innerJoin.right
+    val innerCond = innerJoin.condition.get
+
+    val parsed = parseSelfJoinCondition(innerCond, innerLeft, innerRight)
+    if (parsed.isEmpty) return None
+    // parseSelfJoinCondition has validated column correspondence and equi-key 
uniqueness.
+    val (equiPairs, neqPairs) = parsed.get
+
+    val innerLeftEquiAttrs: Seq[Attribute] = equiPairs.map(_._1)
+    val innerLeftNeqAttr: Attribute = neqPairs.head._1
+    val filtered = buildAggregateHavingDistinctGt1(innerLeftEquiAttrs, 
innerLeftNeqAttr, innerLeft)
+
+    // Fail-closed on bare-Join subqueries: without a wrapping Project the 
subquery output
+    // is the full self-join output (both sides' columns). Replacing that with
+    // `Project(equiKeys, filtered)` shrinks the output; if the enclosing 
InSubquery
+    // referenced a non-equi column by position, 
`values.zip(sub.output).map(EqualTo.tupled)`
+    // inside RewritePredicateSubquery would build an incorrect semi 
condition. Q95's
+    // subqueries all have an explicit Project wrapper, so this branch does 
not affect it.
+    projectListOpt match {
+      case None =>
+        None
+      case Some(pl) =>
+        canonicalizeWrapper(pl, equiPairs, filtered).map {
+          case (newWrapper, _) =>
+            logDebug(
+              s"Pattern A' - 
equiKeys=[${innerLeftEquiAttrs.map(_.name).mkString(",")}]" +
+                s", neqCol=${innerLeftNeqAttr.name}" +
+                s", 
outCols=[${newWrapper.projectList.map(_.name).mkString(",")}]")
+            newWrapper
+        }
+    }
+  }
+
+  // 
============================================================================
+  //  Pattern A2 : self-join nested inside another InnerJoin in the subquery
+  // 
============================================================================
+
+  private def rewriteNestedSelfJoin(
+      projectListOpt: Option[Seq[NamedExpression]],
+      outerJoin: Join): Option[LogicalPlan] = {
+    val outerCond = outerJoin.condition.get
+
+    val (selfJoinSide, selfJoinOnRight) =
+      tryExtractSelfJoin(outerJoin.right) match {
+        case Some(_) => (outerJoin.right, true)
+        case None =>
+          tryExtractSelfJoin(outerJoin.left) match {
+            case Some(_) => (outerJoin.left, false)
+            case None => return None
+          }
+      }
+
+    val (selfJoinProjectOpt, selfJoin) = selfJoinSide match {
+      case p @ Project(_, j: Join) if j.joinType == Inner && 
j.condition.isDefined =>
+        (Some(p), j)
+      case j: Join if j.joinType == Inner && j.condition.isDefined =>
+        (None, j)
+      case _ => return None
+    }
+
+    val sjLeft = selfJoin.left
+    val sjRight = selfJoin.right
+    val sjCond = selfJoin.condition.get
+    if (!isSameBaseRelation(sjLeft, sjRight)) return None
+
+    val parsed = parseSelfJoinCondition(sjCond, sjLeft, sjRight)
+    if (parsed.isEmpty) return None

Review Comment:
   `tryExtractSelfJoin` does non-trivial work (base-relation check + condition 
parsing), but its result is discarded here and the same subtree is later 
re-matched and re-parsed again. Consider changing `tryExtractSelfJoin` to 
return the extracted wrapper/join (and ideally the parsed `(equiPairs, 
neqPairs)`) and thread that through, so the nested-path only 
canonicalizes/parses once.



##########
backends-velox/src/main/scala/org/apache/gluten/extension/RewriteSelfJoinInequalityToAggregate.scala:
##########
@@ -0,0 +1,674 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.gluten.extension
+
+import org.apache.gluten.config.VeloxConfig
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.RowOrdering
+import org.apache.spark.sql.catalyst.expressions.aggregate._
+import org.apache.spark.sql.catalyst.plans._
+import org.apache.spark.sql.catalyst.plans.logical._
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, 
LogicalRelation}
+import org.apache.spark.sql.execution.datasources.parquet.ParquetFileFormat
+import org.apache.spark.sql.types.LongType
+
+/**
+ * Rewrites self-join with inequality into GROUP BY + HAVING COUNT(DISTINCT) > 
1.
+ *
+ * Targets the two uncorrelated InSubquery shapes exercised by TPC-DS Q95:
+ *
+ *   - Pattern A': the subquery top-level InnerJoin is a direct self-join.
+ *   - Pattern A2: the subquery contains an outer InnerJoin with a self-join 
child; only the
+ *     self-join child is replaced with Aggregate and the outer join is 
preserved.
+ *
+ * Both patterns require an existence-only membership context so row-count 
multiplicity from the
+ * original self-join cross-product does not affect semantics. Correlated 
InSubquery expressions are
+ * intentionally fail-closed because the ExprId remapping performed here does 
not rewrite correlated
+ * predicates.
+ *
+ * Both patterns share:
+ *   - [[buildAggregateHavingDistinctGt1]] to construct `Filter(cnt > 1, 
Aggregate)`
+ *   - [[canonicalizeWrapper]] to rebuild a wrapping Project so every equi-key 
reference points to
+ *     the sjLeft-side attribute, with **fresh exprIds** (Spark's SPARK-21835 
style -- no reuse of
+ *     original exprIds), returning an old->new attribute remap for downstream 
rewrite.
+ *
+ * Controlled by `spark.gluten.sql.rewrite.selfJoinInequality` (default false, 
opt-in).
+ */
+case class RewriteSelfJoinInequalityToAggregate(spark: SparkSession)
+  extends Rule[LogicalPlan]
+  with PredicateHelper
+  with Logging {
+
+  private val CountDistinctAliasName = "_gluten_rw_selfjoin_cnt_distinct"
+
+  override def apply(plan: LogicalPlan): LogicalPlan = {
+    if (!VeloxConfig.get.enableRewriteSelfJoinInequality) {
+      logDebug("RewriteSelfJoinInequalityToAggregate: disabled via config, 
skipping")
+      return plan
+    }
+
+    // Pattern A' / A2: rewrite uncorrelated InSubquery plans.
+    // Correlated subqueries carry outer references / correlated join 
conditions in
+    // `SubqueryExpression.children`; fail closed because this rule does not 
remap them.
+    val rewritten = plan.transformAllExpressions {
+      case in @ InSubquery(_, lq: ListQuery) if lq.children.isEmpty =>
+        rewriteSubqueryPlan(lq.plan) match {
+          case Some(newSub) => in.copy(query = lq.copy(plan = newSub))
+          case None => in
+        }
+    }
+    if (!(rewritten eq plan)) {
+      logDebug(
+        "RewriteSelfJoinInequalityToAggregate: rewrote self-join to " +
+          "GROUP BY + HAVING COUNT(DISTINCT) > 1")
+    }
+    rewritten
+  }
+  // 
============================================================================
+  //  Shared helpers
+  // 
============================================================================
+
+  /**
+   * Build `Filter(cnt > 1, Aggregate(equiKeys, [equiKeys, cnt_alias], 
Filter(IsNotNull(equiKeys),
+   * child)))`. Returns the Filter node whose output is `equiKeys ++ 
[count_alias_attr]`.
+   *
+   * The extra `IsNotNull(equiKeys)` filter is essential to preserve the 
original equi-join's NULL
+   * semantics. Under SQL 3VL, `left.k = right.k` never matches when either 
side is NULL, so the
+   * original self-join drops rows with NULL equi-keys. Aggregate, in 
contrast, groups NULL keys
+   * together into a single "NULL group" -- if that group has >= 2 distinct 
non-null neq values,
+   * COUNT(DISTINCT) > 1 fires and injects NULL into the subquery output. That 
leaked NULL then
+   * turns `NOT IN` into a spurious empty result (Spark's null-aware anti-join 
uses
+   * `Or(equi, IsNull(equi))` which any NULL sub-row satisfies) and can flip 
IN/NOT IN outcomes. The
+   * neq column needs no such filter: `COUNT(DISTINCT col)` already ignores 
NULL.
+   */
+  private def buildAggregateHavingDistinctGt1(
+      equiKeys: Seq[Attribute],
+      neqCol: Attribute,
+      child: LogicalPlan): LogicalPlan = {
+    val countExpr = AggregateExpression(
+      Count(Seq(neqCol)),
+      mode = Complete,
+      isDistinct = true,
+      filter = None,
+      NamedExpression.newExprId)
+    val countAlias = Alias(countExpr, CountDistinctAliasName)()
+    // Seq[Attribute] is a Seq[NamedExpression] via covariance; no cast needed.
+    val aggExprs: Seq[NamedExpression] = equiKeys :+ countAlias
+    val nonNullChild = equiKeys
+      .map(a => IsNotNull(a): Expression)
+      .reduceOption(And)
+      .map(Filter(_, child))
+      .getOrElse(child)
+    val agg = Aggregate(equiKeys, aggExprs, nonNullChild)
+    Filter(GreaterThan(countAlias.toAttribute, Literal(1L, LongType)), agg)
+  }
+
+  /**
+   * Canonicalize a Project so every equi-key reference points at the 
sjLeft-side attribute.
+   * [[parseSelfJoinCondition]] has already verified that each pair refers to 
the same output
+   * position on the two structurally identical self-join sides. Uses **fresh 
exprIds** (no reuse of
+   * original wrapper output exprIds) -- the same technique Spark's own 
`dedupSubqueryOnSelfJoin`
+   * uses when it needs to change subquery output.
+   *
+   * Returns the rebuilt Project and a map `oldWrapperOutputExprId -> 
newWrapperOutputAttr`, so
+   * downstream references (outer join condition, top-level Project) can be 
updated consistently.
+   *
+   * `equiPairs` provides the definitive ExprId-based lookup: `equiPair (l, 
r)` binds
+   * `l.exprId -> l` (identity) and `r.exprId -> l` (sjRight -> sjLeft). 
Attribute identity in
+   * Catalyst is ExprId, not name; two columns can share a name with distinct 
ExprIds. Name-based
+   * lookup would silently drop such entries via `.toMap`.
+   *
+   * Fails (returns None) when a projectList entry is neither an equi-key 
Attribute (by ExprId) nor
+   * `Alias(equi-key Attribute, _)`. Fail-closed.
+   */
+  private def canonicalizeWrapper(
+      projectList: Seq[NamedExpression],
+      equiPairs: Seq[(Attribute, Attribute)],
+      newChild: LogicalPlan): Option[(Project, Map[ExprId, Attribute])] = {
+    // ExprId-based canonical map: any equi-key attribute (either side) -> 
sjLeft attribute.
+    val exprIdToLeft: Map[ExprId, Attribute] =
+      equiPairs.flatMap { case (l, r) => Seq(l.exprId -> l, r.exprId -> l) 
}.toMap
+    val oldOutput: Seq[Attribute] = projectList.map(_.toAttribute)
+    val mapped: Seq[Option[NamedExpression]] = projectList.map {
+      case a: Attribute if exprIdToLeft.contains(a.exprId) =>
+        // Wrap every rewritten output slot in a fresh Alias.
+        //
+        // When a wrapper reprojects BOTH sides of the same equi pair (e.g.
+        // `SELECT s1.k, s2.k FROM T s1 JOIN T s2 ON s1.k = s2.k AND s1.v <> 
s2.v`),
+        // both entries collapse to the same sjLeft Attribute after the 
self-join is
+        // rewritten. Duplicate output ExprIds are not illegal in Spark 
(`SELECT a, a`
+        // is a valid Project), but fresh Aliases give each output slot an 
independent
+        // identity, which keeps the `oldOutput -> newOutput` remap 1-to-1 and 
lets
+        // downstream references (outer join condition, top-level Project) be 
updated
+        // unambiguously via ExprId.
+        //
+        // The fresh ExprId is on the Alias ITSELF; the referenced child keeps 
its
+        // original ExprId. Spark's logical-plan integrity checks reject 
reusing a
+        // referenced ExprId as the Alias's own ExprId, not duplication across 
slots.
+        Some(Alias(exprIdToLeft(a.exprId), a.name)(): NamedExpression)
+      case al @ Alias(a: Attribute, _) if exprIdToLeft.contains(a.exprId) =>
+        // Fresh exprId; do NOT reuse `al.exprId`. Reusing another 
expression's exprId
+        // is the pattern that Spark 3.3 flags via structural-integrity checks.
+        Some(Alias(exprIdToLeft(a.exprId), al.name)(): NamedExpression)
+      case _ => None
+    }
+    if (mapped.exists(_.isEmpty)) {
+      None
+    } else {
+      val newProjectList = mapped.flatten
+      val newWrapper = Project(newProjectList, newChild)
+      val newOutput = newWrapper.output
+      val remap: Map[ExprId, Attribute] =
+        oldOutput.zip(newOutput).map { case (o, n) => o.exprId -> n }.toMap
+      Some((newWrapper, remap))
+    }
+  }
+
+  /**
+   * Replace equi-key attribute references inside a NamedExpression according 
to `remap`, while
+   * preserving the NamedExpression shape.
+   *
+   * `Expression.transformUp` returns `Expression`, not `NamedExpression`. We 
avoid a blanket
+   * `asInstanceOf[NamedExpression]` by handling the two shapes that can 
appear in a Project's
+   * `projectList` explicitly: a bare Attribute (whose top-level may itself be 
replaced) and an
+   * Alias (which stays an Alias while its child is transformed). Anything 
else in a projectList --
+   * e.g. computed expressions we don't own -- is passed through unchanged.
+   */
+  private def remapNamedExpressionAttributes(
+      ne: NamedExpression,
+      remap: Map[ExprId, Attribute]): NamedExpression = ne match {
+    case a: Attribute if remap.contains(a.exprId) => remap(a.exprId)
+    case a: Attribute => a
+    case al: Alias =>
+      val newChild = al.child.transformUp {
+        case a: Attribute if remap.contains(a.exprId) => remap(a.exprId)
+      }
+      if (newChild eq al.child) al
+      else Alias(newChild, al.name)(al.exprId, al.qualifier, 
al.explicitMetadata)
+    case other => other
+  }
+
+  // 
============================================================================
+  //  Pattern A' / A2 dispatch (subquery plans of InSubquery)
+  // 
============================================================================
+
+  private def rewriteSubqueryPlan(plan: LogicalPlan): Option[LogicalPlan] = {
+    // Candidate-level nondeterminism guard: reject if ANY node in the whole 
subquery plan
+    // is non-repeatable (Rand, LIMIT-without-ORDER-BY, Sample, Offset, 
streaming). This
+    // catches nondeterminism that has been hoisted above the self-join by an 
earlier
+    // optimizer rule -- the per-side `isSameBaseRelation` check alone would 
miss it because
+    // both innerLeft/innerRight can look deterministic after such a hoist.
+    if (!isRepeatablePlan(plan)) return None
+
+    val (projectListOpt, innerJoin): (Option[Seq[NamedExpression]], Join) = 
plan match {
+      case Project(pl, j: Join) if j.joinType == Inner && 
j.condition.isDefined =>
+        (Some(pl), j)
+      case j: Join if j.joinType == Inner && j.condition.isDefined =>
+        (None, j)
+      case _ => return None
+    }
+
+    if (isSameBaseRelation(innerJoin.left, innerJoin.right)) {
+      rewriteDirectSelfJoin(projectListOpt, innerJoin)
+    } else {
+      rewriteNestedSelfJoin(projectListOpt, innerJoin)
+    }
+  }
+
+  // 
============================================================================
+  //  Pattern A' : direct self-join at subquery top level
+  // 
============================================================================
+
+  private def rewriteDirectSelfJoin(
+      projectListOpt: Option[Seq[NamedExpression]],
+      innerJoin: Join): Option[LogicalPlan] = {
+    val innerLeft = innerJoin.left
+    val innerRight = innerJoin.right
+    val innerCond = innerJoin.condition.get
+
+    val parsed = parseSelfJoinCondition(innerCond, innerLeft, innerRight)
+    if (parsed.isEmpty) return None
+    // parseSelfJoinCondition has validated column correspondence and equi-key 
uniqueness.
+    val (equiPairs, neqPairs) = parsed.get
+
+    val innerLeftEquiAttrs: Seq[Attribute] = equiPairs.map(_._1)
+    val innerLeftNeqAttr: Attribute = neqPairs.head._1
+    val filtered = buildAggregateHavingDistinctGt1(innerLeftEquiAttrs, 
innerLeftNeqAttr, innerLeft)
+
+    // Fail-closed on bare-Join subqueries: without a wrapping Project the 
subquery output
+    // is the full self-join output (both sides' columns). Replacing that with
+    // `Project(equiKeys, filtered)` shrinks the output; if the enclosing 
InSubquery
+    // referenced a non-equi column by position, 
`values.zip(sub.output).map(EqualTo.tupled)`
+    // inside RewritePredicateSubquery would build an incorrect semi 
condition. Q95's
+    // subqueries all have an explicit Project wrapper, so this branch does 
not affect it.
+    projectListOpt match {
+      case None =>
+        None
+      case Some(pl) =>
+        canonicalizeWrapper(pl, equiPairs, filtered).map {
+          case (newWrapper, _) =>
+            logDebug(
+              s"Pattern A' - 
equiKeys=[${innerLeftEquiAttrs.map(_.name).mkString(",")}]" +
+                s", neqCol=${innerLeftNeqAttr.name}" +
+                s", 
outCols=[${newWrapper.projectList.map(_.name).mkString(",")}]")
+            newWrapper

Review Comment:
   Using `isEmpty` + `get` is safe here but makes control flow more brittle and 
verbose. Consider pattern matching on the `Option` (or using `fold`/`map`) to 
keep the success path scoped without `get`.



##########
docs/velox-configuration.md:
##########
@@ -92,6 +92,7 @@ nav_order: 16
 | spark.gluten.sql.columnar.backend.velox.valueStream.dynamicFilter.enabled    
    | 🔄 Dynamic    | false             | Whether to apply dynamic filters 
pushed down from hash probe in the ValueStream (shuffle reader) operator to 
filter rows before they reach the hash join.                                    
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                             
                 |
 | spark.gluten.sql.enable.enhancedFeatures                                     
    | 🔄 Dynamic    | true              | Enable some features including iceberg 
native write and other features.                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                   
                 |
 | spark.gluten.sql.rewrite.castArrayToString                                   
    | 🔄 Dynamic    | true              | When true, rewrite `cast(array as 
String)` to `concat('[', array_join(array, ', ', null), ']')` to allow 
offloading to Velox.                                                            
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                 
                 |
+| spark.gluten.sql.rewrite.selfJoinInequality                                  
    | 🔄 Dynamic    | false             | When true, rewrite supported 
uncorrelated InSubquery self-joins with inequality predicate into GROUP BY + 
HAVING COUNT(DISTINCT) > 1. Currently targets Parquet-backed direct and nested 
self-join shapes exercised by TPC-DS Q95. Opt-in default (false) until the 
rewrite has been exercised more broadly across workloads.                       
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                                                
                                                      
                 |

Review Comment:
   This is user-facing documentation, but it uses the Catalyst internal name 
`InSubquery`. Consider rephrasing to SQL terminology (e.g., “uncorrelated `IN 
(subquery)` self-joins”) and formatting SQL fragments with backticks for 
readability/consistency in the docs table.



##########
backends-velox/src/main/scala/org/apache/gluten/extension/RewriteSelfJoinInequalityToAggregate.scala:
##########
@@ -0,0 +1,674 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.gluten.extension
+
+import org.apache.gluten.config.VeloxConfig
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.RowOrdering
+import org.apache.spark.sql.catalyst.expressions.aggregate._
+import org.apache.spark.sql.catalyst.plans._
+import org.apache.spark.sql.catalyst.plans.logical._
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, 
LogicalRelation}
+import org.apache.spark.sql.execution.datasources.parquet.ParquetFileFormat
+import org.apache.spark.sql.types.LongType
+
+/**
+ * Rewrites self-join with inequality into GROUP BY + HAVING COUNT(DISTINCT) > 
1.
+ *
+ * Targets the two uncorrelated InSubquery shapes exercised by TPC-DS Q95:
+ *
+ *   - Pattern A': the subquery top-level InnerJoin is a direct self-join.
+ *   - Pattern A2: the subquery contains an outer InnerJoin with a self-join 
child; only the
+ *     self-join child is replaced with Aggregate and the outer join is 
preserved.
+ *
+ * Both patterns require an existence-only membership context so row-count 
multiplicity from the
+ * original self-join cross-product does not affect semantics. Correlated 
InSubquery expressions are
+ * intentionally fail-closed because the ExprId remapping performed here does 
not rewrite correlated
+ * predicates.
+ *
+ * Both patterns share:
+ *   - [[buildAggregateHavingDistinctGt1]] to construct `Filter(cnt > 1, 
Aggregate)`
+ *   - [[canonicalizeWrapper]] to rebuild a wrapping Project so every equi-key 
reference points to
+ *     the sjLeft-side attribute, with **fresh exprIds** (Spark's SPARK-21835 
style -- no reuse of
+ *     original exprIds), returning an old->new attribute remap for downstream 
rewrite.
+ *
+ * Controlled by `spark.gluten.sql.rewrite.selfJoinInequality` (default false, 
opt-in).
+ */
+case class RewriteSelfJoinInequalityToAggregate(spark: SparkSession)
+  extends Rule[LogicalPlan]
+  with PredicateHelper
+  with Logging {
+
+  private val CountDistinctAliasName = "_gluten_rw_selfjoin_cnt_distinct"
+
+  override def apply(plan: LogicalPlan): LogicalPlan = {
+    if (!VeloxConfig.get.enableRewriteSelfJoinInequality) {
+      logDebug("RewriteSelfJoinInequalityToAggregate: disabled via config, 
skipping")
+      return plan
+    }
+
+    // Pattern A' / A2: rewrite uncorrelated InSubquery plans.
+    // Correlated subqueries carry outer references / correlated join 
conditions in
+    // `SubqueryExpression.children`; fail closed because this rule does not 
remap them.
+    val rewritten = plan.transformAllExpressions {
+      case in @ InSubquery(_, lq: ListQuery) if lq.children.isEmpty =>
+        rewriteSubqueryPlan(lq.plan) match {
+          case Some(newSub) => in.copy(query = lq.copy(plan = newSub))
+          case None => in
+        }
+    }
+    if (!(rewritten eq plan)) {
+      logDebug(
+        "RewriteSelfJoinInequalityToAggregate: rewrote self-join to " +
+          "GROUP BY + HAVING COUNT(DISTINCT) > 1")
+    }
+    rewritten
+  }
+  // 
============================================================================
+  //  Shared helpers
+  // 
============================================================================
+
+  /**
+   * Build `Filter(cnt > 1, Aggregate(equiKeys, [equiKeys, cnt_alias], 
Filter(IsNotNull(equiKeys),
+   * child)))`. Returns the Filter node whose output is `equiKeys ++ 
[count_alias_attr]`.
+   *
+   * The extra `IsNotNull(equiKeys)` filter is essential to preserve the 
original equi-join's NULL
+   * semantics. Under SQL 3VL, `left.k = right.k` never matches when either 
side is NULL, so the
+   * original self-join drops rows with NULL equi-keys. Aggregate, in 
contrast, groups NULL keys
+   * together into a single "NULL group" -- if that group has >= 2 distinct 
non-null neq values,
+   * COUNT(DISTINCT) > 1 fires and injects NULL into the subquery output. That 
leaked NULL then
+   * turns `NOT IN` into a spurious empty result (Spark's null-aware anti-join 
uses
+   * `Or(equi, IsNull(equi))` which any NULL sub-row satisfies) and can flip 
IN/NOT IN outcomes. The
+   * neq column needs no such filter: `COUNT(DISTINCT col)` already ignores 
NULL.
+   */
+  private def buildAggregateHavingDistinctGt1(
+      equiKeys: Seq[Attribute],
+      neqCol: Attribute,
+      child: LogicalPlan): LogicalPlan = {
+    val countExpr = AggregateExpression(
+      Count(Seq(neqCol)),
+      mode = Complete,
+      isDistinct = true,
+      filter = None,
+      NamedExpression.newExprId)
+    val countAlias = Alias(countExpr, CountDistinctAliasName)()
+    // Seq[Attribute] is a Seq[NamedExpression] via covariance; no cast needed.
+    val aggExprs: Seq[NamedExpression] = equiKeys :+ countAlias
+    val nonNullChild = equiKeys
+      .map(a => IsNotNull(a): Expression)
+      .reduceOption(And)
+      .map(Filter(_, child))
+      .getOrElse(child)
+    val agg = Aggregate(equiKeys, aggExprs, nonNullChild)
+    Filter(GreaterThan(countAlias.toAttribute, Literal(1L, LongType)), agg)
+  }
+
+  /**
+   * Canonicalize a Project so every equi-key reference points at the 
sjLeft-side attribute.
+   * [[parseSelfJoinCondition]] has already verified that each pair refers to 
the same output
+   * position on the two structurally identical self-join sides. Uses **fresh 
exprIds** (no reuse of
+   * original wrapper output exprIds) -- the same technique Spark's own 
`dedupSubqueryOnSelfJoin`
+   * uses when it needs to change subquery output.
+   *
+   * Returns the rebuilt Project and a map `oldWrapperOutputExprId -> 
newWrapperOutputAttr`, so
+   * downstream references (outer join condition, top-level Project) can be 
updated consistently.
+   *
+   * `equiPairs` provides the definitive ExprId-based lookup: `equiPair (l, 
r)` binds
+   * `l.exprId -> l` (identity) and `r.exprId -> l` (sjRight -> sjLeft). 
Attribute identity in
+   * Catalyst is ExprId, not name; two columns can share a name with distinct 
ExprIds. Name-based
+   * lookup would silently drop such entries via `.toMap`.
+   *
+   * Fails (returns None) when a projectList entry is neither an equi-key 
Attribute (by ExprId) nor
+   * `Alias(equi-key Attribute, _)`. Fail-closed.
+   */
+  private def canonicalizeWrapper(
+      projectList: Seq[NamedExpression],
+      equiPairs: Seq[(Attribute, Attribute)],
+      newChild: LogicalPlan): Option[(Project, Map[ExprId, Attribute])] = {
+    // ExprId-based canonical map: any equi-key attribute (either side) -> 
sjLeft attribute.
+    val exprIdToLeft: Map[ExprId, Attribute] =
+      equiPairs.flatMap { case (l, r) => Seq(l.exprId -> l, r.exprId -> l) 
}.toMap
+    val oldOutput: Seq[Attribute] = projectList.map(_.toAttribute)
+    val mapped: Seq[Option[NamedExpression]] = projectList.map {
+      case a: Attribute if exprIdToLeft.contains(a.exprId) =>
+        // Wrap every rewritten output slot in a fresh Alias.
+        //
+        // When a wrapper reprojects BOTH sides of the same equi pair (e.g.
+        // `SELECT s1.k, s2.k FROM T s1 JOIN T s2 ON s1.k = s2.k AND s1.v <> 
s2.v`),
+        // both entries collapse to the same sjLeft Attribute after the 
self-join is
+        // rewritten. Duplicate output ExprIds are not illegal in Spark 
(`SELECT a, a`
+        // is a valid Project), but fresh Aliases give each output slot an 
independent
+        // identity, which keeps the `oldOutput -> newOutput` remap 1-to-1 and 
lets
+        // downstream references (outer join condition, top-level Project) be 
updated
+        // unambiguously via ExprId.
+        //
+        // The fresh ExprId is on the Alias ITSELF; the referenced child keeps 
its
+        // original ExprId. Spark's logical-plan integrity checks reject 
reusing a
+        // referenced ExprId as the Alias's own ExprId, not duplication across 
slots.
+        Some(Alias(exprIdToLeft(a.exprId), a.name)(): NamedExpression)
+      case al @ Alias(a: Attribute, _) if exprIdToLeft.contains(a.exprId) =>
+        // Fresh exprId; do NOT reuse `al.exprId`. Reusing another 
expression's exprId
+        // is the pattern that Spark 3.3 flags via structural-integrity checks.
+        Some(Alias(exprIdToLeft(a.exprId), al.name)(): NamedExpression)
+      case _ => None
+    }
+    if (mapped.exists(_.isEmpty)) {
+      None
+    } else {
+      val newProjectList = mapped.flatten
+      val newWrapper = Project(newProjectList, newChild)
+      val newOutput = newWrapper.output
+      val remap: Map[ExprId, Attribute] =
+        oldOutput.zip(newOutput).map { case (o, n) => o.exprId -> n }.toMap
+      Some((newWrapper, remap))
+    }
+  }
+
+  /**
+   * Replace equi-key attribute references inside a NamedExpression according 
to `remap`, while
+   * preserving the NamedExpression shape.
+   *
+   * `Expression.transformUp` returns `Expression`, not `NamedExpression`. We 
avoid a blanket
+   * `asInstanceOf[NamedExpression]` by handling the two shapes that can 
appear in a Project's
+   * `projectList` explicitly: a bare Attribute (whose top-level may itself be 
replaced) and an
+   * Alias (which stays an Alias while its child is transformed). Anything 
else in a projectList --
+   * e.g. computed expressions we don't own -- is passed through unchanged.
+   */
+  private def remapNamedExpressionAttributes(
+      ne: NamedExpression,
+      remap: Map[ExprId, Attribute]): NamedExpression = ne match {
+    case a: Attribute if remap.contains(a.exprId) => remap(a.exprId)
+    case a: Attribute => a
+    case al: Alias =>
+      val newChild = al.child.transformUp {
+        case a: Attribute if remap.contains(a.exprId) => remap(a.exprId)
+      }
+      if (newChild eq al.child) al
+      else Alias(newChild, al.name)(al.exprId, al.qualifier, 
al.explicitMetadata)
+    case other => other
+  }
+
+  // 
============================================================================
+  //  Pattern A' / A2 dispatch (subquery plans of InSubquery)
+  // 
============================================================================
+
+  private def rewriteSubqueryPlan(plan: LogicalPlan): Option[LogicalPlan] = {
+    // Candidate-level nondeterminism guard: reject if ANY node in the whole 
subquery plan
+    // is non-repeatable (Rand, LIMIT-without-ORDER-BY, Sample, Offset, 
streaming). This
+    // catches nondeterminism that has been hoisted above the self-join by an 
earlier
+    // optimizer rule -- the per-side `isSameBaseRelation` check alone would 
miss it because
+    // both innerLeft/innerRight can look deterministic after such a hoist.
+    if (!isRepeatablePlan(plan)) return None
+
+    val (projectListOpt, innerJoin): (Option[Seq[NamedExpression]], Join) = 
plan match {
+      case Project(pl, j: Join) if j.joinType == Inner && 
j.condition.isDefined =>
+        (Some(pl), j)
+      case j: Join if j.joinType == Inner && j.condition.isDefined =>
+        (None, j)
+      case _ => return None
+    }
+
+    if (isSameBaseRelation(innerJoin.left, innerJoin.right)) {
+      rewriteDirectSelfJoin(projectListOpt, innerJoin)
+    } else {
+      rewriteNestedSelfJoin(projectListOpt, innerJoin)
+    }
+  }
+
+  // 
============================================================================
+  //  Pattern A' : direct self-join at subquery top level
+  // 
============================================================================
+
+  private def rewriteDirectSelfJoin(
+      projectListOpt: Option[Seq[NamedExpression]],
+      innerJoin: Join): Option[LogicalPlan] = {
+    val innerLeft = innerJoin.left
+    val innerRight = innerJoin.right
+    val innerCond = innerJoin.condition.get
+
+    val parsed = parseSelfJoinCondition(innerCond, innerLeft, innerRight)
+    if (parsed.isEmpty) return None
+    // parseSelfJoinCondition has validated column correspondence and equi-key 
uniqueness.
+    val (equiPairs, neqPairs) = parsed.get
+
+    val innerLeftEquiAttrs: Seq[Attribute] = equiPairs.map(_._1)
+    val innerLeftNeqAttr: Attribute = neqPairs.head._1
+    val filtered = buildAggregateHavingDistinctGt1(innerLeftEquiAttrs, 
innerLeftNeqAttr, innerLeft)
+
+    // Fail-closed on bare-Join subqueries: without a wrapping Project the 
subquery output
+    // is the full self-join output (both sides' columns). Replacing that with
+    // `Project(equiKeys, filtered)` shrinks the output; if the enclosing 
InSubquery
+    // referenced a non-equi column by position, 
`values.zip(sub.output).map(EqualTo.tupled)`
+    // inside RewritePredicateSubquery would build an incorrect semi 
condition. Q95's
+    // subqueries all have an explicit Project wrapper, so this branch does 
not affect it.
+    projectListOpt match {
+      case None =>
+        None
+      case Some(pl) =>
+        canonicalizeWrapper(pl, equiPairs, filtered).map {
+          case (newWrapper, _) =>
+            logDebug(
+              s"Pattern A' - 
equiKeys=[${innerLeftEquiAttrs.map(_.name).mkString(",")}]" +
+                s", neqCol=${innerLeftNeqAttr.name}" +
+                s", 
outCols=[${newWrapper.projectList.map(_.name).mkString(",")}]")
+            newWrapper
+        }
+    }
+  }
+
+  // 
============================================================================
+  //  Pattern A2 : self-join nested inside another InnerJoin in the subquery
+  // 
============================================================================
+
+  private def rewriteNestedSelfJoin(
+      projectListOpt: Option[Seq[NamedExpression]],
+      outerJoin: Join): Option[LogicalPlan] = {
+    val outerCond = outerJoin.condition.get
+
+    val (selfJoinSide, selfJoinOnRight) =
+      tryExtractSelfJoin(outerJoin.right) match {
+        case Some(_) => (outerJoin.right, true)
+        case None =>
+          tryExtractSelfJoin(outerJoin.left) match {
+            case Some(_) => (outerJoin.left, false)
+            case None => return None
+          }
+      }
+
+    val (selfJoinProjectOpt, selfJoin) = selfJoinSide match {
+      case p @ Project(_, j: Join) if j.joinType == Inner && 
j.condition.isDefined =>
+        (Some(p), j)
+      case j: Join if j.joinType == Inner && j.condition.isDefined =>
+        (None, j)
+      case _ => return None
+    }
+
+    val sjLeft = selfJoin.left
+    val sjRight = selfJoin.right
+    val sjCond = selfJoin.condition.get
+    if (!isSameBaseRelation(sjLeft, sjRight)) return None
+
+    val parsed = parseSelfJoinCondition(sjCond, sjLeft, sjRight)
+    if (parsed.isEmpty) return None
+    // parseSelfJoinCondition has validated column correspondence and equi-key 
uniqueness.
+    val (equiPairs, neqPairs) = parsed.get
+
+    val sjLeftEquiAttrs: Seq[Attribute] = equiPairs.map(_._1)
+    val sjLeftNeqAttr: Attribute = neqPairs.head._1
+
+    val selfJoinOutputSet = selfJoinSide.outputSet
+    val sjEquiExprIds: Set[ExprId] =
+      equiPairs.flatMap { case (l, r) => Seq(l.exprId, r.exprId) }.toSet
+    // wrapper Project may reproject equi-keys under fresh alias exprIds; 
include those.
+    val wrapperEquiExprIds: Set[ExprId] = selfJoinProjectOpt.toSeq.flatMap {
+      p =>
+        p.projectList.flatMap {
+          case a: Attribute if sjEquiExprIds.contains(a.exprId) => 
Some(a.exprId)
+          case al @ Alias(a: Attribute, _) if sjEquiExprIds.contains(a.exprId) 
=> Some(al.exprId)
+          case _ => None
+        }
+    }.toSet
+    val allEquiExprIds = sjEquiExprIds ++ wrapperEquiExprIds
+
+    // Outer join condition may reference only equi-key attrs from the 
self-join side.
+    val outerCondRefs = outerCond.references.filter(selfJoinOutputSet.contains)
+    if (!outerCondRefs.forall(a => allEquiExprIds.contains(a.exprId))) return 
None
+
+    // Top-level subquery Project may reference only equi-key attrs from the 
self-join side.
+    val projectOk = projectListOpt.forall {
+      pl =>
+        val refs = pl.flatMap(_.references).filter(selfJoinOutputSet.contains)
+        refs.forall(a => allEquiExprIds.contains(a.exprId))
+    }
+    if (!projectOk) return None
+
+    val filtered = buildAggregateHavingDistinctGt1(sjLeftEquiAttrs, 
sjLeftNeqAttr, sjLeft)
+
+    val (newSelfJoinSide, outputRemap): (LogicalPlan, Map[ExprId, Attribute]) =
+      selfJoinProjectOpt match {
+        case Some(wp) =>
+          canonicalizeWrapper(wp.projectList, equiPairs, filtered) match {
+            case Some((newWrapper, remap)) => (newWrapper, remap)
+            case None => return None
+          }
+        case None if projectListOpt.isEmpty =>
+          // Fail-closed: with neither a wrapper Project around the self-join 
nor a top-level
+          // subquery Project, the outer join currently exposes every 
self-join column, and
+          // replacing the self-join with `Project(equiKeys, filtered)` would 
shrink the outer
+          // join's right-hand output arity. RewritePredicateSubquery's 
positional zip
+          // (`values.zip(sub.output).map(EqualTo.tupled)`) would then bind 
semi predicates to
+          // the wrong attributes -- silently dropping components of a tuple 
IN. A
+          // top-level Project (`projectListOpt`) is what would let the arity 
be preserved
+          // by the top-level rewrite loop; without one, refuse to rewrite.
+          return None
+        case None =>
+          // No wrapper Project but there IS a top-level subquery Project: 
shrinking the outer
+          // join's self-join-side output is safe because the top-level 
Project is rewritten
+          // consistently via `outputRemap` below and the top-level rewrite 
loop ensures
+          // subquery output arity matches what the enclosing InSubquery 
expects.
+          // Outer references may point at sjRight equi-attributes; remap them 
to sjLeft
+          // (same output position in a valid self-join).
+          val newP = Project(sjLeftEquiAttrs, filtered)
+          val remap: Map[ExprId, Attribute] =
+            equiPairs.map { case (l, r) => r.exprId -> l }.toMap
+          (newP, remap)
+      }
+
+    // Rewrite outer join condition to use new wrapper output attributes.
+    val newOuterCond = outerCond.transformUp {
+      case a: Attribute if outputRemap.contains(a.exprId) => 
outputRemap(a.exprId)
+    }
+
+    val newOuterJoin = if (selfJoinOnRight) {
+      outerJoin.copy(right = newSelfJoinSide, condition = Some(newOuterCond))
+    } else {
+      outerJoin.copy(left = newSelfJoinSide, condition = Some(newOuterCond))
+    }
+
+    // Rewrite top-level Project references.
+    val result = projectListOpt match {
+      case Some(pl) =>
+        val newPl = pl.map(ne => remapNamedExpressionAttributes(ne, 
outputRemap))
+        Project(newPl, newOuterJoin)
+      case None => newOuterJoin
+    }
+
+    logDebug(
+      s"Pattern A2 - equiKeys=[${sjLeftEquiAttrs.map(_.name).mkString(",")}]" +
+        s", neqCol=${sjLeftNeqAttr.name}")
+    Some(result)
+  }
+
+  private def tryExtractSelfJoin(plan: LogicalPlan): Option[Join] = {
+    val join = plan match {
+      case Project(_, j: Join) if j.joinType == Inner && j.condition.isDefined 
=> j
+      case j: Join if j.joinType == Inner && j.condition.isDefined => j
+      case _ => return None
+    }
+    if (!isSameBaseRelation(join.left, join.right)) return None
+    val parsed = parseSelfJoinCondition(join.condition.get, join.left, 
join.right)
+    if (parsed.isEmpty) return None
+    Some(join)
+  }
+
+  // 
============================================================================
+  //  parseSelfJoinCondition + isSameBaseRelation
+  // 
============================================================================
+
+  private def outputOrdinal(plan: LogicalPlan, attr: Attribute): Int =
+    plan.output.indexWhere(_.exprId == attr.exprId)
+
+  private def sameOutputPosition(
+      leftPlan: LogicalPlan,
+      rightPlan: LogicalPlan,
+      leftAttr: Attribute,
+      rightAttr: Attribute): Boolean = {
+    val leftPos = outputOrdinal(leftPlan, leftAttr)
+    val rightPos = outputOrdinal(rightPlan, rightAttr)
+    leftPos >= 0 && rightPos >= 0 && leftPos == rightPos
+  }

Review Comment:
   `outputOrdinal` is linear in `plan.output` and is called repeatedly during 
condition parsing/validation. Precomputing `Map[ExprId, Int]` for 
`leftPlan.output` and `rightPlan.output` (once per candidate) would avoid 
repeated scans and keep the check O(1) per predicate.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to