PostgreSQL Source Code  git master
cost.h File Reference
#include "nodes/plannodes.h"
#include "nodes/relation.h"
Include dependency graph for cost.h:
This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Macros

#define DEFAULT_SEQ_PAGE_COST   1.0
 
#define DEFAULT_RANDOM_PAGE_COST   4.0
 
#define DEFAULT_CPU_TUPLE_COST   0.01
 
#define DEFAULT_CPU_INDEX_TUPLE_COST   0.005
 
#define DEFAULT_CPU_OPERATOR_COST   0.0025
 
#define DEFAULT_PARALLEL_TUPLE_COST   0.1
 
#define DEFAULT_PARALLEL_SETUP_COST   1000.0
 
#define DEFAULT_EFFECTIVE_CACHE_SIZE   524288 /* measured in pages */
 

Enumerations

enum  ConstraintExclusionType { CONSTRAINT_EXCLUSION_OFF, CONSTRAINT_EXCLUSION_ON, CONSTRAINT_EXCLUSION_PARTITION }
 

Functions

double clamp_row_est (double nrows)
 
double index_pages_fetched (double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
 
void cost_seqscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_samplescan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_index (IndexPath *path, PlannerInfo *root, double loop_count, bool partial_path)
 
void cost_bitmap_heap_scan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, Path *bitmapqual, double loop_count)
 
void cost_bitmap_and_node (BitmapAndPath *path, PlannerInfo *root)
 
void cost_bitmap_or_node (BitmapOrPath *path, PlannerInfo *root)
 
void cost_bitmap_tree_node (Path *path, Cost *cost, Selectivity *selec)
 
void cost_tidscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
 
void cost_subqueryscan (SubqueryScanPath *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_functionscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_tableexprscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_valuesscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_tablefuncscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_ctescan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_namedtuplestorescan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_recursive_union (Path *runion, Path *nrterm, Path *rterm)
 
void cost_sort (Path *path, PlannerInfo *root, List *pathkeys, Cost input_cost, double tuples, int width, Cost comparison_cost, int sort_mem, double limit_tuples)
 
void cost_merge_append (Path *path, PlannerInfo *root, List *pathkeys, int n_streams, Cost input_startup_cost, Cost input_total_cost, double tuples)
 
void cost_material (Path *path, Cost input_startup_cost, Cost input_total_cost, double tuples, int width)
 
void cost_agg (Path *path, PlannerInfo *root, AggStrategy aggstrategy, const AggClauseCosts *aggcosts, int numGroupCols, double numGroups, List *quals, Cost input_startup_cost, Cost input_total_cost, double input_tuples)
 
void cost_windowagg (Path *path, PlannerInfo *root, List *windowFuncs, int numPartCols, int numOrderCols, Cost input_startup_cost, Cost input_total_cost, double input_tuples)
 
void cost_group (Path *path, PlannerInfo *root, int numGroupCols, double numGroups, List *quals, Cost input_startup_cost, Cost input_total_cost, double input_tuples)
 
void initial_cost_nestloop (PlannerInfo *root, JoinCostWorkspace *workspace, JoinType jointype, Path *outer_path, Path *inner_path, JoinPathExtraData *extra)
 
void final_cost_nestloop (PlannerInfo *root, NestPath *path, JoinCostWorkspace *workspace, JoinPathExtraData *extra)
 
void initial_cost_mergejoin (PlannerInfo *root, JoinCostWorkspace *workspace, JoinType jointype, List *mergeclauses, Path *outer_path, Path *inner_path, List *outersortkeys, List *innersortkeys, JoinPathExtraData *extra)
 
void final_cost_mergejoin (PlannerInfo *root, MergePath *path, JoinCostWorkspace *workspace, JoinPathExtraData *extra)
 
void initial_cost_hashjoin (PlannerInfo *root, JoinCostWorkspace *workspace, JoinType jointype, List *hashclauses, Path *outer_path, Path *inner_path, JoinPathExtraData *extra)
 
void final_cost_hashjoin (PlannerInfo *root, HashPath *path, JoinCostWorkspace *workspace, JoinPathExtraData *extra)
 
void cost_gather (GatherPath *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, double *rows)
 
void cost_subplan (PlannerInfo *root, SubPlan *subplan, Plan *plan)
 
void cost_qual_eval (QualCost *cost, List *quals, PlannerInfo *root)
 
void cost_qual_eval_node (QualCost *cost, Node *qual, PlannerInfo *root)
 
void compute_semi_anti_join_factors (PlannerInfo *root, RelOptInfo *outerrel, RelOptInfo *innerrel, JoinType jointype, SpecialJoinInfo *sjinfo, List *restrictlist, SemiAntiJoinFactors *semifactors)
 
void set_baserel_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
double get_parameterized_baserel_size (PlannerInfo *root, RelOptInfo *rel, List *param_clauses)
 
double get_parameterized_joinrel_size (PlannerInfo *root, RelOptInfo *rel, Path *outer_path, Path *inner_path, SpecialJoinInfo *sjinfo, List *restrict_clauses)
 
void set_joinrel_size_estimates (PlannerInfo *root, RelOptInfo *rel, RelOptInfo *outer_rel, RelOptInfo *inner_rel, SpecialJoinInfo *sjinfo, List *restrictlist)
 
void set_subquery_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_function_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_values_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_cte_size_estimates (PlannerInfo *root, RelOptInfo *rel, double cte_rows)
 
void set_tablefunc_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_namedtuplestore_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_foreign_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
PathTargetset_pathtarget_cost_width (PlannerInfo *root, PathTarget *target)
 
double compute_bitmap_pages (PlannerInfo *root, RelOptInfo *baserel, Path *bitmapqual, int loop_count, Cost *cost, double *tuple)
 
Selectivity clauselist_selectivity (PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
 
Selectivity clause_selectivity (PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
 
void cost_gather_merge (GatherMergePath *path, PlannerInfo *root, RelOptInfo *rel, ParamPathInfo *param_info, Cost input_startup_cost, Cost input_total_cost, double *rows)
 

Variables

PGDLLIMPORT double seq_page_cost
 
PGDLLIMPORT double random_page_cost
 
PGDLLIMPORT double cpu_tuple_cost
 
PGDLLIMPORT double cpu_index_tuple_cost
 
PGDLLIMPORT double cpu_operator_cost
 
PGDLLIMPORT double parallel_tuple_cost
 
PGDLLIMPORT double parallel_setup_cost
 
PGDLLIMPORT int effective_cache_size
 
Cost disable_cost
 
int max_parallel_workers_per_gather
 
bool enable_seqscan
 
bool enable_indexscan
 
bool enable_indexonlyscan
 
bool enable_bitmapscan
 
bool enable_tidscan
 
bool enable_sort
 
bool enable_hashagg
 
bool enable_nestloop
 
bool enable_material
 
bool enable_mergejoin
 
bool enable_hashjoin
 
bool enable_gathermerge
 
bool enable_partition_wise_join
 
int constraint_exclusion
 

Macro Definition Documentation

◆ DEFAULT_CPU_INDEX_TUPLE_COST

#define DEFAULT_CPU_INDEX_TUPLE_COST   0.005

Definition at line 27 of file cost.h.

◆ DEFAULT_CPU_OPERATOR_COST

#define DEFAULT_CPU_OPERATOR_COST   0.0025

Definition at line 28 of file cost.h.

◆ DEFAULT_CPU_TUPLE_COST

#define DEFAULT_CPU_TUPLE_COST   0.01

Definition at line 26 of file cost.h.

◆ DEFAULT_EFFECTIVE_CACHE_SIZE

#define DEFAULT_EFFECTIVE_CACHE_SIZE   524288 /* measured in pages */

Definition at line 32 of file cost.h.

◆ DEFAULT_PARALLEL_SETUP_COST

#define DEFAULT_PARALLEL_SETUP_COST   1000.0

Definition at line 30 of file cost.h.

◆ DEFAULT_PARALLEL_TUPLE_COST

#define DEFAULT_PARALLEL_TUPLE_COST   0.1

Definition at line 29 of file cost.h.

◆ DEFAULT_RANDOM_PAGE_COST

#define DEFAULT_RANDOM_PAGE_COST   4.0

Definition at line 25 of file cost.h.

◆ DEFAULT_SEQ_PAGE_COST

#define DEFAULT_SEQ_PAGE_COST   1.0

Definition at line 24 of file cost.h.

Enumeration Type Documentation

◆ ConstraintExclusionType

Enumerator
CONSTRAINT_EXCLUSION_OFF 
CONSTRAINT_EXCLUSION_ON 
CONSTRAINT_EXCLUSION_PARTITION 

Definition at line 34 of file cost.h.

35 {
36  CONSTRAINT_EXCLUSION_OFF, /* do not use c_e */
37  CONSTRAINT_EXCLUSION_ON, /* apply c_e to all rels */
38  CONSTRAINT_EXCLUSION_PARTITION /* apply c_e to otherrels only */
ConstraintExclusionType
Definition: cost.h:34

Function Documentation

◆ clamp_row_est()

double clamp_row_est ( double  nrows)

Definition at line 174 of file costsize.c.

References rint().

Referenced by approx_tuple_count(), bernoulli_samplescangetsamplesize(), calc_joinrel_size_estimate(), compute_bitmap_pages(), cost_agg(), cost_bitmap_heap_scan(), cost_group(), cost_index(), cost_seqscan(), cost_subplan(), create_bitmap_subplan(), create_limit_path(), estimate_hash_bucket_stats(), estimate_num_groups(), estimate_path_cost_size(), estimate_size(), expression_returns_set_rows(), final_cost_hashjoin(), final_cost_mergejoin(), final_cost_nestloop(), get_parameterized_baserel_size(), get_variable_numdistinct(), initial_cost_mergejoin(), set_baserel_size_estimates(), set_foreign_size(), system_rows_samplescangetsamplesize(), system_samplescangetsamplesize(), and system_time_samplescangetsamplesize().

175 {
176  /*
177  * Force estimate to be at least one row, to make explain output look
178  * better and to avoid possible divide-by-zero when interpolating costs.
179  * Make it an integer, too.
180  */
181  if (nrows <= 1.0)
182  nrows = 1.0;
183  else
184  nrows = rint(nrows);
185 
186  return nrows;
187 }
double rint(double x)
Definition: rint.c:22

◆ clause_selectivity()

Selectivity clause_selectivity ( PlannerInfo root,
Node clause,
int  varRelid,
JoinType  jointype,
SpecialJoinInfo sjinfo 
)

Definition at line 574 of file clausesel.c.

References and_clause(), arg, generate_unaccent_rules::args, OpExpr::args, bms_is_subset_singleton(), booltestsel(), boolvarsel(), RestrictInfo::clause, RestrictInfo::clause_relids, clause_selectivity(), clauselist_selectivity(), Const::constisnull, Const::constvalue, CurrentOfExpr::cvarno, DatumGetBool, DEBUG4, elog, estimate_expression_value(), find_base_rel(), get_notclausearg(), OpExpr::inputcollid, is_opclause, IsA, JOIN_INNER, join_selectivity(), lfirst, RestrictInfo::norm_selec, not_clause(), nulltestsel(), OpExpr::opno, or_clause(), RestrictInfo::orclause, RestrictInfo::outer_selec, RestrictInfo::pseudoconstant, restriction_selectivity(), rowcomparesel(), s1, s2, scalararraysel(), treat_as_join_clause(), RelOptInfo::tuples, RangeQueryClause::var, Var::varlevelsup, and Var::varno.

Referenced by approx_tuple_count(), booltestsel(), clause_selectivity(), clauselist_selectivity(), consider_new_or_clause(), and dependencies_clauselist_selectivity().

579 {
580  Selectivity s1 = 0.5; /* default for any unhandled clause type */
581  RestrictInfo *rinfo = NULL;
582  bool cacheable = false;
583 
584  if (clause == NULL) /* can this still happen? */
585  return s1;
586 
587  if (IsA(clause, RestrictInfo))
588  {
589  rinfo = (RestrictInfo *) clause;
590 
591  /*
592  * If the clause is marked pseudoconstant, then it will be used as a
593  * gating qual and should not affect selectivity estimates; hence
594  * return 1.0. The only exception is that a constant FALSE may be
595  * taken as having selectivity 0.0, since it will surely mean no rows
596  * out of the plan. This case is simple enough that we need not
597  * bother caching the result.
598  */
599  if (rinfo->pseudoconstant)
600  {
601  if (!IsA(rinfo->clause, Const))
602  return (Selectivity) 1.0;
603  }
604 
605  /*
606  * If the clause is marked redundant, always return 1.0.
607  */
608  if (rinfo->norm_selec > 1)
609  return (Selectivity) 1.0;
610 
611  /*
612  * If possible, cache the result of the selectivity calculation for
613  * the clause. We can cache if varRelid is zero or the clause
614  * contains only vars of that relid --- otherwise varRelid will affect
615  * the result, so mustn't cache. Outer join quals might be examined
616  * with either their join's actual jointype or JOIN_INNER, so we need
617  * two cache variables to remember both cases. Note: we assume the
618  * result won't change if we are switching the input relations or
619  * considering a unique-ified case, so we only need one cache variable
620  * for all non-JOIN_INNER cases.
621  */
622  if (varRelid == 0 ||
623  bms_is_subset_singleton(rinfo->clause_relids, varRelid))
624  {
625  /* Cacheable --- do we already have the result? */
626  if (jointype == JOIN_INNER)
627  {
628  if (rinfo->norm_selec >= 0)
629  return rinfo->norm_selec;
630  }
631  else
632  {
633  if (rinfo->outer_selec >= 0)
634  return rinfo->outer_selec;
635  }
636  cacheable = true;
637  }
638 
639  /*
640  * Proceed with examination of contained clause. If the clause is an
641  * OR-clause, we want to look at the variant with sub-RestrictInfos,
642  * so that per-subclause selectivities can be cached.
643  */
644  if (rinfo->orclause)
645  clause = (Node *) rinfo->orclause;
646  else
647  clause = (Node *) rinfo->clause;
648  }
649 
650  if (IsA(clause, Var))
651  {
652  Var *var = (Var *) clause;
653 
654  /*
655  * We probably shouldn't ever see an uplevel Var here, but if we do,
656  * return the default selectivity...
657  */
658  if (var->varlevelsup == 0 &&
659  (varRelid == 0 || varRelid == (int) var->varno))
660  {
661  /* Use the restriction selectivity function for a bool Var */
662  s1 = boolvarsel(root, (Node *) var, varRelid);
663  }
664  }
665  else if (IsA(clause, Const))
666  {
667  /* bool constant is pretty easy... */
668  Const *con = (Const *) clause;
669 
670  s1 = con->constisnull ? 0.0 :
671  DatumGetBool(con->constvalue) ? 1.0 : 0.0;
672  }
673  else if (IsA(clause, Param))
674  {
675  /* see if we can replace the Param */
676  Node *subst = estimate_expression_value(root, clause);
677 
678  if (IsA(subst, Const))
679  {
680  /* bool constant is pretty easy... */
681  Const *con = (Const *) subst;
682 
683  s1 = con->constisnull ? 0.0 :
684  DatumGetBool(con->constvalue) ? 1.0 : 0.0;
685  }
686  else
687  {
688  /* XXX any way to do better than default? */
689  }
690  }
691  else if (not_clause(clause))
692  {
693  /* inverse of the selectivity of the underlying clause */
694  s1 = 1.0 - clause_selectivity(root,
695  (Node *) get_notclausearg((Expr *) clause),
696  varRelid,
697  jointype,
698  sjinfo);
699  }
700  else if (and_clause(clause))
701  {
702  /* share code with clauselist_selectivity() */
703  s1 = clauselist_selectivity(root,
704  ((BoolExpr *) clause)->args,
705  varRelid,
706  jointype,
707  sjinfo);
708  }
709  else if (or_clause(clause))
710  {
711  /*
712  * Selectivities for an OR clause are computed as s1+s2 - s1*s2 to
713  * account for the probable overlap of selected tuple sets.
714  *
715  * XXX is this too conservative?
716  */
717  ListCell *arg;
718 
719  s1 = 0.0;
720  foreach(arg, ((BoolExpr *) clause)->args)
721  {
723  (Node *) lfirst(arg),
724  varRelid,
725  jointype,
726  sjinfo);
727 
728  s1 = s1 + s2 - s1 * s2;
729  }
730  }
731  else if (is_opclause(clause) || IsA(clause, DistinctExpr))
732  {
733  OpExpr *opclause = (OpExpr *) clause;
734  Oid opno = opclause->opno;
735 
736  if (treat_as_join_clause(clause, rinfo, varRelid, sjinfo))
737  {
738  /* Estimate selectivity for a join clause. */
739  s1 = join_selectivity(root, opno,
740  opclause->args,
741  opclause->inputcollid,
742  jointype,
743  sjinfo);
744  }
745  else
746  {
747  /* Estimate selectivity for a restriction clause. */
748  s1 = restriction_selectivity(root, opno,
749  opclause->args,
750  opclause->inputcollid,
751  varRelid);
752  }
753 
754  /*
755  * DistinctExpr has the same representation as OpExpr, but the
756  * contained operator is "=" not "<>", so we must negate the result.
757  * This estimation method doesn't give the right behavior for nulls,
758  * but it's better than doing nothing.
759  */
760  if (IsA(clause, DistinctExpr))
761  s1 = 1.0 - s1;
762  }
763  else if (IsA(clause, ScalarArrayOpExpr))
764  {
765  /* Use node specific selectivity calculation function */
766  s1 = scalararraysel(root,
767  (ScalarArrayOpExpr *) clause,
768  treat_as_join_clause(clause, rinfo,
769  varRelid, sjinfo),
770  varRelid,
771  jointype,
772  sjinfo);
773  }
774  else if (IsA(clause, RowCompareExpr))
775  {
776  /* Use node specific selectivity calculation function */
777  s1 = rowcomparesel(root,
778  (RowCompareExpr *) clause,
779  varRelid,
780  jointype,
781  sjinfo);
782  }
783  else if (IsA(clause, NullTest))
784  {
785  /* Use node specific selectivity calculation function */
786  s1 = nulltestsel(root,
787  ((NullTest *) clause)->nulltesttype,
788  (Node *) ((NullTest *) clause)->arg,
789  varRelid,
790  jointype,
791  sjinfo);
792  }
793  else if (IsA(clause, BooleanTest))
794  {
795  /* Use node specific selectivity calculation function */
796  s1 = booltestsel(root,
797  ((BooleanTest *) clause)->booltesttype,
798  (Node *) ((BooleanTest *) clause)->arg,
799  varRelid,
800  jointype,
801  sjinfo);
802  }
803  else if (IsA(clause, CurrentOfExpr))
804  {
805  /* CURRENT OF selects at most one row of its table */
806  CurrentOfExpr *cexpr = (CurrentOfExpr *) clause;
807  RelOptInfo *crel = find_base_rel(root, cexpr->cvarno);
808 
809  if (crel->tuples > 0)
810  s1 = 1.0 / crel->tuples;
811  }
812  else if (IsA(clause, RelabelType))
813  {
814  /* Not sure this case is needed, but it can't hurt */
815  s1 = clause_selectivity(root,
816  (Node *) ((RelabelType *) clause)->arg,
817  varRelid,
818  jointype,
819  sjinfo);
820  }
821  else if (IsA(clause, CoerceToDomain))
822  {
823  /* Not sure this case is needed, but it can't hurt */
824  s1 = clause_selectivity(root,
825  (Node *) ((CoerceToDomain *) clause)->arg,
826  varRelid,
827  jointype,
828  sjinfo);
829  }
830  else
831  {
832  /*
833  * For anything else, see if we can consider it as a boolean variable.
834  * This only works if it's an immutable expression in Vars of a single
835  * relation; but there's no point in us checking that here because
836  * boolvarsel() will do it internally, and return a suitable default
837  * selectivity if not.
838  */
839  s1 = boolvarsel(root, clause, varRelid);
840  }
841 
842  /* Cache the result if possible */
843  if (cacheable)
844  {
845  if (jointype == JOIN_INNER)
846  rinfo->norm_selec = s1;
847  else
848  rinfo->outer_selec = s1;
849  }
850 
851 #ifdef SELECTIVITY_DEBUG
852  elog(DEBUG4, "clause_selectivity: s1 %f", s1);
853 #endif /* SELECTIVITY_DEBUG */
854 
855  return s1;
856 }
Datum constvalue
Definition: primnodes.h:196
Expr * get_notclausearg(Expr *notclause)
Definition: clauses.c:265
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
Index varlevelsup
Definition: primnodes.h:173
Node * estimate_expression_value(PlannerInfo *root, Node *node)
Definition: clauses.c:2492
Expr * orclause
Definition: relation.h:1866
Selectivity restriction_selectivity(PlannerInfo *root, Oid operatorid, List *args, Oid inputcollid, int varRelid)
Definition: plancat.c:1665
double tuples
Definition: relation.h:625
Selectivity rowcomparesel(PlannerInfo *root, RowCompareExpr *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:2205
Relids clause_relids
Definition: relation.h:1850
bool pseudoconstant
Definition: relation.h:1843
Definition: nodes.h:511
double Selectivity
Definition: nodes.h:641
unsigned int Oid
Definition: postgres_ext.h:31
Definition: primnodes.h:163
#define DEBUG4
Definition: elog.h:22
Selectivity scalararraysel(PlannerInfo *root, ScalarArrayOpExpr *clause, bool is_join_clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1848
Selectivity norm_selec
Definition: relation.h:1873
static bool treat_as_join_clause(Node *clause, RestrictInfo *rinfo, int varRelid, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:496
static bool bms_is_subset_singleton(const Bitmapset *s, int x)
Definition: clausesel.c:475
char * s1
#define is_opclause(clause)
Definition: clauses.h:20
Selectivity nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1739
bool and_clause(Node *clause)
Definition: clauses.c:314
Selectivity clause_selectivity(PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:574
#define DatumGetBool(X)
Definition: postgres.h:399
Selectivity outer_selec
Definition: relation.h:1876
bool not_clause(Node *clause)
Definition: clauses.c:236
Expr * clause
Definition: relation.h:1835
Index varno
Definition: primnodes.h:166
char * s2
bool or_clause(Node *clause)
Definition: clauses.c:280
#define lfirst(lc)
Definition: pg_list.h:106
Oid inputcollid
Definition: primnodes.h:501
void * arg
Selectivity join_selectivity(PlannerInfo *root, Oid operatorid, List *args, Oid inputcollid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: plancat.c:1702
Oid opno
Definition: primnodes.h:496
#define elog
Definition: elog.h:219
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:99
RelOptInfo * find_base_rel(PlannerInfo *root, int relid)
Definition: relnode.c:277
List * args
Definition: primnodes.h:502
Selectivity booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1581
Selectivity boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
Definition: selfuncs.c:1542
bool constisnull
Definition: primnodes.h:197

◆ clauselist_selectivity()

Selectivity clauselist_selectivity ( PlannerInfo root,
List clauses,
int  varRelid,
JoinType  jointype,
SpecialJoinInfo sjinfo 
)

Definition at line 99 of file clausesel.c.

References addRangeClause(), generate_unaccent_rules::args, OpExpr::args, bms_is_member(), bms_membership(), BMS_SINGLETON, RestrictInfo::clause, RestrictInfo::clause_relids, clause_selectivity(), DEFAULT_INEQ_SEL, DEFAULT_RANGE_INEQ_SEL, dependencies_clauselist_selectivity(), find_single_rel_for_clauses(), get_oprrest(), RangeQueryClause::have_hibound, RangeQueryClause::have_lobound, RangeQueryClause::hibound, IS_NULL, is_opclause, is_pseudo_constant_clause(), is_pseudo_constant_clause_relids(), IsA, RestrictInfo::left_relids, lfirst, linitial, list_length(), RangeQueryClause::lobound, lsecond, RangeQueryClause::next, NIL, nulltestsel(), NumRelids(), OpExpr::opno, pfree(), RestrictInfo::pseudoconstant, RestrictInfo::right_relids, RTE_RELATION, RelOptInfo::rtekind, s1, s2, RelOptInfo::statlist, and RangeQueryClause::var.

Referenced by brincostestimate(), btcostestimate(), calc_joinrel_size_estimate(), clause_selectivity(), compute_semi_anti_join_factors(), cost_agg(), cost_group(), estimate_path_cost_size(), estimate_size(), genericcostestimate(), get_parameterized_baserel_size(), gincostestimate(), postgresGetForeignJoinPaths(), postgresGetForeignRelSize(), and set_baserel_size_estimates().

104 {
105  Selectivity s1 = 1.0;
106  RelOptInfo *rel;
107  Bitmapset *estimatedclauses = NULL;
108  RangeQueryClause *rqlist = NULL;
109  ListCell *l;
110  int listidx;
111 
112  /*
113  * If there's exactly one clause, just go directly to
114  * clause_selectivity(). None of what we might do below is relevant.
115  */
116  if (list_length(clauses) == 1)
117  return clause_selectivity(root, (Node *) linitial(clauses),
118  varRelid, jointype, sjinfo);
119 
120  /*
121  * Determine if these clauses reference a single relation. If so, and if
122  * it has extended statistics, try to apply those.
123  */
124  rel = find_single_rel_for_clauses(root, clauses);
125  if (rel && rel->rtekind == RTE_RELATION && rel->statlist != NIL)
126  {
127  /*
128  * Perform selectivity estimations on any clauses found applicable by
129  * dependencies_clauselist_selectivity. 'estimatedclauses' will be
130  * filled with the 0-based list positions of clauses used that way, so
131  * that we can ignore them below.
132  */
133  s1 *= dependencies_clauselist_selectivity(root, clauses, varRelid,
134  jointype, sjinfo, rel,
135  &estimatedclauses);
136 
137  /*
138  * This would be the place to apply any other types of extended
139  * statistics selectivity estimations for remaining clauses.
140  */
141  }
142 
143  /*
144  * Apply normal selectivity estimates for remaining clauses. We'll be
145  * careful to skip any clauses which were already estimated above.
146  *
147  * Anything that doesn't look like a potential rangequery clause gets
148  * multiplied into s1 and forgotten. Anything that does gets inserted into
149  * an rqlist entry.
150  */
151  listidx = -1;
152  foreach(l, clauses)
153  {
154  Node *clause = (Node *) lfirst(l);
155  RestrictInfo *rinfo;
156  Selectivity s2;
157 
158  listidx++;
159 
160  /*
161  * Skip this clause if it's already been estimated by some other
162  * statistics above.
163  */
164  if (bms_is_member(listidx, estimatedclauses))
165  continue;
166 
167  /* Always compute the selectivity using clause_selectivity */
168  s2 = clause_selectivity(root, clause, varRelid, jointype, sjinfo);
169 
170  /*
171  * Check for being passed a RestrictInfo.
172  *
173  * If it's a pseudoconstant RestrictInfo, then s2 is either 1.0 or
174  * 0.0; just use that rather than looking for range pairs.
175  */
176  if (IsA(clause, RestrictInfo))
177  {
178  rinfo = (RestrictInfo *) clause;
179  if (rinfo->pseudoconstant)
180  {
181  s1 = s1 * s2;
182  continue;
183  }
184  clause = (Node *) rinfo->clause;
185  }
186  else
187  rinfo = NULL;
188 
189  /*
190  * See if it looks like a restriction clause with a pseudoconstant on
191  * one side. (Anything more complicated than that might not behave in
192  * the simple way we are expecting.) Most of the tests here can be
193  * done more efficiently with rinfo than without.
194  */
195  if (is_opclause(clause) && list_length(((OpExpr *) clause)->args) == 2)
196  {
197  OpExpr *expr = (OpExpr *) clause;
198  bool varonleft = true;
199  bool ok;
200 
201  if (rinfo)
202  {
203  ok = (bms_membership(rinfo->clause_relids) == BMS_SINGLETON) &&
205  rinfo->right_relids) ||
206  (varonleft = false,
208  rinfo->left_relids)));
209  }
210  else
211  {
212  ok = (NumRelids(clause) == 1) &&
214  (varonleft = false,
216  }
217 
218  if (ok)
219  {
220  /*
221  * If it's not a "<"/"<="/">"/">=" operator, just merge the
222  * selectivity in generically. But if it's the right oprrest,
223  * add the clause to rqlist for later processing.
224  */
225  switch (get_oprrest(expr->opno))
226  {
227  case F_SCALARLTSEL:
228  case F_SCALARLESEL:
229  addRangeClause(&rqlist, clause,
230  varonleft, true, s2);
231  break;
232  case F_SCALARGTSEL:
233  case F_SCALARGESEL:
234  addRangeClause(&rqlist, clause,
235  varonleft, false, s2);
236  break;
237  default:
238  /* Just merge the selectivity in generically */
239  s1 = s1 * s2;
240  break;
241  }
242  continue; /* drop to loop bottom */
243  }
244  }
245 
246  /* Not the right form, so treat it generically. */
247  s1 = s1 * s2;
248  }
249 
250  /*
251  * Now scan the rangequery pair list.
252  */
253  while (rqlist != NULL)
254  {
255  RangeQueryClause *rqnext;
256 
257  if (rqlist->have_lobound && rqlist->have_hibound)
258  {
259  /* Successfully matched a pair of range clauses */
260  Selectivity s2;
261 
262  /*
263  * Exact equality to the default value probably means the
264  * selectivity function punted. This is not airtight but should
265  * be good enough.
266  */
267  if (rqlist->hibound == DEFAULT_INEQ_SEL ||
268  rqlist->lobound == DEFAULT_INEQ_SEL)
269  {
271  }
272  else
273  {
274  s2 = rqlist->hibound + rqlist->lobound - 1.0;
275 
276  /* Adjust for double-exclusion of NULLs */
277  s2 += nulltestsel(root, IS_NULL, rqlist->var,
278  varRelid, jointype, sjinfo);
279 
280  /*
281  * A zero or slightly negative s2 should be converted into a
282  * small positive value; we probably are dealing with a very
283  * tight range and got a bogus result due to roundoff errors.
284  * However, if s2 is very negative, then we probably have
285  * default selectivity estimates on one or both sides of the
286  * range that we failed to recognize above for some reason.
287  */
288  if (s2 <= 0.0)
289  {
290  if (s2 < -0.01)
291  {
292  /*
293  * No data available --- use a default estimate that
294  * is small, but not real small.
295  */
297  }
298  else
299  {
300  /*
301  * It's just roundoff error; use a small positive
302  * value
303  */
304  s2 = 1.0e-10;
305  }
306  }
307  }
308  /* Merge in the selectivity of the pair of clauses */
309  s1 *= s2;
310  }
311  else
312  {
313  /* Only found one of a pair, merge it in generically */
314  if (rqlist->have_lobound)
315  s1 *= rqlist->lobound;
316  else
317  s1 *= rqlist->hibound;
318  }
319  /* release storage and advance */
320  rqnext = rqlist->next;
321  pfree(rqlist);
322  rqlist = rqnext;
323  }
324 
325  return s1;
326 }
#define NIL
Definition: pg_list.h:69
Selectivity dependencies_clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo, RelOptInfo *rel, Bitmapset **estimatedclauses)
Definition: dependencies.c:919
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
List * statlist
Definition: relation.h:623
bool is_pseudo_constant_clause_relids(Node *clause, Relids relids)
Definition: clauses.c:2250
#define DEFAULT_INEQ_SEL
Definition: selfuncs.h:37
Relids clause_relids
Definition: relation.h:1850
bool pseudoconstant
Definition: relation.h:1843
Definition: nodes.h:511
Relids left_relids
Definition: relation.h:1862
double Selectivity
Definition: nodes.h:641
#define lsecond(l)
Definition: pg_list.h:116
void pfree(void *pointer)
Definition: mcxt.c:949
#define linitial(l)
Definition: pg_list.h:111
char * s1
#define is_opclause(clause)
Definition: clauses.h:20
Selectivity nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1739
Selectivity clause_selectivity(PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:574
bool is_pseudo_constant_clause(Node *clause)
Definition: clauses.c:2230
struct RangeQueryClause * next
Definition: clausesel.c:34
static void addRangeClause(RangeQueryClause **rqlist, Node *clause, bool varonleft, bool isLTsel, Selectivity s2)
Definition: clausesel.c:334
Selectivity hibound
Definition: clausesel.c:39
Expr * clause
Definition: relation.h:1835
Selectivity lobound
Definition: clausesel.c:38
RegProcedure get_oprrest(Oid opno)
Definition: lsyscache.c:1361
BMS_Membership bms_membership(const Bitmapset *a)
Definition: bitmapset.c:634
char * s2
RTEKind rtekind
Definition: relation.h:615
static RelOptInfo * find_single_rel_for_clauses(PlannerInfo *root, List *clauses)
Definition: clausesel.c:430
Relids right_relids
Definition: relation.h:1863
#define lfirst(lc)
Definition: pg_list.h:106
static int list_length(const List *l)
Definition: pg_list.h:89
#define DEFAULT_RANGE_INEQ_SEL
Definition: selfuncs.h:40
Oid opno
Definition: primnodes.h:496
List * args
Definition: primnodes.h:502
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:420
int NumRelids(Node *clause)
Definition: clauses.c:2272

◆ compute_bitmap_pages()

double compute_bitmap_pages ( PlannerInfo root,
RelOptInfo baserel,
Path bitmapqual,
int  loop_count,
Cost cost,
double *  tuple 
)

Definition at line 5170 of file costsize.c.

References clamp_row_est(), cost_bitmap_tree_node(), get_indexpath_pages(), index_pages_fetched(), Max, Min, RelOptInfo::pages, T, tbm_calculate_entries(), RelOptInfo::tuples, and work_mem.

Referenced by cost_bitmap_heap_scan(), and create_partial_bitmap_paths().

5172 {
5173  Cost indexTotalCost;
5174  Selectivity indexSelectivity;
5175  double T;
5176  double pages_fetched;
5177  double tuples_fetched;
5178  double heap_pages;
5179  long maxentries;
5180 
5181  /*
5182  * Fetch total cost of obtaining the bitmap, as well as its total
5183  * selectivity.
5184  */
5185  cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
5186 
5187  /*
5188  * Estimate number of main-table pages fetched.
5189  */
5190  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
5191 
5192  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
5193 
5194  /*
5195  * For a single scan, the number of heap pages that need to be fetched is
5196  * the same as the Mackert and Lohman formula for the case T <= b (ie, no
5197  * re-reads needed).
5198  */
5199  pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
5200 
5201  /*
5202  * Calculate the number of pages fetched from the heap. Then based on
5203  * current work_mem estimate get the estimated maxentries in the bitmap.
5204  * (Note that we always do this calculation based on the number of pages
5205  * that would be fetched in a single iteration, even if loop_count > 1.
5206  * That's correct, because only that number of entries will be stored in
5207  * the bitmap at one time.)
5208  */
5209  heap_pages = Min(pages_fetched, baserel->pages);
5210  maxentries = tbm_calculate_entries(work_mem * 1024L);
5211 
5212  if (loop_count > 1)
5213  {
5214  /*
5215  * For repeated bitmap scans, scale up the number of tuples fetched in
5216  * the Mackert and Lohman formula by the number of scans, so that we
5217  * estimate the number of pages fetched by all the scans. Then
5218  * pro-rate for one scan.
5219  */
5220  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
5221  baserel->pages,
5222  get_indexpath_pages(bitmapqual),
5223  root);
5224  pages_fetched /= loop_count;
5225  }
5226 
5227  if (pages_fetched >= T)
5228  pages_fetched = T;
5229  else
5230  pages_fetched = ceil(pages_fetched);
5231 
5232  if (maxentries < heap_pages)
5233  {
5234  double exact_pages;
5235  double lossy_pages;
5236 
5237  /*
5238  * Crude approximation of the number of lossy pages. Because of the
5239  * way tbm_lossify() is coded, the number of lossy pages increases
5240  * very sharply as soon as we run short of memory; this formula has
5241  * that property and seems to perform adequately in testing, but it's
5242  * possible we could do better somehow.
5243  */
5244  lossy_pages = Max(0, heap_pages - maxentries / 2);
5245  exact_pages = heap_pages - lossy_pages;
5246 
5247  /*
5248  * If there are lossy pages then recompute the number of tuples
5249  * processed by the bitmap heap node. We assume here that the chance
5250  * of a given tuple coming from an exact page is the same as the
5251  * chance that a given page is exact. This might not be true, but
5252  * it's not clear how we can do any better.
5253  */
5254  if (lossy_pages > 0)
5255  tuples_fetched =
5256  clamp_row_est(indexSelectivity *
5257  (exact_pages / heap_pages) * baserel->tuples +
5258  (lossy_pages / heap_pages) * baserel->tuples);
5259  }
5260 
5261  if (cost)
5262  *cost = indexTotalCost;
5263  if (tuple)
5264  *tuple = tuples_fetched;
5265 
5266  return pages_fetched;
5267 }
double tuples
Definition: relation.h:625
#define Min(x, y)
Definition: c.h:802
double Selectivity
Definition: nodes.h:641
static const uint32 T[65]
Definition: md5.c:101
int work_mem
Definition: globals.c:113
#define Max(x, y)
Definition: c.h:796
BlockNumber pages
Definition: relation.h:624
static double get_indexpath_pages(Path *bitmapqual)
Definition: costsize.c:879
long tbm_calculate_entries(double maxbytes)
Definition: tidbitmap.c:1545
double clamp_row_est(double nrows)
Definition: costsize.c:174
double index_pages_fetched(double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
Definition: costsize.c:814
double Cost
Definition: nodes.h:642
void cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
Definition: costsize.c:1030

◆ compute_semi_anti_join_factors()

void compute_semi_anti_join_factors ( PlannerInfo root,
RelOptInfo outerrel,
RelOptInfo innerrel,
JoinType  jointype,
SpecialJoinInfo sjinfo,
List restrictlist,
SemiAntiJoinFactors semifactors 
)

Definition at line 3828 of file costsize.c.

References clauselist_selectivity(), SpecialJoinInfo::delay_upper_joins, IS_OUTER_JOIN, RestrictInfo::is_pushed_down, JOIN_ANTI, JOIN_INNER, JOIN_SEMI, SpecialJoinInfo::jointype, lappend(), lfirst_node, SpecialJoinInfo::lhs_strict, list_free(), SemiAntiJoinFactors::match_count, Max, SpecialJoinInfo::min_lefthand, SpecialJoinInfo::min_righthand, NIL, SemiAntiJoinFactors::outer_match_frac, RelOptInfo::relids, RelOptInfo::rows, SpecialJoinInfo::semi_can_btree, SpecialJoinInfo::semi_can_hash, SpecialJoinInfo::semi_operators, SpecialJoinInfo::semi_rhs_exprs, SpecialJoinInfo::syn_lefthand, SpecialJoinInfo::syn_righthand, T_SpecialJoinInfo, and SpecialJoinInfo::type.

Referenced by add_paths_to_joinrel().

3835 {
3836  Selectivity jselec;
3837  Selectivity nselec;
3838  Selectivity avgmatch;
3839  SpecialJoinInfo norm_sjinfo;
3840  List *joinquals;
3841  ListCell *l;
3842 
3843  /*
3844  * In an ANTI join, we must ignore clauses that are "pushed down", since
3845  * those won't affect the match logic. In a SEMI join, we do not
3846  * distinguish joinquals from "pushed down" quals, so just use the whole
3847  * restrictinfo list. For other outer join types, we should consider only
3848  * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
3849  */
3850  if (IS_OUTER_JOIN(jointype))
3851  {
3852  joinquals = NIL;
3853  foreach(l, restrictlist)
3854  {
3855  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
3856 
3857  if (!rinfo->is_pushed_down)
3858  joinquals = lappend(joinquals, rinfo);
3859  }
3860  }
3861  else
3862  joinquals = restrictlist;
3863 
3864  /*
3865  * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
3866  */
3867  jselec = clauselist_selectivity(root,
3868  joinquals,
3869  0,
3870  (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
3871  sjinfo);
3872 
3873  /*
3874  * Also get the normal inner-join selectivity of the join clauses.
3875  */
3876  norm_sjinfo.type = T_SpecialJoinInfo;
3877  norm_sjinfo.min_lefthand = outerrel->relids;
3878  norm_sjinfo.min_righthand = innerrel->relids;
3879  norm_sjinfo.syn_lefthand = outerrel->relids;
3880  norm_sjinfo.syn_righthand = innerrel->relids;
3881  norm_sjinfo.jointype = JOIN_INNER;
3882  /* we don't bother trying to make the remaining fields valid */
3883  norm_sjinfo.lhs_strict = false;
3884  norm_sjinfo.delay_upper_joins = false;
3885  norm_sjinfo.semi_can_btree = false;
3886  norm_sjinfo.semi_can_hash = false;
3887  norm_sjinfo.semi_operators = NIL;
3888  norm_sjinfo.semi_rhs_exprs = NIL;
3889 
3890  nselec = clauselist_selectivity(root,
3891  joinquals,
3892  0,
3893  JOIN_INNER,
3894  &norm_sjinfo);
3895 
3896  /* Avoid leaking a lot of ListCells */
3897  if (IS_OUTER_JOIN(jointype))
3898  list_free(joinquals);
3899 
3900  /*
3901  * jselec can be interpreted as the fraction of outer-rel rows that have
3902  * any matches (this is true for both SEMI and ANTI cases). And nselec is
3903  * the fraction of the Cartesian product that matches. So, the average
3904  * number of matches for each outer-rel row that has at least one match is
3905  * nselec * inner_rows / jselec.
3906  *
3907  * Note: it is correct to use the inner rel's "rows" count here, even
3908  * though we might later be considering a parameterized inner path with
3909  * fewer rows. This is because we have included all the join clauses in
3910  * the selectivity estimate.
3911  */
3912  if (jselec > 0) /* protect against zero divide */
3913  {
3914  avgmatch = nselec * innerrel->rows / jselec;
3915  /* Clamp to sane range */
3916  avgmatch = Max(1.0, avgmatch);
3917  }
3918  else
3919  avgmatch = 1.0;
3920 
3921  semifactors->outer_match_frac = jselec;
3922  semifactors->match_count = avgmatch;
3923 }
#define NIL
Definition: pg_list.h:69
bool semi_can_btree
Definition: relation.h:2015
Relids min_righthand
Definition: relation.h:2008
Selectivity outer_match_frac
Definition: relation.h:2251
NodeTag type
Definition: relation.h:2006
#define IS_OUTER_JOIN(jointype)
Definition: nodes.h:724
double Selectivity
Definition: nodes.h:641
Relids syn_lefthand
Definition: relation.h:2009
Relids syn_righthand
Definition: relation.h:2010
List * semi_rhs_exprs
Definition: relation.h:2018
bool semi_can_hash
Definition: relation.h:2016
#define lfirst_node(type, lc)
Definition: pg_list.h:109
Relids relids
Definition: relation.h:585
List * lappend(List *list, void *datum)
Definition: list.c:128
bool delay_upper_joins
Definition: relation.h:2013
double rows
Definition: relation.h:588
bool is_pushed_down
Definition: relation.h:1837
#define Max(x, y)
Definition: c.h:796
JoinType jointype
Definition: relation.h:2011
Selectivity match_count
Definition: relation.h:2252
List * semi_operators
Definition: relation.h:2017
void list_free(List *list)
Definition: list.c:1133
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:99
Definition: pg_list.h:45
Relids min_lefthand
Definition: relation.h:2007

◆ cost_agg()

void cost_agg ( Path path,
PlannerInfo root,
AggStrategy  aggstrategy,
const AggClauseCosts aggcosts,
int  numGroupCols,
double  numGroups,
List quals,
Cost  input_startup_cost,
Cost  input_total_cost,
double  input_tuples 
)

Definition at line 1874 of file costsize.c.

References AGG_HASHED, AGG_MIXED, AGG_PLAIN, AGG_SORTED, Assert, clamp_row_est(), clauselist_selectivity(), cost_qual_eval(), cpu_operator_cost, cpu_tuple_cost, disable_cost, enable_hashagg, AggClauseCosts::finalCost, JOIN_INNER, MemSet, QualCost::per_tuple, Path::rows, QualCost::startup, Path::startup_cost, Path::total_cost, and AggClauseCosts::transCost.

Referenced by choose_hashed_setop(), create_agg_path(), create_groupingsets_path(), and create_unique_path().

1880 {
1881  double output_tuples;
1882  Cost startup_cost;
1883  Cost total_cost;
1884  AggClauseCosts dummy_aggcosts;
1885 
1886  /* Use all-zero per-aggregate costs if NULL is passed */
1887  if (aggcosts == NULL)
1888  {
1889  Assert(aggstrategy == AGG_HASHED);
1890  MemSet(&dummy_aggcosts, 0, sizeof(AggClauseCosts));
1891  aggcosts = &dummy_aggcosts;
1892  }
1893 
1894  /*
1895  * The transCost.per_tuple component of aggcosts should be charged once
1896  * per input tuple, corresponding to the costs of evaluating the aggregate
1897  * transfns and their input expressions (with any startup cost of course
1898  * charged but once). The finalCost component is charged once per output
1899  * tuple, corresponding to the costs of evaluating the finalfns.
1900  *
1901  * If we are grouping, we charge an additional cpu_operator_cost per
1902  * grouping column per input tuple for grouping comparisons.
1903  *
1904  * We will produce a single output tuple if not grouping, and a tuple per
1905  * group otherwise. We charge cpu_tuple_cost for each output tuple.
1906  *
1907  * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
1908  * same total CPU cost, but AGG_SORTED has lower startup cost. If the
1909  * input path is already sorted appropriately, AGG_SORTED should be
1910  * preferred (since it has no risk of memory overflow). This will happen
1911  * as long as the computed total costs are indeed exactly equal --- but if
1912  * there's roundoff error we might do the wrong thing. So be sure that
1913  * the computations below form the same intermediate values in the same
1914  * order.
1915  */
1916  if (aggstrategy == AGG_PLAIN)
1917  {
1918  startup_cost = input_total_cost;
1919  startup_cost += aggcosts->transCost.startup;
1920  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
1921  startup_cost += aggcosts->finalCost;
1922  /* we aren't grouping */
1923  total_cost = startup_cost + cpu_tuple_cost;
1924  output_tuples = 1;
1925  }
1926  else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
1927  {
1928  /* Here we are able to deliver output on-the-fly */
1929  startup_cost = input_startup_cost;
1930  total_cost = input_total_cost;
1931  if (aggstrategy == AGG_MIXED && !enable_hashagg)
1932  {
1933  startup_cost += disable_cost;
1934  total_cost += disable_cost;
1935  }
1936  /* calcs phrased this way to match HASHED case, see note above */
1937  total_cost += aggcosts->transCost.startup;
1938  total_cost += aggcosts->transCost.per_tuple * input_tuples;
1939  total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
1940  total_cost += aggcosts->finalCost * numGroups;
1941  total_cost += cpu_tuple_cost * numGroups;
1942  output_tuples = numGroups;
1943  }
1944  else
1945  {
1946  /* must be AGG_HASHED */
1947  startup_cost = input_total_cost;
1948  if (!enable_hashagg)
1949  startup_cost += disable_cost;
1950  startup_cost += aggcosts->transCost.startup;
1951  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
1952  startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
1953  total_cost = startup_cost;
1954  total_cost += aggcosts->finalCost * numGroups;
1955  total_cost += cpu_tuple_cost * numGroups;
1956  output_tuples = numGroups;
1957  }
1958 
1959  /*
1960  * If there are quals (HAVING quals), account for their cost and
1961  * selectivity.
1962  */
1963  if (quals)
1964  {
1965  QualCost qual_cost;
1966 
1967  cost_qual_eval(&qual_cost, quals, root);
1968  startup_cost += qual_cost.startup;
1969  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
1970 
1971  output_tuples = clamp_row_est(output_tuples *
1973  quals,
1974  0,
1975  JOIN_INNER,
1976  NULL));
1977  }
1978 
1979  path->rows = output_tuples;
1980  path->startup_cost = startup_cost;
1981  path->total_cost = total_cost;
1982 }
#define MemSet(start, val, len)
Definition: c.h:853
QualCost transCost
Definition: relation.h:62
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
double cpu_operator_cost
Definition: costsize.c:108
Cost finalCost
Definition: relation.h:63
Cost total_cost
Definition: relation.h:1054
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
double cpu_tuple_cost
Definition: costsize.c:106
bool enable_hashagg
Definition: costsize.c:124
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:99
double clamp_row_est(double nrows)
Definition: costsize.c:174
double Cost
Definition: nodes.h:642

◆ cost_bitmap_and_node()

void cost_bitmap_and_node ( BitmapAndPath path,
PlannerInfo root 
)

Definition at line 1073 of file costsize.c.

References BitmapAndPath::bitmapquals, BitmapAndPath::bitmapselectivity, cost_bitmap_tree_node(), cpu_operator_cost, lfirst, list_head(), BitmapAndPath::path, Path::rows, Path::startup_cost, subpath(), and Path::total_cost.

Referenced by bitmap_and_cost_est(), and create_bitmap_and_path().

1074 {
1075  Cost totalCost;
1076  Selectivity selec;
1077  ListCell *l;
1078 
1079  /*
1080  * We estimate AND selectivity on the assumption that the inputs are
1081  * independent. This is probably often wrong, but we don't have the info
1082  * to do better.
1083  *
1084  * The runtime cost of the BitmapAnd itself is estimated at 100x
1085  * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1086  * definitely too simplistic?
1087  */
1088  totalCost = 0.0;
1089  selec = 1.0;
1090  foreach(l, path->bitmapquals)
1091  {
1092  Path *subpath = (Path *) lfirst(l);
1093  Cost subCost;
1094  Selectivity subselec;
1095 
1096  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1097 
1098  selec *= subselec;
1099 
1100  totalCost += subCost;
1101  if (l != list_head(path->bitmapquals))
1102  totalCost += 100.0 * cpu_operator_cost;
1103  }
1104  path->bitmapselectivity = selec;
1105  path->path.rows = 0; /* per above, not used */
1106  path->path.startup_cost = totalCost;
1107  path->path.total_cost = totalCost;
1108 }
double Selectivity
Definition: nodes.h:641
Selectivity bitmapselectivity
Definition: relation.h:1163
List * bitmapquals
Definition: relation.h:1162
Cost startup_cost
Definition: relation.h:1053
static ListCell * list_head(const List *l)
Definition: pg_list.h:77
double cpu_operator_cost
Definition: costsize.c:108
Cost total_cost
Definition: relation.h:1054
#define lfirst(lc)
Definition: pg_list.h:106
double rows
Definition: relation.h:1052
double Cost
Definition: nodes.h:642
Datum subpath(PG_FUNCTION_ARGS)
Definition: ltree_op.c:234
void cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
Definition: costsize.c:1030

◆ cost_bitmap_heap_scan()

void cost_bitmap_heap_scan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info,
Path bitmapqual,
double  loop_count 
)

Definition at line 929 of file costsize.c.

References Assert, clamp_row_est(), compute_bitmap_pages(), PathTarget::cost, cpu_tuple_cost, disable_cost, enable_bitmapscan, get_parallel_divisor(), get_restriction_qual_cost(), get_tablespace_page_costs(), IsA, RelOptInfo::pages, Path::parallel_workers, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::reltablespace, RelOptInfo::rows, Path::rows, RTE_RELATION, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, T, and Path::total_cost.

Referenced by bitmap_and_cost_est(), bitmap_scan_cost_est(), and create_bitmap_heap_path().

932 {
933  Cost startup_cost = 0;
934  Cost run_cost = 0;
935  Cost indexTotalCost;
936  QualCost qpqual_cost;
937  Cost cpu_per_tuple;
938  Cost cost_per_page;
939  Cost cpu_run_cost;
940  double tuples_fetched;
941  double pages_fetched;
942  double spc_seq_page_cost,
943  spc_random_page_cost;
944  double T;
945 
946  /* Should only be applied to base relations */
947  Assert(IsA(baserel, RelOptInfo));
948  Assert(baserel->relid > 0);
949  Assert(baserel->rtekind == RTE_RELATION);
950 
951  /* Mark the path with the correct row estimate */
952  if (param_info)
953  path->rows = param_info->ppi_rows;
954  else
955  path->rows = baserel->rows;
956 
957  if (!enable_bitmapscan)
958  startup_cost += disable_cost;
959 
960  pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
961  loop_count, &indexTotalCost,
962  &tuples_fetched);
963 
964  startup_cost += indexTotalCost;
965  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
966 
967  /* Fetch estimated page costs for tablespace containing table. */
969  &spc_random_page_cost,
970  &spc_seq_page_cost);
971 
972  /*
973  * For small numbers of pages we should charge spc_random_page_cost
974  * apiece, while if nearly all the table's pages are being read, it's more
975  * appropriate to charge spc_seq_page_cost apiece. The effect is
976  * nonlinear, too. For lack of a better idea, interpolate like this to
977  * determine the cost per page.
978  */
979  if (pages_fetched >= 2.0)
980  cost_per_page = spc_random_page_cost -
981  (spc_random_page_cost - spc_seq_page_cost)
982  * sqrt(pages_fetched / T);
983  else
984  cost_per_page = spc_random_page_cost;
985 
986  run_cost += pages_fetched * cost_per_page;
987 
988  /*
989  * Estimate CPU costs per tuple.
990  *
991  * Often the indexquals don't need to be rechecked at each tuple ... but
992  * not always, especially not if there are enough tuples involved that the
993  * bitmaps become lossy. For the moment, just assume they will be
994  * rechecked always. This means we charge the full freight for all the
995  * scan clauses.
996  */
997  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
998 
999  startup_cost += qpqual_cost.startup;
1000  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1001  cpu_run_cost = cpu_per_tuple * tuples_fetched;
1002 
1003  /* Adjust costing for parallelism, if used. */
1004  if (path->parallel_workers > 0)
1005  {
1006  double parallel_divisor = get_parallel_divisor(path);
1007 
1008  /* The CPU cost is divided among all the workers. */
1009  cpu_run_cost /= parallel_divisor;
1010 
1011  path->rows = clamp_row_est(path->rows / parallel_divisor);
1012  }
1013 
1014 
1015  run_cost += cpu_run_cost;
1016 
1017  /* tlist eval costs are paid per output row, not per tuple scanned */
1018  startup_cost += path->pathtarget->cost.startup;
1019  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1020 
1021  path->startup_cost = startup_cost;
1022  path->total_cost = startup_cost + run_cost;
1023 }
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
PathTarget * pathtarget
Definition: relation.h:1043
Oid reltablespace
Definition: relation.h:614
int parallel_workers
Definition: relation.h:1049
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
static const uint32 T[65]
Definition: md5.c:101
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5137
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
Index relid
Definition: relation.h:613
bool enable_bitmapscan
Definition: costsize.c:121
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
BlockNumber pages
Definition: relation.h:624
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
double clamp_row_est(double nrows)
Definition: costsize.c:174
double compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel, Path *bitmapqual, int loop_count, Cost *cost, double *tuple)
Definition: costsize.c:5170
double Cost
Definition: nodes.h:642

◆ cost_bitmap_or_node()

void cost_bitmap_or_node ( BitmapOrPath path,
PlannerInfo root 
)

Definition at line 1117 of file costsize.c.

References BitmapOrPath::bitmapquals, BitmapOrPath::bitmapselectivity, cost_bitmap_tree_node(), cpu_operator_cost, IsA, lfirst, list_head(), Min, BitmapOrPath::path, Path::rows, Path::startup_cost, subpath(), and Path::total_cost.

Referenced by create_bitmap_or_path().

1118 {
1119  Cost totalCost;
1120  Selectivity selec;
1121  ListCell *l;
1122 
1123  /*
1124  * We estimate OR selectivity on the assumption that the inputs are
1125  * non-overlapping, since that's often the case in "x IN (list)" type
1126  * situations. Of course, we clamp to 1.0 at the end.
1127  *
1128  * The runtime cost of the BitmapOr itself is estimated at 100x
1129  * cpu_operator_cost for each tbm_union needed. Probably too small,
1130  * definitely too simplistic? We are aware that the tbm_unions are
1131  * optimized out when the inputs are BitmapIndexScans.
1132  */
1133  totalCost = 0.0;
1134  selec = 0.0;
1135  foreach(l, path->bitmapquals)
1136  {
1137  Path *subpath = (Path *) lfirst(l);
1138  Cost subCost;
1139  Selectivity subselec;
1140 
1141  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1142 
1143  selec += subselec;
1144 
1145  totalCost += subCost;
1146  if (l != list_head(path->bitmapquals) &&
1147  !IsA(subpath, IndexPath))
1148  totalCost += 100.0 * cpu_operator_cost;
1149  }
1150  path->bitmapselectivity = Min(selec, 1.0);
1151  path->path.rows = 0; /* per above, not used */
1152  path->path.startup_cost = totalCost;
1153  path->path.total_cost = totalCost;
1154 }
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
#define Min(x, y)
Definition: c.h:802
double Selectivity
Definition: nodes.h:641
List * bitmapquals
Definition: relation.h:1175
Cost startup_cost
Definition: relation.h:1053
static ListCell * list_head(const List *l)
Definition: pg_list.h:77
double cpu_operator_cost
Definition: costsize.c:108
Selectivity bitmapselectivity
Definition: relation.h:1176
Cost total_cost
Definition: relation.h:1054
#define lfirst(lc)
Definition: pg_list.h:106
double rows
Definition: relation.h:1052
double Cost
Definition: nodes.h:642
Datum subpath(PG_FUNCTION_ARGS)
Definition: ltree_op.c:234
void cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
Definition: costsize.c:1030

◆ cost_bitmap_tree_node()

void cost_bitmap_tree_node ( Path path,
Cost cost,
Selectivity selec 
)

Definition at line 1030 of file costsize.c.

References cpu_operator_cost, elog, ERROR, IsA, nodeTag, Path::rows, and Path::total_cost.

Referenced by choose_bitmap_and(), compute_bitmap_pages(), cost_bitmap_and_node(), cost_bitmap_or_node(), and path_usage_comparator().

1031 {
1032  if (IsA(path, IndexPath))
1033  {
1034  *cost = ((IndexPath *) path)->indextotalcost;
1035  *selec = ((IndexPath *) path)->indexselectivity;
1036 
1037  /*
1038  * Charge a small amount per retrieved tuple to reflect the costs of
1039  * manipulating the bitmap. This is mostly to make sure that a bitmap
1040  * scan doesn't look to be the same cost as an indexscan to retrieve a
1041  * single tuple.
1042  */
1043  *cost += 0.1 * cpu_operator_cost * path->rows;
1044  }
1045  else if (IsA(path, BitmapAndPath))
1046  {
1047  *cost = path->total_cost;
1048  *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1049  }
1050  else if (IsA(path, BitmapOrPath))
1051  {
1052  *cost = path->total_cost;
1053  *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1054  }
1055  else
1056  {
1057  elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1058  *cost = *selec = 0; /* keep compiler quiet */
1059  }
1060 }
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
#define ERROR
Definition: elog.h:43
double cpu_operator_cost
Definition: costsize.c:108
Cost total_cost
Definition: relation.h:1054
double rows
Definition: relation.h:1052
#define nodeTag(nodeptr)
Definition: nodes.h:516
#define elog
Definition: elog.h:219

◆ cost_ctescan()

void cost_ctescan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 1484 of file costsize.c.

References Assert, PathTarget::cost, cpu_tuple_cost, get_restriction_qual_cost(), Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::rows, Path::rows, RTE_CTE, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_ctescan_path(), and create_worktablescan_path().

1486 {
1487  Cost startup_cost = 0;
1488  Cost run_cost = 0;
1489  QualCost qpqual_cost;
1490  Cost cpu_per_tuple;
1491 
1492  /* Should only be applied to base relations that are CTEs */
1493  Assert(baserel->relid > 0);
1494  Assert(baserel->rtekind == RTE_CTE);
1495 
1496  /* Mark the path with the correct row estimate */
1497  if (param_info)
1498  path->rows = param_info->ppi_rows;
1499  else
1500  path->rows = baserel->rows;
1501 
1502  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1503  cpu_per_tuple = cpu_tuple_cost;
1504 
1505  /* Add scanning CPU costs */
1506  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1507 
1508  startup_cost += qpqual_cost.startup;
1509  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1510  run_cost += cpu_per_tuple * baserel->tuples;
1511 
1512  /* tlist eval costs are paid per output row, not per tuple scanned */
1513  startup_cost += path->pathtarget->cost.startup;
1514  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1515 
1516  path->startup_cost = startup_cost;
1517  path->total_cost = startup_cost + run_cost;
1518 }
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
Index relid
Definition: relation.h:613
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
double Cost
Definition: nodes.h:642

◆ cost_functionscan()

void cost_functionscan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 1317 of file costsize.c.

References Assert, PathTarget::cost, cost_qual_eval_node(), cpu_tuple_cost, RangeTblEntry::functions, get_restriction_qual_cost(), Path::pathtarget, QualCost::per_tuple, planner_rt_fetch, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::rows, Path::rows, RTE_FUNCTION, RangeTblEntry::rtekind, QualCost::startup, Path::startup_cost, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_functionscan_path().

1319 {
1320  Cost startup_cost = 0;
1321  Cost run_cost = 0;
1322  QualCost qpqual_cost;
1323  Cost cpu_per_tuple;
1324  RangeTblEntry *rte;
1325  QualCost exprcost;
1326 
1327  /* Should only be applied to base relations that are functions */
1328  Assert(baserel->relid > 0);
1329  rte = planner_rt_fetch(baserel->relid, root);
1330  Assert(rte->rtekind == RTE_FUNCTION);
1331 
1332  /* Mark the path with the correct row estimate */
1333  if (param_info)
1334  path->rows = param_info->ppi_rows;
1335  else
1336  path->rows = baserel->rows;
1337 
1338  /*
1339  * Estimate costs of executing the function expression(s).
1340  *
1341  * Currently, nodeFunctionscan.c always executes the functions to
1342  * completion before returning any rows, and caches the results in a
1343  * tuplestore. So the function eval cost is all startup cost, and per-row
1344  * costs are minimal.
1345  *
1346  * XXX in principle we ought to charge tuplestore spill costs if the
1347  * number of rows is large. However, given how phony our rowcount
1348  * estimates for functions tend to be, there's not a lot of point in that
1349  * refinement right now.
1350  */
1351  cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1352 
1353  startup_cost += exprcost.startup + exprcost.per_tuple;
1354 
1355  /* Add scanning CPU costs */
1356  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1357 
1358  startup_cost += qpqual_cost.startup;
1359  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1360  run_cost += cpu_per_tuple * baserel->tuples;
1361 
1362  /* tlist eval costs are paid per output row, not per tuple scanned */
1363  startup_cost += path->pathtarget->cost.startup;
1364  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1365 
1366  path->startup_cost = startup_cost;
1367  path->total_cost = startup_cost + run_cost;
1368 }
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:3534
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Definition: nodes.h:511
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
#define planner_rt_fetch(rti, root)
Definition: relation.h:328
Cost startup_cost
Definition: relation.h:1053
Index relid
Definition: relation.h:613
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
List * functions
Definition: parsenodes.h:1005
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
RTEKind rtekind
Definition: parsenodes.h:951
double Cost
Definition: nodes.h:642

◆ cost_gather()

void cost_gather ( GatherPath path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info,
double *  rows 
)

Definition at line 350 of file costsize.c.

References parallel_setup_cost, parallel_tuple_cost, GatherPath::path, ParamPathInfo::ppi_rows, RelOptInfo::rows, Path::rows, Path::startup_cost, GatherPath::subpath, and Path::total_cost.

Referenced by create_gather_path().

353 {
354  Cost startup_cost = 0;
355  Cost run_cost = 0;
356 
357  /* Mark the path with the correct row estimate */
358  if (rows)
359  path->path.rows = *rows;
360  else if (param_info)
361  path->path.rows = param_info->ppi_rows;
362  else
363  path->path.rows = rel->rows;
364 
365  startup_cost = path->subpath->startup_cost;
366 
367  run_cost = path->subpath->total_cost - path->subpath->startup_cost;
368 
369  /* Parallel setup and communication cost. */
370  startup_cost += parallel_setup_cost;
371  run_cost += parallel_tuple_cost * path->path.rows;
372 
373  path->path.startup_cost = startup_cost;
374  path->path.total_cost = (startup_cost + run_cost);
375 }
double parallel_setup_cost
Definition: costsize.c:110
Cost startup_cost
Definition: relation.h:1053
Path * subpath
Definition: relation.h:1353
Cost total_cost
Definition: relation.h:1054
double rows
Definition: relation.h:1052
double ppi_rows
Definition: relation.h:1002
Path path
Definition: relation.h:1352
double Cost
Definition: nodes.h:642
double parallel_tuple_cost
Definition: costsize.c:109

◆ cost_gather_merge()

void cost_gather_merge ( GatherMergePath path,
PlannerInfo root,
RelOptInfo rel,
ParamPathInfo param_info,
Cost  input_startup_cost,
Cost  input_total_cost,
double *  rows 
)

Definition at line 388 of file costsize.c.

References Assert, cpu_operator_cost, disable_cost, enable_gathermerge, LOG2, GatherMergePath::num_workers, parallel_setup_cost, parallel_tuple_cost, GatherMergePath::path, ParamPathInfo::ppi_rows, RelOptInfo::rows, Path::rows, Path::startup_cost, and Path::total_cost.

Referenced by create_gather_merge_path().

392 {
393  Cost startup_cost = 0;
394  Cost run_cost = 0;
395  Cost comparison_cost;
396  double N;
397  double logN;
398 
399  /* Mark the path with the correct row estimate */
400  if (rows)
401  path->path.rows = *rows;
402  else if (param_info)
403  path->path.rows = param_info->ppi_rows;
404  else
405  path->path.rows = rel->rows;
406 
407  if (!enable_gathermerge)
408  startup_cost += disable_cost;
409 
410  /*
411  * Add one to the number of workers to account for the leader. This might
412  * be overgenerous since the leader will do less work than other workers
413  * in typical cases, but we'll go with it for now.
414  */
415  Assert(path->num_workers > 0);
416  N = (double) path->num_workers + 1;
417  logN = LOG2(N);
418 
419  /* Assumed cost per tuple comparison */
420  comparison_cost = 2.0 * cpu_operator_cost;
421 
422  /* Heap creation cost */
423  startup_cost += comparison_cost * N * logN;
424 
425  /* Per-tuple heap maintenance cost */
426  run_cost += path->path.rows * comparison_cost * logN;
427 
428  /* small cost for heap management, like cost_merge_append */
429  run_cost += cpu_operator_cost * path->path.rows;
430 
431  /*
432  * Parallel setup and communication cost. Since Gather Merge, unlike
433  * Gather, requires us to block until a tuple is available from every
434  * worker, we bump the IPC cost up a little bit as compared with Gather.
435  * For lack of a better idea, charge an extra 5%.
436  */
437  startup_cost += parallel_setup_cost;
438  run_cost += parallel_tuple_cost * path->path.rows * 1.05;
439 
440  path->path.startup_cost = startup_cost + input_startup_cost;
441  path->path.total_cost = (startup_cost + run_cost + input_total_cost);
442 }
double parallel_setup_cost
Definition: costsize.c:110
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
double cpu_operator_cost
Definition: costsize.c:108
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
#define LOG2(x)
Definition: costsize.c:101
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
double ppi_rows
Definition: relation.h:1002
bool enable_gathermerge
Definition: costsize.c:129
double Cost
Definition: nodes.h:642
double parallel_tuple_cost
Definition: costsize.c:109

◆ cost_group()

void cost_group ( Path path,
PlannerInfo root,
int  numGroupCols,
double  numGroups,
List quals,
Cost  input_startup_cost,
Cost  input_total_cost,
double  input_tuples 
)

Definition at line 2062 of file costsize.c.

References clamp_row_est(), clauselist_selectivity(), cost_qual_eval(), cpu_operator_cost, JOIN_INNER, QualCost::per_tuple, Path::rows, QualCost::startup, Path::startup_cost, and Path::total_cost.

Referenced by choose_hashed_setop(), and create_group_path().

2067 {
2068  double output_tuples;
2069  Cost startup_cost;
2070  Cost total_cost;
2071 
2072  output_tuples = numGroups;
2073  startup_cost = input_startup_cost;
2074  total_cost = input_total_cost;
2075 
2076  /*
2077  * Charge one cpu_operator_cost per comparison per input tuple. We assume
2078  * all columns get compared at most of the tuples.
2079  */
2080  total_cost += cpu_operator_cost * input_tuples * numGroupCols;
2081 
2082  /*
2083  * If there are quals (HAVING quals), account for their cost and
2084  * selectivity.
2085  */
2086  if (quals)
2087  {
2088  QualCost qual_cost;
2089 
2090  cost_qual_eval(&qual_cost, quals, root);
2091  startup_cost += qual_cost.startup;
2092  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2093 
2094  output_tuples = clamp_row_est(output_tuples *
2096  quals,
2097  0,
2098  JOIN_INNER,
2099  NULL));
2100  }
2101 
2102  path->rows = output_tuples;
2103  path->startup_cost = startup_cost;
2104  path->total_cost = total_cost;
2105 }
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: relation.h:1053
double cpu_operator_cost
Definition: costsize.c:108
Cost total_cost
Definition: relation.h:1054
double rows
Definition: relation.h:1052
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:99
double clamp_row_est(double nrows)
Definition: costsize.c:174
double Cost
Definition: nodes.h:642

◆ cost_index()

void cost_index ( IndexPath path,
PlannerInfo root,
double  loop_count,
bool  partial_path 
)

Definition at line 463 of file costsize.c.

References RelOptInfo::allvisfrac, IndexOptInfo::amcostestimate, Assert, clamp_row_est(), compute_parallel_worker(), PathTarget::cost, cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_indexscan, extract_nonindex_conditions(), get_parallel_divisor(), get_tablespace_page_costs(), index_pages_fetched(), IndexPath::indexinfo, IndexPath::indexquals, IndexPath::indexselectivity, IndexPath::indextotalcost, IndexOptInfo::indrestrictinfo, IsA, list_concat(), RelOptInfo::pages, IndexOptInfo::pages, Path::parallel_aware, Path::parallel_workers, Path::param_info, IndexPath::path, Path::pathtarget, Path::pathtype, QualCost::per_tuple, ParamPathInfo::ppi_clauses, ParamPathInfo::ppi_rows, IndexOptInfo::rel, RelOptInfo::relid, RelOptInfo::reltablespace, RelOptInfo::rows, Path::rows, RTE_RELATION, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, T_IndexOnlyScan, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_index_path(), and reparameterize_path().

465 {
466  IndexOptInfo *index = path->indexinfo;
467  RelOptInfo *baserel = index->rel;
468  bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
469  amcostestimate_function amcostestimate;
470  List *qpquals;
471  Cost startup_cost = 0;
472  Cost run_cost = 0;
473  Cost cpu_run_cost = 0;
474  Cost indexStartupCost;
475  Cost indexTotalCost;
476  Selectivity indexSelectivity;
477  double indexCorrelation,
478  csquared;
479  double spc_seq_page_cost,
480  spc_random_page_cost;
481  Cost min_IO_cost,
482  max_IO_cost;
483  QualCost qpqual_cost;
484  Cost cpu_per_tuple;
485  double tuples_fetched;
486  double pages_fetched;
487  double rand_heap_pages;
488  double index_pages;
489 
490  /* Should only be applied to base relations */
491  Assert(IsA(baserel, RelOptInfo) &&
492  IsA(index, IndexOptInfo));
493  Assert(baserel->relid > 0);
494  Assert(baserel->rtekind == RTE_RELATION);
495 
496  /*
497  * Mark the path with the correct row estimate, and identify which quals
498  * will need to be enforced as qpquals. We need not check any quals that
499  * are implied by the index's predicate, so we can use indrestrictinfo not
500  * baserestrictinfo as the list of relevant restriction clauses for the
501  * rel.
502  */
503  if (path->path.param_info)
504  {
505  path->path.rows = path->path.param_info->ppi_rows;
506  /* qpquals come from the rel's restriction clauses and ppi_clauses */
507  qpquals = list_concat(
509  path->indexquals),
511  path->indexquals));
512  }
513  else
514  {
515  path->path.rows = baserel->rows;
516  /* qpquals come from just the rel's restriction clauses */
518  path->indexquals);
519  }
520 
521  if (!enable_indexscan)
522  startup_cost += disable_cost;
523  /* we don't need to check enable_indexonlyscan; indxpath.c does that */
524 
525  /*
526  * Call index-access-method-specific code to estimate the processing cost
527  * for scanning the index, as well as the selectivity of the index (ie,
528  * the fraction of main-table tuples we will have to retrieve) and its
529  * correlation to the main-table tuple order. We need a cast here because
530  * relation.h uses a weak function type to avoid including amapi.h.
531  */
532  amcostestimate = (amcostestimate_function) index->amcostestimate;
533  amcostestimate(root, path, loop_count,
534  &indexStartupCost, &indexTotalCost,
535  &indexSelectivity, &indexCorrelation,
536  &index_pages);
537 
538  /*
539  * Save amcostestimate's results for possible use in bitmap scan planning.
540  * We don't bother to save indexStartupCost or indexCorrelation, because a
541  * bitmap scan doesn't care about either.
542  */
543  path->indextotalcost = indexTotalCost;
544  path->indexselectivity = indexSelectivity;
545 
546  /* all costs for touching index itself included here */
547  startup_cost += indexStartupCost;
548  run_cost += indexTotalCost - indexStartupCost;
549 
550  /* estimate number of main-table tuples fetched */
551  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
552 
553  /* fetch estimated page costs for tablespace containing table */
555  &spc_random_page_cost,
556  &spc_seq_page_cost);
557 
558  /*----------
559  * Estimate number of main-table pages fetched, and compute I/O cost.
560  *
561  * When the index ordering is uncorrelated with the table ordering,
562  * we use an approximation proposed by Mackert and Lohman (see
563  * index_pages_fetched() for details) to compute the number of pages
564  * fetched, and then charge spc_random_page_cost per page fetched.
565  *
566  * When the index ordering is exactly correlated with the table ordering
567  * (just after a CLUSTER, for example), the number of pages fetched should
568  * be exactly selectivity * table_size. What's more, all but the first
569  * will be sequential fetches, not the random fetches that occur in the
570  * uncorrelated case. So if the number of pages is more than 1, we
571  * ought to charge
572  * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
573  * For partially-correlated indexes, we ought to charge somewhere between
574  * these two estimates. We currently interpolate linearly between the
575  * estimates based on the correlation squared (XXX is that appropriate?).
576  *
577  * If it's an index-only scan, then we will not need to fetch any heap
578  * pages for which the visibility map shows all tuples are visible.
579  * Hence, reduce the estimated number of heap fetches accordingly.
580  * We use the measured fraction of the entire heap that is all-visible,
581  * which might not be particularly relevant to the subset of the heap
582  * that this query will fetch; but it's not clear how to do better.
583  *----------
584  */
585  if (loop_count > 1)
586  {
587  /*
588  * For repeated indexscans, the appropriate estimate for the
589  * uncorrelated case is to scale up the number of tuples fetched in
590  * the Mackert and Lohman formula by the number of scans, so that we
591  * estimate the number of pages fetched by all the scans; then
592  * pro-rate the costs for one scan. In this case we assume all the
593  * fetches are random accesses.
594  */
595  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
596  baserel->pages,
597  (double) index->pages,
598  root);
599 
600  if (indexonly)
601  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
602 
603  rand_heap_pages = pages_fetched;
604 
605  max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
606 
607  /*
608  * In the perfectly correlated case, the number of pages touched by
609  * each scan is selectivity * table_size, and we can use the Mackert
610  * and Lohman formula at the page level to estimate how much work is
611  * saved by caching across scans. We still assume all the fetches are
612  * random, though, which is an overestimate that's hard to correct for
613  * without double-counting the cache effects. (But in most cases
614  * where such a plan is actually interesting, only one page would get
615  * fetched per scan anyway, so it shouldn't matter much.)
616  */
617  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
618 
619  pages_fetched = index_pages_fetched(pages_fetched * loop_count,
620  baserel->pages,
621  (double) index->pages,
622  root);
623 
624  if (indexonly)
625  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
626 
627  min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
628  }
629  else
630  {
631  /*
632  * Normal case: apply the Mackert and Lohman formula, and then
633  * interpolate between that and the correlation-derived result.
634  */
635  pages_fetched = index_pages_fetched(tuples_fetched,
636  baserel->pages,
637  (double) index->pages,
638  root);
639 
640  if (indexonly)
641  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
642 
643  rand_heap_pages = pages_fetched;
644 
645  /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
646  max_IO_cost = pages_fetched * spc_random_page_cost;
647 
648  /* min_IO_cost is for the perfectly correlated case (csquared=1) */
649  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
650 
651  if (indexonly)
652  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
653 
654  if (pages_fetched > 0)
655  {
656  min_IO_cost = spc_random_page_cost;
657  if (pages_fetched > 1)
658  min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
659  }
660  else
661  min_IO_cost = 0;
662  }
663 
664  if (partial_path)
665  {
666  /*
667  * For index only scans compute workers based on number of index pages
668  * fetched; the number of heap pages we fetch might be so small as to
669  * effectively rule out parallelism, which we don't want to do.
670  */
671  if (indexonly)
672  rand_heap_pages = -1;
673 
674  /*
675  * Estimate the number of parallel workers required to scan index. Use
676  * the number of heap pages computed considering heap fetches won't be
677  * sequential as for parallel scans the pages are accessed in random
678  * order.
679  */
681  rand_heap_pages, index_pages);
682 
683  /*
684  * Fall out if workers can't be assigned for parallel scan, because in
685  * such a case this path will be rejected. So there is no benefit in
686  * doing extra computation.
687  */
688  if (path->path.parallel_workers <= 0)
689  return;
690 
691  path->path.parallel_aware = true;
692  }
693 
694  /*
695  * Now interpolate based on estimated index order correlation to get total
696  * disk I/O cost for main table accesses.
697  */
698  csquared = indexCorrelation * indexCorrelation;
699 
700  run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
701 
702  /*
703  * Estimate CPU costs per tuple.
704  *
705  * What we want here is cpu_tuple_cost plus the evaluation costs of any
706  * qual clauses that we have to evaluate as qpquals.
707  */
708  cost_qual_eval(&qpqual_cost, qpquals, root);
709 
710  startup_cost += qpqual_cost.startup;
711  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
712 
713  cpu_run_cost += cpu_per_tuple * tuples_fetched;
714 
715  /* tlist eval costs are paid per output row, not per tuple scanned */
716  startup_cost += path->path.pathtarget->cost.startup;
717  cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
718 
719  /* Adjust costing for parallelism, if used. */
720  if (path->path.parallel_workers > 0)
721  {
722  double parallel_divisor = get_parallel_divisor(&path->path);
723 
724  path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
725 
726  /* The CPU cost is divided among all the workers. */
727  cpu_run_cost /= parallel_divisor;
728  }
729 
730  run_cost += cpu_run_cost;
731 
732  path->path.startup_cost = startup_cost;
733  path->path.total_cost = startup_cost + run_cost;
734 }
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
PathTarget * pathtarget
Definition: relation.h:1043
Path path
Definition: relation.h:1118
IndexOptInfo * indexinfo
Definition: relation.h:1119
double tuples
Definition: relation.h:625
Oid reltablespace
Definition: relation.h:614
int parallel_workers
Definition: relation.h:1049
ParamPathInfo * param_info
Definition: relation.h:1045
List * list_concat(List *list1, List *list2)
Definition: list.c:321
static List * extract_nonindex_conditions(List *qual_clauses, List *indexquals)
Definition: costsize.c:754
double Selectivity
Definition: nodes.h:641
Cost startup
Definition: relation.h:45
double allvisfrac
Definition: relation.h:626
Definition: type.h:89
BlockNumber pages
Definition: relation.h:724
NodeTag pathtype
Definition: relation.h:1040
Cost per_tuple
Definition: relation.h:46
List * indexquals
Definition: relation.h:1121
RelOptInfo * rel
Definition: relation.h:721
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: relation.h:1053
Cost indextotalcost
Definition: relation.h:1126
Cost disable_cost
Definition: costsize.c:114
Selectivity indexselectivity
Definition: relation.h:1127
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5137
void(* amcostestimate)()
Definition: relation.h:766
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
Index relid
Definition: relation.h:613
List * indrestrictinfo
Definition: relation.h:746
int compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages)
Definition: allpaths.c:3169
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
BlockNumber pages
Definition: relation.h:624
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
List * ppi_clauses
Definition: relation.h:1003
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
void(* amcostestimate_function)(struct PlannerInfo *root, struct IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: amapi.h:92
bool parallel_aware
Definition: relation.h:1047
double clamp_row_est(double nrows)
Definition: costsize.c:174
Definition: pg_list.h:45
bool enable_indexscan
Definition: costsize.c:119
double index_pages_fetched(double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
Definition: costsize.c:814
double Cost
Definition: nodes.h:642

◆ cost_material()

void cost_material ( Path path,
Cost  input_startup_cost,
Cost  input_total_cost,
double  tuples,
int  width 
)

Definition at line 1820 of file costsize.c.

References cpu_operator_cost, relation_byte_size(), Path::rows, seq_page_cost, Path::startup_cost, Path::total_cost, and work_mem.

Referenced by create_material_path(), and materialize_finished_plan().

1823 {
1824  Cost startup_cost = input_startup_cost;
1825  Cost run_cost = input_total_cost - input_startup_cost;
1826  double nbytes = relation_byte_size(tuples, width);
1827  long work_mem_bytes = work_mem * 1024L;
1828 
1829  path->rows = tuples;
1830 
1831  /*
1832  * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
1833  * reflect bookkeeping overhead. (This rate must be more than what
1834  * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
1835  * if it is exactly the same then there will be a cost tie between
1836  * nestloop with A outer, materialized B inner and nestloop with B outer,
1837  * materialized A inner. The extra cost ensures we'll prefer
1838  * materializing the smaller rel.) Note that this is normally a good deal
1839  * less than cpu_tuple_cost; which is OK because a Material plan node
1840  * doesn't do qual-checking or projection, so it's got less overhead than
1841  * most plan nodes.
1842  */
1843  run_cost += 2 * cpu_operator_cost * tuples;
1844 
1845  /*
1846  * If we will spill to disk, charge at the rate of seq_page_cost per page.
1847  * This cost is assumed to be evenly spread through the plan run phase,
1848  * which isn't exactly accurate but our cost model doesn't allow for
1849  * nonuniform costs within the run phase.
1850  */
1851  if (nbytes > work_mem_bytes)
1852  {
1853  double npages = ceil(nbytes / BLCKSZ);
1854 
1855  run_cost += seq_page_cost * npages;
1856  }
1857 
1858  path->startup_cost = startup_cost;
1859  path->total_cost = startup_cost + run_cost;
1860 }
Cost startup_cost
Definition: relation.h:1053
double cpu_operator_cost
Definition: costsize.c:108
static double relation_byte_size(double tuples, int width)
Definition: costsize.c:5116
int work_mem
Definition: globals.c:113
Cost total_cost
Definition: relation.h:1054
double rows
Definition: relation.h:1052
double seq_page_cost
Definition: costsize.c:104
double Cost
Definition: nodes.h:642

◆ cost_merge_append()

void cost_merge_append ( Path path,
PlannerInfo root,
List pathkeys,
int  n_streams,
Cost  input_startup_cost,
Cost  input_total_cost,
double  tuples 
)

Definition at line 1769 of file costsize.c.

References cpu_operator_cost, LOG2, Path::startup_cost, and Path::total_cost.

Referenced by create_merge_append_path().

1773 {
1774  Cost startup_cost = 0;
1775  Cost run_cost = 0;
1776  Cost comparison_cost;
1777  double N;
1778  double logN;
1779 
1780  /*
1781  * Avoid log(0)...
1782  */
1783  N = (n_streams < 2) ? 2.0 : (double) n_streams;
1784  logN = LOG2(N);
1785 
1786  /* Assumed cost per tuple comparison */
1787  comparison_cost = 2.0 * cpu_operator_cost;
1788 
1789  /* Heap creation cost */
1790  startup_cost += comparison_cost * N * logN;
1791 
1792  /* Per-tuple heap maintenance cost */
1793  run_cost += tuples * comparison_cost * logN;
1794 
1795  /*
1796  * Also charge a small amount (arbitrarily set equal to operator cost) per
1797  * extracted tuple. We don't charge cpu_tuple_cost because a MergeAppend
1798  * node doesn't do qual-checking or projection, so it has less overhead
1799  * than most plan nodes.
1800  */
1801  run_cost += cpu_operator_cost * tuples;
1802 
1803  path->startup_cost = startup_cost + input_startup_cost;
1804  path->total_cost = startup_cost + run_cost + input_total_cost;
1805 }
Cost startup_cost
Definition: relation.h:1053
double cpu_operator_cost
Definition: costsize.c:108
Cost total_cost
Definition: relation.h:1054
#define LOG2(x)
Definition: costsize.c:101
double Cost
Definition: nodes.h:642

◆ cost_namedtuplestorescan()

void cost_namedtuplestorescan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 1525 of file costsize.c.

References Assert, cpu_tuple_cost, get_restriction_qual_cost(), QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::rows, Path::rows, RTE_NAMEDTUPLESTORE, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_namedtuplestorescan_path().

1527 {
1528  Cost startup_cost = 0;
1529  Cost run_cost = 0;
1530  QualCost qpqual_cost;
1531  Cost cpu_per_tuple;
1532 
1533  /* Should only be applied to base relations that are Tuplestores */
1534  Assert(baserel->relid > 0);
1535  Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1536 
1537  /* Mark the path with the correct row estimate */
1538  if (param_info)
1539  path->rows = param_info->ppi_rows;
1540  else
1541  path->rows = baserel->rows;
1542 
1543  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1544  cpu_per_tuple = cpu_tuple_cost;
1545 
1546  /* Add scanning CPU costs */
1547  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1548 
1549  startup_cost += qpqual_cost.startup;
1550  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1551  run_cost += cpu_per_tuple * baserel->tuples;
1552 
1553  path->startup_cost = startup_cost;
1554  path->total_cost = startup_cost + run_cost;
1555 }
double tuples
Definition: relation.h:625
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
Index relid
Definition: relation.h:613
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
double Cost
Definition: nodes.h:642

◆ cost_qual_eval()

void cost_qual_eval ( QualCost cost,
List quals,
PlannerInfo root 
)

Definition at line 3508 of file costsize.c.

References cost_qual_eval_walker(), lfirst, QualCost::per_tuple, cost_qual_eval_context::root, QualCost::startup, and cost_qual_eval_context::total.

Referenced by cost_agg(), cost_group(), cost_index(), cost_subplan(), cost_tidscan(), create_minmaxagg_path(), create_result_path(), estimate_path_cost_size(), final_cost_hashjoin(), final_cost_mergejoin(), final_cost_nestloop(), get_restriction_qual_cost(), inline_function(), plan_cluster_use_sort(), postgresGetForeignJoinPaths(), postgresGetForeignRelSize(), set_baserel_size_estimates(), and set_foreign_size_estimates().

3509 {
3510  cost_qual_eval_context context;
3511  ListCell *l;
3512 
3513  context.root = root;
3514  context.total.startup = 0;
3515  context.total.per_tuple = 0;
3516 
3517  /* We don't charge any cost for the implicit ANDing at top level ... */
3518 
3519  foreach(l, quals)
3520  {
3521  Node *qual = (Node *) lfirst(l);
3522 
3523  cost_qual_eval_walker(qual, &context);
3524  }
3525 
3526  *cost = context.total;
3527 }
PlannerInfo * root
Definition: costsize.c:134
Definition: nodes.h:511
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
Definition: costsize.c:3548
#define lfirst(lc)
Definition: pg_list.h:106

◆ cost_qual_eval_node()

void cost_qual_eval_node ( QualCost cost,
Node qual,
PlannerInfo root 
)

Definition at line 3534 of file costsize.c.

References cost_qual_eval_walker(), QualCost::per_tuple, cost_qual_eval_context::root, QualCost::startup, and cost_qual_eval_context::total.

Referenced by add_placeholders_to_joinrel(), cost_functionscan(), cost_qual_eval_walker(), cost_tablefuncscan(), cost_windowagg(), get_agg_clause_costs_walker(), make_sort_input_target(), order_qual_clauses(), orderby_operands_eval_cost(), other_operands_eval_cost(), set_pathtarget_cost_width(), and set_rel_width().

3535 {
3536  cost_qual_eval_context context;
3537 
3538  context.root = root;
3539  context.total.startup = 0;
3540  context.total.per_tuple = 0;
3541 
3542  cost_qual_eval_walker(qual, &context);
3543 
3544  *cost = context.total;
3545 }
PlannerInfo * root
Definition: costsize.c:134
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
Definition: costsize.c:3548

◆ cost_recursive_union()

void cost_recursive_union ( Path runion,
Path nrterm,
Path rterm 
)

Definition at line 1565 of file costsize.c.

References cpu_tuple_cost, Max, Path::pathtarget, Path::rows, Path::startup_cost, Path::total_cost, and PathTarget::width.

Referenced by create_recursiveunion_path().

1566 {
1567  Cost startup_cost;
1568  Cost total_cost;
1569  double total_rows;
1570 
1571  /* We probably have decent estimates for the non-recursive term */
1572  startup_cost = nrterm->startup_cost;
1573  total_cost = nrterm->total_cost;
1574  total_rows = nrterm->rows;
1575 
1576  /*
1577  * We arbitrarily assume that about 10 recursive iterations will be
1578  * needed, and that we've managed to get a good fix on the cost and output
1579  * size of each one of them. These are mighty shaky assumptions but it's
1580  * hard to see how to do better.
1581  */
1582  total_cost += 10 * rterm->total_cost;
1583  total_rows += 10 * rterm->rows;
1584 
1585  /*
1586  * Also charge cpu_tuple_cost per row to account for the costs of
1587  * manipulating the tuplestores. (We don't worry about possible
1588  * spill-to-disk costs.)
1589  */
1590  total_cost += cpu_tuple_cost * total_rows;
1591 
1592  runion->startup_cost = startup_cost;
1593  runion->total_cost = total_cost;
1594  runion->rows = total_rows;
1595  runion->pathtarget->width = Max(nrterm->pathtarget->width,
1596  rterm->pathtarget->width);
1597 }
PathTarget * pathtarget
Definition: relation.h:1043
Cost startup_cost
Definition: relation.h:1053
Cost total_cost
Definition: relation.h:1054
#define Max(x, y)
Definition: c.h:796
double rows
Definition: relation.h:1052
double cpu_tuple_cost
Definition: costsize.c:106
int width
Definition: relation.h:975
double Cost
Definition: nodes.h:642

◆ cost_samplescan()

void cost_samplescan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 275 of file costsize.c.

References Assert, PathTarget::cost, cpu_tuple_cost, get_restriction_qual_cost(), get_tablespace_page_costs(), GetTsmRoutine(), TsmRoutine::NextSampleBlock, RelOptInfo::pages, Path::pathtarget, QualCost::per_tuple, planner_rt_fetch, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::reltablespace, RelOptInfo::rows, Path::rows, RTE_RELATION, RangeTblEntry::rtekind, QualCost::startup, Path::startup_cost, RangeTblEntry::tablesample, Path::total_cost, TableSampleClause::tsmhandler, and RelOptInfo::tuples.

Referenced by create_samplescan_path().

277 {
278  Cost startup_cost = 0;
279  Cost run_cost = 0;
280  RangeTblEntry *rte;
281  TableSampleClause *tsc;
282  TsmRoutine *tsm;
283  double spc_seq_page_cost,
284  spc_random_page_cost,
285  spc_page_cost;
286  QualCost qpqual_cost;
287  Cost cpu_per_tuple;
288 
289  /* Should only be applied to base relations with tablesample clauses */
290  Assert(baserel->relid > 0);
291  rte = planner_rt_fetch(baserel->relid, root);
292  Assert(rte->rtekind == RTE_RELATION);
293  tsc = rte->tablesample;
294  Assert(tsc != NULL);
295  tsm = GetTsmRoutine(tsc->tsmhandler);
296 
297  /* Mark the path with the correct row estimate */
298  if (param_info)
299  path->rows = param_info->ppi_rows;
300  else
301  path->rows = baserel->rows;
302 
303  /* fetch estimated page cost for tablespace containing table */
305  &spc_random_page_cost,
306  &spc_seq_page_cost);
307 
308  /* if NextSampleBlock is used, assume random access, else sequential */
309  spc_page_cost = (tsm->NextSampleBlock != NULL) ?
310  spc_random_page_cost : spc_seq_page_cost;
311 
312  /*
313  * disk costs (recall that baserel->pages has already been set to the
314  * number of pages the sampling method will visit)
315  */
316  run_cost += spc_page_cost * baserel->pages;
317 
318  /*
319  * CPU costs (recall that baserel->tuples has already been set to the
320  * number of tuples the sampling method will select). Note that we ignore
321  * execution cost of the TABLESAMPLE parameter expressions; they will be
322  * evaluated only once per scan, and in most usages they'll likely be
323  * simple constants anyway. We also don't charge anything for the
324  * calculations the sampling method might do internally.
325  */
326  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
327 
328  startup_cost += qpqual_cost.startup;
329  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
330  run_cost += cpu_per_tuple * baserel->tuples;
331  /* tlist eval costs are paid per output row, not per tuple scanned */
332  startup_cost += path->pathtarget->cost.startup;
333  run_cost += path->pathtarget->cost.per_tuple * path->rows;
334 
335  path->startup_cost = startup_cost;
336  path->total_cost = startup_cost + run_cost;
337 }
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Oid reltablespace
Definition: relation.h:614
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
#define planner_rt_fetch(rti, root)
Definition: relation.h:328
Cost startup_cost
Definition: relation.h:1053
NextSampleBlock_function NextSampleBlock
Definition: tsmapi.h:72
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
Index relid
Definition: relation.h:613
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
TsmRoutine * GetTsmRoutine(Oid tsmhandler)
Definition: tablesample.c:27
BlockNumber pages
Definition: relation.h:624
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
RTEKind rtekind
Definition: parsenodes.h:951
struct TableSampleClause * tablesample
Definition: parsenodes.h:969
double Cost
Definition: nodes.h:642

◆ cost_seqscan()

void cost_seqscan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 198 of file costsize.c.

References Assert, clamp_row_est(), PathTarget::cost, cpu_tuple_cost, disable_cost, enable_seqscan, get_parallel_divisor(), get_restriction_qual_cost(), get_tablespace_page_costs(), RelOptInfo::pages, Path::parallel_workers, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::reltablespace, RelOptInfo::rows, Path::rows, RTE_RELATION, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_seqscan_path().

200 {
201  Cost startup_cost = 0;
202  Cost cpu_run_cost;
203  Cost disk_run_cost;
204  double spc_seq_page_cost;
205  QualCost qpqual_cost;
206  Cost cpu_per_tuple;
207 
208  /* Should only be applied to base relations */
209  Assert(baserel->relid > 0);
210  Assert(baserel->rtekind == RTE_RELATION);
211 
212  /* Mark the path with the correct row estimate */
213  if (param_info)
214  path->rows = param_info->ppi_rows;
215  else
216  path->rows = baserel->rows;
217 
218  if (!enable_seqscan)
219  startup_cost += disable_cost;
220 
221  /* fetch estimated page cost for tablespace containing table */
223  NULL,
224  &spc_seq_page_cost);
225 
226  /*
227  * disk costs
228  */
229  disk_run_cost = spc_seq_page_cost * baserel->pages;
230 
231  /* CPU costs */
232  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
233 
234  startup_cost += qpqual_cost.startup;
235  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
236  cpu_run_cost = cpu_per_tuple * baserel->tuples;
237  /* tlist eval costs are paid per output row, not per tuple scanned */
238  startup_cost += path->pathtarget->cost.startup;
239  cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
240 
241  /* Adjust costing for parallelism, if used. */
242  if (path->parallel_workers > 0)
243  {
244  double parallel_divisor = get_parallel_divisor(path);
245 
246  /* The CPU cost is divided among all the workers. */
247  cpu_run_cost /= parallel_divisor;
248 
249  /*
250  * It may be possible to amortize some of the I/O cost, but probably
251  * not very much, because most operating systems already do aggressive
252  * prefetching. For now, we assume that the disk run cost can't be
253  * amortized at all.
254  */
255 
256  /*
257  * In the case of a parallel plan, the row count needs to represent
258  * the number of tuples processed per worker.
259  */
260  path->rows = clamp_row_est(path->rows / parallel_divisor);
261  }
262 
263  path->startup_cost = startup_cost;
264  path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
265 }
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Oid reltablespace
Definition: relation.h:614
int parallel_workers
Definition: relation.h:1049
bool enable_seqscan
Definition: costsize.c:118
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5137
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
Index relid
Definition: relation.h:613
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
BlockNumber pages
Definition: relation.h:624
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
double clamp_row_est(double nrows)
Definition: costsize.c:174
double Cost
Definition: nodes.h:642

◆ cost_sort()

void cost_sort ( Path path,
PlannerInfo root,
List pathkeys,
Cost  input_cost,
double  tuples,
int  width,
Cost  comparison_cost,
int  sort_mem,
double  limit_tuples 
)

Definition at line 1645 of file costsize.c.

References cpu_operator_cost, disable_cost, enable_sort, LOG2, random_page_cost, relation_byte_size(), Path::rows, seq_page_cost, Path::startup_cost, Path::total_cost, and tuplesort_merge_order().

Referenced by choose_hashed_setop(), create_gather_merge_path(), create_groupingsets_path(), create_merge_append_path(), create_sort_path(), create_unique_path(), initial_cost_mergejoin(), label_sort_with_costsize(), and plan_cluster_use_sort().

1649 {
1650  Cost startup_cost = input_cost;
1651  Cost run_cost = 0;
1652  double input_bytes = relation_byte_size(tuples, width);
1653  double output_bytes;
1654  double output_tuples;
1655  long sort_mem_bytes = sort_mem * 1024L;
1656 
1657  if (!enable_sort)
1658  startup_cost += disable_cost;
1659 
1660  path->rows = tuples;
1661 
1662  /*
1663  * We want to be sure the cost of a sort is never estimated as zero, even
1664  * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1665  */
1666  if (tuples < 2.0)
1667  tuples = 2.0;
1668 
1669  /* Include the default cost-per-comparison */
1670  comparison_cost += 2.0 * cpu_operator_cost;
1671 
1672  /* Do we have a useful LIMIT? */
1673  if (limit_tuples > 0 && limit_tuples < tuples)
1674  {
1675  output_tuples = limit_tuples;
1676  output_bytes = relation_byte_size(output_tuples, width);
1677  }
1678  else
1679  {
1680  output_tuples = tuples;
1681  output_bytes = input_bytes;
1682  }
1683 
1684  if (output_bytes > sort_mem_bytes)
1685  {
1686  /*
1687  * We'll have to use a disk-based sort of all the tuples
1688  */
1689  double npages = ceil(input_bytes / BLCKSZ);
1690  double nruns = input_bytes / sort_mem_bytes;
1691  double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1692  double log_runs;
1693  double npageaccesses;
1694 
1695  /*
1696  * CPU costs
1697  *
1698  * Assume about N log2 N comparisons
1699  */
1700  startup_cost += comparison_cost * tuples * LOG2(tuples);
1701 
1702  /* Disk costs */
1703 
1704  /* Compute logM(r) as log(r) / log(M) */
1705  if (nruns > mergeorder)
1706  log_runs = ceil(log(nruns) / log(mergeorder));
1707  else
1708  log_runs = 1.0;
1709  npageaccesses = 2.0 * npages * log_runs;
1710  /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1711  startup_cost += npageaccesses *
1712  (seq_page_cost * 0.75 + random_page_cost * 0.25);
1713  }
1714  else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1715  {
1716  /*
1717  * We'll use a bounded heap-sort keeping just K tuples in memory, for
1718  * a total number of tuple comparisons of N log2 K; but the constant
1719  * factor is a bit higher than for quicksort. Tweak it so that the
1720  * cost curve is continuous at the crossover point.
1721  */
1722  startup_cost += comparison_cost * tuples * LOG2(2.0 * output_tuples);
1723  }
1724  else
1725  {
1726  /* We'll use plain quicksort on all the input tuples */
1727  startup_cost += comparison_cost * tuples * LOG2(tuples);
1728  }
1729 
1730  /*
1731  * Also charge a small amount (arbitrarily set equal to operator cost) per
1732  * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1733  * doesn't do qual-checking or projection, so it has less overhead than
1734  * most plan nodes. Note it's correct to use tuples not output_tuples
1735  * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1736  * counting the LIMIT otherwise.
1737  */
1738  run_cost += cpu_operator_cost * tuples;
1739 
1740  path->startup_cost = startup_cost;
1741  path->total_cost = startup_cost + run_cost;
1742 }
bool enable_sort
Definition: costsize.c:123
double random_page_cost
Definition: costsize.c:105
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
double cpu_operator_cost
Definition: costsize.c:108
static double relation_byte_size(double tuples, int width)
Definition: costsize.c:5116
Cost total_cost
Definition: relation.h:1054
#define LOG2(x)
Definition: costsize.c:101
double rows
Definition: relation.h:1052
int tuplesort_merge_order(int64 allowedMem)
Definition: tuplesort.c:2188
double seq_page_cost
Definition: costsize.c:104
double Cost
Definition: nodes.h:642

◆ cost_subplan()

void cost_subplan ( PlannerInfo root,
SubPlan subplan,
Plan plan 
)

Definition at line 3308 of file costsize.c.

References ALL_SUBLINK, ANY_SUBLINK, clamp_row_est(), cost_qual_eval(), cpu_operator_cost, ExecMaterializesOutput(), EXISTS_SUBLINK, make_ands_implicit(), NIL, nodeTag, SubPlan::parParam, SubPlan::per_call_cost, QualCost::per_tuple, Plan::plan_rows, QualCost::startup, Plan::startup_cost, SubPlan::startup_cost, SubPlan::subLinkType, SubPlan::testexpr, Plan::total_cost, and SubPlan::useHashTable.

Referenced by build_subplan(), SS_make_initplan_from_plan(), and SS_process_ctes().

3309 {
3310  QualCost sp_cost;
3311 
3312  /* Figure any cost for evaluating the testexpr */
3313  cost_qual_eval(&sp_cost,
3314  make_ands_implicit((Expr *) subplan->testexpr),
3315  root);
3316 
3317  if (subplan->useHashTable)
3318  {
3319  /*
3320  * If we are using a hash table for the subquery outputs, then the
3321  * cost of evaluating the query is a one-time cost. We charge one
3322  * cpu_operator_cost per tuple for the work of loading the hashtable,
3323  * too.
3324  */
3325  sp_cost.startup += plan->total_cost +
3326  cpu_operator_cost * plan->plan_rows;
3327 
3328  /*
3329  * The per-tuple costs include the cost of evaluating the lefthand
3330  * expressions, plus the cost of probing the hashtable. We already
3331  * accounted for the lefthand expressions as part of the testexpr, and
3332  * will also have counted one cpu_operator_cost for each comparison
3333  * operator. That is probably too low for the probing cost, but it's
3334  * hard to make a better estimate, so live with it for now.
3335  */
3336  }
3337  else
3338  {
3339  /*
3340  * Otherwise we will be rescanning the subplan output on each
3341  * evaluation. We need to estimate how much of the output we will
3342  * actually need to scan. NOTE: this logic should agree with the
3343  * tuple_fraction estimates used by make_subplan() in
3344  * plan/subselect.c.
3345  */
3346  Cost plan_run_cost = plan->total_cost - plan->startup_cost;
3347 
3348  if (subplan->subLinkType == EXISTS_SUBLINK)
3349  {
3350  /* we only need to fetch 1 tuple; clamp to avoid zero divide */
3351  sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
3352  }
3353  else if (subplan->subLinkType == ALL_SUBLINK ||
3354  subplan->subLinkType == ANY_SUBLINK)
3355  {
3356  /* assume we need 50% of the tuples */
3357  sp_cost.per_tuple += 0.50 * plan_run_cost;
3358  /* also charge a cpu_operator_cost per row examined */
3359  sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
3360  }
3361  else
3362  {
3363  /* assume we need all tuples */
3364  sp_cost.per_tuple += plan_run_cost;
3365  }
3366 
3367  /*
3368  * Also account for subplan's startup cost. If the subplan is
3369  * uncorrelated or undirect correlated, AND its topmost node is one
3370  * that materializes its output, assume that we'll only need to pay
3371  * its startup cost once; otherwise assume we pay the startup cost
3372  * every time.
3373  */
3374  if (subplan->parParam == NIL &&
3376  sp_cost.startup += plan->startup_cost;
3377  else
3378  sp_cost.per_tuple += plan->startup_cost;
3379  }
3380 
3381  subplan->startup_cost = sp_cost.startup;
3382  subplan->per_call_cost = sp_cost.per_tuple;
3383 }
#define NIL
Definition: pg_list.h:69
double plan_rows
Definition: plannodes.h:131
SubLinkType subLinkType
Definition: primnodes.h:684
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
List * make_ands_implicit(Expr *clause)
Definition: clauses.c:378
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: plannodes.h:125
double cpu_operator_cost
Definition: costsize.c:108
Node * testexpr
Definition: primnodes.h:686
Cost per_call_cost
Definition: primnodes.h:713
List * parParam
Definition: primnodes.h:709
#define nodeTag(nodeptr)
Definition: nodes.h:516
Cost total_cost
Definition: plannodes.h:126
bool ExecMaterializesOutput(NodeTag plantype)
Definition: execAmi.c:577
bool useHashTable
Definition: primnodes.h:698
Cost startup_cost
Definition: primnodes.h:712
double clamp_row_est(double nrows)
Definition: costsize.c:174
double Cost
Definition: nodes.h:642

◆ cost_subqueryscan()

void cost_subqueryscan ( SubqueryScanPath path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 1268 of file costsize.c.

References Assert, PathTarget::cost, cpu_tuple_cost, get_restriction_qual_cost(), SubqueryScanPath::path, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::rows, Path::rows, RTE_SUBQUERY, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, SubqueryScanPath::subpath, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_subqueryscan_path().

1270 {
1271  Cost startup_cost;
1272  Cost run_cost;
1273  QualCost qpqual_cost;
1274  Cost cpu_per_tuple;
1275 
1276  /* Should only be applied to base relations that are subqueries */
1277  Assert(baserel->relid > 0);
1278  Assert(baserel->rtekind == RTE_SUBQUERY);
1279 
1280  /* Mark the path with the correct row estimate */
1281  if (param_info)
1282  path->path.rows = param_info->ppi_rows;
1283  else
1284  path->path.rows = baserel->rows;
1285 
1286  /*
1287  * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1288  * any restriction clauses and tlist that will be attached to the
1289  * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1290  * projection overhead.
1291  */
1292  path->path.startup_cost = path->subpath->startup_cost;
1293  path->path.total_cost = path->subpath->total_cost;
1294 
1295  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1296 
1297  startup_cost = qpqual_cost.startup;
1298  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1299  run_cost = cpu_per_tuple * baserel->tuples;
1300 
1301  /* tlist eval costs are paid per output row, not per tuple scanned */
1302  startup_cost += path->path.pathtarget->cost.startup;
1303  run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1304 
1305  path->path.startup_cost += startup_cost;
1306  path->path.total_cost += startup_cost + run_cost;
1307 }
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
Index relid
Definition: relation.h:613
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
double Cost
Definition: nodes.h:642

◆ cost_tableexprscan()

void cost_tableexprscan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

◆ cost_tablefuncscan()

void cost_tablefuncscan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 1378 of file costsize.c.

References Assert, PathTarget::cost, cost_qual_eval_node(), cpu_tuple_cost, get_restriction_qual_cost(), Path::pathtarget, QualCost::per_tuple, planner_rt_fetch, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::rows, Path::rows, RTE_TABLEFUNC, RangeTblEntry::rtekind, QualCost::startup, Path::startup_cost, RangeTblEntry::tablefunc, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_tablefuncscan_path().

1380 {
1381  Cost startup_cost = 0;
1382  Cost run_cost = 0;
1383  QualCost qpqual_cost;
1384  Cost cpu_per_tuple;
1385  RangeTblEntry *rte;
1386  QualCost exprcost;
1387 
1388  /* Should only be applied to base relations that are functions */
1389  Assert(baserel->relid > 0);
1390  rte = planner_rt_fetch(baserel->relid, root);
1391  Assert(rte->rtekind == RTE_TABLEFUNC);
1392 
1393  /* Mark the path with the correct row estimate */
1394  if (param_info)
1395  path->rows = param_info->ppi_rows;
1396  else
1397  path->rows = baserel->rows;
1398 
1399  /*
1400  * Estimate costs of executing the table func expression(s).
1401  *
1402  * XXX in principle we ought to charge tuplestore spill costs if the
1403  * number of rows is large. However, given how phony our rowcount
1404  * estimates for tablefuncs tend to be, there's not a lot of point in that
1405  * refinement right now.
1406  */
1407  cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1408 
1409  startup_cost += exprcost.startup + exprcost.per_tuple;
1410 
1411  /* Add scanning CPU costs */
1412  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1413 
1414  startup_cost += qpqual_cost.startup;
1415  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1416  run_cost += cpu_per_tuple * baserel->tuples;
1417 
1418  /* tlist eval costs are paid per output row, not per tuple scanned */
1419  startup_cost += path->pathtarget->cost.startup;
1420  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1421 
1422  path->startup_cost = startup_cost;
1423  path->total_cost = startup_cost + run_cost;
1424 }
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:3534
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Definition: nodes.h:511
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
#define planner_rt_fetch(rti, root)
Definition: relation.h:328
TableFunc * tablefunc
Definition: parsenodes.h:1011
Cost startup_cost
Definition: relation.h:1053
Index relid
Definition: relation.h:613
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
RTEKind rtekind
Definition: parsenodes.h:951
double Cost
Definition: nodes.h:642

◆ cost_tidscan()

void cost_tidscan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
List tidquals,
ParamPathInfo param_info 
)

Definition at line 1165 of file costsize.c.

References ScalarArrayOpExpr::args, Assert, RelOptInfo::baserestrictcost, PathTarget::cost, cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_tidscan, estimate_array_length(), get_restriction_qual_cost(), get_tablespace_page_costs(), IsA, lfirst, lsecond, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::reltablespace, RelOptInfo::rows, Path::rows, RTE_RELATION, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, and Path::total_cost.

Referenced by create_tidscan_path().

1167 {
1168  Cost startup_cost = 0;
1169  Cost run_cost = 0;
1170  bool isCurrentOf = false;
1171  QualCost qpqual_cost;
1172  Cost cpu_per_tuple;
1173  QualCost tid_qual_cost;
1174  int ntuples;
1175  ListCell *l;
1176  double spc_random_page_cost;
1177 
1178  /* Should only be applied to base relations */
1179  Assert(baserel->relid > 0);
1180  Assert(baserel->rtekind == RTE_RELATION);
1181 
1182  /* Mark the path with the correct row estimate */
1183  if (param_info)
1184  path->rows = param_info->ppi_rows;
1185  else
1186  path->rows = baserel->rows;
1187 
1188  /* Count how many tuples we expect to retrieve */
1189  ntuples = 0;
1190  foreach(l, tidquals)
1191  {
1192  if (IsA(lfirst(l), ScalarArrayOpExpr))
1193  {
1194  /* Each element of the array yields 1 tuple */
1196  Node *arraynode = (Node *) lsecond(saop->args);
1197 
1198  ntuples += estimate_array_length(arraynode);
1199  }
1200  else if (IsA(lfirst(l), CurrentOfExpr))
1201  {
1202  /* CURRENT OF yields 1 tuple */
1203  isCurrentOf = true;
1204  ntuples++;
1205  }
1206  else
1207  {
1208  /* It's just CTID = something, count 1 tuple */
1209  ntuples++;
1210  }
1211  }
1212 
1213  /*
1214  * We must force TID scan for WHERE CURRENT OF, because only nodeTidscan.c
1215  * understands how to do it correctly. Therefore, honor enable_tidscan
1216  * only when CURRENT OF isn't present. Also note that cost_qual_eval
1217  * counts a CurrentOfExpr as having startup cost disable_cost, which we
1218  * subtract off here; that's to prevent other plan types such as seqscan
1219  * from winning.
1220  */
1221  if (isCurrentOf)
1222  {
1224  startup_cost -= disable_cost;
1225  }
1226  else if (!enable_tidscan)
1227  startup_cost += disable_cost;
1228 
1229  /*
1230  * The TID qual expressions will be computed once, any other baserestrict
1231  * quals once per retrieved tuple.
1232  */
1233  cost_qual_eval(&tid_qual_cost, tidquals, root);
1234 
1235  /* fetch estimated page cost for tablespace containing table */
1237  &spc_random_page_cost,
1238  NULL);
1239 
1240  /* disk costs --- assume each tuple on a different page */
1241  run_cost += spc_random_page_cost * ntuples;
1242 
1243  /* Add scanning CPU costs */
1244  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1245 
1246  /* XXX currently we assume TID quals are a subset of qpquals */
1247  startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1248  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1249  tid_qual_cost.per_tuple;
1250  run_cost += cpu_per_tuple * ntuples;
1251 
1252  /* tlist eval costs are paid per output row, not per tuple scanned */
1253  startup_cost += path->pathtarget->cost.startup;
1254  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1255 
1256  path->startup_cost = startup_cost;
1257  path->total_cost = startup_cost + run_cost;
1258 }
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
PathTarget * pathtarget
Definition: relation.h:1043
bool enable_tidscan
Definition: costsize.c:122
Oid reltablespace
Definition: relation.h:614
Definition: nodes.h:511
#define lsecond(l)
Definition: pg_list.h:116
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
int estimate_array_length(Node *arrayexpr)
Definition: selfuncs.c:2167
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
Index relid
Definition: relation.h:613
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
#define lfirst(lc)
Definition: pg_list.h:106
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
QualCost baserestrictcost
Definition: relation.h:646
double Cost
Definition: nodes.h:642

◆ cost_valuesscan()

void cost_valuesscan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
ParamPathInfo param_info 
)

Definition at line 1434 of file costsize.c.

References Assert, PathTarget::cost, cpu_operator_cost, cpu_tuple_cost, get_restriction_qual_cost(), Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, RelOptInfo::relid, RelOptInfo::rows, Path::rows, RTE_VALUES, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, Path::total_cost, and RelOptInfo::tuples.

Referenced by create_valuesscan_path().

1436 {
1437  Cost startup_cost = 0;
1438  Cost run_cost = 0;
1439  QualCost qpqual_cost;
1440  Cost cpu_per_tuple;
1441 
1442  /* Should only be applied to base relations that are values lists */
1443  Assert(baserel->relid > 0);
1444  Assert(baserel->rtekind == RTE_VALUES);
1445 
1446  /* Mark the path with the correct row estimate */
1447  if (param_info)
1448  path->rows = param_info->ppi_rows;
1449  else
1450  path->rows = baserel->rows;
1451 
1452  /*
1453  * For now, estimate list evaluation cost at one operator eval per list
1454  * (probably pretty bogus, but is it worth being smarter?)
1455  */
1456  cpu_per_tuple = cpu_operator_cost;
1457 
1458  /* Add scanning CPU costs */
1459  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1460 
1461  startup_cost += qpqual_cost.startup;
1462  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1463  run_cost += cpu_per_tuple * baserel->tuples;
1464 
1465  /* tlist eval costs are paid per output row, not per tuple scanned */
1466  startup_cost += path->pathtarget->cost.startup;
1467  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1468 
1469  path->startup_cost = startup_cost;
1470  path->total_cost = startup_cost + run_cost;
1471 }
PathTarget * pathtarget
Definition: relation.h:1043
double tuples
Definition: relation.h:625
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
double cpu_operator_cost
Definition: costsize.c:108
Index relid
Definition: relation.h:613
RTEKind rtekind
Definition: relation.h:615
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:3787
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
double Cost
Definition: nodes.h:642

◆ cost_windowagg()

void cost_windowagg ( Path path,
PlannerInfo root,
List windowFuncs,
int  numPartCols,
int  numOrderCols,
Cost  input_startup_cost,
Cost  input_total_cost,
double  input_tuples 
)

Definition at line 1992 of file costsize.c.

References WindowFunc::aggfilter, WindowFunc::args, cost_qual_eval_node(), cpu_operator_cost, cpu_tuple_cost, get_func_cost(), lfirst_node, QualCost::per_tuple, Path::rows, QualCost::startup, Path::startup_cost, Path::total_cost, and WindowFunc::winfnoid.

Referenced by create_windowagg_path().

1996 {
1997  Cost startup_cost;
1998  Cost total_cost;
1999  ListCell *lc;
2000 
2001  startup_cost = input_startup_cost;
2002  total_cost = input_total_cost;
2003 
2004  /*
2005  * Window functions are assumed to cost their stated execution cost, plus
2006  * the cost of evaluating their input expressions, per tuple. Since they
2007  * may in fact evaluate their inputs at multiple rows during each cycle,
2008  * this could be a drastic underestimate; but without a way to know how
2009  * many rows the window function will fetch, it's hard to do better. In
2010  * any case, it's a good estimate for all the built-in window functions,
2011  * so we'll just do this for now.
2012  */
2013  foreach(lc, windowFuncs)
2014  {
2015  WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
2016  Cost wfunccost;
2017  QualCost argcosts;
2018 
2019  wfunccost = get_func_cost(wfunc->winfnoid) * cpu_operator_cost;
2020 
2021  /* also add the input expressions' cost to per-input-row costs */
2022  cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
2023  startup_cost += argcosts.startup;
2024  wfunccost += argcosts.per_tuple;
2025 
2026  /*
2027  * Add the filter's cost to per-input-row costs. XXX We should reduce
2028  * input expression costs according to filter selectivity.
2029  */
2030  cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
2031  startup_cost += argcosts.startup;
2032  wfunccost += argcosts.per_tuple;
2033 
2034  total_cost += wfunccost * input_tuples;
2035  }
2036 
2037  /*
2038  * We also charge cpu_operator_cost per grouping column per tuple for
2039  * grouping comparisons, plus cpu_tuple_cost per tuple for general
2040  * overhead.
2041  *
2042  * XXX this neglects costs of spooling the data to disk when it overflows
2043  * work_mem. Sooner or later that should get accounted for.
2044  */
2045  total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
2046  total_cost += cpu_tuple_cost * input_tuples;
2047 
2048  path->rows = input_tuples;
2049  path->startup_cost = startup_cost;
2050  path->total_cost = total_cost;
2051 }
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:3534
List * args
Definition: primnodes.h:359
Definition: nodes.h:511
float4 get_func_cost(Oid funcid)
Definition: lsyscache.c:1641
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
Cost startup_cost
Definition: relation.h:1053
#define lfirst_node(type, lc)
Definition: pg_list.h:109
double cpu_operator_cost
Definition: costsize.c:108
Oid winfnoid
Definition: primnodes.h:355
Cost total_cost
Definition: relation.h:1054
Expr * aggfilter
Definition: primnodes.h:360
double rows
Definition: relation.h:1052
double cpu_tuple_cost
Definition: costsize.c:106
double Cost
Definition: nodes.h:642

◆ final_cost_hashjoin()

void final_cost_hashjoin ( PlannerInfo root,
HashPath path,
JoinCostWorkspace workspace,
JoinPathExtraData extra 
)

Definition at line 3057 of file costsize.c.

References approx_tuple_count(), Assert, bms_is_subset(), clamp_row_est(), RestrictInfo::clause, PathTarget::cost, cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_hashjoin, estimate_hash_bucket_stats(), get_leftop(), get_parallel_divisor(), get_rightop(), JoinPathExtraData::inner_unique, JoinPath::innerjoinpath, IsA, JOIN_ANTI, JOIN_SEMI, JoinPath::joinrestrictinfo, JoinPath::jointype, HashPath::jpath, RestrictInfo::left_bucketsize, RestrictInfo::left_mcvfreq, RestrictInfo::left_relids, lfirst_node, SemiAntiJoinFactors::match_count, HashPath::num_batches, JoinCostWorkspace::numbatches, JoinCostWorkspace::numbuckets, SemiAntiJoinFactors::outer_match_frac, JoinPath::outerjoinpath, Path::parallel_workers, Path::param_info, Path::parent, JoinPath::path, HashPath::path_hashclauses, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, relation_byte_size(), RelOptInfo::relids, RestrictInfo::right_bucketsize, RestrictInfo::right_mcvfreq, RestrictInfo::right_relids, rint(), RelOptInfo::rows, Path::rows, JoinCostWorkspace::run_cost, JoinPathExtraData::semifactors, QualCost::startup, Path::startup_cost, JoinCostWorkspace::startup_cost, Path::total_cost, PathTarget::width, and work_mem.

Referenced by create_hashjoin_path().

3060 {
3061  Path *outer_path = path->jpath.outerjoinpath;
3062  Path *inner_path = path->jpath.innerjoinpath;
3063  double outer_path_rows = outer_path->rows;
3064  double inner_path_rows = inner_path->rows;
3065  List *hashclauses = path->path_hashclauses;
3066  Cost startup_cost = workspace->startup_cost;
3067  Cost run_cost = workspace->run_cost;
3068  int numbuckets = workspace->numbuckets;
3069  int numbatches = workspace->numbatches;
3070  Cost cpu_per_tuple;
3071  QualCost hash_qual_cost;
3072  QualCost qp_qual_cost;
3073  double hashjointuples;
3074  double virtualbuckets;
3075  Selectivity innerbucketsize;
3076  Selectivity innermcvfreq;
3077  ListCell *hcl;
3078 
3079  /* Mark the path with the correct row estimate */
3080  if (path->jpath.path.param_info)
3081  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3082  else
3083  path->jpath.path.rows = path->jpath.path.parent->rows;
3084 
3085  /* For partial paths, scale row estimate. */
3086  if (path->jpath.path.parallel_workers > 0)
3087  {
3088  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3089 
3090  path->jpath.path.rows =
3091  clamp_row_est(path->jpath.path.rows / parallel_divisor);
3092  }
3093 
3094  /*
3095  * We could include disable_cost in the preliminary estimate, but that
3096  * would amount to optimizing for the case where the join method is
3097  * disabled, which doesn't seem like the way to bet.
3098  */
3099  if (!enable_hashjoin)
3100  startup_cost += disable_cost;
3101 
3102  /* mark the path with estimated # of batches */
3103  path->num_batches = numbatches;
3104 
3105  /* and compute the number of "virtual" buckets in the whole join */
3106  virtualbuckets = (double) numbuckets * (double) numbatches;
3107 
3108  /*
3109  * Determine bucketsize fraction and MCV frequency for the inner relation.
3110  * We use the smallest bucketsize or MCV frequency estimated for any
3111  * individual hashclause; this is undoubtedly conservative.
3112  *
3113  * BUT: if inner relation has been unique-ified, we can assume it's good
3114  * for hashing. This is important both because it's the right answer, and
3115  * because we avoid contaminating the cache with a value that's wrong for
3116  * non-unique-ified paths.
3117  */
3118  if (IsA(inner_path, UniquePath))
3119  {
3120  innerbucketsize = 1.0 / virtualbuckets;
3121  innermcvfreq = 0.0;
3122  }
3123  else
3124  {
3125  innerbucketsize = 1.0;
3126  innermcvfreq = 1.0;
3127  foreach(hcl, hashclauses)
3128  {
3129  RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
3130  Selectivity thisbucketsize;
3131  Selectivity thismcvfreq;
3132 
3133  /*
3134  * First we have to figure out which side of the hashjoin clause
3135  * is the inner side.
3136  *
3137  * Since we tend to visit the same clauses over and over when
3138  * planning a large query, we cache the bucket stats estimates in
3139  * the RestrictInfo node to avoid repeated lookups of statistics.
3140  */
3141  if (bms_is_subset(restrictinfo->right_relids,
3142  inner_path->parent->relids))
3143  {
3144  /* righthand side is inner */
3145  thisbucketsize = restrictinfo->right_bucketsize;
3146  if (thisbucketsize < 0)
3147  {
3148  /* not cached yet */
3150  get_rightop(restrictinfo->clause),
3151  virtualbuckets,
3152  &restrictinfo->right_mcvfreq,
3153  &restrictinfo->right_bucketsize);
3154  thisbucketsize = restrictinfo->right_bucketsize;
3155  }
3156  thismcvfreq = restrictinfo->right_mcvfreq;
3157  }
3158  else
3159  {
3160  Assert(bms_is_subset(restrictinfo->left_relids,
3161  inner_path->parent->relids));
3162  /* lefthand side is inner */
3163  thisbucketsize = restrictinfo->left_bucketsize;
3164  if (thisbucketsize < 0)
3165  {
3166  /* not cached yet */
3168  get_leftop(restrictinfo->clause),
3169  virtualbuckets,
3170  &restrictinfo->left_mcvfreq,
3171  &restrictinfo->left_bucketsize);
3172  thisbucketsize = restrictinfo->left_bucketsize;
3173  }
3174  thismcvfreq = restrictinfo->left_mcvfreq;
3175  }
3176 
3177  if (innerbucketsize > thisbucketsize)
3178  innerbucketsize = thisbucketsize;
3179  if (innermcvfreq > thismcvfreq)
3180  innermcvfreq = thismcvfreq;
3181  }
3182  }
3183 
3184  /*
3185  * If the bucket holding the inner MCV would exceed work_mem, we don't
3186  * want to hash unless there is really no other alternative, so apply
3187  * disable_cost. (The executor normally copes with excessive memory usage
3188  * by splitting batches, but obviously it cannot separate equal values
3189  * that way, so it will be unable to drive the batch size below work_mem
3190  * when this is true.)
3191  */
3192  if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
3193  inner_path->pathtarget->width) >
3194  (work_mem * 1024L))
3195  startup_cost += disable_cost;
3196 
3197  /*
3198  * Compute cost of the hashquals and qpquals (other restriction clauses)
3199  * separately.
3200  */
3201  cost_qual_eval(&hash_qual_cost, hashclauses, root);
3202  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3203  qp_qual_cost.startup -= hash_qual_cost.startup;
3204  qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
3205 
3206  /* CPU costs */
3207 
3208  if (path->jpath.jointype == JOIN_SEMI ||
3209  path->jpath.jointype == JOIN_ANTI ||
3210  extra->inner_unique)
3211  {
3212  double outer_matched_rows;
3213  Selectivity inner_scan_frac;
3214 
3215  /*
3216  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3217  * executor will stop after the first match.
3218  *
3219  * For an outer-rel row that has at least one match, we can expect the
3220  * bucket scan to stop after a fraction 1/(match_count+1) of the
3221  * bucket's rows, if the matches are evenly distributed. Since they
3222  * probably aren't quite evenly distributed, we apply a fuzz factor of
3223  * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
3224  * to clamp inner_scan_frac to at most 1.0; but since match_count is
3225  * at least 1, no such clamp is needed now.)
3226  */
3227  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3228  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3229 
3230  startup_cost += hash_qual_cost.startup;
3231  run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
3232  clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
3233 
3234  /*
3235  * For unmatched outer-rel rows, the picture is quite a lot different.
3236  * In the first place, there is no reason to assume that these rows
3237  * preferentially hit heavily-populated buckets; instead assume they
3238  * are uncorrelated with the inner distribution and so they see an
3239  * average bucket size of inner_path_rows / virtualbuckets. In the
3240  * second place, it seems likely that they will have few if any exact
3241  * hash-code matches and so very few of the tuples in the bucket will
3242  * actually require eval of the hash quals. We don't have any good
3243  * way to estimate how many will, but for the moment assume that the
3244  * effective cost per bucket entry is one-tenth what it is for
3245  * matchable tuples.
3246  */
3247  run_cost += hash_qual_cost.per_tuple *
3248  (outer_path_rows - outer_matched_rows) *
3249  clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
3250 
3251  /* Get # of tuples that will pass the basic join */
3252  if (path->jpath.jointype == JOIN_SEMI)
3253  hashjointuples = outer_matched_rows;
3254  else
3255  hashjointuples = outer_path_rows - outer_matched_rows;
3256  }
3257  else
3258  {
3259  /*
3260  * The number of tuple comparisons needed is the number of outer
3261  * tuples times the typical number of tuples in a hash bucket, which
3262  * is the inner relation size times its bucketsize fraction. At each
3263  * one, we need to evaluate the hashjoin quals. But actually,
3264  * charging the full qual eval cost at each tuple is pessimistic,
3265  * since we don't evaluate the quals unless the hash values match
3266  * exactly. For lack of a better idea, halve the cost estimate to
3267  * allow for that.
3268  */
3269  startup_cost += hash_qual_cost.startup;
3270  run_cost += hash_qual_cost.per_tuple * outer_path_rows *
3271  clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
3272 
3273  /*
3274  * Get approx # tuples passing the hashquals. We use
3275  * approx_tuple_count here because we need an estimate done with
3276  * JOIN_INNER semantics.
3277  */
3278  hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
3279  }
3280 
3281  /*
3282  * For each tuple that gets through the hashjoin proper, we charge
3283  * cpu_tuple_cost plus the cost of evaluating additional restriction
3284  * clauses that are to be applied at the join. (This is pessimistic since
3285  * not all of the quals may get evaluated at each tuple.)
3286  */
3287  startup_cost += qp_qual_cost.startup;
3288  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
3289  run_cost += cpu_per_tuple * hashjointuples;
3290 
3291  /* tlist eval costs are paid per output row, not per tuple scanned */
3292  startup_cost += path->jpath.path.pathtarget->cost.startup;
3293  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3294 
3295  path->jpath.path.startup_cost = startup_cost;
3296  path->jpath.path.total_cost = startup_cost + run_cost;
3297 }
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
JoinPath jpath
Definition: relation.h:1458
PathTarget * pathtarget
Definition: relation.h:1043
SemiAntiJoinFactors semifactors
Definition: relation.h:2274
int num_batches
Definition: relation.h:1460
Selectivity right_mcvfreq
Definition: relation.h:1899
Selectivity outer_match_frac
Definition: relation.h:2251
Path * innerjoinpath
Definition: relation.h:1385
static double approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
Definition: costsize.c:4029
int parallel_workers
Definition: relation.h:1049
ParamPathInfo * param_info
Definition: relation.h:1045
Relids left_relids
Definition: relation.h:1862
double Selectivity
Definition: nodes.h:641
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Node * get_leftop(const Expr *clause)
Definition: clauses.c:199
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
List * joinrestrictinfo
Definition: relation.h:1387
RelOptInfo * parent
Definition: relation.h:1042
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:308
#define lfirst_node(type, lc)
Definition: pg_list.h:109
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5137
Relids relids
Definition: relation.h:585
double rint(double x)
Definition: rint.c:22
Expr * clause
Definition: relation.h:1835
static double relation_byte_size(double tuples, int width)
Definition: costsize.c:5116
Path * outerjoinpath
Definition: relation.h:1384
int work_mem
Definition: globals.c:113
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
Selectivity left_bucketsize
Definition: relation.h:1896
Relids right_relids
Definition: relation.h:1863
Path path
Definition: relation.h:1377
#define Assert(condition)
Definition: c.h:670
double rows
Definition: relation.h:1052
Selectivity left_mcvfreq
Definition: relation.h:1898
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
Node * get_rightop(const Expr *clause)
Definition: clauses.c:216
double ppi_rows
Definition: relation.h:1002
bool enable_hashjoin
Definition: costsize.c:128
int width
Definition: relation.h:975
Selectivity match_count
Definition: relation.h:2252
Selectivity right_bucketsize
Definition: relation.h:1897
JoinType jointype
Definition: relation.h:1379
void estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets, Selectivity *mcv_freq, Selectivity *bucketsize_frac)
Definition: selfuncs.c:3696
List * path_hashclauses
Definition: relation.h:1459
double clamp_row_est(double nrows)
Definition: costsize.c:174
Definition: pg_list.h:45
double Cost
Definition: nodes.h:642

◆ final_cost_mergejoin()

void final_cost_mergejoin ( PlannerInfo root,
MergePath path,
JoinCostWorkspace workspace,
JoinPathExtraData extra 
)

Definition at line 2644 of file costsize.c.

References approx_tuple_count(), clamp_row_est(), PathTarget::cost, cost_qual_eval(), cpu_operator_cost, cpu_tuple_cost, disable_cost, enable_material, enable_mergejoin, ExecSupportsMarkRestore(), get_parallel_divisor(), JoinCostWorkspace::inner_rows, JoinCostWorkspace::inner_run_cost, JoinCostWorkspace::inner_skip_rows, JoinPathExtraData::inner_unique, JoinPath::innerjoinpath, MergePath::innersortkeys, IsA, JOIN_ANTI, JOIN_SEMI, JoinPath::joinrestrictinfo, JoinPath::jointype, MergePath::jpath, list_length(), MergePath::materialize_inner, NIL, JoinCostWorkspace::outer_rows, JoinCostWorkspace::outer_skip_rows, JoinPath::outerjoinpath, Path::parallel_workers, Path::param_info, Path::parent, JoinPath::path, MergePath::path_mergeclauses, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, relation_byte_size(), RelOptInfo::rows, Path::rows, JoinCostWorkspace::run_cost, MergePath::skip_mark_restore, QualCost::startup, Path::startup_cost, JoinCostWorkspace::startup_cost, Path::total_cost, PathTarget::width, and work_mem.

Referenced by create_mergejoin_path().

2647 {
2648  Path *outer_path = path->jpath.outerjoinpath;
2649  Path *inner_path = path->jpath.innerjoinpath;
2650  double inner_path_rows = inner_path->rows;
2651  List *mergeclauses = path->path_mergeclauses;
2652  List *innersortkeys = path->innersortkeys;
2653  Cost startup_cost = workspace->startup_cost;
2654  Cost run_cost = workspace->run_cost;
2655  Cost inner_run_cost = workspace->inner_run_cost;
2656  double outer_rows = workspace->outer_rows;
2657  double inner_rows = workspace->inner_rows;
2658  double outer_skip_rows = workspace->outer_skip_rows;
2659  double inner_skip_rows = workspace->inner_skip_rows;
2660  Cost cpu_per_tuple,
2661  bare_inner_cost,
2662  mat_inner_cost;
2663  QualCost merge_qual_cost;
2664  QualCost qp_qual_cost;
2665  double mergejointuples,
2666  rescannedtuples;
2667  double rescanratio;
2668 
2669  /* Protect some assumptions below that rowcounts aren't zero or NaN */
2670  if (inner_path_rows <= 0 || isnan(inner_path_rows))
2671  inner_path_rows = 1;
2672 
2673  /* Mark the path with the correct row estimate */
2674  if (path->jpath.path.param_info)
2675  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
2676  else
2677  path->jpath.path.rows = path->jpath.path.parent->rows;
2678 
2679  /* For partial paths, scale row estimate. */
2680  if (path->jpath.path.parallel_workers > 0)
2681  {
2682  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
2683 
2684  path->jpath.path.rows =
2685  clamp_row_est(path->jpath.path.rows / parallel_divisor);
2686  }
2687 
2688  /*
2689  * We could include disable_cost in the preliminary estimate, but that
2690  * would amount to optimizing for the case where the join method is
2691  * disabled, which doesn't seem like the way to bet.
2692  */
2693  if (!enable_mergejoin)
2694  startup_cost += disable_cost;
2695 
2696  /*
2697  * Compute cost of the mergequals and qpquals (other restriction clauses)
2698  * separately.
2699  */
2700  cost_qual_eval(&merge_qual_cost, mergeclauses, root);
2701  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
2702  qp_qual_cost.startup -= merge_qual_cost.startup;
2703  qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
2704 
2705  /*
2706  * With a SEMI or ANTI join, or if the innerrel is known unique, the
2707  * executor will stop scanning for matches after the first match. When
2708  * all the joinclauses are merge clauses, this means we don't ever need to
2709  * back up the merge, and so we can skip mark/restore overhead.
2710  */
2711  if ((path->jpath.jointype == JOIN_SEMI ||
2712  path->jpath.jointype == JOIN_ANTI ||
2713  extra->inner_unique) &&
2716  path->skip_mark_restore = true;
2717  else
2718  path->skip_mark_restore = false;
2719 
2720  /*
2721  * Get approx # tuples passing the mergequals. We use approx_tuple_count
2722  * here because we need an estimate done with JOIN_INNER semantics.
2723  */
2724  mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
2725 
2726  /*
2727  * When there are equal merge keys in the outer relation, the mergejoin
2728  * must rescan any matching tuples in the inner relation. This means
2729  * re-fetching inner tuples; we have to estimate how often that happens.
2730  *
2731  * For regular inner and outer joins, the number of re-fetches can be
2732  * estimated approximately as size of merge join output minus size of
2733  * inner relation. Assume that the distinct key values are 1, 2, ..., and
2734  * denote the number of values of each key in the outer relation as m1,
2735  * m2, ...; in the inner relation, n1, n2, ... Then we have
2736  *
2737  * size of join = m1 * n1 + m2 * n2 + ...
2738  *
2739  * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
2740  * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
2741  * relation
2742  *
2743  * This equation works correctly for outer tuples having no inner match
2744  * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
2745  * are effectively subtracting those from the number of rescanned tuples,
2746  * when we should not. Can we do better without expensive selectivity
2747  * computations?
2748  *
2749  * The whole issue is moot if we are working from a unique-ified outer
2750  * input, or if we know we don't need to mark/restore at all.
2751  */
2752  if (IsA(outer_path, UniquePath) ||path->skip_mark_restore)
2753  rescannedtuples = 0;
2754  else
2755  {
2756  rescannedtuples = mergejointuples - inner_path_rows;
2757  /* Must clamp because of possible underestimate */
2758  if (rescannedtuples < 0)
2759  rescannedtuples = 0;
2760  }
2761  /* We'll inflate various costs this much to account for rescanning */
2762  rescanratio = 1.0 + (rescannedtuples / inner_path_rows);
2763 
2764  /*
2765  * Decide whether we want to materialize the inner input to shield it from
2766  * mark/restore and performing re-fetches. Our cost model for regular
2767  * re-fetches is that a re-fetch costs the same as an original fetch,
2768  * which is probably an overestimate; but on the other hand we ignore the
2769  * bookkeeping costs of mark/restore. Not clear if it's worth developing
2770  * a more refined model. So we just need to inflate the inner run cost by
2771  * rescanratio.
2772  */
2773  bare_inner_cost = inner_run_cost * rescanratio;
2774 
2775  /*
2776  * When we interpose a Material node the re-fetch cost is assumed to be
2777  * just cpu_operator_cost per tuple, independently of the underlying
2778  * plan's cost; and we charge an extra cpu_operator_cost per original
2779  * fetch as well. Note that we're assuming the materialize node will
2780  * never spill to disk, since it only has to remember tuples back to the
2781  * last mark. (If there are a huge number of duplicates, our other cost
2782  * factors will make the path so expensive that it probably won't get
2783  * chosen anyway.) So we don't use cost_rescan here.
2784  *
2785  * Note: keep this estimate in sync with create_mergejoin_plan's labeling
2786  * of the generated Material node.
2787  */
2788  mat_inner_cost = inner_run_cost +
2789  cpu_operator_cost * inner_path_rows * rescanratio;
2790 
2791  /*
2792  * If we don't need mark/restore at all, we don't need materialization.
2793  */
2794  if (path->skip_mark_restore)
2795  path->materialize_inner = false;
2796 
2797  /*
2798  * Prefer materializing if it looks cheaper, unless the user has asked to
2799  * suppress materialization.
2800  */
2801  else if (enable_material && mat_inner_cost < bare_inner_cost)
2802  path->materialize_inner = true;
2803 
2804  /*
2805  * Even if materializing doesn't look cheaper, we *must* do it if the
2806  * inner path is to be used directly (without sorting) and it doesn't
2807  * support mark/restore.
2808  *
2809  * Since the inner side must be ordered, and only Sorts and IndexScans can
2810  * create order to begin with, and they both support mark/restore, you
2811  * might think there's no problem --- but you'd be wrong. Nestloop and
2812  * merge joins can *preserve* the order of their inputs, so they can be
2813  * selected as the input of a mergejoin, and they don't support
2814  * mark/restore at present.
2815  *
2816  * We don't test the value of enable_material here, because
2817  * materialization is required for correctness in this case, and turning
2818  * it off does not entitle us to deliver an invalid plan.
2819  */
2820  else if (innersortkeys == NIL &&
2821  !ExecSupportsMarkRestore(inner_path))
2822  path->materialize_inner = true;
2823 
2824  /*
2825  * Also, force materializing if the inner path is to be sorted and the
2826  * sort is expected to spill to disk. This is because the final merge
2827  * pass can be done on-the-fly if it doesn't have to support mark/restore.
2828  * We don't try to adjust the cost estimates for this consideration,
2829  * though.
2830  *
2831  * Since materialization is a performance optimization in this case,
2832  * rather than necessary for correctness, we skip it if enable_material is
2833  * off.
2834  */
2835  else if (enable_material && innersortkeys != NIL &&
2836  relation_byte_size(inner_path_rows,
2837  inner_path->pathtarget->width) >
2838  (work_mem * 1024L))
2839  path->materialize_inner = true;
2840  else
2841  path->materialize_inner = false;
2842 
2843  /* Charge the right incremental cost for the chosen case */
2844  if (path->materialize_inner)
2845  run_cost += mat_inner_cost;
2846  else
2847  run_cost += bare_inner_cost;
2848 
2849  /* CPU costs */
2850 
2851  /*
2852  * The number of tuple comparisons needed is approximately number of outer
2853  * rows plus number of inner rows plus number of rescanned tuples (can we
2854  * refine this?). At each one, we need to evaluate the mergejoin quals.
2855  */
2856  startup_cost += merge_qual_cost.startup;
2857  startup_cost += merge_qual_cost.per_tuple *
2858  (outer_skip_rows + inner_skip_rows * rescanratio);
2859  run_cost += merge_qual_cost.per_tuple *
2860  ((outer_rows - outer_skip_rows) +
2861  (inner_rows - inner_skip_rows) * rescanratio);
2862 
2863  /*
2864  * For each tuple that gets through the mergejoin proper, we charge
2865  * cpu_tuple_cost plus the cost of evaluating additional restriction
2866  * clauses that are to be applied at the join. (This is pessimistic since
2867  * not all of the quals may get evaluated at each tuple.)
2868  *
2869  * Note: we could adjust for SEMI/ANTI joins skipping some qual
2870  * evaluations here, but it's probably not worth the trouble.
2871  */
2872  startup_cost += qp_qual_cost.startup;
2873  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
2874  run_cost += cpu_per_tuple * mergejointuples;
2875 
2876  /* tlist eval costs are paid per output row, not per tuple scanned */
2877  startup_cost += path->jpath.path.pathtarget->cost.startup;
2878  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
2879 
2880  path->jpath.path.startup_cost = startup_cost;
2881  path->jpath.path.total_cost = startup_cost + run_cost;
2882 }
#define NIL
Definition: pg_list.h:69
List * path_mergeclauses
Definition: relation.h:1440
#define IsA(nodeptr, _type_)
Definition: nodes.h:562
PathTarget * pathtarget
Definition: relation.h:1043
bool ExecSupportsMarkRestore(Path *pathnode)
Definition: execAmi.c:405
bool materialize_inner
Definition: relation.h:1444
Path * innerjoinpath
Definition: relation.h:1385
static double approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
Definition: costsize.c:4029
int parallel_workers
Definition: relation.h:1049
ParamPathInfo * param_info
Definition: relation.h:1045
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
bool skip_mark_restore
Definition: relation.h:1443
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
List * joinrestrictinfo
Definition: relation.h:1387
RelOptInfo * parent
Definition: relation.h:1042
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5137
double cpu_operator_cost
Definition: costsize.c:108
static double relation_byte_size(double tuples, int width)
Definition: costsize.c:5116
Path * outerjoinpath
Definition: relation.h:1384
int work_mem
Definition: globals.c:113
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
double outer_skip_rows
Definition: relation.h:2306
bool enable_mergejoin
Definition: costsize.c:127
Path path
Definition: relation.h:1377
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
static int list_length(const List *l)
Definition: pg_list.h:89
List * innersortkeys
Definition: relation.h:1442
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
int width
Definition: relation.h:975
JoinType jointype
Definition: relation.h:1379
JoinPath jpath
Definition: relation.h:1439
double inner_skip_rows
Definition: relation.h:2307
double clamp_row_est(double nrows)
Definition: costsize.c:174
Definition: pg_list.h:45
double Cost
Definition: nodes.h:642
bool enable_material
Definition: costsize.c:126

◆ final_cost_nestloop()

void final_cost_nestloop ( PlannerInfo root,
NestPath path,
JoinCostWorkspace workspace,
JoinPathExtraData extra 
)

Definition at line 2207 of file costsize.c.

References clamp_row_est(), PathTarget::cost, cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_nestloop, get_parallel_divisor(), has_indexed_join_quals(), JoinCostWorkspace::inner_rescan_run_cost, JoinCostWorkspace::inner_run_cost, JoinPathExtraData::inner_unique, JoinPath::innerjoinpath, JOIN_ANTI, JOIN_SEMI, JoinPath::joinrestrictinfo, JoinPath::jointype, SemiAntiJoinFactors::match_count, SemiAntiJoinFactors::outer_match_frac, JoinPath::outerjoinpath, Path::parallel_workers, Path::param_info, Path::parent, JoinPath::path, Path::pathtarget, QualCost::per_tuple, ParamPathInfo::ppi_rows, rint(), RelOptInfo::rows, Path::rows, JoinCostWorkspace::run_cost, JoinPathExtraData::semifactors, QualCost::startup, Path::startup_cost, JoinCostWorkspace::startup_cost, and Path::total_cost.

Referenced by create_nestloop_path().

2210 {
2211  Path *outer_path = path->outerjoinpath;
2212  Path *inner_path = path->innerjoinpath;
2213  double outer_path_rows = outer_path->rows;
2214  double inner_path_rows = inner_path->rows;
2215  Cost startup_cost = workspace->startup_cost;
2216  Cost run_cost = workspace->run_cost;
2217  Cost cpu_per_tuple;
2218  QualCost restrict_qual_cost;
2219  double ntuples;
2220 
2221  /* Protect some assumptions below that rowcounts aren't zero or NaN */
2222  if (outer_path_rows <= 0 || isnan(outer_path_rows))
2223  outer_path_rows = 1;
2224  if (inner_path_rows <= 0 || isnan(inner_path_rows))
2225  inner_path_rows = 1;
2226 
2227  /* Mark the path with the correct row estimate */
2228  if (path->path.param_info)
2229  path->path.rows = path->path.param_info->ppi_rows;
2230  else
2231  path->path.rows = path->path.parent->rows;
2232 
2233  /* For partial paths, scale row estimate. */
2234  if (path->path.parallel_workers > 0)
2235  {
2236  double parallel_divisor = get_parallel_divisor(&path->path);
2237 
2238  path->path.rows =
2239  clamp_row_est(path->path.rows / parallel_divisor);
2240  }
2241 
2242  /*
2243  * We could include disable_cost in the preliminary estimate, but that
2244  * would amount to optimizing for the case where the join method is
2245  * disabled, which doesn't seem like the way to bet.
2246  */
2247  if (!enable_nestloop)
2248  startup_cost += disable_cost;
2249 
2250  /* cost of inner-relation source data (we already dealt with outer rel) */
2251 
2252  if (path->jointype == JOIN_SEMI || path->jointype == JOIN_ANTI ||
2253  extra->inner_unique)
2254  {
2255  /*
2256  * With a SEMI or ANTI join, or if the innerrel is known unique, the
2257  * executor will stop after the first match.
2258  */
2259  Cost inner_run_cost = workspace->inner_run_cost;
2260  Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
2261  double outer_matched_rows;
2262  double outer_unmatched_rows;
2263  Selectivity inner_scan_frac;
2264 
2265  /*
2266  * For an outer-rel row that has at least one match, we can expect the
2267  * inner scan to stop after a fraction 1/(match_count+1) of the inner
2268  * rows, if the matches are evenly distributed. Since they probably
2269  * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
2270  * that fraction. (If we used a larger fuzz factor, we'd have to
2271  * clamp inner_scan_frac to at most 1.0; but since match_count is at
2272  * least 1, no such clamp is needed now.)
2273  */
2274  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
2275  outer_unmatched_rows = outer_path_rows - outer_matched_rows;
2276  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
2277 
2278  /*
2279  * Compute number of tuples processed (not number emitted!). First,
2280  * account for successfully-matched outer rows.
2281  */
2282  ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
2283 
2284  /*
2285  * Now we need to estimate the actual costs of scanning the inner
2286  * relation, which may be quite a bit less than N times inner_run_cost
2287  * due to early scan stops. We consider two cases. If the inner path
2288  * is an indexscan using all the joinquals as indexquals, then an
2289  * unmatched outer row results in an indexscan returning no rows,
2290  * which is probably quite cheap. Otherwise, the executor will have
2291  * to scan the whole inner rel for an unmatched row; not so cheap.
2292  */
2293  if (has_indexed_join_quals(path))
2294  {
2295  /*
2296  * Successfully-matched outer rows will only require scanning
2297  * inner_scan_frac of the inner relation. In this case, we don't
2298  * need to charge the full inner_run_cost even when that's more
2299  * than inner_rescan_run_cost, because we can assume that none of
2300  * the inner scans ever scan the whole inner relation. So it's
2301  * okay to assume that all the inner scan executions can be
2302  * fractions of the full cost, even if materialization is reducing
2303  * the rescan cost. At this writing, it's impossible to get here
2304  * for a materialized inner scan, so inner_run_cost and
2305  * inner_rescan_run_cost will be the same anyway; but just in
2306  * case, use inner_run_cost for the first matched tuple and
2307  * inner_rescan_run_cost for additional ones.
2308  */
2309  run_cost += inner_run_cost * inner_scan_frac;
2310  if (outer_matched_rows > 1)
2311  run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
2312 
2313  /*
2314  * Add the cost of inner-scan executions for unmatched outer rows.
2315  * We estimate this as the same cost as returning the first tuple
2316  * of a nonempty scan. We consider that these are all rescans,
2317  * since we used inner_run_cost once already.
2318  */
2319  run_cost += outer_unmatched_rows *
2320  inner_rescan_run_cost / inner_path_rows;
2321 
2322  /*
2323  * We won't be evaluating any quals at all for unmatched rows, so
2324  * don't add them to ntuples.
2325  */
2326  }
2327  else
2328  {
2329  /*
2330  * Here, a complicating factor is that rescans may be cheaper than
2331  * first scans. If we never scan all the way to the end of the
2332  * inner rel, it might be (depending on the plan type) that we'd
2333  * never pay the whole inner first-scan run cost. However it is
2334  * difficult to estimate whether that will happen (and it could
2335  * not happen if there are any unmatched outer rows!), so be
2336  * conservative and always charge the whole first-scan cost once.
2337  * We consider this charge to correspond to the first unmatched
2338  * outer row, unless there isn't one in our estimate, in which
2339  * case blame it on the first matched row.
2340  */
2341 
2342  /* First, count all unmatched join tuples as being processed */
2343  ntuples += outer_unmatched_rows * inner_path_rows;
2344 
2345  /* Now add the forced full scan, and decrement appropriate count */
2346  run_cost += inner_run_cost;
2347  if (outer_unmatched_rows >= 1)
2348  outer_unmatched_rows -= 1;
2349  else
2350  outer_matched_rows -= 1;
2351 
2352  /* Add inner run cost for additional outer tuples having matches */
2353  if (outer_matched_rows > 0)
2354  run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
2355 
2356  /* Add inner run cost for additional unmatched outer tuples */
2357  if (outer_unmatched_rows > 0)
2358  run_cost += outer_unmatched_rows * inner_rescan_run_cost;
2359  }
2360  }
2361  else
2362  {
2363  /* Normal-case source costs were included in preliminary estimate */
2364 
2365  /* Compute number of tuples processed (not number emitted!) */
2366  ntuples = outer_path_rows * inner_path_rows;
2367  }
2368 
2369  /* CPU costs */
2370  cost_qual_eval(&restrict_qual_cost, path->joinrestrictinfo, root);
2371  startup_cost += restrict_qual_cost.startup;
2372  cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
2373  run_cost += cpu_per_tuple * ntuples;
2374 
2375  /* tlist eval costs are paid per output row, not per tuple scanned */
2376  startup_cost += path->path.pathtarget->cost.startup;
2377  run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
2378 
2379  path->path.startup_cost = startup_cost;
2380  path->path.total_cost = startup_cost + run_cost;
2381 }
PathTarget * pathtarget
Definition: relation.h:1043
SemiAntiJoinFactors semifactors
Definition: relation.h:2274
bool enable_nestloop
Definition: costsize.c:125
Selectivity outer_match_frac
Definition: relation.h:2251
Path * innerjoinpath
Definition: relation.h:1385
int parallel_workers
Definition: relation.h:1049
ParamPathInfo * param_info
Definition: relation.h:1045
double Selectivity
Definition: nodes.h:641
Cost inner_rescan_run_cost
Definition: relation.h:2301
Cost startup
Definition: relation.h:45
Cost per_tuple
Definition: relation.h:46
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3508
Cost startup_cost
Definition: relation.h:1053
Cost disable_cost
Definition: costsize.c:114
List * joinrestrictinfo
Definition: relation.h:1387
RelOptInfo * parent
Definition: relation.h:1042
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5137
double rint(double x)
Definition: rint.c:22
Path * outerjoinpath
Definition: relation.h:1384
double rows
Definition: relation.h:588
Cost total_cost
Definition: relation.h:1054
Path path
Definition: relation.h:1377
static bool has_indexed_join_quals(NestPath *joinpath)
Definition: costsize.c:3936
double rows
Definition: relation.h:1052
QualCost cost
Definition: relation.h:974
double cpu_tuple_cost
Definition: costsize.c:106
double ppi_rows
Definition: relation.h:1002
Selectivity match_count
Definition: relation.h:2252
JoinType jointype
Definition: relation.h:1379
double clamp_row_est(double nrows)
Definition: costsize.c:174
double Cost
Definition: nodes.h:642

◆ get_parameterized_baserel_size()

double get_parameterized_baserel_size ( PlannerInfo root,
RelOptInfo rel,
List param_clauses 
)

Definition at line 4115 of file costsize.c.

References RelOptInfo::baserestrictinfo, clamp_row_est(), clauselist_selectivity(), JOIN_INNER, list_concat(), list_copy(), RelOptInfo::relid, RelOptInfo::rows, and RelOptInfo::tuples.

Referenced by get_baserel_parampathinfo().

4117 {
4118  List *allclauses;
4119  double nrows;
4120 
4121  /*
4122  * Estimate the number of rows returned by the parameterized scan, knowing
4123  * that it will apply all the extra join clauses as well as the rel's own
4124  * restriction clauses. Note that we force the clauses to be treated as
4125  * non-join clauses during selectivity estimation.
4126  */
4127  allclauses = list_concat(list_copy(param_clauses),
4128  rel->baserestrictinfo);
4129  nrows = rel->tuples *
4131  allclauses,
4132  rel->relid, /* do not use 0! */
4133  JOIN_INNER,
4134  NULL);
4135  nrows = clamp_row_est(nrows);
4136  /* For safety, make sure result is not more than the base estimate */
4137  if (nrows > rel->rows)
4138  nrows = rel->rows;
4139  return nrows;
4140 }
double tuples
Definition: relation.h:625
List * baserestrictinfo
Definition: relation.h:645
List * list_copy(const List *oldlist)
Definition: list.c:1160
List * list_concat(List *list1, List *list2)
Definition: list.c:321
Index relid
Definition: relation.h:613
double rows
Definition: relation.h:588
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:99
double clamp_row_est(double nrows)
Definition: costsize.c:174
Definition: pg_list.h:45

◆ get_parameterized_joinrel_size()

double get_parameterized_joinrel_size ( PlannerInfo root,
RelOptInfo rel,
Path outer_path,
Path inner_path,
SpecialJoinInfo sjinfo,
List restrict_clauses 
)

Definition at line 4196 of file costsize.c.

References calc_joinrel_size_estimate(), Path::parent, RelOptInfo::rows, and Path::rows.

Referenced by get_joinrel_parampathinfo().

4201 {
4202  double nrows;
4203 
4204  /*
4205  * Estimate the number of rows returned by the parameterized join as the
4206  * sizes of the input paths times the selectivity of the clauses that have
4207  * ended up at this join node.
4208  *
4209  * As with set_joinrel_size_estimates, the rowcount estimate could depend
4210  * on the pair of input paths provided, though ideally we'd get the same
4211  * estimate for any pair with the same parameterization.
4212  */
4213  nrows = calc_joinrel_size_estimate(root,
4214  outer_path->parent,
4215  inner_path->parent,
4216  outer_path->rows,
4217  inner_path->rows,
4218  sjinfo,
4219  restrict_clauses);
4220  /* For safety, make sure result is not more than the base estimate */
4221  if (nrows > rel->rows)
4222  nrows = rel->rows;
4223  return nrows;
4224 }
RelOptInfo * parent
Definition: relation.h:1042
double rows
Definition: relation.h:588
double rows
Definition: relation.h:1052
static double calc_joinrel_size_estimate(PlannerInfo *root, RelOptInfo *outer_rel, RelOptInfo *inner_rel, double outer_rows, double inner_rows, SpecialJoinInfo *sjinfo, List *restrictlist)
Definition: costsize.c:4236

◆ index_pages_fetched()

double index_pages_fetched ( double  tuples_fetched,
BlockNumber  pages,
double  index_pages,
PlannerInfo root 
)

Definition at line 814 of file costsize.c.

References Assert, effective_cache_size, Max, T, and PlannerInfo::total_table_pages.

Referenced by compute_bitmap_pages(), cost_index(), genericcostestimate(), and gincostestimate().

816 {
817  double pages_fetched;
818  double total_pages;
819  double T,
820  b;
821 
822  /* T is # pages in table, but don't allow it to be zero */
823  T = (pages > 1) ? (double) pages : 1.0;
824 
825  /* Compute number of pages assumed to be competing for cache space */
826  total_pages = root->total_table_pages + index_pages;
827  total_pages = Max(total_pages, 1.0);
828  Assert(T <= total_pages);
829 
830  /* b is pro-rated share of effective_cache_size */
831  b = (double) effective_cache_size * T / total_pages;
832 
833  /* force it positive and integral */
834  if (b <= 1.0)
835  b = 1.0;
836  else
837  b = ceil(b);
838 
839  /* This part is the Mackert and Lohman formula */
840  if (T <= b)
841  {
842  pages_fetched =
843  (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
844  if (pages_fetched >= T)
845  pages_fetched = T;
846  else
847  pages_fetched = ceil(pages_fetched);
848  }
849  else
850  {
851  double lim;
852 
853  lim = (2.0 * T * b) / (2.0 * T - b);
854  if (tuples_fetched <= lim)
855  {
856  pages_fetched =
857  (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
858  }
859  else
860  {
861  pages_fetched =
862  b + (tuples_fetched - lim) * (T - b) / T;
863  }
864  pages_fetched = ceil(pages_fetched);
865  }
866  return pages_fetched;
867 }
int effective_cache_size
Definition: costsize.c:112
static const uint32 T[65]
Definition: md5.c:101
double total_table_pages
Definition: relation.h:292
#define Max(x, y)
Definition: c.h:796
#define Assert(condition)
Definition: c.h:670

◆ initial_cost_hashjoin()

void initial_cost_hashjoin ( PlannerInfo root,
JoinCostWorkspace workspace,
JoinType  jointype,
List hashclauses,
Path outer_path,
Path inner_path,
JoinPathExtraData extra 
)

Definition at line 2965 of file costsize.c.

References cpu_operator_cost, cpu_tuple_cost, ExecChooseHashTableSize(), list_length(), JoinCostWorkspace::numbatches, JoinCostWorkspace::numbuckets, page_size(), Path::pathtarget, Path::rows, JoinCostWorkspace::run_cost, seq_page_cost, Path::startup_cost, JoinCostWorkspace::startup_cost, Path::total_cost, JoinCostWorkspace::total_cost, and PathTarget::width.

Referenced by try_hashjoin_path(), and try_partial_hashjoin_path().

2970 {
2971  Cost startup_cost = 0;
2972  Cost run_cost = 0;
2973  double outer_path_rows = outer_path->rows;
2974  double inner_path_rows = inner_path->rows;
2975  int num_hashclauses = list_length(hashclauses);
2976  int numbuckets;
2977  int numbatches;
2978  int num_skew_mcvs;
2979 
2980  /* cost of source data */
2981  startup_cost += outer_path->startup_cost;
2982  run_cost += outer_path->total_cost - outer_path->startup_cost;
2983  startup_cost += inner_path->total_cost;
2984 
2985  /*
2986  * Cost of computing hash function: must do it once per input tuple. We
2987  * charge one cpu_operator_cost for each column's hash function. Also,
2988  * tack on one cpu_tuple_cost per inner row, to model the costs of
2989  * inserting the row into the hashtable.
2990  *
2991  * XXX when a hashclause is more complex than a single operator, we really
2992  * should charge the extra eval costs of the left or right side, as
2993  * appropriate, here. This seems more work than it's worth at the moment.
2994  */
2995  startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
2996  * inner_path_rows;
2997  run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
2998 
2999  /*
3000  * Get hash table size that executor would use for inner relation.
3001  *
3002  * XXX for the moment, always assume that skew optimization will be
3003  * performed. As long as SKEW_WORK_MEM_PERCENT is small, it's not worth
3004  * trying to determine that for sure.
3005  *
3006  * XXX at some point it might be interesting to try to account for skew
3007  * optimization in the cost estimate, but for now, we don't.
3008  */
3009  ExecChooseHashTableSize(inner_path_rows,
3010  inner_path->pathtarget->width,
3011  true, /* useskew */
3012  &numbuckets,
3013  &numbatches,
3014  &num_skew_mcvs);
3015 
3016  /*
3017  * If inner relation is too big then we will need to "batch" the join,
3018  * which implies writing and reading most of the tuples to disk an extra
3019  * time. Charge seq_page_cost per page, since the I/O should be nice and
3020  * sequential. Writing the inner rel counts as startup cost, all the rest
3021  * as run cost.
3022  */
3023  if (numbatches > 1)
3024  {
3025  double outerpages = page_size(outer_path_rows,
3026  outer_path->pathtarget->width);
3027  double innerpages = page_size(inner_path_rows,
3028  inner_path->pathtarget->width);
3029 
3030  startup_cost += seq_page_cost * innerpages;
3031  run_cost += seq_page_cost * (innerpages + 2 * outerpages);
3032  }
3033 
3034  /* CPU costs left for later */
3035 
3036  /* Public result fields */
3037  workspace->startup_cost = startup_cost;
3038  workspace->total_cost = startup_cost + run_cost;
3039  /* Save private data for final_cost_hashjoin */
3040  workspace->run_cost = run_cost;
3041  workspace->numbuckets = numbuckets;
3042  workspace->numbatches = numbatches;
3043 }
PathTarget * pathtarget
Definition: relation.h:1043
static double page_size(double tuples, int width)
Definition: costsize.c:5127
void ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew, int *numbuckets, int *numbatches, int *num_skew_mcvs)
Definition: nodeHash.c:401
Cost startup_cost
Definition: relation.h:1053
double cpu_operator_cost
Definition: costsize.c:108
Cost total_cost
Definition: relation.h:1054
double rows
Definition: relation.h:1052
static int list_length(const List *l)
Definition: pg_list.h:89
double cpu_tuple_cost
Definition: costsize.c:106
int width
Definition: relation.h:975
double seq_page_cost
Definition: costsize.c:104
double Cost
Definition: nodes.h:642

◆ initial_cost_mergejoin()

void initial_cost_mergejoin ( PlannerInfo root,
JoinCostWorkspace workspace,
JoinType  jointype,
List mergeclauses,
Path outer_path,
Path inner_path,
List outersortkeys,
List innersortkeys,
JoinPathExtraData extra 
)

Definition at line 2414 of file costsize.c.

References Assert, bms_is_subset(), cached_scansel(), clamp_row_est(), cost_sort(), EquivalenceClass::ec_collation, elog, ERROR, JoinCostWorkspace::inner_rows, JoinCostWorkspace::inner_run_cost, JoinCostWorkspace::inner_skip_rows, JOIN_ANTI, JOIN_FULL, JOIN_LEFT, JOIN_RIGHT, RestrictInfo::left_relids, MergeScanSelCache::leftendsel, MergeScanSelCache::leftstartsel, linitial, JoinCostWorkspace::outer_rows, JoinCostWorkspace::outer_skip_rows, Path::parent, Path::pathkeys, Path::pathtarget, PathKey::pk_eclass, PathKey::pk_nulls_first, PathKey::pk_opfamily, PathKey::pk_strategy, RelOptInfo::relids, MergeScanSelCache::rightendsel, MergeScanSelCache::rightstartsel, rint(), Path::rows, JoinCostWorkspace::run_cost, Path::startup_cost, JoinCostWorkspace::startup_cost, Path::total_cost, JoinCostWorkspace::total_cost, PathTarget::width, and work_mem.

Referenced by try_mergejoin_path(), and try_partial_mergejoin_path().

2420 {
2421  Cost startup_cost = 0;
2422  Cost run_cost = 0;
2423  double outer_path_rows = outer_path->rows;
2424  double inner_path_rows = inner_path->rows;
2425  Cost inner_run_cost;
2426  double outer_rows,
2427  inner_rows,
2428  outer_skip_rows,
2429  inner_skip_rows;
2430  Selectivity outerstartsel,
2431  outerendsel,
2432  innerstartsel,
2433  innerendsel;
2434  Path sort_path; /* dummy for result of cost_sort */
2435 
2436  /* Protect some assumptions below that rowcounts aren't zero or NaN */
2437  if (outer_path_rows <= 0 || isnan(outer_path_rows))
2438  outer_path_rows = 1;
2439  if (inner_path_rows <= 0 || isnan(inner_path_rows))
2440  inner_path_rows = 1;
2441 
2442  /*
2443  * A merge join will stop as soon as it exhausts either input stream
2444  * (unless it's an outer join, in which case the outer side has to be
2445  * scanned all the way anyway). Estimate fraction of the left and right
2446  * inputs that will actually need to be scanned. Likewise, we can
2447  * estimate the number of rows that will be skipped before the first join
2448  * pair is found, which should be factored into startup cost. We use only
2449  * the first (most significant) merge clause for this purpose. Since
2450  * mergejoinscansel() is a fairly expensive computation, we cache the
2451  * results in the merge clause RestrictInfo.
2452  */
2453  if (mergeclauses && jointype != JOIN_FULL)
2454  {
2455  RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
2456  List *opathkeys;
2457  List *ipathkeys;
2458  PathKey *opathkey;
2459  PathKey *ipathkey;
2460  MergeScanSelCache *cache;
2461 
2462  /* Get the input pathkeys to determine the sort-order details */
2463  opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
2464  ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
2465  Assert(opathkeys);
2466  Assert(ipathkeys);
2467  opathkey = (PathKey *) linitial(opathkeys);
2468  ipathkey = (PathKey *) linitial(ipathkeys);
2469  /* debugging check */
2470  if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
2471  opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
2472  opathkey->pk_strategy != ipathkey->pk_strategy ||
2473  opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
2474  elog(ERROR, "left and right pathkeys do not match in mergejoin");
2475 
2476  /* Get the selectivity with caching */
2477  cache = cached_scansel(root, firstclause, opathkey);
2478 
2479  if (bms_is_subset(firstclause->left_relids,
2480  outer_path->parent->relids))
2481  {
2482  /* left side of clause is outer */
2483  outerstartsel = cache->leftstartsel;
2484  outerendsel = cache->leftendsel;
2485  innerstartsel = cache->rightstartsel;
2486  innerendsel = cache->rightendsel;
2487  }
2488  else
2489  {
2490  /* left side of clause is inner */
2491  outerstartsel = cache->rightstartsel;
2492  outerendsel = cache->rightendsel;
2493  innerstartsel = cache->leftstartsel;
2494  innerendsel = cache->leftendsel;
2495  }
2496  if (jointype == JOIN_LEFT ||
2497  jointype == JOIN_ANTI)
2498  {
2499  outerstartsel = 0.0;
2500  outerendsel = 1.0;
2501  }
2502  else if (jointype == JOIN_RIGHT)
2503  {
2504  innerstartsel = 0.0;
2505  innerendsel = 1.0;
2506  }
2507  }
2508  else
2509  {
2510  /* cope with clauseless or full mergejoin */
2511  outerstartsel = innerstartsel = 0.0;
2512  outerendsel = innerendsel = 1.0;
2513  }
2514 
2515  /*
2516  * Convert selectivities to row counts. We force outer_rows and
2517  * inner_rows to be at least 1, but the skip_rows estimates can be zero.
2518  */
2519  outer_skip_rows = rint(outer_path_rows * outerstartsel);
2520  inner_skip_rows = rint(inner_path_rows * innerstartsel);
2521  outer_rows = clamp_row_est(outer_path_rows * outerendsel);
2522  inner_rows = clamp_row_est(inner_path_rows * innerendsel);
2523 
2524  Assert(outer_skip_rows <= outer_rows);
2525  Assert(inner_skip_rows <= inner_rows);
2526 
2527  /*
2528  * Readjust scan selectivities to account for above rounding. This is
2529  * normally an insignificant effect, but when there are only a few rows in
2530  * the inputs, failing to do this makes for a large percentage error.
2531  */
2532  outerstartsel = outer_skip_rows / outer_path_rows;
2533  innerstartsel = inner_skip_rows / inner_path_rows;
2534  outerendsel = outer_rows / outer_path_rows;
2535  innerendsel = inner_rows / inner_path_rows;
2536 
2537  Assert(outerstartsel <= outerendsel);
2538  Assert(innerstartsel <= innerendsel);
2539 
2540  /* cost of source data */
2541 
2542  if (outersortkeys) /* do we need to sort outer? */
2543  {
2544  cost_sort(&sort_path,
2545  root,
2546  outersortkeys,
2547  outer_path->total_cost,
2548  outer_path_rows,
2549  outer_path->pathtarget->width,
2550  0.0,
2551  work_mem,
2552  -1.0);
2553  startup_cost += sort_path.startup_cost;
2554  startup_cost += (sort_path.total_cost - sort_path.startup_cost)
2555  * outerstartsel;
2556  run_cost += (sort_path.total_cost - sort_path.startup_cost)
2557  * (outerendsel - outerstartsel);
2558  }
2559  else
2560  {
2561  startup_cost += outer_path->startup_cost;
2562  startup_cost += (outer_path->total_cost - outer_path->startup_cost)
2563  * outerstartsel;
2564  run_cost += (outer_path->total_cost - outer_path->startup_cost)
2565  * (outerendsel - outerstartsel);
2566  }
2567 
2568  if (innersortkeys) /* do we need to sort inner? */
2569  {
2570  cost_sort(&sort_path,
2571  root,
2572  innersortkeys,
2573  inner_path->total_cost,
2574  inner_path_rows,
2575  inner_path->pathtarget->width,
2576  0.0,
2577  work_mem,
2578  -1.0);
2579  startup_cost += sort_path.startup_cost;
2580  startup_cost += (sort_path.total_cost - sort_path.startup_cost)
2581  * innerstartsel;
2582  inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
2583  * (innerendsel - innerstartsel);
2584  }
2585  else
2586  {
2587  startup_cost += inner_path->startup_cost;
2588  startup_cost += (inner_path->total_cost - inner_path->startup_cost)
2589  * innerstartsel;
2590  inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
2591  * (innerendsel - innerstartsel);
2592  }
2593 
2594  /*
2595  * We can't yet determine whether rescanning occurs, or whether
2596  * materialization of the inner input should be done. The minimum
2597  * possible inner input cost, regardless of rescan and materialization
2598  * considerations, is inner_run_cost. We include that in
2599  * workspace->total_cost, but not yet in run_cost.
2600  */
2601 
2602  /* CPU costs left for later */
2603 
2604  /* Public result fields */
2605  workspace->startup_cost = startup_cost;
2606  workspace->total_cost = startup_cost + run_cost + inner_run_cost;
2607  /* Save private data for final_cost_mergejoin */
2608  workspace->run_cost = run_cost;
2609  workspace->inner_run_cost = inner_run_cost;
2610  workspace->outer_rows = outer_rows;
2611  workspace->inner_rows = inner_rows;
2612  workspace->outer_skip_rows = outer_skip_rows;
2613  workspace->inner_skip_rows = inner_skip_rows;
2614 }
Selectivity leftendsel
Definition: relation.h:1918
PathTarget * pathtarget
Definition: relation.h:1043
static MergeScanSelCache * cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
Definition: costsize.c:2888
Relids left_relids
Definition: relation.h:1862
double Selectivity
Definition: nodes.h:641
int pk_strategy
Definition: relation.h:941
#define linitial(l)
Definition: pg_list.h:111
bool pk_nulls_first
Definition: relation.h:942
#define ERROR
Definition: elog.h:43
Cost startup_cost
Definition: relation.h:1053
RelOptInfo * parent
Definition: