diff --git a/src/backend/catalog/dependency.c src/backend/catalog/dependency.c
index a0334764830..1b5de765768 100644
--- a/src/backend/catalog/dependency.c
+++ src/backend/catalog/dependency.c
@@ -303,6 +303,27 @@ performDeletion(const ObjectAddress *object,
 	Relation	depRel;
 	ObjectAddresses *targetObjects;
 
+	if (flags & PERFORM_DELETION_CONCURRENTLY)
+	{
+		/*
+		 * We must commit our transaction in order to make the first pg_index
+		 * state update visible to other sessions.  If the DROP machinery has
+		 * already performed any other actions (removal of other objects,
+		 * pg_depend entries, etc), the commit would make those actions
+		 * permanent, which would leave us with inconsistent catalog state if
+		 * we fail partway through the following sequence.  Since DROP INDEX
+		 * CONCURRENTLY is restricted to dropping just one index that has no
+		 * dependencies, we should get here before anything's been done ---
+		 * but let's check that to be sure.  We can verify that the current
+		 * transaction has not executed any transactional updates by checking
+		 * that no XID has been assigned.
+		 */
+		if (GetTopTransactionIdIfAny() != InvalidTransactionId)
+			ereport(ERROR,
+					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+					 errmsg("DROP INDEX CONCURRENTLY must be first action in transaction")));
+	}
+
 	/*
 	 * We save some cycles by opening pg_depend just once and passing the
 	 * Relation pointer down to all the recursive deletion steps.
diff --git a/src/backend/catalog/index.c src/backend/catalog/index.c
index d959b7cc819..044f5e52fcf 100644
--- a/src/backend/catalog/index.c
+++ src/backend/catalog/index.c
@@ -52,6 +52,7 @@
 #include "commands/tablecmds.h"
 #include "commands/event_trigger.h"
 #include "commands/trigger.h"
+#include "commands/typecmds.h"
 #include "executor/executor.h"
 #include "miscadmin.h"
 #include "nodes/makefuncs.h"
@@ -145,7 +146,7 @@ static void ResetReindexProcessing(void);
 static void SetReindexPending(List *indexes);
 static void RemoveReindexPending(Oid indexOid);
 static void ResetReindexPending(void);
-
+static Oid IndexTypeCreate(Relation indexRelation);
 
 /*
  * relationHasPrimaryKey
@@ -699,6 +700,110 @@ UpdateIndexRelation(Oid indexoid,
 	heap_freetuple(tuple);
 }
 
+/*
+ * We only need to create reltype for multicolumn user-defined
+ * B-tree indexes that don't have a reltype yet.
+ */
+#define INDEX_NEEDS_RELTYPE(indexRelation, indexInfo) ( \
+	!IsSystemRelation(indexRelation)				\
+	&& indexInfo->ii_NumIndexAttrs > 1				\
+	&& indexInfo->ii_Am == BTREE_AM_OID				\
+	&& indexRelation->rd_rel->reltype == InvalidOid	\
+	&& (!IsBinaryUpgrade || binary_upgrade_next_pg_type_oid != InvalidOid))
+
+/*
+ * IndexTypeCreate
+ *
+ * Create type for specified index.
+ */
+Oid
+IndexTypeCreate(Relation indexRelation)
+{
+	Oid			ownerId = GetUserId();
+	Oid			namespaceId = RelationGetNamespace(indexRelation);
+	Oid			new_array_oid = AssignTypeArrayOid();
+	ObjectAddress new_type_addr;
+	char	   *relarrayname;
+
+	/* Index must not have a reltype yet */
+	Assert(indexRelation->rd_rel->reltype == InvalidOid);
+
+	/*
+	* Build compound type for compound index to be able to use it in statistic.
+	* We need to collect statistic for compound indexes to be able to better
+	* predict selectivity of multicolumn joins.
+	*/
+	new_type_addr = TypeCreate(InvalidOid,
+			RelationGetRelationName(indexRelation),
+			namespaceId,
+			RelationGetRelid(indexRelation),
+			RELKIND_INDEX,
+			ownerId,			/* owner's ID */
+			-1,					/* internal size (varlena) */
+			TYPTYPE_COMPOSITE,	/* type-type (composite) */
+			TYPCATEGORY_COMPOSITE,	/* type-category (ditto) */
+			false,				/* composite types are never preferred */
+			DEFAULT_TYPDELIM,	/* default array delimiter */
+			F_RECORD_IN,		/* input procedure */
+			F_RECORD_OUT,		/* output procedure */
+			F_RECORD_RECV,		/* receive procedure */
+			F_RECORD_SEND,		/* send procedure */
+			InvalidOid,			/* typmodin procedure - none */
+			InvalidOid,			/* typmodout procedure - none */
+			InvalidOid,			/* analyze procedure - default */
+			InvalidOid,			/* array element type - irrelevant */
+			false,				/* this is not an array type */
+			new_array_oid,		/* array type if any */
+			InvalidOid,			/* domain base type - irrelevant */
+			NULL,				/* default value - none */
+			NULL,				/* default binary representation */
+			false,				/* passed by reference */
+			'd',				/* alignment - must be the largest! */
+			'x',				/* fully TOASTable */
+			-1,					/* typmod */
+			0,					/* array dimensions for typBaseType */
+			false,				/* Type NOT NULL */
+			InvalidOid);		/* rowtypes never have a collation */
+
+	relarrayname = makeArrayTypeName(RelationGetRelationName(indexRelation),
+										namespaceId);
+
+	TypeCreate(new_array_oid,	/* force the type's OID to this */
+				relarrayname,	/* Array type name */
+				namespaceId,	/* Same namespace as parent */
+				InvalidOid,		/* Not composite, no relationOid */
+				0,				/* relkind, also N/A here */
+				ownerId,		/* owner's ID */
+				-1,				/* Internal size (varlena) */
+				TYPTYPE_BASE,	/* Not composite - typelem is */
+				TYPCATEGORY_ARRAY,	/* type-category (array) */
+				false,			/* array types are never preferred */
+				DEFAULT_TYPDELIM,	/* default array delimiter */
+				F_ARRAY_IN,		/* array input proc */
+				F_ARRAY_OUT,	/* array output proc */
+				F_ARRAY_RECV,	/* array recv (bin) proc */
+				F_ARRAY_SEND,	/* array send (bin) proc */
+				InvalidOid,		/* typmodin procedure - none */
+				InvalidOid,		/* typmodout procedure - none */
+				F_ARRAY_TYPANALYZE,	/* array analyze procedure */
+				new_type_addr.objectId,	/* array element type - the rowtype */
+				true,			/* yes, this is an array type */
+				InvalidOid,		/* this has no array type */
+				InvalidOid,		/* domain base type - irrelevant */
+				NULL,			/* default value - none */
+				NULL,			/* default binary representation */
+				false,			/* passed by reference */
+				'd',			/* alignment - must be the largest! */
+				'x',			/* fully TOASTable */
+				-1,				/* typmod */
+				0,				/* array dimensions for typBaseType */
+				false,			/* Type NOT NULL */
+				InvalidOid);	/* rowtypes never have a collation */
+
+	pfree(relarrayname);
+
+	return new_type_addr.objectId;
+}
 
 /*
  * index_create
@@ -783,6 +888,7 @@ index_create(Relation heapRelation,
 	bool		concurrent = (flags & INDEX_CREATE_CONCURRENT) != 0;
 	bool		partitioned = (flags & INDEX_CREATE_PARTITIONED) != 0;
 	char		relkind;
+	Oid			new_reltype = InvalidOid;
 
 	/* constraint flags can only be set when a constraint is requested */
 	Assert((constr_flags == 0) ||
@@ -946,6 +1052,10 @@ index_create(Relation heapRelation,
 
 	Assert(indexRelationId == RelationGetRelid(indexRelation));
 
+	/* Create a reltype for index if it is needed */
+	if (INDEX_NEEDS_RELTYPE(indexRelation, indexInfo))
+		new_reltype = IndexTypeCreate(indexRelation);
+
 	/*
 	 * Obtain exclusive lock on it.  Although no other transactions can see it
 	 * until we commit, this prevents deadlock-risk complaints from lock
@@ -959,6 +1069,7 @@ index_create(Relation heapRelation,
 	 *
 	 * XXX should have a cleaner way to create cataloged indexes
 	 */
+	indexRelation->rd_rel->reltype = new_reltype;
 	indexRelation->rd_rel->relowner = heapRelation->rd_rel->relowner;
 	indexRelation->rd_rel->relam = accessMethodObjectId;
 	indexRelation->rd_rel->relhasoids = false;
@@ -1480,6 +1591,7 @@ index_drop(Oid indexId, bool concurrent)
 	Relation	indexRelation;
 	HeapTuple	tuple;
 	bool		hasexprs;
+	bool		remove_statistics;
 	LockRelId	heaprelid,
 				indexrelid;
 	LOCKTAG		heaplocktag;
@@ -1547,24 +1659,6 @@ index_drop(Oid indexId, bool concurrent)
 	 */
 	if (concurrent)
 	{
-		/*
-		 * We must commit our transaction in order to make the first pg_index
-		 * state update visible to other sessions.  If the DROP machinery has
-		 * already performed any other actions (removal of other objects,
-		 * pg_depend entries, etc), the commit would make those actions
-		 * permanent, which would leave us with inconsistent catalog state if
-		 * we fail partway through the following sequence.  Since DROP INDEX
-		 * CONCURRENTLY is restricted to dropping just one index that has no
-		 * dependencies, we should get here before anything's been done ---
-		 * but let's check that to be sure.  We can verify that the current
-		 * transaction has not executed any transactional updates by checking
-		 * that no XID has been assigned.
-		 */
-		if (GetTopTransactionIdIfAny() != InvalidTransactionId)
-			ereport(ERROR,
-					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-					 errmsg("DROP INDEX CONCURRENTLY must be first action in transaction")));
-
 		/*
 		 * Mark index invalid by updating its pg_index entry
 		 */
@@ -1685,6 +1779,16 @@ index_drop(Oid indexId, bool concurrent)
 	if (userIndexRelation->rd_rel->relkind != RELKIND_PARTITIONED_INDEX)
 		RelationDropStorage(userIndexRelation);
 
+	/*
+	 * We might have stored multicolumn statistics for btree indexes. They are
+	 * created only for non-system and non-TOAST indexes, so check only for such
+	 * such indexes.
+	 */
+	remove_statistics =
+		IndexRelationGetNumberOfAttributes(userIndexRelation) > 1 &&
+		userIndexRelation->rd_rel->relam == BTREE_AM_OID &&
+		!IsSystemRelation(userIndexRelation);
+
 	/*
 	 * Close and flush the index's relcache entry, to ensure relcache doesn't
 	 * try to rebuild it while we're deleting catalog entries. We keep the
@@ -1712,10 +1816,10 @@ index_drop(Oid indexId, bool concurrent)
 	heap_close(indexRelation, RowExclusiveLock);
 
 	/*
-	 * if it has any expression columns, we might have stored statistics about
-	 * them.
+	 * if it has any expression columns or whole index stat, we might have
+	 * stored statistics about them.
 	 */
-	if (hasexprs)
+	if (hasexprs || remove_statistics)
 		RemoveStatistics(indexId, 0);
 
 	/*
@@ -2189,6 +2293,14 @@ index_update_stats(Relation rel,
 		dirty = true;
 	}
 
+	/* If index's reltype has been created, update it in pg_class. */
+	// if (rel->rd_rel->relkind == RELKIND_INDEX &&
+	// 	rd_rel->reltype != rel->rd_rel->reltype)
+	// {
+	// 	rd_rel->reltype = rel->rd_rel->reltype;
+	// 	dirty = true;
+	// }
+
 	if (reltuples >= 0)
 	{
 		BlockNumber relpages = RelationGetNumberOfBlocks(rel);
@@ -3547,9 +3659,6 @@ index_set_state_flags(Oid indexId, IndexStateFlagsAction action)
 	HeapTuple	indexTuple;
 	Form_pg_index indexForm;
 
-	/* Assert that current xact hasn't done any transactional updates */
-	Assert(GetTopTransactionIdIfAny() == InvalidTransactionId);
-
 	/* Open pg_index and fetch a writable copy of the index's tuple */
 	pg_index = heap_open(IndexRelationId, RowExclusiveLock);
 
@@ -3736,6 +3845,41 @@ reindex_index(Oid indexId, bool skip_constraint_checks, char persistence,
 	PG_END_TRY();
 	ResetReindexProcessing();
 
+	/*
+	 * We might have stored multicolumn statistics for btree indexes. They are
+	 * created only for non-system and non-TOAST indexes, so check only for such
+	 * such indexes.
+	 */
+	if (indexInfo->ii_NumIndexAttrs > 1 && indexInfo->ii_Am == BTREE_AM_OID &&
+		!IsSystemRelation(iRel))
+		RemoveStatistics(indexId, 0);
+
+	/* Create a reltype for index if it is needed */
+	if (INDEX_NEEDS_RELTYPE(iRel, indexInfo))
+	{
+		Relation	pg_class;
+		HeapTuple	tuple;
+		Form_pg_class rd_rel;
+		Oid			new_reltype;
+
+		new_reltype = IndexTypeCreate(iRel);
+
+		pg_class = heap_open(RelationRelationId, RowExclusiveLock);
+		tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(indexId));
+		if (!HeapTupleIsValid(tuple))
+			elog(ERROR, "could not find tuple for relation %u", indexId);
+		rd_rel = (Form_pg_class) GETSTRUCT(tuple);
+
+		rd_rel->reltype = new_reltype;
+
+		CatalogTupleUpdate(pg_class, &tuple->t_self, tuple);
+
+		heap_freetuple(tuple);
+		heap_close(pg_class, RowExclusiveLock);
+
+		iRel->rd_rel->reltype = new_reltype;
+	}
+
 	/*
 	 * If the index is marked invalid/not-ready/dead (ie, it's from a failed
 	 * CREATE INDEX CONCURRENTLY, or a DROP INDEX CONCURRENTLY failed midway),
diff --git a/src/backend/commands/analyze.c src/backend/commands/analyze.c
index 5d52c94c63d..521907c1132 100644
--- a/src/backend/commands/analyze.c
+++ src/backend/commands/analyze.c
@@ -35,8 +35,11 @@
 #include "commands/vacuum.h"
 #include "executor/executor.h"
 #include "foreign/fdwapi.h"
+#include "funcapi.h"
 #include "miscadmin.h"
 #include "nodes/nodeFuncs.h"
+#include "nodes/makefuncs.h"
+#include "nodes/pg_list.h"
 #include "parser/parse_oper.h"
 #include "parser/parse_relation.h"
 #include "pgstat.h"
@@ -61,6 +64,7 @@
 #include "utils/syscache.h"
 #include "utils/timestamp.h"
 #include "utils/tqual.h"
+#include "utils/typcache.h"
 
 
 /* Per-index data for ANALYZE */
@@ -70,6 +74,7 @@ typedef struct AnlIndexData
 	double		tupleFract;		/* fraction of rows for partial index */
 	VacAttrStats **vacattrstats;	/* index attrs to analyze */
 	int			attr_cnt;
+	bool        multicolumn;    /* Collect compound row statistic for multicolumn index */
 } AnlIndexData;
 
 
@@ -514,6 +519,21 @@ do_analyze_rel(Relation onerel, int options, VacuumParams *params,
 				}
 				thisdata->attr_cnt = tcnt;
 			}
+			else if (indexInfo->ii_NumIndexAttrs > 1 && va_cols == NIL &&
+					 Irel[ind]->rd_rel->reltype != InvalidOid)
+			{
+				/* Collect statistic for multicolumn index for better predicting selectivity of multicolumn joins */
+				RowExpr* row = makeNode(RowExpr);
+				row->row_typeid = Irel[ind]->rd_rel->reltype;
+				row->row_format = COERCE_EXPLICIT_CAST;
+				row->location = -1;
+				row->colnames = NULL;
+				thisdata->vacattrstats = (VacAttrStats **)palloc(sizeof(VacAttrStats *));
+				thisdata->vacattrstats[0] = examine_attribute(Irel[ind], 1, (Node*)row);
+				thisdata->vacattrstats[0]->tupDesc = lookup_type_cache(row->row_typeid, TYPECACHE_TUPDESC)->tupDesc;
+				thisdata->attr_cnt = 1;
+				thisdata->multicolumn = true;
+			}
 		}
 	}
 
@@ -837,28 +857,41 @@ compute_index_stats(Relation onerel, double totalrows,
 							   values,
 							   isnull);
 
-				/*
-				 * Save just the columns we care about.  We copy the values
-				 * into ind_context from the estate's per-tuple context.
-				 */
-				for (i = 0; i < attr_cnt; i++)
+				if (thisdata->multicolumn)
 				{
-					VacAttrStats *stats = thisdata->vacattrstats[i];
-					int			attnum = stats->attr->attnum;
-
-					if (isnull[attnum - 1])
-					{
-						exprvals[tcnt] = (Datum) 0;
-						exprnulls[tcnt] = true;
-					}
-					else
+					/* For multicolumn index construct compound value */
+					VacAttrStats *stats = thisdata->vacattrstats[0];
+					exprvals[tcnt] = HeapTupleGetDatum(heap_form_tuple(stats->tupDesc,
+																	   values,
+																	   isnull));
+					exprnulls[tcnt] = false;
+					tcnt++;
+				}
+				else
+				{
+					/*
+					 * Save just the columns we care about.  We copy the values
+					 * into ind_context from the estate's per-tuple context.
+					 */
+					for (i = 0; i < attr_cnt; i++)
 					{
-						exprvals[tcnt] = datumCopy(values[attnum - 1],
-												   stats->attrtype->typbyval,
-												   stats->attrtype->typlen);
-						exprnulls[tcnt] = false;
+						VacAttrStats *stats = thisdata->vacattrstats[i];
+						int			attnum = stats->attr->attnum;
+
+						if (isnull[attnum - 1])
+						{
+							exprvals[tcnt] = (Datum) 0;
+							exprnulls[tcnt] = true;
+						}
+						else
+						{
+							exprvals[tcnt] = datumCopy(values[attnum - 1],
+													   stats->attrtype->typbyval,
+													   stats->attrtype->typlen);
+							exprnulls[tcnt] = false;
+						}
+						tcnt++;
 					}
-					tcnt++;
 				}
 			}
 		}
@@ -2658,6 +2691,7 @@ compute_scalar_stats(VacAttrStatsP stats,
 		 * histogram won't collapse to empty or a singleton.)
 		 */
 		num_hist = ndistinct - num_mcv;
+
 		if (num_hist > num_bins)
 			num_hist = num_bins + 1;
 		if (num_hist >= 2)
diff --git a/src/backend/optimizer/path/clausesel.c src/backend/optimizer/path/clausesel.c
index f4717942c3a..fa8c5088185 100644
--- a/src/backend/optimizer/path/clausesel.c
+++ src/backend/optimizer/path/clausesel.c
@@ -14,16 +14,31 @@
  */
 #include "postgres.h"
 
+#include "access/genam.h"
+#include "access/htup_details.h"
+#include "catalog/pg_collation.h"
+#include "commands/vacuum.h"
+#include "funcapi.h"
 #include "nodes/makefuncs.h"
+#include "nodes/nodeFuncs.h"
 #include "optimizer/clauses.h"
 #include "optimizer/cost.h"
 #include "optimizer/pathnode.h"
 #include "optimizer/plancat.h"
+#include "optimizer/var.h"
+#include "parser/parsetree.h"
+#include "utils/array.h"
+#include "utils/builtins.h"
 #include "utils/fmgroids.h"
 #include "utils/lsyscache.h"
+#include "utils/rel.h"
 #include "utils/selfuncs.h"
+#include "utils/syscache.h"
+#include "utils/typcache.h"
 #include "statistics/statistics.h"
 
+#define EXHAUSTIVE_IN_SELECTIVITY_THRESHOLD (default_statistics_target/4)
+#define RANGE_IN_SELECTIVITY_THRESHOLD (default_statistics_target/20)
 
 /*
  * Data structure for accumulating info about possible range-query
@@ -43,6 +58,948 @@ static void addRangeClause(RangeQueryClause **rqlist, Node *clause,
 			   bool varonleft, bool isLTsel, Selectivity s2);
 static RelOptInfo *find_single_rel_for_clauses(PlannerInfo *root,
 							List *clauses);
+static bool treat_as_join_clause(Node *clause, RestrictInfo *rinfo,
+					 int varRelid, SpecialJoinInfo *sjinfo);
+
+typedef enum CorrelationKind {
+	CKRestrict = 0,
+	CKIndepend,		/* unknown correlation */
+	CKLikelySelf,	/* Seems, should be close to be correlated, like agg with
+					   self join */
+	CKSelf,			/* 100% correlation because of self join */
+	CKMul			/* product of all CKLikelySelf * CKSelf */
+} CorrelationKind;
+static CorrelationKind get_correlation_kind(PlannerInfo *root, int varRelid,
+											OpExpr* expr);
+
+/*
+ *  Get variabe node. Returns null if node is not a Var node.
+ */
+static inline Var*
+get_var(Node* node)
+{
+	if (IsA(node, RelabelType))
+		node = (Node *) ((RelabelType *) node)->arg;
+
+	return IsA(node, Var) ? (Var*)node : NULL;
+}
+
+/*
+ * Locate compound index which can be used for multicolumn clauses/join.
+ */
+static IndexOptInfo*
+locate_inner_multicolumn_index(PlannerInfo *root, Index varno, List* vars,
+						 int n_clauses,
+						 int **permutation, List **missed_vars, int* n_keys)
+{
+	ListCell		*ilist;
+	RelOptInfo		*rel = find_base_rel(root, varno);
+	IndexOptInfo	*index_opt = NULL;
+	List			*missed_vars_opt = NIL;
+	int				*permutation_opt = NULL;
+	int				n_index_cols_opt = 0;
+	bool			used[INDEX_MAX_KEYS];
+	int				posvars[INDEX_MAX_KEYS];
+
+	*n_keys = 0;
+	*missed_vars = NIL;
+
+	Assert(list_length(vars) >= 1);
+	Assert(list_length(vars) <= n_clauses);
+
+	foreach(ilist, rel->indexlist)
+	{
+		IndexOptInfo	*index = (IndexOptInfo *) lfirst(ilist);
+		ListCell		*vlist;
+		int				i, n_index_cols = 0;
+		List			*missed = NIL;
+		int				*perm = NULL;
+
+		memset(used, 0, sizeof(used));
+		perm = palloc(n_clauses * sizeof(*perm));
+		for(i=0; i<n_clauses; i++)
+			perm[i] = -1;
+
+		i = 0;
+		foreach (vlist, vars)
+		{
+			Var* var = lfirst(vlist);
+			int pos;
+
+			for (pos = 0; pos < index->nkeycolumns; pos++)
+			{
+				if (index->indexkeys[pos] == var->varattno)
+				{
+					if (used[pos])
+						missed = lappend(missed, var);
+					else
+					{
+						used[pos] = true;
+						posvars[pos] = i;
+						perm[i] = pos;
+						n_index_cols++;
+						break;
+					}
+				}
+			}
+
+			/* var isn't found in index columns */
+			if (pos == index->nkeycolumns && !list_member_ptr(missed, var))
+				missed = lappend(missed, var);
+
+			i += 1;
+		}
+
+		if (n_index_cols == 0)
+			continue;
+
+		/* check that found columns are first columns in index */
+		if (index->nkeycolumns != n_index_cols)
+		{
+			int old_n_index_cols = n_index_cols;
+
+			for (i = 0; i < old_n_index_cols; i++)
+			{
+				if (n_index_cols != old_n_index_cols)
+				{
+					/*
+					 * We will use only first n_index_cols columns instead of
+					 * found old_n_index_cols, so, all other columns should be
+					 * added to missed list
+					 */
+					if (used[i])
+					{
+						Var *var = list_nth(vars, posvars[i]);
+
+						missed = lappend(missed, var);
+					}
+				}
+				else if (!used[i])
+				{
+					if (i==0)
+						/* there isn't useful prefix */
+						goto TryNextIndex;
+
+					/* we will use only first i columns, save as new n_index_cols */
+					n_index_cols = i;
+				}
+			}
+		}
+
+		/* found exact match vars - index, immediately return */
+		if (vlist == NULL && list_length(missed) == 0 && n_index_cols == index->nkeycolumns)
+		{
+			*permutation = perm;
+			*n_keys = n_index_cols;
+			return index;
+		}
+
+		/* save partially matched index */
+		if (index_opt == NULL ||
+			n_index_cols > n_index_cols_opt ||
+			(n_index_cols == n_index_cols_opt && index->nkeycolumns < index_opt->nkeycolumns))
+		{
+			index_opt = index;
+			missed_vars_opt = missed;
+			if (permutation_opt)
+				pfree(permutation_opt);
+			permutation_opt = perm;
+			perm = NULL;
+			n_index_cols_opt = n_index_cols;
+		}
+TryNextIndex:
+		if (perm)
+			pfree(perm);
+	}
+
+	if (index_opt)
+	{
+		*missed_vars = list_concat_unique(*missed_vars, missed_vars_opt);
+		*permutation = permutation_opt;
+		*n_keys = n_index_cols_opt;
+	}
+
+	return index_opt;
+}
+
+/*
+ * verify that used vars are leading columns
+ */
+static bool
+check_leading_vars_index(IndexOptInfo *index, int n_vars,
+						 bool used[INDEX_MAX_KEYS])
+{
+	int	i;
+
+	if (index->nkeycolumns == n_vars)
+		return true;
+
+	for(i=0; i<n_vars; i++)
+		if (used[i] == false)
+			return false;
+
+	return true;
+}
+
+
+/*
+ * Locate index which exactly match joins vars
+ */
+static IndexOptInfo*
+locate_outer_multicolumn_index(PlannerInfo *root, Index varno, List* vars,
+							   int *permutation)
+{
+	ListCell   *ilist;
+	RelOptInfo* rel = find_base_rel(root, varno);
+	int n_vars = list_length(vars);
+	bool	used[INDEX_MAX_KEYS];
+	IndexOptInfo *index_opt = NULL;
+
+	Assert(n_vars >= 1);
+
+	foreach(ilist, rel->indexlist)
+	{
+		IndexOptInfo	*index = (IndexOptInfo *) lfirst(ilist);
+		ListCell		*vlist;
+		int				i;
+
+		if (index->nkeycolumns < n_vars)
+			continue;
+
+		memset(used, 0, sizeof(used));
+
+		i = 0;
+		foreach (vlist, vars)
+		{
+			Var* var = lfirst(vlist);
+
+			if (permutation[i] < 0 ||
+				index->nkeycolumns <= permutation[i] ||
+				index->indexkeys[permutation[i]] != var->varattno)
+				break;
+
+			used[i] = true;
+			i += 1;
+		}
+
+		if (vlist == NULL && check_leading_vars_index(index, n_vars, used))
+		{
+			if (index->nkeycolumns == n_vars)
+				/* found exact match vars - index, immediately return */
+				return index;
+			else if (index_opt == NULL ||
+					 index_opt->nkeycolumns > index->nkeycolumns)
+				/* found better candidate - store it */
+				index_opt = index;
+		}
+	}
+
+	return index_opt;
+}
+
+typedef struct InArrayClause
+{
+	ArrayType*	array;
+	Datum*		elems;
+	bool*		nulls;
+	int			index;
+	int			n_elems;
+	int			curr_elem;
+} InArrayClause;
+
+typedef struct TupleIterator
+{
+	Datum	values	[INDEX_MAX_KEYS];
+	bool	isnull[INDEX_MAX_KEYS];
+	int		n_variants;
+	int		i_variant;
+	int		*permutation;
+	List	*in_clauses;
+	bool	isExhaustive;
+} TupleIterator;
+
+static void
+initTupleIterator(TupleIterator *it, List *consts, int *permutation,
+				  List *in_clauses)
+{
+	ListCell	*l;
+	int			i;
+
+	it->n_variants = 1;
+	it->permutation = permutation;
+	it->in_clauses = in_clauses;
+	it->isExhaustive = false;
+	for(i = 0; i < INDEX_MAX_KEYS; i++)
+		it->isnull[i] = true;
+
+	i = 0;
+	foreach (l, consts)
+	{
+		Const* c = (Const*) lfirst(l);
+		int j = permutation[i++];
+
+		if (j<0)
+			continue;
+		it->values[j] = c->constvalue;
+		it->isnull[j] = c->constisnull;
+	}
+
+	foreach (l, in_clauses)
+	{
+		InArrayClause*	iac = (InArrayClause*) lfirst(l);
+		int16			elmlen;
+		bool			elmbyval;
+		char			elmalign;
+
+		get_typlenbyvalalign(iac->array->elemtype,
+							 &elmlen, &elmbyval, &elmalign);
+		deconstruct_array(iac->array, iac->array->elemtype,
+						  elmlen, elmbyval, elmalign,
+						  &iac->elems, &iac->nulls, &iac->n_elems);
+		iac->curr_elem = 0;
+		it->n_variants *= iac->n_elems;
+	}
+
+	if (it->n_variants > EXHAUSTIVE_IN_SELECTIVITY_THRESHOLD)
+	{
+		it->isExhaustive = true;
+		it->n_variants = EXHAUSTIVE_IN_SELECTIVITY_THRESHOLD;
+	}
+
+	it->i_variant = it->n_variants;
+}
+
+static void
+resetTupleIterator(TupleIterator *it)
+{
+	ListCell	*l;
+
+	it->i_variant = it->n_variants;
+
+	foreach (l, it->in_clauses)
+	{
+		InArrayClause*  iac = (InArrayClause*) lfirst(l);
+
+		iac->curr_elem = 0;
+	}
+}
+
+static bool
+getTupleIterator(TupleIterator *it)
+{
+	ListCell	*l;
+	int			carry = 1;
+
+	if (it->i_variant == 0)
+		return false;
+
+	it->i_variant--;
+
+	foreach (l, it->in_clauses)
+	{
+		InArrayClause* iac = (InArrayClause*) lfirst(l);
+		int j = it->permutation[iac->index];
+
+		if (j<0)
+			continue;
+
+		if (it->isExhaustive)
+		{
+			/* use random subset of IN list(s) */
+			iac->curr_elem = random() % iac->n_elems;
+		}
+		else if ((iac->curr_elem += carry) >= iac->n_elems)
+		{
+			iac->curr_elem = 0;
+			carry = 1;
+		}
+		else
+			carry = 0;
+
+		it->values[j] = iac->elems[iac->curr_elem];
+		it->isnull[j] = iac->nulls[iac->curr_elem];
+	}
+
+	return true;
+}
+
+static Selectivity
+estimate_selectivity_by_index(PlannerInfo *root, IndexOptInfo* index,
+							  VariableStatData *vardata,
+							  List *consts, List** missed_vars, int *permutation,
+							  List *in_clauses, int n_keys,
+							  bool *usedEqSel)
+{
+	TupleIterator	it;
+	Selectivity		sum = 0.0;
+	TypeCacheEntry	*typentry;
+	Datum			constant;
+	int				nBins;
+
+	/*
+	 * Assume that two compound types are coherent, so we can use equality
+	 * function from one type to compare it with other type. Use >= and <= range
+	 * definition.
+	 */
+	typentry = lookup_type_cache(vardata->atttype,
+								 TYPECACHE_EQ_OPR | TYPECACHE_TUPDESC);
+	initTupleIterator(&it, consts, permutation, in_clauses);
+
+	/*
+	 * Try to  simplify calculations: if all variants matches to small amount of
+	 * bins histogram the we don't need to check tuples separately, it's enough
+	 * to checck min and max tuples and compute selecivity by range of bins
+	 */
+
+	if (n_keys != index->nkeycolumns &&
+		it.n_variants > RANGE_IN_SELECTIVITY_THRESHOLD)
+	{
+		Datum	constantMax = 0,
+				constantMin = 0;
+		FmgrInfo		opprocLT, opprocGT;
+
+		fmgr_info(F_RECORD_GT, &opprocGT);
+		fmgr_info(F_RECORD_LT, &opprocLT);
+
+		/*
+		 * Find min and max tuples
+		 */
+		while(getTupleIterator(&it))
+		{
+			constant = HeapTupleGetDatum(heap_form_tuple(typentry->tupDesc,
+														 it.values, it.isnull));
+
+			if (constantMax == 0 ||
+				DatumGetBool(FunctionCall2Coll(&opprocGT,
+											   DEFAULT_COLLATION_OID,
+											   constant, constantMax)))
+			{
+				constantMax = constant;
+				if (constantMin != 0)
+					continue;
+			}
+			if (constantMin == 0 ||
+				DatumGetBool(FunctionCall2Coll(&opprocLT,
+											   DEFAULT_COLLATION_OID,
+											   constant, constantMin)))
+			{
+				constantMin = constant;
+			}
+		}
+
+		sum = prefix_record_histogram_selectivity(vardata,
+												  constantMin, constantMax,
+												  n_keys, &nBins);
+
+		if (sum > 0 && nBins <= it.n_variants)
+			/*
+			 * conclude that all  tuples are in the same, rather small, range of
+			 * bins
+			 */
+			goto finish;
+
+		/*
+		 * let try tuples one by one
+		 */
+		sum = 0.0;
+		resetTupleIterator(&it);
+	}
+
+	while(getTupleIterator(&it))
+	{
+		Selectivity	s;
+
+		constant = HeapTupleGetDatum(heap_form_tuple(typentry->tupDesc,
+													 it.values, it.isnull));
+
+		if (n_keys != index->nkeycolumns)
+		{
+			s = prefix_record_histogram_selectivity(vardata,
+													constant, constant,
+													n_keys, &nBins);
+
+			if (s < 0)
+			{
+				/*
+				 * There is no histogram, fallback to single available option
+				 */
+				s = eqconst_selectivity(typentry->eq_opr, vardata,
+											 constant, false, true, false,
+											 n_keys);
+
+				if (usedEqSel)
+					*usedEqSel = true;
+			}
+		}
+		else
+		{
+			s = eqconst_selectivity(typentry->eq_opr, vardata,
+									constant, false, true, false,
+									-1);
+		}
+
+		sum += s - s*sum;
+	}
+
+finish:
+	if (it.isExhaustive)
+		sum *= ((double)(it.n_variants))/EXHAUSTIVE_IN_SELECTIVITY_THRESHOLD;
+
+	return sum;
+}
+
+typedef struct ClauseVarPair
+{
+	Var		*var;
+	int		idx;
+} ClauseVarPair;
+
+static void
+appendCVP(List **cvp, Var *var, int idx)
+{
+	ClauseVarPair	*e;
+
+	e = palloc(sizeof(*e));
+	e->var = var;
+	e->idx = idx;
+
+	*cvp = lappend(*cvp, e);
+}
+
+static bool
+initVarData(IndexOptInfo *index, VariableStatData *vardata)
+{
+	Relation	indexRel = index_open(index->indexoid, AccessShareLock);
+
+	if (!indexRel->rd_rel->reltype)
+	{
+		index_close(indexRel, AccessShareLock);
+
+		return false;
+	}
+
+	memset(vardata, 0, sizeof(*vardata));
+	vardata->isunique = index->unique;
+	vardata->atttype = indexRel->rd_rel->reltype;
+	vardata->rel = index->rel;
+	vardata->acl_ok = true;
+	vardata->statsTuple = SearchSysCache3(STATRELATTINH,
+										  ObjectIdGetDatum(index->indexoid),
+										  Int16GetDatum(1),
+										  BoolGetDatum(false));
+	vardata->freefunc = ReleaseSysCache;
+
+	index_close(indexRel, AccessShareLock);
+
+	if (!HeapTupleIsValid(vardata->statsTuple))
+	{
+		ReleaseVariableStats(*vardata);
+		return false;
+	}
+
+	vardata->sslots = index->sslots;
+
+	return true;
+}
+
+static int
+markEstimatedColumns(Bitmapset **estimatedclauses, List	*pairs,
+					 List	*vars, List	*missed_vars)
+{
+	ListCell	*l;
+	int			n_estimated = 0;
+
+	foreach(l, vars)
+	{
+		Var* var = (Var *) lfirst(l);
+		ListCell	*ll;
+
+		if (list_member_ptr(missed_vars, var))
+			continue;
+
+		foreach(ll, pairs)
+		{
+			ClauseVarPair *cvp=(ClauseVarPair*)lfirst(ll);
+
+			if (cvp->var == var)
+			{
+				*estimatedclauses = bms_add_member(*estimatedclauses, cvp->idx);
+				n_estimated += 1;
+				break;
+			}
+		}
+
+		Assert(ll != NULL);
+	}
+
+	return n_estimated;
+}
+
+#define SET_VARNOS(vn) do {										\
+	if ((vn) != 0)												\
+	{															\
+		if (data[0].varno == 0)									\
+			data[0].varno = (vn);								\
+		else if (data[1].varno == 0 && data[0].varno != (vn))	\
+			data[1].varno = (vn);								\
+	}															\
+} while(0)
+
+#define GET_RELBY_NO(vn)	\
+((data[0].varno == (vn) && (vn) != 0) ? &data[0] : ((data[1].varno == (vn) && (vn) != 0) ? &data[1] : NULL))
+
+#define SET_CURDATA(vn)	((cur = GET_RELBY_NO(vn)) != NULL)
+
+/*
+ * Check if clauses represent multicolumn join with compound indexes available
+ * for both side of comparison of indexed columns of one relation with constant
+ * values. If so, calculates selectivity of compound type comparison and returns
+ * true.
+ */
+static bool
+use_multicolumn_statistic(PlannerInfo *root, List *clauses, int varRelid,
+						  JoinType jointype, SpecialJoinInfo *sjinfo,
+						  Selectivity* restrict_selectivity, Selectivity *join_selectivity,
+						  Bitmapset	**estimatedclauses, CorrelationKind
+						  *correlationKind)
+{
+	ListCell			*l;
+	List*				var_clause_map = NIL;
+	List*				missed_vars = NIL;
+	int					i;
+	int					*permutation = NULL;
+	int					n_estimated = 0;
+	int					n_keys;
+	TypeCacheEntry		*typentry;
+
+	struct	{
+		Index				varno;
+
+		List				*restrictionColumns;
+		List				*restrictionConsts;
+		List				*in_clauses;
+		List				*ineqRestrictionClauses;
+
+		List				*joinColumns;
+
+		IndexOptInfo		*index;
+		VariableStatData	vardata;
+	} data[2], *cur;
+
+	if (list_length(clauses) < 1)
+		return false;
+
+	/*
+	 * Do not use expensive machinery for simple cases, we believe that default
+	 * selectivity estimator works well enough
+	 */
+	if (root->join_rel_list == NIL && root->parent_root == NULL)
+		return false;
+
+	*correlationKind = CKIndepend;
+	memset(data, 0, sizeof(data));
+
+	i=-1;
+	foreach(l, clauses)
+	{
+		Node* clause = (Node *) lfirst(l);
+		RestrictInfo* rinfo = NULL;
+		OpExpr	   *opclause = NULL;
+
+		i++;
+
+		/* do not use already estimated clauses */
+		if (bms_is_member(i, *estimatedclauses))
+			continue;
+
+		if (IsA(clause, RestrictInfo))
+		{
+			rinfo = (RestrictInfo *) clause;
+			if (!rinfo->orclause)
+				clause = (Node*)rinfo->clause;
+		}
+		if (IsA(clause, OpExpr))
+			opclause = (OpExpr*)clause;
+
+		if (IsA(clause, Var)) /* boolean variable */
+		{
+			Var* var1 = (Var*)clause;
+
+			SET_VARNOS(var1->varno);
+			if (SET_CURDATA(var1->varno))
+			{
+				cur->restrictionColumns = lappend(cur->restrictionColumns, var1);
+				appendCVP(&var_clause_map, var1, i);
+				cur->restrictionConsts = lappend(cur->restrictionConsts,
+												 makeBoolConst(true, false));
+			}
+		}
+		else if (IsA(clause, BoolExpr) && ((BoolExpr*)clause)->boolop == NOT_EXPR) /* (NOT bool_expr) */
+		{
+			Node* arg1 = (Node*) linitial( ((BoolExpr*)clause)->args);
+			Var* var1 = get_var(arg1);
+
+			if (var1 == NULL)
+				continue;
+
+			SET_VARNOS(var1->varno);
+			if (SET_CURDATA(var1->varno))
+			{
+				cur->restrictionColumns = lappend(cur->restrictionColumns, var1);
+				appendCVP(&var_clause_map, var1, i);
+				cur->restrictionConsts = lappend(cur->restrictionConsts,
+												 makeBoolConst(false, false));
+			}
+		}
+		else if (IsA(clause, ScalarArrayOpExpr))
+		{
+			ScalarArrayOpExpr* in = (ScalarArrayOpExpr*)clause;
+			Var* var1;
+			Node* arg2;
+			InArrayClause* iac;
+
+			var1 = get_var((Node*)linitial(in->args));
+			arg2 = (Node*) lsecond(in->args);
+
+			if (!in->useOr
+				|| list_length(in->args) != 2
+				|| get_oprrest(in->opno) != F_EQSEL
+				|| var1 == NULL
+				|| !IsA(arg2, Const))
+			{
+				continue;
+			}
+
+			SET_VARNOS(var1->varno);
+			if (SET_CURDATA(var1->varno))
+			{
+				cur->restrictionColumns = lappend(cur->restrictionColumns, var1);
+				appendCVP(&var_clause_map, var1, i);
+				cur->restrictionConsts = lappend(cur->restrictionConsts, arg2);
+
+				iac = (InArrayClause*)palloc(sizeof(InArrayClause));
+				iac->array = (ArrayType*)DatumGetPointer(((Const*)arg2)->constvalue);
+				iac->index = list_length(cur->restrictionConsts) - 1;
+
+				cur->in_clauses = lappend(cur->in_clauses, iac);
+			}
+		}
+		else if (opclause
+				 && list_length(opclause->args) == 2)
+		{
+			int oprrest = get_oprrest(opclause->opno);
+			Node* arg1 = (Node*) linitial(opclause->args);
+			Node* arg2 = (Node*) lsecond(opclause->args);
+			Var* var1 = get_var(arg1);
+			Var* var2 = get_var(arg2);
+
+			if (oprrest == F_EQSEL && treat_as_join_clause((Node*)opclause, NULL, varRelid, sjinfo))
+			{
+				if (var1 == NULL || var2 == NULL || var1->vartype != var2->vartype)
+					continue;
+
+				SET_VARNOS(var1->varno);
+				SET_VARNOS(var2->varno);
+
+				if (var1->varno == data[0].varno && var2->varno == data[1].varno)
+				{
+					data[0].joinColumns = lappend(data[0].joinColumns, var1);
+					appendCVP(&var_clause_map, var1, i);
+					data[1].joinColumns = lappend(data[1].joinColumns, var2);
+					appendCVP(&var_clause_map, var2, i);
+				}
+				else if (var1->varno == data[1].varno && var2->varno == data[0].varno)
+				{
+					data[0].joinColumns = lappend(data[0].joinColumns, var2);
+					appendCVP(&var_clause_map, var2, i);
+					data[1].joinColumns = lappend(data[1].joinColumns, var1);
+					appendCVP(&var_clause_map, var1, i);
+				}
+			}
+			else /* Estimate selectivity for a restriction clause. */
+			{
+				/*
+				 * Give up if it is not equality comparison of variable with
+				 * constant or some other clause is treated as join condition
+				 */
+				if (((var1 == NULL) == (var2 == NULL)))
+					continue;
+
+				if (var1 == NULL)
+				{
+					/* swap var1 and var2 */
+					var1 = var2;
+					arg2 = arg1;
+				}
+
+				SET_VARNOS(var1->varno);
+
+				if (SET_CURDATA(var1->varno))
+				{
+					if ((rinfo && is_pseudo_constant_clause_relids(arg2, rinfo->right_relids))
+						|| (!rinfo && NumRelids(clause) == 1 && is_pseudo_constant_clause(arg2)))
+					{
+						/* Restriction clause with a pseudoconstant . */
+						Node* const_val = estimate_expression_value(root, arg2);
+
+						if (IsA(const_val, Const))
+						{
+							switch (oprrest)
+							{
+								case F_EQSEL:
+									cur->restrictionColumns =
+										lappend(cur->restrictionColumns, var1);
+									cur->restrictionConsts =
+										lappend(cur->restrictionConsts, const_val);
+									appendCVP(&var_clause_map, var1, i);
+									break;
+								case F_SCALARGTSEL:
+								case F_SCALARGESEL:
+								case F_SCALARLTSEL:
+								case F_SCALARLESEL:
+									/*
+									 * We do not consider range predicates now,
+									 * but we can mark them as estimated
+									 * if their variables are covered by index.
+									 */
+									appendCVP(&var_clause_map, var1, i);
+									cur->ineqRestrictionClauses =
+										lappend(cur->ineqRestrictionClauses, var1);
+									break;
+								default:
+									break;
+							}
+						}
+					}
+
+				}
+			}
+		}
+		/* else just skip clause to work with it later in caller */
+	}
+
+	*restrict_selectivity = 1.0;
+	*join_selectivity = 1.0;
+
+	/*
+	 * First, try to estimate selectivity by restrictions
+	 */
+	for(i=0; i<lengthof(data); i++)
+	{
+		cur = &data[i];
+
+		/* compute restriction clauses if applicable */
+		if (cur->varno == 0 || list_length(cur->restrictionColumns) < 1)
+			continue;
+
+		cur->index = locate_inner_multicolumn_index(
+					root, cur->varno, cur->restrictionColumns,
+					list_length(clauses), &permutation, &missed_vars, &n_keys);
+
+		if (cur->index && n_keys > 0 &&
+			initVarData(cur->index, &cur->vardata))
+		{
+			bool	usedEqSel= false;
+
+			*restrict_selectivity *= estimate_selectivity_by_index(
+									root, cur->index, &cur->vardata,
+									cur->restrictionConsts, &missed_vars, permutation,
+									cur->in_clauses, n_keys, &usedEqSel);
+
+			ReleaseVariableStats(cur->vardata);
+
+			/*
+			 * mark inequality clauses as used, see estimate_selectivity_by_index()
+			 */
+			if (usedEqSel)
+			{
+				foreach(l, cur->ineqRestrictionClauses)
+				{
+					Var* var = (Var *) lfirst(l);
+
+					/*
+					 * Note, restrictionColumns will contains extra columns !
+					 */
+					for(i=0; i<cur->index->nkeycolumns; i++)
+						if (cur->index->indexkeys[i] == var->varattno)
+							cur->restrictionColumns =
+								lappend(cur->restrictionColumns, var);
+				}
+			}
+
+			n_estimated +=
+				markEstimatedColumns(estimatedclauses, var_clause_map,
+									 cur->restrictionColumns, missed_vars);
+		}
+
+		if (permutation)
+			pfree(permutation);
+		permutation = NULL;
+	}
+
+	/* Deal with join clauses, if possible */
+	if (list_length(data[0].joinColumns) < 1)
+		goto cleanup;
+
+	data[0].index = locate_inner_multicolumn_index(
+						root,
+						data[0].varno, data[0].joinColumns,
+						list_length(clauses), &permutation, &missed_vars, &n_keys);
+
+	if (!data[0].index || n_keys < 1)
+		goto cleanup;
+
+	Assert(permutation != NULL);
+	Assert(data[1].varno != 0);
+	Assert(list_length(data[0].joinColumns) == list_length(data[1].joinColumns));
+
+	data[1].index = locate_outer_multicolumn_index(
+										root,
+										data[1].varno, data[1].joinColumns,
+										permutation);
+
+	if (!data[1].index)
+		goto cleanup;
+
+	if (!initVarData(data[0].index, &data[0].vardata))
+		goto cleanup;
+
+	if (!initVarData(data[1].index, &data[1].vardata))
+	{
+		ReleaseVariableStats(data[0].vardata);
+		goto cleanup;
+	}
+
+	typentry = lookup_type_cache(data[0].vardata.atttype, TYPECACHE_EQ_OPR);
+	*join_selectivity *= eqjoin_selectivity(root, typentry->eq_opr,
+											&data[0].vardata, &data[1].vardata,
+									   sjinfo, n_keys);
+
+	/* for self join */
+	if (data[0].index->indexoid == data[1].index->indexoid)
+		*correlationKind = CKSelf;
+	else
+	{
+		RangeTblEntry *lrte = planner_rt_fetch(data[0].index->rel->relid, root),
+					  *rrte = planner_rt_fetch(data[1].index->rel->relid, root);
+
+		if (lrte->relid == rrte->relid)
+			*correlationKind = CKSelf;
+	}
+
+	for (i = 0; i < lengthof(data); i++)
+		ReleaseVariableStats(data[i].vardata);
+
+	n_estimated +=
+				markEstimatedColumns(estimatedclauses, var_clause_map,
+									 data[0].joinColumns, missed_vars);
+
+cleanup:
+	if (permutation)
+		pfree(permutation);
+
+	return n_estimated != 0;
+}
 
 /****************************************************************************
  *		ROUTINES TO COMPUTE SELECTIVITIES
@@ -95,6 +1052,28 @@ static RelOptInfo *find_single_rel_for_clauses(PlannerInfo *root,
  * Of course this is all very dependent on the behavior of the inequality
  * selectivity functions; perhaps some day we can generalize the approach.
  */
+
+static void
+appendSelectivityRes(Selectivity s[5], Selectivity sel, CorrelationKind ck)
+{
+	switch(ck)
+	{
+		case CKRestrict:
+			s[ck] *= sel;
+			break;
+		case CKSelf:
+		case CKLikelySelf:
+			s[CKMul] *= sel;
+			if (s[ck] > sel)
+				s[ck] = sel;
+		case CKIndepend:
+			s[CKIndepend] *= sel;
+			break;
+		default:
+			elog(ERROR, "unknown selectivity kind: %d", ck);
+	}
+}
+
 Selectivity
 clauselist_selectivity(PlannerInfo *root,
 					   List *clauses,
@@ -102,12 +1081,14 @@ clauselist_selectivity(PlannerInfo *root,
 					   JoinType jointype,
 					   SpecialJoinInfo *sjinfo)
 {
-	Selectivity s1 = 1.0;
+	Selectivity s[5 /* per CorrelationKind */]  = {1.0, 1.0, 1.0, 1.0, 1.0};
+	Selectivity s2 = 1.0, s3 = 1.0;
 	RelOptInfo *rel;
 	Bitmapset  *estimatedclauses = NULL;
 	RangeQueryClause *rqlist = NULL;
 	ListCell   *l;
 	int			listidx;
+	CorrelationKind	ck;
 
 	/*
 	 * If there's exactly one clause, just go directly to
@@ -130,9 +1111,10 @@ clauselist_selectivity(PlannerInfo *root,
 		 * filled with the 0-based list positions of clauses used that way, so
 		 * that we can ignore them below.
 		 */
-		s1 *= dependencies_clauselist_selectivity(root, clauses, varRelid,
+		s2 = dependencies_clauselist_selectivity(root, clauses, varRelid,
 												  jointype, sjinfo, rel,
 												  &estimatedclauses);
+		appendSelectivityRes(s, s2, CKRestrict);
 
 		/*
 		 * This would be the place to apply any other types of extended
@@ -140,12 +1122,25 @@ clauselist_selectivity(PlannerInfo *root,
 		 */
 	}
 
+	/*
+	 * Check if join conjuncts corresponds to some compound indexes on left and
+	 * right joined relations or indexed columns of one relation is compared
+	 * with constant values. In this case selectivity of join can be calculated
+	 * based on statistic of this compound index.
+	 */
+	while(use_multicolumn_statistic(root, clauses, varRelid, jointype, sjinfo,
+									&s2, &s3, &estimatedclauses, &ck))
+	{
+		appendSelectivityRes(s, s2, CKRestrict);
+		appendSelectivityRes(s, s3, ck);
+	}
+
 	/*
 	 * Apply normal selectivity estimates for remaining clauses. We'll be
 	 * careful to skip any clauses which were already estimated above.
 	 *
 	 * Anything that doesn't look like a potential rangequery clause gets
-	 * multiplied into s1 and forgotten. Anything that does gets inserted into
+	 * multiplied into s and forgotten. Anything that does gets inserted into
 	 * an rqlist entry.
 	 */
 	listidx = -1;
@@ -153,7 +1148,6 @@ clauselist_selectivity(PlannerInfo *root,
 	{
 		Node	   *clause = (Node *) lfirst(l);
 		RestrictInfo *rinfo;
-		Selectivity s2;
 
 		listidx++;
 
@@ -178,7 +1172,7 @@ clauselist_selectivity(PlannerInfo *root,
 			rinfo = (RestrictInfo *) clause;
 			if (rinfo->pseudoconstant)
 			{
-				s1 = s1 * s2;
+				appendSelectivityRes(s, s2, CKRestrict);
 				continue;
 			}
 			clause = (Node *) rinfo->clause;
@@ -192,12 +1186,17 @@ clauselist_selectivity(PlannerInfo *root,
 		 * the simple way we are expecting.)  Most of the tests here can be
 		 * done more efficiently with rinfo than without.
 		 */
+		ck = treat_as_join_clause(clause, rinfo, varRelid, sjinfo) ?
+				CKIndepend : CKRestrict;
 		if (is_opclause(clause) && list_length(((OpExpr *) clause)->args) == 2)
 		{
 			OpExpr	   *expr = (OpExpr *) clause;
 			bool		varonleft = true;
 			bool		ok;
 
+			if (ck == CKIndepend)
+				ck = get_correlation_kind(root, varRelid, expr);
+
 			if (rinfo)
 			{
 				ok = (bms_membership(rinfo->clause_relids) == BMS_SINGLETON) &&
@@ -236,7 +1235,7 @@ clauselist_selectivity(PlannerInfo *root,
 						break;
 					default:
 						/* Just merge the selectivity in generically */
-						s1 = s1 * s2;
+						appendSelectivityRes(s, s2, ck);
 						break;
 				}
 				continue;		/* drop to loop bottom */
@@ -244,7 +1243,7 @@ clauselist_selectivity(PlannerInfo *root,
 		}
 
 		/* Not the right form, so treat it generically. */
-		s1 = s1 * s2;
+		appendSelectivityRes(s, s2, ck);
 	}
 
 	/*
@@ -306,15 +1305,13 @@ clauselist_selectivity(PlannerInfo *root,
 				}
 			}
 			/* Merge in the selectivity of the pair of clauses */
-			s1 *= s2;
+			appendSelectivityRes(s, s2, CKRestrict);
 		}
 		else
 		{
 			/* Only found one of a pair, merge it in generically */
-			if (rqlist->have_lobound)
-				s1 *= rqlist->lobound;
-			else
-				s1 *= rqlist->hibound;
+			appendSelectivityRes(s, (rqlist->have_lobound) ? rqlist->lobound :
+								 rqlist->hibound, CKRestrict);
 		}
 		/* release storage and advance */
 		rqnext = rqlist->next;
@@ -322,7 +1319,25 @@ clauselist_selectivity(PlannerInfo *root,
 		rqlist = rqnext;
 	}
 
-	return s1;
+	/* count final selectivity */
+	s2 = s[CKRestrict] * s[CKIndepend];
+
+	if (s[CKIndepend] != s[CKMul])
+	{
+		/* we hahe both independ and correlated - fallback */
+		s2 *= s[CKMul];
+	}
+	else
+	{
+		/* we have only correlated join clauses */
+		if (s[CKLikelySelf] != 1.0 && s2 < s[CKLikelySelf])
+			s2 = s2 + (s[CKLikelySelf] - s2) * 0.25;
+
+		if (s[CKSelf] != 1.0 && s2 < s[CKSelf])
+			s2 = s2 + (s[CKSelf] - s2) * 1.0;
+	}
+
+	return s2;
 }
 
 /*
@@ -531,6 +1546,137 @@ treat_as_join_clause(Node *clause, RestrictInfo *rinfo,
 	}
 }
 
+typedef struct RangeTblEntryContext {
+	RangeTblEntry	*rte;
+	int				count;
+} RangeTblEntryContext;
+
+static bool
+find_rte_walker(Node *node, RangeTblEntryContext *context)
+{
+	if (node == NULL)
+		return false;
+
+	if (context->count > 1)
+		return true; /* skip rest */
+
+	if (IsA(node, RangeTblEntry)) {
+		RangeTblEntry	*rte = (RangeTblEntry*)node;
+
+		if (rte->rtekind == RTE_RELATION)
+		{
+			if (context->count == 0)
+			{
+				context->count++;
+				context->rte=rte;
+			}
+			else if (rte->relid != context->rte->relid)
+			{
+				context->count++;
+				return true; /* more that one relation in subtree */
+			}
+		}
+		else if (!(rte->rtekind == RTE_SUBQUERY || rte->rtekind == RTE_JOIN ||
+				   rte->rtekind == RTE_CTE))
+		{
+			context->count++;
+			return true; /* more that one relation in subtree */
+		}
+
+		return false; /* allow range_table_walker to continue */
+	}
+
+	if (IsA(node, Query))
+		return query_tree_walker((Query *) node, find_rte_walker,
+								 (void *) context, QTW_EXAMINE_RTES);
+
+	return expression_tree_walker(node, find_rte_walker, (void *) context);
+}
+
+static RangeTblEntry*
+find_single_rte(RangeTblEntry *node)
+{
+	RangeTblEntryContext	context;
+
+	context.rte = NULL;
+	context.count = 0;
+
+	(void)range_table_walker(list_make1(node),
+							 find_rte_walker,
+							 (void *) &context, QTW_EXAMINE_RTES);
+
+	return context.count == 1 ? context.rte : NULL;
+}
+
+#define IsSameRelationRTE(a, b)	( \
+	(a)->rtekind == (b)->rtekind && \
+	(a)->rtekind == RTE_RELATION && \
+	(a)->relid == (b)->relid \
+)
+
+
+/*
+ * Any self join or join with aggregation over the same table
+ */
+
+static CorrelationKind
+get_correlation_kind(PlannerInfo *root, int varRelid, OpExpr* expr)
+{
+	Node	*left_arg, *right_arg;
+	Relids	left_varnos, right_varnos;
+	int		left_varno, right_varno;
+	RangeTblEntry	*left_rte, *right_rte;
+
+	if (varRelid != 0)
+		/* We consider only case of joins, not restriction mode */
+		return CKIndepend;
+
+	/* Check if it is equality comparison */
+	if (get_oprrest(expr->opno) != F_EQSEL)
+		return CKIndepend;
+
+	left_arg = linitial(expr->args);
+	right_arg = lsecond(expr->args);
+
+	/*
+	 * Check if it is join of two different relations
+	 */
+	left_varnos = pull_varnos(left_arg);
+	right_varnos = pull_varnos(right_arg);
+	if (!bms_get_singleton_member(left_varnos, &left_varno) ||
+		!bms_get_singleton_member(right_varnos, &right_varno) ||
+		left_varno == right_varno)
+		return CKIndepend;
+
+	left_rte = planner_rt_fetch(left_varno, root);
+	right_rte = planner_rt_fetch(right_varno, root);
+
+	if (IsSameRelationRTE(left_rte, right_rte))
+	{
+		Var *lvar = get_var(left_arg),
+			*rvar = get_var(right_arg);
+
+		/* self join detected, check if it simple a=b clause */
+		if (lvar == NULL || rvar == NULL)
+			return CKLikelySelf;
+		return (lvar->varattno == rvar->varattno) ?
+											CKSelf : CKLikelySelf;
+	}
+
+	if ((left_rte = find_single_rte(left_rte)) == NULL)
+		return CKIndepend;
+	if ((right_rte = find_single_rte(right_rte)) == NULL)
+		return CKIndepend;
+
+	if (IsSameRelationRTE(left_rte, right_rte))
+	{
+		/* self join detected, but over some transformation which cannot be
+		 * flatten */
+		return CKLikelySelf;
+	}
+
+	return CKIndepend;
+}
 
 /*
  * clause_selectivity -
diff --git a/src/backend/optimizer/util/plancat.c src/backend/optimizer/util/plancat.c
index d99d1393fb4..382756bd561 100644
--- a/src/backend/optimizer/util/plancat.c
+++ src/backend/optimizer/util/plancat.c
@@ -421,6 +421,8 @@ get_relation_info(PlannerInfo *root, Oid relationObjectId, bool inhparent,
 
 			index_close(indexRelation, NoLock);
 
+			memset(info->sslots, 0, sizeof(info->sslots));
+
 			indexinfos = lcons(info, indexinfos);
 		}
 
diff --git a/src/backend/tcop/postgres.c src/backend/tcop/postgres.c
index 015d7120472..36ce8817ea8 100644
--- a/src/backend/tcop/postgres.c
+++ src/backend/tcop/postgres.c
@@ -71,6 +71,7 @@
 #include "tcop/pquery.h"
 #include "tcop/tcopprot.h"
 #include "tcop/utility.h"
+#include "utils/builtins.h"
 #include "utils/lsyscache.h"
 #include "utils/memutils.h"
 #include "utils/ps_status.h"
diff --git a/src/backend/utils/adt/rowtypes.c src/backend/utils/adt/rowtypes.c
index 5f729342f8d..0c0acab7c92 100644
--- a/src/backend/utils/adt/rowtypes.c
+++ src/backend/utils/adt/rowtypes.c
@@ -26,7 +26,6 @@
 #include "utils/lsyscache.h"
 #include "utils/typcache.h"
 
-
 /*
  * structure to cache metadata needed for record I/O
  */
@@ -785,6 +784,9 @@ record_cmp(FunctionCallInfo fcinfo)
 {
 	HeapTupleHeader record1 = PG_GETARG_HEAPTUPLEHEADER(0);
 	HeapTupleHeader record2 = PG_GETARG_HEAPTUPLEHEADER(1);
+	int				record_cmp_prefix =
+							(PG_NARGS() == 3 && PG_GETARG_INT32(2) > 0) ?
+									PG_GETARG_INT32(2) : INT_MAX;
 	int			result = 0;
 	Oid			tupType1;
 	Oid			tupType2;
@@ -869,6 +871,9 @@ record_cmp(FunctionCallInfo fcinfo)
 	nulls2 = (bool *) palloc(ncolumns2 * sizeof(bool));
 	heap_deform_tuple(&tuple2, tupdesc2, values2, nulls2);
 
+	ncolumns1 = Min(ncolumns1, record_cmp_prefix);
+	ncolumns2 = Min(ncolumns2, record_cmp_prefix);
+
 	/*
 	 * Scan corresponding columns, allowing for dropped columns in different
 	 * places in the two rows.  i1 and i2 are physical column indexes, j is
@@ -1027,6 +1032,9 @@ record_eq(PG_FUNCTION_ARGS)
 {
 	HeapTupleHeader record1 = PG_GETARG_HEAPTUPLEHEADER(0);
 	HeapTupleHeader record2 = PG_GETARG_HEAPTUPLEHEADER(1);
+	int				record_cmp_prefix =
+							(PG_NARGS() == 3 && PG_GETARG_INT32(2) > 0) ?
+									PG_GETARG_INT32(2) : INT_MAX;
 	bool		result = true;
 	Oid			tupType1;
 	Oid			tupType2;
@@ -1111,6 +1119,9 @@ record_eq(PG_FUNCTION_ARGS)
 	nulls2 = (bool *) palloc(ncolumns2 * sizeof(bool));
 	heap_deform_tuple(&tuple2, tupdesc2, values2, nulls2);
 
+	ncolumns1 = Min(ncolumns1, record_cmp_prefix);
+	ncolumns2 = Min(ncolumns2, record_cmp_prefix);
+
 	/*
 	 * Scan corresponding columns, allowing for dropped columns in different
 	 * places in the two rows.  i1 and i2 are physical column indexes, j is
diff --git a/src/backend/utils/adt/selfuncs.c src/backend/utils/adt/selfuncs.c
index 8592e6cb87c..9e4ffe0e2ad 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ src/backend/utils/adt/selfuncs.c
@@ -164,15 +164,12 @@ static double var_eq_const(VariableStatData *vardata, Oid operator,
 static double var_eq_non_const(VariableStatData *vardata, Oid operator,
 				 Node *other,
 				 bool varonleft, bool negate);
-static double ineq_histogram_selectivity(PlannerInfo *root,
-						   VariableStatData *vardata,
-						   FmgrInfo *opproc, bool isgt, bool iseq,
-						   Datum constval, Oid consttype);
 static double eqjoinsel_inner(Oid operator,
-				VariableStatData *vardata1, VariableStatData *vardata2);
+				VariableStatData *vardata1, VariableStatData *vardata2,
+				int record_cmp_prefix);
 static double eqjoinsel_semi(Oid operator,
 			   VariableStatData *vardata1, VariableStatData *vardata2,
-			   RelOptInfo *inner_rel);
+			   RelOptInfo *inner_rel, int record_cmp_prefix);
 static bool estimate_multivariate_ndistinct(PlannerInfo *root,
 								RelOptInfo *rel, List **varinfos, double *ndistinct);
 static bool convert_to_scalar(Datum value, Oid valuetypid, double *scaledvalue,
@@ -299,6 +296,31 @@ eqsel_internal(PG_FUNCTION_ARGS, bool negate)
 	return selec;
 }
 
+static bool
+get_cached_attstatsslot(AttStatsSlot *sslot, VariableStatData *vardata,
+						int reqkind, Oid reqop, int flags)
+{
+	if (vardata->sslots)
+	{
+		/*
+		 * vardata has somewhere cache
+		 */
+		AttStatsSlot	*sslotp;
+
+		sslotp = fill_attstatsslot(vardata->sslots, vardata->statsTuple,
+								   reqkind, reqop, flags);
+
+		if (sslotp)
+		{
+			*sslot = *sslotp;
+			return true;
+		}
+	}
+
+	return get_attstatsslot(sslot, vardata->statsTuple,
+							reqkind, reqop, flags);
+}
+
 /*
  * var_eq_const --- eqsel for var = const case
  *
@@ -308,6 +330,18 @@ static double
 var_eq_const(VariableStatData *vardata, Oid operator,
 			 Datum constval, bool constisnull,
 			 bool varonleft, bool negate)
+{
+	return eqconst_selectivity(operator, vardata, constval, constisnull,
+				varonleft, negate, -1);
+}
+
+
+Selectivity
+eqconst_selectivity(Oid operator,
+					VariableStatData *vardata,
+					Datum constval, bool constisnull,
+					bool varonleft, bool negate,
+					int	record_cmp_prefix)
 {
 	double		selec;
 	double		nullfrac = 0.0;
@@ -340,7 +374,8 @@ var_eq_const(VariableStatData *vardata, Oid operator,
 	 * different from ours, but it's much more likely to be right than
 	 * ignoring the information.)
 	 */
-	if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+	if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0 &&
+		record_cmp_prefix <= 0)
 	{
 		selec = 1.0 / vardata->rel->tuples;
 	}
@@ -359,9 +394,9 @@ var_eq_const(VariableStatData *vardata, Oid operator,
 		 * don't like this, maybe you shouldn't be using eqsel for your
 		 * operator...)
 		 */
-		if (get_attstatsslot(&sslot, vardata->statsTuple,
-							 STATISTIC_KIND_MCV, InvalidOid,
-							 ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
+		if (get_cached_attstatsslot(&sslot, vardata,
+									STATISTIC_KIND_MCV, InvalidOid,
+									ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
 		{
 			FmgrInfo	eqproc;
 
@@ -371,15 +406,17 @@ var_eq_const(VariableStatData *vardata, Oid operator,
 			{
 				/* be careful to apply operator right way 'round */
 				if (varonleft)
-					match = DatumGetBool(FunctionCall2Coll(&eqproc,
+					match = DatumGetBool(FunctionCall3Coll(&eqproc,
 														   DEFAULT_COLLATION_OID,
 														   sslot.values[i],
-														   constval));
+														   constval,
+														   Int32GetDatum(record_cmp_prefix)));
 				else
-					match = DatumGetBool(FunctionCall2Coll(&eqproc,
+					match = DatumGetBool(FunctionCall3Coll(&eqproc,
 														   DEFAULT_COLLATION_OID,
 														   constval,
-														   sslot.values[i]));
+														   sslot.values[i],
+														   Int32GetDatum(record_cmp_prefix)));
 				if (match)
 					break;
 			}
@@ -605,7 +642,8 @@ scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
 	 */
 	hist_selec = ineq_histogram_selectivity(root, vardata,
 											&opproc, isgt, iseq,
-											constval, consttype);
+											constval, consttype,
+											-1);
 
 	/*
 	 * Now merge the results from the MCV and histogram calculations,
@@ -770,6 +808,121 @@ histogram_selectivity(VariableStatData *vardata, FmgrInfo *opproc,
 	return result;
 }
 
+static int
+prefix_record_histogram_search(AttStatsSlot	*sslot, int start,
+							   Datum constval,  int record_cmp_prefix,
+							   FmgrInfo	*opproc, bool isgt)
+{
+	int			lobound = start;	/* first possible slot to search */
+	int			hibound = sslot->nvalues;	/* last+1 slot to search */
+
+	while (lobound < hibound)
+	{
+		int			probe = (lobound + hibound) / 2;
+		bool		ltcmp;
+
+		ltcmp = DatumGetBool(FunctionCall3Coll(opproc,
+											   DEFAULT_COLLATION_OID,
+											   sslot->values[probe],
+											   constval,
+											   Int32GetDatum(record_cmp_prefix)));
+		if (isgt)
+			ltcmp = !ltcmp;
+		if (ltcmp)
+			lobound = probe + 1;
+		else
+			hibound = probe;
+	}
+
+	return lobound;
+}
+
+/*
+ * Simple function to estimate selctivity by prefix of record, it just counts
+ * number of histogram bins matched by record prefix - similar to
+ * histogram_selectivity() but it knows about sortability of record
+ */
+double
+prefix_record_histogram_selectivity(VariableStatData *vardata,
+									Datum constvalLeft, Datum constvalRight,
+									int record_cmp_prefix, int *n_bins)
+{
+	double		result = -1.0;
+	AttStatsSlot sslot;
+
+	if (HeapTupleIsValid(vardata->statsTuple) &&
+		get_cached_attstatsslot(&sslot, vardata,
+								STATISTIC_KIND_HISTOGRAM, InvalidOid,
+								ATTSTATSSLOT_VALUES))
+	{
+		FmgrInfo	opprocLT, opprocGT;
+		int			start = -1,
+					end = -1;
+
+
+		if (sslot.nvalues > 2)
+		{
+			fmgr_info(F_RECORD_GE, &opprocGT);
+			fmgr_info(F_RECORD_LE, &opprocLT);
+
+			start = prefix_record_histogram_search(&sslot, 0, constvalLeft,
+												   record_cmp_prefix,
+												   &opprocGT, true);
+			if (start < 0)
+				start = 0;
+
+			end   = prefix_record_histogram_search(&sslot, start, constvalRight,
+												   -1,
+												   &opprocLT, false);
+			if (end >= sslot.nvalues)
+				end = sslot.nvalues - 1;
+		}
+		else
+		{
+			fmgr_info(F_RECORD_GT, &opprocGT);
+			fmgr_info(F_RECORD_LE, &opprocLT);
+
+			/*
+			 *  Find first bin which start border is less than constant
+			 */
+			for (start = sslot.nvalues - 1; start >= 0; start--)
+			{
+				if (DatumGetBool(FunctionCall3Coll(&opprocGT,
+												   DEFAULT_COLLATION_OID,
+												   constvalLeft,
+												   sslot.values[start],
+												   Int32GetDatum(record_cmp_prefix))))
+				break;
+			}
+
+			if (start < 0)
+				start=0;
+
+			/*
+			 * Find last bin which end border is less than constant
+			 */
+			for (end = start; end <= sslot.nvalues - 2; end ++)
+			{
+				if (DatumGetBool(FunctionCall3Coll(&opprocLT,
+												   DEFAULT_COLLATION_OID,
+												   constvalRight,
+												   sslot.values[end + 1],
+												   Int32GetDatum(-1))))
+					break;
+			}
+		}
+
+		*n_bins = (start >= end) ? 0 : end - start;
+		result = (start >= end) ? 0.5 :  end - start;
+		result /= ((double) (sslot.nvalues));
+
+		free_attstatsslot(&sslot);
+	}
+
+	return result;
+
+}
+
 /*
  *	ineq_histogram_selectivity	- Examine the histogram for scalarineqsel
  *
@@ -783,11 +936,12 @@ histogram_selectivity(VariableStatData *vardata, FmgrInfo *opproc,
  * null entries.  The caller is expected to combine this result with
  * statistics for those portions of the column population.
  */
-static double
+Selectivity
 ineq_histogram_selectivity(PlannerInfo *root,
 						   VariableStatData *vardata,
 						   FmgrInfo *opproc, bool isgt, bool iseq,
-						   Datum constval, Oid consttype)
+						   Datum constval, Oid consttype,
+						   int record_cmp_prefix)
 {
 	double		hist_selec;
 	AttStatsSlot sslot;
@@ -874,10 +1028,11 @@ ineq_histogram_selectivity(PlannerInfo *root,
 														 NULL,
 														 &sslot.values[probe]);
 
-				ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
+				ltcmp = DatumGetBool(FunctionCall3Coll(opproc,
 													   DEFAULT_COLLATION_OID,
 													   sslot.values[probe],
-													   constval));
+													   constval,
+													   Int32GetDatum(record_cmp_prefix)));
 				if (isgt)
 					ltcmp = !ltcmp;
 				if (ltcmp)
@@ -2301,21 +2456,39 @@ eqjoinsel(PG_FUNCTION_ARGS)
 	JoinType	jointype = (JoinType) PG_GETARG_INT16(3);
 #endif
 	SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
-	double		selec;
 	VariableStatData vardata1;
 	VariableStatData vardata2;
 	bool		join_is_reversed;
-	RelOptInfo *inner_rel;
+	double		selec;
 
 	get_join_variables(root, args, sjinfo,
 					   &vardata1, &vardata2, &join_is_reversed);
 
+	selec =	join_is_reversed
+		? eqjoin_selectivity(root, operator, &vardata2, &vardata1, sjinfo, -1)
+		: eqjoin_selectivity(root, operator, &vardata1, &vardata2, sjinfo, -1);
+
+	ReleaseVariableStats(vardata1);
+	ReleaseVariableStats(vardata2);
+
+	PG_RETURN_FLOAT8((float8)selec);
+}
+
+Selectivity
+eqjoin_selectivity(PlannerInfo *root, Oid operator, VariableStatData* vardata1,
+				   VariableStatData* vardata2, SpecialJoinInfo *sjinfo,
+				   int record_cmp_prefix)
+{
+	Selectivity	selec;
+	RelOptInfo *inner_rel;
+
 	switch (sjinfo->jointype)
 	{
 		case JOIN_INNER:
 		case JOIN_LEFT:
 		case JOIN_FULL:
-			selec = eqjoinsel_inner(operator, &vardata1, &vardata2);
+			selec = eqjoinsel_inner(operator, vardata1, vardata2,
+									record_cmp_prefix);
 			break;
 		case JOIN_SEMI:
 		case JOIN_ANTI:
@@ -2328,13 +2501,8 @@ eqjoinsel(PG_FUNCTION_ARGS)
 			 */
 			inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
 
-			if (!join_is_reversed)
-				selec = eqjoinsel_semi(operator, &vardata1, &vardata2,
-									   inner_rel);
-			else
-				selec = eqjoinsel_semi(get_commutator(operator),
-									   &vardata2, &vardata1,
-									   inner_rel);
+			selec = eqjoinsel_semi(operator, vardata1, vardata2, inner_rel,
+								   record_cmp_prefix);
 			break;
 		default:
 			/* other values not expected here */
@@ -2344,12 +2512,9 @@ eqjoinsel(PG_FUNCTION_ARGS)
 			break;
 	}
 
-	ReleaseVariableStats(vardata1);
-	ReleaseVariableStats(vardata2);
-
 	CLAMP_PROBABILITY(selec);
 
-	PG_RETURN_FLOAT8((float8) selec);
+	return selec;
 }
 
 /*
@@ -2360,7 +2525,8 @@ eqjoinsel(PG_FUNCTION_ARGS)
  */
 static double
 eqjoinsel_inner(Oid operator,
-				VariableStatData *vardata1, VariableStatData *vardata2)
+				VariableStatData *vardata1, VariableStatData *vardata2,
+				int record_cmp_prefix)
 {
 	double		selec;
 	double		nd1;
@@ -2388,9 +2554,9 @@ eqjoinsel_inner(Oid operator,
 		/* note we allow use of nullfrac regardless of security check */
 		stats1 = (Form_pg_statistic) GETSTRUCT(vardata1->statsTuple);
 		if (statistic_proc_security_check(vardata1, opfuncoid))
-			have_mcvs1 = get_attstatsslot(&sslot1, vardata1->statsTuple,
-										  STATISTIC_KIND_MCV, InvalidOid,
-										  ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+			have_mcvs1 = get_cached_attstatsslot(&sslot1, vardata1,
+												 STATISTIC_KIND_MCV, InvalidOid,
+												 ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
 	}
 
 	if (HeapTupleIsValid(vardata2->statsTuple))
@@ -2398,9 +2564,9 @@ eqjoinsel_inner(Oid operator,
 		/* note we allow use of nullfrac regardless of security check */
 		stats2 = (Form_pg_statistic) GETSTRUCT(vardata2->statsTuple);
 		if (statistic_proc_security_check(vardata2, opfuncoid))
-			have_mcvs2 = get_attstatsslot(&sslot2, vardata2->statsTuple,
-										  STATISTIC_KIND_MCV, InvalidOid,
-										  ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+			have_mcvs2 = get_cached_attstatsslot(&sslot2, vardata2,
+												 STATISTIC_KIND_MCV, InvalidOid,
+												 ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
 	}
 
 	if (have_mcvs1 && have_mcvs2)
@@ -2454,10 +2620,11 @@ eqjoinsel_inner(Oid operator,
 			{
 				if (hasmatch2[j])
 					continue;
-				if (DatumGetBool(FunctionCall2Coll(&eqproc,
+				if (DatumGetBool(FunctionCall3Coll(&eqproc,
 												   DEFAULT_COLLATION_OID,
 												   sslot1.values[i],
-												   sslot2.values[j])))
+												   sslot2.values[j],
+												   Int32GetDatum(record_cmp_prefix))))
 				{
 					hasmatch1[i] = hasmatch2[j] = true;
 					matchprodfreq += sslot1.numbers[i] * sslot2.numbers[j];
@@ -2578,7 +2745,7 @@ eqjoinsel_inner(Oid operator,
 static double
 eqjoinsel_semi(Oid operator,
 			   VariableStatData *vardata1, VariableStatData *vardata2,
-			   RelOptInfo *inner_rel)
+			   RelOptInfo *inner_rel, int record_cmp_prefix)
 {
 	double		selec;
 	double		nd1;
@@ -2638,17 +2805,17 @@ eqjoinsel_semi(Oid operator,
 		/* note we allow use of nullfrac regardless of security check */
 		stats1 = (Form_pg_statistic) GETSTRUCT(vardata1->statsTuple);
 		if (statistic_proc_security_check(vardata1, opfuncoid))
-			have_mcvs1 = get_attstatsslot(&sslot1, vardata1->statsTuple,
-										  STATISTIC_KIND_MCV, InvalidOid,
-										  ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+			have_mcvs1 = get_cached_attstatsslot(&sslot1, vardata1,
+												 STATISTIC_KIND_MCV, InvalidOid,
+												 ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
 	}
 
 	if (HeapTupleIsValid(vardata2->statsTuple) &&
 		statistic_proc_security_check(vardata2, opfuncoid))
 	{
-		have_mcvs2 = get_attstatsslot(&sslot2, vardata2->statsTuple,
-									  STATISTIC_KIND_MCV, InvalidOid,
-									  ATTSTATSSLOT_VALUES);
+		have_mcvs2 = get_cached_attstatsslot(&sslot2, vardata2,
+											 STATISTIC_KIND_MCV, InvalidOid,
+											 ATTSTATSSLOT_VALUES);
 		/* note: currently don't need stanumbers from RHS */
 	}
 
@@ -2701,10 +2868,11 @@ eqjoinsel_semi(Oid operator,
 			{
 				if (hasmatch2[j])
 					continue;
-				if (DatumGetBool(FunctionCall2Coll(&eqproc,
+				if (DatumGetBool(FunctionCall3Coll(&eqproc,
 												   DEFAULT_COLLATION_OID,
 												   sslot1.values[i],
-												   sslot2.values[j])))
+												   sslot2.values[j],
+												   Int32GetDatum(record_cmp_prefix))))
 				{
 					hasmatch1[i] = hasmatch2[j] = true;
 					nmatches++;
@@ -6110,7 +6278,8 @@ prefix_selectivity(PlannerInfo *root, VariableStatData *vardata,
 	prefixsel = ineq_histogram_selectivity(root, vardata,
 										   &opproc, true, true,
 										   prefixcon->constvalue,
-										   prefixcon->consttype);
+										   prefixcon->consttype,
+										   -1);
 
 	if (prefixsel < 0.0)
 	{
@@ -6137,7 +6306,8 @@ prefix_selectivity(PlannerInfo *root, VariableStatData *vardata,
 		topsel = ineq_histogram_selectivity(root, vardata,
 											&opproc, false, false,
 											greaterstrcon->constvalue,
-											greaterstrcon->consttype);
+											greaterstrcon->consttype,
+											-1);
 
 		/* ineq_histogram_selectivity worked before, it shouldn't fail now */
 		Assert(topsel >= 0.0);
diff --git a/src/backend/utils/cache/lsyscache.c src/backend/utils/cache/lsyscache.c
index 82d39e94980..88ea9ec74ba 100644
--- a/src/backend/utils/cache/lsyscache.c
+++ src/backend/utils/cache/lsyscache.c
@@ -42,6 +42,7 @@
 #include "utils/datum.h"
 #include "utils/fmgroids.h"
 #include "utils/lsyscache.h"
+#include "utils/memutils.h"
 #include "utils/rel.h"
 #include "utils/syscache.h"
 #include "utils/typcache.h"
@@ -3045,6 +3046,53 @@ get_attstatsslot(AttStatsSlot *sslot, HeapTuple statstuple,
 	return true;
 }
 
+AttStatsSlot*
+fill_attstatsslot(AttStatsSlot *sslots, HeapTuple statstuple,
+				  int reqkind, Oid reqop, int flags)
+{
+	int	add_flags = 0, has_flags = 0;
+	AttStatsSlot	*sslot;
+	MemoryContext	cntx;
+
+	if (reqkind >= STATISTIC_NUM_SLOTS)
+		return NULL;			/* not there */
+
+	sslot = sslots + reqkind;
+
+	if (sslot->values != NULL)
+		has_flags |= ATTSTATSSLOT_VALUES;
+	if (sslot->numbers != NULL)
+		has_flags |= ATTSTATSSLOT_NUMBERS;
+
+	if ((flags & ATTSTATSSLOT_VALUES) && !(has_flags & ATTSTATSSLOT_VALUES))
+		add_flags |= ATTSTATSSLOT_VALUES;
+
+	if ((flags & ATTSTATSSLOT_NUMBERS) && !(has_flags & ATTSTATSSLOT_NUMBERS))
+		add_flags |= ATTSTATSSLOT_NUMBERS;
+
+	if (add_flags == 0 && (reqop == InvalidOid || sslot->staop == reqop))
+		return sslot;
+
+	sslot->incache = false;
+	free_attstatsslot(sslot);
+
+	/* 
+	 * GEQO could call us in short-lived memory context, use rather long-lived
+	 * context to cache statstic data
+	 */
+	cntx = MemoryContextSwitchTo(MessageContext);
+
+	if (get_attstatsslot(sslot, statstuple, reqkind, reqop,
+						 add_flags | has_flags))
+		sslot->incache = true;
+	else
+		sslot = NULL;
+
+	MemoryContextSwitchTo(cntx);
+
+	return sslot;
+}
+
 /*
  * free_attstatsslot
  *		Free data allocated by get_attstatsslot
@@ -3052,6 +3100,10 @@ get_attstatsslot(AttStatsSlot *sslot, HeapTuple statstuple,
 void
 free_attstatsslot(AttStatsSlot *sslot)
 {
+	/* do not free cached slot */
+	if (sslot->incache)
+		return;
+
 	/* The values[] array was separately palloc'd by deconstruct_array */
 	if (sslot->values)
 		pfree(sslot->values);
diff --git a/src/bin/pg_dump/common.c src/bin/pg_dump/common.c
index 9b5869add8b..372aa154f32 100644
--- a/src/bin/pg_dump/common.c
+++ src/bin/pg_dump/common.c
@@ -54,6 +54,7 @@ static DumpableObject **oprinfoindex;
 static DumpableObject **collinfoindex;
 static DumpableObject **nspinfoindex;
 static DumpableObject **extinfoindex;
+static DumpableObject **idxinfoindex;
 static int	numTables;
 static int	numTypes;
 static int	numFuncs;
@@ -61,6 +62,7 @@ static int	numOperators;
 static int	numCollations;
 static int	numNamespaces;
 static int	numExtensions;
+static int	numIndexes;
 
 /* This is an array of object identities, not actual DumpableObjects */
 static ExtensionMemberId *extmembers;
@@ -77,9 +79,8 @@ static int	ExtensionMemberIdCompare(const void *p1, const void *p2);
 static void findParentsByOid(TableInfo *self,
 				 InhInfo *inhinfo, int numInherits);
 static int	strInArray(const char *pattern, char **arr, int arr_size);
-static IndxInfo *findIndexByOid(Oid oid, DumpableObject **idxinfoindex,
-			   int numIndexes);
-
+static IndxInfo *findTableIndexByOid(Oid oid,
+				DumpableObject **idxinfoTableIndex, int numTableIndexes);
 
 /*
  * getSchemaData
@@ -96,6 +97,7 @@ getSchemaData(Archive *fout, int *numTablesPtr)
 	NamespaceInfo *nspinfo;
 	ExtensionInfo *extinfo;
 	InhInfo    *inhinfo;
+	IndxInfo   *idxinfo;
 	int			numAggregates;
 	int			numInherits;
 	int			numRules;
@@ -258,7 +260,8 @@ getSchemaData(Archive *fout, int *numTablesPtr)
 
 	if (g_verbose)
 		write_msg(NULL, "reading indexes\n");
-	getIndexes(fout, tblinfo, numTables);
+	idxinfo = getIndexes(fout, tblinfo, numTables, &numIndexes);
+	idxinfoindex = buildIndexArray(idxinfo, numIndexes, sizeof(IndxInfo));
 
 	if (g_verbose)
 		write_msg(NULL, "flagging indexes in partitioned tables\n");
@@ -414,7 +417,7 @@ flagInhIndexes(Archive *fout, TableInfo tblinfo[], int numTables)
 			if (index->parentidx == 0)
 				continue;
 
-			parentidx = findIndexByOid(index->parentidx,
+			parentidx = findTableIndexByOid(index->parentidx,
 									   parentIndexArray[parenttbl->dobj.dumpId],
 									   parenttbl->numIndexes);
 			if (parentidx == NULL)
@@ -936,15 +939,26 @@ findExtensionByOid(Oid oid)
 
 /*
  * findIndexByOid
+ *	  find the entry (in idxinfo) of the index with the given oid
+ *	  returns NULL if not found
+ */
+IndxInfo *
+findIndexByOid(Oid oid)
+{
+	return (IndxInfo *) findObjectByOid(oid, idxinfoindex, numIndexes);
+}
+
+/*
+ * findTableIndexByOid
  *		find the entry of the index with the given oid
  *
- * This one's signature is different from the previous ones because we lack a
- * global array of all indexes, so caller must pass their array as argument.
+ * This one's signature is different from the previous ones because we use
+ * it to find an index of specific table who passes its index array as argument.
  */
 static IndxInfo *
-findIndexByOid(Oid oid, DumpableObject **idxinfoindex, int numIndexes)
+findTableIndexByOid(Oid oid, DumpableObject **tbl_idxinfoindex, int tblNumIndexes)
 {
-	return (IndxInfo *) findObjectByOid(oid, idxinfoindex, numIndexes);
+	return (IndxInfo *) findObjectByOid(oid, tbl_idxinfoindex, tblNumIndexes);
 }
 
 /*
diff --git a/src/bin/pg_dump/pg_dump.c src/bin/pg_dump/pg_dump.c
index 54ed71ca4c1..f4fb4da7d3a 100644
--- a/src/bin/pg_dump/pg_dump.c
+++ src/bin/pg_dump/pg_dump.c
@@ -1523,11 +1523,22 @@ selectDumpableType(TypeInfo *tyinfo, Archive *fout)
 	if (OidIsValid(tyinfo->typrelid) &&
 		tyinfo->typrelkind != RELKIND_COMPOSITE_TYPE)
 	{
-		TableInfo  *tytable = findTableByOid(tyinfo->typrelid);
+		DumpableObject *parentRel;
 
 		tyinfo->dobj.objType = DO_DUMMY_TYPE;
-		if (tytable != NULL)
-			tyinfo->dobj.dump = tytable->dobj.dump;
+
+		/* Get associated relation */
+		if (tyinfo->typrelkind == RELKIND_INDEX)
+			parentRel = (DumpableObject *) findIndexByOid(tyinfo->typrelid);
+		else
+			parentRel = (DumpableObject *) findTableByOid(tyinfo->typrelid);
+
+		/*
+		 * If associated relation found, dump based on if the
+		 * contents of the associated relation are being dumped.
+		 */
+		if (parentRel != NULL)
+			tyinfo->dobj.dump = parentRel->dump;
 		else
 			tyinfo->dobj.dump = DUMP_COMPONENT_NONE;
 		return;
@@ -4247,6 +4258,9 @@ binary_upgrade_set_type_oids_by_type_oid(Archive *fout,
 	PGresult   *res;
 	Oid			pg_type_array_oid;
 
+	if (pg_type_oid == InvalidOid)
+		return;
+
 	appendPQExpBufferStr(upgrade_buffer, "\n-- For binary upgrade, must preserve pg_type oid\n");
 	appendPQExpBuffer(upgrade_buffer,
 					  "SELECT pg_catalog.binary_upgrade_set_next_pg_type_oid('%u'::pg_catalog.oid);\n\n",
@@ -6745,8 +6759,8 @@ getInherits(Archive *fout, int *numInherits)
  * Note: index data is not returned directly to the caller, but it
  * does get entered into the DumpableObject tables.
  */
-void
-getIndexes(Archive *fout, TableInfo tblinfo[], int numTables)
+IndxInfo *
+getIndexes(Archive *fout, TableInfo tblinfo[], int numTables, int *numIndexes)
 {
 	int			i,
 				j;
@@ -6777,6 +6791,9 @@ getIndexes(Archive *fout, TableInfo tblinfo[], int numTables)
 				i_indstatcols,
 				i_indstatvals;
 	int			ntups;
+	Size		off = 0;
+
+	*numIndexes = 0;
 
 	for (i = 0; i < numTables; i++)
 	{
@@ -7022,6 +7039,7 @@ getIndexes(Archive *fout, TableInfo tblinfo[], int numTables)
 			(IndxInfo *) pg_malloc(ntups * sizeof(IndxInfo));
 		constrinfo = (ConstraintInfo *) pg_malloc(ntups * sizeof(ConstraintInfo));
 		tbinfo->numIndexes = ntups;
+		*numIndexes += ntups;
 
 		for (j = 0; j < ntups; j++)
 		{
@@ -7091,6 +7109,27 @@ getIndexes(Archive *fout, TableInfo tblinfo[], int numTables)
 	}
 
 	destroyPQExpBuffer(query);
+
+	/*
+	 * A second pass to form an array of all index infos.
+	 * Now that we know the total number of indexes after the first pass,
+	 * we can allocate all needed memory in one call instead of using realloc.
+	 */
+	indxinfo = (IndxInfo *) pg_malloc(*numIndexes * sizeof(IndxInfo));
+
+	for (i = 0; i < numTables; i++)
+	{
+		TableInfo  *tbinfo = &tblinfo[i];
+		int			copynum = tbinfo->numIndexes;
+
+		if (copynum < 1)
+			continue;
+
+		memcpy(indxinfo + off, tbinfo->indexes, copynum * sizeof(IndxInfo));
+		off += copynum;
+	}
+
+	return indxinfo;
 }
 
 /*
@@ -16291,8 +16330,13 @@ dumpIndex(Archive *fout, IndxInfo *indxinfo)
 		int			nstatvals;
 
 		if (dopt->binary_upgrade)
+		{
 			binary_upgrade_set_pg_class_oids(fout, q,
 											 indxinfo->dobj.catId.oid, true);
+			if (indxinfo->indnkeyattrs > 1)
+				binary_upgrade_set_type_oids_by_rel_oid(fout, q,
+														indxinfo->dobj.catId.oid);
+		}
 
 		/* Plain secondary index */
 		appendPQExpBuffer(q, "%s;\n", indxinfo->indexdef);
@@ -18041,6 +18085,27 @@ addBoundaryDependencies(DumpableObject **dobjs, int numObjs,
 		 */
 		switch (dobj->objType)
 		{
+			case DO_DUMMY_TYPE:
+			{
+				/*
+				 * In Vanilla, dummy types were only created for tables.
+				 * In Postgres Pro for improving join selectivity estimation
+				 * we also create two types for each composite index:
+				 *   1) a type for attributes of the index
+				 *   2) a type which is an array containing elements of type (1)
+				 * These types depend on indexes, so adding preDataBound -> type
+				 * dependency would create a loop; don't do that.
+				 */
+				TypeInfo *tyinfo = (TypeInfo *) dobj;
+				if (tyinfo->isArray)
+					/* If it's an array, take its element type */
+					tyinfo = findTypeByOid(tyinfo->typelem);
+
+				if (OidIsValid(tyinfo->typrelid) &&
+					(tyinfo->typrelkind == RELKIND_INDEX ||
+					tyinfo->typrelkind == RELKIND_PARTITIONED_INDEX))
+					break;
+			}
 			case DO_NAMESPACE:
 			case DO_EXTENSION:
 			case DO_TYPE:
@@ -18057,7 +18122,6 @@ addBoundaryDependencies(DumpableObject **dobjs, int numObjs,
 			case DO_ATTRDEF:
 			case DO_PROCLANG:
 			case DO_CAST:
-			case DO_DUMMY_TYPE:
 			case DO_TSPARSER:
 			case DO_TSDICT:
 			case DO_TSTEMPLATE:
diff --git a/src/bin/pg_dump/pg_dump.h src/bin/pg_dump/pg_dump.h
index b1f635a32e6..118172b9561 100644
--- a/src/bin/pg_dump/pg_dump.h
+++ src/bin/pg_dump/pg_dump.h
@@ -670,6 +670,7 @@ extern OprInfo *findOprByOid(Oid oid);
 extern CollInfo *findCollationByOid(Oid oid);
 extern NamespaceInfo *findNamespaceByOid(Oid oid);
 extern ExtensionInfo *findExtensionByOid(Oid oid);
+extern IndxInfo *findIndexByOid(Oid oid);
 
 extern void setExtensionMembership(ExtensionMemberId *extmems, int nextmems);
 extern ExtensionInfo *findOwningExtension(CatalogId catalogId);
@@ -698,7 +699,8 @@ extern ConvInfo *getConversions(Archive *fout, int *numConversions);
 extern TableInfo *getTables(Archive *fout, int *numTables);
 extern void getOwnedSeqs(Archive *fout, TableInfo tblinfo[], int numTables);
 extern InhInfo *getInherits(Archive *fout, int *numInherits);
-extern void getIndexes(Archive *fout, TableInfo tblinfo[], int numTables);
+extern IndxInfo *getIndexes(Archive *fout, TableInfo tblinfo[], int numTables,
+						int *numIndexes);
 extern void getExtendedStatistics(Archive *fout);
 extern void getConstraints(Archive *fout, TableInfo tblinfo[], int numTables);
 extern RuleInfo *getRules(Archive *fout, int *numRules);
diff --git a/src/include/nodes/relation.h src/include/nodes/relation.h
index 20cbb78abde..68981fc2b97 100644
--- a/src/include/nodes/relation.h
+++ src/include/nodes/relation.h
@@ -15,11 +15,13 @@
 #define RELATION_H
 
 #include "access/sdir.h"
+#include "catalog/pg_statistic.h"
 #include "fmgr.h"
 #include "lib/stringinfo.h"
 #include "nodes/params.h"
 #include "nodes/parsenodes.h"
 #include "storage/block.h"
+#include "utils/lsyscache.h"
 
 
 /*
@@ -811,6 +813,10 @@ typedef struct IndexOptInfo
 	bool		amcanparallel;	/* does AM support parallel scan? */
 	/* Rather than include amapi.h here, we declare amcostestimate like this */
 	void		(*amcostestimate) ();	/* AM's cost estimator */
+
+	/* cache for per-tuple index statistic. That stats could be large and it
+	 * will be expensive to uncomress it every time */
+	AttStatsSlot	sslots[STATISTIC_NUM_SLOTS + 1];
 } IndexOptInfo;
 
 /*
diff --git a/src/include/utils/builtins.h src/include/utils/builtins.h
index d0416e90fcc..142ae49c0ad 100644
--- a/src/include/utils/builtins.h
+++ src/include/utils/builtins.h
@@ -18,7 +18,6 @@
 #include "nodes/nodes.h"
 #include "utils/fmgrprotos.h"
 
-
 /* bool.c */
 extern bool parse_bool(const char *value, bool *result);
 extern bool parse_bool_with_len(const char *value, size_t len, bool *result);
diff --git a/src/include/utils/lsyscache.h src/include/utils/lsyscache.h
index e868d84cef6..8d3998c19b0 100644
--- a/src/include/utils/lsyscache.h
+++ src/include/utils/lsyscache.h
@@ -55,6 +55,8 @@ typedef struct AttStatsSlot
 	/* Remaining fields are private to get_attstatsslot/free_attstatsslot */
 	void	   *values_arr;		/* palloc'd values array, if any */
 	void	   *numbers_arr;	/* palloc'd numbers array, if any */
+
+	bool		incache;		/* do not free because struct is cached */
 } AttStatsSlot;
 
 /* Hook for plugins to get control in get_attavgwidth() */
@@ -173,6 +175,8 @@ extern int32 get_typavgwidth(Oid typid, int32 typmod);
 extern int32 get_attavgwidth(Oid relid, AttrNumber attnum);
 extern bool get_attstatsslot(AttStatsSlot *sslot, HeapTuple statstuple,
 				 int reqkind, Oid reqop, int flags);
+extern AttStatsSlot* fill_attstatsslot(AttStatsSlot *sslots, HeapTuple statstuple,
+				 int reqkind, Oid reqop, int flags);
 extern void free_attstatsslot(AttStatsSlot *sslot);
 extern char *get_namespace_name(Oid nspid);
 extern char *get_namespace_name_or_temp(Oid nspid);
diff --git a/src/include/utils/selfuncs.h src/include/utils/selfuncs.h
index 95e44280c4c..a52a6da5472 100644
--- a/src/include/utils/selfuncs.h
+++ src/include/utils/selfuncs.h
@@ -76,6 +76,7 @@ typedef struct VariableStatData
 	int32		atttypmod;		/* actual typmod (after stripping relabel) */
 	bool		isunique;		/* matches unique index or DISTINCT clause */
 	bool		acl_ok;			/* result of ACL check on table or column */
+	AttStatsSlot	*sslots;
 } VariableStatData;
 
 #define ReleaseVariableStats(vardata)  \
@@ -176,6 +177,9 @@ extern double histogram_selectivity(VariableStatData *vardata, FmgrInfo *opproc,
 					  Datum constval, bool varonleft,
 					  int min_hist_size, int n_skip,
 					  int *hist_size);
+double prefix_record_histogram_selectivity(VariableStatData *vardata,
+									Datum constvalLeft, Datum constvalRight,
+									int record_cmp_prefix, int *n_bins);
 
 extern Pattern_Prefix_Status pattern_fixed_prefix(Const *patt,
 					 Pattern_Type ptype,
@@ -226,5 +230,19 @@ extern Selectivity scalararraysel_containment(PlannerInfo *root,
 						   Node *leftop, Node *rightop,
 						   Oid elemtype, bool isEquality, bool useOr,
 						   int varRelid);
+extern Selectivity eqjoin_selectivity(PlannerInfo *root, Oid operator,
+									  VariableStatData* vardata1,
+									  VariableStatData* vardata2,
+									  SpecialJoinInfo *sjinfo,
+									  int record_cmp_prefix);
+extern Selectivity eqconst_selectivity(Oid operator, VariableStatData *vardata,
+									   Datum constval, bool constisnull,
+									   bool varonleft, bool negate,
+									   int record_cmp_prefix);
+extern Selectivity ineq_histogram_selectivity(PlannerInfo *root,
+											  VariableStatData *vardata,
+											  FmgrInfo *opproc, bool isgt, bool iseq,
+											  Datum constval, Oid consttype,
+											  int record_cmp_prefix);
 
 #endif							/* SELFUNCS_H */
diff --git a/src/test/regress/expected/insert_conflict.out src/test/regress/expected/insert_conflict.out
index a28611745c1..44306281940 100644
--- a/src/test/regress/expected/insert_conflict.out
+++ src/test/regress/expected/insert_conflict.out
@@ -891,6 +891,17 @@ create table parted_conflict_1 (drp text, c int, a int, b text);
 alter table parted_conflict_1 drop column drp;
 create unique index on parted_conflict (a, b);
 alter table parted_conflict attach partition parted_conflict_1 for values from (0) to (1000);
+-- test that index types were created after ALTER TABLE
+select p.reltype > 0
+from pg_index i
+join pg_inherits inh on inh.inhparent = i.indexrelid
+join pg_class p on p.oid = inh.inhrelid
+where i.indrelid = 'parted_conflict'::regclass::oid;
+ ?column? 
+----------
+ t
+(1 row)
+
 truncate parted_conflict;
 insert into parted_conflict values (50, 'cincuenta', 1);
 insert into parted_conflict values (50, 'cincuenta', 2)
diff --git a/src/test/regress/expected/stats_ext.out src/test/regress/expected/stats_ext.out
index eebf250998e..309c8942ed1 100644
--- a/src/test/regress/expected/stats_ext.out
+++ src/test/regress/expected/stats_ext.out
@@ -402,6 +402,14 @@ CREATE TABLE functional_dependencies (
 SET random_page_cost = 1.2;
 CREATE INDEX fdeps_ab_idx ON functional_dependencies (a, b);
 CREATE INDEX fdeps_abc_idx ON functional_dependencies (a, b, c);
+-- test that index types were created
+SELECT reltype > 0 FROM pg_class where relname in ('fdeps_ab_idx', 'fdeps_abc_idx');
+ ?column? 
+----------
+ t
+ t
+(2 rows)
+
 -- random data (no functional dependencies)
 INSERT INTO functional_dependencies (a, b, c, filler1)
      SELECT mod(i, 23), mod(i, 29), mod(i, 31), i FROM generate_series(1,5000) s(i);
@@ -493,6 +501,14 @@ EXPLAIN (COSTS OFF)
 
 -- check change of column type doesn't break it
 ALTER TABLE functional_dependencies ALTER COLUMN c TYPE numeric;
+-- test that index types were created after ALTER TABLE
+SELECT reltype > 0 FROM pg_class where relname in ('fdeps_ab_idx', 'fdeps_abc_idx');
+ ?column? 
+----------
+ t
+ t
+(2 rows)
+
 EXPLAIN (COSTS OFF)
  SELECT * FROM functional_dependencies WHERE a = 1 AND b = '1' AND c = 1;
                     QUERY PLAN                     
diff --git a/src/test/regress/sql/insert_conflict.sql src/test/regress/sql/insert_conflict.sql
index c68013e1795..3a5fa16e7a2 100644
--- a/src/test/regress/sql/insert_conflict.sql
+++ src/test/regress/sql/insert_conflict.sql
@@ -566,6 +566,14 @@ create table parted_conflict_1 (drp text, c int, a int, b text);
 alter table parted_conflict_1 drop column drp;
 create unique index on parted_conflict (a, b);
 alter table parted_conflict attach partition parted_conflict_1 for values from (0) to (1000);
+
+-- test that index types were created after ALTER TABLE
+select p.reltype > 0
+from pg_index i
+join pg_inherits inh on inh.inhparent = i.indexrelid
+join pg_class p on p.oid = inh.inhrelid
+where i.indrelid = 'parted_conflict'::regclass::oid;
+
 truncate parted_conflict;
 insert into parted_conflict values (50, 'cincuenta', 1);
 insert into parted_conflict values (50, 'cincuenta', 2)
diff --git a/src/test/regress/sql/stats_ext.sql src/test/regress/sql/stats_ext.sql
index 43ff77c5344..2d29190aa7a 100644
--- a/src/test/regress/sql/stats_ext.sql
+++ src/test/regress/sql/stats_ext.sql
@@ -229,6 +229,9 @@ SET random_page_cost = 1.2;
 CREATE INDEX fdeps_ab_idx ON functional_dependencies (a, b);
 CREATE INDEX fdeps_abc_idx ON functional_dependencies (a, b, c);
 
+-- test that index types were created
+SELECT reltype > 0 FROM pg_class where relname in ('fdeps_ab_idx', 'fdeps_abc_idx');
+
 -- random data (no functional dependencies)
 INSERT INTO functional_dependencies (a, b, c, filler1)
      SELECT mod(i, 23), mod(i, 29), mod(i, 31), i FROM generate_series(1,5000) s(i);
@@ -281,6 +284,9 @@ EXPLAIN (COSTS OFF)
 -- check change of column type doesn't break it
 ALTER TABLE functional_dependencies ALTER COLUMN c TYPE numeric;
 
+-- test that index types were created after ALTER TABLE
+SELECT reltype > 0 FROM pg_class where relname in ('fdeps_ab_idx', 'fdeps_abc_idx');
+
 EXPLAIN (COSTS OFF)
  SELECT * FROM functional_dependencies WHERE a = 1 AND b = '1' AND c = 1;