PostgreSQL Source Code  git master
costsize.c File Reference
#include "postgres.h"
#include <limits.h>
#include <math.h>
#include "access/amapi.h"
#include "access/htup_details.h"
#include "access/tsmapi.h"
#include "executor/executor.h"
#include "executor/nodeAgg.h"
#include "executor/nodeHash.h"
#include "executor/nodeMemoize.h"
#include "miscadmin.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "optimizer/clauses.h"
#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "optimizer/placeholder.h"
#include "optimizer/plancat.h"
#include "optimizer/planmain.h"
#include "optimizer/restrictinfo.h"
#include "parser/parsetree.h"
#include "utils/lsyscache.h"
#include "utils/selfuncs.h"
#include "utils/spccache.h"
#include "utils/tuplesort.h"
Include dependency graph for costsize.c:

Go to the source code of this file.

Data Structures

struct  cost_qual_eval_context
 

Macros

#define LOG2(x)   (log(x) / 0.693147180559945)
 
#define APPEND_CPU_COST_MULTIPLIER   0.5
 
#define MAXIMUM_ROWCOUNT   1e100
 

Functions

static Listextract_nonindex_conditions (List *qual_clauses, List *indexclauses)
 
static MergeScanSelCachecached_scansel (PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
 
static void cost_rescan (PlannerInfo *root, Path *path, Cost *rescan_startup_cost, Cost *rescan_total_cost)
 
static bool cost_qual_eval_walker (Node *node, cost_qual_eval_context *context)
 
static void get_restriction_qual_cost (PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
 
static bool has_indexed_join_quals (NestPath *path)
 
static double approx_tuple_count (PlannerInfo *root, JoinPath *path, List *quals)
 
static double calc_joinrel_size_estimate (PlannerInfo *root, RelOptInfo *joinrel, RelOptInfo *outer_rel, RelOptInfo *inner_rel, double outer_rows, double inner_rows, SpecialJoinInfo *sjinfo, List *restrictlist)
 
static Selectivity get_foreign_key_join_selectivity (PlannerInfo *root, Relids outer_relids, Relids inner_relids, SpecialJoinInfo *sjinfo, List **restrictlist)
 
static Cost append_nonpartial_cost (List *subpaths, int numpaths, int parallel_workers)
 
static void set_rel_width (PlannerInfo *root, RelOptInfo *rel)
 
static int32 get_expr_width (PlannerInfo *root, const Node *expr)
 
static double relation_byte_size (double tuples, int width)
 
static double page_size (double tuples, int width)
 
static double get_parallel_divisor (Path *path)
 
double clamp_row_est (double nrows)
 
long clamp_cardinality_to_long (Cardinality x)
 
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_gather (GatherPath *path, PlannerInfo *root, RelOptInfo *rel, ParamPathInfo *param_info, double *rows)
 
void cost_gather_merge (GatherMergePath *path, PlannerInfo *root, RelOptInfo *rel, ParamPathInfo *param_info, Cost input_startup_cost, Cost input_total_cost, double *rows)
 
void cost_index (IndexPath *path, PlannerInfo *root, double loop_count, bool partial_path)
 
double index_pages_fetched (double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
 
static double get_indexpath_pages (Path *bitmapqual)
 
void cost_bitmap_heap_scan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, Path *bitmapqual, double loop_count)
 
void cost_bitmap_tree_node (Path *path, Cost *cost, Selectivity *selec)
 
void cost_bitmap_and_node (BitmapAndPath *path, PlannerInfo *root)
 
void cost_bitmap_or_node (BitmapOrPath *path, PlannerInfo *root)
 
void cost_tidscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
 
void cost_tidrangescan (Path *path, PlannerInfo *root, RelOptInfo *baserel, List *tidrangequals, ParamPathInfo *param_info)
 
void cost_subqueryscan (SubqueryScanPath *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, bool trivial_pathtarget)
 
void cost_functionscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_tablefuncscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_valuesscan (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_resultscan (Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
 
void cost_recursive_union (Path *runion, Path *nrterm, Path *rterm)
 
static void cost_tuplesort (Cost *startup_cost, Cost *run_cost, double tuples, int width, Cost comparison_cost, int sort_mem, double limit_tuples)
 
void cost_incremental_sort (Path *path, PlannerInfo *root, List *pathkeys, int presorted_keys, Cost input_startup_cost, Cost input_total_cost, double input_tuples, int width, Cost comparison_cost, int sort_mem, double limit_tuples)
 
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_append (AppendPath *apath)
 
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)
 
static void cost_memoize_rescan (PlannerInfo *root, MemoizePath *mpath, Cost *rescan_startup_cost, Cost *rescan_total_cost)
 
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, double input_width)
 
static double get_windowclause_startup_tuples (PlannerInfo *root, WindowClause *wc, double input_tuples)
 
void cost_windowagg (Path *path, PlannerInfo *root, List *windowFuncs, WindowClause *winclause, 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, bool parallel_hash)
 
void final_cost_hashjoin (PlannerInfo *root, HashPath *path, JoinCostWorkspace *workspace, JoinPathExtraData *extra)
 
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 *joinrel, 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)
 
void set_joinrel_size_estimates (PlannerInfo *root, RelOptInfo *rel, RelOptInfo *outer_rel, RelOptInfo *inner_rel, SpecialJoinInfo *sjinfo, List *restrictlist)
 
double get_parameterized_joinrel_size (PlannerInfo *root, RelOptInfo *rel, Path *outer_path, Path *inner_path, SpecialJoinInfo *sjinfo, List *restrict_clauses)
 
void set_subquery_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_function_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_tablefunc_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_namedtuplestore_size_estimates (PlannerInfo *root, RelOptInfo *rel)
 
void set_result_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)
 

Variables

double seq_page_cost = DEFAULT_SEQ_PAGE_COST
 
double random_page_cost = DEFAULT_RANDOM_PAGE_COST
 
double cpu_tuple_cost = DEFAULT_CPU_TUPLE_COST
 
double cpu_index_tuple_cost = DEFAULT_CPU_INDEX_TUPLE_COST
 
double cpu_operator_cost = DEFAULT_CPU_OPERATOR_COST
 
double parallel_tuple_cost = DEFAULT_PARALLEL_TUPLE_COST
 
double parallel_setup_cost = DEFAULT_PARALLEL_SETUP_COST
 
double recursive_worktable_factor = DEFAULT_RECURSIVE_WORKTABLE_FACTOR
 
int effective_cache_size = DEFAULT_EFFECTIVE_CACHE_SIZE
 
Cost disable_cost = 1.0e10
 
int max_parallel_workers_per_gather = 2
 
bool enable_seqscan = true
 
bool enable_indexscan = true
 
bool enable_indexonlyscan = true
 
bool enable_bitmapscan = true
 
bool enable_tidscan = true
 
bool enable_sort = true
 
bool enable_incremental_sort = true
 
bool enable_hashagg = true
 
bool enable_nestloop = true
 
bool enable_material = true
 
bool enable_memoize = true
 
bool enable_mergejoin = true
 
bool enable_hashjoin = true
 
bool enable_gathermerge = true
 
bool enable_partitionwise_join = false
 
bool enable_partitionwise_aggregate = false
 
bool enable_parallel_append = true
 
bool enable_parallel_hash = true
 
bool enable_partition_pruning = true
 
bool enable_presorted_aggregate = true
 
bool enable_async_append = true
 

Macro Definition Documentation

◆ APPEND_CPU_COST_MULTIPLIER

#define APPEND_CPU_COST_MULTIPLIER   0.5

Definition at line 110 of file costsize.c.

◆ LOG2

#define LOG2 (   x)    (log(x) / 0.693147180559945)

Definition at line 103 of file costsize.c.

◆ MAXIMUM_ROWCOUNT

#define MAXIMUM_ROWCOUNT   1e100

Definition at line 118 of file costsize.c.

Function Documentation

◆ append_nonpartial_cost()

static Cost append_nonpartial_cost ( List subpaths,
int  numpaths,
int  parallel_workers 
)
static

Definition at line 2127 of file costsize.c.

2128 {
2129  Cost *costarr;
2130  int arrlen;
2131  ListCell *l;
2132  ListCell *cell;
2133  int path_index;
2134  int min_index;
2135  int max_index;
2136 
2137  if (numpaths == 0)
2138  return 0;
2139 
2140  /*
2141  * Array length is number of workers or number of relevant paths,
2142  * whichever is less.
2143  */
2144  arrlen = Min(parallel_workers, numpaths);
2145  costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2146 
2147  /* The first few paths will each be claimed by a different worker. */
2148  path_index = 0;
2149  foreach(cell, subpaths)
2150  {
2151  Path *subpath = (Path *) lfirst(cell);
2152 
2153  if (path_index == arrlen)
2154  break;
2155  costarr[path_index++] = subpath->total_cost;
2156  }
2157 
2158  /*
2159  * Since subpaths are sorted by decreasing cost, the last one will have
2160  * the minimum cost.
2161  */
2162  min_index = arrlen - 1;
2163 
2164  /*
2165  * For each of the remaining subpaths, add its cost to the array element
2166  * with minimum cost.
2167  */
2168  for_each_cell(l, subpaths, cell)
2169  {
2170  Path *subpath = (Path *) lfirst(l);
2171 
2172  /* Consider only the non-partial paths */
2173  if (path_index++ == numpaths)
2174  break;
2175 
2176  costarr[min_index] += subpath->total_cost;
2177 
2178  /* Update the new min cost array index */
2179  min_index = 0;
2180  for (int i = 0; i < arrlen; i++)
2181  {
2182  if (costarr[i] < costarr[min_index])
2183  min_index = i;
2184  }
2185  }
2186 
2187  /* Return the highest cost from the array */
2188  max_index = 0;
2189  for (int i = 0; i < arrlen; i++)
2190  {
2191  if (costarr[i] > costarr[max_index])
2192  max_index = i;
2193  }
2194 
2195  return costarr[max_index];
2196 }
#define Min(x, y)
Definition: c.h:993
int i
Definition: isn.c:73
Datum subpath(PG_FUNCTION_ARGS)
Definition: ltree_op.c:241
void * palloc(Size size)
Definition: mcxt.c:1226
double Cost
Definition: nodes.h:262
#define lfirst(lc)
Definition: pg_list.h:172
#define for_each_cell(cell, lst, initcell)
Definition: pg_list.h:438

References for_each_cell, i, lfirst, Min, palloc(), and subpath().

Referenced by cost_append().

◆ approx_tuple_count()

static double approx_tuple_count ( PlannerInfo root,
JoinPath path,
List quals 
)
static

Definition at line 5181 of file costsize.c.

5182 {
5183  double tuples;
5184  double outer_tuples = path->outerjoinpath->rows;
5185  double inner_tuples = path->innerjoinpath->rows;
5186  SpecialJoinInfo sjinfo;
5187  Selectivity selec = 1.0;
5188  ListCell *l;
5189 
5190  /*
5191  * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5192  */
5193  sjinfo.type = T_SpecialJoinInfo;
5194  sjinfo.min_lefthand = path->outerjoinpath->parent->relids;
5195  sjinfo.min_righthand = path->innerjoinpath->parent->relids;
5196  sjinfo.syn_lefthand = path->outerjoinpath->parent->relids;
5197  sjinfo.syn_righthand = path->innerjoinpath->parent->relids;
5198  sjinfo.jointype = JOIN_INNER;
5199  sjinfo.ojrelid = 0;
5200  sjinfo.commute_above_l = NULL;
5201  sjinfo.commute_above_r = NULL;
5202  sjinfo.commute_below_l = NULL;
5203  sjinfo.commute_below_r = NULL;
5204  /* we don't bother trying to make the remaining fields valid */
5205  sjinfo.lhs_strict = false;
5206  sjinfo.semi_can_btree = false;
5207  sjinfo.semi_can_hash = false;
5208  sjinfo.semi_operators = NIL;
5209  sjinfo.semi_rhs_exprs = NIL;
5210 
5211  /* Get the approximate selectivity */
5212  foreach(l, quals)
5213  {
5214  Node *qual = (Node *) lfirst(l);
5215 
5216  /* Note that clause_selectivity will be able to cache its result */
5217  selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5218  }
5219 
5220  /* Apply it to the input relation sizes */
5221  tuples = selec * outer_tuples * inner_tuples;
5222 
5223  return clamp_row_est(tuples);
5224 }
Selectivity clause_selectivity(PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:669
double clamp_row_est(double nrows)
Definition: costsize.c:203
double Selectivity
Definition: nodes.h:261
@ JOIN_INNER
Definition: nodes.h:304
#define NIL
Definition: pg_list.h:68
Path * outerjoinpath
Definition: pathnodes.h:2042
Path * innerjoinpath
Definition: pathnodes.h:2043
Definition: nodes.h:129
Cardinality rows
Definition: pathnodes.h:1628
Relids commute_above_r
Definition: pathnodes.h:2860
Relids syn_lefthand
Definition: pathnodes.h:2855
Relids min_righthand
Definition: pathnodes.h:2854
List * semi_rhs_exprs
Definition: pathnodes.h:2868
Relids commute_above_l
Definition: pathnodes.h:2859
JoinType jointype
Definition: pathnodes.h:2857
Relids commute_below_l
Definition: pathnodes.h:2861
Relids min_lefthand
Definition: pathnodes.h:2853
Relids syn_righthand
Definition: pathnodes.h:2856
Relids commute_below_r
Definition: pathnodes.h:2862
List * semi_operators
Definition: pathnodes.h:2867

References clamp_row_est(), clause_selectivity(), SpecialJoinInfo::commute_above_l, SpecialJoinInfo::commute_above_r, SpecialJoinInfo::commute_below_l, SpecialJoinInfo::commute_below_r, JoinPath::innerjoinpath, JOIN_INNER, SpecialJoinInfo::jointype, lfirst, SpecialJoinInfo::lhs_strict, SpecialJoinInfo::min_lefthand, SpecialJoinInfo::min_righthand, NIL, SpecialJoinInfo::ojrelid, JoinPath::outerjoinpath, Path::rows, SpecialJoinInfo::semi_can_btree, SpecialJoinInfo::semi_can_hash, SpecialJoinInfo::semi_operators, SpecialJoinInfo::semi_rhs_exprs, SpecialJoinInfo::syn_lefthand, and SpecialJoinInfo::syn_righthand.

Referenced by final_cost_hashjoin(), and final_cost_mergejoin().

◆ cached_scansel()

static MergeScanSelCache * cached_scansel ( PlannerInfo root,
RestrictInfo rinfo,
PathKey pathkey 
)
static

Definition at line 3966 of file costsize.c.

3967 {
3968  MergeScanSelCache *cache;
3969  ListCell *lc;
3970  Selectivity leftstartsel,
3971  leftendsel,
3972  rightstartsel,
3973  rightendsel;
3974  MemoryContext oldcontext;
3975 
3976  /* Do we have this result already? */
3977  foreach(lc, rinfo->scansel_cache)
3978  {
3979  cache = (MergeScanSelCache *) lfirst(lc);
3980  if (cache->opfamily == pathkey->pk_opfamily &&
3981  cache->collation == pathkey->pk_eclass->ec_collation &&
3982  cache->strategy == pathkey->pk_strategy &&
3983  cache->nulls_first == pathkey->pk_nulls_first)
3984  return cache;
3985  }
3986 
3987  /* Nope, do the computation */
3988  mergejoinscansel(root,
3989  (Node *) rinfo->clause,
3990  pathkey->pk_opfamily,
3991  pathkey->pk_strategy,
3992  pathkey->pk_nulls_first,
3993  &leftstartsel,
3994  &leftendsel,
3995  &rightstartsel,
3996  &rightendsel);
3997 
3998  /* Cache the result in suitably long-lived workspace */
3999  oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4000 
4001  cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4002  cache->opfamily = pathkey->pk_opfamily;
4003  cache->collation = pathkey->pk_eclass->ec_collation;
4004  cache->strategy = pathkey->pk_strategy;
4005  cache->nulls_first = pathkey->pk_nulls_first;
4006  cache->leftstartsel = leftstartsel;
4007  cache->leftendsel = leftendsel;
4008  cache->rightstartsel = rightstartsel;
4009  cache->rightendsel = rightendsel;
4010 
4011  rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4012 
4013  MemoryContextSwitchTo(oldcontext);
4014 
4015  return cache;
4016 }
List * lappend(List *list, void *datum)
Definition: list.c:338
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:138
void mergejoinscansel(PlannerInfo *root, Node *clause, Oid opfamily, int strategy, bool nulls_first, Selectivity *leftstart, Selectivity *leftend, Selectivity *rightstart, Selectivity *rightend)
Definition: selfuncs.c:2922
Selectivity leftstartsel
Definition: pathnodes.h:2702
Selectivity leftendsel
Definition: pathnodes.h:2703
Selectivity rightendsel
Definition: pathnodes.h:2705
Selectivity rightstartsel
Definition: pathnodes.h:2704
bool pk_nulls_first
Definition: pathnodes.h:1456
int pk_strategy
Definition: pathnodes.h:1455
Oid pk_opfamily
Definition: pathnodes.h:1454
Expr * clause
Definition: pathnodes.h:2529

References RestrictInfo::clause, MergeScanSelCache::collation, lappend(), MergeScanSelCache::leftendsel, MergeScanSelCache::leftstartsel, lfirst, MemoryContextSwitchTo(), mergejoinscansel(), MergeScanSelCache::nulls_first, MergeScanSelCache::opfamily, palloc(), PathKey::pk_nulls_first, PathKey::pk_opfamily, PathKey::pk_strategy, MergeScanSelCache::rightendsel, MergeScanSelCache::rightstartsel, and MergeScanSelCache::strategy.

Referenced by initial_cost_mergejoin().

◆ calc_joinrel_size_estimate()

static double calc_joinrel_size_estimate ( PlannerInfo root,
RelOptInfo joinrel,
RelOptInfo outer_rel,
RelOptInfo inner_rel,
double  outer_rows,
double  inner_rows,
SpecialJoinInfo sjinfo,
List restrictlist 
)
static

Definition at line 5393 of file costsize.c.

5401 {
5402  JoinType jointype = sjinfo->jointype;
5403  Selectivity fkselec;
5404  Selectivity jselec;
5405  Selectivity pselec;
5406  double nrows;
5407 
5408  /*
5409  * Compute joinclause selectivity. Note that we are only considering
5410  * clauses that become restriction clauses at this join level; we are not
5411  * double-counting them because they were not considered in estimating the
5412  * sizes of the component rels.
5413  *
5414  * First, see whether any of the joinclauses can be matched to known FK
5415  * constraints. If so, drop those clauses from the restrictlist, and
5416  * instead estimate their selectivity using FK semantics. (We do this
5417  * without regard to whether said clauses are local or "pushed down".
5418  * Probably, an FK-matching clause could never be seen as pushed down at
5419  * an outer join, since it would be strict and hence would be grounds for
5420  * join strength reduction.) fkselec gets the net selectivity for
5421  * FK-matching clauses, or 1.0 if there are none.
5422  */
5423  fkselec = get_foreign_key_join_selectivity(root,
5424  outer_rel->relids,
5425  inner_rel->relids,
5426  sjinfo,
5427  &restrictlist);
5428 
5429  /*
5430  * For an outer join, we have to distinguish the selectivity of the join's
5431  * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5432  * down". For inner joins we just count them all as joinclauses.
5433  */
5434  if (IS_OUTER_JOIN(jointype))
5435  {
5436  List *joinquals = NIL;
5437  List *pushedquals = NIL;
5438  ListCell *l;
5439 
5440  /* Grovel through the clauses to separate into two lists */
5441  foreach(l, restrictlist)
5442  {
5443  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5444 
5445  if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5446  pushedquals = lappend(pushedquals, rinfo);
5447  else
5448  joinquals = lappend(joinquals, rinfo);
5449  }
5450 
5451  /* Get the separate selectivities */
5452  jselec = clauselist_selectivity(root,
5453  joinquals,
5454  0,
5455  jointype,
5456  sjinfo);
5457  pselec = clauselist_selectivity(root,
5458  pushedquals,
5459  0,
5460  jointype,
5461  sjinfo);
5462 
5463  /* Avoid leaking a lot of ListCells */
5464  list_free(joinquals);
5465  list_free(pushedquals);
5466  }
5467  else
5468  {
5469  jselec = clauselist_selectivity(root,
5470  restrictlist,
5471  0,
5472  jointype,
5473  sjinfo);
5474  pselec = 0.0; /* not used, keep compiler quiet */
5475  }
5476 
5477  /*
5478  * Basically, we multiply size of Cartesian product by selectivity.
5479  *
5480  * If we are doing an outer join, take that into account: the joinqual
5481  * selectivity has to be clamped using the knowledge that the output must
5482  * be at least as large as the non-nullable input. However, any
5483  * pushed-down quals are applied after the outer join, so their
5484  * selectivity applies fully.
5485  *
5486  * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5487  * of LHS rows that have matches, and we apply that straightforwardly.
5488  */
5489  switch (jointype)
5490  {
5491  case JOIN_INNER:
5492  nrows = outer_rows * inner_rows * fkselec * jselec;
5493  /* pselec not used */
5494  break;
5495  case JOIN_LEFT:
5496  nrows = outer_rows * inner_rows * fkselec * jselec;
5497  if (nrows < outer_rows)
5498  nrows = outer_rows;
5499  nrows *= pselec;
5500  break;
5501  case JOIN_FULL:
5502  nrows = outer_rows * inner_rows * fkselec * jselec;
5503  if (nrows < outer_rows)
5504  nrows = outer_rows;
5505  if (nrows < inner_rows)
5506  nrows = inner_rows;
5507  nrows *= pselec;
5508  break;
5509  case JOIN_SEMI:
5510  nrows = outer_rows * fkselec * jselec;
5511  /* pselec not used */
5512  break;
5513  case JOIN_ANTI:
5514  nrows = outer_rows * (1.0 - fkselec * jselec);
5515  nrows *= pselec;
5516  break;
5517  default:
5518  /* other values not expected here */
5519  elog(ERROR, "unrecognized join type: %d", (int) jointype);
5520  nrows = 0; /* keep compiler quiet */
5521  break;
5522  }
5523 
5524  return clamp_row_est(nrows);
5525 }
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:102
static Selectivity get_foreign_key_join_selectivity(PlannerInfo *root, Relids outer_relids, Relids inner_relids, SpecialJoinInfo *sjinfo, List **restrictlist)
Definition: costsize.c:5543
#define ERROR
Definition: elog.h:39
void list_free(List *list)
Definition: list.c:1545
#define IS_OUTER_JOIN(jointype)
Definition: nodes.h:348
JoinType
Definition: nodes.h:299
@ JOIN_SEMI
Definition: nodes.h:318
@ JOIN_FULL
Definition: nodes.h:306
@ JOIN_LEFT
Definition: nodes.h:305
@ JOIN_ANTI
Definition: nodes.h:319
#define RINFO_IS_PUSHED_DOWN(rinfo, joinrelids)
Definition: pathnodes.h:2683
#define lfirst_node(type, lc)
Definition: pg_list.h:176
Definition: pg_list.h:54
Relids relids
Definition: pathnodes.h:856

References clamp_row_est(), clauselist_selectivity(), elog(), ERROR, get_foreign_key_join_selectivity(), IS_OUTER_JOIN, JOIN_ANTI, JOIN_FULL, JOIN_INNER, JOIN_LEFT, JOIN_SEMI, SpecialJoinInfo::jointype, lappend(), lfirst_node, list_free(), NIL, RelOptInfo::relids, and RINFO_IS_PUSHED_DOWN.

Referenced by get_parameterized_joinrel_size(), and set_joinrel_size_estimates().

◆ clamp_cardinality_to_long()

long clamp_cardinality_to_long ( Cardinality  x)

Definition at line 226 of file costsize.c.

227 {
228  /*
229  * Just for paranoia's sake, ensure we do something sane with negative or
230  * NaN values.
231  */
232  if (isnan(x))
233  return LONG_MAX;
234  if (x <= 0)
235  return 0;
236 
237  /*
238  * If "long" is 64 bits, then LONG_MAX cannot be represented exactly as a
239  * double. Casting it to double and back may well result in overflow due
240  * to rounding, so avoid doing that. We trust that any double value that
241  * compares strictly less than "(double) LONG_MAX" will cast to a
242  * representable "long" value.
243  */
244  return (x < (double) LONG_MAX) ? (long) x : LONG_MAX;
245 }
int x
Definition: isn.c:71

References x.

Referenced by buildSubPlanHash(), create_recursiveunion_plan(), create_setop_plan(), and make_agg().

◆ clamp_row_est()

double clamp_row_est ( double  nrows)

Definition at line 203 of file costsize.c.

204 {
205  /*
206  * Avoid infinite and NaN row estimates. Costs derived from such values
207  * are going to be useless. Also force the estimate to be at least one
208  * row, to make explain output look better and to avoid possible
209  * divide-by-zero when interpolating costs. Make it an integer, too.
210  */
211  if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
212  nrows = MAXIMUM_ROWCOUNT;
213  else if (nrows <= 1.0)
214  nrows = 1.0;
215  else
216  nrows = rint(nrows);
217 
218  return nrows;
219 }
#define MAXIMUM_ROWCOUNT
Definition: costsize.c:118

References MAXIMUM_ROWCOUNT.

Referenced by adjust_limit_rows_costs(), approx_tuple_count(), bernoulli_samplescangetsamplesize(), calc_joinrel_size_estimate(), compute_bitmap_pages(), cost_agg(), cost_append(), cost_bitmap_heap_scan(), cost_group(), cost_index(), cost_seqscan(), cost_subplan(), cost_subqueryscan(), create_bitmap_subplan(), 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(), get_windowclause_startup_tuples(), initial_cost_mergejoin(), set_baserel_size_estimates(), set_cte_size_estimates(), set_foreign_size(), system_rows_samplescangetsamplesize(), system_samplescangetsamplesize(), and system_time_samplescangetsamplesize().

◆ compute_bitmap_pages()

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

Definition at line 6400 of file costsize.c.

6402 {
6403  Cost indexTotalCost;
6404  Selectivity indexSelectivity;
6405  double T;
6406  double pages_fetched;
6407  double tuples_fetched;
6408  double heap_pages;
6409  long maxentries;
6410 
6411  /*
6412  * Fetch total cost of obtaining the bitmap, as well as its total
6413  * selectivity.
6414  */
6415  cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
6416 
6417  /*
6418  * Estimate number of main-table pages fetched.
6419  */
6420  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
6421 
6422  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
6423 
6424  /*
6425  * For a single scan, the number of heap pages that need to be fetched is
6426  * the same as the Mackert and Lohman formula for the case T <= b (ie, no
6427  * re-reads needed).
6428  */
6429  pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
6430 
6431  /*
6432  * Calculate the number of pages fetched from the heap. Then based on
6433  * current work_mem estimate get the estimated maxentries in the bitmap.
6434  * (Note that we always do this calculation based on the number of pages
6435  * that would be fetched in a single iteration, even if loop_count > 1.
6436  * That's correct, because only that number of entries will be stored in
6437  * the bitmap at one time.)
6438  */
6439  heap_pages = Min(pages_fetched, baserel->pages);
6440  maxentries = tbm_calculate_entries(work_mem * 1024L);
6441 
6442  if (loop_count > 1)
6443  {
6444  /*
6445  * For repeated bitmap scans, scale up the number of tuples fetched in
6446  * the Mackert and Lohman formula by the number of scans, so that we
6447  * estimate the number of pages fetched by all the scans. Then
6448  * pro-rate for one scan.
6449  */
6450  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
6451  baserel->pages,
6452  get_indexpath_pages(bitmapqual),
6453  root);
6454  pages_fetched /= loop_count;
6455  }
6456 
6457  if (pages_fetched >= T)
6458  pages_fetched = T;
6459  else
6460  pages_fetched = ceil(pages_fetched);
6461 
6462  if (maxentries < heap_pages)
6463  {
6464  double exact_pages;
6465  double lossy_pages;
6466 
6467  /*
6468  * Crude approximation of the number of lossy pages. Because of the
6469  * way tbm_lossify() is coded, the number of lossy pages increases
6470  * very sharply as soon as we run short of memory; this formula has
6471  * that property and seems to perform adequately in testing, but it's
6472  * possible we could do better somehow.
6473  */
6474  lossy_pages = Max(0, heap_pages - maxentries / 2);
6475  exact_pages = heap_pages - lossy_pages;
6476 
6477  /*
6478  * If there are lossy pages then recompute the number of tuples
6479  * processed by the bitmap heap node. We assume here that the chance
6480  * of a given tuple coming from an exact page is the same as the
6481  * chance that a given page is exact. This might not be true, but
6482  * it's not clear how we can do any better.
6483  */
6484  if (lossy_pages > 0)
6485  tuples_fetched =
6486  clamp_row_est(indexSelectivity *
6487  (exact_pages / heap_pages) * baserel->tuples +
6488  (lossy_pages / heap_pages) * baserel->tuples);
6489  }
6490 
6491  if (cost)
6492  *cost = indexTotalCost;
6493  if (tuple)
6494  *tuple = tuples_fetched;
6495 
6496  return pages_fetched;
6497 }
#define Max(x, y)
Definition: c.h:987
double index_pages_fetched(double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
Definition: costsize.c:870
void cost_bitmap_tree_node(Path *path, Cost *cost, Selectivity *selec)
Definition: costsize.c:1086
static double get_indexpath_pages(Path *bitmapqual)
Definition: costsize.c:935
int work_mem
Definition: globals.c:125
static const uint32 T[65]
Definition: md5.c:119
Cardinality tuples
Definition: pathnodes.h:928
BlockNumber pages
Definition: pathnodes.h:927
long tbm_calculate_entries(double maxbytes)
Definition: tidbitmap.c:1545

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().

◆ compute_semi_anti_join_factors()

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

Definition at line 4975 of file costsize.c.

4983 {
4984  Selectivity jselec;
4985  Selectivity nselec;
4986  Selectivity avgmatch;
4987  SpecialJoinInfo norm_sjinfo;
4988  List *joinquals;
4989  ListCell *l;
4990 
4991  /*
4992  * In an ANTI join, we must ignore clauses that are "pushed down", since
4993  * those won't affect the match logic. In a SEMI join, we do not
4994  * distinguish joinquals from "pushed down" quals, so just use the whole
4995  * restrictinfo list. For other outer join types, we should consider only
4996  * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
4997  */
4998  if (IS_OUTER_JOIN(jointype))
4999  {
5000  joinquals = NIL;
5001  foreach(l, restrictlist)
5002  {
5003  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5004 
5005  if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5006  joinquals = lappend(joinquals, rinfo);
5007  }
5008  }
5009  else
5010  joinquals = restrictlist;
5011 
5012  /*
5013  * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5014  */
5015  jselec = clauselist_selectivity(root,
5016  joinquals,
5017  0,
5018  (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5019  sjinfo);
5020 
5021  /*
5022  * Also get the normal inner-join selectivity of the join clauses.
5023  */
5024  norm_sjinfo.type = T_SpecialJoinInfo;
5025  norm_sjinfo.min_lefthand = outerrel->relids;
5026  norm_sjinfo.min_righthand = innerrel->relids;
5027  norm_sjinfo.syn_lefthand = outerrel->relids;
5028  norm_sjinfo.syn_righthand = innerrel->relids;
5029  norm_sjinfo.jointype = JOIN_INNER;
5030  norm_sjinfo.ojrelid = 0;
5031  norm_sjinfo.commute_above_l = NULL;
5032  norm_sjinfo.commute_above_r = NULL;
5033  norm_sjinfo.commute_below_l = NULL;
5034  norm_sjinfo.commute_below_r = NULL;
5035  /* we don't bother trying to make the remaining fields valid */
5036  norm_sjinfo.lhs_strict = false;
5037  norm_sjinfo.semi_can_btree = false;
5038  norm_sjinfo.semi_can_hash = false;
5039  norm_sjinfo.semi_operators = NIL;
5040  norm_sjinfo.semi_rhs_exprs = NIL;
5041 
5042  nselec = clauselist_selectivity(root,
5043  joinquals,
5044  0,
5045  JOIN_INNER,
5046  &norm_sjinfo);
5047 
5048  /* Avoid leaking a lot of ListCells */
5049  if (IS_OUTER_JOIN(jointype))
5050  list_free(joinquals);
5051 
5052  /*
5053  * jselec can be interpreted as the fraction of outer-rel rows that have
5054  * any matches (this is true for both SEMI and ANTI cases). And nselec is
5055  * the fraction of the Cartesian product that matches. So, the average
5056  * number of matches for each outer-rel row that has at least one match is
5057  * nselec * inner_rows / jselec.
5058  *
5059  * Note: it is correct to use the inner rel's "rows" count here, even
5060  * though we might later be considering a parameterized inner path with
5061  * fewer rows. This is because we have included all the join clauses in
5062  * the selectivity estimate.
5063  */
5064  if (jselec > 0) /* protect against zero divide */
5065  {
5066  avgmatch = nselec * innerrel->rows / jselec;
5067  /* Clamp to sane range */
5068  avgmatch = Max(1.0, avgmatch);
5069  }
5070  else
5071  avgmatch = 1.0;
5072 
5073  semifactors->outer_match_frac = jselec;
5074  semifactors->match_count = avgmatch;
5075 }
Cardinality rows
Definition: pathnodes.h:862
Selectivity outer_match_frac
Definition: pathnodes.h:3170
Selectivity match_count
Definition: pathnodes.h:3171

References clauselist_selectivity(), SpecialJoinInfo::commute_above_l, SpecialJoinInfo::commute_above_r, SpecialJoinInfo::commute_below_l, SpecialJoinInfo::commute_below_r, IS_OUTER_JOIN, 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, SpecialJoinInfo::ojrelid, SemiAntiJoinFactors::outer_match_frac, RelOptInfo::relids, RINFO_IS_PUSHED_DOWN, RelOptInfo::rows, SpecialJoinInfo::semi_can_btree, SpecialJoinInfo::semi_can_hash, SpecialJoinInfo::semi_operators, SpecialJoinInfo::semi_rhs_exprs, SpecialJoinInfo::syn_lefthand, and SpecialJoinInfo::syn_righthand.

Referenced by add_paths_to_joinrel().

◆ 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,
double  input_width 
)

Definition at line 2622 of file costsize.c.

2628 {
2629  double output_tuples;
2630  Cost startup_cost;
2631  Cost total_cost;
2632  AggClauseCosts dummy_aggcosts;
2633 
2634  /* Use all-zero per-aggregate costs if NULL is passed */
2635  if (aggcosts == NULL)
2636  {
2637  Assert(aggstrategy == AGG_HASHED);
2638  MemSet(&dummy_aggcosts, 0, sizeof(AggClauseCosts));
2639  aggcosts = &dummy_aggcosts;
2640  }
2641 
2642  /*
2643  * The transCost.per_tuple component of aggcosts should be charged once
2644  * per input tuple, corresponding to the costs of evaluating the aggregate
2645  * transfns and their input expressions. The finalCost.per_tuple component
2646  * is charged once per output tuple, corresponding to the costs of
2647  * evaluating the finalfns. Startup costs are of course charged but once.
2648  *
2649  * If we are grouping, we charge an additional cpu_operator_cost per
2650  * grouping column per input tuple for grouping comparisons.
2651  *
2652  * We will produce a single output tuple if not grouping, and a tuple per
2653  * group otherwise. We charge cpu_tuple_cost for each output tuple.
2654  *
2655  * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2656  * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2657  * input path is already sorted appropriately, AGG_SORTED should be
2658  * preferred (since it has no risk of memory overflow). This will happen
2659  * as long as the computed total costs are indeed exactly equal --- but if
2660  * there's roundoff error we might do the wrong thing. So be sure that
2661  * the computations below form the same intermediate values in the same
2662  * order.
2663  */
2664  if (aggstrategy == AGG_PLAIN)
2665  {
2666  startup_cost = input_total_cost;
2667  startup_cost += aggcosts->transCost.startup;
2668  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2669  startup_cost += aggcosts->finalCost.startup;
2670  startup_cost += aggcosts->finalCost.per_tuple;
2671  /* we aren't grouping */
2672  total_cost = startup_cost + cpu_tuple_cost;
2673  output_tuples = 1;
2674  }
2675  else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2676  {
2677  /* Here we are able to deliver output on-the-fly */
2678  startup_cost = input_startup_cost;
2679  total_cost = input_total_cost;
2680  if (aggstrategy == AGG_MIXED && !enable_hashagg)
2681  {
2682  startup_cost += disable_cost;
2683  total_cost += disable_cost;
2684  }
2685  /* calcs phrased this way to match HASHED case, see note above */
2686  total_cost += aggcosts->transCost.startup;
2687  total_cost += aggcosts->transCost.per_tuple * input_tuples;
2688  total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2689  total_cost += aggcosts->finalCost.startup;
2690  total_cost += aggcosts->finalCost.per_tuple * numGroups;
2691  total_cost += cpu_tuple_cost * numGroups;
2692  output_tuples = numGroups;
2693  }
2694  else
2695  {
2696  /* must be AGG_HASHED */
2697  startup_cost = input_total_cost;
2698  if (!enable_hashagg)
2699  startup_cost += disable_cost;
2700  startup_cost += aggcosts->transCost.startup;
2701  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2702  /* cost of computing hash value */
2703  startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2704  startup_cost += aggcosts->finalCost.startup;
2705 
2706  total_cost = startup_cost;
2707  total_cost += aggcosts->finalCost.per_tuple * numGroups;
2708  /* cost of retrieving from hash table */
2709  total_cost += cpu_tuple_cost * numGroups;
2710  output_tuples = numGroups;
2711  }
2712 
2713  /*
2714  * Add the disk costs of hash aggregation that spills to disk.
2715  *
2716  * Groups that go into the hash table stay in memory until finalized, so
2717  * spilling and reprocessing tuples doesn't incur additional invocations
2718  * of transCost or finalCost. Furthermore, the computed hash value is
2719  * stored with the spilled tuples, so we don't incur extra invocations of
2720  * the hash function.
2721  *
2722  * Hash Agg begins returning tuples after the first batch is complete.
2723  * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2724  * accrue reads only to total_cost.
2725  */
2726  if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2727  {
2728  double pages;
2729  double pages_written = 0.0;
2730  double pages_read = 0.0;
2731  double spill_cost;
2732  double hashentrysize;
2733  double nbatches;
2734  Size mem_limit;
2735  uint64 ngroups_limit;
2736  int num_partitions;
2737  int depth;
2738 
2739  /*
2740  * Estimate number of batches based on the computed limits. If less
2741  * than or equal to one, all groups are expected to fit in memory;
2742  * otherwise we expect to spill.
2743  */
2744  hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2745  input_width,
2746  aggcosts->transitionSpace);
2747  hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2748  &ngroups_limit, &num_partitions);
2749 
2750  nbatches = Max((numGroups * hashentrysize) / mem_limit,
2751  numGroups / ngroups_limit);
2752 
2753  nbatches = Max(ceil(nbatches), 1.0);
2754  num_partitions = Max(num_partitions, 2);
2755 
2756  /*
2757  * The number of partitions can change at different levels of
2758  * recursion; but for the purposes of this calculation assume it stays
2759  * constant.
2760  */
2761  depth = ceil(log(nbatches) / log(num_partitions));
2762 
2763  /*
2764  * Estimate number of pages read and written. For each level of
2765  * recursion, a tuple must be written and then later read.
2766  */
2767  pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2768  pages_written = pages_read = pages * depth;
2769 
2770  /*
2771  * HashAgg has somewhat worse IO behavior than Sort on typical
2772  * hardware/OS combinations. Account for this with a generic penalty.
2773  */
2774  pages_read *= 2.0;
2775  pages_written *= 2.0;
2776 
2777  startup_cost += pages_written * random_page_cost;
2778  total_cost += pages_written * random_page_cost;
2779  total_cost += pages_read * seq_page_cost;
2780 
2781  /* account for CPU cost of spilling a tuple and reading it back */
2782  spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2783  startup_cost += spill_cost;
2784  total_cost += spill_cost;
2785  }
2786 
2787  /*
2788  * If there are quals (HAVING quals), account for their cost and
2789  * selectivity.
2790  */
2791  if (quals)
2792  {
2793  QualCost qual_cost;
2794 
2795  cost_qual_eval(&qual_cost, quals, root);
2796  startup_cost += qual_cost.startup;
2797  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2798 
2799  output_tuples = clamp_row_est(output_tuples *
2801  quals,
2802  0,
2803  JOIN_INNER,
2804  NULL));
2805  }
2806 
2807  path->rows = output_tuples;
2808  path->startup_cost = startup_cost;
2809  path->total_cost = total_cost;
2810 }
#define MemSet(start, val, len)
Definition: c.h:1009
size_t Size
Definition: c.h:594
double random_page_cost
Definition: costsize.c:121
double cpu_operator_cost
Definition: costsize.c:124
static double relation_byte_size(double tuples, int width)
Definition: costsize.c:6346
double cpu_tuple_cost
Definition: costsize.c:122
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:4612
double seq_page_cost
Definition: costsize.c:120
bool enable_hashagg
Definition: costsize.c:142
Cost disable_cost
Definition: costsize.c:131
Assert(fmt[strlen(fmt) - 1] !='\n')
Size hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
Definition: nodeAgg.c:1695
void hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits, Size *mem_limit, uint64 *ngroups_limit, int *num_partitions)
Definition: nodeAgg.c:1799
@ AGG_SORTED
Definition: nodes.h:365
@ AGG_HASHED
Definition: nodes.h:366
@ AGG_MIXED
Definition: nodes.h:367
@ AGG_PLAIN
Definition: nodes.h:364
static int list_length(const List *l)
Definition: pg_list.h:152
QualCost finalCost
Definition: pathnodes.h:61
Size transitionSpace
Definition: pathnodes.h:62
QualCost transCost
Definition: pathnodes.h:60
Cost startup_cost
Definition: pathnodes.h:1629
Cost total_cost
Definition: pathnodes.h:1630
List * aggtransinfos
Definition: pathnodes.h:509
Cost per_tuple
Definition: pathnodes.h:48
Cost startup
Definition: pathnodes.h:47

References AGG_HASHED, AGG_MIXED, AGG_PLAIN, AGG_SORTED, PlannerInfo::aggtransinfos, Assert(), clamp_row_est(), clauselist_selectivity(), cost_qual_eval(), cpu_operator_cost, cpu_tuple_cost, disable_cost, enable_hashagg, AggClauseCosts::finalCost, hash_agg_entry_size(), hash_agg_set_limits(), JOIN_INNER, list_length(), Max, MemSet, QualCost::per_tuple, random_page_cost, relation_byte_size(), Path::rows, seq_page_cost, QualCost::startup, Path::startup_cost, Path::total_cost, AggClauseCosts::transCost, and AggClauseCosts::transitionSpace.

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

◆ cost_append()

void cost_append ( AppendPath apath)

Definition at line 2203 of file costsize.c.

2204 {
2205  ListCell *l;
2206 
2207  apath->path.startup_cost = 0;
2208  apath->path.total_cost = 0;
2209  apath->path.rows = 0;
2210 
2211  if (apath->subpaths == NIL)
2212  return;
2213 
2214  if (!apath->path.parallel_aware)
2215  {
2216  List *pathkeys = apath->path.pathkeys;
2217 
2218  if (pathkeys == NIL)
2219  {
2220  Path *firstsubpath = (Path *) linitial(apath->subpaths);
2221 
2222  /*
2223  * For an unordered, non-parallel-aware Append we take the startup
2224  * cost as the startup cost of the first subpath.
2225  */
2226  apath->path.startup_cost = firstsubpath->startup_cost;
2227 
2228  /* Compute rows and costs as sums of subplan rows and costs. */
2229  foreach(l, apath->subpaths)
2230  {
2231  Path *subpath = (Path *) lfirst(l);
2232 
2233  apath->path.rows += subpath->rows;
2234  apath->path.total_cost += subpath->total_cost;
2235  }
2236  }
2237  else
2238  {
2239  /*
2240  * For an ordered, non-parallel-aware Append we take the startup
2241  * cost as the sum of the subpath startup costs. This ensures
2242  * that we don't underestimate the startup cost when a query's
2243  * LIMIT is such that several of the children have to be run to
2244  * satisfy it. This might be overkill --- another plausible hack
2245  * would be to take the Append's startup cost as the maximum of
2246  * the child startup costs. But we don't want to risk believing
2247  * that an ORDER BY LIMIT query can be satisfied at small cost
2248  * when the first child has small startup cost but later ones
2249  * don't. (If we had the ability to deal with nonlinear cost
2250  * interpolation for partial retrievals, we would not need to be
2251  * so conservative about this.)
2252  *
2253  * This case is also different from the above in that we have to
2254  * account for possibly injecting sorts into subpaths that aren't
2255  * natively ordered.
2256  */
2257  foreach(l, apath->subpaths)
2258  {
2259  Path *subpath = (Path *) lfirst(l);
2260  Path sort_path; /* dummy for result of cost_sort */
2261 
2262  if (!pathkeys_contained_in(pathkeys, subpath->pathkeys))
2263  {
2264  /*
2265  * We'll need to insert a Sort node, so include costs for
2266  * that. We can use the parent's LIMIT if any, since we
2267  * certainly won't pull more than that many tuples from
2268  * any child.
2269  */
2270  cost_sort(&sort_path,
2271  NULL, /* doesn't currently need root */
2272  pathkeys,
2273  subpath->total_cost,
2274  subpath->rows,
2275  subpath->pathtarget->width,
2276  0.0,
2277  work_mem,
2278  apath->limit_tuples);
2279  subpath = &sort_path;
2280  }
2281 
2282  apath->path.rows += subpath->rows;
2283  apath->path.startup_cost += subpath->startup_cost;
2284  apath->path.total_cost += subpath->total_cost;
2285  }
2286  }
2287  }
2288  else /* parallel-aware */
2289  {
2290  int i = 0;
2291  double parallel_divisor = get_parallel_divisor(&apath->path);
2292 
2293  /* Parallel-aware Append never produces ordered output. */
2294  Assert(apath->path.pathkeys == NIL);
2295 
2296  /* Calculate startup cost. */
2297  foreach(l, apath->subpaths)
2298  {
2299  Path *subpath = (Path *) lfirst(l);
2300 
2301  /*
2302  * Append will start returning tuples when the child node having
2303  * lowest startup cost is done setting up. We consider only the
2304  * first few subplans that immediately get a worker assigned.
2305  */
2306  if (i == 0)
2307  apath->path.startup_cost = subpath->startup_cost;
2308  else if (i < apath->path.parallel_workers)
2309  apath->path.startup_cost = Min(apath->path.startup_cost,
2310  subpath->startup_cost);
2311 
2312  /*
2313  * Apply parallel divisor to subpaths. Scale the number of rows
2314  * for each partial subpath based on the ratio of the parallel
2315  * divisor originally used for the subpath to the one we adopted.
2316  * Also add the cost of partial paths to the total cost, but
2317  * ignore non-partial paths for now.
2318  */
2319  if (i < apath->first_partial_path)
2320  apath->path.rows += subpath->rows / parallel_divisor;
2321  else
2322  {
2323  double subpath_parallel_divisor;
2324 
2325  subpath_parallel_divisor = get_parallel_divisor(subpath);
2326  apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2327  parallel_divisor);
2328  apath->path.total_cost += subpath->total_cost;
2329  }
2330 
2331  apath->path.rows = clamp_row_est(apath->path.rows);
2332 
2333  i++;
2334  }
2335 
2336  /* Add cost for non-partial subpaths. */
2337  apath->path.total_cost +=
2339  apath->first_partial_path,
2340  apath->path.parallel_workers);
2341  }
2342 
2343  /*
2344  * Although Append does not do any selection or projection, it's not free;
2345  * add a small per-tuple overhead.
2346  */
2347  apath->path.total_cost +=
2349 }
#define APPEND_CPU_COST_MULTIPLIER
Definition: costsize.c:110
static Cost append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
Definition: costsize.c:2127
static double get_parallel_divisor(Path *path)
Definition: costsize.c:6367
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: costsize.c:2096
bool pathkeys_contained_in(List *keys1, List *keys2)
Definition: pathkeys.c:340
#define linitial(l)
Definition: pg_list.h:178
int first_partial_path
Definition: pathnodes.h:1902
Cardinality limit_tuples
Definition: pathnodes.h:1903
List * subpaths
Definition: pathnodes.h:1900
List * pathkeys
Definition: pathnodes.h:1633
int parallel_workers
Definition: pathnodes.h:1625
bool parallel_aware
Definition: pathnodes.h:1621

References APPEND_CPU_COST_MULTIPLIER, append_nonpartial_cost(), Assert(), clamp_row_est(), cost_sort(), cpu_tuple_cost, AppendPath::first_partial_path, get_parallel_divisor(), i, lfirst, AppendPath::limit_tuples, linitial, Min, NIL, Path::parallel_aware, Path::parallel_workers, AppendPath::path, Path::pathkeys, pathkeys_contained_in(), Path::rows, Path::startup_cost, subpath(), AppendPath::subpaths, Path::total_cost, and work_mem.

Referenced by create_append_path().

◆ cost_bitmap_and_node()

void cost_bitmap_and_node ( BitmapAndPath path,
PlannerInfo root 
)

Definition at line 1129 of file costsize.c.

1130 {
1131  Cost totalCost;
1132  Selectivity selec;
1133  ListCell *l;
1134 
1135  /*
1136  * We estimate AND selectivity on the assumption that the inputs are
1137  * independent. This is probably often wrong, but we don't have the info
1138  * to do better.
1139  *
1140  * The runtime cost of the BitmapAnd itself is estimated at 100x
1141  * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1142  * definitely too simplistic?
1143  */
1144  totalCost = 0.0;
1145  selec = 1.0;
1146  foreach(l, path->bitmapquals)
1147  {
1148  Path *subpath = (Path *) lfirst(l);
1149  Cost subCost;
1150  Selectivity subselec;
1151 
1152  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1153 
1154  selec *= subselec;
1155 
1156  totalCost += subCost;
1157  if (l != list_head(path->bitmapquals))
1158  totalCost += 100.0 * cpu_operator_cost;
1159  }
1160  path->bitmapselectivity = selec;
1161  path->path.rows = 0; /* per above, not used */
1162  path->path.startup_cost = totalCost;
1163  path->path.total_cost = totalCost;
1164 }
static ListCell * list_head(const List *l)
Definition: pg_list.h:128
Selectivity bitmapselectivity
Definition: pathnodes.h:1766
List * bitmapquals
Definition: pathnodes.h:1765

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 create_bitmap_and_path().

◆ 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 985 of file costsize.c.

988 {
989  Cost startup_cost = 0;
990  Cost run_cost = 0;
991  Cost indexTotalCost;
992  QualCost qpqual_cost;
993  Cost cpu_per_tuple;
994  Cost cost_per_page;
995  Cost cpu_run_cost;
996  double tuples_fetched;
997  double pages_fetched;
998  double spc_seq_page_cost,
999  spc_random_page_cost;
1000  double T;
1001 
1002  /* Should only be applied to base relations */
1003  Assert(IsA(baserel, RelOptInfo));
1004  Assert(baserel->relid > 0);
1005  Assert(baserel->rtekind == RTE_RELATION);
1006 
1007  /* Mark the path with the correct row estimate */
1008  if (param_info)
1009  path->rows = param_info->ppi_rows;
1010  else
1011  path->rows = baserel->rows;
1012 
1013  if (!enable_bitmapscan)
1014  startup_cost += disable_cost;
1015 
1016  pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1017  loop_count, &indexTotalCost,
1018  &tuples_fetched);
1019 
1020  startup_cost += indexTotalCost;
1021  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1022 
1023  /* Fetch estimated page costs for tablespace containing table. */
1025  &spc_random_page_cost,
1026  &spc_seq_page_cost);
1027 
1028  /*
1029  * For small numbers of pages we should charge spc_random_page_cost
1030  * apiece, while if nearly all the table's pages are being read, it's more
1031  * appropriate to charge spc_seq_page_cost apiece. The effect is
1032  * nonlinear, too. For lack of a better idea, interpolate like this to
1033  * determine the cost per page.
1034  */
1035  if (pages_fetched >= 2.0)
1036  cost_per_page = spc_random_page_cost -
1037  (spc_random_page_cost - spc_seq_page_cost)
1038  * sqrt(pages_fetched / T);
1039  else
1040  cost_per_page = spc_random_page_cost;
1041 
1042  run_cost += pages_fetched * cost_per_page;
1043 
1044  /*
1045  * Estimate CPU costs per tuple.
1046  *
1047  * Often the indexquals don't need to be rechecked at each tuple ... but
1048  * not always, especially not if there are enough tuples involved that the
1049  * bitmaps become lossy. For the moment, just assume they will be
1050  * rechecked always. This means we charge the full freight for all the
1051  * scan clauses.
1052  */
1053  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1054 
1055  startup_cost += qpqual_cost.startup;
1056  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1057  cpu_run_cost = cpu_per_tuple * tuples_fetched;
1058 
1059  /* Adjust costing for parallelism, if used. */
1060  if (path->parallel_workers > 0)
1061  {
1062  double parallel_divisor = get_parallel_divisor(path);
1063 
1064  /* The CPU cost is divided among all the workers. */
1065  cpu_run_cost /= parallel_divisor;
1066 
1067  path->rows = clamp_row_est(path->rows / parallel_divisor);
1068  }
1069 
1070 
1071  run_cost += cpu_run_cost;
1072 
1073  /* tlist eval costs are paid per output row, not per tuple scanned */
1074  startup_cost += path->pathtarget->cost.startup;
1075  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1076 
1077  path->startup_cost = startup_cost;
1078  path->total_cost = startup_cost + run_cost;
1079 }
static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, QualCost *qpqual_cost)
Definition: costsize.c:4933
double compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel, Path *bitmapqual, int loop_count, Cost *cost, double *tuple)
Definition: costsize.c:6400
bool enable_bitmapscan
Definition: costsize.c:138
#define IsA(nodeptr, _type_)
Definition: nodes.h:179
@ RTE_RELATION
Definition: parsenodes.h:1013
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
Cardinality ppi_rows
Definition: pathnodes.h:1548
Index relid
Definition: pathnodes.h:903
Oid reltablespace
Definition: pathnodes.h:905
RTEKind rtekind
Definition: pathnodes.h:907

References Assert(), clamp_row_est(), compute_bitmap_pages(), cpu_tuple_cost, disable_cost, enable_bitmapscan, get_parallel_divisor(), get_restriction_qual_cost(), get_tablespace_page_costs(), IsA, RelOptInfo::pages, Path::parallel_workers, 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_scan_cost_est(), and create_bitmap_heap_path().

◆ cost_bitmap_or_node()

void cost_bitmap_or_node ( BitmapOrPath path,
PlannerInfo root 
)

Definition at line 1173 of file costsize.c.

1174 {
1175  Cost totalCost;
1176  Selectivity selec;
1177  ListCell *l;
1178 
1179  /*
1180  * We estimate OR selectivity on the assumption that the inputs are
1181  * non-overlapping, since that's often the case in "x IN (list)" type
1182  * situations. Of course, we clamp to 1.0 at the end.
1183  *
1184  * The runtime cost of the BitmapOr itself is estimated at 100x
1185  * cpu_operator_cost for each tbm_union needed. Probably too small,
1186  * definitely too simplistic? We are aware that the tbm_unions are
1187  * optimized out when the inputs are BitmapIndexScans.
1188  */
1189  totalCost = 0.0;
1190  selec = 0.0;
1191  foreach(l, path->bitmapquals)
1192  {
1193  Path *subpath = (Path *) lfirst(l);
1194  Cost subCost;
1195  Selectivity subselec;
1196 
1197  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1198 
1199  selec += subselec;
1200 
1201  totalCost += subCost;
1202  if (l != list_head(path->bitmapquals) &&
1203  !IsA(subpath, IndexPath))
1204  totalCost += 100.0 * cpu_operator_cost;
1205  }
1206  path->bitmapselectivity = Min(selec, 1.0);
1207  path->path.rows = 0; /* per above, not used */
1208  path->path.startup_cost = totalCost;
1209  path->path.total_cost = totalCost;
1210 }
Selectivity bitmapselectivity
Definition: pathnodes.h:1779
List * bitmapquals
Definition: pathnodes.h:1778

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().

◆ cost_bitmap_tree_node()

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

Definition at line 1086 of file costsize.c.

1087 {
1088  if (IsA(path, IndexPath))
1089  {
1090  *cost = ((IndexPath *) path)->indextotalcost;
1091  *selec = ((IndexPath *) path)->indexselectivity;
1092 
1093  /*
1094  * Charge a small amount per retrieved tuple to reflect the costs of
1095  * manipulating the bitmap. This is mostly to make sure that a bitmap
1096  * scan doesn't look to be the same cost as an indexscan to retrieve a
1097  * single tuple.
1098  */
1099  *cost += 0.1 * cpu_operator_cost * path->rows;
1100  }
1101  else if (IsA(path, BitmapAndPath))
1102  {
1103  *cost = path->total_cost;
1104  *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1105  }
1106  else if (IsA(path, BitmapOrPath))
1107  {
1108  *cost = path->total_cost;
1109  *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1110  }
1111  else
1112  {
1113  elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1114  *cost = *selec = 0; /* keep compiler quiet */
1115  }
1116 }
#define nodeTag(nodeptr)
Definition: nodes.h:133

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().

◆ cost_ctescan()

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

Definition at line 1670 of file costsize.c.

1672 {
1673  Cost startup_cost = 0;
1674  Cost run_cost = 0;
1675  QualCost qpqual_cost;
1676  Cost cpu_per_tuple;
1677 
1678  /* Should only be applied to base relations that are CTEs */
1679  Assert(baserel->relid > 0);
1680  Assert(baserel->rtekind == RTE_CTE);
1681 
1682  /* Mark the path with the correct row estimate */
1683  if (param_info)
1684  path->rows = param_info->ppi_rows;
1685  else
1686  path->rows = baserel->rows;
1687 
1688  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1689  cpu_per_tuple = cpu_tuple_cost;
1690 
1691  /* Add scanning CPU costs */
1692  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1693 
1694  startup_cost += qpqual_cost.startup;
1695  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1696  run_cost += cpu_per_tuple * baserel->tuples;
1697 
1698  /* tlist eval costs are paid per output row, not per tuple scanned */
1699  startup_cost += path->pathtarget->cost.startup;
1700  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1701 
1702  path->startup_cost = startup_cost;
1703  path->total_cost = startup_cost + run_cost;
1704 }
@ RTE_CTE
Definition: parsenodes.h:1019

References Assert(), cpu_tuple_cost, get_restriction_qual_cost(), 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().

◆ cost_functionscan()

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

Definition at line 1503 of file costsize.c.

1505 {
1506  Cost startup_cost = 0;
1507  Cost run_cost = 0;
1508  QualCost qpqual_cost;
1509  Cost cpu_per_tuple;
1510  RangeTblEntry *rte;
1511  QualCost exprcost;
1512 
1513  /* Should only be applied to base relations that are functions */
1514  Assert(baserel->relid > 0);
1515  rte = planner_rt_fetch(baserel->relid, root);
1516  Assert(rte->rtekind == RTE_FUNCTION);
1517 
1518  /* Mark the path with the correct row estimate */
1519  if (param_info)
1520  path->rows = param_info->ppi_rows;
1521  else
1522  path->rows = baserel->rows;
1523 
1524  /*
1525  * Estimate costs of executing the function expression(s).
1526  *
1527  * Currently, nodeFunctionscan.c always executes the functions to
1528  * completion before returning any rows, and caches the results in a
1529  * tuplestore. So the function eval cost is all startup cost, and per-row
1530  * costs are minimal.
1531  *
1532  * XXX in principle we ought to charge tuplestore spill costs if the
1533  * number of rows is large. However, given how phony our rowcount
1534  * estimates for functions tend to be, there's not a lot of point in that
1535  * refinement right now.
1536  */
1537  cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1538 
1539  startup_cost += exprcost.startup + exprcost.per_tuple;
1540 
1541  /* Add scanning CPU costs */
1542  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1543 
1544  startup_cost += qpqual_cost.startup;
1545  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1546  run_cost += cpu_per_tuple * baserel->tuples;
1547 
1548  /* tlist eval costs are paid per output row, not per tuple scanned */
1549  startup_cost += path->pathtarget->cost.startup;
1550  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1551 
1552  path->startup_cost = startup_cost;
1553  path->total_cost = startup_cost + run_cost;
1554 }
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:4638
@ RTE_FUNCTION
Definition: parsenodes.h:1016
#define planner_rt_fetch(rti, root)
Definition: pathnodes.h:555
List * functions
Definition: parsenodes.h:1147
RTEKind rtekind
Definition: parsenodes.h:1032

References Assert(), cost_qual_eval_node(), cpu_tuple_cost, RangeTblEntry::functions, get_restriction_qual_cost(), 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().

◆ cost_gather()

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

Definition at line 408 of file costsize.c.

411 {
412  Cost startup_cost = 0;
413  Cost run_cost = 0;
414 
415  /* Mark the path with the correct row estimate */
416  if (rows)
417  path->path.rows = *rows;
418  else if (param_info)
419  path->path.rows = param_info->ppi_rows;
420  else
421  path->path.rows = rel->rows;
422 
423  startup_cost = path->subpath->startup_cost;
424 
425  run_cost = path->subpath->total_cost - path->subpath->startup_cost;
426 
427  /* Parallel setup and communication cost. */
428  startup_cost += parallel_setup_cost;
429  run_cost += parallel_tuple_cost * path->path.rows;
430 
431  path->path.startup_cost = startup_cost;
432  path->path.total_cost = (startup_cost + run_cost);
433 }
double parallel_setup_cost
Definition: costsize.c:126
double parallel_tuple_cost
Definition: costsize.c:125
Path * subpath
Definition: pathnodes.h:2010

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().

◆ 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 446 of file costsize.c.

450 {
451  Cost startup_cost = 0;
452  Cost run_cost = 0;
453  Cost comparison_cost;
454  double N;
455  double logN;
456 
457  /* Mark the path with the correct row estimate */
458  if (rows)
459  path->path.rows = *rows;
460  else if (param_info)
461  path->path.rows = param_info->ppi_rows;
462  else
463  path->path.rows = rel->rows;
464 
465  if (!enable_gathermerge)
466  startup_cost += disable_cost;
467 
468  /*
469  * Add one to the number of workers to account for the leader. This might
470  * be overgenerous since the leader will do less work than other workers
471  * in typical cases, but we'll go with it for now.
472  */
473  Assert(path->num_workers > 0);
474  N = (double) path->num_workers + 1;
475  logN = LOG2(N);
476 
477  /* Assumed cost per tuple comparison */
478  comparison_cost = 2.0 * cpu_operator_cost;
479 
480  /* Heap creation cost */
481  startup_cost += comparison_cost * N * logN;
482 
483  /* Per-tuple heap maintenance cost */
484  run_cost += path->path.rows * comparison_cost * logN;
485 
486  /* small cost for heap management, like cost_merge_append */
487  run_cost += cpu_operator_cost * path->path.rows;
488 
489  /*
490  * Parallel setup and communication cost. Since Gather Merge, unlike
491  * Gather, requires us to block until a tuple is available from every
492  * worker, we bump the IPC cost up a little bit as compared with Gather.
493  * For lack of a better idea, charge an extra 5%.
494  */
495  startup_cost += parallel_setup_cost;
496  run_cost += parallel_tuple_cost * path->path.rows * 1.05;
497 
498  path->path.startup_cost = startup_cost + input_startup_cost;
499  path->path.total_cost = (startup_cost + run_cost + input_total_cost);
500 }
#define LOG2(x)
Definition: costsize.c:103
bool enable_gathermerge
Definition: costsize.c:148

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().

◆ 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 3135 of file costsize.c.

3140 {
3141  double output_tuples;
3142  Cost startup_cost;
3143  Cost total_cost;
3144 
3145  output_tuples = numGroups;
3146  startup_cost = input_startup_cost;
3147  total_cost = input_total_cost;
3148 
3149  /*
3150  * Charge one cpu_operator_cost per comparison per input tuple. We assume
3151  * all columns get compared at most of the tuples.
3152  */
3153  total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3154 
3155  /*
3156  * If there are quals (HAVING quals), account for their cost and
3157  * selectivity.
3158  */
3159  if (quals)
3160  {
3161  QualCost qual_cost;
3162 
3163  cost_qual_eval(&qual_cost, quals, root);
3164  startup_cost += qual_cost.startup;
3165  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3166 
3167  output_tuples = clamp_row_est(output_tuples *
3169  quals,
3170  0,
3171  JOIN_INNER,
3172  NULL));
3173  }
3174 
3175  path->rows = output_tuples;
3176  path->startup_cost = startup_cost;
3177  path->total_cost = total_cost;
3178 }

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().

◆ cost_incremental_sort()

void cost_incremental_sort ( Path path,
PlannerInfo root,
List pathkeys,
int  presorted_keys,
Cost  input_startup_cost,
Cost  input_total_cost,
double  input_tuples,
int  width,
Cost  comparison_cost,
int  sort_mem,
double  limit_tuples 
)

Definition at line 1958 of file costsize.c.

1963 {
1964  Cost startup_cost,
1965  run_cost,
1966  input_run_cost = input_total_cost - input_startup_cost;
1967  double group_tuples,
1968  input_groups;
1969  Cost group_startup_cost,
1970  group_run_cost,
1971  group_input_run_cost;
1972  List *presortedExprs = NIL;
1973  ListCell *l;
1974  bool unknown_varno = false;
1975 
1976  Assert(presorted_keys > 0 && presorted_keys < list_length(pathkeys));
1977 
1978  /*
1979  * We want to be sure the cost of a sort is never estimated as zero, even
1980  * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1981  */
1982  if (input_tuples < 2.0)
1983  input_tuples = 2.0;
1984 
1985  /* Default estimate of number of groups, capped to one group per row. */
1986  input_groups = Min(input_tuples, DEFAULT_NUM_DISTINCT);
1987 
1988  /*
1989  * Extract presorted keys as list of expressions.
1990  *
1991  * We need to be careful about Vars containing "varno 0" which might have
1992  * been introduced by generate_append_tlist, which would confuse
1993  * estimate_num_groups (in fact it'd fail for such expressions). See
1994  * recurse_set_operations which has to deal with the same issue.
1995  *
1996  * Unlike recurse_set_operations we can't access the original target list
1997  * here, and even if we could it's not very clear how useful would that be
1998  * for a set operation combining multiple tables. So we simply detect if
1999  * there are any expressions with "varno 0" and use the default
2000  * DEFAULT_NUM_DISTINCT in that case.
2001  *
2002  * We might also use either 1.0 (a single group) or input_tuples (each row
2003  * being a separate group), pretty much the worst and best case for
2004  * incremental sort. But those are extreme cases and using something in
2005  * between seems reasonable. Furthermore, generate_append_tlist is used
2006  * for set operations, which are likely to produce mostly unique output
2007  * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
2008  * while maintaining lower startup cost.
2009  */
2010  foreach(l, pathkeys)
2011  {
2012  PathKey *key = (PathKey *) lfirst(l);
2013  EquivalenceMember *member = (EquivalenceMember *)
2014  linitial(key->pk_eclass->ec_members);
2015 
2016  /*
2017  * Check if the expression contains Var with "varno 0" so that we
2018  * don't call estimate_num_groups in that case.
2019  */
2020  if (bms_is_member(0, pull_varnos(root, (Node *) member->em_expr)))
2021  {
2022  unknown_varno = true;
2023  break;
2024  }
2025 
2026  /* expression not containing any Vars with "varno 0" */
2027  presortedExprs = lappend(presortedExprs, member->em_expr);
2028 
2029  if (foreach_current_index(l) + 1 >= presorted_keys)
2030  break;
2031  }
2032 
2033  /* Estimate the number of groups with equal presorted keys. */
2034  if (!unknown_varno)
2035  input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2036  NULL, NULL);
2037 
2038  group_tuples = input_tuples / input_groups;
2039  group_input_run_cost = input_run_cost / input_groups;
2040 
2041  /*
2042  * Estimate the average cost of sorting of one group where presorted keys
2043  * are equal.
2044  */
2045  cost_tuplesort(&group_startup_cost, &group_run_cost,
2046  group_tuples, width, comparison_cost, sort_mem,
2047  limit_tuples);
2048 
2049  /*
2050  * Startup cost of incremental sort is the startup cost of its first group
2051  * plus the cost of its input.
2052  */
2053  startup_cost = group_startup_cost + input_startup_cost +
2054  group_input_run_cost;
2055 
2056  /*
2057  * After we started producing tuples from the first group, the cost of
2058  * producing all the tuples is given by the cost to finish processing this
2059  * group, plus the total cost to process the remaining groups, plus the
2060  * remaining cost of input.
2061  */
2062  run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2063  (input_groups - 1) + group_input_run_cost * (input_groups - 1);
2064 
2065  /*
2066  * Incremental sort adds some overhead by itself. Firstly, it has to
2067  * detect the sort groups. This is roughly equal to one extra copy and
2068  * comparison per tuple.
2069  */
2070  run_cost += (cpu_tuple_cost + comparison_cost) * input_tuples;
2071 
2072  /*
2073  * Additionally, we charge double cpu_tuple_cost for each input group to
2074  * account for the tuplesort_reset that's performed after each group.
2075  */
2076  run_cost += 2.0 * cpu_tuple_cost * input_groups;
2077 
2078  path->rows = input_tuples;
2079  path->startup_cost = startup_cost;
2080  path->total_cost = startup_cost + run_cost;
2081 }
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:460
static void cost_tuplesort(Cost *startup_cost, Cost *run_cost, double tuples, int width, Cost comparison_cost, int sort_mem, double limit_tuples)
Definition: costsize.c:1856
#define foreach_current_index(cell)
Definition: pg_list.h:403
double estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows, List **pgset, EstimationInfo *estinfo)
Definition: selfuncs.c:3386
#define DEFAULT_NUM_DISTINCT
Definition: selfuncs.h:52
Relids pull_varnos(PlannerInfo *root, Node *node)
Definition: var.c:108

References Assert(), bms_is_member(), cost_tuplesort(), cpu_tuple_cost, DEFAULT_NUM_DISTINCT, EquivalenceMember::em_expr, estimate_num_groups(), foreach_current_index, sort-test::key, lappend(), lfirst, linitial, list_length(), Min, NIL, pull_varnos(), Path::rows, Path::startup_cost, and Path::total_cost.

Referenced by create_incremental_sort_path().

◆ cost_index()

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

Definition at line 521 of file costsize.c.

523 {
524  IndexOptInfo *index = path->indexinfo;
525  RelOptInfo *baserel = index->rel;
526  bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
527  amcostestimate_function amcostestimate;
528  List *qpquals;
529  Cost startup_cost = 0;
530  Cost run_cost = 0;
531  Cost cpu_run_cost = 0;
532  Cost indexStartupCost;
533  Cost indexTotalCost;
534  Selectivity indexSelectivity;
535  double indexCorrelation,
536  csquared;
537  double spc_seq_page_cost,
538  spc_random_page_cost;
539  Cost min_IO_cost,
540  max_IO_cost;
541  QualCost qpqual_cost;
542  Cost cpu_per_tuple;
543  double tuples_fetched;
544  double pages_fetched;
545  double rand_heap_pages;
546  double index_pages;
547 
548  /* Should only be applied to base relations */
549  Assert(IsA(baserel, RelOptInfo) &&
551  Assert(baserel->relid > 0);
552  Assert(baserel->rtekind == RTE_RELATION);
553 
554  /*
555  * Mark the path with the correct row estimate, and identify which quals
556  * will need to be enforced as qpquals. We need not check any quals that
557  * are implied by the index's predicate, so we can use indrestrictinfo not
558  * baserestrictinfo as the list of relevant restriction clauses for the
559  * rel.
560  */
561  if (path->path.param_info)
562  {
563  path->path.rows = path->path.param_info->ppi_rows;
564  /* qpquals come from the rel's restriction clauses and ppi_clauses */
566  path->indexclauses),
567  extract_nonindex_conditions(path->path.param_info->ppi_clauses,
568  path->indexclauses));
569  }
570  else
571  {
572  path->path.rows = baserel->rows;
573  /* qpquals come from just the rel's restriction clauses */
575  path->indexclauses);
576  }
577 
578  if (!enable_indexscan)
579  startup_cost += disable_cost;
580  /* we don't need to check enable_indexonlyscan; indxpath.c does that */
581 
582  /*
583  * Call index-access-method-specific code to estimate the processing cost
584  * for scanning the index, as well as the selectivity of the index (ie,
585  * the fraction of main-table tuples we will have to retrieve) and its
586  * correlation to the main-table tuple order. We need a cast here because
587  * pathnodes.h uses a weak function type to avoid including amapi.h.
588  */
589  amcostestimate = (amcostestimate_function) index->amcostestimate;
590  amcostestimate(root, path, loop_count,
591  &indexStartupCost, &indexTotalCost,
592  &indexSelectivity, &indexCorrelation,
593  &index_pages);
594 
595  /*
596  * Save amcostestimate's results for possible use in bitmap scan planning.
597  * We don't bother to save indexStartupCost or indexCorrelation, because a
598  * bitmap scan doesn't care about either.
599  */
600  path->indextotalcost = indexTotalCost;
601  path->indexselectivity = indexSelectivity;
602 
603  /* all costs for touching index itself included here */
604  startup_cost += indexStartupCost;
605  run_cost += indexTotalCost - indexStartupCost;
606 
607  /* estimate number of main-table tuples fetched */
608  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
609 
610  /* fetch estimated page costs for tablespace containing table */
612  &spc_random_page_cost,
613  &spc_seq_page_cost);
614 
615  /*----------
616  * Estimate number of main-table pages fetched, and compute I/O cost.
617  *
618  * When the index ordering is uncorrelated with the table ordering,
619  * we use an approximation proposed by Mackert and Lohman (see
620  * index_pages_fetched() for details) to compute the number of pages
621  * fetched, and then charge spc_random_page_cost per page fetched.
622  *
623  * When the index ordering is exactly correlated with the table ordering
624  * (just after a CLUSTER, for example), the number of pages fetched should
625  * be exactly selectivity * table_size. What's more, all but the first
626  * will be sequential fetches, not the random fetches that occur in the
627  * uncorrelated case. So if the number of pages is more than 1, we
628  * ought to charge
629  * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
630  * For partially-correlated indexes, we ought to charge somewhere between
631  * these two estimates. We currently interpolate linearly between the
632  * estimates based on the correlation squared (XXX is that appropriate?).
633  *
634  * If it's an index-only scan, then we will not need to fetch any heap
635  * pages for which the visibility map shows all tuples are visible.
636  * Hence, reduce the estimated number of heap fetches accordingly.
637  * We use the measured fraction of the entire heap that is all-visible,
638  * which might not be particularly relevant to the subset of the heap
639  * that this query will fetch; but it's not clear how to do better.
640  *----------
641  */
642  if (loop_count > 1)
643  {
644  /*
645  * For repeated indexscans, the appropriate estimate for the
646  * uncorrelated case is to scale up the number of tuples fetched in
647  * the Mackert and Lohman formula by the number of scans, so that we
648  * estimate the number of pages fetched by all the scans; then
649  * pro-rate the costs for one scan. In this case we assume all the
650  * fetches are random accesses.
651  */
652  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
653  baserel->pages,
654  (double) index->pages,
655  root);
656 
657  if (indexonly)
658  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
659 
660  rand_heap_pages = pages_fetched;
661 
662  max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
663 
664  /*
665  * In the perfectly correlated case, the number of pages touched by
666  * each scan is selectivity * table_size, and we can use the Mackert
667  * and Lohman formula at the page level to estimate how much work is
668  * saved by caching across scans. We still assume all the fetches are
669  * random, though, which is an overestimate that's hard to correct for
670  * without double-counting the cache effects. (But in most cases
671  * where such a plan is actually interesting, only one page would get
672  * fetched per scan anyway, so it shouldn't matter much.)
673  */
674  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
675 
676  pages_fetched = index_pages_fetched(pages_fetched * loop_count,
677  baserel->pages,
678  (double) index->pages,
679  root);
680 
681  if (indexonly)
682  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
683 
684  min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
685  }
686  else
687  {
688  /*
689  * Normal case: apply the Mackert and Lohman formula, and then
690  * interpolate between that and the correlation-derived result.
691  */
692  pages_fetched = index_pages_fetched(tuples_fetched,
693  baserel->pages,
694  (double) index->pages,
695  root);
696 
697  if (indexonly)
698  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
699 
700  rand_heap_pages = pages_fetched;
701 
702  /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
703  max_IO_cost = pages_fetched * spc_random_page_cost;
704 
705  /* min_IO_cost is for the perfectly correlated case (csquared=1) */
706  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
707 
708  if (indexonly)
709  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
710 
711  if (pages_fetched > 0)
712  {
713  min_IO_cost = spc_random_page_cost;
714  if (pages_fetched > 1)
715  min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
716  }
717  else
718  min_IO_cost = 0;
719  }
720 
721  if (partial_path)
722  {
723  /*
724  * For index only scans compute workers based on number of index pages
725  * fetched; the number of heap pages we fetch might be so small as to
726  * effectively rule out parallelism, which we don't want to do.
727  */
728  if (indexonly)
729  rand_heap_pages = -1;
730 
731  /*
732  * Estimate the number of parallel workers required to scan index. Use
733  * the number of heap pages computed considering heap fetches won't be
734  * sequential as for parallel scans the pages are accessed in random
735  * order.
736  */
738  rand_heap_pages,
739  index_pages,
741 
742  /*
743  * Fall out if workers can't be assigned for parallel scan, because in
744  * such a case this path will be rejected. So there is no benefit in
745  * doing extra computation.
746  */
747  if (path->path.parallel_workers <= 0)
748  return;
749 
750  path->path.parallel_aware = true;
751  }
752 
753  /*
754  * Now interpolate based on estimated index order correlation to get total
755  * disk I/O cost for main table accesses.
756  */
757  csquared = indexCorrelation * indexCorrelation;
758 
759  run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
760 
761  /*
762  * Estimate CPU costs per tuple.
763  *
764  * What we want here is cpu_tuple_cost plus the evaluation costs of any
765  * qual clauses that we have to evaluate as qpquals.
766  */
767  cost_qual_eval(&qpqual_cost, qpquals, root);
768 
769  startup_cost += qpqual_cost.startup;
770  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
771 
772  cpu_run_cost += cpu_per_tuple * tuples_fetched;
773 
774  /* tlist eval costs are paid per output row, not per tuple scanned */
775  startup_cost += path->path.pathtarget->cost.startup;
776  cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
777 
778  /* Adjust costing for parallelism, if used. */
779  if (path->path.parallel_workers > 0)
780  {
781  double parallel_divisor = get_parallel_divisor(&path->path);
782 
783  path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
784 
785  /* The CPU cost is divided among all the workers. */
786  cpu_run_cost /= parallel_divisor;
787  }
788 
789  run_cost += cpu_run_cost;
790 
791  path->path.startup_cost = startup_cost;
792  path->path.total_cost = startup_cost + run_cost;
793 }
int compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages, int max_workers)
Definition: allpaths.c:4211
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:130
int max_parallel_workers_per_gather
Definition: costsize.c:133
static List * extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
Definition: costsize.c:812
bool enable_indexscan
Definition: costsize.c:136
List * list_concat(List *list1, const List *list2)
Definition: list.c:560
List * indrestrictinfo
Definition: pathnodes.h:1160
List * indexclauses
Definition: pathnodes.h:1679
Path path
Definition: pathnodes.h:1677
Selectivity indexselectivity
Definition: pathnodes.h:1684
Cost indextotalcost
Definition: pathnodes.h:1683
IndexOptInfo * indexinfo
Definition: pathnodes.h:1678
NodeTag pathtype
Definition: pathnodes.h:1594
double allvisfrac
Definition: pathnodes.h:929
Definition: type.h:95

References RelOptInfo::allvisfrac, Assert(), clamp_row_est(), compute_parallel_worker(), cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_indexscan, extract_nonindex_conditions(), get_parallel_divisor(), get_tablespace_page_costs(), index_pages_fetched(), IndexPath::indexclauses, IndexPath::indexinfo, IndexPath::indexselectivity, IndexPath::indextotalcost, IndexOptInfo::indrestrictinfo, IsA, list_concat(), max_parallel_workers_per_gather, RelOptInfo::pages, Path::parallel_aware, Path::parallel_workers, IndexPath::path, Path::pathtype, QualCost::per_tuple, 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_index_path(), and reparameterize_path().

◆ cost_material()

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

Definition at line 2425 of file costsize.c.

2428 {
2429  Cost startup_cost = input_startup_cost;
2430  Cost run_cost = input_total_cost - input_startup_cost;
2431  double nbytes = relation_byte_size(tuples, width);
2432  long work_mem_bytes = work_mem * 1024L;
2433 
2434  path->rows = tuples;
2435 
2436  /*
2437  * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2438  * reflect bookkeeping overhead. (This rate must be more than what
2439  * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2440  * if it is exactly the same then there will be a cost tie between
2441  * nestloop with A outer, materialized B inner and nestloop with B outer,
2442  * materialized A inner. The extra cost ensures we'll prefer
2443  * materializing the smaller rel.) Note that this is normally a good deal
2444  * less than cpu_tuple_cost; which is OK because a Material plan node
2445  * doesn't do qual-checking or projection, so it's got less overhead than
2446  * most plan nodes.
2447  */
2448  run_cost += 2 * cpu_operator_cost * tuples;
2449 
2450  /*
2451  * If we will spill to disk, charge at the rate of seq_page_cost per page.
2452  * This cost is assumed to be evenly spread through the plan run phase,
2453  * which isn't exactly accurate but our cost model doesn't allow for
2454  * nonuniform costs within the run phase.
2455  */
2456  if (nbytes > work_mem_bytes)
2457  {
2458  double npages = ceil(nbytes / BLCKSZ);
2459 
2460  run_cost += seq_page_cost * npages;
2461  }
2462 
2463  path->startup_cost = startup_cost;
2464  path->total_cost = startup_cost + run_cost;
2465 }

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().

◆ cost_memoize_rescan()

static void cost_memoize_rescan ( PlannerInfo root,
MemoizePath mpath,
Cost rescan_startup_cost,
Cost rescan_total_cost 
)
static

Definition at line 2481 of file costsize.c.

2483 {
2484  EstimationInfo estinfo;
2485  ListCell *lc;
2486  Cost input_startup_cost = mpath->subpath->startup_cost;
2487  Cost input_total_cost = mpath->subpath->total_cost;
2488  double tuples = mpath->subpath->rows;
2489  double calls = mpath->calls;
2490  int width = mpath->subpath->pathtarget->width;
2491 
2492  double hash_mem_bytes;
2493  double est_entry_bytes;
2494  double est_cache_entries;
2495  double ndistinct;
2496  double evict_ratio;
2497  double hit_ratio;
2498  Cost startup_cost;
2499  Cost total_cost;
2500 
2501  /* available cache space */
2502  hash_mem_bytes = get_hash_memory_limit();
2503 
2504  /*
2505  * Set the number of bytes each cache entry should consume in the cache.
2506  * To provide us with better estimations on how many cache entries we can
2507  * store at once, we make a call to the executor here to ask it what
2508  * memory overheads there are for a single cache entry.
2509  */
2510  est_entry_bytes = relation_byte_size(tuples, width) +
2512 
2513  /* include the estimated width for the cache keys */
2514  foreach(lc, mpath->param_exprs)
2515  est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2516 
2517  /* estimate on the upper limit of cache entries we can hold at once */
2518  est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2519 
2520  /* estimate on the distinct number of parameter values */
2521  ndistinct = estimate_num_groups(root, mpath->param_exprs, calls, NULL,
2522  &estinfo);
2523 
2524  /*
2525  * When the estimation fell back on using a default value, it's a bit too
2526  * risky to assume that it's ok to use a Memoize node. The use of a
2527  * default could cause us to use a Memoize node when it's really
2528  * inappropriate to do so. If we see that this has been done, then we'll
2529  * assume that every call will have unique parameters, which will almost
2530  * certainly mean a MemoizePath will never survive add_path().
2531  */
2532  if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2533  ndistinct = calls;
2534 
2535  /*
2536  * Since we've already estimated the maximum number of entries we can
2537  * store at once and know the estimated number of distinct values we'll be
2538  * called with, we'll take this opportunity to set the path's est_entries.
2539  * This will ultimately determine the hash table size that the executor
2540  * will use. If we leave this at zero, the executor will just choose the
2541  * size itself. Really this is not the right place to do this, but it's
2542  * convenient since everything is already calculated.
2543  */
2544  mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2545  PG_UINT32_MAX);
2546 
2547  /*
2548  * When the number of distinct parameter values is above the amount we can
2549  * store in the cache, then we'll have to evict some entries from the
2550  * cache. This is not free. Here we estimate how often we'll incur the
2551  * cost of that eviction.
2552  */
2553  evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2554 
2555  /*
2556  * In order to estimate how costly a single scan will be, we need to
2557  * attempt to estimate what the cache hit ratio will be. To do that we
2558  * must look at how many scans are estimated in total for this node and
2559  * how many of those scans we expect to get a cache hit.
2560  */
2561  hit_ratio = ((calls - ndistinct) / calls) *
2562  (est_cache_entries / Max(ndistinct, est_cache_entries));
2563 
2564  Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2565 
2566  /*
2567  * Set the total_cost accounting for the expected cache hit ratio. We
2568  * also add on a cpu_operator_cost to account for a cache lookup. This
2569  * will happen regardless of whether it's a cache hit or not.
2570  */
2571  total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2572 
2573  /* Now adjust the total cost to account for cache evictions */
2574 
2575  /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2576  total_cost += cpu_tuple_cost * evict_ratio;
2577 
2578  /*
2579  * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2580  * The per-tuple eviction is really just a pfree, so charging a whole
2581  * cpu_operator_cost seems a little excessive.
2582  */
2583  total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2584 
2585  /*
2586  * Now adjust for storing things in the cache, since that's not free
2587  * either. Everything must go in the cache. We don't proportion this
2588  * over any ratio, just apply it once for the scan. We charge a
2589  * cpu_tuple_cost for the creation of the cache entry and also a
2590  * cpu_operator_cost for each tuple we expect to cache.
2591  */
2592  total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2593 
2594  /*
2595  * Getting the first row must be also be proportioned according to the
2596  * expected cache hit ratio.
2597  */
2598  startup_cost = input_startup_cost * (1.0 - hit_ratio);
2599 
2600  /*
2601  * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2602  * which we'll do regardless of whether it was a cache hit or not.
2603  */
2604  startup_cost += cpu_tuple_cost;
2605 
2606  *rescan_startup_cost = startup_cost;
2607  *rescan_total_cost = total_cost;
2608 }
#define PG_UINT32_MAX
Definition: c.h:579
static int32 get_expr_width(PlannerInfo *root, const Node *expr)
Definition: costsize.c:6298
size_t get_hash_memory_limit(void)
Definition: nodeHash.c:3587
double ExecEstimateCacheEntryOverheadBytes(double ntuples)
Definition: nodeMemoize.c:1132
#define SELFLAG_USED_DEFAULT
Definition: selfuncs.h:76
uint32 flags
Definition: selfuncs.h:80
uint32 est_entries
Definition: pathnodes.h:1969
Cardinality calls
Definition: pathnodes.h:1968
Path * subpath
Definition: pathnodes.h:1961
List * param_exprs
Definition: pathnodes.h:1963

References Assert(), MemoizePath::calls, cpu_operator_cost, cpu_tuple_cost, MemoizePath::est_entries, estimate_num_groups(), ExecEstimateCacheEntryOverheadBytes(), EstimationInfo::flags, get_expr_width(), get_hash_memory_limit(), lfirst, Max, Min, MemoizePath::param_exprs, PG_UINT32_MAX, relation_byte_size(), Path::rows, SELFLAG_USED_DEFAULT, Path::startup_cost, MemoizePath::subpath, and Path::total_cost.

Referenced by cost_rescan().

◆ 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 2376 of file costsize.c.

2380 {
2381  Cost startup_cost = 0;
2382  Cost run_cost = 0;
2383  Cost comparison_cost;
2384  double N;
2385  double logN;
2386 
2387  /*
2388  * Avoid log(0)...
2389  */
2390  N = (n_streams < 2) ? 2.0 : (double) n_streams;
2391  logN = LOG2(N);
2392 
2393  /* Assumed cost per tuple comparison */
2394  comparison_cost = 2.0 * cpu_operator_cost;
2395 
2396  /* Heap creation cost */
2397  startup_cost += comparison_cost * N * logN;
2398 
2399  /* Per-tuple heap maintenance cost */
2400  run_cost += tuples * comparison_cost * logN;
2401 
2402  /*
2403  * Although MergeAppend does not do any selection or projection, it's not
2404  * free; add a small per-tuple overhead.
2405  */
2406  run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2407 
2408  path->startup_cost = startup_cost + input_startup_cost;
2409  path->total_cost = startup_cost + run_cost + input_total_cost;
2410 }

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

Referenced by create_merge_append_path().

◆ cost_namedtuplestorescan()

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

Definition at line 1711 of file costsize.c.

1713 {
1714  Cost startup_cost = 0;
1715  Cost run_cost = 0;
1716  QualCost qpqual_cost;
1717  Cost cpu_per_tuple;
1718 
1719  /* Should only be applied to base relations that are Tuplestores */
1720  Assert(baserel->relid > 0);
1721  Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1722 
1723  /* Mark the path with the correct row estimate */
1724  if (param_info)
1725  path->rows = param_info->ppi_rows;
1726  else
1727  path->rows = baserel->rows;
1728 
1729  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1730  cpu_per_tuple = cpu_tuple_cost;
1731 
1732  /* Add scanning CPU costs */
1733  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1734 
1735  startup_cost += qpqual_cost.startup;
1736  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1737  run_cost += cpu_per_tuple * baserel->tuples;
1738 
1739  path->startup_cost = startup_cost;
1740  path->total_cost = startup_cost + run_cost;
1741 }
@ RTE_NAMEDTUPLESTORE
Definition: parsenodes.h:1020

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().

◆ cost_qual_eval()

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

Definition at line 4612 of file costsize.c.

4613 {
4614  cost_qual_eval_context context;
4615  ListCell *l;
4616 
4617  context.root = root;
4618  context.total.startup = 0;
4619  context.total.per_tuple = 0;
4620 
4621  /* We don't charge any cost for the implicit ANDing at top level ... */
4622 
4623  foreach(l, quals)
4624  {
4625  Node *qual = (Node *) lfirst(l);
4626 
4627  cost_qual_eval_walker(qual, &context);
4628  }
4629 
4630  *cost = context.total;
4631 }
static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context)
Definition: costsize.c:4652
PlannerInfo * root
Definition: costsize.c:159

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

Referenced by add_foreign_grouping_paths(), cost_agg(), cost_group(), cost_index(), cost_subplan(), cost_tidrangescan(), cost_tidscan(), create_group_result_path(), create_minmaxagg_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().

◆ cost_qual_eval_node()

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

◆ cost_qual_eval_walker()

static bool cost_qual_eval_walker ( Node node,
cost_qual_eval_context context 
)
static

Definition at line 4652 of file costsize.c.

4653 {
4654  if (node == NULL)
4655  return false;
4656 
4657  /*
4658  * RestrictInfo nodes contain an eval_cost field reserved for this
4659  * routine's use, so that it's not necessary to evaluate the qual clause's
4660  * cost more than once. If the clause's cost hasn't been computed yet,
4661  * the field's startup value will contain -1.
4662  */
4663  if (IsA(node, RestrictInfo))
4664  {
4665  RestrictInfo *rinfo = (RestrictInfo *) node;
4666 
4667  if (rinfo->eval_cost.startup < 0)
4668  {
4669  cost_qual_eval_context locContext;
4670 
4671  locContext.root = context->root;
4672  locContext.total.startup = 0;
4673  locContext.total.per_tuple = 0;
4674 
4675  /*
4676  * For an OR clause, recurse into the marked-up tree so that we
4677  * set the eval_cost for contained RestrictInfos too.
4678  */
4679  if (rinfo->orclause)
4680  cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4681  else
4682  cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4683 
4684  /*
4685  * If the RestrictInfo is marked pseudoconstant, it will be tested
4686  * only once, so treat its cost as all startup cost.
4687  */
4688  if (rinfo->pseudoconstant)
4689  {
4690  /* count one execution during startup */
4691  locContext.total.startup += locContext.total.per_tuple;
4692  locContext.total.per_tuple = 0;
4693  }
4694  rinfo->eval_cost = locContext.total;
4695  }
4696  context->total.startup += rinfo->eval_cost.startup;
4697  context->total.per_tuple += rinfo->eval_cost.per_tuple;
4698  /* do NOT recurse into children */
4699  return false;
4700  }
4701 
4702  /*
4703  * For each operator or function node in the given tree, we charge the
4704  * estimated execution cost given by pg_proc.procost (remember to multiply
4705  * this by cpu_operator_cost).
4706  *
4707  * Vars and Consts are charged zero, and so are boolean operators (AND,
4708  * OR, NOT). Simplistic, but a lot better than no model at all.
4709  *
4710  * Should we try to account for the possibility of short-circuit
4711  * evaluation of AND/OR? Probably *not*, because that would make the
4712  * results depend on the clause ordering, and we are not in any position
4713  * to expect that the current ordering of the clauses is the one that's
4714  * going to end up being used. The above per-RestrictInfo caching would
4715  * not mix well with trying to re-order clauses anyway.
4716  *
4717  * Another issue that is entirely ignored here is that if a set-returning
4718  * function is below top level in the tree, the functions/operators above
4719  * it will need to be evaluated multiple times. In practical use, such
4720  * cases arise so seldom as to not be worth the added complexity needed;
4721  * moreover, since our rowcount estimates for functions tend to be pretty
4722  * phony, the results would also be pretty phony.
4723  */
4724  if (IsA(node, FuncExpr))
4725  {
4726  add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4727  &context->total);
4728  }
4729  else if (IsA(node, OpExpr) ||
4730  IsA(node, DistinctExpr) ||
4731  IsA(node, NullIfExpr))
4732  {
4733  /* rely on struct equivalence to treat these all alike */
4734  set_opfuncid((OpExpr *) node);
4735  add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4736  &context->total);
4737  }
4738  else if (IsA(node, ScalarArrayOpExpr))
4739  {
4740  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4741  Node *arraynode = (Node *) lsecond(saop->args);
4742  QualCost sacosts;
4743  QualCost hcosts;
4744  int estarraylen = estimate_array_length(arraynode);
4745 
4746  set_sa_opfuncid(saop);
4747  sacosts.startup = sacosts.per_tuple = 0;
4748  add_function_cost(context->root, saop->opfuncid, NULL,
4749  &sacosts);
4750 
4751  if (OidIsValid(saop->hashfuncid))
4752  {
4753  /* Handle costs for hashed ScalarArrayOpExpr */
4754  hcosts.startup = hcosts.per_tuple = 0;
4755 
4756  add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4757  context->total.startup += sacosts.startup + hcosts.startup;
4758 
4759  /* Estimate the cost of building the hashtable. */
4760  context->total.startup += estarraylen * hcosts.per_tuple;
4761 
4762  /*
4763  * XXX should we charge a little bit for sacosts.per_tuple when
4764  * building the table, or is it ok to assume there will be zero
4765  * hash collision?
4766  */
4767 
4768  /*
4769  * Charge for hashtable lookups. Charge a single hash and a
4770  * single comparison.
4771  */
4772  context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4773  }
4774  else
4775  {
4776  /*
4777  * Estimate that the operator will be applied to about half of the
4778  * array elements before the answer is determined.
4779  */
4780  context->total.startup += sacosts.startup;
4781  context->total.per_tuple += sacosts.per_tuple *
4782  estimate_array_length(arraynode) * 0.5;
4783  }
4784  }
4785  else if (IsA(node, Aggref) ||
4786  IsA(node, WindowFunc))
4787  {
4788  /*
4789  * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4790  * ie, zero execution cost in the current model, because they behave
4791  * essentially like Vars at execution. We disregard the costs of
4792  * their input expressions for the same reason. The actual execution
4793  * costs of the aggregate/window functions and their arguments have to
4794  * be factored into plan-node-specific costing of the Agg or WindowAgg
4795  * plan node.
4796  */
4797  return false; /* don't recurse into children */
4798  }
4799  else if (IsA(node, GroupingFunc))
4800  {
4801  /* Treat this as having cost 1 */
4802  context->total.per_tuple += cpu_operator_cost;
4803  return false; /* don't recurse into children */
4804  }
4805  else if (IsA(node, CoerceViaIO))
4806  {
4807  CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4808  Oid iofunc;
4809  Oid typioparam;
4810  bool typisvarlena;
4811 
4812  /* check the result type's input function */
4813  getTypeInputInfo(iocoerce->resulttype,
4814  &iofunc, &typioparam);
4815  add_function_cost(context->root, iofunc, NULL,
4816  &context->total);
4817  /* check the input type's output function */
4818  getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4819  &iofunc, &typisvarlena);
4820  add_function_cost(context->root, iofunc, NULL,
4821  &context->total);
4822  }
4823  else if (IsA(node, ArrayCoerceExpr))
4824  {
4825  ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4826  QualCost perelemcost;
4827 
4828  cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4829  context->root);
4830  context->total.startup += perelemcost.startup;
4831  if (perelemcost.per_tuple > 0)
4832  context->total.per_tuple += perelemcost.per_tuple *
4833  estimate_array_length((Node *) acoerce->arg);
4834  }
4835  else if (IsA(node, RowCompareExpr))
4836  {
4837  /* Conservatively assume we will check all the columns */
4838  RowCompareExpr *rcexpr = (RowCompareExpr *) node;
4839  ListCell *lc;
4840 
4841  foreach(lc, rcexpr->opnos)
4842  {
4843  Oid opid = lfirst_oid(lc);
4844 
4845  add_function_cost(context->root, get_opcode(opid), NULL,
4846  &context->total);
4847  }
4848  }
4849  else if (IsA(node, MinMaxExpr) ||
4850  IsA(node, SQLValueFunction) ||
4851  IsA(node, XmlExpr) ||
4852  IsA(node, CoerceToDomain) ||
4853  IsA(node, NextValueExpr))
4854  {
4855  /* Treat all these as having cost 1 */
4856  context->total.per_tuple += cpu_operator_cost;
4857  }
4858  else if (IsA(node, CurrentOfExpr))
4859  {
4860  /* Report high cost to prevent selection of anything but TID scan */
4861  context->total.startup += disable_cost;
4862  }
4863  else if (IsA(node, SubLink))
4864  {
4865  /* This routine should not be applied to un-planned expressions */
4866  elog(ERROR, "cannot handle unplanned sub-select");
4867  }
4868  else if (IsA(node, SubPlan))
4869  {
4870  /*
4871  * A subplan node in an expression typically indicates that the
4872  * subplan will be executed on each evaluation, so charge accordingly.
4873  * (Sub-selects that can be executed as InitPlans have already been
4874  * removed from the expression.)
4875  */
4876  SubPlan *subplan = (SubPlan *) node;
4877 
4878  context->total.startup += subplan->startup_cost;
4879  context->total.per_tuple += subplan->per_call_cost;
4880 
4881  /*
4882  * We don't want to recurse into the testexpr, because it was already
4883  * counted in the SubPlan node's costs. So we're done.
4884  */
4885  return false;
4886  }
4887  else if (IsA(node, AlternativeSubPlan))
4888  {
4889  /*
4890  * Arbitrarily use the first alternative plan for costing. (We should
4891  * certainly only include one alternative, and we don't yet have
4892  * enough information to know which one the executor is most likely to
4893  * use.)
4894  */
4895  AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
4896 
4897  return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
4898  context);
4899  }
4900  else if (IsA(node, PlaceHolderVar))
4901  {
4902  /*
4903  * A PlaceHolderVar should be given cost zero when considering general
4904  * expression evaluation costs. The expense of doing the contained
4905  * expression is charged as part of the tlist eval costs of the scan
4906  * or join where the PHV is first computed (see set_rel_width and
4907  * add_placeholders_to_joinrel). If we charged it again here, we'd be
4908  * double-counting the cost for each level of plan that the PHV
4909  * bubbles up through. Hence, return without recursing into the
4910  * phexpr.
4911  */
4912  return false;
4913  }
4914 
4915  /* recurse into children */
4917  (void *) context);
4918 }
#define OidIsValid(objectId)
Definition: c.h:764
void getTypeOutputInfo(Oid type, Oid *typOutput, bool *typIsVarlena)
Definition: lsyscache.c:2889
RegProcedure get_opcode(Oid opno)
Definition: lsyscache.c:1289
void getTypeInputInfo(Oid type, Oid *typInput, Oid *typIOParam)
Definition: lsyscache.c:2856
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:43
void set_sa_opfuncid(ScalarArrayOpExpr *opexpr)
Definition: nodeFuncs.c:1790
void set_opfuncid(OpExpr *opexpr)
Definition: nodeFuncs.c:1779
#define expression_tree_walker(n, w, c)
Definition: nodeFuncs.h:151
#define lsecond(l)
Definition: pg_list.h:183
#define lfirst_oid(lc)
Definition: pg_list.h:174
void add_function_cost(PlannerInfo *root, Oid funcid, Node *node, QualCost *cost)
Definition: plancat.c:2043
unsigned int Oid
Definition: postgres_ext.h:31
int estimate_array_length(Node *arrayexpr)
Definition: selfuncs.c:2134
Expr * arg
Definition: primnodes.h:1134
Oid resulttype
Definition: primnodes.h:1135
Cost startup_cost
Definition: primnodes.h:1020
Cost per_call_cost
Definition: primnodes.h:1021

References add_function_cost(), CoerceViaIO::arg, ArrayCoerceExpr::arg, ScalarArrayOpExpr::args, RestrictInfo::clause, cost_qual_eval_node(), cpu_operator_cost, disable_cost, ArrayCoerceExpr::elemexpr, elog(), ERROR, estimate_array_length(), expression_tree_walker, exprType(), get_opcode(), getTypeInputInfo(), getTypeOutputInfo(), IsA, lfirst_oid, linitial, lsecond, OidIsValid, SubPlan::per_call_cost, QualCost::per_tuple, CoerceViaIO::resulttype, cost_qual_eval_context::root, set_opfuncid(), set_sa_opfuncid(), QualCost::startup, SubPlan::startup_cost, AlternativeSubPlan::subplans, and cost_qual_eval_context::total.

Referenced by cost_qual_eval(), and cost_qual_eval_node().

◆ cost_recursive_union()

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

Definition at line 1785 of file costsize.c.

1786 {
1787  Cost startup_cost;
1788  Cost total_cost;
1789  double total_rows;
1790 
1791  /* We probably have decent estimates for the non-recursive term */
1792  startup_cost = nrterm->startup_cost;
1793  total_cost = nrterm->total_cost;
1794  total_rows = nrterm->rows;
1795 
1796  /*
1797  * We arbitrarily assume that about 10 recursive iterations will be
1798  * needed, and that we've managed to get a good fix on the cost and output
1799  * size of each one of them. These are mighty shaky assumptions but it's
1800  * hard to see how to do better.
1801  */
1802  total_cost += 10 * rterm->total_cost;
1803  total_rows += 10 * rterm->rows;
1804 
1805  /*
1806  * Also charge cpu_tuple_cost per row to account for the costs of
1807  * manipulating the tuplestores. (We don't worry about possible
1808  * spill-to-disk costs.)
1809  */
1810  total_cost += cpu_tuple_cost * total_rows;
1811 
1812  runion->startup_cost = startup_cost;
1813  runion->total_cost = total_cost;
1814  runion->rows = total_rows;
1815  runion->pathtarget->width = Max(nrterm->pathtarget->width,
1816  rterm->pathtarget->width);
1817 }

References cpu_tuple_cost, Max, Path::rows, Path::startup_cost, and Path::total_cost.

Referenced by create_recursiveunion_path().

◆ cost_rescan()

static void cost_rescan ( PlannerInfo root,
Path path,
Cost rescan_startup_cost,
Cost rescan_total_cost 
)
static

Definition at line 4500 of file costsize.c.

4503 {
4504  switch (path->pathtype)
4505  {
4506  case T_FunctionScan:
4507 
4508  /*
4509  * Currently, nodeFunctionscan.c always executes the function to
4510  * completion before returning any rows, and caches the results in
4511  * a tuplestore. So the function eval cost is all startup cost
4512  * and isn't paid over again on rescans. However, all run costs
4513  * will be paid over again.
4514  */
4515  *rescan_startup_cost = 0;
4516  *rescan_total_cost = path->total_cost - path->startup_cost;
4517  break;
4518  case T_HashJoin:
4519 
4520  /*
4521  * If it's a single-batch join, we don't need to rebuild the hash
4522  * table during a rescan.
4523  */
4524  if (((HashPath *) path)->num_batches == 1)
4525  {
4526  /* Startup cost is exactly the cost of hash table building */
4527  *rescan_startup_cost = 0;
4528  *rescan_total_cost = path->total_cost - path->startup_cost;
4529  }
4530  else
4531  {
4532  /* Otherwise, no special treatment */
4533  *rescan_startup_cost = path->startup_cost;
4534  *rescan_total_cost = path->total_cost;
4535  }
4536  break;
4537  case T_CteScan:
4538  case T_WorkTableScan:
4539  {
4540  /*
4541  * These plan types materialize their final result in a
4542  * tuplestore or tuplesort object. So the rescan cost is only
4543  * cpu_tuple_cost per tuple, unless the result is large enough
4544  * to spill to disk.
4545  */
4546  Cost run_cost = cpu_tuple_cost * path->rows;
4547  double nbytes = relation_byte_size(path->rows,
4548  path->pathtarget->width);
4549  long work_mem_bytes = work_mem * 1024L;
4550 
4551  if (nbytes > work_mem_bytes)
4552  {
4553  /* It will spill, so account for re-read cost */
4554  double npages = ceil(nbytes / BLCKSZ);
4555 
4556  run_cost += seq_page_cost * npages;
4557  }
4558  *rescan_startup_cost = 0;
4559  *rescan_total_cost = run_cost;
4560  }
4561  break;
4562  case T_Material:
4563  case T_Sort:
4564  {
4565  /*
4566  * These plan types not only materialize their results, but do
4567  * not implement qual filtering or projection. So they are
4568  * even cheaper to rescan than the ones above. We charge only
4569  * cpu_operator_cost per tuple. (Note: keep that in sync with
4570  * the run_cost charge in cost_sort, and also see comments in
4571  * cost_material before you change it.)
4572  */
4573  Cost run_cost = cpu_operator_cost * path->rows;
4574  double nbytes = relation_byte_size(path->rows,
4575  path->pathtarget->width);
4576  long work_mem_bytes = work_mem * 1024L;
4577 
4578  if (nbytes > work_mem_bytes)
4579  {
4580  /* It will spill, so account for re-read cost */
4581  double npages = ceil(nbytes / BLCKSZ);
4582 
4583  run_cost += seq_page_cost * npages;
4584  }
4585  *rescan_startup_cost = 0;
4586  *rescan_total_cost = run_cost;
4587  }
4588  break;
4589  case T_Memoize:
4590  /* All the hard work is done by cost_memoize_rescan */
4591  cost_memoize_rescan(root, (MemoizePath *) path,
4592  rescan_startup_cost, rescan_total_cost);
4593  break;
4594  default:
4595  *rescan_startup_cost = path->startup_cost;
4596  *rescan_total_cost = path->total_cost;
4597  break;
4598  }
4599 }
static void cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath, Cost *rescan_startup_cost, Cost *rescan_total_cost)
Definition: costsize.c:2481

References cost_memoize_rescan(), cpu_operator_cost, cpu_tuple_cost, Path::pathtype, relation_byte_size(), seq_page_cost, Path::startup_cost, Path::total_cost, and work_mem.

Referenced by initial_cost_nestloop().

◆ cost_resultscan()

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

Definition at line 1748 of file costsize.c.

1750 {
1751  Cost startup_cost = 0;
1752  Cost run_cost = 0;
1753  QualCost qpqual_cost;
1754  Cost cpu_per_tuple;
1755 
1756  /* Should only be applied to RTE_RESULT base relations */
1757  Assert(baserel->relid > 0);
1758  Assert(baserel->rtekind == RTE_RESULT);
1759 
1760  /* Mark the path with the correct row estimate */
1761  if (param_info)
1762  path->rows = param_info->ppi_rows;
1763  else
1764  path->rows = baserel->rows;
1765 
1766  /* We charge qual cost plus cpu_tuple_cost */
1767  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1768 
1769  startup_cost += qpqual_cost.startup;
1770  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1771  run_cost += cpu_per_tuple * baserel->tuples;
1772 
1773  path->startup_cost = startup_cost;
1774  path->total_cost = startup_cost + run_cost;
1775 }
@ RTE_RESULT
Definition: parsenodes.h:1021

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

Referenced by create_resultscan_path().

◆ cost_samplescan()

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

Definition at line 333 of file costsize.c.

335 {
336  Cost startup_cost = 0;
337  Cost run_cost = 0;
338  RangeTblEntry *rte;
339  TableSampleClause *tsc;
340  TsmRoutine *tsm;
341  double spc_seq_page_cost,
342  spc_random_page_cost,
343  spc_page_cost;
344  QualCost qpqual_cost;
345  Cost cpu_per_tuple;
346 
347  /* Should only be applied to base relations with tablesample clauses */
348  Assert(baserel->relid > 0);
349  rte = planner_rt_fetch(baserel->relid, root);
350  Assert(rte->rtekind == RTE_RELATION);
351  tsc = rte->tablesample;
352  Assert(tsc != NULL);
353  tsm = GetTsmRoutine(tsc->tsmhandler);
354 
355  /* Mark the path with the correct row estimate */
356  if (param_info)
357  path->rows = param_info->ppi_rows;
358  else
359  path->rows = baserel->rows;
360 
361  /* fetch estimated page cost for tablespace containing table */
363  &spc_random_page_cost,
364  &spc_seq_page_cost);
365 
366  /* if NextSampleBlock is used, assume random access, else sequential */
367  spc_page_cost = (tsm->NextSampleBlock != NULL) ?
368  spc_random_page_cost : spc_seq_page_cost;
369 
370  /*
371  * disk costs (recall that baserel->pages has already been set to the
372  * number of pages the sampling method will visit)
373  */
374  run_cost += spc_page_cost * baserel->pages;
375 
376  /*
377  * CPU costs (recall that baserel->tuples has already been set to the
378  * number of tuples the sampling method will select). Note that we ignore
379  * execution cost of the TABLESAMPLE parameter expressions; they will be
380  * evaluated only once per scan, and in most usages they'll likely be
381  * simple constants anyway. We also don't charge anything for the
382  * calculations the sampling method might do internally.
383  */
384  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
385 
386  startup_cost += qpqual_cost.startup;
387  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
388  run_cost += cpu_per_tuple * baserel->tuples;
389  /* tlist eval costs are paid per output row, not per tuple scanned */
390  startup_cost += path->pathtarget->cost.startup;
391  run_cost += path->pathtarget->cost.per_tuple * path->rows;
392 
393  path->startup_cost = startup_cost;
394  path->total_cost = startup_cost + run_cost;
395 }
struct TableSampleClause * tablesample
Definition: parsenodes.h:1074
NextSampleBlock_function NextSampleBlock
Definition: tsmapi.h:73
TsmRoutine * GetTsmRoutine(Oid tsmhandler)
Definition: tablesample.c:27

References Assert(), cpu_tuple_cost, get_restriction_qual_cost(), get_tablespace_page_costs(), GetTsmRoutine(), TsmRoutine::NextSampleBlock, RelOptInfo::pages, 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().

◆ cost_seqscan()

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

Definition at line 256 of file costsize.c.

258 {
259  Cost startup_cost = 0;
260  Cost cpu_run_cost;
261  Cost disk_run_cost;
262  double spc_seq_page_cost;
263  QualCost qpqual_cost;
264  Cost cpu_per_tuple;
265 
266  /* Should only be applied to base relations */
267  Assert(baserel->relid > 0);
268  Assert(baserel->rtekind == RTE_RELATION);
269 
270  /* Mark the path with the correct row estimate */
271  if (param_info)
272  path->rows = param_info->ppi_rows;
273  else
274  path->rows = baserel->rows;
275 
276  if (!enable_seqscan)
277  startup_cost += disable_cost;
278 
279  /* fetch estimated page cost for tablespace containing table */
281  NULL,
282  &spc_seq_page_cost);
283 
284  /*
285  * disk costs
286  */
287  disk_run_cost = spc_seq_page_cost * baserel->pages;
288 
289  /* CPU costs */
290  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
291 
292  startup_cost += qpqual_cost.startup;
293  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
294  cpu_run_cost = cpu_per_tuple * baserel->tuples;
295  /* tlist eval costs are paid per output row, not per tuple scanned */
296  startup_cost += path->pathtarget->cost.startup;
297  cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
298 
299  /* Adjust costing for parallelism, if used. */
300  if (path->parallel_workers > 0)
301  {
302  double parallel_divisor = get_parallel_divisor(path);
303 
304  /* The CPU cost is divided among all the workers. */
305  cpu_run_cost /= parallel_divisor;
306 
307  /*
308  * It may be possible to amortize some of the I/O cost, but probably
309  * not very much, because most operating systems already do aggressive
310  * prefetching. For now, we assume that the disk run cost can't be
311  * amortized at all.
312  */
313 
314  /*
315  * In the case of a parallel plan, the row count needs to represent
316  * the number of tuples processed per worker.
317  */
318  path->rows = clamp_row_est(path->rows / parallel_divisor);
319  }
320 
321  path->startup_cost = startup_cost;
322  path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
323 }
bool enable_seqscan
Definition: costsize.c:135

References Assert(), clamp_row_est(), cpu_tuple_cost, disable_cost, enable_seqscan, get_parallel_divisor(), get_restriction_qual_cost(), get_tablespace_page_costs(), RelOptInfo::pages, Path::parallel_workers, 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().

◆ 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 2096 of file costsize.c.

2101 {
2102  Cost startup_cost;
2103  Cost run_cost;
2104 
2105  cost_tuplesort(&startup_cost, &run_cost,
2106  tuples, width,
2107  comparison_cost, sort_mem,
2108  limit_tuples);
2109 
2110  if (!enable_sort)
2111  startup_cost += disable_cost;
2112 
2113  startup_cost += input_cost;
2114 
2115  path->rows = tuples;
2116  path->startup_cost = startup_cost;
2117  path->total_cost = startup_cost + run_cost;
2118 }
bool enable_sort
Definition: costsize.c:140

References cost_tuplesort(), disable_cost, enable_sort, Path::rows, Path::startup_cost, and Path::total_cost.

Referenced by adjust_foreign_grouping_path_cost(), choose_hashed_setop(), cost_append(), 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().

◆ cost_subplan()

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

Definition at line 4407 of file costsize.c.

4408 {
4409  QualCost sp_cost;
4410 
4411  /* Figure any cost for evaluating the testexpr */
4412  cost_qual_eval(&sp_cost,
4413  make_ands_implicit((Expr *) subplan->testexpr),
4414  root);
4415 
4416  if (subplan->useHashTable)
4417  {
4418  /*
4419  * If we are using a hash table for the subquery outputs, then the
4420  * cost of evaluating the query is a one-time cost. We charge one
4421  * cpu_operator_cost per tuple for the work of loading the hashtable,
4422  * too.
4423  */
4424  sp_cost.startup += plan->total_cost +
4425  cpu_operator_cost * plan->plan_rows;
4426 
4427  /*
4428  * The per-tuple costs include the cost of evaluating the lefthand
4429  * expressions, plus the cost of probing the hashtable. We already
4430  * accounted for the lefthand expressions as part of the testexpr, and
4431  * will also have counted one cpu_operator_cost for each comparison
4432  * operator. That is probably too low for the probing cost, but it's
4433  * hard to make a better estimate, so live with it for now.
4434  */
4435  }
4436  else
4437  {
4438  /*
4439  * Otherwise we will be rescanning the subplan output on each
4440  * evaluation. We need to estimate how much of the output we will
4441  * actually need to scan. NOTE: this logic should agree with the
4442  * tuple_fraction estimates used by make_subplan() in
4443  * plan/subselect.c.
4444  */
4445  Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4446 
4447  if (subplan->subLinkType == EXISTS_SUBLINK)
4448  {
4449  /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4450  sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4451  }
4452  else if (subplan->subLinkType == ALL_SUBLINK ||
4453  subplan->subLinkType == ANY_SUBLINK)
4454  {
4455  /* assume we need 50% of the tuples */
4456  sp_cost.per_tuple += 0.50 * plan_run_cost;
4457  /* also charge a cpu_operator_cost per row examined */
4458  sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4459  }
4460  else
4461  {
4462  /* assume we need all tuples */
4463  sp_cost.per_tuple += plan_run_cost;
4464  }
4465 
4466  /*
4467  * Also account for subplan's startup cost. If the subplan is
4468  * uncorrelated or undirect correlated, AND its topmost node is one
4469  * that materializes its output, assume that we'll only need to pay
4470  * its startup cost once; otherwise assume we pay the startup cost
4471  * every time.
4472  */
4473  if (subplan->parParam == NIL &&
4475  sp_cost.startup += plan->startup_cost;
4476  else
4477  sp_cost.per_tuple += plan->startup_cost;
4478  }
4479 
4480  subplan->startup_cost = sp_cost.startup;
4481  subplan->per_call_cost = sp_cost.per_tuple;
4482 }
bool ExecMaterializesOutput(NodeTag plantype)
Definition: execAmi.c:637
List * make_ands_implicit(Expr *clause)
Definition: makefuncs.c:722
#define plan(x)
Definition: pg_regress.c:154
@ ANY_SUBLINK
Definition: primnodes.h:926
@ ALL_SUBLINK
Definition: primnodes.h:925
@ EXISTS_SUBLINK
Definition: primnodes.h:924
bool useHashTable
Definition: primnodes.h:1006
Node * testexpr
Definition: primnodes.h:994
List * parParam
Definition: primnodes.h:1017
SubLinkType subLinkType
Definition: primnodes.h:992

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, QualCost::startup, SubPlan::startup_cost, SubPlan::subLinkType, SubPlan::testexpr, and SubPlan::useHashTable.

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

◆ cost_subqueryscan()

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

Definition at line 1423 of file costsize.c.

1426 {
1427  Cost startup_cost;
1428  Cost run_cost;
1429  List *qpquals;
1430  QualCost qpqual_cost;
1431  Cost cpu_per_tuple;
1432 
1433  /* Should only be applied to base relations that are subqueries */
1434  Assert(baserel->relid > 0);
1435  Assert(baserel->rtekind == RTE_SUBQUERY);
1436 
1437  /*
1438  * We compute the rowcount estimate as the subplan's estimate times the
1439  * selectivity of relevant restriction clauses. In simple cases this will
1440  * come out the same as baserel->rows; but when dealing with parallelized
1441  * paths we must do it like this to get the right answer.
1442  */
1443  if (param_info)
1444  qpquals = list_concat_copy(param_info->ppi_clauses,
1445  baserel->baserestrictinfo);
1446  else
1447  qpquals = baserel->baserestrictinfo;
1448 
1449  path->path.rows = clamp_row_est(path->subpath->rows *
1451  qpquals,
1452  0,
1453  JOIN_INNER,
1454  NULL));
1455 
1456  /*
1457  * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1458  * any restriction clauses and tlist that will be attached to the
1459  * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1460  * projection overhead.
1461  */
1462  path->path.startup_cost = path->subpath->startup_cost;
1463  path->path.total_cost = path->subpath->total_cost;
1464 
1465  /*
1466  * However, if there are no relevant restriction clauses and the
1467  * pathtarget is trivial, then we expect that setrefs.c will optimize away
1468  * the SubqueryScan plan node altogether, so we should just make its cost
1469  * and rowcount equal to the input path's.
1470  *
1471  * Note: there are some edge cases where createplan.c will apply a
1472  * different targetlist to the SubqueryScan node, thus falsifying our
1473  * current estimate of whether the target is trivial, and making the cost
1474  * estimate (though not the rowcount) wrong. It does not seem worth the
1475  * extra complication to try to account for that exactly, especially since
1476  * that behavior falsifies other cost estimates as well.
1477  */
1478  if (qpquals == NIL && trivial_pathtarget)
1479  return;
1480 
1481  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1482 
1483  startup_cost = qpqual_cost.startup;
1484  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1485  run_cost = cpu_per_tuple * path->subpath->rows;
1486 
1487  /* tlist eval costs are paid per output row, not per tuple scanned */
1488  startup_cost += path->path.pathtarget->cost.startup;
1489  run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1490 
1491  path->path.startup_cost += startup_cost;
1492  path->path.total_cost += startup_cost + run_cost;
1493 }
List * list_concat_copy(const List *list1, const List *list2)
Definition: list.c:597
@ RTE_SUBQUERY
Definition: parsenodes.h:1014
List * ppi_clauses
Definition: pathnodes.h:1549
List * baserestrictinfo
Definition: pathnodes.h:964

References Assert(), RelOptInfo::baserestrictinfo, clamp_row_est(), clauselist_selectivity(), cpu_tuple_cost, get_restriction_qual_cost(), JOIN_INNER, list_concat_copy(), NIL, SubqueryScanPath::path, QualCost::per_tuple, ParamPathInfo::ppi_clauses, RelOptInfo::relid, Path::rows, RTE_SUBQUERY, RelOptInfo::rtekind, QualCost::startup, Path::startup_cost, SubqueryScanPath::subpath, and Path::total_cost.

Referenced by create_subqueryscan_path().

◆ cost_tablefuncscan()

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

Definition at line 1564 of file costsize.c.

1566 {
1567  Cost startup_cost = 0;
1568  Cost run_cost = 0;
1569  QualCost qpqual_cost;
1570  Cost cpu_per_tuple;
1571  RangeTblEntry *rte;
1572  QualCost exprcost;
1573 
1574  /* Should only be applied to base relations that are functions */
1575  Assert(baserel->relid > 0);
1576  rte = planner_rt_fetch(baserel->relid, root);
1577  Assert(rte->rtekind == RTE_TABLEFUNC);
1578 
1579  /* Mark the path with the correct row estimate */
1580  if (param_info)
1581  path->rows = param_info->ppi_rows;
1582  else
1583  path->rows = baserel->rows;
1584 
1585  /*
1586  * Estimate costs of executing the table func expression(s).
1587  *
1588  * XXX in principle we ought to charge tuplestore spill costs if the
1589  * number of rows is large. However, given how phony our rowcount
1590  * estimates for tablefuncs tend to be, there's not a lot of point in that
1591  * refinement right now.
1592  */
1593  cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1594 
1595  startup_cost += exprcost.startup + exprcost.per_tuple;
1596 
1597  /* Add scanning CPU costs */
1598  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1599 
1600  startup_cost += qpqual_cost.startup;
1601  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1602  run_cost += cpu_per_tuple * baserel->tuples;
1603 
1604  /* tlist eval costs are paid per output row, not per tuple scanned */
1605  startup_cost += path->pathtarget->cost.startup;
1606  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1607 
1608  path->startup_cost = startup_cost;
1609  path->total_cost = startup_cost + run_cost;
1610 }
@ RTE_TABLEFUNC
Definition: parsenodes.h:1017
TableFunc * tablefunc
Definition: parsenodes.h:1153

References Assert(), cost_qual_eval_node(), cpu_tuple_cost, get_restriction_qual_cost(), 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().

◆ cost_tidrangescan()

void cost_tidrangescan ( Path path,
PlannerInfo root,
RelOptInfo baserel,
List tidrangequals,
ParamPathInfo param_info 
)

Definition at line 1329 of file costsize.c.

1332 {
1333  Selectivity selectivity;
1334  double pages;
1335  Cost startup_cost = 0;
1336  Cost run_cost = 0;
1337  QualCost qpqual_cost;
1338  Cost cpu_per_tuple;
1339  QualCost tid_qual_cost;
1340  double ntuples;
1341  double nseqpages;
1342  double spc_random_page_cost;
1343  double spc_seq_page_cost;
1344 
1345  /* Should only be applied to base relations */
1346  Assert(baserel->relid > 0);
1347  Assert(baserel->rtekind == RTE_RELATION);
1348 
1349  /* Mark the path with the correct row estimate */
1350  if (param_info)
1351  path->rows = param_info->ppi_rows;
1352  else
1353  path->rows = baserel->rows;
1354 
1355  /* Count how many tuples and pages we expect to scan */
1356  selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1357  JOIN_INNER, NULL);
1358  pages = ceil(selectivity * baserel->pages);
1359 
1360  if (pages <= 0.0)
1361  pages = 1.0;
1362 
1363  /*
1364  * The first page in a range requires a random seek, but each subsequent
1365  * page is just a normal sequential page read. NOTE: it's desirable for
1366  * TID Range Scans to cost more than the equivalent Sequential Scans,
1367  * because Seq Scans have some performance advantages such as scan
1368  * synchronization and parallelizability, and we'd prefer one of them to
1369  * be picked unless a TID Range Scan really is better.
1370  */
1371  ntuples = selectivity * baserel->tuples;
1372  nseqpages = pages - 1.0;
1373 
1374  if (!enable_tidscan)
1375  startup_cost += disable_cost;
1376 
1377  /*
1378  * The TID qual expressions will be computed once, any other baserestrict
1379  * quals once per retrieved tuple.
1380  */
1381  cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1382 
1383  /* fetch estimated page cost for tablespace containing table */
1385  &spc_random_page_cost,
1386  &spc_seq_page_cost);
1387 
1388  /* disk costs; 1 random page and the remainder as seq pages */
1389  run_cost += spc_random_page_cost + spc_seq_page_cost * nseqpages;
1390 
1391  /* Add scanning CPU costs */
1392  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1393 
1394  /*
1395  * XXX currently we assume TID quals are a subset of qpquals at this
1396  * point; they will be removed (if possible) when we create the plan, so
1397  * we subtract their cost from the total qpqual cost. (If the TID quals
1398  * can't be removed, this is a mistake and we're going to underestimate
1399  * the CPU cost a bit.)
1400  */
1401  startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1402  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1403  tid_qual_cost.per_tuple;
1404  run_cost += cpu_per_tuple * ntuples;
1405 
1406  /* tlist eval costs are paid per output row, not per tuple scanned */
1407  startup_cost += path->pathtarget->cost.startup;
1408  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1409 
1410  path->startup_cost = startup_cost;
1411  path->total_cost = startup_cost + run_cost;
1412 }
bool enable_tidscan
Definition: costsize.c:139

References Assert(), clauselist_selectivity(), cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_tidscan, get_restriction_qual_cost(), get_tablespace_page_costs(), JOIN_INNER, RelOptInfo::pages, 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_tidrangescan_path().

◆ cost_tidscan()

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

Definition at line 1221 of file costsize.c.

1223 {
1224  Cost startup_cost = 0;
1225  Cost run_cost = 0;
1226  bool isCurrentOf = false;
1227  QualCost qpqual_cost;
1228  Cost cpu_per_tuple;
1229  QualCost tid_qual_cost;
1230  int ntuples;
1231  ListCell *l;
1232  double spc_random_page_cost;
1233 
1234  /* Should only be applied to base relations */
1235  Assert(baserel->relid > 0);
1236  Assert(baserel->rtekind == RTE_RELATION);
1237 
1238  /* Mark the path with the correct row estimate */
1239  if (param_info)
1240  path->rows = param_info->ppi_rows;
1241  else
1242  path->rows = baserel->rows;
1243 
1244  /* Count how many tuples we expect to retrieve */
1245  ntuples = 0;
1246  foreach(l, tidquals)
1247  {
1248  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1249  Expr *qual = rinfo->clause;
1250 
1251  if (IsA(qual, ScalarArrayOpExpr))
1252  {
1253  /* Each element of the array yields 1 tuple */
1254  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) qual;
1255  Node *arraynode = (Node *) lsecond(saop->args);
1256 
1257  ntuples += estimate_array_length(arraynode);
1258  }
1259  else if (IsA(qual, CurrentOfExpr))
1260  {
1261  /* CURRENT OF yields 1 tuple */
1262  isCurrentOf = true;
1263  ntuples++;
1264  }
1265  else
1266  {
1267  /* It's just CTID = something, count 1 tuple */
1268  ntuples++;
1269  }
1270  }
1271 
1272  /*
1273  * We must force TID scan for WHERE CURRENT OF, because only nodeTidscan.c
1274  * understands how to do it correctly. Therefore, honor enable_tidscan
1275  * only when CURRENT OF isn't present. Also note that cost_qual_eval
1276  * counts a CurrentOfExpr as having startup cost disable_cost, which we
1277  * subtract off here; that's to prevent other plan types such as seqscan
1278  * from winning.
1279  */
1280  if (isCurrentOf)
1281  {
1283  startup_cost -= disable_cost;
1284  }
1285  else if (!enable_tidscan)
1286  startup_cost += disable_cost;
1287 
1288  /*
1289  * The TID qual expressions will be computed once, any other baserestrict
1290  * quals once per retrieved tuple.
1291  */
1292  cost_qual_eval(&tid_qual_cost, tidquals, root);
1293 
1294  /* fetch estimated page cost for tablespace containing table */
1296  &spc_random_page_cost,
1297  NULL);
1298 
1299  /* disk costs --- assume each tuple on a different page */
1300  run_cost += spc_random_page_cost * ntuples;
1301 
1302  /* Add scanning CPU costs */
1303  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1304 
1305  /* XXX currently we assume TID quals are a subset of qpquals */
1306  startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1307  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1308  tid_qual_cost.per_tuple;
1309  run_cost += cpu_per_tuple * ntuples;
1310 
1311  /* tlist eval costs are paid per output row, not per tuple scanned */
1312  startup_cost += path->pathtarget->cost.startup;
1313  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1314 
1315  path->startup_cost = startup_cost;
1316  path->total_cost = startup_cost + run_cost;
1317 }
QualCost baserestrictcost
Definition: pathnodes.h:966

References ScalarArrayOpExpr::args, Assert(), RelOptInfo::baserestrictcost, RestrictInfo::clause, cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_tidscan, estimate_array_length(), get_restriction_qual_cost(), get_tablespace_page_costs(), IsA, lfirst_node, lsecond, 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().

◆ cost_tuplesort()

static void cost_tuplesort ( Cost startup_cost,
Cost run_cost,
double  tuples,
int  width,
Cost  comparison_cost,
int  sort_mem,
double  limit_tuples 
)
static

Definition at line 1856 of file costsize.c.

1860 {
1861  double input_bytes = relation_byte_size(tuples, width);
1862  double output_bytes;
1863  double output_tuples;
1864  long sort_mem_bytes = sort_mem * 1024L;
1865 
1866  /*
1867  * We want to be sure the cost of a sort is never estimated as zero, even
1868  * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1869  */
1870  if (tuples < 2.0)
1871  tuples = 2.0;
1872 
1873  /* Include the default cost-per-comparison */
1874  comparison_cost += 2.0 * cpu_operator_cost;
1875 
1876  /* Do we have a useful LIMIT? */
1877  if (limit_tuples > 0 && limit_tuples < tuples)
1878  {
1879  output_tuples = limit_tuples;
1880  output_bytes = relation_byte_size(output_tuples, width);
1881  }
1882  else
1883  {
1884  output_tuples = tuples;
1885  output_bytes = input_bytes;
1886  }
1887 
1888  if (output_bytes > sort_mem_bytes)
1889  {
1890  /*
1891  * We'll have to use a disk-based sort of all the tuples
1892  */
1893  double npages = ceil(input_bytes / BLCKSZ);
1894  double nruns = input_bytes / sort_mem_bytes;
1895  double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1896  double log_runs;
1897  double npageaccesses;
1898 
1899  /*
1900  * CPU costs
1901  *
1902  * Assume about N log2 N comparisons
1903  */
1904  *startup_cost = comparison_cost * tuples * LOG2(tuples);
1905 
1906  /* Disk costs */
1907 
1908  /* Compute logM(r) as log(r) / log(M) */
1909  if (nruns > mergeorder)
1910  log_runs = ceil(log(nruns) / log(mergeorder));
1911  else
1912  log_runs = 1.0;
1913  npageaccesses = 2.0 * npages * log_runs;
1914  /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1915  *startup_cost += npageaccesses *
1916  (seq_page_cost * 0.75 + random_page_cost * 0.25);
1917  }
1918  else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1919  {
1920  /*
1921  * We'll use a bounded heap-sort keeping just K tuples in memory, for
1922  * a total number of tuple comparisons of N log2 K; but the constant
1923  * factor is a bit higher than for quicksort. Tweak it so that the
1924  * cost curve is continuous at the crossover point.
1925  */
1926  *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
1927  }
1928  else
1929  {
1930  /* We'll use plain quicksort on all the input tuples */
1931  *startup_cost = comparison_cost * tuples * LOG2(tuples);
1932  }
1933 
1934  /*
1935  * Also charge a small amount (arbitrarily set equal to operator cost) per
1936  * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1937  * doesn't do qual-checking or projection, so it has less overhead than
1938  * most plan nodes. Note it's correct to use tuples not output_tuples
1939  * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1940  * counting the LIMIT otherwise.
1941  */
1942  *run_cost = cpu_operator_cost * tuples;
1943 }
int tuplesort_merge_order(int64 allowedMem)
Definition: tuplesort.c:1801

References cpu_operator_cost, LOG2, random_page_cost, relation_byte_size(), seq_page_cost, and tuplesort_merge_order().

Referenced by cost_incremental_sort(), and cost_sort().

◆ cost_valuesscan()

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

Definition at line 1620 of file costsize.c.

1622 {
1623  Cost startup_cost = 0;
1624  Cost run_cost = 0;
1625  QualCost qpqual_cost;
1626  Cost cpu_per_tuple;
1627 
1628  /* Should only be applied to base relations that are values lists */
1629  Assert(baserel->relid > 0);
1630  Assert(baserel->rtekind == RTE_VALUES);
1631 
1632  /* Mark the path with the correct row estimate */
1633  if (param_info)
1634  path->rows = param_info->ppi_rows;
1635  else
1636  path->rows = baserel->rows;
1637 
1638  /*
1639  * For now, estimate list evaluation cost at one operator eval per list
1640  * (probably pretty bogus, but is it worth being smarter?)
1641  */
1642  cpu_per_tuple = cpu_operator_cost;
1643 
1644  /* Add scanning CPU costs */
1645  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1646 
1647  startup_cost += qpqual_cost.startup;
1648  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1649  run_cost += cpu_per_tuple * baserel->tuples;
1650 
1651  /* tlist eval costs are paid per output row, not per tuple scanned */
1652  startup_cost += path->pathtarget->cost.startup;
1653  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1654 
1655  path->startup_cost = startup_cost;
1656  path->total_cost = startup_cost + run_cost;
1657 }
@ RTE_VALUES
Definition: parsenodes.h:1018

References Assert(), cpu_operator_cost, cpu_tuple_cost, get_restriction_qual_cost(), 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().

◆ cost_windowagg()

void cost_windowagg ( Path path,
PlannerInfo root,
List windowFuncs,
WindowClause winclause,
Cost  input_startup_cost,
Cost  input_total_cost,
double  input_tuples 
)

Definition at line 3040 of file costsize.c.

3044 {
3045  Cost startup_cost;
3046  Cost total_cost;
3047  double startup_tuples;
3048  int numPartCols;
3049  int numOrderCols;
3050  ListCell *lc;
3051 
3052  numPartCols = list_length(winclause->partitionClause);
3053  numOrderCols = list_length(winclause->orderClause);
3054 
3055  startup_cost = input_startup_cost;
3056  total_cost = input_total_cost;
3057 
3058  /*
3059  * Window functions are assumed to cost their stated execution cost, plus
3060  * the cost of evaluating their input expressions, per tuple. Since they
3061  * may in fact evaluate their inputs at multiple rows during each cycle,
3062  * this could be a drastic underestimate; but without a way to know how
3063  * many rows the window function will fetch, it's hard to do better. In
3064  * any case, it's a good estimate for all the built-in window functions,
3065  * so we'll just do this for now.
3066  */
3067  foreach(lc, windowFuncs)
3068  {
3069  WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3070  Cost wfunccost;
3071  QualCost argcosts;
3072 
3073  argcosts.startup = argcosts.per_tuple = 0;
3074  add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3075  &argcosts);
3076  startup_cost += argcosts.startup;
3077  wfunccost = argcosts.per_tuple;
3078 
3079  /* also add the input expressions' cost to per-input-row costs */
3080  cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3081  startup_cost += argcosts.startup;
3082  wfunccost += argcosts.per_tuple;
3083 
3084  /*
3085  * Add the filter's cost to per-input-row costs. XXX We should reduce
3086  * input expression costs according to filter selectivity.
3087  */
3088  cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3089  startup_cost += argcosts.startup;
3090  wfunccost += argcosts.per_tuple;
3091 
3092  total_cost += wfunccost * input_tuples;
3093  }
3094 
3095  /*
3096  * We also charge cpu_operator_cost per grouping column per tuple for
3097  * grouping comparisons, plus cpu_tuple_cost per tuple for general
3098  * overhead.
3099  *
3100  * XXX this neglects costs of spooling the data to disk when it overflows
3101  * work_mem. Sooner or later that should get accounted for.
3102  */
3103  total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3104  total_cost += cpu_tuple_cost * input_tuples;
3105 
3106  path->rows = input_tuples;
3107  path->startup_cost = startup_cost;
3108  path->total_cost = total_cost;
3109 
3110  /*
3111  * Also, take into account how many tuples we need to read from the
3112  * subnode in order to produce the first tuple from the WindowAgg. To do
3113  * this we proportion the run cost (total cost not including startup cost)
3114  * over the estimated startup tuples. We already included the startup
3115  * cost of the subnode, so we only need to do this when the estimated
3116  * startup tuples is above 1.0.
3117  */
3118  startup_tuples = get_windowclause_startup_tuples(root, winclause,
3119  input_tuples);
3120 
3121  if (startup_tuples > 1.0)
3122  path->startup_cost += (total_cost - startup_cost) / input_tuples *
3123  (startup_tuples - 1.0);
3124 }
static double get_windowclause_startup_tuples(PlannerInfo *root, WindowClause *wc, double input_tuples)
Definition: costsize.c:2826
List * partitionClause
Definition: parsenodes.h:1496
List * orderClause
Definition: parsenodes.h:1498
List * args
Definition: primnodes.h:553
Expr * aggfilter
Definition: primnodes.h:555
Oid winfnoid
Definition: primnodes.h:545

References add_function_cost(), WindowFunc::aggfilter, WindowFunc::args, cost_qual_eval_node(), cpu_operator_cost, cpu_tuple_cost, get_windowclause_startup_tuples(), lfirst_node, list_length(), WindowClause::orderClause, WindowClause::partitionClause, QualCost::per_tuple, Path::rows, QualCost::startup, Path::startup_cost, Path::total_cost, and WindowFunc::winfnoid.

Referenced by create_windowagg_path().

◆ extract_nonindex_conditions()

static List * extract_nonindex_conditions ( List qual_clauses,
List indexclauses 
)
static

Definition at line 812 of file costsize.c.

813 {
814  List *result = NIL;
815  ListCell *lc;
816 
817  foreach(lc, qual_clauses)
818  {
819  RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
820 
821  if (rinfo->pseudoconstant)
822  continue; /* we may drop pseudoconstants here */
823  if (is_redundant_with_indexclauses(rinfo, indexclauses))
824  continue; /* dup or derived from same EquivalenceClass */
825  /* ... skip the predicate proof attempt createplan.c will try ... */
826  result = lappend(result, rinfo);
827  }
828  return result;
829 }
bool is_redundant_with_indexclauses(RestrictInfo *rinfo, List *indexclauses)
Definition: equivclass.c:3221

References is_redundant_with_indexclauses(), lappend(), lfirst_node, and NIL.

Referenced by cost_index().

◆ final_cost_hashjoin()

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

Definition at line 4153 of file costsize.c.

4156 {
4157  Path *outer_path = path->jpath.outerjoinpath;
4158  Path *inner_path = path->jpath.innerjoinpath;
4159  double outer_path_rows = outer_path->rows;
4160  double inner_path_rows = inner_path->rows;
4161  double inner_path_rows_total = workspace->inner_rows_total;
4162  List *hashclauses = path->path_hashclauses;
4163  Cost startup_cost = workspace->startup_cost;
4164  Cost run_cost = workspace->run_cost;
4165  int numbuckets = workspace->numbuckets;
4166  int numbatches = workspace->numbatches;
4167  Cost cpu_per_tuple;
4168  QualCost hash_qual_cost;
4169  QualCost qp_qual_cost;
4170  double hashjointuples;
4171  double virtualbuckets;
4172  Selectivity innerbucketsize;
4173  Selectivity innermcvfreq;
4174  ListCell *hcl;
4175 
4176  /* Mark the path with the correct row estimate */
4177  if (path->jpath.path.param_info)
4178  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4179  else
4180  path->jpath.path.rows = path->jpath.path.parent->rows;
4181 
4182  /* For partial paths, scale row estimate. */
4183  if (path->jpath.path.parallel_workers > 0)
4184  {
4185  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4186 
4187  path->jpath.path.rows =
4188  clamp_row_est(path->jpath.path.rows / parallel_divisor);
4189  }
4190 
4191  /*
4192  * We could include disable_cost in the preliminary estimate, but that
4193  * would amount to optimizing for the case where the join method is
4194  * disabled, which doesn't seem like the way to bet.
4195  */
4196  if (!enable_hashjoin)
4197  startup_cost += disable_cost;
4198 
4199  /* mark the path with estimated # of batches */
4200  path->num_batches = numbatches;
4201 
4202  /* store the total number of tuples (sum of partial row estimates) */
4203  path->inner_rows_total = inner_path_rows_total;
4204 
4205  /* and compute the number of "virtual" buckets in the whole join */
4206  virtualbuckets = (double) numbuckets * (double) numbatches;
4207 
4208  /*
4209  * Determine bucketsize fraction and MCV frequency for the inner relation.
4210  * We use the smallest bucketsize or MCV frequency estimated for any
4211  * individual hashclause; this is undoubtedly conservative.
4212  *
4213  * BUT: if inner relation has been unique-ified, we can assume it's good
4214  * for hashing. This is important both because it's the right answer, and
4215  * because we avoid contaminating the cache with a value that's wrong for
4216  * non-unique-ified paths.
4217  */
4218  if (IsA(inner_path, UniquePath))
4219  {
4220  innerbucketsize = 1.0 / virtualbuckets;
4221  innermcvfreq = 0.0;
4222  }
4223  else
4224  {
4225  innerbucketsize = 1.0;
4226  innermcvfreq = 1.0;
4227  foreach(hcl, hashclauses)
4228  {
4229  RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4230  Selectivity thisbucketsize;
4231  Selectivity thismcvfreq;
4232 
4233  /*
4234  * First we have to figure out which side of the hashjoin clause
4235  * is the inner side.
4236  *
4237  * Since we tend to visit the same clauses over and over when
4238  * planning a large query, we cache the bucket stats estimates in
4239  * the RestrictInfo node to avoid repeated lookups of statistics.
4240  */
4241  if (bms_is_subset(restrictinfo->right_relids,
4242  inner_path->parent->relids))
4243  {
4244  /* righthand side is inner */
4245  thisbucketsize = restrictinfo->right_bucketsize;
4246  if (thisbucketsize < 0)
4247  {
4248  /* not cached yet */
4250  get_rightop(restrictinfo->clause),
4251  virtualbuckets,
4252  &restrictinfo->right_mcvfreq,
4253  &restrictinfo->right_bucketsize);
4254  thisbucketsize = restrictinfo->right_bucketsize;
4255  }
4256  thismcvfreq = restrictinfo->right_mcvfreq;
4257  }
4258  else
4259  {
4260  Assert(bms_is_subset(restrictinfo->left_relids,
4261  inner_path->parent->relids));
4262  /* lefthand side is inner */
4263  thisbucketsize = restrictinfo->left_bucketsize;
4264  if (thisbucketsize < 0)
4265  {
4266  /* not cached yet */
4268  get_leftop(restrictinfo->clause),
4269  virtualbuckets,
4270  &restrictinfo->left_mcvfreq,
4271  &restrictinfo->left_bucketsize);
4272  thisbucketsize = restrictinfo->left_bucketsize;
4273  }
4274  thismcvfreq = restrictinfo->left_mcvfreq;
4275  }
4276 
4277  if (innerbucketsize > thisbucketsize)
4278  innerbucketsize = thisbucketsize;
4279  if (innermcvfreq > thismcvfreq)
4280  innermcvfreq = thismcvfreq;
4281  }
4282  }
4283 
4284  /*
4285  * If the bucket holding the inner MCV would exceed hash_mem, we don't
4286  * want to hash unless there is really no other alternative, so apply
4287  * disable_cost. (The executor normally copes with excessive memory usage
4288  * by splitting batches, but obviously it cannot separate equal values
4289  * that way, so it will be unable to drive the batch size below hash_mem
4290  * when this is true.)
4291  */
4292  if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4293  inner_path->pathtarget->width) > get_hash_memory_limit())
4294  startup_cost += disable_cost;
4295 
4296  /*
4297  * Compute cost of the hashquals and qpquals (other restriction clauses)
4298  * separately.
4299  */
4300  cost_qual_eval(&hash_qual_cost, hashclauses, root);
4301  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4302  qp_qual_cost.startup -= hash_qual_cost.startup;
4303  qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4304 
4305  /* CPU costs */
4306 
4307  if (path->jpath.jointype == JOIN_SEMI ||
4308  path->jpath.jointype == JOIN_ANTI ||
4309  extra->inner_unique)
4310  {
4311  double outer_matched_rows;
4312  Selectivity inner_scan_frac;
4313 
4314  /*
4315  * With a SEMI or ANTI join, or if the innerrel is known unique, the
4316  * executor will stop after the first match.
4317  *
4318  * For an outer-rel row that has at least one match, we can expect the
4319  * bucket scan to stop after a fraction 1/(match_count+1) of the
4320  * bucket's rows, if the matches are evenly distributed. Since they
4321  * probably aren't quite evenly distributed, we apply a fuzz factor of
4322  * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4323  * to clamp inner_scan_frac to at most 1.0; but since match_count is
4324  * at least 1, no such clamp is needed now.)
4325  */
4326  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4327  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4328 
4329  startup_cost += hash_qual_cost.startup;
4330  run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4331  clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4332 
4333  /*
4334  * For unmatched outer-rel rows, the picture is quite a lot different.
4335  * In the first place, there is no reason to assume that these rows
4336  * preferentially hit heavily-populated buckets; instead assume they
4337  * are uncorrelated with the inner distribution and so they see an
4338  * average bucket size of inner_path_rows / virtualbuckets. In the
4339  * second place, it seems likely that they will have few if any exact
4340  * hash-code matches and so very few of the tuples in the bucket will
4341  * actually require eval of the hash quals. We don't have any good
4342  * way to estimate how many will, but for the moment assume that the
4343  * effective cost per bucket entry is one-tenth what it is for
4344  * matchable tuples.
4345  */
4346  run_cost += hash_qual_cost.per_tuple *
4347  (outer_path_rows - outer_matched_rows) *
4348  clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4349 
4350  /* Get # of tuples that will pass the basic join */
4351  if (path->jpath.jointype == JOIN_ANTI)
4352  hashjointuples = outer_path_rows - outer_matched_rows;
4353  else
4354  hashjointuples = outer_matched_rows;
4355  }
4356  else
4357  {
4358  /*
4359  * The number of tuple comparisons needed is the number of outer
4360  * tuples times the typical number of tuples in a hash bucket, which
4361  * is the inner relation size times its bucketsize fraction. At each
4362  * one, we need to evaluate the hashjoin quals. But actually,
4363  * charging the full qual eval cost at each tuple is pessimistic,
4364  * since we don't evaluate the quals unless the hash values match
4365  * exactly. For lack of a better idea, halve the cost estimate to
4366  * allow for that.
4367  */
4368  startup_cost += hash_qual_cost.startup;
4369  run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4370  clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4371 
4372  /*
4373  * Get approx # tuples passing the hashquals. We use
4374  * approx_tuple_count here because we need an estimate done with
4375  * JOIN_INNER semantics.
4376  */
4377  hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4378  }
4379 
4380  /*
4381  * For each tuple that gets through the hashjoin proper, we charge
4382  * cpu_tuple_cost plus the cost of evaluating additional restriction
4383  * clauses that are to be applied at the join. (This is pessimistic since
4384  * not all of the quals may get evaluated at each tuple.)
4385  */
4386  startup_cost += qp_qual_cost.startup;
4387  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4388  run_cost += cpu_per_tuple * hashjointuples;
4389 
4390  /* tlist eval costs are paid per output row, not per tuple scanned */
4391  startup_cost += path->jpath.path.pathtarget->cost.startup;
4392  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4393 
4394  path->jpath.path.startup_cost = startup_cost;
4395  path->jpath.path.total_cost = startup_cost + run_cost;
4396 }
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:363
bool enable_hashjoin
Definition: costsize.c:147
static double approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
Definition: costsize.c:5181
static Node * get_rightop(const void *clause)
Definition: nodeFuncs.h:93
static Node * get_leftop(const void *clause)
Definition: nodeFuncs.h:81
void estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets, Selectivity *mcv_freq, Selectivity *bucketsize_frac)
Definition: selfuncs.c:3768
List * path_hashclauses
Definition: pathnodes.h:2120
Cardinality inner_rows_total
Definition: pathnodes.h:2122
int num_batches
Definition: pathnodes.h:2121
JoinPath jpath
Definition: pathnodes.h:2119
Cardinality inner_rows_total
Definition: pathnodes.h:3314
SemiAntiJoinFactors semifactors
Definition: pathnodes.h:3193
JoinType jointype
Definition: pathnodes.h:2037
List * joinrestrictinfo
Definition: pathnodes.h:2045

References approx_tuple_count(), Assert(), bms_is_subset(), clamp_row_est(), RestrictInfo::clause, cost_qual_eval(), cpu_tuple_cost, disable_cost, enable_hashjoin, estimate_hash_bucket_stats(), get_hash_memory_limit(), get_leftop(), get_parallel_divisor(), get_rightop(), HashPath::inner_rows_total, JoinCostWorkspace::inner_rows_total, JoinPathExtraData::inner_unique, JoinPath::innerjoinpath, IsA, JOIN_ANTI, JOIN_SEMI, JoinPath::joinrestrictinfo, JoinPath::jointype, HashPath::jpath, lfirst_node, SemiAntiJoinFactors::match_count, HashPath::num_batches, JoinCostWorkspace::numbatches, JoinCostWorkspace::numbuckets, SemiAntiJoinFactors::outer_match_frac, JoinPath::outerjoinpath, HashPath::path_hashclauses, QualCost::per_tuple, relation_byte_size(), Path::rows, JoinCostWorkspace::run_cost, JoinPathExtraData::semifactors, QualCost::startup, and JoinCostWorkspace::startup_cost.

Referenced by create_hashjoin_path().

◆ final_cost_mergejoin()

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

Definition at line 3717 of file costsize.c.

3720 {
3721  Path *outer_path = path->jpath.outerjoinpath;
3722  Path *inner_path = path->jpath.innerjoinpath;
3723  double inner_path_rows = inner_path->rows;
3724  List *mergeclauses = path->path_mergeclauses;
3725  List *innersortkeys = path->innersortkeys;
3726  Cost startup_cost = workspace->startup_cost;
3727  Cost run_cost = workspace->run_cost;
3728  Cost inner_run_cost = workspace->inner_run_cost;
3729  double outer_rows = workspace->outer_rows;
3730  double inner_rows = workspace->inner_rows;
3731  double outer_skip_rows = workspace->outer_skip_rows;
3732  double inner_skip_rows = workspace->inner_skip_rows;
3733  Cost cpu_per_tuple,
3734  bare_inner_cost,
3735  mat_inner_cost;
3736  QualCost merge_qual_cost;
3737  QualCost qp_qual_cost;
3738  double mergejointuples,
3739  rescannedtuples;
3740  double rescanratio;
3741 
3742  /* Protect some assumptions below that rowcounts aren't zero */
3743  if (inner_path_rows <= 0)
3744  inner_path_rows = 1;
3745 
3746  /* Mark the path with the correct row estimate */
3747  if (path->jpath.path.param_info)
3748  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3749  else
3750  path->jpath.path.rows = path->jpath.path.parent->rows;
3751 
3752  /* For partial paths, scale row estimate. */
3753  if (path->jpath.path.parallel_workers > 0)
3754  {
3755  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3756 
3757  path->jpath.path.rows =
3758  clamp_row_est(path->jpath.path.rows / parallel_divisor);
3759  }
3760 
3761  /*
3762  * We could include disable_cost in the preliminary estimate, but that
3763  * would amount to optimizing for the case where the join method is
3764  * disabled, which doesn't seem like the way to bet.
3765  */
3766  if (!enable_mergejoin)
3767  startup_cost += disable_cost;
3768 
3769  /*
3770  * Compute cost of the mergequals and qpquals (other restriction clauses)
3771  * separately.
3772  */
3773  cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3774  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3775  qp_qual_cost.startup -= merge_qual_cost.startup;
3776  qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3777 
3778  /*
3779  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3780  * executor will stop scanning for matches after the first match. When
3781  * all the joinclauses are merge clauses, this means we don't ever need to
3782  * back up the merge, and so we can skip mark/restore overhead.
3783  */
3784  if ((path->jpath.jointype == JOIN_SEMI ||
3785  path->jpath.jointype == JOIN_ANTI ||
3786  extra->inner_unique) &&
3789  path->skip_mark_restore = true;
3790  else
3791  path->skip_mark_restore = false;
3792 
3793  /*
3794  * Get approx # tuples passing the mergequals. We use approx_tuple_count
3795  * here because we need an estimate done with JOIN_INNER semantics.
3796  */
3797  mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3798 
3799  /*
3800  * When there are equal merge keys in the outer relation, the mergejoin
3801  * must rescan any matching tuples in the inner relation. This means
3802  * re-fetching inner tuples; we have to estimate how often that happens.
3803  *
3804  * For regular inner and outer joins, the number of re-fetches can be
3805  * estimated approximately as size of merge join output minus size of
3806  * inner relation. Assume that the distinct key values are 1, 2, ..., and
3807  * denote the number of values of each key in the outer relation as m1,
3808  * m2, ...; in the inner relation, n1, n2, ... Then we have
3809  *
3810  * size of join = m1 * n1 + m2 * n2 + ...
3811  *
3812  * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3813  * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3814  * relation
3815  *
3816  * This equation works correctly for outer tuples having no inner match
3817  * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3818  * are effectively subtracting those from the number of rescanned tuples,
3819  * when we should not. Can we do better without expensive selectivity
3820  * computations?
3821  *
3822  * The whole issue is moot if we are working from a unique-ified outer
3823  * input, or if we know we don't need to mark/restore at all.
3824  */
3825  if (IsA(outer_path, UniquePath) || path->skip_mark_restore)
3826  rescannedtuples = 0;
3827  else
3828  {
3829  rescannedtuples = mergejointuples - inner_path_rows;
3830  /* Must clamp because of possible underestimate */
3831  if (rescannedtuples < 0)
3832  rescannedtuples = 0;
3833  }
3834 
3835  /*
3836  * We'll inflate various costs this much to account for rescanning. Note
3837  * that this is to be multiplied by something involving inner_rows, or
3838  * another number related to the portion of the inner rel we'll scan.
3839  */
3840  rescanratio = 1.0 + (rescannedtuples / inner_rows);
3841 
3842  /*
3843  * Decide whether we want to materialize the inner input to shield it from
3844  * mark/restore and performing re-fetches. Our cost model for regular
3845  * re-fetches is that a re-fetch costs the same as an original fetch,
3846  * which is probably an overestimate; but on the other hand we ignore the
3847  * bookkeeping costs of mark/restore. Not clear if it's worth developing
3848  * a more refined model. So we just need to inflate the inner run cost by
3849  * rescanratio.
3850  */
3851  bare_inner_cost = inner_run_cost * rescanratio;
3852 
3853  /*
3854  * When we interpose a Material node the re-fetch cost is assumed to be
3855  * just cpu_operator_cost per tuple, independently of the underlying
3856  * plan's cost; and we charge an extra cpu_operator_cost per original
3857  * fetch as well. Note that we're assuming the materialize node will
3858  * never spill to disk, since it only has to remember tuples back to the
3859  * last mark. (If there are a huge number of duplicates, our other cost
3860  * factors will make the path so expensive that it probably won't get
3861  * chosen anyway.) So we don't use cost_rescan here.
3862  *
3863  * Note: keep this estimate in sync with create_mergejoin_plan's labeling
3864  * of the generated Material node.
3865  */
3866  mat_inner_cost = inner_run_cost +
3867  cpu_operator_cost * inner_rows * rescanratio;
3868 
3869  /*
3870  * If we don't need mark/restore at all, we don't need materialization.
3871  */
3872  if (path->skip_mark_restore)
3873  path->materialize_inner = false;
3874 
3875  /*
3876  * Prefer materializing if it looks cheaper, unless the user has asked to
3877  * suppress materialization.
3878  */
3879  else if (enable_material && mat_inner_cost < bare_inner_cost)
3880  path->materialize_inner = true;
3881 
3882  /*
3883  * Even if materializing doesn't look cheaper, we *must* do it if the
3884  * inner path is to be used directly (without sorting) and it doesn't
3885  * support mark/restore.
3886  *
3887  * Since the inner side must be ordered, and only Sorts and IndexScans can
3888  * create order to begin with, and they both support mark/restore, you
3889  * might think there's no problem --- but you'd be wrong. Nestloop and
3890  * merge joins can *preserve* the order of their inputs, so they can be
3891  * selected as the input of a mergejoin, and they don't support
3892  * mark/restore at present.
3893  *
3894  * We don't test the value of enable_material here, because
3895  * materialization is required for correctness in this case, and turning
3896  * it off does not entitle us to deliver an invalid plan.
3897  */
3898  else if (innersortkeys == NIL &&
3899  !ExecSupportsMarkRestore(inner_path))
3900  path->materialize_inner = true;
3901 
3902  /*
3903  * Also, force materializing if the inner path is to be sorted and the
3904  * sort is expected to spill to disk. This is because the final merge
3905  * pass can be done on-the-fly if it doesn't have to support mark/restore.
3906  * We don't try to adjust the cost estimates for this consideration,
3907  * though.
3908  *
3909  * Since materialization is a performance optimization in this case,
3910  * rather than necessary for correctness, we skip it if enable_material is
3911  * off.
3912  */
3913  else if (enable_material && innersortkeys != NIL &&
3914  relation_byte_size(inner_path_rows,
3915  inner_path->pathtarget->width) >
3916  (work_mem * 1024L))
3917  path->materialize_inner = true;
3918  else
3919  path->materialize_inner = false;
3920 
3921  /* Charge the right incremental cost for the chosen case */
3922  if (path->materialize_inner)
3923  run_cost += mat_inner_cost;
3924  else
3925  run_cost += bare_inner_cost;
3926 
3927  /* CPU costs */
3928 
3929  /*
3930  * The number of tuple comparisons needed is approximately number of outer
3931  * rows plus number of inner rows plus number of rescanned tuples (can we
3932  * refine this?). At each one, we need to evaluate the mergejoin quals.
3933  */
3934  startup_cost += merge_qual_cost.startup;
3935  startup_cost += merge_qual_cost.per_tuple *
3936  (outer_skip_rows + inner_skip_rows * rescanratio);
3937  run_cost += merge_qual_cost.per_tuple *
3938  ((outer_rows - outer_skip_rows) +
3939  (inner_rows - inner_skip_rows) * rescanratio);
3940 
3941  /*
3942  * For each tuple that gets through the mergejoin proper, we charge
3943  * cpu_tuple_cost plus the cost of evaluating additional restriction
3944  * clauses that are to be applied at the join. (This is pessimistic since
3945  * not all of the quals may get evaluated at each tuple.)
3946  *
3947  * Note: we could adjust for SEMI/ANTI joins skipping some qual
3948  * evaluations here, but it's probably not worth the trouble.
3949  */
3950  startup_cost += qp_qual_cost.startup;
3951  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
3952  run_cost += cpu_per_tuple * mergejointuples;
3953 
3954  /* tlist eval costs are paid per output row, not per tuple scanned */
3955  startup_cost += path->jpath.path.pathtarget->cost.startup;
3956  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3957 
3958  path->jpath.path.startup_cost = startup_cost;
3959  path->jpath.path.total_cost = startup_cost + run_cost;
3960 }
bool enable_material
Definition: costsize.c:144
bool enable_mergejoin
Definition: costsize.c:146
bool ExecSupportsMarkRestore(Path *pathnode)
Definition: execAmi.c:419
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:77
Cardinality inner_rows
Definition: pathnodes.h:3307
Cardinality outer_rows
Definition: pathnodes.h:3306
Cardinality inner_skip_rows
Definition: pathnodes.h:3309
Cardinality outer_skip_rows
Definition: pathnodes.h:3308
bool skip_mark_restore
Definition: pathnodes.h:2104
List * innersortkeys
Definition: pathnodes.h:2103
JoinPath jpath
Definition: pathnodes.h:2100
bool materialize_inner
Definition: pathnodes.h:2105
List * path_mergeclauses
Definition: pathnodes.h:2101

References approx_tuple_count(), clamp_row_est(), cost_qual_eval(), cpu_operator_cost, cpu_tuple_cost, disable_cost, enable_material, enable_mergejoin, ExecSupportsMarkRestore(), get_parallel_divisor(), if(), 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, MergePath::path_mergeclauses, QualCost::per_tuple, relation_byte_size(), Path::rows, JoinCostWorkspace::run_cost, MergePath::skip_mark_restore, QualCost::startup, JoinCostWorkspace::startup_cost, and work_mem.

Referenced by create_mergejoin_path().

◆ final_cost_nestloop()

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

Definition at line 3280 of file costsize.c.

3283 {
3284  Path *outer_path = path->jpath.outerjoinpath;
3285  Path *inner_path = path->jpath.innerjoinpath;
3286  double outer_path_rows = outer_path->rows;
3287  double inner_path_rows = inner_path->rows;
3288  Cost startup_cost = workspace->startup_cost;
3289  Cost run_cost = workspace->run_cost;
3290  Cost cpu_per_tuple;
3291  QualCost restrict_qual_cost;
3292  double ntuples;
3293 
3294  /* Protect some assumptions below that rowcounts aren't zero */
3295  if (outer_path_rows <= 0)
3296  outer_path_rows = 1;
3297  if (inner_path_rows <= 0)
3298  inner_path_rows = 1;
3299  /* Mark the path with the correct row estimate */
3300  if (path->jpath.path.param_info)
3301  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3302  else
3303  path->jpath.path.rows = path->jpath.path.parent->rows;
3304 
3305  /* For partial paths, scale row estimate. */
3306  if (path->jpath.path.parallel_workers > 0)
3307  {
3308  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3309 
3310  path->jpath.path.rows =
3311  clamp_row_est(path->jpath.path.rows / parallel_divisor);
3312  }
3313 
3314  /*
3315  * We could include disable_cost in the preliminary estimate, but that
3316  * would amount to optimizing for the case where the join method is
3317  * disabled, which doesn't seem like the way to bet.
3318  */
3319  if (!enable_nestloop)
3320  startup_cost += disable_cost;
3321 
3322  /* cost of inner-relation source data (we already dealt with outer rel) */
3323 
3324  if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3325  extra->inner_unique)
3326  {
3327  /*
3328  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3329  * executor will stop after the first match.
3330  */
3331  Cost inner_run_cost = workspace->inner_run_cost;
3332  Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3333  double outer_matched_rows;
3334  double outer_unmatched_rows;
3335  Selectivity inner_scan_frac;
3336 
3337  /*
3338  * For an outer-rel row that has at least one match, we can expect the
3339  * inner scan to stop after a fraction 1/(match_count+1) of the inner
3340  * rows, if the matches are evenly distributed. Since they probably
3341  * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3342  * that fraction. (If we used a larger fuzz factor, we'd have to
3343  * clamp inner_scan_frac to at most 1.0; but since match_count is at
3344  * least 1, no such clamp is needed now.)
3345  */
3346  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3347  outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3348  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3349 
3350  /*
3351  * Compute number of tuples processed (not number emitted!). First,
3352  * account for successfully-matched outer rows.
3353  */
3354  ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3355 
3356  /*
3357  * Now we need to estimate the actual costs of scanning the inner
3358  * relation, which may be quite a bit less than N times inner_run_cost
3359  * due to early scan stops. We consider two cases. If the inner path
3360  * is an indexscan using all the joinquals as indexquals, then an
3361  * unmatched outer row results in an indexscan returning no rows,
3362  * which is probably quite cheap. Otherwise, the executor will have
3363  * to scan the whole inner rel for an unmatched row; not so cheap.
3364  */
3365  if (has_indexed_join_quals(path))
3366  {
3367  /*
3368  * Successfully-matched outer rows will only require scanning
3369  * inner_scan_frac of the inner relation. In this case, we don't
3370  * need to charge the full inner_run_cost even when that's more
3371  * than inner_rescan_run_cost, because we can assume that none of
3372  * the inner scans ever scan the whole inner relation. So it's
3373  * okay to assume that all the inner scan executions can be
3374  * fractions of the full cost, even if materialization is reducing
3375  * the rescan cost. At this writing, it's impossible to get here
3376  * for a materialized inner scan, so inner_run_cost and
3377  * inner_rescan_run_cost will be the same anyway; but just in
3378  * case, use inner_run_cost for the first matched tuple and
3379  * inner_rescan_run_cost for additional ones.
3380  */
3381  run_cost += inner_run_cost * inner_scan_frac;
3382  if (outer_matched_rows > 1)
3383  run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3384 
3385  /*
3386  * Add the cost of inner-scan executions for unmatched outer rows.
3387  * We estimate this as the same cost as returning the first tuple
3388  * of a nonempty scan. We consider that these are all rescans,
3389  * since we used inner_run_cost once already.
3390  */
3391  run_cost += outer_unmatched_rows *
3392  inner_rescan_run_cost / inner_path_rows;
3393 
3394  /*
3395  * We won't be evaluating any quals at all for unmatched rows, so
3396  * don't add them to ntuples.
3397  */
3398  }
3399  else
3400  {
3401  /*
3402  * Here, a complicating factor is that rescans may be cheaper than
3403  * first scans. If we never scan all the way to the end of the
3404  * inner rel, it might be (depending on the plan type) that we'd
3405  * never pay the whole inner first-scan run cost. However it is
3406  * difficult to estimate whether that will happen (and it could
3407  * not happen if there are any unmatched outer rows!), so be
3408  * conservative and always charge the whole first-scan cost once.
3409  * We consider this charge to correspond to the first unmatched
3410  * outer row, unless there isn't one in our estimate, in which
3411  * case blame it on the first matched row.
3412  */
3413 
3414  /* First, count all unmatched join tuples as being processed */
3415  ntuples += outer_unmatched_rows * inner_path_rows;
3416 
3417  /* Now add the forced full scan, and decrement appropriate count */
3418  run_cost += inner_run_cost;
3419  if (outer_unmatched_rows >= 1)
3420  outer_unmatched_rows -= 1;
3421  else
3422  outer_matched_rows -= 1;
3423 
3424  /* Add inner run cost for additional outer tuples having matches */
3425  if (outer_matched_rows > 0)
3426  run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3427 
3428  /* Add inner run cost for additional unmatched outer tuples */
3429  if (outer_unmatched_rows > 0)
3430  run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3431  }
3432  }
3433  else
3434  {
3435  /* Normal-case source costs were included in preliminary estimate */
3436 
3437  /* Compute number of tuples processed (not number emitted!) */
3438  ntuples = outer_path_rows * inner_path_rows;
3439  }
3440 
3441  /* CPU costs */
3442  cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3443  startup_cost += restrict_qual_cost.startup;
3444  cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3445  run_cost += cpu_per_tuple * ntuples;
3446 
3447  /* tlist eval costs are paid per output row, not per tuple scanned */
3448  startup_cost += path->jpath.path.pathtarget->cost.startup;
3449  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3450 
3451  path->jpath.path.startup_cost = startup_cost;
3452  path->jpath.path.total_cost = startup_cost + run_cost;
3453 }
static bool has_indexed_join_quals(NestPath *path)
Definition: costsize.c:5088
bool enable_nestloop
Definition: costsize.c:143
Cost inner_rescan_run_cost
Definition: pathnodes.h:3303
JoinPath jpath
Definition: pathnodes.h:2060

References clamp_row_est(), 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, NestPath::jpath, SemiAntiJoinFactors::match_count, SemiAntiJoinFactors::outer_match_frac, JoinPath::outerjoinpath, QualCost::per_tuple, Path::rows, JoinCostWorkspace::run_cost, JoinPathExtraData::semifactors, QualCost::startup, and JoinCostWorkspace::startup_cost.

Referenced by create_nestloop_path().

◆ get_expr_width()

static int32 get_expr_width ( PlannerInfo root,
const Node expr 
)
static

Definition at line 6298 of file costsize.c.

6299 {
6300  int32 width;
6301 
6302  if (IsA(expr, Var))
6303  {
6304  const Var *var = (const Var *) expr;
6305 
6306  /* We should not see any upper-level Vars here */
6307  Assert(var->varlevelsup == 0);
6308 
6309  /* Try to get data from RelOptInfo cache */
6310  if (!IS_SPECIAL_VARNO(var->varno) &&
6311  var->varno < root->simple_rel_array_size)
6312  {
6313  RelOptInfo *rel = root->simple_rel_array[var->varno];
6314 
6315  if (rel != NULL &&
6316  var->varattno >= rel->min_attr &&
6317  var->varattno <= rel->max_attr)
6318  {
6319  int ndx = var->varattno - rel->min_attr;
6320 
6321  if (rel->attr_widths[ndx] > 0)
6322  return rel->attr_widths[ndx];
6323  }
6324  }
6325 
6326  /*
6327  * No cached data available, so estimate using just the type info.
6328  */
6329  width = get_typavgwidth(var->vartype, var->vartypmod);
6330  Assert(width > 0);
6331 
6332  return width;
6333  }
6334 
6335  width = get_typavgwidth(exprType(expr), exprTypmod(expr));
6336  Assert(width > 0);
6337  return width;
6338 }
signed int int32
Definition: c.h:483
int32 get_typavgwidth(Oid typid, int32 typmod)
Definition: lsyscache.c:2560
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:282
#define IS_SPECIAL_VARNO(varno)
Definition: primnodes.h:219
int simple_rel_array_size
Definition: pathnodes.h:229
AttrNumber max_attr
Definition: pathnodes.h:911
AttrNumber min_attr
Definition: pathnodes.h:909
Definition: primnodes.h:226
AttrNumber varattno
Definition: primnodes.h:238
int varno
Definition: primnodes.h:233
Index varlevelsup
Definition: primnodes.h:258

References Assert(), exprType(), exprTypmod(), get_typavgwidth(), IS_SPECIAL_VARNO, IsA, RelOptInfo::max_attr, RelOptInfo::min_attr, PlannerInfo::simple_rel_array_size, Var::varattno, Var::varlevelsup, and Var::varno.

Referenced by cost_memoize_rescan(), and set_pathtarget_cost_width().

◆ get_foreign_key_join_selectivity()

static Selectivity get_foreign_key_join_selectivity ( PlannerInfo root,
Relids  outer_relids,
Relids  inner_relids,
SpecialJoinInfo sjinfo,
List **  restrictlist 
)
static

Definition at line 5543 of file costsize.c.

5548 {
5549  Selectivity fkselec = 1.0;
5550  JoinType jointype = sjinfo->jointype;
5551  List *worklist = *restrictlist;
5552  ListCell *lc;
5553 
5554  /* Consider each FK constraint that is known to match the query */
5555  foreach(lc, root->fkey_list)
5556  {
5557  ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5558  bool ref_is_outer;
5559  List *removedlist;
5560  ListCell *cell;
5561 
5562  /*
5563  * This FK is not relevant unless it connects a baserel on one side of
5564  * this join to a baserel on the other side.
5565  */
5566  if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5567  bms_is_member(fkinfo->ref_relid, inner_relids))
5568  ref_is_outer = false;
5569  else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5570  bms_is_member(fkinfo->con_relid, inner_relids))
5571  ref_is_outer = true;
5572  else
5573  continue;
5574 
5575  /*
5576  * If we're dealing with a semi/anti join, and the FK's referenced
5577  * relation is on the outside, then knowledge of the FK doesn't help
5578  * us figure out what we need to know (which is the fraction of outer
5579  * rows that have matches). On the other hand, if the referenced rel
5580  * is on the inside, then all outer rows must have matches in the
5581  * referenced table (ignoring nulls). But any restriction or join
5582  * clauses that filter that table will reduce the fraction of matches.
5583  * We can account for restriction clauses, but it's too hard to guess
5584  * how many table rows would get through a join that's inside the RHS.
5585  * Hence, if either case applies, punt and ignore the FK.
5586  */
5587  if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5588  (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5589  continue;
5590 
5591  /*
5592  * Modify the restrictlist by removing clauses that match the FK (and
5593  * putting them into removedlist instead). It seems unsafe to modify
5594  * the originally-passed List structure, so we make a shallow copy the
5595  * first time through.
5596  */
5597  if (worklist == *restrictlist)
5598  worklist = list_copy(worklist);
5599 
5600  removedlist = NIL;
5601  foreach(cell, worklist)
5602  {
5603  RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5604  bool remove_it = false;
5605  int i;
5606 
5607  /* Drop this clause if it matches any column of the FK */
5608  for (i = 0; i < fkinfo->nkeys; i++)
5609  {
5610  if (rinfo->parent_ec)
5611  {
5612  /*
5613  * EC-derived clauses can only match by EC. It is okay to
5614  * consider any clause derived from the same EC as
5615  * matching the FK: even if equivclass.c chose to generate
5616  * a clause equating some other pair of Vars, it could
5617  * have generated one equating the FK's Vars. So for
5618  * purposes of estimation, we can act as though it did so.
5619  *
5620  * Note: checking parent_ec is a bit of a cheat because
5621  * there are EC-derived clauses that don't have parent_ec
5622  * set; but such clauses must compare expressions that
5623  * aren't just Vars, so they cannot match the FK anyway.
5624  */
5625  if (fkinfo->eclass[i] == rinfo->parent_ec)
5626  {
5627  remove_it = true;
5628  break;
5629  }
5630  }
5631  else
5632  {
5633  /*
5634  * Otherwise, see if rinfo was previously matched to FK as
5635  * a "loose" clause.
5636  */
5637  if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5638  {
5639  remove_it = true;
5640  break;
5641  }
5642  }
5643  }
5644  if (remove_it)
5645  {
5646  worklist = foreach_delete_current(worklist, cell);
5647  removedlist = lappend(removedlist, rinfo);
5648  }
5649  }
5650 
5651  /*
5652  * If we failed to remove all the matching clauses we expected to
5653  * find, chicken out and ignore this FK; applying its selectivity
5654  * might result in double-counting. Put any clauses we did manage to
5655  * remove back into the worklist.
5656  *
5657  * Since the matching clauses are known not outerjoin-delayed, they
5658  * would normally have appeared in the initial joinclause list. If we
5659  * didn't find them, there are two possibilities:
5660  *
5661  * 1. If the FK match is based on an EC that is ec_has_const, it won't
5662  * have generated any join clauses at all. We discount such ECs while
5663  * checking to see if we have "all" the clauses. (Below, we'll adjust
5664  * the selectivity estimate for this case.)
5665  *
5666  * 2. The clauses were matched to some other FK in a previous
5667  * iteration of this loop, and thus removed from worklist. (A likely
5668  * case is that two FKs are matched to the same EC; there will be only
5669  * one EC-derived clause in the initial list, so the first FK will
5670  * consume it.) Applying both FKs' selectivity independently risks
5671  * underestimating the join size; in particular, this would undo one
5672  * of the main things that ECs were invented for, namely to avoid
5673  * double-counting the selectivity of redundant equality conditions.
5674  * Later we might think of a reasonable way to combine the estimates,
5675  * but for now, just punt, since this is a fairly uncommon situation.
5676  */
5677  if (removedlist == NIL ||
5678  list_length(removedlist) !=
5679  (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5680  {
5681  worklist = list_concat(worklist, removedlist);
5682  continue;
5683  }
5684 
5685  /*
5686  * Finally we get to the payoff: estimate selectivity using the
5687  * knowledge that each referencing row will match exactly one row in
5688  * the referenced table.
5689  *
5690  * XXX that's not true in the presence of nulls in the referencing
5691  * column(s), so in principle we should derate the estimate for those.
5692  * However (1) if there are any strict restriction clauses for the
5693  * referencing column(s) elsewhere in the query, derating here would
5694  * be double-counting the null fraction, and (2) it's not very clear
5695  * how to combine null fractions for multiple referencing columns. So
5696  * we do nothing for now about correcting for nulls.
5697  *
5698  * XXX another point here is that if either side of an FK constraint
5699  * is an inheritance parent, we estimate as though the constraint
5700  * covers all its children as well. This is not an unreasonable
5701  * assumption for a referencing table, ie the user probably applied
5702  * identical constraints to all child tables (though perhaps we ought
5703  * to check that). But it's not possible to have done that for a
5704  * referenced table. Fortunately, precisely because that doesn't
5705  * work, it is uncommon in practice to have an FK referencing a parent
5706  * table. So, at least for now, disregard inheritance here.
5707  */
5708  if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5709  {
5710  /*
5711  * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5712  * referenced table is exactly the inside of the join. The join
5713  * selectivity is defined as the fraction of LHS rows that have
5714  * matches. The FK implies that every LHS row has a match *in the
5715  * referenced table*; but any restriction clauses on it will
5716  * reduce the number of matches. Hence we take the join
5717  * selectivity as equal to the selectivity of the table's
5718  * restriction clauses, which is rows / tuples; but we must guard
5719  * against tuples == 0.
5720  */
5721  RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5722  double ref_tuples = Max(ref_rel->tuples, 1.0);
5723 
5724  fkselec *= ref_rel->rows / ref_tuples;
5725  }
5726  else
5727  {
5728  /*
5729  * Otherwise, selectivity is exactly 1/referenced-table-size; but
5730  * guard against tuples == 0. Note we should use the raw table
5731  * tuple count, not any estimate of its filtered or joined size.
5732  */
5733  RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5734  double ref_tuples = Max(ref_rel->tuples, 1.0);
5735 
5736  fkselec *= 1.0 / ref_tuples;
5737  }
5738 
5739  /*
5740  * If any of the FK columns participated in ec_has_const ECs, then
5741  * equivclass.c will have generated "var = const" restrictions for
5742  * each side of the join, thus reducing the sizes of both input
5743  * relations. Taking the fkselec at face value would amount to
5744  * double-counting the selectivity of the constant restriction for the
5745  * referencing Var. Hence, look for the restriction clause(s) that
5746  * were applied to the referencing Var(s), and divide out their
5747  * selectivity to correct for this.
5748  */
5749  if (fkinfo->nconst_ec > 0)
5750  {
5751  for (int i = 0; i < fkinfo->nkeys; i++)
5752  {
5753  EquivalenceClass *ec = fkinfo->eclass[i];
5754 
5755  if (ec && ec->ec_has_const)
5756  {
5757  EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5759  em);
5760 
5761  if (rinfo)
5762  {
5763  Selectivity s0;
5764 
5765  s0 = clause_selectivity(root,
5766  (Node *) rinfo,
5767  0,
5768  jointype,
5769  sjinfo);
5770  if (s0 > 0)
5771  fkselec /= s0;
5772  }
5773  }
5774  }
5775  }
5776  }
5777 
5778  *restrictlist = worklist;
5779  CLAMP_PROBABILITY(fkselec);
5780  return fkselec;
5781 }
BMS_Membership bms_membership(const Bitmapset *a)
Definition: bitmapset.c:712
@ BMS_SINGLETON
Definition: bitmapset.h:72
RestrictInfo * find_derived_clause_for_ec_member(EquivalenceClass *ec, EquivalenceMember *em)
Definition: equivclass.c:2581
List * list_copy(const List *oldlist)
Definition: list.c:1572
bool list_member_ptr(const List *list, const void *datum)
Definition: list.c:681
#define foreach_delete_current(lst, cell)
Definition: pg_list.h:390
RelOptInfo * find_base_rel(PlannerInfo *root, int relid)
Definition: relnode.c:405
#define CLAMP_PROBABILITY(p)
Definition: selfuncs.h:63
struct EquivalenceClass * eclass[INDEX_MAX_KEYS]
Definition: pathnodes.h:1235
List * rinfos[INDEX_MAX_KEYS]
Definition: pathnodes.h:1239
struct EquivalenceMember * fk_eclass_member[INDEX_MAX_KEYS]
Definition: pathnodes.h:1237
List * fkey_list
Definition: pathnodes.h:379

References bms_is_member(), bms_membership(), BMS_SINGLETON, CLAMP_PROBABILITY, clause_selectivity(), ForeignKeyOptInfo::con_relid, EquivalenceClass::ec_has_const, ForeignKeyOptInfo::eclass, find_base_rel(), find_derived_clause_for_ec_member(), ForeignKeyOptInfo::fk_eclass_member, PlannerInfo::fkey_list, foreach_delete_current, i, JOIN_ANTI, JOIN_SEMI, SpecialJoinInfo::jointype, lappend(),