Adding upstream version 14.5.upstream/14.5upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

8 files changed, 7755 insertions, 0 deletions

diff --git a/src/backend/statistics/Makefile b/src/backend/statistics/Makefile
new file mode 100644
index 0000000..89cf8c2
--- /dev/null
+++ b/src/backend/statistics/Makefile
@@ -0,0 +1,21 @@
+#-------------------------------------------------------------------------
+#
+# Makefile--
+#    Makefile for statistics
+#
+# IDENTIFICATION
+#    src/backend/statistics/Makefile
+#
+#-------------------------------------------------------------------------
+
+subdir = src/backend/statistics
+top_builddir = ../../..
+include $(top_builddir)/src/Makefile.global
+
+OBJS = \
+	dependencies.o \
+	extended_stats.o \
+	mcv.o \
+	mvdistinct.o
+
+include $(top_srcdir)/src/backend/common.mk
diff --git a/src/backend/statistics/README b/src/backend/statistics/README
new file mode 100644
index 0000000..7fda13e
--- /dev/null
+++ b/src/backend/statistics/README
@@ -0,0 +1,101 @@
+Extended statistics
+===================
+
+When estimating various quantities (e.g. condition selectivities) the default
+approach relies on the assumption of independence. In practice that's often
+not true, resulting in estimation errors.
+
+Extended statistics track different types of dependencies between the columns,
+hopefully improving the estimates and producing better plans.
+
+
+Types of statistics
+-------------------
+
+There are currently two kinds of extended statistics:
+
+    (a) ndistinct coefficients
+
+    (b) soft functional dependencies (README.dependencies)
+
+    (c) MCV lists (README.mcv)
+
+
+Compatible clause types
+-----------------------
+
+Each type of statistics may be used to estimate some subset of clause types.
+
+    (a) functional dependencies - equality clauses (AND), possibly IS NULL
+
+    (b) MCV lists - equality and inequality clauses (AND, OR, NOT), IS [NOT] NULL
+
+Currently, only OpExprs in the form Var op Const, or Const op Var are
+supported, however it's feasible to expand the code later to also estimate the
+selectivities on clauses such as Var op Var.
+
+
+Complex clauses
+---------------
+
+We also support estimating more complex clauses - essentially AND/OR clauses
+with (Var op Const) as leaves, as long as all the referenced attributes are
+covered by a single statistics object.
+
+For example this condition
+
+    (a=1) AND ((b=2) OR ((c=3) AND (d=4)))
+
+may be estimated using statistics on (a,b,c,d). If we only have statistics on
+(b,c,d) we may estimate the second part, and estimate (a=1) using simple stats.
+
+If we only have statistics on (a,b,c) we can't apply it at all at this point,
+but it's worth pointing out clauselist_selectivity() works recursively and when
+handling the second part (the OR-clause), we'll be able to apply the statistics.
+
+Note: The multi-statistics estimation patch also makes it possible to pass some
+clauses as 'conditions' into the deeper parts of the expression tree.
+
+
+Selectivity estimation
+----------------------
+
+Throughout the planner clauselist_selectivity() still remains in charge of
+most selectivity estimate requests. clauselist_selectivity() can be instructed
+to try to make use of any extended statistics on the given RelOptInfo, which
+it will do if:
+
+    (a) An actual valid RelOptInfo was given. Join relations are passed in as
+        NULL, therefore are invalid.
+
+    (b) The relation given actually has any extended statistics defined which
+        are actually built.
+
+When the above conditions are met, clauselist_selectivity() first attempts to
+pass the clause list off to the extended statistics selectivity estimation
+function. This functions may not find any clauses which is can perform any
+estimations on. In such cases these clauses are simply ignored. When actual
+estimation work is performed in these functions they're expected to mark which
+clauses they've performed estimations for so that any other function
+performing estimations knows which clauses are to be skipped.
+
+Size of sample in ANALYZE
+-------------------------
+
+When performing ANALYZE, the number of rows to sample is determined as
+
+    (300 * statistics_target)
+
+That works reasonably well for statistics on individual columns, but perhaps
+it's not enough for extended statistics. Papers analyzing estimation errors
+all use samples proportional to the table (usually finding that 1-3% of the
+table is enough to build accurate stats).
+
+The requested accuracy (number of MCV items or histogram bins) should also
+be considered when determining the sample size, and in extended statistics
+those are not necessarily limited by statistics_target.
+
+This however merits further discussion, because collecting the sample is quite
+expensive and increasing it further would make ANALYZE even more painful.
+Judging by the experiments with the current implementation, the fixed size
+seems to work reasonably well for now, so we leave this as future work.
diff --git a/src/backend/statistics/README.dependencies b/src/backend/statistics/README.dependencies
new file mode 100644
index 0000000..6c446bd
--- /dev/null
+++ b/src/backend/statistics/README.dependencies
@@ -0,0 +1,116 @@
+Soft functional dependencies
+============================
+
+Functional dependencies are a concept well described in relational theory,
+particularly in the definition of normalization and "normal forms". Wikipedia
+has a nice definition of a functional dependency [1]:
+
+    In a given table, an attribute Y is said to have a functional dependency
+    on a set of attributes X (written X -> Y) if and only if each X value is
+    associated with precisely one Y value. For example, in an "Employee"
+    table that includes the attributes "Employee ID" and "Employee Date of
+    Birth", the functional dependency
+
+        {Employee ID} -> {Employee Date of Birth}
+
+    would hold. It follows from the previous two sentences that each
+    {Employee ID} is associated with precisely one {Employee Date of Birth}.
+
+    [1] https://en.wikipedia.org/wiki/Functional_dependency
+
+In practical terms, functional dependencies mean that a value in one column
+determines values in some other column. Consider for example this trivial
+table with two integer columns:
+
+    CREATE TABLE t (a INT, b INT)
+        AS SELECT i, i/10 FROM generate_series(1,100000) s(i);
+
+Clearly, knowledge of the value in column 'a' is sufficient to determine the
+value in column 'b', as it's simply (a/10). A more practical example may be
+addresses, where the knowledge of a ZIP code (usually) determines city. Larger
+cities may have multiple ZIP codes, so the dependency can't be reversed.
+
+Many datasets might be normalized not to contain such dependencies, but often
+it's not practical for various reasons. In some cases, it's actually a conscious
+design choice to model the dataset in a denormalized way, either because of
+performance or to make querying easier.
+
+
+Soft dependencies
+-----------------
+
+Real-world data sets often contain data errors, either because of data entry
+mistakes (user mistyping the ZIP code) or perhaps issues in generating the
+data (e.g. a ZIP code mistakenly assigned to two cities in different states).
+
+A strict implementation would either ignore dependencies in such cases,
+rendering the approach mostly useless even for slightly noisy data sets, or
+result in sudden changes in behavior depending on minor differences between
+samples provided to ANALYZE.
+
+For this reason, extended statistics implement "soft" functional dependencies,
+associating each functional dependency with a degree of validity (a number
+between 0 and 1). This degree is then used to combine selectivities in a
+smooth manner.
+
+
+Mining dependencies (ANALYZE)
+-----------------------------
+
+The current algorithm is fairly simple - generate all possible functional
+dependencies, and for each one count the number of rows consistent with it.
+Then use the fraction of rows (supporting/total) as the degree.
+
+To count the rows consistent with the dependency (a => b):
+
+ (a) Sort the data lexicographically, i.e. first by 'a' then 'b'.
+
+ (b) For each group of rows with the same 'a' value, count the number of
+     distinct values in 'b'.
+
+ (c) If there's a single distinct value in 'b', the rows are consistent with
+     the functional dependency, otherwise they contradict it.
+
+
+Clause reduction (planner/optimizer)
+------------------------------------
+
+Applying the functional dependencies is fairly simple: given a list of
+equality clauses, we compute selectivities of each clause and then use the
+degree to combine them using this formula
+
+    P(a=?,b=?) = P(a=?) * (d + (1-d) * P(b=?))
+
+Where 'd' is the degree of functional dependency (a => b).
+
+With more than two equality clauses, this process happens recursively. For
+example for (a,b,c) we first use (a,b => c) to break the computation into
+
+    P(a=?,b=?,c=?) = P(a=?,b=?) * (e + (1-e) * P(c=?))
+
+where 'e' is the degree of functional dependency (a,b => c); then we can
+apply (a=>b) the same way on P(a=?,b=?).
+
+
+Consistency of clauses
+----------------------
+
+Functional dependencies only express general dependencies between columns,
+without referencing particular values. This assumes that the equality clauses
+are in fact consistent with the functional dependency, i.e. that given a
+dependency (a=>b), the value in (b=?) clause is the value determined by (a=?).
+If that's not the case, the clauses are "inconsistent" with the functional
+dependency and the result will be over-estimation.
+
+This may happen, for example, when using conditions on the ZIP code and city
+name with mismatching values (ZIP code for a different city), etc. In such a
+case, the result set will be empty, but we'll estimate the selectivity using
+the ZIP code condition.
+
+In this case, the default estimation based on AVIA principle happens to work
+better, but mostly by chance.
+
+This issue is the price for the simplicity of functional dependencies. If the
+application frequently constructs queries with clauses inconsistent with
+functional dependencies present in the data, the best solution is not to
+use functional dependencies, but one of the more complex types of statistics.
diff --git a/src/backend/statistics/README.mcv b/src/backend/statistics/README.mcv
new file mode 100644
index 0000000..8455b0d
--- /dev/null
+++ b/src/backend/statistics/README.mcv
@@ -0,0 +1,103 @@
+MCV lists
+=========
+
+Multivariate MCV (most-common values) lists are a straightforward extension of
+regular MCV list, tracking most frequent combinations of values for a group of
+attributes.
+
+This works particularly well for columns with a small number of distinct values,
+as the list may include all the combinations and approximate the distribution
+very accurately.
+
+For columns with a large number of distinct values (e.g. those with continuous
+domains), the list will only track the most frequent combinations. If the
+distribution is mostly uniform (all combinations about equally frequent), the
+MCV list will be empty.
+
+Estimates of some clauses (e.g. equality) based on MCV lists are more accurate
+than when using histograms.
+
+Also, MCV lists don't necessarily require sorting of the values (the fact that
+we use sorting when building them is implementation detail), but even more
+importantly the ordering is not built into the approximation (while histograms
+are built on ordering). So MCV lists work well even for attributes where the
+ordering of the data type is disconnected from the meaning of the data. For
+example we know how to sort strings, but it's unlikely to make much sense for
+city names (or other label-like attributes).
+
+
+Selectivity estimation
+----------------------
+
+The estimation, implemented in mcv_clauselist_selectivity(), is quite simple
+in principle - we need to identify MCV items matching all the clauses and sum
+frequencies of all those items.
+
+Currently MCV lists support estimation of the following clause types:
+
+    (a) equality clauses      WHERE (a = 1) AND (b = 2)
+    (b) inequality clauses    WHERE (a < 1) AND (b >= 2)
+    (c) NULL clauses          WHERE (a IS NULL) AND (b IS NOT NULL)
+    (d) OR clauses            WHERE (a < 1) OR (b >= 2)
+
+It's possible to add support for additional clauses, for example:
+
+    (e) multi-var clauses     WHERE (a > b)
+
+and possibly others. These are tasks for the future, not yet implemented.
+
+
+Hashed MCV (not yet implemented)
+--------------------------------
+
+Regular MCV lists have to include actual values for each item, so if those items
+are large the list may be quite large. This is especially true for multivariate
+MCV lists, although the current implementation partially mitigates this by
+performing de-duplicating the values before storing them on disk.
+
+It's possible to only store hashes (32-bit values) instead of the actual values,
+significantly reducing the space requirements. Obviously, this would only make
+the MCV lists useful for estimating equality conditions (assuming the 32-bit
+hashes make the collisions rare enough).
+
+This might also complicate matching the columns to available stats.
+
+
+TODO Consider implementing hashed MCV list, storing just 32-bit hashes instead
+     of the actual values. This type of MCV list will be useful only for
+     estimating equality clauses, and will reduce space requirements for large
+     varlena types (in such cases we usually only want equality anyway).
+
+
+Inspecting the MCV list
+-----------------------
+
+Inspecting the regular (per-attribute) MCV lists is trivial, as it's enough
+to select the columns from pg_stats. The data is encoded as anyarrays, and
+all the items have the same data type, so anyarray provides a simple way to
+get a text representation.
+
+With multivariate MCV lists the columns may use different data types, making
+it impossible to use anyarrays. It might be possible to produce a similar
+array-like representation, but that would complicate further processing and
+analysis of the MCV list.
+
+So instead the MCV lists are stored in a custom data type (pg_mcv_list),
+which however makes it more difficult to inspect the contents. To make that
+easier, there's a SRF returning detailed information about the MCV lists.
+
+    SELECT m.* FROM pg_statistic_ext s,
+                    pg_statistic_ext_data d,
+                    pg_mcv_list_items(stxdmcv) m
+              WHERE s.stxname = 'stts2'
+                AND d.stxoid = s.oid;
+
+It accepts one parameter - a pg_mcv_list value (which can only be obtained
+from pg_statistic_ext_data catalog, to defend against malicious input), and
+returns these columns:
+
+    - item index (0, ..., (nitems-1))
+    - values (string array)
+    - nulls only (boolean array)
+    - frequency (double precision)
+    - base_frequency (double precision)
diff --git a/src/backend/statistics/dependencies.c b/src/backend/statistics/dependencies.c
new file mode 100644
index 0000000..a9187dd
--- /dev/null
+++ b/src/backend/statistics/dependencies.c
@@ -0,0 +1,1860 @@
+/*-------------------------------------------------------------------------
+ *
+ * dependencies.c
+ *	  POSTGRES functional dependencies
+ *
+ * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * IDENTIFICATION
+ *	  src/backend/statistics/dependencies.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include "access/htup_details.h"
+#include "access/sysattr.h"
+#include "catalog/pg_operator.h"
+#include "catalog/pg_statistic_ext.h"
+#include "catalog/pg_statistic_ext_data.h"
+#include "lib/stringinfo.h"
+#include "nodes/nodeFuncs.h"
+#include "nodes/nodes.h"
+#include "nodes/pathnodes.h"
+#include "optimizer/clauses.h"
+#include "optimizer/optimizer.h"
+#include "parser/parsetree.h"
+#include "statistics/extended_stats_internal.h"
+#include "statistics/statistics.h"
+#include "utils/bytea.h"
+#include "utils/fmgroids.h"
+#include "utils/fmgrprotos.h"
+#include "utils/lsyscache.h"
+#include "utils/memutils.h"
+#include "utils/selfuncs.h"
+#include "utils/syscache.h"
+#include "utils/typcache.h"
+
+/* size of the struct header fields (magic, type, ndeps) */
+#define SizeOfHeader		(3 * sizeof(uint32))
+
+/* size of a serialized dependency (degree, natts, atts) */
+#define SizeOfItem(natts) \
+	(sizeof(double) + sizeof(AttrNumber) * (1 + (natts)))
+
+/* minimal size of a dependency (with two attributes) */
+#define MinSizeOfItem	SizeOfItem(2)
+
+/* minimal size of dependencies, when all deps are minimal */
+#define MinSizeOfItems(ndeps) \
+	(SizeOfHeader + (ndeps) * MinSizeOfItem)
+
+/*
+ * Internal state for DependencyGenerator of dependencies. Dependencies are similar to
+ * k-permutations of n elements, except that the order does not matter for the
+ * first (k-1) elements. That is, (a,b=>c) and (b,a=>c) are equivalent.
+ */
+typedef struct DependencyGeneratorData
+{
+	int			k;				/* size of the dependency */
+	int			n;				/* number of possible attributes */
+	int			current;		/* next dependency to return (index) */
+	AttrNumber	ndependencies;	/* number of dependencies generated */
+	AttrNumber *dependencies;	/* array of pre-generated dependencies	*/
+} DependencyGeneratorData;
+
+typedef DependencyGeneratorData *DependencyGenerator;
+
+static void generate_dependencies_recurse(DependencyGenerator state,
+										  int index, AttrNumber start, AttrNumber *current);
+static void generate_dependencies(DependencyGenerator state);
+static DependencyGenerator DependencyGenerator_init(int n, int k);
+static void DependencyGenerator_free(DependencyGenerator state);
+static AttrNumber *DependencyGenerator_next(DependencyGenerator state);
+static double dependency_degree(StatsBuildData *data, int k, AttrNumber *dependency);
+static bool dependency_is_fully_matched(MVDependency *dependency,
+										Bitmapset *attnums);
+static bool dependency_is_compatible_clause(Node *clause, Index relid,
+											AttrNumber *attnum);
+static bool dependency_is_compatible_expression(Node *clause, Index relid,
+												List *statlist, Node **expr);
+static MVDependency *find_strongest_dependency(MVDependencies **dependencies,
+											   int ndependencies, Bitmapset *attnums);
+static Selectivity clauselist_apply_dependencies(PlannerInfo *root, List *clauses,
+												 int varRelid, JoinType jointype,
+												 SpecialJoinInfo *sjinfo,
+												 MVDependency **dependencies,
+												 int ndependencies,
+												 AttrNumber *list_attnums,
+												 Bitmapset **estimatedclauses);
+
+static void
+generate_dependencies_recurse(DependencyGenerator state, int index,
+							  AttrNumber start, AttrNumber *current)
+{
+	/*
+	 * The generator handles the first (k-1) elements differently from the
+	 * last element.
+	 */
+	if (index < (state->k - 1))
+	{
+		AttrNumber	i;
+
+		/*
+		 * The first (k-1) values have to be in ascending order, which we
+		 * generate recursively.
+		 */
+
+		for (i = start; i < state->n; i++)
+		{
+			current[index] = i;
+			generate_dependencies_recurse(state, (index + 1), (i + 1), current);
+		}
+	}
+	else
+	{
+		int			i;
+
+		/*
+		 * the last element is the implied value, which does not respect the
+		 * ascending order. We just need to check that the value is not in the
+		 * first (k-1) elements.
+		 */
+
+		for (i = 0; i < state->n; i++)
+		{
+			int			j;
+			bool		match = false;
+
+			current[index] = i;
+
+			for (j = 0; j < index; j++)
+			{
+				if (current[j] == i)
+				{
+					match = true;
+					break;
+				}
+			}
+
+			/*
+			 * If the value is not found in the first part of the dependency,
+			 * we're done.
+			 */
+			if (!match)
+			{
+				state->dependencies = (AttrNumber *) repalloc(state->dependencies,
+															  state->k * (state->ndependencies + 1) * sizeof(AttrNumber));
+				memcpy(&state->dependencies[(state->k * state->ndependencies)],
+					   current, state->k * sizeof(AttrNumber));
+				state->ndependencies++;
+			}
+		}
+	}
+}
+
+/* generate all dependencies (k-permutations of n elements) */
+static void
+generate_dependencies(DependencyGenerator state)
+{
+	AttrNumber *current = (AttrNumber *) palloc0(sizeof(AttrNumber) * state->k);
+
+	generate_dependencies_recurse(state, 0, 0, current);
+
+	pfree(current);
+}
+
+/*
+ * initialize the DependencyGenerator of variations, and prebuild the variations
+ *
+ * This pre-builds all the variations. We could also generate them in
+ * DependencyGenerator_next(), but this seems simpler.
+ */
+static DependencyGenerator
+DependencyGenerator_init(int n, int k)
+{
+	DependencyGenerator state;
+
+	Assert((n >= k) && (k > 0));
+
+	/* allocate the DependencyGenerator state */
+	state = (DependencyGenerator) palloc0(sizeof(DependencyGeneratorData));
+	state->dependencies = (AttrNumber *) palloc(k * sizeof(AttrNumber));
+
+	state->ndependencies = 0;
+	state->current = 0;
+	state->k = k;
+	state->n = n;
+
+	/* now actually pre-generate all the variations */
+	generate_dependencies(state);
+
+	return state;
+}
+
+/* free the DependencyGenerator state */
+static void
+DependencyGenerator_free(DependencyGenerator state)
+{
+	pfree(state->dependencies);
+	pfree(state);
+
+}
+
+/* generate next combination */
+static AttrNumber *
+DependencyGenerator_next(DependencyGenerator state)
+{
+	if (state->current == state->ndependencies)
+		return NULL;
+
+	return &state->dependencies[state->k * state->current++];
+}
+
+
+/*
+ * validates functional dependency on the data
+ *
+ * An actual work horse of detecting functional dependencies. Given a variation
+ * of k attributes, it checks that the first (k-1) are sufficient to determine
+ * the last one.
+ */
+static double
+dependency_degree(StatsBuildData *data, int k, AttrNumber *dependency)
+{
+	int			i,
+				nitems;
+	MultiSortSupport mss;
+	SortItem   *items;
+	AttrNumber *attnums_dep;
+
+	/* counters valid within a group */
+	int			group_size = 0;
+	int			n_violations = 0;
+
+	/* total number of rows supporting (consistent with) the dependency */
+	int			n_supporting_rows = 0;
+
+	/* Make sure we have at least two input attributes. */
+	Assert(k >= 2);
+
+	/* sort info for all attributes columns */
+	mss = multi_sort_init(k);
+
+	/*
+	 * Translate the array of indexes to regular attnums for the dependency
+	 * (we will need this to identify the columns in StatsBuildData).
+	 */
+	attnums_dep = (AttrNumber *) palloc(k * sizeof(AttrNumber));
+	for (i = 0; i < k; i++)
+		attnums_dep[i] = data->attnums[dependency[i]];
+
+	/*
+	 * Verify the dependency (a,b,...)->z, using a rather simple algorithm:
+	 *
+	 * (a) sort the data lexicographically
+	 *
+	 * (b) split the data into groups by first (k-1) columns
+	 *
+	 * (c) for each group count different values in the last column
+	 *
+	 * We use the column data types' default sort operators and collations;
+	 * perhaps at some point it'd be worth using column-specific collations?
+	 */
+
+	/* prepare the sort function for the dimensions */
+	for (i = 0; i < k; i++)
+	{
+		VacAttrStats *colstat = data->stats[dependency[i]];
+		TypeCacheEntry *type;
+
+		type = lookup_type_cache(colstat->attrtypid, TYPECACHE_LT_OPR);
+		if (type->lt_opr == InvalidOid) /* shouldn't happen */
+			elog(ERROR, "cache lookup failed for ordering operator for type %u",
+				 colstat->attrtypid);
+
+		/* prepare the sort function for this dimension */
+		multi_sort_add_dimension(mss, i, type->lt_opr, colstat->attrcollid);
+	}
+
+	/*
+	 * build an array of SortItem(s) sorted using the multi-sort support
+	 *
+	 * XXX This relies on all stats entries pointing to the same tuple
+	 * descriptor.  For now that assumption holds, but it might change in the
+	 * future for example if we support statistics on multiple tables.
+	 */
+	items = build_sorted_items(data, &nitems, mss, k, attnums_dep);
+
+	/*
+	 * Walk through the sorted array, split it into rows according to the
+	 * first (k-1) columns. If there's a single value in the last column, we
+	 * count the group as 'supporting' the functional dependency. Otherwise we
+	 * count it as contradicting.
+	 */
+
+	/* start with the first row forming a group */
+	group_size = 1;
+
+	/* loop 1 beyond the end of the array so that we count the final group */
+	for (i = 1; i <= nitems; i++)
+	{
+		/*
+		 * Check if the group ended, which may be either because we processed
+		 * all the items (i==nitems), or because the i-th item is not equal to
+		 * the preceding one.
+		 */
+		if (i == nitems ||
+			multi_sort_compare_dims(0, k - 2, &items[i - 1], &items[i], mss) != 0)
+		{
+			/*
+			 * If no violations were found in the group then track the rows of
+			 * the group as supporting the functional dependency.
+			 */
+			if (n_violations == 0)
+				n_supporting_rows += group_size;
+
+			/* Reset counters for the new group */
+			n_violations = 0;
+			group_size = 1;
+			continue;
+		}
+		/* first columns match, but the last one does not (so contradicting) */
+		else if (multi_sort_compare_dim(k - 1, &items[i - 1], &items[i], mss) != 0)
+			n_violations++;
+
+		group_size++;
+	}
+
+	/* Compute the 'degree of validity' as (supporting/total). */
+	return (n_supporting_rows * 1.0 / data->numrows);
+}
+
+/*
+ * detects functional dependencies between groups of columns
+ *
+ * Generates all possible subsets of columns (variations) and computes
+ * the degree of validity for each one. For example when creating statistics
+ * on three columns (a,b,c) there are 9 possible dependencies
+ *
+ *	   two columns			  three columns
+ *	   -----------			  -------------
+ *	   (a) -> b				  (a,b) -> c
+ *	   (a) -> c				  (a,c) -> b
+ *	   (b) -> a				  (b,c) -> a
+ *	   (b) -> c
+ *	   (c) -> a
+ *	   (c) -> b
+ */
+MVDependencies *
+statext_dependencies_build(StatsBuildData *data)
+{
+	int			i,
+				k;
+
+	/* result */
+	MVDependencies *dependencies = NULL;
+	MemoryContext	cxt;
+
+	Assert(data->nattnums >= 2);
+
+	/* tracks memory allocated by dependency_degree calls */
+	cxt = AllocSetContextCreate(CurrentMemoryContext,
+								"dependency_degree cxt",
+								ALLOCSET_DEFAULT_SIZES);
+
+	/*
+	 * We'll try build functional dependencies starting from the smallest ones
+	 * covering just 2 columns, to the largest ones, covering all columns
+	 * included in the statistics object.  We start from the smallest ones
+	 * because we want to be able to skip already implied ones.
+	 */
+	for (k = 2; k <= data->nattnums; k++)
+	{
+		AttrNumber *dependency; /* array with k elements */
+
+		/* prepare a DependencyGenerator of variation */
+		DependencyGenerator DependencyGenerator = DependencyGenerator_init(data->nattnums, k);
+
+		/* generate all possible variations of k values (out of n) */
+		while ((dependency = DependencyGenerator_next(DependencyGenerator)))
+		{
+			double		degree;
+			MVDependency *d;
+			MemoryContext oldcxt;
+
+			/* release memory used by dependency degree calculation */
+			oldcxt = MemoryContextSwitchTo(cxt);
+
+			/* compute how valid the dependency seems */
+			degree = dependency_degree(data, k, dependency);
+
+			MemoryContextSwitchTo(oldcxt);
+			MemoryContextReset(cxt);
+
+			/*
+			 * if the dependency seems entirely invalid, don't store it
+			 */
+			if (degree == 0.0)
+				continue;
+
+			d = (MVDependency *) palloc0(offsetof(MVDependency, attributes)
+										 + k * sizeof(AttrNumber));
+
+			/* copy the dependency (and keep the indexes into stxkeys) */
+			d->degree = degree;
+			d->nattributes = k;
+			for (i = 0; i < k; i++)
+				d->attributes[i] = data->attnums[dependency[i]];
+
+			/* initialize the list of dependencies */
+			if (dependencies == NULL)
+			{
+				dependencies
+					= (MVDependencies *) palloc0(sizeof(MVDependencies));
+
+				dependencies->magic = STATS_DEPS_MAGIC;
+				dependencies->type = STATS_DEPS_TYPE_BASIC;
+				dependencies->ndeps = 0;
+			}
+
+			dependencies->ndeps++;
+			dependencies = (MVDependencies *) repalloc(dependencies,
+													   offsetof(MVDependencies, deps)
+													   + dependencies->ndeps * sizeof(MVDependency *));
+
+			dependencies->deps[dependencies->ndeps - 1] = d;
+		}
+
+		/*
+		 * we're done with variations of k elements, so free the
+		 * DependencyGenerator
+		 */
+		DependencyGenerator_free(DependencyGenerator);
+	}
+
+	MemoryContextDelete(cxt);
+
+	return dependencies;
+}
+
+
+/*
+ * Serialize list of dependencies into a bytea value.
+ */
+bytea *
+statext_dependencies_serialize(MVDependencies *dependencies)
+{
+	int			i;
+	bytea	   *output;
+	char	   *tmp;
+	Size		len;
+
+	/* we need to store ndeps, with a number of attributes for each one */
+	len = VARHDRSZ + SizeOfHeader;
+
+	/* and also include space for the actual attribute numbers and degrees */
+	for (i = 0; i < dependencies->ndeps; i++)
+		len += SizeOfItem(dependencies->deps[i]->nattributes);
+
+	output = (bytea *) palloc0(len);
+	SET_VARSIZE(output, len);
+
+	tmp = VARDATA(output);
+
+	/* Store the base struct values (magic, type, ndeps) */
+	memcpy(tmp, &dependencies->magic, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(tmp, &dependencies->type, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(tmp, &dependencies->ndeps, sizeof(uint32));
+	tmp += sizeof(uint32);
+
+	/* store number of attributes and attribute numbers for each dependency */
+	for (i = 0; i < dependencies->ndeps; i++)
+	{
+		MVDependency *d = dependencies->deps[i];
+
+		memcpy(tmp, &d->degree, sizeof(double));
+		tmp += sizeof(double);
+
+		memcpy(tmp, &d->nattributes, sizeof(AttrNumber));
+		tmp += sizeof(AttrNumber);
+
+		memcpy(tmp, d->attributes, sizeof(AttrNumber) * d->nattributes);
+		tmp += sizeof(AttrNumber) * d->nattributes;
+
+		/* protect against overflow */
+		Assert(tmp <= ((char *) output + len));
+	}
+
+	/* make sure we've produced exactly the right amount of data */
+	Assert(tmp == ((char *) output + len));
+
+	return output;
+}
+
+/*
+ * Reads serialized dependencies into MVDependencies structure.
+ */
+MVDependencies *
+statext_dependencies_deserialize(bytea *data)
+{
+	int			i;
+	Size		min_expected_size;
+	MVDependencies *dependencies;
+	char	   *tmp;
+
+	if (data == NULL)
+		return NULL;
+
+	if (VARSIZE_ANY_EXHDR(data) < SizeOfHeader)
+		elog(ERROR, "invalid MVDependencies size %zd (expected at least %zd)",
+			 VARSIZE_ANY_EXHDR(data), SizeOfHeader);
+
+	/* read the MVDependencies header */
+	dependencies = (MVDependencies *) palloc0(sizeof(MVDependencies));
+
+	/* initialize pointer to the data part (skip the varlena header) */
+	tmp = VARDATA_ANY(data);
+
+	/* read the header fields and perform basic sanity checks */
+	memcpy(&dependencies->magic, tmp, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(&dependencies->type, tmp, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(&dependencies->ndeps, tmp, sizeof(uint32));
+	tmp += sizeof(uint32);
+
+	if (dependencies->magic != STATS_DEPS_MAGIC)
+		elog(ERROR, "invalid dependency magic %d (expected %d)",
+			 dependencies->magic, STATS_DEPS_MAGIC);
+
+	if (dependencies->type != STATS_DEPS_TYPE_BASIC)
+		elog(ERROR, "invalid dependency type %d (expected %d)",
+			 dependencies->type, STATS_DEPS_TYPE_BASIC);
+
+	if (dependencies->ndeps == 0)
+		elog(ERROR, "invalid zero-length item array in MVDependencies");
+
+	/* what minimum bytea size do we expect for those parameters */
+	min_expected_size = SizeOfItem(dependencies->ndeps);
+
+	if (VARSIZE_ANY_EXHDR(data) < min_expected_size)
+		elog(ERROR, "invalid dependencies size %zd (expected at least %zd)",
+			 VARSIZE_ANY_EXHDR(data), min_expected_size);
+
+	/* allocate space for the MCV items */
+	dependencies = repalloc(dependencies, offsetof(MVDependencies, deps)
+							+ (dependencies->ndeps * sizeof(MVDependency *)));
+
+	for (i = 0; i < dependencies->ndeps; i++)
+	{
+		double		degree;
+		AttrNumber	k;
+		MVDependency *d;
+
+		/* degree of validity */
+		memcpy(&degree, tmp, sizeof(double));
+		tmp += sizeof(double);
+
+		/* number of attributes */
+		memcpy(&k, tmp, sizeof(AttrNumber));
+		tmp += sizeof(AttrNumber);
+
+		/* is the number of attributes valid? */
+		Assert((k >= 2) && (k <= STATS_MAX_DIMENSIONS));
+
+		/* now that we know the number of attributes, allocate the dependency */
+		d = (MVDependency *) palloc0(offsetof(MVDependency, attributes)
+									 + (k * sizeof(AttrNumber)));
+
+		d->degree = degree;
+		d->nattributes = k;
+
+		/* copy attribute numbers */
+		memcpy(d->attributes, tmp, sizeof(AttrNumber) * d->nattributes);
+		tmp += sizeof(AttrNumber) * d->nattributes;
+
+		dependencies->deps[i] = d;
+
+		/* still within the bytea */
+		Assert(tmp <= ((char *) data + VARSIZE_ANY(data)));
+	}
+
+	/* we should have consumed the whole bytea exactly */
+	Assert(tmp == ((char *) data + VARSIZE_ANY(data)));
+
+	return dependencies;
+}
+
+/*
+ * dependency_is_fully_matched
+ *		checks that a functional dependency is fully matched given clauses on
+ *		attributes (assuming the clauses are suitable equality clauses)
+ */
+static bool
+dependency_is_fully_matched(MVDependency *dependency, Bitmapset *attnums)
+{
+	int			j;
+
+	/*
+	 * Check that the dependency actually is fully covered by clauses. We have
+	 * to translate all attribute numbers, as those are referenced
+	 */
+	for (j = 0; j < dependency->nattributes; j++)
+	{
+		int			attnum = dependency->attributes[j];
+
+		if (!bms_is_member(attnum, attnums))
+			return false;
+	}
+
+	return true;
+}
+
+/*
+ * statext_dependencies_load
+ *		Load the functional dependencies for the indicated pg_statistic_ext tuple
+ */
+MVDependencies *
+statext_dependencies_load(Oid mvoid)
+{
+	MVDependencies *result;
+	bool		isnull;
+	Datum		deps;
+	HeapTuple	htup;
+
+	htup = SearchSysCache1(STATEXTDATASTXOID, ObjectIdGetDatum(mvoid));
+	if (!HeapTupleIsValid(htup))
+		elog(ERROR, "cache lookup failed for statistics object %u", mvoid);
+
+	deps = SysCacheGetAttr(STATEXTDATASTXOID, htup,
+						   Anum_pg_statistic_ext_data_stxddependencies, &isnull);
+	if (isnull)
+		elog(ERROR,
+			 "requested statistics kind \"%c\" is not yet built for statistics object %u",
+			 STATS_EXT_DEPENDENCIES, mvoid);
+
+	result = statext_dependencies_deserialize(DatumGetByteaPP(deps));
+
+	ReleaseSysCache(htup);
+
+	return result;
+}
+
+/*
+ * pg_dependencies_in		- input routine for type pg_dependencies.
+ *
+ * pg_dependencies is real enough to be a table column, but it has no operations
+ * of its own, and disallows input too
+ */
+Datum
+pg_dependencies_in(PG_FUNCTION_ARGS)
+{
+	/*
+	 * pg_node_list stores the data in binary form and parsing text input is
+	 * not needed, so disallow this.
+	 */
+	ereport(ERROR,
+			(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+			 errmsg("cannot accept a value of type %s", "pg_dependencies")));
+
+	PG_RETURN_VOID();			/* keep compiler quiet */
+}
+
+/*
+ * pg_dependencies		- output routine for type pg_dependencies.
+ */
+Datum
+pg_dependencies_out(PG_FUNCTION_ARGS)
+{
+	bytea	   *data = PG_GETARG_BYTEA_PP(0);
+	MVDependencies *dependencies = statext_dependencies_deserialize(data);
+	int			i,
+				j;
+	StringInfoData str;
+
+	initStringInfo(&str);
+	appendStringInfoChar(&str, '{');
+
+	for (i = 0; i < dependencies->ndeps; i++)
+	{
+		MVDependency *dependency = dependencies->deps[i];
+
+		if (i > 0)
+			appendStringInfoString(&str, ", ");
+
+		appendStringInfoChar(&str, '"');
+		for (j = 0; j < dependency->nattributes; j++)
+		{
+			if (j == dependency->nattributes - 1)
+				appendStringInfoString(&str, " => ");
+			else if (j > 0)
+				appendStringInfoString(&str, ", ");
+
+			appendStringInfo(&str, "%d", dependency->attributes[j]);
+		}
+		appendStringInfo(&str, "\": %f", dependency->degree);
+	}
+
+	appendStringInfoChar(&str, '}');
+
+	PG_RETURN_CSTRING(str.data);
+}
+
+/*
+ * pg_dependencies_recv		- binary input routine for type pg_dependencies.
+ */
+Datum
+pg_dependencies_recv(PG_FUNCTION_ARGS)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+			 errmsg("cannot accept a value of type %s", "pg_dependencies")));
+
+	PG_RETURN_VOID();			/* keep compiler quiet */
+}
+
+/*
+ * pg_dependencies_send		- binary output routine for type pg_dependencies.
+ *
+ * Functional dependencies are serialized in a bytea value (although the type
+ * is named differently), so let's just send that.
+ */
+Datum
+pg_dependencies_send(PG_FUNCTION_ARGS)
+{
+	return byteasend(fcinfo);
+}
+
+/*
+ * dependency_is_compatible_clause
+ *		Determines if the clause is compatible with functional dependencies
+ *
+ * Only clauses that have the form of equality to a pseudoconstant, or can be
+ * interpreted that way, are currently accepted.  Furthermore the variable
+ * part of the clause must be a simple Var belonging to the specified
+ * relation, whose attribute number we return in *attnum on success.
+ */
+static bool
+dependency_is_compatible_clause(Node *clause, Index relid, AttrNumber *attnum)
+{
+	Var		   *var;
+	Node	   *clause_expr;
+
+	if (IsA(clause, RestrictInfo))
+	{
+		RestrictInfo *rinfo = (RestrictInfo *) clause;
+
+		/* Pseudoconstants are not interesting (they couldn't contain a Var) */
+		if (rinfo->pseudoconstant)
+			return false;
+
+		/* Clauses referencing multiple, or no, varnos are incompatible */
+		if (bms_membership(rinfo->clause_relids) != BMS_SINGLETON)
+			return false;
+
+		clause = (Node *) rinfo->clause;
+	}
+
+	if (is_opclause(clause))
+	{
+		/* If it's an opclause, check for Var = Const or Const = Var. */
+		OpExpr	   *expr = (OpExpr *) clause;
+
+		/* Only expressions with two arguments are candidates. */
+		if (list_length(expr->args) != 2)
+			return false;
+
+		/* Make sure non-selected argument is a pseudoconstant. */
+		if (is_pseudo_constant_clause(lsecond(expr->args)))
+			clause_expr = linitial(expr->args);
+		else if (is_pseudo_constant_clause(linitial(expr->args)))
+			clause_expr = lsecond(expr->args);
+		else
+			return false;
+
+		/*
+		 * If it's not an "=" operator, just ignore the clause, as it's not
+		 * compatible with functional dependencies.
+		 *
+		 * This uses the function for estimating selectivity, not the operator
+		 * directly (a bit awkward, but well ...).
+		 *
+		 * XXX this is pretty dubious; probably it'd be better to check btree
+		 * or hash opclass membership, so as not to be fooled by custom
+		 * selectivity functions, and to be more consistent with decisions
+		 * elsewhere in the planner.
+		 */
+		if (get_oprrest(expr->opno) != F_EQSEL)
+			return false;
+
+		/* OK to proceed with checking "var" */
+	}
+	else if (IsA(clause, ScalarArrayOpExpr))
+	{
+		/* If it's an scalar array operator, check for Var IN Const. */
+		ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) clause;
+
+		/*
+		 * Reject ALL() variant, we only care about ANY/IN.
+		 *
+		 * XXX Maybe we should check if all the values are the same, and allow
+		 * ALL in that case? Doesn't seem very practical, though.
+		 */
+		if (!expr->useOr)
+			return false;
+
+		/* Only expressions with two arguments are candidates. */
+		if (list_length(expr->args) != 2)
+			return false;
+
+		/*
+		 * We know it's always (Var IN Const), so we assume the var is the
+		 * first argument, and pseudoconstant is the second one.
+		 */
+		if (!is_pseudo_constant_clause(lsecond(expr->args)))
+			return false;
+
+		clause_expr = linitial(expr->args);
+
+		/*
+		 * If it's not an "=" operator, just ignore the clause, as it's not
+		 * compatible with functional dependencies. The operator is identified
+		 * simply by looking at which function it uses to estimate
+		 * selectivity. That's a bit strange, but it's what other similar
+		 * places do.
+		 */
+		if (get_oprrest(expr->opno) != F_EQSEL)
+			return false;
+
+		/* OK to proceed with checking "var" */
+	}
+	else if (is_orclause(clause))
+	{
+		BoolExpr   *bool_expr = (BoolExpr *) clause;
+		ListCell   *lc;
+
+		/* start with no attribute number */
+		*attnum = InvalidAttrNumber;
+
+		foreach(lc, bool_expr->args)
+		{
+			AttrNumber	clause_attnum;
+
+			/*
+			 * Had we found incompatible clause in the arguments, treat the
+			 * whole clause as incompatible.
+			 */
+			if (!dependency_is_compatible_clause((Node *) lfirst(lc),
+												 relid, &clause_attnum))
+				return false;
+
+			if (*attnum == InvalidAttrNumber)
+				*attnum = clause_attnum;
+
+			/* ensure all the variables are the same (same attnum) */
+			if (*attnum != clause_attnum)
+				return false;
+		}
+
+		/* the Var is already checked by the recursive call */
+		return true;
+	}
+	else if (is_notclause(clause))
+	{
+		/*
+		 * "NOT x" can be interpreted as "x = false", so get the argument and
+		 * proceed with seeing if it's a suitable Var.
+		 */
+		clause_expr = (Node *) get_notclausearg(clause);
+	}
+	else
+	{
+		/*
+		 * A boolean expression "x" can be interpreted as "x = true", so
+		 * proceed with seeing if it's a suitable Var.
+		 */
+		clause_expr = (Node *) clause;
+	}
+
+	/*
+	 * We may ignore any RelabelType node above the operand.  (There won't be
+	 * more than one, since eval_const_expressions has been applied already.)
+	 */
+	if (IsA(clause_expr, RelabelType))
+		clause_expr = (Node *) ((RelabelType *) clause_expr)->arg;
+
+	/* We only support plain Vars for now */
+	if (!IsA(clause_expr, Var))
+		return false;
+
+	/* OK, we know we have a Var */
+	var = (Var *) clause_expr;
+
+	/* Ensure Var is from the correct relation */
+	if (var->varno != relid)
+		return false;
+
+	/* We also better ensure the Var is from the current level */
+	if (var->varlevelsup != 0)
+		return false;
+
+	/* Also ignore system attributes (we don't allow stats on those) */
+	if (!AttrNumberIsForUserDefinedAttr(var->varattno))
+		return false;
+
+	*attnum = var->varattno;
+	return true;
+}
+
+/*
+ * find_strongest_dependency
+ *		find the strongest dependency on the attributes
+ *
+ * When applying functional dependencies, we start with the strongest
+ * dependencies. That is, we select the dependency that:
+ *
+ * (a) has all attributes covered by equality clauses
+ *
+ * (b) has the most attributes
+ *
+ * (c) has the highest degree of validity
+ *
+ * This guarantees that we eliminate the most redundant conditions first
+ * (see the comment in dependencies_clauselist_selectivity).
+ */
+static MVDependency *
+find_strongest_dependency(MVDependencies **dependencies, int ndependencies,
+						  Bitmapset *attnums)
+{
+	int			i,
+				j;
+	MVDependency *strongest = NULL;
+
+	/* number of attnums in clauses */
+	int			nattnums = bms_num_members(attnums);
+
+	/*
+	 * Iterate over the MVDependency items and find the strongest one from the
+	 * fully-matched dependencies. We do the cheap checks first, before
+	 * matching it against the attnums.
+	 */
+	for (i = 0; i < ndependencies; i++)
+	{
+		for (j = 0; j < dependencies[i]->ndeps; j++)
+		{
+			MVDependency *dependency = dependencies[i]->deps[j];
+
+			/*
+			 * Skip dependencies referencing more attributes than available
+			 * clauses, as those can't be fully matched.
+			 */
+			if (dependency->nattributes > nattnums)
+				continue;
+
+			if (strongest)
+			{
+				/* skip dependencies on fewer attributes than the strongest. */
+				if (dependency->nattributes < strongest->nattributes)
+					continue;
+
+				/* also skip weaker dependencies when attribute count matches */
+				if (strongest->nattributes == dependency->nattributes &&
+					strongest->degree > dependency->degree)
+					continue;
+			}
+
+			/*
+			 * this dependency is stronger, but we must still check that it's
+			 * fully matched to these attnums. We perform this check last as
+			 * it's slightly more expensive than the previous checks.
+			 */
+			if (dependency_is_fully_matched(dependency, attnums))
+				strongest = dependency; /* save new best match */
+		}
+	}
+
+	return strongest;
+}
+
+/*
+ * clauselist_apply_dependencies
+ *		Apply the specified functional dependencies to a list of clauses and
+ *		return the estimated selectivity of the clauses that are compatible
+ *		with any of the given dependencies.
+ *
+ * This will estimate all not-already-estimated clauses that are compatible
+ * with functional dependencies, and which have an attribute mentioned by any
+ * of the given dependencies (either as an implying or implied attribute).
+ *
+ * Given (lists of) clauses on attributes (a,b) and a functional dependency
+ * (a=>b), the per-column selectivities P(a) and P(b) are notionally combined
+ * using the formula
+ *
+ *		P(a,b) = f * P(a) + (1-f) * P(a) * P(b)
+ *
+ * where 'f' is the degree of dependency.  This reflects the fact that we
+ * expect a fraction f of all rows to be consistent with the dependency
+ * (a=>b), and so have a selectivity of P(a), while the remaining rows are
+ * treated as independent.
+ *
+ * In practice, we use a slightly modified version of this formula, which uses
+ * a selectivity of Min(P(a), P(b)) for the dependent rows, since the result
+ * should obviously not exceed either column's individual selectivity.  I.e.,
+ * we actually combine selectivities using the formula
+ *
+ *		P(a,b) = f * Min(P(a), P(b)) + (1-f) * P(a) * P(b)
+ *
+ * This can make quite a difference if the specific values matching the
+ * clauses are not consistent with the functional dependency.
+ */
+static Selectivity
+clauselist_apply_dependencies(PlannerInfo *root, List *clauses,
+							  int varRelid, JoinType jointype,
+							  SpecialJoinInfo *sjinfo,
+							  MVDependency **dependencies, int ndependencies,
+							  AttrNumber *list_attnums,
+							  Bitmapset **estimatedclauses)
+{
+	Bitmapset  *attnums;
+	int			i;
+	int			j;
+	int			nattrs;
+	Selectivity *attr_sel;
+	int			attidx;
+	int			listidx;
+	ListCell   *l;
+	Selectivity s1;
+
+	/*
+	 * Extract the attnums of all implying and implied attributes from all the
+	 * given dependencies.  Each of these attributes is expected to have at
+	 * least 1 not-already-estimated compatible clause that we will estimate
+	 * here.
+	 */
+	attnums = NULL;
+	for (i = 0; i < ndependencies; i++)
+	{
+		for (j = 0; j < dependencies[i]->nattributes; j++)
+		{
+			AttrNumber	attnum = dependencies[i]->attributes[j];
+
+			attnums = bms_add_member(attnums, attnum);
+		}
+	}
+
+	/*
+	 * Compute per-column selectivity estimates for each of these attributes,
+	 * and mark all the corresponding clauses as estimated.
+	 */
+	nattrs = bms_num_members(attnums);
+	attr_sel = (Selectivity *) palloc(sizeof(Selectivity) * nattrs);
+
+	attidx = 0;
+	i = -1;
+	while ((i = bms_next_member(attnums, i)) >= 0)
+	{
+		List	   *attr_clauses = NIL;
+		Selectivity simple_sel;
+
+		listidx = -1;
+		foreach(l, clauses)
+		{
+			Node	   *clause = (Node *) lfirst(l);
+
+			listidx++;
+			if (list_attnums[listidx] == i)
+			{
+				attr_clauses = lappend(attr_clauses, clause);
+				*estimatedclauses = bms_add_member(*estimatedclauses, listidx);
+			}
+		}
+
+		simple_sel = clauselist_selectivity_ext(root, attr_clauses, varRelid,
+												jointype, sjinfo, false);
+		attr_sel[attidx++] = simple_sel;
+	}
+
+	/*
+	 * Now combine these selectivities using the dependency information.  For
+	 * chains of dependencies such as a -> b -> c, the b -> c dependency will
+	 * come before the a -> b dependency in the array, so we traverse the
+	 * array backwards to ensure such chains are computed in the right order.
+	 *
+	 * As explained above, pairs of selectivities are combined using the
+	 * formula
+	 *
+	 * P(a,b) = f * Min(P(a), P(b)) + (1-f) * P(a) * P(b)
+	 *
+	 * to ensure that the combined selectivity is never greater than either
+	 * individual selectivity.
+	 *
+	 * Where multiple dependencies apply (e.g., a -> b -> c), we use
+	 * conditional probabilities to compute the overall result as follows:
+	 *
+	 * P(a,b,c) = P(c|a,b) * P(a,b) = P(c|a,b) * P(b|a) * P(a)
+	 *
+	 * so we replace the selectivities of all implied attributes with
+	 * conditional probabilities, that are conditional on all their implying
+	 * attributes.  The selectivities of all other non-implied attributes are
+	 * left as they are.
+	 */
+	for (i = ndependencies - 1; i >= 0; i--)
+	{
+		MVDependency *dependency = dependencies[i];
+		AttrNumber	attnum;
+		Selectivity s2;
+		double		f;
+
+		/* Selectivity of all the implying attributes */
+		s1 = 1.0;
+		for (j = 0; j < dependency->nattributes - 1; j++)
+		{
+			attnum = dependency->attributes[j];
+			attidx = bms_member_index(attnums, attnum);
+			s1 *= attr_sel[attidx];
+		}
+
+		/* Original selectivity of the implied attribute */
+		attnum = dependency->attributes[j];
+		attidx = bms_member_index(attnums, attnum);
+		s2 = attr_sel[attidx];
+
+		/*
+		 * Replace s2 with the conditional probability s2 given s1, computed
+		 * using the formula P(b|a) = P(a,b) / P(a), which simplifies to
+		 *
+		 * P(b|a) = f * Min(P(a), P(b)) / P(a) + (1-f) * P(b)
+		 *
+		 * where P(a) = s1, the selectivity of the implying attributes, and
+		 * P(b) = s2, the selectivity of the implied attribute.
+		 */
+		f = dependency->degree;
+
+		if (s1 <= s2)
+			attr_sel[attidx] = f + (1 - f) * s2;
+		else
+			attr_sel[attidx] = f * s2 / s1 + (1 - f) * s2;
+	}
+
+	/*
+	 * The overall selectivity of all the clauses on all these attributes is
+	 * then the product of all the original (non-implied) probabilities and
+	 * the new conditional (implied) probabilities.
+	 */
+	s1 = 1.0;
+	for (i = 0; i < nattrs; i++)
+		s1 *= attr_sel[i];
+
+	CLAMP_PROBABILITY(s1);
+
+	pfree(attr_sel);
+	bms_free(attnums);
+
+	return s1;
+}
+
+/*
+ * dependency_is_compatible_expression
+ *		Determines if the expression is compatible with functional dependencies
+ *
+ * Similar to dependency_is_compatible_clause, but doesn't enforce that the
+ * expression is a simple Var. OTOH we check that there's at least one
+ * statistics object matching the expression.
+ */
+static bool
+dependency_is_compatible_expression(Node *clause, Index relid, List *statlist, Node **expr)
+{
+	List	   *vars;
+	ListCell   *lc,
+			   *lc2;
+	Node	   *clause_expr;
+
+	if (IsA(clause, RestrictInfo))
+	{
+		RestrictInfo *rinfo = (RestrictInfo *) clause;
+
+		/* Pseudoconstants are not interesting (they couldn't contain a Var) */
+		if (rinfo->pseudoconstant)
+			return false;
+
+		/* Clauses referencing multiple, or no, varnos are incompatible */
+		if (bms_membership(rinfo->clause_relids) != BMS_SINGLETON)
+			return false;
+
+		clause = (Node *) rinfo->clause;
+	}
+
+	if (is_opclause(clause))
+	{
+		/* If it's an opclause, check for Var = Const or Const = Var. */
+		OpExpr	   *expr = (OpExpr *) clause;
+
+		/* Only expressions with two arguments are candidates. */
+		if (list_length(expr->args) != 2)
+			return false;
+
+		/* Make sure non-selected argument is a pseudoconstant. */
+		if (is_pseudo_constant_clause(lsecond(expr->args)))
+			clause_expr = linitial(expr->args);
+		else if (is_pseudo_constant_clause(linitial(expr->args)))
+			clause_expr = lsecond(expr->args);
+		else
+			return false;
+
+		/*
+		 * If it's not an "=" operator, just ignore the clause, as it's not
+		 * compatible with functional dependencies.
+		 *
+		 * This uses the function for estimating selectivity, not the operator
+		 * directly (a bit awkward, but well ...).
+		 *
+		 * XXX this is pretty dubious; probably it'd be better to check btree
+		 * or hash opclass membership, so as not to be fooled by custom
+		 * selectivity functions, and to be more consistent with decisions
+		 * elsewhere in the planner.
+		 */
+		if (get_oprrest(expr->opno) != F_EQSEL)
+			return false;
+
+		/* OK to proceed with checking "var" */
+	}
+	else if (IsA(clause, ScalarArrayOpExpr))
+	{
+		/* If it's an scalar array operator, check for Var IN Const. */
+		ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) clause;
+
+		/*
+		 * Reject ALL() variant, we only care about ANY/IN.
+		 *
+		 * FIXME Maybe we should check if all the values are the same, and
+		 * allow ALL in that case? Doesn't seem very practical, though.
+		 */
+		if (!expr->useOr)
+			return false;
+
+		/* Only expressions with two arguments are candidates. */
+		if (list_length(expr->args) != 2)
+			return false;
+
+		/*
+		 * We know it's always (Var IN Const), so we assume the var is the
+		 * first argument, and pseudoconstant is the second one.
+		 */
+		if (!is_pseudo_constant_clause(lsecond(expr->args)))
+			return false;
+
+		clause_expr = linitial(expr->args);
+
+		/*
+		 * If it's not an "=" operator, just ignore the clause, as it's not
+		 * compatible with functional dependencies. The operator is identified
+		 * simply by looking at which function it uses to estimate
+		 * selectivity. That's a bit strange, but it's what other similar
+		 * places do.
+		 */
+		if (get_oprrest(expr->opno) != F_EQSEL)
+			return false;
+
+		/* OK to proceed with checking "var" */
+	}
+	else if (is_orclause(clause))
+	{
+		BoolExpr   *bool_expr = (BoolExpr *) clause;
+		ListCell   *lc;
+
+		/* start with no expression (we'll use the first match) */
+		*expr = NULL;
+
+		foreach(lc, bool_expr->args)
+		{
+			Node	   *or_expr = NULL;
+
+			/*
+			 * Had we found incompatible expression in the arguments, treat
+			 * the whole expression as incompatible.
+			 */
+			if (!dependency_is_compatible_expression((Node *) lfirst(lc), relid,
+													 statlist, &or_expr))
+				return false;
+
+			if (*expr == NULL)
+				*expr = or_expr;
+
+			/* ensure all the expressions are the same */
+			if (!equal(or_expr, *expr))
+				return false;
+		}
+
+		/* the expression is already checked by the recursive call */
+		return true;
+	}
+	else if (is_notclause(clause))
+	{
+		/*
+		 * "NOT x" can be interpreted as "x = false", so get the argument and
+		 * proceed with seeing if it's a suitable Var.
+		 */
+		clause_expr = (Node *) get_notclausearg(clause);
+	}
+	else
+	{
+		/*
+		 * A boolean expression "x" can be interpreted as "x = true", so
+		 * proceed with seeing if it's a suitable Var.
+		 */
+		clause_expr = (Node *) clause;
+	}
+
+	/*
+	 * We may ignore any RelabelType node above the operand.  (There won't be
+	 * more than one, since eval_const_expressions has been applied already.)
+	 */
+	if (IsA(clause_expr, RelabelType))
+		clause_expr = (Node *) ((RelabelType *) clause_expr)->arg;
+
+	vars = pull_var_clause(clause_expr, 0);
+
+	foreach(lc, vars)
+	{
+		Var		   *var = (Var *) lfirst(lc);
+
+		/* Ensure Var is from the correct relation */
+		if (var->varno != relid)
+			return false;
+
+		/* We also better ensure the Var is from the current level */
+		if (var->varlevelsup != 0)
+			return false;
+
+		/* Also ignore system attributes (we don't allow stats on those) */
+		if (!AttrNumberIsForUserDefinedAttr(var->varattno))
+			return false;
+	}
+
+	/*
+	 * Check if we actually have a matching statistics for the expression.
+	 *
+	 * XXX Maybe this is an overkill. We'll eliminate the expressions later.
+	 */
+	foreach(lc, statlist)
+	{
+		StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
+
+		/* ignore stats without dependencies */
+		if (info->kind != STATS_EXT_DEPENDENCIES)
+			continue;
+
+		foreach(lc2, info->exprs)
+		{
+			Node	   *stat_expr = (Node *) lfirst(lc2);
+
+			if (equal(clause_expr, stat_expr))
+			{
+				*expr = stat_expr;
+				return true;
+			}
+		}
+	}
+
+	return false;
+}
+
+/*
+ * dependencies_clauselist_selectivity
+ *		Return the estimated selectivity of (a subset of) the given clauses
+ *		using functional dependency statistics, or 1.0 if no useful functional
+ *		dependency statistic exists.
+ *
+ * 'estimatedclauses' is an input/output argument that gets a bit set
+ * corresponding to the (zero-based) list index of each clause that is included
+ * in the estimated selectivity.
+ *
+ * Given equality clauses on attributes (a,b) we find the strongest dependency
+ * between them, i.e. either (a=>b) or (b=>a). Assuming (a=>b) is the selected
+ * dependency, we then combine the per-clause selectivities using the formula
+ *
+ *	   P(a,b) = f * P(a) + (1-f) * P(a) * P(b)
+ *
+ * where 'f' is the degree of the dependency.  (Actually we use a slightly
+ * modified version of this formula -- see clauselist_apply_dependencies()).
+ *
+ * With clauses on more than two attributes, the dependencies are applied
+ * recursively, starting with the widest/strongest dependencies. For example
+ * P(a,b,c) is first split like this:
+ *
+ *	   P(a,b,c) = f * P(a,b) + (1-f) * P(a,b) * P(c)
+ *
+ * assuming (a,b=>c) is the strongest dependency.
+ */
+Selectivity
+dependencies_clauselist_selectivity(PlannerInfo *root,
+									List *clauses,
+									int varRelid,
+									JoinType jointype,
+									SpecialJoinInfo *sjinfo,
+									RelOptInfo *rel,
+									Bitmapset **estimatedclauses)
+{
+	Selectivity s1 = 1.0;
+	ListCell   *l;
+	Bitmapset  *clauses_attnums = NULL;
+	AttrNumber *list_attnums;
+	int			listidx;
+	MVDependencies **func_dependencies;
+	int			nfunc_dependencies;
+	int			total_ndeps;
+	MVDependency **dependencies;
+	int			ndependencies;
+	int			i;
+	AttrNumber	attnum_offset;
+	RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
+
+	/* unique expressions */
+	Node	  **unique_exprs;
+	int			unique_exprs_cnt;
+
+	/*
+	 * When dealing with regular inheritance trees, ignore extended stats
+	 * (which were built without data from child rels, and thus do not
+	 * represent them). For partitioned tables data there's no data in the
+	 * non-leaf relations, so we build stats only for the inheritance tree.
+	 * So for partitioned tables we do consider extended stats.
+	 */
+	if (rte->inh && rte->relkind != RELKIND_PARTITIONED_TABLE)
+		return 1.0;
+
+	/* check if there's any stats that might be useful for us. */
+	if (!has_stats_of_kind(rel->statlist, STATS_EXT_DEPENDENCIES))
+		return 1.0;
+
+	list_attnums = (AttrNumber *) palloc(sizeof(AttrNumber) *
+										 list_length(clauses));
+
+	/*
+	 * We allocate space as if every clause was a unique expression, although
+	 * that's probably overkill. Some will be simple column references that
+	 * we'll translate to attnums, and there might be duplicates. But it's
+	 * easier and cheaper to just do one allocation than repalloc later.
+	 */
+	unique_exprs = (Node **) palloc(sizeof(Node *) * list_length(clauses));
+	unique_exprs_cnt = 0;
+
+	/*
+	 * Pre-process the clauses list to extract the attnums seen in each item.
+	 * We need to determine if there's any clauses which will be useful for
+	 * dependency selectivity estimations. Along the way we'll record all of
+	 * the attnums for each clause in a list which we'll reference later so we
+	 * don't need to repeat the same work again. We'll also keep track of all
+	 * attnums seen.
+	 *
+	 * We also skip clauses that we already estimated using different types of
+	 * statistics (we treat them as incompatible).
+	 *
+	 * To handle expressions, we assign them negative attnums, as if it was a
+	 * system attribute (this is fine, as we only allow extended stats on user
+	 * attributes). And then we offset everything by the number of
+	 * expressions, so that we can store the values in a bitmapset.
+	 */
+	listidx = 0;
+	foreach(l, clauses)
+	{
+		Node	   *clause = (Node *) lfirst(l);
+		AttrNumber	attnum;
+		Node	   *expr = NULL;
+
+		/* ignore clause by default */
+		list_attnums[listidx] = InvalidAttrNumber;
+
+		if (!bms_is_member(listidx, *estimatedclauses))
+		{
+			/*
+			 * If it's a simple column reference, just extract the attnum. If
+			 * it's an expression, assign a negative attnum as if it was a
+			 * system attribute.
+			 */
+			if (dependency_is_compatible_clause(clause, rel->relid, &attnum))
+			{
+				list_attnums[listidx] = attnum;
+			}
+			else if (dependency_is_compatible_expression(clause, rel->relid,
+														 rel->statlist,
+														 &expr))
+			{
+				/* special attnum assigned to this expression */
+				attnum = InvalidAttrNumber;
+
+				Assert(expr != NULL);
+
+				/* If the expression is duplicate, use the same attnum. */
+				for (i = 0; i < unique_exprs_cnt; i++)
+				{
+					if (equal(unique_exprs[i], expr))
+					{
+						/* negative attribute number to expression */
+						attnum = -(i + 1);
+						break;
+					}
+				}
+
+				/* not found in the list, so add it */
+				if (attnum == InvalidAttrNumber)
+				{
+					unique_exprs[unique_exprs_cnt++] = expr;
+
+					/* after incrementing the value, to get -1, -2, ... */
+					attnum = (-unique_exprs_cnt);
+				}
+
+				/* remember which attnum was assigned to this clause */
+				list_attnums[listidx] = attnum;
+			}
+		}
+
+		listidx++;
+	}
+
+	Assert(listidx == list_length(clauses));
+
+	/*
+	 * How much we need to offset the attnums? If there are no expressions,
+	 * then no offset is needed. Otherwise we need to offset enough for the
+	 * lowest value (-unique_exprs_cnt) to become 1.
+	 */
+	if (unique_exprs_cnt > 0)
+		attnum_offset = (unique_exprs_cnt + 1);
+	else
+		attnum_offset = 0;
+
+	/*
+	 * Now that we know how many expressions there are, we can offset the
+	 * values just enough to build the bitmapset.
+	 */
+	for (i = 0; i < list_length(clauses); i++)
+	{
+		AttrNumber	attnum;
+
+		/* ignore incompatible or already estimated clauses */
+		if (list_attnums[i] == InvalidAttrNumber)
+			continue;
+
+		/* make sure the attnum is in the expected range */
+		Assert(list_attnums[i] >= (-unique_exprs_cnt));
+		Assert(list_attnums[i] <= MaxHeapAttributeNumber);
+
+		/* make sure the attnum is positive (valid AttrNumber) */
+		attnum = list_attnums[i] + attnum_offset;
+
+		/*
+		 * Either it's a regular attribute, or it's an expression, in which
+		 * case we must not have seen it before (expressions are unique).
+		 *
+		 * XXX Check whether it's a regular attribute has to be done using the
+		 * original attnum, while the second check has to use the value with
+		 * an offset.
+		 */
+		Assert(AttrNumberIsForUserDefinedAttr(list_attnums[i]) ||
+			   !bms_is_member(attnum, clauses_attnums));
+
+		/*
+		 * Remember the offset attnum, both for attributes and expressions.
+		 * We'll pass list_attnums to clauselist_apply_dependencies, which
+		 * uses it to identify clauses in a bitmap. We could also pass the
+		 * offset, but this is more convenient.
+		 */
+		list_attnums[i] = attnum;
+
+		clauses_attnums = bms_add_member(clauses_attnums, attnum);
+	}
+
+	/*
+	 * If there's not at least two distinct attnums and expressions, then
+	 * reject the whole list of clauses. We must return 1.0 so the calling
+	 * function's selectivity is unaffected.
+	 */
+	if (bms_membership(clauses_attnums) != BMS_MULTIPLE)
+	{
+		bms_free(clauses_attnums);
+		pfree(list_attnums);
+		return 1.0;
+	}
+
+	/*
+	 * Load all functional dependencies matching at least two parameters. We
+	 * can simply consider all dependencies at once, without having to search
+	 * for the best statistics object.
+	 *
+	 * To not waste cycles and memory, we deserialize dependencies only for
+	 * statistics that match at least two attributes. The array is allocated
+	 * with the assumption that all objects match - we could grow the array to
+	 * make it just the right size, but it's likely wasteful anyway thanks to
+	 * moving the freed chunks to freelists etc.
+	 */
+	func_dependencies = (MVDependencies **) palloc(sizeof(MVDependencies *) *
+												   list_length(rel->statlist));
+	nfunc_dependencies = 0;
+	total_ndeps = 0;
+
+	foreach(l, rel->statlist)
+	{
+		StatisticExtInfo *stat = (StatisticExtInfo *) lfirst(l);
+		int			nmatched;
+		int			nexprs;
+		int			k;
+		MVDependencies *deps;
+
+		/* skip statistics that are not of the correct type */
+		if (stat->kind != STATS_EXT_DEPENDENCIES)
+			continue;
+
+		/*
+		 * Count matching attributes - we have to undo the attnum offsets. The
+		 * input attribute numbers are not offset (expressions are not
+		 * included in stat->keys, so it's not necessary). But we need to
+		 * offset it before checking against clauses_attnums.
+		 */
+		nmatched = 0;
+		k = -1;
+		while ((k = bms_next_member(stat->keys, k)) >= 0)
+		{
+			AttrNumber	attnum = (AttrNumber) k;
+
+			/* skip expressions */
+			if (!AttrNumberIsForUserDefinedAttr(attnum))
+				continue;
+
+			/* apply the same offset as above */
+			attnum += attnum_offset;
+
+			if (bms_is_member(attnum, clauses_attnums))
+				nmatched++;
+		}
+
+		/* count matching expressions */
+		nexprs = 0;
+		for (i = 0; i < unique_exprs_cnt; i++)
+		{
+			ListCell   *lc;
+
+			foreach(lc, stat->exprs)
+			{
+				Node	   *stat_expr = (Node *) lfirst(lc);
+
+				/* try to match it */
+				if (equal(stat_expr, unique_exprs[i]))
+					nexprs++;
+			}
+		}
+
+		/*
+		 * Skip objects matching fewer than two attributes/expressions from
+		 * clauses.
+		 */
+		if (nmatched + nexprs < 2)
+			continue;
+
+		deps = statext_dependencies_load(stat->statOid);
+
+		/*
+		 * The expressions may be represented by different attnums in the
+		 * stats, we need to remap them to be consistent with the clauses.
+		 * That will make the later steps (e.g. picking the strongest item and
+		 * so on) much simpler and cheaper, because it won't need to care
+		 * about the offset at all.
+		 *
+		 * When we're at it, we can ignore dependencies that are not fully
+		 * matched by clauses (i.e. referencing attributes or expressions that
+		 * are not in the clauses).
+		 *
+		 * We have to do this for all statistics, as long as there are any
+		 * expressions - we need to shift the attnums in all dependencies.
+		 *
+		 * XXX Maybe we should do this always, because it also eliminates some
+		 * of the dependencies early. It might be cheaper than having to walk
+		 * the longer list in find_strongest_dependency later, especially as
+		 * we need to do that repeatedly?
+		 *
+		 * XXX We have to do this even when there are no expressions in
+		 * clauses, otherwise find_strongest_dependency may fail for stats
+		 * with expressions (due to lookup of negative value in bitmap). So we
+		 * need to at least filter out those dependencies. Maybe we could do
+		 * it in a cheaper way (if there are no expr clauses, we can just
+		 * discard all negative attnums without any lookups).
+		 */
+		if (unique_exprs_cnt > 0 || stat->exprs != NIL)
+		{
+			int			ndeps = 0;
+
+			for (i = 0; i < deps->ndeps; i++)
+			{
+				bool		skip = false;
+				MVDependency *dep = deps->deps[i];
+				int			j;
+
+				for (j = 0; j < dep->nattributes; j++)
+				{
+					int			idx;
+					Node	   *expr;
+					int			k;
+					AttrNumber	unique_attnum = InvalidAttrNumber;
+					AttrNumber	attnum;
+
+					/* undo the per-statistics offset */
+					attnum = dep->attributes[j];
+
+					/*
+					 * For regular attributes we can simply check if it
+					 * matches any clause. If there's no matching clause, we
+					 * can just ignore it. We need to offset the attnum
+					 * though.
+					 */
+					if (AttrNumberIsForUserDefinedAttr(attnum))
+					{
+						dep->attributes[j] = attnum + attnum_offset;
+
+						if (!bms_is_member(dep->attributes[j], clauses_attnums))
+						{
+							skip = true;
+							break;
+						}
+
+						continue;
+					}
+
+					/*
+					 * the attnum should be a valid system attnum (-1, -2,
+					 * ...)
+					 */
+					Assert(AttributeNumberIsValid(attnum));
+
+					/*
+					 * For expressions, we need to do two translations. First
+					 * we have to translate the negative attnum to index in
+					 * the list of expressions (in the statistics object).
+					 * Then we need to see if there's a matching clause. The
+					 * index of the unique expression determines the attnum
+					 * (and we offset it).
+					 */
+					idx = -(1 + attnum);
+
+					/* Is the expression index is valid? */
+					Assert((idx >= 0) && (idx < list_length(stat->exprs)));
+
+					expr = (Node *) list_nth(stat->exprs, idx);
+
+					/* try to find the expression in the unique list */
+					for (k = 0; k < unique_exprs_cnt; k++)
+					{
+						/*
+						 * found a matching unique expression, use the attnum
+						 * (derived from index of the unique expression)
+						 */
+						if (equal(unique_exprs[k], expr))
+						{
+							unique_attnum = -(k + 1) + attnum_offset;
+							break;
+						}
+					}
+
+					/*
+					 * Found no matching expression, so we can simply skip
+					 * this dependency, because there's no chance it will be
+					 * fully covered.
+					 */
+					if (unique_attnum == InvalidAttrNumber)
+					{
+						skip = true;
+						break;
+					}
+
+					/* otherwise remap it to the new attnum */
+					dep->attributes[j] = unique_attnum;
+				}
+
+				/* if found a matching dependency, keep it */
+				if (!skip)
+				{
+					/* maybe we've skipped something earlier, so move it */
+					if (ndeps != i)
+						deps->deps[ndeps] = deps->deps[i];
+
+					ndeps++;
+				}
+			}
+
+			deps->ndeps = ndeps;
+		}
+
+		/*
+		 * It's possible we've removed all dependencies, in which case we
+		 * don't bother adding it to the list.
+		 */
+		if (deps->ndeps > 0)
+		{
+			func_dependencies[nfunc_dependencies] = deps;
+			total_ndeps += deps->ndeps;
+			nfunc_dependencies++;
+		}
+	}
+
+	/* if no matching stats could be found then we've nothing to do */
+	if (nfunc_dependencies == 0)
+	{
+		pfree(func_dependencies);
+		bms_free(clauses_attnums);
+		pfree(list_attnums);
+		pfree(unique_exprs);
+		return 1.0;
+	}
+
+	/*
+	 * Work out which dependencies we can apply, starting with the
+	 * widest/strongest ones, and proceeding to smaller/weaker ones.
+	 */
+	dependencies = (MVDependency **) palloc(sizeof(MVDependency *) *
+											total_ndeps);
+	ndependencies = 0;
+
+	while (true)
+	{
+		MVDependency *dependency;
+		AttrNumber	attnum;
+
+		/* the widest/strongest dependency, fully matched by clauses */
+		dependency = find_strongest_dependency(func_dependencies,
+											   nfunc_dependencies,
+											   clauses_attnums);
+		if (!dependency)
+			break;
+
+		dependencies[ndependencies++] = dependency;
+
+		/* Ignore dependencies using this implied attribute in later loops */
+		attnum = dependency->attributes[dependency->nattributes - 1];
+		clauses_attnums = bms_del_member(clauses_attnums, attnum);
+	}
+
+	/*
+	 * If we found applicable dependencies, use them to estimate all
+	 * compatible clauses on attributes that they refer to.
+	 */
+	if (ndependencies != 0)
+		s1 = clauselist_apply_dependencies(root, clauses, varRelid, jointype,
+										   sjinfo, dependencies, ndependencies,
+										   list_attnums, estimatedclauses);
+
+	/* free deserialized functional dependencies (and then the array) */
+	for (i = 0; i < nfunc_dependencies; i++)
+		pfree(func_dependencies[i]);
+
+	pfree(dependencies);
+	pfree(func_dependencies);
+	bms_free(clauses_attnums);
+	pfree(list_attnums);
+	pfree(unique_exprs);
+
+	return s1;
+}
diff --git a/src/backend/statistics/extended_stats.c b/src/backend/statistics/extended_stats.c
new file mode 100644
index 0000000..d28c9d9
--- /dev/null
+++ b/src/backend/statistics/extended_stats.c
@@ -0,0 +1,2675 @@
+/*-------------------------------------------------------------------------
+ *
+ * extended_stats.c
+ *	  POSTGRES extended statistics
+ *
+ * Generic code supporting statistics objects created via CREATE STATISTICS.
+ *
+ *
+ * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * IDENTIFICATION
+ *	  src/backend/statistics/extended_stats.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include "access/detoast.h"
+#include "access/genam.h"
+#include "access/htup_details.h"
+#include "access/table.h"
+#include "catalog/indexing.h"
+#include "catalog/pg_collation.h"
+#include "catalog/pg_statistic_ext.h"
+#include "catalog/pg_statistic_ext_data.h"
+#include "executor/executor.h"
+#include "commands/progress.h"
+#include "miscadmin.h"
+#include "nodes/nodeFuncs.h"
+#include "optimizer/clauses.h"
+#include "optimizer/optimizer.h"
+#include "parser/parsetree.h"
+#include "pgstat.h"
+#include "postmaster/autovacuum.h"
+#include "statistics/extended_stats_internal.h"
+#include "statistics/statistics.h"
+#include "utils/acl.h"
+#include "utils/array.h"
+#include "utils/attoptcache.h"
+#include "utils/builtins.h"
+#include "utils/datum.h"
+#include "utils/fmgroids.h"
+#include "utils/lsyscache.h"
+#include "utils/memutils.h"
+#include "utils/rel.h"
+#include "utils/selfuncs.h"
+#include "utils/syscache.h"
+#include "utils/typcache.h"
+
+/*
+ * To avoid consuming too much memory during analysis and/or too much space
+ * in the resulting pg_statistic rows, we ignore varlena datums that are wider
+ * than WIDTH_THRESHOLD (after detoasting!).  This is legitimate for MCV
+ * and distinct-value calculations since a wide value is unlikely to be
+ * duplicated at all, much less be a most-common value.  For the same reason,
+ * ignoring wide values will not affect our estimates of histogram bin
+ * boundaries very much.
+ */
+#define WIDTH_THRESHOLD  1024
+
+/*
+ * Used internally to refer to an individual statistics object, i.e.,
+ * a pg_statistic_ext entry.
+ */
+typedef struct StatExtEntry
+{
+	Oid			statOid;		/* OID of pg_statistic_ext entry */
+	char	   *schema;			/* statistics object's schema */
+	char	   *name;			/* statistics object's name */
+	Bitmapset  *columns;		/* attribute numbers covered by the object */
+	List	   *types;			/* 'char' list of enabled statistics kinds */
+	int			stattarget;		/* statistics target (-1 for default) */
+	List	   *exprs;			/* expressions */
+} StatExtEntry;
+
+
+static List *fetch_statentries_for_relation(Relation pg_statext, Oid relid);
+static VacAttrStats **lookup_var_attr_stats(Relation rel, Bitmapset *attrs, List *exprs,
+											int nvacatts, VacAttrStats **vacatts);
+static void statext_store(Oid statOid,
+						  MVNDistinct *ndistinct, MVDependencies *dependencies,
+						  MCVList *mcv, Datum exprs, VacAttrStats **stats);
+static int	statext_compute_stattarget(int stattarget,
+									   int natts, VacAttrStats **stats);
+
+/* Information needed to analyze a single simple expression. */
+typedef struct AnlExprData
+{
+	Node	   *expr;			/* expression to analyze */
+	VacAttrStats *vacattrstat;	/* statistics attrs to analyze */
+} AnlExprData;
+
+static void compute_expr_stats(Relation onerel, double totalrows,
+							   AnlExprData *exprdata, int nexprs,
+							   HeapTuple *rows, int numrows);
+static Datum serialize_expr_stats(AnlExprData *exprdata, int nexprs);
+static Datum expr_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
+static AnlExprData *build_expr_data(List *exprs, int stattarget);
+
+static StatsBuildData *make_build_data(Relation onerel, StatExtEntry *stat,
+									   int numrows, HeapTuple *rows,
+									   VacAttrStats **stats, int stattarget);
+
+
+/*
+ * Compute requested extended stats, using the rows sampled for the plain
+ * (single-column) stats.
+ *
+ * This fetches a list of stats types from pg_statistic_ext, computes the
+ * requested stats, and serializes them back into the catalog.
+ */
+void
+BuildRelationExtStatistics(Relation onerel, double totalrows,
+						   int numrows, HeapTuple *rows,
+						   int natts, VacAttrStats **vacattrstats)
+{
+	Relation	pg_stext;
+	ListCell   *lc;
+	List	   *statslist;
+	MemoryContext cxt;
+	MemoryContext oldcxt;
+	int64		ext_cnt;
+
+	/* Do nothing if there are no columns to analyze. */
+	if (!natts)
+		return;
+
+	/* the list of stats has to be allocated outside the memory context */
+	pg_stext = table_open(StatisticExtRelationId, RowExclusiveLock);
+	statslist = fetch_statentries_for_relation(pg_stext, RelationGetRelid(onerel));
+
+	/* memory context for building each statistics object */
+	cxt = AllocSetContextCreate(CurrentMemoryContext,
+								"BuildRelationExtStatistics",
+								ALLOCSET_DEFAULT_SIZES);
+	oldcxt = MemoryContextSwitchTo(cxt);
+
+	/* report this phase */
+	if (statslist != NIL)
+	{
+		const int	index[] = {
+			PROGRESS_ANALYZE_PHASE,
+			PROGRESS_ANALYZE_EXT_STATS_TOTAL
+		};
+		const int64 val[] = {
+			PROGRESS_ANALYZE_PHASE_COMPUTE_EXT_STATS,
+			list_length(statslist)
+		};
+
+		pgstat_progress_update_multi_param(2, index, val);
+	}
+
+	ext_cnt = 0;
+	foreach(lc, statslist)
+	{
+		StatExtEntry *stat = (StatExtEntry *) lfirst(lc);
+		MVNDistinct *ndistinct = NULL;
+		MVDependencies *dependencies = NULL;
+		MCVList    *mcv = NULL;
+		Datum		exprstats = (Datum) 0;
+		VacAttrStats **stats;
+		ListCell   *lc2;
+		int			stattarget;
+		StatsBuildData *data;
+
+		/*
+		 * Check if we can build these stats based on the column analyzed. If
+		 * not, report this fact (except in autovacuum) and move on.
+		 */
+		stats = lookup_var_attr_stats(onerel, stat->columns, stat->exprs,
+									  natts, vacattrstats);
+		if (!stats)
+		{
+			if (!IsAutoVacuumWorkerProcess())
+				ereport(WARNING,
+						(errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+						 errmsg("statistics object \"%s.%s\" could not be computed for relation \"%s.%s\"",
+								stat->schema, stat->name,
+								get_namespace_name(onerel->rd_rel->relnamespace),
+								RelationGetRelationName(onerel)),
+						 errtable(onerel)));
+			continue;
+		}
+
+		/* compute statistics target for this statistics object */
+		stattarget = statext_compute_stattarget(stat->stattarget,
+												bms_num_members(stat->columns),
+												stats);
+
+		/*
+		 * Don't rebuild statistics objects with statistics target set to 0
+		 * (we just leave the existing values around, just like we do for
+		 * regular per-column statistics).
+		 */
+		if (stattarget == 0)
+			continue;
+
+		/* evaluate expressions (if the statistics object has any) */
+		data = make_build_data(onerel, stat, numrows, rows, stats, stattarget);
+
+		/* compute statistic of each requested type */
+		foreach(lc2, stat->types)
+		{
+			char		t = (char) lfirst_int(lc2);
+
+			if (t == STATS_EXT_NDISTINCT)
+				ndistinct = statext_ndistinct_build(totalrows, data);
+			else if (t == STATS_EXT_DEPENDENCIES)
+				dependencies = statext_dependencies_build(data);
+			else if (t == STATS_EXT_MCV)
+				mcv = statext_mcv_build(data, totalrows, stattarget);
+			else if (t == STATS_EXT_EXPRESSIONS)
+			{
+				AnlExprData *exprdata;
+				int			nexprs;
+
+				/* should not happen, thanks to checks when defining stats */
+				if (!stat->exprs)
+					elog(ERROR, "requested expression stats, but there are no expressions");
+
+				exprdata = build_expr_data(stat->exprs, stattarget);
+				nexprs = list_length(stat->exprs);
+
+				compute_expr_stats(onerel, totalrows,
+								   exprdata, nexprs,
+								   rows, numrows);
+
+				exprstats = serialize_expr_stats(exprdata, nexprs);
+			}
+		}
+
+		/* store the statistics in the catalog */
+		statext_store(stat->statOid, ndistinct, dependencies, mcv, exprstats, stats);
+
+		/* for reporting progress */
+		pgstat_progress_update_param(PROGRESS_ANALYZE_EXT_STATS_COMPUTED,
+									 ++ext_cnt);
+
+		/* free the data used for building this statistics object */
+		MemoryContextReset(cxt);
+	}
+
+	MemoryContextSwitchTo(oldcxt);
+	MemoryContextDelete(cxt);
+
+	list_free(statslist);
+
+	table_close(pg_stext, RowExclusiveLock);
+}
+
+/*
+ * ComputeExtStatisticsRows
+ *		Compute number of rows required by extended statistics on a table.
+ *
+ * Computes number of rows we need to sample to build extended statistics on a
+ * table. This only looks at statistics we can actually build - for example
+ * when analyzing only some of the columns, this will skip statistics objects
+ * that would require additional columns.
+ *
+ * See statext_compute_stattarget for details about how we compute the
+ * statistics target for a statistics object (from the object target,
+ * attribute targets and default statistics target).
+ */
+int
+ComputeExtStatisticsRows(Relation onerel,
+						 int natts, VacAttrStats **vacattrstats)
+{
+	Relation	pg_stext;
+	ListCell   *lc;
+	List	   *lstats;
+	MemoryContext cxt;
+	MemoryContext oldcxt;
+	int			result = 0;
+
+	/* If there are no columns to analyze, just return 0. */
+	if (!natts)
+		return 0;
+
+	cxt = AllocSetContextCreate(CurrentMemoryContext,
+								"ComputeExtStatisticsRows",
+								ALLOCSET_DEFAULT_SIZES);
+	oldcxt = MemoryContextSwitchTo(cxt);
+
+	pg_stext = table_open(StatisticExtRelationId, RowExclusiveLock);
+	lstats = fetch_statentries_for_relation(pg_stext, RelationGetRelid(onerel));
+
+	foreach(lc, lstats)
+	{
+		StatExtEntry *stat = (StatExtEntry *) lfirst(lc);
+		int			stattarget;
+		VacAttrStats **stats;
+		int			nattrs = bms_num_members(stat->columns);
+
+		/*
+		 * Check if we can build this statistics object based on the columns
+		 * analyzed. If not, ignore it (don't report anything, we'll do that
+		 * during the actual build BuildRelationExtStatistics).
+		 */
+		stats = lookup_var_attr_stats(onerel, stat->columns, stat->exprs,
+									  natts, vacattrstats);
+
+		if (!stats)
+			continue;
+
+		/*
+		 * Compute statistics target, based on what's set for the statistic
+		 * object itself, and for its attributes.
+		 */
+		stattarget = statext_compute_stattarget(stat->stattarget,
+												nattrs, stats);
+
+		/* Use the largest value for all statistics objects. */
+		if (stattarget > result)
+			result = stattarget;
+	}
+
+	table_close(pg_stext, RowExclusiveLock);
+
+	MemoryContextSwitchTo(oldcxt);
+	MemoryContextDelete(cxt);
+
+	/* compute sample size based on the statistics target */
+	return (300 * result);
+}
+
+/*
+ * statext_compute_stattarget
+ *		compute statistics target for an extended statistic
+ *
+ * When computing target for extended statistics objects, we consider three
+ * places where the target may be set - the statistics object itself,
+ * attributes the statistics object is defined on, and then the default
+ * statistics target.
+ *
+ * First we look at what's set for the statistics object itself, using the
+ * ALTER STATISTICS ... SET STATISTICS command. If we find a valid value
+ * there (i.e. not -1) we're done. Otherwise we look at targets set for any
+ * of the attributes the statistic is defined on, and if there are columns
+ * with defined target, we use the maximum value. We do this mostly for
+ * backwards compatibility, because this is what we did before having
+ * statistics target for extended statistics.
+ *
+ * And finally, if we still don't have a statistics target, we use the value
+ * set in default_statistics_target.
+ */
+static int
+statext_compute_stattarget(int stattarget, int nattrs, VacAttrStats **stats)
+{
+	int			i;
+
+	/*
+	 * If there's statistics target set for the statistics object, use it. It
+	 * may be set to 0 which disables building of that statistic.
+	 */
+	if (stattarget >= 0)
+		return stattarget;
+
+	/*
+	 * The target for the statistics object is set to -1, in which case we
+	 * look at the maximum target set for any of the attributes the object is
+	 * defined on.
+	 */
+	for (i = 0; i < nattrs; i++)
+	{
+		/* keep the maximum statistics target */
+		if (stats[i]->attr->attstattarget > stattarget)
+			stattarget = stats[i]->attr->attstattarget;
+	}
+
+	/*
+	 * If the value is still negative (so neither the statistics object nor
+	 * any of the columns have custom statistics target set), use the global
+	 * default target.
+	 */
+	if (stattarget < 0)
+		stattarget = default_statistics_target;
+
+	/* As this point we should have a valid statistics target. */
+	Assert((stattarget >= 0) && (stattarget <= 10000));
+
+	return stattarget;
+}
+
+/*
+ * statext_is_kind_built
+ *		Is this stat kind built in the given pg_statistic_ext_data tuple?
+ */
+bool
+statext_is_kind_built(HeapTuple htup, char type)
+{
+	AttrNumber	attnum;
+
+	switch (type)
+	{
+		case STATS_EXT_NDISTINCT:
+			attnum = Anum_pg_statistic_ext_data_stxdndistinct;
+			break;
+
+		case STATS_EXT_DEPENDENCIES:
+			attnum = Anum_pg_statistic_ext_data_stxddependencies;
+			break;
+
+		case STATS_EXT_MCV:
+			attnum = Anum_pg_statistic_ext_data_stxdmcv;
+			break;
+
+		case STATS_EXT_EXPRESSIONS:
+			attnum = Anum_pg_statistic_ext_data_stxdexpr;
+			break;
+
+		default:
+			elog(ERROR, "unexpected statistics type requested: %d", type);
+	}
+
+	return !heap_attisnull(htup, attnum, NULL);
+}
+
+/*
+ * Return a list (of StatExtEntry) of statistics objects for the given relation.
+ */
+static List *
+fetch_statentries_for_relation(Relation pg_statext, Oid relid)
+{
+	SysScanDesc scan;
+	ScanKeyData skey;
+	HeapTuple	htup;
+	List	   *result = NIL;
+
+	/*
+	 * Prepare to scan pg_statistic_ext for entries having stxrelid = this
+	 * rel.
+	 */
+	ScanKeyInit(&skey,
+				Anum_pg_statistic_ext_stxrelid,
+				BTEqualStrategyNumber, F_OIDEQ,
+				ObjectIdGetDatum(relid));
+
+	scan = systable_beginscan(pg_statext, StatisticExtRelidIndexId, true,
+							  NULL, 1, &skey);
+
+	while (HeapTupleIsValid(htup = systable_getnext(scan)))
+	{
+		StatExtEntry *entry;
+		Datum		datum;
+		bool		isnull;
+		int			i;
+		ArrayType  *arr;
+		char	   *enabled;
+		Form_pg_statistic_ext staForm;
+		List	   *exprs = NIL;
+
+		entry = palloc0(sizeof(StatExtEntry));
+		staForm = (Form_pg_statistic_ext) GETSTRUCT(htup);
+		entry->statOid = staForm->oid;
+		entry->schema = get_namespace_name(staForm->stxnamespace);
+		entry->name = pstrdup(NameStr(staForm->stxname));
+		entry->stattarget = staForm->stxstattarget;
+		for (i = 0; i < staForm->stxkeys.dim1; i++)
+		{
+			entry->columns = bms_add_member(entry->columns,
+											staForm->stxkeys.values[i]);
+		}
+
+		/* decode the stxkind char array into a list of chars */
+		datum = SysCacheGetAttr(STATEXTOID, htup,
+								Anum_pg_statistic_ext_stxkind, &isnull);
+		Assert(!isnull);
+		arr = DatumGetArrayTypeP(datum);
+		if (ARR_NDIM(arr) != 1 ||
+			ARR_HASNULL(arr) ||
+			ARR_ELEMTYPE(arr) != CHAROID)
+			elog(ERROR, "stxkind is not a 1-D char array");
+		enabled = (char *) ARR_DATA_PTR(arr);
+		for (i = 0; i < ARR_DIMS(arr)[0]; i++)
+		{
+			Assert((enabled[i] == STATS_EXT_NDISTINCT) ||
+				   (enabled[i] == STATS_EXT_DEPENDENCIES) ||
+				   (enabled[i] == STATS_EXT_MCV) ||
+				   (enabled[i] == STATS_EXT_EXPRESSIONS));
+			entry->types = lappend_int(entry->types, (int) enabled[i]);
+		}
+
+		/* decode expression (if any) */
+		datum = SysCacheGetAttr(STATEXTOID, htup,
+								Anum_pg_statistic_ext_stxexprs, &isnull);
+
+		if (!isnull)
+		{
+			char	   *exprsString;
+
+			exprsString = TextDatumGetCString(datum);
+			exprs = (List *) stringToNode(exprsString);
+
+			pfree(exprsString);
+
+			/*
+			 * Run the expressions through eval_const_expressions. This is not
+			 * just an optimization, but is necessary, because the planner
+			 * will be comparing them to similarly-processed qual clauses, and
+			 * may fail to detect valid matches without this.  We must not use
+			 * canonicalize_qual, however, since these aren't qual
+			 * expressions.
+			 */
+			exprs = (List *) eval_const_expressions(NULL, (Node *) exprs);
+
+			/* May as well fix opfuncids too */
+			fix_opfuncids((Node *) exprs);
+		}
+
+		entry->exprs = exprs;
+
+		result = lappend(result, entry);
+	}
+
+	systable_endscan(scan);
+
+	return result;
+}
+
+/*
+ * examine_attribute -- pre-analysis of a single column
+ *
+ * Determine whether the column is analyzable; if so, create and initialize
+ * a VacAttrStats struct for it.  If not, return NULL.
+ */
+static VacAttrStats *
+examine_attribute(Node *expr)
+{
+	HeapTuple	typtuple;
+	VacAttrStats *stats;
+	int			i;
+	bool		ok;
+
+	/*
+	 * Create the VacAttrStats struct.  Note that we only have a copy of the
+	 * fixed fields of the pg_attribute tuple.
+	 */
+	stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
+
+	/* fake the attribute */
+	stats->attr = (Form_pg_attribute) palloc0(ATTRIBUTE_FIXED_PART_SIZE);
+	stats->attr->attstattarget = -1;
+
+	/*
+	 * When analyzing an expression, believe the expression tree's type not
+	 * the column datatype --- the latter might be the opckeytype storage type
+	 * of the opclass, which is not interesting for our purposes.  (Note: if
+	 * we did anything with non-expression statistics columns, we'd need to
+	 * figure out where to get the correct type info from, but for now that's
+	 * not a problem.)	It's not clear whether anyone will care about the
+	 * typmod, but we store that too just in case.
+	 */
+	stats->attrtypid = exprType(expr);
+	stats->attrtypmod = exprTypmod(expr);
+	stats->attrcollid = exprCollation(expr);
+
+	typtuple = SearchSysCacheCopy1(TYPEOID,
+								   ObjectIdGetDatum(stats->attrtypid));
+	if (!HeapTupleIsValid(typtuple))
+		elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
+	stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
+
+	/*
+	 * We don't actually analyze individual attributes, so no need to set the
+	 * memory context.
+	 */
+	stats->anl_context = NULL;
+	stats->tupattnum = InvalidAttrNumber;
+
+	/*
+	 * The fields describing the stats->stavalues[n] element types default to
+	 * the type of the data being analyzed, but the type-specific typanalyze
+	 * function can change them if it wants to store something else.
+	 */
+	for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
+	{
+		stats->statypid[i] = stats->attrtypid;
+		stats->statyplen[i] = stats->attrtype->typlen;
+		stats->statypbyval[i] = stats->attrtype->typbyval;
+		stats->statypalign[i] = stats->attrtype->typalign;
+	}
+
+	/*
+	 * Call the type-specific typanalyze function.  If none is specified, use
+	 * std_typanalyze().
+	 */
+	if (OidIsValid(stats->attrtype->typanalyze))
+		ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
+										   PointerGetDatum(stats)));
+	else
+		ok = std_typanalyze(stats);
+
+	if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
+	{
+		heap_freetuple(typtuple);
+		pfree(stats->attr);
+		pfree(stats);
+		return NULL;
+	}
+
+	return stats;
+}
+
+/*
+ * examine_expression -- pre-analysis of a single expression
+ *
+ * Determine whether the expression is analyzable; if so, create and initialize
+ * a VacAttrStats struct for it.  If not, return NULL.
+ */
+static VacAttrStats *
+examine_expression(Node *expr, int stattarget)
+{
+	HeapTuple	typtuple;
+	VacAttrStats *stats;
+	int			i;
+	bool		ok;
+
+	Assert(expr != NULL);
+
+	/*
+	 * Create the VacAttrStats struct.
+	 */
+	stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
+
+	/*
+	 * When analyzing an expression, believe the expression tree's type.
+	 */
+	stats->attrtypid = exprType(expr);
+	stats->attrtypmod = exprTypmod(expr);
+
+	/*
+	 * We don't allow collation to be specified in CREATE STATISTICS, so we
+	 * have to use the collation specified for the expression. It's possible
+	 * to specify the collation in the expression "(col COLLATE "en_US")" in
+	 * which case exprCollation() does the right thing.
+	 */
+	stats->attrcollid = exprCollation(expr);
+
+	/*
+	 * We don't have any pg_attribute for expressions, so let's fake something
+	 * reasonable into attstattarget, which is the only thing std_typanalyze
+	 * needs.
+	 */
+	stats->attr = (Form_pg_attribute) palloc(ATTRIBUTE_FIXED_PART_SIZE);
+
+	/*
+	 * We can't have statistics target specified for the expression, so we
+	 * could use either the default_statistics_target, or the target computed
+	 * for the extended statistics. The second option seems more reasonable.
+	 */
+	stats->attr->attstattarget = stattarget;
+
+	/* initialize some basic fields */
+	stats->attr->attrelid = InvalidOid;
+	stats->attr->attnum = InvalidAttrNumber;
+	stats->attr->atttypid = stats->attrtypid;
+
+	typtuple = SearchSysCacheCopy1(TYPEOID,
+								   ObjectIdGetDatum(stats->attrtypid));
+	if (!HeapTupleIsValid(typtuple))
+		elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
+
+	stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
+	stats->anl_context = CurrentMemoryContext;	/* XXX should be using
+												 * something else? */
+	stats->tupattnum = InvalidAttrNumber;
+
+	/*
+	 * The fields describing the stats->stavalues[n] element types default to
+	 * the type of the data being analyzed, but the type-specific typanalyze
+	 * function can change them if it wants to store something else.
+	 */
+	for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
+	{
+		stats->statypid[i] = stats->attrtypid;
+		stats->statyplen[i] = stats->attrtype->typlen;
+		stats->statypbyval[i] = stats->attrtype->typbyval;
+		stats->statypalign[i] = stats->attrtype->typalign;
+	}
+
+	/*
+	 * Call the type-specific typanalyze function.  If none is specified, use
+	 * std_typanalyze().
+	 */
+	if (OidIsValid(stats->attrtype->typanalyze))
+		ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
+										   PointerGetDatum(stats)));
+	else
+		ok = std_typanalyze(stats);
+
+	if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
+	{
+		heap_freetuple(typtuple);
+		pfree(stats);
+		return NULL;
+	}
+
+	return stats;
+}
+
+/*
+ * Using 'vacatts' of size 'nvacatts' as input data, return a newly built
+ * VacAttrStats array which includes only the items corresponding to
+ * attributes indicated by 'stxkeys'. If we don't have all of the per column
+ * stats available to compute the extended stats, then we return NULL to indicate
+ * to the caller that the stats should not be built.
+ */
+static VacAttrStats **
+lookup_var_attr_stats(Relation rel, Bitmapset *attrs, List *exprs,
+					  int nvacatts, VacAttrStats **vacatts)
+{
+	int			i = 0;
+	int			x = -1;
+	int			natts;
+	VacAttrStats **stats;
+	ListCell   *lc;
+
+	natts = bms_num_members(attrs) + list_length(exprs);
+
+	stats = (VacAttrStats **) palloc(natts * sizeof(VacAttrStats *));
+
+	/* lookup VacAttrStats info for the requested columns (same attnum) */
+	while ((x = bms_next_member(attrs, x)) >= 0)
+	{
+		int			j;
+
+		stats[i] = NULL;
+		for (j = 0; j < nvacatts; j++)
+		{
+			if (x == vacatts[j]->tupattnum)
+			{
+				stats[i] = vacatts[j];
+				break;
+			}
+		}
+
+		if (!stats[i])
+		{
+			/*
+			 * Looks like stats were not gathered for one of the columns
+			 * required. We'll be unable to build the extended stats without
+			 * this column.
+			 */
+			pfree(stats);
+			return NULL;
+		}
+
+		/*
+		 * Sanity check that the column is not dropped - stats should have
+		 * been removed in this case.
+		 */
+		Assert(!stats[i]->attr->attisdropped);
+
+		i++;
+	}
+
+	/* also add info for expressions */
+	foreach(lc, exprs)
+	{
+		Node	   *expr = (Node *) lfirst(lc);
+
+		stats[i] = examine_attribute(expr);
+
+		/*
+		 * XXX We need tuple descriptor later, and we just grab it from
+		 * stats[0]->tupDesc (see e.g. statext_mcv_build). But as coded
+		 * examine_attribute does not set that, so just grab it from the first
+		 * vacatts element.
+		 */
+		stats[i]->tupDesc = vacatts[0]->tupDesc;
+
+		i++;
+	}
+
+	return stats;
+}
+
+/*
+ * statext_store
+ *	Serializes the statistics and stores them into the pg_statistic_ext_data
+ *	tuple.
+ */
+static void
+statext_store(Oid statOid,
+			  MVNDistinct *ndistinct, MVDependencies *dependencies,
+			  MCVList *mcv, Datum exprs, VacAttrStats **stats)
+{
+	Relation	pg_stextdata;
+	HeapTuple	stup,
+				oldtup;
+	Datum		values[Natts_pg_statistic_ext_data];
+	bool		nulls[Natts_pg_statistic_ext_data];
+	bool		replaces[Natts_pg_statistic_ext_data];
+
+	pg_stextdata = table_open(StatisticExtDataRelationId, RowExclusiveLock);
+
+	memset(nulls, true, sizeof(nulls));
+	memset(replaces, false, sizeof(replaces));
+	memset(values, 0, sizeof(values));
+
+	/*
+	 * Construct a new pg_statistic_ext_data tuple, replacing the calculated
+	 * stats.
+	 */
+	if (ndistinct != NULL)
+	{
+		bytea	   *data = statext_ndistinct_serialize(ndistinct);
+
+		nulls[Anum_pg_statistic_ext_data_stxdndistinct - 1] = (data == NULL);
+		values[Anum_pg_statistic_ext_data_stxdndistinct - 1] = PointerGetDatum(data);
+	}
+
+	if (dependencies != NULL)
+	{
+		bytea	   *data = statext_dependencies_serialize(dependencies);
+
+		nulls[Anum_pg_statistic_ext_data_stxddependencies - 1] = (data == NULL);
+		values[Anum_pg_statistic_ext_data_stxddependencies - 1] = PointerGetDatum(data);
+	}
+	if (mcv != NULL)
+	{
+		bytea	   *data = statext_mcv_serialize(mcv, stats);
+
+		nulls[Anum_pg_statistic_ext_data_stxdmcv - 1] = (data == NULL);
+		values[Anum_pg_statistic_ext_data_stxdmcv - 1] = PointerGetDatum(data);
+	}
+	if (exprs != (Datum) 0)
+	{
+		nulls[Anum_pg_statistic_ext_data_stxdexpr - 1] = false;
+		values[Anum_pg_statistic_ext_data_stxdexpr - 1] = exprs;
+	}
+
+	/* always replace the value (either by bytea or NULL) */
+	replaces[Anum_pg_statistic_ext_data_stxdndistinct - 1] = true;
+	replaces[Anum_pg_statistic_ext_data_stxddependencies - 1] = true;
+	replaces[Anum_pg_statistic_ext_data_stxdmcv - 1] = true;
+	replaces[Anum_pg_statistic_ext_data_stxdexpr - 1] = true;
+
+	/* there should already be a pg_statistic_ext_data tuple */
+	oldtup = SearchSysCache1(STATEXTDATASTXOID, ObjectIdGetDatum(statOid));
+	if (!HeapTupleIsValid(oldtup))
+		elog(ERROR, "cache lookup failed for statistics object %u", statOid);
+
+	/* replace it */
+	stup = heap_modify_tuple(oldtup,
+							 RelationGetDescr(pg_stextdata),
+							 values,
+							 nulls,
+							 replaces);
+	ReleaseSysCache(oldtup);
+	CatalogTupleUpdate(pg_stextdata, &stup->t_self, stup);
+
+	heap_freetuple(stup);
+
+	table_close(pg_stextdata, RowExclusiveLock);
+}
+
+/* initialize multi-dimensional sort */
+MultiSortSupport
+multi_sort_init(int ndims)
+{
+	MultiSortSupport mss;
+
+	Assert(ndims >= 2);
+
+	mss = (MultiSortSupport) palloc0(offsetof(MultiSortSupportData, ssup)
+									 + sizeof(SortSupportData) * ndims);
+
+	mss->ndims = ndims;
+
+	return mss;
+}
+
+/*
+ * Prepare sort support info using the given sort operator and collation
+ * at the position 'sortdim'
+ */
+void
+multi_sort_add_dimension(MultiSortSupport mss, int sortdim,
+						 Oid oper, Oid collation)
+{
+	SortSupport ssup = &mss->ssup[sortdim];
+
+	ssup->ssup_cxt = CurrentMemoryContext;
+	ssup->ssup_collation = collation;
+	ssup->ssup_nulls_first = false;
+
+	PrepareSortSupportFromOrderingOp(oper, ssup);
+}
+
+/* compare all the dimensions in the selected order */
+int
+multi_sort_compare(const void *a, const void *b, void *arg)
+{
+	MultiSortSupport mss = (MultiSortSupport) arg;
+	SortItem   *ia = (SortItem *) a;
+	SortItem   *ib = (SortItem *) b;
+	int			i;
+
+	for (i = 0; i < mss->ndims; i++)
+	{
+		int			compare;
+
+		compare = ApplySortComparator(ia->values[i], ia->isnull[i],
+									  ib->values[i], ib->isnull[i],
+									  &mss->ssup[i]);
+
+		if (compare != 0)
+			return compare;
+	}
+
+	/* equal by default */
+	return 0;
+}
+
+/* compare selected dimension */
+int
+multi_sort_compare_dim(int dim, const SortItem *a, const SortItem *b,
+					   MultiSortSupport mss)
+{
+	return ApplySortComparator(a->values[dim], a->isnull[dim],
+							   b->values[dim], b->isnull[dim],
+							   &mss->ssup[dim]);
+}
+
+int
+multi_sort_compare_dims(int start, int end,
+						const SortItem *a, const SortItem *b,
+						MultiSortSupport mss)
+{
+	int			dim;
+
+	for (dim = start; dim <= end; dim++)
+	{
+		int			r = ApplySortComparator(a->values[dim], a->isnull[dim],
+											b->values[dim], b->isnull[dim],
+											&mss->ssup[dim]);
+
+		if (r != 0)
+			return r;
+	}
+
+	return 0;
+}
+
+int
+compare_scalars_simple(const void *a, const void *b, void *arg)
+{
+	return compare_datums_simple(*(Datum *) a,
+								 *(Datum *) b,
+								 (SortSupport) arg);
+}
+
+int
+compare_datums_simple(Datum a, Datum b, SortSupport ssup)
+{
+	return ApplySortComparator(a, false, b, false, ssup);
+}
+
+/*
+ * build_attnums_array
+ *		Transforms a bitmap into an array of AttrNumber values.
+ *
+ * This is used for extended statistics only, so all the attribute must be
+ * user-defined. That means offsetting by FirstLowInvalidHeapAttributeNumber
+ * is not necessary here (and when querying the bitmap).
+ */
+AttrNumber *
+build_attnums_array(Bitmapset *attrs, int nexprs, int *numattrs)
+{
+	int			i,
+				j;
+	AttrNumber *attnums;
+	int			num = bms_num_members(attrs);
+
+	if (numattrs)
+		*numattrs = num;
+
+	/* build attnums from the bitmapset */
+	attnums = (AttrNumber *) palloc(sizeof(AttrNumber) * num);
+	i = 0;
+	j = -1;
+	while ((j = bms_next_member(attrs, j)) >= 0)
+	{
+		int			attnum = (j - nexprs);
+
+		/*
+		 * Make sure the bitmap contains only user-defined attributes. As
+		 * bitmaps can't contain negative values, this can be violated in two
+		 * ways. Firstly, the bitmap might contain 0 as a member, and secondly
+		 * the integer value might be larger than MaxAttrNumber.
+		 */
+		Assert(AttributeNumberIsValid(attnum));
+		Assert(attnum <= MaxAttrNumber);
+		Assert(attnum >= (-nexprs));
+
+		attnums[i++] = (AttrNumber) attnum;
+
+		/* protect against overflows */
+		Assert(i <= num);
+	}
+
+	return attnums;
+}
+
+/*
+ * build_sorted_items
+ *		build a sorted array of SortItem with values from rows
+ *
+ * Note: All the memory is allocated in a single chunk, so that the caller
+ * can simply pfree the return value to release all of it.
+ */
+SortItem *
+build_sorted_items(StatsBuildData *data, int *nitems,
+				   MultiSortSupport mss,
+				   int numattrs, AttrNumber *attnums)
+{
+	int			i,
+				j,
+				len,
+				nrows;
+	int			nvalues = data->numrows * numattrs;
+
+	SortItem   *items;
+	Datum	   *values;
+	bool	   *isnull;
+	char	   *ptr;
+	int		   *typlen;
+
+	/* Compute the total amount of memory we need (both items and values). */
+	len = data->numrows * sizeof(SortItem) + nvalues * (sizeof(Datum) + sizeof(bool));
+
+	/* Allocate the memory and split it into the pieces. */
+	ptr = palloc0(len);
+
+	/* items to sort */
+	items = (SortItem *) ptr;
+	ptr += data->numrows * sizeof(SortItem);
+
+	/* values and null flags */
+	values = (Datum *) ptr;
+	ptr += nvalues * sizeof(Datum);
+
+	isnull = (bool *) ptr;
+	ptr += nvalues * sizeof(bool);
+
+	/* make sure we consumed the whole buffer exactly */
+	Assert((ptr - (char *) items) == len);
+
+	/* fix the pointers to Datum and bool arrays */
+	nrows = 0;
+	for (i = 0; i < data->numrows; i++)
+	{
+		items[nrows].values = &values[nrows * numattrs];
+		items[nrows].isnull = &isnull[nrows * numattrs];
+
+		nrows++;
+	}
+
+	/* build a local cache of typlen for all attributes */
+	typlen = (int *) palloc(sizeof(int) * data->nattnums);
+	for (i = 0; i < data->nattnums; i++)
+		typlen[i] = get_typlen(data->stats[i]->attrtypid);
+
+	nrows = 0;
+	for (i = 0; i < data->numrows; i++)
+	{
+		bool		toowide = false;
+
+		/* load the values/null flags from sample rows */
+		for (j = 0; j < numattrs; j++)
+		{
+			Datum		value;
+			bool		isnull;
+			int			attlen;
+			AttrNumber	attnum = attnums[j];
+
+			int			idx;
+
+			/* match attnum to the pre-calculated data */
+			for (idx = 0; idx < data->nattnums; idx++)
+			{
+				if (attnum == data->attnums[idx])
+					break;
+			}
+
+			Assert(idx < data->nattnums);
+
+			value = data->values[idx][i];
+			isnull = data->nulls[idx][i];
+			attlen = typlen[idx];
+
+			/*
+			 * If this is a varlena value, check if it's too wide and if yes
+			 * then skip the whole item. Otherwise detoast the value.
+			 *
+			 * XXX It may happen that we've already detoasted some preceding
+			 * values for the current item. We don't bother to cleanup those
+			 * on the assumption that those are small (below WIDTH_THRESHOLD)
+			 * and will be discarded at the end of analyze.
+			 */
+			if ((!isnull) && (attlen == -1))
+			{
+				if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
+				{
+					toowide = true;
+					break;
+				}
+
+				value = PointerGetDatum(PG_DETOAST_DATUM(value));
+			}
+
+			items[nrows].values[j] = value;
+			items[nrows].isnull[j] = isnull;
+		}
+
+		if (toowide)
+			continue;
+
+		nrows++;
+	}
+
+	/* store the actual number of items (ignoring the too-wide ones) */
+	*nitems = nrows;
+
+	/* all items were too wide */
+	if (nrows == 0)
+	{
+		/* everything is allocated as a single chunk */
+		pfree(items);
+		return NULL;
+	}
+
+	/* do the sort, using the multi-sort */
+	qsort_interruptible((void *) items, nrows, sizeof(SortItem),
+						multi_sort_compare, mss);
+
+	return items;
+}
+
+/*
+ * has_stats_of_kind
+ *		Check whether the list contains statistic of a given kind
+ */
+bool
+has_stats_of_kind(List *stats, char requiredkind)
+{
+	ListCell   *l;
+
+	foreach(l, stats)
+	{
+		StatisticExtInfo *stat = (StatisticExtInfo *) lfirst(l);
+
+		if (stat->kind == requiredkind)
+			return true;
+	}
+
+	return false;
+}
+
+/*
+ * stat_find_expression
+ *		Search for an expression in statistics object's list of expressions.
+ *
+ * Returns the index of the expression in the statistics object's list of
+ * expressions, or -1 if not found.
+ */
+static int
+stat_find_expression(StatisticExtInfo *stat, Node *expr)
+{
+	ListCell   *lc;
+	int			idx;
+
+	idx = 0;
+	foreach(lc, stat->exprs)
+	{
+		Node	   *stat_expr = (Node *) lfirst(lc);
+
+		if (equal(stat_expr, expr))
+			return idx;
+		idx++;
+	}
+
+	/* Expression not found */
+	return -1;
+}
+
+/*
+ * stat_covers_expressions
+ * 		Test whether a statistics object covers all expressions in a list.
+ *
+ * Returns true if all expressions are covered.  If expr_idxs is non-NULL, it
+ * is populated with the indexes of the expressions found.
+ */
+static bool
+stat_covers_expressions(StatisticExtInfo *stat, List *exprs,
+						Bitmapset **expr_idxs)
+{
+	ListCell   *lc;
+
+	foreach(lc, exprs)
+	{
+		Node	   *expr = (Node *) lfirst(lc);
+		int			expr_idx;
+
+		expr_idx = stat_find_expression(stat, expr);
+		if (expr_idx == -1)
+			return false;
+
+		if (expr_idxs != NULL)
+			*expr_idxs = bms_add_member(*expr_idxs, expr_idx);
+	}
+
+	/* If we reach here, all expressions are covered */
+	return true;
+}
+
+/*
+ * choose_best_statistics
+ *		Look for and return statistics with the specified 'requiredkind' which
+ *		have keys that match at least two of the given attnums.  Return NULL if
+ *		there's no match.
+ *
+ * The current selection criteria is very simple - we choose the statistics
+ * object referencing the most attributes in covered (and still unestimated
+ * clauses), breaking ties in favor of objects with fewer keys overall.
+ *
+ * The clause_attnums is an array of bitmaps, storing attnums for individual
+ * clauses. A NULL element means the clause is either incompatible or already
+ * estimated.
+ *
+ * XXX If multiple statistics objects tie on both criteria, then which object
+ * is chosen depends on the order that they appear in the stats list. Perhaps
+ * further tiebreakers are needed.
+ */
+StatisticExtInfo *
+choose_best_statistics(List *stats, char requiredkind,
+					   Bitmapset **clause_attnums, List **clause_exprs,
+					   int nclauses)
+{
+	ListCell   *lc;
+	StatisticExtInfo *best_match = NULL;
+	int			best_num_matched = 2;	/* goal #1: maximize */
+	int			best_match_keys = (STATS_MAX_DIMENSIONS + 1);	/* goal #2: minimize */
+
+	foreach(lc, stats)
+	{
+		int			i;
+		StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
+		Bitmapset  *matched_attnums = NULL;
+		Bitmapset  *matched_exprs = NULL;
+		int			num_matched;
+		int			numkeys;
+
+		/* skip statistics that are not of the correct type */
+		if (info->kind != requiredkind)
+			continue;
+
+		/*
+		 * Collect attributes and expressions in remaining (unestimated)
+		 * clauses fully covered by this statistic object.
+		 *
+		 * We know already estimated clauses have both clause_attnums and
+		 * clause_exprs set to NULL. We leave the pointers NULL if already
+		 * estimated, or we reset them to NULL after estimating the clause.
+		 */
+		for (i = 0; i < nclauses; i++)
+		{
+			Bitmapset  *expr_idxs = NULL;
+
+			/* ignore incompatible/estimated clauses */
+			if (!clause_attnums[i] && !clause_exprs[i])
+				continue;
+
+			/* ignore clauses that are not covered by this object */
+			if (!bms_is_subset(clause_attnums[i], info->keys) ||
+				!stat_covers_expressions(info, clause_exprs[i], &expr_idxs))
+				continue;
+
+			/* record attnums and indexes of expressions covered */
+			matched_attnums = bms_add_members(matched_attnums, clause_attnums[i]);
+			matched_exprs = bms_add_members(matched_exprs, expr_idxs);
+		}
+
+		num_matched = bms_num_members(matched_attnums) + bms_num_members(matched_exprs);
+
+		bms_free(matched_attnums);
+		bms_free(matched_exprs);
+
+		/*
+		 * save the actual number of keys in the stats so that we can choose
+		 * the narrowest stats with the most matching keys.
+		 */
+		numkeys = bms_num_members(info->keys) + list_length(info->exprs);
+
+		/*
+		 * Use this object when it increases the number of matched attributes
+		 * and expressions or when it matches the same number of attributes
+		 * and expressions but these stats have fewer keys than any previous
+		 * match.
+		 */
+		if (num_matched > best_num_matched ||
+			(num_matched == best_num_matched && numkeys < best_match_keys))
+		{
+			best_match = info;
+			best_num_matched = num_matched;
+			best_match_keys = numkeys;
+		}
+	}
+
+	return best_match;
+}
+
+/*
+ * statext_is_compatible_clause_internal
+ *		Determines if the clause is compatible with MCV lists.
+ *
+ * To be compatible, the given clause must be a combination of supported
+ * clauses built from Vars or sub-expressions (where a sub-expression is
+ * something that exactly matches an expression found in statistics objects).
+ * This function recursively examines the clause and extracts any
+ * sub-expressions that will need to be matched against statistics.
+ *
+ * Currently, we only support the following types of clauses:
+ *
+ * (a) OpExprs of the form (Var/Expr op Const), or (Const op Var/Expr), where
+ * the op is one of ("=", "<", ">", ">=", "<=")
+ *
+ * (b) (Var/Expr IS [NOT] NULL)
+ *
+ * (c) combinations using AND/OR/NOT
+ *
+ * (d) ScalarArrayOpExprs of the form (Var/Expr op ANY (Const)) or
+ * (Var/Expr op ALL (Const))
+ *
+ * In the future, the range of supported clauses may be expanded to more
+ * complex cases, for example (Var op Var).
+ *
+ * Arguments:
+ * clause: (sub)clause to be inspected (bare clause, not a RestrictInfo)
+ * relid: rel that all Vars in clause must belong to
+ * *attnums: input/output parameter collecting attribute numbers of all
+ *		mentioned Vars.  Note that we do not offset the attribute numbers,
+ *		so we can't cope with system columns.
+ * *exprs: input/output parameter collecting primitive subclauses within
+ *		the clause tree
+ *
+ * Returns false if there is something we definitively can't handle.
+ * On true return, we can proceed to match the *exprs against statistics.
+ */
+static bool
+statext_is_compatible_clause_internal(PlannerInfo *root, Node *clause,
+									  Index relid, Bitmapset **attnums,
+									  List **exprs)
+{
+	/* Look inside any binary-compatible relabeling (as in examine_variable) */
+	if (IsA(clause, RelabelType))
+		clause = (Node *) ((RelabelType *) clause)->arg;
+
+	/* plain Var references (boolean Vars or recursive checks) */
+	if (IsA(clause, Var))
+	{
+		Var		   *var = (Var *) clause;
+
+		/* Ensure var is from the correct relation */
+		if (var->varno != relid)
+			return false;
+
+		/* we also better ensure the Var is from the current level */
+		if (var->varlevelsup > 0)
+			return false;
+
+		/*
+		 * Also reject system attributes and whole-row Vars (we don't allow
+		 * stats on those).
+		 */
+		if (!AttrNumberIsForUserDefinedAttr(var->varattno))
+			return false;
+
+		/* OK, record the attnum for later permissions checks. */
+		*attnums = bms_add_member(*attnums, var->varattno);
+
+		return true;
+	}
+
+	/* (Var/Expr op Const) or (Const op Var/Expr) */
+	if (is_opclause(clause))
+	{
+		RangeTblEntry *rte = root->simple_rte_array[relid];
+		OpExpr	   *expr = (OpExpr *) clause;
+		Node	   *clause_expr;
+
+		/* Only expressions with two arguments are considered compatible. */
+		if (list_length(expr->args) != 2)
+			return false;
+
+		/* Check if the expression has the right shape */
+		if (!examine_opclause_args(expr->args, &clause_expr, NULL, NULL))
+			return false;
+
+		/*
+		 * If it's not one of the supported operators ("=", "<", ">", etc.),
+		 * just ignore the clause, as it's not compatible with MCV lists.
+		 *
+		 * This uses the function for estimating selectivity, not the operator
+		 * directly (a bit awkward, but well ...).
+		 */
+		switch (get_oprrest(expr->opno))
+		{
+			case F_EQSEL:
+			case F_NEQSEL:
+			case F_SCALARLTSEL:
+			case F_SCALARLESEL:
+			case F_SCALARGTSEL:
+			case F_SCALARGESEL:
+				/* supported, will continue with inspection of the Var/Expr */
+				break;
+
+			default:
+				/* other estimators are considered unknown/unsupported */
+				return false;
+		}
+
+		/*
+		 * If there are any securityQuals on the RTE from security barrier
+		 * views or RLS policies, then the user may not have access to all the
+		 * table's data, and we must check that the operator is leak-proof.
+		 *
+		 * If the operator is leaky, then we must ignore this clause for the
+		 * purposes of estimating with MCV lists, otherwise the operator might
+		 * reveal values from the MCV list that the user doesn't have
+		 * permission to see.
+		 */
+		if (rte->securityQuals != NIL &&
+			!get_func_leakproof(get_opcode(expr->opno)))
+			return false;
+
+		/* Check (Var op Const) or (Const op Var) clauses by recursing. */
+		if (IsA(clause_expr, Var))
+			return statext_is_compatible_clause_internal(root, clause_expr,
+														 relid, attnums, exprs);
+
+		/* Otherwise we have (Expr op Const) or (Const op Expr). */
+		*exprs = lappend(*exprs, clause_expr);
+		return true;
+	}
+
+	/* Var/Expr IN Array */
+	if (IsA(clause, ScalarArrayOpExpr))
+	{
+		RangeTblEntry *rte = root->simple_rte_array[relid];
+		ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) clause;
+		Node	   *clause_expr;
+		bool		expronleft;
+
+		/* Only expressions with two arguments are considered compatible. */
+		if (list_length(expr->args) != 2)
+			return false;
+
+		/* Check if the expression has the right shape (one Var, one Const) */
+		if (!examine_opclause_args(expr->args, &clause_expr, NULL, &expronleft))
+			return false;
+
+		/* We only support Var on left, Const on right */
+		if (!expronleft)
+			return false;
+
+		/*
+		 * If it's not one of the supported operators ("=", "<", ">", etc.),
+		 * just ignore the clause, as it's not compatible with MCV lists.
+		 *
+		 * This uses the function for estimating selectivity, not the operator
+		 * directly (a bit awkward, but well ...).
+		 */
+		switch (get_oprrest(expr->opno))
+		{
+			case F_EQSEL:
+			case F_NEQSEL:
+			case F_SCALARLTSEL:
+			case F_SCALARLESEL:
+			case F_SCALARGTSEL:
+			case F_SCALARGESEL:
+				/* supported, will continue with inspection of the Var/Expr */
+				break;
+
+			default:
+				/* other estimators are considered unknown/unsupported */
+				return false;
+		}
+
+		/*
+		 * If there are any securityQuals on the RTE from security barrier
+		 * views or RLS policies, then the user may not have access to all the
+		 * table's data, and we must check that the operator is leak-proof.
+		 *
+		 * If the operator is leaky, then we must ignore this clause for the
+		 * purposes of estimating with MCV lists, otherwise the operator might
+		 * reveal values from the MCV list that the user doesn't have
+		 * permission to see.
+		 */
+		if (rte->securityQuals != NIL &&
+			!get_func_leakproof(get_opcode(expr->opno)))
+			return false;
+
+		/* Check Var IN Array clauses by recursing. */
+		if (IsA(clause_expr, Var))
+			return statext_is_compatible_clause_internal(root, clause_expr,
+														 relid, attnums, exprs);
+
+		/* Otherwise we have Expr IN Array. */
+		*exprs = lappend(*exprs, clause_expr);
+		return true;
+	}
+
+	/* AND/OR/NOT clause */
+	if (is_andclause(clause) ||
+		is_orclause(clause) ||
+		is_notclause(clause))
+	{
+		/*
+		 * AND/OR/NOT-clauses are supported if all sub-clauses are supported
+		 *
+		 * Perhaps we could improve this by handling mixed cases, when some of
+		 * the clauses are supported and some are not. Selectivity for the
+		 * supported subclauses would be computed using extended statistics,
+		 * and the remaining clauses would be estimated using the traditional
+		 * algorithm (product of selectivities).
+		 *
+		 * It however seems overly complex, and in a way we already do that
+		 * because if we reject the whole clause as unsupported here, it will
+		 * be eventually passed to clauselist_selectivity() which does exactly
+		 * this (split into supported/unsupported clauses etc).
+		 */
+		BoolExpr   *expr = (BoolExpr *) clause;
+		ListCell   *lc;
+
+		foreach(lc, expr->args)
+		{
+			/*
+			 * If we find an incompatible clause in the arguments, treat the
+			 * whole clause as incompatible.
+			 */
+			if (!statext_is_compatible_clause_internal(root,
+													   (Node *) lfirst(lc),
+													   relid, attnums, exprs))
+				return false;
+		}
+
+		return true;
+	}
+
+	/* Var/Expr IS NULL */
+	if (IsA(clause, NullTest))
+	{
+		NullTest   *nt = (NullTest *) clause;
+
+		/* Check Var IS NULL clauses by recursing. */
+		if (IsA(nt->arg, Var))
+			return statext_is_compatible_clause_internal(root, (Node *) (nt->arg),
+														 relid, attnums, exprs);
+
+		/* Otherwise we have Expr IS NULL. */
+		*exprs = lappend(*exprs, nt->arg);
+		return true;
+	}
+
+	/*
+	 * Treat any other expressions as bare expressions to be matched against
+	 * expressions in statistics objects.
+	 */
+	*exprs = lappend(*exprs, clause);
+	return true;
+}
+
+/*
+ * statext_is_compatible_clause
+ *		Determines if the clause is compatible with MCV lists.
+ *
+ * See statext_is_compatible_clause_internal, above, for the basic rules.
+ * This layer deals with RestrictInfo superstructure and applies permissions
+ * checks to verify that it's okay to examine all mentioned Vars.
+ *
+ * Arguments:
+ * clause: clause to be inspected (in RestrictInfo form)
+ * relid: rel that all Vars in clause must belong to
+ * *attnums: input/output parameter collecting attribute numbers of all
+ *		mentioned Vars.  Note that we do not offset the attribute numbers,
+ *		so we can't cope with system columns.
+ * *exprs: input/output parameter collecting primitive subclauses within
+ *		the clause tree
+ *
+ * Returns false if there is something we definitively can't handle.
+ * On true return, we can proceed to match the *exprs against statistics.
+ */
+static bool
+statext_is_compatible_clause(PlannerInfo *root, Node *clause, Index relid,
+							 Bitmapset **attnums, List **exprs)
+{
+	RangeTblEntry *rte = root->simple_rte_array[relid];
+	RestrictInfo *rinfo;
+	int			clause_relid;
+	Oid			userid;
+
+	/*
+	 * Special-case handling for bare BoolExpr AND clauses, because the
+	 * restrictinfo machinery doesn't build RestrictInfos on top of AND
+	 * clauses.
+	 */
+	if (is_andclause(clause))
+	{
+		BoolExpr   *expr = (BoolExpr *) clause;
+		ListCell   *lc;
+
+		/*
+		 * Check that each sub-clause is compatible.  We expect these to be
+		 * RestrictInfos.
+		 */
+		foreach(lc, expr->args)
+		{
+			if (!statext_is_compatible_clause(root, (Node *) lfirst(lc),
+											  relid, attnums, exprs))
+				return false;
+		}
+
+		return true;
+	}
+
+	/* Otherwise it must be a RestrictInfo. */
+	if (!IsA(clause, RestrictInfo))
+		return false;
+	rinfo = (RestrictInfo *) clause;
+
+	/* Pseudoconstants are not really interesting here. */
+	if (rinfo->pseudoconstant)
+		return false;
+
+	/* Clauses referencing other varnos are incompatible. */
+	if (!bms_get_singleton_member(rinfo->clause_relids, &clause_relid) ||
+		clause_relid != relid)
+		return false;
+
+	/* Check the clause and determine what attributes it references. */
+	if (!statext_is_compatible_clause_internal(root, (Node *) rinfo->clause,
+											   relid, attnums, exprs))
+		return false;
+
+	/*
+	 * Check that the user has permission to read all required attributes. Use
+	 * checkAsUser if it's set, in case we're accessing the table via a view.
+	 */
+	userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
+
+	/* Table-level SELECT privilege is sufficient for all columns */
+	if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) != ACLCHECK_OK)
+	{
+		Bitmapset  *clause_attnums = NULL;
+		int			attnum = -1;
+
+		/*
+		 * We have to check per-column privileges.  *attnums has the attnums
+		 * for individual Vars we saw, but there may also be Vars within
+		 * subexpressions in *exprs.  We can use pull_varattnos() to extract
+		 * those, but there's an impedance mismatch: attnums returned by
+		 * pull_varattnos() are offset by FirstLowInvalidHeapAttributeNumber,
+		 * while attnums within *attnums aren't.  Convert *attnums to the
+		 * offset style so we can combine the results.
+		 */
+		while ((attnum = bms_next_member(*attnums, attnum)) >= 0)
+		{
+			clause_attnums =
+				bms_add_member(clause_attnums,
+							   attnum - FirstLowInvalidHeapAttributeNumber);
+		}
+
+		/* Now merge attnums from *exprs into clause_attnums */
+		if (*exprs != NIL)
+			pull_varattnos((Node *) *exprs, relid, &clause_attnums);
+
+		attnum = -1;
+		while ((attnum = bms_next_member(clause_attnums, attnum)) >= 0)
+		{
+			/* Undo the offset */
+			AttrNumber	attno = attnum + FirstLowInvalidHeapAttributeNumber;
+
+			if (attno == InvalidAttrNumber)
+			{
+				/* Whole-row reference, so must have access to all columns */
+				if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
+											  ACLMASK_ALL) != ACLCHECK_OK)
+					return false;
+			}
+			else
+			{
+				if (pg_attribute_aclcheck(rte->relid, attno, userid,
+										  ACL_SELECT) != ACLCHECK_OK)
+					return false;
+			}
+		}
+	}
+
+	/* If we reach here, the clause is OK */
+	return true;
+}
+
+/*
+ * statext_mcv_clauselist_selectivity
+ *		Estimate clauses using the best multi-column statistics.
+ *
+ * Applies available extended (multi-column) statistics on a table. There may
+ * be multiple applicable statistics (with respect to the clauses), in which
+ * case we use greedy approach. In each round we select the best statistic on
+ * a table (measured by the number of attributes extracted from the clauses
+ * and covered by it), and compute the selectivity for the supplied clauses.
+ * We repeat this process with the remaining clauses (if any), until none of
+ * the available statistics can be used.
+ *
+ * One of the main challenges with using MCV lists is how to extrapolate the
+ * estimate to the data not covered by the MCV list. To do that, we compute
+ * not only the "MCV selectivity" (selectivities for MCV items matching the
+ * supplied clauses), but also the following related selectivities:
+ *
+ * - simple selectivity:  Computed without extended statistics, i.e. as if the
+ * columns/clauses were independent.
+ *
+ * - base selectivity:  Similar to simple selectivity, but is computed using
+ * the extended statistic by adding up the base frequencies (that we compute
+ * and store for each MCV item) of matching MCV items.
+ *
+ * - total selectivity: Selectivity covered by the whole MCV list.
+ *
+ * These are passed to mcv_combine_selectivities() which combines them to
+ * produce a selectivity estimate that makes use of both per-column statistics
+ * and the multi-column MCV statistics.
+ *
+ * 'estimatedclauses' is an input/output parameter.  We set bits for the
+ * 0-based 'clauses' indexes we estimate for and also skip clause items that
+ * already have a bit set.
+ */
+static Selectivity
+statext_mcv_clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid,
+								   JoinType jointype, SpecialJoinInfo *sjinfo,
+								   RelOptInfo *rel, Bitmapset **estimatedclauses,
+								   bool is_or)
+{
+	ListCell   *l;
+	Bitmapset **list_attnums;	/* attnums extracted from the clause */
+	List	  **list_exprs;		/* expressions matched to any statistic */
+	int			listidx;
+	Selectivity sel = (is_or) ? 0.0 : 1.0;
+	RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
+
+	/*
+	 * When dealing with regular inheritance trees, ignore extended stats
+	 * (which were built without data from child rels, and thus do not
+	 * represent them). For partitioned tables data there's no data in the
+	 * non-leaf relations, so we build stats only for the inheritance tree.
+	 * So for partitioned tables we do consider extended stats.
+	 */
+	if (rte->inh && rte->relkind != RELKIND_PARTITIONED_TABLE)
+		return sel;
+
+	/* check if there's any stats that might be useful for us. */
+	if (!has_stats_of_kind(rel->statlist, STATS_EXT_MCV))
+		return sel;
+
+	list_attnums = (Bitmapset **) palloc(sizeof(Bitmapset *) *
+										 list_length(clauses));
+
+	/* expressions extracted from complex expressions */
+	list_exprs = (List **) palloc(sizeof(Node *) * list_length(clauses));
+
+	/*
+	 * Pre-process the clauses list to extract the attnums and expressions
+	 * seen in each item.  We need to determine if there are any clauses which
+	 * will be useful for selectivity estimations with extended stats.  Along
+	 * the way we'll record all of the attnums and expressions for each clause
+	 * in lists which we'll reference later so we don't need to repeat the
+	 * same work again.
+	 *
+	 * We also skip clauses that we already estimated using different types of
+	 * statistics (we treat them as incompatible).
+	 */
+	listidx = 0;
+	foreach(l, clauses)
+	{
+		Node	   *clause = (Node *) lfirst(l);
+		Bitmapset  *attnums = NULL;
+		List	   *exprs = NIL;
+
+		if (!bms_is_member(listidx, *estimatedclauses) &&
+			statext_is_compatible_clause(root, clause, rel->relid, &attnums, &exprs))
+		{
+			list_attnums[listidx] = attnums;
+			list_exprs[listidx] = exprs;
+		}
+		else
+		{
+			list_attnums[listidx] = NULL;
+			list_exprs[listidx] = NIL;
+		}
+
+		listidx++;
+	}
+
+	/* apply as many extended statistics as possible */
+	while (true)
+	{
+		StatisticExtInfo *stat;
+		List	   *stat_clauses;
+		Bitmapset  *simple_clauses;
+
+		/* find the best suited statistics object for these attnums */
+		stat = choose_best_statistics(rel->statlist, STATS_EXT_MCV,
+									  list_attnums, list_exprs,
+									  list_length(clauses));
+
+		/*
+		 * if no (additional) matching stats could be found then we've nothing
+		 * to do
+		 */
+		if (!stat)
+			break;
+
+		/* Ensure choose_best_statistics produced an expected stats type. */
+		Assert(stat->kind == STATS_EXT_MCV);
+
+		/* now filter the clauses to be estimated using the selected MCV */
+		stat_clauses = NIL;
+
+		/* record which clauses are simple (single column or expression) */
+		simple_clauses = NULL;
+
+		listidx = -1;
+		foreach(l, clauses)
+		{
+			/* Increment the index before we decide if to skip the clause. */
+			listidx++;
+
+			/*
+			 * Ignore clauses from which we did not extract any attnums or
+			 * expressions (this needs to be consistent with what we do in
+			 * choose_best_statistics).
+			 *
+			 * This also eliminates already estimated clauses - both those
+			 * estimated before and during applying extended statistics.
+			 *
+			 * XXX This check is needed because both bms_is_subset and
+			 * stat_covers_expressions return true for empty attnums and
+			 * expressions.
+			 */
+			if (!list_attnums[listidx] && !list_exprs[listidx])
+				continue;
+
+			/*
+			 * The clause was not estimated yet, and we've extracted either
+			 * attnums or expressions from it. Ignore it if it's not fully
+			 * covered by the chosen statistics object.
+			 *
+			 * We need to check both attributes and expressions, and reject if
+			 * either is not covered.
+			 */
+			if (!bms_is_subset(list_attnums[listidx], stat->keys) ||
+				!stat_covers_expressions(stat, list_exprs[listidx], NULL))
+				continue;
+
+			/*
+			 * Now we know the clause is compatible (we have either attnums or
+			 * expressions extracted from it), and was not estimated yet.
+			 */
+
+			/* record simple clauses (single column or expression) */
+			if ((list_attnums[listidx] == NULL &&
+				 list_length(list_exprs[listidx]) == 1) ||
+				(list_exprs[listidx] == NIL &&
+				 bms_membership(list_attnums[listidx]) == BMS_SINGLETON))
+				simple_clauses = bms_add_member(simple_clauses,
+												list_length(stat_clauses));
+
+			/* add clause to list and mark it as estimated */
+			stat_clauses = lappend(stat_clauses, (Node *) lfirst(l));
+			*estimatedclauses = bms_add_member(*estimatedclauses, listidx);
+
+			/*
+			 * Reset the pointers, so that choose_best_statistics knows this
+			 * clause was estimated and does not consider it again.
+			 */
+			bms_free(list_attnums[listidx]);
+			list_attnums[listidx] = NULL;
+
+			list_free(list_exprs[listidx]);
+			list_exprs[listidx] = NULL;
+		}
+
+		if (is_or)
+		{
+			bool	   *or_matches = NULL;
+			Selectivity simple_or_sel = 0.0,
+						stat_sel = 0.0;
+			MCVList    *mcv_list;
+
+			/* Load the MCV list stored in the statistics object */
+			mcv_list = statext_mcv_load(stat->statOid);
+
+			/*
+			 * Compute the selectivity of the ORed list of clauses covered by
+			 * this statistics object by estimating each in turn and combining
+			 * them using the formula P(A OR B) = P(A) + P(B) - P(A AND B).
+			 * This allows us to use the multivariate MCV stats to better
+			 * estimate the individual terms and their overlap.
+			 *
+			 * Each time we iterate this formula, the clause "A" above is
+			 * equal to all the clauses processed so far, combined with "OR".
+			 */
+			listidx = 0;
+			foreach(l, stat_clauses)
+			{
+				Node	   *clause = (Node *) lfirst(l);
+				Selectivity simple_sel,
+							overlap_simple_sel,
+							mcv_sel,
+							mcv_basesel,
+							overlap_mcvsel,
+							overlap_basesel,
+							mcv_totalsel,
+							clause_sel,
+							overlap_sel;
+
+				/*
+				 * "Simple" selectivity of the next clause and its overlap
+				 * with any of the previous clauses.  These are our initial
+				 * estimates of P(B) and P(A AND B), assuming independence of
+				 * columns/clauses.
+				 */
+				simple_sel = clause_selectivity_ext(root, clause, varRelid,
+													jointype, sjinfo, false);
+
+				overlap_simple_sel = simple_or_sel * simple_sel;
+
+				/*
+				 * New "simple" selectivity of all clauses seen so far,
+				 * assuming independence.
+				 */
+				simple_or_sel += simple_sel - overlap_simple_sel;
+				CLAMP_PROBABILITY(simple_or_sel);
+
+				/*
+				 * Multi-column estimate of this clause using MCV statistics,
+				 * along with base and total selectivities, and corresponding
+				 * selectivities for the overlap term P(A AND B).
+				 */
+				mcv_sel = mcv_clause_selectivity_or(root, stat, mcv_list,
+													clause, &or_matches,
+													&mcv_basesel,
+													&overlap_mcvsel,
+													&overlap_basesel,
+													&mcv_totalsel);
+
+				/*
+				 * Combine the simple and multi-column estimates.
+				 *
+				 * If this clause is a simple single-column clause, then we
+				 * just use the simple selectivity estimate for it, since the
+				 * multi-column statistics are unlikely to improve on that
+				 * (and in fact could make it worse).  For the overlap, we
+				 * always make use of the multi-column statistics.
+				 */
+				if (bms_is_member(listidx, simple_clauses))
+					clause_sel = simple_sel;
+				else
+					clause_sel = mcv_combine_selectivities(simple_sel,
+														   mcv_sel,
+														   mcv_basesel,
+														   mcv_totalsel);
+
+				overlap_sel = mcv_combine_selectivities(overlap_simple_sel,
+														overlap_mcvsel,
+														overlap_basesel,
+														mcv_totalsel);
+
+				/* Factor these into the result for this statistics object */
+				stat_sel += clause_sel - overlap_sel;
+				CLAMP_PROBABILITY(stat_sel);
+
+				listidx++;
+			}
+
+			/*
+			 * Factor the result for this statistics object into the overall
+			 * result.  We treat the results from each separate statistics
+			 * object as independent of one another.
+			 */
+			sel = sel + stat_sel - sel * stat_sel;
+		}
+		else					/* Implicitly-ANDed list of clauses */
+		{
+			Selectivity simple_sel,
+						mcv_sel,
+						mcv_basesel,
+						mcv_totalsel,
+						stat_sel;
+
+			/*
+			 * "Simple" selectivity, i.e. without any extended statistics,
+			 * essentially assuming independence of the columns/clauses.
+			 */
+			simple_sel = clauselist_selectivity_ext(root, stat_clauses,
+													varRelid, jointype,
+													sjinfo, false);
+
+			/*
+			 * Multi-column estimate using MCV statistics, along with base and
+			 * total selectivities.
+			 */
+			mcv_sel = mcv_clauselist_selectivity(root, stat, stat_clauses,
+												 varRelid, jointype, sjinfo,
+												 rel, &mcv_basesel,
+												 &mcv_totalsel);
+
+			/* Combine the simple and multi-column estimates. */
+			stat_sel = mcv_combine_selectivities(simple_sel,
+												 mcv_sel,
+												 mcv_basesel,
+												 mcv_totalsel);
+
+			/* Factor this into the overall result */
+			sel *= stat_sel;
+		}
+	}
+
+	return sel;
+}
+
+/*
+ * statext_clauselist_selectivity
+ *		Estimate clauses using the best multi-column statistics.
+ */
+Selectivity
+statext_clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid,
+							   JoinType jointype, SpecialJoinInfo *sjinfo,
+							   RelOptInfo *rel, Bitmapset **estimatedclauses,
+							   bool is_or)
+{
+	Selectivity sel;
+
+	/* First, try estimating clauses using a multivariate MCV list. */
+	sel = statext_mcv_clauselist_selectivity(root, clauses, varRelid, jointype,
+											 sjinfo, rel, estimatedclauses, is_or);
+
+	/*
+	 * Functional dependencies only work for clauses connected by AND, so for
+	 * OR clauses we're done.
+	 */
+	if (is_or)
+		return sel;
+
+	/*
+	 * Then, apply functional dependencies on the remaining clauses by calling
+	 * dependencies_clauselist_selectivity.  Pass 'estimatedclauses' so the
+	 * function can properly skip clauses already estimated above.
+	 *
+	 * The reasoning for applying dependencies last is that the more complex
+	 * stats can track more complex correlations between the attributes, and
+	 * so may be considered more reliable.
+	 *
+	 * For example, MCV list can give us an exact selectivity for values in
+	 * two columns, while functional dependencies can only provide information
+	 * about the overall strength of the dependency.
+	 */
+	sel *= dependencies_clauselist_selectivity(root, clauses, varRelid,
+											   jointype, sjinfo, rel,
+											   estimatedclauses);
+
+	return sel;
+}
+
+/*
+ * examine_opclause_args
+ *		Split an operator expression's arguments into Expr and Const parts.
+ *
+ * Attempts to match the arguments to either (Expr op Const) or (Const op
+ * Expr), possibly with a RelabelType on top. When the expression matches this
+ * form, returns true, otherwise returns false.
+ *
+ * Optionally returns pointers to the extracted Expr/Const nodes, when passed
+ * non-null pointers (exprp, cstp and expronleftp). The expronleftp flag
+ * specifies on which side of the operator we found the expression node.
+ */
+bool
+examine_opclause_args(List *args, Node **exprp, Const **cstp,
+					  bool *expronleftp)
+{
+	Node	   *expr;
+	Const	   *cst;
+	bool		expronleft;
+	Node	   *leftop,
+			   *rightop;
+
+	/* enforced by statext_is_compatible_clause_internal */
+	Assert(list_length(args) == 2);
+
+	leftop = linitial(args);
+	rightop = lsecond(args);
+
+	/* strip RelabelType from either side of the expression */
+	if (IsA(leftop, RelabelType))
+		leftop = (Node *) ((RelabelType *) leftop)->arg;
+
+	if (IsA(rightop, RelabelType))
+		rightop = (Node *) ((RelabelType *) rightop)->arg;
+
+	if (IsA(rightop, Const))
+	{
+		expr = (Node *) leftop;
+		cst = (Const *) rightop;
+		expronleft = true;
+	}
+	else if (IsA(leftop, Const))
+	{
+		expr = (Node *) rightop;
+		cst = (Const *) leftop;
+		expronleft = false;
+	}
+	else
+		return false;
+
+	/* return pointers to the extracted parts if requested */
+	if (exprp)
+		*exprp = expr;
+
+	if (cstp)
+		*cstp = cst;
+
+	if (expronleftp)
+		*expronleftp = expronleft;
+
+	return true;
+}
+
+
+/*
+ * Compute statistics about expressions of a relation.
+ */
+static void
+compute_expr_stats(Relation onerel, double totalrows,
+				   AnlExprData *exprdata, int nexprs,
+				   HeapTuple *rows, int numrows)
+{
+	MemoryContext expr_context,
+				old_context;
+	int			ind,
+				i;
+
+	expr_context = AllocSetContextCreate(CurrentMemoryContext,
+										 "Analyze Expression",
+										 ALLOCSET_DEFAULT_SIZES);
+	old_context = MemoryContextSwitchTo(expr_context);
+
+	for (ind = 0; ind < nexprs; ind++)
+	{
+		AnlExprData *thisdata = &exprdata[ind];
+		VacAttrStats *stats = thisdata->vacattrstat;
+		Node	   *expr = thisdata->expr;
+		TupleTableSlot *slot;
+		EState	   *estate;
+		ExprContext *econtext;
+		Datum	   *exprvals;
+		bool	   *exprnulls;
+		ExprState  *exprstate;
+		int			tcnt;
+
+		/* Are we still in the main context? */
+		Assert(CurrentMemoryContext == expr_context);
+
+		/*
+		 * Need an EState for evaluation of expressions.  Create it in the
+		 * per-expression context to be sure it gets cleaned up at the bottom
+		 * of the loop.
+		 */
+		estate = CreateExecutorState();
+		econtext = GetPerTupleExprContext(estate);
+
+		/* Set up expression evaluation state */
+		exprstate = ExecPrepareExpr((Expr *) expr, estate);
+
+		/* Need a slot to hold the current heap tuple, too */
+		slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
+										&TTSOpsHeapTuple);
+
+		/* Arrange for econtext's scan tuple to be the tuple under test */
+		econtext->ecxt_scantuple = slot;
+
+		/* Compute and save expression values */
+		exprvals = (Datum *) palloc(numrows * sizeof(Datum));
+		exprnulls = (bool *) palloc(numrows * sizeof(bool));
+
+		tcnt = 0;
+		for (i = 0; i < numrows; i++)
+		{
+			Datum		datum;
+			bool		isnull;
+
+			/*
+			 * Reset the per-tuple context each time, to reclaim any cruft
+			 * left behind by evaluating the statistics expressions.
+			 */
+			ResetExprContext(econtext);
+
+			/* Set up for expression evaluation */
+			ExecStoreHeapTuple(rows[i], slot, false);
+
+			/*
+			 * Evaluate the expression. We do this in the per-tuple context so
+			 * as not to leak memory, and then copy the result into the
+			 * context created at the beginning of this function.
+			 */
+			datum = ExecEvalExprSwitchContext(exprstate,
+											  GetPerTupleExprContext(estate),
+											  &isnull);
+			if (isnull)
+			{
+				exprvals[tcnt] = (Datum) 0;
+				exprnulls[tcnt] = true;
+			}
+			else
+			{
+				/* Make sure we copy the data into the context. */
+				Assert(CurrentMemoryContext == expr_context);
+
+				exprvals[tcnt] = datumCopy(datum,
+										   stats->attrtype->typbyval,
+										   stats->attrtype->typlen);
+				exprnulls[tcnt] = false;
+			}
+
+			tcnt++;
+		}
+
+		/*
+		 * Now we can compute the statistics for the expression columns.
+		 *
+		 * XXX Unlike compute_index_stats we don't need to switch and reset
+		 * memory contexts here, because we're only computing stats for a
+		 * single expression (and not iterating over many indexes), so we just
+		 * do it in expr_context. Note that compute_stats copies the result
+		 * into stats->anl_context, so it does not disappear.
+		 */
+		if (tcnt > 0)
+		{
+			AttributeOpts *aopt =
+			get_attribute_options(stats->attr->attrelid,
+								  stats->attr->attnum);
+
+			stats->exprvals = exprvals;
+			stats->exprnulls = exprnulls;
+			stats->rowstride = 1;
+			stats->compute_stats(stats,
+								 expr_fetch_func,
+								 tcnt,
+								 tcnt);
+
+			/*
+			 * If the n_distinct option is specified, it overrides the above
+			 * computation.
+			 */
+			if (aopt != NULL && aopt->n_distinct != 0.0)
+				stats->stadistinct = aopt->n_distinct;
+		}
+
+		/* And clean up */
+		MemoryContextSwitchTo(expr_context);
+
+		ExecDropSingleTupleTableSlot(slot);
+		FreeExecutorState(estate);
+		MemoryContextResetAndDeleteChildren(expr_context);
+	}
+
+	MemoryContextSwitchTo(old_context);
+	MemoryContextDelete(expr_context);
+}
+
+
+/*
+ * Fetch function for analyzing statistics object expressions.
+ *
+ * We have not bothered to construct tuples from the data, instead the data
+ * is just in Datum arrays.
+ */
+static Datum
+expr_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
+{
+	int			i;
+
+	/* exprvals and exprnulls are already offset for proper column */
+	i = rownum * stats->rowstride;
+	*isNull = stats->exprnulls[i];
+	return stats->exprvals[i];
+}
+
+/*
+ * Build analyze data for a list of expressions. As this is not tied
+ * directly to a relation (table or index), we have to fake some of
+ * the fields in examine_expression().
+ */
+static AnlExprData *
+build_expr_data(List *exprs, int stattarget)
+{
+	int			idx;
+	int			nexprs = list_length(exprs);
+	AnlExprData *exprdata;
+	ListCell   *lc;
+
+	exprdata = (AnlExprData *) palloc0(nexprs * sizeof(AnlExprData));
+
+	idx = 0;
+	foreach(lc, exprs)
+	{
+		Node	   *expr = (Node *) lfirst(lc);
+		AnlExprData *thisdata = &exprdata[idx];
+
+		thisdata->expr = expr;
+		thisdata->vacattrstat = examine_expression(expr, stattarget);
+		idx++;
+	}
+
+	return exprdata;
+}
+
+/* form an array of pg_statistic rows (per update_attstats) */
+static Datum
+serialize_expr_stats(AnlExprData *exprdata, int nexprs)
+{
+	int			exprno;
+	Oid			typOid;
+	Relation	sd;
+
+	ArrayBuildState *astate = NULL;
+
+	sd = table_open(StatisticRelationId, RowExclusiveLock);
+
+	/* lookup OID of composite type for pg_statistic */
+	typOid = get_rel_type_id(StatisticRelationId);
+	if (!OidIsValid(typOid))
+		ereport(ERROR,
+				(errcode(ERRCODE_WRONG_OBJECT_TYPE),
+				 errmsg("relation \"%s\" does not have a composite type",
+						"pg_statistic")));
+
+	for (exprno = 0; exprno < nexprs; exprno++)
+	{
+		int			i,
+					k;
+		VacAttrStats *stats = exprdata[exprno].vacattrstat;
+
+		Datum		values[Natts_pg_statistic];
+		bool		nulls[Natts_pg_statistic];
+		HeapTuple	stup;
+
+		if (!stats->stats_valid)
+		{
+			astate = accumArrayResult(astate,
+									  (Datum) 0,
+									  true,
+									  typOid,
+									  CurrentMemoryContext);
+			continue;
+		}
+
+		/*
+		 * Construct a new pg_statistic tuple
+		 */
+		for (i = 0; i < Natts_pg_statistic; ++i)
+		{
+			nulls[i] = false;
+		}
+
+		values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(InvalidOid);
+		values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(InvalidAttrNumber);
+		values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(false);
+		values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
+		values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
+		values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
+		i = Anum_pg_statistic_stakind1 - 1;
+		for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
+		{
+			values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
+		}
+		i = Anum_pg_statistic_staop1 - 1;
+		for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
+		{
+			values[i++] = ObjectIdGetDatum(stats->staop[k]);	/* staopN */
+		}
+		i = Anum_pg_statistic_stacoll1 - 1;
+		for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
+		{
+			values[i++] = ObjectIdGetDatum(stats->stacoll[k]);	/* stacollN */
+		}
+		i = Anum_pg_statistic_stanumbers1 - 1;
+		for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
+		{
+			int			nnum = stats->numnumbers[k];
+
+			if (nnum > 0)
+			{
+				int			n;
+				Datum	   *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
+				ArrayType  *arry;
+
+				for (n = 0; n < nnum; n++)
+					numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
+				/* XXX knows more than it should about type float4: */
+				arry = construct_array(numdatums, nnum,
+									   FLOAT4OID,
+									   sizeof(float4), true, TYPALIGN_INT);
+				values[i++] = PointerGetDatum(arry);	/* stanumbersN */
+			}
+			else
+			{
+				nulls[i] = true;
+				values[i++] = (Datum) 0;
+			}
+		}
+		i = Anum_pg_statistic_stavalues1 - 1;
+		for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
+		{
+			if (stats->numvalues[k] > 0)
+			{
+				ArrayType  *arry;
+
+				arry = construct_array(stats->stavalues[k],
+									   stats->numvalues[k],
+									   stats->statypid[k],
+									   stats->statyplen[k],
+									   stats->statypbyval[k],
+									   stats->statypalign[k]);
+				values[i++] = PointerGetDatum(arry);	/* stavaluesN */
+			}
+			else
+			{
+				nulls[i] = true;
+				values[i++] = (Datum) 0;
+			}
+		}
+
+		stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
+
+		astate = accumArrayResult(astate,
+								  heap_copy_tuple_as_datum(stup, RelationGetDescr(sd)),
+								  false,
+								  typOid,
+								  CurrentMemoryContext);
+	}
+
+	table_close(sd, RowExclusiveLock);
+
+	return makeArrayResult(astate, CurrentMemoryContext);
+}
+
+/*
+ * Loads pg_statistic record from expression statistics for expression
+ * identified by the supplied index.
+ */
+HeapTuple
+statext_expressions_load(Oid stxoid, int idx)
+{
+	bool		isnull;
+	Datum		value;
+	HeapTuple	htup;
+	ExpandedArrayHeader *eah;
+	HeapTupleHeader td;
+	HeapTupleData tmptup;
+	HeapTuple	tup;
+
+	htup = SearchSysCache1(STATEXTDATASTXOID, ObjectIdGetDatum(stxoid));
+	if (!HeapTupleIsValid(htup))
+		elog(ERROR, "cache lookup failed for statistics object %u", stxoid);
+
+	value = SysCacheGetAttr(STATEXTDATASTXOID, htup,
+							Anum_pg_statistic_ext_data_stxdexpr, &isnull);
+	if (isnull)
+		elog(ERROR,
+			 "requested statistics kind \"%c\" is not yet built for statistics object %u",
+			 STATS_EXT_DEPENDENCIES, stxoid);
+
+	eah = DatumGetExpandedArray(value);
+
+	deconstruct_expanded_array(eah);
+
+	td = DatumGetHeapTupleHeader(eah->dvalues[idx]);
+
+	/* Build a temporary HeapTuple control structure */
+	tmptup.t_len = HeapTupleHeaderGetDatumLength(td);
+	ItemPointerSetInvalid(&(tmptup.t_self));
+	tmptup.t_tableOid = InvalidOid;
+	tmptup.t_data = td;
+
+	tup = heap_copytuple(&tmptup);
+
+	ReleaseSysCache(htup);
+
+	return tup;
+}
+
+/*
+ * Evaluate the expressions, so that we can use the results to build
+ * all the requested statistics types. This matters especially for
+ * expensive expressions, of course.
+ */
+static StatsBuildData *
+make_build_data(Relation rel, StatExtEntry *stat, int numrows, HeapTuple *rows,
+				VacAttrStats **stats, int stattarget)
+{
+	/* evaluated expressions */
+	StatsBuildData *result;
+	char	   *ptr;
+	Size		len;
+
+	int			i;
+	int			k;
+	int			idx;
+	TupleTableSlot *slot;
+	EState	   *estate;
+	ExprContext *econtext;
+	List	   *exprstates = NIL;
+	int			nkeys = bms_num_members(stat->columns) + list_length(stat->exprs);
+	ListCell   *lc;
+
+	/* allocate everything as a single chunk, so we can free it easily */
+	len = MAXALIGN(sizeof(StatsBuildData));
+	len += MAXALIGN(sizeof(AttrNumber) * nkeys);	/* attnums */
+	len += MAXALIGN(sizeof(VacAttrStats *) * nkeys);	/* stats */
+
+	/* values */
+	len += MAXALIGN(sizeof(Datum *) * nkeys);
+	len += nkeys * MAXALIGN(sizeof(Datum) * numrows);
+
+	/* nulls */
+	len += MAXALIGN(sizeof(bool *) * nkeys);
+	len += nkeys * MAXALIGN(sizeof(bool) * numrows);
+
+	ptr = palloc(len);
+
+	/* set the pointers */
+	result = (StatsBuildData *) ptr;
+	ptr += MAXALIGN(sizeof(StatsBuildData));
+
+	/* attnums */
+	result->attnums = (AttrNumber *) ptr;
+	ptr += MAXALIGN(sizeof(AttrNumber) * nkeys);
+
+	/* stats */
+	result->stats = (VacAttrStats **) ptr;
+	ptr += MAXALIGN(sizeof(VacAttrStats *) * nkeys);
+
+	/* values */
+	result->values = (Datum **) ptr;
+	ptr += MAXALIGN(sizeof(Datum *) * nkeys);
+
+	/* nulls */
+	result->nulls = (bool **) ptr;
+	ptr += MAXALIGN(sizeof(bool *) * nkeys);
+
+	for (i = 0; i < nkeys; i++)
+	{
+		result->values[i] = (Datum *) ptr;
+		ptr += MAXALIGN(sizeof(Datum) * numrows);
+
+		result->nulls[i] = (bool *) ptr;
+		ptr += MAXALIGN(sizeof(bool) * numrows);
+	}
+
+	Assert((ptr - (char *) result) == len);
+
+	/* we have it allocated, so let's fill the values */
+	result->nattnums = nkeys;
+	result->numrows = numrows;
+
+	/* fill the attribute info - first attributes, then expressions */
+	idx = 0;
+	k = -1;
+	while ((k = bms_next_member(stat->columns, k)) >= 0)
+	{
+		result->attnums[idx] = k;
+		result->stats[idx] = stats[idx];
+
+		idx++;
+	}
+
+	k = -1;
+	foreach(lc, stat->exprs)
+	{
+		Node	   *expr = (Node *) lfirst(lc);
+
+		result->attnums[idx] = k;
+		result->stats[idx] = examine_expression(expr, stattarget);
+
+		idx++;
+		k--;
+	}
+
+	/* first extract values for all the regular attributes */
+	for (i = 0; i < numrows; i++)
+	{
+		idx = 0;
+		k = -1;
+		while ((k = bms_next_member(stat->columns, k)) >= 0)
+		{
+			result->values[idx][i] = heap_getattr(rows[i], k,
+												  result->stats[idx]->tupDesc,
+												  &result->nulls[idx][i]);
+
+			idx++;
+		}
+	}
+
+	/* Need an EState for evaluation expressions. */
+	estate = CreateExecutorState();
+	econtext = GetPerTupleExprContext(estate);
+
+	/* Need a slot to hold the current heap tuple, too */
+	slot = MakeSingleTupleTableSlot(RelationGetDescr(rel),
+									&TTSOpsHeapTuple);
+
+	/* Arrange for econtext's scan tuple to be the tuple under test */
+	econtext->ecxt_scantuple = slot;
+
+	/* Set up expression evaluation state */
+	exprstates = ExecPrepareExprList(stat->exprs, estate);
+
+	for (i = 0; i < numrows; i++)
+	{
+		/*
+		 * Reset the per-tuple context each time, to reclaim any cruft left
+		 * behind by evaluating the statistics object expressions.
+		 */
+		ResetExprContext(econtext);
+
+		/* Set up for expression evaluation */
+		ExecStoreHeapTuple(rows[i], slot, false);
+
+		idx = bms_num_members(stat->columns);
+		foreach(lc, exprstates)
+		{
+			Datum		datum;
+			bool		isnull;
+			ExprState  *exprstate = (ExprState *) lfirst(lc);
+
+			/*
+			 * XXX This probably leaks memory. Maybe we should use
+			 * ExecEvalExprSwitchContext but then we need to copy the result
+			 * somewhere else.
+			 */
+			datum = ExecEvalExpr(exprstate,
+								 GetPerTupleExprContext(estate),
+								 &isnull);
+			if (isnull)
+			{
+				result->values[idx][i] = (Datum) 0;
+				result->nulls[idx][i] = true;
+			}
+			else
+			{
+				result->values[idx][i] = (Datum) datum;
+				result->nulls[idx][i] = false;
+			}
+
+			idx++;
+		}
+	}
+
+	ExecDropSingleTupleTableSlot(slot);
+	FreeExecutorState(estate);
+
+	return result;
+}
diff --git a/src/backend/statistics/mcv.c b/src/backend/statistics/mcv.c
new file mode 100644
index 0000000..0183775
--- /dev/null
+++ b/src/backend/statistics/mcv.c
@@ -0,0 +1,2180 @@
+/*-------------------------------------------------------------------------
+ *
+ * mcv.c
+ *	  POSTGRES multivariate MCV lists
+ *
+ *
+ * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * IDENTIFICATION
+ *	  src/backend/statistics/mcv.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include <math.h>
+
+#include "access/htup_details.h"
+#include "catalog/pg_collation.h"
+#include "catalog/pg_statistic_ext.h"
+#include "catalog/pg_statistic_ext_data.h"
+#include "fmgr.h"
+#include "funcapi.h"
+#include "nodes/nodeFuncs.h"
+#include "optimizer/clauses.h"
+#include "statistics/extended_stats_internal.h"
+#include "statistics/statistics.h"
+#include "utils/array.h"
+#include "utils/builtins.h"
+#include "utils/bytea.h"
+#include "utils/fmgroids.h"
+#include "utils/fmgrprotos.h"
+#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
+#include "utils/syscache.h"
+#include "utils/typcache.h"
+
+/*
+ * Computes size of a serialized MCV item, depending on the number of
+ * dimensions (columns) the statistic is defined on. The datum values are
+ * stored in a separate array (deduplicated, to minimize the size), and
+ * so the serialized items only store uint16 indexes into that array.
+ *
+ * Each serialized item stores (in this order):
+ *
+ * - indexes to values	  (ndim * sizeof(uint16))
+ * - null flags			  (ndim * sizeof(bool))
+ * - frequency			  (sizeof(double))
+ * - base_frequency		  (sizeof(double))
+ *
+ * There is no alignment padding within an MCV item.
+ * So in total each MCV item requires this many bytes:
+ *
+ *	 ndim * (sizeof(uint16) + sizeof(bool)) + 2 * sizeof(double)
+ */
+#define ITEM_SIZE(ndims)	\
+	((ndims) * (sizeof(uint16) + sizeof(bool)) + 2 * sizeof(double))
+
+/*
+ * Used to compute size of serialized MCV list representation.
+ */
+#define MinSizeOfMCVList		\
+	(VARHDRSZ + sizeof(uint32) * 3 + sizeof(AttrNumber))
+
+/*
+ * Size of the serialized MCV list, excluding the space needed for
+ * deduplicated per-dimension values. The macro is meant to be used
+ * when it's not yet safe to access the serialized info about amount
+ * of data for each column.
+ */
+#define SizeOfMCVList(ndims,nitems)	\
+	((MinSizeOfMCVList + sizeof(Oid) * (ndims)) + \
+	 ((ndims) * sizeof(DimensionInfo)) + \
+	 ((nitems) * ITEM_SIZE(ndims)))
+
+static MultiSortSupport build_mss(StatsBuildData *data);
+
+static SortItem *build_distinct_groups(int numrows, SortItem *items,
+									   MultiSortSupport mss, int *ndistinct);
+
+static SortItem **build_column_frequencies(SortItem *groups, int ngroups,
+										   MultiSortSupport mss, int *ncounts);
+
+static int	count_distinct_groups(int numrows, SortItem *items,
+								  MultiSortSupport mss);
+
+/*
+ * Compute new value for bitmap item, considering whether it's used for
+ * clauses connected by AND/OR.
+ */
+#define RESULT_MERGE(value, is_or, match) \
+	((is_or) ? ((value) || (match)) : ((value) && (match)))
+
+/*
+ * When processing a list of clauses, the bitmap item may get set to a value
+ * such that additional clauses can't change it. For example, when processing
+ * a list of clauses connected to AND, as soon as the item gets set to 'false'
+ * then it'll remain like that. Similarly clauses connected by OR and 'true'.
+ *
+ * Returns true when the value in the bitmap can't change no matter how the
+ * remaining clauses are evaluated.
+ */
+#define RESULT_IS_FINAL(value, is_or)	((is_or) ? (value) : (!(value)))
+
+/*
+ * get_mincount_for_mcv_list
+ * 		Determine the minimum number of times a value needs to appear in
+ * 		the sample for it to be included in the MCV list.
+ *
+ * We want to keep only values that appear sufficiently often in the
+ * sample that it is reasonable to extrapolate their sample frequencies to
+ * the entire table.  We do this by placing an upper bound on the relative
+ * standard error of the sample frequency, so that any estimates the
+ * planner generates from the MCV statistics can be expected to be
+ * reasonably accurate.
+ *
+ * Since we are sampling without replacement, the sample frequency of a
+ * particular value is described by a hypergeometric distribution.  A
+ * common rule of thumb when estimating errors in this situation is to
+ * require at least 10 instances of the value in the sample, in which case
+ * the distribution can be approximated by a normal distribution, and
+ * standard error analysis techniques can be applied.  Given a sample size
+ * of n, a population size of N, and a sample frequency of p=cnt/n, the
+ * standard error of the proportion p is given by
+ *		SE = sqrt(p*(1-p)/n) * sqrt((N-n)/(N-1))
+ * where the second term is the finite population correction.  To get
+ * reasonably accurate planner estimates, we impose an upper bound on the
+ * relative standard error of 20% -- i.e., SE/p < 0.2.  This 20% relative
+ * error bound is fairly arbitrary, but has been found empirically to work
+ * well.  Rearranging this formula gives a lower bound on the number of
+ * instances of the value seen:
+ *		cnt > n*(N-n) / (N-n+0.04*n*(N-1))
+ * This bound is at most 25, and approaches 0 as n approaches 0 or N. The
+ * case where n approaches 0 cannot happen in practice, since the sample
+ * size is at least 300.  The case where n approaches N corresponds to
+ * sampling the whole the table, in which case it is reasonable to keep
+ * the whole MCV list (have no lower bound), so it makes sense to apply
+ * this formula for all inputs, even though the above derivation is
+ * technically only valid when the right hand side is at least around 10.
+ *
+ * An alternative way to look at this formula is as follows -- assume that
+ * the number of instances of the value seen scales up to the entire
+ * table, so that the population count is K=N*cnt/n. Then the distribution
+ * in the sample is a hypergeometric distribution parameterised by N, n
+ * and K, and the bound above is mathematically equivalent to demanding
+ * that the standard deviation of that distribution is less than 20% of
+ * its mean.  Thus the relative errors in any planner estimates produced
+ * from the MCV statistics are likely to be not too large.
+ */
+static double
+get_mincount_for_mcv_list(int samplerows, double totalrows)
+{
+	double		n = samplerows;
+	double		N = totalrows;
+	double		numer,
+				denom;
+
+	numer = n * (N - n);
+	denom = N - n + 0.04 * n * (N - 1);
+
+	/* Guard against division by zero (possible if n = N = 1) */
+	if (denom == 0.0)
+		return 0.0;
+
+	return numer / denom;
+}
+
+/*
+ * Builds MCV list from the set of sampled rows.
+ *
+ * The algorithm is quite simple:
+ *
+ *	   (1) sort the data (default collation, '<' for the data type)
+ *
+ *	   (2) count distinct groups, decide how many to keep
+ *
+ *	   (3) build the MCV list using the threshold determined in (2)
+ *
+ *	   (4) remove rows represented by the MCV from the sample
+ *
+ */
+MCVList *
+statext_mcv_build(StatsBuildData *data, double totalrows, int stattarget)
+{
+	int			i,
+				numattrs,
+				numrows,
+				ngroups,
+				nitems;
+	double		mincount;
+	SortItem   *items;
+	SortItem   *groups;
+	MCVList    *mcvlist = NULL;
+	MultiSortSupport mss;
+
+	/* comparator for all the columns */
+	mss = build_mss(data);
+
+	/* sort the rows */
+	items = build_sorted_items(data, &nitems, mss,
+							   data->nattnums, data->attnums);
+
+	if (!items)
+		return NULL;
+
+	/* for convenience */
+	numattrs = data->nattnums;
+	numrows = data->numrows;
+
+	/* transform the sorted rows into groups (sorted by frequency) */
+	groups = build_distinct_groups(nitems, items, mss, &ngroups);
+
+	/*
+	 * The maximum number of MCV items to store, based on the statistics
+	 * target we computed for the statistics object (from the target set for
+	 * the object itself, attributes and the system default). In any case, we
+	 * can't keep more groups than we have available.
+	 */
+	nitems = stattarget;
+	if (nitems > ngroups)
+		nitems = ngroups;
+
+	/*
+	 * Decide how many items to keep in the MCV list. We can't use the same
+	 * algorithm as per-column MCV lists, because that only considers the
+	 * actual group frequency - but we're primarily interested in how the
+	 * actual frequency differs from the base frequency (product of simple
+	 * per-column frequencies, as if the columns were independent).
+	 *
+	 * Using the same algorithm might exclude items that are close to the
+	 * "average" frequency of the sample. But that does not say whether the
+	 * observed frequency is close to the base frequency or not. We also need
+	 * to consider unexpectedly uncommon items (again, compared to the base
+	 * frequency), and the single-column algorithm does not have to.
+	 *
+	 * We simply decide how many items to keep by computing the minimum count
+	 * using get_mincount_for_mcv_list() and then keep all items that seem to
+	 * be more common than that.
+	 */
+	mincount = get_mincount_for_mcv_list(numrows, totalrows);
+
+	/*
+	 * Walk the groups until we find the first group with a count below the
+	 * mincount threshold (the index of that group is the number of groups we
+	 * want to keep).
+	 */
+	for (i = 0; i < nitems; i++)
+	{
+		if (groups[i].count < mincount)
+		{
+			nitems = i;
+			break;
+		}
+	}
+
+	/*
+	 * At this point, we know the number of items for the MCV list. There
+	 * might be none (for uniform distribution with many groups), and in that
+	 * case, there will be no MCV list. Otherwise, construct the MCV list.
+	 */
+	if (nitems > 0)
+	{
+		int			j;
+		SortItem	key;
+		MultiSortSupport tmp;
+
+		/* frequencies for values in each attribute */
+		SortItem  **freqs;
+		int		   *nfreqs;
+
+		/* used to search values */
+		tmp = (MultiSortSupport) palloc(offsetof(MultiSortSupportData, ssup)
+										+ sizeof(SortSupportData));
+
+		/* compute frequencies for values in each column */
+		nfreqs = (int *) palloc0(sizeof(int) * numattrs);
+		freqs = build_column_frequencies(groups, ngroups, mss, nfreqs);
+
+		/*
+		 * Allocate the MCV list structure, set the global parameters.
+		 */
+		mcvlist = (MCVList *) palloc0(offsetof(MCVList, items) +
+									  sizeof(MCVItem) * nitems);
+
+		mcvlist->magic = STATS_MCV_MAGIC;
+		mcvlist->type = STATS_MCV_TYPE_BASIC;
+		mcvlist->ndimensions = numattrs;
+		mcvlist->nitems = nitems;
+
+		/* store info about data type OIDs */
+		for (i = 0; i < numattrs; i++)
+			mcvlist->types[i] = data->stats[i]->attrtypid;
+
+		/* Copy the first chunk of groups into the result. */
+		for (i = 0; i < nitems; i++)
+		{
+			/* just pointer to the proper place in the list */
+			MCVItem    *item = &mcvlist->items[i];
+
+			item->values = (Datum *) palloc(sizeof(Datum) * numattrs);
+			item->isnull = (bool *) palloc(sizeof(bool) * numattrs);
+
+			/* copy values for the group */
+			memcpy(item->values, groups[i].values, sizeof(Datum) * numattrs);
+			memcpy(item->isnull, groups[i].isnull, sizeof(bool) * numattrs);
+
+			/* groups should be sorted by frequency in descending order */
+			Assert((i == 0) || (groups[i - 1].count >= groups[i].count));
+
+			/* group frequency */
+			item->frequency = (double) groups[i].count / numrows;
+
+			/* base frequency, if the attributes were independent */
+			item->base_frequency = 1.0;
+			for (j = 0; j < numattrs; j++)
+			{
+				SortItem   *freq;
+
+				/* single dimension */
+				tmp->ndims = 1;
+				tmp->ssup[0] = mss->ssup[j];
+
+				/* fill search key */
+				key.values = &groups[i].values[j];
+				key.isnull = &groups[i].isnull[j];
+
+				freq = (SortItem *) bsearch_arg(&key, freqs[j], nfreqs[j],
+												sizeof(SortItem),
+												multi_sort_compare, tmp);
+
+				item->base_frequency *= ((double) freq->count) / numrows;
+			}
+		}
+
+		pfree(nfreqs);
+		pfree(freqs);
+	}
+
+	pfree(items);
+	pfree(groups);
+
+	return mcvlist;
+}
+
+/*
+ * build_mss
+ *		Build a MultiSortSupport for the given StatsBuildData.
+ */
+static MultiSortSupport
+build_mss(StatsBuildData *data)
+{
+	int			i;
+	int			numattrs = data->nattnums;
+
+	/* Sort by multiple columns (using array of SortSupport) */
+	MultiSortSupport mss = multi_sort_init(numattrs);
+
+	/* prepare the sort functions for all the attributes */
+	for (i = 0; i < numattrs; i++)
+	{
+		VacAttrStats *colstat = data->stats[i];
+		TypeCacheEntry *type;
+
+		type = lookup_type_cache(colstat->attrtypid, TYPECACHE_LT_OPR);
+		if (type->lt_opr == InvalidOid) /* shouldn't happen */
+			elog(ERROR, "cache lookup failed for ordering operator for type %u",
+				 colstat->attrtypid);
+
+		multi_sort_add_dimension(mss, i, type->lt_opr, colstat->attrcollid);
+	}
+
+	return mss;
+}
+
+/*
+ * count_distinct_groups
+ *		Count distinct combinations of SortItems in the array.
+ *
+ * The array is assumed to be sorted according to the MultiSortSupport.
+ */
+static int
+count_distinct_groups(int numrows, SortItem *items, MultiSortSupport mss)
+{
+	int			i;
+	int			ndistinct;
+
+	ndistinct = 1;
+	for (i = 1; i < numrows; i++)
+	{
+		/* make sure the array really is sorted */
+		Assert(multi_sort_compare(&items[i], &items[i - 1], mss) >= 0);
+
+		if (multi_sort_compare(&items[i], &items[i - 1], mss) != 0)
+			ndistinct += 1;
+	}
+
+	return ndistinct;
+}
+
+/*
+ * compare_sort_item_count
+ *		Comparator for sorting items by count (frequencies) in descending
+ *		order.
+ */
+static int
+compare_sort_item_count(const void *a, const void *b, void *arg)
+{
+	SortItem   *ia = (SortItem *) a;
+	SortItem   *ib = (SortItem *) b;
+
+	if (ia->count == ib->count)
+		return 0;
+	else if (ia->count > ib->count)
+		return -1;
+
+	return 1;
+}
+
+/*
+ * build_distinct_groups
+ *		Build an array of SortItems for distinct groups and counts matching
+ *		items.
+ *
+ * The 'items' array is assumed to be sorted.
+ */
+static SortItem *
+build_distinct_groups(int numrows, SortItem *items, MultiSortSupport mss,
+					  int *ndistinct)
+{
+	int			i,
+				j;
+	int			ngroups = count_distinct_groups(numrows, items, mss);
+
+	SortItem   *groups = (SortItem *) palloc(ngroups * sizeof(SortItem));
+
+	j = 0;
+	groups[0] = items[0];
+	groups[0].count = 1;
+
+	for (i = 1; i < numrows; i++)
+	{
+		/* Assume sorted in ascending order. */
+		Assert(multi_sort_compare(&items[i], &items[i - 1], mss) >= 0);
+
+		/* New distinct group detected. */
+		if (multi_sort_compare(&items[i], &items[i - 1], mss) != 0)
+		{
+			groups[++j] = items[i];
+			groups[j].count = 0;
+		}
+
+		groups[j].count++;
+	}
+
+	/* ensure we filled the expected number of distinct groups */
+	Assert(j + 1 == ngroups);
+
+	/* Sort the distinct groups by frequency (in descending order). */
+	qsort_interruptible((void *) groups, ngroups, sizeof(SortItem),
+						compare_sort_item_count, NULL);
+
+	*ndistinct = ngroups;
+	return groups;
+}
+
+/* compare sort items (single dimension) */
+static int
+sort_item_compare(const void *a, const void *b, void *arg)
+{
+	SortSupport ssup = (SortSupport) arg;
+	SortItem   *ia = (SortItem *) a;
+	SortItem   *ib = (SortItem *) b;
+
+	return ApplySortComparator(ia->values[0], ia->isnull[0],
+							   ib->values[0], ib->isnull[0],
+							   ssup);
+}
+
+/*
+ * build_column_frequencies
+ *		Compute frequencies of values in each column.
+ *
+ * This returns an array of SortItems for each attribute the MCV is built
+ * on, with a frequency (number of occurrences) for each value. This is
+ * then used to compute "base" frequency of MCV items.
+ *
+ * All the memory is allocated in a single chunk, so that a single pfree
+ * is enough to release it. We do not allocate space for values/isnull
+ * arrays in the SortItems, because we can simply point into the input
+ * groups directly.
+ */
+static SortItem **
+build_column_frequencies(SortItem *groups, int ngroups,
+						 MultiSortSupport mss, int *ncounts)
+{
+	int			i,
+				dim;
+	SortItem  **result;
+	char	   *ptr;
+
+	Assert(groups);
+	Assert(ncounts);
+
+	/* allocate arrays for all columns as a single chunk */
+	ptr = palloc(MAXALIGN(sizeof(SortItem *) * mss->ndims) +
+				 mss->ndims * MAXALIGN(sizeof(SortItem) * ngroups));
+
+	/* initial array of pointers */
+	result = (SortItem **) ptr;
+	ptr += MAXALIGN(sizeof(SortItem *) * mss->ndims);
+
+	for (dim = 0; dim < mss->ndims; dim++)
+	{
+		SortSupport ssup = &mss->ssup[dim];
+
+		/* array of values for a single column */
+		result[dim] = (SortItem *) ptr;
+		ptr += MAXALIGN(sizeof(SortItem) * ngroups);
+
+		/* extract data for the dimension */
+		for (i = 0; i < ngroups; i++)
+		{
+			/* point into the input groups */
+			result[dim][i].values = &groups[i].values[dim];
+			result[dim][i].isnull = &groups[i].isnull[dim];
+			result[dim][i].count = groups[i].count;
+		}
+
+		/* sort the values, deduplicate */
+		qsort_interruptible((void *) result[dim], ngroups, sizeof(SortItem),
+							sort_item_compare, ssup);
+
+		/*
+		 * Identify distinct values, compute frequency (there might be
+		 * multiple MCV items containing this value, so we need to sum counts
+		 * from all of them.
+		 */
+		ncounts[dim] = 1;
+		for (i = 1; i < ngroups; i++)
+		{
+			if (sort_item_compare(&result[dim][i - 1], &result[dim][i], ssup) == 0)
+			{
+				result[dim][ncounts[dim] - 1].count += result[dim][i].count;
+				continue;
+			}
+
+			result[dim][ncounts[dim]] = result[dim][i];
+
+			ncounts[dim]++;
+		}
+	}
+
+	return result;
+}
+
+/*
+ * statext_mcv_load
+ *		Load the MCV list for the indicated pg_statistic_ext tuple.
+ */
+MCVList *
+statext_mcv_load(Oid mvoid)
+{
+	MCVList    *result;
+	bool		isnull;
+	Datum		mcvlist;
+	HeapTuple	htup = SearchSysCache1(STATEXTDATASTXOID, ObjectIdGetDatum(mvoid));
+
+	if (!HeapTupleIsValid(htup))
+		elog(ERROR, "cache lookup failed for statistics object %u", mvoid);
+
+	mcvlist = SysCacheGetAttr(STATEXTDATASTXOID, htup,
+							  Anum_pg_statistic_ext_data_stxdmcv, &isnull);
+
+	if (isnull)
+		elog(ERROR,
+			 "requested statistics kind \"%c\" is not yet built for statistics object %u",
+			 STATS_EXT_DEPENDENCIES, mvoid);
+
+	result = statext_mcv_deserialize(DatumGetByteaP(mcvlist));
+
+	ReleaseSysCache(htup);
+
+	return result;
+}
+
+
+/*
+ * statext_mcv_serialize
+ *		Serialize MCV list into a pg_mcv_list value.
+ *
+ * The MCV items may include values of various data types, and it's reasonable
+ * to expect redundancy (values for a given attribute, repeated for multiple
+ * MCV list items). So we deduplicate the values into arrays, and then replace
+ * the values by indexes into those arrays.
+ *
+ * The overall structure of the serialized representation looks like this:
+ *
+ * +---------------+----------------+---------------------+-------+
+ * | header fields | dimension info | deduplicated values | items |
+ * +---------------+----------------+---------------------+-------+
+ *
+ * Where dimension info stores information about the type of the K-th
+ * attribute (e.g. typlen, typbyval and length of deduplicated values).
+ * Deduplicated values store deduplicated values for each attribute.  And
+ * items store the actual MCV list items, with values replaced by indexes into
+ * the arrays.
+ *
+ * When serializing the items, we use uint16 indexes. The number of MCV items
+ * is limited by the statistics target (which is capped to 10k at the moment).
+ * We might increase this to 65k and still fit into uint16, so there's a bit of
+ * slack. Furthermore, this limit is on the number of distinct values per column,
+ * and we usually have few of those (and various combinations of them for the
+ * those MCV list). So uint16 seems fine for now.
+ *
+ * We don't really expect the serialization to save as much space as for
+ * histograms, as we are not doing any bucket splits (which is the source
+ * of high redundancy in histograms).
+ *
+ * TODO: Consider packing boolean flags (NULL) for each item into a single char
+ * (or a longer type) instead of using an array of bool items.
+ */
+bytea *
+statext_mcv_serialize(MCVList *mcvlist, VacAttrStats **stats)
+{
+	int			i;
+	int			dim;
+	int			ndims = mcvlist->ndimensions;
+
+	SortSupport ssup;
+	DimensionInfo *info;
+
+	Size		total_length;
+
+	/* serialized items (indexes into arrays, etc.) */
+	bytea	   *raw;
+	char	   *ptr;
+	char	   *endptr PG_USED_FOR_ASSERTS_ONLY;
+
+	/* values per dimension (and number of non-NULL values) */
+	Datum	  **values = (Datum **) palloc0(sizeof(Datum *) * ndims);
+	int		   *counts = (int *) palloc0(sizeof(int) * ndims);
+
+	/*
+	 * We'll include some rudimentary information about the attribute types
+	 * (length, by-val flag), so that we don't have to look them up while
+	 * deserializing the MCV list (we already have the type OID in the
+	 * header).  This is safe because when changing the type of the attribute
+	 * the statistics gets dropped automatically.  We need to store the info
+	 * about the arrays of deduplicated values anyway.
+	 */
+	info = (DimensionInfo *) palloc0(sizeof(DimensionInfo) * ndims);
+
+	/* sort support data for all attributes included in the MCV list */
+	ssup = (SortSupport) palloc0(sizeof(SortSupportData) * ndims);
+
+	/* collect and deduplicate values for each dimension (attribute) */
+	for (dim = 0; dim < ndims; dim++)
+	{
+		int			ndistinct;
+		TypeCacheEntry *typentry;
+
+		/*
+		 * Lookup the LT operator (can't get it from stats extra_data, as we
+		 * don't know how to interpret that - scalar vs. array etc.).
+		 */
+		typentry = lookup_type_cache(stats[dim]->attrtypid, TYPECACHE_LT_OPR);
+
+		/* copy important info about the data type (length, by-value) */
+		info[dim].typlen = stats[dim]->attrtype->typlen;
+		info[dim].typbyval = stats[dim]->attrtype->typbyval;
+
+		/* allocate space for values in the attribute and collect them */
+		values[dim] = (Datum *) palloc0(sizeof(Datum) * mcvlist->nitems);
+
+		for (i = 0; i < mcvlist->nitems; i++)
+		{
+			/* skip NULL values - we don't need to deduplicate those */
+			if (mcvlist->items[i].isnull[dim])
+				continue;
+
+			/* append the value at the end */
+			values[dim][counts[dim]] = mcvlist->items[i].values[dim];
+			counts[dim] += 1;
+		}
+
+		/* if there are just NULL values in this dimension, we're done */
+		if (counts[dim] == 0)
+			continue;
+
+		/* sort and deduplicate the data */
+		ssup[dim].ssup_cxt = CurrentMemoryContext;
+		ssup[dim].ssup_collation = stats[dim]->attrcollid;
+		ssup[dim].ssup_nulls_first = false;
+
+		PrepareSortSupportFromOrderingOp(typentry->lt_opr, &ssup[dim]);
+
+		qsort_interruptible(values[dim], counts[dim], sizeof(Datum),
+							compare_scalars_simple, &ssup[dim]);
+
+		/*
+		 * Walk through the array and eliminate duplicate values, but keep the
+		 * ordering (so that we can do a binary search later). We know there's
+		 * at least one item as (counts[dim] != 0), so we can skip the first
+		 * element.
+		 */
+		ndistinct = 1;			/* number of distinct values */
+		for (i = 1; i < counts[dim]; i++)
+		{
+			/* expect sorted array */
+			Assert(compare_datums_simple(values[dim][i - 1], values[dim][i], &ssup[dim]) <= 0);
+
+			/* if the value is the same as the previous one, we can skip it */
+			if (!compare_datums_simple(values[dim][i - 1], values[dim][i], &ssup[dim]))
+				continue;
+
+			values[dim][ndistinct] = values[dim][i];
+			ndistinct += 1;
+		}
+
+		/* we must not exceed PG_UINT16_MAX, as we use uint16 indexes */
+		Assert(ndistinct <= PG_UINT16_MAX);
+
+		/*
+		 * Store additional info about the attribute - number of deduplicated
+		 * values, and also size of the serialized data. For fixed-length data
+		 * types this is trivial to compute, for varwidth types we need to
+		 * actually walk the array and sum the sizes.
+		 */
+		info[dim].nvalues = ndistinct;
+
+		if (info[dim].typbyval) /* by-value data types */
+		{
+			info[dim].nbytes = info[dim].nvalues * info[dim].typlen;
+
+			/*
+			 * We copy the data into the MCV item during deserialization, so
+			 * we don't need to allocate any extra space.
+			 */
+			info[dim].nbytes_aligned = 0;
+		}
+		else if (info[dim].typlen > 0)	/* fixed-length by-ref */
+		{
+			/*
+			 * We don't care about alignment in the serialized data, so we
+			 * pack the data as much as possible. But we also track how much
+			 * data will be needed after deserialization, and in that case we
+			 * need to account for alignment of each item.
+			 *
+			 * Note: As the items are fixed-length, we could easily compute
+			 * this during deserialization, but we do it here anyway.
+			 */
+			info[dim].nbytes = info[dim].nvalues * info[dim].typlen;
+			info[dim].nbytes_aligned = info[dim].nvalues * MAXALIGN(info[dim].typlen);
+		}
+		else if (info[dim].typlen == -1)	/* varlena */
+		{
+			info[dim].nbytes = 0;
+			info[dim].nbytes_aligned = 0;
+			for (i = 0; i < info[dim].nvalues; i++)
+			{
+				Size		len;
+
+				/*
+				 * For varlena values, we detoast the values and store the
+				 * length and data separately. We don't bother with alignment
+				 * here, which means that during deserialization we need to
+				 * copy the fields and only access the copies.
+				 */
+				values[dim][i] = PointerGetDatum(PG_DETOAST_DATUM(values[dim][i]));
+
+				/* serialized length (uint32 length + data) */
+				len = VARSIZE_ANY_EXHDR(values[dim][i]);
+				info[dim].nbytes += sizeof(uint32); /* length */
+				info[dim].nbytes += len;	/* value (no header) */
+
+				/*
+				 * During deserialization we'll build regular varlena values
+				 * with full headers, and we need to align them properly.
+				 */
+				info[dim].nbytes_aligned += MAXALIGN(VARHDRSZ + len);
+			}
+		}
+		else if (info[dim].typlen == -2)	/* cstring */
+		{
+			info[dim].nbytes = 0;
+			info[dim].nbytes_aligned = 0;
+			for (i = 0; i < info[dim].nvalues; i++)
+			{
+				Size		len;
+
+				/*
+				 * cstring is handled similar to varlena - first we store the
+				 * length as uint32 and then the data. We don't care about
+				 * alignment, which means that during deserialization we need
+				 * to copy the fields and only access the copies.
+				 */
+
+				/* c-strings include terminator, so +1 byte */
+				len = strlen(DatumGetCString(values[dim][i])) + 1;
+				info[dim].nbytes += sizeof(uint32); /* length */
+				info[dim].nbytes += len;	/* value */
+
+				/* space needed for properly aligned deserialized copies */
+				info[dim].nbytes_aligned += MAXALIGN(len);
+			}
+		}
+
+		/* we know (count>0) so there must be some data */
+		Assert(info[dim].nbytes > 0);
+	}
+
+	/*
+	 * Now we can finally compute how much space we'll actually need for the
+	 * whole serialized MCV list (varlena header, MCV header, dimension info
+	 * for each attribute, deduplicated values and items).
+	 */
+	total_length = (3 * sizeof(uint32)) /* magic + type + nitems */
+		+ sizeof(AttrNumber)	/* ndimensions */
+		+ (ndims * sizeof(Oid));	/* attribute types */
+
+	/* dimension info */
+	total_length += ndims * sizeof(DimensionInfo);
+
+	/* add space for the arrays of deduplicated values */
+	for (i = 0; i < ndims; i++)
+		total_length += info[i].nbytes;
+
+	/*
+	 * And finally account for the items (those are fixed-length, thanks to
+	 * replacing values with uint16 indexes into the deduplicated arrays).
+	 */
+	total_length += mcvlist->nitems * ITEM_SIZE(dim);
+
+	/*
+	 * Allocate space for the whole serialized MCV list (we'll skip bytes, so
+	 * we set them to zero to make the result more compressible).
+	 */
+	raw = (bytea *) palloc0(VARHDRSZ + total_length);
+	SET_VARSIZE(raw, VARHDRSZ + total_length);
+
+	ptr = VARDATA(raw);
+	endptr = ptr + total_length;
+
+	/* copy the MCV list header fields, one by one */
+	memcpy(ptr, &mcvlist->magic, sizeof(uint32));
+	ptr += sizeof(uint32);
+
+	memcpy(ptr, &mcvlist->type, sizeof(uint32));
+	ptr += sizeof(uint32);
+
+	memcpy(ptr, &mcvlist->nitems, sizeof(uint32));
+	ptr += sizeof(uint32);
+
+	memcpy(ptr, &mcvlist->ndimensions, sizeof(AttrNumber));
+	ptr += sizeof(AttrNumber);
+
+	memcpy(ptr, mcvlist->types, sizeof(Oid) * ndims);
+	ptr += (sizeof(Oid) * ndims);
+
+	/* store information about the attributes (data amounts, ...) */
+	memcpy(ptr, info, sizeof(DimensionInfo) * ndims);
+	ptr += sizeof(DimensionInfo) * ndims;
+
+	/* Copy the deduplicated values for all attributes to the output. */
+	for (dim = 0; dim < ndims; dim++)
+	{
+		/* remember the starting point for Asserts later */
+		char	   *start PG_USED_FOR_ASSERTS_ONLY = ptr;
+
+		for (i = 0; i < info[dim].nvalues; i++)
+		{
+			Datum		value = values[dim][i];
+
+			if (info[dim].typbyval) /* passed by value */
+			{
+				Datum		tmp;
+
+				/*
+				 * For byval types, we need to copy just the significant bytes
+				 * - we can't use memcpy directly, as that assumes
+				 * little-endian behavior.  store_att_byval does almost what
+				 * we need, but it requires a properly aligned buffer - the
+				 * output buffer does not guarantee that. So we simply use a
+				 * local Datum variable (which guarantees proper alignment),
+				 * and then copy the value from it.
+				 */
+				store_att_byval(&tmp, value, info[dim].typlen);
+
+				memcpy(ptr, &tmp, info[dim].typlen);
+				ptr += info[dim].typlen;
+			}
+			else if (info[dim].typlen > 0)	/* passed by reference */
+			{
+				/* no special alignment needed, treated as char array */
+				memcpy(ptr, DatumGetPointer(value), info[dim].typlen);
+				ptr += info[dim].typlen;
+			}
+			else if (info[dim].typlen == -1)	/* varlena */
+			{
+				uint32		len = VARSIZE_ANY_EXHDR(DatumGetPointer(value));
+
+				/* copy the length */
+				memcpy(ptr, &len, sizeof(uint32));
+				ptr += sizeof(uint32);
+
+				/* data from the varlena value (without the header) */
+				memcpy(ptr, VARDATA_ANY(DatumGetPointer(value)), len);
+				ptr += len;
+			}
+			else if (info[dim].typlen == -2)	/* cstring */
+			{
+				uint32		len = (uint32) strlen(DatumGetCString(value)) + 1;
+
+				/* copy the length */
+				memcpy(ptr, &len, sizeof(uint32));
+				ptr += sizeof(uint32);
+
+				/* value */
+				memcpy(ptr, DatumGetCString(value), len);
+				ptr += len;
+			}
+
+			/* no underflows or overflows */
+			Assert((ptr > start) && ((ptr - start) <= info[dim].nbytes));
+		}
+
+		/* we should get exactly nbytes of data for this dimension */
+		Assert((ptr - start) == info[dim].nbytes);
+	}
+
+	/* Serialize the items, with uint16 indexes instead of the values. */
+	for (i = 0; i < mcvlist->nitems; i++)
+	{
+		MCVItem    *mcvitem = &mcvlist->items[i];
+
+		/* don't write beyond the allocated space */
+		Assert(ptr <= (endptr - ITEM_SIZE(dim)));
+
+		/* copy NULL and frequency flags into the serialized MCV */
+		memcpy(ptr, mcvitem->isnull, sizeof(bool) * ndims);
+		ptr += sizeof(bool) * ndims;
+
+		memcpy(ptr, &mcvitem->frequency, sizeof(double));
+		ptr += sizeof(double);
+
+		memcpy(ptr, &mcvitem->base_frequency, sizeof(double));
+		ptr += sizeof(double);
+
+		/* store the indexes last */
+		for (dim = 0; dim < ndims; dim++)
+		{
+			uint16		index = 0;
+			Datum	   *value;
+
+			/* do the lookup only for non-NULL values */
+			if (!mcvitem->isnull[dim])
+			{
+				value = (Datum *) bsearch_arg(&mcvitem->values[dim], values[dim],
+											  info[dim].nvalues, sizeof(Datum),
+											  compare_scalars_simple, &ssup[dim]);
+
+				Assert(value != NULL);	/* serialization or deduplication
+										 * error */
+
+				/* compute index within the deduplicated array */
+				index = (uint16) (value - values[dim]);
+
+				/* check the index is within expected bounds */
+				Assert(index < info[dim].nvalues);
+			}
+
+			/* copy the index into the serialized MCV */
+			memcpy(ptr, &index, sizeof(uint16));
+			ptr += sizeof(uint16);
+		}
+
+		/* make sure we don't overflow the allocated value */
+		Assert(ptr <= endptr);
+	}
+
+	/* at this point we expect to match the total_length exactly */
+	Assert(ptr == endptr);
+
+	pfree(values);
+	pfree(counts);
+
+	return raw;
+}
+
+/*
+ * statext_mcv_deserialize
+ *		Reads serialized MCV list into MCVList structure.
+ *
+ * All the memory needed by the MCV list is allocated as a single chunk, so
+ * it's possible to simply pfree() it at once.
+ */
+MCVList *
+statext_mcv_deserialize(bytea *data)
+{
+	int			dim,
+				i;
+	Size		expected_size;
+	MCVList    *mcvlist;
+	char	   *raw;
+	char	   *ptr;
+	char	   *endptr PG_USED_FOR_ASSERTS_ONLY;
+
+	int			ndims,
+				nitems;
+	DimensionInfo *info = NULL;
+
+	/* local allocation buffer (used only for deserialization) */
+	Datum	  **map = NULL;
+
+	/* MCV list */
+	Size		mcvlen;
+
+	/* buffer used for the result */
+	Size		datalen;
+	char	   *dataptr;
+	char	   *valuesptr;
+	char	   *isnullptr;
+
+	if (data == NULL)
+		return NULL;
+
+	/*
+	 * We can't possibly deserialize a MCV list if there's not even a complete
+	 * header. We need an explicit formula here, because we serialize the
+	 * header fields one by one, so we need to ignore struct alignment.
+	 */
+	if (VARSIZE_ANY(data) < MinSizeOfMCVList)
+		elog(ERROR, "invalid MCV size %zd (expected at least %zu)",
+			 VARSIZE_ANY(data), MinSizeOfMCVList);
+
+	/* read the MCV list header */
+	mcvlist = (MCVList *) palloc0(offsetof(MCVList, items));
+
+	/* pointer to the data part (skip the varlena header) */
+	raw = (char *) data;
+	ptr = VARDATA_ANY(raw);
+	endptr = (char *) raw + VARSIZE_ANY(data);
+
+	/* get the header and perform further sanity checks */
+	memcpy(&mcvlist->magic, ptr, sizeof(uint32));
+	ptr += sizeof(uint32);
+
+	memcpy(&mcvlist->type, ptr, sizeof(uint32));
+	ptr += sizeof(uint32);
+
+	memcpy(&mcvlist->nitems, ptr, sizeof(uint32));
+	ptr += sizeof(uint32);
+
+	memcpy(&mcvlist->ndimensions, ptr, sizeof(AttrNumber));
+	ptr += sizeof(AttrNumber);
+
+	if (mcvlist->magic != STATS_MCV_MAGIC)
+		elog(ERROR, "invalid MCV magic %u (expected %u)",
+			 mcvlist->magic, STATS_MCV_MAGIC);
+
+	if (mcvlist->type != STATS_MCV_TYPE_BASIC)
+		elog(ERROR, "invalid MCV type %u (expected %u)",
+			 mcvlist->type, STATS_MCV_TYPE_BASIC);
+
+	if (mcvlist->ndimensions == 0)
+		elog(ERROR, "invalid zero-length dimension array in MCVList");
+	else if ((mcvlist->ndimensions > STATS_MAX_DIMENSIONS) ||
+			 (mcvlist->ndimensions < 0))
+		elog(ERROR, "invalid length (%d) dimension array in MCVList",
+			 mcvlist->ndimensions);
+
+	if (mcvlist->nitems == 0)
+		elog(ERROR, "invalid zero-length item array in MCVList");
+	else if (mcvlist->nitems > STATS_MCVLIST_MAX_ITEMS)
+		elog(ERROR, "invalid length (%u) item array in MCVList",
+			 mcvlist->nitems);
+
+	nitems = mcvlist->nitems;
+	ndims = mcvlist->ndimensions;
+
+	/*
+	 * Check amount of data including DimensionInfo for all dimensions and
+	 * also the serialized items (including uint16 indexes). Also, walk
+	 * through the dimension information and add it to the sum.
+	 */
+	expected_size = SizeOfMCVList(ndims, nitems);
+
+	/*
+	 * Check that we have at least the dimension and info records, along with
+	 * the items. We don't know the size of the serialized values yet. We need
+	 * to do this check first, before accessing the dimension info.
+	 */
+	if (VARSIZE_ANY(data) < expected_size)
+		elog(ERROR, "invalid MCV size %zd (expected %zu)",
+			 VARSIZE_ANY(data), expected_size);
+
+	/* Now copy the array of type Oids. */
+	memcpy(mcvlist->types, ptr, sizeof(Oid) * ndims);
+	ptr += (sizeof(Oid) * ndims);
+
+	/* Now it's safe to access the dimension info. */
+	info = palloc(ndims * sizeof(DimensionInfo));
+
+	memcpy(info, ptr, ndims * sizeof(DimensionInfo));
+	ptr += (ndims * sizeof(DimensionInfo));
+
+	/* account for the value arrays */
+	for (dim = 0; dim < ndims; dim++)
+	{
+		/*
+		 * XXX I wonder if we can/should rely on asserts here. Maybe those
+		 * checks should be done every time?
+		 */
+		Assert(info[dim].nvalues >= 0);
+		Assert(info[dim].nbytes >= 0);
+
+		expected_size += info[dim].nbytes;
+	}
+
+	/*
+	 * Now we know the total expected MCV size, including all the pieces
+	 * (header, dimension info. items and deduplicated data). So do the final
+	 * check on size.
+	 */
+	if (VARSIZE_ANY(data) != expected_size)
+		elog(ERROR, "invalid MCV size %zd (expected %zu)",
+			 VARSIZE_ANY(data), expected_size);
+
+	/*
+	 * We need an array of Datum values for each dimension, so that we can
+	 * easily translate the uint16 indexes later. We also need a top-level
+	 * array of pointers to those per-dimension arrays.
+	 *
+	 * While allocating the arrays for dimensions, compute how much space we
+	 * need for a copy of the by-ref data, as we can't simply point to the
+	 * original values (it might go away).
+	 */
+	datalen = 0;				/* space for by-ref data */
+	map = (Datum **) palloc(ndims * sizeof(Datum *));
+
+	for (dim = 0; dim < ndims; dim++)
+	{
+		map[dim] = (Datum *) palloc(sizeof(Datum) * info[dim].nvalues);
+
+		/* space needed for a copy of data for by-ref types */
+		datalen += info[dim].nbytes_aligned;
+	}
+
+	/*
+	 * Now resize the MCV list so that the allocation includes all the data.
+	 *
+	 * Allocate space for a copy of the data, as we can't simply reference the
+	 * serialized data - it's not aligned properly, and it may disappear while
+	 * we're still using the MCV list, e.g. due to catcache release.
+	 *
+	 * We do care about alignment here, because we will allocate all the
+	 * pieces at once, but then use pointers to different parts.
+	 */
+	mcvlen = MAXALIGN(offsetof(MCVList, items) + (sizeof(MCVItem) * nitems));
+
+	/* arrays of values and isnull flags for all MCV items */
+	mcvlen += nitems * MAXALIGN(sizeof(Datum) * ndims);
+	mcvlen += nitems * MAXALIGN(sizeof(bool) * ndims);
+
+	/* we don't quite need to align this, but it makes some asserts easier */
+	mcvlen += MAXALIGN(datalen);
+
+	/* now resize the deserialized MCV list, and compute pointers to parts */
+	mcvlist = repalloc(mcvlist, mcvlen);
+
+	/* pointer to the beginning of values/isnull arrays */
+	valuesptr = (char *) mcvlist
+		+ MAXALIGN(offsetof(MCVList, items) + (sizeof(MCVItem) * nitems));
+
+	isnullptr = valuesptr + (nitems * MAXALIGN(sizeof(Datum) * ndims));
+
+	dataptr = isnullptr + (nitems * MAXALIGN(sizeof(bool) * ndims));
+
+	/*
+	 * Build mapping (index => value) for translating the serialized data into
+	 * the in-memory representation.
+	 */
+	for (dim = 0; dim < ndims; dim++)
+	{
+		/* remember start position in the input array */
+		char	   *start PG_USED_FOR_ASSERTS_ONLY = ptr;
+
+		if (info[dim].typbyval)
+		{
+			/* for by-val types we simply copy data into the mapping */
+			for (i = 0; i < info[dim].nvalues; i++)
+			{
+				Datum		v = 0;
+
+				memcpy(&v, ptr, info[dim].typlen);
+				ptr += info[dim].typlen;
+
+				map[dim][i] = fetch_att(&v, true, info[dim].typlen);
+
+				/* no under/overflow of input array */
+				Assert(ptr <= (start + info[dim].nbytes));
+			}
+		}
+		else
+		{
+			/* for by-ref types we need to also make a copy of the data */
+
+			/* passed by reference, but fixed length (name, tid, ...) */
+			if (info[dim].typlen > 0)
+			{
+				for (i = 0; i < info[dim].nvalues; i++)
+				{
+					memcpy(dataptr, ptr, info[dim].typlen);
+					ptr += info[dim].typlen;
+
+					/* just point into the array */
+					map[dim][i] = PointerGetDatum(dataptr);
+					dataptr += MAXALIGN(info[dim].typlen);
+				}
+			}
+			else if (info[dim].typlen == -1)
+			{
+				/* varlena */
+				for (i = 0; i < info[dim].nvalues; i++)
+				{
+					uint32		len;
+
+					/* read the uint32 length */
+					memcpy(&len, ptr, sizeof(uint32));
+					ptr += sizeof(uint32);
+
+					/* the length is data-only */
+					SET_VARSIZE(dataptr, len + VARHDRSZ);
+					memcpy(VARDATA(dataptr), ptr, len);
+					ptr += len;
+
+					/* just point into the array */
+					map[dim][i] = PointerGetDatum(dataptr);
+
+					/* skip to place of the next deserialized value */
+					dataptr += MAXALIGN(len + VARHDRSZ);
+				}
+			}
+			else if (info[dim].typlen == -2)
+			{
+				/* cstring */
+				for (i = 0; i < info[dim].nvalues; i++)
+				{
+					uint32		len;
+
+					memcpy(&len, ptr, sizeof(uint32));
+					ptr += sizeof(uint32);
+
+					memcpy(dataptr, ptr, len);
+					ptr += len;
+
+					/* just point into the array */
+					map[dim][i] = PointerGetDatum(dataptr);
+					dataptr += MAXALIGN(len);
+				}
+			}
+
+			/* no under/overflow of input array */
+			Assert(ptr <= (start + info[dim].nbytes));
+
+			/* no overflow of the output mcv value */
+			Assert(dataptr <= ((char *) mcvlist + mcvlen));
+		}
+
+		/* check we consumed input data for this dimension exactly */
+		Assert(ptr == (start + info[dim].nbytes));
+	}
+
+	/* we should have also filled the MCV list exactly */
+	Assert(dataptr == ((char *) mcvlist + mcvlen));
+
+	/* deserialize the MCV items and translate the indexes to Datums */
+	for (i = 0; i < nitems; i++)
+	{
+		MCVItem    *item = &mcvlist->items[i];
+
+		item->values = (Datum *) valuesptr;
+		valuesptr += MAXALIGN(sizeof(Datum) * ndims);
+
+		item->isnull = (bool *) isnullptr;
+		isnullptr += MAXALIGN(sizeof(bool) * ndims);
+
+		memcpy(item->isnull, ptr, sizeof(bool) * ndims);
+		ptr += sizeof(bool) * ndims;
+
+		memcpy(&item->frequency, ptr, sizeof(double));
+		ptr += sizeof(double);
+
+		memcpy(&item->base_frequency, ptr, sizeof(double));
+		ptr += sizeof(double);
+
+		/* finally translate the indexes (for non-NULL only) */
+		for (dim = 0; dim < ndims; dim++)
+		{
+			uint16		index;
+
+			memcpy(&index, ptr, sizeof(uint16));
+			ptr += sizeof(uint16);
+
+			if (item->isnull[dim])
+				continue;
+
+			item->values[dim] = map[dim][index];
+		}
+
+		/* check we're not overflowing the input */
+		Assert(ptr <= endptr);
+	}
+
+	/* check that we processed all the data */
+	Assert(ptr == endptr);
+
+	/* release the buffers used for mapping */
+	for (dim = 0; dim < ndims; dim++)
+		pfree(map[dim]);
+
+	pfree(map);
+
+	return mcvlist;
+}
+
+/*
+ * SRF with details about buckets of a histogram:
+ *
+ * - item ID (0...nitems)
+ * - values (string array)
+ * - nulls only (boolean array)
+ * - frequency (double precision)
+ * - base_frequency (double precision)
+ *
+ * The input is the OID of the statistics, and there are no rows returned if
+ * the statistics contains no histogram.
+ */
+Datum
+pg_stats_ext_mcvlist_items(PG_FUNCTION_ARGS)
+{
+	FuncCallContext *funcctx;
+
+	/* stuff done only on the first call of the function */
+	if (SRF_IS_FIRSTCALL())
+	{
+		MemoryContext oldcontext;
+		MCVList    *mcvlist;
+		TupleDesc	tupdesc;
+
+		/* create a function context for cross-call persistence */
+		funcctx = SRF_FIRSTCALL_INIT();
+
+		/* switch to memory context appropriate for multiple function calls */
+		oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
+
+		mcvlist = statext_mcv_deserialize(PG_GETARG_BYTEA_P(0));
+
+		funcctx->user_fctx = mcvlist;
+
+		/* total number of tuples to be returned */
+		funcctx->max_calls = 0;
+		if (funcctx->user_fctx != NULL)
+			funcctx->max_calls = mcvlist->nitems;
+
+		/* Build a tuple descriptor for our result type */
+		if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
+			ereport(ERROR,
+					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+					 errmsg("function returning record called in context "
+							"that cannot accept type record")));
+		tupdesc = BlessTupleDesc(tupdesc);
+
+		/*
+		 * generate attribute metadata needed later to produce tuples from raw
+		 * C strings
+		 */
+		funcctx->attinmeta = TupleDescGetAttInMetadata(tupdesc);
+
+		MemoryContextSwitchTo(oldcontext);
+	}
+
+	/* stuff done on every call of the function */
+	funcctx = SRF_PERCALL_SETUP();
+
+	if (funcctx->call_cntr < funcctx->max_calls)	/* do when there is more
+													 * left to send */
+	{
+		Datum		values[5];
+		bool		nulls[5];
+		HeapTuple	tuple;
+		Datum		result;
+		ArrayBuildState *astate_values = NULL;
+		ArrayBuildState *astate_nulls = NULL;
+
+		int			i;
+		MCVList    *mcvlist;
+		MCVItem    *item;
+
+		mcvlist = (MCVList *) funcctx->user_fctx;
+
+		Assert(funcctx->call_cntr < mcvlist->nitems);
+
+		item = &mcvlist->items[funcctx->call_cntr];
+
+		for (i = 0; i < mcvlist->ndimensions; i++)
+		{
+
+			astate_nulls = accumArrayResult(astate_nulls,
+											BoolGetDatum(item->isnull[i]),
+											false,
+											BOOLOID,
+											CurrentMemoryContext);
+
+			if (!item->isnull[i])
+			{
+				bool		isvarlena;
+				Oid			outfunc;
+				FmgrInfo	fmgrinfo;
+				Datum		val;
+				text	   *txt;
+
+				/* lookup output func for the type */
+				getTypeOutputInfo(mcvlist->types[i], &outfunc, &isvarlena);
+				fmgr_info(outfunc, &fmgrinfo);
+
+				val = FunctionCall1(&fmgrinfo, item->values[i]);
+				txt = cstring_to_text(DatumGetPointer(val));
+
+				astate_values = accumArrayResult(astate_values,
+												 PointerGetDatum(txt),
+												 false,
+												 TEXTOID,
+												 CurrentMemoryContext);
+			}
+			else
+				astate_values = accumArrayResult(astate_values,
+												 (Datum) 0,
+												 true,
+												 TEXTOID,
+												 CurrentMemoryContext);
+		}
+
+		values[0] = Int32GetDatum(funcctx->call_cntr);
+		values[1] = PointerGetDatum(makeArrayResult(astate_values, CurrentMemoryContext));
+		values[2] = PointerGetDatum(makeArrayResult(astate_nulls, CurrentMemoryContext));
+		values[3] = Float8GetDatum(item->frequency);
+		values[4] = Float8GetDatum(item->base_frequency);
+
+		/* no NULLs in the tuple */
+		memset(nulls, 0, sizeof(nulls));
+
+		/* build a tuple */
+		tuple = heap_form_tuple(funcctx->attinmeta->tupdesc, values, nulls);
+
+		/* make the tuple into a datum */
+		result = HeapTupleGetDatum(tuple);
+
+		SRF_RETURN_NEXT(funcctx, result);
+	}
+	else						/* do when there is no more left */
+	{
+		SRF_RETURN_DONE(funcctx);
+	}
+}
+
+/*
+ * pg_mcv_list_in		- input routine for type pg_mcv_list.
+ *
+ * pg_mcv_list is real enough to be a table column, but it has no operations
+ * of its own, and disallows input too
+ */
+Datum
+pg_mcv_list_in(PG_FUNCTION_ARGS)
+{
+	/*
+	 * pg_mcv_list stores the data in binary form and parsing text input is
+	 * not needed, so disallow this.
+	 */
+	ereport(ERROR,
+			(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+			 errmsg("cannot accept a value of type %s", "pg_mcv_list")));
+
+	PG_RETURN_VOID();			/* keep compiler quiet */
+}
+
+
+/*
+ * pg_mcv_list_out		- output routine for type pg_mcv_list.
+ *
+ * MCV lists are serialized into a bytea value, so we simply call byteaout()
+ * to serialize the value into text. But it'd be nice to serialize that into
+ * a meaningful representation (e.g. for inspection by people).
+ *
+ * XXX This should probably return something meaningful, similar to what
+ * pg_dependencies_out does. Not sure how to deal with the deduplicated
+ * values, though - do we want to expand that or not?
+ */
+Datum
+pg_mcv_list_out(PG_FUNCTION_ARGS)
+{
+	return byteaout(fcinfo);
+}
+
+/*
+ * pg_mcv_list_recv		- binary input routine for type pg_mcv_list.
+ */
+Datum
+pg_mcv_list_recv(PG_FUNCTION_ARGS)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+			 errmsg("cannot accept a value of type %s", "pg_mcv_list")));
+
+	PG_RETURN_VOID();			/* keep compiler quiet */
+}
+
+/*
+ * pg_mcv_list_send		- binary output routine for type pg_mcv_list.
+ *
+ * MCV lists are serialized in a bytea value (although the type is named
+ * differently), so let's just send that.
+ */
+Datum
+pg_mcv_list_send(PG_FUNCTION_ARGS)
+{
+	return byteasend(fcinfo);
+}
+
+/*
+ * match the attribute/expression to a dimension of the statistic
+ *
+ * Returns the zero-based index of the matching statistics dimension.
+ * Optionally determines the collation.
+ */
+static int
+mcv_match_expression(Node *expr, Bitmapset *keys, List *exprs, Oid *collid)
+{
+	int			idx;
+
+	if (IsA(expr, Var))
+	{
+		/* simple Var, so just lookup using varattno */
+		Var		   *var = (Var *) expr;
+
+		if (collid)
+			*collid = var->varcollid;
+
+		idx = bms_member_index(keys, var->varattno);
+
+		if (idx < 0)
+			elog(ERROR, "variable not found in statistics object");
+	}
+	else
+	{
+		/* expression - lookup in stats expressions */
+		ListCell   *lc;
+
+		if (collid)
+			*collid = exprCollation(expr);
+
+		/* expressions are stored after the simple columns */
+		idx = bms_num_members(keys);
+		foreach(lc, exprs)
+		{
+			Node	   *stat_expr = (Node *) lfirst(lc);
+
+			if (equal(expr, stat_expr))
+				break;
+
+			idx++;
+		}
+
+		if (lc == NULL)
+			elog(ERROR, "expression not found in statistics object");
+	}
+
+	return idx;
+}
+
+/*
+ * mcv_get_match_bitmap
+ *	Evaluate clauses using the MCV list, and update the match bitmap.
+ *
+ * A match bitmap keeps match/mismatch status for each MCV item, and we
+ * update it based on additional clauses. We also use it to skip items
+ * that can't possibly match (e.g. item marked as "mismatch" can't change
+ * to "match" when evaluating AND clause list).
+ *
+ * The function also returns a flag indicating whether there was an
+ * equality condition for all attributes, the minimum frequency in the MCV
+ * list, and a total MCV frequency (sum of frequencies for all items).
+ *
+ * XXX Currently the match bitmap uses a bool for each MCV item, which is
+ * somewhat wasteful as we could do with just a single bit, thus reducing
+ * the size to ~1/8. It would also allow us to combine bitmaps simply using
+ * & and |, which should be faster than min/max. The bitmaps are fairly
+ * small, though (thanks to the cap on the MCV list size).
+ */
+static bool *
+mcv_get_match_bitmap(PlannerInfo *root, List *clauses,
+					 Bitmapset *keys, List *exprs,
+					 MCVList *mcvlist, bool is_or)
+{
+	int			i;
+	ListCell   *l;
+	bool	   *matches;
+
+	/* The bitmap may be partially built. */
+	Assert(clauses != NIL);
+	Assert(list_length(clauses) >= 1);
+	Assert(mcvlist != NULL);
+	Assert(mcvlist->nitems > 0);
+	Assert(mcvlist->nitems <= STATS_MCVLIST_MAX_ITEMS);
+
+	matches = palloc(sizeof(bool) * mcvlist->nitems);
+	memset(matches, (is_or) ? false : true,
+		   sizeof(bool) * mcvlist->nitems);
+
+	/*
+	 * Loop through the list of clauses, and for each of them evaluate all the
+	 * MCV items not yet eliminated by the preceding clauses.
+	 */
+	foreach(l, clauses)
+	{
+		Node	   *clause = (Node *) lfirst(l);
+
+		/* if it's a RestrictInfo, then extract the clause */
+		if (IsA(clause, RestrictInfo))
+			clause = (Node *) ((RestrictInfo *) clause)->clause;
+
+		/*
+		 * Handle the various types of clauses - OpClause, NullTest and
+		 * AND/OR/NOT
+		 */
+		if (is_opclause(clause))
+		{
+			OpExpr	   *expr = (OpExpr *) clause;
+			FmgrInfo	opproc;
+
+			/* valid only after examine_opclause_args returns true */
+			Node	   *clause_expr;
+			Const	   *cst;
+			bool		expronleft;
+			int			idx;
+			Oid			collid;
+
+			fmgr_info(get_opcode(expr->opno), &opproc);
+
+			/* extract the var/expr and const from the expression */
+			if (!examine_opclause_args(expr->args, &clause_expr, &cst, &expronleft))
+				elog(ERROR, "incompatible clause");
+
+			/* match the attribute/expression to a dimension of the statistic */
+			idx = mcv_match_expression(clause_expr, keys, exprs, &collid);
+
+			/*
+			 * Walk through the MCV items and evaluate the current clause. We
+			 * can skip items that were already ruled out, and terminate if
+			 * there are no remaining MCV items that might possibly match.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+			{
+				bool		match = true;
+				MCVItem    *item = &mcvlist->items[i];
+
+				Assert(idx >= 0);
+
+				/*
+				 * When the MCV item or the Const value is NULL we can treat
+				 * this as a mismatch. We must not call the operator because
+				 * of strictness.
+				 */
+				if (item->isnull[idx] || cst->constisnull)
+				{
+					matches[i] = RESULT_MERGE(matches[i], is_or, false);
+					continue;
+				}
+
+				/*
+				 * Skip MCV items that can't change result in the bitmap. Once
+				 * the value gets false for AND-lists, or true for OR-lists,
+				 * we don't need to look at more clauses.
+				 */
+				if (RESULT_IS_FINAL(matches[i], is_or))
+					continue;
+
+				/*
+				 * First check whether the constant is below the lower
+				 * boundary (in that case we can skip the bucket, because
+				 * there's no overlap).
+				 *
+				 * We don't store collations used to build the statistics, but
+				 * we can use the collation for the attribute itself, as
+				 * stored in varcollid. We do reset the statistics after a
+				 * type change (including collation change), so this is OK.
+				 * For expressions, we use the collation extracted from the
+				 * expression itself.
+				 */
+				if (expronleft)
+					match = DatumGetBool(FunctionCall2Coll(&opproc,
+														   collid,
+														   item->values[idx],
+														   cst->constvalue));
+				else
+					match = DatumGetBool(FunctionCall2Coll(&opproc,
+														   collid,
+														   cst->constvalue,
+														   item->values[idx]));
+
+				/* update the match bitmap with the result */
+				matches[i] = RESULT_MERGE(matches[i], is_or, match);
+			}
+		}
+		else if (IsA(clause, ScalarArrayOpExpr))
+		{
+			ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) clause;
+			FmgrInfo	opproc;
+
+			/* valid only after examine_opclause_args returns true */
+			Node	   *clause_expr;
+			Const	   *cst;
+			bool		expronleft;
+			Oid			collid;
+			int			idx;
+
+			/* array evaluation */
+			ArrayType  *arrayval;
+			int16		elmlen;
+			bool		elmbyval;
+			char		elmalign;
+			int			num_elems;
+			Datum	   *elem_values;
+			bool	   *elem_nulls;
+
+			fmgr_info(get_opcode(expr->opno), &opproc);
+
+			/* extract the var/expr and const from the expression */
+			if (!examine_opclause_args(expr->args, &clause_expr, &cst, &expronleft))
+				elog(ERROR, "incompatible clause");
+
+			/* We expect Var on left */
+			if (!expronleft)
+				elog(ERROR, "incompatible clause");
+
+			/*
+			 * Deconstruct the array constant, unless it's NULL (we'll cover
+			 * that case below)
+			 */
+			if (!cst->constisnull)
+			{
+				arrayval = DatumGetArrayTypeP(cst->constvalue);
+				get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
+									 &elmlen, &elmbyval, &elmalign);
+				deconstruct_array(arrayval,
+								  ARR_ELEMTYPE(arrayval),
+								  elmlen, elmbyval, elmalign,
+								  &elem_values, &elem_nulls, &num_elems);
+			}
+
+			/* match the attribute/expression to a dimension of the statistic */
+			idx = mcv_match_expression(clause_expr, keys, exprs, &collid);
+
+			/*
+			 * Walk through the MCV items and evaluate the current clause. We
+			 * can skip items that were already ruled out, and terminate if
+			 * there are no remaining MCV items that might possibly match.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+			{
+				int			j;
+				bool		match = (expr->useOr ? false : true);
+				MCVItem    *item = &mcvlist->items[i];
+
+				/*
+				 * When the MCV item or the Const value is NULL we can treat
+				 * this as a mismatch. We must not call the operator because
+				 * of strictness.
+				 */
+				if (item->isnull[idx] || cst->constisnull)
+				{
+					matches[i] = RESULT_MERGE(matches[i], is_or, false);
+					continue;
+				}
+
+				/*
+				 * Skip MCV items that can't change result in the bitmap. Once
+				 * the value gets false for AND-lists, or true for OR-lists,
+				 * we don't need to look at more clauses.
+				 */
+				if (RESULT_IS_FINAL(matches[i], is_or))
+					continue;
+
+				for (j = 0; j < num_elems; j++)
+				{
+					Datum		elem_value = elem_values[j];
+					bool		elem_isnull = elem_nulls[j];
+					bool		elem_match;
+
+					/* NULL values always evaluate as not matching. */
+					if (elem_isnull)
+					{
+						match = RESULT_MERGE(match, expr->useOr, false);
+						continue;
+					}
+
+					/*
+					 * Stop evaluating the array elements once we reach a
+					 * matching value that can't change - ALL() is the same as
+					 * AND-list, ANY() is the same as OR-list.
+					 */
+					if (RESULT_IS_FINAL(match, expr->useOr))
+						break;
+
+					elem_match = DatumGetBool(FunctionCall2Coll(&opproc,
+																collid,
+																item->values[idx],
+																elem_value));
+
+					match = RESULT_MERGE(match, expr->useOr, elem_match);
+				}
+
+				/* update the match bitmap with the result */
+				matches[i] = RESULT_MERGE(matches[i], is_or, match);
+			}
+		}
+		else if (IsA(clause, NullTest))
+		{
+			NullTest   *expr = (NullTest *) clause;
+			Node	   *clause_expr = (Node *) (expr->arg);
+
+			/* match the attribute/expression to a dimension of the statistic */
+			int			idx = mcv_match_expression(clause_expr, keys, exprs, NULL);
+
+			/*
+			 * Walk through the MCV items and evaluate the current clause. We
+			 * can skip items that were already ruled out, and terminate if
+			 * there are no remaining MCV items that might possibly match.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+			{
+				bool		match = false;	/* assume mismatch */
+				MCVItem    *item = &mcvlist->items[i];
+
+				/* if the clause mismatches the MCV item, update the bitmap */
+				switch (expr->nulltesttype)
+				{
+					case IS_NULL:
+						match = (item->isnull[idx]) ? true : match;
+						break;
+
+					case IS_NOT_NULL:
+						match = (!item->isnull[idx]) ? true : match;
+						break;
+				}
+
+				/* now, update the match bitmap, depending on OR/AND type */
+				matches[i] = RESULT_MERGE(matches[i], is_or, match);
+			}
+		}
+		else if (is_orclause(clause) || is_andclause(clause))
+		{
+			/* AND/OR clause, with all subclauses being compatible */
+
+			int			i;
+			BoolExpr   *bool_clause = ((BoolExpr *) clause);
+			List	   *bool_clauses = bool_clause->args;
+
+			/* match/mismatch bitmap for each MCV item */
+			bool	   *bool_matches = NULL;
+
+			Assert(bool_clauses != NIL);
+			Assert(list_length(bool_clauses) >= 2);
+
+			/* build the match bitmap for the OR-clauses */
+			bool_matches = mcv_get_match_bitmap(root, bool_clauses, keys, exprs,
+												mcvlist, is_orclause(clause));
+
+			/*
+			 * Merge the bitmap produced by mcv_get_match_bitmap into the
+			 * current one. We need to consider if we're evaluating AND or OR
+			 * condition when merging the results.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+				matches[i] = RESULT_MERGE(matches[i], is_or, bool_matches[i]);
+
+			pfree(bool_matches);
+		}
+		else if (is_notclause(clause))
+		{
+			/* NOT clause, with all subclauses compatible */
+
+			int			i;
+			BoolExpr   *not_clause = ((BoolExpr *) clause);
+			List	   *not_args = not_clause->args;
+
+			/* match/mismatch bitmap for each MCV item */
+			bool	   *not_matches = NULL;
+
+			Assert(not_args != NIL);
+			Assert(list_length(not_args) == 1);
+
+			/* build the match bitmap for the NOT-clause */
+			not_matches = mcv_get_match_bitmap(root, not_args, keys, exprs,
+											   mcvlist, false);
+
+			/*
+			 * Merge the bitmap produced by mcv_get_match_bitmap into the
+			 * current one. We're handling a NOT clause, so invert the result
+			 * before merging it into the global bitmap.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+				matches[i] = RESULT_MERGE(matches[i], is_or, !not_matches[i]);
+
+			pfree(not_matches);
+		}
+		else if (IsA(clause, Var))
+		{
+			/* Var (has to be a boolean Var, possibly from below NOT) */
+
+			Var		   *var = (Var *) (clause);
+
+			/* match the attribute to a dimension of the statistic */
+			int			idx = bms_member_index(keys, var->varattno);
+
+			Assert(var->vartype == BOOLOID);
+
+			/*
+			 * Walk through the MCV items and evaluate the current clause. We
+			 * can skip items that were already ruled out, and terminate if
+			 * there are no remaining MCV items that might possibly match.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+			{
+				MCVItem    *item = &mcvlist->items[i];
+				bool		match = false;
+
+				/* if the item is NULL, it's a mismatch */
+				if (!item->isnull[idx] && DatumGetBool(item->values[idx]))
+					match = true;
+
+				/* update the result bitmap */
+				matches[i] = RESULT_MERGE(matches[i], is_or, match);
+			}
+		}
+		else
+		{
+			/* Otherwise, it must be a bare boolean-returning expression */
+			int			idx;
+
+			/* match the expression to a dimension of the statistic */
+			idx = mcv_match_expression(clause, keys, exprs, NULL);
+
+			/*
+			 * Walk through the MCV items and evaluate the current clause. We
+			 * can skip items that were already ruled out, and terminate if
+			 * there are no remaining MCV items that might possibly match.
+			 */
+			for (i = 0; i < mcvlist->nitems; i++)
+			{
+				bool		match;
+				MCVItem    *item = &mcvlist->items[i];
+
+				/* "match" just means it's bool TRUE */
+				match = !item->isnull[idx] && DatumGetBool(item->values[idx]);
+
+				/* now, update the match bitmap, depending on OR/AND type */
+				matches[i] = RESULT_MERGE(matches[i], is_or, match);
+			}
+		}
+	}
+
+	return matches;
+}
+
+
+/*
+ * mcv_combine_selectivities
+ * 		Combine per-column and multi-column MCV selectivity estimates.
+ *
+ * simple_sel is a "simple" selectivity estimate (produced without using any
+ * extended statistics, essentially assuming independence of columns/clauses).
+ *
+ * mcv_sel and mcv_basesel are sums of the frequencies and base frequencies of
+ * all matching MCV items.  The difference (mcv_sel - mcv_basesel) is then
+ * essentially interpreted as a correction to be added to simple_sel, as
+ * described below.
+ *
+ * mcv_totalsel is the sum of the frequencies of all MCV items (not just the
+ * matching ones).  This is used as an upper bound on the portion of the
+ * selectivity estimates not covered by the MCV statistics.
+ *
+ * Note: While simple and base selectivities are defined in a quite similar
+ * way, the values are computed differently and are not therefore equal. The
+ * simple selectivity is computed as a product of per-clause estimates, while
+ * the base selectivity is computed by adding up base frequencies of matching
+ * items of the multi-column MCV list. So the values may differ for two main
+ * reasons - (a) the MCV list may not cover 100% of the data and (b) some of
+ * the MCV items did not match the estimated clauses.
+ *
+ * As both (a) and (b) reduce the base selectivity value, it generally holds
+ * that (simple_sel >= mcv_basesel). If the MCV list covers all the data, the
+ * values may be equal.
+ *
+ * So, other_sel = (simple_sel - mcv_basesel) is an estimate for the part not
+ * covered by the MCV list, and (mcv_sel - mcv_basesel) may be seen as a
+ * correction for the part covered by the MCV list. Those two statements are
+ * actually equivalent.
+ */
+Selectivity
+mcv_combine_selectivities(Selectivity simple_sel,
+						  Selectivity mcv_sel,
+						  Selectivity mcv_basesel,
+						  Selectivity mcv_totalsel)
+{
+	Selectivity other_sel;
+	Selectivity sel;
+
+	/* estimated selectivity of values not covered by MCV matches */
+	other_sel = simple_sel - mcv_basesel;
+	CLAMP_PROBABILITY(other_sel);
+
+	/* this non-MCV selectivity cannot exceed 1 - mcv_totalsel */
+	if (other_sel > 1.0 - mcv_totalsel)
+		other_sel = 1.0 - mcv_totalsel;
+
+	/* overall selectivity is the sum of the MCV and non-MCV parts */
+	sel = mcv_sel + other_sel;
+	CLAMP_PROBABILITY(sel);
+
+	return sel;
+}
+
+
+/*
+ * mcv_clauselist_selectivity
+ *		Use MCV statistics to estimate the selectivity of an implicitly-ANDed
+ *		list of clauses.
+ *
+ * This determines which MCV items match every clause in the list and returns
+ * the sum of the frequencies of those items.
+ *
+ * In addition, it returns the sum of the base frequencies of each of those
+ * items (that is the sum of the selectivities that each item would have if
+ * the columns were independent of one another), and the total selectivity of
+ * all the MCV items (not just the matching ones).  These are expected to be
+ * used together with a "simple" selectivity estimate (one based only on
+ * per-column statistics) to produce an overall selectivity estimate that
+ * makes use of both per-column and multi-column statistics --- see
+ * mcv_combine_selectivities().
+ */
+Selectivity
+mcv_clauselist_selectivity(PlannerInfo *root, StatisticExtInfo *stat,
+						   List *clauses, int varRelid,
+						   JoinType jointype, SpecialJoinInfo *sjinfo,
+						   RelOptInfo *rel,
+						   Selectivity *basesel, Selectivity *totalsel)
+{
+	int			i;
+	MCVList    *mcv;
+	Selectivity s = 0.0;
+
+	/* match/mismatch bitmap for each MCV item */
+	bool	   *matches = NULL;
+
+	/* load the MCV list stored in the statistics object */
+	mcv = statext_mcv_load(stat->statOid);
+
+	/* build a match bitmap for the clauses */
+	matches = mcv_get_match_bitmap(root, clauses, stat->keys, stat->exprs,
+								   mcv, false);
+
+	/* sum frequencies for all the matching MCV items */
+	*basesel = 0.0;
+	*totalsel = 0.0;
+	for (i = 0; i < mcv->nitems; i++)
+	{
+		*totalsel += mcv->items[i].frequency;
+
+		if (matches[i] != false)
+		{
+			*basesel += mcv->items[i].base_frequency;
+			s += mcv->items[i].frequency;
+		}
+	}
+
+	return s;
+}
+
+
+/*
+ * mcv_clause_selectivity_or
+ *		Use MCV statistics to estimate the selectivity of a clause that
+ *		appears in an ORed list of clauses.
+ *
+ * As with mcv_clauselist_selectivity() this determines which MCV items match
+ * the clause and returns both the sum of the frequencies and the sum of the
+ * base frequencies of those items, as well as the sum of the frequencies of
+ * all MCV items (not just the matching ones) so that this information can be
+ * used by mcv_combine_selectivities() to produce a selectivity estimate that
+ * makes use of both per-column and multi-column statistics.
+ *
+ * Additionally, we return information to help compute the overall selectivity
+ * of the ORed list of clauses assumed to contain this clause.  This function
+ * is intended to be called for each clause in the ORed list of clauses,
+ * allowing the overall selectivity to be computed using the following
+ * algorithm:
+ *
+ * Suppose P[n] = P(C[1] OR C[2] OR ... OR C[n]) is the combined selectivity
+ * of the first n clauses in the list.  Then the combined selectivity taking
+ * into account the next clause C[n+1] can be written as
+ *
+ *		P[n+1] = P[n] + P(C[n+1]) - P((C[1] OR ... OR C[n]) AND C[n+1])
+ *
+ * The final term above represents the overlap between the clauses examined so
+ * far and the (n+1)'th clause.  To estimate its selectivity, we track the
+ * match bitmap for the ORed list of clauses examined so far and examine its
+ * intersection with the match bitmap for the (n+1)'th clause.
+ *
+ * We then also return the sums of the MCV item frequencies and base
+ * frequencies for the match bitmap intersection corresponding to the overlap
+ * term above, so that they can be combined with a simple selectivity estimate
+ * for that term.
+ *
+ * The parameter "or_matches" is an in/out parameter tracking the match bitmap
+ * for the clauses examined so far.  The caller is expected to set it to NULL
+ * the first time it calls this function.
+ */
+Selectivity
+mcv_clause_selectivity_or(PlannerInfo *root, StatisticExtInfo *stat,
+						  MCVList *mcv, Node *clause, bool **or_matches,
+						  Selectivity *basesel, Selectivity *overlap_mcvsel,
+						  Selectivity *overlap_basesel, Selectivity *totalsel)
+{
+	Selectivity s = 0.0;
+	bool	   *new_matches;
+	int			i;
+
+	/* build the OR-matches bitmap, if not built already */
+	if (*or_matches == NULL)
+		*or_matches = palloc0(sizeof(bool) * mcv->nitems);
+
+	/* build the match bitmap for the new clause */
+	new_matches = mcv_get_match_bitmap(root, list_make1(clause), stat->keys,
+									   stat->exprs, mcv, false);
+
+	/*
+	 * Sum the frequencies for all the MCV items matching this clause and also
+	 * those matching the overlap between this clause and any of the preceding
+	 * clauses as described above.
+	 */
+	*basesel = 0.0;
+	*overlap_mcvsel = 0.0;
+	*overlap_basesel = 0.0;
+	*totalsel = 0.0;
+	for (i = 0; i < mcv->nitems; i++)
+	{
+		*totalsel += mcv->items[i].frequency;
+
+		if (new_matches[i])
+		{
+			s += mcv->items[i].frequency;
+			*basesel += mcv->items[i].base_frequency;
+
+			if ((*or_matches)[i])
+			{
+				*overlap_mcvsel += mcv->items[i].frequency;
+				*overlap_basesel += mcv->items[i].base_frequency;
+			}
+		}
+
+		/* update the OR-matches bitmap for the next clause */
+		(*or_matches)[i] = (*or_matches)[i] || new_matches[i];
+	}
+
+	pfree(new_matches);
+
+	return s;
+}
diff --git a/src/backend/statistics/mvdistinct.c b/src/backend/statistics/mvdistinct.c
new file mode 100644
index 0000000..4b4ecec
--- /dev/null
+++ b/src/backend/statistics/mvdistinct.c
@@ -0,0 +1,699 @@
+/*-------------------------------------------------------------------------
+ *
+ * mvdistinct.c
+ *	  POSTGRES multivariate ndistinct coefficients
+ *
+ * Estimating number of groups in a combination of columns (e.g. for GROUP BY)
+ * is tricky, and the estimation error is often significant.
+
+ * The multivariate ndistinct coefficients address this by storing ndistinct
+ * estimates for combinations of the user-specified columns.  So for example
+ * given a statistics object on three columns (a,b,c), this module estimates
+ * and stores n-distinct for (a,b), (a,c), (b,c) and (a,b,c).  The per-column
+ * estimates are already available in pg_statistic.
+ *
+ *
+ * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * IDENTIFICATION
+ *	  src/backend/statistics/mvdistinct.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include <math.h>
+
+#include "access/htup_details.h"
+#include "catalog/pg_statistic_ext.h"
+#include "catalog/pg_statistic_ext_data.h"
+#include "lib/stringinfo.h"
+#include "statistics/extended_stats_internal.h"
+#include "statistics/statistics.h"
+#include "utils/fmgrprotos.h"
+#include "utils/lsyscache.h"
+#include "utils/syscache.h"
+#include "utils/typcache.h"
+
+static double ndistinct_for_combination(double totalrows, StatsBuildData *data,
+										int k, int *combination);
+static double estimate_ndistinct(double totalrows, int numrows, int d, int f1);
+static int	n_choose_k(int n, int k);
+static int	num_combinations(int n);
+
+/* size of the struct header fields (magic, type, nitems) */
+#define SizeOfHeader		(3 * sizeof(uint32))
+
+/* size of a serialized ndistinct item (coefficient, natts, atts) */
+#define SizeOfItem(natts) \
+	(sizeof(double) + sizeof(int) + (natts) * sizeof(AttrNumber))
+
+/* minimal size of a ndistinct item (with two attributes) */
+#define MinSizeOfItem	SizeOfItem(2)
+
+/* minimal size of mvndistinct, when all items are minimal */
+#define MinSizeOfItems(nitems)	\
+	(SizeOfHeader + (nitems) * MinSizeOfItem)
+
+/* Combination generator API */
+
+/* internal state for generator of k-combinations of n elements */
+typedef struct CombinationGenerator
+{
+	int			k;				/* size of the combination */
+	int			n;				/* total number of elements */
+	int			current;		/* index of the next combination to return */
+	int			ncombinations;	/* number of combinations (size of array) */
+	int		   *combinations;	/* array of pre-built combinations */
+} CombinationGenerator;
+
+static CombinationGenerator *generator_init(int n, int k);
+static void generator_free(CombinationGenerator *state);
+static int *generator_next(CombinationGenerator *state);
+static void generate_combinations(CombinationGenerator *state);
+
+
+/*
+ * statext_ndistinct_build
+ *		Compute ndistinct coefficient for the combination of attributes.
+ *
+ * This computes the ndistinct estimate using the same estimator used
+ * in analyze.c and then computes the coefficient.
+ *
+ * To handle expressions easily, we treat them as system attributes with
+ * negative attnums, and offset everything by number of expressions to
+ * allow using Bitmapsets.
+ */
+MVNDistinct *
+statext_ndistinct_build(double totalrows, StatsBuildData *data)
+{
+	MVNDistinct *result;
+	int			k;
+	int			itemcnt;
+	int			numattrs = data->nattnums;
+	int			numcombs = num_combinations(numattrs);
+
+	result = palloc(offsetof(MVNDistinct, items) +
+					numcombs * sizeof(MVNDistinctItem));
+	result->magic = STATS_NDISTINCT_MAGIC;
+	result->type = STATS_NDISTINCT_TYPE_BASIC;
+	result->nitems = numcombs;
+
+	itemcnt = 0;
+	for (k = 2; k <= numattrs; k++)
+	{
+		int		   *combination;
+		CombinationGenerator *generator;
+
+		/* generate combinations of K out of N elements */
+		generator = generator_init(numattrs, k);
+
+		while ((combination = generator_next(generator)))
+		{
+			MVNDistinctItem *item = &result->items[itemcnt];
+			int			j;
+
+			item->attributes = palloc(sizeof(AttrNumber) * k);
+			item->nattributes = k;
+
+			/* translate the indexes to attnums */
+			for (j = 0; j < k; j++)
+			{
+				item->attributes[j] = data->attnums[combination[j]];
+
+				Assert(AttributeNumberIsValid(item->attributes[j]));
+			}
+
+			item->ndistinct =
+				ndistinct_for_combination(totalrows, data, k, combination);
+
+			itemcnt++;
+			Assert(itemcnt <= result->nitems);
+		}
+
+		generator_free(generator);
+	}
+
+	/* must consume exactly the whole output array */
+	Assert(itemcnt == result->nitems);
+
+	return result;
+}
+
+/*
+ * statext_ndistinct_load
+ *		Load the ndistinct value for the indicated pg_statistic_ext tuple
+ */
+MVNDistinct *
+statext_ndistinct_load(Oid mvoid)
+{
+	MVNDistinct *result;
+	bool		isnull;
+	Datum		ndist;
+	HeapTuple	htup;
+
+	htup = SearchSysCache1(STATEXTDATASTXOID, ObjectIdGetDatum(mvoid));
+	if (!HeapTupleIsValid(htup))
+		elog(ERROR, "cache lookup failed for statistics object %u", mvoid);
+
+	ndist = SysCacheGetAttr(STATEXTDATASTXOID, htup,
+							Anum_pg_statistic_ext_data_stxdndistinct, &isnull);
+	if (isnull)
+		elog(ERROR,
+			 "requested statistics kind \"%c\" is not yet built for statistics object %u",
+			 STATS_EXT_NDISTINCT, mvoid);
+
+	result = statext_ndistinct_deserialize(DatumGetByteaPP(ndist));
+
+	ReleaseSysCache(htup);
+
+	return result;
+}
+
+/*
+ * statext_ndistinct_serialize
+ *		serialize ndistinct to the on-disk bytea format
+ */
+bytea *
+statext_ndistinct_serialize(MVNDistinct *ndistinct)
+{
+	int			i;
+	bytea	   *output;
+	char	   *tmp;
+	Size		len;
+
+	Assert(ndistinct->magic == STATS_NDISTINCT_MAGIC);
+	Assert(ndistinct->type == STATS_NDISTINCT_TYPE_BASIC);
+
+	/*
+	 * Base size is size of scalar fields in the struct, plus one base struct
+	 * for each item, including number of items for each.
+	 */
+	len = VARHDRSZ + SizeOfHeader;
+
+	/* and also include space for the actual attribute numbers */
+	for (i = 0; i < ndistinct->nitems; i++)
+	{
+		int			nmembers;
+
+		nmembers = ndistinct->items[i].nattributes;
+		Assert(nmembers >= 2);
+
+		len += SizeOfItem(nmembers);
+	}
+
+	output = (bytea *) palloc(len);
+	SET_VARSIZE(output, len);
+
+	tmp = VARDATA(output);
+
+	/* Store the base struct values (magic, type, nitems) */
+	memcpy(tmp, &ndistinct->magic, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(tmp, &ndistinct->type, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(tmp, &ndistinct->nitems, sizeof(uint32));
+	tmp += sizeof(uint32);
+
+	/*
+	 * store number of attributes and attribute numbers for each entry
+	 */
+	for (i = 0; i < ndistinct->nitems; i++)
+	{
+		MVNDistinctItem item = ndistinct->items[i];
+		int			nmembers = item.nattributes;
+
+		memcpy(tmp, &item.ndistinct, sizeof(double));
+		tmp += sizeof(double);
+		memcpy(tmp, &nmembers, sizeof(int));
+		tmp += sizeof(int);
+
+		memcpy(tmp, item.attributes, sizeof(AttrNumber) * nmembers);
+		tmp += nmembers * sizeof(AttrNumber);
+
+		/* protect against overflows */
+		Assert(tmp <= ((char *) output + len));
+	}
+
+	/* check we used exactly the expected space */
+	Assert(tmp == ((char *) output + len));
+
+	return output;
+}
+
+/*
+ * statext_ndistinct_deserialize
+ *		Read an on-disk bytea format MVNDistinct to in-memory format
+ */
+MVNDistinct *
+statext_ndistinct_deserialize(bytea *data)
+{
+	int			i;
+	Size		minimum_size;
+	MVNDistinct ndist;
+	MVNDistinct *ndistinct;
+	char	   *tmp;
+
+	if (data == NULL)
+		return NULL;
+
+	/* we expect at least the basic fields of MVNDistinct struct */
+	if (VARSIZE_ANY_EXHDR(data) < SizeOfHeader)
+		elog(ERROR, "invalid MVNDistinct size %zd (expected at least %zd)",
+			 VARSIZE_ANY_EXHDR(data), SizeOfHeader);
+
+	/* initialize pointer to the data part (skip the varlena header) */
+	tmp = VARDATA_ANY(data);
+
+	/* read the header fields and perform basic sanity checks */
+	memcpy(&ndist.magic, tmp, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(&ndist.type, tmp, sizeof(uint32));
+	tmp += sizeof(uint32);
+	memcpy(&ndist.nitems, tmp, sizeof(uint32));
+	tmp += sizeof(uint32);
+
+	if (ndist.magic != STATS_NDISTINCT_MAGIC)
+		elog(ERROR, "invalid ndistinct magic %08x (expected %08x)",
+			 ndist.magic, STATS_NDISTINCT_MAGIC);
+	if (ndist.type != STATS_NDISTINCT_TYPE_BASIC)
+		elog(ERROR, "invalid ndistinct type %d (expected %d)",
+			 ndist.type, STATS_NDISTINCT_TYPE_BASIC);
+	if (ndist.nitems == 0)
+		elog(ERROR, "invalid zero-length item array in MVNDistinct");
+
+	/* what minimum bytea size do we expect for those parameters */
+	minimum_size = MinSizeOfItems(ndist.nitems);
+	if (VARSIZE_ANY_EXHDR(data) < minimum_size)
+		elog(ERROR, "invalid MVNDistinct size %zd (expected at least %zd)",
+			 VARSIZE_ANY_EXHDR(data), minimum_size);
+
+	/*
+	 * Allocate space for the ndistinct items (no space for each item's
+	 * attnos: those live in bitmapsets allocated separately)
+	 */
+	ndistinct = palloc0(MAXALIGN(offsetof(MVNDistinct, items)) +
+						(ndist.nitems * sizeof(MVNDistinctItem)));
+	ndistinct->magic = ndist.magic;
+	ndistinct->type = ndist.type;
+	ndistinct->nitems = ndist.nitems;
+
+	for (i = 0; i < ndistinct->nitems; i++)
+	{
+		MVNDistinctItem *item = &ndistinct->items[i];
+
+		/* ndistinct value */
+		memcpy(&item->ndistinct, tmp, sizeof(double));
+		tmp += sizeof(double);
+
+		/* number of attributes */
+		memcpy(&item->nattributes, tmp, sizeof(int));
+		tmp += sizeof(int);
+		Assert((item->nattributes >= 2) && (item->nattributes <= STATS_MAX_DIMENSIONS));
+
+		item->attributes
+			= (AttrNumber *) palloc(item->nattributes * sizeof(AttrNumber));
+
+		memcpy(item->attributes, tmp, sizeof(AttrNumber) * item->nattributes);
+		tmp += sizeof(AttrNumber) * item->nattributes;
+
+		/* still within the bytea */
+		Assert(tmp <= ((char *) data + VARSIZE_ANY(data)));
+	}
+
+	/* we should have consumed the whole bytea exactly */
+	Assert(tmp == ((char *) data + VARSIZE_ANY(data)));
+
+	return ndistinct;
+}
+
+/*
+ * pg_ndistinct_in
+ *		input routine for type pg_ndistinct
+ *
+ * pg_ndistinct is real enough to be a table column, but it has no
+ * operations of its own, and disallows input (just like pg_node_tree).
+ */
+Datum
+pg_ndistinct_in(PG_FUNCTION_ARGS)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+			 errmsg("cannot accept a value of type %s", "pg_ndistinct")));
+
+	PG_RETURN_VOID();			/* keep compiler quiet */
+}
+
+/*
+ * pg_ndistinct
+ *		output routine for type pg_ndistinct
+ *
+ * Produces a human-readable representation of the value.
+ */
+Datum
+pg_ndistinct_out(PG_FUNCTION_ARGS)
+{
+	bytea	   *data = PG_GETARG_BYTEA_PP(0);
+	MVNDistinct *ndist = statext_ndistinct_deserialize(data);
+	int			i;
+	StringInfoData str;
+
+	initStringInfo(&str);
+	appendStringInfoChar(&str, '{');
+
+	for (i = 0; i < ndist->nitems; i++)
+	{
+		int			j;
+		MVNDistinctItem item = ndist->items[i];
+
+		if (i > 0)
+			appendStringInfoString(&str, ", ");
+
+		for (j = 0; j < item.nattributes; j++)
+		{
+			AttrNumber	attnum = item.attributes[j];
+
+			appendStringInfo(&str, "%s%d", (j == 0) ? "\"" : ", ", attnum);
+		}
+		appendStringInfo(&str, "\": %d", (int) item.ndistinct);
+	}
+
+	appendStringInfoChar(&str, '}');
+
+	PG_RETURN_CSTRING(str.data);
+}
+
+/*
+ * pg_ndistinct_recv
+ *		binary input routine for type pg_ndistinct
+ */
+Datum
+pg_ndistinct_recv(PG_FUNCTION_ARGS)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+			 errmsg("cannot accept a value of type %s", "pg_ndistinct")));
+
+	PG_RETURN_VOID();			/* keep compiler quiet */
+}
+
+/*
+ * pg_ndistinct_send
+ *		binary output routine for type pg_ndistinct
+ *
+ * n-distinct is serialized into a bytea value, so let's send that.
+ */
+Datum
+pg_ndistinct_send(PG_FUNCTION_ARGS)
+{
+	return byteasend(fcinfo);
+}
+
+/*
+ * ndistinct_for_combination
+ *		Estimates number of distinct values in a combination of columns.
+ *
+ * This uses the same ndistinct estimator as compute_scalar_stats() in
+ * ANALYZE, i.e.,
+ *		n*d / (n - f1 + f1*n/N)
+ *
+ * except that instead of values in a single column we are dealing with
+ * combination of multiple columns.
+ */
+static double
+ndistinct_for_combination(double totalrows, StatsBuildData *data,
+						  int k, int *combination)
+{
+	int			i,
+				j;
+	int			f1,
+				cnt,
+				d;
+	bool	   *isnull;
+	Datum	   *values;
+	SortItem   *items;
+	MultiSortSupport mss;
+	int			numrows = data->numrows;
+
+	mss = multi_sort_init(k);
+
+	/*
+	 * In order to determine the number of distinct elements, create separate
+	 * values[]/isnull[] arrays with all the data we have, then sort them
+	 * using the specified column combination as dimensions.  We could try to
+	 * sort in place, but it'd probably be more complex and bug-prone.
+	 */
+	items = (SortItem *) palloc(numrows * sizeof(SortItem));
+	values = (Datum *) palloc0(sizeof(Datum) * numrows * k);
+	isnull = (bool *) palloc0(sizeof(bool) * numrows * k);
+
+	for (i = 0; i < numrows; i++)
+	{
+		items[i].values = &values[i * k];
+		items[i].isnull = &isnull[i * k];
+	}
+
+	/*
+	 * For each dimension, set up sort-support and fill in the values from the
+	 * sample data.
+	 *
+	 * We use the column data types' default sort operators and collations;
+	 * perhaps at some point it'd be worth using column-specific collations?
+	 */
+	for (i = 0; i < k; i++)
+	{
+		Oid			typid;
+		TypeCacheEntry *type;
+		Oid			collid = InvalidOid;
+		VacAttrStats *colstat = data->stats[combination[i]];
+
+		typid = colstat->attrtypid;
+		collid = colstat->attrcollid;
+
+		type = lookup_type_cache(typid, TYPECACHE_LT_OPR);
+		if (type->lt_opr == InvalidOid) /* shouldn't happen */
+			elog(ERROR, "cache lookup failed for ordering operator for type %u",
+				 typid);
+
+		/* prepare the sort function for this dimension */
+		multi_sort_add_dimension(mss, i, type->lt_opr, collid);
+
+		/* accumulate all the data for this dimension into the arrays */
+		for (j = 0; j < numrows; j++)
+		{
+			items[j].values[i] = data->values[combination[i]][j];
+			items[j].isnull[i] = data->nulls[combination[i]][j];
+		}
+	}
+
+	/* We can sort the array now ... */
+	qsort_interruptible((void *) items, numrows, sizeof(SortItem),
+						multi_sort_compare, mss);
+
+	/* ... and count the number of distinct combinations */
+
+	f1 = 0;
+	cnt = 1;
+	d = 1;
+	for (i = 1; i < numrows; i++)
+	{
+		if (multi_sort_compare(&items[i], &items[i - 1], mss) != 0)
+		{
+			if (cnt == 1)
+				f1 += 1;
+
+			d++;
+			cnt = 0;
+		}
+
+		cnt += 1;
+	}
+
+	if (cnt == 1)
+		f1 += 1;
+
+	return estimate_ndistinct(totalrows, numrows, d, f1);
+}
+
+/* The Duj1 estimator (already used in analyze.c). */
+static double
+estimate_ndistinct(double totalrows, int numrows, int d, int f1)
+{
+	double		numer,
+				denom,
+				ndistinct;
+
+	numer = (double) numrows * (double) d;
+
+	denom = (double) (numrows - f1) +
+		(double) f1 * (double) numrows / totalrows;
+
+	ndistinct = numer / denom;
+
+	/* Clamp to sane range in case of roundoff error */
+	if (ndistinct < (double) d)
+		ndistinct = (double) d;
+
+	if (ndistinct > totalrows)
+		ndistinct = totalrows;
+
+	return floor(ndistinct + 0.5);
+}
+
+/*
+ * n_choose_k
+ *		computes binomial coefficients using an algorithm that is both
+ *		efficient and prevents overflows
+ */
+static int
+n_choose_k(int n, int k)
+{
+	int			d,
+				r;
+
+	Assert((k > 0) && (n >= k));
+
+	/* use symmetry of the binomial coefficients */
+	k = Min(k, n - k);
+
+	r = 1;
+	for (d = 1; d <= k; ++d)
+	{
+		r *= n--;
+		r /= d;
+	}
+
+	return r;
+}
+
+/*
+ * num_combinations
+ *		number of combinations, excluding single-value combinations
+ */
+static int
+num_combinations(int n)
+{
+	return (1 << n) - (n + 1);
+}
+
+/*
+ * generator_init
+ *		initialize the generator of combinations
+ *
+ * The generator produces combinations of K elements in the interval (0..N).
+ * We prebuild all the combinations in this method, which is simpler than
+ * generating them on the fly.
+ */
+static CombinationGenerator *
+generator_init(int n, int k)
+{
+	CombinationGenerator *state;
+
+	Assert((n >= k) && (k > 0));
+
+	/* allocate the generator state as a single chunk of memory */
+	state = (CombinationGenerator *) palloc(sizeof(CombinationGenerator));
+
+	state->ncombinations = n_choose_k(n, k);
+
+	/* pre-allocate space for all combinations */
+	state->combinations = (int *) palloc(sizeof(int) * k * state->ncombinations);
+
+	state->current = 0;
+	state->k = k;
+	state->n = n;
+
+	/* now actually pre-generate all the combinations of K elements */
+	generate_combinations(state);
+
+	/* make sure we got the expected number of combinations */
+	Assert(state->current == state->ncombinations);
+
+	/* reset the number, so we start with the first one */
+	state->current = 0;
+
+	return state;
+}
+
+/*
+ * generator_next
+ *		returns the next combination from the prebuilt list
+ *
+ * Returns a combination of K array indexes (0 .. N), as specified to
+ * generator_init), or NULL when there are no more combination.
+ */
+static int *
+generator_next(CombinationGenerator *state)
+{
+	if (state->current == state->ncombinations)
+		return NULL;
+
+	return &state->combinations[state->k * state->current++];
+}
+
+/*
+ * generator_free
+ *		free the internal state of the generator
+ *
+ * Releases the generator internal state (pre-built combinations).
+ */
+static void
+generator_free(CombinationGenerator *state)
+{
+	pfree(state->combinations);
+	pfree(state);
+}
+
+/*
+ * generate_combinations_recurse
+ *		given a prefix, generate all possible combinations
+ *
+ * Given a prefix (first few elements of the combination), generate following
+ * elements recursively. We generate the combinations in lexicographic order,
+ * which eliminates permutations of the same combination.
+ */
+static void
+generate_combinations_recurse(CombinationGenerator *state,
+							  int index, int start, int *current)
+{
+	/* If we haven't filled all the elements, simply recurse. */
+	if (index < state->k)
+	{
+		int			i;
+
+		/*
+		 * The values have to be in ascending order, so make sure we start
+		 * with the value passed by parameter.
+		 */
+
+		for (i = start; i < state->n; i++)
+		{
+			current[index] = i;
+			generate_combinations_recurse(state, (index + 1), (i + 1), current);
+		}
+
+		return;
+	}
+	else
+	{
+		/* we got a valid combination, add it to the array */
+		memcpy(&state->combinations[(state->k * state->current)],
+			   current, state->k * sizeof(int));
+		state->current++;
+	}
+}
+
+/*
+ * generate_combinations
+ *		generate all k-combinations of N elements
+ */
+static void
+generate_combinations(CombinationGenerator *state)
+{
+	int		   *current = (int *) palloc0(sizeof(int) * state->k);
+
+	generate_combinations_recurse(state, 0, 0, current);
+
+	pfree(current);
+}