PostgreSQL Source Code  git master
selfuncs.c File Reference
#include "postgres.h"
#include <ctype.h>
#include <math.h>
#include "access/brin.h"
#include "access/brin_page.h"
#include "access/gin.h"
#include "access/table.h"
#include "access/tableam.h"
#include "access/visibilitymap.h"
#include "catalog/pg_am.h"
#include "catalog/pg_collation.h"
#include "catalog/pg_operator.h"
#include "catalog/pg_statistic.h"
#include "catalog/pg_statistic_ext.h"
#include "executor/nodeAgg.h"
#include "miscadmin.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "optimizer/clauses.h"
#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "optimizer/plancat.h"
#include "parser/parse_clause.h"
#include "parser/parse_relation.h"
#include "parser/parsetree.h"
#include "statistics/statistics.h"
#include "storage/bufmgr.h"
#include "utils/acl.h"
#include "utils/array.h"
#include "utils/builtins.h"
#include "utils/date.h"
#include "utils/datum.h"
#include "utils/fmgroids.h"
#include "utils/index_selfuncs.h"
#include "utils/lsyscache.h"
#include "utils/memutils.h"
#include "utils/pg_locale.h"
#include "utils/rel.h"
#include "utils/selfuncs.h"
#include "utils/snapmgr.h"
#include "utils/spccache.h"
#include "utils/syscache.h"
#include "utils/timestamp.h"
#include "utils/typcache.h"
Include dependency graph for selfuncs.c:

Go to the source code of this file.

Data Structures

struct  GroupVarInfo
 
struct  GinQualCounts
 

Macros

#define DEFAULT_PAGE_CPU_MULTIPLIER   50.0
 
#define VISITED_PAGES_LIMIT   100
 

Functions

static double eqsel_internal (PG_FUNCTION_ARGS, bool negate)
 
static double eqjoinsel_inner (Oid opfuncoid, Oid collation, VariableStatData *vardata1, VariableStatData *vardata2, double nd1, double nd2, bool isdefault1, bool isdefault2, AttStatsSlot *sslot1, AttStatsSlot *sslot2, Form_pg_statistic stats1, Form_pg_statistic stats2, bool have_mcvs1, bool have_mcvs2)
 
static double eqjoinsel_semi (Oid opfuncoid, Oid collation, VariableStatData *vardata1, VariableStatData *vardata2, double nd1, double nd2, bool isdefault1, bool isdefault2, AttStatsSlot *sslot1, AttStatsSlot *sslot2, Form_pg_statistic stats1, Form_pg_statistic stats2, bool have_mcvs1, bool have_mcvs2, RelOptInfo *inner_rel)
 
static bool estimate_multivariate_ndistinct (PlannerInfo *root, RelOptInfo *rel, List **varinfos, double *ndistinct)
 
static bool convert_to_scalar (Datum value, Oid valuetypid, Oid collid, double *scaledvalue, Datum lobound, Datum hibound, Oid boundstypid, double *scaledlobound, double *scaledhibound)
 
static double convert_numeric_to_scalar (Datum value, Oid typid, bool *failure)
 
static void convert_string_to_scalar (char *value, double *scaledvalue, char *lobound, double *scaledlobound, char *hibound, double *scaledhibound)
 
static void convert_bytea_to_scalar (Datum value, double *scaledvalue, Datum lobound, double *scaledlobound, Datum hibound, double *scaledhibound)
 
static double convert_one_string_to_scalar (char *value, int rangelo, int rangehi)
 
static double convert_one_bytea_to_scalar (unsigned char *value, int valuelen, int rangelo, int rangehi)
 
static char * convert_string_datum (Datum value, Oid typid, Oid collid, bool *failure)
 
static double convert_timevalue_to_scalar (Datum value, Oid typid, bool *failure)
 
static void examine_simple_variable (PlannerInfo *root, Var *var, VariableStatData *vardata)
 
static bool get_variable_range (PlannerInfo *root, VariableStatData *vardata, Oid sortop, Oid collation, Datum *min, Datum *max)
 
static void get_stats_slot_range (AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc, Oid collation, int16 typLen, bool typByVal, Datum *min, Datum *max, bool *p_have_data)
 
static bool get_actual_variable_range (PlannerInfo *root, VariableStatData *vardata, Oid sortop, Oid collation, Datum *min, Datum *max)
 
static bool get_actual_variable_endpoint (Relation heapRel, Relation indexRel, ScanDirection indexscandir, ScanKey scankeys, int16 typLen, bool typByVal, TupleTableSlot *tableslot, MemoryContext outercontext, Datum *endpointDatum)
 
static RelOptInfofind_join_input_rel (PlannerInfo *root, Relids relids)
 
Datum eqsel (PG_FUNCTION_ARGS)
 
double var_eq_const (VariableStatData *vardata, Oid oproid, Oid collation, Datum constval, bool constisnull, bool varonleft, bool negate)
 
double var_eq_non_const (VariableStatData *vardata, Oid oproid, Oid collation, Node *other, bool varonleft, bool negate)
 
Datum neqsel (PG_FUNCTION_ARGS)
 
static double scalarineqsel (PlannerInfo *root, Oid operator, bool isgt, bool iseq, Oid collation, VariableStatData *vardata, Datum constval, Oid consttype)
 
double mcv_selectivity (VariableStatData *vardata, FmgrInfo *opproc, Oid collation, Datum constval, bool varonleft, double *sumcommonp)
 
double histogram_selectivity (VariableStatData *vardata, FmgrInfo *opproc, Oid collation, Datum constval, bool varonleft, int min_hist_size, int n_skip, int *hist_size)
 
double generic_restriction_selectivity (PlannerInfo *root, Oid oproid, Oid collation, List *args, int varRelid, double default_selectivity)
 
double ineq_histogram_selectivity (PlannerInfo *root, VariableStatData *vardata, Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq, Oid collation, Datum constval, Oid consttype)
 
static Datum scalarineqsel_wrapper (PG_FUNCTION_ARGS, bool isgt, bool iseq)
 
Datum scalarltsel (PG_FUNCTION_ARGS)
 
Datum scalarlesel (PG_FUNCTION_ARGS)
 
Datum scalargtsel (PG_FUNCTION_ARGS)
 
Datum scalargesel (PG_FUNCTION_ARGS)
 
Selectivity boolvarsel (PlannerInfo *root, Node *arg, int varRelid)
 
Selectivity booltestsel (PlannerInfo *root, BoolTestType booltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
 
Selectivity nulltestsel (PlannerInfo *root, NullTestType nulltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
 
static Nodestrip_array_coercion (Node *node)
 
Selectivity scalararraysel (PlannerInfo *root, ScalarArrayOpExpr *clause, bool is_join_clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
 
double estimate_array_length (PlannerInfo *root, Node *arrayexpr)
 
Selectivity rowcomparesel (PlannerInfo *root, RowCompareExpr *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
 
Datum eqjoinsel (PG_FUNCTION_ARGS)
 
Datum neqjoinsel (PG_FUNCTION_ARGS)
 
Datum scalarltjoinsel (PG_FUNCTION_ARGS)
 
Datum scalarlejoinsel (PG_FUNCTION_ARGS)
 
Datum scalargtjoinsel (PG_FUNCTION_ARGS)
 
Datum scalargejoinsel (PG_FUNCTION_ARGS)
 
void mergejoinscansel (PlannerInfo *root, Node *clause, Oid opfamily, int strategy, bool nulls_first, Selectivity *leftstart, Selectivity *leftend, Selectivity *rightstart, Selectivity *rightend)
 
Datum matchingsel (PG_FUNCTION_ARGS)
 
Datum matchingjoinsel (PG_FUNCTION_ARGS)
 
static Listadd_unique_group_var (PlannerInfo *root, List *varinfos, Node *var, VariableStatData *vardata)
 
double estimate_num_groups (PlannerInfo *root, List *groupExprs, double input_rows, List **pgset, EstimationInfo *estinfo)
 
void estimate_hash_bucket_stats (PlannerInfo *root, Node *hashkey, double nbuckets, Selectivity *mcv_freq, Selectivity *bucketsize_frac)
 
double estimate_hashagg_tablesize (PlannerInfo *root, Path *path, const AggClauseCosts *agg_costs, double dNumGroups)
 
bool get_restriction_variable (PlannerInfo *root, List *args, int varRelid, VariableStatData *vardata, Node **other, bool *varonleft)
 
void get_join_variables (PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo, VariableStatData *vardata1, VariableStatData *vardata2, bool *join_is_reversed)
 
static void ReleaseDummy (HeapTuple tuple)
 
void examine_variable (PlannerInfo *root, Node *node, int varRelid, VariableStatData *vardata)
 
bool statistic_proc_security_check (VariableStatData *vardata, Oid func_oid)
 
double get_variable_numdistinct (VariableStatData *vardata, bool *isdefault)
 
Listget_quals_from_indexclauses (List *indexclauses)
 
Cost index_other_operands_eval_cost (PlannerInfo *root, List *indexquals)
 
void genericcostestimate (PlannerInfo *root, IndexPath *path, double loop_count, GenericCosts *costs)
 
Listadd_predicate_to_index_quals (IndexOptInfo *index, List *indexQuals)
 
void btcostestimate (PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
 
void hashcostestimate (PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
 
void gistcostestimate (PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
 
void spgcostestimate (PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
 
static bool gincost_pattern (IndexOptInfo *index, int indexcol, Oid clause_op, Datum query, GinQualCounts *counts)
 
static bool gincost_opexpr (PlannerInfo *root, IndexOptInfo *index, int indexcol, OpExpr *clause, GinQualCounts *counts)
 
static bool gincost_scalararrayopexpr (PlannerInfo *root, IndexOptInfo *index, int indexcol, ScalarArrayOpExpr *clause, double numIndexEntries, GinQualCounts *counts)
 
void gincostestimate (PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
 
void brincostestimate (PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
 

Variables

get_relation_stats_hook_type get_relation_stats_hook = NULL
 
get_index_stats_hook_type get_index_stats_hook = NULL
 

Macro Definition Documentation

◆ DEFAULT_PAGE_CPU_MULTIPLIER

#define DEFAULT_PAGE_CPU_MULTIPLIER   50.0

Definition at line 144 of file selfuncs.c.

◆ VISITED_PAGES_LIMIT

#define VISITED_PAGES_LIMIT   100

Function Documentation

◆ add_predicate_to_index_quals()

List* add_predicate_to_index_quals ( IndexOptInfo index,
List indexQuals 
)

Definition at line 6762 of file selfuncs.c.

6763 {
6764  List *predExtraQuals = NIL;
6765  ListCell *lc;
6766 
6767  if (index->indpred == NIL)
6768  return indexQuals;
6769 
6770  foreach(lc, index->indpred)
6771  {
6772  Node *predQual = (Node *) lfirst(lc);
6773  List *oneQual = list_make1(predQual);
6774 
6775  if (!predicate_implied_by(oneQual, indexQuals, false))
6776  predExtraQuals = list_concat(predExtraQuals, oneQual);
6777  }
6778  return list_concat(predExtraQuals, indexQuals);
6779 }
List * list_concat(List *list1, const List *list2)
Definition: list.c:561
#define lfirst(lc)
Definition: pg_list.h:172
#define NIL
Definition: pg_list.h:68
#define list_make1(x1)
Definition: pg_list.h:212
bool predicate_implied_by(List *predicate_list, List *clause_list, bool weak)
Definition: predtest.c:152
Definition: pg_list.h:54
Definition: nodes.h:129
Definition: type.h:95

References lfirst, list_concat(), list_make1, NIL, and predicate_implied_by().

Referenced by btcostestimate(), genericcostestimate(), and gincostestimate().

◆ add_unique_group_var()

static List* add_unique_group_var ( PlannerInfo root,
List varinfos,
Node var,
VariableStatData vardata 
)
static

Definition at line 3296 of file selfuncs.c.

3298 {
3299  GroupVarInfo *varinfo;
3300  double ndistinct;
3301  bool isdefault;
3302  ListCell *lc;
3303 
3304  ndistinct = get_variable_numdistinct(vardata, &isdefault);
3305 
3306  foreach(lc, varinfos)
3307  {
3308  varinfo = (GroupVarInfo *) lfirst(lc);
3309 
3310  /* Drop exact duplicates */
3311  if (equal(var, varinfo->var))
3312  return varinfos;
3313 
3314  /*
3315  * Drop known-equal vars, but only if they belong to different
3316  * relations (see comments for estimate_num_groups)
3317  */
3318  if (vardata->rel != varinfo->rel &&
3319  exprs_known_equal(root, var, varinfo->var))
3320  {
3321  if (varinfo->ndistinct <= ndistinct)
3322  {
3323  /* Keep older item, forget new one */
3324  return varinfos;
3325  }
3326  else
3327  {
3328  /* Delete the older item */
3329  varinfos = foreach_delete_current(varinfos, lc);
3330  }
3331  }
3332  }
3333 
3334  varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3335 
3336  varinfo->var = var;
3337  varinfo->rel = vardata->rel;
3338  varinfo->ndistinct = ndistinct;
3339  varinfo->isdefault = isdefault;
3340  varinfos = lappend(varinfos, varinfo);
3341  return varinfos;
3342 }
bool equal(const void *a, const void *b)
Definition: equalfuncs.c:223
bool exprs_known_equal(PlannerInfo *root, Node *item1, Node *item2)
Definition: equivclass.c:2450
List * lappend(List *list, void *datum)
Definition: list.c:339
void * palloc(Size size)
Definition: mcxt.c:1304
#define foreach_delete_current(lst, var_or_cell)
Definition: pg_list.h:391
double get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
Definition: selfuncs.c:5764
RelOptInfo * rel
Definition: selfuncs.c:3290
double ndistinct
Definition: selfuncs.c:3291
bool isdefault
Definition: selfuncs.c:3292
Node * var
Definition: selfuncs.c:3289
RelOptInfo * rel
Definition: selfuncs.h:88

References equal(), exprs_known_equal(), foreach_delete_current, get_variable_numdistinct(), GroupVarInfo::isdefault, lappend(), lfirst, GroupVarInfo::ndistinct, palloc(), GroupVarInfo::rel, VariableStatData::rel, and GroupVarInfo::var.

Referenced by estimate_num_groups().

◆ booltestsel()

Selectivity booltestsel ( PlannerInfo root,
BoolTestType  booltesttype,
Node arg,
int  varRelid,
JoinType  jointype,
SpecialJoinInfo sjinfo 
)

Definition at line 1540 of file selfuncs.c.

1542 {
1543  VariableStatData vardata;
1544  double selec;
1545 
1546  examine_variable(root, arg, varRelid, &vardata);
1547 
1548  if (HeapTupleIsValid(vardata.statsTuple))
1549  {
1550  Form_pg_statistic stats;
1551  double freq_null;
1552  AttStatsSlot sslot;
1553 
1554  stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1555  freq_null = stats->stanullfrac;
1556 
1557  if (get_attstatsslot(&sslot, vardata.statsTuple,
1558  STATISTIC_KIND_MCV, InvalidOid,
1560  && sslot.nnumbers > 0)
1561  {
1562  double freq_true;
1563  double freq_false;
1564 
1565  /*
1566  * Get first MCV frequency and derive frequency for true.
1567  */
1568  if (DatumGetBool(sslot.values[0]))
1569  freq_true = sslot.numbers[0];
1570  else
1571  freq_true = 1.0 - sslot.numbers[0] - freq_null;
1572 
1573  /*
1574  * Next derive frequency for false. Then use these as appropriate
1575  * to derive frequency for each case.
1576  */
1577  freq_false = 1.0 - freq_true - freq_null;
1578 
1579  switch (booltesttype)
1580  {
1581  case IS_UNKNOWN:
1582  /* select only NULL values */
1583  selec = freq_null;
1584  break;
1585  case IS_NOT_UNKNOWN:
1586  /* select non-NULL values */
1587  selec = 1.0 - freq_null;
1588  break;
1589  case IS_TRUE:
1590  /* select only TRUE values */
1591  selec = freq_true;
1592  break;
1593  case IS_NOT_TRUE:
1594  /* select non-TRUE values */
1595  selec = 1.0 - freq_true;
1596  break;
1597  case IS_FALSE:
1598  /* select only FALSE values */
1599  selec = freq_false;
1600  break;
1601  case IS_NOT_FALSE:
1602  /* select non-FALSE values */
1603  selec = 1.0 - freq_false;
1604  break;
1605  default:
1606  elog(ERROR, "unrecognized booltesttype: %d",
1607  (int) booltesttype);
1608  selec = 0.0; /* Keep compiler quiet */
1609  break;
1610  }
1611 
1612  free_attstatsslot(&sslot);
1613  }
1614  else
1615  {
1616  /*
1617  * No most-common-value info available. Still have null fraction
1618  * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1619  * for null fraction and assume a 50-50 split of TRUE and FALSE.
1620  */
1621  switch (booltesttype)
1622  {
1623  case IS_UNKNOWN:
1624  /* select only NULL values */
1625  selec = freq_null;
1626  break;
1627  case IS_NOT_UNKNOWN:
1628  /* select non-NULL values */
1629  selec = 1.0 - freq_null;
1630  break;
1631  case IS_TRUE:
1632  case IS_FALSE:
1633  /* Assume we select half of the non-NULL values */
1634  selec = (1.0 - freq_null) / 2.0;
1635  break;
1636  case IS_NOT_TRUE:
1637  case IS_NOT_FALSE:
1638  /* Assume we select NULLs plus half of the non-NULLs */
1639  /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1640  selec = (freq_null + 1.0) / 2.0;
1641  break;
1642  default:
1643  elog(ERROR, "unrecognized booltesttype: %d",
1644  (int) booltesttype);
1645  selec = 0.0; /* Keep compiler quiet */
1646  break;
1647  }
1648  }
1649  }
1650  else
1651  {
1652  /*
1653  * If we can't get variable statistics for the argument, perhaps
1654  * clause_selectivity can do something with it. We ignore the
1655  * possibility of a NULL value when using clause_selectivity, and just
1656  * assume the value is either TRUE or FALSE.
1657  */
1658  switch (booltesttype)
1659  {
1660  case IS_UNKNOWN:
1661  selec = DEFAULT_UNK_SEL;
1662  break;
1663  case IS_NOT_UNKNOWN:
1664  selec = DEFAULT_NOT_UNK_SEL;
1665  break;
1666  case IS_TRUE:
1667  case IS_NOT_FALSE:
1668  selec = (double) clause_selectivity(root, arg,
1669  varRelid,
1670  jointype, sjinfo);
1671  break;
1672  case IS_FALSE:
1673  case IS_NOT_TRUE:
1674  selec = 1.0 - (double) clause_selectivity(root, arg,
1675  varRelid,
1676  jointype, sjinfo);
1677  break;
1678  default:
1679  elog(ERROR, "unrecognized booltesttype: %d",
1680  (int) booltesttype);
1681  selec = 0.0; /* Keep compiler quiet */
1682  break;
1683  }
1684  }
1685 
1686  ReleaseVariableStats(vardata);
1687 
1688  /* result should be in range, but make sure... */
1689  CLAMP_PROBABILITY(selec);
1690 
1691  return (Selectivity) selec;
1692 }
Selectivity clause_selectivity(PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:667
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:224
#define HeapTupleIsValid(tuple)
Definition: htup.h:78
#define GETSTRUCT(TUP)
Definition: htup_details.h:653
void free_attstatsslot(AttStatsSlot *sslot)
Definition: lsyscache.c:3300
bool get_attstatsslot(AttStatsSlot *sslot, HeapTuple statstuple, int reqkind, Oid reqop, int flags)
Definition: lsyscache.c:3190
#define ATTSTATSSLOT_NUMBERS
Definition: lsyscache.h:43
#define ATTSTATSSLOT_VALUES
Definition: lsyscache.h:42
double Selectivity
Definition: nodes.h:240
void * arg
FormData_pg_statistic * Form_pg_statistic
Definition: pg_statistic.h:135
static bool DatumGetBool(Datum X)
Definition: postgres.h:90
#define InvalidOid
Definition: postgres_ext.h:36
@ IS_NOT_TRUE
Definition: primnodes.h:1739
@ IS_NOT_FALSE
Definition: primnodes.h:1739
@ IS_NOT_UNKNOWN
Definition: primnodes.h:1739
@ IS_TRUE
Definition: primnodes.h:1739
@ IS_UNKNOWN
Definition: primnodes.h:1739
@ IS_FALSE
Definition: primnodes.h:1739
void examine_variable(PlannerInfo *root, Node *node, int varRelid, VariableStatData *vardata)
Definition: selfuncs.c:5012
#define DEFAULT_NOT_UNK_SEL
Definition: selfuncs.h:56
#define ReleaseVariableStats(vardata)
Definition: selfuncs.h:99
#define CLAMP_PROBABILITY(p)
Definition: selfuncs.h:63
#define DEFAULT_UNK_SEL
Definition: selfuncs.h:55
Datum * values
Definition: lsyscache.h:53
float4 * numbers
Definition: lsyscache.h:56
int nnumbers
Definition: lsyscache.h:57
HeapTuple statsTuple
Definition: selfuncs.h:89

References arg, ATTSTATSSLOT_NUMBERS, ATTSTATSSLOT_VALUES, CLAMP_PROBABILITY, clause_selectivity(), DatumGetBool(), DEFAULT_NOT_UNK_SEL, DEFAULT_UNK_SEL, elog, ERROR, examine_variable(), free_attstatsslot(), get_attstatsslot(), GETSTRUCT, HeapTupleIsValid, InvalidOid, IS_FALSE, IS_NOT_FALSE, IS_NOT_TRUE, IS_NOT_UNKNOWN, IS_TRUE, IS_UNKNOWN, AttStatsSlot::nnumbers, AttStatsSlot::numbers, ReleaseVariableStats, VariableStatData::statsTuple, and AttStatsSlot::values.

Referenced by clause_selectivity_ext().

◆ boolvarsel()

Selectivity boolvarsel ( PlannerInfo root,
Node arg,
int  varRelid 
)

Definition at line 1512 of file selfuncs.c.

1513 {
1514  VariableStatData vardata;
1515  double selec;
1516 
1517  examine_variable(root, arg, varRelid, &vardata);
1518  if (HeapTupleIsValid(vardata.statsTuple))
1519  {
1520  /*
1521  * A boolean variable V is equivalent to the clause V = 't', so we
1522  * compute the selectivity as if that is what we have.
1523  */
1524  selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1525  BoolGetDatum(true), false, true, false);
1526  }
1527  else
1528  {
1529  /* Otherwise, the default estimate is 0.5 */
1530  selec = 0.5;
1531  }
1532  ReleaseVariableStats(vardata);
1533  return selec;
1534 }
static Datum BoolGetDatum(bool X)
Definition: postgres.h:102
double var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation, Datum constval, bool constisnull, bool varonleft, bool negate)
Definition: selfuncs.c:295

References arg, BoolGetDatum(), examine_variable(), HeapTupleIsValid, InvalidOid, ReleaseVariableStats, VariableStatData::statsTuple, and var_eq_const().

Referenced by clause_selectivity_ext().

◆ brincostestimate()

void brincostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
Cost indexStartupCost,
Cost indexTotalCost,
Selectivity indexSelectivity,
double *  indexCorrelation,
double *  indexPages 
)

Definition at line 7928 of file selfuncs.c.

7932 {
7933  IndexOptInfo *index = path->indexinfo;
7934  List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7935  double numPages = index->pages;
7936  RelOptInfo *baserel = index->rel;
7937  RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
7938  Cost spc_seq_page_cost;
7939  Cost spc_random_page_cost;
7940  double qual_arg_cost;
7941  double qualSelectivity;
7942  BrinStatsData statsData;
7943  double indexRanges;
7944  double minimalRanges;
7945  double estimatedRanges;
7946  double selec;
7947  Relation indexRel;
7948  ListCell *l;
7949  VariableStatData vardata;
7950 
7951  Assert(rte->rtekind == RTE_RELATION);
7952 
7953  /* fetch estimated page cost for the tablespace containing the index */
7954  get_tablespace_page_costs(index->reltablespace,
7955  &spc_random_page_cost,
7956  &spc_seq_page_cost);
7957 
7958  /*
7959  * Obtain some data from the index itself, if possible. Otherwise invent
7960  * some plausible internal statistics based on the relation page count.
7961  */
7962  if (!index->hypothetical)
7963  {
7964  /*
7965  * A lock should have already been obtained on the index in plancat.c.
7966  */
7967  indexRel = index_open(index->indexoid, NoLock);
7968  brinGetStats(indexRel, &statsData);
7969  index_close(indexRel, NoLock);
7970 
7971  /* work out the actual number of ranges in the index */
7972  indexRanges = Max(ceil((double) baserel->pages /
7973  statsData.pagesPerRange), 1.0);
7974  }
7975  else
7976  {
7977  /*
7978  * Assume default number of pages per range, and estimate the number
7979  * of ranges based on that.
7980  */
7981  indexRanges = Max(ceil((double) baserel->pages /
7983 
7985  statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
7986  }
7987 
7988  /*
7989  * Compute index correlation
7990  *
7991  * Because we can use all index quals equally when scanning, we can use
7992  * the largest correlation (in absolute value) among columns used by the
7993  * query. Start at zero, the worst possible case. If we cannot find any
7994  * correlation statistics, we will keep it as 0.
7995  */
7996  *indexCorrelation = 0;
7997 
7998  foreach(l, path->indexclauses)
7999  {
8000  IndexClause *iclause = lfirst_node(IndexClause, l);
8001  AttrNumber attnum = index->indexkeys[iclause->indexcol];
8002 
8003  /* attempt to lookup stats in relation for this index column */
8004  if (attnum != 0)
8005  {
8006  /* Simple variable -- look to stats for the underlying table */
8008  (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8009  {
8010  /*
8011  * The hook took control of acquiring a stats tuple. If it
8012  * did supply a tuple, it'd better have supplied a freefunc.
8013  */
8014  if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8015  elog(ERROR,
8016  "no function provided to release variable stats with");
8017  }
8018  else
8019  {
8020  vardata.statsTuple =
8021  SearchSysCache3(STATRELATTINH,
8022  ObjectIdGetDatum(rte->relid),
8024  BoolGetDatum(false));
8025  vardata.freefunc = ReleaseSysCache;
8026  }
8027  }
8028  else
8029  {
8030  /*
8031  * Looks like we've found an expression column in the index. Let's
8032  * see if there's any stats for it.
8033  */
8034 
8035  /* get the attnum from the 0-based index. */
8036  attnum = iclause->indexcol + 1;
8037 
8038  if (get_index_stats_hook &&
8039  (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8040  {
8041  /*
8042  * The hook took control of acquiring a stats tuple. If it
8043  * did supply a tuple, it'd better have supplied a freefunc.
8044  */
8045  if (HeapTupleIsValid(vardata.statsTuple) &&
8046  !vardata.freefunc)
8047  elog(ERROR, "no function provided to release variable stats with");
8048  }
8049  else
8050  {
8051  vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8052  ObjectIdGetDatum(index->indexoid),
8054  BoolGetDatum(false));
8055  vardata.freefunc = ReleaseSysCache;
8056  }
8057  }
8058 
8059  if (HeapTupleIsValid(vardata.statsTuple))
8060  {
8061  AttStatsSlot sslot;
8062 
8063  if (get_attstatsslot(&sslot, vardata.statsTuple,
8064  STATISTIC_KIND_CORRELATION, InvalidOid,
8066  {
8067  double varCorrelation = 0.0;
8068 
8069  if (sslot.nnumbers > 0)
8070  varCorrelation = fabs(sslot.numbers[0]);
8071 
8072  if (varCorrelation > *indexCorrelation)
8073  *indexCorrelation = varCorrelation;
8074 
8075  free_attstatsslot(&sslot);
8076  }
8077  }
8078 
8079  ReleaseVariableStats(vardata);
8080  }
8081 
8082  qualSelectivity = clauselist_selectivity(root, indexQuals,
8083  baserel->relid,
8084  JOIN_INNER, NULL);
8085 
8086  /*
8087  * Now calculate the minimum possible ranges we could match with if all of
8088  * the rows were in the perfect order in the table's heap.
8089  */
8090  minimalRanges = ceil(indexRanges * qualSelectivity);
8091 
8092  /*
8093  * Now estimate the number of ranges that we'll touch by using the
8094  * indexCorrelation from the stats. Careful not to divide by zero (note
8095  * we're using the absolute value of the correlation).
8096  */
8097  if (*indexCorrelation < 1.0e-10)
8098  estimatedRanges = indexRanges;
8099  else
8100  estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8101 
8102  /* we expect to visit this portion of the table */
8103  selec = estimatedRanges / indexRanges;
8104 
8105  CLAMP_PROBABILITY(selec);
8106 
8107  *indexSelectivity = selec;
8108 
8109  /*
8110  * Compute the index qual costs, much as in genericcostestimate, to add to
8111  * the index costs. We can disregard indexorderbys, since BRIN doesn't
8112  * support those.
8113  */
8114  qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8115 
8116  /*
8117  * Compute the startup cost as the cost to read the whole revmap
8118  * sequentially, including the cost to execute the index quals.
8119  */
8120  *indexStartupCost =
8121  spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8122  *indexStartupCost += qual_arg_cost;
8123 
8124  /*
8125  * To read a BRIN index there might be a bit of back and forth over
8126  * regular pages, as revmap might point to them out of sequential order;
8127  * calculate the total cost as reading the whole index in random order.
8128  */
8129  *indexTotalCost = *indexStartupCost +
8130  spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8131 
8132  /*
8133  * Charge a small amount per range tuple which we expect to match to. This
8134  * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8135  * will set a bit for each page in the range when we find a matching
8136  * range, so we must multiply the charge by the number of pages in the
8137  * range.
8138  */
8139  *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8140  statsData.pagesPerRange;
8141 
8142  *indexPages = index->pages;
8143 }
int16 AttrNumber
Definition: attnum.h:21
void brinGetStats(Relation index, BrinStatsData *stats)
Definition: brin.c:1634
#define BRIN_DEFAULT_PAGES_PER_RANGE
Definition: brin.h:39
#define REVMAP_PAGE_MAXITEMS
Definition: brin_page.h:93
#define Min(x, y)
Definition: c.h:991
#define Max(x, y)
Definition: c.h:985
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:100
double cpu_operator_cost
Definition: costsize.c:123
void index_close(Relation relation, LOCKMODE lockmode)
Definition: indexam.c:177
Relation index_open(Oid relationId, LOCKMODE lockmode)
Definition: indexam.c:133
Assert(fmt[strlen(fmt) - 1] !='\n')
#define NoLock
Definition: lockdefs.h:34
double Cost
Definition: nodes.h:241
@ JOIN_INNER
Definition: nodes.h:283
@ RTE_RELATION
Definition: parsenodes.h:1011
#define planner_rt_fetch(rti, root)
Definition: pathnodes.h:555
int16 attnum
Definition: pg_attribute.h:74
#define lfirst_node(type, lc)
Definition: pg_list.h:176
static Datum Int16GetDatum(int16 X)
Definition: postgres.h:172
static Datum ObjectIdGetDatum(Oid X)
Definition: postgres.h:252
List * get_quals_from_indexclauses(List *indexclauses)
Definition: selfuncs.c:6460
get_index_stats_hook_type get_index_stats_hook
Definition: selfuncs.c:148
Cost index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
Definition: selfuncs.c:6490
get_relation_stats_hook_type get_relation_stats_hook
Definition: selfuncs.c:147
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
BlockNumber revmapNumPages
Definition: brin.h:35
BlockNumber pagesPerRange
Definition: brin.h:34
AttrNumber indexcol
Definition: pathnodes.h:1741
List * indexclauses
Definition: pathnodes.h:1691
IndexOptInfo * indexinfo
Definition: pathnodes.h:1690
RTEKind rtekind
Definition: parsenodes.h:1030
Index relid
Definition: pathnodes.h:903
BlockNumber pages
Definition: pathnodes.h:929
void(* freefunc)(HeapTuple tuple)
Definition: selfuncs.h:91
void ReleaseSysCache(HeapTuple tuple)
Definition: syscache.c:266
HeapTuple SearchSysCache3(int cacheId, Datum key1, Datum key2, Datum key3)
Definition: syscache.c:240

References Assert(), attnum, ATTSTATSSLOT_NUMBERS, BoolGetDatum(), BRIN_DEFAULT_PAGES_PER_RANGE, brinGetStats(), CLAMP_PROBABILITY, clauselist_selectivity(), cpu_operator_cost, elog, ERROR, free_attstatsslot(), VariableStatData::freefunc, get_attstatsslot(), get_index_stats_hook, get_quals_from_indexclauses(), get_relation_stats_hook, get_tablespace_page_costs(), HeapTupleIsValid, index_close(), index_open(), index_other_operands_eval_cost(), IndexPath::indexclauses, IndexClause::indexcol, IndexPath::indexinfo, Int16GetDatum(), InvalidOid, JOIN_INNER, lfirst_node, Max, Min, AttStatsSlot::nnumbers, NoLock, AttStatsSlot::numbers, ObjectIdGetDatum(), RelOptInfo::pages, BrinStatsData::pagesPerRange, planner_rt_fetch, ReleaseSysCache(), ReleaseVariableStats, RangeTblEntry::relid, RelOptInfo::relid, REVMAP_PAGE_MAXITEMS, BrinStatsData::revmapNumPages, RTE_RELATION, RangeTblEntry::rtekind, SearchSysCache3(), and VariableStatData::statsTuple.

Referenced by brinhandler().

◆ btcostestimate()

void btcostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
Cost indexStartupCost,
Cost indexTotalCost,
Selectivity indexSelectivity,
double *  indexCorrelation,
double *  indexPages 
)

Definition at line 6783 of file selfuncs.c.

6787 {
6788  IndexOptInfo *index = path->indexinfo;
6789  GenericCosts costs = {0};
6790  Oid relid;
6791  AttrNumber colnum;
6792  VariableStatData vardata = {0};
6793  double numIndexTuples;
6794  Cost descentCost;
6795  List *indexBoundQuals;
6796  int indexcol;
6797  bool eqQualHere;
6798  bool found_saop;
6799  bool found_is_null_op;
6800  double num_sa_scans;
6801  ListCell *lc;
6802 
6803  /*
6804  * For a btree scan, only leading '=' quals plus inequality quals for the
6805  * immediately next attribute contribute to index selectivity (these are
6806  * the "boundary quals" that determine the starting and stopping points of
6807  * the index scan). Additional quals can suppress visits to the heap, so
6808  * it's OK to count them in indexSelectivity, but they should not count
6809  * for estimating numIndexTuples. So we must examine the given indexquals
6810  * to find out which ones count as boundary quals. We rely on the
6811  * knowledge that they are given in index column order.
6812  *
6813  * For a RowCompareExpr, we consider only the first column, just as
6814  * rowcomparesel() does.
6815  *
6816  * If there's a ScalarArrayOpExpr in the quals, we'll actually perform N
6817  * index scans not one, but the ScalarArrayOpExpr's operator can be
6818  * considered to act the same as it normally does.
6819  */
6820  indexBoundQuals = NIL;
6821  indexcol = 0;
6822  eqQualHere = false;
6823  found_saop = false;
6824  found_is_null_op = false;
6825  num_sa_scans = 1;
6826  foreach(lc, path->indexclauses)
6827  {
6828  IndexClause *iclause = lfirst_node(IndexClause, lc);
6829  ListCell *lc2;
6830 
6831  if (indexcol != iclause->indexcol)
6832  {
6833  /* Beginning of a new column's quals */
6834  if (!eqQualHere)
6835  break; /* done if no '=' qual for indexcol */
6836  eqQualHere = false;
6837  indexcol++;
6838  if (indexcol != iclause->indexcol)
6839  break; /* no quals at all for indexcol */
6840  }
6841 
6842  /* Examine each indexqual associated with this index clause */
6843  foreach(lc2, iclause->indexquals)
6844  {
6845  RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6846  Expr *clause = rinfo->clause;
6847  Oid clause_op = InvalidOid;
6848  int op_strategy;
6849 
6850  if (IsA(clause, OpExpr))
6851  {
6852  OpExpr *op = (OpExpr *) clause;
6853 
6854  clause_op = op->opno;
6855  }
6856  else if (IsA(clause, RowCompareExpr))
6857  {
6858  RowCompareExpr *rc = (RowCompareExpr *) clause;
6859 
6860  clause_op = linitial_oid(rc->opnos);
6861  }
6862  else if (IsA(clause, ScalarArrayOpExpr))
6863  {
6864  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6865  Node *other_operand = (Node *) lsecond(saop->args);
6866  double alength = estimate_array_length(root, other_operand);
6867 
6868  clause_op = saop->opno;
6869  found_saop = true;
6870  /* count number of SA scans induced by indexBoundQuals only */
6871  if (alength > 1)
6872  num_sa_scans *= alength;
6873  }
6874  else if (IsA(clause, NullTest))
6875  {
6876  NullTest *nt = (NullTest *) clause;
6877 
6878  if (nt->nulltesttype == IS_NULL)
6879  {
6880  found_is_null_op = true;
6881  /* IS NULL is like = for selectivity purposes */
6882  eqQualHere = true;
6883  }
6884  }
6885  else
6886  elog(ERROR, "unsupported indexqual type: %d",
6887  (int) nodeTag(clause));
6888 
6889  /* check for equality operator */
6890  if (OidIsValid(clause_op))
6891  {
6892  op_strategy = get_op_opfamily_strategy(clause_op,
6893  index->opfamily[indexcol]);
6894  Assert(op_strategy != 0); /* not a member of opfamily?? */
6895  if (op_strategy == BTEqualStrategyNumber)
6896  eqQualHere = true;
6897  }
6898 
6899  indexBoundQuals = lappend(indexBoundQuals, rinfo);
6900  }
6901  }
6902 
6903  /*
6904  * If index is unique and we found an '=' clause for each column, we can
6905  * just assume numIndexTuples = 1 and skip the expensive
6906  * clauselist_selectivity calculations. However, a ScalarArrayOp or
6907  * NullTest invalidates that theory, even though it sets eqQualHere.
6908  */
6909  if (index->unique &&
6910  indexcol == index->nkeycolumns - 1 &&
6911  eqQualHere &&
6912  !found_saop &&
6913  !found_is_null_op)
6914  numIndexTuples = 1.0;
6915  else
6916  {
6917  List *selectivityQuals;
6918  Selectivity btreeSelectivity;
6919 
6920  /*
6921  * If the index is partial, AND the index predicate with the
6922  * index-bound quals to produce a more accurate idea of the number of
6923  * rows covered by the bound conditions.
6924  */
6925  selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
6926 
6927  btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
6928  index->rel->relid,
6929  JOIN_INNER,
6930  NULL);
6931  numIndexTuples = btreeSelectivity * index->rel->tuples;
6932 
6933  /*
6934  * As in genericcostestimate(), we have to adjust for any
6935  * ScalarArrayOpExpr quals included in indexBoundQuals, and then round
6936  * to integer.
6937  */
6938  numIndexTuples = rint(numIndexTuples / num_sa_scans);
6939  }
6940 
6941  /*
6942  * Now do generic index cost estimation.
6943  */
6944  costs.numIndexTuples = numIndexTuples;
6945 
6946  genericcostestimate(root, path, loop_count, &costs);
6947 
6948  /*
6949  * Add a CPU-cost component to represent the costs of initial btree
6950  * descent. We don't charge any I/O cost for touching upper btree levels,
6951  * since they tend to stay in cache, but we still have to do about log2(N)
6952  * comparisons to descend a btree of N leaf tuples. We charge one
6953  * cpu_operator_cost per comparison.
6954  *
6955  * If there are ScalarArrayOpExprs, charge this once per SA scan. The
6956  * ones after the first one are not startup cost so far as the overall
6957  * plan is concerned, so add them only to "total" cost.
6958  */
6959  if (index->tuples > 1) /* avoid computing log(0) */
6960  {
6961  descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
6962  costs.indexStartupCost += descentCost;
6963  costs.indexTotalCost += costs.num_sa_scans * descentCost;
6964  }
6965 
6966  /*
6967  * Even though we're not charging I/O cost for touching upper btree pages,
6968  * it's still reasonable to charge some CPU cost per page descended
6969  * through. Moreover, if we had no such charge at all, bloated indexes
6970  * would appear to have the same search cost as unbloated ones, at least
6971  * in cases where only a single leaf page is expected to be visited. This
6972  * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
6973  * touched. The number of such pages is btree tree height plus one (ie,
6974  * we charge for the leaf page too). As above, charge once per SA scan.
6975  */
6976  descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
6977  costs.indexStartupCost += descentCost;
6978  costs.indexTotalCost += costs.num_sa_scans * descentCost;
6979 
6980  /*
6981  * If we can get an estimate of the first column's ordering correlation C
6982  * from pg_statistic, estimate the index correlation as C for a
6983  * single-column index, or C * 0.75 for multiple columns. (The idea here
6984  * is that multiple columns dilute the importance of the first column's
6985  * ordering, but don't negate it entirely. Before 8.0 we divided the
6986  * correlation by the number of columns, but that seems too strong.)
6987  */
6988  if (index->indexkeys[0] != 0)
6989  {
6990  /* Simple variable --- look to stats for the underlying table */
6991  RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6992 
6993  Assert(rte->rtekind == RTE_RELATION);
6994  relid = rte->relid;
6995  Assert(relid != InvalidOid);
6996  colnum = index->indexkeys[0];
6997 
6999  (*get_relation_stats_hook) (root, rte, colnum, &vardata))
7000  {
7001  /*
7002  * The hook took control of acquiring a stats tuple. If it did
7003  * supply a tuple, it'd better have supplied a freefunc.
7004  */
7005  if (HeapTupleIsValid(vardata.statsTuple) &&
7006  !vardata.freefunc)
7007  elog(ERROR, "no function provided to release variable stats with");
7008  }
7009  else
7010  {
7011  vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7012  ObjectIdGetDatum(relid),
7013  Int16GetDatum(colnum),
7014  BoolGetDatum(rte->inh));
7015  vardata.freefunc = ReleaseSysCache;
7016  }
7017  }
7018  else
7019  {
7020  /* Expression --- maybe there are stats for the index itself */
7021  relid = index->indexoid;
7022  colnum = 1;
7023 
7024  if (get_index_stats_hook &&
7025  (*get_index_stats_hook) (root, relid, colnum, &vardata))
7026  {
7027  /*
7028  * The hook took control of acquiring a stats tuple. If it did
7029  * supply a tuple, it'd better have supplied a freefunc.
7030  */
7031  if (HeapTupleIsValid(vardata.statsTuple) &&
7032  !vardata.freefunc)
7033  elog(ERROR, "no function provided to release variable stats with");
7034  }
7035  else
7036  {
7037  vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7038  ObjectIdGetDatum(relid),
7039  Int16GetDatum(colnum),
7040  BoolGetDatum(false));
7041  vardata.freefunc = ReleaseSysCache;
7042  }
7043  }
7044 
7045  if (HeapTupleIsValid(vardata.statsTuple))
7046  {
7047  Oid sortop;
7048  AttStatsSlot sslot;
7049 
7050  sortop = get_opfamily_member(index->opfamily[0],
7051  index->opcintype[0],
7052  index->opcintype[0],
7054  if (OidIsValid(sortop) &&
7055  get_attstatsslot(&sslot, vardata.statsTuple,
7056  STATISTIC_KIND_CORRELATION, sortop,
7058  {
7059  double varCorrelation;
7060 
7061  Assert(sslot.nnumbers == 1);
7062  varCorrelation = sslot.numbers[0];
7063 
7064  if (index->reverse_sort[0])
7065  varCorrelation = -varCorrelation;
7066 
7067  if (index->nkeycolumns > 1)
7068  costs.indexCorrelation = varCorrelation * 0.75;
7069  else
7070  costs.indexCorrelation = varCorrelation;
7071 
7072  free_attstatsslot(&sslot);
7073  }
7074  }
7075 
7076  ReleaseVariableStats(vardata);
7077 
7078  *indexStartupCost = costs.indexStartupCost;
7079  *indexTotalCost = costs.indexTotalCost;
7080  *indexSelectivity = costs.indexSelectivity;
7081  *indexCorrelation = costs.indexCorrelation;
7082  *indexPages = costs.numIndexPages;
7083 }
#define OidIsValid(objectId)
Definition: c.h:762
int get_op_opfamily_strategy(Oid opno, Oid opfamily)
Definition: lsyscache.c:83
Oid get_opfamily_member(Oid opfamily, Oid lefttype, Oid righttype, int16 strategy)
Definition: lsyscache.c:166
#define IsA(nodeptr, _type_)
Definition: nodes.h:158
#define nodeTag(nodeptr)
Definition: nodes.h:133
#define lsecond(l)
Definition: pg_list.h:183
#define linitial_oid(l)
Definition: pg_list.h:180
unsigned int Oid
Definition: postgres_ext.h:31
@ IS_NULL
Definition: primnodes.h:1715
#define DEFAULT_PAGE_CPU_MULTIPLIER
Definition: selfuncs.c:144
double estimate_array_length(PlannerInfo *root, Node *arrayexpr)
Definition: selfuncs.c:2136
void genericcostestimate(PlannerInfo *root, IndexPath *path, double loop_count, GenericCosts *costs)
Definition: selfuncs.c:6544
List * add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
Definition: selfuncs.c:6762
#define BTLessStrategyNumber
Definition: stratnum.h:29
#define BTEqualStrategyNumber
Definition: stratnum.h:31
Selectivity indexSelectivity
Definition: selfuncs.h:124
Cost indexStartupCost
Definition: selfuncs.h:122
double indexCorrelation
Definition: selfuncs.h:125
double num_sa_scans
Definition: selfuncs.h:131
Cost indexTotalCost
Definition: selfuncs.h:123
double numIndexPages
Definition: selfuncs.h:128
double numIndexTuples
Definition: selfuncs.h:129
List * indexquals
Definition: pathnodes.h:1739
NullTestType nulltesttype
Definition: primnodes.h:1722
Oid opno
Definition: primnodes.h:774
Expr * clause
Definition: pathnodes.h:2541

References add_predicate_to_index_quals(), ScalarArrayOpExpr::args, Assert(), ATTSTATSSLOT_NUMBERS, BoolGetDatum(), BTEqualStrategyNumber, BTLessStrategyNumber, RestrictInfo::clause, clauselist_selectivity(), cpu_operator_cost, DEFAULT_PAGE_CPU_MULTIPLIER, elog, ERROR, estimate_array_length(), free_attstatsslot(), VariableStatData::freefunc, genericcostestimate(), get_attstatsslot(), get_index_stats_hook, get_op_opfamily_strategy(), get_opfamily_member(), get_relation_stats_hook, HeapTupleIsValid, IndexPath::indexclauses, IndexClause::indexcol, GenericCosts::indexCorrelation, IndexPath::indexinfo, IndexClause::indexquals, GenericCosts::indexSelectivity, GenericCosts::indexStartupCost, GenericCosts::indexTotalCost, RangeTblEntry::inh, Int16GetDatum(), InvalidOid, IS_NULL, IsA, JOIN_INNER, lappend(), lfirst_node, linitial_oid, lsecond, NIL, AttStatsSlot::nnumbers, nodeTag, NullTest::nulltesttype, GenericCosts::num_sa_scans, AttStatsSlot::numbers, GenericCosts::numIndexPages, GenericCosts::numIndexTuples, ObjectIdGetDatum(), OidIsValid, OpExpr::opno, ScalarArrayOpExpr::opno, planner_rt_fetch, ReleaseSysCache(), ReleaseVariableStats, RangeTblEntry::relid, RTE_RELATION, RangeTblEntry::rtekind, SearchSysCache3(), and VariableStatData::statsTuple.

Referenced by bthandler().

◆ convert_bytea_to_scalar()

static void convert_bytea_to_scalar ( Datum  value,
double *  scaledvalue,
Datum  lobound,
double *  scaledlobound,
Datum  hibound,
double *  scaledhibound 
)
static

Definition at line 4726 of file selfuncs.c.

4732 {
4733  bytea *valuep = DatumGetByteaPP(value);
4734  bytea *loboundp = DatumGetByteaPP(lobound);
4735  bytea *hiboundp = DatumGetByteaPP(hibound);
4736  int rangelo,
4737  rangehi,
4738  valuelen = VARSIZE_ANY_EXHDR(valuep),
4739  loboundlen = VARSIZE_ANY_EXHDR(loboundp),
4740  hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
4741  i,
4742  minlen;
4743  unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
4744  unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
4745  unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
4746 
4747  /*
4748  * Assume bytea data is uniformly distributed across all byte values.
4749  */
4750  rangelo = 0;
4751  rangehi = 255;
4752 
4753  /*
4754  * Now strip any common prefix of the three strings.
4755  */
4756  minlen = Min(Min(valuelen, loboundlen), hiboundlen);
4757  for (i = 0; i < minlen; i++)
4758  {
4759  if (*lostr != *histr || *lostr != *valstr)
4760  break;
4761  lostr++, histr++, valstr++;
4762  loboundlen--, hiboundlen--, valuelen--;
4763  }
4764 
4765  /*
4766  * Now we can do the conversions.
4767  */
4768  *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
4769  *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
4770  *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
4771 }
#define DatumGetByteaPP(X)
Definition: fmgr.h:291
static struct @150 value
int i
Definition: isn.c:73
static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen, int rangelo, int rangehi)
Definition: selfuncs.c:4774
Definition: c.h:674
#define VARDATA_ANY(PTR)
Definition: varatt.h:324
#define VARSIZE_ANY_EXHDR(PTR)
Definition: varatt.h:317

References convert_one_bytea_to_scalar(), DatumGetByteaPP, i, Min, value, VARDATA_ANY, and VARSIZE_ANY_EXHDR.

Referenced by convert_to_scalar().

◆ convert_numeric_to_scalar()

static double convert_numeric_to_scalar ( Datum  value,
Oid  typid,
bool failure 
)
static

Definition at line 4452 of file selfuncs.c.

4453 {
4454  switch (typid)
4455  {
4456  case BOOLOID:
4457  return (double) DatumGetBool(value);
4458  case INT2OID:
4459  return (double) DatumGetInt16(value);
4460  case INT4OID:
4461  return (double) DatumGetInt32(value);
4462  case INT8OID:
4463  return (double) DatumGetInt64(value);
4464  case FLOAT4OID:
4465  return (double) DatumGetFloat4(value);
4466  case FLOAT8OID:
4467  return (double) DatumGetFloat8(value);
4468  case NUMERICOID:
4469  /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4470  return (double)
4472  value));
4473  case OIDOID:
4474  case REGPROCOID:
4475  case REGPROCEDUREOID:
4476  case REGOPEROID:
4477  case REGOPERATOROID:
4478  case REGCLASSOID:
4479  case REGTYPEOID:
4480  case REGCOLLATIONOID:
4481  case REGCONFIGOID:
4482  case REGDICTIONARYOID:
4483  case REGROLEOID:
4484  case REGNAMESPACEOID:
4485  /* we can treat OIDs as integers... */
4486  return (double) DatumGetObjectId(value);
4487  }
4488 
4489  *failure = true;
4490  return 0;
4491 }
Datum numeric_float8_no_overflow(PG_FUNCTION_ARGS)
Definition: numeric.c:4609
#define DirectFunctionCall1(func, arg1)
Definition: fmgr.h:642
static int64 DatumGetInt64(Datum X)
Definition: postgres.h:385
static float4 DatumGetFloat4(Datum X)
Definition: postgres.h:458
static Oid DatumGetObjectId(Datum X)
Definition: postgres.h:242
static float8 DatumGetFloat8(Datum X)
Definition: postgres.h:494
static int16 DatumGetInt16(Datum X)
Definition: postgres.h:162
static int32 DatumGetInt32(Datum X)
Definition: postgres.h:202

References DatumGetBool(), DatumGetFloat4(), DatumGetFloat8(), DatumGetInt16(), DatumGetInt32(), DatumGetInt64(), DatumGetObjectId(), DirectFunctionCall1, numeric_float8_no_overflow(), and value.

Referenced by convert_to_scalar().

◆ convert_one_bytea_to_scalar()

static double convert_one_bytea_to_scalar ( unsigned char *  value,
int  valuelen,
int  rangelo,
int  rangehi 
)
static

Definition at line 4774 of file selfuncs.c.

4776 {
4777  double num,
4778  denom,
4779  base;
4780 
4781  if (valuelen <= 0)
4782  return 0.0; /* empty string has scalar value 0 */
4783 
4784  /*
4785  * Since base is 256, need not consider more than about 10 chars (even
4786  * this many seems like overkill)
4787  */
4788  if (valuelen > 10)
4789  valuelen = 10;
4790 
4791  /* Convert initial characters to fraction */
4792  base = rangehi - rangelo + 1;
4793  num = 0.0;
4794  denom = base;
4795  while (valuelen-- > 0)
4796  {
4797  int ch = *value++;
4798 
4799  if (ch < rangelo)
4800  ch = rangelo - 1;
4801  else if (ch > rangehi)
4802  ch = rangehi + 1;
4803  num += ((double) (ch - rangelo)) / denom;
4804  denom *= base;
4805  }
4806 
4807  return num;
4808 }

References value.

Referenced by convert_bytea_to_scalar().

◆ convert_one_string_to_scalar()

static double convert_one_string_to_scalar ( char *  value,
int  rangelo,
int  rangehi 
)
static

Definition at line 4594 of file selfuncs.c.

4595 {
4596  int slen = strlen(value);
4597  double num,
4598  denom,
4599  base;
4600 
4601  if (slen <= 0)
4602  return 0.0; /* empty string has scalar value 0 */
4603 
4604  /*
4605  * There seems little point in considering more than a dozen bytes from
4606  * the string. Since base is at least 10, that will give us nominal
4607  * resolution of at least 12 decimal digits, which is surely far more
4608  * precision than this estimation technique has got anyway (especially in
4609  * non-C locales). Also, even with the maximum possible base of 256, this
4610  * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4611  * overflow on any known machine.
4612  */
4613  if (slen > 12)
4614  slen = 12;
4615 
4616  /* Convert initial characters to fraction */
4617  base = rangehi - rangelo + 1;
4618  num = 0.0;
4619  denom = base;
4620  while (slen-- > 0)
4621  {
4622  int ch = (unsigned char) *value++;
4623 
4624  if (ch < rangelo)
4625  ch = rangelo - 1;
4626  else if (ch > rangehi)
4627  ch = rangehi + 1;
4628  num += ((double) (ch - rangelo)) / denom;
4629  denom *= base;
4630  }
4631 
4632  return num;
4633 }

References value.

Referenced by convert_string_to_scalar().

◆ convert_string_datum()

static char * convert_string_datum ( Datum  value,
Oid  typid,
Oid  collid,
bool failure 
)
static

Definition at line 4645 of file selfuncs.c.

4646 {
4647  char *val;
4648 
4649  switch (typid)
4650  {
4651  case CHAROID:
4652  val = (char *) palloc(2);
4653  val[0] = DatumGetChar(value);
4654  val[1] = '\0';
4655  break;
4656  case BPCHAROID:
4657  case VARCHAROID:
4658  case TEXTOID:
4660  break;
4661  case NAMEOID:
4662  {
4664 
4665  val = pstrdup(NameStr(*nm));
4666  break;
4667  }
4668  default:
4669  *failure = true;
4670  return NULL;
4671  }
4672 
4673  if (!lc_collate_is_c(collid))
4674  {
4675  char *xfrmstr;
4676  size_t xfrmlen;
4677  size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4678 
4679  /*
4680  * XXX: We could guess at a suitable output buffer size and only call
4681  * strxfrm twice if our guess is too small.
4682  *
4683  * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4684  * bogus data or set an error. This is not really a problem unless it
4685  * crashes since it will only give an estimation error and nothing
4686  * fatal.
4687  */
4688  xfrmlen = strxfrm(NULL, val, 0);
4689 #ifdef WIN32
4690 
4691  /*
4692  * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4693  * of trying to allocate this much memory (and fail), just return the
4694  * original string unmodified as if we were in the C locale.
4695  */
4696  if (xfrmlen == INT_MAX)
4697  return val;
4698 #endif
4699  xfrmstr = (char *) palloc(xfrmlen + 1);
4700  xfrmlen2 = strxfrm(xfrmstr, val, xfrmlen + 1);
4701 
4702  /*
4703  * Some systems (e.g., glibc) can return a smaller value from the
4704  * second call than the first; thus the Assert must be <= not ==.
4705  */
4706  Assert(xfrmlen2 <= xfrmlen);
4707  pfree(val);
4708  val = xfrmstr;
4709  }
4710 
4711  return val;
4712 }
#define TextDatumGetCString(d)
Definition: builtins.h:98
#define NameStr(name)
Definition: c.h:733
#define PG_USED_FOR_ASSERTS_ONLY
Definition: c.h:169
Oid collid
long val
Definition: informix.c:664
char * pstrdup(const char *in)
Definition: mcxt.c:1683
void pfree(void *pointer)
Definition: mcxt.c:1508
bool lc_collate_is_c(Oid collation)
Definition: pg_locale.c:1311
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:312
static char DatumGetChar(Datum X)
Definition: postgres.h:112
Definition: c.h:728

References Assert(), collid, DatumGetChar(), DatumGetPointer(), lc_collate_is_c(), NameStr, palloc(), pfree(), PG_USED_FOR_ASSERTS_ONLY, pstrdup(), TextDatumGetCString, val, and value.

Referenced by convert_to_scalar().

◆ convert_string_to_scalar()

static void convert_string_to_scalar ( char *  value,
double *  scaledvalue,
char *  lobound,
double *  scaledlobound,
char *  hibound,
double *  scaledhibound 
)
static

Definition at line 4514 of file selfuncs.c.

4520 {
4521  int rangelo,
4522  rangehi;
4523  char *sptr;
4524 
4525  rangelo = rangehi = (unsigned char) hibound[0];
4526  for (sptr = lobound; *sptr; sptr++)
4527  {
4528  if (rangelo > (unsigned char) *sptr)
4529  rangelo = (unsigned char) *sptr;
4530  if (rangehi < (unsigned char) *sptr)
4531  rangehi = (unsigned char) *sptr;
4532  }
4533  for (sptr = hibound; *sptr; sptr++)
4534  {
4535  if (rangelo > (unsigned char) *sptr)
4536  rangelo = (unsigned char) *sptr;
4537  if (rangehi < (unsigned char) *sptr)
4538  rangehi = (unsigned char) *sptr;
4539  }
4540  /* If range includes any upper-case ASCII chars, make it include all */
4541  if (rangelo <= 'Z' && rangehi >= 'A')
4542  {
4543  if (rangelo > 'A')
4544  rangelo = 'A';
4545  if (rangehi < 'Z')
4546  rangehi = 'Z';
4547  }
4548  /* Ditto lower-case */
4549  if (rangelo <= 'z' && rangehi >= 'a')
4550  {
4551  if (rangelo > 'a')
4552  rangelo = 'a';
4553  if (rangehi < 'z')
4554  rangehi = 'z';
4555  }
4556  /* Ditto digits */
4557  if (rangelo <= '9' && rangehi >= '0')
4558  {
4559  if (rangelo > '0')
4560  rangelo = '0';
4561  if (rangehi < '9')
4562  rangehi = '9';
4563  }
4564 
4565  /*
4566  * If range includes less than 10 chars, assume we have not got enough
4567  * data, and make it include regular ASCII set.
4568  */
4569  if (rangehi - rangelo < 9)
4570  {
4571  rangelo = ' ';
4572  rangehi = 127;
4573  }
4574 
4575  /*
4576  * Now strip any common prefix of the three strings.
4577  */
4578  while (*lobound)
4579  {
4580  if (*lobound != *hibound || *lobound != *value)
4581  break;
4582  lobound++, hibound++, value++;
4583  }
4584 
4585  /*
4586  * Now we can do the conversions.
4587  */
4588  *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4589  *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4590  *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4591 }
static double convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
Definition: selfuncs.c:4594

References convert_one_string_to_scalar(), and value.

Referenced by convert_to_scalar().

◆ convert_timevalue_to_scalar()

static double convert_timevalue_to_scalar ( Datum  value,
Oid  typid,
bool failure 
)
static

Definition at line 4817 of file selfuncs.c.

4818 {
4819  switch (typid)
4820  {
4821  case TIMESTAMPOID:
4822  return DatumGetTimestamp(value);
4823  case TIMESTAMPTZOID:
4824  return DatumGetTimestampTz(value);
4825  case DATEOID:
4827  case INTERVALOID:
4828  {
4830 
4831  /*
4832  * Convert the month part of Interval to days using assumed
4833  * average month length of 365.25/12.0 days. Not too
4834  * accurate, but plenty good enough for our purposes.
4835  *
4836  * This also works for infinite intervals, which just have all
4837  * fields set to INT_MIN/INT_MAX, and so will produce a result
4838  * smaller/larger than any finite interval.
4839  */
4840  return interval->time + interval->day * (double) USECS_PER_DAY +
4842  }
4843  case TIMEOID:
4844  return DatumGetTimeADT(value);
4845  case TIMETZOID:
4846  {
4847  TimeTzADT *timetz = DatumGetTimeTzADTP(value);
4848 
4849  /* use GMT-equivalent time */
4850  return (double) (timetz->time + (timetz->zone * 1000000.0));
4851  }
4852  }
4853 
4854  *failure = true;
4855  return 0;
4856 }
#define MONTHS_PER_YEAR
Definition: timestamp.h:108
#define USECS_PER_DAY
Definition: timestamp.h:131
#define DAYS_PER_YEAR
Definition: timestamp.h:107
double date2timestamp_no_overflow(DateADT dateVal)
Definition: date.c:720
static DateADT DatumGetDateADT(Datum X)
Definition: date.h:54
static TimeADT DatumGetTimeADT(Datum X)
Definition: date.h:60
static TimeTzADT * DatumGetTimeTzADTP(Datum X)
Definition: date.h:66
Definition: date.h:28
TimeADT time
Definition: date.h:29
int32 zone
Definition: date.h:30
static Interval * DatumGetIntervalP(Datum X)
Definition: timestamp.h:40
static Timestamp DatumGetTimestamp(Datum X)
Definition: timestamp.h:28
static TimestampTz DatumGetTimestampTz(Datum X)
Definition: timestamp.h:34

References date2timestamp_no_overflow(), DatumGetDateADT(), DatumGetIntervalP(), DatumGetTimeADT(), DatumGetTimestamp(), DatumGetTimestampTz(), DatumGetTimeTzADTP(), DAYS_PER_YEAR, interval::month, MONTHS_PER_YEAR, TimeTzADT::time, interval::time, USECS_PER_DAY, value, and TimeTzADT::zone.

Referenced by convert_to_scalar().

◆ convert_to_scalar()

static bool convert_to_scalar ( Datum  value,
Oid  valuetypid,
Oid  collid,
double *  scaledvalue,
Datum  lobound,
Datum  hibound,
Oid  boundstypid,
double *  scaledlobound,
double *  scaledhibound 
)
static

Definition at line 4305 of file selfuncs.c.

4308 {
4309  bool failure = false;
4310 
4311  /*
4312  * Both the valuetypid and the boundstypid should exactly match the
4313  * declared input type(s) of the operator we are invoked for. However,
4314  * extensions might try to use scalarineqsel as estimator for operators
4315  * with input type(s) we don't handle here; in such cases, we want to
4316  * return false, not fail. In any case, we mustn't assume that valuetypid
4317  * and boundstypid are identical.
4318  *
4319  * XXX The histogram we are interpolating between points of could belong
4320  * to a column that's only binary-compatible with the declared type. In
4321  * essence we are assuming that the semantics of binary-compatible types
4322  * are enough alike that we can use a histogram generated with one type's
4323  * operators to estimate selectivity for the other's. This is outright
4324  * wrong in some cases --- in particular signed versus unsigned
4325  * interpretation could trip us up. But it's useful enough in the
4326  * majority of cases that we do it anyway. Should think about more
4327  * rigorous ways to do it.
4328  */
4329  switch (valuetypid)
4330  {
4331  /*
4332  * Built-in numeric types
4333  */
4334  case BOOLOID:
4335  case INT2OID:
4336  case INT4OID:
4337  case INT8OID:
4338  case FLOAT4OID:
4339  case FLOAT8OID:
4340  case NUMERICOID:
4341  case OIDOID:
4342  case REGPROCOID:
4343  case REGPROCEDUREOID:
4344  case REGOPEROID:
4345  case REGOPERATOROID:
4346  case REGCLASSOID:
4347  case REGTYPEOID:
4348  case REGCOLLATIONOID:
4349  case REGCONFIGOID:
4350  case REGDICTIONARYOID:
4351  case REGROLEOID:
4352  case REGNAMESPACEOID:
4353  *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4354  &failure);
4355  *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4356  &failure);
4357  *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4358  &failure);
4359  return !failure;
4360 
4361  /*
4362  * Built-in string types
4363  */
4364  case CHAROID:
4365  case BPCHAROID:
4366  case VARCHAROID:
4367  case TEXTOID:
4368  case NAMEOID:
4369  {
4370  char *valstr = convert_string_datum(value, valuetypid,
4371  collid, &failure);
4372  char *lostr = convert_string_datum(lobound, boundstypid,
4373  collid, &failure);
4374  char *histr = convert_string_datum(hibound, boundstypid,
4375  collid, &failure);
4376 
4377  /*
4378  * Bail out if any of the values is not of string type. We
4379  * might leak converted strings for the other value(s), but
4380  * that's not worth troubling over.
4381  */
4382  if (failure)
4383  return false;
4384 
4385  convert_string_to_scalar(valstr, scaledvalue,
4386  lostr, scaledlobound,
4387  histr, scaledhibound);
4388  pfree(valstr);
4389  pfree(lostr);
4390  pfree(histr);
4391  return true;
4392  }
4393 
4394  /*
4395  * Built-in bytea type
4396  */
4397  case BYTEAOID:
4398  {
4399  /* We only support bytea vs bytea comparison */
4400  if (boundstypid != BYTEAOID)
4401  return false;
4402  convert_bytea_to_scalar(value, scaledvalue,
4403  lobound, scaledlobound,
4404  hibound, scaledhibound);
4405  return true;
4406  }
4407 
4408  /*
4409  * Built-in time types
4410  */
4411  case TIMESTAMPOID:
4412  case TIMESTAMPTZOID:
4413  case DATEOID:
4414  case INTERVALOID:
4415  case TIMEOID:
4416  case TIMETZOID:
4417  *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4418  &failure);
4419  *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4420  &failure);
4421  *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4422  &failure);
4423  return !failure;
4424 
4425  /*
4426  * Built-in network types
4427  */
4428  case INETOID:
4429  case CIDROID:
4430  case MACADDROID:
4431  case MACADDR8OID:
4432  *scaledvalue = convert_network_to_scalar(value, valuetypid,
4433  &failure);
4434  *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4435  &failure);
4436  *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4437  &failure);
4438  return !failure;
4439  }
4440  /* Don't know how to convert */
4441  *scaledvalue = *scaledlobound = *scaledhibound = 0;
4442  return false;
4443 }
double convert_network_to_scalar(Datum value, Oid typid, bool *failure)
Definition: network.c:1502
static void convert_string_to_scalar(char *value, double *scaledvalue, char *lobound, double *scaledlobound, char *hibound, double *scaledhibound)
Definition: selfuncs.c:4514
static double convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
Definition: selfuncs.c:4817
static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
Definition: selfuncs.c:4452
static void convert_bytea_to_scalar(Datum value, double *scaledvalue, Datum lobound, double *scaledlobound, Datum hibound, double *scaledhibound)
Definition: selfuncs.c:4726
static char * convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
Definition: selfuncs.c:4645

References collid, convert_bytea_to_scalar(), convert_network_to_scalar(), convert_numeric_to_scalar(), convert_string_datum(), convert_string_to_scalar(), convert_timevalue_to_scalar(), pfree(), and value.

Referenced by ineq_histogram_selectivity().

◆ eqjoinsel()

Datum eqjoinsel ( PG_FUNCTION_ARGS  )

Definition at line 2269 of file selfuncs.c.

2270 {
2271  PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2272  Oid operator = PG_GETARG_OID(1);
2273  List *args = (List *) PG_GETARG_POINTER(2);
2274 
2275 #ifdef NOT_USED
2276  JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2277 #endif
2279  Oid collation = PG_GET_COLLATION();
2280  double selec;
2281  double selec_inner;
2282  VariableStatData vardata1;
2283  VariableStatData vardata2;
2284  double nd1;
2285  double nd2;
2286  bool isdefault1;
2287  bool isdefault2;
2288  Oid opfuncoid;
2289  AttStatsSlot sslot1;
2290  AttStatsSlot sslot2;
2291  Form_pg_statistic stats1 = NULL;
2292  Form_pg_statistic stats2 = NULL;
2293  bool have_mcvs1 = false;
2294  bool have_mcvs2 = false;
2295  bool get_mcv_stats;
2296  bool join_is_reversed;
2297  RelOptInfo *inner_rel;
2298 
2299  get_join_variables(root, args, sjinfo,
2300  &vardata1, &vardata2, &join_is_reversed);
2301 
2302  nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2303  nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2304 
2305  opfuncoid = get_opcode(operator);
2306 
2307  memset(&sslot1, 0, sizeof(sslot1));
2308  memset(&sslot2, 0, sizeof(sslot2));
2309 
2310  /*
2311  * There is no use in fetching one side's MCVs if we lack MCVs for the
2312  * other side, so do a quick check to verify that both stats exist.
2313  */
2314  get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2315  HeapTupleIsValid(vardata2.statsTuple) &&
2316  get_attstatsslot(&sslot1, vardata1.statsTuple,
2317  STATISTIC_KIND_MCV, InvalidOid,
2318  0) &&
2319  get_attstatsslot(&sslot2, vardata2.statsTuple,
2320  STATISTIC_KIND_MCV, InvalidOid,
2321  0));
2322 
2323  if (HeapTupleIsValid(vardata1.statsTuple))
2324  {
2325  /* note we allow use of nullfrac regardless of security check */
2326  stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2327  if (get_mcv_stats &&
2328  statistic_proc_security_check(&vardata1, opfuncoid))
2329  have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2330  STATISTIC_KIND_MCV, InvalidOid,
2332  }
2333 
2334  if (HeapTupleIsValid(vardata2.statsTuple))
2335  {
2336  /* note we allow use of nullfrac regardless of security check */
2337  stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2338  if (get_mcv_stats &&
2339  statistic_proc_security_check(&vardata2, opfuncoid))
2340  have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2341  STATISTIC_KIND_MCV, InvalidOid,
2343  }
2344 
2345  /* We need to compute the inner-join selectivity in all cases */
2346  selec_inner = eqjoinsel_inner(opfuncoid, collation,
2347  &vardata1, &vardata2,
2348  nd1, nd2,
2349  isdefault1, isdefault2,
2350  &sslot1, &sslot2,
2351  stats1, stats2,
2352  have_mcvs1, have_mcvs2);
2353 
2354  switch (sjinfo->jointype)
2355  {
2356  case JOIN_INNER:
2357  case JOIN_LEFT:
2358  case JOIN_FULL:
2359  selec = selec_inner;
2360  break;
2361  case JOIN_SEMI:
2362  case JOIN_ANTI:
2363 
2364  /*
2365  * Look up the join's inner relation. min_righthand is sufficient
2366  * information because neither SEMI nor ANTI joins permit any
2367  * reassociation into or out of their RHS, so the righthand will
2368  * always be exactly that set of rels.
2369  */
2370  inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2371 
2372  if (!join_is_reversed)
2373  selec = eqjoinsel_semi(opfuncoid, collation,
2374  &vardata1, &vardata2,
2375  nd1, nd2,
2376  isdefault1, isdefault2,
2377  &sslot1, &sslot2,
2378  stats1, stats2,
2379  have_mcvs1, have_mcvs2,
2380  inner_rel);
2381  else
2382  {
2383  Oid commop = get_commutator(operator);
2384  Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2385 
2386  selec = eqjoinsel_semi(commopfuncoid, collation,
2387  &vardata2, &vardata1,
2388  nd2, nd1,
2389  isdefault2, isdefault1,
2390  &sslot2, &sslot1,
2391  stats2, stats1,
2392  have_mcvs2, have_mcvs1,
2393  inner_rel);
2394  }
2395 
2396  /*
2397  * We should never estimate the output of a semijoin to be more
2398  * rows than we estimate for an inner join with the same input
2399  * rels and join condition; it's obviously impossible for that to
2400  * happen. The former estimate is N1 * Ssemi while the latter is
2401  * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2402  * this is worthwhile because of the shakier estimation rules we
2403  * use in eqjoinsel_semi, particularly in cases where it has to
2404  * punt entirely.
2405  */
2406  selec = Min(selec, inner_rel->rows * selec_inner);
2407  break;
2408  default:
2409  /* other values not expected here */
2410  elog(ERROR, "unrecognized join type: %d",
2411  (int) sjinfo->jointype);
2412  selec = 0; /* keep compiler quiet */
2413  break;
2414  }
2415 
2416  free_attstatsslot(&sslot1);
2417  free_attstatsslot(&sslot2);
2418 
2419  ReleaseVariableStats(vardata1);
2420  ReleaseVariableStats(vardata2);
2421 
2422  CLAMP_PROBABILITY(selec);
2423 
2424  PG_RETURN_FLOAT8((float8) selec);
2425 }
double float8
Definition: c.h:617
#define PG_GETARG_OID(n)
Definition: fmgr.h:275
#define PG_RETURN_FLOAT8(x)
Definition: fmgr.h:367
#define PG_GETARG_POINTER(n)
Definition: fmgr.h:276
#define PG_GET_COLLATION()
Definition: fmgr.h:198
#define PG_GETARG_INT16(n)
Definition: fmgr.h:271
RegProcedure get_opcode(Oid opno)
Definition: lsyscache.c:1263
Oid get_commutator(Oid opno)
Definition: lsyscache.c:1487
JoinType
Definition: nodes.h:278
@ JOIN_SEMI
Definition: nodes.h:297
@ JOIN_FULL
Definition: nodes.h:285
@ JOIN_LEFT
Definition: nodes.h:284
@ JOIN_ANTI
Definition: nodes.h:298
static RelOptInfo * find_join_input_rel(PlannerInfo *root, Relids relids)
Definition: selfuncs.c:6428
static double eqjoinsel_inner(Oid opfuncoid, Oid collation, VariableStatData *vardata1, VariableStatData *vardata2, double nd1, double nd2, bool isdefault1, bool isdefault2, AttStatsSlot *sslot1, AttStatsSlot *sslot2, Form_pg_statistic stats1, Form_pg_statistic stats2, bool have_mcvs1, bool have_mcvs2)
Definition: selfuncs.c:2434
static double eqjoinsel_semi(Oid opfuncoid, Oid collation, VariableStatData *vardata1, VariableStatData *vardata2, double nd1, double nd2, bool isdefault1, bool isdefault2, AttStatsSlot *sslot1, AttStatsSlot *sslot2, Form_pg_statistic stats1, Form_pg_statistic stats2, bool have_mcvs1, bool have_mcvs2, RelOptInfo *inner_rel)
Definition: selfuncs.c:2631
bool statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
Definition: selfuncs.c:5735
void get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo, VariableStatData *vardata1, VariableStatData *vardata2, bool *join_is_reversed)
Definition: selfuncs.c:4943
Cardinality rows
Definition: pathnodes.h:862
Relids min_righthand
Definition: pathnodes.h:2869
JoinType jointype
Definition: pathnodes.h:2872

References generate_unaccent_rules::args, ATTSTATSSLOT_NUMBERS, ATTSTATSSLOT_VALUES, CLAMP_PROBABILITY, elog, eqjoinsel_inner(), eqjoinsel_semi(), ERROR, find_join_input_rel(), free_attstatsslot(), get_attstatsslot(), get_commutator(), get_join_variables(), get_opcode(), get_variable_numdistinct(), GETSTRUCT, HeapTupleIsValid, InvalidOid, JOIN_ANTI, JOIN_FULL, JOIN_INNER, JOIN_LEFT, JOIN_SEMI, SpecialJoinInfo::jointype, Min, SpecialJoinInfo::min_righthand, OidIsValid, PG_GET_COLLATION, PG_GETARG_INT16, PG_GETARG_OID, PG_GETARG_POINTER, PG_RETURN_FLOAT8, ReleaseVariableStats, RelOptInfo::rows, statistic_proc_security_check(), and VariableStatData::statsTuple.

Referenced by neqjoinsel().

◆ eqjoinsel_inner()

static double eqjoinsel_inner ( Oid  opfuncoid,
Oid  collation,
VariableStatData vardata1,
VariableStatData vardata2,
double  nd1,
double  nd2,
bool  isdefault1,
bool  isdefault2,
AttStatsSlot sslot1,
AttStatsSlot sslot2,
Form_pg_statistic  stats1,
Form_pg_statistic  stats2,
bool  have_mcvs1,
bool  have_mcvs2 
)
static

Definition at line 2434 of file selfuncs.c.

2441 {
2442  double selec;
2443 
2444  if (have_mcvs1 && have_mcvs2)
2445  {
2446  /*
2447  * We have most-common-value lists for both relations. Run through
2448  * the lists to see which MCVs actually join to each other with the
2449  * given operator. This allows us to determine the exact join
2450  * selectivity for the portion of the relations represented by the MCV
2451  * lists. We still have to estimate for the remaining population, but
2452  * in a skewed distribution this gives us a big leg up in accuracy.
2453  * For motivation see the analysis in Y. Ioannidis and S.
2454  * Christodoulakis, "On the propagation of errors in the size of join
2455  * results", Technical Report 1018, Computer Science Dept., University
2456  * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2457  */
2458  LOCAL_FCINFO(fcinfo, 2);
2459  FmgrInfo eqproc;
2460  bool *hasmatch1;
2461  bool *hasmatch2;
2462  double nullfrac1 = stats1->stanullfrac;
2463  double nullfrac2 = stats2->stanullfrac;
2464  double matchprodfreq,
2465  matchfreq1,
2466  matchfreq2,
2467  unmatchfreq1,
2468  unmatchfreq2,
2469  otherfreq1,
2470  otherfreq2,
2471  totalsel1,
2472  totalsel2;
2473  int i,
2474  nmatches;
2475 
2476  fmgr_info(opfuncoid, &eqproc);
2477 
2478  /*
2479  * Save a few cycles by setting up the fcinfo struct just once. Using
2480  * FunctionCallInvoke directly also avoids failure if the eqproc
2481  * returns NULL, though really equality functions should never do
2482  * that.
2483  */
2484  InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2485  NULL, NULL);
2486  fcinfo->args[0].isnull = false;
2487  fcinfo->args[1].isnull = false;
2488 
2489  hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2490  hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2491 
2492  /*
2493  * Note we assume that each MCV will match at most one member of the
2494  * other MCV list. If the operator isn't really equality, there could
2495  * be multiple matches --- but we don't look for them, both for speed
2496  * and because the math wouldn't add up...
2497  */
2498  matchprodfreq = 0.0;
2499  nmatches = 0;
2500  for (i = 0; i < sslot1->nvalues; i++)
2501  {
2502  int j;
2503 
2504  fcinfo->args[0].value = sslot1->values[i];
2505 
2506  for (j = 0; j < sslot2->nvalues; j++)
2507  {
2508  Datum fresult;
2509 
2510  if (hasmatch2[j])
2511  continue;
2512  fcinfo->args[1].value = sslot2->values[j];
2513  fcinfo->isnull = false;
2514  fresult = FunctionCallInvoke(fcinfo);
2515  if (!fcinfo->isnull && DatumGetBool(fresult))
2516  {
2517  hasmatch1[i] = hasmatch2[j] = true;
2518  matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2519  nmatches++;
2520  break;
2521  }
2522  }
2523  }
2524  CLAMP_PROBABILITY(matchprodfreq);
2525  /* Sum up frequencies of matched and unmatched MCVs */
2526  matchfreq1 = unmatchfreq1 = 0.0;
2527  for (i = 0; i < sslot1->nvalues; i++)
2528  {
2529  if (hasmatch1[i])
2530  matchfreq1 += sslot1->numbers[i];
2531  else
2532  unmatchfreq1 += sslot1->numbers[i];
2533  }
2534  CLAMP_PROBABILITY(matchfreq1);
2535  CLAMP_PROBABILITY(unmatchfreq1);
2536  matchfreq2 = unmatchfreq2 = 0.0;
2537  for (i = 0; i < sslot2->nvalues; i++)
2538  {
2539  if (hasmatch2[i])
2540  matchfreq2 += sslot2->numbers[i];
2541  else
2542  unmatchfreq2 += sslot2->numbers[i];
2543  }
2544  CLAMP_PROBABILITY(matchfreq2);
2545  CLAMP_PROBABILITY(unmatchfreq2);
2546  pfree(hasmatch1);
2547  pfree(hasmatch2);
2548 
2549  /*
2550  * Compute total frequency of non-null values that are not in the MCV
2551  * lists.
2552  */
2553  otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2554  otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2555  CLAMP_PROBABILITY(otherfreq1);
2556  CLAMP_PROBABILITY(otherfreq2);
2557 
2558  /*
2559  * We can estimate the total selectivity from the point of view of
2560  * relation 1 as: the known selectivity for matched MCVs, plus
2561  * unmatched MCVs that are assumed to match against random members of
2562  * relation 2's non-MCV population, plus non-MCV values that are
2563  * assumed to match against random members of relation 2's unmatched
2564  * MCVs plus non-MCV values.
2565  */
2566  totalsel1 = matchprodfreq;
2567  if (nd2 > sslot2->nvalues)
2568  totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2569  if (nd2 > nmatches)
2570  totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2571  (nd2 - nmatches);
2572  /* Same estimate from the point of view of relation 2. */
2573  totalsel2 = matchprodfreq;
2574  if (nd1 > sslot1->nvalues)
2575  totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2576  if (nd1 > nmatches)
2577  totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2578  (nd1 - nmatches);
2579 
2580  /*
2581  * Use the smaller of the two estimates. This can be justified in
2582  * essentially the same terms as given below for the no-stats case: to
2583  * a first approximation, we are estimating from the point of view of
2584  * the relation with smaller nd.
2585  */
2586  selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2587  }
2588  else
2589  {
2590  /*
2591  * We do not have MCV lists for both sides. Estimate the join
2592  * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2593  * is plausible if we assume that the join operator is strict and the
2594  * non-null values are about equally distributed: a given non-null
2595  * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2596  * of rel2, so total join rows are at most
2597  * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2598  * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2599  * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2600  * with MIN() is an upper bound. Using the MIN() means we estimate
2601  * from the point of view of the relation with smaller nd (since the
2602  * larger nd is determining the MIN). It is reasonable to assume that
2603  * most tuples in this rel will have join partners, so the bound is
2604  * probably reasonably tight and should be taken as-is.
2605  *
2606  * XXX Can we be smarter if we have an MCV list for just one side? It
2607  * seems that if we assume equal distribution for the other side, we
2608  * end up with the same answer anyway.
2609  */
2610  double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2611  double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2612 
2613  selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2614  if (nd1 > nd2)
2615  selec /= nd1;
2616  else
2617  selec /= nd2;
2618  }
2619 
2620  return selec;
2621 }
void fmgr_info(Oid functionId, FmgrInfo *finfo)
Definition: fmgr.c:127
#define InitFunctionCallInfoData(Fcinfo, Flinfo, Nargs, Collation, Context, Resultinfo)
Definition: fmgr.h:150
#define LOCAL_FCINFO(name, nargs)
Definition: fmgr.h:110
#define FunctionCallInvoke(fcinfo)
Definition: fmgr.h:172
int j
Definition: isn.c:74
void * palloc0(Size size)
Definition: mcxt.c:1334
uintptr_t Datum
Definition: postgres.h:64
Definition: fmgr.h:57

References CLAMP_PROBABILITY, DatumGetBool(), fmgr_info(), FunctionCallInvoke, i, InitFunctionCallInfoData, j, LOCAL_FCINFO, AttStatsSlot::numbers, AttStatsSlot::nvalues, palloc0(), pfree(), and AttStatsSlot::values.

Referenced by eqjoinsel().

◆ eqjoinsel_semi()

static double eqjoinsel_semi ( Oid  opfuncoid,
Oid  collation,
VariableStatData vardata1,
VariableStatData vardata2,
double  nd1,
double  nd2,
bool  isdefault1,
bool  isdefault2,
AttStatsSlot sslot1,
AttStatsSlot sslot2,
Form_pg_statistic  stats1,
Form_pg_statistic  stats2,
bool  have_mcvs1,
bool  have_mcvs2,
RelOptInfo inner_rel 
)
static

Definition at line 2631 of file selfuncs.c.

2639 {
2640  double selec;
2641 
2642  /*
2643  * We clamp nd2 to be not more than what we estimate the inner relation's
2644  * size to be. This is intuitively somewhat reasonable since obviously
2645  * there can't be more than that many distinct values coming from the
2646  * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2647  * likewise) is that this is the only pathway by which restriction clauses
2648  * applied to the inner rel will affect the join result size estimate,
2649  * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2650  * only the outer rel's size. If we clamped nd1 we'd be double-counting
2651  * the selectivity of outer-rel restrictions.
2652  *
2653  * We can apply this clamping both with respect to the base relation from
2654  * which the join variable comes (if there is just one), and to the
2655  * immediate inner input relation of the current join.
2656  *
2657  * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2658  * great, maybe, but it didn't come out of nowhere either. This is most
2659  * helpful when the inner relation is empty and consequently has no stats.
2660  */
2661  if (vardata2->rel)
2662  {
2663  if (nd2 >= vardata2->rel->rows)
2664  {
2665  nd2 = vardata2->rel->rows;
2666  isdefault2 = false;
2667  }
2668  }
2669  if (nd2 >= inner_rel->rows)
2670  {
2671  nd2 = inner_rel->rows;
2672  isdefault2 = false;
2673  }
2674 
2675  if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2676  {
2677  /*
2678  * We have most-common-value lists for both relations. Run through
2679  * the lists to see which MCVs actually join to each other with the
2680  * given operator. This allows us to determine the exact join
2681  * selectivity for the portion of the relations represented by the MCV
2682  * lists. We still have to estimate for the remaining population, but
2683  * in a skewed distribution this gives us a big leg up in accuracy.
2684  */
2685  LOCAL_FCINFO(fcinfo, 2);
2686  FmgrInfo eqproc;
2687  bool *hasmatch1;
2688  bool *hasmatch2;
2689  double nullfrac1 = stats1->stanullfrac;
2690  double matchfreq1,
2691  uncertainfrac,
2692  uncertain;
2693  int i,
2694  nmatches,
2695  clamped_nvalues2;
2696 
2697  /*
2698  * The clamping above could have resulted in nd2 being less than
2699  * sslot2->nvalues; in which case, we assume that precisely the nd2
2700  * most common values in the relation will appear in the join input,
2701  * and so compare to only the first nd2 members of the MCV list. Of
2702  * course this is frequently wrong, but it's the best bet we can make.
2703  */
2704  clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2705 
2706  fmgr_info(opfuncoid, &eqproc);
2707 
2708  /*
2709  * Save a few cycles by setting up the fcinfo struct just once. Using
2710  * FunctionCallInvoke directly also avoids failure if the eqproc
2711  * returns NULL, though really equality functions should never do
2712  * that.
2713  */
2714  InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2715  NULL, NULL);
2716  fcinfo->args[0].isnull = false;
2717  fcinfo->args[1].isnull = false;
2718 
2719  hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2720  hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2721 
2722  /*
2723  * Note we assume that each MCV will match at most one member of the
2724  * other MCV list. If the operator isn't really equality, there could
2725  * be multiple matches --- but we don't look for them, both for speed
2726  * and because the math wouldn't add up...
2727  */
2728  nmatches = 0;
2729  for (i = 0; i < sslot1->nvalues; i++)
2730  {
2731  int j;
2732 
2733  fcinfo->args[0].value = sslot1->values[i];
2734 
2735  for (j = 0; j < clamped_nvalues2; j++)
2736  {
2737  Datum fresult;
2738 
2739  if (hasmatch2[j])
2740  continue;
2741  fcinfo->args[1].value = sslot2->values[j];
2742  fcinfo->isnull = false;
2743  fresult = FunctionCallInvoke(fcinfo);
2744  if (!fcinfo->isnull && DatumGetBool(fresult))
2745  {
2746  hasmatch1[i] = hasmatch2[j] = true;
2747  nmatches++;
2748  break;
2749  }
2750  }
2751  }
2752  /* Sum up frequencies of matched MCVs */
2753  matchfreq1 = 0.0;
2754  for (i = 0; i < sslot1->nvalues; i++)
2755  {
2756  if (hasmatch1[i])
2757  matchfreq1 += sslot1->numbers[i];
2758  }
2759  CLAMP_PROBABILITY(matchfreq1);
2760  pfree(hasmatch1);
2761  pfree(hasmatch2);
2762 
2763  /*
2764  * Now we need to estimate the fraction of relation 1 that has at
2765  * least one join partner. We know for certain that the matched MCVs
2766  * do, so that gives us a lower bound, but we're really in the dark
2767  * about everything else. Our crude approach is: if nd1 <= nd2 then
2768  * assume all non-null rel1 rows have join partners, else assume for
2769  * the uncertain rows that a fraction nd2/nd1 have join partners. We
2770  * can discount the known-matched MCVs from the distinct-values counts
2771  * before doing the division.
2772  *
2773  * Crude as the above is, it's completely useless if we don't have
2774  * reliable ndistinct values for both sides. Hence, if either nd1 or
2775  * nd2 is default, punt and assume half of the uncertain rows have
2776  * join partners.
2777  */
2778  if (!isdefault1 && !isdefault2)
2779  {
2780  nd1 -= nmatches;
2781  nd2 -= nmatches;
2782  if (nd1 <= nd2 || nd2 < 0)
2783  uncertainfrac = 1.0;
2784  else
2785  uncertainfrac = nd2 / nd1;
2786  }
2787  else
2788  uncertainfrac = 0.5;
2789  uncertain = 1.0 - matchfreq1 - nullfrac1;
2790  CLAMP_PROBABILITY(uncertain);
2791  selec = matchfreq1 + uncertainfrac * uncertain;
2792  }
2793  else
2794  {
2795  /*
2796  * Without MCV lists for both sides, we can only use the heuristic
2797  * about nd1 vs nd2.
2798  */
2799  double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2800 
2801  if (!isdefault1 && !isdefault2)
2802  {
2803  if (nd1 <= nd2 || nd2 < 0)
2804  selec = 1.0 - nullfrac1;
2805  else
2806  selec = (nd2 / nd1) * (1.0 - nullfrac1);
2807  }
2808  else
2809  selec = 0.5 * (1.0 - nullfrac1);
2810  }
2811 
2812  return selec;
2813 }

References CLAMP_PROBABILITY, DatumGetBool(), fmgr_info(), FunctionCallInvoke, i, InitFunctionCallInfoData, j, LOCAL_FCINFO, Min, AttStatsSlot::numbers, AttStatsSlot::nvalues, OidIsValid, palloc0(), pfree(), VariableStatData::rel, RelOptInfo::rows, and AttStatsSlot::values.

Referenced by eqjoinsel().

◆ eqsel()

Datum eqsel ( PG_FUNCTION_ARGS  )

Definition at line 227 of file selfuncs.c.

228 {
229  PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
230 }
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate)
Definition: selfuncs.c:236

References eqsel_internal(), and PG_RETURN_FLOAT8.

◆ eqsel_internal()

static double eqsel_internal ( PG_FUNCTION_ARGS  ,
bool  negate 
)
static

Definition at line 236 of file selfuncs.c.

237 {
239  Oid operator = PG_GETARG_OID(1);
240  List *args = (List *) PG_GETARG_POINTER(2);
241  int varRelid = PG_GETARG_INT32(3);
242  Oid collation = PG_GET_COLLATION();
243  VariableStatData vardata;
244  Node *other;
245  bool varonleft;
246  double selec;
247 
248  /*
249  * When asked about <>, we do the estimation using the corresponding =
250  * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
251  */
252  if (negate)
253  {
254  operator = get_negator(operator);
255  if (!OidIsValid(operator))
256  {
257  /* Use default selectivity (should we raise an error instead?) */
258  return 1.0 - DEFAULT_EQ_SEL;
259  }
260  }
261 
262  /*
263  * If expression is not variable = something or something = variable, then
264  * punt and return a default estimate.
265  */
266  if (!get_restriction_variable(root, args, varRelid,
267  &vardata, &other, &varonleft))
268  return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
269 
270  /*
271  * We can do a lot better if the something is a constant. (Note: the
272  * Const might result from estimation rather than being a simple constant
273  * in the query.)
274  */
275  if (IsA(other, Const))
276  selec = var_eq_const(&vardata, operator, collation,
277  ((Const *) other)->constvalue,
278  ((Const *) other)->constisnull,
279  varonleft, negate);
280  else
281  selec = var_eq_non_const(&vardata, operator, collation, other,
282  varonleft, negate);
283 
284  ReleaseVariableStats(vardata);
285 
286  return selec;
287 }
#define PG_GETARG_INT32(n)
Definition: fmgr.h:269
Oid get_negator(Oid opno)
Definition: lsyscache.c:1511
bool get_restriction_variable(PlannerInfo *root, List *args, int varRelid, VariableStatData *vardata, Node **other, bool *varonleft)
Definition: selfuncs.c:4883
double var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation, Node *other, bool varonleft, bool negate)
Definition: selfuncs.c:466
#define DEFAULT_EQ_SEL
Definition: selfuncs.h:34

References generate_unaccent_rules::args, DEFAULT_EQ_SEL, get_negator(), get_restriction_variable(), IsA, OidIsValid, PG_GET_COLLATION, PG_GETARG_INT32, PG_GETARG_OID, PG_GETARG_POINTER, ReleaseVariableStats, var_eq_const(), and var_eq_non_const().

Referenced by eqsel(), and neqsel().

◆ estimate_array_length()

double estimate_array_length ( PlannerInfo root,
Node arrayexpr 
)

Definition at line 2136 of file selfuncs.c.

2137 {
2138  /* look through any binary-compatible relabeling of arrayexpr */
2139  arrayexpr = strip_array_coercion(arrayexpr);
2140 
2141  if (arrayexpr && IsA(arrayexpr, Const))
2142  {
2143  Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2144  bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2145  ArrayType *arrayval;
2146 
2147  if (arrayisnull)
2148  return 0;
2149  arrayval = DatumGetArrayTypeP(arraydatum);
2150  return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2151  }
2152  else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2153  !((ArrayExpr *) arrayexpr)->multidims)
2154  {
2155  return list_length(((ArrayExpr *) arrayexpr)->elements);
2156  }
2157  else if (arrayexpr)
2158  {
2159  /* See if we can find any statistics about it */
2160  VariableStatData vardata;
2161  AttStatsSlot sslot;
2162  double nelem = 0;
2163 
2164  examine_variable(root, arrayexpr, 0, &vardata);
2165  if (HeapTupleIsValid(vardata.statsTuple))
2166  {
2167  /*
2168  * Found stats, so use the average element count, which is stored
2169  * in the last stanumbers element of the DECHIST statistics.
2170  * Actually that is the average count of *distinct* elements;
2171  * perhaps we should scale it up somewhat?
2172  */
2173  if (get_attstatsslot(&sslot, vardata.statsTuple,
2174  STATISTIC_KIND_DECHIST, InvalidOid,
2176  {
2177  if (sslot.nnumbers > 0)
2178  nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2179  free_attstatsslot(&sslot);
2180  }
2181  }
2182  ReleaseVariableStats(vardata);
2183 
2184  if (nelem > 0)
2185  return nelem;
2186  }
2187 
2188  /* Else use a default guess --- this should match scalararraysel */
2189  return 10;
2190 }
#define ARR_NDIM(a)
Definition: array.h:290
#define DatumGetArrayTypeP(X)
Definition: array.h:261
#define ARR_DIMS(a)
Definition: array.h:294
int ArrayGetNItems(int ndim, const int *dims)
Definition: arrayutils.c:57
double clamp_row_est(double nrows)
Definition: costsize.c:202
static int list_length(const List *l)
Definition: pg_list.h:152
static Node * strip_array_coercion(Node *node)
Definition: selfuncs.c:1783

References ARR_DIMS, ARR_NDIM, ArrayGetNItems(), ATTSTATSSLOT_NUMBERS, clamp_row_est(), DatumGetArrayTypeP, examine_variable(), free_attstatsslot(), get_attstatsslot(), HeapTupleIsValid, InvalidOid, IsA, list_length(), AttStatsSlot::nnumbers, AttStatsSlot::numbers, ReleaseVariableStats, VariableStatData::statsTuple, and strip_array_coercion().

Referenced by array_unnest_support(), btcostestimate(), cost_qual_eval_walker(), cost_tidscan(), genericcostestimate(), and gincost_scalararrayopexpr().

◆ estimate_hash_bucket_stats()

void estimate_hash_bucket_stats ( PlannerInfo root,
Node hashkey,
double  nbuckets,
Selectivity mcv_freq,
Selectivity bucketsize_frac 
)

Definition at line 3798 of file selfuncs.c.

3801 {
3802  VariableStatData vardata;
3803  double estfract,
3804  ndistinct,
3805  stanullfrac,
3806  avgfreq;
3807  bool isdefault;
3808  AttStatsSlot sslot;
3809 
3810  examine_variable(root, hashkey, 0, &vardata);
3811 
3812  /* Look up the frequency of the most common value, if available */
3813  *mcv_freq = 0.0;
3814 
3815  if (HeapTupleIsValid(vardata.statsTuple))
3816  {
3817  if (get_attstatsslot(&sslot, vardata.statsTuple,
3818  STATISTIC_KIND_MCV, InvalidOid,
3820  {
3821  /*
3822  * The first MCV stat is for the most common value.
3823  */
3824  if (sslot.nnumbers > 0)
3825  *mcv_freq = sslot.numbers[0];
3826  free_attstatsslot(&sslot);
3827  }
3828  }
3829 
3830  /* Get number of distinct values */
3831  ndistinct = get_variable_numdistinct(&vardata, &isdefault);
3832 
3833  /*
3834  * If ndistinct isn't real, punt. We normally return 0.1, but if the
3835  * mcv_freq is known to be even higher than that, use it instead.
3836  */
3837  if (isdefault)
3838  {
3839  *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
3840  ReleaseVariableStats(vardata);
3841  return;
3842  }
3843 
3844  /* Get fraction that are null */
3845  if (HeapTupleIsValid(vardata.statsTuple))
3846  {
3847  Form_pg_statistic stats;
3848 
3849  stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
3850  stanullfrac = stats->stanullfrac;
3851  }
3852  else
3853  stanullfrac = 0.0;
3854 
3855  /* Compute avg freq of all distinct data values in raw relation */
3856  avgfreq = (1.0 - stanullfrac) / ndistinct;
3857 
3858  /*
3859  * Adjust ndistinct to account for restriction clauses. Observe we are
3860  * assuming that the data distribution is affected uniformly by the
3861  * restriction clauses!
3862  *
3863  * XXX Possibly better way, but much more expensive: multiply by
3864  * selectivity of rel's restriction clauses that mention the target Var.
3865  */
3866  if (vardata.rel && vardata.rel->tuples > 0)
3867  {
3868  ndistinct *= vardata.rel->rows / vardata.rel->tuples;
3869  ndistinct = clamp_row_est(ndistinct);
3870  }
3871 
3872  /*
3873  * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
3874  * number of buckets is less than the expected number of distinct values;
3875  * otherwise it is 1/ndistinct.
3876  */
3877  if (ndistinct > nbuckets)
3878  estfract = 1.0 / nbuckets;
3879  else
3880  estfract = 1.0 / ndistinct;
3881 
3882  /*
3883  * Adjust estimated bucketsize upward to account for skewed distribution.
3884  */
3885  if (avgfreq > 0.0 && *mcv_freq > avgfreq)
3886  estfract *= *mcv_freq / avgfreq;
3887 
3888  /*
3889  * Clamp bucketsize to sane range (the above adjustment could easily
3890  * produce an out-of-range result). We set the lower bound a little above
3891  * zero, since zero isn't a very sane result.
3892  */
3893  if (estfract < 1.0e-6)
3894  estfract = 1.0e-6;
3895  else if (estfract > 1.0)
3896  estfract = 1.0;
3897 
3898  *bucketsize_frac = (Selectivity) estfract;
3899 
3900  ReleaseVariableStats(vardata);
3901 }
Cardinality tuples
Definition: pathnodes.h:930

References ATTSTATSSLOT_NUMBERS, clamp_row_est(), examine_variable(), free_attstatsslot(), get_attstatsslot(), get_variable_numdistinct(), GETSTRUCT, HeapTupleIsValid, InvalidOid, Max, AttStatsSlot::nnumbers, AttStatsSlot::numbers, VariableStatData::rel, ReleaseVariableStats, RelOptInfo::rows, VariableStatData::statsTuple, and RelOptInfo::tuples.

Referenced by final_cost_hashjoin().

◆ estimate_hashagg_tablesize()

double estimate_hashagg_tablesize ( PlannerInfo root,
Path path,
const AggClauseCosts agg_costs,
double  dNumGroups 
)

Definition at line 3917 of file selfuncs.c.

3919 {
3920  Size hashentrysize;
3921 
3922  hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
3923  path->pathtarget->width,
3924  agg_costs->transitionSpace);
3925 
3926  /*
3927  * Note that this disregards the effect of fill-factor and growth policy
3928  * of the hash table. That's probably ok, given that the default
3929  * fill-factor is relatively high. It'd be hard to meaningfully factor in
3930  * "double-in-size" growth policies here.
3931  */
3932  return hashentrysize * dNumGroups;
3933 }
size_t Size
Definition: c.h:592
Size hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
Definition: nodeAgg.c:1694
Size transitionSpace
Definition: pathnodes.h:62
List * aggtransinfos
Definition: pathnodes.h:509

References PlannerInfo::aggtransinfos, hash_agg_entry_size(), list_length(), and AggClauseCosts::transitionSpace.

Referenced by consider_groupingsets_paths().

◆ estimate_multivariate_ndistinct()

static bool estimate_multivariate_ndistinct ( PlannerInfo root,
RelOptInfo rel,
List **  varinfos,
double *  ndistinct 
)
static

Definition at line 3954 of file selfuncs.c.

3956 {
3957  ListCell *lc;
3958  int nmatches_vars;
3959  int nmatches_exprs;
3960  Oid statOid = InvalidOid;
3961  MVNDistinct *stats;
3962  StatisticExtInfo *matched_info = NULL;
3963  RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
3964 
3965  /* bail out immediately if the table has no extended statistics */
3966  if (!rel->statlist)
3967  return false;
3968 
3969  /* look for the ndistinct statistics object matching the most vars */
3970  nmatches_vars = 0; /* we require at least two matches */
3971  nmatches_exprs = 0;
3972  foreach(lc, rel->statlist)
3973  {
3974  ListCell *lc2;
3975  StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
3976  int nshared_vars = 0;
3977  int nshared_exprs = 0;
3978 
3979  /* skip statistics of other kinds */
3980  if (info->kind != STATS_EXT_NDISTINCT)
3981  continue;
3982 
3983  /* skip statistics with mismatching stxdinherit value */
3984  if (info->inherit != rte->inh)
3985  continue;
3986 
3987  /*
3988  * Determine how many expressions (and variables in non-matched
3989  * expressions) match. We'll then use these numbers to pick the
3990  * statistics object that best matches the clauses.
3991  */
3992  foreach(lc2, *varinfos)
3993  {
3994  ListCell *lc3;
3995  GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
3997 
3998  Assert(varinfo->rel == rel);
3999 
4000  /* simple Var, search in statistics keys directly */
4001  if (IsA(varinfo->var, Var))
4002  {
4003  attnum = ((Var *) varinfo->var)->varattno;
4004 
4005  /*
4006  * Ignore system attributes - we don't support statistics on
4007  * them, so can't match them (and it'd fail as the values are
4008  * negative).
4009  */
4011  continue;
4012 
4013  if (bms_is_member(attnum, info->keys))
4014  nshared_vars++;
4015 
4016  continue;
4017  }
4018 
4019  /* expression - see if it's in the statistics object */
4020  foreach(lc3, info->exprs)
4021  {
4022  Node *expr = (Node *) lfirst(lc3);
4023 
4024  if (equal(varinfo->var, expr))
4025  {
4026  nshared_exprs++;
4027  break;
4028  }
4029  }
4030  }
4031 
4032  if (nshared_vars + nshared_exprs < 2)
4033  continue;
4034 
4035  /*
4036  * Does this statistics object match more columns than the currently
4037  * best object? If so, use this one instead.
4038  *
4039  * XXX This should break ties using name of the object, or something
4040  * like that, to make the outcome stable.
4041  */
4042  if ((nshared_exprs > nmatches_exprs) ||
4043  (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4044  {
4045  statOid = info->statOid;
4046  nmatches_vars = nshared_vars;
4047  nmatches_exprs = nshared_exprs;
4048  matched_info = info;
4049  }
4050  }
4051 
4052  /* No match? */
4053  if (statOid == InvalidOid)
4054  return false;
4055 
4056  Assert(nmatches_vars + nmatches_exprs > 1);
4057 
4058  stats = statext_ndistinct_load(statOid, rte->inh);
4059 
4060  /*
4061  * If we have a match, search it for the specific item that matches (there
4062  * must be one), and construct the output values.
4063  */
4064  if (stats)
4065  {
4066  int i;
4067  List *newlist = NIL;
4068  MVNDistinctItem *item = NULL;
4069  ListCell *lc2;
4070  Bitmapset *matched = NULL;
4071  AttrNumber attnum_offset;
4072 
4073  /*
4074  * How much we need to offset the attnums? If there are no
4075  * expressions, no offset is needed. Otherwise offset enough to move
4076  * the lowest one (which is equal to number of expressions) to 1.
4077  */
4078  if (matched_info->exprs)
4079  attnum_offset = (list_length(matched_info->exprs) + 1);
4080  else
4081  attnum_offset = 0;
4082 
4083  /* see what actually matched */
4084  foreach(lc2, *varinfos)
4085  {
4086  ListCell *lc3;
4087  int idx;
4088  bool found = false;
4089 
4090  GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4091 
4092  /*
4093  * Process a simple Var expression, by matching it to keys
4094  * directly. If there's a matching expression, we'll try matching
4095  * it later.
4096  */
4097  if (IsA(varinfo->var, Var))
4098  {
4099  AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4100 
4101  /*
4102  * Ignore expressions on system attributes. Can't rely on the
4103  * bms check for negative values.
4104  */
4106  continue;
4107 
4108  /* Is the variable covered by the statistics object? */
4109  if (!bms_is_member(attnum, matched_info->keys))
4110  continue;
4111 
4112  attnum = attnum + attnum_offset;
4113 
4114  /* ensure sufficient offset */
4116 
4117  matched = bms_add_member(matched, attnum);
4118 
4119  found = true;
4120  }
4121 
4122  /*
4123  * XXX Maybe we should allow searching the expressions even if we
4124  * found an attribute matching the expression? That would handle
4125  * trivial expressions like "(a)" but it seems fairly useless.
4126  */
4127  if (found)
4128  continue;
4129 
4130  /* expression - see if it's in the statistics object */
4131  idx = 0;
4132  foreach(lc3, matched_info->exprs)
4133  {
4134  Node *expr = (Node *) lfirst(lc3);
4135 
4136  if (equal(varinfo->var, expr))
4137  {
4138  AttrNumber attnum = -(idx + 1);
4139 
4140  attnum = attnum + attnum_offset;
4141 
4142  /* ensure sufficient offset */
4144 
4145  matched = bms_add_member(matched, attnum);
4146 
4147  /* there should be just one matching expression */
4148  break;
4149  }
4150 
4151  idx++;
4152  }
4153  }
4154 
4155  /* Find the specific item that exactly matches the combination */
4156  for (i = 0; i < stats->nitems; i++)
4157  {
4158  int j;
4159  MVNDistinctItem *tmpitem = &stats->items[i];
4160 
4161  if (tmpitem->nattributes != bms_num_members(matched))
4162  continue;
4163 
4164  /* assume it's the right item */
4165  item = tmpitem;
4166 
4167  /* check that all item attributes/expressions fit the match */
4168  for (j = 0; j < tmpitem->nattributes; j++)
4169  {
4170  AttrNumber attnum = tmpitem->attributes[j];
4171 
4172  /*
4173  * Thanks to how we constructed the matched bitmap above, we
4174  * can just offset all attnums the same way.
4175  */
4176  attnum = attnum + attnum_offset;
4177 
4178  if (!bms_is_member(attnum, matched))
4179  {
4180  /* nah, it's not this item */
4181  item = NULL;
4182  break;
4183  }
4184  }
4185 
4186  /*
4187  * If the item has all the matched attributes, we know it's the
4188  * right one - there can't be a better one. matching more.
4189  */
4190  if (item)
4191  break;
4192  }
4193 
4194  /*
4195  * Make sure we found an item. There has to be one, because ndistinct
4196  * statistics includes all combinations of attributes.
4197  */
4198  if (!item)
4199  elog(ERROR, "corrupt MVNDistinct entry");
4200 
4201  /* Form the output varinfo list, keeping only unmatched ones */
4202  foreach(lc, *varinfos)
4203  {
4204  GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4205  ListCell *lc3;
4206  bool found = false;
4207 
4208  /*
4209  * Let's look at plain variables first, because it's the most
4210  * common case and the check is quite cheap. We can simply get the
4211  * attnum and check (with an offset) matched bitmap.
4212  */
4213  if (IsA(varinfo->var, Var))
4214  {
4215  AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4216 
4217  /*
4218  * If it's a system attribute, we're done. We don't support
4219  * extended statistics on system attributes, so it's clearly
4220  * not matched. Just keep the expression and continue.
4221  */
4223  {
4224  newlist = lappend(newlist, varinfo);
4225  continue;
4226  }
4227 
4228  /* apply the same offset as above */
4229  attnum += attnum_offset;
4230 
4231  /* if it's not matched, keep the varinfo */
4232  if (!bms_is_member(attnum, matched))
4233  newlist = lappend(newlist, varinfo);
4234 
4235  /* The rest of the loop deals with complex expressions. */
4236  continue;
4237  }
4238 
4239  /*
4240  * Process complex expressions, not just simple Vars.
4241  *
4242  * First, we search for an exact match of an expression. If we
4243  * find one, we can just discard the whole GroupVarInfo, with all
4244  * the variables we extracted from it.
4245  *
4246  * Otherwise we inspect the individual vars, and try matching it
4247  * to variables in the item.
4248  */
4249  foreach(lc3, matched_info->exprs)
4250  {
4251  Node *expr = (Node *) lfirst(lc3);
4252 
4253  if (equal(varinfo->var, expr))
4254  {
4255  found = true;
4256  break;
4257  }
4258  }
4259 
4260  /* found exact match, skip */
4261  if (found)
4262  continue;
4263 
4264  newlist = lappend(newlist, varinfo);
4265  }
4266 
4267  *varinfos = newlist;
4268  *ndistinct = item->ndistinct;
4269  return true;
4270  }
4271 
4272  return false;
4273 }
Datum idx(PG_FUNCTION_ARGS)
Definition: _int_op.c:259
#define AttrNumberIsForUserDefinedAttr(attributeNumber)
Definition: attnum.h:41
int bms_num_members(const Bitmapset *a)
Definition: bitmapset.c:751
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:510
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:815
MVNDistinct * statext_ndistinct_load(Oid mvoid, bool inh)
Definition: mvdistinct.c:148
double ndistinct
Definition: statistics.h:28
AttrNumber * attributes
Definition: statistics.h:30
uint32 nitems
Definition: statistics.h:38
MVNDistinctItem items[FLEXIBLE_ARRAY_MEMBER]
Definition: statistics.h:39
List * statlist
Definition: pathnodes.h:927
Bitmapset * keys
Definition: pathnodes.h:1270
Definition: primnodes.h:234

References Assert(), attnum, MVNDistinctItem::attributes, AttrNumberIsForUserDefinedAttr, bms_add_member(), bms_is_member(), bms_num_members(), elog, equal(), ERROR, StatisticExtInfo::exprs, i, idx(), RangeTblEntry::inh, StatisticExtInfo::inherit, InvalidOid, IsA, MVNDistinct::items, j, StatisticExtInfo::keys, StatisticExtInfo::kind, lappend(), lfirst, list_length(), MVNDistinctItem::nattributes, MVNDistinctItem::ndistinct, NIL, MVNDistinct::nitems, planner_rt_fetch, GroupVarInfo::rel, RelOptInfo::relid, statext_ndistinct_load(), RelOptInfo::statlist, StatisticExtInfo::statOid, and GroupVarInfo::var.

Referenced by estimate_num_groups().

◆ estimate_num_groups()

double estimate_num_groups ( PlannerInfo root,
List groupExprs,
double  input_rows,
List **  pgset,
EstimationInfo estinfo 
)

Definition at line 3416 of file selfuncs.c.

3418 {
3419  List *varinfos = NIL;
3420  double srf_multiplier = 1.0;
3421  double numdistinct;
3422  ListCell *l;
3423  int i;
3424 
3425  /* Zero the estinfo output parameter, if non-NULL */
3426  if (estinfo != NULL)
3427  memset(estinfo, 0, sizeof(EstimationInfo));
3428 
3429  /*
3430  * We don't ever want to return an estimate of zero groups, as that tends
3431  * to lead to division-by-zero and other unpleasantness. The input_rows
3432  * estimate is usually already at least 1, but clamp it just in case it
3433  * isn't.
3434  */
3435  input_rows = clamp_row_est(input_rows);
3436 
3437  /*
3438  * If no grouping columns, there's exactly one group. (This can't happen
3439  * for normal cases with GROUP BY or DISTINCT, but it is possible for
3440  * corner cases with set operations.)
3441  */
3442  if (groupExprs == NIL || (pgset && *pgset == NIL))
3443  return 1.0;
3444 
3445  /*
3446  * Count groups derived from boolean grouping expressions. For other
3447  * expressions, find the unique Vars used, treating an expression as a Var
3448  * if we can find stats for it. For each one, record the statistical
3449  * estimate of number of distinct values (total in its table, without
3450  * regard for filtering).
3451  */
3452  numdistinct = 1.0;
3453 
3454  i = 0;
3455  foreach(l, groupExprs)
3456  {
3457  Node *groupexpr = (Node *) lfirst(l);
3458  double this_srf_multiplier;
3459  VariableStatData vardata;
3460  List *varshere;
3461  ListCell *l2;
3462 
3463  /* is expression in this grouping set? */
3464  if (pgset && !list_member_int(*pgset, i++))
3465  continue;
3466 
3467  /*
3468  * Set-returning functions in grouping columns are a bit problematic.
3469  * The code below will effectively ignore their SRF nature and come up
3470  * with a numdistinct estimate as though they were scalar functions.
3471  * We compensate by scaling up the end result by the largest SRF
3472  * rowcount estimate. (This will be an overestimate if the SRF
3473  * produces multiple copies of any output value, but it seems best to
3474  * assume the SRF's outputs are distinct. In any case, it's probably
3475  * pointless to worry too much about this without much better
3476  * estimates for SRF output rowcounts than we have today.)
3477  */
3478  this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3479  if (srf_multiplier < this_srf_multiplier)
3480  srf_multiplier = this_srf_multiplier;
3481 
3482  /* Short-circuit for expressions returning boolean */
3483  if (exprType(groupexpr) == BOOLOID)
3484  {
3485  numdistinct *= 2.0;
3486  continue;
3487  }
3488 
3489  /*
3490  * If examine_variable is able to deduce anything about the GROUP BY
3491  * expression, treat it as a single variable even if it's really more
3492  * complicated.
3493  *
3494  * XXX This has the consequence that if there's a statistics object on
3495  * the expression, we don't split it into individual Vars. This
3496  * affects our selection of statistics in
3497  * estimate_multivariate_ndistinct, because it's probably better to
3498  * use more accurate estimate for each expression and treat them as
3499  * independent, than to combine estimates for the extracted variables
3500  * when we don't know how that relates to the expressions.
3501  */
3502  examine_variable(root, groupexpr, 0, &vardata);
3503  if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3504  {
3505  varinfos = add_unique_group_var(root, varinfos,
3506  groupexpr, &vardata);
3507  ReleaseVariableStats(vardata);
3508  continue;
3509  }
3510  ReleaseVariableStats(vardata);
3511 
3512  /*
3513  * Else pull out the component Vars. Handle PlaceHolderVars by
3514  * recursing into their arguments (effectively assuming that the
3515  * PlaceHolderVar doesn't change the number of groups, which boils
3516  * down to ignoring the possible addition of nulls to the result set).
3517  */
3518  varshere = pull_var_clause(groupexpr,
3522 
3523  /*
3524  * If we find any variable-free GROUP BY item, then either it is a
3525  * constant (and we can ignore it) or it contains a volatile function;
3526  * in the latter case we punt and assume that each input row will
3527  * yield a distinct group.
3528  */
3529  if (varshere == NIL)
3530  {
3531  if (contain_volatile_functions(groupexpr))
3532  return input_rows;
3533  continue;
3534  }
3535 
3536  /*
3537  * Else add variables to varinfos list
3538  */
3539  foreach(l2, varshere)
3540  {
3541  Node *var = (Node *) lfirst(l2);
3542 
3543  examine_variable(root, var, 0, &vardata);
3544  varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3545  ReleaseVariableStats(vardata);
3546  }
3547  }
3548 
3549  /*
3550  * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3551  * list.
3552  */
3553  if (varinfos == NIL)
3554  {
3555  /* Apply SRF multiplier as we would do in the long path */
3556  numdistinct *= srf_multiplier;
3557  /* Round off */
3558  numdistinct = ceil(numdistinct);
3559  /* Guard against out-of-range answers */
3560  if (numdistinct > input_rows)
3561  numdistinct = input_rows;
3562  if (numdistinct < 1.0)
3563  numdistinct = 1.0;
3564  return numdistinct;
3565  }
3566 
3567  /*
3568  * Group Vars by relation and estimate total numdistinct.
3569  *
3570  * For each iteration of the outer loop, we process the frontmost Var in
3571  * varinfos, plus all other Vars in the same relation. We remove these
3572  * Vars from the newvarinfos list for the next iteration. This is the
3573  * easiest way to group Vars of same rel together.
3574  */
3575  do
3576  {
3577  GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3578  RelOptInfo *rel = varinfo1->rel;
3579  double reldistinct = 1;
3580  double relmaxndistinct = reldistinct;
3581  int relvarcount = 0;
3582  List *newvarinfos = NIL;
3583  List *relvarinfos = NIL;
3584 
3585  /*
3586  * Split the list of varinfos in two - one for the current rel, one
3587  * for remaining Vars on other rels.
3588  */
3589  relvarinfos = lappend(relvarinfos, varinfo1);
3590  for_each_from(l, varinfos, 1)
3591  {
3592  GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3593 
3594  if (varinfo2->rel == varinfo1->rel)
3595  {
3596  /* varinfos on current rel */
3597  relvarinfos = lappend(relvarinfos, varinfo2);
3598  }
3599  else
3600  {
3601  /* not time to process varinfo2 yet */
3602  newvarinfos = lappend(newvarinfos, varinfo2);
3603  }
3604  }
3605 
3606  /*
3607  * Get the numdistinct estimate for the Vars of this rel. We
3608  * iteratively search for multivariate n-distinct with maximum number
3609  * of vars; assuming that each var group is independent of the others,
3610  * we multiply them together. Any remaining relvarinfos after no more
3611  * multivariate matches are found are assumed independent too, so
3612  * their individual ndistinct estimates are multiplied also.
3613  *
3614  * While iterating, count how many separate numdistinct values we
3615  * apply. We apply a fudge factor below, but only if we multiplied
3616  * more than one such values.
3617  */
3618  while (relvarinfos)
3619  {
3620  double mvndistinct;
3621 
3622  if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3623  &mvndistinct))
3624  {
3625  reldistinct *= mvndistinct;
3626  if (relmaxndistinct < mvndistinct)
3627  relmaxndistinct = mvndistinct;
3628  relvarcount++;
3629  }
3630  else
3631  {
3632  foreach(l, relvarinfos)
3633  {
3634  GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3635 
3636  reldistinct *= varinfo2->ndistinct;
3637  if (relmaxndistinct < varinfo2->ndistinct)
3638  relmaxndistinct = varinfo2->ndistinct;
3639  relvarcount++;
3640 
3641  /*
3642  * When varinfo2's isdefault is set then we'd better set
3643  * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3644  */
3645  if (estinfo != NULL && varinfo2->isdefault)
3646  estinfo->flags |= SELFLAG_USED_DEFAULT;
3647  }
3648 
3649  /* we're done with this relation */
3650  relvarinfos = NIL;
3651  }
3652  }
3653 
3654  /*
3655  * Sanity check --- don't divide by zero if empty relation.
3656  */
3657  Assert(IS_SIMPLE_REL(rel));
3658  if (rel->tuples > 0)
3659  {
3660  /*
3661  * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3662  * fudge factor is because the Vars are probably correlated but we
3663  * don't know by how much. We should never clamp to less than the
3664  * largest ndistinct value for any of the Vars, though, since
3665  * there will surely be at least that many groups.
3666  */
3667  double clamp = rel->tuples;
3668 
3669  if (relvarcount > 1)
3670  {
3671  clamp *= 0.1;
3672  if (clamp < relmaxndistinct)
3673  {
3674  clamp = relmaxndistinct;
3675  /* for sanity in case some ndistinct is too large: */
3676  if (clamp > rel->tuples)
3677  clamp = rel->tuples;
3678  }
3679  }
3680  if (reldistinct > clamp)
3681  reldistinct = clamp;
3682 
3683  /*
3684  * Update the estimate based on the restriction selectivity,
3685  * guarding against division by zero when reldistinct is zero.
3686  * Also skip this if we know that we are returning all rows.
3687  */
3688  if (reldistinct > 0 && rel->rows < rel->tuples)
3689  {
3690  /*
3691  * Given a table containing N rows with n distinct values in a
3692  * uniform distribution, if we select p rows at random then
3693  * the expected number of distinct values selected is
3694  *
3695  * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3696  *
3697  * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3698  *
3699  * See "Approximating block accesses in database
3700  * organizations", S. B. Yao, Communications of the ACM,
3701  * Volume 20 Issue 4, April 1977 Pages 260-261.
3702  *
3703  * Alternatively, re-arranging the terms from the factorials,
3704  * this may be written as
3705  *
3706  * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3707  *
3708  * This form of the formula is more efficient to compute in
3709  * the common case where p is larger than N/n. Additionally,
3710  * as pointed out by Dell'Era, if i << N for all terms in the
3711  * product, it can be approximated by
3712  *
3713  * n * (1 - ((N-p)/N)^(N/n))
3714  *
3715  * See "Expected distinct values when selecting from a bag
3716  * without replacement", Alberto Dell'Era,
3717  * http://www.adellera.it/investigations/distinct_balls/.
3718  *
3719  * The condition i << N is equivalent to n >> 1, so this is a
3720  * good approximation when the number of distinct values in
3721  * the table is large. It turns out that this formula also
3722  * works well even when n is small.
3723  */
3724  reldistinct *=
3725  (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3726  rel->tuples / reldistinct));
3727  }
3728  reldistinct = clamp_row_est(reldistinct);
3729 
3730  /*
3731  * Update estimate of total distinct groups.
3732  */
3733  numdistinct *= reldistinct;
3734  }
3735 
3736  varinfos = newvarinfos;
3737  } while (varinfos != NIL);
3738 
3739  /* Now we can account for the effects of any SRFs */
3740  numdistinct *= srf_multiplier;
3741 
3742  /* Round off */
3743  numdistinct = ceil(numdistinct);
3744 
3745  /* Guard against out-of-range answers */
3746  if (numdistinct > input_rows)
3747  numdistinct = input_rows;
3748  if (numdistinct < 1.0)
3749  numdistinct = 1.0;
3750 
3751  return numdistinct;
3752 }
bool contain_volatile_functions(Node *clause)
Definition: clauses.c:518
double expression_returns_set_rows(PlannerInfo *root, Node *clause)
Definition: clauses.c:288
bool list_member_int(const List *list, int datum)
Definition: list.c:702
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:42
#define PVC_RECURSE_AGGREGATES
Definition: optimizer.h:187
#define PVC_RECURSE_PLACEHOLDERS
Definition: optimizer.h:191
#define PVC_RECURSE_WINDOWFUNCS
Definition: optimizer.h:189
#define IS_SIMPLE_REL(rel)
Definition: pathnodes.h:824
#define for_each_from(cell, lst, N)
Definition: pg_list.h:414
#define linitial(l)
Definition: pg_list.h:178
static bool estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel, List **varinfos, double *ndistinct)
Definition: selfuncs.c:3954
static List * add_unique_group_var(PlannerInfo *root, List *varinfos, Node *var, VariableStatData *vardata)
Definition: selfuncs.c:3296
#define SELFLAG_USED_DEFAULT
Definition: selfuncs.h:76
uint32 flags
Definition: selfuncs.h:80
List * pull_var_clause(Node *node, int flags)
Definition: var.c:607

References add_unique_group_var(), Assert(), clamp_row_est(), contain_volatile_functions(), estimate_multivariate_ndistinct(), examine_variable(), expression_returns_set_rows(), exprType(), EstimationInfo::flags, for_each_from, HeapTupleIsValid, i, IS_SIMPLE_REL, GroupVarInfo::isdefault, VariableStatData::isunique, lappend(), lfirst, linitial, list_member_int(), GroupVarInfo::ndistinct, NIL, pull_var_clause(), PVC_RECURSE_AGGREGATES, PVC_RECURSE_PLACEHOLDERS, PVC_RECURSE_WINDOWFUNCS, GroupVarInfo::rel, ReleaseVariableStats, RelOptInfo::rows, SELFLAG_USED_DEFAULT, VariableStatData::statsTuple, and RelOptInfo::tuples.

Referenced by adjust_rowcount_for_semijoins(), cost_incremental_sort(), cost_memoize_rescan(), create_final_distinct_paths(), create_partial_distinct_paths(), create_unique_path(), estimate_path_cost_size(), get_number_of_groups(), get_windowclause_startup_tuples(), and recurse_set_operations().

◆ examine_simple_variable()

static void examine_simple_variable ( PlannerInfo root,
Var var,
VariableStatData vardata 
)
static

Definition at line 5406 of file selfuncs.c.

5408 {
5409  RangeTblEntry *rte = root->simple_rte_array[var->varno];
5410 
5411  Assert(IsA(rte, RangeTblEntry));
5412 
5414  (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5415  {
5416  /*
5417  * The hook took control of acquiring a stats tuple. If it did supply
5418  * a tuple, it'd better have supplied a freefunc.
5419  */
5420  if (HeapTupleIsValid(vardata->statsTuple) &&
5421  !vardata->freefunc)
5422  elog(ERROR, "no function provided to release variable stats with");
5423  }
5424  else if (rte->rtekind == RTE_RELATION)
5425  {
5426  /*
5427  * Plain table or parent of an inheritance appendrel, so look up the
5428  * column in pg_statistic
5429  */
5430  vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5431  ObjectIdGetDatum(rte->relid),
5432  Int16GetDatum(var->varattno),
5433  BoolGetDatum(rte->inh));
5434  vardata->freefunc = ReleaseSysCache;
5435 
5436  if (HeapTupleIsValid(vardata->statsTuple))
5437  {
5438  RelOptInfo *onerel = find_base_rel_noerr(root, var->varno);
5439  Oid userid;
5440 
5441  /*
5442  * Check if user has permission to read this column. We require
5443  * all rows to be accessible, so there must be no securityQuals
5444  * from security barrier views or RLS policies.
5445  *
5446  * Normally the Var will have an associated RelOptInfo from which
5447  * we can find out which userid to do the check as; but it might
5448  * not if it's a RETURNING Var for an INSERT target relation. In
5449  * that case use the RTEPermissionInfo associated with the RTE.
5450  */
5451  if (onerel)
5452  userid = onerel->userid;
5453  else
5454  {
5455  RTEPermissionInfo *perminfo;
5456 
5457  perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5458  userid = perminfo->checkAsUser;
5459  }
5460  if (!OidIsValid(userid))
5461  userid = GetUserId();
5462 
5463  vardata->acl_ok =
5464  rte->securityQuals == NIL &&
5465  ((pg_class_aclcheck(rte->relid, userid,
5466  ACL_SELECT) == ACLCHECK_OK) ||
5467  (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
5468  ACL_SELECT) == ACLCHECK_OK));
5469 
5470  /*
5471  * If the user doesn't have permissions to access an inheritance
5472  * child relation or specifically this attribute, check the
5473  * permissions of the table/column actually mentioned in the
5474  * query, since most likely the user does have that permission
5475  * (else the query will fail at runtime), and if the user can read
5476  * the column there then he can get the values of the child table
5477  * too. To do that, we must find out which of the root parent's
5478  * attributes the child relation's attribute corresponds to.
5479  */
5480  if (!vardata->acl_ok && var->varattno > 0 &&
5481  root->append_rel_array != NULL)
5482  {
5483  AppendRelInfo *appinfo;
5484  Index varno = var->varno;
5485  int varattno = var->varattno;
5486  bool found = false;
5487 
5488  appinfo = root->append_rel_array[varno];
5489 
5490  /*
5491  * Partitions are mapped to their immediate parent, not the
5492  * root parent, so must be ready to walk up multiple
5493  * AppendRelInfos. But stop if we hit a parent that is not
5494  * RTE_RELATION --- that's a flattened UNION ALL subquery, not
5495  * an inheritance parent.
5496  */
5497  while (appinfo &&
5498  planner_rt_fetch(appinfo->parent_relid,
5499  root)->rtekind == RTE_RELATION)
5500  {
5501  int parent_varattno;
5502 
5503  found = false;
5504  if (varattno <= 0 || varattno > appinfo->num_child_cols)
5505  break; /* safety check */
5506  parent_varattno = appinfo->parent_colnos[varattno - 1];
5507  if (parent_varattno == 0)
5508  break; /* Var is local to child */
5509 
5510  varno = appinfo->parent_relid;
5511  varattno = parent_varattno;
5512  found = true;
5513 
5514  /* If the parent is itself a child, continue up. */
5515  appinfo = root->append_rel_array[varno];
5516  }
5517 
5518  /*
5519  * In rare cases, the Var may be local to the child table, in
5520  * which case, we've got to live with having no access to this
5521  * column's stats.
5522  */
5523  if (!found)
5524  return;
5525 
5526  /* Repeat the access check on this parent rel & column */
5527  rte = planner_rt_fetch(varno, root);
5528  Assert(rte->rtekind == RTE_RELATION);
5529 
5530  /*
5531  * Fine to use the same userid as it's the same in all
5532  * relations of a given inheritance tree.
5533  */
5534  vardata->acl_ok =
5535  rte->securityQuals == NIL &&
5536  ((pg_class_aclcheck(rte->relid, userid,
5537  ACL_SELECT) == ACLCHECK_OK) ||
5538  (pg_attribute_aclcheck(rte->relid, varattno, userid,
5539  ACL_SELECT) == ACLCHECK_OK));
5540  }
5541  }
5542  else
5543  {
5544  /* suppress any possible leakproofness checks later */
5545  vardata->acl_ok = true;
5546  }
5547  }
5548  else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5549  (rte->rtekind == RTE_CTE && !rte->self_reference))
5550  {
5551  /*
5552  * Plain subquery (not one that was converted to an appendrel) or
5553  * non-recursive CTE. In either case, we can try to find out what the
5554  * Var refers to within the subquery. We skip this for appendrel and
5555  * recursive-CTE cases because any column stats we did find would
5556  * likely not be very relevant.
5557  */
5558  PlannerInfo *subroot;
5559  Query *subquery;
5560  List *subtlist;
5561  TargetEntry *ste;
5562 
5563  /*
5564  * Punt if it's a whole-row var rather than a plain column reference.
5565  */
5566  if (var->varattno == InvalidAttrNumber)
5567  return;
5568 
5569  /*
5570  * Otherwise, find the subquery's planner subroot.
5571  */
5572  if (rte->rtekind == RTE_SUBQUERY)
5573  {
5574  RelOptInfo *rel;
5575 
5576  /*
5577  * Fetch RelOptInfo for subquery. Note that we don't change the
5578  * rel returned in vardata, since caller expects it to be a rel of
5579  * the caller's query level. Because we might already be
5580  * recursing, we can't use that rel pointer either, but have to
5581  * look up the Var's rel afresh.
5582  */
5583  rel = find_base_rel(root, var->varno);
5584 
5585  subroot = rel->subroot;
5586  }
5587  else
5588  {
5589  /* CTE case is more difficult */
5590  PlannerInfo *cteroot;
5591  Index levelsup;
5592  int ndx;
5593  int plan_id;
5594  ListCell *lc;
5595 
5596  /*
5597  * Find the referenced CTE, and locate the subroot previously made
5598  * for it.
5599  */
5600  levelsup = rte->ctelevelsup;
5601  cteroot = root;
5602  while (levelsup-- > 0)
5603  {
5604  cteroot = cteroot->parent_root;
5605  if (!cteroot) /* shouldn't happen */
5606  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5607  }
5608 
5609  /*
5610  * Note: cte_plan_ids can be shorter than cteList, if we are still
5611  * working on planning the CTEs (ie, this is a side-reference from
5612  * another CTE). So we mustn't use forboth here.
5613  */
5614  ndx = 0;
5615  foreach(lc, cteroot->parse->cteList)
5616  {
5617  CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5618 
5619  if (strcmp(cte->ctename, rte->ctename) == 0)
5620  break;
5621  ndx++;
5622  }
5623  if (lc == NULL) /* shouldn't happen */
5624  elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5625  if (ndx >= list_length(cteroot->cte_plan_ids))
5626  elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5627  plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5628  if (plan_id <= 0)
5629  elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5630  subroot = list_nth(root->glob->subroots, plan_id - 1);
5631  }
5632 
5633  /* If the subquery hasn't been planned yet, we have to punt */
5634  if (subroot == NULL)
5635  return;
5636  Assert(IsA(subroot, PlannerInfo));
5637 
5638  /*
5639  * We must use the subquery parsetree as mangled by the planner, not
5640  * the raw version from the RTE, because we need a Var that will refer
5641  * to the subroot's live RelOptInfos. For instance, if any subquery
5642  * pullup happened during planning, Vars in the targetlist might have
5643  * gotten replaced, and we need to see the replacement expressions.
5644  */
5645  subquery = subroot->parse;
5646  Assert(IsA(subquery, Query));
5647 
5648  /*
5649  * Punt if subquery uses set operations or GROUP BY, as these will
5650  * mash underlying columns' stats beyond recognition. (Set ops are
5651  * particularly nasty; if we forged ahead, we would return stats
5652  * relevant to only the leftmost subselect...) DISTINCT is also
5653  * problematic, but we check that later because there is a possibility
5654  * of learning something even with it.
5655  */
5656  if (subquery->setOperations ||
5657  subquery->groupClause ||
5658  subquery->groupingSets)
5659  return;
5660 
5661  /* Get the subquery output expression referenced by the upper Var */
5662  if (subquery->returningList)
5663  subtlist = subquery->returningList;
5664  else
5665  subtlist = subquery->targetList;
5666  ste = get_tle_by_resno(subtlist, var->varattno);
5667  if (ste == NULL || ste->resjunk)
5668  elog(ERROR, "subquery %s does not have attribute %d",
5669  rte->eref->aliasname, var->varattno);
5670  var = (Var *) ste->expr;
5671 
5672  /*
5673  * If subquery uses DISTINCT, we can't make use of any stats for the
5674  * variable ... but, if it's the only DISTINCT column, we are entitled
5675  * to consider it unique. We do the test this way so that it works
5676  * for cases involving DISTINCT ON.
5677  */
5678  if (subquery->distinctClause)
5679  {
5680  if (list_length(subquery->distinctClause) == 1 &&
5681  targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5682  vardata->isunique = true;
5683  /* cannot go further */
5684  return;
5685  }
5686 
5687  /*
5688  * If the sub-query originated from a view with the security_barrier
5689  * attribute, we must not look at the variable's statistics, though it
5690  * seems all right to notice the existence of a DISTINCT clause. So
5691  * stop here.
5692  *
5693  * This is probably a harsher restriction than necessary; it's
5694  * certainly OK for the selectivity estimator (which is a C function,
5695  * and therefore omnipotent anyway) to look at the statistics. But
5696  * many selectivity estimators will happily *invoke the operator
5697  * function* to try to work out a good estimate - and that's not OK.
5698  * So for now, don't dig down for stats.
5699  */
5700  if (rte->security_barrier)
5701  return;
5702 
5703  /* Can only handle a simple Var of subquery's query level */
5704  if (var && IsA(var, Var) &&
5705  var->varlevelsup == 0)
5706  {
5707  /*
5708  * OK, recurse into the subquery. Note that the original setting
5709  * of vardata->isunique (which will surely be false) is left
5710  * unchanged in this situation. That's what we want, since even
5711  * if the underlying column is unique, the subquery may have
5712  * joined to other tables in a way that creates duplicates.
5713  */
5714  examine_simple_variable(subroot, var, vardata);
5715  }
5716  }
5717  else
5718  {
5719  /*
5720  * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5721  * see RTE_JOIN here because join alias Vars have already been
5722  * flattened.) There's not much we can do with function outputs, but
5723  * maybe someday try to be smarter about VALUES.
5724  */
5725  }
5726 }
@ ACLCHECK_OK
Definition: acl.h:183
AclResult pg_attribute_aclcheck(Oid table_oid, AttrNumber attnum, Oid roleid, AclMode mode)
Definition: aclchk.c:3908
AclResult pg_class_aclcheck(Oid table_oid, Oid roleid, AclMode mode)
Definition: aclchk.c:4079
#define InvalidAttrNumber
Definition: attnum.h:23
unsigned int Index
Definition: c.h:601
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:77
Oid GetUserId(void)
Definition: miscinit.c:514
bool targetIsInSortList(TargetEntry *tle, Oid sortop, List *sortList)
RTEPermissionInfo * getRTEPermissionInfo(List *rteperminfos, RangeTblEntry *rte)
TargetEntry * get_tle_by_resno(List *tlist, AttrNumber resno)
@ RTE_CTE
Definition: parsenodes.h:1017
@ RTE_SUBQUERY
Definition: parsenodes.h:1012
#define ACL_SELECT
Definition: parsenodes.h:77
static void * list_nth(const List *list, int n)
Definition: pg_list.h:299
static int list_nth_int(const List *list, int n)
Definition: pg_list.h:310
RelOptInfo * find_base_rel(PlannerInfo *root, int relid)
Definition: relnode.c:407
RelOptInfo * find_base_rel_noerr(PlannerInfo *root, int relid)
Definition: relnode.c:429
static void examine_simple_variable(PlannerInfo *root, Var *var, VariableStatData *vardata)
Definition: selfuncs.c:5406
char * aliasname
Definition: primnodes.h:50
Index parent_relid
Definition: pathnodes.h:2943
int num_child_cols
Definition: pathnodes.h:2979
List * cte_plan_ids
Definition: pathnodes.h:302
PlannerGlobal * glob
Definition: pathnodes.h:202
Query * parse
Definition: pathnodes.h:199
List * returningList
Definition: parsenodes.h:197
Node * setOperations
Definition: parsenodes.h:218
List * cteList
Definition: parsenodes.h:166
List * groupClause
Definition: parsenodes.h:199
List * targetList
Definition: parsenodes.h:190
List * groupingSets
Definition: parsenodes.h:202
List * distinctClause
Definition: parsenodes.h:208
char * ctename
Definition: parsenodes.h:1169
bool self_reference
Definition: parsenodes.h:1171
bool security_barrier
Definition: parsenodes.h:1087
Index ctelevelsup
Definition: parsenodes.h:1170
List * securityQuals
Definition: parsenodes.h:1208
Alias * eref
Definition: parsenodes.h:1205
Oid userid
Definition: pathnodes.h:947
PlannerInfo * subroot
Definition: pathnodes.h:934
Expr * expr
Definition: primnodes.h:1943
AttrNumber varattno
Definition: primnodes.h:246
int varno
Definition: primnodes.h:241
Index varlevelsup
Definition: primnodes.h:266

References VariableStatData::acl_ok, ACL_SELECT, ACLCHECK_OK, Alias::aliasname, Assert(), BoolGetDatum(), RTEPermissionInfo::checkAsUser, PlannerInfo::cte_plan_ids, RangeTblEntry::ctelevelsup, Query::cteList, RangeTblEntry::ctename, CommonTableExpr::ctename, Query::distinctClause, elog, RangeTblEntry::eref, ERROR, TargetEntry::expr, find_base_rel(), find_base_rel_noerr(), VariableStatData::freefunc, get_relation_stats_hook, get_tle_by_resno(), getRTEPermissionInfo(), GetUserId(), PlannerInfo::glob, Query::groupClause, Query::groupingSets, HeapTupleIsValid, if(), RangeTblEntry::inh, Int16GetDatum(), InvalidAttrNumber, InvalidOid, IsA, VariableStatData::isunique, lfirst, list_length(), list_nth(), list_nth_int(), NIL, AppendRelInfo::num_child_cols, ObjectIdGetDatum(), OidIsValid, AppendRelInfo::parent_relid, PlannerInfo::parse, pg_attribute_aclcheck(), pg_class_aclcheck(), planner_rt_fetch, ReleaseSysCache(), RangeTblEntry::relid, Query::returningList, RTE_CTE, RTE_RELATION, RTE_SUBQUERY, RangeTblEntry::rtekind, SearchSysCache3(), RangeTblEntry::security_barrier, RangeTblEntry::securityQuals, RangeTblEntry::self_reference, Query::setOperations, VariableStatData::statsTuple, RelOptInfo::subroot, targetIsInSortList(), Query::targetList, RelOptInfo::userid, Var::varattno, Var::varlevelsup, and Var::varno.

Referenced by examine_variable().

◆ examine_variable()

void examine_variable ( PlannerInfo root,
Node node,
int  varRelid,
VariableStatData vardata 
)

Definition at line 5012 of file selfuncs.c.

5014 {
5015  Node *basenode;
5016  Relids varnos;
5017  RelOptInfo *onerel;
5018 
5019  /* Make sure we don't return dangling pointers in vardata */
5020  MemSet(vardata, 0, sizeof(VariableStatData));
5021 
5022  /* Save the exposed type of the expression */
5023  vardata->vartype = exprType(node);
5024 
5025  /* Look inside any binary-compatible relabeling */
5026 
5027  if (IsA(node, RelabelType))
5028  basenode = (Node *) ((RelabelType *) node)->arg;
5029  else
5030  basenode = node;
5031 
5032  /* Fast path for a simple Var */
5033 
5034  if (IsA(basenode, Var) &&
5035  (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5036  {
5037  Var *var = (Var *) basenode;
5038 
5039  /* Set up result fields other than the stats tuple */
5040  vardata->var = basenode; /* return Var without relabeling */
5041  vardata->rel = find_base_rel(root, var->varno);
5042  vardata->atttype = var->vartype;
5043  vardata->atttypmod = var->vartypmod;
5044  vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5045 
5046  /* Try to locate some stats */
5047  examine_simple_variable(root, var, vardata);
5048 
5049  return;
5050  }
5051 
5052  /*
5053  * Okay, it's a more complicated expression. Determine variable
5054  * membership. Note that when varRelid isn't zero, only vars of that
5055  * relation are considered "real" vars.
5056  */
5057  varnos = pull_varnos(root, basenode);
5058 
5059  onerel = NULL;
5060 
5061  if (bms_is_empty(varnos))
5062  {
5063  /* No Vars at all ... must be pseudo-constant clause */
5064  }
5065  else
5066  {
5067  int relid;
5068 
5069  if (bms_get_singleton_member(varnos, &relid))
5070  {
5071  if (varRelid == 0 || varRelid == relid)
5072  {
5073  onerel = find_base_rel(root, relid);
5074  vardata->rel = onerel;
5075  node = basenode; /* strip any relabeling */
5076  }
5077  /* else treat it as a constant */
5078  }
5079  else
5080  {
5081  /* varnos has multiple relids */
5082  if (varRelid == 0)
5083  {
5084  /* treat it as a variable of a join relation */
5085  vardata->rel = find_join_rel(root, varnos);
5086  node = basenode; /* strip any relabeling */
5087  }
5088  else if (bms_is_member(varRelid, varnos))
5089  {
5090  /* ignore the vars belonging to other relations */
5091  vardata->rel = find_base_rel(root, varRelid);
5092  node = basenode; /* strip any relabeling */
5093  /* note: no point in expressional-index search here */
5094  }
5095  /* else treat it as a constant */
5096  }
5097  }
5098 
5099  bms_free(varnos);
5100 
5101  vardata->var = node;
5102  vardata->atttype = exprType(node);
5103  vardata->atttypmod = exprTypmod(node);
5104 
5105  if (onerel)
5106  {
5107  /*
5108  * We have an expression in vars of a single relation. Try to match
5109  * it to expressional index columns, in hopes of finding some
5110  * statistics.
5111  *
5112  * Note that we consider all index columns including INCLUDE columns,
5113  * since there could be stats for such columns. But the test for
5114  * uniqueness needs to be warier.
5115  *
5116  * XXX it's conceivable that there are multiple matches with different
5117  * index opfamilies; if so, we need to pick one that matches the
5118  * operator we are estimating for. FIXME later.
5119  */
5120  ListCell *ilist;
5121  ListCell *slist;
5122  Oid userid;
5123 
5124  /*
5125  * Determine the user ID to use for privilege checks: either
5126  * onerel->userid if it's set (e.g., in case we're accessing the table
5127  * via a view), or the current user otherwise.
5128  *
5129  * If we drill down to child relations, we keep using the same userid:
5130  * it's going to be the same anyway, due to how we set up the relation
5131  * tree (q.v. build_simple_rel).
5132  */
5133  userid = OidIsValid(onerel->userid) ? onerel->userid : GetUserId();
5134 
5135  foreach(ilist, onerel->indexlist)
5136  {
5137  IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5138  ListCell *indexpr_item;
5139  int pos;
5140 
5141  indexpr_item = list_head(index->indexprs);
5142  if (indexpr_item == NULL)
5143  continue; /* no expressions here... */
5144 
5145  for (pos = 0; pos < index->ncolumns; pos++)
5146  {
5147  if (index->indexkeys[pos] == 0)
5148  {
5149  Node *indexkey;
5150 
5151  if (indexpr_item == NULL)
5152  elog(ERROR, "too few entries in indexprs list");
5153  indexkey = (Node *) lfirst(indexpr_item);
5154  if (indexkey && IsA(indexkey, RelabelType))
5155  indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5156  if (equal(node, indexkey))
5157  {
5158  /*
5159  * Found a match ... is it a unique index? Tests here
5160  * should match has_unique_index().
5161  */
5162  if (index->unique &&
5163  index->nkeycolumns == 1 &&
5164  pos == 0 &&
5165  (index->indpred == NIL || index->predOK))
5166  vardata->isunique = true;
5167 
5168  /*
5169  * Has it got stats? We only consider stats for
5170  * non-partial indexes, since partial indexes probably
5171  * don't reflect whole-relation statistics; the above
5172  * check for uniqueness is the only info we take from
5173  * a partial index.
5174  *
5175  * An index stats hook, however, must make its own
5176  * decisions about what to do with partial indexes.
5177  */
5179  (*get_index_stats_hook) (root, index->indexoid,
5180  pos + 1, vardata))
5181  {
5182  /*
5183  * The hook took control of acquiring a stats
5184  * tuple. If it did supply a tuple, it'd better
5185  * have supplied a freefunc.
5186  */
5187  if (HeapTupleIsValid(vardata->statsTuple) &&
5188  !vardata->freefunc)
5189  elog(ERROR, "no function provided to release variable stats with");
5190  }
5191  else if (index->indpred == NIL)
5192  {
5193  vardata->statsTuple =
5194  SearchSysCache3(STATRELATTINH,
5195  ObjectIdGetDatum(index->indexoid),
5196  Int16GetDatum(pos + 1),
5197  BoolGetDatum(false));
5198  vardata->freefunc = ReleaseSysCache;
5199 
5200  if (HeapTupleIsValid(vardata->statsTuple))
5201  {
5202  /* Get index's table for permission check */
5203  RangeTblEntry *rte;
5204 
5205  rte = planner_rt_fetch(index->rel->relid, root);
5206  Assert(rte->rtekind == RTE_RELATION);
5207 
5208  /*
5209  * For simplicity, we insist on the whole
5210  * table being selectable, rather than trying
5211  * to identify which column(s) the index
5212  * depends on. Also require all rows to be
5213  * selectable --- there must be no
5214  * securityQuals from security barrier views
5215  * or RLS policies.
5216  */
5217  vardata->acl_ok =
5218  rte->securityQuals == NIL &&
5219  (pg_class_aclcheck(rte->relid, userid,
5220  ACL_SELECT) == ACLCHECK_OK);
5221 
5222  /*
5223  * If the user doesn't have permissions to
5224  * access an inheritance child relation, check
5225  * the permissions of the table actually
5226  * mentioned in the query, since most likely
5227  * the user does have that permission. Note
5228  * that whole-table select privilege on the
5229  * parent doesn't quite guarantee that the
5230  * user could read all columns of the child.
5231  * But in practice it's unlikely that any
5232  * interesting security violation could result
5233  * from allowing access to the expression
5234  * index's stats, so we allow it anyway. See
5235  * similar code in examine_simple_variable()
5236  * for additional comments.
5237  */
5238  if (!vardata->acl_ok &&
5239  root->append_rel_array != NULL)
5240  {
5241  AppendRelInfo *appinfo;
5242  Index varno = index->rel->relid;
5243 
5244  appinfo = root->append_rel_array[varno];
5245  while (appinfo &&
5246  planner_rt_fetch(appinfo->parent_relid,
5247  root)->rtekind == RTE_RELATION)
5248  {
5249  varno = appinfo->parent_relid;
5250  appinfo = root->append_rel_array[varno];
5251  }
5252  if (varno != index->rel->relid)
5253  {
5254  /* Repeat access check on this rel */
5255  rte = planner_rt_fetch(varno, root);
5256  Assert(rte->rtekind == RTE_RELATION);
5257 
5258  vardata->acl_ok =
5259  rte->securityQuals == NIL &&
5260  (pg_class_aclcheck(rte->relid,
5261  userid,
5262  ACL_SELECT) == ACLCHECK_OK);
5263  }
5264  }
5265  }
5266  else
5267  {
5268  /* suppress leakproofness checks later */
5269  vardata->acl_ok = true;
5270  }
5271  }
5272  if (vardata->statsTuple)
5273  break;
5274  }
5275  indexpr_item = lnext(index->indexprs, indexpr_item);
5276  }
5277  }
5278  if (vardata->statsTuple)
5279  break;
5280  }
5281 
5282  /*
5283  * Search extended statistics for one with a matching expression.
5284  * There might be multiple ones, so just grab the first one. In the
5285  * future, we might consider the statistics target (and pick the most
5286  * accurate statistics) and maybe some other parameters.
5287  */
5288  foreach(slist, onerel->statlist)
5289  {
5290  StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5291  RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5292  ListCell *expr_item;
5293  int pos;
5294 
5295  /*
5296  * Stop once we've found statistics for the expression (either
5297  * from extended stats, or for an index in the preceding loop).
5298  */
5299  if (vardata->statsTuple)
5300  break;
5301 
5302  /* skip stats without per-expression stats */
5303  if (info->kind != STATS_EXT_EXPRESSIONS)
5304  continue;
5305 
5306  /* skip stats with mismatching stxdinherit value */
5307  if (info->inherit != rte->inh)
5308  continue;
5309 
5310  pos = 0;
5311  foreach(expr_item, info->exprs)
5312  {
5313  Node *expr = (Node *) lfirst(expr_item);
5314 
5315  Assert(expr);
5316 
5317  /* strip RelabelType before comparing it */
5318  if (expr && IsA(expr, RelabelType))
5319  expr = (Node *) ((RelabelType *) expr)->arg;
5320 
5321  /* found a match, see if we can extract pg_statistic row */
5322  if (equal(node, expr))
5323  {
5324  /*
5325  * XXX Not sure if we should cache the tuple somewhere.
5326  * Now we just create a new copy every time.
5327  */
5328  vardata->statsTuple =
5329  statext_expressions_load(info->statOid, rte->inh, pos);
5330 
5331  vardata->freefunc = ReleaseDummy;
5332 
5333  /*
5334  * For simplicity, we insist on the whole table being
5335  * selectable, rather than trying to identify which
5336  * column(s) the statistics object depends on. Also
5337  * require all rows to be selectable --- there must be no
5338  * securityQuals from security barrier views or RLS
5339  * policies.
5340  */
5341  vardata->acl_ok =
5342  rte->securityQuals == NIL &&
5343  (pg_class_aclcheck(rte->relid, userid,
5344  ACL_SELECT) == ACLCHECK_OK);
5345 
5346  /*
5347  * If the user doesn't have permissions to access an
5348  * inheritance child relation, check the permissions of
5349  * the table actually mentioned in the query, since most
5350  * likely the user does have that permission. Note that
5351  * whole-table select privilege on the parent doesn't
5352  * quite guarantee that the user could read all columns of
5353  * the child. But in practice it's unlikely that any
5354  * interesting security violation could result from
5355  * allowing access to the expression stats, so we allow it
5356  * anyway. See similar code in examine_simple_variable()
5357  * for additional comments.
5358  */
5359  if (!vardata->acl_ok &&
5360  root->append_rel_array != NULL)
5361  {
5362  AppendRelInfo *appinfo;
5363  Index varno = onerel->relid;
5364 
5365  appinfo = root->append_rel_array[varno];
5366  while (appinfo &&
5367  planner_rt_fetch(appinfo->parent_relid,
5368  root)->rtekind == RTE_RELATION)
5369  {
5370  varno = appinfo->parent_relid;
5371  appinfo = root->append_rel_array[varno];
5372  }
5373  if (varno != onerel->relid)
5374  {
5375  /* Repeat access check on this rel */
5376  rte = planner_rt_fetch(varno, root);
5377  Assert(rte->rtekind == RTE_RELATION);
5378 
5379  vardata->acl_ok =
5380  rte->securityQuals == NIL &&
5381  (pg_class_aclcheck(rte->relid,
5382  userid,
5383  ACL_SELECT) == ACLCHECK_OK);
5384  }
5385  }
5386 
5387  break;
5388  }
5389 
5390  pos++;
5391  }
5392  }
5393  }
5394 }
void bms_free(Bitmapset *a)
Definition: bitmapset.c:239
bool bms_get_singleton_member(const Bitmapset *a, int *member)
Definition: bitmapset.c:715
#define bms_is_empty(a)
Definition: bitmapset.h:118
#define MemSet(start, val, len)
Definition: c.h:1007
HeapTuple statext_expressions_load(Oid stxoid, bool inh, int idx)
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:284
static ListCell * list_head(const List *l)
Definition: pg_list.h:128
static ListCell * lnext(const List *l, const ListCell *c)
Definition: pg_list.h:343
bool has_unique_index(RelOptInfo *rel, AttrNumber attno)
Definition: plancat.c:2180
RelOptInfo * find_join_rel(PlannerInfo *root, Relids relids)
Definition: relnode.c:520
static void ReleaseDummy(HeapTuple tuple)
Definition: selfuncs.c:4971
List * indexlist
Definition: pathnodes.h:925
int32 atttypmod
Definition: selfuncs.h:94
Relids pull_varnos(PlannerInfo *root, Node *node)
Definition: var.c:108

References VariableStatData::acl_ok, ACL_SELECT, ACLCHECK_OK, arg, Assert(), VariableStatData::atttype, VariableStatData::atttypmod, bms_free(), bms_get_singleton_member(), bms_is_empty, bms_is_member(), BoolGetDatum(), elog, equal(), ERROR, examine_simple_variable(), StatisticExtInfo::exprs, exprType(), exprTypmod(), find_base_rel(), find_join_rel(), VariableStatData::freefunc, get_index_stats_hook, GetUserId(), has_unique_index(), HeapTupleIsValid, if(), RelOptInfo::indexlist, RangeTblEntry::inh, StatisticExtInfo::inherit, Int16GetDatum(), IsA, VariableStatData::isunique, StatisticExtInfo::kind, lfirst, list_head(), lnext(), MemSet, NIL, ObjectIdGetDatum(), OidIsValid, AppendRelInfo::parent_relid, pg_class_aclcheck(), planner_rt_fetch, pull_varnos(), VariableStatData::rel, ReleaseDummy(), ReleaseSysCache(), RangeTblEntry::relid, RelOptInfo::relid, RTE_RELATION, RangeTblEntry::rtekind, SearchSysCache3(), RangeTblEntry::securityQuals, statext_expressions_load(), RelOptInfo::statlist, StatisticExtInfo::statOid, VariableStatData::statsTuple, RelOptInfo::userid, VariableStatData::var, Var::varattno, Var::varno, and VariableStatData::vartype.

Referenced by booltestsel(), boolvarsel(), estimate_array_length(), estimate_hash_bucket_stats(), estimate_num_groups(), get_join_variables(), get_restriction_variable(), mergejoinscansel(), nulltestsel(), and scalararraysel_containment().

◆ find_join_input_rel()

static RelOptInfo * find_join_input_rel ( PlannerInfo root,
Relids  relids 
)
static

Definition at line 6428 of file selfuncs.c.

6429 {
6430  RelOptInfo *rel = NULL;
6431 
6432  if (!bms_is_empty(relids))
6433  {
6434  int relid;
6435 
6436  if (bms_get_singleton_member(relids, &relid))
6437  rel = find_base_rel(root, relid);
6438  else
6439  rel = find_join_rel(root, relids);
6440  }
6441 
6442  if (rel == NULL)
6443  elog(ERROR, "could not find RelOptInfo for given relids");
6444 
6445  return rel;
6446 }

References bms_get_singleton_member(), bms_is_empty, elog, ERROR, find_base_rel(), and find_join_rel().

Referenced by eqjoinsel().

◆ generic_restriction_selectivity()

double generic_restriction_selectivity ( PlannerInfo root,
Oid  oproid,
Oid  collation,
List args,
int  varRelid,
double  default_selectivity 
)

Definition at line 914 of file selfuncs.c.

917 {
918  double selec;
919  VariableStatData vardata;
920  Node *other;
921  bool varonleft;
922 
923  /*
924  * If expression is not variable OP something or something OP variable,
925  * then punt and return the default estimate.
926  */
927  if (!get_restriction_variable(root, args, varRelid,
928  &vardata, &other, &varonleft))
929  return default_selectivity;
930 
931  /*
932  * If the something is a NULL constant, assume operator is strict and
933  * return zero, ie, operator will never return TRUE.
934  */
935  if (IsA(other, Const) &&
936  ((Const *) other)->constisnull)
937  {
938  ReleaseVariableStats(vardata);
939  return 0.0;
940  }
941 
942  if (IsA(other, Const))
943  {
944  /* Variable is being compared to a known non-null constant */
945  Datum constval = ((Const *) other)->constvalue;
946  FmgrInfo opproc;
947  double mcvsum;
948  double mcvsel;
949  double nullfrac;
950  int hist_size;
951 
952  fmgr_info(get_opcode(oproid), &opproc);
953 
954  /*
955  * Calculate the selectivity for the column's most common values.
956  */
957  mcvsel = mcv_selectivity(&vardata, &opproc, collation,
958  constval, varonleft,
959  &mcvsum);
960 
961  /*
962  * If the histogram is large enough, see what fraction of it matches
963  * the query, and assume that's representative of the non-MCV
964  * population. Otherwise use the default selectivity for the non-MCV
965  * population.
966  */
967  selec = histogram_selectivity(&vardata, &opproc, collation,
968  constval, varonleft,
969  10, 1, &hist_size);
970  if (selec < 0)
971  {
972  /* Nope, fall back on default */
973  selec = default_selectivity;
974  }
975  else if (hist_size < 100)
976  {
977  /*
978  * For histogram sizes from 10 to 100, we combine the histogram
979  * and default selectivities, putting increasingly more trust in
980  * the histogram for larger sizes.
981  */
982  double hist_weight = hist_size / 100.0;
983 
984  selec = selec * hist_weight +
985  default_selectivity * (1.0 - hist_weight);
986  }
987 
988  /* In any case, don't believe extremely small or large estimates. */
989  if (selec < 0.0001)
990  selec = 0.0001;
991  else if (selec > 0.9999)
992  selec = 0.9999;
993 
994  /* Don't forget to account for nulls. */
995  if (HeapTupleIsValid(vardata.statsTuple))
996  nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
997  else
998  nullfrac = 0.0;
999 
1000  /*
1001  * Now merge the results from the MCV and histogram calculations,
1002  * realizing that the histogram covers only the non-null values that
1003  * are not listed in MCV.
1004  */
1005  selec *= 1.0 - nullfrac - mcvsum;
1006  selec += mcvsel;
1007  }
1008  else
1009  {
1010  /* Comparison value is not constant, so we can't do anything */
1011  selec = default_selectivity;
1012  }
1013 
1014  ReleaseVariableStats(vardata);
1015 
1016  /* result should be in range, but make sure... */
1017  CLAMP_PROBABILITY(selec);
1018 
1019  return selec;
1020 }
double mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation, Datum constval, bool varonleft, double *sumcommonp)
Definition: selfuncs.c:732
double histogram_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation, Datum constval, bool varonleft, int min_hist_size, int n_skip, int *hist_size)
Definition: selfuncs.c:823

References generate_unaccent_rules::args, CLAMP_PROBABILITY, fmgr_info(), get_opcode(), get_restriction_variable(), GETSTRUCT, HeapTupleIsValid, histogram_selectivity(), IsA, mcv_selectivity(), ReleaseVariableStats, and VariableStatData::statsTuple.

Referenced by ltreeparentsel(), and matchingsel().

◆ genericcostestimate()

void genericcostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
GenericCosts costs 
)

Definition at line 6544 of file selfuncs.c.

6548 {
6549  IndexOptInfo *index = path->indexinfo;
6550  List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6551  List *indexOrderBys = path->indexorderbys;
6552  Cost indexStartupCost;
6553  Cost indexTotalCost;
6554  Selectivity indexSelectivity;
6555  double indexCorrelation;
6556  double numIndexPages;
6557  double numIndexTuples;
6558  double spc_random_page_cost;
6559  double num_sa_scans;
6560  double num_outer_scans;
6561  double num_scans;
6562  double qual_op_cost;
6563  double qual_arg_cost;
6564  List *selectivityQuals;
6565  ListCell *l;
6566 
6567  /*
6568  * If the index is partial, AND the index predicate with the explicitly
6569  * given indexquals to produce a more accurate idea of the index
6570  * selectivity.
6571  */
6572  selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6573 
6574  /*
6575  * Check for ScalarArrayOpExpr index quals, and estimate the number of
6576  * index scans that will be performed.
6577  */
6578  num_sa_scans = 1;
6579  foreach(l, indexQuals)
6580  {
6581  RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
6582 
6583  if (IsA(rinfo->clause, ScalarArrayOpExpr))
6584  {
6585  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
6586  double alength = estimate_array_length(root, lsecond(saop->args));
6587 
6588  if (alength > 1)
6589  num_sa_scans *= alength;
6590  }
6591  }
6592 
6593  /* Estimate the fraction of main-table tuples that will be visited */
6594  indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6595  index->rel->relid,
6596  JOIN_INNER,
6597  NULL);
6598 
6599  /*
6600  * If caller didn't give us an estimate, estimate the number of index
6601  * tuples that will be visited. We do it in this rather peculiar-looking
6602  * way in order to get the right answer for partial indexes.
6603  */
6604  numIndexTuples = costs->numIndexTuples;
6605  if (numIndexTuples <= 0.0)
6606  {
6607  numIndexTuples = indexSelectivity * index->rel->tuples;
6608 
6609  /*
6610  * The above calculation counts all the tuples visited across all
6611  * scans induced by ScalarArrayOpExpr nodes. We want to consider the
6612  * average per-indexscan number, so adjust. This is a handy place to
6613  * round to integer, too. (If caller supplied tuple estimate, it's
6614  * responsible for handling these considerations.)
6615  */
6616  numIndexTuples = rint(numIndexTuples / num_sa_scans);
6617  }
6618 
6619  /*
6620  * We can bound the number of tuples by the index size in any case. Also,
6621  * always estimate at least one tuple is touched, even when
6622  * indexSelectivity estimate is tiny.
6623  */
6624  if (numIndexTuples > index->tuples)
6625  numIndexTuples = index->tuples;
6626  if (numIndexTuples < 1.0)
6627  numIndexTuples = 1.0;
6628 
6629  /*
6630  * Estimate the number of index pages that will be retrieved.
6631  *
6632  * We use the simplistic method of taking a pro-rata fraction of the total
6633  * number of index pages. In effect, this counts only leaf pages and not
6634  * any overhead such as index metapage or upper tree levels.
6635  *
6636  * In practice access to upper index levels is often nearly free because
6637  * those tend to stay in cache under load; moreover, the cost involved is
6638  * highly dependent on index type. We therefore ignore such costs here
6639  * and leave it to the caller to add a suitable charge if needed.
6640  */
6641  if (index->pages > 1 && index->tuples > 1)
6642  numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
6643  else
6644  numIndexPages = 1.0;
6645 
6646  /* fetch estimated page cost for tablespace containing index */
6647  get_tablespace_page_costs(index->reltablespace,
6648  &spc_random_page_cost,
6649  NULL);
6650 
6651  /*
6652  * Now compute the disk access costs.
6653  *
6654  * The above calculations are all per-index-scan. However, if we are in a
6655  * nestloop inner scan, we can expect the scan to be repeated (with
6656  * different search keys) for each row of the outer relation. Likewise,
6657  * ScalarArrayOpExpr quals result in multiple index scans. This creates
6658  * the potential for cache effects to reduce the number of disk page
6659  * fetches needed. We want to estimate the average per-scan I/O cost in
6660  * the presence of caching.
6661  *
6662  * We use the Mackert-Lohman formula (see costsize.c for details) to
6663  * estimate the total number of page fetches that occur. While this
6664  * wasn't what it was designed for, it seems a reasonable model anyway.
6665  * Note that we are counting pages not tuples anymore, so we take N = T =
6666  * index size, as if there were one "tuple" per page.
6667  */
6668  num_outer_scans = loop_count;
6669  num_scans = num_sa_scans * num_outer_scans;
6670 
6671  if (num_scans > 1)
6672  {
6673  double pages_fetched;
6674 
6675  /* total page fetches ignoring cache effects */
6676  pages_fetched = numIndexPages * num_scans;
6677 
6678  /* use Mackert and Lohman formula to adjust for cache effects */
6679  pages_fetched = index_pages_fetched(pages_fetched,
6680  index->pages,
6681  (double) index->pages,
6682  root);
6683 
6684  /*
6685  * Now compute the total disk access cost, and then report a pro-rated
6686  * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
6687  * since that's internal to the indexscan.)
6688  */
6689  indexTotalCost = (pages_fetched * spc_random_page_cost)
6690  / num_outer_scans;
6691  }
6692  else
6693  {
6694  /*
6695  * For a single index scan, we just charge spc_random_page_cost per
6696  * page touched.
6697  */
6698  indexTotalCost = numIndexPages * spc_random_page_cost;
6699  }
6700 
6701  /*
6702  * CPU cost: any complex expressions in the indexquals will need to be
6703  * evaluated once at the start of the scan to reduce them to runtime keys
6704  * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
6705  * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
6706  * indexqual operator. Because we have numIndexTuples as a per-scan
6707  * number, we have to multiply by num_sa_scans to get the correct result
6708  * for ScalarArrayOpExpr cases. Similarly add in costs for any index
6709  * ORDER BY expressions.
6710  *
6711  * Note: this neglects the possible costs of rechecking lossy operators.
6712  * Detecting that that might be needed seems more expensive than it's
6713  * worth, though, considering all the other inaccuracies here ...
6714  */
6715  qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
6716  index_other_operands_eval_cost(root, indexOrderBys);
6717  qual_op_cost = cpu_operator_cost *
6718  (list_length(indexQuals) + list_length(indexOrderBys));
6719 
6720  indexStartupCost = qual_arg_cost;
6721  indexTotalCost += qual_arg_cost;
6722  indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
6723 
6724  /*
6725  * Generic assumption about index correlation: there isn't any.
6726  */
6727  indexCorrelation = 0.0;
6728 
6729  /*
6730  * Return everything to caller.
6731  */
6732  costs->indexStartupCost = indexStartupCost;
6733  costs->indexTotalCost = indexTotalCost;
6734  costs->indexSelectivity = indexSelectivity;
6735  costs->indexCorrelation = indexCorrelation;
6736  costs->numIndexPages = numIndexPages;
6737  costs->numIndexTuples = numIndexTuples;
6738  costs->spc_random_page_cost = spc_random_page_cost;
6739  costs->num_sa_scans = num_sa_scans;
6740 }
double index_pages_fetched(double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
Definition: costsize.c:898
double cpu_index_tuple_cost
Definition: costsize.c:122
double spc_random_page_cost
Definition: selfuncs.h:130
List * indexorderbys
Definition: pathnodes.h:1692

References add_predicate_to_index_quals(), ScalarArrayOpExpr::args, RestrictInfo::clause, clauselist_selectivity(), cpu_index_tuple_cost, cpu_operator_cost, estimate_array_length(), get_quals_from_indexclauses(), get_tablespace_page_costs(), index_other_operands_eval_cost(), index_pages_fetched(), IndexPath::indexclauses, GenericCosts::indexCorrelation, IndexPath::indexinfo, IndexPath::indexorderbys, GenericCosts::indexSelectivity, GenericCosts::indexStartupCost, GenericCosts::indexTotalCost, IsA, JOIN_INNER, lfirst, list_length(), lsecond, GenericCosts::num_sa_scans, GenericCosts::numIndexPages, GenericCosts::numIndexTuples, and GenericCosts::spc_random_page_cost.

Referenced by blcostestimate(), btcostestimate(), gistcostestimate(), hashcostestimate(), and spgcostestimate().

◆ get_actual_variable_endpoint()

static bool get_actual_variable_endpoint ( Relation  heapRel,
Relation  indexRel,
ScanDirection  indexscandir,
ScanKey  scankeys,
int16  typLen,
bool  typByVal,
TupleTableSlot tableslot,
MemoryContext  outercontext,
Datum endpointDatum 
)
static

Definition at line 6267 of file selfuncs.c.

6276 {
6277  bool have_data = false;
6278  SnapshotData SnapshotNonVacuumable;
6279  IndexScanDesc index_scan;
6280  Buffer vmbuffer = InvalidBuffer;
6281  BlockNumber last_heap_block = InvalidBlockNumber;
6282  int n_visited_heap_pages = 0;
6283  ItemPointer tid;
6285  bool isnull[INDEX_MAX_KEYS];
6286  MemoryContext oldcontext;
6287 
6288  /*
6289  * We use the index-only-scan machinery for this. With mostly-static
6290  * tables that's a win because it avoids a heap visit. It's also a win
6291  * for dynamic data, but the reason is less obvious; read on for details.
6292  *
6293  * In principle, we should scan the index with our current active
6294  * snapshot, which is the best approximation we've got to what the query
6295  * will see when executed. But that won't be exact if a new snap is taken
6296  * before running the query, and it can be very expensive if a lot of
6297  * recently-dead or uncommitted rows exist at the beginning or end of the
6298  * index (because we'll laboriously fetch each one and reject it).
6299  * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6300  * and uncommitted rows as well as normal visible rows. On the other
6301  * hand, it will reject known-dead rows, and thus not give a bogus answer
6302  * when the extreme value has been deleted (unless the deletion was quite
6303  * recent); that case motivates not using SnapshotAny here.
6304  *
6305  * A crucial point here is that SnapshotNonVacuumable, with
6306  * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6307  * condition that the indexscan will use to decide that index entries are
6308  * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6309  * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6310  * have to continue scanning past it, we know that the indexscan will mark
6311  * that index entry killed. That means that the next
6312  * get_actual_variable_endpoint() call will not have to re-consider that
6313  * index entry. In this way we avoid repetitive work when this function
6314  * is used a lot during planning.
6315  *
6316  * But using SnapshotNonVacuumable creates a hazard of its own. In a
6317  * recently-created index, some index entries may point at "broken" HOT
6318  * chains in which not all the tuple versions contain data matching the
6319  * index entry. The live tuple version(s) certainly do match the index,
6320  * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6321  * don't match. Hence, if we took data from the selected heap tuple, we
6322  * might get a bogus answer that's not close to the index extremal value,
6323  * or could even be NULL. We avoid this hazard because we take the data
6324  * from the index entry not the heap.
6325  *
6326  * Despite all this care, there are situations where we might find many
6327  * non-visible tuples near the end of the index. We don't want to expend
6328  * a huge amount of time here, so we give up once we've read too many heap
6329  * pages. When we fail for that reason, the caller will end up using
6330  * whatever extremal value is recorded in pg_statistic.
6331  */
6332  InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6333  GlobalVisTestFor(heapRel));
6334 
6335  index_scan = index_beginscan(heapRel, indexRel,
6336  &SnapshotNonVacuumable,
6337  1, 0);
6338  /* Set it up for index-only scan */
6339  index_scan->xs_want_itup = true;
6340  index_rescan(index_scan, scankeys, 1, NULL, 0);
6341 
6342  /* Fetch first/next tuple in specified direction */
6343  while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6344  {
6346 
6347  if (!VM_ALL_VISIBLE(heapRel,
6348  block,
6349  &vmbuffer))
6350  {
6351  /* Rats, we have to visit the heap to check visibility */
6352  if (!index_fetch_heap(index_scan, tableslot))
6353  {
6354  /*
6355  * No visible tuple for this index entry, so we need to
6356  * advance to the next entry. Before doing so, count heap
6357  * page fetches and give up if we've done too many.
6358  *
6359  * We don't charge a page fetch if this is the same heap page
6360  * as the previous tuple. This is on the conservative side,
6361  * since other recently-accessed pages are probably still in
6362  * buffers too; but it's good enough for this heuristic.
6363  */
6364 #define VISITED_PAGES_LIMIT 100
6365 
6366  if (block != last_heap_block)
6367  {
6368  last_heap_block = block;
6369  n_visited_heap_pages++;
6370  if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6371  break;
6372  }
6373 
6374  continue; /* no visible tuple, try next index entry */
6375  }
6376 
6377  /* We don't actually need the heap tuple for anything */
6378  ExecClearTuple(tableslot);
6379 
6380  /*
6381  * We don't care whether there's more than one visible tuple in
6382  * the HOT chain; if any are visible, that's good enough.
6383  */
6384  }
6385 
6386  /*
6387  * We expect that btree will return data in IndexTuple not HeapTuple
6388  * format. It's not lossy either.
6389  */
6390  if (!index_scan->xs_itup)
6391  elog(ERROR, "no data returned for index-only scan");
6392  if (index_scan->xs_recheck)
6393  elog(ERROR, "unexpected recheck indication from btree");
6394 
6395  /* OK to deconstruct the index tuple */
6396  index_deform_tuple(index_scan->xs_itup,
6397  index_scan->xs_itupdesc,
6398  values, isnull);
6399 
6400  /* Shouldn't have got a null, but be careful */
6401  if (isnull[0])
6402  elog(ERROR, "found unexpected null value in index \"%s\"",
6403  RelationGetRelationName(indexRel));
6404 
6405  /* Copy the index column value out to caller's context */
6406  oldcontext = MemoryContextSwitchTo(outercontext);
6407  *endpointDatum = datumCopy(values[0], typByVal, typLen);
6408  MemoryContextSwitchTo(oldcontext);
6409  have_data = true;
6410  break;
6411  }
6412 
6413  if (vmbuffer != InvalidBuffer)
6414  ReleaseBuffer(vmbuffer);
6415  index_endscan(index_scan);
6416 
6417  return have_data;
6418 }
uint32 BlockNumber
Definition: block.h:31
#define InvalidBlockNumber
Definition: block.h:33
static Datum values[MAXATTR]
Definition: bootstrap.c:152
int Buffer
Definition: buf.h:23
#define InvalidBuffer
Definition: buf.h:25
void ReleaseBuffer(Buffer buffer)
Definition: bufmgr.c:4560
Datum datumCopy(Datum value, bool typByVal, int typLen)
Definition: datum.c:132
ItemPointer index_getnext_tid(IndexScanDesc scan, ScanDirection direction)
Definition: indexam.c:577
IndexScanDesc index_beginscan(Relation heapRelation, Relation indexRelation, Snapshot snapshot, int nkeys, int norderbys)
Definition: indexam.c:257
bool index_fetch_heap(IndexScanDesc scan, TupleTableSlot *slot)
Definition: indexam.c:635
void index_endscan(IndexScanDesc scan)
Definition: indexam.c:379
void index_rescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int norderbys)
Definition: indexam.c:353
void index_deform_tuple(IndexTuple tup, TupleDesc tupleDescriptor, Datum *values, bool *isnull)
Definition: indextuple.c:456
static BlockNumber ItemPointerGetBlockNumber(const ItemPointerData *pointer)
Definition: itemptr.h:103
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:124
#define INDEX_MAX_KEYS
GlobalVisState * GlobalVisTestFor(Relation rel)
Definition: procarray.c:4091
#define RelationGetRelationName(relation)
Definition: rel.h:539
#define VISITED_PAGES_LIMIT
#define InitNonVacuumableSnapshot(snapshotdata, vistestp)
Definition: snapmgr.h:48
IndexTuple xs_itup
Definition: relscan.h:142
struct TupleDescData * xs_itupdesc
Definition: relscan.h:143
static TupleTableSlot * ExecClearTuple(TupleTableSlot *slot)
Definition: tuptable.h:433
#define VM_ALL_VISIBLE(r, b, v)
Definition: visibilitymap.h:24

References datumCopy(), elog, ERROR, ExecClearTuple(), GlobalVisTestFor(), index_beginscan(), index_deform_tuple(), index_endscan(), index_fetch_heap(), index_getnext_tid(), INDEX_MAX_KEYS, index_rescan(), InitNonVacuumableSnapshot, InvalidBlockNumber, InvalidBuffer, ItemPointerGetBlockNumber(), MemoryContextSwitchTo(), RelationGetRelationName, ReleaseBuffer(), values, VISITED_PAGES_LIMIT, VM_ALL_VISIBLE, IndexScanDescData::xs_itup, IndexScanDescData::xs_itupdesc, IndexScanDescData::xs_recheck, and IndexScanDescData::xs_want_itup.

Referenced by get_actual_variable_range().

◆ get_actual_variable_range()

static bool get_actual_variable_range ( PlannerInfo root,
VariableStatData vardata,
Oid  sortop,
Oid  collation,
Datum min,
Datum max 
)
static

Definition at line 6087 of file selfuncs.c.

6090 {
6091  bool have_data = false;
6092  RelOptInfo *rel = vardata->rel;
6093  RangeTblEntry *rte;
6094  ListCell *lc;
6095 
6096  /* No hope if no relation or it doesn't have indexes */
6097  if (rel == NULL || rel->indexlist == NIL)
6098  return false;
6099  /* If it has indexes it must be a plain relation */
6100  rte = root->simple_rte_array[rel->relid];
6101  Assert(rte->rtekind == RTE_RELATION);
6102 
6103  /* ignore partitioned tables. Any indexes here are not real indexes */
6104  if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6105  return false;
6106 
6107  /* Search through the indexes to see if any match our problem */
6108  foreach(lc, rel->indexlist)
6109  {
6111  ScanDirection indexscandir;
6112 
6113  /* Ignore non-btree indexes */
6114  if (index->relam != BTREE_AM_OID)
6115  continue;
6116 
6117  /*
6118  * Ignore partial indexes --- we only want stats that cover the entire
6119  * relation.
6120  */
6121  if (index->indpred != NIL)
6122  continue;
6123 
6124  /*
6125  * The index list might include hypothetical indexes inserted by a
6126  * get_relation_info hook --- don't try to access them.
6127  */
6128  if (index->hypothetical)
6129  continue;
6130 
6131  /*
6132  * The first index column must match the desired variable, sortop, and
6133  * collation --- but we can use a descending-order index.
6134  */
6135  if (collation != index->indexcollations[0])
6136  continue; /* test first 'cause it's cheapest */
6137  if (!match_index_to_operand(vardata->var, 0, index))
6138  continue;
6139  switch (get_op_opfamily_strategy(sortop, index->sortopfamily[0]))
6140  {
6141  case BTLessStrategyNumber:
6142  if (index->reverse_sort[0])
6143  indexscandir = BackwardScanDirection;
6144  else
6145  indexscandir = ForwardScanDirection;
6146  break;
6148  if (index->reverse_sort[0])
6149  indexscandir = ForwardScanDirection;
6150  else
6151  indexscandir = BackwardScanDirection;
6152  break;
6153  default:
6154  /* index doesn't match the sortop */
6155  continue;
6156  }
6157 
6158  /*
6159  * Found a suitable index to extract data from. Set up some data that
6160  * can be used by both invocations of get_actual_variable_endpoint.
6161  */
6162  {
6163  MemoryContext tmpcontext;
6164  MemoryContext oldcontext;
6165  Relation heapRel;
6166  Relation indexRel;
6167  TupleTableSlot *slot;
6168  int16 typLen;
6169  bool typByVal;
6170  ScanKeyData scankeys[1];
6171 
6172  /* Make sure any cruft gets recycled when we're done */
6174  "get_actual_variable_range workspace",
6176  oldcontext = MemoryContextSwitchTo(tmpcontext);
6177 
6178  /*
6179  * Open the table and index so we can read from them. We should
6180  * already have some type of lock on each.
6181  */
6182  heapRel = table_open(rte->relid, NoLock);
6183  indexRel = index_open(index->indexoid, NoLock);
6184 
6185  /* build some stuff needed for indexscan execution */
6186  slot = table_slot_create(heapRel, NULL);
6187  get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6188 
6189  /* set up an IS NOT NULL scan key so that we ignore nulls */
6190  ScanKeyEntryInitialize(&scankeys[0],
6192  1, /* index col to scan */
6193  InvalidStrategy, /* no strategy */
6194  InvalidOid, /* no strategy subtype */
6195  InvalidOid, /* no collation */
6196  InvalidOid, /* no reg proc for this */
6197  (Datum) 0); /* constant */
6198 
6199  /* If min is requested ... */
6200  if (min)
6201  {
6202  have_data = get_actual_variable_endpoint(heapRel,
6203  indexRel,
6204  indexscandir,
6205  scankeys,
6206  typLen,
6207  typByVal,
6208  slot,
6209  oldcontext,
6210  min);
6211  }
6212  else
6213  {
6214  /* If min not requested, still want to fetch max */
6215  have_data = true;
6216  }
6217 
6218  /* If max is requested, and we didn't already fail ... */
6219  if (max && have_data)
6220  {
6221  /* scan in the opposite direction; all else is the same */
6222  have_data = get_actual_variable_endpoint(heapRel,
6223  indexRel,
6224  -indexscandir,
6225  scankeys,
6226  typLen,
6227  typByVal,
6228  slot,
6229  oldcontext,
6230  max);
6231  }
6232 
6233  /* Clean everything up */
6235 
6236  index_close(indexRel, NoLock);
6237  table_close(heapRel, NoLock);
6238 
6239  MemoryContextSwitchTo(oldcontext);
6240  MemoryContextDelete(tmpcontext);
6241 
6242  /* And we're done */
6243  break;
6244  }
6245  }
6246 
6247  return have_data;
6248 }
signed short int16
Definition: c.h:480
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:1253
bool match_index_to_operand(Node *operand, int indexcol, IndexOptInfo *index)
Definition: indxpath.c:3758
void get_typlenbyval(Oid typid, int16 *typlen, bool *typbyval)
Definition: lsyscache.c:2207
MemoryContext CurrentMemoryContext
Definition: mcxt.c:131
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:442
#define AllocSetContextCreate
Definition: memutils.h:129
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:153
void ScanKeyEntryInitialize(ScanKey entry, int flags, AttrNumber attributeNumber, StrategyNumber strategy, Oid subtype, Oid collation, RegProcedure procedure, Datum argument)
Definition: scankey.c:32
ScanDirection
Definition: sdir.h:25
@ BackwardScanDirection
Definition: sdir.h:26
@ ForwardScanDirection
Definition: sdir.h:28
static bool get_actual_variable_endpoint(Relation heapRel, Relation indexRel, ScanDirection indexscandir, ScanKey scankeys, int16 typLen, bool typByVal, TupleTableSlot *tableslot, MemoryContext outercontext, Datum *endpointDatum)
Definition: selfuncs.c:6267
#define SK_SEARCHNOTNULL
Definition: skey.h:122
#define SK_ISNULL
Definition: skey.h:115
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
#define InvalidStrategy
Definition: stratnum.h:24
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:126
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:40
TupleTableSlot * table_slot_create(Relation relation, List **reglist)
Definition: tableam.c:91

References ALLOCSET_DEFAULT_SIZES, AllocSetContextCreate, Assert(), VariableStatData::atttype, BackwardScanDirection, BTGreaterStrategyNumber, BTLessStrategyNumber, CurrentMemoryContext, ExecDropSingleTupleTableSlot(), ForwardScanDirection, get_actual_variable_endpoint(), get_op_opfamily_strategy(), get_typlenbyval(), index_close(), index_open(), RelOptInfo::indexlist, InvalidOid, InvalidStrategy, lfirst, match_index_to_operand(), MemoryContextDelete(), MemoryContextSwitchTo(), NIL, NoLock, VariableStatData::rel, RangeTblEntry::relid, RelOptInfo::relid, RangeTblEntry::relkind, RTE_RELATION, RangeTblEntry::rtekind, ScanKeyEntryInitialize(), SK_ISNULL, SK_SEARCHNOTNULL, table_close(), table_open(), table_slot_create(), and VariableStatData::var.

Referenced by get_variable_range(), and ineq_histogram_selectivity().

◆ get_join_variables()

void get_join_variables ( PlannerInfo root,
List args,
SpecialJoinInfo sjinfo,
VariableStatData vardata1,
VariableStatData vardata2,
bool join_is_reversed 
)

Definition at line 4943 of file selfuncs.c.

4946 {
4947  Node *left,
4948  *right;
4949 
4950  if (list_length(args) != 2)
4951  elog(ERROR, "join operator should take two arguments");
4952 
4953  left = (Node *) linitial(args);
4954  right = (Node *) lsecond(args);
4955 
4956  examine_variable(root, left, 0, vardata1);
4957  examine_variable(root, right, 0, vardata2);
4958 
4959  if (vardata1->rel &&
4960  bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
4961  *join_is_reversed = true; /* var1 is on RHS */
4962  else if (vardata2->rel &&
4963  bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
4964  *join_is_reversed = true; /* var2 is on LHS */
4965  else
4966  *join_is_reversed = false;
4967 }
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:412
Relids relids
Definition: pathnodes.h:856
Relids syn_lefthand
Definition: pathnodes.h:2870
Relids syn_righthand
Definition: pathnodes.h:2871

References generate_unaccent_rules::args, bms_is_subset(), elog, ERROR, examine_variable(), linitial, list_length(), lsecond, VariableStatData::rel, RelOptInfo::relids, SpecialJoinInfo::syn_lefthand, and SpecialJoinInfo::syn_righthand.

Referenced by eqjoinsel(), neqjoinsel(), and networkjoinsel().

◆ get_quals_from_indexclauses()

List* get_quals_from_indexclauses ( List indexclauses)

Definition at line 6460 of file selfuncs.c.

6461 {
6462  List *result = NIL;
6463  ListCell *lc;
6464 
6465  foreach(lc, indexclauses)
6466  {
6467  IndexClause *iclause = lfirst_node(IndexClause, lc);
6468  ListCell *lc2;
6469 
6470  foreach(lc2, iclause->indexquals)
6471  {
6472  RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6473 
6474  result = lappend(result, rinfo);
6475  }
6476  }
6477  return result;
6478 }

References IndexClause::indexquals, lappend(), lfirst_node, and NIL.

Referenced by brincostestimate(), genericcostestimate(), and gincostestimate().

◆ get_restriction_variable()

bool get_restriction_variable ( PlannerInfo root,
List args,
int  varRelid,
VariableStatData vardata,
Node **  other,
bool varonleft 
)

Definition at line 4883 of file selfuncs.c.

4886 {
4887  Node *left,
4888  *right;
4889  VariableStatData rdata;
4890 
4891  /* Fail if not a binary opclause (probably shouldn't happen) */
4892  if (list_length(args) != 2)
4893  return false;
4894 
4895  left = (Node *) linitial(args);
4896  right = (Node *) lsecond(args);
4897 
4898  /*
4899  * Examine both sides. Note that when varRelid is nonzero, Vars of other
4900  * relations will be treated as pseudoconstants.
4901  */
4902  examine_variable(root, left, varRelid, vardata);
4903  examine_variable(root, right, varRelid, &rdata);
4904 
4905  /*
4906  * If one side is a variable and the other not, we win.
4907  */
4908  if (vardata->rel && rdata.rel == NULL)
4909  {
4910  *varonleft = true;
4911  *other = estimate_expression_value(root, rdata.var);
4912  /* Assume we need no ReleaseVariableStats(rdata) here */
4913  return true;
4914  }
4915 
4916  if (vardata->rel == NULL && rdata.rel)
4917  {
4918  *varonleft = false;
4919  *other = estimate_expression_value(root, vardata->var);
4920  /* Assume we need no ReleaseVariableStats(*vardata) here */
4921  *vardata = rdata;
4922  return true;
4923  }
4924 
4925  /* Oops, clause has wrong structure (probably var op var) */
4926  ReleaseVariableStats(*vardata);
4927  ReleaseVariableStats(rdata);
4928 
4929  return false;
4930 }
Node * estimate_expression_value(PlannerInfo *root, Node *node)
Definition: clauses.c:2375

References generate_unaccent_rules::args, estimate_expression_value(), examine_variable(), linitial, list_length(), lsecond, VariableStatData::rel, ReleaseVariableStats, and VariableStatData::var.

Referenced by _int_matchsel(), arraycontsel(), eqsel_internal(), generic_restriction_selectivity(), multirangesel(), networksel(), patternsel_common(), rangesel(), scalarineqsel_wrapper(), and tsmatchsel().

◆ get_stats_slot_range()

static void get_stats_slot_range ( AttStatsSlot sslot,
Oid  opfuncoid,
FmgrInfo opproc,
Oid  collation,
int16  typLen,
bool  typByVal,
Datum min,
Datum max,
bool p_have_data 
)
static

Definition at line 6024 of file selfuncs.c.

6027 {
6028  Datum tmin = *min;
6029  Datum tmax = *max;
6030  bool have_data = *p_have_data;
6031  bool found_tmin = false;
6032  bool found_tmax = false;
6033 
6034  /* Look up the comparison function, if we didn't already do so */
6035  if (opproc->fn_oid != opfuncoid)
6036  fmgr_info(opfuncoid, opproc);
6037 
6038  /* Scan all the slot's values */
6039  for (int i = 0; i < sslot->nvalues; i++)
6040  {
6041  if (!have_data)
6042  {
6043  tmin = tmax = sslot->values[i];
6044  found_tmin = found_tmax = true;
6045  *p_have_data = have_data = true;
6046  continue;
6047  }
6048  if (DatumGetBool(FunctionCall2Coll(opproc,
6049  collation,
6050  sslot->values[i], tmin)))
6051  {
6052  tmin = sslot->values[i];
6053  found_tmin = true;
6054  }
6055  if (DatumGetBool(FunctionCall2Coll(opproc,
6056  collation,
6057  tmax, sslot->values[i])))
6058  {
6059  tmax = sslot->values[i];
6060  found_tmax = true;
6061  }
6062  }
6063 
6064  /*
6065  * Copy the slot's values, if we found new extreme values.
6066  */
6067  if (found_tmin)
6068  *min = datumCopy(tmin, typByVal, typLen);
6069  if (found_tmax)
6070  *max = datumCopy(tmax, typByVal, typLen);
6071 }
Datum FunctionCall2Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2)
Definition: fmgr.c:1149
Oid fn_oid
Definition: fmgr.h:59

References datumCopy(), DatumGetBool(), fmgr_info(), FmgrInfo::fn_oid, FunctionCall2Coll(), i, AttStatsSlot::nvalues, and AttStatsSlot::values.

Referenced by get_variable_range().

◆ get_variable_numdistinct()

double get_variable_numdistinct ( VariableStatData vardata,
bool isdefault 
)

Definition at line 5764 of file selfuncs.c.

5765 {
5766  double stadistinct;
5767  double stanullfrac = 0.0;
5768  double ntuples;
5769 
5770  *isdefault = false;
5771 
5772  /*
5773  * Determine the stadistinct value to use. There are cases where we can
5774  * get an estimate even without a pg_statistic entry, or can get a better
5775  * value than is in pg_statistic. Grab stanullfrac too if we can find it
5776  * (otherwise, assume no nulls, for lack of any better idea).
5777  */
5778  if (HeapTupleIsValid(vardata->statsTuple))
5779  {
5780  /* Use the pg_statistic entry */
5781  Form_pg_statistic stats;
5782 
5783  stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
5784  stadistinct = stats->stadistinct;
5785  stanullfrac = stats->stanullfrac;
5786  }
5787  else if (vardata->vartype == BOOLOID)
5788  {
5789  /*
5790  * Special-case boolean columns: presumably, two distinct values.
5791  *
5792  * Are there any other datatypes we should wire in special estimates
5793  * for?
5794  */
5795  stadistinct = 2.0;
5796  }
5797  else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
5798  {
5799  /*
5800  * If the Var represents a column of a VALUES RTE, assume it's unique.
5801  * This could of course be very wrong, but it should tend to be true
5802  * in well-written queries. We could consider examining the VALUES'
5803  * contents to get some real statistics; but that only works if the
5804  * entries are all constants, and it would be pretty expensive anyway.
5805  */
5806  stadistinct = -1.0; /* unique (and all non null) */
5807  }
5808  else
5809  {
5810  /*
5811  * We don't keep statistics for system columns, but in some cases we
5812  * can infer distinctness anyway.
5813  */
5814  if (vardata->var && IsA(vardata->var, Var))
5815  {
5816  switch (((Var *) vardata->var)->varattno)
5817  {
5819  stadistinct = -1.0; /* unique (and all non null) */
5820  break;
5822  stadistinct = 1.0; /* only 1 value */
5823  break;
5824  default:
5825  stadistinct = 0.0; /* means "unknown" */
5826  break;
5827  }
5828  }
5829  else
5830  stadistinct = 0.0; /* means "unknown" */
5831 
5832  /*
5833  * XXX consider using estimate_num_groups on expressions?
5834  */
5835  }
5836 
5837  /*
5838  * If there is a unique index or DISTINCT clause for the variable, assume
5839  * it is unique no matter what pg_statistic says; the statistics could be
5840  * out of date, or we might have found a partial unique index that proves
5841  * the var is unique for this query. However, we'd better still believe
5842  * the null-fraction statistic.
5843  */
5844  if (vardata->isunique)
5845  stadistinct = -1.0 * (1.0 - stanullfrac);
5846 
5847  /*
5848  * If we had an absolute estimate, use that.
5849  */
5850  if (stadistinct > 0.0)
5851  return clamp_row_est(stadistinct);
5852 
5853  /*
5854  * Otherwise we need to get the relation size; punt if not available.
5855  */
5856  if (vardata->rel == NULL)
5857  {
5858  *isdefault = true;
5859  return DEFAULT_NUM_DISTINCT;
5860  }
5861  ntuples = vardata->rel->tuples;
5862  if (ntuples <= 0.0)
5863  {
5864  *isdefault = true;
5865  return DEFAULT_NUM_DISTINCT;
5866  }
5867 
5868  /*
5869  * If we had a relative estimate, use that.
5870  */
5871  if (stadistinct < 0.0)
5872  return clamp_row_est(-stadistinct * ntuples);
5873 
5874  /*
5875  * With no data, estimate ndistinct = ntuples if the table is small, else
5876  * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
5877  * that the behavior isn't discontinuous.
5878  */
5879  if (ntuples < DEFAULT_NUM_DISTINCT)
5880  return clamp_row_est(ntuples);
5881 
5882  *isdefault = true;
5883  return DEFAULT_NUM_DISTINCT;
5884 }
@ RTE_VALUES
Definition: parsenodes.h:1016
#define DEFAULT_NUM_DISTINCT
Definition: selfuncs.h:52
RTEKind rtekind
Definition: pathnodes.h:907
#define TableOidAttributeNumber
Definition: sysattr.h:26
#define SelfItemPointerAttributeNumber
Definition: sysattr.h:21

References clamp_row_est(), DEFAULT_NUM_DISTINCT, GETSTRUCT, HeapTupleIsValid, IsA, VariableStatData::isunique, VariableStatData::rel, RTE_VALUES, RelOptInfo::rtekind, SelfItemPointerAttributeNumber, VariableStatData::statsTuple, TableOidAttributeNumber, RelOptInfo::tuples, VariableStatData::var, and VariableStatData::vartype.

Referenced by add_unique_group_var(), eqjoinsel(), estimate_hash_bucket_stats(), ineq_histogram_selectivity(), var_eq_const(), and var_eq_non_const().

◆ get_variable_range()

static bool get_variable_range ( PlannerInfo root,
VariableStatData vardata,
Oid  sortop,
Oid  collation,
Datum min,
Datum max 
)
static

Definition at line 5897 of file selfuncs.c.

5900 {
5901  Datum tmin = 0;
5902  Datum tmax = 0;
5903  bool have_data = false;
5904  int16 typLen;
5905  bool typByVal;
5906  Oid opfuncoid;
5907  FmgrInfo opproc;
5908  AttStatsSlot sslot;
5909 
5910  /*
5911  * XXX It's very tempting to try to use the actual column min and max, if
5912  * we can get them relatively-cheaply with an index probe. However, since
5913  * this function is called many times during join planning, that could
5914  * have unpleasant effects on planning speed. Need more investigation
5915  * before enabling this.
5916  */
5917 #ifdef NOT_USED
5918  if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
5919  return true;
5920 #endif
5921 
5922  if (!HeapTupleIsValid(vardata->statsTuple))
5923  {
5924  /* no stats available, so default result */
5925  return false;
5926  }
5927 
5928  /*
5929  * If we can't apply the sortop to the stats data, just fail. In
5930  * principle, if there's a histogram and no MCVs, we could return the
5931  * histogram endpoints without ever applying the sortop ... but it's
5932  * probably not worth trying, because whatever the caller wants to do with
5933  * the endpoints would likely fail the security check too.
5934  */
5935  if (!statistic_proc_security_check(vardata,
5936  (opfuncoid = get_opcode(sortop))))
5937  return false;
5938 
5939  opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
5940 
5941  get_typlenbyval(vardata->atttype, &typLen, &typByVal);
5942 
5943  /*
5944  * If there is a histogram with the ordering we want, grab the first and
5945  * last values.
5946  */
5947  if (get_attstatsslot(&sslot, vardata->statsTuple,
5948  STATISTIC_KIND_HISTOGRAM, sortop,
5950  {
5951  if (sslot.stacoll == collation && sslot.nvalues > 0)
5952  {
5953  tmin = datumCopy(sslot.values[0], typByVal, typLen);
5954  tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
5955  have_data = true;
5956  }
5957  free_attstatsslot(&sslot);
5958  }
5959 
5960  /*
5961  * Otherwise, if there is a histogram with some other ordering, scan it
5962  * and get the min and max values according to the ordering we want. This
5963  * of course may not find values that are really extremal according to our
5964  * ordering, but it beats ignoring available data.
5965  */
5966  if (!have_data &&
5967  get_attstatsslot(&sslot, vardata->statsTuple,
5968  STATISTIC_KIND_HISTOGRAM, InvalidOid,
5970  {
5971  get_stats_slot_range(&sslot, opfuncoid, &opproc,
5972  collation, typLen, typByVal,
5973  &tmin, &tmax, &have_data);
5974  free_attstatsslot(&sslot);
5975  }
5976 
5977  /*
5978  * If we have most-common-values info, look for extreme MCVs. This is
5979  * needed even if we also have a histogram, since the histogram excludes
5980  * the MCVs. However, if we *only* have MCVs and no histogram, we should
5981  * be pretty wary of deciding that that is a full representation of the
5982  * data. Proceed only if the MCVs represent the whole table (to within
5983  * roundoff error).
5984  */
5985  if (get_attstatsslot(&sslot, vardata->statsTuple,
5986  STATISTIC_KIND_MCV, InvalidOid,
5987  have_data ? ATTSTATSSLOT_VALUES :
5989  {
5990  bool use_mcvs = have_data;
5991 
5992  if (!have_data)
5993  {
5994  double sumcommon = 0.0;
5995  double nullfrac;
5996  int i;
5997 
5998  for (i = 0; i < sslot.nnumbers; i++)
5999  sumcommon += sslot.numbers[i];
6000  nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6001  if (sumcommon + nullfrac > 0.99999)
6002  use_mcvs = true;
6003  }
6004 
6005  if (use_mcvs)
6006  get_stats_slot_range(&sslot, opfuncoid, &opproc,
6007  collation, typLen, typByVal,
6008  &tmin, &tmax, &have_data);
6009  free_attstatsslot(&sslot);
6010  }
6011 
6012  *min = tmin;
6013  *max = tmax;
6014  return have_data;
6015 }
static bool get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata, Oid sortop, Oid collation, Datum *min, Datum *max)
Definition: selfuncs.c:6087
static void get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc, Oid collation, int16 typLen, bool typByVal, Datum *min, Datum *max, bool *p_have_data)
Definition: selfuncs.c:6024

References ATTSTATSSLOT_NUMBERS, ATTSTATSSLOT_VALUES, VariableStatData::atttype, datumCopy(), FmgrInfo::fn_oid, free_attstatsslot(), get_actual_variable_range(), get_attstatsslot(), get_opcode(), get_stats_slot_range(), get_typlenbyval(), GETSTRUCT, HeapTupleIsValid, i, InvalidOid, AttStatsSlot::nnumbers, AttStatsSlot::numbers, AttStatsSlot::nvalues, AttStatsSlot::stacoll, statistic_proc_security_check(), VariableStatData::statsTuple, and AttStatsSlot::values.

Referenced by mergejoinscansel().

◆ gincost_opexpr()

static bool gincost_opexpr ( PlannerInfo root,
IndexOptInfo index,
int  indexcol,
OpExpr clause,
GinQualCounts counts 
)
static

Definition at line 7372 of file selfuncs.c.

7377 {
7378  Oid clause_op = clause->opno;
7379  Node *operand = (Node *) lsecond(clause->args);
7380 
7381  /* aggressively reduce to a constant, and look through relabeling */
7382  operand = estimate_expression_value(root, operand);
7383 
7384  if (IsA(operand, RelabelType))
7385  operand = (Node *) ((RelabelType *) operand)->arg;
7386 
7387  /*
7388  * It's impossible to call extractQuery method for unknown operand. So
7389  * unless operand is a Const we can't do much; just assume there will be
7390  * one ordinary search entry from the operand at runtime.
7391  */
7392  if (!IsA(operand, Const))
7393  {
7394  counts->exactEntries++;
7395  counts->searchEntries++;
7396  return true;
7397  }
7398 
7399  /* If Const is null, there can be no matches */
7400  if (((Const *) operand)->constisnull)
7401  return false;
7402 
7403  /* Otherwise, apply extractQuery and get the actual term counts */
7404  return gincost_pattern(index, indexcol, clause_op,
7405  ((Const *) operand)->constvalue,
7406  counts);
7407 }
static bool gincost_pattern(IndexOptInfo *index, int indexcol, Oid clause_op, Datum query, GinQualCounts *counts)
Definition: selfuncs.c:7258
double exactEntries
Definition: selfuncs.c:7247
double searchEntries
Definition: selfuncs.c:7248
List * args
Definition: primnodes.h:792

References arg, OpExpr::args, estimate_expression_value(), GinQualCounts::exactEntries, gincost_pattern(), IsA, lsecond, OpExpr::opno, and GinQualCounts::searchEntries.

Referenced by gincostestimate().

◆ gincost_pattern()

static bool gincost_pattern ( IndexOptInfo index,
int  indexcol,
Oid  clause_op,
Datum  query,
GinQualCounts counts 
)
static

Definition at line 7258 of file selfuncs.c.

7261 {
7262  FmgrInfo flinfo;
7263  Oid extractProcOid;
7264  Oid collation;
7265  int strategy_op;
7266  Oid lefttype,
7267  righttype;
7268  int32 nentries = 0;
7269  bool *partial_matches = NULL;
7270  Pointer *extra_data = NULL;
7271  bool *nullFlags = NULL;
7272  int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7273  int32 i;
7274 
7275  Assert(indexcol < index->nkeycolumns);
7276 
7277  /*
7278  * Get the operator's strategy number and declared input data types within
7279  * the index opfamily. (We don't need the latter, but we use
7280  * get_op_opfamily_properties because it will throw error if it fails to
7281  * find a matching pg_amop entry.)
7282  */
7283  get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7284  &strategy_op, &lefttype, &righttype);
7285 
7286  /*
7287  * GIN always uses the "default" support functions, which are those with
7288  * lefttype == righttype == the opclass' opcintype (see
7289  * IndexSupportInitialize in relcache.c).
7290  */
7291  extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7292  index->opcintype[indexcol],
7293  index->opcintype[indexcol],
7295 
7296  if (!OidIsValid(extractProcOid))
7297  {
7298  /* should not happen; throw same error as index_getprocinfo */
7299  elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7300  GIN_EXTRACTQUERY_PROC, indexcol + 1,
7301  get_rel_name(index->indexoid));
7302  }
7303 
7304  /*
7305  * Choose collation to pass to extractProc (should match initGinState).
7306  */
7307  if (OidIsValid(index->indexcollations[indexcol]))
7308  collation = index->indexcollations[indexcol];
7309  else
7310  collation = DEFAULT_COLLATION_OID;
7311 
7312  fmgr_info(extractProcOid, &flinfo);
7313 
7314  set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7315 
7316  FunctionCall7Coll(&flinfo,
7317  collation,
7318  query,
7319  PointerGetDatum(&nentries),
7320  UInt16GetDatum(strategy_op),
7321  PointerGetDatum(&partial_matches),
7322  PointerGetDatum(&extra_data),
7323  PointerGetDatum(&nullFlags),
7324  PointerGetDatum(&searchMode));
7325 
7326  if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7327  {
7328  /* No match is possible */
7329  return false;
7330  }
7331 
7332  for (i = 0; i < nentries; i++)
7333  {
7334  /*
7335  * For partial match we haven't any information to estimate number of
7336  * matched entries in index, so, we just estimate it as 100
7337  */
7338  if (partial_matches && partial_matches[i])
7339  counts->partialEntries += 100;
7340  else
7341  counts->exactEntries++;
7342 
7343  counts->searchEntries++;
7344  }
7345 
7346  if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7347  {
7348  counts->attHasNormalScan[indexcol] = true;
7349  }
7350  else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
7351  {
7352  /* Treat "include empty" like an exact-match item */
7353  counts->attHasNormalScan[indexcol] = true;
7354  counts->exactEntries++;
7355  counts->searchEntries++;
7356  }
7357  else
7358  {
7359  /* It's GIN_SEARCH_MODE_ALL */
7360  counts->attHasFullScan[indexcol] = true;
7361  }
7362 
7363  return true;
7364 }
signed int int32
Definition: c.h:481
char * Pointer
Definition: c.h:470
void set_fn_opclass_options(FmgrInfo *flinfo, bytea *options)
Definition: fmgr.c:2070
Datum FunctionCall7Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4, Datum arg5, Datum arg6, Datum arg7)
Definition: fmgr.c:1284
#define GIN_EXTRACTQUERY_PROC
Definition: gin.h:24
#define GIN_SEARCH_MODE_DEFAULT
Definition: gin.h:34
#define GIN_SEARCH_MODE_INCLUDE_EMPTY
Definition: gin.h:35
void get_op_opfamily_properties(Oid opno, Oid opfamily, bool ordering_op, int *strategy, Oid *lefttype, Oid *righttype)
Definition: lsyscache.c:136
Oid get_opfamily_proc(Oid opfamily, Oid lefttype, Oid righttype, int16 procnum)
Definition: lsyscache.c:796
char * get_rel_name(Oid relid)
Definition: lsyscache.c:1906
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:322
static Datum UInt16GetDatum(uint16 X)
Definition: postgres.h:192
bool attHasNormalScan[INDEX_MAX_KEYS]
Definition: selfuncs.c:7245
double partialEntries
Definition: selfuncs.c:7246
bool attHasFullScan[INDEX_MAX_KEYS]
Definition: selfuncs.c:7244

References Assert(), GinQualCounts::attHasFullScan, GinQualCounts::attHasNormalScan, elog, ERROR, GinQualCounts::exactEntries, fmgr_info(), FunctionCall7Coll(), get_op_opfamily_properties(), get_opfamily_proc(), get_rel_name(), GIN_EXTRACTQUERY_PROC, GIN_SEARCH_MODE_DEFAULT, GIN_SEARCH_MODE_INCLUDE_EMPTY, i, OidIsValid, GinQualCounts::partialEntries, PointerGetDatum(), GinQualCounts::searchEntries, set_fn_opclass_options(), and UInt16GetDatum().

Referenced by gincost_opexpr(), and gincost_scalararrayopexpr().

◆ gincost_scalararrayopexpr()

static bool gincost_scalararrayopexpr ( PlannerInfo root,
IndexOptInfo index,
int  indexcol,
ScalarArrayOpExpr clause,
double  numIndexEntries,
GinQualCounts counts 
)
static

Definition at line 7422 of file selfuncs.c.

7428 {
7429  Oid clause_op = clause->opno;
7430  Node *rightop = (Node *) lsecond(clause->args);
7431  ArrayType *arrayval;
7432  int16 elmlen;
7433  bool elmbyval;
7434  char elmalign;
7435  int numElems;
7436  Datum *elemValues;
7437  bool *elemNulls;
7438  GinQualCounts arraycounts;
7439  int numPossible = 0;
7440  int i;
7441 
7442  Assert(clause->useOr);
7443 
7444  /* aggressively reduce to a constant, and look through relabeling */
7445  rightop = estimate_expression_value(root, rightop);
7446 
7447  if (IsA(rightop, RelabelType))
7448  rightop = (Node *) ((RelabelType *) rightop)->arg;
7449 
7450  /*
7451  * It's impossible to call extractQuery method for unknown operand. So
7452  * unless operand is a Const we can't do much; just assume there will be
7453  * one ordinary search entry from each array entry at runtime, and fall
7454  * back on a probably-bad estimate of the number of array entries.
7455  */
7456  if (!IsA(rightop, Const))
7457  {
7458  counts->exactEntries++;
7459  counts->searchEntries++;
7460  counts->arrayScans *= estimate_array_length(root, rightop);
7461  return true;
7462  }
7463 
7464  /* If Const is null, there can be no matches */
7465  if (((Const *) rightop)->constisnull)
7466  return false;
7467 
7468  /* Otherwise, extract the array elements and iterate over them */
7469  arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
7471  &elmlen, &elmbyval, &elmalign);
7472  deconstruct_array(arrayval,
7473  ARR_ELEMTYPE(arrayval),
7474  elmlen, elmbyval, elmalign,
7475  &elemValues, &elemNulls, &numElems);
7476 
7477  memset(&arraycounts, 0, sizeof(arraycounts));
7478 
7479  for (i = 0; i < numElems; i++)
7480  {
7481  GinQualCounts elemcounts;
7482 
7483  /* NULL can't match anything, so ignore, as the executor will */
7484  if (elemNulls[i])
7485  continue;
7486 
7487  /* Otherwise, apply extractQuery and get the actual term counts */
7488  memset(&elemcounts, 0, sizeof(elemcounts));
7489 
7490  if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
7491  &elemcounts))
7492  {
7493  /* We ignore array elements that are unsatisfiable patterns */
7494  numPossible++;
7495 
7496  if (elemcounts.attHasFullScan[indexcol] &&
7497  !elemcounts.attHasNormalScan[indexcol])
7498  {
7499  /*
7500  * Full index scan will be required. We treat this as if
7501  * every key in the index had been listed in the query; is
7502  * that reasonable?
7503  */
7504  elemcounts.partialEntries = 0;
7505  elemcounts.exactEntries = numIndexEntries;
7506  elemcounts.searchEntries = numIndexEntries;
7507  }
7508  arraycounts.partialEntries += elemcounts.partialEntries;
7509  arraycounts.exactEntries += elemcounts.exactEntries;
7510  arraycounts.searchEntries += elemcounts.searchEntries;
7511  }
7512  }
7513 
7514  if (numPossible == 0)
7515  {
7516  /* No satisfiable patterns in the array */
7517  return false;
7518  }
7519 
7520  /*
7521  * Now add the averages to the global counts. This will give us an
7522  * estimate of the average number of terms searched for in each indexscan,
7523  * including contributions from both array and non-array quals.
7524  */
7525  counts->partialEntries += arraycounts.partialEntries / numPossible;
7526  counts->exactEntries += arraycounts.exactEntries / numPossible;
7527  counts->searchEntries += arraycounts.searchEntries / numPossible;
7528 
7529  counts->arrayScans *= numPossible;
7530 
7531  return true;
7532 }
#define ARR_ELEMTYPE(a)
Definition: array.h:292
void deconstruct_array(ArrayType *array, Oid elmtype, int elmlen, bool elmbyval, char elmalign, Datum **elemsp, bool **nullsp, int *nelemsp)
Definition: arrayfuncs.c:3612
void get_typlenbyvalalign(Oid typid, int16 *typlen, bool *typbyval, char *typalign)
Definition: lsyscache.c:2227
double arrayScans
Definition: selfuncs.c:7249

References arg, ScalarArrayOpExpr::args, ARR_ELEMTYPE, GinQualCounts::arrayScans, Assert(), GinQualCounts::attHasFullScan, GinQualCounts::attHasNormalScan, DatumGetArrayTypeP, deconstruct_array(), estimate_array_length(), estimate_expression_value(), GinQualCounts::exactEntries, get_typlenbyvalalign(), gincost_pattern(), i, IsA, lsecond, ScalarArrayOpExpr::opno, GinQualCounts::partialEntries, GinQualCounts::searchEntries, and ScalarArrayOpExpr::useOr.

Referenced by gincostestimate().

◆ gincostestimate()

void gincostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
Cost indexStartupCost,
Cost indexTotalCost,
Selectivity indexSelectivity,
double *  indexCorrelation,
double *  indexPages 
)

Definition at line 7538 of file selfuncs.c.

7542 {
7543  IndexOptInfo *index = path->indexinfo;
7544  List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7545  List *selectivityQuals;
7546  double numPages = index->pages,
7547  numTuples = index->tuples;
7548  double numEntryPages,
7549  numDataPages,
7550  numPendingPages,
7551  numEntries;
7552  GinQualCounts counts;
7553  bool matchPossible;
7554  bool fullIndexScan;
7555  double partialScale;
7556  double entryPagesFetched,
7557  dataPagesFetched,
7558  dataPagesFetchedBySel;
7559  double qual_op_cost,
7560  qual_arg_cost,
7561  spc_random_page_cost,
7562  outer_scans;
7563  Cost descentCost;
7564  Relation indexRel;
7565  GinStatsData ginStats;
7566  ListCell *lc;
7567  int i;
7568 
7569  /*
7570  * Obtain statistical information from the meta page, if possible. Else
7571  * set ginStats to zeroes, and we'll cope below.
7572  */
7573  if (!index->hypothetical)
7574  {
7575  /* Lock should have already been obtained in plancat.c */
7576  indexRel = index_open(index->indexoid, NoLock);
7577  ginGetStats(indexRel, &ginStats);
7578  index_close(indexRel, NoLock);
7579  }
7580  else
7581  {
7582  memset(&ginStats, 0, sizeof(ginStats));
7583  }
7584 
7585  /*
7586  * Assuming we got valid (nonzero) stats at all, nPendingPages can be
7587  * trusted, but the other fields are data as of the last VACUUM. We can
7588  * scale them up to account for growth since then, but that method only
7589  * goes so far; in the worst case, the stats might be for a completely
7590  * empty index, and scaling them will produce pretty bogus numbers.
7591  * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
7592  * it's grown more than that, fall back to estimating things only from the
7593  * assumed-accurate index size. But we'll trust nPendingPages in any case
7594  * so long as it's not clearly insane, ie, more than the index size.
7595  */
7596  if (ginStats.nPendingPages < numPages)
7597  numPendingPages = ginStats.nPendingPages;
7598  else
7599  numPendingPages = 0;
7600 
7601  if (numPages > 0 && ginStats.nTotalPages <= numPages &&
7602  ginStats.nTotalPages > numPages / 4 &&
7603  ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
7604  {
7605  /*
7606  * OK, the stats seem close enough to sane to be trusted. But we
7607  * still need to scale them by the ratio numPages / nTotalPages to
7608  * account for growth since the last VACUUM.
7609  */
7610  double scale = numPages / ginStats.nTotalPages;
7611 
7612  numEntryPages = ceil(ginStats.nEntryPages * scale);
7613  numDataPages = ceil(ginStats.nDataPages * scale);
7614  numEntries = ceil(ginStats.nEntries * scale);
7615  /* ensure we didn't round up too much */
7616  numEntryPages = Min(numEntryPages, numPages - numPendingPages);
7617  numDataPages = Min(numDataPages,
7618  numPages - numPendingPages - numEntryPages);
7619  }
7620  else
7621  {
7622  /*
7623  * We might get here because it's a hypothetical index, or an index
7624  * created pre-9.1 and never vacuumed since upgrading (in which case
7625  * its stats would read as zeroes), or just because it's grown too
7626  * much since the last VACUUM for us to put our faith in scaling.
7627  *
7628  * Invent some plausible internal statistics based on the index page
7629  * count (and clamp that to at least 10 pages, just in case). We
7630  * estimate that 90% of the index is entry pages, and the rest is data
7631  * pages. Estimate 100 entries per entry page; this is rather bogus
7632  * since it'll depend on the size of the keys, but it's more robust
7633  * than trying to predict the number of entries per heap tuple.
7634  */
7635  numPages = Max(numPages, 10);
7636  numEntryPages = floor((numPages - numPendingPages) * 0.90);
7637  numDataPages = numPages - numPendingPages - numEntryPages;
7638  numEntries = floor(numEntryPages * 100);
7639  }
7640 
7641  /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
7642  if (numEntries < 1)
7643  numEntries = 1;
7644 
7645  /*
7646  * If the index is partial, AND the index predicate with the index-bound
7647  * quals to produce a more accurate idea of the number of rows covered by
7648  * the bound conditions.
7649  */
7650  selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7651 
7652  /* Estimate the fraction of main-table tuples that will be visited */
7653  *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7654  index->rel->relid,
7655  JOIN_INNER,
7656  NULL);
7657 
7658  /* fetch estimated page cost for tablespace containing index */
7659  get_tablespace_page_costs(index->reltablespace,
7660  &spc_random_page_cost,
7661  NULL);
7662 
7663  /*
7664  * Generic assumption about index correlation: there isn't any.
7665  */
7666  *indexCorrelation = 0.0;
7667 
7668  /*
7669  * Examine quals to estimate number of search entries & partial matches
7670  */
7671  memset(&counts, 0, sizeof(counts));
7672  counts.arrayScans = 1;
7673  matchPossible = true;
7674 
7675  foreach(lc, path->indexclauses)
7676  {
7677  IndexClause *iclause = lfirst_node(IndexClause, lc);
7678  ListCell *lc2;
7679 
7680  foreach(lc2, iclause->indexquals)
7681  {
7682  RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7683  Expr *clause = rinfo->clause;
7684 
7685  if (IsA(clause, OpExpr))
7686  {
7687  matchPossible = gincost_opexpr(root,
7688  index,
7689  iclause->indexcol,
7690  (OpExpr *) clause,
7691  &counts);
7692  if (!matchPossible)
7693  break;
7694  }
7695  else if (IsA(clause, ScalarArrayOpExpr))
7696  {
7697  matchPossible = gincost_scalararrayopexpr(root,
7698  index,
7699  iclause->indexcol,
7700  (ScalarArrayOpExpr *) clause,
7701  numEntries,
7702  &counts);
7703  if (!matchPossible)
7704  break;
7705  }
7706  else
7707  {
7708  /* shouldn't be anything else for a GIN index */
7709  elog(ERROR, "unsupported GIN indexqual type: %d",
7710  (int) nodeTag(clause));
7711  }
7712  }
7713  }
7714 
7715  /* Fall out if there were any provably-unsatisfiable quals */
7716  if (!matchPossible)
7717  {
7718  *indexStartupCost = 0;
7719  *indexTotalCost = 0;
7720  *indexSelectivity = 0;
7721  return;
7722  }
7723 
7724  /*
7725  * If attribute has a full scan and at the same time doesn't have normal
7726  * scan, then we'll have to scan all non-null entries of that attribute.
7727  * Currently, we don't have per-attribute statistics for GIN. Thus, we
7728  * must assume the whole GIN index has to be scanned in this case.
7729  */
7730  fullIndexScan = false;
7731  for (i = 0; i < index->nkeycolumns; i++)
7732  {
7733  if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
7734  {
7735  fullIndexScan = true;
7736  break;
7737  }
7738  }
7739 
7740  if (fullIndexScan || indexQuals == NIL)
7741  {
7742  /*
7743  * Full index scan will be required. We treat this as if every key in
7744  * the index had been listed in the query; is that reasonable?
7745  */
7746  counts.partialEntries = 0;
7747  counts.exactEntries = numEntries;
7748  counts.searchEntries = numEntries;
7749  }
7750 
7751  /* Will we have more than one iteration of a nestloop scan? */
7752  outer_scans = loop_count;
7753 
7754  /*
7755  * Compute cost to begin scan, first of all, pay attention to pending
7756  * list.
7757  */
7758  entryPagesFetched = numPendingPages;
7759 
7760  /*
7761  * Estimate number of entry pages read. We need to do
7762  * counts.searchEntries searches. Use a power function as it should be,
7763  * but tuples on leaf pages usually is much greater. Here we include all
7764  * searches in entry tree, including search of first entry in partial
7765  * match algorithm
7766  */
7767  entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
7768 
7769  /*
7770  * Add an estimate of entry pages read by partial match algorithm. It's a
7771  * scan over leaf pages in entry tree. We haven't any useful stats here,
7772  * so estimate it as proportion. Because counts.partialEntries is really
7773  * pretty bogus (see code above), it's possible that it is more than
7774  * numEntries; clamp the proportion to ensure sanity.
7775  */
7776  partialScale = counts.partialEntries / numEntries;
7777  partialScale = Min(partialScale, 1.0);
7778 
7779  entryPagesFetched += ceil(numEntryPages * partialScale);
7780 
7781  /*
7782  * Partial match algorithm reads all data pages before doing actual scan,
7783  * so it's a startup cost. Again, we haven't any useful stats here, so
7784  * estimate it as proportion.
7785  */
7786  dataPagesFetched = ceil(numDataPages * partialScale);
7787 
7788  *indexStartupCost = 0;
7789  *indexTotalCost = 0;
7790 
7791  /*
7792  * Add a CPU-cost component to represent the costs of initial entry btree
7793  * descent. We don't charge any I/O cost for touching upper btree levels,
7794  * since they tend to stay in cache, but we still have to do about log2(N)
7795  * comparisons to descend a btree of N leaf tuples. We charge one
7796  * cpu_operator_cost per comparison.
7797  *
7798  * If there are ScalarArrayOpExprs, charge this once per SA scan. The
7799  * ones after the first one are not startup cost so far as the overall
7800  * plan is concerned, so add them only to "total" cost.
7801  */
7802  if (numEntries > 1) /* avoid computing log(0) */
7803  {
7804  descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
7805  *indexStartupCost += descentCost * counts.searchEntries;
7806  *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
7807  }
7808 
7809  /*
7810  * Add a cpu cost per entry-page fetched. This is not amortized over a
7811  * loop.
7812  */
7813  *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7814  *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7815 
7816  /*
7817  * Add a cpu cost per data-page fetched. This is also not amortized over a
7818  * loop. Since those are the data pages from the partial match algorithm,
7819  * charge them as startup cost.
7820  */
7821  *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
7822 
7823  /*
7824  * Since we add the startup cost to the total cost later on, remove the
7825  * initial arrayscan from the total.
7826  */
7827  *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7828 
7829  /*
7830  * Calculate cache effects if more than one scan due to nestloops or array
7831  * quals. The result is pro-rated per nestloop scan, but the array qual
7832  * factor shouldn't be pro-rated (compare genericcostestimate).
7833  */
7834  if (outer_scans > 1 || counts.arrayScans > 1)
7835  {
7836  entryPagesFetched *= outer_scans * counts.arrayScans;
7837  entryPagesFetched = index_pages_fetched(entryPagesFetched,
7838  (BlockNumber) numEntryPages,
7839  numEntryPages, root);
7840  entryPagesFetched /= outer_scans;
7841  dataPagesFetched *= outer_scans * counts.arrayScans;
7842  dataPagesFetched = index_pages_fetched(dataPagesFetched,
7843  (BlockNumber) numDataPages,
7844  numDataPages, root);
7845  dataPagesFetched /= outer_scans;
7846  }
7847 
7848  /*
7849  * Here we use random page cost because logically-close pages could be far
7850  * apart on disk.
7851  */
7852  *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
7853 
7854  /*
7855  * Now compute the number of data pages fetched during the scan.
7856  *
7857  * We assume every entry to have the same number of items, and that there
7858  * is no overlap between them. (XXX: tsvector and array opclasses collect
7859  * statistics on the frequency of individual keys; it would be nice to use
7860  * those here.)
7861  */
7862  dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
7863 
7864  /*
7865  * If there is a lot of overlap among the entries, in particular if one of
7866  * the entries is very frequent, the above calculation can grossly
7867  * under-estimate. As a simple cross-check, calculate a lower bound based
7868  * on the overall selectivity of the quals. At a minimum, we must read
7869  * one item pointer for each matching entry.
7870  *
7871  * The width of each item pointer varies, based on the level of
7872  * compression. We don't have statistics on that, but an average of
7873  * around 3 bytes per item is fairly typical.
7874  */
7875  dataPagesFetchedBySel = ceil(*indexSelectivity *
7876  (numTuples / (BLCKSZ / 3)));
7877  if (dataPagesFetchedBySel > dataPagesFetched)
7878  dataPagesFetched = dataPagesFetchedBySel;
7879 
7880  /* Add one page cpu-cost to the startup cost */
7881  *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
7882 
7883  /*
7884  * Add once again a CPU-cost for those data pages, before amortizing for
7885  * cache.
7886  */
7887  *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7888 
7889  /* Account for cache effects, the same as above */
7890  if (outer_scans > 1 || counts.arrayScans > 1)
7891  {
7892  dataPagesFetched *= outer_scans * counts.arrayScans;
7893  dataPagesFetched = index_pages_fetched(dataPagesFetched,
7894  (BlockNumber) numDataPages,
7895  numDataPages, root);
7896  dataPagesFetched /= outer_scans;
7897  }
7898 
7899  /* And apply random_page_cost as the cost per page */
7900  *indexTotalCost += *indexStartupCost +
7901  dataPagesFetched * spc_random_page_cost;
7902 
7903  /*
7904  * Add on index qual eval costs, much as in genericcostestimate. We charge
7905  * cpu but we can disregard indexorderbys, since GIN doesn't support
7906  * those.
7907  */
7908  qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
7909  qual_op_cost = cpu_operator_cost * list_length(indexQuals);
7910 
7911  *indexStartupCost += qual_arg_cost;
7912  *indexTotalCost += qual_arg_cost;
7913 
7914  /*
7915  * Add a cpu cost per search entry, corresponding to the actual visited
7916  * entries.
7917  */
7918  *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
7919  /* Now add a cpu cost per tuple in the posting lists / trees */
7920  *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
7921  *indexPages = dataPagesFetched;
7922 }
void ginGetStats(Relation index, GinStatsData *stats)
Definition: ginutil.c:623
int scale
Definition: pgbench.c:181
static bool gincost_scalararrayopexpr(PlannerInfo *root, IndexOptInfo *index, int indexcol, ScalarArrayOpExpr *clause, double numIndexEntries, GinQualCounts *counts)
Definition: selfuncs.c:7422
static bool gincost_opexpr(PlannerInfo *root, IndexOptInfo *index, int indexcol, OpExpr *clause, GinQualCounts *counts)
Definition: selfuncs.c:7372
BlockNumber nDataPages
Definition: gin.h:47
BlockNumber nPendingPages
Definition: gin.h:44
BlockNumber nEntryPages
Definition: gin.h:46
int64 nEntries
Definition: gin.h:48
BlockNumber nTotalPages
Definition: gin.h:45

References add_predicate_to_index_quals(), GinQualCounts::arrayScans, GinQualCounts::attHasFullScan, GinQualCounts::attHasNormalScan, RestrictInfo::clause, clauselist_selectivity(), cpu_index_tuple_cost, cpu_operator_cost, DEFAULT_PAGE_CPU_MULTIPLIER, elog, ERROR, GinQualCounts::exactEntries, get_quals_from_indexclauses(), get_tablespace_page_costs(), gincost_opexpr(), gincost_scalararrayopexpr(), ginGetStats(), i, index_close(), index_open(), index_other_operands_eval_cost(), index_pages_fetched(), IndexPath::indexclauses, IndexClause::indexcol, IndexPath::indexinfo, IndexClause::indexquals, IsA, JOIN_INNER, lfirst_node, list_length(), Max, Min, GinStatsData::nDataPages, GinStatsData::nEntries, GinStatsData::nEntryPages, NIL, nodeTag, NoLock, GinStatsData::nPendingPages, GinStatsData::nTotalPages, GinQualCounts::partialEntries, scale, and GinQualCounts::searchEntries.

Referenced by ginhandler().

◆ gistcostestimate()

void gistcostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
Cost indexStartupCost,
Cost indexTotalCost,
Selectivity indexSelectivity,
double *  indexCorrelation,
double *  indexPages 
)

Definition at line 7128 of file selfuncs.c.

7132 {
7133  IndexOptInfo *index = path->indexinfo;
7134  GenericCosts costs = {0};
7135  Cost descentCost;
7136 
7137  genericcostestimate(root, path, loop_count, &costs);
7138 
7139  /*
7140  * We model index descent costs similarly to those for btree, but to do
7141  * that we first need an idea of the tree height. We somewhat arbitrarily
7142  * assume that the fanout is 100, meaning the tree height is at most
7143  * log100(index->pages).
7144  *
7145  * Although this computation isn't really expensive enough to require
7146  * caching, we might as well use index->tree_height to cache it.
7147  */
7148  if (index->tree_height < 0) /* unknown? */
7149  {
7150  if (index->pages > 1) /* avoid computing log(0) */
7151  index->tree_height = (int) (log(index->pages) / log(100.0));
7152  else
7153  index->tree_height = 0;
7154  }
7155 
7156  /*
7157  * Add a CPU-cost component to represent the costs of initial descent. We
7158  * just use log(N) here not log2(N) since the branching factor isn't
7159  * necessarily two anyway. As for btree, charge once per SA scan.
7160  */
7161  if (index->tuples > 1) /* avoid computing log(0) */
7162  {
7163  descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7164  costs.indexStartupCost += descentCost;
7165  costs.indexTotalCost += costs.num_sa_scans * descentCost;
7166  }
7167 
7168  /*
7169  * Likewise add a per-page charge, calculated the same as for btrees.
7170  */
7171  descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7172  costs.indexStartupCost += descentCost;
7173  costs.indexTotalCost += costs.num_sa_scans * descentCost;
7174 
7175  *indexStartupCost = costs.indexStartupCost;
7176  *indexTotalCost = costs.indexTotalCost;
7177  *indexSelectivity = costs.indexSelectivity;
7178  *indexCorrelation = costs.indexCorrelation;
7179  *indexPages = costs.numIndexPages;
7180 }

References cpu_operator_cost, DEFAULT_PAGE_CPU_MULTIPLIER, genericcostestimate(), GenericCosts::indexCorrelation, IndexPath::indexinfo, GenericCosts::indexSelectivity, GenericCosts::indexStartupCost, GenericCosts::indexTotalCost, GenericCosts::num_sa_scans, and GenericCosts::numIndexPages.

Referenced by gisthandler().

◆ hashcostestimate()

void hashcostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
Cost indexStartupCost,
Cost indexTotalCost,
Selectivity indexSelectivity,
double *  indexCorrelation,
double *  indexPages 
)

Definition at line 7086 of file selfuncs.c.

7090 {
7091  GenericCosts costs = {0};
7092 
7093  genericcostestimate(root, path, loop_count, &costs);
7094 
7095  /*
7096  * A hash index has no descent costs as such, since the index AM can go
7097  * directly to the target bucket after computing the hash value. There
7098  * are a couple of other hash-specific costs that we could conceivably add
7099  * here, though:
7100  *
7101  * Ideally we'd charge spc_random_page_cost for each page in the target
7102  * bucket, not just the numIndexPages pages that genericcostestimate
7103  * thought we'd visit. However in most cases we don't know which bucket
7104  * that will be. There's no point in considering the average bucket size
7105  * because the hash AM makes sure that's always one page.
7106  *
7107  * Likewise, we could consider charging some CPU for each index tuple in
7108  * the bucket, if we knew how many there were. But the per-tuple cost is
7109  * just a hash value comparison, not a general datatype-dependent
7110  * comparison, so any such charge ought to be quite a bit less than
7111  * cpu_operator_cost; which makes it probably not worth worrying about.
7112  *
7113  * A bigger issue is that chance hash-value collisions will result in
7114  * wasted probes into the heap. We don't currently attempt to model this
7115  * cost on the grounds that it's rare, but maybe it's not rare enough.
7116  * (Any fix for this ought to consider the generic lossy-operator problem,
7117  * though; it's not entirely hash-specific.)
7118  */
7119 
7120  *indexStartupCost = costs.indexStartupCost;
7121  *indexTotalCost = costs.indexTotalCost;
7122  *indexSelectivity = costs.indexSelectivity;
7123  *indexCorrelation = costs.indexCorrelation;
7124  *indexPages = costs.numIndexPages;
7125 }

References genericcostestimate(), GenericCosts::indexCorrelation, GenericCosts::indexSelectivity, GenericCosts::indexStartupCost, GenericCosts::indexTotalCost, and GenericCosts::numIndexPages.

Referenced by hashhandler().

◆ histogram_selectivity()

double histogram_selectivity ( VariableStatData vardata,
FmgrInfo opproc,
Oid  collation,
Datum  constval,
bool  varonleft,
int  min_hist_size,
int  n_skip,
int *  hist_size 
)

Definition at line 823 of file selfuncs.c.

828 {
829  double result;
830  AttStatsSlot sslot;
831 
832  /* check sanity of parameters */
833  Assert(n_skip >= 0);
834  Assert(min_hist_size > 2 * n_skip);
835 
836  if (HeapTupleIsValid(vardata->statsTuple) &&
837  statistic_proc_security_check(vardata, opproc->fn_oid) &&
838  get_attstatsslot(&sslot, vardata->statsTuple,
839  STATISTIC_KIND_HISTOGRAM, InvalidOid,
841  {
842  *hist_size = sslot.nvalues;
843  if (sslot.nvalues >= min_hist_size)
844  {
845  LOCAL_FCINFO(fcinfo, 2);
846  int nmatch = 0;
847  int i;
848 
849  /*
850  * We invoke the opproc "by hand" so that we won't fail on NULL
851  * results. Such cases won't arise for normal comparison
852  * functions, but generic_restriction_selectivity could perhaps be
853  * used with operators that can return NULL. A small side benefit
854  * is to not need to re-initialize the fcinfo struct from scratch
855  * each time.
856  */
857  InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
858  NULL, NULL);
859  fcinfo->args[0].isnull = false;
860  fcinfo->args[1].isnull = false;
861  /* be careful to apply operator right way 'round */
862  if (varonleft)
863  fcinfo->args[1].value = constval;
864  else
865  fcinfo->args[0].value = constval;
866 
867  for (i = n_skip; i < sslot.nvalues - n_skip; i++)
868  {
869  Datum fresult;
870 
871  if (varonleft)
872  fcinfo->args[0].value = sslot.values[i];
873  else
874  fcinfo->args[1].value = sslot.values[i];
875  fcinfo->isnull = false;
876  fresult = FunctionCallInvoke(fcinfo);
877  if (!fcinfo->isnull && DatumGetBool(fresult))
878  nmatch++;
879  }
880  result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
881  }
882  else
883  result = -1;
884  free_attstatsslot(&sslot);
885  }
886  else
887  {
888  *hist_size = 0;
889  result = -1;
890  }
891 
892  return result;
893 }

References Assert(), ATTSTATSSLOT_VALUES, DatumGetBool(), FmgrInfo::fn_oid, free_attstatsslot(), FunctionCallInvoke, get_attstatsslot(), HeapTupleIsValid, i, InitFunctionCallInfoData, InvalidOid, LOCAL_FCINFO, AttStatsSlot::nvalues, statistic_proc_security_check(), VariableStatData::statsTuple, and AttStatsSlot::values.

Referenced by generic_restriction_selectivity(), and patternsel_common().

◆ index_other_operands_eval_cost()

Cost index_other_operands_eval_cost ( PlannerInfo root,
List indexquals 
)

Definition at line 6490 of file selfuncs.c.

6491 {
6492  Cost qual_arg_cost = 0;
6493  ListCell *lc;
6494 
6495  foreach(lc, indexquals)
6496  {
6497  Expr *clause = (Expr *) lfirst(lc);
6498  Node *other_operand;
6499  QualCost index_qual_cost;
6500 
6501  /*
6502  * Index quals will have RestrictInfos, indexorderbys won't. Look
6503  * through RestrictInfo if present.
6504  */
6505  if (IsA(clause, RestrictInfo))
6506  clause = ((RestrictInfo *) clause)->clause;
6507 
6508  if (IsA(clause, OpExpr))
6509  {
6510  OpExpr *op = (OpExpr *) clause;
6511 
6512  other_operand = (Node *) lsecond(op->args);
6513  }
6514  else if (IsA(clause, RowCompareExpr))
6515  {
6516  RowCompareExpr *rc = (RowCompareExpr *) clause;
6517 
6518  other_operand = (Node *) rc->rargs;
6519  }
6520  else if (IsA(clause, ScalarArrayOpExpr))
6521  {
6522  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6523 
6524  other_operand = (Node *) lsecond(saop->args);
6525  }
6526  else if (IsA(clause, NullTest))
6527  {
6528  other_operand = NULL;
6529  }
6530  else
6531  {
6532  elog(ERROR, "unsupported indexqual type: %d",
6533  (int) nodeTag(clause));
6534  other_operand = NULL; /* keep compiler quiet */
6535  }
6536 
6537  cost_qual_eval_node(&index_qual_cost, other_operand, root);
6538  qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6539  }
6540  return qual_arg_cost;
6541 }
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:4666
Cost per_tuple
Definition: pathnodes.h:48
Cost startup
Definition: pathnodes.h:47

References OpExpr::args, ScalarArrayOpExpr::args, cost_qual_eval_node(), elog, ERROR, IsA, lfirst, lsecond, nodeTag, QualCost::per_tuple, RowCompareExpr::rargs, and QualCost::startup.

Referenced by brincostestimate(), genericcostestimate(), and gincostestimate().

◆ ineq_histogram_selectivity()

double ineq_histogram_selectivity ( PlannerInfo root,
VariableStatData vardata,
Oid  opoid,
FmgrInfo opproc,
bool  isgt,
bool  iseq,
Oid  collation,
Datum  constval,
Oid  consttype 
)

Definition at line 1041 of file selfuncs.c.

1046 {
1047  double hist_selec;
1048  AttStatsSlot sslot;
1049 
1050  hist_selec = -1.0;
1051 
1052  /*
1053  * Someday, ANALYZE might store more than one histogram per rel/att,
1054  * corresponding to more than one possible sort ordering defined for the
1055  * column type. Right now, we know there is only one, so just grab it and
1056  * see if it matches the query.
1057  *
1058  * Note that we can't use opoid as search argument; the staop appearing in
1059  * pg_statistic will be for the relevant '<' operator, but what we have
1060  * might be some other inequality operator such as '>='. (Even if opoid
1061  * is a '<' operator, it could be cross-type.) Hence we must use
1062  * comparison_ops_are_compatible() to see if the operators match.
1063  */
1064  if (HeapTupleIsValid(vardata->statsTuple) &&
1065  statistic_proc_security_check(vardata, opproc->fn_oid) &&
1066  get_attstatsslot(&sslot, vardata->statsTuple,
1067  STATISTIC_KIND_HISTOGRAM, InvalidOid,
1069  {
1070  if (sslot.nvalues > 1 &&
1071  sslot.stacoll == collation &&
1072  comparison_ops_are_compatible(sslot.staop, opoid))
1073  {
1074  /*
1075  * Use binary search to find the desired location, namely the
1076  * right end of the histogram bin containing the comparison value,
1077  * which is the leftmost entry for which the comparison operator
1078  * succeeds (if isgt) or fails (if !isgt).
1079  *
1080  * In this loop, we pay no attention to whether the operator iseq
1081  * or not; that detail will be mopped up below. (We cannot tell,
1082  * anyway, whether the operator thinks the values are equal.)
1083  *
1084  * If the binary search accesses the first or last histogram
1085  * entry, we try to replace that endpoint with the true column min
1086  * or max as found by get_actual_variable_range(). This
1087  * ameliorates misestimates when the min or max is moving as a
1088  * result of changes since the last ANALYZE. Note that this could
1089  * result in effectively including MCVs into the histogram that
1090  * weren't there before, but we don't try to correct for that.
1091  */
1092  double histfrac;
1093  int lobound = 0; /* first possible slot to search */
1094  int hibound = sslot.nvalues; /* last+1 slot to search */
1095  bool have_end = false;
1096 
1097  /*
1098  * If there are only two histogram entries, we'll want up-to-date
1099  * values for both. (If there are more than two, we need at most
1100  * one of them to be updated, so we deal with that within the
1101  * loop.)
1102  */
1103  if (sslot.nvalues == 2)
1104  have_end = get_actual_variable_range(root,
1105  vardata,
1106  sslot.staop,
1107  collation,
1108  &sslot.values[0],
1109  &sslot.values[1]);
1110 
1111  while (lobound < hibound)
1112  {
1113  int probe = (lobound + hibound) / 2;
1114  bool ltcmp;
1115 
1116  /*
1117  * If we find ourselves about to compare to the first or last
1118  * histogram entry, first try to replace it with the actual
1119  * current min or max (unless we already did so above).
1120  */
1121  if (probe == 0 && sslot.nvalues > 2)
1122  have_end = get_actual_variable_range(root,
1123  vardata,
1124  sslot.staop,
1125  collation,
1126  &sslot.values[0],
1127  NULL);
1128  else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1129  have_end = get_actual_variable_range(root,
1130  vardata,
1131  sslot.staop,
1132  collation,
1133  NULL,
1134  &sslot.values[probe]);
1135 
1136  ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1137  collation,
1138  sslot.values[probe],
1139  constval));
1140  if (isgt)
1141  ltcmp = !ltcmp;
1142  if (ltcmp)
1143  lobound = probe + 1;
1144  else
1145  hibound = probe;
1146  }
1147 
1148  if (lobound <= 0)
1149  {
1150  /*
1151  * Constant is below lower histogram boundary. More
1152  * precisely, we have found that no entry in the histogram
1153  * satisfies the inequality clause (if !isgt) or they all do
1154  * (if isgt). We estimate that that's true of the entire
1155  * table, so set histfrac to 0.0 (which we'll flip to 1.0
1156  * below, if isgt).
1157  */
1158  histfrac = 0.0;
1159  }
1160  else if (lobound >= sslot.nvalues)
1161  {
1162  /*
1163  * Inverse case: constant is above upper histogram boundary.
1164  */
1165  histfrac = 1.0;
1166  }
1167  else
1168  {
1169  /* We have values[i-1] <= constant <= values[i]. */
1170  int i = lobound;
1171  double eq_selec = 0;
1172  double val,
1173  high,
1174  low;
1175  double binfrac;
1176 
1177  /*
1178  * In the cases where we'll need it below, obtain an estimate
1179  * of the selectivity of "x = constval". We use a calculation
1180  * similar to what var_eq_const() does for a non-MCV constant,
1181  * ie, estimate that all distinct non-MCV values occur equally
1182  * often. But multiplication by "1.0 - sumcommon - nullfrac"
1183  * will be done by our caller, so we shouldn't do that here.
1184  * Therefore we can't try to clamp the estimate by reference
1185  * to the least common MCV; the result would be too small.
1186  *
1187  * Note: since this is effectively assuming that constval
1188  * isn't an MCV, it's logically dubious if constval in fact is
1189  * one. But we have to apply *some* correction for equality,
1190  * and anyway we cannot tell if constval is an MCV, since we
1191  * don't have a suitable equality operator at hand.
1192  */
1193  if (i == 1 || isgt == iseq)
1194  {
1195  double otherdistinct;
1196  bool isdefault;
1197  AttStatsSlot mcvslot;
1198 
1199  /* Get estimated number of distinct values */
1200  otherdistinct = get_variable_numdistinct(vardata,
1201  &isdefault);
1202 
1203  /* Subtract off the number of known MCVs */
1204  if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1205  STATISTIC_KIND_MCV, InvalidOid,
1207  {
1208  otherdistinct -= mcvslot.nnumbers;
1209  free_attstatsslot(&mcvslot);
1210  }
1211 
1212  /* If result doesn't seem sane, leave eq_selec at 0 */
1213  if (otherdistinct > 1)
1214  eq_selec = 1.0 / otherdistinct;
1215  }
1216 
1217  /*
1218  * Convert the constant and the two nearest bin boundary
1219  * values to a uniform comparison scale, and do a linear
1220  * interpolation within this bin.
1221  */
1222  if (convert_to_scalar(constval, consttype, collation,
1223  &val,
1224  sslot.values[i - 1], sslot.values[i],
1225  vardata->vartype,
1226  &low, &high))
1227  {
1228  if (high <= low)
1229  {
1230  /* cope if bin boundaries appear identical */
1231  binfrac = 0.5;
1232  }
1233  else if (val <= low)
1234  binfrac = 0.0;
1235  else if (val >= high)
1236  binfrac = 1.0;
1237  else
1238  {
1239  binfrac = (val - low) / (high - low);
1240 
1241  /*
1242  * Watch out for the possibility that we got a NaN or
1243  * Infinity from the division. This can happen
1244  * despite the previous checks, if for example "low"
1245  * is -Infinity.
1246  */
1247  if (isnan(binfrac) ||
1248  binfrac < 0.0 || binfrac > 1.0)
1249  binfrac = 0.5;
1250  }
1251  }
1252  else
1253  {
1254  /*
1255  * Ideally we'd produce an error here, on the grounds that
1256  * the given operator shouldn't have scalarXXsel
1257  * registered as its selectivity func unless we can deal
1258  * with its operand types. But currently, all manner of
1259  * stuff is invoking scalarXXsel, so give a default
1260  * estimate until that can be fixed.
1261  */
1262  binfrac = 0.5;
1263  }
1264 
1265  /*
1266  * Now, compute the overall selectivity across the values
1267  * represented by the histogram. We have i-1 full bins and
1268  * binfrac partial bin below the constant.
1269  */
1270  histfrac = (double) (i - 1) + binfrac;
1271  histfrac /= (double) (sslot.nvalues - 1);
1272 
1273  /*
1274  * At this point, histfrac is an estimate of the fraction of
1275  * the population represented by the histogram that satisfies
1276  * "x <= constval". Somewhat remarkably, this statement is
1277  * true regardless of which operator we were doing the probes
1278  * with, so long as convert_to_scalar() delivers reasonable
1279  * results. If the probe constant is equal to some histogram
1280  * entry, we would have considered the bin to the left of that
1281  * entry if probing with "<" or ">=", or the bin to the right
1282  * if probing with "<=" or ">"; but binfrac would have come
1283  * out as 1.0 in the first case and 0.0 in the second, leading
1284  * to the same histfrac in either case. For probe constants
1285  * between histogram entries, we find the same bin and get the
1286  * same estimate with any operator.
1287  *
1288  * The fact that the estimate corresponds to "x <= constval"
1289  * and not "x < constval" is because of the way that ANALYZE
1290  * constructs the histogram: each entry is, effectively, the
1291  * rightmost value in its sample bucket. So selectivity
1292  * values that are exact multiples of 1/(histogram_size-1)
1293  * should be understood as estimates including a histogram
1294  * entry plus everything to its left.
1295  *
1296  * However, that breaks down for the first histogram entry,
1297  * which necessarily is the leftmost value in its sample
1298  * bucket. That means the first histogram bin is slightly
1299  * narrower than the rest, by an amount equal to eq_selec.
1300  * Another way to say that is that we want "x <= leftmost" to
1301  * be estimated as eq_selec not zero. So, if we're dealing
1302  * with the first bin (i==1), rescale to make that true while
1303  * adjusting the rest of that bin linearly.
1304  */
1305  if (i == 1)
1306  histfrac += eq_selec * (1.0 - binfrac);
1307 
1308  /*
1309  * "x <= constval" is good if we want an estimate for "<=" or
1310  * ">", but if we are estimating for "<" or ">=", we now need
1311  * to decrease the estimate by eq_selec.
1312  */
1313  if (isgt == iseq)
1314  histfrac -= eq_selec;
1315  }
1316 
1317  /*
1318  * Now the estimate is finished for "<" and "<=" cases. If we are
1319  * estimating for ">" or ">=", flip it.
1320  */
1321  hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1322 
1323  /*
1324  * The histogram boundaries are only approximate to begin with,
1325  * and may well be out of date anyway. Therefore, don't believe
1326  * extremely small or large selectivity estimates --- unless we
1327  * got actual current endpoint values from the table, in which
1328  * case just do the usual sanity clamp. Somewhat arbitrarily, we
1329  * set the cutoff for other cases at a hundredth of the histogram
1330  * resolution.
1331  */
1332  if (have_end)
1333  CLAMP_PROBABILITY(hist_selec);
1334  else
1335  {
1336  double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1337 
1338  if (hist_selec < cutoff)
1339  hist_selec = cutoff;
1340  else if (hist_selec > 1.0 - cutoff)
1341  hist_selec = 1.0 - cutoff;
1342  }
1343  }
1344  else if (sslot.nvalues > 1)
1345  {
1346  /*
1347  * If we get here, we have a histogram but it's not sorted the way
1348  * we want. Do a brute-force search to see how many of the
1349  * entries satisfy the comparison condition, and take that
1350  * fraction as our estimate. (This is identical to the inner loop
1351  * of histogram_selectivity; maybe share code?)
1352  */
1353  LOCAL_FCINFO(fcinfo, 2);
1354  int nmatch = 0;
1355 
1356  InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1357  NULL, NULL);
1358  fcinfo->args[0].isnull = false;
1359  fcinfo->args[1].isnull = false;
1360  fcinfo->args[1].value = constval;
1361  for (int i = 0; i < sslot.nvalues; i++)
1362  {
1363  Datum fresult;
1364 
1365  fcinfo->args[0].value = sslot.values[i];
1366  fcinfo->isnull = false;
1367  fresult = FunctionCallInvoke(fcinfo);
1368  if (!fcinfo->isnull && DatumGetBool(fresult))
1369  nmatch++;
1370  }
1371  hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1372 
1373  /*
1374  * As above, clamp to a hundredth of the histogram resolution.
1375  * This case is surely even less trustworthy than the normal one,
1376  * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1377  * clamp should be more restrictive in this case?)
1378  */
1379  {
1380  double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1381 
1382  if (hist_selec < cutoff)
1383  hist_selec = cutoff;
1384  else if (hist_selec > 1.0 - cutoff)
1385  hist_selec = 1.0 - cutoff;
1386  }
1387  }
1388 
1389  free_attstatsslot(&sslot);
1390  }
1391 
1392  return hist_selec;
1393 }
bool comparison_ops_are_compatible(Oid opno1, Oid opno2)
Definition: lsyscache.c:749
static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue, Datum lobound, Datum hibound, Oid boundstypid, double *scaledlobound, double *scaledhibound)
Definition: selfuncs.c:4305

References ATTSTATSSLOT_NUMBERS, ATTSTATSSLOT_VALUES, CLAMP_PROBABILITY, comparison_ops_are_compatible(), convert_to_scalar(), DatumGetBool(), FmgrInfo::fn_oid, free_attstatsslot(), FunctionCall2Coll(), FunctionCallInvoke, get_actual_variable_range(), get_attstatsslot(), get_variable_numdistinct(), HeapTupleIsValid, i, InitFunctionCallInfoData, InvalidOid, LOCAL_FCINFO, AttStatsSlot::nnumbers, AttStatsSlot::nvalues, AttStatsSlot::stacoll, AttStatsSlot::staop, statistic_proc_security_check(), VariableStatData::statsTuple, val, AttStatsSlot::values, and VariableStatData::vartype.

Referenced by prefix_selectivity(), and scalarineqsel().

◆ matchingjoinsel()

Datum matchingjoinsel ( PG_FUNCTION_ARGS  )

Definition at line 3275 of file selfuncs.c.

3276 {
3277  /* Just punt, for the moment. */
3279 }
#define DEFAULT_MATCHING_SEL
Definition: selfuncs.h:49

References DEFAULT_MATCHING_SEL, and PG_RETURN_FLOAT8.

◆ matchingsel()

Datum matchingsel ( PG_FUNCTION_ARGS  )

Definition at line 3257 of file selfuncs.c.

3258 {
3259  PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3260  Oid operator = PG_GETARG_OID(1);
3261  List *args = (List *) PG_GETARG_POINTER(2);
3262  int varRelid = PG_GETARG_INT32(3);
3263  Oid collation = PG_GET_COLLATION();
3264  double selec;
3265 
3266  /* Use generic restriction selectivity logic. */
3267  selec = generic_restriction_selectivity(root, operator, collation,
3268  args, varRelid,
3270 
3271  PG_RETURN_FLOAT8((float8) selec);
3272 }
double generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation, List *args, int varRelid, double default_selectivity)
Definition: selfuncs.c:914

References generate_unaccent_rules::args, DEFAULT_MATCHING_SEL, generic_restriction_selectivity(), PG_GET_COLLATION, PG_GETARG_INT32, PG_GETARG_OID, PG_GETARG_POINTER, and PG_RETURN_FLOAT8.

◆ mcv_selectivity()

double mcv_selectivity ( VariableStatData vardata,
FmgrInfo opproc,
Oid  collation,
Datum  constval,
bool  varonleft,
double *  sumcommonp 
)

Definition at line 732 of file selfuncs.c.

735 {
736  double mcv_selec,
737  sumcommon;
738  AttStatsSlot sslot;
739  int i;
740 
741  mcv_selec = 0.0;
742  sumcommon = 0.0;
743 
744  if (HeapTupleIsValid(vardata->statsTuple) &&
745  statistic_proc_security_check(vardata, opproc->fn_oid) &&
746  get_attstatsslot(&sslot, vardata->statsTuple,
747  STATISTIC_KIND_MCV, InvalidOid,
749  {
750  LOCAL_FCINFO(fcinfo, 2);
751 
752  /*
753  * We invoke the opproc "by hand" so that we won't fail on NULL
754  * results. Such cases won't arise for normal comparison functions,
755  * but generic_restriction_selectivity could perhaps be used with
756  * operators that can return NULL. A small side benefit is to not
757  * need to re-initialize the fcinfo struct from scratch each time.
758  */
759  InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
760  NULL, NULL);
761  fcinfo->args[0].isnull = false;
762  fcinfo->args[1].isnull = false;
763  /* be careful to apply operator right way 'round */
764  if (varonleft)
765  fcinfo->args[1].value = constval;
766  else
767  fcinfo->args[0].value = constval;
768 
769  for (i = 0; i < sslot.nvalues; i++)
770  {
771  Datum fresult;
772 
773  if (varonleft)
774  fcinfo->args[0].value = sslot.values[i];
775  else
776  fcinfo->args[1].value = sslot.values[i];
777  fcinfo->isnull = false;
778  fresult = FunctionCallInvoke(fcinfo);
779  if (!fcinfo->isnull && DatumGetBool(fresult))
780  mcv_selec += sslot.numbers[i];
781  sumcommon += sslot.numbers[i];
782  }
783  free_attstatsslot(&sslot);
784  }
785 
786  *sumcommonp = sumcommon;
787  return mcv_selec;
788 }

References ATTSTATSSLOT_NUMBERS, ATTSTATSSLOT_VALUES, DatumGetBool(), FmgrInfo::fn_oid, free_attstatsslot(), FunctionCallInvoke, get_attstatsslot(), HeapTupleIsValid, i, InitFunctionCallInfoData, InvalidOid, LOCAL_FCINFO, AttStatsSlot::numbers, AttStatsSlot::nvalues, statistic_proc_security_check(), VariableStatData::statsTuple, and AttStatsSlot::values.

Referenced by generic_restriction_selectivity(), networksel(), patternsel_common(), and scalarineqsel().

◆ mergejoinscansel()

void mergejoinscansel ( PlannerInfo root,
Node clause,
Oid  opfamily,
int  strategy,
bool  nulls_first,
Selectivity leftstart,
Selectivity leftend,
Selectivity rightstart,
Selectivity rightend 
)

Definition at line 2952 of file selfuncs.c.

2956 {
2957  Node *left,
2958  *right;
2959  VariableStatData leftvar,
2960  rightvar;
2961  int op_strategy;
2962  Oid op_lefttype;
2963  Oid op_righttype;
2964  Oid opno,
2965  collation,
2966  lsortop,
2967  rsortop,
2968  lstatop,
2969  rstatop,
2970  ltop,
2971  leop,
2972  revltop,
2973  revleop;
2974  bool isgt;
2975  Datum leftmin,
2976  leftmax,
2977  rightmin,
2978  rightmax;
2979  double selec;
2980 
2981  /* Set default results if we can't figure anything out. */
2982  /* XXX should default "start" fraction be a bit more than 0? */
2983  *leftstart = *rightstart = 0.0;
2984  *leftend = *rightend = 1.0;
2985 
2986  /* Deconstruct the merge clause */
2987  if (!is_opclause(clause))
2988  return; /* shouldn't happen */
2989  opno = ((OpExpr *) clause)->opno;
2990  collation = ((OpExpr *) clause)->inputcollid;
2991  left = get_leftop((Expr *) clause);
2992  right = get_rightop((Expr *) clause);
2993  if (!right)
2994  return; /* shouldn't happen */
2995 
2996  /* Look for stats for the inputs */
2997  examine_variable(root, left, 0, &leftvar);
2998  examine_variable(root, right, 0, &rightvar);
2999 
3000  /* Extract the operator's declared left/right datatypes */
3001  get_op_opfamily_properties(opno, opfamily, false,
3002  &op_strategy,
3003  &op_lefttype,
3004  &op_righttype);
3005  Assert(op_strategy == BTEqualStrategyNumber);
3006 
3007  /*
3008  * Look up the various operators we need. If we don't find them all, it
3009  * probably means the opfamily is broken, but we just fail silently.
3010  *
3011  * Note: we expect that pg_statistic histograms will be sorted by the '<'
3012  * operator, regardless of which sort direction we are considering.
3013  */
3014  switch (strategy)
3015  {
3016  case BTLessStrategyNumber:
3017  isgt = false;
3018  if (op_lefttype == op_righttype)
3019  {
3020  /* easy case */
3021  ltop = get_opfamily_member(opfamily,
3022  op_lefttype, op_righttype,
3024  leop = get_opfamily_member(opfamily,
3025  op_lefttype, op_righttype,
3027  lsortop = ltop;
3028  rsortop = ltop;
3029  lstatop = lsortop;
3030  rstatop = rsortop;
3031  revltop = ltop;
3032  revleop = leop;
3033  }
3034  else
3035  {
3036  ltop = get_opfamily_member(opfamily,
3037  op_lefttype, op_righttype,
3039  leop = get_opfamily_member(opfamily,
3040  op_lefttype, op_righttype,
3042  lsortop = get_opfamily_member(opfamily,
3043  op_lefttype, op_lefttype,
3045  rsortop = get_opfamily_member(opfamily,
3046  op_righttype, op_righttype,
3048  lstatop = lsortop;
3049  rstatop = rsortop;
3050  revltop = get_opfamily_member(opfamily,
3051  op_righttype, op_lefttype,
3053  revleop = get_opfamily_member(opfamily,
3054  op_righttype, op_lefttype,
3056  }
3057  break;
3059  /* descending-order case */
3060  isgt = true;
3061  if (op_lefttype == op_righttype)
3062  {
3063  /* easy case */
3064  ltop = get_opfamily_member(opfamily,
3065  op_lefttype, op_righttype,
3067  leop = get_opfamily_member(opfamily,
3068  op_lefttype, op_righttype,
3070  lsortop = ltop;
3071  rsortop = ltop;
3072  lstatop = get_opfamily_member(opfamily,
3073  op_lefttype, op_lefttype,
3075  rstatop = lstatop;
3076  revltop = ltop;
3077  revleop = leop;
3078  }
3079  else
3080  {
3081  ltop = get_opfamily_member(opfamily,
3082  op_lefttype, op_righttype,
3084  leop = get_opfamily_member(opfamily,
3085  op_lefttype, op_righttype,
3087  lsortop = get_opfamily_member(opfamily,
3088  op_lefttype, op_lefttype,
3090  rsortop = get_opfamily_member(opfamily,
3091  op_righttype, op_righttype,
3093  lstatop = get_opfamily_member(opfamily,
3094  op_lefttype, op_lefttype,
3096  rstatop = get_opfamily_member(opfamily,
3097  op_righttype, op_righttype,
3099  revltop = get_opfamily_member(opfamily,
3100  op_righttype, op_lefttype,
3102  revleop = get_opfamily_member(opfamily,
3103  op_righttype, op_lefttype,
3105  }
3106  break;
3107  default:
3108  goto fail; /* shouldn't get here */
3109  }
3110 
3111  if (!OidIsValid(lsortop) ||
3112  !OidIsValid(rsortop) ||
3113  !OidIsValid(lstatop) ||
3114  !OidIsValid(rstatop) ||
3115  !OidIsValid(ltop) ||
3116  !OidIsValid(leop) ||
3117  !OidIsValid(revltop) ||
3118  !OidIsValid(revleop))
3119  goto fail; /* insufficient info in catalogs */
3120 
3121  /* Try to get ranges of both inputs */
3122  if (!isgt)
3123  {
3124  if (!get_variable_range(root, &leftvar, lstatop, collation,
3125  &leftmin, &leftmax))
3126  goto fail; /* no range available from stats */
3127  if (!get_variable_range(root, &rightvar, rstatop, collation,
3128  &rightmin, &rightmax))
3129  goto fail; /* no range available from stats */
3130  }
3131  else
3132  {
3133  /* need to swap the max and min */
3134  if (!get_variable_range(root, &leftvar, lstatop, collation,
3135  &leftmax, &leftmin))
3136  goto fail; /* no range available from stats */
3137  if (!get_variable_range(root, &rightvar, rstatop, collation,
3138  &rightmax, &rightmin))
3139  goto fail; /* no range available from stats */
3140  }
3141 
3142  /*
3143  * Now, the fraction of the left variable that will be scanned is the
3144  * fraction that's <= the right-side maximum value. But only believe
3145  * non-default estimates, else stick with our 1.0.
3146  */
3147  selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3148  rightmax, op_righttype);
3149  if (selec != DEFAULT_INEQ_SEL)
3150  *leftend = selec;
3151 
3152  /* And similarly for the right variable. */
3153  selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3154  leftmax, op_lefttype);
3155  if (selec != DEFAULT_INEQ_SEL)
3156  *rightend = selec;
3157 
3158  /*
3159  * Only one of the two "end" fractions can really be less than 1.0;
3160  * believe the smaller estimate and reset the other one to exactly 1.0. If
3161  * we get exactly equal estimates (as can easily happen with self-joins),
3162  * believe neither.
3163  */
3164  if (*leftend > *rightend)
3165  *leftend = 1.0;
3166  else if (*leftend < *rightend)
3167  *rightend = 1.0;
3168  else
3169  *leftend = *rightend = 1.0;
3170 
3171  /*
3172  * Also, the fraction of the left variable that will be scanned before the
3173  * first join pair is found is the fraction that's < the right-side
3174  * minimum value. But only believe non-default estimates, else stick with
3175  * our own default.
3176  */
3177  selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3178  rightmin, op_righttype);
3179  if (selec != DEFAULT_INEQ_SEL)
3180  *leftstart = selec;
3181 
3182  /* And similarly for the right variable. */
3183  selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3184  leftmin, op_lefttype);
3185  if (selec != DEFAULT_INEQ_SEL)
3186  *rightstart = selec;
3187 
3188  /*
3189  * Only one of the two "start" fractions can really be more than zero;
3190  * believe the larger estimate and reset the other one to exactly 0.0. If
3191  * we get exactly equal estimates (as can easily happen with self-joins),
3192  * believe neither.
3193  */
3194  if (*leftstart < *rightstart)
3195  *leftstart = 0.0;
3196  else if (*leftstart > *rightstart)
3197  *rightstart = 0.0;
3198  else
3199  *leftstart = *rightstart = 0.0;
3200 
3201  /*
3202  * If the sort order is nulls-first, we're going to have to skip over any
3203  * nulls too. These would not have been counted by scalarineqsel, and we
3204  * can safely add in this fraction regardless of whether we believe
3205  * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3206  */
3207  if (nulls_first)
3208  {
3209  Form_pg_statistic stats;
3210 
3211  if (HeapTupleIsValid(leftvar.statsTuple))
3212  {
3213  stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3214  *leftstart += stats->stanullfrac;
3215  CLAMP_PROBABILITY(*leftstart);
3216  *leftend += stats->stanullfrac;
3217  CLAMP_PROBABILITY(*leftend);
3218  }
3219  if (HeapTupleIsValid(rightvar.statsTuple))
3220  {
3221  stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3222  *rightstart += stats->stanullfrac;
3223  CLAMP_PROBABILITY(*rightstart);
3224  *rightend += stats->stanullfrac;
3225  CLAMP_PROBABILITY(*rightend);
3226  }
3227  }
3228 
3229  /* Disbelieve start >= end, just in case that can happen */
3230  if (*leftstart >= *leftend)
3231  {
3232  *leftstart = 0.0;
3233  *leftend = 1.0;
3234  }
3235  if (*rightstart >= *rightend)
3236  {
3237  *rightstart = 0.0;
3238  *rightend = 1.0;
3239  }
3240 
3241 fail:
3242  ReleaseVariableStats(leftvar);
3243  ReleaseVariableStats(rightvar);
3244 }
static bool is_opclause(const void *clause)
Definition: nodeFuncs.h:74
static Node * get_rightop(const void *clause)
Definition: nodeFuncs.h:93
static Node * get_leftop(const void *clause)
Definition: nodeFuncs.h:81
static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata, Oid sortop, Oid collation, Datum *min, Datum *max)
Definition: selfuncs.c:5897
static double scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq, Oid collation, VariableStatData *vardata, Datum constval, Oid consttype)
Definition: selfuncs.c:580
#define DEFAULT_INEQ_SEL
Definition: selfuncs.h:37
#define BTLessEqualStrategyNumber
Definition: stratnum.h:30
#define BTGreaterEqualStrategyNumber
Definition: stratnum.h:32

References Assert(), BTEqualStrategyNumber, BTGreaterEqualStrategyNumber, BTGreaterStrategyNumber, BTLessEqualStrategyNumber, BTLessStrategyNumber, CLAMP_PROBABILITY, DEFAULT_INEQ_SEL, examine_variable(), get_leftop(), get_op_opfamily_properties(), get_opfamily_member(), get_rightop(), get_variable_range(), GETSTRUCT, HeapTupleIsValid, is_opclause(), OidIsValid, ReleaseVariableStats, scalarineqsel(), and VariableStatData::statsTuple.

Referenced by cached_scansel().

◆ neqjoinsel()

Datum neqjoinsel ( PG_FUNCTION_ARGS  )

Definition at line 2819 of file selfuncs.c.

2820 {
2821  PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2822  Oid operator = PG_GETARG_OID(1);
2823  List *args = (List *) PG_GETARG_POINTER(2);
2824  JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2826  Oid collation = PG_GET_COLLATION();
2827  float8 result;
2828 
2829  if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2830  {
2831  /*
2832  * For semi-joins, if there is more than one distinct value in the RHS
2833  * relation then every non-null LHS row must find a row to join since
2834  * it can only be equal to one of them. We'll assume that there is
2835  * always more than one distinct RHS value for the sake of stability,
2836  * though in theory we could have special cases for empty RHS
2837  * (selectivity = 0) and single-distinct-value RHS (selectivity =
2838  * fraction of LHS that has the same value as the single RHS value).
2839  *
2840  * For anti-joins, if we use the same assumption that there is more
2841  * than one distinct key in the RHS relation, then every non-null LHS
2842  * row must be suppressed by the anti-join.
2843  *
2844  * So either way, the selectivity estimate should be 1 - nullfrac.
2845  */
2846  VariableStatData leftvar;
2847  VariableStatData rightvar;
2848  bool reversed;
2849  HeapTuple statsTuple;
2850  double nullfrac;
2851 
2852  get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2853  statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2854  if (HeapTupleIsValid(statsTuple))
2855  nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2856  else
2857  nullfrac = 0.0;
2858  ReleaseVariableStats(leftvar);
2859  ReleaseVariableStats(rightvar);
2860 
2861  result = 1.0 - nullfrac;
2862  }
2863  else
2864  {
2865  /*
2866  * We want 1 - eqjoinsel() where the equality operator is the one
2867  * associated with this != operator, that is, its negator.
2868  */
2869  Oid eqop = get_negator(operator);
2870 
2871  if (eqop)
2872  {
2873  result =
2875  collation,
2876  PointerGetDatum(root),
2877  ObjectIdGetDatum(eqop),
2879  Int16GetDatum(jointype),
2880  PointerGetDatum(sjinfo)));
2881  }
2882  else
2883  {
2884  /* Use default selectivity (should we raise an error instead?) */
2885  result = DEFAULT_EQ_SEL;
2886  }
2887  result = 1.0 - result;
2888  }
2889 
2890  PG_RETURN_FLOAT8(result);
2891 }
Datum DirectFunctionCall5Coll(PGFunction func, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4, Datum arg5)
Definition: fmgr.c:886
Datum eqjoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2269

References generate_unaccent_rules::args, DatumGetFloat8(), DEFAULT_EQ_SEL, DirectFunctionCall5Coll(), eqjoinsel(), get_join_variables(), get_negator(), GETSTRUCT, HeapTupleIsValid, Int16GetDatum(), JOIN_ANTI, JOIN_SEMI, ObjectIdGetDatum(), PG_GET_COLLATION, PG_GETARG_INT16, PG_GETARG_OID, PG_GETARG_POINTER, PG_RETURN_FLOAT8, PointerGetDatum(), ReleaseVariableStats, and VariableStatData::statsTuple.

◆ neqsel()

Datum neqsel ( PG_FUNCTION_ARGS  )

Definition at line 557 of file selfuncs.c.

558 {
559  PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
560 }

References eqsel_internal(), and PG_RETURN_FLOAT8.

◆ nulltestsel()

Selectivity nulltestsel ( PlannerInfo root,
NullTestType  nulltesttype,
Node arg,
int  varRelid,
JoinType  jointype,
SpecialJoinInfo sjinfo 
)

Definition at line 1698 of file selfuncs.c.

1700 {
1701  VariableStatData vardata;
1702  double selec;
1703 
1704  examine_variable(root, arg, varRelid, &vardata);
1705 
1706  if (HeapTupleIsValid(vardata.statsTuple))
1707  {
1708  Form_pg_statistic stats;
1709  double freq_null;
1710 
1711  stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1712  freq_null = stats->stanullfrac;
1713 
1714  switch (nulltesttype)
1715  {
1716  case IS_NULL:
1717 
1718  /*
1719  * Use freq_null directly.
1720  */
1721  selec = freq_null;
1722  break;
1723  case IS_NOT_NULL:
1724 
1725  /*
1726  * Select not unknown (not null) values. Calculate from
1727  * freq_null.
1728  */
1729  selec = 1.0 - freq_null;
1730  break;
1731  default:
1732  elog(ERROR, "unrecognized nulltesttype: %d",
1733  (int) nulltesttype);
1734  return (Selectivity) 0; /* keep compiler quiet */
1735  }
1736  }
1737  else if (vardata.var && IsA(vardata.var, Var) &&
1738  ((Var *) vardata.var)->varattno < 0)
1739  {
1740  /*
1741  * There are no stats for system columns, but we know they are never
1742  * NULL.
1743  */
1744  selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1745  }
1746  else
1747  {
1748  /*
1749  * No ANALYZE stats available, so make a guess
1750  */
1751  switch (nulltesttype)
1752  {
1753  case IS_NULL:
1754  selec = DEFAULT_UNK_SEL;
1755  break;
1756  case IS_NOT_NULL:
1757  selec = DEFAULT_NOT_UNK_SEL;
1758  break;
1759  default:
1760  elog(ERROR, "unrecognized nulltesttype: %d",
1761  (int) nulltesttype);
1762  return (Selectivity) 0; /* keep compiler quiet */
1763  }
1764  }
1765 
1766  ReleaseVariableStats(vardata);
1767 
1768  /* result should be in range, but make sure... */
1769  CLAMP_PROBABILITY(selec);
1770 
1771  return (Selectivity) selec;
1772 }
@ IS_NOT_NULL
Definition: primnodes.h:1715

References arg, CLAMP_PROBABILITY, DEFAULT_NOT_UNK_SEL, DEFAULT_UNK_SEL, elog, ERROR, examine_variable(), GETSTRUCT, HeapTupleIsValid, IS_NOT_NULL, IS_NULL, IsA, ReleaseVariableStats, VariableStatData::statsTuple, and VariableStatData::var.

Referenced by clause_selectivity_ext(), and clauselist_selectivity_ext().

◆ ReleaseDummy()

static void ReleaseDummy ( HeapTuple  tuple)
static

Definition at line 4971 of file selfuncs.c.

4972 {
4973  pfree(tuple);
4974 }

References pfree().

Referenced by examine_variable().

◆ rowcomparesel()

Selectivity rowcomparesel ( PlannerInfo root,
RowCompareExpr clause,
int  varRelid,
JoinType  jointype,
SpecialJoinInfo sjinfo 
)

Definition at line 2202 of file selfuncs.c.

2205 {
2206  Selectivity s1;
2207  Oid opno = linitial_oid(clause->opnos);
2208  Oid inputcollid = linitial_oid(clause->inputcollids);
2209  List *opargs;
2210  bool is_join_clause;
2211 
2212  /* Build equivalent arg list for single operator */
2213  opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2214 
2215  /*
2216  * Decide if it's a join clause. This should match clausesel.c's
2217  * treat_as_join_clause(), except that we intentionally consider only the
2218  * leading columns and not the rest of the clause.
2219  */
2220  if (varRelid != 0)
2221  {
2222  /*
2223  * Caller is forcing restriction mode (eg, because we are examining an
2224  * inner indexscan qual).
2225  */
2226  is_join_clause = false;
2227  }
2228  else if (sjinfo == NULL)
2229  {
2230  /*
2231  * It must be a restriction clause, since it's being evaluated at a
2232  * scan node.
2233  */
2234  is_join_clause = false;
2235  }
2236  else
2237  {
2238  /*
2239  * Otherwise, it's a join if there's more than one base relation used.
2240  */
2241  is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2242  }
2243 
2244  if (is_join_clause)
2245  {
2246  /* Estimate selectivity for a join clause. */
2247  s1 = join_selectivity(root, opno,
2248  opargs,
2249  inputcollid,
2250  jointype,
2251  sjinfo);
2252  }
2253  else
2254  {
2255  /* Estimate selectivity for a restriction clause. */
2256  s1 = restriction_selectivity(root, opno,
2257  opargs,
2258  inputcollid,
2259  varRelid);
2260  }
2261 
2262  return s1;
2263 }
int NumRelids(PlannerInfo *root, Node *clause)
Definition: clauses.c:2110
#define list_make2(x1, x2)
Definition: pg_list.h:214
Selectivity restriction_selectivity(PlannerInfo *root, Oid operatorid, List *args, Oid inputcollid, int varRelid)
Definition: plancat.c:1919
Selectivity join_selectivity(PlannerInfo *root, Oid operatorid, List *args, Oid inputcollid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: plancat.c:1958
char * s1

References join_selectivity(), RowCompareExpr::largs, linitial, linitial_oid, list_make2, NumRelids(), RowCompareExpr::rargs, restriction_selectivity(), and s1.

Referenced by clause_selectivity_ext().

◆ scalararraysel()

Selectivity scalararraysel ( PlannerInfo root,
ScalarArrayOpExpr clause,
bool  is_join_clause,
int  varRelid,
JoinType  jointype,
SpecialJoinInfo sjinfo 
)

Definition at line 1816 of file selfuncs.c.

1822 {
1823  Oid operator = clause->opno;
1824  bool useOr = clause->useOr;
1825  bool isEquality = false;
1826  bool isInequality = false;
1827  Node *leftop;
1828  Node *rightop;
1829  Oid nominal_element_type;
1830  Oid nominal_element_collation;
1831  TypeCacheEntry *typentry;
1832  RegProcedure oprsel;
1833  FmgrInfo oprselproc;
1834  Selectivity s1;
1835  Selectivity s1disjoint;
1836 
1837  /* First, deconstruct the expression */
1838  Assert(list_length(clause->args) == 2);
1839  leftop = (Node *) linitial(clause->args);
1840  rightop = (Node *) lsecond(clause->args);
1841 
1842  /* aggressively reduce both sides to constants */
1843  leftop = estimate_expression_value(root, leftop);
1844  rightop = estimate_expression_value(root, rightop);
1845 
1846  /* get nominal (after relabeling) element type of rightop */
1847  nominal_element_type = get_base_element_type(exprType(rightop));
1848  if (!OidIsValid(nominal_element_type))
1849  return (Selectivity) 0.5; /* probably shouldn't happen */
1850  /* get nominal collation, too, for generating constants */
1851  nominal_element_collation = exprCollation(rightop);
1852 
1853  /* look through any binary-compatible relabeling of rightop */
1854  rightop = strip_array_coercion(rightop);
1855 
1856  /*
1857  * Detect whether the operator is the default equality or inequality
1858  * operator of the array element type.
1859  */
1860  typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1861  if (OidIsValid(typentry->eq_opr))
1862  {
1863  if (operator == typentry->eq_opr)
1864  isEquality = true;
1865  else if (get_negator(operator) == typentry->eq_opr)
1866  isInequality = true;
1867  }
1868 
1869  /*
1870  * If it is equality or inequality, we might be able to estimate this as a
1871  * form of array containment; for instance "const = ANY(column)" can be
1872  * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1873  * that, and returns the selectivity estimate if successful, or -1 if not.
1874  */
1875  if ((isEquality || isInequality) && !is_join_clause)
1876  {
1877  s1 = scalararraysel_containment(root, leftop, rightop,
1878  nominal_element_type,
1879  isEquality, useOr, varRelid);
1880  if (s1 >= 0.0)
1881  return s1;
1882  }
1883 
1884  /*
1885  * Look up the underlying operator's selectivity estimator. Punt if it
1886  * hasn't got one.
1887  */
1888  if (is_join_clause)
1889  oprsel = get_oprjoin(operator);
1890  else
1891  oprsel = get_oprrest(operator);
1892  if (!oprsel)
1893  return (Selectivity) 0.5;
1894  fmgr_info(oprsel, &oprselproc);
1895 
1896  /*
1897  * In the array-containment check above, we must only believe that an
1898  * operator is equality or inequality if it is the default btree equality
1899  * operator (or its negator) for the element type, since those are the
1900  * operators that array containment will use. But in what follows, we can
1901  * be a little laxer, and also believe that any operators using eqsel() or
1902  * neqsel() as selectivity estimator act like equality or inequality.
1903  */
1904  if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1905  isEquality = true;
1906  else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1907  isInequality = true;
1908 
1909  /*
1910  * We consider three cases:
1911  *
1912  * 1. rightop is an Array constant: deconstruct the array, apply the
1913  * operator's selectivity function for each array element, and merge the
1914  * results in the same way that clausesel.c does for AND/OR combinations.
1915  *
1916  * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1917  * function for each element of the ARRAY[] construct, and merge.
1918  *
1919  * 3. otherwise, make a guess ...
1920  */
1921  if (rightop && IsA(rightop, Const))
1922  {
1923  Datum arraydatum = ((Const *) rightop)->constvalue;
1924  bool arrayisnull = ((Const *) rightop)->constisnull;
1925  ArrayType *arrayval;
1926  int16 elmlen;
1927  bool elmbyval;
1928  char elmalign;
1929  int num_elems;
1930  Datum *elem_values;
1931  bool *elem_nulls;
1932  int i;
1933 
1934  if (arrayisnull) /* qual can't succeed if null array */
1935  return (Selectivity) 0.0;
1936  arrayval = DatumGetArrayTypeP(arraydatum);
1938  &elmlen, &elmbyval, &elmalign);
1939  deconstruct_array(arrayval,
1940  ARR_ELEMTYPE(arrayval),
1941  elmlen, elmbyval, elmalign,
1942  &elem_values, &elem_nulls, &num_elems);
1943 
1944  /*
1945  * For generic operators, we assume the probability of success is
1946  * independent for each array element. But for "= ANY" or "<> ALL",
1947  * if the array elements are distinct (which'd typically be the case)
1948  * then the probabilities are disjoint, and we should just sum them.
1949  *
1950  * If we were being really tense we would try to confirm that the
1951  * elements are all distinct, but that would be expensive and it
1952  * doesn't seem to be worth the cycles; it would amount to penalizing
1953  * well-written queries in favor of poorly-written ones. However, we
1954  * do protect ourselves a little bit by checking whether the
1955  * disjointness assumption leads to an impossible (out of range)
1956  * probability; if so, we fall back to the normal calculation.
1957  */
1958  s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1959 
1960  for (i = 0; i < num_elems; i++)
1961  {
1962  List *args;
1963  Selectivity s2;
1964 
1965  args = list_make2(leftop,
1966  makeConst(nominal_element_type,
1967  -1,
1968  nominal_element_collation,
1969  elmlen,
1970  elem_values[i],
1971  elem_nulls[i],
1972  elmbyval));
1973  if (is_join_clause)
1974  s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1975  clause->inputcollid,
1976  PointerGetDatum(root),
1977  ObjectIdGetDatum(operator),
1979  Int16GetDatum(jointype),
1980  PointerGetDatum(sjinfo)));
1981  else
1982  s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1983  clause->inputcollid,
1984  PointerGetDatum(root),
1985  ObjectIdGetDatum(operator),
1987  Int32GetDatum(varRelid)));
1988 
1989  if (useOr)
1990  {
1991  s1 = s1 + s2 - s1 * s2;
1992  if (isEquality)
1993  s1disjoint += s2;
1994  }
1995  else
1996  {
1997  s1 = s1 * s2;
1998  if (isInequality)
1999  s1disjoint += s2 - 1.0;
2000  }
2001  }
2002 
2003  /* accept disjoint-probability estimate if in range */
2004  if ((useOr ? isEquality : isInequality) &&
2005  s1disjoint >= 0.0 && s1disjoint <= 1.0)
2006  s1 = s1disjoint;
2007  }
2008  else if (rightop && IsA(rightop, ArrayExpr) &&
2009  !((ArrayExpr *) rightop)->multidims)
2010  {
2011  ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2012  int16 elmlen;
2013  bool elmbyval;
2014  ListCell *l;
2015 
2016  get_typlenbyval(arrayexpr->element_typeid,
2017  &elmlen, &elmbyval);
2018 
2019  /*
2020  * We use the assumption of disjoint probabilities here too, although
2021  * the odds of equal array elements are rather higher if the elements
2022  * are not all constants (which they won't be, else constant folding
2023  * would have reduced the ArrayExpr to a Const). In this path it's
2024  * critical to have the sanity check on the s1disjoint estimate.
2025  */
2026  s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2027 
2028  foreach(l, arrayexpr->elements)
2029  {
2030  Node *elem = (Node *) lfirst(l);
2031  List *args;
2032  Selectivity s2;
2033 
2034  /*
2035  * Theoretically, if elem isn't of nominal_element_type we should
2036  * insert a RelabelType, but it seems unlikely that any operator
2037  * estimation function would really care ...
2038  */
2039  args = list_make2(leftop, elem);
2040  if (is_join_clause)
2041  s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2042  clause->inputcollid,
2043  PointerGetDatum(root),
2044  ObjectIdGetDatum(operator),
2046  Int16GetDatum(jointype),
2047  PointerGetDatum(sjinfo)));
2048  else
2049  s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2050  clause->inputcollid,
2051  PointerGetDatum(root),
2052  ObjectIdGetDatum(operator),
2054  Int32GetDatum(varRelid)));
2055 
2056  if (useOr)
2057  {
2058  s1 = s1 + s2 - s1 * s2;
2059  if (isEquality)
2060  s1disjoint += s2;
2061  }
2062  else
2063  {
2064  s1 = s1 * s2;
2065  if (isInequality)
2066  s1disjoint += s2 - 1.0;
2067  }
2068  }
2069 
2070  /* accept disjoint-probability estimate if in range */
2071  if ((useOr ? isEquality : isInequality) &&
2072  s1disjoint >= 0.0 && s1disjoint <= 1.0)
2073  s1 = s1disjoint;
2074  }
2075  else
2076  {
2077  CaseTestExpr *dummyexpr;
2078  List *args;
2079  Selectivity s2;
2080  int i;
2081 
2082  /*
2083  * We need a dummy rightop to pass to the operator selectivity
2084  * routine. It can be pretty much anything that doesn't look like a
2085  * constant; CaseTestExpr is a convenient choice.
2086  */
2087  dummyexpr = makeNode(CaseTestExpr);
2088  dummyexpr->typeId = nominal_element_type;
2089  dummyexpr->typeMod = -1;
2090  dummyexpr->collation = clause->inputcollid;
2091  args = list_make2(leftop, dummyexpr);
2092  if (is_join_clause)
2093  s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2094  clause->inputcollid,
2095  PointerGetDatum(root),
2096  ObjectIdGetDatum(operator),
2098  Int16GetDatum(jointype),
2099  PointerGetDatum(sjinfo)));
2100  else
2101  s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2102  clause->inputcollid,
2103  PointerGetDatum(root),
2104  ObjectIdGetDatum(operator),
2106  Int32GetDatum(varRelid)));
2107  s1 = useOr ? 0.0 : 1.0;
2108 
2109  /*
2110  * Arbitrarily assume 10 elements in the eventual array value (see
2111  * also estimate_array_length). We don't risk an assumption of
2112  * disjoint probabilities here.
2113  */
2114  for (i = 0; i < 10; i++)
2115  {
2116  if (useOr)
2117  s1 = s1 + s2 - s1 * s2;
2118  else
2119  s1 = s1 * s2;
2120  }
2121  }
2122 
2123  /* result should be in range, but make sure... */
2125 
2126  return s1;
2127 }
Selectivity scalararraysel_containment(PlannerInfo *root, Node *leftop, Node *rightop, Oid elemtype, bool isEquality, bool useOr, int varRelid)
regproc RegProcedure
Definition: c.h:637
Datum FunctionCall4Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4)
Definition: fmgr.c:1196
Datum FunctionCall5Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4, Datum arg5)
Definition: fmgr.c:1223
RegProcedure get_oprrest(Oid opno)
Definition: lsyscache.c:1535
RegProcedure get_oprjoin(Oid opno)
Definition: lsyscache.c:1559
Oid get_base_element_type(Oid typid)
Definition: lsyscache.c:2788
Const * makeConst(Oid consttype, int32 consttypmod, Oid constcollid, int constlen, Datum constvalue, bool constisnull, bool constbyval)
Definition: makefuncs.c:301
Oid exprCollation(const Node *expr)
Definition: nodeFuncs.c:788
#define makeNode(_type_)
Definition: nodes.h:155
static Datum Int32GetDatum(int32 X)
Definition: postgres.h:212
char * s2
List * elements
Definition: primnodes.h:1336
TypeCacheEntry * lookup_type_cache(Oid type_id, int flags)
Definition: typcache.c:346
#define TYPECACHE_EQ_OPR
Definition: typcache.h:137

References generate_unaccent_rules::args, ScalarArrayOpExpr::args, ARR_ELEMTYPE, Assert(), CLAMP_PROBABILITY, DatumGetArrayTypeP, DatumGetFloat8(), deconstruct_array(), ArrayExpr::elements, TypeCacheEntry::eq_opr, estimate_expression_value(), exprCollation(), exprType(), fmgr_info(), FunctionCall4Coll(), FunctionCall5Coll(), get_base_element_type(), get_negator(), get_oprjoin(), get_oprrest(), get_typlenbyval(), get_typlenbyvalalign(), i, Int16GetDatum(), Int32GetDatum(), IsA, lfirst, linitial, list_length(), list_make2, lookup_type_cache(), lsecond, makeConst(), makeNode, ObjectIdGetDatum(), OidIsValid, ScalarArrayOpExpr::opno, PointerGetDatum(), s1, s2, scalararraysel_containment(), strip_array_coercion(), TYPECACHE_EQ_OPR, CaseTestExpr::typeId, and ScalarArrayOpExpr::useOr.

Referenced by clause_selectivity_ext().

◆ scalargejoinsel()

Datum scalargejoinsel ( PG_FUNCTION_ARGS  )

Definition at line 2924 of file selfuncs.c.

2925 {
2927 }

References DEFAULT_INEQ_SEL, and PG_RETURN_FLOAT8.

◆ scalargesel()

Datum scalargesel ( PG_FUNCTION_ARGS  )

Definition at line 1498 of file selfuncs.c.

1499 {
1500  return scalarineqsel_wrapper(fcinfo, true, true);
1501 }
static Datum scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
Definition: selfuncs.c:1400

References scalarineqsel_wrapper().

◆ scalargtjoinsel()

Datum scalargtjoinsel ( PG_FUNCTION_ARGS  )

Definition at line 2915 of file selfuncs.c.

2916 {
2918 }

References DEFAULT_INEQ_SEL, and PG_RETURN_FLOAT8.

◆ scalargtsel()

Datum scalargtsel ( PG_FUNCTION_ARGS  )

Definition at line 1489 of file selfuncs.c.

1490 {
1491  return scalarineqsel_wrapper(fcinfo, true, false);
1492 }

References scalarineqsel_wrapper().

◆ scalarineqsel()

static double scalarineqsel ( PlannerInfo root,
Oid  operator,
bool  isgt,
bool  iseq,
Oid  collation,
VariableStatData vardata,
Datum  constval,
Oid  consttype 
)
static

Definition at line 580 of file selfuncs.c.

583 {
584  Form_pg_statistic stats;
585  FmgrInfo opproc;
586  double mcv_selec,
587  hist_selec,
588  sumcommon;
589  double selec;
590 
591  if (!HeapTupleIsValid(vardata->statsTuple))
592  {
593  /*
594  * No stats are available. Typically this means we have to fall back
595  * on the default estimate; but if the variable is CTID then we can
596  * make an estimate based on comparing the constant to the table size.
597  */
598  if (vardata->var && IsA(vardata->var, Var) &&
599  ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
600  {
601  ItemPointer itemptr;
602  double block;
603  double density;
604 
605  /*
606  * If the relation's empty, we're going to include all of it.
607  * (This is mostly to avoid divide-by-zero below.)
608  */
609  if (vardata->rel->pages == 0)
610  return 1.0;
611 
612  itemptr = (ItemPointer) DatumGetPointer(constval);
613  block = ItemPointerGetBlockNumberNoCheck(itemptr);
614 
615  /*
616  * Determine the average number of tuples per page (density).
617  *
618  * Since the last page will, on average, be only half full, we can
619  * estimate it to have half as many tuples as earlier pages. So
620  * give it half the weight of a regular page.
621  */
622  density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
623 
624  /* If target is the last page, use half the density. */
625  if (block >= vardata->rel->pages - 1)
626  density *= 0.5;
627 
628  /*
629  * Using the average tuples per page, calculate how far into the
630  * page the itemptr is likely to be and adjust block accordingly,
631  * by adding that fraction of a whole block (but never more than a
632  * whole block, no matter how high the itemptr's offset is). Here
633  * we are ignoring the possibility of dead-tuple line pointers,
634  * which is fairly bogus, but we lack the info to do better.
635  */
636  if (density > 0.0)
637  {
639 
640  block += Min(offset / density, 1.0);
641  }
642 
643  /*
644  * Convert relative block number to selectivity. Again, the last
645  * page has only half weight.
646  */
647  selec = block / (vardata->rel->pages - 0.5);
648 
649  /*
650  * The calculation so far gave us a selectivity for the "<=" case.
651  * We'll have one fewer tuple for "<" and one additional tuple for
652  * ">=", the latter of which we'll reverse the selectivity for
653  * below, so we can simply subtract one tuple for both cases. The
654  * cases that need this adjustment can be identified by iseq being
655  * equal to isgt.
656  */
657  if (iseq == isgt && vardata->rel->tuples >= 1.0)
658  selec -= (1.0 / vardata->rel->tuples);
659 
660  /* Finally, reverse the selectivity for the ">", ">=" cases. */
661  if (isgt)
662  selec = 1.0 - selec;
663 
664  CLAMP_PROBABILITY(selec);
665  return selec;
666  }
667 
668  /* no stats available, so default result */
669  return DEFAULT_INEQ_SEL;
670  }
671  stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
672 
673  fmgr_info(get_opcode(operator), &opproc);
674 
675  /*
676  * If we have most-common-values info, add up the fractions of the MCV
677  * entries that satisfy MCV OP CONST. These fractions contribute directly
678  * to the result selectivity. Also add up the total fraction represented
679  * by MCV entries.
680  */
681  mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
682  &sumcommon);
683 
684  /*
685  * If there is a histogram, determine which bin the constant falls in, and
686  * compute the resulting contribution to selectivity.
687  */
688  hist_selec = ineq_histogram_selectivity(root, vardata,
689  operator, &opproc, isgt, iseq,
690  collation,
691  constval, consttype);
692 
693  /*
694  * Now merge the results from the MCV and histogram calculations,
695  * realizing that the histogram covers only the non-null values that are
696  * not listed in MCV.
697  */
698  selec = 1.0 - stats->stanullfrac - sumcommon;
699 
700  if (hist_selec >= 0.0)
701  selec *= hist_selec;
702  else
703  {
704  /*
705  * If no histogram but there are values not accounted for by MCV,
706  * arbitrarily assume half of them will match.
707  */
708  selec *= 0.5;
709  }
710 
711  selec += mcv_selec;
712 
713  /* result should be in range, but make sure... */
714  CLAMP_PROBABILITY(selec);
715 
716  return selec;
717 }
static OffsetNumber ItemPointerGetOffsetNumberNoCheck(const ItemPointerData *pointer)
Definition: itemptr.h:114
static BlockNumber ItemPointerGetBlockNumberNoCheck(const ItemPointerData *pointer)
Definition: itemptr.h:93
ItemPointerData * ItemPointer
Definition: itemptr.h:49
uint16 OffsetNumber
Definition: off.h:24
double ineq_histogram_selectivity(PlannerInfo *root, VariableStatData *vardata, Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq, Oid collation, Datum constval, Oid consttype)
Definition: selfuncs.c:1041

References CLAMP_PROBABILITY, DatumGetPointer(), DEFAULT_INEQ_SEL, fmgr_info(), get_opcode(), GETSTRUCT, HeapTupleIsValid, ineq_histogram_selectivity(), IsA, ItemPointerGetBlockNumberNoCheck(), ItemPointerGetOffsetNumberNoCheck(), mcv_selectivity(), Min, RelOptInfo::pages, VariableStatData::rel, SelfItemPointerAttributeNumber, VariableStatData::statsTuple, RelOptInfo::tuples, and VariableStatData::var.

Referenced by mergejoinscansel(), and scalarineqsel_wrapper().

◆ scalarineqsel_wrapper()

static Datum scalarineqsel_wrapper ( PG_FUNCTION_ARGS  ,
bool  isgt,
bool  iseq 
)
static

Definition at line 1400 of file selfuncs.c.

1401 {
1402  PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1403  Oid operator = PG_GETARG_OID(1);
1404  List *args = (List *) PG_GETARG_POINTER(2);
1405  int varRelid = PG_GETARG_INT32(3);
1406  Oid collation = PG_GET_COLLATION();
1407  VariableStatData vardata;
1408  Node *other;
1409  bool varonleft;
1410  Datum constval;
1411  Oid consttype;
1412  double selec;
1413 
1414  /*
1415  * If expression is not variable op something or something op variable,
1416  * then punt and return a default estimate.
1417  */
1418  if (!get_restriction_variable(root, args, varRelid,
1419  &vardata, &other, &varonleft))
1421 
1422  /*
1423  * Can't do anything useful if the something is not a constant, either.
1424  */
1425  if (!IsA(other, Const))
1426  {
1427  ReleaseVariableStats(vardata);
1429  }
1430 
1431  /*
1432  * If the constant is NULL, assume operator is strict and return zero, ie,
1433  * operator will never return TRUE.
1434  */
1435  if (((Const *) other)->constisnull)
1436  {
1437  ReleaseVariableStats(vardata);
1438  PG_RETURN_FLOAT8(0.0);
1439  }
1440  constval = ((Const *) other)->constvalue;
1441  consttype = ((Const *) other)->consttype;
1442 
1443  /*
1444  * Force the var to be on the left to simplify logic in scalarineqsel.
1445  */
1446  if (!varonleft)
1447  {
1448  operator = get_commutator(operator);
1449  if (!operator)
1450  {
1451  /* Use default selectivity (should we raise an error instead?) */
1452  ReleaseVariableStats(vardata);
1454  }
1455  isgt = !isgt;
1456  }
1457 
1458  /* The rest of the work is done by scalarineqsel(). */
1459  selec = scalarineqsel(root, operator, isgt, iseq, collation,
1460  &vardata, constval, consttype);
1461 
1462  ReleaseVariableStats(vardata);
1463 
1464  PG_RETURN_FLOAT8((float8) selec);
1465 }

References generate_unaccent_rules::args, DEFAULT_INEQ_SEL, get_commutator(), get_restriction_variable(), IsA, PG_GET_COLLATION, PG_GETARG_INT32, PG_GETARG_OID, PG_GETARG_POINTER, PG_RETURN_FLOAT8, ReleaseVariableStats, and scalarineqsel().

Referenced by scalargesel(), scalargtsel(), scalarlesel(), and scalarltsel().

◆ scalarlejoinsel()

Datum scalarlejoinsel ( PG_FUNCTION_ARGS  )

Definition at line 2906 of file selfuncs.c.

2907 {
2909 }

References DEFAULT_INEQ_SEL, and PG_RETURN_FLOAT8.

◆ scalarlesel()

Datum scalarlesel ( PG_FUNCTION_ARGS  )

Definition at line 1480 of file selfuncs.c.

1481 {
1482  return scalarineqsel_wrapper(fcinfo, false, true);
1483 }

References scalarineqsel_wrapper().

◆ scalarltjoinsel()

Datum scalarltjoinsel ( PG_FUNCTION_ARGS  )

Definition at line 2897 of file selfuncs.c.

2898 {
2900 }

References DEFAULT_INEQ_SEL, and PG_RETURN_FLOAT8.

◆ scalarltsel()

Datum scalarltsel ( PG_FUNCTION_ARGS  )

Definition at line 1471 of file selfuncs.c.

1472 {
1473  return scalarineqsel_wrapper(fcinfo, false, false);
1474 }

References scalarineqsel_wrapper().

◆ spgcostestimate()

void spgcostestimate ( PlannerInfo root,
IndexPath path,
double  loop_count,
Cost indexStartupCost,
Cost indexTotalCost,
Selectivity indexSelectivity,
double *  indexCorrelation,
double *  indexPages 
)

Definition at line 7183 of file selfuncs.c.

7187 {
7188  IndexOptInfo *index = path->indexinfo;
7189  GenericCosts costs = {0};
7190  Cost descentCost;
7191 
7192  genericcostestimate(root, path, loop_count, &costs);
7193 
7194  /*
7195  * We model index descent costs similarly to those for btree, but to do
7196  * that we first need an idea of the tree height. We somewhat arbitrarily
7197  * assume that the fanout is 100, meaning the tree height is at most
7198  * log100(index->pages).
7199  *
7200  * Although this computation isn't really expensive enough to require
7201  * caching, we might as well use index->tree_height to cache it.
7202  */
7203  if (index->tree_height < 0) /* unknown? */
7204  {
7205  if (index->pages > 1) /* avoid computing log(0) */
7206  index->tree_height = (int) (log(index->pages) / log(100.0));
7207  else
7208  index->tree_height = 0;
7209  }
7210 
7211  /*
7212  * Add a CPU-cost component to represent the costs of initial descent. We
7213  * just use log(N) here not log2(N) since the branching factor isn't
7214  * necessarily two anyway. As for btree, charge once per SA scan.
7215  */
7216  if (index->tuples > 1) /* avoid computing log(0) */
7217  {
7218  descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7219  costs.indexStartupCost += descentCost;
7220  costs.indexTotalCost += costs.num_sa_scans * descentCost;
7221  }
7222 
7223  /*
7224  * Likewise add a per-page charge, calculated the same as for btrees.
7225  */
7226  descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7227  costs.indexStartupCost += descentCost;
7228  costs.indexTotalCost += costs.num_sa_scans * descentCost;
7229 
7230  *indexStartupCost = costs.indexStartupCost;
7231  *indexTotalCost = costs.indexTotalCost;
7232  *indexSelectivity = costs.indexSelectivity;
7233  *indexCorrelation = costs.indexCorrelation;
7234  *indexPages = costs.numIndexPages;
7235 }

References cpu_operator_cost, DEFAULT_PAGE_CPU_MULTIPLIER, genericcostestimate(), GenericCosts::indexCorrelation, IndexPath::indexinfo, GenericCosts::indexSelectivity, GenericCosts::indexStartupCost, GenericCosts::indexTotalCost, GenericCosts::num_sa_scans, and GenericCosts::numIndexPages.

Referenced by spghandler().

◆ statistic_proc_security_check()

bool statistic_proc_security_check ( VariableStatData vardata,
Oid  func_oid 
)

Definition at line 5735 of file selfuncs.c.

5736 {
5737  if (vardata->acl_ok)
5738  return true;
5739 
5740  if (!OidIsValid(func_oid))
5741  return false;
5742 
5743  if (get_func_leakproof(func_oid))
5744  return true;
5745 
5746  ereport(DEBUG2,
5747  (errmsg_internal("not using statistics because function \"%s\" is not leak-proof",
5748  get_func_name(func_oid))));
5749  return false;
5750 }
int errmsg_internal(const char *fmt,...)
Definition: elog.c:1159
#define DEBUG2
Definition: elog.h:29
#define ereport(elevel,...)
Definition: elog.h:149
char * get_func_name(Oid funcid)
Definition: lsyscache.c:1586
bool get_func_leakproof(Oid funcid)
Definition: lsyscache.c:1815

References VariableStatData::acl_ok, DEBUG2, ereport, errmsg_internal(), get_func_leakproof(), get_func_name(), and OidIsValid.

Referenced by calc_arraycontsel(), calc_hist_selectivity(), eqjoinsel(), get_variable_range(), histogram_selectivity(), ineq_histogram_selectivity(), mcv_selectivity(), scalararraysel_containment(), and var_eq_const().

◆ strip_array_coercion()

static Node* strip_array_coercion ( Node node)
static

Definition at line 1783 of file selfuncs.c.

1784 {
1785  for (;;)
1786  {
1787  if (node && IsA(node, ArrayCoerceExpr))
1788  {
1789  ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1790 
1791  /*
1792  * If the per-element expression is just a RelabelType on top of
1793  * CaseTestExpr, then we know it's a binary-compatible relabeling.
1794  */
1795  if (IsA(acoerce->elemexpr, RelabelType) &&
1796  IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1797  node = (Node *) acoerce->arg;
1798  else
1799  break;
1800  }
1801  else if (node && IsA(node, RelabelType))
1802  {
1803  /* We don't really expect this case, but may as well cope */
1804  node = (Node *) ((RelabelType *) node)->arg;
1805  }
1806  else
1807  break;
1808  }
1809  return node;
1810 }

References arg, ArrayCoerceExpr::arg, ArrayCoerceExpr::elemexpr, and IsA.

Referenced by estimate_array_length(), and scalararraysel().

◆ var_eq_const()

double var_eq_const ( VariableStatData vardata,
Oid  oproid,
Oid  collation,
Datum  constval,
bool  constisnull,
bool  varonleft,
bool  negate 
)

Definition at line 295 of file selfuncs.c.

298 {
299  double selec;
300  double nullfrac = 0.0;
301  bool isdefault;
302  Oid opfuncoid;
303 
304  /*
305  * If the constant is NULL, assume operator is strict and return zero, ie,
306  * operator will never return TRUE. (It's zero even for a negator op.)
307  */
308  if (constisnull)
309  return 0.0;
310 
311  /*
312  * Grab the nullfrac for use below. Note we allow use of nullfrac
313  * regardless of security check.
314  */
315  if (HeapTupleIsValid(vardata->statsTuple))
316  {
317  Form_pg_statistic stats;
318 
319  stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
320  nullfrac = stats->stanullfrac;
321  }
322 
323  /*
324  * If we matched the var to a unique index or DISTINCT clause, assume
325  * there is exactly one match regardless of anything else. (This is
326  * slightly bogus, since the index or clause's equality operator might be
327  * different from ours, but it's much more likely to be right than
328  * ignoring the information.)
329  */
330  if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
331  {
332  selec = 1.0 / vardata->rel->tuples;
333  }
334  else if (HeapTupleIsValid(vardata->statsTuple) &&
336  (opfuncoid = get_opcode(oproid))))
337  {
338  AttStatsSlot sslot;
339  bool match = false;
340  int i;
341 
342  /*
343  * Is the constant "=" to any of the column's most common values?
344  * (Although the given operator may not really be "=", we will assume
345  * that seeing whether it returns TRUE is an appropriate test. If you
346  * don't like this, maybe you shouldn't be using eqsel for your
347  * operator...)
348  */
349  if (get_attstatsslot(&sslot, vardata->statsTuple,
350  STATISTIC_KIND_MCV, InvalidOid,
352  {
353  LOCAL_FCINFO(fcinfo, 2);
354  FmgrInfo eqproc;
355 
356  fmgr_info(opfuncoid, &eqproc);
357 
358  /*
359  * Save a few cycles by setting up the fcinfo struct just once.
360  * Using FunctionCallInvoke directly also avoids failure if the
361  * eqproc returns NULL, though really equality functions should
362  * never do that.
363  */
364  InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
365  NULL, NULL);
366  fcinfo->args[0].isnull = false;
367  fcinfo->args[1].isnull = false;
368  /* be careful to apply operator right way 'round */
369  if (varonleft)
370  fcinfo->args[1].value = constval;
371  else
372  fcinfo->args[0].value = constval;
373 
374  for (i = 0; i < sslot.nvalues; i++)
375  {
376  Datum fresult;
377 
378  if (varonleft)
379  fcinfo->args[0].value = sslot.values[i];
380  else
381  fcinfo->args[1].value = sslot.values[i];
382  fcinfo->isnull = false;
383  fresult = FunctionCallInvoke(fcinfo);
384  if (!fcinfo->isnull && DatumGetBool(fresult))
385  {
386  match = true;
387  break;
388  }
389  }
390  }
391  else
392  {
393  /* no most-common-value info available */
394  i = 0; /* keep compiler quiet */
395  }
396 
397  if (match)
398  {
399  /*
400  * Constant is "=" to this common value. We know selectivity
401  * exactly (or as exactly as ANALYZE could calculate it, anyway).
402  */
403  selec = sslot.numbers[i];
404  }
405  else
406  {
407  /*
408  * Comparison is against a constant that is neither NULL nor any
409  * of the common values. Its selectivity cannot be more than
410  * this:
411  */
412  double sumcommon = 0.0;
413  double otherdistinct;
414 
415  for (i = 0; i < sslot.nnumbers; i++)
416  sumcommon += sslot.numbers[i];
417  selec = 1.0 - sumcommon - nullfrac;
418  CLAMP_PROBABILITY(selec);
419 
420  /*
421  * and in fact it's probably a good deal less. We approximate that
422  * all the not-common values share this remaining fraction
423  * equally, so we divide by the number of other distinct values.
424  */
425  otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
426  sslot.nnumbers;
427  if (otherdistinct > 1)
428  selec /= otherdistinct;
429 
430  /*
431  * Another cross-check: selectivity shouldn't be estimated as more
432  * than the least common "most common value".
433  */
434  if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
435  selec = sslot.numbers[sslot.nnumbers - 1];
436  }
437 
438  free_attstatsslot(&sslot);
439  }
440  else
441  {
442  /*
443  * No ANALYZE stats available, so make a guess using estimated number
444  * of distinct values and assuming they are equally common. (The guess
445  * is unlikely to be very good, but we do know a few special cases.)
446  */
447  selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
448  }
449 
450  /* now adjust if we wanted <> rather than = */
451  if (negate)
452  selec = 1.0 - selec - nullfrac;
453 
454  /* result should be in range, but make sure... */
455  CLAMP_PROBABILITY(selec);
456 
457  return selec;
458 }

References ATTSTATSSLOT_NUMBERS, ATTSTATSSLOT_VALUES, CLAMP_PROBABILITY, DatumGetBool(), fmgr_info(), free_attstatsslot(), FunctionCallInvoke, get_attstatsslot(), get_opcode(), get_variable_numdistinct(), GETSTRUCT, HeapTupleIsValid, i, InitFunctionCallInfoData, InvalidOid, VariableStatData::isunique, LOCAL_FCINFO, AttStatsSlot::nnumbers, AttStatsSlot::numbers, AttStatsSlot::nvalues, VariableStatData::rel, statistic_proc_security_check(), VariableStatData::statsTuple, RelOptInfo::tuples, and AttStatsSlot::values.

Referenced by boolvarsel(), eqsel_internal(), patternsel_common(), and prefix_selectivity().

◆ var_eq_non_const()

double var_eq_non_const ( VariableStatData vardata,
Oid  oproid,
Oid  collation,
Node other,
bool  varonleft,
bool  negate 
)

Definition at line 466 of file selfuncs.c.

469 {
470  double selec;
471  double nullfrac = 0.0;
472  bool isdefault;
473 
474  /*
475  * Grab the nullfrac for use below.
476  */
477  if (HeapTupleIsValid(vardata->statsTuple))
478  {
479  Form_pg_statistic stats;
480 
481  stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
482  nullfrac = stats->stanullfrac;
483  }
484 
485  /*
486  * If we matched the var to a unique index or DISTINCT clause, assume
487  * there is exactly one match regardless of anything else. (This is
488  * slightly bogus, since the index or clause's equality operator might be
489  * different from ours, but it's much more likely to be right than
490  * ignoring the information.)
491  */
492  if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
493  {
494  selec = 1.0 / vardata->rel->tuples;
495  }
496  else if (HeapTupleIsValid(vardata->statsTuple))
497  {
498  double ndistinct;
499  AttStatsSlot sslot;
500 
501  /*
502  * Search is for a value that we do not know a priori, but we will
503  * assume it is not NULL. Estimate the selectivity as non-null
504  * fraction divided by number of distinct values, so that we get a
505  * result averaged over all possible values whether common or
506  * uncommon. (Essentially, we are assuming that the not-yet-known
507  * comparison value is equally likely to be any of the possible
508  * values, regardless of their frequency in the table. Is that a good
509  * idea?)
510  */
511  selec = 1.0 - nullfrac;
512  ndistinct = get_variable_numdistinct(vardata, &isdefault);
513  if (ndistinct > 1)
514  selec /= ndistinct;
515 
516  /*
517  * Cross-check: selectivity should never be estimated as more than the
518  * most common value's.
519  */
520  if (get_attstatsslot(&sslot, vardata->statsTuple,
521  STATISTIC_KIND_MCV, InvalidOid,
523  {
524  if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
525  selec = sslot.numbers[0];
526  free_attstatsslot(&sslot);
527  }
528  }
529  else
530  {
531  /*
532  * No ANALYZE stats available, so make a guess using estimated number
533  * of distinct values and assuming they are equally common. (The guess
534  * is unlikely to be very good, but we do know a few special cases.)
535  */
536  selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
537  }
538 
539  /* now adjust if we wanted <> rather than = */
540  if (negate)
541  selec = 1.0 - selec - nullfrac;
542 
543  /* result should be in range, but make sure... */
544  CLAMP_PROBABILITY(selec);
545 
546  return selec;
547 }

References ATTSTATSSLOT_NUMBERS, CLAMP_PROBABILITY, free_attstatsslot(), get_attstatsslot(), get_variable_numdistinct(), GETSTRUCT, HeapTupleIsValid, InvalidOid, VariableStatData::isunique, AttStatsSlot::nnumbers, AttStatsSlot::numbers, VariableStatData::rel, VariableStatData::statsTuple, and RelOptInfo::tuples.

Referenced by eqsel_internal().

Variable Documentation

◆ get_index_stats_hook

get_index_stats_hook_type get_index_stats_hook = NULL

Definition at line 148 of file selfuncs.c.

Referenced by brincostestimate(), btcostestimate(), and examine_variable().

◆ get_relation_stats_hook

get_relation_stats_hook_type get_relation_stats_hook = NULL

Definition at line 147 of file selfuncs.c.

Referenced by brincostestimate(), btcostestimate(), and examine_simple_variable().