PostgreSQL Source Code  git master
costsize.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * costsize.c
4  * Routines to compute (and set) relation sizes and path costs
5  *
6  * Path costs are measured in arbitrary units established by these basic
7  * parameters:
8  *
9  * seq_page_cost Cost of a sequential page fetch
10  * random_page_cost Cost of a non-sequential page fetch
11  * cpu_tuple_cost Cost of typical CPU time to process a tuple
12  * cpu_index_tuple_cost Cost of typical CPU time to process an index tuple
13  * cpu_operator_cost Cost of CPU time to execute an operator or function
14  * parallel_tuple_cost Cost of CPU time to pass a tuple from worker to leader backend
15  * parallel_setup_cost Cost of setting up shared memory for parallelism
16  *
17  * We expect that the kernel will typically do some amount of read-ahead
18  * optimization; this in conjunction with seek costs means that seq_page_cost
19  * is normally considerably less than random_page_cost. (However, if the
20  * database is fully cached in RAM, it is reasonable to set them equal.)
21  *
22  * We also use a rough estimate "effective_cache_size" of the number of
23  * disk pages in Postgres + OS-level disk cache. (We can't simply use
24  * NBuffers for this purpose because that would ignore the effects of
25  * the kernel's disk cache.)
26  *
27  * Obviously, taking constants for these values is an oversimplification,
28  * but it's tough enough to get any useful estimates even at this level of
29  * detail. Note that all of these parameters are user-settable, in case
30  * the default values are drastically off for a particular platform.
31  *
32  * seq_page_cost and random_page_cost can also be overridden for an individual
33  * tablespace, in case some data is on a fast disk and other data is on a slow
34  * disk. Per-tablespace overrides never apply to temporary work files such as
35  * an external sort or a materialize node that overflows work_mem.
36  *
37  * We compute two separate costs for each path:
38  * total_cost: total estimated cost to fetch all tuples
39  * startup_cost: cost that is expended before first tuple is fetched
40  * In some scenarios, such as when there is a LIMIT or we are implementing
41  * an EXISTS(...) sub-select, it is not necessary to fetch all tuples of the
42  * path's result. A caller can estimate the cost of fetching a partial
43  * result by interpolating between startup_cost and total_cost. In detail:
44  * actual_cost = startup_cost +
45  * (total_cost - startup_cost) * tuples_to_fetch / path->rows;
46  * Note that a base relation's rows count (and, by extension, plan_rows for
47  * plan nodes below the LIMIT node) are set without regard to any LIMIT, so
48  * that this equation works properly. (Note: while path->rows is never zero
49  * for ordinary relations, it is zero for paths for provably-empty relations,
50  * so beware of division-by-zero.) The LIMIT is applied as a top-level
51  * plan node.
52  *
53  * For largely historical reasons, most of the routines in this module use
54  * the passed result Path only to store their results (rows, startup_cost and
55  * total_cost) into. All the input data they need is passed as separate
56  * parameters, even though much of it could be extracted from the Path.
57  * An exception is made for the cost_XXXjoin() routines, which expect all
58  * the other fields of the passed XXXPath to be filled in, and similarly
59  * cost_index() assumes the passed IndexPath is valid except for its output
60  * values.
61  *
62  *
63  * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
64  * Portions Copyright (c) 1994, Regents of the University of California
65  *
66  * IDENTIFICATION
67  * src/backend/optimizer/path/costsize.c
68  *
69  *-------------------------------------------------------------------------
70  */
71 
72 #include "postgres.h"
73 
74 #include <limits.h>
75 #include <math.h>
76 
77 #include "access/amapi.h"
78 #include "access/htup_details.h"
79 #include "access/tsmapi.h"
80 #include "executor/executor.h"
81 #include "executor/nodeAgg.h"
82 #include "executor/nodeHash.h"
83 #include "executor/nodeMemoize.h"
84 #include "miscadmin.h"
85 #include "nodes/makefuncs.h"
86 #include "nodes/nodeFuncs.h"
87 #include "optimizer/clauses.h"
88 #include "optimizer/cost.h"
89 #include "optimizer/optimizer.h"
90 #include "optimizer/pathnode.h"
91 #include "optimizer/paths.h"
92 #include "optimizer/placeholder.h"
93 #include "optimizer/plancat.h"
94 #include "optimizer/restrictinfo.h"
95 #include "parser/parsetree.h"
96 #include "utils/lsyscache.h"
97 #include "utils/selfuncs.h"
98 #include "utils/spccache.h"
99 #include "utils/tuplesort.h"
100 
101 
102 #define LOG2(x) (log(x) / 0.693147180559945)
103 
104 /*
105  * Append and MergeAppend nodes are less expensive than some other operations
106  * which use cpu_tuple_cost; instead of adding a separate GUC, estimate the
107  * per-tuple cost as cpu_tuple_cost multiplied by this value.
108  */
109 #define APPEND_CPU_COST_MULTIPLIER 0.5
110 
111 /*
112  * Maximum value for row estimates. We cap row estimates to this to help
113  * ensure that costs based on these estimates remain within the range of what
114  * double can represent. add_path() wouldn't act sanely given infinite or NaN
115  * cost values.
116  */
117 #define MAXIMUM_ROWCOUNT 1e100
118 
127 
129 
131 
133 
134 bool enable_seqscan = true;
135 bool enable_indexscan = true;
137 bool enable_bitmapscan = true;
138 bool enable_tidscan = true;
139 bool enable_sort = true;
141 bool enable_hashagg = true;
142 bool enable_nestloop = true;
143 bool enable_material = true;
144 bool enable_memoize = true;
145 bool enable_mergejoin = true;
146 bool enable_hashjoin = true;
147 bool enable_gathermerge = true;
155 
156 typedef struct
157 {
161 
162 static List *extract_nonindex_conditions(List *qual_clauses, List *indexclauses);
164  RestrictInfo *rinfo,
165  PathKey *pathkey);
166 static void cost_rescan(PlannerInfo *root, Path *path,
167  Cost *rescan_startup_cost, Cost *rescan_total_cost);
170  ParamPathInfo *param_info,
171  QualCost *qpqual_cost);
172 static bool has_indexed_join_quals(NestPath *path);
173 static double approx_tuple_count(PlannerInfo *root, JoinPath *path,
174  List *quals);
176  RelOptInfo *joinrel,
177  RelOptInfo *outer_rel,
178  RelOptInfo *inner_rel,
179  double outer_rows,
180  double inner_rows,
181  SpecialJoinInfo *sjinfo,
182  List *restrictlist);
184  Relids outer_relids,
185  Relids inner_relids,
186  SpecialJoinInfo *sjinfo,
187  List **restrictlist);
188 static Cost append_nonpartial_cost(List *subpaths, int numpaths,
189  int parallel_workers);
190 static void set_rel_width(PlannerInfo *root, RelOptInfo *rel);
191 static int32 get_expr_width(PlannerInfo *root, const Node *expr);
192 static double relation_byte_size(double tuples, int width);
193 static double page_size(double tuples, int width);
194 static double get_parallel_divisor(Path *path);
195 
196 
197 /*
198  * clamp_row_est
199  * Force a row-count estimate to a sane value.
200  */
201 double
202 clamp_row_est(double nrows)
203 {
204  /*
205  * Avoid infinite and NaN row estimates. Costs derived from such values
206  * are going to be useless. Also force the estimate to be at least one
207  * row, to make explain output look better and to avoid possible
208  * divide-by-zero when interpolating costs. Make it an integer, too.
209  */
210  if (nrows > MAXIMUM_ROWCOUNT || isnan(nrows))
211  nrows = MAXIMUM_ROWCOUNT;
212  else if (nrows <= 1.0)
213  nrows = 1.0;
214  else
215  nrows = rint(nrows);
216 
217  return nrows;
218 }
219 
220 /*
221  * clamp_width_est
222  * Force a tuple-width estimate to a sane value.
223  *
224  * The planner represents datatype width and tuple width estimates as int32.
225  * When summing column width estimates to create a tuple width estimate,
226  * it's possible to reach integer overflow in edge cases. To ensure sane
227  * behavior, we form such sums in int64 arithmetic and then apply this routine
228  * to clamp to int32 range.
229  */
230 int32
231 clamp_width_est(int64 tuple_width)
232 {
233  /*
234  * Anything more than MaxAllocSize is clearly bogus, since we could not
235  * create a tuple that large.
236  */
237  if (tuple_width > MaxAllocSize)
238  return (int32) MaxAllocSize;
239 
240  /*
241  * Unlike clamp_row_est, we just Assert that the value isn't negative,
242  * rather than masking such errors.
243  */
244  Assert(tuple_width >= 0);
245 
246  return (int32) tuple_width;
247 }
248 
249 /*
250  * clamp_cardinality_to_long
251  * Cast a Cardinality value to a sane long value.
252  */
253 long
255 {
256  /*
257  * Just for paranoia's sake, ensure we do something sane with negative or
258  * NaN values.
259  */
260  if (isnan(x))
261  return LONG_MAX;
262  if (x <= 0)
263  return 0;
264 
265  /*
266  * If "long" is 64 bits, then LONG_MAX cannot be represented exactly as a
267  * double. Casting it to double and back may well result in overflow due
268  * to rounding, so avoid doing that. We trust that any double value that
269  * compares strictly less than "(double) LONG_MAX" will cast to a
270  * representable "long" value.
271  */
272  return (x < (double) LONG_MAX) ? (long) x : LONG_MAX;
273 }
274 
275 
276 /*
277  * cost_seqscan
278  * Determines and returns the cost of scanning a relation sequentially.
279  *
280  * 'baserel' is the relation to be scanned
281  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
282  */
283 void
285  RelOptInfo *baserel, ParamPathInfo *param_info)
286 {
287  Cost startup_cost = 0;
288  Cost cpu_run_cost;
289  Cost disk_run_cost;
290  double spc_seq_page_cost;
291  QualCost qpqual_cost;
292  Cost cpu_per_tuple;
293 
294  /* Should only be applied to base relations */
295  Assert(baserel->relid > 0);
296  Assert(baserel->rtekind == RTE_RELATION);
297 
298  /* Mark the path with the correct row estimate */
299  if (param_info)
300  path->rows = param_info->ppi_rows;
301  else
302  path->rows = baserel->rows;
303 
304  if (!enable_seqscan)
305  startup_cost += disable_cost;
306 
307  /* fetch estimated page cost for tablespace containing table */
309  NULL,
310  &spc_seq_page_cost);
311 
312  /*
313  * disk costs
314  */
315  disk_run_cost = spc_seq_page_cost * baserel->pages;
316 
317  /* CPU costs */
318  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
319 
320  startup_cost += qpqual_cost.startup;
321  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
322  cpu_run_cost = cpu_per_tuple * baserel->tuples;
323  /* tlist eval costs are paid per output row, not per tuple scanned */
324  startup_cost += path->pathtarget->cost.startup;
325  cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
326 
327  /* Adjust costing for parallelism, if used. */
328  if (path->parallel_workers > 0)
329  {
330  double parallel_divisor = get_parallel_divisor(path);
331 
332  /* The CPU cost is divided among all the workers. */
333  cpu_run_cost /= parallel_divisor;
334 
335  /*
336  * It may be possible to amortize some of the I/O cost, but probably
337  * not very much, because most operating systems already do aggressive
338  * prefetching. For now, we assume that the disk run cost can't be
339  * amortized at all.
340  */
341 
342  /*
343  * In the case of a parallel plan, the row count needs to represent
344  * the number of tuples processed per worker.
345  */
346  path->rows = clamp_row_est(path->rows / parallel_divisor);
347  }
348 
349  path->startup_cost = startup_cost;
350  path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
351 }
352 
353 /*
354  * cost_samplescan
355  * Determines and returns the cost of scanning a relation using sampling.
356  *
357  * 'baserel' is the relation to be scanned
358  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
359  */
360 void
362  RelOptInfo *baserel, ParamPathInfo *param_info)
363 {
364  Cost startup_cost = 0;
365  Cost run_cost = 0;
366  RangeTblEntry *rte;
367  TableSampleClause *tsc;
368  TsmRoutine *tsm;
369  double spc_seq_page_cost,
370  spc_random_page_cost,
371  spc_page_cost;
372  QualCost qpqual_cost;
373  Cost cpu_per_tuple;
374 
375  /* Should only be applied to base relations with tablesample clauses */
376  Assert(baserel->relid > 0);
377  rte = planner_rt_fetch(baserel->relid, root);
378  Assert(rte->rtekind == RTE_RELATION);
379  tsc = rte->tablesample;
380  Assert(tsc != NULL);
381  tsm = GetTsmRoutine(tsc->tsmhandler);
382 
383  /* Mark the path with the correct row estimate */
384  if (param_info)
385  path->rows = param_info->ppi_rows;
386  else
387  path->rows = baserel->rows;
388 
389  /* fetch estimated page cost for tablespace containing table */
391  &spc_random_page_cost,
392  &spc_seq_page_cost);
393 
394  /* if NextSampleBlock is used, assume random access, else sequential */
395  spc_page_cost = (tsm->NextSampleBlock != NULL) ?
396  spc_random_page_cost : spc_seq_page_cost;
397 
398  /*
399  * disk costs (recall that baserel->pages has already been set to the
400  * number of pages the sampling method will visit)
401  */
402  run_cost += spc_page_cost * baserel->pages;
403 
404  /*
405  * CPU costs (recall that baserel->tuples has already been set to the
406  * number of tuples the sampling method will select). Note that we ignore
407  * execution cost of the TABLESAMPLE parameter expressions; they will be
408  * evaluated only once per scan, and in most usages they'll likely be
409  * simple constants anyway. We also don't charge anything for the
410  * calculations the sampling method might do internally.
411  */
412  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
413 
414  startup_cost += qpqual_cost.startup;
415  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
416  run_cost += cpu_per_tuple * baserel->tuples;
417  /* tlist eval costs are paid per output row, not per tuple scanned */
418  startup_cost += path->pathtarget->cost.startup;
419  run_cost += path->pathtarget->cost.per_tuple * path->rows;
420 
421  path->startup_cost = startup_cost;
422  path->total_cost = startup_cost + run_cost;
423 }
424 
425 /*
426  * cost_gather
427  * Determines and returns the cost of gather path.
428  *
429  * 'rel' is the relation to be operated upon
430  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
431  * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
432  * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
433  * correspond to any particular RelOptInfo.
434  */
435 void
437  RelOptInfo *rel, ParamPathInfo *param_info,
438  double *rows)
439 {
440  Cost startup_cost = 0;
441  Cost run_cost = 0;
442 
443  /* Mark the path with the correct row estimate */
444  if (rows)
445  path->path.rows = *rows;
446  else if (param_info)
447  path->path.rows = param_info->ppi_rows;
448  else
449  path->path.rows = rel->rows;
450 
451  startup_cost = path->subpath->startup_cost;
452 
453  run_cost = path->subpath->total_cost - path->subpath->startup_cost;
454 
455  /* Parallel setup and communication cost. */
456  startup_cost += parallel_setup_cost;
457  run_cost += parallel_tuple_cost * path->path.rows;
458 
459  path->path.startup_cost = startup_cost;
460  path->path.total_cost = (startup_cost + run_cost);
461 }
462 
463 /*
464  * cost_gather_merge
465  * Determines and returns the cost of gather merge path.
466  *
467  * GatherMerge merges several pre-sorted input streams, using a heap that at
468  * any given instant holds the next tuple from each stream. If there are N
469  * streams, we need about N*log2(N) tuple comparisons to construct the heap at
470  * startup, and then for each output tuple, about log2(N) comparisons to
471  * replace the top heap entry with the next tuple from the same stream.
472  */
473 void
475  RelOptInfo *rel, ParamPathInfo *param_info,
476  Cost input_startup_cost, Cost input_total_cost,
477  double *rows)
478 {
479  Cost startup_cost = 0;
480  Cost run_cost = 0;
481  Cost comparison_cost;
482  double N;
483  double logN;
484 
485  /* Mark the path with the correct row estimate */
486  if (rows)
487  path->path.rows = *rows;
488  else if (param_info)
489  path->path.rows = param_info->ppi_rows;
490  else
491  path->path.rows = rel->rows;
492 
493  if (!enable_gathermerge)
494  startup_cost += disable_cost;
495 
496  /*
497  * Add one to the number of workers to account for the leader. This might
498  * be overgenerous since the leader will do less work than other workers
499  * in typical cases, but we'll go with it for now.
500  */
501  Assert(path->num_workers > 0);
502  N = (double) path->num_workers + 1;
503  logN = LOG2(N);
504 
505  /* Assumed cost per tuple comparison */
506  comparison_cost = 2.0 * cpu_operator_cost;
507 
508  /* Heap creation cost */
509  startup_cost += comparison_cost * N * logN;
510 
511  /* Per-tuple heap maintenance cost */
512  run_cost += path->path.rows * comparison_cost * logN;
513 
514  /* small cost for heap management, like cost_merge_append */
515  run_cost += cpu_operator_cost * path->path.rows;
516 
517  /*
518  * Parallel setup and communication cost. Since Gather Merge, unlike
519  * Gather, requires us to block until a tuple is available from every
520  * worker, we bump the IPC cost up a little bit as compared with Gather.
521  * For lack of a better idea, charge an extra 5%.
522  */
523  startup_cost += parallel_setup_cost;
524  run_cost += parallel_tuple_cost * path->path.rows * 1.05;
525 
526  path->path.startup_cost = startup_cost + input_startup_cost;
527  path->path.total_cost = (startup_cost + run_cost + input_total_cost);
528 }
529 
530 /*
531  * cost_index
532  * Determines and returns the cost of scanning a relation using an index.
533  *
534  * 'path' describes the indexscan under consideration, and is complete
535  * except for the fields to be set by this routine
536  * 'loop_count' is the number of repetitions of the indexscan to factor into
537  * estimates of caching behavior
538  *
539  * In addition to rows, startup_cost and total_cost, cost_index() sets the
540  * path's indextotalcost and indexselectivity fields. These values will be
541  * needed if the IndexPath is used in a BitmapIndexScan.
542  *
543  * NOTE: path->indexquals must contain only clauses usable as index
544  * restrictions. Any additional quals evaluated as qpquals may reduce the
545  * number of returned tuples, but they won't reduce the number of tuples
546  * we have to fetch from the table, so they don't reduce the scan cost.
547  */
548 void
549 cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
550  bool partial_path)
551 {
552  IndexOptInfo *index = path->indexinfo;
553  RelOptInfo *baserel = index->rel;
554  bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
555  amcostestimate_function amcostestimate;
556  List *qpquals;
557  Cost startup_cost = 0;
558  Cost run_cost = 0;
559  Cost cpu_run_cost = 0;
560  Cost indexStartupCost;
561  Cost indexTotalCost;
562  Selectivity indexSelectivity;
563  double indexCorrelation,
564  csquared;
565  double spc_seq_page_cost,
566  spc_random_page_cost;
567  Cost min_IO_cost,
568  max_IO_cost;
569  QualCost qpqual_cost;
570  Cost cpu_per_tuple;
571  double tuples_fetched;
572  double pages_fetched;
573  double rand_heap_pages;
574  double index_pages;
575 
576  /* Should only be applied to base relations */
577  Assert(IsA(baserel, RelOptInfo) &&
579  Assert(baserel->relid > 0);
580  Assert(baserel->rtekind == RTE_RELATION);
581 
582  /*
583  * Mark the path with the correct row estimate, and identify which quals
584  * will need to be enforced as qpquals. We need not check any quals that
585  * are implied by the index's predicate, so we can use indrestrictinfo not
586  * baserestrictinfo as the list of relevant restriction clauses for the
587  * rel.
588  */
589  if (path->path.param_info)
590  {
591  path->path.rows = path->path.param_info->ppi_rows;
592  /* qpquals come from the rel's restriction clauses and ppi_clauses */
594  path->indexclauses),
595  extract_nonindex_conditions(path->path.param_info->ppi_clauses,
596  path->indexclauses));
597  }
598  else
599  {
600  path->path.rows = baserel->rows;
601  /* qpquals come from just the rel's restriction clauses */
603  path->indexclauses);
604  }
605 
606  if (!enable_indexscan)
607  startup_cost += disable_cost;
608  /* we don't need to check enable_indexonlyscan; indxpath.c does that */
609 
610  /*
611  * Call index-access-method-specific code to estimate the processing cost
612  * for scanning the index, as well as the selectivity of the index (ie,
613  * the fraction of main-table tuples we will have to retrieve) and its
614  * correlation to the main-table tuple order. We need a cast here because
615  * pathnodes.h uses a weak function type to avoid including amapi.h.
616  */
617  amcostestimate = (amcostestimate_function) index->amcostestimate;
618  amcostestimate(root, path, loop_count,
619  &indexStartupCost, &indexTotalCost,
620  &indexSelectivity, &indexCorrelation,
621  &index_pages);
622 
623  /*
624  * Save amcostestimate's results for possible use in bitmap scan planning.
625  * We don't bother to save indexStartupCost or indexCorrelation, because a
626  * bitmap scan doesn't care about either.
627  */
628  path->indextotalcost = indexTotalCost;
629  path->indexselectivity = indexSelectivity;
630 
631  /* all costs for touching index itself included here */
632  startup_cost += indexStartupCost;
633  run_cost += indexTotalCost - indexStartupCost;
634 
635  /* estimate number of main-table tuples fetched */
636  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
637 
638  /* fetch estimated page costs for tablespace containing table */
640  &spc_random_page_cost,
641  &spc_seq_page_cost);
642 
643  /*----------
644  * Estimate number of main-table pages fetched, and compute I/O cost.
645  *
646  * When the index ordering is uncorrelated with the table ordering,
647  * we use an approximation proposed by Mackert and Lohman (see
648  * index_pages_fetched() for details) to compute the number of pages
649  * fetched, and then charge spc_random_page_cost per page fetched.
650  *
651  * When the index ordering is exactly correlated with the table ordering
652  * (just after a CLUSTER, for example), the number of pages fetched should
653  * be exactly selectivity * table_size. What's more, all but the first
654  * will be sequential fetches, not the random fetches that occur in the
655  * uncorrelated case. So if the number of pages is more than 1, we
656  * ought to charge
657  * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
658  * For partially-correlated indexes, we ought to charge somewhere between
659  * these two estimates. We currently interpolate linearly between the
660  * estimates based on the correlation squared (XXX is that appropriate?).
661  *
662  * If it's an index-only scan, then we will not need to fetch any heap
663  * pages for which the visibility map shows all tuples are visible.
664  * Hence, reduce the estimated number of heap fetches accordingly.
665  * We use the measured fraction of the entire heap that is all-visible,
666  * which might not be particularly relevant to the subset of the heap
667  * that this query will fetch; but it's not clear how to do better.
668  *----------
669  */
670  if (loop_count > 1)
671  {
672  /*
673  * For repeated indexscans, the appropriate estimate for the
674  * uncorrelated case is to scale up the number of tuples fetched in
675  * the Mackert and Lohman formula by the number of scans, so that we
676  * estimate the number of pages fetched by all the scans; then
677  * pro-rate the costs for one scan. In this case we assume all the
678  * fetches are random accesses.
679  */
680  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
681  baserel->pages,
682  (double) index->pages,
683  root);
684 
685  if (indexonly)
686  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
687 
688  rand_heap_pages = pages_fetched;
689 
690  max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
691 
692  /*
693  * In the perfectly correlated case, the number of pages touched by
694  * each scan is selectivity * table_size, and we can use the Mackert
695  * and Lohman formula at the page level to estimate how much work is
696  * saved by caching across scans. We still assume all the fetches are
697  * random, though, which is an overestimate that's hard to correct for
698  * without double-counting the cache effects. (But in most cases
699  * where such a plan is actually interesting, only one page would get
700  * fetched per scan anyway, so it shouldn't matter much.)
701  */
702  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
703 
704  pages_fetched = index_pages_fetched(pages_fetched * loop_count,
705  baserel->pages,
706  (double) index->pages,
707  root);
708 
709  if (indexonly)
710  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
711 
712  min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
713  }
714  else
715  {
716  /*
717  * Normal case: apply the Mackert and Lohman formula, and then
718  * interpolate between that and the correlation-derived result.
719  */
720  pages_fetched = index_pages_fetched(tuples_fetched,
721  baserel->pages,
722  (double) index->pages,
723  root);
724 
725  if (indexonly)
726  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
727 
728  rand_heap_pages = pages_fetched;
729 
730  /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
731  max_IO_cost = pages_fetched * spc_random_page_cost;
732 
733  /* min_IO_cost is for the perfectly correlated case (csquared=1) */
734  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
735 
736  if (indexonly)
737  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
738 
739  if (pages_fetched > 0)
740  {
741  min_IO_cost = spc_random_page_cost;
742  if (pages_fetched > 1)
743  min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
744  }
745  else
746  min_IO_cost = 0;
747  }
748 
749  if (partial_path)
750  {
751  /*
752  * For index only scans compute workers based on number of index pages
753  * fetched; the number of heap pages we fetch might be so small as to
754  * effectively rule out parallelism, which we don't want to do.
755  */
756  if (indexonly)
757  rand_heap_pages = -1;
758 
759  /*
760  * Estimate the number of parallel workers required to scan index. Use
761  * the number of heap pages computed considering heap fetches won't be
762  * sequential as for parallel scans the pages are accessed in random
763  * order.
764  */
766  rand_heap_pages,
767  index_pages,
769 
770  /*
771  * Fall out if workers can't be assigned for parallel scan, because in
772  * such a case this path will be rejected. So there is no benefit in
773  * doing extra computation.
774  */
775  if (path->path.parallel_workers <= 0)
776  return;
777 
778  path->path.parallel_aware = true;
779  }
780 
781  /*
782  * Now interpolate based on estimated index order correlation to get total
783  * disk I/O cost for main table accesses.
784  */
785  csquared = indexCorrelation * indexCorrelation;
786 
787  run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
788 
789  /*
790  * Estimate CPU costs per tuple.
791  *
792  * What we want here is cpu_tuple_cost plus the evaluation costs of any
793  * qual clauses that we have to evaluate as qpquals.
794  */
795  cost_qual_eval(&qpqual_cost, qpquals, root);
796 
797  startup_cost += qpqual_cost.startup;
798  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
799 
800  cpu_run_cost += cpu_per_tuple * tuples_fetched;
801 
802  /* tlist eval costs are paid per output row, not per tuple scanned */
803  startup_cost += path->path.pathtarget->cost.startup;
804  cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
805 
806  /* Adjust costing for parallelism, if used. */
807  if (path->path.parallel_workers > 0)
808  {
809  double parallel_divisor = get_parallel_divisor(&path->path);
810 
811  path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
812 
813  /* The CPU cost is divided among all the workers. */
814  cpu_run_cost /= parallel_divisor;
815  }
816 
817  run_cost += cpu_run_cost;
818 
819  path->path.startup_cost = startup_cost;
820  path->path.total_cost = startup_cost + run_cost;
821 }
822 
823 /*
824  * extract_nonindex_conditions
825  *
826  * Given a list of quals to be enforced in an indexscan, extract the ones that
827  * will have to be applied as qpquals (ie, the index machinery won't handle
828  * them). Here we detect only whether a qual clause is directly redundant
829  * with some indexclause. If the index path is chosen for use, createplan.c
830  * will try a bit harder to get rid of redundant qual conditions; specifically
831  * it will see if quals can be proven to be implied by the indexquals. But
832  * it does not seem worth the cycles to try to factor that in at this stage,
833  * since we're only trying to estimate qual eval costs. Otherwise this must
834  * match the logic in create_indexscan_plan().
835  *
836  * qual_clauses, and the result, are lists of RestrictInfos.
837  * indexclauses is a list of IndexClauses.
838  */
839 static List *
840 extract_nonindex_conditions(List *qual_clauses, List *indexclauses)
841 {
842  List *result = NIL;
843  ListCell *lc;
844 
845  foreach(lc, qual_clauses)
846  {
847  RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
848 
849  if (rinfo->pseudoconstant)
850  continue; /* we may drop pseudoconstants here */
851  if (is_redundant_with_indexclauses(rinfo, indexclauses))
852  continue; /* dup or derived from same EquivalenceClass */
853  /* ... skip the predicate proof attempt createplan.c will try ... */
854  result = lappend(result, rinfo);
855  }
856  return result;
857 }
858 
859 /*
860  * index_pages_fetched
861  * Estimate the number of pages actually fetched after accounting for
862  * cache effects.
863  *
864  * We use an approximation proposed by Mackert and Lohman, "Index Scans
865  * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
866  * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
867  * The Mackert and Lohman approximation is that the number of pages
868  * fetched is
869  * PF =
870  * min(2TNs/(2T+Ns), T) when T <= b
871  * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
872  * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
873  * where
874  * T = # pages in table
875  * N = # tuples in table
876  * s = selectivity = fraction of table to be scanned
877  * b = # buffer pages available (we include kernel space here)
878  *
879  * We assume that effective_cache_size is the total number of buffer pages
880  * available for the whole query, and pro-rate that space across all the
881  * tables in the query and the index currently under consideration. (This
882  * ignores space needed for other indexes used by the query, but since we
883  * don't know which indexes will get used, we can't estimate that very well;
884  * and in any case counting all the tables may well be an overestimate, since
885  * depending on the join plan not all the tables may be scanned concurrently.)
886  *
887  * The product Ns is the number of tuples fetched; we pass in that
888  * product rather than calculating it here. "pages" is the number of pages
889  * in the object under consideration (either an index or a table).
890  * "index_pages" is the amount to add to the total table space, which was
891  * computed for us by make_one_rel.
892  *
893  * Caller is expected to have ensured that tuples_fetched is greater than zero
894  * and rounded to integer (see clamp_row_est). The result will likewise be
895  * greater than zero and integral.
896  */
897 double
898 index_pages_fetched(double tuples_fetched, BlockNumber pages,
899  double index_pages, PlannerInfo *root)
900 {
901  double pages_fetched;
902  double total_pages;
903  double T,
904  b;
905 
906  /* T is # pages in table, but don't allow it to be zero */
907  T = (pages > 1) ? (double) pages : 1.0;
908 
909  /* Compute number of pages assumed to be competing for cache space */
910  total_pages = root->total_table_pages + index_pages;
911  total_pages = Max(total_pages, 1.0);
912  Assert(T <= total_pages);
913 
914  /* b is pro-rated share of effective_cache_size */
915  b = (double) effective_cache_size * T / total_pages;
916 
917  /* force it positive and integral */
918  if (b <= 1.0)
919  b = 1.0;
920  else
921  b = ceil(b);
922 
923  /* This part is the Mackert and Lohman formula */
924  if (T <= b)
925  {
926  pages_fetched =
927  (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
928  if (pages_fetched >= T)
929  pages_fetched = T;
930  else
931  pages_fetched = ceil(pages_fetched);
932  }
933  else
934  {
935  double lim;
936 
937  lim = (2.0 * T * b) / (2.0 * T - b);
938  if (tuples_fetched <= lim)
939  {
940  pages_fetched =
941  (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
942  }
943  else
944  {
945  pages_fetched =
946  b + (tuples_fetched - lim) * (T - b) / T;
947  }
948  pages_fetched = ceil(pages_fetched);
949  }
950  return pages_fetched;
951 }
952 
953 /*
954  * get_indexpath_pages
955  * Determine the total size of the indexes used in a bitmap index path.
956  *
957  * Note: if the same index is used more than once in a bitmap tree, we will
958  * count it multiple times, which perhaps is the wrong thing ... but it's
959  * not completely clear, and detecting duplicates is difficult, so ignore it
960  * for now.
961  */
962 static double
964 {
965  double result = 0;
966  ListCell *l;
967 
968  if (IsA(bitmapqual, BitmapAndPath))
969  {
970  BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
971 
972  foreach(l, apath->bitmapquals)
973  {
974  result += get_indexpath_pages((Path *) lfirst(l));
975  }
976  }
977  else if (IsA(bitmapqual, BitmapOrPath))
978  {
979  BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
980 
981  foreach(l, opath->bitmapquals)
982  {
983  result += get_indexpath_pages((Path *) lfirst(l));
984  }
985  }
986  else if (IsA(bitmapqual, IndexPath))
987  {
988  IndexPath *ipath = (IndexPath *) bitmapqual;
989 
990  result = (double) ipath->indexinfo->pages;
991  }
992  else
993  elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
994 
995  return result;
996 }
997 
998 /*
999  * cost_bitmap_heap_scan
1000  * Determines and returns the cost of scanning a relation using a bitmap
1001  * index-then-heap plan.
1002  *
1003  * 'baserel' is the relation to be scanned
1004  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1005  * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
1006  * 'loop_count' is the number of repetitions of the indexscan to factor into
1007  * estimates of caching behavior
1008  *
1009  * Note: the component IndexPaths in bitmapqual should have been costed
1010  * using the same loop_count.
1011  */
1012 void
1014  ParamPathInfo *param_info,
1015  Path *bitmapqual, double loop_count)
1016 {
1017  Cost startup_cost = 0;
1018  Cost run_cost = 0;
1019  Cost indexTotalCost;
1020  QualCost qpqual_cost;
1021  Cost cpu_per_tuple;
1022  Cost cost_per_page;
1023  Cost cpu_run_cost;
1024  double tuples_fetched;
1025  double pages_fetched;
1026  double spc_seq_page_cost,
1027  spc_random_page_cost;
1028  double T;
1029 
1030  /* Should only be applied to base relations */
1031  Assert(IsA(baserel, RelOptInfo));
1032  Assert(baserel->relid > 0);
1033  Assert(baserel->rtekind == RTE_RELATION);
1034 
1035  /* Mark the path with the correct row estimate */
1036  if (param_info)
1037  path->rows = param_info->ppi_rows;
1038  else
1039  path->rows = baserel->rows;
1040 
1041  if (!enable_bitmapscan)
1042  startup_cost += disable_cost;
1043 
1044  pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
1045  loop_count, &indexTotalCost,
1046  &tuples_fetched);
1047 
1048  startup_cost += indexTotalCost;
1049  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
1050 
1051  /* Fetch estimated page costs for tablespace containing table. */
1053  &spc_random_page_cost,
1054  &spc_seq_page_cost);
1055 
1056  /*
1057  * For small numbers of pages we should charge spc_random_page_cost
1058  * apiece, while if nearly all the table's pages are being read, it's more
1059  * appropriate to charge spc_seq_page_cost apiece. The effect is
1060  * nonlinear, too. For lack of a better idea, interpolate like this to
1061  * determine the cost per page.
1062  */
1063  if (pages_fetched >= 2.0)
1064  cost_per_page = spc_random_page_cost -
1065  (spc_random_page_cost - spc_seq_page_cost)
1066  * sqrt(pages_fetched / T);
1067  else
1068  cost_per_page = spc_random_page_cost;
1069 
1070  run_cost += pages_fetched * cost_per_page;
1071 
1072  /*
1073  * Estimate CPU costs per tuple.
1074  *
1075  * Often the indexquals don't need to be rechecked at each tuple ... but
1076  * not always, especially not if there are enough tuples involved that the
1077  * bitmaps become lossy. For the moment, just assume they will be
1078  * rechecked always. This means we charge the full freight for all the
1079  * scan clauses.
1080  */
1081  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1082 
1083  startup_cost += qpqual_cost.startup;
1084  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1085  cpu_run_cost = cpu_per_tuple * tuples_fetched;
1086 
1087  /* Adjust costing for parallelism, if used. */
1088  if (path->parallel_workers > 0)
1089  {
1090  double parallel_divisor = get_parallel_divisor(path);
1091 
1092  /* The CPU cost is divided among all the workers. */
1093  cpu_run_cost /= parallel_divisor;
1094 
1095  path->rows = clamp_row_est(path->rows / parallel_divisor);
1096  }
1097 
1098 
1099  run_cost += cpu_run_cost;
1100 
1101  /* tlist eval costs are paid per output row, not per tuple scanned */
1102  startup_cost += path->pathtarget->cost.startup;
1103  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1104 
1105  path->startup_cost = startup_cost;
1106  path->total_cost = startup_cost + run_cost;
1107 }
1108 
1109 /*
1110  * cost_bitmap_tree_node
1111  * Extract cost and selectivity from a bitmap tree node (index/and/or)
1112  */
1113 void
1115 {
1116  if (IsA(path, IndexPath))
1117  {
1118  *cost = ((IndexPath *) path)->indextotalcost;
1119  *selec = ((IndexPath *) path)->indexselectivity;
1120 
1121  /*
1122  * Charge a small amount per retrieved tuple to reflect the costs of
1123  * manipulating the bitmap. This is mostly to make sure that a bitmap
1124  * scan doesn't look to be the same cost as an indexscan to retrieve a
1125  * single tuple.
1126  */
1127  *cost += 0.1 * cpu_operator_cost * path->rows;
1128  }
1129  else if (IsA(path, BitmapAndPath))
1130  {
1131  *cost = path->total_cost;
1132  *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1133  }
1134  else if (IsA(path, BitmapOrPath))
1135  {
1136  *cost = path->total_cost;
1137  *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1138  }
1139  else
1140  {
1141  elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1142  *cost = *selec = 0; /* keep compiler quiet */
1143  }
1144 }
1145 
1146 /*
1147  * cost_bitmap_and_node
1148  * Estimate the cost of a BitmapAnd node
1149  *
1150  * Note that this considers only the costs of index scanning and bitmap
1151  * creation, not the eventual heap access. In that sense the object isn't
1152  * truly a Path, but it has enough path-like properties (costs in particular)
1153  * to warrant treating it as one. We don't bother to set the path rows field,
1154  * however.
1155  */
1156 void
1158 {
1159  Cost totalCost;
1160  Selectivity selec;
1161  ListCell *l;
1162 
1163  /*
1164  * We estimate AND selectivity on the assumption that the inputs are
1165  * independent. This is probably often wrong, but we don't have the info
1166  * to do better.
1167  *
1168  * The runtime cost of the BitmapAnd itself is estimated at 100x
1169  * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1170  * definitely too simplistic?
1171  */
1172  totalCost = 0.0;
1173  selec = 1.0;
1174  foreach(l, path->bitmapquals)
1175  {
1176  Path *subpath = (Path *) lfirst(l);
1177  Cost subCost;
1178  Selectivity subselec;
1179 
1180  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1181 
1182  selec *= subselec;
1183 
1184  totalCost += subCost;
1185  if (l != list_head(path->bitmapquals))
1186  totalCost += 100.0 * cpu_operator_cost;
1187  }
1188  path->bitmapselectivity = selec;
1189  path->path.rows = 0; /* per above, not used */
1190  path->path.startup_cost = totalCost;
1191  path->path.total_cost = totalCost;
1192 }
1193 
1194 /*
1195  * cost_bitmap_or_node
1196  * Estimate the cost of a BitmapOr node
1197  *
1198  * See comments for cost_bitmap_and_node.
1199  */
1200 void
1202 {
1203  Cost totalCost;
1204  Selectivity selec;
1205  ListCell *l;
1206 
1207  /*
1208  * We estimate OR selectivity on the assumption that the inputs are
1209  * non-overlapping, since that's often the case in "x IN (list)" type
1210  * situations. Of course, we clamp to 1.0 at the end.
1211  *
1212  * The runtime cost of the BitmapOr itself is estimated at 100x
1213  * cpu_operator_cost for each tbm_union needed. Probably too small,
1214  * definitely too simplistic? We are aware that the tbm_unions are
1215  * optimized out when the inputs are BitmapIndexScans.
1216  */
1217  totalCost = 0.0;
1218  selec = 0.0;
1219  foreach(l, path->bitmapquals)
1220  {
1221  Path *subpath = (Path *) lfirst(l);
1222  Cost subCost;
1223  Selectivity subselec;
1224 
1225  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1226 
1227  selec += subselec;
1228 
1229  totalCost += subCost;
1230  if (l != list_head(path->bitmapquals) &&
1231  !IsA(subpath, IndexPath))
1232  totalCost += 100.0 * cpu_operator_cost;
1233  }
1234  path->bitmapselectivity = Min(selec, 1.0);
1235  path->path.rows = 0; /* per above, not used */
1236  path->path.startup_cost = totalCost;
1237  path->path.total_cost = totalCost;
1238 }
1239 
1240 /*
1241  * cost_tidscan
1242  * Determines and returns the cost of scanning a relation using TIDs.
1243  *
1244  * 'baserel' is the relation to be scanned
1245  * 'tidquals' is the list of TID-checkable quals
1246  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1247  */
1248 void
1250  RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1251 {
1252  Cost startup_cost = 0;
1253  Cost run_cost = 0;
1254  bool isCurrentOf = false;
1255  QualCost qpqual_cost;
1256  Cost cpu_per_tuple;
1257  QualCost tid_qual_cost;
1258  double ntuples;
1259  ListCell *l;
1260  double spc_random_page_cost;
1261 
1262  /* Should only be applied to base relations */
1263  Assert(baserel->relid > 0);
1264  Assert(baserel->rtekind == RTE_RELATION);
1265 
1266  /* Mark the path with the correct row estimate */
1267  if (param_info)
1268  path->rows = param_info->ppi_rows;
1269  else
1270  path->rows = baserel->rows;
1271 
1272  /* Count how many tuples we expect to retrieve */
1273  ntuples = 0;
1274  foreach(l, tidquals)
1275  {
1276  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
1277  Expr *qual = rinfo->clause;
1278 
1279  if (IsA(qual, ScalarArrayOpExpr))
1280  {
1281  /* Each element of the array yields 1 tuple */
1282  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) qual;
1283  Node *arraynode = (Node *) lsecond(saop->args);
1284 
1285  ntuples += estimate_array_length(root, arraynode);
1286  }
1287  else if (IsA(qual, CurrentOfExpr))
1288  {
1289  /* CURRENT OF yields 1 tuple */
1290  isCurrentOf = true;
1291  ntuples++;
1292  }
1293  else
1294  {
1295  /* It's just CTID = something, count 1 tuple */
1296  ntuples++;
1297  }
1298  }
1299 
1300  /*
1301  * We must force TID scan for WHERE CURRENT OF, because only nodeTidscan.c
1302  * understands how to do it correctly. Therefore, honor enable_tidscan
1303  * only when CURRENT OF isn't present. Also note that cost_qual_eval
1304  * counts a CurrentOfExpr as having startup cost disable_cost, which we
1305  * subtract off here; that's to prevent other plan types such as seqscan
1306  * from winning.
1307  */
1308  if (isCurrentOf)
1309  {
1311  startup_cost -= disable_cost;
1312  }
1313  else if (!enable_tidscan)
1314  startup_cost += disable_cost;
1315 
1316  /*
1317  * The TID qual expressions will be computed once, any other baserestrict
1318  * quals once per retrieved tuple.
1319  */
1320  cost_qual_eval(&tid_qual_cost, tidquals, root);
1321 
1322  /* fetch estimated page cost for tablespace containing table */
1324  &spc_random_page_cost,
1325  NULL);
1326 
1327  /* disk costs --- assume each tuple on a different page */
1328  run_cost += spc_random_page_cost * ntuples;
1329 
1330  /* Add scanning CPU costs */
1331  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1332 
1333  /* XXX currently we assume TID quals are a subset of qpquals */
1334  startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1335  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1336  tid_qual_cost.per_tuple;
1337  run_cost += cpu_per_tuple * ntuples;
1338 
1339  /* tlist eval costs are paid per output row, not per tuple scanned */
1340  startup_cost += path->pathtarget->cost.startup;
1341  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1342 
1343  path->startup_cost = startup_cost;
1344  path->total_cost = startup_cost + run_cost;
1345 }
1346 
1347 /*
1348  * cost_tidrangescan
1349  * Determines and sets the costs of scanning a relation using a range of
1350  * TIDs for 'path'
1351  *
1352  * 'baserel' is the relation to be scanned
1353  * 'tidrangequals' is the list of TID-checkable range quals
1354  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1355  */
1356 void
1358  RelOptInfo *baserel, List *tidrangequals,
1359  ParamPathInfo *param_info)
1360 {
1361  Selectivity selectivity;
1362  double pages;
1363  Cost startup_cost = 0;
1364  Cost run_cost = 0;
1365  QualCost qpqual_cost;
1366  Cost cpu_per_tuple;
1367  QualCost tid_qual_cost;
1368  double ntuples;
1369  double nseqpages;
1370  double spc_random_page_cost;
1371  double spc_seq_page_cost;
1372 
1373  /* Should only be applied to base relations */
1374  Assert(baserel->relid > 0);
1375  Assert(baserel->rtekind == RTE_RELATION);
1376 
1377  /* Mark the path with the correct row estimate */
1378  if (param_info)
1379  path->rows = param_info->ppi_rows;
1380  else
1381  path->rows = baserel->rows;
1382 
1383  /* Count how many tuples and pages we expect to scan */
1384  selectivity = clauselist_selectivity(root, tidrangequals, baserel->relid,
1385  JOIN_INNER, NULL);
1386  pages = ceil(selectivity * baserel->pages);
1387 
1388  if (pages <= 0.0)
1389  pages = 1.0;
1390 
1391  /*
1392  * The first page in a range requires a random seek, but each subsequent
1393  * page is just a normal sequential page read. NOTE: it's desirable for
1394  * TID Range Scans to cost more than the equivalent Sequential Scans,
1395  * because Seq Scans have some performance advantages such as scan
1396  * synchronization and parallelizability, and we'd prefer one of them to
1397  * be picked unless a TID Range Scan really is better.
1398  */
1399  ntuples = selectivity * baserel->tuples;
1400  nseqpages = pages - 1.0;
1401 
1402  if (!enable_tidscan)
1403  startup_cost += disable_cost;
1404 
1405  /*
1406  * The TID qual expressions will be computed once, any other baserestrict
1407  * quals once per retrieved tuple.
1408  */
1409  cost_qual_eval(&tid_qual_cost, tidrangequals, root);
1410 
1411  /* fetch estimated page cost for tablespace containing table */
1413  &spc_random_page_cost,
1414  &spc_seq_page_cost);
1415 
1416  /* disk costs; 1 random page and the remainder as seq pages */
1417  run_cost += spc_random_page_cost + spc_seq_page_cost * nseqpages;
1418 
1419  /* Add scanning CPU costs */
1420  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1421 
1422  /*
1423  * XXX currently we assume TID quals are a subset of qpquals at this
1424  * point; they will be removed (if possible) when we create the plan, so
1425  * we subtract their cost from the total qpqual cost. (If the TID quals
1426  * can't be removed, this is a mistake and we're going to underestimate
1427  * the CPU cost a bit.)
1428  */
1429  startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1430  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1431  tid_qual_cost.per_tuple;
1432  run_cost += cpu_per_tuple * ntuples;
1433 
1434  /* tlist eval costs are paid per output row, not per tuple scanned */
1435  startup_cost += path->pathtarget->cost.startup;
1436  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1437 
1438  path->startup_cost = startup_cost;
1439  path->total_cost = startup_cost + run_cost;
1440 }
1441 
1442 /*
1443  * cost_subqueryscan
1444  * Determines and returns the cost of scanning a subquery RTE.
1445  *
1446  * 'baserel' is the relation to be scanned
1447  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1448  * 'trivial_pathtarget' is true if the pathtarget is believed to be trivial.
1449  */
1450 void
1452  RelOptInfo *baserel, ParamPathInfo *param_info,
1453  bool trivial_pathtarget)
1454 {
1455  Cost startup_cost;
1456  Cost run_cost;
1457  List *qpquals;
1458  QualCost qpqual_cost;
1459  Cost cpu_per_tuple;
1460 
1461  /* Should only be applied to base relations that are subqueries */
1462  Assert(baserel->relid > 0);
1463  Assert(baserel->rtekind == RTE_SUBQUERY);
1464 
1465  /*
1466  * We compute the rowcount estimate as the subplan's estimate times the
1467  * selectivity of relevant restriction clauses. In simple cases this will
1468  * come out the same as baserel->rows; but when dealing with parallelized
1469  * paths we must do it like this to get the right answer.
1470  */
1471  if (param_info)
1472  qpquals = list_concat_copy(param_info->ppi_clauses,
1473  baserel->baserestrictinfo);
1474  else
1475  qpquals = baserel->baserestrictinfo;
1476 
1477  path->path.rows = clamp_row_est(path->subpath->rows *
1479  qpquals,
1480  0,
1481  JOIN_INNER,
1482  NULL));
1483 
1484  /*
1485  * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1486  * any restriction clauses and tlist that will be attached to the
1487  * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1488  * projection overhead.
1489  */
1490  path->path.startup_cost = path->subpath->startup_cost;
1491  path->path.total_cost = path->subpath->total_cost;
1492 
1493  /*
1494  * However, if there are no relevant restriction clauses and the
1495  * pathtarget is trivial, then we expect that setrefs.c will optimize away
1496  * the SubqueryScan plan node altogether, so we should just make its cost
1497  * and rowcount equal to the input path's.
1498  *
1499  * Note: there are some edge cases where createplan.c will apply a
1500  * different targetlist to the SubqueryScan node, thus falsifying our
1501  * current estimate of whether the target is trivial, and making the cost
1502  * estimate (though not the rowcount) wrong. It does not seem worth the
1503  * extra complication to try to account for that exactly, especially since
1504  * that behavior falsifies other cost estimates as well.
1505  */
1506  if (qpquals == NIL && trivial_pathtarget)
1507  return;
1508 
1509  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1510 
1511  startup_cost = qpqual_cost.startup;
1512  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1513  run_cost = cpu_per_tuple * path->subpath->rows;
1514 
1515  /* tlist eval costs are paid per output row, not per tuple scanned */
1516  startup_cost += path->path.pathtarget->cost.startup;
1517  run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1518 
1519  path->path.startup_cost += startup_cost;
1520  path->path.total_cost += startup_cost + run_cost;
1521 }
1522 
1523 /*
1524  * cost_functionscan
1525  * Determines and returns the cost of scanning a function RTE.
1526  *
1527  * 'baserel' is the relation to be scanned
1528  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1529  */
1530 void
1532  RelOptInfo *baserel, ParamPathInfo *param_info)
1533 {
1534  Cost startup_cost = 0;
1535  Cost run_cost = 0;
1536  QualCost qpqual_cost;
1537  Cost cpu_per_tuple;
1538  RangeTblEntry *rte;
1539  QualCost exprcost;
1540 
1541  /* Should only be applied to base relations that are functions */
1542  Assert(baserel->relid > 0);
1543  rte = planner_rt_fetch(baserel->relid, root);
1544  Assert(rte->rtekind == RTE_FUNCTION);
1545 
1546  /* Mark the path with the correct row estimate */
1547  if (param_info)
1548  path->rows = param_info->ppi_rows;
1549  else
1550  path->rows = baserel->rows;
1551 
1552  /*
1553  * Estimate costs of executing the function expression(s).
1554  *
1555  * Currently, nodeFunctionscan.c always executes the functions to
1556  * completion before returning any rows, and caches the results in a
1557  * tuplestore. So the function eval cost is all startup cost, and per-row
1558  * costs are minimal.
1559  *
1560  * XXX in principle we ought to charge tuplestore spill costs if the
1561  * number of rows is large. However, given how phony our rowcount
1562  * estimates for functions tend to be, there's not a lot of point in that
1563  * refinement right now.
1564  */
1565  cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1566 
1567  startup_cost += exprcost.startup + exprcost.per_tuple;
1568 
1569  /* Add scanning CPU costs */
1570  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1571 
1572  startup_cost += qpqual_cost.startup;
1573  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1574  run_cost += cpu_per_tuple * baserel->tuples;
1575 
1576  /* tlist eval costs are paid per output row, not per tuple scanned */
1577  startup_cost += path->pathtarget->cost.startup;
1578  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1579 
1580  path->startup_cost = startup_cost;
1581  path->total_cost = startup_cost + run_cost;
1582 }
1583 
1584 /*
1585  * cost_tablefuncscan
1586  * Determines and returns the cost of scanning a table function.
1587  *
1588  * 'baserel' is the relation to be scanned
1589  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1590  */
1591 void
1593  RelOptInfo *baserel, ParamPathInfo *param_info)
1594 {
1595  Cost startup_cost = 0;
1596  Cost run_cost = 0;
1597  QualCost qpqual_cost;
1598  Cost cpu_per_tuple;
1599  RangeTblEntry *rte;
1600  QualCost exprcost;
1601 
1602  /* Should only be applied to base relations that are functions */
1603  Assert(baserel->relid > 0);
1604  rte = planner_rt_fetch(baserel->relid, root);
1605  Assert(rte->rtekind == RTE_TABLEFUNC);
1606 
1607  /* Mark the path with the correct row estimate */
1608  if (param_info)
1609  path->rows = param_info->ppi_rows;
1610  else
1611  path->rows = baserel->rows;
1612 
1613  /*
1614  * Estimate costs of executing the table func expression(s).
1615  *
1616  * XXX in principle we ought to charge tuplestore spill costs if the
1617  * number of rows is large. However, given how phony our rowcount
1618  * estimates for tablefuncs tend to be, there's not a lot of point in that
1619  * refinement right now.
1620  */
1621  cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1622 
1623  startup_cost += exprcost.startup + exprcost.per_tuple;
1624 
1625  /* Add scanning CPU costs */
1626  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1627 
1628  startup_cost += qpqual_cost.startup;
1629  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1630  run_cost += cpu_per_tuple * baserel->tuples;
1631 
1632  /* tlist eval costs are paid per output row, not per tuple scanned */
1633  startup_cost += path->pathtarget->cost.startup;
1634  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1635 
1636  path->startup_cost = startup_cost;
1637  path->total_cost = startup_cost + run_cost;
1638 }
1639 
1640 /*
1641  * cost_valuesscan
1642  * Determines and returns the cost of scanning a VALUES RTE.
1643  *
1644  * 'baserel' is the relation to be scanned
1645  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1646  */
1647 void
1649  RelOptInfo *baserel, ParamPathInfo *param_info)
1650 {
1651  Cost startup_cost = 0;
1652  Cost run_cost = 0;
1653  QualCost qpqual_cost;
1654  Cost cpu_per_tuple;
1655 
1656  /* Should only be applied to base relations that are values lists */
1657  Assert(baserel->relid > 0);
1658  Assert(baserel->rtekind == RTE_VALUES);
1659 
1660  /* Mark the path with the correct row estimate */
1661  if (param_info)
1662  path->rows = param_info->ppi_rows;
1663  else
1664  path->rows = baserel->rows;
1665 
1666  /*
1667  * For now, estimate list evaluation cost at one operator eval per list
1668  * (probably pretty bogus, but is it worth being smarter?)
1669  */
1670  cpu_per_tuple = cpu_operator_cost;
1671 
1672  /* Add scanning CPU costs */
1673  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1674 
1675  startup_cost += qpqual_cost.startup;
1676  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1677  run_cost += cpu_per_tuple * baserel->tuples;
1678 
1679  /* tlist eval costs are paid per output row, not per tuple scanned */
1680  startup_cost += path->pathtarget->cost.startup;
1681  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1682 
1683  path->startup_cost = startup_cost;
1684  path->total_cost = startup_cost + run_cost;
1685 }
1686 
1687 /*
1688  * cost_ctescan
1689  * Determines and returns the cost of scanning a CTE RTE.
1690  *
1691  * Note: this is used for both self-reference and regular CTEs; the
1692  * possible cost differences are below the threshold of what we could
1693  * estimate accurately anyway. Note that the costs of evaluating the
1694  * referenced CTE query are added into the final plan as initplan costs,
1695  * and should NOT be counted here.
1696  */
1697 void
1699  RelOptInfo *baserel, ParamPathInfo *param_info)
1700 {
1701  Cost startup_cost = 0;
1702  Cost run_cost = 0;
1703  QualCost qpqual_cost;
1704  Cost cpu_per_tuple;
1705 
1706  /* Should only be applied to base relations that are CTEs */
1707  Assert(baserel->relid > 0);
1708  Assert(baserel->rtekind == RTE_CTE);
1709 
1710  /* Mark the path with the correct row estimate */
1711  if (param_info)
1712  path->rows = param_info->ppi_rows;
1713  else
1714  path->rows = baserel->rows;
1715 
1716  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1717  cpu_per_tuple = cpu_tuple_cost;
1718 
1719  /* Add scanning CPU costs */
1720  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1721 
1722  startup_cost += qpqual_cost.startup;
1723  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1724  run_cost += cpu_per_tuple * baserel->tuples;
1725 
1726  /* tlist eval costs are paid per output row, not per tuple scanned */
1727  startup_cost += path->pathtarget->cost.startup;
1728  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1729 
1730  path->startup_cost = startup_cost;
1731  path->total_cost = startup_cost + run_cost;
1732 }
1733 
1734 /*
1735  * cost_namedtuplestorescan
1736  * Determines and returns the cost of scanning a named tuplestore.
1737  */
1738 void
1740  RelOptInfo *baserel, ParamPathInfo *param_info)
1741 {
1742  Cost startup_cost = 0;
1743  Cost run_cost = 0;
1744  QualCost qpqual_cost;
1745  Cost cpu_per_tuple;
1746 
1747  /* Should only be applied to base relations that are Tuplestores */
1748  Assert(baserel->relid > 0);
1749  Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1750 
1751  /* Mark the path with the correct row estimate */
1752  if (param_info)
1753  path->rows = param_info->ppi_rows;
1754  else
1755  path->rows = baserel->rows;
1756 
1757  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1758  cpu_per_tuple = cpu_tuple_cost;
1759 
1760  /* Add scanning CPU costs */
1761  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1762 
1763  startup_cost += qpqual_cost.startup;
1764  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1765  run_cost += cpu_per_tuple * baserel->tuples;
1766 
1767  path->startup_cost = startup_cost;
1768  path->total_cost = startup_cost + run_cost;
1769 }
1770 
1771 /*
1772  * cost_resultscan
1773  * Determines and returns the cost of scanning an RTE_RESULT relation.
1774  */
1775 void
1777  RelOptInfo *baserel, ParamPathInfo *param_info)
1778 {
1779  Cost startup_cost = 0;
1780  Cost run_cost = 0;
1781  QualCost qpqual_cost;
1782  Cost cpu_per_tuple;
1783 
1784  /* Should only be applied to RTE_RESULT base relations */
1785  Assert(baserel->relid > 0);
1786  Assert(baserel->rtekind == RTE_RESULT);
1787 
1788  /* Mark the path with the correct row estimate */
1789  if (param_info)
1790  path->rows = param_info->ppi_rows;
1791  else
1792  path->rows = baserel->rows;
1793 
1794  /* We charge qual cost plus cpu_tuple_cost */
1795  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1796 
1797  startup_cost += qpqual_cost.startup;
1798  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1799  run_cost += cpu_per_tuple * baserel->tuples;
1800 
1801  path->startup_cost = startup_cost;
1802  path->total_cost = startup_cost + run_cost;
1803 }
1804 
1805 /*
1806  * cost_recursive_union
1807  * Determines and returns the cost of performing a recursive union,
1808  * and also the estimated output size.
1809  *
1810  * We are given Paths for the nonrecursive and recursive terms.
1811  */
1812 void
1813 cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1814 {
1815  Cost startup_cost;
1816  Cost total_cost;
1817  double total_rows;
1818 
1819  /* We probably have decent estimates for the non-recursive term */
1820  startup_cost = nrterm->startup_cost;
1821  total_cost = nrterm->total_cost;
1822  total_rows = nrterm->rows;
1823 
1824  /*
1825  * We arbitrarily assume that about 10 recursive iterations will be
1826  * needed, and that we've managed to get a good fix on the cost and output
1827  * size of each one of them. These are mighty shaky assumptions but it's
1828  * hard to see how to do better.
1829  */
1830  total_cost += 10 * rterm->total_cost;
1831  total_rows += 10 * rterm->rows;
1832 
1833  /*
1834  * Also charge cpu_tuple_cost per row to account for the costs of
1835  * manipulating the tuplestores. (We don't worry about possible
1836  * spill-to-disk costs.)
1837  */
1838  total_cost += cpu_tuple_cost * total_rows;
1839 
1840  runion->startup_cost = startup_cost;
1841  runion->total_cost = total_cost;
1842  runion->rows = total_rows;
1843  runion->pathtarget->width = Max(nrterm->pathtarget->width,
1844  rterm->pathtarget->width);
1845 }
1846 
1847 /*
1848  * cost_tuplesort
1849  * Determines and returns the cost of sorting a relation using tuplesort,
1850  * not including the cost of reading the input data.
1851  *
1852  * If the total volume of data to sort is less than sort_mem, we will do
1853  * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1854  * comparisons for t tuples.
1855  *
1856  * If the total volume exceeds sort_mem, we switch to a tape-style merge
1857  * algorithm. There will still be about t*log2(t) tuple comparisons in
1858  * total, but we will also need to write and read each tuple once per
1859  * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1860  * number of initial runs formed and M is the merge order used by tuplesort.c.
1861  * Since the average initial run should be about sort_mem, we have
1862  * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1863  * cpu = comparison_cost * t * log2(t)
1864  *
1865  * If the sort is bounded (i.e., only the first k result tuples are needed)
1866  * and k tuples can fit into sort_mem, we use a heap method that keeps only
1867  * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1868  *
1869  * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1870  * accesses (XXX can't we refine that guess?)
1871  *
1872  * By default, we charge two operator evals per tuple comparison, which should
1873  * be in the right ballpark in most cases. The caller can tweak this by
1874  * specifying nonzero comparison_cost; typically that's used for any extra
1875  * work that has to be done to prepare the inputs to the comparison operators.
1876  *
1877  * 'tuples' is the number of tuples in the relation
1878  * 'width' is the average tuple width in bytes
1879  * 'comparison_cost' is the extra cost per comparison, if any
1880  * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1881  * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1882  */
1883 static void
1884 cost_tuplesort(Cost *startup_cost, Cost *run_cost,
1885  double tuples, int width,
1886  Cost comparison_cost, int sort_mem,
1887  double limit_tuples)
1888 {
1889  double input_bytes = relation_byte_size(tuples, width);
1890  double output_bytes;
1891  double output_tuples;
1892  long sort_mem_bytes = sort_mem * 1024L;
1893 
1894  /*
1895  * We want to be sure the cost of a sort is never estimated as zero, even
1896  * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1897  */
1898  if (tuples < 2.0)
1899  tuples = 2.0;
1900 
1901  /* Include the default cost-per-comparison */
1902  comparison_cost += 2.0 * cpu_operator_cost;
1903 
1904  /* Do we have a useful LIMIT? */
1905  if (limit_tuples > 0 && limit_tuples < tuples)
1906  {
1907  output_tuples = limit_tuples;
1908  output_bytes = relation_byte_size(output_tuples, width);
1909  }
1910  else
1911  {
1912  output_tuples = tuples;
1913  output_bytes = input_bytes;
1914  }
1915 
1916  if (output_bytes > sort_mem_bytes)
1917  {
1918  /*
1919  * We'll have to use a disk-based sort of all the tuples
1920  */
1921  double npages = ceil(input_bytes / BLCKSZ);
1922  double nruns = input_bytes / sort_mem_bytes;
1923  double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1924  double log_runs;
1925  double npageaccesses;
1926 
1927  /*
1928  * CPU costs
1929  *
1930  * Assume about N log2 N comparisons
1931  */
1932  *startup_cost = comparison_cost * tuples * LOG2(tuples);
1933 
1934  /* Disk costs */
1935 
1936  /* Compute logM(r) as log(r) / log(M) */
1937  if (nruns > mergeorder)
1938  log_runs = ceil(log(nruns) / log(mergeorder));
1939  else
1940  log_runs = 1.0;
1941  npageaccesses = 2.0 * npages * log_runs;
1942  /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1943  *startup_cost += npageaccesses *
1944  (seq_page_cost * 0.75 + random_page_cost * 0.25);
1945  }
1946  else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1947  {
1948  /*
1949  * We'll use a bounded heap-sort keeping just K tuples in memory, for
1950  * a total number of tuple comparisons of N log2 K; but the constant
1951  * factor is a bit higher than for quicksort. Tweak it so that the
1952  * cost curve is continuous at the crossover point.
1953  */
1954  *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
1955  }
1956  else
1957  {
1958  /* We'll use plain quicksort on all the input tuples */
1959  *startup_cost = comparison_cost * tuples * LOG2(tuples);
1960  }
1961 
1962  /*
1963  * Also charge a small amount (arbitrarily set equal to operator cost) per
1964  * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1965  * doesn't do qual-checking or projection, so it has less overhead than
1966  * most plan nodes. Note it's correct to use tuples not output_tuples
1967  * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1968  * counting the LIMIT otherwise.
1969  */
1970  *run_cost = cpu_operator_cost * tuples;
1971 }
1972 
1973 /*
1974  * cost_incremental_sort
1975  * Determines and returns the cost of sorting a relation incrementally, when
1976  * the input path is presorted by a prefix of the pathkeys.
1977  *
1978  * 'presorted_keys' is the number of leading pathkeys by which the input path
1979  * is sorted.
1980  *
1981  * We estimate the number of groups into which the relation is divided by the
1982  * leading pathkeys, and then calculate the cost of sorting a single group
1983  * with tuplesort using cost_tuplesort().
1984  */
1985 void
1987  PlannerInfo *root, List *pathkeys, int presorted_keys,
1988  Cost input_startup_cost, Cost input_total_cost,
1989  double input_tuples, int width, Cost comparison_cost, int sort_mem,
1990  double limit_tuples)
1991 {
1992  Cost startup_cost,
1993  run_cost,
1994  input_run_cost = input_total_cost - input_startup_cost;
1995  double group_tuples,
1996  input_groups;
1997  Cost group_startup_cost,
1998  group_run_cost,
1999  group_input_run_cost;
2000  List *presortedExprs = NIL;
2001  ListCell *l;
2002  bool unknown_varno = false;
2003 
2004  Assert(presorted_keys > 0 && presorted_keys < list_length(pathkeys));
2005 
2006  /*
2007  * We want to be sure the cost of a sort is never estimated as zero, even
2008  * if passed-in tuple count is zero. Besides, mustn't do log(0)...
2009  */
2010  if (input_tuples < 2.0)
2011  input_tuples = 2.0;
2012 
2013  /* Default estimate of number of groups, capped to one group per row. */
2014  input_groups = Min(input_tuples, DEFAULT_NUM_DISTINCT);
2015 
2016  /*
2017  * Extract presorted keys as list of expressions.
2018  *
2019  * We need to be careful about Vars containing "varno 0" which might have
2020  * been introduced by generate_append_tlist, which would confuse
2021  * estimate_num_groups (in fact it'd fail for such expressions). See
2022  * recurse_set_operations which has to deal with the same issue.
2023  *
2024  * Unlike recurse_set_operations we can't access the original target list
2025  * here, and even if we could it's not very clear how useful would that be
2026  * for a set operation combining multiple tables. So we simply detect if
2027  * there are any expressions with "varno 0" and use the default
2028  * DEFAULT_NUM_DISTINCT in that case.
2029  *
2030  * We might also use either 1.0 (a single group) or input_tuples (each row
2031  * being a separate group), pretty much the worst and best case for
2032  * incremental sort. But those are extreme cases and using something in
2033  * between seems reasonable. Furthermore, generate_append_tlist is used
2034  * for set operations, which are likely to produce mostly unique output
2035  * anyway - from that standpoint the DEFAULT_NUM_DISTINCT is defensive
2036  * while maintaining lower startup cost.
2037  */
2038  foreach(l, pathkeys)
2039  {
2040  PathKey *key = (PathKey *) lfirst(l);
2041  EquivalenceMember *member = (EquivalenceMember *)
2042  linitial(key->pk_eclass->ec_members);
2043 
2044  /*
2045  * Check if the expression contains Var with "varno 0" so that we
2046  * don't call estimate_num_groups in that case.
2047  */
2048  if (bms_is_member(0, pull_varnos(root, (Node *) member->em_expr)))
2049  {
2050  unknown_varno = true;
2051  break;
2052  }
2053 
2054  /* expression not containing any Vars with "varno 0" */
2055  presortedExprs = lappend(presortedExprs, member->em_expr);
2056 
2057  if (foreach_current_index(l) + 1 >= presorted_keys)
2058  break;
2059  }
2060 
2061  /* Estimate the number of groups with equal presorted keys. */
2062  if (!unknown_varno)
2063  input_groups = estimate_num_groups(root, presortedExprs, input_tuples,
2064  NULL, NULL);
2065 
2066  group_tuples = input_tuples / input_groups;
2067  group_input_run_cost = input_run_cost / input_groups;
2068 
2069  /*
2070  * Estimate the average cost of sorting of one group where presorted keys
2071  * are equal.
2072  */
2073  cost_tuplesort(&group_startup_cost, &group_run_cost,
2074  group_tuples, width, comparison_cost, sort_mem,
2075  limit_tuples);
2076 
2077  /*
2078  * Startup cost of incremental sort is the startup cost of its first group
2079  * plus the cost of its input.
2080  */
2081  startup_cost = group_startup_cost + input_startup_cost +
2082  group_input_run_cost;
2083 
2084  /*
2085  * After we started producing tuples from the first group, the cost of
2086  * producing all the tuples is given by the cost to finish processing this
2087  * group, plus the total cost to process the remaining groups, plus the
2088  * remaining cost of input.
2089  */
2090  run_cost = group_run_cost + (group_run_cost + group_startup_cost) *
2091  (input_groups - 1) + group_input_run_cost * (input_groups - 1);
2092 
2093  /*
2094  * Incremental sort adds some overhead by itself. Firstly, it has to
2095  * detect the sort groups. This is roughly equal to one extra copy and
2096  * comparison per tuple.
2097  */
2098  run_cost += (cpu_tuple_cost + comparison_cost) * input_tuples;
2099 
2100  /*
2101  * Additionally, we charge double cpu_tuple_cost for each input group to
2102  * account for the tuplesort_reset that's performed after each group.
2103  */
2104  run_cost += 2.0 * cpu_tuple_cost * input_groups;
2105 
2106  path->rows = input_tuples;
2107  path->startup_cost = startup_cost;
2108  path->total_cost = startup_cost + run_cost;
2109 }
2110 
2111 /*
2112  * cost_sort
2113  * Determines and returns the cost of sorting a relation, including
2114  * the cost of reading the input data.
2115  *
2116  * NOTE: some callers currently pass NIL for pathkeys because they
2117  * can't conveniently supply the sort keys. Since this routine doesn't
2118  * currently do anything with pathkeys anyway, that doesn't matter...
2119  * but if it ever does, it should react gracefully to lack of key data.
2120  * (Actually, the thing we'd most likely be interested in is just the number
2121  * of sort keys, which all callers *could* supply.)
2122  */
2123 void
2125  List *pathkeys, Cost input_cost, double tuples, int width,
2126  Cost comparison_cost, int sort_mem,
2127  double limit_tuples)
2128 
2129 {
2130  Cost startup_cost;
2131  Cost run_cost;
2132 
2133  cost_tuplesort(&startup_cost, &run_cost,
2134  tuples, width,
2135  comparison_cost, sort_mem,
2136  limit_tuples);
2137 
2138  if (!enable_sort)
2139  startup_cost += disable_cost;
2140 
2141  startup_cost += input_cost;
2142 
2143  path->rows = tuples;
2144  path->startup_cost = startup_cost;
2145  path->total_cost = startup_cost + run_cost;
2146 }
2147 
2148 /*
2149  * append_nonpartial_cost
2150  * Estimate the cost of the non-partial paths in a Parallel Append.
2151  * The non-partial paths are assumed to be the first "numpaths" paths
2152  * from the subpaths list, and to be in order of decreasing cost.
2153  */
2154 static Cost
2155 append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
2156 {
2157  Cost *costarr;
2158  int arrlen;
2159  ListCell *l;
2160  ListCell *cell;
2161  int path_index;
2162  int min_index;
2163  int max_index;
2164 
2165  if (numpaths == 0)
2166  return 0;
2167 
2168  /*
2169  * Array length is number of workers or number of relevant paths,
2170  * whichever is less.
2171  */
2172  arrlen = Min(parallel_workers, numpaths);
2173  costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
2174 
2175  /* The first few paths will each be claimed by a different worker. */
2176  path_index = 0;
2177  foreach(cell, subpaths)
2178  {
2179  Path *subpath = (Path *) lfirst(cell);
2180 
2181  if (path_index == arrlen)
2182  break;
2183  costarr[path_index++] = subpath->total_cost;
2184  }
2185 
2186  /*
2187  * Since subpaths are sorted by decreasing cost, the last one will have
2188  * the minimum cost.
2189  */
2190  min_index = arrlen - 1;
2191 
2192  /*
2193  * For each of the remaining subpaths, add its cost to the array element
2194  * with minimum cost.
2195  */
2196  for_each_cell(l, subpaths, cell)
2197  {
2198  Path *subpath = (Path *) lfirst(l);
2199 
2200  /* Consider only the non-partial paths */
2201  if (path_index++ == numpaths)
2202  break;
2203 
2204  costarr[min_index] += subpath->total_cost;
2205 
2206  /* Update the new min cost array index */
2207  min_index = 0;
2208  for (int i = 0; i < arrlen; i++)
2209  {
2210  if (costarr[i] < costarr[min_index])
2211  min_index = i;
2212  }
2213  }
2214 
2215  /* Return the highest cost from the array */
2216  max_index = 0;
2217  for (int i = 0; i < arrlen; i++)
2218  {
2219  if (costarr[i] > costarr[max_index])
2220  max_index = i;
2221  }
2222 
2223  return costarr[max_index];
2224 }
2225 
2226 /*
2227  * cost_append
2228  * Determines and returns the cost of an Append node.
2229  */
2230 void
2232 {
2233  ListCell *l;
2234 
2235  apath->path.startup_cost = 0;
2236  apath->path.total_cost = 0;
2237  apath->path.rows = 0;
2238 
2239  if (apath->subpaths == NIL)
2240  return;
2241 
2242  if (!apath->path.parallel_aware)
2243  {
2244  List *pathkeys = apath->path.pathkeys;
2245 
2246  if (pathkeys == NIL)
2247  {
2248  Path *firstsubpath = (Path *) linitial(apath->subpaths);
2249 
2250  /*
2251  * For an unordered, non-parallel-aware Append we take the startup
2252  * cost as the startup cost of the first subpath.
2253  */
2254  apath->path.startup_cost = firstsubpath->startup_cost;
2255 
2256  /* Compute rows and costs as sums of subplan rows and costs. */
2257  foreach(l, apath->subpaths)
2258  {
2259  Path *subpath = (Path *) lfirst(l);
2260 
2261  apath->path.rows += subpath->rows;
2262  apath->path.total_cost += subpath->total_cost;
2263  }
2264  }
2265  else
2266  {
2267  /*
2268  * For an ordered, non-parallel-aware Append we take the startup
2269  * cost as the sum of the subpath startup costs. This ensures
2270  * that we don't underestimate the startup cost when a query's
2271  * LIMIT is such that several of the children have to be run to
2272  * satisfy it. This might be overkill --- another plausible hack
2273  * would be to take the Append's startup cost as the maximum of
2274  * the child startup costs. But we don't want to risk believing
2275  * that an ORDER BY LIMIT query can be satisfied at small cost
2276  * when the first child has small startup cost but later ones
2277  * don't. (If we had the ability to deal with nonlinear cost
2278  * interpolation for partial retrievals, we would not need to be
2279  * so conservative about this.)
2280  *
2281  * This case is also different from the above in that we have to
2282  * account for possibly injecting sorts into subpaths that aren't
2283  * natively ordered.
2284  */
2285  foreach(l, apath->subpaths)
2286  {
2287  Path *subpath = (Path *) lfirst(l);
2288  Path sort_path; /* dummy for result of cost_sort */
2289 
2290  if (!pathkeys_contained_in(pathkeys, subpath->pathkeys))
2291  {
2292  /*
2293  * We'll need to insert a Sort node, so include costs for
2294  * that. We can use the parent's LIMIT if any, since we
2295  * certainly won't pull more than that many tuples from
2296  * any child.
2297  */
2298  cost_sort(&sort_path,
2299  NULL, /* doesn't currently need root */
2300  pathkeys,
2301  subpath->total_cost,
2302  subpath->rows,
2303  subpath->pathtarget->width,
2304  0.0,
2305  work_mem,
2306  apath->limit_tuples);
2307  subpath = &sort_path;
2308  }
2309 
2310  apath->path.rows += subpath->rows;
2311  apath->path.startup_cost += subpath->startup_cost;
2312  apath->path.total_cost += subpath->total_cost;
2313  }
2314  }
2315  }
2316  else /* parallel-aware */
2317  {
2318  int i = 0;
2319  double parallel_divisor = get_parallel_divisor(&apath->path);
2320 
2321  /* Parallel-aware Append never produces ordered output. */
2322  Assert(apath->path.pathkeys == NIL);
2323 
2324  /* Calculate startup cost. */
2325  foreach(l, apath->subpaths)
2326  {
2327  Path *subpath = (Path *) lfirst(l);
2328 
2329  /*
2330  * Append will start returning tuples when the child node having
2331  * lowest startup cost is done setting up. We consider only the
2332  * first few subplans that immediately get a worker assigned.
2333  */
2334  if (i == 0)
2335  apath->path.startup_cost = subpath->startup_cost;
2336  else if (i < apath->path.parallel_workers)
2337  apath->path.startup_cost = Min(apath->path.startup_cost,
2338  subpath->startup_cost);
2339 
2340  /*
2341  * Apply parallel divisor to subpaths. Scale the number of rows
2342  * for each partial subpath based on the ratio of the parallel
2343  * divisor originally used for the subpath to the one we adopted.
2344  * Also add the cost of partial paths to the total cost, but
2345  * ignore non-partial paths for now.
2346  */
2347  if (i < apath->first_partial_path)
2348  apath->path.rows += subpath->rows / parallel_divisor;
2349  else
2350  {
2351  double subpath_parallel_divisor;
2352 
2353  subpath_parallel_divisor = get_parallel_divisor(subpath);
2354  apath->path.rows += subpath->rows * (subpath_parallel_divisor /
2355  parallel_divisor);
2356  apath->path.total_cost += subpath->total_cost;
2357  }
2358 
2359  apath->path.rows = clamp_row_est(apath->path.rows);
2360 
2361  i++;
2362  }
2363 
2364  /* Add cost for non-partial subpaths. */
2365  apath->path.total_cost +=
2367  apath->first_partial_path,
2368  apath->path.parallel_workers);
2369  }
2370 
2371  /*
2372  * Although Append does not do any selection or projection, it's not free;
2373  * add a small per-tuple overhead.
2374  */
2375  apath->path.total_cost +=
2377 }
2378 
2379 /*
2380  * cost_merge_append
2381  * Determines and returns the cost of a MergeAppend node.
2382  *
2383  * MergeAppend merges several pre-sorted input streams, using a heap that
2384  * at any given instant holds the next tuple from each stream. If there
2385  * are N streams, we need about N*log2(N) tuple comparisons to construct
2386  * the heap at startup, and then for each output tuple, about log2(N)
2387  * comparisons to replace the top entry.
2388  *
2389  * (The effective value of N will drop once some of the input streams are
2390  * exhausted, but it seems unlikely to be worth trying to account for that.)
2391  *
2392  * The heap is never spilled to disk, since we assume N is not very large.
2393  * So this is much simpler than cost_sort.
2394  *
2395  * As in cost_sort, we charge two operator evals per tuple comparison.
2396  *
2397  * 'pathkeys' is a list of sort keys
2398  * 'n_streams' is the number of input streams
2399  * 'input_startup_cost' is the sum of the input streams' startup costs
2400  * 'input_total_cost' is the sum of the input streams' total costs
2401  * 'tuples' is the number of tuples in all the streams
2402  */
2403 void
2405  List *pathkeys, int n_streams,
2406  Cost input_startup_cost, Cost input_total_cost,
2407  double tuples)
2408 {
2409  Cost startup_cost = 0;
2410  Cost run_cost = 0;
2411  Cost comparison_cost;
2412  double N;
2413  double logN;
2414 
2415  /*
2416  * Avoid log(0)...
2417  */
2418  N = (n_streams < 2) ? 2.0 : (double) n_streams;
2419  logN = LOG2(N);
2420 
2421  /* Assumed cost per tuple comparison */
2422  comparison_cost = 2.0 * cpu_operator_cost;
2423 
2424  /* Heap creation cost */
2425  startup_cost += comparison_cost * N * logN;
2426 
2427  /* Per-tuple heap maintenance cost */
2428  run_cost += tuples * comparison_cost * logN;
2429 
2430  /*
2431  * Although MergeAppend does not do any selection or projection, it's not
2432  * free; add a small per-tuple overhead.
2433  */
2434  run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
2435 
2436  path->startup_cost = startup_cost + input_startup_cost;
2437  path->total_cost = startup_cost + run_cost + input_total_cost;
2438 }
2439 
2440 /*
2441  * cost_material
2442  * Determines and returns the cost of materializing a relation, including
2443  * the cost of reading the input data.
2444  *
2445  * If the total volume of data to materialize exceeds work_mem, we will need
2446  * to write it to disk, so the cost is much higher in that case.
2447  *
2448  * Note that here we are estimating the costs for the first scan of the
2449  * relation, so the materialization is all overhead --- any savings will
2450  * occur only on rescan, which is estimated in cost_rescan.
2451  */
2452 void
2454  Cost input_startup_cost, Cost input_total_cost,
2455  double tuples, int width)
2456 {
2457  Cost startup_cost = input_startup_cost;
2458  Cost run_cost = input_total_cost - input_startup_cost;
2459  double nbytes = relation_byte_size(tuples, width);
2460  long work_mem_bytes = work_mem * 1024L;
2461 
2462  path->rows = tuples;
2463 
2464  /*
2465  * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2466  * reflect bookkeeping overhead. (This rate must be more than what
2467  * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2468  * if it is exactly the same then there will be a cost tie between
2469  * nestloop with A outer, materialized B inner and nestloop with B outer,
2470  * materialized A inner. The extra cost ensures we'll prefer
2471  * materializing the smaller rel.) Note that this is normally a good deal
2472  * less than cpu_tuple_cost; which is OK because a Material plan node
2473  * doesn't do qual-checking or projection, so it's got less overhead than
2474  * most plan nodes.
2475  */
2476  run_cost += 2 * cpu_operator_cost * tuples;
2477 
2478  /*
2479  * If we will spill to disk, charge at the rate of seq_page_cost per page.
2480  * This cost is assumed to be evenly spread through the plan run phase,
2481  * which isn't exactly accurate but our cost model doesn't allow for
2482  * nonuniform costs within the run phase.
2483  */
2484  if (nbytes > work_mem_bytes)
2485  {
2486  double npages = ceil(nbytes / BLCKSZ);
2487 
2488  run_cost += seq_page_cost * npages;
2489  }
2490 
2491  path->startup_cost = startup_cost;
2492  path->total_cost = startup_cost + run_cost;
2493 }
2494 
2495 /*
2496  * cost_memoize_rescan
2497  * Determines the estimated cost of rescanning a Memoize node.
2498  *
2499  * In order to estimate this, we must gain knowledge of how often we expect to
2500  * be called and how many distinct sets of parameters we are likely to be
2501  * called with. If we expect a good cache hit ratio, then we can set our
2502  * costs to account for that hit ratio, plus a little bit of cost for the
2503  * caching itself. Caching will not work out well if we expect to be called
2504  * with too many distinct parameter values. The worst-case here is that we
2505  * never see any parameter value twice, in which case we'd never get a cache
2506  * hit and caching would be a complete waste of effort.
2507  */
2508 static void
2510  Cost *rescan_startup_cost, Cost *rescan_total_cost)
2511 {
2512  EstimationInfo estinfo;
2513  ListCell *lc;
2514  Cost input_startup_cost = mpath->subpath->startup_cost;
2515  Cost input_total_cost = mpath->subpath->total_cost;
2516  double tuples = mpath->subpath->rows;
2517  double calls = mpath->calls;
2518  int width = mpath->subpath->pathtarget->width;
2519 
2520  double hash_mem_bytes;
2521  double est_entry_bytes;
2522  double est_cache_entries;
2523  double ndistinct;
2524  double evict_ratio;
2525  double hit_ratio;
2526  Cost startup_cost;
2527  Cost total_cost;
2528 
2529  /* available cache space */
2530  hash_mem_bytes = get_hash_memory_limit();
2531 
2532  /*
2533  * Set the number of bytes each cache entry should consume in the cache.
2534  * To provide us with better estimations on how many cache entries we can
2535  * store at once, we make a call to the executor here to ask it what
2536  * memory overheads there are for a single cache entry.
2537  */
2538  est_entry_bytes = relation_byte_size(tuples, width) +
2540 
2541  /* include the estimated width for the cache keys */
2542  foreach(lc, mpath->param_exprs)
2543  est_entry_bytes += get_expr_width(root, (Node *) lfirst(lc));
2544 
2545  /* estimate on the upper limit of cache entries we can hold at once */
2546  est_cache_entries = floor(hash_mem_bytes / est_entry_bytes);
2547 
2548  /* estimate on the distinct number of parameter values */
2549  ndistinct = estimate_num_groups(root, mpath->param_exprs, calls, NULL,
2550  &estinfo);
2551 
2552  /*
2553  * When the estimation fell back on using a default value, it's a bit too
2554  * risky to assume that it's ok to use a Memoize node. The use of a
2555  * default could cause us to use a Memoize node when it's really
2556  * inappropriate to do so. If we see that this has been done, then we'll
2557  * assume that every call will have unique parameters, which will almost
2558  * certainly mean a MemoizePath will never survive add_path().
2559  */
2560  if ((estinfo.flags & SELFLAG_USED_DEFAULT) != 0)
2561  ndistinct = calls;
2562 
2563  /*
2564  * Since we've already estimated the maximum number of entries we can
2565  * store at once and know the estimated number of distinct values we'll be
2566  * called with, we'll take this opportunity to set the path's est_entries.
2567  * This will ultimately determine the hash table size that the executor
2568  * will use. If we leave this at zero, the executor will just choose the
2569  * size itself. Really this is not the right place to do this, but it's
2570  * convenient since everything is already calculated.
2571  */
2572  mpath->est_entries = Min(Min(ndistinct, est_cache_entries),
2573  PG_UINT32_MAX);
2574 
2575  /*
2576  * When the number of distinct parameter values is above the amount we can
2577  * store in the cache, then we'll have to evict some entries from the
2578  * cache. This is not free. Here we estimate how often we'll incur the
2579  * cost of that eviction.
2580  */
2581  evict_ratio = 1.0 - Min(est_cache_entries, ndistinct) / ndistinct;
2582 
2583  /*
2584  * In order to estimate how costly a single scan will be, we need to
2585  * attempt to estimate what the cache hit ratio will be. To do that we
2586  * must look at how many scans are estimated in total for this node and
2587  * how many of those scans we expect to get a cache hit.
2588  */
2589  hit_ratio = ((calls - ndistinct) / calls) *
2590  (est_cache_entries / Max(ndistinct, est_cache_entries));
2591 
2592  Assert(hit_ratio >= 0 && hit_ratio <= 1.0);
2593 
2594  /*
2595  * Set the total_cost accounting for the expected cache hit ratio. We
2596  * also add on a cpu_operator_cost to account for a cache lookup. This
2597  * will happen regardless of whether it's a cache hit or not.
2598  */
2599  total_cost = input_total_cost * (1.0 - hit_ratio) + cpu_operator_cost;
2600 
2601  /* Now adjust the total cost to account for cache evictions */
2602 
2603  /* Charge a cpu_tuple_cost for evicting the actual cache entry */
2604  total_cost += cpu_tuple_cost * evict_ratio;
2605 
2606  /*
2607  * Charge a 10th of cpu_operator_cost to evict every tuple in that entry.
2608  * The per-tuple eviction is really just a pfree, so charging a whole
2609  * cpu_operator_cost seems a little excessive.
2610  */
2611  total_cost += cpu_operator_cost / 10.0 * evict_ratio * tuples;
2612 
2613  /*
2614  * Now adjust for storing things in the cache, since that's not free
2615  * either. Everything must go in the cache. We don't proportion this
2616  * over any ratio, just apply it once for the scan. We charge a
2617  * cpu_tuple_cost for the creation of the cache entry and also a
2618  * cpu_operator_cost for each tuple we expect to cache.
2619  */
2620  total_cost += cpu_tuple_cost + cpu_operator_cost * tuples;
2621 
2622  /*
2623  * Getting the first row must be also be proportioned according to the
2624  * expected cache hit ratio.
2625  */
2626  startup_cost = input_startup_cost * (1.0 - hit_ratio);
2627 
2628  /*
2629  * Additionally we charge a cpu_tuple_cost to account for cache lookups,
2630  * which we'll do regardless of whether it was a cache hit or not.
2631  */
2632  startup_cost += cpu_tuple_cost;
2633 
2634  *rescan_startup_cost = startup_cost;
2635  *rescan_total_cost = total_cost;
2636 }
2637 
2638 /*
2639  * cost_agg
2640  * Determines and returns the cost of performing an Agg plan node,
2641  * including the cost of its input.
2642  *
2643  * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2644  * we are using a hashed Agg node just to do grouping).
2645  *
2646  * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2647  * are for appropriately-sorted input.
2648  */
2649 void
2651  AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2652  int numGroupCols, double numGroups,
2653  List *quals,
2654  Cost input_startup_cost, Cost input_total_cost,
2655  double input_tuples, double input_width)
2656 {
2657  double output_tuples;
2658  Cost startup_cost;
2659  Cost total_cost;
2660  AggClauseCosts dummy_aggcosts;
2661 
2662  /* Use all-zero per-aggregate costs if NULL is passed */
2663  if (aggcosts == NULL)
2664  {
2665  Assert(aggstrategy == AGG_HASHED);
2666  MemSet(&dummy_aggcosts, 0, sizeof(AggClauseCosts));
2667  aggcosts = &dummy_aggcosts;
2668  }
2669 
2670  /*
2671  * The transCost.per_tuple component of aggcosts should be charged once
2672  * per input tuple, corresponding to the costs of evaluating the aggregate
2673  * transfns and their input expressions. The finalCost.per_tuple component
2674  * is charged once per output tuple, corresponding to the costs of
2675  * evaluating the finalfns. Startup costs are of course charged but once.
2676  *
2677  * If we are grouping, we charge an additional cpu_operator_cost per
2678  * grouping column per input tuple for grouping comparisons.
2679  *
2680  * We will produce a single output tuple if not grouping, and a tuple per
2681  * group otherwise. We charge cpu_tuple_cost for each output tuple.
2682  *
2683  * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2684  * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2685  * input path is already sorted appropriately, AGG_SORTED should be
2686  * preferred (since it has no risk of memory overflow). This will happen
2687  * as long as the computed total costs are indeed exactly equal --- but if
2688  * there's roundoff error we might do the wrong thing. So be sure that
2689  * the computations below form the same intermediate values in the same
2690  * order.
2691  */
2692  if (aggstrategy == AGG_PLAIN)
2693  {
2694  startup_cost = input_total_cost;
2695  startup_cost += aggcosts->transCost.startup;
2696  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2697  startup_cost += aggcosts->finalCost.startup;
2698  startup_cost += aggcosts->finalCost.per_tuple;
2699  /* we aren't grouping */
2700  total_cost = startup_cost + cpu_tuple_cost;
2701  output_tuples = 1;
2702  }
2703  else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2704  {
2705  /* Here we are able to deliver output on-the-fly */
2706  startup_cost = input_startup_cost;
2707  total_cost = input_total_cost;
2708  if (aggstrategy == AGG_MIXED && !enable_hashagg)
2709  {
2710  startup_cost += disable_cost;
2711  total_cost += disable_cost;
2712  }
2713  /* calcs phrased this way to match HASHED case, see note above */
2714  total_cost += aggcosts->transCost.startup;
2715  total_cost += aggcosts->transCost.per_tuple * input_tuples;
2716  total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2717  total_cost += aggcosts->finalCost.startup;
2718  total_cost += aggcosts->finalCost.per_tuple * numGroups;
2719  total_cost += cpu_tuple_cost * numGroups;
2720  output_tuples = numGroups;
2721  }
2722  else
2723  {
2724  /* must be AGG_HASHED */
2725  startup_cost = input_total_cost;
2726  if (!enable_hashagg)
2727  startup_cost += disable_cost;
2728  startup_cost += aggcosts->transCost.startup;
2729  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2730  /* cost of computing hash value */
2731  startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2732  startup_cost += aggcosts->finalCost.startup;
2733 
2734  total_cost = startup_cost;
2735  total_cost += aggcosts->finalCost.per_tuple * numGroups;
2736  /* cost of retrieving from hash table */
2737  total_cost += cpu_tuple_cost * numGroups;
2738  output_tuples = numGroups;
2739  }
2740 
2741  /*
2742  * Add the disk costs of hash aggregation that spills to disk.
2743  *
2744  * Groups that go into the hash table stay in memory until finalized, so
2745  * spilling and reprocessing tuples doesn't incur additional invocations
2746  * of transCost or finalCost. Furthermore, the computed hash value is
2747  * stored with the spilled tuples, so we don't incur extra invocations of
2748  * the hash function.
2749  *
2750  * Hash Agg begins returning tuples after the first batch is complete.
2751  * Accrue writes (spilled tuples) to startup_cost and to total_cost;
2752  * accrue reads only to total_cost.
2753  */
2754  if (aggstrategy == AGG_HASHED || aggstrategy == AGG_MIXED)
2755  {
2756  double pages;
2757  double pages_written = 0.0;
2758  double pages_read = 0.0;
2759  double spill_cost;
2760  double hashentrysize;
2761  double nbatches;
2762  Size mem_limit;
2763  uint64 ngroups_limit;
2764  int num_partitions;
2765  int depth;
2766 
2767  /*
2768  * Estimate number of batches based on the computed limits. If less
2769  * than or equal to one, all groups are expected to fit in memory;
2770  * otherwise we expect to spill.
2771  */
2772  hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
2773  input_width,
2774  aggcosts->transitionSpace);
2775  hash_agg_set_limits(hashentrysize, numGroups, 0, &mem_limit,
2776  &ngroups_limit, &num_partitions);
2777 
2778  nbatches = Max((numGroups * hashentrysize) / mem_limit,
2779  numGroups / ngroups_limit);
2780 
2781  nbatches = Max(ceil(nbatches), 1.0);
2782  num_partitions = Max(num_partitions, 2);
2783 
2784  /*
2785  * The number of partitions can change at different levels of
2786  * recursion; but for the purposes of this calculation assume it stays
2787  * constant.
2788  */
2789  depth = ceil(log(nbatches) / log(num_partitions));
2790 
2791  /*
2792  * Estimate number of pages read and written. For each level of
2793  * recursion, a tuple must be written and then later read.
2794  */
2795  pages = relation_byte_size(input_tuples, input_width) / BLCKSZ;
2796  pages_written = pages_read = pages * depth;
2797 
2798  /*
2799  * HashAgg has somewhat worse IO behavior than Sort on typical
2800  * hardware/OS combinations. Account for this with a generic penalty.
2801  */
2802  pages_read *= 2.0;
2803  pages_written *= 2.0;
2804 
2805  startup_cost += pages_written * random_page_cost;
2806  total_cost += pages_written * random_page_cost;
2807  total_cost += pages_read * seq_page_cost;
2808 
2809  /* account for CPU cost of spilling a tuple and reading it back */
2810  spill_cost = depth * input_tuples * 2.0 * cpu_tuple_cost;
2811  startup_cost += spill_cost;
2812  total_cost += spill_cost;
2813  }
2814 
2815  /*
2816  * If there are quals (HAVING quals), account for their cost and
2817  * selectivity.
2818  */
2819  if (quals)
2820  {
2821  QualCost qual_cost;
2822 
2823  cost_qual_eval(&qual_cost, quals, root);
2824  startup_cost += qual_cost.startup;
2825  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2826 
2827  output_tuples = clamp_row_est(output_tuples *
2829  quals,
2830  0,
2831  JOIN_INNER,
2832  NULL));
2833  }
2834 
2835  path->rows = output_tuples;
2836  path->startup_cost = startup_cost;
2837  path->total_cost = total_cost;
2838 }
2839 
2840 /*
2841  * get_windowclause_startup_tuples
2842  * Estimate how many tuples we'll need to fetch from a WindowAgg's
2843  * subnode before we can output the first WindowAgg tuple.
2844  *
2845  * How many tuples need to be read depends on the WindowClause. For example,
2846  * a WindowClause with no PARTITION BY and no ORDER BY requires that all
2847  * subnode tuples are read and aggregated before the WindowAgg can output
2848  * anything. If there's a PARTITION BY, then we only need to look at tuples
2849  * in the first partition. Here we attempt to estimate just how many
2850  * 'input_tuples' the WindowAgg will need to read for the given WindowClause
2851  * before the first tuple can be output.
2852  */
2853 static double
2855  double input_tuples)
2856 {
2857  int frameOptions = wc->frameOptions;
2858  double partition_tuples;
2859  double return_tuples;
2860  double peer_tuples;
2861 
2862  /*
2863  * First, figure out how many partitions there are likely to be and set
2864  * partition_tuples according to that estimate.
2865  */
2866  if (wc->partitionClause != NIL)
2867  {
2868  double num_partitions;
2869  List *partexprs = get_sortgrouplist_exprs(wc->partitionClause,
2870  root->parse->targetList);
2871 
2872  num_partitions = estimate_num_groups(root, partexprs, input_tuples,
2873  NULL, NULL);
2874  list_free(partexprs);
2875 
2876  partition_tuples = input_tuples / num_partitions;
2877  }
2878  else
2879  {
2880  /* all tuples belong to the same partition */
2881  partition_tuples = input_tuples;
2882  }
2883 
2884  /* estimate the number of tuples in each peer group */
2885  if (wc->orderClause != NIL)
2886  {
2887  double num_groups;
2888  List *orderexprs;
2889 
2890  orderexprs = get_sortgrouplist_exprs(wc->orderClause,
2891  root->parse->targetList);
2892 
2893  /* estimate out how many peer groups there are in the partition */
2894  num_groups = estimate_num_groups(root, orderexprs,
2895  partition_tuples, NULL,
2896  NULL);
2897  list_free(orderexprs);
2898  peer_tuples = partition_tuples / num_groups;
2899  }
2900  else
2901  {
2902  /* no ORDER BY so only 1 tuple belongs in each peer group */
2903  peer_tuples = 1.0;
2904  }
2905 
2906  if (frameOptions & FRAMEOPTION_END_UNBOUNDED_FOLLOWING)
2907  {
2908  /* include all partition rows */
2909  return_tuples = partition_tuples;
2910  }
2911  else if (frameOptions & FRAMEOPTION_END_CURRENT_ROW)
2912  {
2913  if (frameOptions & FRAMEOPTION_ROWS)
2914  {
2915  /* just count the current row */
2916  return_tuples = 1.0;
2917  }
2918  else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
2919  {
2920  /*
2921  * When in RANGE/GROUPS mode, it's more complex. If there's no
2922  * ORDER BY, then all rows in the partition are peers, otherwise
2923  * we'll need to read the first group of peers.
2924  */
2925  if (wc->orderClause == NIL)
2926  return_tuples = partition_tuples;
2927  else
2928  return_tuples = peer_tuples;
2929  }
2930  else
2931  {
2932  /*
2933  * Something new we don't support yet? This needs attention.
2934  * We'll just return 1.0 in the meantime.
2935  */
2936  Assert(false);
2937  return_tuples = 1.0;
2938  }
2939  }
2940  else if (frameOptions & FRAMEOPTION_END_OFFSET_PRECEDING)
2941  {
2942  /*
2943  * BETWEEN ... AND N PRECEDING will only need to read the WindowAgg's
2944  * subnode after N ROWS/RANGES/GROUPS. N can be 0, but not negative,
2945  * so we'll just assume only the current row needs to be read to fetch
2946  * the first WindowAgg row.
2947  */
2948  return_tuples = 1.0;
2949  }
2950  else if (frameOptions & FRAMEOPTION_END_OFFSET_FOLLOWING)
2951  {
2952  Const *endOffset = (Const *) wc->endOffset;
2953  double end_offset_value;
2954 
2955  /* try and figure out the value specified in the endOffset. */
2956  if (IsA(endOffset, Const))
2957  {
2958  if (endOffset->constisnull)
2959  {
2960  /*
2961  * NULLs are not allowed, but currently, there's no code to
2962  * error out if there's a NULL Const. We'll only discover
2963  * this during execution. For now, just pretend everything is
2964  * fine and assume that just the first row/range/group will be
2965  * needed.
2966  */
2967  end_offset_value = 1.0;
2968  }
2969  else
2970  {
2971  switch (endOffset->consttype)
2972  {
2973  case INT2OID:
2974  end_offset_value =
2975  (double) DatumGetInt16(endOffset->constvalue);
2976  break;
2977  case INT4OID:
2978  end_offset_value =
2979  (double) DatumGetInt32(endOffset->constvalue);
2980  break;
2981  case INT8OID:
2982  end_offset_value =
2983  (double) DatumGetInt64(endOffset->constvalue);
2984  break;
2985  default:
2986  end_offset_value =
2987  partition_tuples / peer_tuples *
2989  break;
2990  }
2991  }
2992  }
2993  else
2994  {
2995  /*
2996  * When the end bound is not a Const, we'll just need to guess. We
2997  * just make use of DEFAULT_INEQ_SEL.
2998  */
2999  end_offset_value =
3000  partition_tuples / peer_tuples * DEFAULT_INEQ_SEL;
3001  }
3002 
3003  if (frameOptions & FRAMEOPTION_ROWS)
3004  {
3005  /* include the N FOLLOWING and the current row */
3006  return_tuples = end_offset_value + 1.0;
3007  }
3008  else if (frameOptions & (FRAMEOPTION_RANGE | FRAMEOPTION_GROUPS))
3009  {
3010  /* include N FOLLOWING ranges/group and the initial range/group */
3011  return_tuples = peer_tuples * (end_offset_value + 1.0);
3012  }
3013  else
3014  {
3015  /*
3016  * Something new we don't support yet? This needs attention.
3017  * We'll just return 1.0 in the meantime.
3018  */
3019  Assert(false);
3020  return_tuples = 1.0;
3021  }
3022  }
3023  else
3024  {
3025  /*
3026  * Something new we don't support yet? This needs attention. We'll
3027  * just return 1.0 in the meantime.
3028  */
3029  Assert(false);
3030  return_tuples = 1.0;
3031  }
3032 
3033  if (wc->partitionClause != NIL || wc->orderClause != NIL)
3034  {
3035  /*
3036  * Cap the return value to the estimated partition tuples and account
3037  * for the extra tuple WindowAgg will need to read to confirm the next
3038  * tuple does not belong to the same partition or peer group.
3039  */
3040  return_tuples = Min(return_tuples + 1.0, partition_tuples);
3041  }
3042  else
3043  {
3044  /*
3045  * Cap the return value so it's never higher than the expected tuples
3046  * in the partition.
3047  */
3048  return_tuples = Min(return_tuples, partition_tuples);
3049  }
3050 
3051  /*
3052  * We needn't worry about any EXCLUDE options as those only exclude rows
3053  * from being aggregated, not from being read from the WindowAgg's
3054  * subnode.
3055  */
3056 
3057  return clamp_row_est(return_tuples);
3058 }
3059 
3060 /*
3061  * cost_windowagg
3062  * Determines and returns the cost of performing a WindowAgg plan node,
3063  * including the cost of its input.
3064  *
3065  * Input is assumed already properly sorted.
3066  */
3067 void
3069  List *windowFuncs, WindowClause *winclause,
3070  Cost input_startup_cost, Cost input_total_cost,
3071  double input_tuples)
3072 {
3073  Cost startup_cost;
3074  Cost total_cost;
3075  double startup_tuples;
3076  int numPartCols;
3077  int numOrderCols;
3078  ListCell *lc;
3079 
3080  numPartCols = list_length(winclause->partitionClause);
3081  numOrderCols = list_length(winclause->orderClause);
3082 
3083  startup_cost = input_startup_cost;
3084  total_cost = input_total_cost;
3085 
3086  /*
3087  * Window functions are assumed to cost their stated execution cost, plus
3088  * the cost of evaluating their input expressions, per tuple. Since they
3089  * may in fact evaluate their inputs at multiple rows during each cycle,
3090  * this could be a drastic underestimate; but without a way to know how
3091  * many rows the window function will fetch, it's hard to do better. In
3092  * any case, it's a good estimate for all the built-in window functions,
3093  * so we'll just do this for now.
3094  */
3095  foreach(lc, windowFuncs)
3096  {
3097  WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
3098  Cost wfunccost;
3099  QualCost argcosts;
3100 
3101  argcosts.startup = argcosts.per_tuple = 0;
3102  add_function_cost(root, wfunc->winfnoid, (Node *) wfunc,
3103  &argcosts);
3104  startup_cost += argcosts.startup;
3105  wfunccost = argcosts.per_tuple;
3106 
3107  /* also add the input expressions' cost to per-input-row costs */
3108  cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
3109  startup_cost += argcosts.startup;
3110  wfunccost += argcosts.per_tuple;
3111 
3112  /*
3113  * Add the filter's cost to per-input-row costs. XXX We should reduce
3114  * input expression costs according to filter selectivity.
3115  */
3116  cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
3117  startup_cost += argcosts.startup;
3118  wfunccost += argcosts.per_tuple;
3119 
3120  total_cost += wfunccost * input_tuples;
3121  }
3122 
3123  /*
3124  * We also charge cpu_operator_cost per grouping column per tuple for
3125  * grouping comparisons, plus cpu_tuple_cost per tuple for general
3126  * overhead.
3127  *
3128  * XXX this neglects costs of spooling the data to disk when it overflows
3129  * work_mem. Sooner or later that should get accounted for.
3130  */
3131  total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
3132  total_cost += cpu_tuple_cost * input_tuples;
3133 
3134  path->rows = input_tuples;
3135  path->startup_cost = startup_cost;
3136  path->total_cost = total_cost;
3137 
3138  /*
3139  * Also, take into account how many tuples we need to read from the
3140  * subnode in order to produce the first tuple from the WindowAgg. To do
3141  * this we proportion the run cost (total cost not including startup cost)
3142  * over the estimated startup tuples. We already included the startup
3143  * cost of the subnode, so we only need to do this when the estimated
3144  * startup tuples is above 1.0.
3145  */
3146  startup_tuples = get_windowclause_startup_tuples(root, winclause,
3147  input_tuples);
3148 
3149  if (startup_tuples > 1.0)
3150  path->startup_cost += (total_cost - startup_cost) / input_tuples *
3151  (startup_tuples - 1.0);
3152 }
3153 
3154 /*
3155  * cost_group
3156  * Determines and returns the cost of performing a Group plan node,
3157  * including the cost of its input.
3158  *
3159  * Note: caller must ensure that input costs are for appropriately-sorted
3160  * input.
3161  */
3162 void
3164  int numGroupCols, double numGroups,
3165  List *quals,
3166  Cost input_startup_cost, Cost input_total_cost,
3167  double input_tuples)
3168 {
3169  double output_tuples;
3170  Cost startup_cost;
3171  Cost total_cost;
3172 
3173  output_tuples = numGroups;
3174  startup_cost = input_startup_cost;
3175  total_cost = input_total_cost;
3176 
3177  /*
3178  * Charge one cpu_operator_cost per comparison per input tuple. We assume
3179  * all columns get compared at most of the tuples.
3180  */
3181  total_cost += cpu_operator_cost * input_tuples * numGroupCols;
3182 
3183  /*
3184  * If there are quals (HAVING quals), account for their cost and
3185  * selectivity.
3186  */
3187  if (quals)
3188  {
3189  QualCost qual_cost;
3190 
3191  cost_qual_eval(&qual_cost, quals, root);
3192  startup_cost += qual_cost.startup;
3193  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
3194 
3195  output_tuples = clamp_row_est(output_tuples *
3197  quals,
3198  0,
3199  JOIN_INNER,
3200  NULL));
3201  }
3202 
3203  path->rows = output_tuples;
3204  path->startup_cost = startup_cost;
3205  path->total_cost = total_cost;
3206 }
3207 
3208 /*
3209  * initial_cost_nestloop
3210  * Preliminary estimate of the cost of a nestloop join path.
3211  *
3212  * This must quickly produce lower-bound estimates of the path's startup and
3213  * total costs. If we are unable to eliminate the proposed path from
3214  * consideration using the lower bounds, final_cost_nestloop will be called
3215  * to obtain the final estimates.
3216  *
3217  * The exact division of labor between this function and final_cost_nestloop
3218  * is private to them, and represents a tradeoff between speed of the initial
3219  * estimate and getting a tight lower bound. We choose to not examine the
3220  * join quals here, since that's by far the most expensive part of the
3221  * calculations. The end result is that CPU-cost considerations must be
3222  * left for the second phase; and for SEMI/ANTI joins, we must also postpone
3223  * incorporation of the inner path's run cost.
3224  *
3225  * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3226  * other data to be used by final_cost_nestloop
3227  * 'jointype' is the type of join to be performed
3228  * 'outer_path' is the outer input to the join
3229  * 'inner_path' is the inner input to the join
3230  * 'extra' contains miscellaneous information about the join
3231  */
3232 void
3234  JoinType jointype,
3235  Path *outer_path, Path *inner_path,
3236  JoinPathExtraData *extra)
3237 {
3238  Cost startup_cost = 0;
3239  Cost run_cost = 0;
3240  double outer_path_rows = outer_path->rows;
3241  Cost inner_rescan_start_cost;
3242  Cost inner_rescan_total_cost;
3243  Cost inner_run_cost;
3244  Cost inner_rescan_run_cost;
3245 
3246  /* estimate costs to rescan the inner relation */
3247  cost_rescan(root, inner_path,
3248  &inner_rescan_start_cost,
3249  &inner_rescan_total_cost);
3250 
3251  /* cost of source data */
3252 
3253  /*
3254  * NOTE: clearly, we must pay both outer and inner paths' startup_cost
3255  * before we can start returning tuples, so the join's startup cost is
3256  * their sum. We'll also pay the inner path's rescan startup cost
3257  * multiple times.
3258  */
3259  startup_cost += outer_path->startup_cost + inner_path->startup_cost;
3260  run_cost += outer_path->total_cost - outer_path->startup_cost;
3261  if (outer_path_rows > 1)
3262  run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
3263 
3264  inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
3265  inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
3266 
3267  if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
3268  extra->inner_unique)
3269  {
3270  /*
3271  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3272  * executor will stop after the first match.
3273  *
3274  * Getting decent estimates requires inspection of the join quals,
3275  * which we choose to postpone to final_cost_nestloop.
3276  */
3277 
3278  /* Save private data for final_cost_nestloop */
3279  workspace->inner_run_cost = inner_run_cost;
3280  workspace->inner_rescan_run_cost = inner_rescan_run_cost;
3281  }
3282  else
3283  {
3284  /* Normal case; we'll scan whole input rel for each outer row */
3285  run_cost += inner_run_cost;
3286  if (outer_path_rows > 1)
3287  run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
3288  }
3289 
3290  /* CPU costs left for later */
3291 
3292  /* Public result fields */
3293  workspace->startup_cost = startup_cost;
3294  workspace->total_cost = startup_cost + run_cost;
3295  /* Save private data for final_cost_nestloop */
3296  workspace->run_cost = run_cost;
3297 }
3298 
3299 /*
3300  * final_cost_nestloop
3301  * Final estimate of the cost and result size of a nestloop join path.
3302  *
3303  * 'path' is already filled in except for the rows and cost fields
3304  * 'workspace' is the result from initial_cost_nestloop
3305  * 'extra' contains miscellaneous information about the join
3306  */
3307 void
3309  JoinCostWorkspace *workspace,
3310  JoinPathExtraData *extra)
3311 {
3312  Path *outer_path = path->jpath.outerjoinpath;
3313  Path *inner_path = path->jpath.innerjoinpath;
3314  double outer_path_rows = outer_path->rows;
3315  double inner_path_rows = inner_path->rows;
3316  Cost startup_cost = workspace->startup_cost;
3317  Cost run_cost = workspace->run_cost;
3318  Cost cpu_per_tuple;
3319  QualCost restrict_qual_cost;
3320  double ntuples;
3321 
3322  /* Protect some assumptions below that rowcounts aren't zero */
3323  if (outer_path_rows <= 0)
3324  outer_path_rows = 1;
3325  if (inner_path_rows <= 0)
3326  inner_path_rows = 1;
3327  /* Mark the path with the correct row estimate */
3328  if (path->jpath.path.param_info)
3329  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3330  else
3331  path->jpath.path.rows = path->jpath.path.parent->rows;
3332 
3333  /* For partial paths, scale row estimate. */
3334  if (path->jpath.path.parallel_workers > 0)
3335  {
3336  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3337 
3338  path->jpath.path.rows =
3339  clamp_row_est(path->jpath.path.rows / parallel_divisor);
3340  }
3341 
3342  /*
3343  * We could include disable_cost in the preliminary estimate, but that
3344  * would amount to optimizing for the case where the join method is
3345  * disabled, which doesn't seem like the way to bet.
3346  */
3347  if (!enable_nestloop)
3348  startup_cost += disable_cost;
3349 
3350  /* cost of inner-relation source data (we already dealt with outer rel) */
3351 
3352  if (path->jpath.jointype == JOIN_SEMI || path->jpath.jointype == JOIN_ANTI ||
3353  extra->inner_unique)
3354  {
3355  /*
3356  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3357  * executor will stop after the first match.
3358  */
3359  Cost inner_run_cost = workspace->inner_run_cost;
3360  Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
3361  double outer_matched_rows;
3362  double outer_unmatched_rows;
3363  Selectivity inner_scan_frac;
3364 
3365  /*
3366  * For an outer-rel row that has at least one match, we can expect the
3367  * inner scan to stop after a fraction 1/(match_count+1) of the inner
3368  * rows, if the matches are evenly distributed. Since they probably
3369  * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
3370  * that fraction. (If we used a larger fuzz factor, we'd have to
3371  * clamp inner_scan_frac to at most 1.0; but since match_count is at
3372  * least 1, no such clamp is needed now.)
3373  */
3374  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3375  outer_unmatched_rows = outer_path_rows - outer_matched_rows;
3376  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3377 
3378  /*
3379  * Compute number of tuples processed (not number emitted!). First,
3380  * account for successfully-matched outer rows.
3381  */
3382  ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
3383 
3384  /*
3385  * Now we need to estimate the actual costs of scanning the inner
3386  * relation, which may be quite a bit less than N times inner_run_cost
3387  * due to early scan stops. We consider two cases. If the inner path
3388  * is an indexscan using all the joinquals as indexquals, then an
3389  * unmatched outer row results in an indexscan returning no rows,
3390  * which is probably quite cheap. Otherwise, the executor will have
3391  * to scan the whole inner rel for an unmatched row; not so cheap.
3392  */
3393  if (has_indexed_join_quals(path))
3394  {
3395  /*
3396  * Successfully-matched outer rows will only require scanning
3397  * inner_scan_frac of the inner relation. In this case, we don't
3398  * need to charge the full inner_run_cost even when that's more
3399  * than inner_rescan_run_cost, because we can assume that none of
3400  * the inner scans ever scan the whole inner relation. So it's
3401  * okay to assume that all the inner scan executions can be
3402  * fractions of the full cost, even if materialization is reducing
3403  * the rescan cost. At this writing, it's impossible to get here
3404  * for a materialized inner scan, so inner_run_cost and
3405  * inner_rescan_run_cost will be the same anyway; but just in
3406  * case, use inner_run_cost for the first matched tuple and
3407  * inner_rescan_run_cost for additional ones.
3408  */
3409  run_cost += inner_run_cost * inner_scan_frac;
3410  if (outer_matched_rows > 1)
3411  run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
3412 
3413  /*
3414  * Add the cost of inner-scan executions for unmatched outer rows.
3415  * We estimate this as the same cost as returning the first tuple
3416  * of a nonempty scan. We consider that these are all rescans,
3417  * since we used inner_run_cost once already.
3418  */
3419  run_cost += outer_unmatched_rows *
3420  inner_rescan_run_cost / inner_path_rows;
3421 
3422  /*
3423  * We won't be evaluating any quals at all for unmatched rows, so
3424  * don't add them to ntuples.
3425  */
3426  }
3427  else
3428  {
3429  /*
3430  * Here, a complicating factor is that rescans may be cheaper than
3431  * first scans. If we never scan all the way to the end of the
3432  * inner rel, it might be (depending on the plan type) that we'd
3433  * never pay the whole inner first-scan run cost. However it is
3434  * difficult to estimate whether that will happen (and it could
3435  * not happen if there are any unmatched outer rows!), so be
3436  * conservative and always charge the whole first-scan cost once.
3437  * We consider this charge to correspond to the first unmatched
3438  * outer row, unless there isn't one in our estimate, in which
3439  * case blame it on the first matched row.
3440  */
3441 
3442  /* First, count all unmatched join tuples as being processed */
3443  ntuples += outer_unmatched_rows * inner_path_rows;
3444 
3445  /* Now add the forced full scan, and decrement appropriate count */
3446  run_cost += inner_run_cost;
3447  if (outer_unmatched_rows >= 1)
3448  outer_unmatched_rows -= 1;
3449  else
3450  outer_matched_rows -= 1;
3451 
3452  /* Add inner run cost for additional outer tuples having matches */
3453  if (outer_matched_rows > 0)
3454  run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
3455 
3456  /* Add inner run cost for additional unmatched outer tuples */
3457  if (outer_unmatched_rows > 0)
3458  run_cost += outer_unmatched_rows * inner_rescan_run_cost;
3459  }
3460  }
3461  else
3462  {
3463  /* Normal-case source costs were included in preliminary estimate */
3464 
3465  /* Compute number of tuples processed (not number emitted!) */
3466  ntuples = outer_path_rows * inner_path_rows;
3467  }
3468 
3469  /* CPU costs */
3470  cost_qual_eval(&restrict_qual_cost, path->jpath.joinrestrictinfo, root);
3471  startup_cost += restrict_qual_cost.startup;
3472  cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
3473  run_cost += cpu_per_tuple * ntuples;
3474 
3475  /* tlist eval costs are paid per output row, not per tuple scanned */
3476  startup_cost += path->jpath.path.pathtarget->cost.startup;
3477  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3478 
3479  path->jpath.path.startup_cost = startup_cost;
3480  path->jpath.path.total_cost = startup_cost + run_cost;
3481 }
3482 
3483 /*
3484  * initial_cost_mergejoin
3485  * Preliminary estimate of the cost of a mergejoin path.
3486  *
3487  * This must quickly produce lower-bound estimates of the path's startup and
3488  * total costs. If we are unable to eliminate the proposed path from
3489  * consideration using the lower bounds, final_cost_mergejoin will be called
3490  * to obtain the final estimates.
3491  *
3492  * The exact division of labor between this function and final_cost_mergejoin
3493  * is private to them, and represents a tradeoff between speed of the initial
3494  * estimate and getting a tight lower bound. We choose to not examine the
3495  * join quals here, except for obtaining the scan selectivity estimate which
3496  * is really essential (but fortunately, use of caching keeps the cost of
3497  * getting that down to something reasonable).
3498  * We also assume that cost_sort is cheap enough to use here.
3499  *
3500  * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3501  * other data to be used by final_cost_mergejoin
3502  * 'jointype' is the type of join to be performed
3503  * 'mergeclauses' is the list of joinclauses to be used as merge clauses
3504  * 'outer_path' is the outer input to the join
3505  * 'inner_path' is the inner input to the join
3506  * 'outersortkeys' is the list of sort keys for the outer path
3507  * 'innersortkeys' is the list of sort keys for the inner path
3508  * 'extra' contains miscellaneous information about the join
3509  *
3510  * Note: outersortkeys and innersortkeys should be NIL if no explicit
3511  * sort is needed because the respective source path is already ordered.
3512  */
3513 void
3515  JoinType jointype,
3516  List *mergeclauses,
3517  Path *outer_path, Path *inner_path,
3518  List *outersortkeys, List *innersortkeys,
3519  JoinPathExtraData *extra)
3520 {
3521  Cost startup_cost = 0;
3522  Cost run_cost = 0;
3523  double outer_path_rows = outer_path->rows;
3524  double inner_path_rows = inner_path->rows;
3525  Cost inner_run_cost;
3526  double outer_rows,
3527  inner_rows,
3528  outer_skip_rows,
3529  inner_skip_rows;
3530  Selectivity outerstartsel,
3531  outerendsel,
3532  innerstartsel,
3533  innerendsel;
3534  Path sort_path; /* dummy for result of cost_sort */
3535 
3536  /* Protect some assumptions below that rowcounts aren't zero */
3537  if (outer_path_rows <= 0)
3538  outer_path_rows = 1;
3539  if (inner_path_rows <= 0)
3540  inner_path_rows = 1;
3541 
3542  /*
3543  * A merge join will stop as soon as it exhausts either input stream
3544  * (unless it's an outer join, in which case the outer side has to be
3545  * scanned all the way anyway). Estimate fraction of the left and right
3546  * inputs that will actually need to be scanned. Likewise, we can
3547  * estimate the number of rows that will be skipped before the first join
3548  * pair is found, which should be factored into startup cost. We use only
3549  * the first (most significant) merge clause for this purpose. Since
3550  * mergejoinscansel() is a fairly expensive computation, we cache the
3551  * results in the merge clause RestrictInfo.
3552  */
3553  if (mergeclauses && jointype != JOIN_FULL)
3554  {
3555  RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
3556  List *opathkeys;
3557  List *ipathkeys;
3558  PathKey *opathkey;
3559  PathKey *ipathkey;
3560  MergeScanSelCache *cache;
3561 
3562  /* Get the input pathkeys to determine the sort-order details */
3563  opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
3564  ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
3565  Assert(opathkeys);
3566  Assert(ipathkeys);
3567  opathkey = (PathKey *) linitial(opathkeys);
3568  ipathkey = (PathKey *) linitial(ipathkeys);
3569  /* debugging check */
3570  if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
3571  opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
3572  opathkey->pk_strategy != ipathkey->pk_strategy ||
3573  opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
3574  elog(ERROR, "left and right pathkeys do not match in mergejoin");
3575 
3576  /* Get the selectivity with caching */
3577  cache = cached_scansel(root, firstclause, opathkey);
3578 
3579  if (bms_is_subset(firstclause->left_relids,
3580  outer_path->parent->relids))
3581  {
3582  /* left side of clause is outer */
3583  outerstartsel = cache->leftstartsel;
3584  outerendsel = cache->leftendsel;
3585  innerstartsel = cache->rightstartsel;
3586  innerendsel = cache->rightendsel;
3587  }
3588  else
3589  {
3590  /* left side of clause is inner */
3591  outerstartsel = cache->rightstartsel;
3592  outerendsel = cache->rightendsel;
3593  innerstartsel = cache->leftstartsel;
3594  innerendsel = cache->leftendsel;
3595  }
3596  if (jointype == JOIN_LEFT ||
3597  jointype == JOIN_ANTI)
3598  {
3599  outerstartsel = 0.0;
3600  outerendsel = 1.0;
3601  }
3602  else if (jointype == JOIN_RIGHT ||
3603  jointype == JOIN_RIGHT_ANTI)
3604  {
3605  innerstartsel = 0.0;
3606  innerendsel = 1.0;
3607  }
3608  }
3609  else
3610  {
3611  /* cope with clauseless or full mergejoin */
3612  outerstartsel = innerstartsel = 0.0;
3613  outerendsel = innerendsel = 1.0;
3614  }
3615 
3616  /*
3617  * Convert selectivities to row counts. We force outer_rows and
3618  * inner_rows to be at least 1, but the skip_rows estimates can be zero.
3619  */
3620  outer_skip_rows = rint(outer_path_rows * outerstartsel);
3621  inner_skip_rows = rint(inner_path_rows * innerstartsel);
3622  outer_rows = clamp_row_est(outer_path_rows * outerendsel);
3623  inner_rows = clamp_row_est(inner_path_rows * innerendsel);
3624 
3625  Assert(outer_skip_rows <= outer_rows);
3626  Assert(inner_skip_rows <= inner_rows);
3627 
3628  /*
3629  * Readjust scan selectivities to account for above rounding. This is
3630  * normally an insignificant effect, but when there are only a few rows in
3631  * the inputs, failing to do this makes for a large percentage error.
3632  */
3633  outerstartsel = outer_skip_rows / outer_path_rows;
3634  innerstartsel = inner_skip_rows / inner_path_rows;
3635  outerendsel = outer_rows / outer_path_rows;
3636  innerendsel = inner_rows / inner_path_rows;
3637 
3638  Assert(outerstartsel <= outerendsel);
3639  Assert(innerstartsel <= innerendsel);
3640 
3641  /* cost of source data */
3642 
3643  if (outersortkeys) /* do we need to sort outer? */
3644  {
3645  cost_sort(&sort_path,
3646  root,
3647  outersortkeys,
3648  outer_path->total_cost,
3649  outer_path_rows,
3650  outer_path->pathtarget->width,
3651  0.0,
3652  work_mem,
3653  -1.0);
3654  startup_cost += sort_path.startup_cost;
3655  startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3656  * outerstartsel;
3657  run_cost += (sort_path.total_cost - sort_path.startup_cost)
3658  * (outerendsel - outerstartsel);
3659  }
3660  else
3661  {
3662  startup_cost += outer_path->startup_cost;
3663  startup_cost += (outer_path->total_cost - outer_path->startup_cost)
3664  * outerstartsel;
3665  run_cost += (outer_path->total_cost - outer_path->startup_cost)
3666  * (outerendsel - outerstartsel);
3667  }
3668 
3669  if (innersortkeys) /* do we need to sort inner? */
3670  {
3671  cost_sort(&sort_path,
3672  root,
3673  innersortkeys,
3674  inner_path->total_cost,
3675  inner_path_rows,
3676  inner_path->pathtarget->width,
3677  0.0,
3678  work_mem,
3679  -1.0);
3680  startup_cost += sort_path.startup_cost;
3681  startup_cost += (sort_path.total_cost - sort_path.startup_cost)
3682  * innerstartsel;
3683  inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
3684  * (innerendsel - innerstartsel);
3685  }
3686  else
3687  {
3688  startup_cost += inner_path->startup_cost;
3689  startup_cost += (inner_path->total_cost - inner_path->startup_cost)
3690  * innerstartsel;
3691  inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
3692  * (innerendsel - innerstartsel);
3693  }
3694 
3695  /*
3696  * We can't yet determine whether rescanning occurs, or whether
3697  * materialization of the inner input should be done. The minimum
3698  * possible inner input cost, regardless of rescan and materialization
3699  * considerations, is inner_run_cost. We include that in
3700  * workspace->total_cost, but not yet in run_cost.
3701  */
3702 
3703  /* CPU costs left for later */
3704 
3705  /* Public result fields */
3706  workspace->startup_cost = startup_cost;
3707  workspace->total_cost = startup_cost + run_cost + inner_run_cost;
3708  /* Save private data for final_cost_mergejoin */
3709  workspace->run_cost = run_cost;
3710  workspace->inner_run_cost = inner_run_cost;
3711  workspace->outer_rows = outer_rows;
3712  workspace->inner_rows = inner_rows;
3713  workspace->outer_skip_rows = outer_skip_rows;
3714  workspace->inner_skip_rows = inner_skip_rows;
3715 }
3716 
3717 /*
3718  * final_cost_mergejoin
3719  * Final estimate of the cost and result size of a mergejoin path.
3720  *
3721  * Unlike other costsize functions, this routine makes two actual decisions:
3722  * whether the executor will need to do mark/restore, and whether we should
3723  * materialize the inner path. It would be logically cleaner to build
3724  * separate paths testing these alternatives, but that would require repeating
3725  * most of the cost calculations, which are not all that cheap. Since the
3726  * choice will not affect output pathkeys or startup cost, only total cost,
3727  * there is no possibility of wanting to keep more than one path. So it seems
3728  * best to make the decisions here and record them in the path's
3729  * skip_mark_restore and materialize_inner fields.
3730  *
3731  * Mark/restore overhead is usually required, but can be skipped if we know
3732  * that the executor need find only one match per outer tuple, and that the
3733  * mergeclauses are sufficient to identify a match.
3734  *
3735  * We materialize the inner path if we need mark/restore and either the inner
3736  * path can't support mark/restore, or it's cheaper to use an interposed
3737  * Material node to handle mark/restore.
3738  *
3739  * 'path' is already filled in except for the rows and cost fields and
3740  * skip_mark_restore and materialize_inner
3741  * 'workspace' is the result from initial_cost_mergejoin
3742  * 'extra' contains miscellaneous information about the join
3743  */
3744 void
3746  JoinCostWorkspace *workspace,
3747  JoinPathExtraData *extra)
3748 {
3749  Path *outer_path = path->jpath.outerjoinpath;
3750  Path *inner_path = path->jpath.innerjoinpath;
3751  double inner_path_rows = inner_path->rows;
3752  List *mergeclauses = path->path_mergeclauses;
3753  List *innersortkeys = path->innersortkeys;
3754  Cost startup_cost = workspace->startup_cost;
3755  Cost run_cost = workspace->run_cost;
3756  Cost inner_run_cost = workspace->inner_run_cost;
3757  double outer_rows = workspace->outer_rows;
3758  double inner_rows = workspace->inner_rows;
3759  double outer_skip_rows = workspace->outer_skip_rows;
3760  double inner_skip_rows = workspace->inner_skip_rows;
3761  Cost cpu_per_tuple,
3762  bare_inner_cost,
3763  mat_inner_cost;
3764  QualCost merge_qual_cost;
3765  QualCost qp_qual_cost;
3766  double mergejointuples,
3767  rescannedtuples;
3768  double rescanratio;
3769 
3770  /* Protect some assumptions below that rowcounts aren't zero */
3771  if (inner_path_rows <= 0)
3772  inner_path_rows = 1;
3773 
3774  /* Mark the path with the correct row estimate */
3775  if (path->jpath.path.param_info)
3776  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3777  else
3778  path->jpath.path.rows = path->jpath.path.parent->rows;
3779 
3780  /* For partial paths, scale row estimate. */
3781  if (path->jpath.path.parallel_workers > 0)
3782  {
3783  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3784 
3785  path->jpath.path.rows =
3786  clamp_row_est(path->jpath.path.rows / parallel_divisor);
3787  }
3788 
3789  /*
3790  * We could include disable_cost in the preliminary estimate, but that
3791  * would amount to optimizing for the case where the join method is
3792  * disabled, which doesn't seem like the way to bet.
3793  */
3794  if (!enable_mergejoin)
3795  startup_cost += disable_cost;
3796 
3797  /*
3798  * Compute cost of the mergequals and qpquals (other restriction clauses)
3799  * separately.
3800  */
3801  cost_qual_eval(&merge_qual_cost, mergeclauses, root);
3802  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3803  qp_qual_cost.startup -= merge_qual_cost.startup;
3804  qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
3805 
3806  /*
3807  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3808  * executor will stop scanning for matches after the first match. When
3809  * all the joinclauses are merge clauses, this means we don't ever need to
3810  * back up the merge, and so we can skip mark/restore overhead.
3811  */
3812  if ((path->jpath.jointype == JOIN_SEMI ||
3813  path->jpath.jointype == JOIN_ANTI ||
3814  extra->inner_unique) &&
3817  path->skip_mark_restore = true;
3818  else
3819  path->skip_mark_restore = false;
3820 
3821  /*
3822  * Get approx # tuples passing the mergequals. We use approx_tuple_count
3823  * here because we need an estimate done with JOIN_INNER semantics.
3824  */
3825  mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
3826 
3827  /*
3828  * When there are equal merge keys in the outer relation, the mergejoin
3829  * must rescan any matching tuples in the inner relation. This means
3830  * re-fetching inner tuples; we have to estimate how often that happens.
3831  *
3832  * For regular inner and outer joins, the number of re-fetches can be
3833  * estimated approximately as size of merge join output minus size of
3834  * inner relation. Assume that the distinct key values are 1, 2, ..., and
3835  * denote the number of values of each key in the outer relation as m1,
3836  * m2, ...; in the inner relation, n1, n2, ... Then we have
3837  *
3838  * size of join = m1 * n1 + m2 * n2 + ...
3839  *
3840  * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
3841  * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
3842  * relation
3843  *
3844  * This equation works correctly for outer tuples having no inner match
3845  * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
3846  * are effectively subtracting those from the number of rescanned tuples,
3847  * when we should not. Can we do better without expensive selectivity
3848  * computations?
3849  *
3850  * The whole issue is moot if we are working from a unique-ified outer
3851  * input, or if we know we don't need to mark/restore at all.
3852  */
3853  if (IsA(outer_path, UniquePath) || path->skip_mark_restore)
3854  rescannedtuples = 0;
3855  else
3856  {
3857  rescannedtuples = mergejointuples - inner_path_rows;
3858  /* Must clamp because of possible underestimate */
3859  if (rescannedtuples < 0)
3860  rescannedtuples = 0;
3861  }
3862 
3863  /*
3864  * We'll inflate various costs this much to account for rescanning. Note
3865  * that this is to be multiplied by something involving inner_rows, or
3866  * another number related to the portion of the inner rel we'll scan.
3867  */
3868  rescanratio = 1.0 + (rescannedtuples / inner_rows);
3869 
3870  /*
3871  * Decide whether we want to materialize the inner input to shield it from
3872  * mark/restore and performing re-fetches. Our cost model for regular
3873  * re-fetches is that a re-fetch costs the same as an original fetch,
3874  * which is probably an overestimate; but on the other hand we ignore the
3875  * bookkeeping costs of mark/restore. Not clear if it's worth developing
3876  * a more refined model. So we just need to inflate the inner run cost by
3877  * rescanratio.
3878  */
3879  bare_inner_cost = inner_run_cost * rescanratio;
3880 
3881  /*
3882  * When we interpose a Material node the re-fetch cost is assumed to be
3883  * just cpu_operator_cost per tuple, independently of the underlying
3884  * plan's cost; and we charge an extra cpu_operator_cost per original
3885  * fetch as well. Note that we're assuming the materialize node will
3886  * never spill to disk, since it only has to remember tuples back to the
3887  * last mark. (If there are a huge number of duplicates, our other cost
3888  * factors will make the path so expensive that it probably won't get
3889  * chosen anyway.) So we don't use cost_rescan here.
3890  *
3891  * Note: keep this estimate in sync with create_mergejoin_plan's labeling
3892  * of the generated Material node.
3893  */
3894  mat_inner_cost = inner_run_cost +
3895  cpu_operator_cost * inner_rows * rescanratio;
3896 
3897  /*
3898  * If we don't need mark/restore at all, we don't need materialization.
3899  */
3900  if (path->skip_mark_restore)
3901  path->materialize_inner = false;
3902 
3903  /*
3904  * Prefer materializing if it looks cheaper, unless the user has asked to
3905  * suppress materialization.
3906  */
3907  else if (enable_material && mat_inner_cost < bare_inner_cost)
3908  path->materialize_inner = true;
3909 
3910  /*
3911  * Even if materializing doesn't look cheaper, we *must* do it if the
3912  * inner path is to be used directly (without sorting) and it doesn't
3913  * support mark/restore.
3914  *
3915  * Since the inner side must be ordered, and only Sorts and IndexScans can
3916  * create order to begin with, and they both support mark/restore, you
3917  * might think there's no problem --- but you'd be wrong. Nestloop and
3918  * merge joins can *preserve* the order of their inputs, so they can be
3919  * selected as the input of a mergejoin, and they don't support
3920  * mark/restore at present.
3921  *
3922  * We don't test the value of enable_material here, because
3923  * materialization is required for correctness in this case, and turning
3924  * it off does not entitle us to deliver an invalid plan.
3925  */
3926  else if (innersortkeys == NIL &&
3927  !ExecSupportsMarkRestore(inner_path))
3928  path->materialize_inner = true;
3929 
3930  /*
3931  * Also, force materializing if the inner path is to be sorted and the
3932  * sort is expected to spill to disk. This is because the final merge
3933  * pass can be done on-the-fly if it doesn't have to support mark/restore.
3934  * We don't try to adjust the cost estimates for this consideration,
3935  * though.
3936  *
3937  * Since materialization is a performance optimization in this case,
3938  * rather than necessary for correctness, we skip it if enable_material is
3939  * off.
3940  */
3941  else if (enable_material && innersortkeys != NIL &&
3942  relation_byte_size(inner_path_rows,
3943  inner_path->pathtarget->width) >
3944  (work_mem * 1024L))
3945  path->materialize_inner = true;
3946  else
3947  path->materialize_inner = false;
3948 
3949  /* Charge the right incremental cost for the chosen case */
3950  if (path->materialize_inner)
3951  run_cost += mat_inner_cost;
3952  else
3953  run_cost += bare_inner_cost;
3954 
3955  /* CPU costs */
3956 
3957  /*
3958  * The number of tuple comparisons needed is approximately number of outer
3959  * rows plus number of inner rows plus number of rescanned tuples (can we
3960  * refine this?). At each one, we need to evaluate the mergejoin quals.
3961  */
3962  startup_cost += merge_qual_cost.startup;
3963  startup_cost += merge_qual_cost.per_tuple *
3964  (outer_skip_rows + inner_skip_rows * rescanratio);
3965  run_cost += merge_qual_cost.per_tuple *
3966  ((outer_rows - outer_skip_rows) +
3967  (inner_rows - inner_skip_rows) * rescanratio);
3968 
3969  /*
3970  * For each tuple that gets through the mergejoin proper, we charge
3971  * cpu_tuple_cost plus the cost of evaluating additional restriction
3972  * clauses that are to be applied at the join. (This is pessimistic since
3973  * not all of the quals may get evaluated at each tuple.)
3974  *
3975  * Note: we could adjust for SEMI/ANTI joins skipping some qual
3976  * evaluations here, but it's probably not worth the trouble.
3977  */
3978  startup_cost += qp_qual_cost.startup;
3979  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
3980  run_cost += cpu_per_tuple * mergejointuples;
3981 
3982  /* tlist eval costs are paid per output row, not per tuple scanned */
3983  startup_cost += path->jpath.path.pathtarget->cost.startup;
3984  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3985 
3986  path->jpath.path.startup_cost = startup_cost;
3987  path->jpath.path.total_cost = startup_cost + run_cost;
3988 }
3989 
3990 /*
3991  * run mergejoinscansel() with caching
3992  */
3993 static MergeScanSelCache *
3995 {
3996  MergeScanSelCache *cache;
3997  ListCell *lc;
3998  Selectivity leftstartsel,
3999  leftendsel,
4000  rightstartsel,
4001  rightendsel;
4002  MemoryContext oldcontext;
4003 
4004  /* Do we have this result already? */
4005  foreach(lc, rinfo->scansel_cache)
4006  {
4007  cache = (MergeScanSelCache *) lfirst(lc);
4008  if (cache->opfamily == pathkey->pk_opfamily &&
4009  cache->collation == pathkey->pk_eclass->ec_collation &&
4010  cache->strategy == pathkey->pk_strategy &&
4011  cache->nulls_first == pathkey->pk_nulls_first)
4012  return cache;
4013  }
4014 
4015  /* Nope, do the computation */
4017  (Node *) rinfo->clause,
4018  pathkey->pk_opfamily,
4019  pathkey->pk_strategy,
4020  pathkey->pk_nulls_first,
4021  &leftstartsel,
4022  &leftendsel,
4023  &rightstartsel,
4024  &rightendsel);
4025 
4026  /* Cache the result in suitably long-lived workspace */
4027  oldcontext = MemoryContextSwitchTo(root->planner_cxt);
4028 
4029  cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
4030  cache->opfamily = pathkey->pk_opfamily;
4031  cache->collation = pathkey->pk_eclass->ec_collation;
4032  cache->strategy = pathkey->pk_strategy;
4033  cache->nulls_first = pathkey->pk_nulls_first;
4034  cache->leftstartsel = leftstartsel;
4035  cache->leftendsel = leftendsel;
4036  cache->rightstartsel = rightstartsel;
4037  cache->rightendsel = rightendsel;
4038 
4039  rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
4040 
4041  MemoryContextSwitchTo(oldcontext);
4042 
4043  return cache;
4044 }
4045 
4046 /*
4047  * initial_cost_hashjoin
4048  * Preliminary estimate of the cost of a hashjoin path.
4049  *
4050  * This must quickly produce lower-bound estimates of the path's startup and
4051  * total costs. If we are unable to eliminate the proposed path from
4052  * consideration using the lower bounds, final_cost_hashjoin will be called
4053  * to obtain the final estimates.
4054  *
4055  * The exact division of labor between this function and final_cost_hashjoin
4056  * is private to them, and represents a tradeoff between speed of the initial
4057  * estimate and getting a tight lower bound. We choose to not examine the
4058  * join quals here (other than by counting the number of hash clauses),
4059  * so we can't do much with CPU costs. We do assume that
4060  * ExecChooseHashTableSize is cheap enough to use here.
4061  *
4062  * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
4063  * other data to be used by final_cost_hashjoin
4064  * 'jointype' is the type of join to be performed
4065  * 'hashclauses' is the list of joinclauses to be used as hash clauses
4066  * 'outer_path' is the outer input to the join
4067  * 'inner_path' is the inner input to the join
4068  * 'extra' contains miscellaneous information about the join
4069  * 'parallel_hash' indicates that inner_path is partial and that a shared
4070  * hash table will be built in parallel
4071  */
4072 void
4074  JoinType jointype,
4075  List *hashclauses,
4076  Path *outer_path, Path *inner_path,
4077  JoinPathExtraData *extra,
4078  bool parallel_hash)
4079 {
4080  Cost startup_cost = 0;
4081  Cost run_cost = 0;
4082  double outer_path_rows = outer_path->rows;
4083  double inner_path_rows = inner_path->rows;
4084  double inner_path_rows_total = inner_path_rows;
4085  int num_hashclauses = list_length(hashclauses);
4086  int numbuckets;
4087  int numbatches;
4088  int num_skew_mcvs;
4089  size_t space_allowed; /* unused */
4090 
4091  /* cost of source data */
4092  startup_cost += outer_path->startup_cost;
4093  run_cost += outer_path->total_cost - outer_path->startup_cost;
4094  startup_cost += inner_path->total_cost;
4095 
4096  /*
4097  * Cost of computing hash function: must do it once per input tuple. We
4098  * charge one cpu_operator_cost for each column's hash function. Also,
4099  * tack on one cpu_tuple_cost per inner row, to model the costs of
4100  * inserting the row into the hashtable.
4101  *
4102  * XXX when a hashclause is more complex than a single operator, we really
4103  * should charge the extra eval costs of the left or right side, as
4104  * appropriate, here. This seems more work than it's worth at the moment.
4105  */
4106  startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
4107  * inner_path_rows;
4108  run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
4109 
4110  /*
4111  * If this is a parallel hash build, then the value we have for
4112  * inner_rows_total currently refers only to the rows returned by each
4113  * participant. For shared hash table size estimation, we need the total
4114  * number, so we need to undo the division.
4115  */
4116  if (parallel_hash)
4117  inner_path_rows_total *= get_parallel_divisor(inner_path);
4118 
4119  /*
4120  * Get hash table size that executor would use for inner relation.
4121  *
4122  * XXX for the moment, always assume that skew optimization will be
4123  * performed. As long as SKEW_HASH_MEM_PERCENT is small, it's not worth
4124  * trying to determine that for sure.
4125  *
4126  * XXX at some point it might be interesting to try to account for skew
4127  * optimization in the cost estimate, but for now, we don't.
4128  */
4129  ExecChooseHashTableSize(inner_path_rows_total,
4130  inner_path->pathtarget->width,
4131  true, /* useskew */
4132  parallel_hash, /* try_combined_hash_mem */
4133  outer_path->parallel_workers,
4134  &space_allowed,
4135  &numbuckets,
4136  &numbatches,
4137  &num_skew_mcvs);
4138 
4139  /*
4140  * If inner relation is too big then we will need to "batch" the join,
4141  * which implies writing and reading most of the tuples to disk an extra
4142  * time. Charge seq_page_cost per page, since the I/O should be nice and
4143  * sequential. Writing the inner rel counts as startup cost, all the rest
4144  * as run cost.
4145  */
4146  if (numbatches > 1)
4147  {
4148  double outerpages = page_size(outer_path_rows,
4149  outer_path->pathtarget->width);
4150  double innerpages = page_size(inner_path_rows,
4151  inner_path->pathtarget->width);
4152 
4153  startup_cost += seq_page_cost * innerpages;
4154  run_cost += seq_page_cost * (innerpages + 2 * outerpages);
4155  }
4156 
4157  /* CPU costs left for later */
4158 
4159  /* Public result fields */
4160  workspace->startup_cost = startup_cost;
4161  workspace->total_cost = startup_cost + run_cost;
4162  /* Save private data for final_cost_hashjoin */
4163  workspace->run_cost = run_cost;
4164  workspace->numbuckets = numbuckets;
4165  workspace->numbatches = numbatches;
4166  workspace->inner_rows_total = inner_path_rows_total;
4167 }
4168 
4169 /*
4170  * final_cost_hashjoin
4171  * Final estimate of the cost and result size of a hashjoin path.
4172  *
4173  * Note: the numbatches estimate is also saved into 'path' for use later
4174  *
4175  * 'path' is already filled in except for the rows and cost fields and
4176  * num_batches
4177  * 'workspace' is the result from initial_cost_hashjoin
4178  * 'extra' contains miscellaneous information about the join
4179  */
4180 void
4182  JoinCostWorkspace *workspace,
4183  JoinPathExtraData *extra)
4184 {
4185  Path *outer_path = path->jpath.outerjoinpath;
4186  Path *inner_path = path->jpath.innerjoinpath;
4187  double outer_path_rows = outer_path->rows;
4188  double inner_path_rows = inner_path->rows;
4189  double inner_path_rows_total = workspace->inner_rows_total;
4190  List *hashclauses = path->path_hashclauses;
4191  Cost startup_cost = workspace->startup_cost;
4192  Cost run_cost = workspace->run_cost;
4193  int numbuckets = workspace->numbuckets;
4194  int numbatches = workspace->numbatches;
4195  Cost cpu_per_tuple;
4196  QualCost hash_qual_cost;
4197  QualCost qp_qual_cost;
4198  double hashjointuples;
4199  double virtualbuckets;
4200  Selectivity innerbucketsize;
4201  Selectivity innermcvfreq;
4202  ListCell *hcl;
4203 
4204  /* Mark the path with the correct row estimate */
4205  if (path->jpath.path.param_info)
4206  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
4207  else
4208  path->jpath.path.rows = path->jpath.path.parent->rows;
4209 
4210  /* For partial paths, scale row estimate. */
4211  if (path->jpath.path.parallel_workers > 0)
4212  {
4213  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
4214 
4215  path->jpath.path.rows =
4216  clamp_row_est(path->jpath.path.rows / parallel_divisor);
4217  }
4218 
4219  /*
4220  * We could include disable_cost in the preliminary estimate, but that
4221  * would amount to optimizing for the case where the join method is
4222  * disabled, which doesn't seem like the way to bet.
4223  */
4224  if (!enable_hashjoin)
4225  startup_cost += disable_cost;
4226 
4227  /* mark the path with estimated # of batches */
4228  path->num_batches = numbatches;
4229 
4230  /* store the total number of tuples (sum of partial row estimates) */
4231  path->inner_rows_total = inner_path_rows_total;
4232 
4233  /* and compute the number of "virtual" buckets in the whole join */
4234  virtualbuckets = (double) numbuckets * (double) numbatches;
4235 
4236  /*
4237  * Determine bucketsize fraction and MCV frequency for the inner relation.
4238  * We use the smallest bucketsize or MCV frequency estimated for any
4239  * individual hashclause; this is undoubtedly conservative.
4240  *
4241  * BUT: if inner relation has been unique-ified, we can assume it's good
4242  * for hashing. This is important both because it's the right answer, and
4243  * because we avoid contaminating the cache with a value that's wrong for
4244  * non-unique-ified paths.
4245  */
4246  if (IsA(inner_path, UniquePath))
4247  {
4248  innerbucketsize = 1.0 / virtualbuckets;
4249  innermcvfreq = 0.0;
4250  }
4251  else
4252  {
4253  innerbucketsize = 1.0;
4254  innermcvfreq = 1.0;
4255  foreach(hcl, hashclauses)
4256  {
4257  RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
4258  Selectivity thisbucketsize;
4259  Selectivity thismcvfreq;
4260 
4261  /*
4262  * First we have to figure out which side of the hashjoin clause
4263  * is the inner side.
4264  *
4265  * Since we tend to visit the same clauses over and over when
4266  * planning a large query, we cache the bucket stats estimates in
4267  * the RestrictInfo node to avoid repeated lookups of statistics.
4268  */
4269  if (bms_is_subset(restrictinfo->right_relids,
4270  inner_path->parent->relids))
4271  {
4272  /* righthand side is inner */
4273  thisbucketsize = restrictinfo->right_bucketsize;
4274  if (thisbucketsize < 0)
4275  {
4276  /* not cached yet */
4278  get_rightop(restrictinfo->clause),
4279  virtualbuckets,
4280  &restrictinfo->right_mcvfreq,
4281  &restrictinfo->right_bucketsize);
4282  thisbucketsize = restrictinfo->right_bucketsize;
4283  }
4284  thismcvfreq = restrictinfo->right_mcvfreq;
4285  }
4286  else
4287  {
4288  Assert(bms_is_subset(restrictinfo->left_relids,
4289  inner_path->parent->relids));
4290  /* lefthand side is inner */
4291  thisbucketsize = restrictinfo->left_bucketsize;
4292  if (thisbucketsize < 0)
4293  {
4294  /* not cached yet */
4296  get_leftop(restrictinfo->clause),
4297  virtualbuckets,
4298  &restrictinfo->left_mcvfreq,
4299  &restrictinfo->left_bucketsize);
4300  thisbucketsize = restrictinfo->left_bucketsize;
4301  }
4302  thismcvfreq = restrictinfo->left_mcvfreq;
4303  }
4304 
4305  if (innerbucketsize > thisbucketsize)
4306  innerbucketsize = thisbucketsize;
4307  if (innermcvfreq > thismcvfreq)
4308  innermcvfreq = thismcvfreq;
4309  }
4310  }
4311 
4312  /*
4313  * If the bucket holding the inner MCV would exceed hash_mem, we don't
4314  * want to hash unless there is really no other alternative, so apply
4315  * disable_cost. (The executor normally copes with excessive memory usage
4316  * by splitting batches, but obviously it cannot separate equal values
4317  * that way, so it will be unable to drive the batch size below hash_mem
4318  * when this is true.)
4319  */
4320  if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
4321  inner_path->pathtarget->width) > get_hash_memory_limit())
4322  startup_cost += disable_cost;
4323 
4324  /*
4325  * Compute cost of the hashquals and qpquals (other restriction clauses)
4326  * separately.
4327  */
4328  cost_qual_eval(&hash_qual_cost, hashclauses, root);
4329  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
4330  qp_qual_cost.startup -= hash_qual_cost.startup;
4331  qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
4332 
4333  /* CPU costs */
4334 
4335  if (path->jpath.jointype == JOIN_SEMI ||
4336  path->jpath.jointype == JOIN_ANTI ||
4337  extra->inner_unique)
4338  {
4339  double outer_matched_rows;
4340  Selectivity inner_scan_frac;
4341 
4342  /*
4343  * With a SEMI or ANTI join, or if the innerrel is known unique, the
4344  * executor will stop after the first match.
4345  *
4346  * For an outer-rel row that has at least one match, we can expect the
4347  * bucket scan to stop after a fraction 1/(match_count+1) of the
4348  * bucket's rows, if the matches are evenly distributed. Since they
4349  * probably aren't quite evenly distributed, we apply a fuzz factor of
4350  * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
4351  * to clamp inner_scan_frac to at most 1.0; but since match_count is
4352  * at least 1, no such clamp is needed now.)
4353  */
4354  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
4355  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
4356 
4357  startup_cost += hash_qual_cost.startup;
4358  run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
4359  clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
4360 
4361  /*
4362  * For unmatched outer-rel rows, the picture is quite a lot different.
4363  * In the first place, there is no reason to assume that these rows
4364  * preferentially hit heavily-populated buckets; instead assume they
4365  * are uncorrelated with the inner distribution and so they see an
4366  * average bucket size of inner_path_rows / virtualbuckets. In the
4367  * second place, it seems likely that they will have few if any exact
4368  * hash-code matches and so very few of the tuples in the bucket will
4369  * actually require eval of the hash quals. We don't have any good
4370  * way to estimate how many will, but for the moment assume that the
4371  * effective cost per bucket entry is one-tenth what it is for
4372  * matchable tuples.
4373  */
4374  run_cost += hash_qual_cost.per_tuple *
4375  (outer_path_rows - outer_matched_rows) *
4376  clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
4377 
4378  /* Get # of tuples that will pass the basic join */
4379  if (path->jpath.jointype == JOIN_ANTI)
4380  hashjointuples = outer_path_rows - outer_matched_rows;
4381  else
4382  hashjointuples = outer_matched_rows;
4383  }
4384  else
4385  {
4386  /*
4387  * The number of tuple comparisons needed is the number of outer
4388  * tuples times the typical number of tuples in a hash bucket, which
4389  * is the inner relation size times its bucketsize fraction. At each
4390  * one, we need to evaluate the hashjoin quals. But actually,
4391  * charging the full qual eval cost at each tuple is pessimistic,
4392  * since we don't evaluate the quals unless the hash values match
4393  * exactly. For lack of a better idea, halve the cost estimate to
4394  * allow for that.
4395  */
4396  startup_cost += hash_qual_cost.startup;
4397  run_cost += hash_qual_cost.per_tuple * outer_path_rows *
4398  clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
4399 
4400  /*
4401  * Get approx # tuples passing the hashquals. We use
4402  * approx_tuple_count here because we need an estimate done with
4403  * JOIN_INNER semantics.
4404  */
4405  hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
4406  }
4407 
4408  /*
4409  * For each tuple that gets through the hashjoin proper, we charge
4410  * cpu_tuple_cost plus the cost of evaluating additional restriction
4411  * clauses that are to be applied at the join. (This is pessimistic since
4412  * not all of the quals may get evaluated at each tuple.)
4413  */
4414  startup_cost += qp_qual_cost.startup;
4415  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
4416  run_cost += cpu_per_tuple * hashjointuples;
4417 
4418  /* tlist eval costs are paid per output row, not per tuple scanned */
4419  startup_cost += path->jpath.path.pathtarget->cost.startup;
4420  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
4421 
4422  path->jpath.path.startup_cost = startup_cost;
4423  path->jpath.path.total_cost = startup_cost + run_cost;
4424 }
4425 
4426 
4427 /*
4428  * cost_subplan
4429  * Figure the costs for a SubPlan (or initplan).
4430  *
4431  * Note: we could dig the subplan's Plan out of the root list, but in practice
4432  * all callers have it handy already, so we make them pass it.
4433  */
4434 void
4436 {
4437  QualCost sp_cost;
4438 
4439  /* Figure any cost for evaluating the testexpr */
4440  cost_qual_eval(&sp_cost,
4441  make_ands_implicit((Expr *) subplan->testexpr),
4442  root);
4443 
4444  if (subplan->useHashTable)
4445  {
4446  /*
4447  * If we are using a hash table for the subquery outputs, then the
4448  * cost of evaluating the query is a one-time cost. We charge one
4449  * cpu_operator_cost per tuple for the work of loading the hashtable,
4450  * too.
4451  */
4452  sp_cost.startup += plan->total_cost +
4453  cpu_operator_cost * plan->plan_rows;
4454 
4455  /*
4456  * The per-tuple costs include the cost of evaluating the lefthand
4457  * expressions, plus the cost of probing the hashtable. We already
4458  * accounted for the lefthand expressions as part of the testexpr, and
4459  * will also have counted one cpu_operator_cost for each comparison
4460  * operator. That is probably too low for the probing cost, but it's
4461  * hard to make a better estimate, so live with it for now.
4462  */
4463  }
4464  else
4465  {
4466  /*
4467  * Otherwise we will be rescanning the subplan output on each
4468  * evaluation. We need to estimate how much of the output we will
4469  * actually need to scan. NOTE: this logic should agree with the
4470  * tuple_fraction estimates used by make_subplan() in
4471  * plan/subselect.c.
4472  */
4473  Cost plan_run_cost = plan->total_cost - plan->startup_cost;
4474 
4475  if (subplan->subLinkType == EXISTS_SUBLINK)
4476  {
4477  /* we only need to fetch 1 tuple; clamp to avoid zero divide */
4478  sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
4479  }
4480  else if (subplan->subLinkType == ALL_SUBLINK ||
4481  subplan->subLinkType == ANY_SUBLINK)
4482  {
4483  /* assume we need 50% of the tuples */
4484  sp_cost.per_tuple += 0.50 * plan_run_cost;
4485  /* also charge a cpu_operator_cost per row examined */
4486  sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
4487  }
4488  else
4489  {
4490  /* assume we need all tuples */
4491  sp_cost.per_tuple += plan_run_cost;
4492  }
4493 
4494  /*
4495  * Also account for subplan's startup cost. If the subplan is
4496  * uncorrelated or undirect correlated, AND its topmost node is one
4497  * that materializes its output, assume that we'll only need to pay
4498  * its startup cost once; otherwise assume we pay the startup cost
4499  * every time.
4500  */
4501  if (subplan->parParam == NIL &&
4503  sp_cost.startup += plan->startup_cost;
4504  else
4505  sp_cost.per_tuple += plan->startup_cost;
4506  }
4507 
4508  subplan->startup_cost = sp_cost.startup;
4509  subplan->per_call_cost = sp_cost.per_tuple;
4510 }
4511 
4512 
4513 /*
4514  * cost_rescan
4515  * Given a finished Path, estimate the costs of rescanning it after
4516  * having done so the first time. For some Path types a rescan is
4517  * cheaper than an original scan (if no parameters change), and this
4518  * function embodies knowledge about that. The default is to return
4519  * the same costs stored in the Path. (Note that the cost estimates
4520  * actually stored in Paths are always for first scans.)
4521  *
4522  * This function is not currently intended to model effects such as rescans
4523  * being cheaper due to disk block caching; what we are concerned with is
4524  * plan types wherein the executor caches results explicitly, or doesn't
4525  * redo startup calculations, etc.
4526  */
4527 static void
4529  Cost *rescan_startup_cost, /* output parameters */
4530  Cost *rescan_total_cost)
4531 {
4532  switch (path->pathtype)
4533  {
4534  case T_FunctionScan:
4535 
4536  /*
4537  * Currently, nodeFunctionscan.c always executes the function to
4538  * completion before returning any rows, and caches the results in
4539  * a tuplestore. So the function eval cost is all startup cost
4540  * and isn't paid over again on rescans. However, all run costs
4541  * will be paid over again.
4542  */
4543  *rescan_startup_cost = 0;
4544  *rescan_total_cost = path->total_cost - path->startup_cost;
4545  break;
4546  case T_HashJoin:
4547 
4548  /*
4549  * If it's a single-batch join, we don't need to rebuild the hash
4550  * table during a rescan.
4551  */
4552  if (((HashPath *) path)->num_batches == 1)
4553  {
4554  /* Startup cost is exactly the cost of hash table building */
4555  *rescan_startup_cost = 0;
4556  *rescan_total_cost = path->total_cost - path->startup_cost;
4557  }
4558  else
4559  {
4560  /* Otherwise, no special treatment */
4561  *rescan_startup_cost = path->startup_cost;
4562  *rescan_total_cost = path->total_cost;
4563  }
4564  break;
4565  case T_CteScan:
4566  case T_WorkTableScan:
4567  {
4568  /*
4569  * These plan types materialize their final result in a
4570  * tuplestore or tuplesort object. So the rescan cost is only
4571  * cpu_tuple_cost per tuple, unless the result is large enough
4572  * to spill to disk.
4573  */
4574  Cost run_cost = cpu_tuple_cost * path->rows;
4575  double nbytes = relation_byte_size(path->rows,
4576  path->pathtarget->width);
4577  long work_mem_bytes = work_mem * 1024L;
4578 
4579  if (nbytes > work_mem_bytes)
4580  {
4581  /* It will spill, so account for re-read cost */
4582  double npages = ceil(nbytes / BLCKSZ);
4583 
4584  run_cost += seq_page_cost * npages;
4585  }
4586  *rescan_startup_cost = 0;
4587  *rescan_total_cost = run_cost;
4588  }
4589  break;
4590  case T_Material:
4591  case T_Sort:
4592  {
4593  /*
4594  * These plan types not only materialize their results, but do
4595  * not implement qual filtering or projection. So they are
4596  * even cheaper to rescan than the ones above. We charge only
4597  * cpu_operator_cost per tuple. (Note: keep that in sync with
4598  * the run_cost charge in cost_sort, and also see comments in
4599  * cost_material before you change it.)
4600  */
4601  Cost run_cost = cpu_operator_cost * path->rows;
4602  double nbytes = relation_byte_size(path->rows,
4603  path->pathtarget->width);
4604  long work_mem_bytes = work_mem * 1024L;
4605 
4606  if (nbytes > work_mem_bytes)
4607  {
4608  /* It will spill, so account for re-read cost */
4609  double npages = ceil(nbytes / BLCKSZ);
4610 
4611  run_cost += seq_page_cost * npages;
4612  }
4613  *rescan_startup_cost = 0;
4614  *rescan_total_cost = run_cost;
4615  }
4616  break;
4617  case T_Memoize:
4618  /* All the hard work is done by cost_memoize_rescan */
4620  rescan_startup_cost, rescan_total_cost);
4621  break;
4622  default:
4623  *rescan_startup_cost = path->startup_cost;
4624  *rescan_total_cost = path->total_cost;
4625  break;
4626  }
4627 }
4628 
4629 
4630 /*
4631  * cost_qual_eval
4632  * Estimate the CPU costs of evaluating a WHERE clause.
4633  * The input can be either an implicitly-ANDed list of boolean
4634  * expressions, or a list of RestrictInfo nodes. (The latter is
4635  * preferred since it allows caching of the results.)
4636  * The result includes both a one-time (startup) component,
4637  * and a per-evaluation component.
4638  */
4639 void
4641 {
4643  ListCell *l;
4644 
4645  context.root = root;
4646  context.total.startup = 0;
4647  context.total.per_tuple = 0;
4648 
4649  /* We don't charge any cost for the implicit ANDing at top level ... */
4650 
4651  foreach(l, quals)
4652  {
4653  Node *qual = (Node *) lfirst(l);
4654 
4656  }
4657 
4658  *cost = context.total;
4659 }
4660 
4661 /*
4662  * cost_qual_eval_node
4663  * As above, for a single RestrictInfo or expression.
4664  */
4665 void
4667 {
4669 
4670  context.root = root;
4671  context.total.startup = 0;
4672  context.total.per_tuple = 0;
4673 
4675 
4676  *cost = context.total;
4677 }
4678 
4679 static bool
4681 {
4682  if (node == NULL)
4683  return false;
4684 
4685  /*
4686  * RestrictInfo nodes contain an eval_cost field reserved for this
4687  * routine's use, so that it's not necessary to evaluate the qual clause's
4688  * cost more than once. If the clause's cost hasn't been computed yet,
4689  * the field's startup value will contain -1.
4690  */
4691  if (IsA(node, RestrictInfo))
4692  {
4693  RestrictInfo *rinfo = (RestrictInfo *) node;
4694 
4695  if (rinfo->eval_cost.startup < 0)
4696  {
4697  cost_qual_eval_context locContext;
4698 
4699  locContext.root = context->root;
4700  locContext.total.startup = 0;
4701  locContext.total.per_tuple = 0;
4702 
4703  /*
4704  * For an OR clause, recurse into the marked-up tree so that we
4705  * set the eval_cost for contained RestrictInfos too.
4706  */
4707  if (rinfo->orclause)
4708  cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
4709  else
4710  cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
4711 
4712  /*
4713  * If the RestrictInfo is marked pseudoconstant, it will be tested
4714  * only once, so treat its cost as all startup cost.
4715  */
4716  if (rinfo->pseudoconstant)
4717  {
4718  /* count one execution during startup */
4719  locContext.total.startup += locContext.total.per_tuple;
4720  locContext.total.per_tuple = 0;
4721  }
4722  rinfo->eval_cost = locContext.total;
4723  }
4724  context->total.startup += rinfo->eval_cost.startup;
4725  context->total.per_tuple += rinfo->eval_cost.per_tuple;
4726  /* do NOT recurse into children */
4727  return false;
4728  }
4729 
4730  /*
4731  * For each operator or function node in the given tree, we charge the
4732  * estimated execution cost given by pg_proc.procost (remember to multiply
4733  * this by cpu_operator_cost).
4734  *
4735  * Vars and Consts are charged zero, and so are boolean operators (AND,
4736  * OR, NOT). Simplistic, but a lot better than no model at all.
4737  *
4738  * Should we try to account for the possibility of short-circuit
4739  * evaluation of AND/OR? Probably *not*, because that would make the
4740  * results depend on the clause ordering, and we are not in any position
4741  * to expect that the current ordering of the clauses is the one that's
4742  * going to end up being used. The above per-RestrictInfo caching would
4743  * not mix well with trying to re-order clauses anyway.
4744  *
4745  * Another issue that is entirely ignored here is that if a set-returning
4746  * function is below top level in the tree, the functions/operators above
4747  * it will need to be evaluated multiple times. In practical use, such
4748  * cases arise so seldom as to not be worth the added complexity needed;
4749  * moreover, since our rowcount estimates for functions tend to be pretty
4750  * phony, the results would also be pretty phony.
4751  */
4752  if (IsA(node, FuncExpr))
4753  {
4754  add_function_cost(context->root, ((FuncExpr *) node)->funcid, node,
4755  &context->total);
4756  }
4757  else if (IsA(node, OpExpr) ||
4758  IsA(node, DistinctExpr) ||
4759  IsA(node, NullIfExpr))
4760  {
4761  /* rely on struct equivalence to treat these all alike */
4762  set_opfuncid((OpExpr *) node);
4763  add_function_cost(context->root, ((OpExpr *) node)->opfuncid, node,
4764  &context->total);
4765  }
4766  else if (IsA(node, ScalarArrayOpExpr))
4767  {
4768  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
4769  Node *arraynode = (Node *) lsecond(saop->args);
4770  QualCost sacosts;
4771  QualCost hcosts;
4772  double estarraylen = estimate_array_length(context->root, arraynode);
4773 
4774  set_sa_opfuncid(saop);
4775  sacosts.startup = sacosts.per_tuple = 0;
4776  add_function_cost(context->root, saop->opfuncid, NULL,
4777  &sacosts);
4778 
4779  if (OidIsValid(saop->hashfuncid))
4780  {
4781  /* Handle costs for hashed ScalarArrayOpExpr */
4782  hcosts.startup = hcosts.per_tuple = 0;
4783 
4784  add_function_cost(context->root, saop->hashfuncid, NULL, &hcosts);
4785  context->total.startup += sacosts.startup + hcosts.startup;
4786 
4787  /* Estimate the cost of building the hashtable. */
4788  context->total.startup += estarraylen * hcosts.per_tuple;
4789 
4790  /*
4791  * XXX should we charge a little bit for sacosts.per_tuple when
4792  * building the table, or is it ok to assume there will be zero
4793  * hash collision?
4794  */
4795 
4796  /*
4797  * Charge for hashtable lookups. Charge a single hash and a
4798  * single comparison.
4799  */
4800  context->total.per_tuple += hcosts.per_tuple + sacosts.per_tuple;
4801  }
4802  else
4803  {
4804  /*
4805  * Estimate that the operator will be applied to about half of the
4806  * array elements before the answer is determined.
4807  */
4808  context->total.startup += sacosts.startup;
4809  context->total.per_tuple += sacosts.per_tuple *
4810  estimate_array_length(context->root, arraynode) * 0.5;
4811  }
4812  }
4813  else if (IsA(node, Aggref) ||
4814  IsA(node, WindowFunc))
4815  {
4816  /*
4817  * Aggref and WindowFunc nodes are (and should be) treated like Vars,
4818  * ie, zero execution cost in the current model, because they behave
4819  * essentially like Vars at execution. We disregard the costs of
4820  * their input expressions for the same reason. The actual execution
4821  * costs of the aggregate/window functions and their arguments have to
4822  * be factored into plan-node-specific costing of the Agg or WindowAgg
4823  * plan node.
4824  */
4825  return false; /* don't recurse into children */
4826  }
4827  else if (IsA(node, GroupingFunc))
4828  {
4829  /* Treat this as having cost 1 */
4830  context->total.per_tuple += cpu_operator_cost;
4831  return false; /* don't recurse into children */
4832  }
4833  else if (IsA(node, CoerceViaIO))
4834  {
4835  CoerceViaIO *iocoerce = (CoerceViaIO *) node;
4836  Oid iofunc;
4837  Oid typioparam;
4838  bool typisvarlena;
4839 
4840  /* check the result type's input function */
4841  getTypeInputInfo(iocoerce->resulttype,
4842  &iofunc, &typioparam);
4843  add_function_cost(context->root, iofunc, NULL,
4844  &context->total);
4845  /* check the input type's output function */
4846  getTypeOutputInfo(exprType((Node *) iocoerce->arg),
4847  &iofunc, &typisvarlena);
4848  add_function_cost(context->root, iofunc, NULL,
4849  &context->total);
4850  }
4851  else if (IsA(node, ArrayCoerceExpr))
4852  {
4853  ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
4854  QualCost perelemcost;
4855 
4856  cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
4857  context->root);
4858  context->total.startup += perelemcost.startup;
4859  if (perelemcost.per_tuple > 0)
4860  context->total.per_tuple += perelemcost.per_tuple *
4861  estimate_array_length(context->root, (Node *) acoerce->arg);
4862  }
4863  else if (IsA(node, RowCompareExpr))
4864  {
4865  /* Conservatively assume we will check all the columns */
4866  RowCompareExpr *rcexpr = (RowCompareExpr *) node;
4867  ListCell *lc;
4868 
4869  foreach(lc, rcexpr->opnos)
4870  {
4871  Oid opid = lfirst_oid(lc);
4872 
4873  add_function_cost(context->root, get_opcode(opid), NULL,
4874  &context->total);
4875  }
4876  }
4877  else if (IsA(node, MinMaxExpr) ||
4878  IsA(node, SQLValueFunction) ||
4879  IsA(node, XmlExpr) ||
4880  IsA(node, CoerceToDomain) ||
4881  IsA(node, NextValueExpr) ||
4882  IsA(node, JsonExpr))
4883  {
4884  /* Treat all these as having cost 1 */
4885  context->total.per_tuple += cpu_operator_cost;
4886  }
4887  else if (IsA(node, CurrentOfExpr))
4888  {
4889  /* Report high cost to prevent selection of anything but TID scan */
4890  context->total.startup += disable_cost;
4891  }
4892  else if (IsA(node, SubLink))
4893  {
4894  /* This routine should not be applied to un-planned expressions */
4895  elog(ERROR, "cannot handle unplanned sub-select");
4896  }
4897  else if (IsA(node, SubPlan))
4898  {
4899  /*
4900  * A subplan node in an expression typically indicates that the
4901  * subplan will be executed on each evaluation, so charge accordingly.
4902  * (Sub-selects that can be executed as InitPlans have already been
4903  * removed from the expression.)
4904  */
4905  SubPlan *subplan = (SubPlan *) node;
4906 
4907  context->total.startup += subplan->startup_cost;
4908  context->total.per_tuple += subplan->per_call_cost;
4909 
4910  /*
4911  * We don't want to recurse into the testexpr, because it was already
4912  * counted in the SubPlan node's costs. So we're done.
4913  */
4914  return false;
4915  }
4916  else if (IsA(node, AlternativeSubPlan))
4917  {
4918  /*
4919  * Arbitrarily use the first alternative plan for costing. (We should
4920  * certainly only include one alternative, and we don't yet have
4921  * enough information to know which one the executor is most likely to
4922  * use.)
4923  */
4924  AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
4925 
4926  return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
4927  context);
4928  }
4929  else if (IsA(node, PlaceHolderVar))
4930  {
4931  /*
4932  * A PlaceHolderVar should be given cost zero when considering general
4933  * expression evaluation costs. The expense of doing the contained
4934  * expression is charged as part of the tlist eval costs of the scan
4935  * or join where the PHV is first computed (see set_rel_width and
4936  * add_placeholders_to_joinrel). If we charged it again here, we'd be
4937  * double-counting the cost for each level of plan that the PHV
4938  * bubbles up through. Hence, return without recursing into the
4939  * phexpr.
4940  */
4941  return false;
4942  }
4943 
4944  /* recurse into children */
4946  (void *) context);
4947 }
4948 
4949 /*
4950  * get_restriction_qual_cost
4951  * Compute evaluation costs of a baserel's restriction quals, plus any
4952  * movable join quals that have been pushed down to the scan.
4953  * Results are returned into *qpqual_cost.
4954  *
4955  * This is a convenience subroutine that works for seqscans and other cases
4956  * where all the given quals will be evaluated the hard way. It's not useful
4957  * for cost_index(), for example, where the index machinery takes care of
4958  * some of the quals. We assume baserestrictcost was previously set by
4959  * set_baserel_size_estimates().
4960  */
4961 static void
4963  ParamPathInfo *param_info,
4964  QualCost *qpqual_cost)
4965 {
4966  if (param_info)
4967  {
4968  /* Include costs of pushed-down clauses */
4969  cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
4970 
4971  qpqual_cost->startup += baserel->baserestrictcost.startup;
4972  qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
4973  }
4974  else
4975  *qpqual_cost = baserel->baserestrictcost;
4976 }
4977 
4978 
4979 /*
4980  * compute_semi_anti_join_factors
4981  * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
4982  * can be expected to scan.
4983  *
4984  * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
4985  * inner rows as soon as it finds a match to the current outer row.
4986  * The same happens if we have detected the inner rel is unique.
4987  * We should therefore adjust some of the cost components for this effect.
4988  * This function computes some estimates needed for these adjustments.
4989  * These estimates will be the same regardless of the particular paths used
4990  * for the outer and inner relation, so we compute these once and then pass
4991  * them to all the join cost estimation functions.
4992  *
4993  * Input parameters:
4994  * joinrel: join relation under consideration
4995  * outerrel: outer relation under consideration
4996  * innerrel: inner relation under consideration
4997  * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
4998  * sjinfo: SpecialJoinInfo relevant to this join
4999  * restrictlist: join quals
5000  * Output parameters:
5001  * *semifactors is filled in (see pathnodes.h for field definitions)
5002  */
5003 void
5005  RelOptInfo *joinrel,
5006  RelOptInfo *outerrel,
5007  RelOptInfo *innerrel,
5008  JoinType jointype,
5009  SpecialJoinInfo *sjinfo,
5010  List *restrictlist,
5011  SemiAntiJoinFactors *semifactors)
5012 {
5013  Selectivity jselec;
5014  Selectivity nselec;
5015  Selectivity avgmatch;
5016  SpecialJoinInfo norm_sjinfo;
5017  List *joinquals;
5018  ListCell *l;
5019 
5020  /*
5021  * In an ANTI join, we must ignore clauses that are "pushed down", since
5022  * those won't affect the match logic. In a SEMI join, we do not
5023  * distinguish joinquals from "pushed down" quals, so just use the whole
5024  * restrictinfo list. For other outer join types, we should consider only
5025  * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
5026  */
5027  if (IS_OUTER_JOIN(jointype))
5028  {
5029  joinquals = NIL;
5030  foreach(l, restrictlist)
5031  {
5032  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5033 
5034  if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5035  joinquals = lappend(joinquals, rinfo);
5036  }
5037  }
5038  else
5039  joinquals = restrictlist;
5040 
5041  /*
5042  * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
5043  */
5044  jselec = clauselist_selectivity(root,
5045  joinquals,
5046  0,
5047  (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
5048  sjinfo);
5049 
5050  /*
5051  * Also get the normal inner-join selectivity of the join clauses.
5052  */
5053  init_dummy_sjinfo(&norm_sjinfo, outerrel->relids, innerrel->relids);
5054 
5055  nselec = clauselist_selectivity(root,
5056  joinquals,
5057  0,
5058  JOIN_INNER,
5059  &norm_sjinfo);
5060 
5061  /* Avoid leaking a lot of ListCells */
5062  if (IS_OUTER_JOIN(jointype))
5063  list_free(joinquals);
5064 
5065  /*
5066  * jselec can be interpreted as the fraction of outer-rel rows that have
5067  * any matches (this is true for both SEMI and ANTI cases). And nselec is
5068  * the fraction of the Cartesian product that matches. So, the average
5069  * number of matches for each outer-rel row that has at least one match is
5070  * nselec * inner_rows / jselec.
5071  *
5072  * Note: it is correct to use the inner rel's "rows" count here, even
5073  * though we might later be considering a parameterized inner path with
5074  * fewer rows. This is because we have included all the join clauses in
5075  * the selectivity estimate.
5076  */
5077  if (jselec > 0) /* protect against zero divide */
5078  {
5079  avgmatch = nselec * innerrel->rows / jselec;
5080  /* Clamp to sane range */
5081  avgmatch = Max(1.0, avgmatch);
5082  }
5083  else
5084  avgmatch = 1.0;
5085 
5086  semifactors->outer_match_frac = jselec;
5087  semifactors->match_count = avgmatch;
5088 }
5089 
5090 /*
5091  * has_indexed_join_quals
5092  * Check whether all the joinquals of a nestloop join are used as
5093  * inner index quals.
5094  *
5095  * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
5096  * indexscan) that uses all the joinquals as indexquals, we can assume that an
5097  * unmatched outer tuple is cheap to process, whereas otherwise it's probably
5098  * expensive.
5099  */
5100 static bool
5102 {
5103  JoinPath *joinpath = &path->jpath;
5104  Relids joinrelids = joinpath->path.parent->relids;
5105  Path *innerpath = joinpath->innerjoinpath;
5106  List *indexclauses;
5107  bool found_one;
5108  ListCell *lc;
5109 
5110  /* If join still has quals to evaluate, it's not fast */
5111  if (joinpath->joinrestrictinfo != NIL)
5112  return false;
5113  /* Nor if the inner path isn't parameterized at all */
5114  if (innerpath->param_info == NULL)
5115  return false;
5116 
5117  /* Find the indexclauses list for the inner scan */
5118  switch (innerpath->pathtype)
5119  {
5120  case T_IndexScan:
5121  case T_IndexOnlyScan:
5122  indexclauses = ((IndexPath *) innerpath)->indexclauses;
5123  break;
5124  case T_BitmapHeapScan:
5125  {
5126  /* Accept only a simple bitmap scan, not AND/OR cases */
5127  Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
5128 
5129  if (IsA(bmqual, IndexPath))
5130  indexclauses = ((IndexPath *) bmqual)->indexclauses;
5131  else
5132  return false;
5133  break;
5134  }
5135  default:
5136 
5137  /*
5138  * If it's not a simple indexscan, it probably doesn't run quickly
5139  * for zero rows out, even if it's a parameterized path using all
5140  * the joinquals.
5141  */
5142  return false;
5143  }
5144 
5145  /*
5146  * Examine the inner path's param clauses. Any that are from the outer
5147  * path must be found in the indexclauses list, either exactly or in an
5148  * equivalent form generated by equivclass.c. Also, we must find at least
5149  * one such clause, else it's a clauseless join which isn't fast.
5150  */
5151  found_one = false;
5152  foreach(lc, innerpath->param_info->ppi_clauses)
5153  {
5154  RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
5155 
5156  if (join_clause_is_movable_into(rinfo,
5157  innerpath->parent->relids,
5158  joinrelids))
5159  {
5160  if (!is_redundant_with_indexclauses(rinfo, indexclauses))
5161  return false;
5162  found_one = true;
5163  }
5164  }
5165  return found_one;
5166 }
5167 
5168 
5169 /*
5170  * approx_tuple_count
5171  * Quick-and-dirty estimation of the number of join rows passing
5172  * a set of qual conditions.
5173  *
5174  * The quals can be either an implicitly-ANDed list of boolean expressions,
5175  * or a list of RestrictInfo nodes (typically the latter).
5176  *
5177  * We intentionally compute the selectivity under JOIN_INNER rules, even
5178  * if it's some type of outer join. This is appropriate because we are
5179  * trying to figure out how many tuples pass the initial merge or hash
5180  * join step.
5181  *
5182  * This is quick-and-dirty because we bypass clauselist_selectivity, and
5183  * simply multiply the independent clause selectivities together. Now
5184  * clauselist_selectivity often can't do any better than that anyhow, but
5185  * for some situations (such as range constraints) it is smarter. However,
5186  * we can't effectively cache the results of clauselist_selectivity, whereas
5187  * the individual clause selectivities can be and are cached.
5188  *
5189  * Since we are only using the results to estimate how many potential
5190  * output tuples are generated and passed through qpqual checking, it
5191  * seems OK to live with the approximation.
5192  */
5193 static double
5195 {
5196  double tuples;
5197  double outer_tuples = path->outerjoinpath->rows;
5198  double inner_tuples = path->innerjoinpath->rows;
5199  SpecialJoinInfo sjinfo;
5200  Selectivity selec = 1.0;
5201  ListCell *l;
5202 
5203  /*
5204  * Make up a SpecialJoinInfo for JOIN_INNER semantics.
5205  */
5206  init_dummy_sjinfo(&sjinfo, path->outerjoinpath->parent->relids,
5207  path->innerjoinpath->parent->relids);
5208 
5209  /* Get the approximate selectivity */
5210  foreach(l, quals)
5211  {
5212  Node *qual = (Node *) lfirst(l);
5213 
5214  /* Note that clause_selectivity will be able to cache its result */
5215  selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
5216  }
5217 
5218  /* Apply it to the input relation sizes */
5219  tuples = selec * outer_tuples * inner_tuples;
5220 
5221  return clamp_row_est(tuples);
5222 }
5223 
5224 
5225 /*
5226  * set_baserel_size_estimates
5227  * Set the size estimates for the given base relation.
5228  *
5229  * The rel's targetlist and restrictinfo list must have been constructed
5230  * already, and rel->tuples must be set.
5231  *
5232  * We set the following fields of the rel node:
5233  * rows: the estimated number of output tuples (after applying
5234  * restriction clauses).
5235  * width: the estimated average output tuple width in bytes.
5236  * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
5237  */
5238 void
5240 {
5241  double nrows;
5242 
5243  /* Should only be applied to base relations */
5244  Assert(rel->relid > 0);
5245 
5246  nrows = rel->tuples *
5248  rel->baserestrictinfo,
5249  0,
5250  JOIN_INNER,
5251  NULL);
5252 
5253  rel->rows = clamp_row_est(nrows);
5254 
5256 
5257  set_rel_width(root, rel);
5258 }
5259 
5260 /*
5261  * get_parameterized_baserel_size
5262  * Make a size estimate for a parameterized scan of a base relation.
5263  *
5264  * 'param_clauses' lists the additional join clauses to be used.
5265  *
5266  * set_baserel_size_estimates must have been applied already.
5267  */
5268 double
5270  List *param_clauses)
5271 {
5272  List *allclauses;
5273  double nrows;
5274 
5275  /*
5276  * Estimate the number of rows returned by the parameterized scan, knowing
5277  * that it will apply all the extra join clauses as well as the rel's own
5278  * restriction clauses. Note that we force the clauses to be treated as
5279  * non-join clauses during selectivity estimation.
5280  */
5281  allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);
5282  nrows = rel->tuples *
5284  allclauses,
5285  rel->relid, /* do not use 0! */
5286  JOIN_INNER,
5287  NULL);
5288  nrows = clamp_row_est(nrows);
5289  /* For safety, make sure result is not more than the base estimate */
5290  if (nrows > rel->rows)
5291  nrows = rel->rows;
5292  return nrows;
5293 }
5294 
5295 /*
5296  * set_joinrel_size_estimates
5297  * Set the size estimates for the given join relation.
5298  *
5299  * The rel's targetlist must have been constructed already, and a
5300  * restriction clause list that matches the given component rels must
5301  * be provided.
5302  *
5303  * Since there is more than one way to make a joinrel for more than two
5304  * base relations, the results we get here could depend on which component
5305  * rel pair is provided. In theory we should get the same answers no matter
5306  * which pair is provided; in practice, since the selectivity estimation
5307  * routines don't handle all cases equally well, we might not. But there's
5308  * not much to be done about it. (Would it make sense to repeat the
5309  * calculations for each pair of input rels that's encountered, and somehow
5310  * average the results? Probably way more trouble than it's worth, and
5311  * anyway we must keep the rowcount estimate the same for all paths for the
5312  * joinrel.)
5313  *
5314  * We set only the rows field here. The reltarget field was already set by
5315  * build_joinrel_tlist, and baserestrictcost is not used for join rels.
5316  */
5317 void
5319  RelOptInfo *outer_rel,
5320  RelOptInfo *inner_rel,
5321  SpecialJoinInfo *sjinfo,
5322  List *restrictlist)
5323 {
5325  rel,
5326  outer_rel,
5327  inner_rel,
5328  outer_rel->rows,
5329  inner_rel->rows,
5330  sjinfo,
5331  restrictlist);
5332 }
5333 
5334 /*
5335  * get_parameterized_joinrel_size
5336  * Make a size estimate for a parameterized scan of a join relation.
5337  *
5338  * 'rel' is the joinrel under consideration.
5339  * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
5340  * produce the relations being joined.
5341  * 'sjinfo' is any SpecialJoinInfo relevant to this join.
5342  * 'restrict_clauses' lists the join clauses that need to be applied at the
5343  * join node (including any movable clauses that were moved down to this join,
5344  * and not including any movable clauses that were pushed down into the
5345  * child paths).
5346  *
5347  * set_joinrel_size_estimates must have been applied already.
5348  */
5349 double
5351  Path *outer_path,
5352  Path *inner_path,
5353  SpecialJoinInfo *sjinfo,
5354  List *restrict_clauses)
5355 {
5356  double nrows;
5357 
5358  /*
5359  * Estimate the number of rows returned by the parameterized join as the
5360  * sizes of the input paths times the selectivity of the clauses that have
5361  * ended up at this join node.
5362  *
5363  * As with set_joinrel_size_estimates, the rowcount estimate could depend
5364  * on the pair of input paths provided, though ideally we'd get the same
5365  * estimate for any pair with the same parameterization.
5366  */
5368  rel,
5369  outer_path->parent,
5370  inner_path->parent,
5371  outer_path->rows,
5372  inner_path->rows,
5373  sjinfo,
5374  restrict_clauses);
5375  /* For safety, make sure result is not more than the base estimate */
5376  if (nrows > rel->rows)
5377  nrows = rel->rows;
5378  return nrows;
5379 }
5380 
5381 /*
5382  * calc_joinrel_size_estimate
5383  * Workhorse for set_joinrel_size_estimates and
5384  * get_parameterized_joinrel_size.
5385  *
5386  * outer_rel/inner_rel are the relations being joined, but they should be
5387  * assumed to have sizes outer_rows/inner_rows; those numbers might be less
5388  * than what rel->rows says, when we are considering parameterized paths.
5389  */
5390 static double
5392  RelOptInfo *joinrel,
5393  RelOptInfo *outer_rel,
5394  RelOptInfo *inner_rel,
5395  double outer_rows,
5396  double inner_rows,
5397  SpecialJoinInfo *sjinfo,
5398  List *restrictlist)
5399 {
5400  JoinType jointype = sjinfo->jointype;
5401  Selectivity fkselec;
5402  Selectivity jselec;
5403  Selectivity pselec;
5404  double nrows;
5405 
5406  /*
5407  * Compute joinclause selectivity. Note that we are only considering
5408  * clauses that become restriction clauses at this join level; we are not
5409  * double-counting them because they were not considered in estimating the
5410  * sizes of the component rels.
5411  *
5412  * First, see whether any of the joinclauses can be matched to known FK
5413  * constraints. If so, drop those clauses from the restrictlist, and
5414  * instead estimate their selectivity using FK semantics. (We do this
5415  * without regard to whether said clauses are local or "pushed down".
5416  * Probably, an FK-matching clause could never be seen as pushed down at
5417  * an outer join, since it would be strict and hence would be grounds for
5418  * join strength reduction.) fkselec gets the net selectivity for
5419  * FK-matching clauses, or 1.0 if there are none.
5420  */
5422  outer_rel->relids,
5423  inner_rel->relids,
5424  sjinfo,
5425  &restrictlist);
5426 
5427  /*
5428  * For an outer join, we have to distinguish the selectivity of the join's
5429  * own clauses (JOIN/ON conditions) from any clauses that were "pushed
5430  * down". For inner joins we just count them all as joinclauses.
5431  */
5432  if (IS_OUTER_JOIN(jointype))
5433  {
5434  List *joinquals = NIL;
5435  List *pushedquals = NIL;
5436  ListCell *l;
5437 
5438  /* Grovel through the clauses to separate into two lists */
5439  foreach(l, restrictlist)
5440  {
5441  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
5442 
5443  if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
5444  pushedquals = lappend(pushedquals, rinfo);
5445  else
5446  joinquals = lappend(joinquals, rinfo);
5447  }
5448 
5449  /* Get the separate selectivities */
5450  jselec = clauselist_selectivity(root,
5451  joinquals,
5452  0,
5453  jointype,
5454  sjinfo);
5455  pselec = clauselist_selectivity(root,
5456  pushedquals,
5457  0,
5458  jointype,
5459  sjinfo);
5460 
5461  /* Avoid leaking a lot of ListCells */
5462  list_free(joinquals);
5463  list_free(pushedquals);
5464  }
5465  else
5466  {
5467  jselec = clauselist_selectivity(root,
5468  restrictlist,
5469  0,
5470  jointype,
5471  sjinfo);
5472  pselec = 0.0; /* not used, keep compiler quiet */
5473  }
5474 
5475  /*
5476  * Basically, we multiply size of Cartesian product by selectivity.
5477  *
5478  * If we are doing an outer join, take that into account: the joinqual
5479  * selectivity has to be clamped using the knowledge that the output must
5480  * be at least as large as the non-nullable input. However, any
5481  * pushed-down quals are applied after the outer join, so their
5482  * selectivity applies fully.
5483  *
5484  * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
5485  * of LHS rows that have matches, and we apply that straightforwardly.
5486  */
5487  switch (jointype)
5488  {
5489  case JOIN_INNER:
5490  nrows = outer_rows * inner_rows * fkselec * jselec;
5491  /* pselec not used */
5492  break;
5493  case JOIN_LEFT:
5494  nrows = outer_rows * inner_rows * fkselec * jselec;
5495  if (nrows < outer_rows)
5496  nrows = outer_rows;
5497  nrows *= pselec;
5498  break;
5499  case JOIN_FULL:
5500  nrows = outer_rows * inner_rows * fkselec * jselec;
5501  if (nrows < outer_rows)
5502  nrows = outer_rows;
5503  if (nrows < inner_rows)
5504  nrows = inner_rows;
5505  nrows *= pselec;
5506  break;
5507  case JOIN_SEMI:
5508  nrows = outer_rows * fkselec * jselec;
5509  /* pselec not used */
5510  break;
5511  case JOIN_ANTI:
5512  nrows = outer_rows * (1.0 - fkselec * jselec);
5513  nrows *= pselec;
5514  break;
5515  default:
5516  /* other values not expected here */
5517  elog(ERROR, "unrecognized join type: %d", (int) jointype);
5518  nrows = 0; /* keep compiler quiet */
5519  break;
5520  }
5521 
5522  return clamp_row_est(nrows);
5523 }
5524 
5525 /*
5526  * get_foreign_key_join_selectivity
5527  * Estimate join selectivity for foreign-key-related clauses.
5528  *
5529  * Remove any clauses that can be matched to FK constraints from *restrictlist,
5530  * and return a substitute estimate of their selectivity. 1.0 is returned
5531  * when there are no such clauses.
5532  *
5533  * The reason for treating such clauses specially is that we can get better
5534  * estimates this way than by relying on clauselist_selectivity(), especially
5535  * for multi-column FKs where that function's assumption that the clauses are
5536  * independent falls down badly. But even with single-column FKs, we may be
5537  * able to get a better answer when the pg_statistic stats are missing or out
5538  * of date.
5539  */
5540 static Selectivity
5542  Relids outer_relids,
5543  Relids inner_relids,
5544  SpecialJoinInfo *sjinfo,
5545  List **restrictlist)
5546 {
5547  Selectivity fkselec = 1.0;
5548  JoinType jointype = sjinfo->jointype;
5549  List *worklist = *restrictlist;
5550  ListCell *lc;
5551 
5552  /* Consider each FK constraint that is known to match the query */
5553  foreach(lc, root->fkey_list)
5554  {
5555  ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
5556  bool ref_is_outer;
5557  List *removedlist;
5558  ListCell *cell;
5559 
5560  /*
5561  * This FK is not relevant unless it connects a baserel on one side of
5562  * this join to a baserel on the other side.
5563  */
5564  if (bms_is_member(fkinfo->con_relid, outer_relids) &&
5565  bms_is_member(fkinfo->ref_relid, inner_relids))
5566  ref_is_outer = false;
5567  else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
5568  bms_is_member(fkinfo->con_relid, inner_relids))
5569  ref_is_outer = true;
5570  else
5571  continue;
5572 
5573  /*
5574  * If we're dealing with a semi/anti join, and the FK's referenced
5575  * relation is on the outside, then knowledge of the FK doesn't help
5576  * us figure out what we need to know (which is the fraction of outer
5577  * rows that have matches). On the other hand, if the referenced rel
5578  * is on the inside, then all outer rows must have matches in the
5579  * referenced table (ignoring nulls). But any restriction or join
5580  * clauses that filter that table will reduce the fraction of matches.
5581  * We can account for restriction clauses, but it's too hard to guess
5582  * how many table rows would get through a join that's inside the RHS.
5583  * Hence, if either case applies, punt and ignore the FK.
5584  */
5585  if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
5586  (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
5587  continue;
5588 
5589  /*
5590  * Modify the restrictlist by removing clauses that match the FK (and
5591  * putting them into removedlist instead). It seems unsafe to modify
5592  * the originally-passed List structure, so we make a shallow copy the
5593  * first time through.
5594  */
5595  if (worklist == *restrictlist)
5596  worklist = list_copy(worklist);
5597 
5598  removedlist = NIL;
5599  foreach(cell, worklist)
5600  {
5601  RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
5602  bool remove_it = false;
5603  int i;
5604 
5605  /* Drop this clause if it matches any column of the FK */
5606  for (i = 0; i < fkinfo->nkeys; i++)
5607  {
5608  if (rinfo->parent_ec)
5609  {
5610  /*
5611  * EC-derived clauses can only match by EC. It is okay to
5612  * consider any clause derived from the same EC as
5613  * matching the FK: even if equivclass.c chose to generate
5614  * a clause equating some other pair of Vars, it could
5615  * have generated one equating the FK's Vars. So for
5616  * purposes of estimation, we can act as though it did so.
5617  *
5618  * Note: checking parent_ec is a bit of a cheat because
5619  * there are EC-derived clauses that don't have parent_ec
5620  * set; but such clauses must compare expressions that
5621  * aren't just Vars, so they cannot match the FK anyway.
5622  */
5623  if (fkinfo->eclass[i] == rinfo->parent_ec)
5624  {
5625  remove_it = true;
5626  break;
5627  }
5628  }
5629  else
5630  {
5631  /*
5632  * Otherwise, see if rinfo was previously matched to FK as
5633  * a "loose" clause.
5634  */
5635  if (list_member_ptr(fkinfo->rinfos[i], rinfo))
5636  {
5637  remove_it = true;
5638  break;
5639  }
5640  }
5641  }
5642  if (remove_it)
5643  {
5644  worklist = foreach_delete_current(worklist, cell);
5645  removedlist = lappend(removedlist, rinfo);
5646  }
5647  }
5648 
5649  /*
5650  * If we failed to remove all the matching clauses we expected to
5651  * find, chicken out and ignore this FK; applying its selectivity
5652  * might result in double-counting. Put any clauses we did manage to
5653  * remove back into the worklist.
5654  *
5655  * Since the matching clauses are known not outerjoin-delayed, they
5656  * would normally have appeared in the initial joinclause list. If we
5657  * didn't find them, there are two possibilities:
5658  *
5659  * 1. If the FK match is based on an EC that is ec_has_const, it won't
5660  * have generated any join clauses at all. We discount such ECs while
5661  * checking to see if we have "all" the clauses. (Below, we'll adjust
5662  * the selectivity estimate for this case.)
5663  *
5664  * 2. The clauses were matched to some other FK in a previous
5665  * iteration of this loop, and thus removed from worklist. (A likely
5666  * case is that two FKs are matched to the same EC; there will be only
5667  * one EC-derived clause in the initial list, so the first FK will
5668  * consume it.) Applying both FKs' selectivity independently risks
5669  * underestimating the join size; in particular, this would undo one
5670  * of the main things that ECs were invented for, namely to avoid
5671  * double-counting the selectivity of redundant equality conditions.
5672  * Later we might think of a reasonable way to combine the estimates,
5673  * but for now, just punt, since this is a fairly uncommon situation.
5674  */
5675  if (removedlist == NIL ||
5676  list_length(removedlist) !=
5677  (fkinfo->nmatched_ec - fkinfo->nconst_ec + fkinfo->nmatched_ri))
5678  {
5679  worklist = list_concat(worklist, removedlist);
5680  continue;
5681  }
5682 
5683  /*
5684  * Finally we get to the payoff: estimate selectivity using the
5685  * knowledge that each referencing row will match exactly one row in
5686  * the referenced table.
5687  *
5688  * XXX that's not true in the presence of nulls in the referencing
5689  * column(s), so in principle we should derate the estimate for those.
5690  * However (1) if there are any strict restriction clauses for the
5691  * referencing column(s) elsewhere in the query, derating here would
5692  * be double-counting the null fraction, and (2) it's not very clear
5693  * how to combine null fractions for multiple referencing columns. So
5694  * we do nothing for now about correcting for nulls.
5695  *
5696  * XXX another point here is that if either side of an FK constraint
5697  * is an inheritance parent, we estimate as though the constraint
5698  * covers all its children as well. This is not an unreasonable
5699  * assumption for a referencing table, ie the user probably applied
5700  * identical constraints to all child tables (though perhaps we ought
5701  * to check that). But it's not possible to have done that for a
5702  * referenced table. Fortunately, precisely because that doesn't
5703  * work, it is uncommon in practice to have an FK referencing a parent
5704  * table. So, at least for now, disregard inheritance here.
5705  */
5706  if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
5707  {
5708  /*
5709  * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
5710  * referenced table is exactly the inside of the join. The join
5711  * selectivity is defined as the fraction of LHS rows that have
5712  * matches. The FK implies that every LHS row has a match *in the
5713  * referenced table*; but any restriction clauses on it will
5714  * reduce the number of matches. Hence we take the join
5715  * selectivity as equal to the selectivity of the table's
5716  * restriction clauses, which is rows / tuples; but we must guard
5717  * against tuples == 0.
5718  */
5719  RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5720  double ref_tuples = Max(ref_rel->tuples, 1.0);
5721 
5722  fkselec *= ref_rel->rows / ref_tuples;
5723  }
5724  else
5725  {
5726  /*
5727  * Otherwise, selectivity is exactly 1/referenced-table-size; but
5728  * guard against tuples == 0. Note we should use the raw table
5729  * tuple count, not any estimate of its filtered or joined size.
5730  */
5731  RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
5732  double ref_tuples = Max(ref_rel->tuples, 1.0);
5733 
5734  fkselec *= 1.0 / ref_tuples;
5735  }
5736 
5737  /*
5738  * If any of the FK columns participated in ec_has_const ECs, then
5739  * equivclass.c will have generated "var = const" restrictions for
5740  * each side of the join, thus reducing the sizes of both input
5741  * relations. Taking the fkselec at face value would amount to
5742  * double-counting the selectivity of the constant restriction for the
5743  * referencing Var. Hence, look for the restriction clause(s) that
5744  * were applied to the referencing Var(s), and divide out their
5745  * selectivity to correct for this.
5746  */
5747  if (fkinfo->nconst_ec > 0)
5748  {
5749  for (int i = 0; i < fkinfo->nkeys; i++)
5750  {
5751  EquivalenceClass *ec = fkinfo->eclass[i];
5752 
5753  if (ec && ec->ec_has_const)
5754  {
5755  EquivalenceMember *em = fkinfo->fk_eclass_member[i];
5757  em);
5758 
5759  if (rinfo)
5760  {
5761  Selectivity s0;
5762 
5763  s0 = clause_selectivity(root,
5764  (Node *) rinfo,
5765  0,
5766  jointype,
5767  sjinfo);
5768  if (s0 > 0)
5769  fkselec /= s0;
5770  }
5771  }
5772  }
5773  }
5774  }
5775 
5776  *restrictlist = worklist;
5777  CLAMP_PROBABILITY(fkselec);
5778  return fkselec;
5779 }
5780 
5781 /*
5782  * set_subquery_size_estimates
5783  * Set the size estimates for a base relation that is a subquery.
5784  *
5785  * The rel's targetlist and restrictinfo list must have been constructed
5786  * already, and the Paths for the subquery must have been completed.
5787  * We look at the subquery's PlannerInfo to extract data.
5788  *
5789  * We set the same fields as set_baserel_size_estimates.
5790  */
5791 void
5793 {
5794  PlannerInfo *subroot = rel->subroot;
5795  RelOptInfo *sub_final_rel;
5796  ListCell *lc;
5797 
5798  /* Should only be applied to base relations that are subqueries */
5799  Assert(rel->relid > 0);
5800  Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
5801 
5802  /*
5803  * Copy raw number of output rows from subquery. All of its paths should
5804  * have the same output rowcount, so just look at cheapest-total.
5805  */
5806  sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
5807  rel->tuples = sub_final_rel->cheapest_total_path->rows;
5808 
5809  /*
5810  * Compute per-output-column width estimates by examining the subquery's
5811  * targetlist. For any output that is a plain Var, get the width estimate
5812  * that was made while planning the subquery. Otherwise, we leave it to
5813  * set_rel_width to fill in a datatype-based default estimate.
5814  */
5815  foreach(lc, subroot->parse->targetList)
5816  {
5817  TargetEntry *te = lfirst_node(TargetEntry, lc);
5818  Node *texpr = (Node *) te->expr;
5819  int32 item_width = 0;
5820 
5821  /* junk columns aren't visible to upper query */
5822  if (te->resjunk)
5823  continue;
5824 
5825  /*
5826  * The subquery could be an expansion of a view that's had columns
5827  * added to it since the current query was parsed, so that there are
5828  * non-junk tlist columns in it that don't correspond to any column
5829  * visible at our query level. Ignore such columns.
5830  */
5831  if (te->resno < rel->min_attr || te->resno > rel->max_attr)
5832  continue;
5833 
5834  /*
5835  * XXX This currently doesn't work for subqueries containing set
5836  * operations, because the Vars in their tlists are bogus references
5837  * to the first leaf subquery, which wouldn't give the right answer
5838  * even if we could still get to its PlannerInfo.
5839  *
5840  * Also, the subquery could be an appendrel for which all branches are
5841  * known empty due to constraint exclusion, in which case
5842  * set_append_rel_pathlist will have left the attr_widths set to zero.
5843  *
5844  * In either case, we just leave the width estimate zero until
5845  * set_rel_width fixes it.
5846  */
5847  if (IsA(texpr, Var) &&
5848  subroot->parse->setOperations == NULL)
5849  {
5850  Var *var = (Var *) texpr;
5851  RelOptInfo *subrel = find_base_rel(subroot, var->varno);
5852 
5853  item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
5854  }
5855  rel->attr_widths[te->resno - rel->min_attr] = item_width;
5856  }
5857 
5858  /* Now estimate number of output rows, etc */
5860 }
5861 
5862 /*
5863  * set_function_size_estimates
5864  * Set the size estimates for a base relation that is a function call.
5865  *
5866  * The rel's targetlist and restrictinfo list must have been constructed
5867  * already.
5868  *
5869  * We set the same fields as set_baserel_size_estimates.
5870  */
5871 void
5873 {
5874  RangeTblEntry *rte;
5875  ListCell *lc;
5876 
5877  /* Should only be applied to base relations that are functions */
5878  Assert(rel->relid > 0);
5879  rte = planner_rt_fetch(rel->relid, root);
5880  Assert(rte->rtekind == RTE_FUNCTION);
5881 
5882  /*
5883  * Estimate number of rows the functions will return. The rowcount of the
5884  * node is that of the largest function result.
5885  */
5886  rel->tuples = 0;
5887  foreach(lc, rte->functions)
5888  {
5889  RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
5890  double ntup = expression_returns_set_rows(root, rtfunc->funcexpr);
5891 
5892  if (ntup > rel->tuples)
5893  rel->tuples = ntup;
5894  }
5895 
5896  /* Now estimate number of output rows, etc */
5898 }
5899 
5900 /*
5901  * set_function_size_estimates
5902  * Set the size estimates for a base relation that is a function call.
5903  *
5904  * The rel's targetlist and restrictinfo list must have been constructed
5905  * already.
5906  *
5907  * We set the same fields as set_tablefunc_size_estimates.
5908  */
5909 void
5911 {
5912  /* Should only be applied to base relations that are functions */
5913  Assert(rel->relid > 0);
5914  Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
5915 
5916  rel->tuples = 100;
5917 
5918  /* Now estimate number of output rows, etc */
5920 }
5921 
5922 /*
5923  * set_values_size_estimates
5924  * Set the size estimates for a base relation that is a values list.
5925  *
5926  * The rel's targetlist and restrictinfo list must have been constructed
5927  * already.
5928  *
5929  * We set the same fields as set_baserel_size_estimates.
5930  */
5931 void
5933 {
5934  RangeTblEntry *rte;
5935 
5936  /* Should only be applied to base relations that are values lists */
5937  Assert(rel->relid > 0);
5938  rte = planner_rt_fetch(rel->relid, root);
5939  Assert(rte->rtekind == RTE_VALUES);
5940 
5941  /*
5942  * Estimate number of rows the values list will return. We know this
5943  * precisely based on the list length (well, barring set-returning
5944  * functions in list items, but that's a refinement not catered for
5945  * anywhere else either).
5946  */
5947  rel->tuples = list_length(rte->values_lists);
5948 
5949  /* Now estimate number of output rows, etc */
5951 }
5952 
5953 /*
5954  * set_cte_size_estimates
5955  * Set the size estimates for a base relation that is a CTE reference.
5956  *
5957  * The rel's targetlist and restrictinfo list must have been constructed
5958  * already, and we need an estimate of the number of rows returned by the CTE
5959  * (if a regular CTE) or the non-recursive term (if a self-reference).
5960  *
5961  * We set the same fields as set_baserel_size_estimates.
5962  */
5963 void
5965 {
5966  RangeTblEntry *rte;
5967 
5968  /* Should only be applied to base relations that are CTE references */
5969  Assert(rel->relid > 0);
5970  rte = planner_rt_fetch(rel->relid, root);
5971  Assert(rte->rtekind == RTE_CTE);
5972 
5973  if (rte->self_reference)
5974  {
5975  /*
5976  * In a self-reference, we assume the average worktable size is a
5977  * multiple of the nonrecursive term's size. The best multiplier will
5978  * vary depending on query "fan-out", so make its value adjustable.
5979  */
5980  rel->tuples = clamp_row_est(recursive_worktable_factor * cte_rows);
5981  }
5982  else
5983  {
5984  /* Otherwise just believe the CTE's rowcount estimate */
5985  rel->tuples = cte_rows;
5986  }
5987 
5988  /* Now estimate number of output rows, etc */
5990 }
5991 
5992 /*
5993  * set_namedtuplestore_size_estimates
5994  * Set the size estimates for a base relation that is a tuplestore reference.
5995  *
5996  * The rel's targetlist and restrictinfo list must have been constructed
5997  * already.
5998  *
5999  * We set the same fields as set_baserel_size_estimates.
6000  */
6001 void