Fix orca preprocess step for query with Select-Project-NaryJoin patte…

…rn (#17423)

Issue:
When a query having pattern of Select-Project-NaryJoin
with Select predicate's condition also containing pattern
ScalarSubquery-Project-Select-Project-NaryJoin
is executed, a crash was happening.

RCA:
During the preprocessing step `PexprTransposeSelectAndProject`,
function `CollapseSelectAndReplaceColref` is called on the expr with the
pattern and Project expr inside the Select's predicate  is being dropped,
but it's column references are not being replaced with equivalent dropped project expr.
That is because `CollapseSelectAndReplaceColref` is called recursively on
the Select's predicate. That function only takes care of removing project but
doesn't re-add the project expr on top of collapsed select.

Fix:
Now function `CollapseSelectAndReplaceColref` is removed.

For creation of collapsed select, 
* First colrefs of the columns projected in the project list are replaced
 with their equivalent Project exprs in the Select's predicate expr.
 This happens in the refactored function `CUtils::ReplaceColrefWithProjectExpr`
* Then the new Select predicate with replaced colrefs and the NaryJoin expr 
 are transposed recursively with `PexprTransposeSelectAndProject`.
* Collapsed Select is created using the transposed NaryJoin and transposed Select predicate.

Collapsed Select is then added as a child to the CLogicalProject creating the new transposed expr.

Input:
+--CLogicalSelect
   |--CLogicalProject
   |  +--CLogicalNAryJoin
   +--...
      +--CLogicalSelect
         +--CLogicalProject
            +--CLogicalNAryJoin

Old Output:
+--CLogicalProject
   |--CLogicalSelect
   |  +--CLogicalNAryJoin
   +--...
         +--CLogicalSelect
            +--CLogicalNAryJoin

Fixed output:
+--CLogicalProject
   |--CLogicalSelect
   |  +--CLogicalNAryJoin
   +--...
      +--CLogicalProject
         +--CLogicalSelect
            +--CLogicalNAryJoin


* New regression testcases are added that tests queries containing the above pattern.
* New mdp testcase is added that tests a query with above pattern.

src/backend/gporca/libgpopt/include/gpopt/base/CUtils.h

@@ -1021,6 +1021,12 @@ class CUtils
 
 	static CTableDescriptorHashSet *RemoveDuplicateMdids(
 		CMemoryPool *mp, CTableDescriptorHashSet *tabdescs);
+
+	static CExpression *ReplaceColrefWithProjectExpr(CMemoryPool *mp,
+													 CExpression *pexpr,
+													 CColRef *pcolref,
+													 CExpression *pprojExpr);
+
 };	// class CUtils
 
 // hash set from expressions

src/backend/gporca/libgpopt/include/gpopt/operators/CExpressionPreprocessor.h

@@ -222,11 +222,6 @@ class CExpressionPreprocessor
 	static CExpression *ConvertSplitUpdateToInPlaceUpdate(CMemoryPool *mp,
 														  CExpression *expr);
 
-	static CExpression *CollapseSelectAndReplaceColref(CMemoryPool *mp,
-													   CExpression *expr,
-													   CColRef *pcolref,
-													   CExpression *pprojExpr);
-
 public:
 	CExpressionPreprocessor() = delete;
 

src/backend/gporca/libgpopt/src/base/CUtils.cpp

@@ -5052,4 +5052,29 @@ CUtils::RemoveDuplicateMdids(CMemoryPool *mp, CTableDescriptorHashSet *tabdescs)
 	mdids->Release();
 	return result;
 }
+
+// Replace column reference with projection expr recursively
+CExpression *
+CUtils::ReplaceColrefWithProjectExpr(CMemoryPool *mp, CExpression *pexpr,
+									 CColRef *pcolref, CExpression *pprojExpr)
+{
+	// replace reference
+	if (pexpr->Pop()->Eopid() == COperator::EopScalarIdent &&
+		CColRef::Equals(CScalarIdent::PopConvert(pexpr->Pop())->Pcr(), pcolref))
+	{
+		pprojExpr->AddRef();
+		return pprojExpr;
+	}
+
+	// recurse to children
+	CExpressionArray *pdrgpexprChildren = GPOS_NEW(mp) CExpressionArray(mp);
+	for (ULONG ul = 0; ul < pexpr->Arity(); ul++)
+	{
+		pdrgpexprChildren->Append(
+			ReplaceColrefWithProjectExpr(mp, (*pexpr)[ul], pcolref, pprojExpr));
+	}
+	COperator *pop = pexpr->Pop();
+	pop->AddRef();
+	return GPOS_NEW(mp) CExpression(mp, pop, pdrgpexprChildren);
+}
 // EOF

src/backend/gporca/libgpopt/src/operators/CExpressionPreprocessor.cpp

@@ -3007,59 +3007,6 @@ CExpressionPreprocessor::PcnstrFromChildPartition(
 	return cnstr;
 }
 
-// Collapse a select over a project and update column reference.
-CExpression *
-CExpressionPreprocessor::CollapseSelectAndReplaceColref(CMemoryPool *mp,
-														CExpression *pexpr,
-														CColRef *pcolref,
-														CExpression *pprojExpr)
-{
-	// remove the logical project
-	//
-	// Input:
-	// +--CLogicalSelect (x = 'meh')
-	//    +--CLogicalProject (col1...n, expr as x)
-	//       +-- CLogicalNAryJoin
-	// Output:
-	// +--CLogicalSelect (expr = 'meh')
-	//    +-- CLogicalNAryJoin
-	if (pexpr->Pop()->Eopid() == COperator::EopLogicalSelect &&
-		(*pexpr)[0]->Pop()->Eopid() == COperator::EopLogicalProject &&
-		(*(*pexpr)[0])[0]->Pop()->Eopid() == COperator::EopLogicalNAryJoin)
-	{
-		(*(*pexpr)[0])[0]->AddRef();
-		CExpression *pexprCollapsedSelect = GPOS_NEW(mp)
-			CExpression(mp, GPOS_NEW(mp) CLogicalSelect(mp), (*(*pexpr)[0])[0],
-						CollapseSelectAndReplaceColref(mp, (*pexpr)[1], pcolref,
-													   pprojExpr));
-
-		CExpression *pexprTransposed =
-			PexprTransposeSelectAndProject(mp, pexprCollapsedSelect);
-		pexprCollapsedSelect->Release();
-		return pexprTransposed;
-	}
-
-	// replace reference
-	if (pexpr->Pop()->Eopid() == COperator::EopScalarIdent &&
-		CColRef::Equals(CScalarIdent::PopConvert(pexpr->Pop())->Pcr(), pcolref))
-	{
-		pprojExpr->AddRef();
-		return pprojExpr;
-	}
-
-	// recurse to children
-	CExpressionArray *pdrgpexprChildren = GPOS_NEW(mp) CExpressionArray(mp);
-	for (ULONG ul = 0; ul < pexpr->Arity(); ul++)
-	{
-		pdrgpexprChildren->Append(CollapseSelectAndReplaceColref(
-			mp, (*pexpr)[ul], pcolref, pprojExpr));
-	}
-
-	COperator *pop = pexpr->Pop();
-	pop->AddRef();
-	return GPOS_NEW(mp) CExpression(mp, pop, pdrgpexprChildren);
-}
-
 // Transpose a select over a project
 //
 // This preprocessing step enables additional opportunities for predicate push
@@ -3139,9 +3086,8 @@ CExpressionPreprocessor::PexprTransposeSelectAndProject(CMemoryPool *mp,
 	{
 		CExpression *pproject = (*pexpr)[0];
 		CExpression *pprojectList = (*pproject)[1];
-		CExpression *pselectNew = pexpr;
-
 		CExpressionArray *pdrgpexpr = GPOS_NEW(mp) CExpressionArray(mp);
+		// Return same pexpr if transpose is not possible
 		for (ULONG ul = 0; ul < pprojectList->Arity(); ul++)
 		{
 			CExpression *pprojexpr =
@@ -3168,25 +3114,66 @@ CExpressionPreprocessor::PexprTransposeSelectAndProject(CMemoryPool *mp,
 				pexpr->AddRef();
 				return pexpr;
 			}
-
-			// TODO: In order to support mixed pushable and non-pushable
-			//       predicates we need to be able to deconstruct a select
-			//       conjunction constraint into pushable and non-pushable
-			//       parts.
-			//
-			//       NB: JoinOnViewWithMixOfPushableAndNonpushablePredicates.mdp
-			CExpression *prevpselectNew = pselectNew;
-			pselectNew = CollapseSelectAndReplaceColref(
-				mp, prevpselectNew,
-				CUtils::PNthProjectElement(pproject, ul)->Pcr(),
-				CUtils::PNthProjectElementExpr(pproject, ul));
-			if (pexpr != prevpselectNew)
-			{
-				prevpselectNew->Release();
-			}
 		}
-		pdrgpexpr->Append(pselectNew);
-
+		// TODO: In order to support mixed pushable and non-pushable
+		//       predicates we need to be able to deconstruct a select
+		//       conjunction constraint into pushable and non-pushable
+		//       parts.
+		//
+		//       NB: JoinOnViewWithMixOfPushableAndNonpushablePredicates.mdp
+
+
+		// Transpose new select containing predicate with updated colrefs and NaryJoin recursively
+		// remove the logical project
+		//
+		// Input of step:
+		// +--CLogicalSelect (x = 'meh')
+		//    +--CLogicalProject (col1...n, expr as x)
+		//       +-- CLogicalNAryJoin
+		// Output of step:
+		// +--CLogicalSelect (expr = 'meh')
+		//    +-- CLogicalNAryJoin
+
+		// Replace colref's of the columns projected in the project list in the Select's predicate expr.
+		CExpression *pexprSelectPred = (*pexpr)[1];
+		pexprSelectPred->AddRef();
+		for (ULONG ul = 0; ul < pprojectList->Arity(); ul++)
+		{
+			CExpression *pprojexpr =
+				CUtils::PNthProjectElementExpr(pproject, ul);
+			gpopt::CColRef *pcolref =
+				CUtils::PNthProjectElement(pproject, ul)->Pcr();
+			CExpression *pexprSelectPredTemp =
+				CUtils::ReplaceColrefWithProjectExpr(mp, pexprSelectPred,
+													 pcolref, pprojexpr);
+			pexprSelectPred->Release();
+			pexprSelectPred = pexprSelectPredTemp;
+		}
+
+		// Transpose Select's Relation expr so it can be used to create collpased select
+		CExpression *pexprSelectRelNew =
+			PexprTransposeSelectAndProject(mp, (*(*pexpr)[0])[0]);
+
+		// Transpose Select's Predicate expr so it can be used to create collpased select
+		CExpression *pexprSelectPredNew =
+			PexprTransposeSelectAndProject(mp, pexprSelectPred);
+		pexprSelectPred->Release();
+
+		// Create collapsed select from updated Select children
+		CExpression *pexprCollapsedSelect =
+			GPOS_NEW(mp) CExpression(mp, GPOS_NEW(mp) CLogicalSelect(mp),
+									 pexprSelectRelNew, pexprSelectPredNew);
+
+		// Re-add the CLogicalProject on top of collapsed select
+		// Input of step:
+		// +--CLogicalSelect (x = 'meh')
+		//    +--CLogicalProject (col1...n, expr as x)
+		//       +-- CLogicalNAryJoin
+		// Output of step:
+		// +--CLogicalProject (col1...n, expr as x)
+		//    +--CLogicalSelect (expr = 'meh')
+		//       +-- CLogicalNAryJoin
+		pdrgpexpr->Append(pexprCollapsedSelect);
 		CExpressionArray *pdrgpprojelems = GPOS_NEW(mp) CExpressionArray(mp);
 		for (ULONG ul = 0; ul < pprojectList->Arity(); ul++)
 		{

src/backend/gporca/server/CMakeLists.txt

@@ -184,7 +184,7 @@ MultipleDampedPredJoinCardinality MultipleIndependentPredJoinCardinality MultiDi
 MultiDistKeyWithOtherPredsJoinCardinality NoDistKeyMultiPredJoinCardinality OneDistKeyMultiPredJoinCardinality
 JoinOnViewWithCastedColumn JoinOnViewWithCastedColumnAndSubqueryInPredicate JoinOnViewWithVolatileColumn
 JoinOnViewWithMixOfPushableAndNonpushablePredicates JoinOnViewWithSetReturningColumn OuterJoinOnViewWithCastedColumn JoinWithSubqueryProjectColumn
-AggSubqCollapseFalseFilter NestedJoinWithCastedColumn PushConstantSelectPredicateThruJoin-1 PushConstantSelectPredicateThruJoin-2
+AggSubqCollapseFalseFilter NestedJoinWithCastedColumn JoinOnViewWithCastedColumnAndJoinWithCastedColumnInPredicate PushConstantSelectPredicateThruJoin-1 PushConstantSelectPredicateThruJoin-2
 PushConstantSelectPredicateThruJoin-3 PushConstantSelectPredicateThruJoin-4 PushConstantSelectPredicateThruJoin-5 PushConstantSelectPredicateThruJoin-6
 PushConstantSelectPredicateThruJoin-7 PushConstantSelectPredicateThruJoin-8 PushConstantSelectPredicateThruJoin-9 PushConstantSelectPredicateThruJoin-10
 PushConstantSelectPredicateThruJoin-11 PushConstantSelectPredicateThruJoin-12 PushConstantSelectPredicateThruJoin-13

src/test/regress/expected/qp_subquery.out

@@ -1918,6 +1918,114 @@ select * from v where not exists (select b from v where b<>2);
 ---+---
 (0 rows)
 
+-- Check that a query having pattern of Select-Project-NaryJoin,
+-- also containing a Select predicate condition with the same pattern nested in a subquery runs
+CREATE TABLE tab1(a TEXT, b TEXT) DISTRIBUTED RANDOMLY;
+INSERT INTO tab1 SELECT i,i FROM GENERATE_SERIES(1,3)i;
+SELECT * FROM (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t1
+WHERE EXISTS
+          (SELECT 1 FROM
+              (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t2
+           WHERE t2.param = t1.param);
+ param 
+-------
+ 1
+ 2
+ 3
+(3 rows)
+
+EXPLAIN (COSTS OFF) SELECT * FROM (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t1
+WHERE EXISTS
+          (SELECT 1 FROM
+              (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t2
+           WHERE t2.param = t1.param);
+                                                       QUERY PLAN                                                       
+------------------------------------------------------------------------------------------------------------------------
+ Gather Motion 3:1  (slice1; segments: 3)
+   ->  Hash Join
+         Hash Cond: (p2.a = p1.a)
+         ->  Redistribute Motion 3:3  (slice2; segments: 3)
+               Hash Key: p2.a
+               ->  Seq Scan on tab1 p2
+         ->  Hash
+               ->  Redistribute Motion 3:3  (slice3; segments: 3)
+                     Hash Key: p1.a
+                     ->  Hash Join
+                           Hash Cond: (btrim(p1.b) = btrim(p1_1.b))
+                           ->  Seq Scan on tab1 p1
+                           ->  Hash
+                                 ->  Broadcast Motion 3:3  (slice4; segments: 3)
+                                       ->  HashAggregate
+                                             Group Key: btrim(p1_1.b)
+                                             ->  Result
+                                                   ->  Redistribute Motion 3:3  (slice5; segments: 3)
+                                                         Hash Key: btrim(p1_1.b)
+                                                         ->  Hash Join
+                                                               Hash Cond: (p1_1.a = p2_1.a)
+                                                               ->  Redistribute Motion 3:3  (slice6; segments: 3)
+                                                                     Hash Key: p1_1.a
+                                                                     ->  Seq Scan on tab1 p1_1
+                                                               ->  Hash
+                                                                     ->  Redistribute Motion 3:3  (slice7; segments: 3)
+                                                                           Hash Key: p2_1.a
+                                                                           ->  Seq Scan on tab1 p2_1
+ Optimizer: Postgres-based planner
+(29 rows)
+
+-- Check that a query having pattern of Select-Project-NaryJoin,
+-- also containing a Select predicate condition with the same pattern nested in a subquery runs when subplan is enforced
+SET optimizer_enforce_subplans TO on;
+SELECT * FROM (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t1
+WHERE EXISTS
+          (SELECT 1 FROM
+              (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t2
+           WHERE t2.param = t1.param);
+ param 
+-------
+ 1
+ 2
+ 3
+(3 rows)
+
+EXPLAIN (COSTS OFF) SELECT * FROM (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t1
+WHERE EXISTS
+          (SELECT 1 FROM
+              (SELECT BTRIM(p1.b) AS param FROM tab1 p1 JOIN tab1 p2 USING(a)) t2
+           WHERE t2.param = t1.param);
+                                                       QUERY PLAN                                                       
+------------------------------------------------------------------------------------------------------------------------
+ Gather Motion 3:1  (slice1; segments: 3)
+   ->  Hash Join
+         Hash Cond: (p2.a = p1.a)
+         ->  Redistribute Motion 3:3  (slice2; segments: 3)
+               Hash Key: p2.a
+               ->  Seq Scan on tab1 p2
+         ->  Hash
+               ->  Redistribute Motion 3:3  (slice3; segments: 3)
+                     Hash Key: p1.a
+                     ->  Hash Join
+                           Hash Cond: (btrim(p1.b) = btrim(p1_1.b))
+                           ->  Seq Scan on tab1 p1
+                           ->  Hash
+                                 ->  Broadcast Motion 3:3  (slice4; segments: 3)
+                                       ->  HashAggregate
+                                             Group Key: btrim(p1_1.b)
+                                             ->  Result
+                                                   ->  Redistribute Motion 3:3  (slice5; segments: 3)
+                                                         Hash Key: btrim(p1_1.b)
+                                                         ->  Hash Join
+                                                               Hash Cond: (p1_1.a = p2_1.a)
+                                                               ->  Redistribute Motion 3:3  (slice6; segments: 3)
+                                                                     Hash Key: p1_1.a
+                                                                     ->  Seq Scan on tab1 p1_1
+                                                               ->  Hash
+                                                                     ->  Redistribute Motion 3:3  (slice7; segments: 3)
+                                                                           Hash Key: p2_1.a
+                                                                           ->  Seq Scan on tab1 p2_1
+ Optimizer: Postgres-based planner
+(29 rows)
+
+reset optimizer_enforce_subplans;
 set client_min_messages='warning';
 drop schema qp_subquery cascade;
 reset optimizer_trace_fallback;