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 master 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) 1996-2018, 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 #ifdef _MSC_VER
75 #include <float.h> /* for _isnan */
76 #endif
77 #include <math.h>
78 
79 #include "access/amapi.h"
80 #include "access/htup_details.h"
81 #include "access/tsmapi.h"
82 #include "executor/executor.h"
83 #include "executor/nodeHash.h"
84 #include "miscadmin.h"
85 #include "nodes/nodeFuncs.h"
86 #include "optimizer/clauses.h"
87 #include "optimizer/cost.h"
88 #include "optimizer/pathnode.h"
89 #include "optimizer/paths.h"
90 #include "optimizer/placeholder.h"
91 #include "optimizer/plancat.h"
92 #include "optimizer/planmain.h"
93 #include "optimizer/restrictinfo.h"
94 #include "parser/parsetree.h"
95 #include "utils/lsyscache.h"
96 #include "utils/selfuncs.h"
97 #include "utils/spccache.h"
98 #include "utils/tuplesort.h"
99 
100 
101 #define LOG2(x) (log(x) / 0.693147180559945)
102 
103 /*
104  * Append and MergeAppend nodes are less expensive than some other operations
105  * which use cpu_tuple_cost; instead of adding a separate GUC, estimate the
106  * per-tuple cost as cpu_tuple_cost multiplied by this value.
107  */
108 #define APPEND_CPU_COST_MULTIPLIER 0.5
109 
110 
118 
120 
122 
124 
125 bool enable_seqscan = true;
126 bool enable_indexscan = true;
128 bool enable_bitmapscan = true;
129 bool enable_tidscan = true;
130 bool enable_sort = true;
131 bool enable_hashagg = true;
132 bool enable_nestloop = true;
133 bool enable_material = true;
134 bool enable_mergejoin = true;
135 bool enable_hashjoin = true;
136 bool enable_gathermerge = true;
142 
143 typedef struct
144 {
148 
149 static List *extract_nonindex_conditions(List *qual_clauses, List *indexquals);
151  RestrictInfo *rinfo,
152  PathKey *pathkey);
153 static void cost_rescan(PlannerInfo *root, Path *path,
154  Cost *rescan_startup_cost, Cost *rescan_total_cost);
155 static bool cost_qual_eval_walker(Node *node, cost_qual_eval_context *context);
156 static void get_restriction_qual_cost(PlannerInfo *root, RelOptInfo *baserel,
157  ParamPathInfo *param_info,
158  QualCost *qpqual_cost);
159 static bool has_indexed_join_quals(NestPath *joinpath);
160 static double approx_tuple_count(PlannerInfo *root, JoinPath *path,
161  List *quals);
162 static double calc_joinrel_size_estimate(PlannerInfo *root,
163  RelOptInfo *joinrel,
164  RelOptInfo *outer_rel,
165  RelOptInfo *inner_rel,
166  double outer_rows,
167  double inner_rows,
168  SpecialJoinInfo *sjinfo,
169  List *restrictlist);
171  Relids outer_relids,
172  Relids inner_relids,
173  SpecialJoinInfo *sjinfo,
174  List **restrictlist);
175 static Cost append_nonpartial_cost(List *subpaths, int numpaths,
176  int parallel_workers);
177 static void set_rel_width(PlannerInfo *root, RelOptInfo *rel);
178 static double relation_byte_size(double tuples, int width);
179 static double page_size(double tuples, int width);
180 static double get_parallel_divisor(Path *path);
181 
182 
183 /*
184  * clamp_row_est
185  * Force a row-count estimate to a sane value.
186  */
187 double
188 clamp_row_est(double nrows)
189 {
190  /*
191  * Force estimate to be at least one row, to make explain output look
192  * better and to avoid possible divide-by-zero when interpolating costs.
193  * Make it an integer, too.
194  */
195  if (nrows <= 1.0)
196  nrows = 1.0;
197  else
198  nrows = rint(nrows);
199 
200  return nrows;
201 }
202 
203 
204 /*
205  * cost_seqscan
206  * Determines and returns the cost of scanning a relation sequentially.
207  *
208  * 'baserel' is the relation to be scanned
209  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
210  */
211 void
213  RelOptInfo *baserel, ParamPathInfo *param_info)
214 {
215  Cost startup_cost = 0;
216  Cost cpu_run_cost;
217  Cost disk_run_cost;
218  double spc_seq_page_cost;
219  QualCost qpqual_cost;
220  Cost cpu_per_tuple;
221 
222  /* Should only be applied to base relations */
223  Assert(baserel->relid > 0);
224  Assert(baserel->rtekind == RTE_RELATION);
225 
226  /* Mark the path with the correct row estimate */
227  if (param_info)
228  path->rows = param_info->ppi_rows;
229  else
230  path->rows = baserel->rows;
231 
232  if (!enable_seqscan)
233  startup_cost += disable_cost;
234 
235  /* fetch estimated page cost for tablespace containing table */
237  NULL,
238  &spc_seq_page_cost);
239 
240  /*
241  * disk costs
242  */
243  disk_run_cost = spc_seq_page_cost * baserel->pages;
244 
245  /* CPU costs */
246  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
247 
248  startup_cost += qpqual_cost.startup;
249  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
250  cpu_run_cost = cpu_per_tuple * baserel->tuples;
251  /* tlist eval costs are paid per output row, not per tuple scanned */
252  startup_cost += path->pathtarget->cost.startup;
253  cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
254 
255  /* Adjust costing for parallelism, if used. */
256  if (path->parallel_workers > 0)
257  {
258  double parallel_divisor = get_parallel_divisor(path);
259 
260  /* The CPU cost is divided among all the workers. */
261  cpu_run_cost /= parallel_divisor;
262 
263  /*
264  * It may be possible to amortize some of the I/O cost, but probably
265  * not very much, because most operating systems already do aggressive
266  * prefetching. For now, we assume that the disk run cost can't be
267  * amortized at all.
268  */
269 
270  /*
271  * In the case of a parallel plan, the row count needs to represent
272  * the number of tuples processed per worker.
273  */
274  path->rows = clamp_row_est(path->rows / parallel_divisor);
275  }
276 
277  path->startup_cost = startup_cost;
278  path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
279 }
280 
281 /*
282  * cost_samplescan
283  * Determines and returns the cost of scanning a relation using sampling.
284  *
285  * 'baserel' is the relation to be scanned
286  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
287  */
288 void
290  RelOptInfo *baserel, ParamPathInfo *param_info)
291 {
292  Cost startup_cost = 0;
293  Cost run_cost = 0;
294  RangeTblEntry *rte;
295  TableSampleClause *tsc;
296  TsmRoutine *tsm;
297  double spc_seq_page_cost,
298  spc_random_page_cost,
299  spc_page_cost;
300  QualCost qpqual_cost;
301  Cost cpu_per_tuple;
302 
303  /* Should only be applied to base relations with tablesample clauses */
304  Assert(baserel->relid > 0);
305  rte = planner_rt_fetch(baserel->relid, root);
306  Assert(rte->rtekind == RTE_RELATION);
307  tsc = rte->tablesample;
308  Assert(tsc != NULL);
309  tsm = GetTsmRoutine(tsc->tsmhandler);
310 
311  /* Mark the path with the correct row estimate */
312  if (param_info)
313  path->rows = param_info->ppi_rows;
314  else
315  path->rows = baserel->rows;
316 
317  /* fetch estimated page cost for tablespace containing table */
319  &spc_random_page_cost,
320  &spc_seq_page_cost);
321 
322  /* if NextSampleBlock is used, assume random access, else sequential */
323  spc_page_cost = (tsm->NextSampleBlock != NULL) ?
324  spc_random_page_cost : spc_seq_page_cost;
325 
326  /*
327  * disk costs (recall that baserel->pages has already been set to the
328  * number of pages the sampling method will visit)
329  */
330  run_cost += spc_page_cost * baserel->pages;
331 
332  /*
333  * CPU costs (recall that baserel->tuples has already been set to the
334  * number of tuples the sampling method will select). Note that we ignore
335  * execution cost of the TABLESAMPLE parameter expressions; they will be
336  * evaluated only once per scan, and in most usages they'll likely be
337  * simple constants anyway. We also don't charge anything for the
338  * calculations the sampling method might do internally.
339  */
340  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
341 
342  startup_cost += qpqual_cost.startup;
343  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
344  run_cost += cpu_per_tuple * baserel->tuples;
345  /* tlist eval costs are paid per output row, not per tuple scanned */
346  startup_cost += path->pathtarget->cost.startup;
347  run_cost += path->pathtarget->cost.per_tuple * path->rows;
348 
349  path->startup_cost = startup_cost;
350  path->total_cost = startup_cost + run_cost;
351 }
352 
353 /*
354  * cost_gather
355  * Determines and returns the cost of gather path.
356  *
357  * 'rel' is the relation to be operated upon
358  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
359  * 'rows' may be used to point to a row estimate; if non-NULL, it overrides
360  * both 'rel' and 'param_info'. This is useful when the path doesn't exactly
361  * correspond to any particular RelOptInfo.
362  */
363 void
365  RelOptInfo *rel, ParamPathInfo *param_info,
366  double *rows)
367 {
368  Cost startup_cost = 0;
369  Cost run_cost = 0;
370 
371  /* Mark the path with the correct row estimate */
372  if (rows)
373  path->path.rows = *rows;
374  else if (param_info)
375  path->path.rows = param_info->ppi_rows;
376  else
377  path->path.rows = rel->rows;
378 
379  startup_cost = path->subpath->startup_cost;
380 
381  run_cost = path->subpath->total_cost - path->subpath->startup_cost;
382 
383  /* Parallel setup and communication cost. */
384  startup_cost += parallel_setup_cost;
385  run_cost += parallel_tuple_cost * path->path.rows;
386 
387  path->path.startup_cost = startup_cost;
388  path->path.total_cost = (startup_cost + run_cost);
389 }
390 
391 /*
392  * cost_gather_merge
393  * Determines and returns the cost of gather merge path.
394  *
395  * GatherMerge merges several pre-sorted input streams, using a heap that at
396  * any given instant holds the next tuple from each stream. If there are N
397  * streams, we need about N*log2(N) tuple comparisons to construct the heap at
398  * startup, and then for each output tuple, about log2(N) comparisons to
399  * replace the top heap entry with the next tuple from the same stream.
400  */
401 void
403  RelOptInfo *rel, ParamPathInfo *param_info,
404  Cost input_startup_cost, Cost input_total_cost,
405  double *rows)
406 {
407  Cost startup_cost = 0;
408  Cost run_cost = 0;
409  Cost comparison_cost;
410  double N;
411  double logN;
412 
413  /* Mark the path with the correct row estimate */
414  if (rows)
415  path->path.rows = *rows;
416  else if (param_info)
417  path->path.rows = param_info->ppi_rows;
418  else
419  path->path.rows = rel->rows;
420 
421  if (!enable_gathermerge)
422  startup_cost += disable_cost;
423 
424  /*
425  * Add one to the number of workers to account for the leader. This might
426  * be overgenerous since the leader will do less work than other workers
427  * in typical cases, but we'll go with it for now.
428  */
429  Assert(path->num_workers > 0);
430  N = (double) path->num_workers + 1;
431  logN = LOG2(N);
432 
433  /* Assumed cost per tuple comparison */
434  comparison_cost = 2.0 * cpu_operator_cost;
435 
436  /* Heap creation cost */
437  startup_cost += comparison_cost * N * logN;
438 
439  /* Per-tuple heap maintenance cost */
440  run_cost += path->path.rows * comparison_cost * logN;
441 
442  /* small cost for heap management, like cost_merge_append */
443  run_cost += cpu_operator_cost * path->path.rows;
444 
445  /*
446  * Parallel setup and communication cost. Since Gather Merge, unlike
447  * Gather, requires us to block until a tuple is available from every
448  * worker, we bump the IPC cost up a little bit as compared with Gather.
449  * For lack of a better idea, charge an extra 5%.
450  */
451  startup_cost += parallel_setup_cost;
452  run_cost += parallel_tuple_cost * path->path.rows * 1.05;
453 
454  path->path.startup_cost = startup_cost + input_startup_cost;
455  path->path.total_cost = (startup_cost + run_cost + input_total_cost);
456 }
457 
458 /*
459  * cost_index
460  * Determines and returns the cost of scanning a relation using an index.
461  *
462  * 'path' describes the indexscan under consideration, and is complete
463  * except for the fields to be set by this routine
464  * 'loop_count' is the number of repetitions of the indexscan to factor into
465  * estimates of caching behavior
466  *
467  * In addition to rows, startup_cost and total_cost, cost_index() sets the
468  * path's indextotalcost and indexselectivity fields. These values will be
469  * needed if the IndexPath is used in a BitmapIndexScan.
470  *
471  * NOTE: path->indexquals must contain only clauses usable as index
472  * restrictions. Any additional quals evaluated as qpquals may reduce the
473  * number of returned tuples, but they won't reduce the number of tuples
474  * we have to fetch from the table, so they don't reduce the scan cost.
475  */
476 void
477 cost_index(IndexPath *path, PlannerInfo *root, double loop_count,
478  bool partial_path)
479 {
480  IndexOptInfo *index = path->indexinfo;
481  RelOptInfo *baserel = index->rel;
482  bool indexonly = (path->path.pathtype == T_IndexOnlyScan);
483  amcostestimate_function amcostestimate;
484  List *qpquals;
485  Cost startup_cost = 0;
486  Cost run_cost = 0;
487  Cost cpu_run_cost = 0;
488  Cost indexStartupCost;
489  Cost indexTotalCost;
490  Selectivity indexSelectivity;
491  double indexCorrelation,
492  csquared;
493  double spc_seq_page_cost,
494  spc_random_page_cost;
495  Cost min_IO_cost,
496  max_IO_cost;
497  QualCost qpqual_cost;
498  Cost cpu_per_tuple;
499  double tuples_fetched;
500  double pages_fetched;
501  double rand_heap_pages;
502  double index_pages;
503 
504  /* Should only be applied to base relations */
505  Assert(IsA(baserel, RelOptInfo) &&
506  IsA(index, IndexOptInfo));
507  Assert(baserel->relid > 0);
508  Assert(baserel->rtekind == RTE_RELATION);
509 
510  /*
511  * Mark the path with the correct row estimate, and identify which quals
512  * will need to be enforced as qpquals. We need not check any quals that
513  * are implied by the index's predicate, so we can use indrestrictinfo not
514  * baserestrictinfo as the list of relevant restriction clauses for the
515  * rel.
516  */
517  if (path->path.param_info)
518  {
519  path->path.rows = path->path.param_info->ppi_rows;
520  /* qpquals come from the rel's restriction clauses and ppi_clauses */
521  qpquals = list_concat(
523  path->indexquals),
525  path->indexquals));
526  }
527  else
528  {
529  path->path.rows = baserel->rows;
530  /* qpquals come from just the rel's restriction clauses */
532  path->indexquals);
533  }
534 
535  if (!enable_indexscan)
536  startup_cost += disable_cost;
537  /* we don't need to check enable_indexonlyscan; indxpath.c does that */
538 
539  /*
540  * Call index-access-method-specific code to estimate the processing cost
541  * for scanning the index, as well as the selectivity of the index (ie,
542  * the fraction of main-table tuples we will have to retrieve) and its
543  * correlation to the main-table tuple order. We need a cast here because
544  * relation.h uses a weak function type to avoid including amapi.h.
545  */
546  amcostestimate = (amcostestimate_function) index->amcostestimate;
547  amcostestimate(root, path, loop_count,
548  &indexStartupCost, &indexTotalCost,
549  &indexSelectivity, &indexCorrelation,
550  &index_pages);
551 
552  /*
553  * Save amcostestimate's results for possible use in bitmap scan planning.
554  * We don't bother to save indexStartupCost or indexCorrelation, because a
555  * bitmap scan doesn't care about either.
556  */
557  path->indextotalcost = indexTotalCost;
558  path->indexselectivity = indexSelectivity;
559 
560  /* all costs for touching index itself included here */
561  startup_cost += indexStartupCost;
562  run_cost += indexTotalCost - indexStartupCost;
563 
564  /* estimate number of main-table tuples fetched */
565  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
566 
567  /* fetch estimated page costs for tablespace containing table */
569  &spc_random_page_cost,
570  &spc_seq_page_cost);
571 
572  /*----------
573  * Estimate number of main-table pages fetched, and compute I/O cost.
574  *
575  * When the index ordering is uncorrelated with the table ordering,
576  * we use an approximation proposed by Mackert and Lohman (see
577  * index_pages_fetched() for details) to compute the number of pages
578  * fetched, and then charge spc_random_page_cost per page fetched.
579  *
580  * When the index ordering is exactly correlated with the table ordering
581  * (just after a CLUSTER, for example), the number of pages fetched should
582  * be exactly selectivity * table_size. What's more, all but the first
583  * will be sequential fetches, not the random fetches that occur in the
584  * uncorrelated case. So if the number of pages is more than 1, we
585  * ought to charge
586  * spc_random_page_cost + (pages_fetched - 1) * spc_seq_page_cost
587  * For partially-correlated indexes, we ought to charge somewhere between
588  * these two estimates. We currently interpolate linearly between the
589  * estimates based on the correlation squared (XXX is that appropriate?).
590  *
591  * If it's an index-only scan, then we will not need to fetch any heap
592  * pages for which the visibility map shows all tuples are visible.
593  * Hence, reduce the estimated number of heap fetches accordingly.
594  * We use the measured fraction of the entire heap that is all-visible,
595  * which might not be particularly relevant to the subset of the heap
596  * that this query will fetch; but it's not clear how to do better.
597  *----------
598  */
599  if (loop_count > 1)
600  {
601  /*
602  * For repeated indexscans, the appropriate estimate for the
603  * uncorrelated case is to scale up the number of tuples fetched in
604  * the Mackert and Lohman formula by the number of scans, so that we
605  * estimate the number of pages fetched by all the scans; then
606  * pro-rate the costs for one scan. In this case we assume all the
607  * fetches are random accesses.
608  */
609  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
610  baserel->pages,
611  (double) index->pages,
612  root);
613 
614  if (indexonly)
615  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
616 
617  rand_heap_pages = pages_fetched;
618 
619  max_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
620 
621  /*
622  * In the perfectly correlated case, the number of pages touched by
623  * each scan is selectivity * table_size, and we can use the Mackert
624  * and Lohman formula at the page level to estimate how much work is
625  * saved by caching across scans. We still assume all the fetches are
626  * random, though, which is an overestimate that's hard to correct for
627  * without double-counting the cache effects. (But in most cases
628  * where such a plan is actually interesting, only one page would get
629  * fetched per scan anyway, so it shouldn't matter much.)
630  */
631  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
632 
633  pages_fetched = index_pages_fetched(pages_fetched * loop_count,
634  baserel->pages,
635  (double) index->pages,
636  root);
637 
638  if (indexonly)
639  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
640 
641  min_IO_cost = (pages_fetched * spc_random_page_cost) / loop_count;
642  }
643  else
644  {
645  /*
646  * Normal case: apply the Mackert and Lohman formula, and then
647  * interpolate between that and the correlation-derived result.
648  */
649  pages_fetched = index_pages_fetched(tuples_fetched,
650  baserel->pages,
651  (double) index->pages,
652  root);
653 
654  if (indexonly)
655  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
656 
657  rand_heap_pages = pages_fetched;
658 
659  /* max_IO_cost is for the perfectly uncorrelated case (csquared=0) */
660  max_IO_cost = pages_fetched * spc_random_page_cost;
661 
662  /* min_IO_cost is for the perfectly correlated case (csquared=1) */
663  pages_fetched = ceil(indexSelectivity * (double) baserel->pages);
664 
665  if (indexonly)
666  pages_fetched = ceil(pages_fetched * (1.0 - baserel->allvisfrac));
667 
668  if (pages_fetched > 0)
669  {
670  min_IO_cost = spc_random_page_cost;
671  if (pages_fetched > 1)
672  min_IO_cost += (pages_fetched - 1) * spc_seq_page_cost;
673  }
674  else
675  min_IO_cost = 0;
676  }
677 
678  if (partial_path)
679  {
680  /*
681  * For index only scans compute workers based on number of index pages
682  * fetched; the number of heap pages we fetch might be so small as to
683  * effectively rule out parallelism, which we don't want to do.
684  */
685  if (indexonly)
686  rand_heap_pages = -1;
687 
688  /*
689  * Estimate the number of parallel workers required to scan index. Use
690  * the number of heap pages computed considering heap fetches won't be
691  * sequential as for parallel scans the pages are accessed in random
692  * order.
693  */
695  rand_heap_pages,
696  index_pages,
698 
699  /*
700  * Fall out if workers can't be assigned for parallel scan, because in
701  * such a case this path will be rejected. So there is no benefit in
702  * doing extra computation.
703  */
704  if (path->path.parallel_workers <= 0)
705  return;
706 
707  path->path.parallel_aware = true;
708  }
709 
710  /*
711  * Now interpolate based on estimated index order correlation to get total
712  * disk I/O cost for main table accesses.
713  */
714  csquared = indexCorrelation * indexCorrelation;
715 
716  run_cost += max_IO_cost + csquared * (min_IO_cost - max_IO_cost);
717 
718  /*
719  * Estimate CPU costs per tuple.
720  *
721  * What we want here is cpu_tuple_cost plus the evaluation costs of any
722  * qual clauses that we have to evaluate as qpquals.
723  */
724  cost_qual_eval(&qpqual_cost, qpquals, root);
725 
726  startup_cost += qpqual_cost.startup;
727  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
728 
729  cpu_run_cost += cpu_per_tuple * tuples_fetched;
730 
731  /* tlist eval costs are paid per output row, not per tuple scanned */
732  startup_cost += path->path.pathtarget->cost.startup;
733  cpu_run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
734 
735  /* Adjust costing for parallelism, if used. */
736  if (path->path.parallel_workers > 0)
737  {
738  double parallel_divisor = get_parallel_divisor(&path->path);
739 
740  path->path.rows = clamp_row_est(path->path.rows / parallel_divisor);
741 
742  /* The CPU cost is divided among all the workers. */
743  cpu_run_cost /= parallel_divisor;
744  }
745 
746  run_cost += cpu_run_cost;
747 
748  path->path.startup_cost = startup_cost;
749  path->path.total_cost = startup_cost + run_cost;
750 }
751 
752 /*
753  * extract_nonindex_conditions
754  *
755  * Given a list of quals to be enforced in an indexscan, extract the ones that
756  * will have to be applied as qpquals (ie, the index machinery won't handle
757  * them). The actual rules for this appear in create_indexscan_plan() in
758  * createplan.c, but the full rules are fairly expensive and we don't want to
759  * go to that much effort for index paths that don't get selected for the
760  * final plan. So we approximate it as quals that don't appear directly in
761  * indexquals and also are not redundant children of the same EquivalenceClass
762  * as some indexqual. This method neglects some infrequently-relevant
763  * considerations, specifically clauses that needn't be checked because they
764  * are implied by an indexqual. It does not seem worth the cycles to try to
765  * factor that in at this stage, even though createplan.c will take pains to
766  * remove such unnecessary clauses from the qpquals list if this path is
767  * selected for use.
768  */
769 static List *
770 extract_nonindex_conditions(List *qual_clauses, List *indexquals)
771 {
772  List *result = NIL;
773  ListCell *lc;
774 
775  foreach(lc, qual_clauses)
776  {
777  RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
778 
779  if (rinfo->pseudoconstant)
780  continue; /* we may drop pseudoconstants here */
781  if (list_member_ptr(indexquals, rinfo))
782  continue; /* simple duplicate */
783  if (is_redundant_derived_clause(rinfo, indexquals))
784  continue; /* derived from same EquivalenceClass */
785  /* ... skip the predicate proof attempt createplan.c will try ... */
786  result = lappend(result, rinfo);
787  }
788  return result;
789 }
790 
791 /*
792  * index_pages_fetched
793  * Estimate the number of pages actually fetched after accounting for
794  * cache effects.
795  *
796  * We use an approximation proposed by Mackert and Lohman, "Index Scans
797  * Using a Finite LRU Buffer: A Validated I/O Model", ACM Transactions
798  * on Database Systems, Vol. 14, No. 3, September 1989, Pages 401-424.
799  * The Mackert and Lohman approximation is that the number of pages
800  * fetched is
801  * PF =
802  * min(2TNs/(2T+Ns), T) when T <= b
803  * 2TNs/(2T+Ns) when T > b and Ns <= 2Tb/(2T-b)
804  * b + (Ns - 2Tb/(2T-b))*(T-b)/T when T > b and Ns > 2Tb/(2T-b)
805  * where
806  * T = # pages in table
807  * N = # tuples in table
808  * s = selectivity = fraction of table to be scanned
809  * b = # buffer pages available (we include kernel space here)
810  *
811  * We assume that effective_cache_size is the total number of buffer pages
812  * available for the whole query, and pro-rate that space across all the
813  * tables in the query and the index currently under consideration. (This
814  * ignores space needed for other indexes used by the query, but since we
815  * don't know which indexes will get used, we can't estimate that very well;
816  * and in any case counting all the tables may well be an overestimate, since
817  * depending on the join plan not all the tables may be scanned concurrently.)
818  *
819  * The product Ns is the number of tuples fetched; we pass in that
820  * product rather than calculating it here. "pages" is the number of pages
821  * in the object under consideration (either an index or a table).
822  * "index_pages" is the amount to add to the total table space, which was
823  * computed for us by query_planner.
824  *
825  * Caller is expected to have ensured that tuples_fetched is greater than zero
826  * and rounded to integer (see clamp_row_est). The result will likewise be
827  * greater than zero and integral.
828  */
829 double
830 index_pages_fetched(double tuples_fetched, BlockNumber pages,
831  double index_pages, PlannerInfo *root)
832 {
833  double pages_fetched;
834  double total_pages;
835  double T,
836  b;
837 
838  /* T is # pages in table, but don't allow it to be zero */
839  T = (pages > 1) ? (double) pages : 1.0;
840 
841  /* Compute number of pages assumed to be competing for cache space */
842  total_pages = root->total_table_pages + index_pages;
843  total_pages = Max(total_pages, 1.0);
844  Assert(T <= total_pages);
845 
846  /* b is pro-rated share of effective_cache_size */
847  b = (double) effective_cache_size * T / total_pages;
848 
849  /* force it positive and integral */
850  if (b <= 1.0)
851  b = 1.0;
852  else
853  b = ceil(b);
854 
855  /* This part is the Mackert and Lohman formula */
856  if (T <= b)
857  {
858  pages_fetched =
859  (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
860  if (pages_fetched >= T)
861  pages_fetched = T;
862  else
863  pages_fetched = ceil(pages_fetched);
864  }
865  else
866  {
867  double lim;
868 
869  lim = (2.0 * T * b) / (2.0 * T - b);
870  if (tuples_fetched <= lim)
871  {
872  pages_fetched =
873  (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
874  }
875  else
876  {
877  pages_fetched =
878  b + (tuples_fetched - lim) * (T - b) / T;
879  }
880  pages_fetched = ceil(pages_fetched);
881  }
882  return pages_fetched;
883 }
884 
885 /*
886  * get_indexpath_pages
887  * Determine the total size of the indexes used in a bitmap index path.
888  *
889  * Note: if the same index is used more than once in a bitmap tree, we will
890  * count it multiple times, which perhaps is the wrong thing ... but it's
891  * not completely clear, and detecting duplicates is difficult, so ignore it
892  * for now.
893  */
894 static double
896 {
897  double result = 0;
898  ListCell *l;
899 
900  if (IsA(bitmapqual, BitmapAndPath))
901  {
902  BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
903 
904  foreach(l, apath->bitmapquals)
905  {
906  result += get_indexpath_pages((Path *) lfirst(l));
907  }
908  }
909  else if (IsA(bitmapqual, BitmapOrPath))
910  {
911  BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
912 
913  foreach(l, opath->bitmapquals)
914  {
915  result += get_indexpath_pages((Path *) lfirst(l));
916  }
917  }
918  else if (IsA(bitmapqual, IndexPath))
919  {
920  IndexPath *ipath = (IndexPath *) bitmapqual;
921 
922  result = (double) ipath->indexinfo->pages;
923  }
924  else
925  elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
926 
927  return result;
928 }
929 
930 /*
931  * cost_bitmap_heap_scan
932  * Determines and returns the cost of scanning a relation using a bitmap
933  * index-then-heap plan.
934  *
935  * 'baserel' is the relation to be scanned
936  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
937  * 'bitmapqual' is a tree of IndexPaths, BitmapAndPaths, and BitmapOrPaths
938  * 'loop_count' is the number of repetitions of the indexscan to factor into
939  * estimates of caching behavior
940  *
941  * Note: the component IndexPaths in bitmapqual should have been costed
942  * using the same loop_count.
943  */
944 void
946  ParamPathInfo *param_info,
947  Path *bitmapqual, double loop_count)
948 {
949  Cost startup_cost = 0;
950  Cost run_cost = 0;
951  Cost indexTotalCost;
952  QualCost qpqual_cost;
953  Cost cpu_per_tuple;
954  Cost cost_per_page;
955  Cost cpu_run_cost;
956  double tuples_fetched;
957  double pages_fetched;
958  double spc_seq_page_cost,
959  spc_random_page_cost;
960  double T;
961 
962  /* Should only be applied to base relations */
963  Assert(IsA(baserel, RelOptInfo));
964  Assert(baserel->relid > 0);
965  Assert(baserel->rtekind == RTE_RELATION);
966 
967  /* Mark the path with the correct row estimate */
968  if (param_info)
969  path->rows = param_info->ppi_rows;
970  else
971  path->rows = baserel->rows;
972 
973  if (!enable_bitmapscan)
974  startup_cost += disable_cost;
975 
976  pages_fetched = compute_bitmap_pages(root, baserel, bitmapqual,
977  loop_count, &indexTotalCost,
978  &tuples_fetched);
979 
980  startup_cost += indexTotalCost;
981  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
982 
983  /* Fetch estimated page costs for tablespace containing table. */
985  &spc_random_page_cost,
986  &spc_seq_page_cost);
987 
988  /*
989  * For small numbers of pages we should charge spc_random_page_cost
990  * apiece, while if nearly all the table's pages are being read, it's more
991  * appropriate to charge spc_seq_page_cost apiece. The effect is
992  * nonlinear, too. For lack of a better idea, interpolate like this to
993  * determine the cost per page.
994  */
995  if (pages_fetched >= 2.0)
996  cost_per_page = spc_random_page_cost -
997  (spc_random_page_cost - spc_seq_page_cost)
998  * sqrt(pages_fetched / T);
999  else
1000  cost_per_page = spc_random_page_cost;
1001 
1002  run_cost += pages_fetched * cost_per_page;
1003 
1004  /*
1005  * Estimate CPU costs per tuple.
1006  *
1007  * Often the indexquals don't need to be rechecked at each tuple ... but
1008  * not always, especially not if there are enough tuples involved that the
1009  * bitmaps become lossy. For the moment, just assume they will be
1010  * rechecked always. This means we charge the full freight for all the
1011  * scan clauses.
1012  */
1013  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1014 
1015  startup_cost += qpqual_cost.startup;
1016  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1017  cpu_run_cost = cpu_per_tuple * tuples_fetched;
1018 
1019  /* Adjust costing for parallelism, if used. */
1020  if (path->parallel_workers > 0)
1021  {
1022  double parallel_divisor = get_parallel_divisor(path);
1023 
1024  /* The CPU cost is divided among all the workers. */
1025  cpu_run_cost /= parallel_divisor;
1026 
1027  path->rows = clamp_row_est(path->rows / parallel_divisor);
1028  }
1029 
1030 
1031  run_cost += cpu_run_cost;
1032 
1033  /* tlist eval costs are paid per output row, not per tuple scanned */
1034  startup_cost += path->pathtarget->cost.startup;
1035  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1036 
1037  path->startup_cost = startup_cost;
1038  path->total_cost = startup_cost + run_cost;
1039 }
1040 
1041 /*
1042  * cost_bitmap_tree_node
1043  * Extract cost and selectivity from a bitmap tree node (index/and/or)
1044  */
1045 void
1047 {
1048  if (IsA(path, IndexPath))
1049  {
1050  *cost = ((IndexPath *) path)->indextotalcost;
1051  *selec = ((IndexPath *) path)->indexselectivity;
1052 
1053  /*
1054  * Charge a small amount per retrieved tuple to reflect the costs of
1055  * manipulating the bitmap. This is mostly to make sure that a bitmap
1056  * scan doesn't look to be the same cost as an indexscan to retrieve a
1057  * single tuple.
1058  */
1059  *cost += 0.1 * cpu_operator_cost * path->rows;
1060  }
1061  else if (IsA(path, BitmapAndPath))
1062  {
1063  *cost = path->total_cost;
1064  *selec = ((BitmapAndPath *) path)->bitmapselectivity;
1065  }
1066  else if (IsA(path, BitmapOrPath))
1067  {
1068  *cost = path->total_cost;
1069  *selec = ((BitmapOrPath *) path)->bitmapselectivity;
1070  }
1071  else
1072  {
1073  elog(ERROR, "unrecognized node type: %d", nodeTag(path));
1074  *cost = *selec = 0; /* keep compiler quiet */
1075  }
1076 }
1077 
1078 /*
1079  * cost_bitmap_and_node
1080  * Estimate the cost of a BitmapAnd node
1081  *
1082  * Note that this considers only the costs of index scanning and bitmap
1083  * creation, not the eventual heap access. In that sense the object isn't
1084  * truly a Path, but it has enough path-like properties (costs in particular)
1085  * to warrant treating it as one. We don't bother to set the path rows field,
1086  * however.
1087  */
1088 void
1090 {
1091  Cost totalCost;
1092  Selectivity selec;
1093  ListCell *l;
1094 
1095  /*
1096  * We estimate AND selectivity on the assumption that the inputs are
1097  * independent. This is probably often wrong, but we don't have the info
1098  * to do better.
1099  *
1100  * The runtime cost of the BitmapAnd itself is estimated at 100x
1101  * cpu_operator_cost for each tbm_intersect needed. Probably too small,
1102  * definitely too simplistic?
1103  */
1104  totalCost = 0.0;
1105  selec = 1.0;
1106  foreach(l, path->bitmapquals)
1107  {
1108  Path *subpath = (Path *) lfirst(l);
1109  Cost subCost;
1110  Selectivity subselec;
1111 
1112  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1113 
1114  selec *= subselec;
1115 
1116  totalCost += subCost;
1117  if (l != list_head(path->bitmapquals))
1118  totalCost += 100.0 * cpu_operator_cost;
1119  }
1120  path->bitmapselectivity = selec;
1121  path->path.rows = 0; /* per above, not used */
1122  path->path.startup_cost = totalCost;
1123  path->path.total_cost = totalCost;
1124 }
1125 
1126 /*
1127  * cost_bitmap_or_node
1128  * Estimate the cost of a BitmapOr node
1129  *
1130  * See comments for cost_bitmap_and_node.
1131  */
1132 void
1134 {
1135  Cost totalCost;
1136  Selectivity selec;
1137  ListCell *l;
1138 
1139  /*
1140  * We estimate OR selectivity on the assumption that the inputs are
1141  * non-overlapping, since that's often the case in "x IN (list)" type
1142  * situations. Of course, we clamp to 1.0 at the end.
1143  *
1144  * The runtime cost of the BitmapOr itself is estimated at 100x
1145  * cpu_operator_cost for each tbm_union needed. Probably too small,
1146  * definitely too simplistic? We are aware that the tbm_unions are
1147  * optimized out when the inputs are BitmapIndexScans.
1148  */
1149  totalCost = 0.0;
1150  selec = 0.0;
1151  foreach(l, path->bitmapquals)
1152  {
1153  Path *subpath = (Path *) lfirst(l);
1154  Cost subCost;
1155  Selectivity subselec;
1156 
1157  cost_bitmap_tree_node(subpath, &subCost, &subselec);
1158 
1159  selec += subselec;
1160 
1161  totalCost += subCost;
1162  if (l != list_head(path->bitmapquals) &&
1163  !IsA(subpath, IndexPath))
1164  totalCost += 100.0 * cpu_operator_cost;
1165  }
1166  path->bitmapselectivity = Min(selec, 1.0);
1167  path->path.rows = 0; /* per above, not used */
1168  path->path.startup_cost = totalCost;
1169  path->path.total_cost = totalCost;
1170 }
1171 
1172 /*
1173  * cost_tidscan
1174  * Determines and returns the cost of scanning a relation using TIDs.
1175  *
1176  * 'baserel' is the relation to be scanned
1177  * 'tidquals' is the list of TID-checkable quals
1178  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1179  */
1180 void
1182  RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
1183 {
1184  Cost startup_cost = 0;
1185  Cost run_cost = 0;
1186  bool isCurrentOf = false;
1187  QualCost qpqual_cost;
1188  Cost cpu_per_tuple;
1189  QualCost tid_qual_cost;
1190  int ntuples;
1191  ListCell *l;
1192  double spc_random_page_cost;
1193 
1194  /* Should only be applied to base relations */
1195  Assert(baserel->relid > 0);
1196  Assert(baserel->rtekind == RTE_RELATION);
1197 
1198  /* Mark the path with the correct row estimate */
1199  if (param_info)
1200  path->rows = param_info->ppi_rows;
1201  else
1202  path->rows = baserel->rows;
1203 
1204  /* Count how many tuples we expect to retrieve */
1205  ntuples = 0;
1206  foreach(l, tidquals)
1207  {
1208  if (IsA(lfirst(l), ScalarArrayOpExpr))
1209  {
1210  /* Each element of the array yields 1 tuple */
1212  Node *arraynode = (Node *) lsecond(saop->args);
1213 
1214  ntuples += estimate_array_length(arraynode);
1215  }
1216  else if (IsA(lfirst(l), CurrentOfExpr))
1217  {
1218  /* CURRENT OF yields 1 tuple */
1219  isCurrentOf = true;
1220  ntuples++;
1221  }
1222  else
1223  {
1224  /* It's just CTID = something, count 1 tuple */
1225  ntuples++;
1226  }
1227  }
1228 
1229  /*
1230  * We must force TID scan for WHERE CURRENT OF, because only nodeTidscan.c
1231  * understands how to do it correctly. Therefore, honor enable_tidscan
1232  * only when CURRENT OF isn't present. Also note that cost_qual_eval
1233  * counts a CurrentOfExpr as having startup cost disable_cost, which we
1234  * subtract off here; that's to prevent other plan types such as seqscan
1235  * from winning.
1236  */
1237  if (isCurrentOf)
1238  {
1240  startup_cost -= disable_cost;
1241  }
1242  else if (!enable_tidscan)
1243  startup_cost += disable_cost;
1244 
1245  /*
1246  * The TID qual expressions will be computed once, any other baserestrict
1247  * quals once per retrieved tuple.
1248  */
1249  cost_qual_eval(&tid_qual_cost, tidquals, root);
1250 
1251  /* fetch estimated page cost for tablespace containing table */
1253  &spc_random_page_cost,
1254  NULL);
1255 
1256  /* disk costs --- assume each tuple on a different page */
1257  run_cost += spc_random_page_cost * ntuples;
1258 
1259  /* Add scanning CPU costs */
1260  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1261 
1262  /* XXX currently we assume TID quals are a subset of qpquals */
1263  startup_cost += qpqual_cost.startup + tid_qual_cost.per_tuple;
1264  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple -
1265  tid_qual_cost.per_tuple;
1266  run_cost += cpu_per_tuple * ntuples;
1267 
1268  /* tlist eval costs are paid per output row, not per tuple scanned */
1269  startup_cost += path->pathtarget->cost.startup;
1270  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1271 
1272  path->startup_cost = startup_cost;
1273  path->total_cost = startup_cost + run_cost;
1274 }
1275 
1276 /*
1277  * cost_subqueryscan
1278  * Determines and returns the cost of scanning a subquery RTE.
1279  *
1280  * 'baserel' is the relation to be scanned
1281  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1282  */
1283 void
1285  RelOptInfo *baserel, ParamPathInfo *param_info)
1286 {
1287  Cost startup_cost;
1288  Cost run_cost;
1289  QualCost qpqual_cost;
1290  Cost cpu_per_tuple;
1291 
1292  /* Should only be applied to base relations that are subqueries */
1293  Assert(baserel->relid > 0);
1294  Assert(baserel->rtekind == RTE_SUBQUERY);
1295 
1296  /* Mark the path with the correct row estimate */
1297  if (param_info)
1298  path->path.rows = param_info->ppi_rows;
1299  else
1300  path->path.rows = baserel->rows;
1301 
1302  /*
1303  * Cost of path is cost of evaluating the subplan, plus cost of evaluating
1304  * any restriction clauses and tlist that will be attached to the
1305  * SubqueryScan node, plus cpu_tuple_cost to account for selection and
1306  * projection overhead.
1307  */
1308  path->path.startup_cost = path->subpath->startup_cost;
1309  path->path.total_cost = path->subpath->total_cost;
1310 
1311  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1312 
1313  startup_cost = qpqual_cost.startup;
1314  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1315  run_cost = cpu_per_tuple * baserel->tuples;
1316 
1317  /* tlist eval costs are paid per output row, not per tuple scanned */
1318  startup_cost += path->path.pathtarget->cost.startup;
1319  run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
1320 
1321  path->path.startup_cost += startup_cost;
1322  path->path.total_cost += startup_cost + run_cost;
1323 }
1324 
1325 /*
1326  * cost_functionscan
1327  * Determines and returns the cost of scanning a function RTE.
1328  *
1329  * 'baserel' is the relation to be scanned
1330  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1331  */
1332 void
1334  RelOptInfo *baserel, ParamPathInfo *param_info)
1335 {
1336  Cost startup_cost = 0;
1337  Cost run_cost = 0;
1338  QualCost qpqual_cost;
1339  Cost cpu_per_tuple;
1340  RangeTblEntry *rte;
1341  QualCost exprcost;
1342 
1343  /* Should only be applied to base relations that are functions */
1344  Assert(baserel->relid > 0);
1345  rte = planner_rt_fetch(baserel->relid, root);
1346  Assert(rte->rtekind == RTE_FUNCTION);
1347 
1348  /* Mark the path with the correct row estimate */
1349  if (param_info)
1350  path->rows = param_info->ppi_rows;
1351  else
1352  path->rows = baserel->rows;
1353 
1354  /*
1355  * Estimate costs of executing the function expression(s).
1356  *
1357  * Currently, nodeFunctionscan.c always executes the functions to
1358  * completion before returning any rows, and caches the results in a
1359  * tuplestore. So the function eval cost is all startup cost, and per-row
1360  * costs are minimal.
1361  *
1362  * XXX in principle we ought to charge tuplestore spill costs if the
1363  * number of rows is large. However, given how phony our rowcount
1364  * estimates for functions tend to be, there's not a lot of point in that
1365  * refinement right now.
1366  */
1367  cost_qual_eval_node(&exprcost, (Node *) rte->functions, root);
1368 
1369  startup_cost += exprcost.startup + exprcost.per_tuple;
1370 
1371  /* Add scanning CPU costs */
1372  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1373 
1374  startup_cost += qpqual_cost.startup;
1375  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1376  run_cost += cpu_per_tuple * baserel->tuples;
1377 
1378  /* tlist eval costs are paid per output row, not per tuple scanned */
1379  startup_cost += path->pathtarget->cost.startup;
1380  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1381 
1382  path->startup_cost = startup_cost;
1383  path->total_cost = startup_cost + run_cost;
1384 }
1385 
1386 /*
1387  * cost_tablefuncscan
1388  * Determines and returns the cost of scanning a table function.
1389  *
1390  * 'baserel' is the relation to be scanned
1391  * 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
1392  */
1393 void
1395  RelOptInfo *baserel, ParamPathInfo *param_info)
1396 {
1397  Cost startup_cost = 0;
1398  Cost run_cost = 0;
1399  QualCost qpqual_cost;
1400  Cost cpu_per_tuple;
1401  RangeTblEntry *rte;
1402  QualCost exprcost;
1403 
1404  /* Should only be applied to base relations that are functions */
1405  Assert(baserel->relid > 0);
1406  rte = planner_rt_fetch(baserel->relid, root);
1407  Assert(rte->rtekind == RTE_TABLEFUNC);
1408 
1409  /* Mark the path with the correct row estimate */
1410  if (param_info)
1411  path->rows = param_info->ppi_rows;
1412  else
1413  path->rows = baserel->rows;
1414 
1415  /*
1416  * Estimate costs of executing the table func expression(s).
1417  *
1418  * XXX in principle we ought to charge tuplestore spill costs if the
1419  * number of rows is large. However, given how phony our rowcount
1420  * estimates for tablefuncs tend to be, there's not a lot of point in that
1421  * refinement right now.
1422  */
1423  cost_qual_eval_node(&exprcost, (Node *) rte->tablefunc, root);
1424 
1425  startup_cost += exprcost.startup + exprcost.per_tuple;
1426 
1427  /* Add scanning CPU costs */
1428  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1429 
1430  startup_cost += qpqual_cost.startup;
1431  cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
1432  run_cost += cpu_per_tuple * baserel->tuples;
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_valuesscan
1444  * Determines and returns the cost of scanning a VALUES 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  */
1449 void
1451  RelOptInfo *baserel, ParamPathInfo *param_info)
1452 {
1453  Cost startup_cost = 0;
1454  Cost run_cost = 0;
1455  QualCost qpqual_cost;
1456  Cost cpu_per_tuple;
1457 
1458  /* Should only be applied to base relations that are values lists */
1459  Assert(baserel->relid > 0);
1460  Assert(baserel->rtekind == RTE_VALUES);
1461 
1462  /* Mark the path with the correct row estimate */
1463  if (param_info)
1464  path->rows = param_info->ppi_rows;
1465  else
1466  path->rows = baserel->rows;
1467 
1468  /*
1469  * For now, estimate list evaluation cost at one operator eval per list
1470  * (probably pretty bogus, but is it worth being smarter?)
1471  */
1472  cpu_per_tuple = cpu_operator_cost;
1473 
1474  /* Add scanning CPU costs */
1475  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1476 
1477  startup_cost += qpqual_cost.startup;
1478  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1479  run_cost += cpu_per_tuple * baserel->tuples;
1480 
1481  /* tlist eval costs are paid per output row, not per tuple scanned */
1482  startup_cost += path->pathtarget->cost.startup;
1483  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1484 
1485  path->startup_cost = startup_cost;
1486  path->total_cost = startup_cost + run_cost;
1487 }
1488 
1489 /*
1490  * cost_ctescan
1491  * Determines and returns the cost of scanning a CTE RTE.
1492  *
1493  * Note: this is used for both self-reference and regular CTEs; the
1494  * possible cost differences are below the threshold of what we could
1495  * estimate accurately anyway. Note that the costs of evaluating the
1496  * referenced CTE query are added into the final plan as initplan costs,
1497  * and should NOT be counted here.
1498  */
1499 void
1501  RelOptInfo *baserel, ParamPathInfo *param_info)
1502 {
1503  Cost startup_cost = 0;
1504  Cost run_cost = 0;
1505  QualCost qpqual_cost;
1506  Cost cpu_per_tuple;
1507 
1508  /* Should only be applied to base relations that are CTEs */
1509  Assert(baserel->relid > 0);
1510  Assert(baserel->rtekind == RTE_CTE);
1511 
1512  /* Mark the path with the correct row estimate */
1513  if (param_info)
1514  path->rows = param_info->ppi_rows;
1515  else
1516  path->rows = baserel->rows;
1517 
1518  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1519  cpu_per_tuple = cpu_tuple_cost;
1520 
1521  /* Add scanning CPU costs */
1522  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1523 
1524  startup_cost += qpqual_cost.startup;
1525  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1526  run_cost += cpu_per_tuple * baserel->tuples;
1527 
1528  /* tlist eval costs are paid per output row, not per tuple scanned */
1529  startup_cost += path->pathtarget->cost.startup;
1530  run_cost += path->pathtarget->cost.per_tuple * path->rows;
1531 
1532  path->startup_cost = startup_cost;
1533  path->total_cost = startup_cost + run_cost;
1534 }
1535 
1536 /*
1537  * cost_namedtuplestorescan
1538  * Determines and returns the cost of scanning a named tuplestore.
1539  */
1540 void
1542  RelOptInfo *baserel, ParamPathInfo *param_info)
1543 {
1544  Cost startup_cost = 0;
1545  Cost run_cost = 0;
1546  QualCost qpqual_cost;
1547  Cost cpu_per_tuple;
1548 
1549  /* Should only be applied to base relations that are Tuplestores */
1550  Assert(baserel->relid > 0);
1551  Assert(baserel->rtekind == RTE_NAMEDTUPLESTORE);
1552 
1553  /* Mark the path with the correct row estimate */
1554  if (param_info)
1555  path->rows = param_info->ppi_rows;
1556  else
1557  path->rows = baserel->rows;
1558 
1559  /* Charge one CPU tuple cost per row for tuplestore manipulation */
1560  cpu_per_tuple = cpu_tuple_cost;
1561 
1562  /* Add scanning CPU costs */
1563  get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
1564 
1565  startup_cost += qpqual_cost.startup;
1566  cpu_per_tuple += cpu_tuple_cost + qpqual_cost.per_tuple;
1567  run_cost += cpu_per_tuple * baserel->tuples;
1568 
1569  path->startup_cost = startup_cost;
1570  path->total_cost = startup_cost + run_cost;
1571 }
1572 
1573 /*
1574  * cost_recursive_union
1575  * Determines and returns the cost of performing a recursive union,
1576  * and also the estimated output size.
1577  *
1578  * We are given Paths for the nonrecursive and recursive terms.
1579  */
1580 void
1581 cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
1582 {
1583  Cost startup_cost;
1584  Cost total_cost;
1585  double total_rows;
1586 
1587  /* We probably have decent estimates for the non-recursive term */
1588  startup_cost = nrterm->startup_cost;
1589  total_cost = nrterm->total_cost;
1590  total_rows = nrterm->rows;
1591 
1592  /*
1593  * We arbitrarily assume that about 10 recursive iterations will be
1594  * needed, and that we've managed to get a good fix on the cost and output
1595  * size of each one of them. These are mighty shaky assumptions but it's
1596  * hard to see how to do better.
1597  */
1598  total_cost += 10 * rterm->total_cost;
1599  total_rows += 10 * rterm->rows;
1600 
1601  /*
1602  * Also charge cpu_tuple_cost per row to account for the costs of
1603  * manipulating the tuplestores. (We don't worry about possible
1604  * spill-to-disk costs.)
1605  */
1606  total_cost += cpu_tuple_cost * total_rows;
1607 
1608  runion->startup_cost = startup_cost;
1609  runion->total_cost = total_cost;
1610  runion->rows = total_rows;
1611  runion->pathtarget->width = Max(nrterm->pathtarget->width,
1612  rterm->pathtarget->width);
1613 }
1614 
1615 /*
1616  * cost_sort
1617  * Determines and returns the cost of sorting a relation, including
1618  * the cost of reading the input data.
1619  *
1620  * If the total volume of data to sort is less than sort_mem, we will do
1621  * an in-memory sort, which requires no I/O and about t*log2(t) tuple
1622  * comparisons for t tuples.
1623  *
1624  * If the total volume exceeds sort_mem, we switch to a tape-style merge
1625  * algorithm. There will still be about t*log2(t) tuple comparisons in
1626  * total, but we will also need to write and read each tuple once per
1627  * merge pass. We expect about ceil(logM(r)) merge passes where r is the
1628  * number of initial runs formed and M is the merge order used by tuplesort.c.
1629  * Since the average initial run should be about sort_mem, we have
1630  * disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
1631  * cpu = comparison_cost * t * log2(t)
1632  *
1633  * If the sort is bounded (i.e., only the first k result tuples are needed)
1634  * and k tuples can fit into sort_mem, we use a heap method that keeps only
1635  * k tuples in the heap; this will require about t*log2(k) tuple comparisons.
1636  *
1637  * The disk traffic is assumed to be 3/4ths sequential and 1/4th random
1638  * accesses (XXX can't we refine that guess?)
1639  *
1640  * By default, we charge two operator evals per tuple comparison, which should
1641  * be in the right ballpark in most cases. The caller can tweak this by
1642  * specifying nonzero comparison_cost; typically that's used for any extra
1643  * work that has to be done to prepare the inputs to the comparison operators.
1644  *
1645  * 'pathkeys' is a list of sort keys
1646  * 'input_cost' is the total cost for reading the input data
1647  * 'tuples' is the number of tuples in the relation
1648  * 'width' is the average tuple width in bytes
1649  * 'comparison_cost' is the extra cost per comparison, if any
1650  * 'sort_mem' is the number of kilobytes of work memory allowed for the sort
1651  * 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
1652  *
1653  * NOTE: some callers currently pass NIL for pathkeys because they
1654  * can't conveniently supply the sort keys. Since this routine doesn't
1655  * currently do anything with pathkeys anyway, that doesn't matter...
1656  * but if it ever does, it should react gracefully to lack of key data.
1657  * (Actually, the thing we'd most likely be interested in is just the number
1658  * of sort keys, which all callers *could* supply.)
1659  */
1660 void
1662  List *pathkeys, Cost input_cost, double tuples, int width,
1663  Cost comparison_cost, int sort_mem,
1664  double limit_tuples)
1665 {
1666  Cost startup_cost = input_cost;
1667  Cost run_cost = 0;
1668  double input_bytes = relation_byte_size(tuples, width);
1669  double output_bytes;
1670  double output_tuples;
1671  long sort_mem_bytes = sort_mem * 1024L;
1672 
1673  if (!enable_sort)
1674  startup_cost += disable_cost;
1675 
1676  path->rows = tuples;
1677 
1678  /*
1679  * We want to be sure the cost of a sort is never estimated as zero, even
1680  * if passed-in tuple count is zero. Besides, mustn't do log(0)...
1681  */
1682  if (tuples < 2.0)
1683  tuples = 2.0;
1684 
1685  /* Include the default cost-per-comparison */
1686  comparison_cost += 2.0 * cpu_operator_cost;
1687 
1688  /* Do we have a useful LIMIT? */
1689  if (limit_tuples > 0 && limit_tuples < tuples)
1690  {
1691  output_tuples = limit_tuples;
1692  output_bytes = relation_byte_size(output_tuples, width);
1693  }
1694  else
1695  {
1696  output_tuples = tuples;
1697  output_bytes = input_bytes;
1698  }
1699 
1700  if (output_bytes > sort_mem_bytes)
1701  {
1702  /*
1703  * We'll have to use a disk-based sort of all the tuples
1704  */
1705  double npages = ceil(input_bytes / BLCKSZ);
1706  double nruns = input_bytes / sort_mem_bytes;
1707  double mergeorder = tuplesort_merge_order(sort_mem_bytes);
1708  double log_runs;
1709  double npageaccesses;
1710 
1711  /*
1712  * CPU costs
1713  *
1714  * Assume about N log2 N comparisons
1715  */
1716  startup_cost += comparison_cost * tuples * LOG2(tuples);
1717 
1718  /* Disk costs */
1719 
1720  /* Compute logM(r) as log(r) / log(M) */
1721  if (nruns > mergeorder)
1722  log_runs = ceil(log(nruns) / log(mergeorder));
1723  else
1724  log_runs = 1.0;
1725  npageaccesses = 2.0 * npages * log_runs;
1726  /* Assume 3/4ths of accesses are sequential, 1/4th are not */
1727  startup_cost += npageaccesses *
1728  (seq_page_cost * 0.75 + random_page_cost * 0.25);
1729  }
1730  else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
1731  {
1732  /*
1733  * We'll use a bounded heap-sort keeping just K tuples in memory, for
1734  * a total number of tuple comparisons of N log2 K; but the constant
1735  * factor is a bit higher than for quicksort. Tweak it so that the
1736  * cost curve is continuous at the crossover point.
1737  */
1738  startup_cost += comparison_cost * tuples * LOG2(2.0 * output_tuples);
1739  }
1740  else
1741  {
1742  /* We'll use plain quicksort on all the input tuples */
1743  startup_cost += comparison_cost * tuples * LOG2(tuples);
1744  }
1745 
1746  /*
1747  * Also charge a small amount (arbitrarily set equal to operator cost) per
1748  * extracted tuple. We don't charge cpu_tuple_cost because a Sort node
1749  * doesn't do qual-checking or projection, so it has less overhead than
1750  * most plan nodes. Note it's correct to use tuples not output_tuples
1751  * here --- the upper LIMIT will pro-rate the run cost so we'd be double
1752  * counting the LIMIT otherwise.
1753  */
1754  run_cost += cpu_operator_cost * tuples;
1755 
1756  path->startup_cost = startup_cost;
1757  path->total_cost = startup_cost + run_cost;
1758 }
1759 
1760 /*
1761  * append_nonpartial_cost
1762  * Estimate the cost of the non-partial paths in a Parallel Append.
1763  * The non-partial paths are assumed to be the first "numpaths" paths
1764  * from the subpaths list, and to be in order of decreasing cost.
1765  */
1766 static Cost
1767 append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
1768 {
1769  Cost *costarr;
1770  int arrlen;
1771  ListCell *l;
1772  ListCell *cell;
1773  int i;
1774  int path_index;
1775  int min_index;
1776  int max_index;
1777 
1778  if (numpaths == 0)
1779  return 0;
1780 
1781  /*
1782  * Array length is number of workers or number of relevants paths,
1783  * whichever is less.
1784  */
1785  arrlen = Min(parallel_workers, numpaths);
1786  costarr = (Cost *) palloc(sizeof(Cost) * arrlen);
1787 
1788  /* The first few paths will each be claimed by a different worker. */
1789  path_index = 0;
1790  foreach(cell, subpaths)
1791  {
1792  Path *subpath = (Path *) lfirst(cell);
1793 
1794  if (path_index == arrlen)
1795  break;
1796  costarr[path_index++] = subpath->total_cost;
1797  }
1798 
1799  /*
1800  * Since subpaths are sorted by decreasing cost, the last one will have
1801  * the minimum cost.
1802  */
1803  min_index = arrlen - 1;
1804 
1805  /*
1806  * For each of the remaining subpaths, add its cost to the array element
1807  * with minimum cost.
1808  */
1809  for_each_cell(l, cell)
1810  {
1811  Path *subpath = (Path *) lfirst(l);
1812  int i;
1813 
1814  /* Consider only the non-partial paths */
1815  if (path_index++ == numpaths)
1816  break;
1817 
1818  costarr[min_index] += subpath->total_cost;
1819 
1820  /* Update the new min cost array index */
1821  for (min_index = i = 0; i < arrlen; i++)
1822  {
1823  if (costarr[i] < costarr[min_index])
1824  min_index = i;
1825  }
1826  }
1827 
1828  /* Return the highest cost from the array */
1829  for (max_index = i = 0; i < arrlen; i++)
1830  {
1831  if (costarr[i] > costarr[max_index])
1832  max_index = i;
1833  }
1834 
1835  return costarr[max_index];
1836 }
1837 
1838 /*
1839  * cost_append
1840  * Determines and returns the cost of an Append node.
1841  */
1842 void
1844 {
1845  ListCell *l;
1846 
1847  apath->path.startup_cost = 0;
1848  apath->path.total_cost = 0;
1849 
1850  if (apath->subpaths == NIL)
1851  return;
1852 
1853  if (!apath->path.parallel_aware)
1854  {
1855  Path *subpath = (Path *) linitial(apath->subpaths);
1856 
1857  /*
1858  * Startup cost of non-parallel-aware Append is the startup cost of
1859  * first subpath.
1860  */
1861  apath->path.startup_cost = subpath->startup_cost;
1862 
1863  /* Compute rows and costs as sums of subplan rows and costs. */
1864  foreach(l, apath->subpaths)
1865  {
1866  Path *subpath = (Path *) lfirst(l);
1867 
1868  apath->path.rows += subpath->rows;
1869  apath->path.total_cost += subpath->total_cost;
1870  }
1871  }
1872  else /* parallel-aware */
1873  {
1874  int i = 0;
1875  double parallel_divisor = get_parallel_divisor(&apath->path);
1876 
1877  /* Calculate startup cost. */
1878  foreach(l, apath->subpaths)
1879  {
1880  Path *subpath = (Path *) lfirst(l);
1881 
1882  /*
1883  * Append will start returning tuples when the child node having
1884  * lowest startup cost is done setting up. We consider only the
1885  * first few subplans that immediately get a worker assigned.
1886  */
1887  if (i == 0)
1888  apath->path.startup_cost = subpath->startup_cost;
1889  else if (i < apath->path.parallel_workers)
1890  apath->path.startup_cost = Min(apath->path.startup_cost,
1891  subpath->startup_cost);
1892 
1893  /*
1894  * Apply parallel divisor to subpaths. Scale the number of rows
1895  * for each partial subpath based on the ratio of the parallel
1896  * divisor originally used for the subpath to the one we adopted.
1897  * Also add the cost of partial paths to the total cost, but
1898  * ignore non-partial paths for now.
1899  */
1900  if (i < apath->first_partial_path)
1901  apath->path.rows += subpath->rows / parallel_divisor;
1902  else
1903  {
1904  double subpath_parallel_divisor;
1905 
1906  subpath_parallel_divisor = get_parallel_divisor(subpath);
1907  apath->path.rows += subpath->rows * (subpath_parallel_divisor /
1908  parallel_divisor);
1909  apath->path.total_cost += subpath->total_cost;
1910  }
1911 
1912  apath->path.rows = clamp_row_est(apath->path.rows);
1913 
1914  i++;
1915  }
1916 
1917  /* Add cost for non-partial subpaths. */
1918  apath->path.total_cost +=
1920  apath->first_partial_path,
1921  apath->path.parallel_workers);
1922  }
1923 
1924  /*
1925  * Although Append does not do any selection or projection, it's not free;
1926  * add a small per-tuple overhead.
1927  */
1928  apath->path.total_cost +=
1930 }
1931 
1932 /*
1933  * cost_merge_append
1934  * Determines and returns the cost of a MergeAppend node.
1935  *
1936  * MergeAppend merges several pre-sorted input streams, using a heap that
1937  * at any given instant holds the next tuple from each stream. If there
1938  * are N streams, we need about N*log2(N) tuple comparisons to construct
1939  * the heap at startup, and then for each output tuple, about log2(N)
1940  * comparisons to replace the top entry.
1941  *
1942  * (The effective value of N will drop once some of the input streams are
1943  * exhausted, but it seems unlikely to be worth trying to account for that.)
1944  *
1945  * The heap is never spilled to disk, since we assume N is not very large.
1946  * So this is much simpler than cost_sort.
1947  *
1948  * As in cost_sort, we charge two operator evals per tuple comparison.
1949  *
1950  * 'pathkeys' is a list of sort keys
1951  * 'n_streams' is the number of input streams
1952  * 'input_startup_cost' is the sum of the input streams' startup costs
1953  * 'input_total_cost' is the sum of the input streams' total costs
1954  * 'tuples' is the number of tuples in all the streams
1955  */
1956 void
1958  List *pathkeys, int n_streams,
1959  Cost input_startup_cost, Cost input_total_cost,
1960  double tuples)
1961 {
1962  Cost startup_cost = 0;
1963  Cost run_cost = 0;
1964  Cost comparison_cost;
1965  double N;
1966  double logN;
1967 
1968  /*
1969  * Avoid log(0)...
1970  */
1971  N = (n_streams < 2) ? 2.0 : (double) n_streams;
1972  logN = LOG2(N);
1973 
1974  /* Assumed cost per tuple comparison */
1975  comparison_cost = 2.0 * cpu_operator_cost;
1976 
1977  /* Heap creation cost */
1978  startup_cost += comparison_cost * N * logN;
1979 
1980  /* Per-tuple heap maintenance cost */
1981  run_cost += tuples * comparison_cost * logN;
1982 
1983  /*
1984  * Although MergeAppend does not do any selection or projection, it's not
1985  * free; add a small per-tuple overhead.
1986  */
1987  run_cost += cpu_tuple_cost * APPEND_CPU_COST_MULTIPLIER * tuples;
1988 
1989  path->startup_cost = startup_cost + input_startup_cost;
1990  path->total_cost = startup_cost + run_cost + input_total_cost;
1991 }
1992 
1993 /*
1994  * cost_material
1995  * Determines and returns the cost of materializing a relation, including
1996  * the cost of reading the input data.
1997  *
1998  * If the total volume of data to materialize exceeds work_mem, we will need
1999  * to write it to disk, so the cost is much higher in that case.
2000  *
2001  * Note that here we are estimating the costs for the first scan of the
2002  * relation, so the materialization is all overhead --- any savings will
2003  * occur only on rescan, which is estimated in cost_rescan.
2004  */
2005 void
2007  Cost input_startup_cost, Cost input_total_cost,
2008  double tuples, int width)
2009 {
2010  Cost startup_cost = input_startup_cost;
2011  Cost run_cost = input_total_cost - input_startup_cost;
2012  double nbytes = relation_byte_size(tuples, width);
2013  long work_mem_bytes = work_mem * 1024L;
2014 
2015  path->rows = tuples;
2016 
2017  /*
2018  * Whether spilling or not, charge 2x cpu_operator_cost per tuple to
2019  * reflect bookkeeping overhead. (This rate must be more than what
2020  * cost_rescan charges for materialize, ie, cpu_operator_cost per tuple;
2021  * if it is exactly the same then there will be a cost tie between
2022  * nestloop with A outer, materialized B inner and nestloop with B outer,
2023  * materialized A inner. The extra cost ensures we'll prefer
2024  * materializing the smaller rel.) Note that this is normally a good deal
2025  * less than cpu_tuple_cost; which is OK because a Material plan node
2026  * doesn't do qual-checking or projection, so it's got less overhead than
2027  * most plan nodes.
2028  */
2029  run_cost += 2 * cpu_operator_cost * tuples;
2030 
2031  /*
2032  * If we will spill to disk, charge at the rate of seq_page_cost per page.
2033  * This cost is assumed to be evenly spread through the plan run phase,
2034  * which isn't exactly accurate but our cost model doesn't allow for
2035  * nonuniform costs within the run phase.
2036  */
2037  if (nbytes > work_mem_bytes)
2038  {
2039  double npages = ceil(nbytes / BLCKSZ);
2040 
2041  run_cost += seq_page_cost * npages;
2042  }
2043 
2044  path->startup_cost = startup_cost;
2045  path->total_cost = startup_cost + run_cost;
2046 }
2047 
2048 /*
2049  * cost_agg
2050  * Determines and returns the cost of performing an Agg plan node,
2051  * including the cost of its input.
2052  *
2053  * aggcosts can be NULL when there are no actual aggregate functions (i.e.,
2054  * we are using a hashed Agg node just to do grouping).
2055  *
2056  * Note: when aggstrategy == AGG_SORTED, caller must ensure that input costs
2057  * are for appropriately-sorted input.
2058  */
2059 void
2061  AggStrategy aggstrategy, const AggClauseCosts *aggcosts,
2062  int numGroupCols, double numGroups,
2063  List *quals,
2064  Cost input_startup_cost, Cost input_total_cost,
2065  double input_tuples)
2066 {
2067  double output_tuples;
2068  Cost startup_cost;
2069  Cost total_cost;
2070  AggClauseCosts dummy_aggcosts;
2071 
2072  /* Use all-zero per-aggregate costs if NULL is passed */
2073  if (aggcosts == NULL)
2074  {
2075  Assert(aggstrategy == AGG_HASHED);
2076  MemSet(&dummy_aggcosts, 0, sizeof(AggClauseCosts));
2077  aggcosts = &dummy_aggcosts;
2078  }
2079 
2080  /*
2081  * The transCost.per_tuple component of aggcosts should be charged once
2082  * per input tuple, corresponding to the costs of evaluating the aggregate
2083  * transfns and their input expressions (with any startup cost of course
2084  * charged but once). The finalCost component is charged once per output
2085  * tuple, corresponding to the costs of evaluating the finalfns.
2086  *
2087  * If we are grouping, we charge an additional cpu_operator_cost per
2088  * grouping column per input tuple for grouping comparisons.
2089  *
2090  * We will produce a single output tuple if not grouping, and a tuple per
2091  * group otherwise. We charge cpu_tuple_cost for each output tuple.
2092  *
2093  * Note: in this cost model, AGG_SORTED and AGG_HASHED have exactly the
2094  * same total CPU cost, but AGG_SORTED has lower startup cost. If the
2095  * input path is already sorted appropriately, AGG_SORTED should be
2096  * preferred (since it has no risk of memory overflow). This will happen
2097  * as long as the computed total costs are indeed exactly equal --- but if
2098  * there's roundoff error we might do the wrong thing. So be sure that
2099  * the computations below form the same intermediate values in the same
2100  * order.
2101  */
2102  if (aggstrategy == AGG_PLAIN)
2103  {
2104  startup_cost = input_total_cost;
2105  startup_cost += aggcosts->transCost.startup;
2106  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2107  startup_cost += aggcosts->finalCost;
2108  /* we aren't grouping */
2109  total_cost = startup_cost + cpu_tuple_cost;
2110  output_tuples = 1;
2111  }
2112  else if (aggstrategy == AGG_SORTED || aggstrategy == AGG_MIXED)
2113  {
2114  /* Here we are able to deliver output on-the-fly */
2115  startup_cost = input_startup_cost;
2116  total_cost = input_total_cost;
2117  if (aggstrategy == AGG_MIXED && !enable_hashagg)
2118  {
2119  startup_cost += disable_cost;
2120  total_cost += disable_cost;
2121  }
2122  /* calcs phrased this way to match HASHED case, see note above */
2123  total_cost += aggcosts->transCost.startup;
2124  total_cost += aggcosts->transCost.per_tuple * input_tuples;
2125  total_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2126  total_cost += aggcosts->finalCost * numGroups;
2127  total_cost += cpu_tuple_cost * numGroups;
2128  output_tuples = numGroups;
2129  }
2130  else
2131  {
2132  /* must be AGG_HASHED */
2133  startup_cost = input_total_cost;
2134  if (!enable_hashagg)
2135  startup_cost += disable_cost;
2136  startup_cost += aggcosts->transCost.startup;
2137  startup_cost += aggcosts->transCost.per_tuple * input_tuples;
2138  startup_cost += (cpu_operator_cost * numGroupCols) * input_tuples;
2139  total_cost = startup_cost;
2140  total_cost += aggcosts->finalCost * numGroups;
2141  total_cost += cpu_tuple_cost * numGroups;
2142  output_tuples = numGroups;
2143  }
2144 
2145  /*
2146  * If there are quals (HAVING quals), account for their cost and
2147  * selectivity.
2148  */
2149  if (quals)
2150  {
2151  QualCost qual_cost;
2152 
2153  cost_qual_eval(&qual_cost, quals, root);
2154  startup_cost += qual_cost.startup;
2155  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2156 
2157  output_tuples = clamp_row_est(output_tuples *
2159  quals,
2160  0,
2161  JOIN_INNER,
2162  NULL));
2163  }
2164 
2165  path->rows = output_tuples;
2166  path->startup_cost = startup_cost;
2167  path->total_cost = total_cost;
2168 }
2169 
2170 /*
2171  * cost_windowagg
2172  * Determines and returns the cost of performing a WindowAgg plan node,
2173  * including the cost of its input.
2174  *
2175  * Input is assumed already properly sorted.
2176  */
2177 void
2179  List *windowFuncs, int numPartCols, int numOrderCols,
2180  Cost input_startup_cost, Cost input_total_cost,
2181  double input_tuples)
2182 {
2183  Cost startup_cost;
2184  Cost total_cost;
2185  ListCell *lc;
2186 
2187  startup_cost = input_startup_cost;
2188  total_cost = input_total_cost;
2189 
2190  /*
2191  * Window functions are assumed to cost their stated execution cost, plus
2192  * the cost of evaluating their input expressions, per tuple. Since they
2193  * may in fact evaluate their inputs at multiple rows during each cycle,
2194  * this could be a drastic underestimate; but without a way to know how
2195  * many rows the window function will fetch, it's hard to do better. In
2196  * any case, it's a good estimate for all the built-in window functions,
2197  * so we'll just do this for now.
2198  */
2199  foreach(lc, windowFuncs)
2200  {
2201  WindowFunc *wfunc = lfirst_node(WindowFunc, lc);
2202  Cost wfunccost;
2203  QualCost argcosts;
2204 
2205  wfunccost = get_func_cost(wfunc->winfnoid) * cpu_operator_cost;
2206 
2207  /* also add the input expressions' cost to per-input-row costs */
2208  cost_qual_eval_node(&argcosts, (Node *) wfunc->args, root);
2209  startup_cost += argcosts.startup;
2210  wfunccost += argcosts.per_tuple;
2211 
2212  /*
2213  * Add the filter's cost to per-input-row costs. XXX We should reduce
2214  * input expression costs according to filter selectivity.
2215  */
2216  cost_qual_eval_node(&argcosts, (Node *) wfunc->aggfilter, root);
2217  startup_cost += argcosts.startup;
2218  wfunccost += argcosts.per_tuple;
2219 
2220  total_cost += wfunccost * input_tuples;
2221  }
2222 
2223  /*
2224  * We also charge cpu_operator_cost per grouping column per tuple for
2225  * grouping comparisons, plus cpu_tuple_cost per tuple for general
2226  * overhead.
2227  *
2228  * XXX this neglects costs of spooling the data to disk when it overflows
2229  * work_mem. Sooner or later that should get accounted for.
2230  */
2231  total_cost += cpu_operator_cost * (numPartCols + numOrderCols) * input_tuples;
2232  total_cost += cpu_tuple_cost * input_tuples;
2233 
2234  path->rows = input_tuples;
2235  path->startup_cost = startup_cost;
2236  path->total_cost = total_cost;
2237 }
2238 
2239 /*
2240  * cost_group
2241  * Determines and returns the cost of performing a Group plan node,
2242  * including the cost of its input.
2243  *
2244  * Note: caller must ensure that input costs are for appropriately-sorted
2245  * input.
2246  */
2247 void
2249  int numGroupCols, double numGroups,
2250  List *quals,
2251  Cost input_startup_cost, Cost input_total_cost,
2252  double input_tuples)
2253 {
2254  double output_tuples;
2255  Cost startup_cost;
2256  Cost total_cost;
2257 
2258  output_tuples = numGroups;
2259  startup_cost = input_startup_cost;
2260  total_cost = input_total_cost;
2261 
2262  /*
2263  * Charge one cpu_operator_cost per comparison per input tuple. We assume
2264  * all columns get compared at most of the tuples.
2265  */
2266  total_cost += cpu_operator_cost * input_tuples * numGroupCols;
2267 
2268  /*
2269  * If there are quals (HAVING quals), account for their cost and
2270  * selectivity.
2271  */
2272  if (quals)
2273  {
2274  QualCost qual_cost;
2275 
2276  cost_qual_eval(&qual_cost, quals, root);
2277  startup_cost += qual_cost.startup;
2278  total_cost += qual_cost.startup + output_tuples * qual_cost.per_tuple;
2279 
2280  output_tuples = clamp_row_est(output_tuples *
2282  quals,
2283  0,
2284  JOIN_INNER,
2285  NULL));
2286  }
2287 
2288  path->rows = output_tuples;
2289  path->startup_cost = startup_cost;
2290  path->total_cost = total_cost;
2291 }
2292 
2293 /*
2294  * initial_cost_nestloop
2295  * Preliminary estimate of the cost of a nestloop join path.
2296  *
2297  * This must quickly produce lower-bound estimates of the path's startup and
2298  * total costs. If we are unable to eliminate the proposed path from
2299  * consideration using the lower bounds, final_cost_nestloop will be called
2300  * to obtain the final estimates.
2301  *
2302  * The exact division of labor between this function and final_cost_nestloop
2303  * is private to them, and represents a tradeoff between speed of the initial
2304  * estimate and getting a tight lower bound. We choose to not examine the
2305  * join quals here, since that's by far the most expensive part of the
2306  * calculations. The end result is that CPU-cost considerations must be
2307  * left for the second phase; and for SEMI/ANTI joins, we must also postpone
2308  * incorporation of the inner path's run cost.
2309  *
2310  * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
2311  * other data to be used by final_cost_nestloop
2312  * 'jointype' is the type of join to be performed
2313  * 'outer_path' is the outer input to the join
2314  * 'inner_path' is the inner input to the join
2315  * 'extra' contains miscellaneous information about the join
2316  */
2317 void
2319  JoinType jointype,
2320  Path *outer_path, Path *inner_path,
2321  JoinPathExtraData *extra)
2322 {
2323  Cost startup_cost = 0;
2324  Cost run_cost = 0;
2325  double outer_path_rows = outer_path->rows;
2326  Cost inner_rescan_start_cost;
2327  Cost inner_rescan_total_cost;
2328  Cost inner_run_cost;
2329  Cost inner_rescan_run_cost;
2330 
2331  /* estimate costs to rescan the inner relation */
2332  cost_rescan(root, inner_path,
2333  &inner_rescan_start_cost,
2334  &inner_rescan_total_cost);
2335 
2336  /* cost of source data */
2337 
2338  /*
2339  * NOTE: clearly, we must pay both outer and inner paths' startup_cost
2340  * before we can start returning tuples, so the join's startup cost is
2341  * their sum. We'll also pay the inner path's rescan startup cost
2342  * multiple times.
2343  */
2344  startup_cost += outer_path->startup_cost + inner_path->startup_cost;
2345  run_cost += outer_path->total_cost - outer_path->startup_cost;
2346  if (outer_path_rows > 1)
2347  run_cost += (outer_path_rows - 1) * inner_rescan_start_cost;
2348 
2349  inner_run_cost = inner_path->total_cost - inner_path->startup_cost;
2350  inner_rescan_run_cost = inner_rescan_total_cost - inner_rescan_start_cost;
2351 
2352  if (jointype == JOIN_SEMI || jointype == JOIN_ANTI ||
2353  extra->inner_unique)
2354  {
2355  /*
2356  * With a SEMI or ANTI join, or if the innerrel is known unique, the
2357  * executor will stop after the first match.
2358  *
2359  * Getting decent estimates requires inspection of the join quals,
2360  * which we choose to postpone to final_cost_nestloop.
2361  */
2362 
2363  /* Save private data for final_cost_nestloop */
2364  workspace->inner_run_cost = inner_run_cost;
2365  workspace->inner_rescan_run_cost = inner_rescan_run_cost;
2366  }
2367  else
2368  {
2369  /* Normal case; we'll scan whole input rel for each outer row */
2370  run_cost += inner_run_cost;
2371  if (outer_path_rows > 1)
2372  run_cost += (outer_path_rows - 1) * inner_rescan_run_cost;
2373  }
2374 
2375  /* CPU costs left for later */
2376 
2377  /* Public result fields */
2378  workspace->startup_cost = startup_cost;
2379  workspace->total_cost = startup_cost + run_cost;
2380  /* Save private data for final_cost_nestloop */
2381  workspace->run_cost = run_cost;
2382 }
2383 
2384 /*
2385  * final_cost_nestloop
2386  * Final estimate of the cost and result size of a nestloop join path.
2387  *
2388  * 'path' is already filled in except for the rows and cost fields
2389  * 'workspace' is the result from initial_cost_nestloop
2390  * 'extra' contains miscellaneous information about the join
2391  */
2392 void
2394  JoinCostWorkspace *workspace,
2395  JoinPathExtraData *extra)
2396 {
2397  Path *outer_path = path->outerjoinpath;
2398  Path *inner_path = path->innerjoinpath;
2399  double outer_path_rows = outer_path->rows;
2400  double inner_path_rows = inner_path->rows;
2401  Cost startup_cost = workspace->startup_cost;
2402  Cost run_cost = workspace->run_cost;
2403  Cost cpu_per_tuple;
2404  QualCost restrict_qual_cost;
2405  double ntuples;
2406 
2407  /* Protect some assumptions below that rowcounts aren't zero or NaN */
2408  if (outer_path_rows <= 0 || isnan(outer_path_rows))
2409  outer_path_rows = 1;
2410  if (inner_path_rows <= 0 || isnan(inner_path_rows))
2411  inner_path_rows = 1;
2412 
2413  /* Mark the path with the correct row estimate */
2414  if (path->path.param_info)
2415  path->path.rows = path->path.param_info->ppi_rows;
2416  else
2417  path->path.rows = path->path.parent->rows;
2418 
2419  /* For partial paths, scale row estimate. */
2420  if (path->path.parallel_workers > 0)
2421  {
2422  double parallel_divisor = get_parallel_divisor(&path->path);
2423 
2424  path->path.rows =
2425  clamp_row_est(path->path.rows / parallel_divisor);
2426  }
2427 
2428  /*
2429  * We could include disable_cost in the preliminary estimate, but that
2430  * would amount to optimizing for the case where the join method is
2431  * disabled, which doesn't seem like the way to bet.
2432  */
2433  if (!enable_nestloop)
2434  startup_cost += disable_cost;
2435 
2436  /* cost of inner-relation source data (we already dealt with outer rel) */
2437 
2438  if (path->jointype == JOIN_SEMI || path->jointype == JOIN_ANTI ||
2439  extra->inner_unique)
2440  {
2441  /*
2442  * With a SEMI or ANTI join, or if the innerrel is known unique, the
2443  * executor will stop after the first match.
2444  */
2445  Cost inner_run_cost = workspace->inner_run_cost;
2446  Cost inner_rescan_run_cost = workspace->inner_rescan_run_cost;
2447  double outer_matched_rows;
2448  double outer_unmatched_rows;
2449  Selectivity inner_scan_frac;
2450 
2451  /*
2452  * For an outer-rel row that has at least one match, we can expect the
2453  * inner scan to stop after a fraction 1/(match_count+1) of the inner
2454  * rows, if the matches are evenly distributed. Since they probably
2455  * aren't quite evenly distributed, we apply a fuzz factor of 2.0 to
2456  * that fraction. (If we used a larger fuzz factor, we'd have to
2457  * clamp inner_scan_frac to at most 1.0; but since match_count is at
2458  * least 1, no such clamp is needed now.)
2459  */
2460  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
2461  outer_unmatched_rows = outer_path_rows - outer_matched_rows;
2462  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
2463 
2464  /*
2465  * Compute number of tuples processed (not number emitted!). First,
2466  * account for successfully-matched outer rows.
2467  */
2468  ntuples = outer_matched_rows * inner_path_rows * inner_scan_frac;
2469 
2470  /*
2471  * Now we need to estimate the actual costs of scanning the inner
2472  * relation, which may be quite a bit less than N times inner_run_cost
2473  * due to early scan stops. We consider two cases. If the inner path
2474  * is an indexscan using all the joinquals as indexquals, then an
2475  * unmatched outer row results in an indexscan returning no rows,
2476  * which is probably quite cheap. Otherwise, the executor will have
2477  * to scan the whole inner rel for an unmatched row; not so cheap.
2478  */
2479  if (has_indexed_join_quals(path))
2480  {
2481  /*
2482  * Successfully-matched outer rows will only require scanning
2483  * inner_scan_frac of the inner relation. In this case, we don't
2484  * need to charge the full inner_run_cost even when that's more
2485  * than inner_rescan_run_cost, because we can assume that none of
2486  * the inner scans ever scan the whole inner relation. So it's
2487  * okay to assume that all the inner scan executions can be
2488  * fractions of the full cost, even if materialization is reducing
2489  * the rescan cost. At this writing, it's impossible to get here
2490  * for a materialized inner scan, so inner_run_cost and
2491  * inner_rescan_run_cost will be the same anyway; but just in
2492  * case, use inner_run_cost for the first matched tuple and
2493  * inner_rescan_run_cost for additional ones.
2494  */
2495  run_cost += inner_run_cost * inner_scan_frac;
2496  if (outer_matched_rows > 1)
2497  run_cost += (outer_matched_rows - 1) * inner_rescan_run_cost * inner_scan_frac;
2498 
2499  /*
2500  * Add the cost of inner-scan executions for unmatched outer rows.
2501  * We estimate this as the same cost as returning the first tuple
2502  * of a nonempty scan. We consider that these are all rescans,
2503  * since we used inner_run_cost once already.
2504  */
2505  run_cost += outer_unmatched_rows *
2506  inner_rescan_run_cost / inner_path_rows;
2507 
2508  /*
2509  * We won't be evaluating any quals at all for unmatched rows, so
2510  * don't add them to ntuples.
2511  */
2512  }
2513  else
2514  {
2515  /*
2516  * Here, a complicating factor is that rescans may be cheaper than
2517  * first scans. If we never scan all the way to the end of the
2518  * inner rel, it might be (depending on the plan type) that we'd
2519  * never pay the whole inner first-scan run cost. However it is
2520  * difficult to estimate whether that will happen (and it could
2521  * not happen if there are any unmatched outer rows!), so be
2522  * conservative and always charge the whole first-scan cost once.
2523  * We consider this charge to correspond to the first unmatched
2524  * outer row, unless there isn't one in our estimate, in which
2525  * case blame it on the first matched row.
2526  */
2527 
2528  /* First, count all unmatched join tuples as being processed */
2529  ntuples += outer_unmatched_rows * inner_path_rows;
2530 
2531  /* Now add the forced full scan, and decrement appropriate count */
2532  run_cost += inner_run_cost;
2533  if (outer_unmatched_rows >= 1)
2534  outer_unmatched_rows -= 1;
2535  else
2536  outer_matched_rows -= 1;
2537 
2538  /* Add inner run cost for additional outer tuples having matches */
2539  if (outer_matched_rows > 0)
2540  run_cost += outer_matched_rows * inner_rescan_run_cost * inner_scan_frac;
2541 
2542  /* Add inner run cost for additional unmatched outer tuples */
2543  if (outer_unmatched_rows > 0)
2544  run_cost += outer_unmatched_rows * inner_rescan_run_cost;
2545  }
2546  }
2547  else
2548  {
2549  /* Normal-case source costs were included in preliminary estimate */
2550 
2551  /* Compute number of tuples processed (not number emitted!) */
2552  ntuples = outer_path_rows * inner_path_rows;
2553  }
2554 
2555  /* CPU costs */
2556  cost_qual_eval(&restrict_qual_cost, path->joinrestrictinfo, root);
2557  startup_cost += restrict_qual_cost.startup;
2558  cpu_per_tuple = cpu_tuple_cost + restrict_qual_cost.per_tuple;
2559  run_cost += cpu_per_tuple * ntuples;
2560 
2561  /* tlist eval costs are paid per output row, not per tuple scanned */
2562  startup_cost += path->path.pathtarget->cost.startup;
2563  run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
2564 
2565  path->path.startup_cost = startup_cost;
2566  path->path.total_cost = startup_cost + run_cost;
2567 }
2568 
2569 /*
2570  * initial_cost_mergejoin
2571  * Preliminary estimate of the cost of a mergejoin path.
2572  *
2573  * This must quickly produce lower-bound estimates of the path's startup and
2574  * total costs. If we are unable to eliminate the proposed path from
2575  * consideration using the lower bounds, final_cost_mergejoin will be called
2576  * to obtain the final estimates.
2577  *
2578  * The exact division of labor between this function and final_cost_mergejoin
2579  * is private to them, and represents a tradeoff between speed of the initial
2580  * estimate and getting a tight lower bound. We choose to not examine the
2581  * join quals here, except for obtaining the scan selectivity estimate which
2582  * is really essential (but fortunately, use of caching keeps the cost of
2583  * getting that down to something reasonable).
2584  * We also assume that cost_sort is cheap enough to use here.
2585  *
2586  * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
2587  * other data to be used by final_cost_mergejoin
2588  * 'jointype' is the type of join to be performed
2589  * 'mergeclauses' is the list of joinclauses to be used as merge clauses
2590  * 'outer_path' is the outer input to the join
2591  * 'inner_path' is the inner input to the join
2592  * 'outersortkeys' is the list of sort keys for the outer path
2593  * 'innersortkeys' is the list of sort keys for the inner path
2594  * 'extra' contains miscellaneous information about the join
2595  *
2596  * Note: outersortkeys and innersortkeys should be NIL if no explicit
2597  * sort is needed because the respective source path is already ordered.
2598  */
2599 void
2601  JoinType jointype,
2602  List *mergeclauses,
2603  Path *outer_path, Path *inner_path,
2604  List *outersortkeys, List *innersortkeys,
2605  JoinPathExtraData *extra)
2606 {
2607  Cost startup_cost = 0;
2608  Cost run_cost = 0;
2609  double outer_path_rows = outer_path->rows;
2610  double inner_path_rows = inner_path->rows;
2611  Cost inner_run_cost;
2612  double outer_rows,
2613  inner_rows,
2614  outer_skip_rows,
2615  inner_skip_rows;
2616  Selectivity outerstartsel,
2617  outerendsel,
2618  innerstartsel,
2619  innerendsel;
2620  Path sort_path; /* dummy for result of cost_sort */
2621 
2622  /* Protect some assumptions below that rowcounts aren't zero or NaN */
2623  if (outer_path_rows <= 0 || isnan(outer_path_rows))
2624  outer_path_rows = 1;
2625  if (inner_path_rows <= 0 || isnan(inner_path_rows))
2626  inner_path_rows = 1;
2627 
2628  /*
2629  * A merge join will stop as soon as it exhausts either input stream
2630  * (unless it's an outer join, in which case the outer side has to be
2631  * scanned all the way anyway). Estimate fraction of the left and right
2632  * inputs that will actually need to be scanned. Likewise, we can
2633  * estimate the number of rows that will be skipped before the first join
2634  * pair is found, which should be factored into startup cost. We use only
2635  * the first (most significant) merge clause for this purpose. Since
2636  * mergejoinscansel() is a fairly expensive computation, we cache the
2637  * results in the merge clause RestrictInfo.
2638  */
2639  if (mergeclauses && jointype != JOIN_FULL)
2640  {
2641  RestrictInfo *firstclause = (RestrictInfo *) linitial(mergeclauses);
2642  List *opathkeys;
2643  List *ipathkeys;
2644  PathKey *opathkey;
2645  PathKey *ipathkey;
2646  MergeScanSelCache *cache;
2647 
2648  /* Get the input pathkeys to determine the sort-order details */
2649  opathkeys = outersortkeys ? outersortkeys : outer_path->pathkeys;
2650  ipathkeys = innersortkeys ? innersortkeys : inner_path->pathkeys;
2651  Assert(opathkeys);
2652  Assert(ipathkeys);
2653  opathkey = (PathKey *) linitial(opathkeys);
2654  ipathkey = (PathKey *) linitial(ipathkeys);
2655  /* debugging check */
2656  if (opathkey->pk_opfamily != ipathkey->pk_opfamily ||
2657  opathkey->pk_eclass->ec_collation != ipathkey->pk_eclass->ec_collation ||
2658  opathkey->pk_strategy != ipathkey->pk_strategy ||
2659  opathkey->pk_nulls_first != ipathkey->pk_nulls_first)
2660  elog(ERROR, "left and right pathkeys do not match in mergejoin");
2661 
2662  /* Get the selectivity with caching */
2663  cache = cached_scansel(root, firstclause, opathkey);
2664 
2665  if (bms_is_subset(firstclause->left_relids,
2666  outer_path->parent->relids))
2667  {
2668  /* left side of clause is outer */
2669  outerstartsel = cache->leftstartsel;
2670  outerendsel = cache->leftendsel;
2671  innerstartsel = cache->rightstartsel;
2672  innerendsel = cache->rightendsel;
2673  }
2674  else
2675  {
2676  /* left side of clause is inner */
2677  outerstartsel = cache->rightstartsel;
2678  outerendsel = cache->rightendsel;
2679  innerstartsel = cache->leftstartsel;
2680  innerendsel = cache->leftendsel;
2681  }
2682  if (jointype == JOIN_LEFT ||
2683  jointype == JOIN_ANTI)
2684  {
2685  outerstartsel = 0.0;
2686  outerendsel = 1.0;
2687  }
2688  else if (jointype == JOIN_RIGHT)
2689  {
2690  innerstartsel = 0.0;
2691  innerendsel = 1.0;
2692  }
2693  }
2694  else
2695  {
2696  /* cope with clauseless or full mergejoin */
2697  outerstartsel = innerstartsel = 0.0;
2698  outerendsel = innerendsel = 1.0;
2699  }
2700 
2701  /*
2702  * Convert selectivities to row counts. We force outer_rows and
2703  * inner_rows to be at least 1, but the skip_rows estimates can be zero.
2704  */
2705  outer_skip_rows = rint(outer_path_rows * outerstartsel);
2706  inner_skip_rows = rint(inner_path_rows * innerstartsel);
2707  outer_rows = clamp_row_est(outer_path_rows * outerendsel);
2708  inner_rows = clamp_row_est(inner_path_rows * innerendsel);
2709 
2710  Assert(outer_skip_rows <= outer_rows);
2711  Assert(inner_skip_rows <= inner_rows);
2712 
2713  /*
2714  * Readjust scan selectivities to account for above rounding. This is
2715  * normally an insignificant effect, but when there are only a few rows in
2716  * the inputs, failing to do this makes for a large percentage error.
2717  */
2718  outerstartsel = outer_skip_rows / outer_path_rows;
2719  innerstartsel = inner_skip_rows / inner_path_rows;
2720  outerendsel = outer_rows / outer_path_rows;
2721  innerendsel = inner_rows / inner_path_rows;
2722 
2723  Assert(outerstartsel <= outerendsel);
2724  Assert(innerstartsel <= innerendsel);
2725 
2726  /* cost of source data */
2727 
2728  if (outersortkeys) /* do we need to sort outer? */
2729  {
2730  cost_sort(&sort_path,
2731  root,
2732  outersortkeys,
2733  outer_path->total_cost,
2734  outer_path_rows,
2735  outer_path->pathtarget->width,
2736  0.0,
2737  work_mem,
2738  -1.0);
2739  startup_cost += sort_path.startup_cost;
2740  startup_cost += (sort_path.total_cost - sort_path.startup_cost)
2741  * outerstartsel;
2742  run_cost += (sort_path.total_cost - sort_path.startup_cost)
2743  * (outerendsel - outerstartsel);
2744  }
2745  else
2746  {
2747  startup_cost += outer_path->startup_cost;
2748  startup_cost += (outer_path->total_cost - outer_path->startup_cost)
2749  * outerstartsel;
2750  run_cost += (outer_path->total_cost - outer_path->startup_cost)
2751  * (outerendsel - outerstartsel);
2752  }
2753 
2754  if (innersortkeys) /* do we need to sort inner? */
2755  {
2756  cost_sort(&sort_path,
2757  root,
2758  innersortkeys,
2759  inner_path->total_cost,
2760  inner_path_rows,
2761  inner_path->pathtarget->width,
2762  0.0,
2763  work_mem,
2764  -1.0);
2765  startup_cost += sort_path.startup_cost;
2766  startup_cost += (sort_path.total_cost - sort_path.startup_cost)
2767  * innerstartsel;
2768  inner_run_cost = (sort_path.total_cost - sort_path.startup_cost)
2769  * (innerendsel - innerstartsel);
2770  }
2771  else
2772  {
2773  startup_cost += inner_path->startup_cost;
2774  startup_cost += (inner_path->total_cost - inner_path->startup_cost)
2775  * innerstartsel;
2776  inner_run_cost = (inner_path->total_cost - inner_path->startup_cost)
2777  * (innerendsel - innerstartsel);
2778  }
2779 
2780  /*
2781  * We can't yet determine whether rescanning occurs, or whether
2782  * materialization of the inner input should be done. The minimum
2783  * possible inner input cost, regardless of rescan and materialization
2784  * considerations, is inner_run_cost. We include that in
2785  * workspace->total_cost, but not yet in run_cost.
2786  */
2787 
2788  /* CPU costs left for later */
2789 
2790  /* Public result fields */
2791  workspace->startup_cost = startup_cost;
2792  workspace->total_cost = startup_cost + run_cost + inner_run_cost;
2793  /* Save private data for final_cost_mergejoin */
2794  workspace->run_cost = run_cost;
2795  workspace->inner_run_cost = inner_run_cost;
2796  workspace->outer_rows = outer_rows;
2797  workspace->inner_rows = inner_rows;
2798  workspace->outer_skip_rows = outer_skip_rows;
2799  workspace->inner_skip_rows = inner_skip_rows;
2800 }
2801 
2802 /*
2803  * final_cost_mergejoin
2804  * Final estimate of the cost and result size of a mergejoin path.
2805  *
2806  * Unlike other costsize functions, this routine makes two actual decisions:
2807  * whether the executor will need to do mark/restore, and whether we should
2808  * materialize the inner path. It would be logically cleaner to build
2809  * separate paths testing these alternatives, but that would require repeating
2810  * most of the cost calculations, which are not all that cheap. Since the
2811  * choice will not affect output pathkeys or startup cost, only total cost,
2812  * there is no possibility of wanting to keep more than one path. So it seems
2813  * best to make the decisions here and record them in the path's
2814  * skip_mark_restore and materialize_inner fields.
2815  *
2816  * Mark/restore overhead is usually required, but can be skipped if we know
2817  * that the executor need find only one match per outer tuple, and that the
2818  * mergeclauses are sufficient to identify a match.
2819  *
2820  * We materialize the inner path if we need mark/restore and either the inner
2821  * path can't support mark/restore, or it's cheaper to use an interposed
2822  * Material node to handle mark/restore.
2823  *
2824  * 'path' is already filled in except for the rows and cost fields and
2825  * skip_mark_restore and materialize_inner
2826  * 'workspace' is the result from initial_cost_mergejoin
2827  * 'extra' contains miscellaneous information about the join
2828  */
2829 void
2831  JoinCostWorkspace *workspace,
2832  JoinPathExtraData *extra)
2833 {
2834  Path *outer_path = path->jpath.outerjoinpath;
2835  Path *inner_path = path->jpath.innerjoinpath;
2836  double inner_path_rows = inner_path->rows;
2837  List *mergeclauses = path->path_mergeclauses;
2838  List *innersortkeys = path->innersortkeys;
2839  Cost startup_cost = workspace->startup_cost;
2840  Cost run_cost = workspace->run_cost;
2841  Cost inner_run_cost = workspace->inner_run_cost;
2842  double outer_rows = workspace->outer_rows;
2843  double inner_rows = workspace->inner_rows;
2844  double outer_skip_rows = workspace->outer_skip_rows;
2845  double inner_skip_rows = workspace->inner_skip_rows;
2846  Cost cpu_per_tuple,
2847  bare_inner_cost,
2848  mat_inner_cost;
2849  QualCost merge_qual_cost;
2850  QualCost qp_qual_cost;
2851  double mergejointuples,
2852  rescannedtuples;
2853  double rescanratio;
2854 
2855  /* Protect some assumptions below that rowcounts aren't zero or NaN */
2856  if (inner_path_rows <= 0 || isnan(inner_path_rows))
2857  inner_path_rows = 1;
2858 
2859  /* Mark the path with the correct row estimate */
2860  if (path->jpath.path.param_info)
2861  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
2862  else
2863  path->jpath.path.rows = path->jpath.path.parent->rows;
2864 
2865  /* For partial paths, scale row estimate. */
2866  if (path->jpath.path.parallel_workers > 0)
2867  {
2868  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
2869 
2870  path->jpath.path.rows =
2871  clamp_row_est(path->jpath.path.rows / parallel_divisor);
2872  }
2873 
2874  /*
2875  * We could include disable_cost in the preliminary estimate, but that
2876  * would amount to optimizing for the case where the join method is
2877  * disabled, which doesn't seem like the way to bet.
2878  */
2879  if (!enable_mergejoin)
2880  startup_cost += disable_cost;
2881 
2882  /*
2883  * Compute cost of the mergequals and qpquals (other restriction clauses)
2884  * separately.
2885  */
2886  cost_qual_eval(&merge_qual_cost, mergeclauses, root);
2887  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
2888  qp_qual_cost.startup -= merge_qual_cost.startup;
2889  qp_qual_cost.per_tuple -= merge_qual_cost.per_tuple;
2890 
2891  /*
2892  * With a SEMI or ANTI join, or if the innerrel is known unique, the
2893  * executor will stop scanning for matches after the first match. When
2894  * all the joinclauses are merge clauses, this means we don't ever need to
2895  * back up the merge, and so we can skip mark/restore overhead.
2896  */
2897  if ((path->jpath.jointype == JOIN_SEMI ||
2898  path->jpath.jointype == JOIN_ANTI ||
2899  extra->inner_unique) &&
2902  path->skip_mark_restore = true;
2903  else
2904  path->skip_mark_restore = false;
2905 
2906  /*
2907  * Get approx # tuples passing the mergequals. We use approx_tuple_count
2908  * here because we need an estimate done with JOIN_INNER semantics.
2909  */
2910  mergejointuples = approx_tuple_count(root, &path->jpath, mergeclauses);
2911 
2912  /*
2913  * When there are equal merge keys in the outer relation, the mergejoin
2914  * must rescan any matching tuples in the inner relation. This means
2915  * re-fetching inner tuples; we have to estimate how often that happens.
2916  *
2917  * For regular inner and outer joins, the number of re-fetches can be
2918  * estimated approximately as size of merge join output minus size of
2919  * inner relation. Assume that the distinct key values are 1, 2, ..., and
2920  * denote the number of values of each key in the outer relation as m1,
2921  * m2, ...; in the inner relation, n1, n2, ... Then we have
2922  *
2923  * size of join = m1 * n1 + m2 * n2 + ...
2924  *
2925  * number of rescanned tuples = (m1 - 1) * n1 + (m2 - 1) * n2 + ... = m1 *
2926  * n1 + m2 * n2 + ... - (n1 + n2 + ...) = size of join - size of inner
2927  * relation
2928  *
2929  * This equation works correctly for outer tuples having no inner match
2930  * (nk = 0), but not for inner tuples having no outer match (mk = 0); we
2931  * are effectively subtracting those from the number of rescanned tuples,
2932  * when we should not. Can we do better without expensive selectivity
2933  * computations?
2934  *
2935  * The whole issue is moot if we are working from a unique-ified outer
2936  * input, or if we know we don't need to mark/restore at all.
2937  */
2938  if (IsA(outer_path, UniquePath) ||path->skip_mark_restore)
2939  rescannedtuples = 0;
2940  else
2941  {
2942  rescannedtuples = mergejointuples - inner_path_rows;
2943  /* Must clamp because of possible underestimate */
2944  if (rescannedtuples < 0)
2945  rescannedtuples = 0;
2946  }
2947  /* We'll inflate various costs this much to account for rescanning */
2948  rescanratio = 1.0 + (rescannedtuples / inner_path_rows);
2949 
2950  /*
2951  * Decide whether we want to materialize the inner input to shield it from
2952  * mark/restore and performing re-fetches. Our cost model for regular
2953  * re-fetches is that a re-fetch costs the same as an original fetch,
2954  * which is probably an overestimate; but on the other hand we ignore the
2955  * bookkeeping costs of mark/restore. Not clear if it's worth developing
2956  * a more refined model. So we just need to inflate the inner run cost by
2957  * rescanratio.
2958  */
2959  bare_inner_cost = inner_run_cost * rescanratio;
2960 
2961  /*
2962  * When we interpose a Material node the re-fetch cost is assumed to be
2963  * just cpu_operator_cost per tuple, independently of the underlying
2964  * plan's cost; and we charge an extra cpu_operator_cost per original
2965  * fetch as well. Note that we're assuming the materialize node will
2966  * never spill to disk, since it only has to remember tuples back to the
2967  * last mark. (If there are a huge number of duplicates, our other cost
2968  * factors will make the path so expensive that it probably won't get
2969  * chosen anyway.) So we don't use cost_rescan here.
2970  *
2971  * Note: keep this estimate in sync with create_mergejoin_plan's labeling
2972  * of the generated Material node.
2973  */
2974  mat_inner_cost = inner_run_cost +
2975  cpu_operator_cost * inner_path_rows * rescanratio;
2976 
2977  /*
2978  * If we don't need mark/restore at all, we don't need materialization.
2979  */
2980  if (path->skip_mark_restore)
2981  path->materialize_inner = false;
2982 
2983  /*
2984  * Prefer materializing if it looks cheaper, unless the user has asked to
2985  * suppress materialization.
2986  */
2987  else if (enable_material && mat_inner_cost < bare_inner_cost)
2988  path->materialize_inner = true;
2989 
2990  /*
2991  * Even if materializing doesn't look cheaper, we *must* do it if the
2992  * inner path is to be used directly (without sorting) and it doesn't
2993  * support mark/restore.
2994  *
2995  * Since the inner side must be ordered, and only Sorts and IndexScans can
2996  * create order to begin with, and they both support mark/restore, you
2997  * might think there's no problem --- but you'd be wrong. Nestloop and
2998  * merge joins can *preserve* the order of their inputs, so they can be
2999  * selected as the input of a mergejoin, and they don't support
3000  * mark/restore at present.
3001  *
3002  * We don't test the value of enable_material here, because
3003  * materialization is required for correctness in this case, and turning
3004  * it off does not entitle us to deliver an invalid plan.
3005  */
3006  else if (innersortkeys == NIL &&
3007  !ExecSupportsMarkRestore(inner_path))
3008  path->materialize_inner = true;
3009 
3010  /*
3011  * Also, force materializing if the inner path is to be sorted and the
3012  * sort is expected to spill to disk. This is because the final merge
3013  * pass can be done on-the-fly if it doesn't have to support mark/restore.
3014  * We don't try to adjust the cost estimates for this consideration,
3015  * though.
3016  *
3017  * Since materialization is a performance optimization in this case,
3018  * rather than necessary for correctness, we skip it if enable_material is
3019  * off.
3020  */
3021  else if (enable_material && innersortkeys != NIL &&
3022  relation_byte_size(inner_path_rows,
3023  inner_path->pathtarget->width) >
3024  (work_mem * 1024L))
3025  path->materialize_inner = true;
3026  else
3027  path->materialize_inner = false;
3028 
3029  /* Charge the right incremental cost for the chosen case */
3030  if (path->materialize_inner)
3031  run_cost += mat_inner_cost;
3032  else
3033  run_cost += bare_inner_cost;
3034 
3035  /* CPU costs */
3036 
3037  /*
3038  * The number of tuple comparisons needed is approximately number of outer
3039  * rows plus number of inner rows plus number of rescanned tuples (can we
3040  * refine this?). At each one, we need to evaluate the mergejoin quals.
3041  */
3042  startup_cost += merge_qual_cost.startup;
3043  startup_cost += merge_qual_cost.per_tuple *
3044  (outer_skip_rows + inner_skip_rows * rescanratio);
3045  run_cost += merge_qual_cost.per_tuple *
3046  ((outer_rows - outer_skip_rows) +
3047  (inner_rows - inner_skip_rows) * rescanratio);
3048 
3049  /*
3050  * For each tuple that gets through the mergejoin proper, we charge
3051  * cpu_tuple_cost plus the cost of evaluating additional restriction
3052  * clauses that are to be applied at the join. (This is pessimistic since
3053  * not all of the quals may get evaluated at each tuple.)
3054  *
3055  * Note: we could adjust for SEMI/ANTI joins skipping some qual
3056  * evaluations here, but it's probably not worth the trouble.
3057  */
3058  startup_cost += qp_qual_cost.startup;
3059  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
3060  run_cost += cpu_per_tuple * mergejointuples;
3061 
3062  /* tlist eval costs are paid per output row, not per tuple scanned */
3063  startup_cost += path->jpath.path.pathtarget->cost.startup;
3064  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3065 
3066  path->jpath.path.startup_cost = startup_cost;
3067  path->jpath.path.total_cost = startup_cost + run_cost;
3068 }
3069 
3070 /*
3071  * run mergejoinscansel() with caching
3072  */
3073 static MergeScanSelCache *
3075 {
3076  MergeScanSelCache *cache;
3077  ListCell *lc;
3078  Selectivity leftstartsel,
3079  leftendsel,
3080  rightstartsel,
3081  rightendsel;
3082  MemoryContext oldcontext;
3083 
3084  /* Do we have this result already? */
3085  foreach(lc, rinfo->scansel_cache)
3086  {
3087  cache = (MergeScanSelCache *) lfirst(lc);
3088  if (cache->opfamily == pathkey->pk_opfamily &&
3089  cache->collation == pathkey->pk_eclass->ec_collation &&
3090  cache->strategy == pathkey->pk_strategy &&
3091  cache->nulls_first == pathkey->pk_nulls_first)
3092  return cache;
3093  }
3094 
3095  /* Nope, do the computation */
3096  mergejoinscansel(root,
3097  (Node *) rinfo->clause,
3098  pathkey->pk_opfamily,
3099  pathkey->pk_strategy,
3100  pathkey->pk_nulls_first,
3101  &leftstartsel,
3102  &leftendsel,
3103  &rightstartsel,
3104  &rightendsel);
3105 
3106  /* Cache the result in suitably long-lived workspace */
3107  oldcontext = MemoryContextSwitchTo(root->planner_cxt);
3108 
3109  cache = (MergeScanSelCache *) palloc(sizeof(MergeScanSelCache));
3110  cache->opfamily = pathkey->pk_opfamily;
3111  cache->collation = pathkey->pk_eclass->ec_collation;
3112  cache->strategy = pathkey->pk_strategy;
3113  cache->nulls_first = pathkey->pk_nulls_first;
3114  cache->leftstartsel = leftstartsel;
3115  cache->leftendsel = leftendsel;
3116  cache->rightstartsel = rightstartsel;
3117  cache->rightendsel = rightendsel;
3118 
3119  rinfo->scansel_cache = lappend(rinfo->scansel_cache, cache);
3120 
3121  MemoryContextSwitchTo(oldcontext);
3122 
3123  return cache;
3124 }
3125 
3126 /*
3127  * initial_cost_hashjoin
3128  * Preliminary estimate of the cost of a hashjoin path.
3129  *
3130  * This must quickly produce lower-bound estimates of the path's startup and
3131  * total costs. If we are unable to eliminate the proposed path from
3132  * consideration using the lower bounds, final_cost_hashjoin will be called
3133  * to obtain the final estimates.
3134  *
3135  * The exact division of labor between this function and final_cost_hashjoin
3136  * is private to them, and represents a tradeoff between speed of the initial
3137  * estimate and getting a tight lower bound. We choose to not examine the
3138  * join quals here (other than by counting the number of hash clauses),
3139  * so we can't do much with CPU costs. We do assume that
3140  * ExecChooseHashTableSize is cheap enough to use here.
3141  *
3142  * 'workspace' is to be filled with startup_cost, total_cost, and perhaps
3143  * other data to be used by final_cost_hashjoin
3144  * 'jointype' is the type of join to be performed
3145  * 'hashclauses' is the list of joinclauses to be used as hash clauses
3146  * 'outer_path' is the outer input to the join
3147  * 'inner_path' is the inner input to the join
3148  * 'extra' contains miscellaneous information about the join
3149  * 'parallel_hash' indicates that inner_path is partial and that a shared
3150  * hash table will be built in parallel
3151  */
3152 void
3154  JoinType jointype,
3155  List *hashclauses,
3156  Path *outer_path, Path *inner_path,
3157  JoinPathExtraData *extra,
3158  bool parallel_hash)
3159 {
3160  Cost startup_cost = 0;
3161  Cost run_cost = 0;
3162  double outer_path_rows = outer_path->rows;
3163  double inner_path_rows = inner_path->rows;
3164  double inner_path_rows_total = inner_path_rows;
3165  int num_hashclauses = list_length(hashclauses);
3166  int numbuckets;
3167  int numbatches;
3168  int num_skew_mcvs;
3169  size_t space_allowed; /* unused */
3170 
3171  /* cost of source data */
3172  startup_cost += outer_path->startup_cost;
3173  run_cost += outer_path->total_cost - outer_path->startup_cost;
3174  startup_cost += inner_path->total_cost;
3175 
3176  /*
3177  * Cost of computing hash function: must do it once per input tuple. We
3178  * charge one cpu_operator_cost for each column's hash function. Also,
3179  * tack on one cpu_tuple_cost per inner row, to model the costs of
3180  * inserting the row into the hashtable.
3181  *
3182  * XXX when a hashclause is more complex than a single operator, we really
3183  * should charge the extra eval costs of the left or right side, as
3184  * appropriate, here. This seems more work than it's worth at the moment.
3185  */
3186  startup_cost += (cpu_operator_cost * num_hashclauses + cpu_tuple_cost)
3187  * inner_path_rows;
3188  run_cost += cpu_operator_cost * num_hashclauses * outer_path_rows;
3189 
3190  /*
3191  * If this is a parallel hash build, then the value we have for
3192  * inner_rows_total currently refers only to the rows returned by each
3193  * participant. For shared hash table size estimation, we need the total
3194  * number, so we need to undo the division.
3195  */
3196  if (parallel_hash)
3197  inner_path_rows_total *= get_parallel_divisor(inner_path);
3198 
3199  /*
3200  * Get hash table size that executor would use for inner relation.
3201  *
3202  * XXX for the moment, always assume that skew optimization will be
3203  * performed. As long as SKEW_WORK_MEM_PERCENT is small, it's not worth
3204  * trying to determine that for sure.
3205  *
3206  * XXX at some point it might be interesting to try to account for skew
3207  * optimization in the cost estimate, but for now, we don't.
3208  */
3209  ExecChooseHashTableSize(inner_path_rows_total,
3210  inner_path->pathtarget->width,
3211  true, /* useskew */
3212  parallel_hash, /* try_combined_work_mem */
3213  outer_path->parallel_workers,
3214  &space_allowed,
3215  &numbuckets,
3216  &numbatches,
3217  &num_skew_mcvs);
3218 
3219  /*
3220  * If inner relation is too big then we will need to "batch" the join,
3221  * which implies writing and reading most of the tuples to disk an extra
3222  * time. Charge seq_page_cost per page, since the I/O should be nice and
3223  * sequential. Writing the inner rel counts as startup cost, all the rest
3224  * as run cost.
3225  */
3226  if (numbatches > 1)
3227  {
3228  double outerpages = page_size(outer_path_rows,
3229  outer_path->pathtarget->width);
3230  double innerpages = page_size(inner_path_rows,
3231  inner_path->pathtarget->width);
3232 
3233  startup_cost += seq_page_cost * innerpages;
3234  run_cost += seq_page_cost * (innerpages + 2 * outerpages);
3235  }
3236 
3237  /* CPU costs left for later */
3238 
3239  /* Public result fields */
3240  workspace->startup_cost = startup_cost;
3241  workspace->total_cost = startup_cost + run_cost;
3242  /* Save private data for final_cost_hashjoin */
3243  workspace->run_cost = run_cost;
3244  workspace->numbuckets = numbuckets;
3245  workspace->numbatches = numbatches;
3246  workspace->inner_rows_total = inner_path_rows_total;
3247 }
3248 
3249 /*
3250  * final_cost_hashjoin
3251  * Final estimate of the cost and result size of a hashjoin path.
3252  *
3253  * Note: the numbatches estimate is also saved into 'path' for use later
3254  *
3255  * 'path' is already filled in except for the rows and cost fields and
3256  * num_batches
3257  * 'workspace' is the result from initial_cost_hashjoin
3258  * 'extra' contains miscellaneous information about the join
3259  */
3260 void
3262  JoinCostWorkspace *workspace,
3263  JoinPathExtraData *extra)
3264 {
3265  Path *outer_path = path->jpath.outerjoinpath;
3266  Path *inner_path = path->jpath.innerjoinpath;
3267  double outer_path_rows = outer_path->rows;
3268  double inner_path_rows = inner_path->rows;
3269  double inner_path_rows_total = workspace->inner_rows_total;
3270  List *hashclauses = path->path_hashclauses;
3271  Cost startup_cost = workspace->startup_cost;
3272  Cost run_cost = workspace->run_cost;
3273  int numbuckets = workspace->numbuckets;
3274  int numbatches = workspace->numbatches;
3275  Cost cpu_per_tuple;
3276  QualCost hash_qual_cost;
3277  QualCost qp_qual_cost;
3278  double hashjointuples;
3279  double virtualbuckets;
3280  Selectivity innerbucketsize;
3281  Selectivity innermcvfreq;
3282  ListCell *hcl;
3283 
3284  /* Mark the path with the correct row estimate */
3285  if (path->jpath.path.param_info)
3286  path->jpath.path.rows = path->jpath.path.param_info->ppi_rows;
3287  else
3288  path->jpath.path.rows = path->jpath.path.parent->rows;
3289 
3290  /* For partial paths, scale row estimate. */
3291  if (path->jpath.path.parallel_workers > 0)
3292  {
3293  double parallel_divisor = get_parallel_divisor(&path->jpath.path);
3294 
3295  path->jpath.path.rows =
3296  clamp_row_est(path->jpath.path.rows / parallel_divisor);
3297  }
3298 
3299  /*
3300  * We could include disable_cost in the preliminary estimate, but that
3301  * would amount to optimizing for the case where the join method is
3302  * disabled, which doesn't seem like the way to bet.
3303  */
3304  if (!enable_hashjoin)
3305  startup_cost += disable_cost;
3306 
3307  /* mark the path with estimated # of batches */
3308  path->num_batches = numbatches;
3309 
3310  /* store the total number of tuples (sum of partial row estimates) */
3311  path->inner_rows_total = inner_path_rows_total;
3312 
3313  /* and compute the number of "virtual" buckets in the whole join */
3314  virtualbuckets = (double) numbuckets * (double) numbatches;
3315 
3316  /*
3317  * Determine bucketsize fraction and MCV frequency for the inner relation.
3318  * We use the smallest bucketsize or MCV frequency estimated for any
3319  * individual hashclause; this is undoubtedly conservative.
3320  *
3321  * BUT: if inner relation has been unique-ified, we can assume it's good
3322  * for hashing. This is important both because it's the right answer, and
3323  * because we avoid contaminating the cache with a value that's wrong for
3324  * non-unique-ified paths.
3325  */
3326  if (IsA(inner_path, UniquePath))
3327  {
3328  innerbucketsize = 1.0 / virtualbuckets;
3329  innermcvfreq = 0.0;
3330  }
3331  else
3332  {
3333  innerbucketsize = 1.0;
3334  innermcvfreq = 1.0;
3335  foreach(hcl, hashclauses)
3336  {
3337  RestrictInfo *restrictinfo = lfirst_node(RestrictInfo, hcl);
3338  Selectivity thisbucketsize;
3339  Selectivity thismcvfreq;
3340 
3341  /*
3342  * First we have to figure out which side of the hashjoin clause
3343  * is the inner side.
3344  *
3345  * Since we tend to visit the same clauses over and over when
3346  * planning a large query, we cache the bucket stats estimates in
3347  * the RestrictInfo node to avoid repeated lookups of statistics.
3348  */
3349  if (bms_is_subset(restrictinfo->right_relids,
3350  inner_path->parent->relids))
3351  {
3352  /* righthand side is inner */
3353  thisbucketsize = restrictinfo->right_bucketsize;
3354  if (thisbucketsize < 0)
3355  {
3356  /* not cached yet */
3358  get_rightop(restrictinfo->clause),
3359  virtualbuckets,
3360  &restrictinfo->right_mcvfreq,
3361  &restrictinfo->right_bucketsize);
3362  thisbucketsize = restrictinfo->right_bucketsize;
3363  }
3364  thismcvfreq = restrictinfo->right_mcvfreq;
3365  }
3366  else
3367  {
3368  Assert(bms_is_subset(restrictinfo->left_relids,
3369  inner_path->parent->relids));
3370  /* lefthand side is inner */
3371  thisbucketsize = restrictinfo->left_bucketsize;
3372  if (thisbucketsize < 0)
3373  {
3374  /* not cached yet */
3376  get_leftop(restrictinfo->clause),
3377  virtualbuckets,
3378  &restrictinfo->left_mcvfreq,
3379  &restrictinfo->left_bucketsize);
3380  thisbucketsize = restrictinfo->left_bucketsize;
3381  }
3382  thismcvfreq = restrictinfo->left_mcvfreq;
3383  }
3384 
3385  if (innerbucketsize > thisbucketsize)
3386  innerbucketsize = thisbucketsize;
3387  if (innermcvfreq > thismcvfreq)
3388  innermcvfreq = thismcvfreq;
3389  }
3390  }
3391 
3392  /*
3393  * If the bucket holding the inner MCV would exceed work_mem, we don't
3394  * want to hash unless there is really no other alternative, so apply
3395  * disable_cost. (The executor normally copes with excessive memory usage
3396  * by splitting batches, but obviously it cannot separate equal values
3397  * that way, so it will be unable to drive the batch size below work_mem
3398  * when this is true.)
3399  */
3400  if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
3401  inner_path->pathtarget->width) >
3402  (work_mem * 1024L))
3403  startup_cost += disable_cost;
3404 
3405  /*
3406  * Compute cost of the hashquals and qpquals (other restriction clauses)
3407  * separately.
3408  */
3409  cost_qual_eval(&hash_qual_cost, hashclauses, root);
3410  cost_qual_eval(&qp_qual_cost, path->jpath.joinrestrictinfo, root);
3411  qp_qual_cost.startup -= hash_qual_cost.startup;
3412  qp_qual_cost.per_tuple -= hash_qual_cost.per_tuple;
3413 
3414  /* CPU costs */
3415 
3416  if (path->jpath.jointype == JOIN_SEMI ||
3417  path->jpath.jointype == JOIN_ANTI ||
3418  extra->inner_unique)
3419  {
3420  double outer_matched_rows;
3421  Selectivity inner_scan_frac;
3422 
3423  /*
3424  * With a SEMI or ANTI join, or if the innerrel is known unique, the
3425  * executor will stop after the first match.
3426  *
3427  * For an outer-rel row that has at least one match, we can expect the
3428  * bucket scan to stop after a fraction 1/(match_count+1) of the
3429  * bucket's rows, if the matches are evenly distributed. Since they
3430  * probably aren't quite evenly distributed, we apply a fuzz factor of
3431  * 2.0 to that fraction. (If we used a larger fuzz factor, we'd have
3432  * to clamp inner_scan_frac to at most 1.0; but since match_count is
3433  * at least 1, no such clamp is needed now.)
3434  */
3435  outer_matched_rows = rint(outer_path_rows * extra->semifactors.outer_match_frac);
3436  inner_scan_frac = 2.0 / (extra->semifactors.match_count + 1.0);
3437 
3438  startup_cost += hash_qual_cost.startup;
3439  run_cost += hash_qual_cost.per_tuple * outer_matched_rows *
3440  clamp_row_est(inner_path_rows * innerbucketsize * inner_scan_frac) * 0.5;
3441 
3442  /*
3443  * For unmatched outer-rel rows, the picture is quite a lot different.
3444  * In the first place, there is no reason to assume that these rows
3445  * preferentially hit heavily-populated buckets; instead assume they
3446  * are uncorrelated with the inner distribution and so they see an
3447  * average bucket size of inner_path_rows / virtualbuckets. In the
3448  * second place, it seems likely that they will have few if any exact
3449  * hash-code matches and so very few of the tuples in the bucket will
3450  * actually require eval of the hash quals. We don't have any good
3451  * way to estimate how many will, but for the moment assume that the
3452  * effective cost per bucket entry is one-tenth what it is for
3453  * matchable tuples.
3454  */
3455  run_cost += hash_qual_cost.per_tuple *
3456  (outer_path_rows - outer_matched_rows) *
3457  clamp_row_est(inner_path_rows / virtualbuckets) * 0.05;
3458 
3459  /* Get # of tuples that will pass the basic join */
3460  if (path->jpath.jointype == JOIN_SEMI)
3461  hashjointuples = outer_matched_rows;
3462  else
3463  hashjointuples = outer_path_rows - outer_matched_rows;
3464  }
3465  else
3466  {
3467  /*
3468  * The number of tuple comparisons needed is the number of outer
3469  * tuples times the typical number of tuples in a hash bucket, which
3470  * is the inner relation size times its bucketsize fraction. At each
3471  * one, we need to evaluate the hashjoin quals. But actually,
3472  * charging the full qual eval cost at each tuple is pessimistic,
3473  * since we don't evaluate the quals unless the hash values match
3474  * exactly. For lack of a better idea, halve the cost estimate to
3475  * allow for that.
3476  */
3477  startup_cost += hash_qual_cost.startup;
3478  run_cost += hash_qual_cost.per_tuple * outer_path_rows *
3479  clamp_row_est(inner_path_rows * innerbucketsize) * 0.5;
3480 
3481  /*
3482  * Get approx # tuples passing the hashquals. We use
3483  * approx_tuple_count here because we need an estimate done with
3484  * JOIN_INNER semantics.
3485  */
3486  hashjointuples = approx_tuple_count(root, &path->jpath, hashclauses);
3487  }
3488 
3489  /*
3490  * For each tuple that gets through the hashjoin proper, we charge
3491  * cpu_tuple_cost plus the cost of evaluating additional restriction
3492  * clauses that are to be applied at the join. (This is pessimistic since
3493  * not all of the quals may get evaluated at each tuple.)
3494  */
3495  startup_cost += qp_qual_cost.startup;
3496  cpu_per_tuple = cpu_tuple_cost + qp_qual_cost.per_tuple;
3497  run_cost += cpu_per_tuple * hashjointuples;
3498 
3499  /* tlist eval costs are paid per output row, not per tuple scanned */
3500  startup_cost += path->jpath.path.pathtarget->cost.startup;
3501  run_cost += path->jpath.path.pathtarget->cost.per_tuple * path->jpath.path.rows;
3502 
3503  path->jpath.path.startup_cost = startup_cost;
3504  path->jpath.path.total_cost = startup_cost + run_cost;
3505 }
3506 
3507 
3508 /*
3509  * cost_subplan
3510  * Figure the costs for a SubPlan (or initplan).
3511  *
3512  * Note: we could dig the subplan's Plan out of the root list, but in practice
3513  * all callers have it handy already, so we make them pass it.
3514  */
3515 void
3516 cost_subplan(PlannerInfo *root, SubPlan *subplan, Plan *plan)
3517 {
3518  QualCost sp_cost;
3519 
3520  /* Figure any cost for evaluating the testexpr */
3521  cost_qual_eval(&sp_cost,
3522  make_ands_implicit((Expr *) subplan->testexpr),
3523  root);
3524 
3525  if (subplan->useHashTable)
3526  {
3527  /*
3528  * If we are using a hash table for the subquery outputs, then the
3529  * cost of evaluating the query is a one-time cost. We charge one
3530  * cpu_operator_cost per tuple for the work of loading the hashtable,
3531  * too.
3532  */
3533  sp_cost.startup += plan->total_cost +
3534  cpu_operator_cost * plan->plan_rows;
3535 
3536  /*
3537  * The per-tuple costs include the cost of evaluating the lefthand
3538  * expressions, plus the cost of probing the hashtable. We already
3539  * accounted for the lefthand expressions as part of the testexpr, and
3540  * will also have counted one cpu_operator_cost for each comparison
3541  * operator. That is probably too low for the probing cost, but it's
3542  * hard to make a better estimate, so live with it for now.
3543  */
3544  }
3545  else
3546  {
3547  /*
3548  * Otherwise we will be rescanning the subplan output on each
3549  * evaluation. We need to estimate how much of the output we will
3550  * actually need to scan. NOTE: this logic should agree with the
3551  * tuple_fraction estimates used by make_subplan() in
3552  * plan/subselect.c.
3553  */
3554  Cost plan_run_cost = plan->total_cost - plan->startup_cost;
3555 
3556  if (subplan->subLinkType == EXISTS_SUBLINK)
3557  {
3558  /* we only need to fetch 1 tuple; clamp to avoid zero divide */
3559  sp_cost.per_tuple += plan_run_cost / clamp_row_est(plan->plan_rows);
3560  }
3561  else if (subplan->subLinkType == ALL_SUBLINK ||
3562  subplan->subLinkType == ANY_SUBLINK)
3563  {
3564  /* assume we need 50% of the tuples */
3565  sp_cost.per_tuple += 0.50 * plan_run_cost;
3566  /* also charge a cpu_operator_cost per row examined */
3567  sp_cost.per_tuple += 0.50 * plan->plan_rows * cpu_operator_cost;
3568  }
3569  else
3570  {
3571  /* assume we need all tuples */
3572  sp_cost.per_tuple += plan_run_cost;
3573  }
3574 
3575  /*
3576  * Also account for subplan's startup cost. If the subplan is
3577  * uncorrelated or undirect correlated, AND its topmost node is one
3578  * that materializes its output, assume that we'll only need to pay
3579  * its startup cost once; otherwise assume we pay the startup cost
3580  * every time.
3581  */
3582  if (subplan->parParam == NIL &&
3584  sp_cost.startup += plan->startup_cost;
3585  else
3586  sp_cost.per_tuple += plan->startup_cost;
3587  }
3588 
3589  subplan->startup_cost = sp_cost.startup;
3590  subplan->per_call_cost = sp_cost.per_tuple;
3591 }
3592 
3593 
3594 /*
3595  * cost_rescan
3596  * Given a finished Path, estimate the costs of rescanning it after
3597  * having done so the first time. For some Path types a rescan is
3598  * cheaper than an original scan (if no parameters change), and this
3599  * function embodies knowledge about that. The default is to return
3600  * the same costs stored in the Path. (Note that the cost estimates
3601  * actually stored in Paths are always for first scans.)
3602  *
3603  * This function is not currently intended to model effects such as rescans
3604  * being cheaper due to disk block caching; what we are concerned with is
3605  * plan types wherein the executor caches results explicitly, or doesn't
3606  * redo startup calculations, etc.
3607  */
3608 static void
3610  Cost *rescan_startup_cost, /* output parameters */
3611  Cost *rescan_total_cost)
3612 {
3613  switch (path->pathtype)
3614  {
3615  case T_FunctionScan:
3616 
3617  /*
3618  * Currently, nodeFunctionscan.c always executes the function to
3619  * completion before returning any rows, and caches the results in
3620  * a tuplestore. So the function eval cost is all startup cost
3621  * and isn't paid over again on rescans. However, all run costs
3622  * will be paid over again.
3623  */
3624  *rescan_startup_cost = 0;
3625  *rescan_total_cost = path->total_cost - path->startup_cost;
3626  break;
3627  case T_HashJoin:
3628 
3629  /*
3630  * If it's a single-batch join, we don't need to rebuild the hash
3631  * table during a rescan.
3632  */
3633  if (((HashPath *) path)->num_batches == 1)
3634  {
3635  /* Startup cost is exactly the cost of hash table building */
3636  *rescan_startup_cost = 0;
3637  *rescan_total_cost = path->total_cost - path->startup_cost;
3638  }
3639  else
3640  {
3641  /* Otherwise, no special treatment */
3642  *rescan_startup_cost = path->startup_cost;
3643  *rescan_total_cost = path->total_cost;
3644  }
3645  break;
3646  case T_CteScan:
3647  case T_WorkTableScan:
3648  {
3649  /*
3650  * These plan types materialize their final result in a
3651  * tuplestore or tuplesort object. So the rescan cost is only
3652  * cpu_tuple_cost per tuple, unless the result is large enough
3653  * to spill to disk.
3654  */
3655  Cost run_cost = cpu_tuple_cost * path->rows;
3656  double nbytes = relation_byte_size(path->rows,
3657  path->pathtarget->width);
3658  long work_mem_bytes = work_mem * 1024L;
3659 
3660  if (nbytes > work_mem_bytes)
3661  {
3662  /* It will spill, so account for re-read cost */
3663  double npages = ceil(nbytes / BLCKSZ);
3664 
3665  run_cost += seq_page_cost * npages;
3666  }
3667  *rescan_startup_cost = 0;
3668  *rescan_total_cost = run_cost;
3669  }
3670  break;
3671  case T_Material:
3672  case T_Sort:
3673  {
3674  /*
3675  * These plan types not only materialize their results, but do
3676  * not implement qual filtering or projection. So they are
3677  * even cheaper to rescan than the ones above. We charge only
3678  * cpu_operator_cost per tuple. (Note: keep that in sync with
3679  * the run_cost charge in cost_sort, and also see comments in
3680  * cost_material before you change it.)
3681  */
3682  Cost run_cost = cpu_operator_cost * path->rows;
3683  double nbytes = relation_byte_size(path->rows,
3684  path->pathtarget->width);
3685  long work_mem_bytes = work_mem * 1024L;
3686 
3687  if (nbytes > work_mem_bytes)
3688  {
3689  /* It will spill, so account for re-read cost */
3690  double npages = ceil(nbytes / BLCKSZ);
3691 
3692  run_cost += seq_page_cost * npages;
3693  }
3694  *rescan_startup_cost = 0;
3695  *rescan_total_cost = run_cost;
3696  }
3697  break;
3698  default:
3699  *rescan_startup_cost = path->startup_cost;
3700  *rescan_total_cost = path->total_cost;
3701  break;
3702  }
3703 }
3704 
3705 
3706 /*
3707  * cost_qual_eval
3708  * Estimate the CPU costs of evaluating a WHERE clause.
3709  * The input can be either an implicitly-ANDed list of boolean
3710  * expressions, or a list of RestrictInfo nodes. (The latter is
3711  * preferred since it allows caching of the results.)
3712  * The result includes both a one-time (startup) component,
3713  * and a per-evaluation component.
3714  */
3715 void
3717 {
3718  cost_qual_eval_context context;
3719  ListCell *l;
3720 
3721  context.root = root;
3722  context.total.startup = 0;
3723  context.total.per_tuple = 0;
3724 
3725  /* We don't charge any cost for the implicit ANDing at top level ... */
3726 
3727  foreach(l, quals)
3728  {
3729  Node *qual = (Node *) lfirst(l);
3730 
3731  cost_qual_eval_walker(qual, &context);
3732  }
3733 
3734  *cost = context.total;
3735 }
3736 
3737 /*
3738  * cost_qual_eval_node
3739  * As above, for a single RestrictInfo or expression.
3740  */
3741 void
3743 {
3744  cost_qual_eval_context context;
3745 
3746  context.root = root;
3747  context.total.startup = 0;
3748  context.total.per_tuple = 0;
3749 
3750  cost_qual_eval_walker(qual, &context);
3751 
3752  *cost = context.total;
3753 }
3754 
3755 static bool
3757 {
3758  if (node == NULL)
3759  return false;
3760 
3761  /*
3762  * RestrictInfo nodes contain an eval_cost field reserved for this
3763  * routine's use, so that it's not necessary to evaluate the qual clause's
3764  * cost more than once. If the clause's cost hasn't been computed yet,
3765  * the field's startup value will contain -1.
3766  */
3767  if (IsA(node, RestrictInfo))
3768  {
3769  RestrictInfo *rinfo = (RestrictInfo *) node;
3770 
3771  if (rinfo->eval_cost.startup < 0)
3772  {
3773  cost_qual_eval_context locContext;
3774 
3775  locContext.root = context->root;
3776  locContext.total.startup = 0;
3777  locContext.total.per_tuple = 0;
3778 
3779  /*
3780  * For an OR clause, recurse into the marked-up tree so that we
3781  * set the eval_cost for contained RestrictInfos too.
3782  */
3783  if (rinfo->orclause)
3784  cost_qual_eval_walker((Node *) rinfo->orclause, &locContext);
3785  else
3786  cost_qual_eval_walker((Node *) rinfo->clause, &locContext);
3787 
3788  /*
3789  * If the RestrictInfo is marked pseudoconstant, it will be tested
3790  * only once, so treat its cost as all startup cost.
3791  */
3792  if (rinfo->pseudoconstant)
3793  {
3794  /* count one execution during startup */
3795  locContext.total.startup += locContext.total.per_tuple;
3796  locContext.total.per_tuple = 0;
3797  }
3798  rinfo->eval_cost = locContext.total;
3799  }
3800  context->total.startup += rinfo->eval_cost.startup;
3801  context->total.per_tuple += rinfo->eval_cost.per_tuple;
3802  /* do NOT recurse into children */
3803  return false;
3804  }
3805 
3806  /*
3807  * For each operator or function node in the given tree, we charge the
3808  * estimated execution cost given by pg_proc.procost (remember to multiply
3809  * this by cpu_operator_cost).
3810  *
3811  * Vars and Consts are charged zero, and so are boolean operators (AND,
3812  * OR, NOT). Simplistic, but a lot better than no model at all.
3813  *
3814  * Should we try to account for the possibility of short-circuit
3815  * evaluation of AND/OR? Probably *not*, because that would make the
3816  * results depend on the clause ordering, and we are not in any position
3817  * to expect that the current ordering of the clauses is the one that's
3818  * going to end up being used. The above per-RestrictInfo caching would
3819  * not mix well with trying to re-order clauses anyway.
3820  *
3821  * Another issue that is entirely ignored here is that if a set-returning
3822  * function is below top level in the tree, the functions/operators above
3823  * it will need to be evaluated multiple times. In practical use, such
3824  * cases arise so seldom as to not be worth the added complexity needed;
3825  * moreover, since our rowcount estimates for functions tend to be pretty
3826  * phony, the results would also be pretty phony.
3827  */
3828  if (IsA(node, FuncExpr))
3829  {
3830  context->total.per_tuple +=
3831  get_func_cost(((FuncExpr *) node)->funcid) * cpu_operator_cost;
3832  }
3833  else if (IsA(node, OpExpr) ||
3834  IsA(node, DistinctExpr) ||
3835  IsA(node, NullIfExpr))
3836  {
3837  /* rely on struct equivalence to treat these all alike */
3838  set_opfuncid((OpExpr *) node);
3839  context->total.per_tuple +=
3840  get_func_cost(((OpExpr *) node)->opfuncid) * cpu_operator_cost;
3841  }
3842  else if (IsA(node, ScalarArrayOpExpr))
3843  {
3844  /*
3845  * Estimate that the operator will be applied to about half of the
3846  * array elements before the answer is determined.
3847  */
3848  ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) node;
3849  Node *arraynode = (Node *) lsecond(saop->args);
3850 
3851  set_sa_opfuncid(saop);
3852  context->total.per_tuple += get_func_cost(saop->opfuncid) *
3853  cpu_operator_cost * estimate_array_length(arraynode) * 0.5;
3854  }
3855  else if (IsA(node, Aggref) ||
3856  IsA(node, WindowFunc))
3857  {
3858  /*
3859  * Aggref and WindowFunc nodes are (and should be) treated like Vars,
3860  * ie, zero execution cost in the current model, because they behave
3861  * essentially like Vars at execution. We disregard the costs of
3862  * their input expressions for the same reason. The actual execution
3863  * costs of the aggregate/window functions and their arguments have to
3864  * be factored into plan-node-specific costing of the Agg or WindowAgg
3865  * plan node.
3866  */
3867  return false; /* don't recurse into children */
3868  }
3869  else if (IsA(node, CoerceViaIO))
3870  {
3871  CoerceViaIO *iocoerce = (CoerceViaIO *) node;
3872  Oid iofunc;
3873  Oid typioparam;
3874  bool typisvarlena;
3875 
3876  /* check the result type's input function */
3877  getTypeInputInfo(iocoerce->resulttype,
3878  &iofunc, &typioparam);
3879  context->total.per_tuple += get_func_cost(iofunc) * cpu_operator_cost;
3880  /* check the input type's output function */
3881  getTypeOutputInfo(exprType((Node *) iocoerce->arg),
3882  &iofunc, &typisvarlena);
3883  context->total.per_tuple += get_func_cost(iofunc) * cpu_operator_cost;
3884  }
3885  else if (IsA(node, ArrayCoerceExpr))
3886  {
3887  ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
3888  QualCost perelemcost;
3889 
3890  cost_qual_eval_node(&perelemcost, (Node *) acoerce->elemexpr,
3891  context->root);
3892  context->total.startup += perelemcost.startup;
3893  if (perelemcost.per_tuple > 0)
3894  context->total.per_tuple += perelemcost.per_tuple *
3895  estimate_array_length((Node *) acoerce->arg);
3896  }
3897  else if (IsA(node, RowCompareExpr))
3898  {
3899  /* Conservatively assume we will check all the columns */
3900  RowCompareExpr *rcexpr = (RowCompareExpr *) node;
3901  ListCell *lc;
3902 
3903  foreach(lc, rcexpr->opnos)
3904  {
3905  Oid opid = lfirst_oid(lc);
3906 
3907  context->total.per_tuple += get_func_cost(get_opcode(opid)) *
3909  }
3910  }
3911  else if (IsA(node, MinMaxExpr) ||
3912  IsA(node, SQLValueFunction) ||
3913  IsA(node, XmlExpr) ||
3914  IsA(node, CoerceToDomain) ||
3915  IsA(node, NextValueExpr))
3916  {
3917  /* Treat all these as having cost 1 */
3918  context->total.per_tuple += cpu_operator_cost;
3919  }
3920  else if (IsA(node, CurrentOfExpr))
3921  {
3922  /* Report high cost to prevent selection of anything but TID scan */
3923  context->total.startup += disable_cost;
3924  }
3925  else if (IsA(node, SubLink))
3926  {
3927  /* This routine should not be applied to un-planned expressions */
3928  elog(ERROR, "cannot handle unplanned sub-select");
3929  }
3930  else if (IsA(node, SubPlan))
3931  {
3932  /*
3933  * A subplan node in an expression typically indicates that the
3934  * subplan will be executed on each evaluation, so charge accordingly.
3935  * (Sub-selects that can be executed as InitPlans have already been
3936  * removed from the expression.)
3937  */
3938  SubPlan *subplan = (SubPlan *) node;
3939 
3940  context->total.startup += subplan->startup_cost;
3941  context->total.per_tuple += subplan->per_call_cost;
3942 
3943  /*
3944  * We don't want to recurse into the testexpr, because it was already
3945  * counted in the SubPlan node's costs. So we're done.
3946  */
3947  return false;
3948  }
3949  else if (IsA(node, AlternativeSubPlan))
3950  {
3951  /*
3952  * Arbitrarily use the first alternative plan for costing. (We should
3953  * certainly only include one alternative, and we don't yet have
3954  * enough information to know which one the executor is most likely to
3955  * use.)
3956  */
3957  AlternativeSubPlan *asplan = (AlternativeSubPlan *) node;
3958 
3959  return cost_qual_eval_walker((Node *) linitial(asplan->subplans),
3960  context);
3961  }
3962  else if (IsA(node, PlaceHolderVar))
3963  {
3964  /*
3965  * A PlaceHolderVar should be given cost zero when considering general
3966  * expression evaluation costs. The expense of doing the contained
3967  * expression is charged as part of the tlist eval costs of the scan
3968  * or join where the PHV is first computed (see set_rel_width and
3969  * add_placeholders_to_joinrel). If we charged it again here, we'd be
3970  * double-counting the cost for each level of plan that the PHV
3971  * bubbles up through. Hence, return without recursing into the
3972  * phexpr.
3973  */
3974  return false;
3975  }
3976 
3977  /* recurse into children */
3979  (void *) context);
3980 }
3981 
3982 /*
3983  * get_restriction_qual_cost
3984  * Compute evaluation costs of a baserel's restriction quals, plus any
3985  * movable join quals that have been pushed down to the scan.
3986  * Results are returned into *qpqual_cost.
3987  *
3988  * This is a convenience subroutine that works for seqscans and other cases
3989  * where all the given quals will be evaluated the hard way. It's not useful
3990  * for cost_index(), for example, where the index machinery takes care of
3991  * some of the quals. We assume baserestrictcost was previously set by
3992  * set_baserel_size_estimates().
3993  */
3994 static void
3996  ParamPathInfo *param_info,
3997  QualCost *qpqual_cost)
3998 {
3999  if (param_info)
4000  {
4001  /* Include costs of pushed-down clauses */
4002  cost_qual_eval(qpqual_cost, param_info->ppi_clauses, root);
4003 
4004  qpqual_cost->startup += baserel->baserestrictcost.startup;
4005  qpqual_cost->per_tuple += baserel->baserestrictcost.per_tuple;
4006  }
4007  else
4008  *qpqual_cost = baserel->baserestrictcost;
4009 }
4010 
4011 
4012 /*
4013  * compute_semi_anti_join_factors
4014  * Estimate how much of the inner input a SEMI, ANTI, or inner_unique join
4015  * can be expected to scan.
4016  *
4017  * In a hash or nestloop SEMI/ANTI join, the executor will stop scanning
4018  * inner rows as soon as it finds a match to the current outer row.
4019  * The same happens if we have detected the inner rel is unique.
4020  * We should therefore adjust some of the cost components for this effect.
4021  * This function computes some estimates needed for these adjustments.
4022  * These estimates will be the same regardless of the particular paths used
4023  * for the outer and inner relation, so we compute these once and then pass
4024  * them to all the join cost estimation functions.
4025  *
4026  * Input parameters:
4027  * joinrel: join relation under consideration
4028  * outerrel: outer relation under consideration
4029  * innerrel: inner relation under consideration
4030  * jointype: if not JOIN_SEMI or JOIN_ANTI, we assume it's inner_unique
4031  * sjinfo: SpecialJoinInfo relevant to this join
4032  * restrictlist: join quals
4033  * Output parameters:
4034  * *semifactors is filled in (see relation.h for field definitions)
4035  */
4036 void
4038  RelOptInfo *joinrel,
4039  RelOptInfo *outerrel,
4040  RelOptInfo *innerrel,
4041  JoinType jointype,
4042  SpecialJoinInfo *sjinfo,
4043  List *restrictlist,
4044  SemiAntiJoinFactors *semifactors)
4045 {
4046  Selectivity jselec;
4047  Selectivity nselec;
4048  Selectivity avgmatch;
4049  SpecialJoinInfo norm_sjinfo;
4050  List *joinquals;
4051  ListCell *l;
4052 
4053  /*
4054  * In an ANTI join, we must ignore clauses that are "pushed down", since
4055  * those won't affect the match logic. In a SEMI join, we do not
4056  * distinguish joinquals from "pushed down" quals, so just use the whole
4057  * restrictinfo list. For other outer join types, we should consider only
4058  * non-pushed-down quals, so that this devolves to an IS_OUTER_JOIN check.
4059  */
4060  if (IS_OUTER_JOIN(jointype))
4061  {
4062  joinquals = NIL;
4063  foreach(l, restrictlist)
4064  {
4065  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
4066 
4067  if (!RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
4068  joinquals = lappend(joinquals, rinfo);
4069  }
4070  }
4071  else
4072  joinquals = restrictlist;
4073 
4074  /*
4075  * Get the JOIN_SEMI or JOIN_ANTI selectivity of the join clauses.
4076  */
4077  jselec = clauselist_selectivity(root,
4078  joinquals,
4079  0,
4080  (jointype == JOIN_ANTI) ? JOIN_ANTI : JOIN_SEMI,
4081  sjinfo);
4082 
4083  /*
4084  * Also get the normal inner-join selectivity of the join clauses.
4085  */
4086  norm_sjinfo.type = T_SpecialJoinInfo;
4087  norm_sjinfo.min_lefthand = outerrel->relids;
4088  norm_sjinfo.min_righthand = innerrel->relids;
4089  norm_sjinfo.syn_lefthand = outerrel->relids;
4090  norm_sjinfo.syn_righthand = innerrel->relids;
4091  norm_sjinfo.jointype = JOIN_INNER;
4092  /* we don't bother trying to make the remaining fields valid */
4093  norm_sjinfo.lhs_strict = false;
4094  norm_sjinfo.delay_upper_joins = false;
4095  norm_sjinfo.semi_can_btree = false;
4096  norm_sjinfo.semi_can_hash = false;
4097  norm_sjinfo.semi_operators = NIL;
4098  norm_sjinfo.semi_rhs_exprs = NIL;
4099 
4100  nselec = clauselist_selectivity(root,
4101  joinquals,
4102  0,
4103  JOIN_INNER,
4104  &norm_sjinfo);
4105 
4106  /* Avoid leaking a lot of ListCells */
4107  if (IS_OUTER_JOIN(jointype))
4108  list_free(joinquals);
4109 
4110  /*
4111  * jselec can be interpreted as the fraction of outer-rel rows that have
4112  * any matches (this is true for both SEMI and ANTI cases). And nselec is
4113  * the fraction of the Cartesian product that matches. So, the average
4114  * number of matches for each outer-rel row that has at least one match is
4115  * nselec * inner_rows / jselec.
4116  *
4117  * Note: it is correct to use the inner rel's "rows" count here, even
4118  * though we might later be considering a parameterized inner path with
4119  * fewer rows. This is because we have included all the join clauses in
4120  * the selectivity estimate.
4121  */
4122  if (jselec > 0) /* protect against zero divide */
4123  {
4124  avgmatch = nselec * innerrel->rows / jselec;
4125  /* Clamp to sane range */
4126  avgmatch = Max(1.0, avgmatch);
4127  }
4128  else
4129  avgmatch = 1.0;
4130 
4131  semifactors->outer_match_frac = jselec;
4132  semifactors->match_count = avgmatch;
4133 }
4134 
4135 /*
4136  * has_indexed_join_quals
4137  * Check whether all the joinquals of a nestloop join are used as
4138  * inner index quals.
4139  *
4140  * If the inner path of a SEMI/ANTI join is an indexscan (including bitmap
4141  * indexscan) that uses all the joinquals as indexquals, we can assume that an
4142  * unmatched outer tuple is cheap to process, whereas otherwise it's probably
4143  * expensive.
4144  */
4145 static bool
4147 {
4148  Relids joinrelids = joinpath->path.parent->relids;
4149  Path *innerpath = joinpath->innerjoinpath;
4150  List *indexclauses;
4151  bool found_one;
4152  ListCell *lc;
4153 
4154  /* If join still has quals to evaluate, it's not fast */
4155  if (joinpath->joinrestrictinfo != NIL)
4156  return false;
4157  /* Nor if the inner path isn't parameterized at all */
4158  if (innerpath->param_info == NULL)
4159  return false;
4160 
4161  /* Find the indexclauses list for the inner scan */
4162  switch (innerpath->pathtype)
4163  {
4164  case T_IndexScan:
4165  case T_IndexOnlyScan:
4166  indexclauses = ((IndexPath *) innerpath)->indexclauses;
4167  break;
4168  case T_BitmapHeapScan:
4169  {
4170  /* Accept only a simple bitmap scan, not AND/OR cases */
4171  Path *bmqual = ((BitmapHeapPath *) innerpath)->bitmapqual;
4172 
4173  if (IsA(bmqual, IndexPath))
4174  indexclauses = ((IndexPath *) bmqual)->indexclauses;
4175  else
4176  return false;
4177  break;
4178  }
4179  default:
4180 
4181  /*
4182  * If it's not a simple indexscan, it probably doesn't run quickly
4183  * for zero rows out, even if it's a parameterized path using all
4184  * the joinquals.
4185  */
4186  return false;
4187  }
4188 
4189  /*
4190  * Examine the inner path's param clauses. Any that are from the outer
4191  * path must be found in the indexclauses list, either exactly or in an
4192  * equivalent form generated by equivclass.c. Also, we must find at least
4193  * one such clause, else it's a clauseless join which isn't fast.
4194  */
4195  found_one = false;
4196  foreach(lc, innerpath->param_info->ppi_clauses)
4197  {
4198  RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4199 
4200  if (join_clause_is_movable_into(rinfo,
4201  innerpath->parent->relids,
4202  joinrelids))
4203  {
4204  if (!(list_member_ptr(indexclauses, rinfo) ||
4205  is_redundant_derived_clause(rinfo, indexclauses)))
4206  return false;
4207  found_one = true;
4208  }
4209  }
4210  return found_one;
4211 }
4212 
4213 
4214 /*
4215  * approx_tuple_count
4216  * Quick-and-dirty estimation of the number of join rows passing
4217  * a set of qual conditions.
4218  *
4219  * The quals can be either an implicitly-ANDed list of boolean expressions,
4220  * or a list of RestrictInfo nodes (typically the latter).
4221  *
4222  * We intentionally compute the selectivity under JOIN_INNER rules, even
4223  * if it's some type of outer join. This is appropriate because we are
4224  * trying to figure out how many tuples pass the initial merge or hash
4225  * join step.
4226  *
4227  * This is quick-and-dirty because we bypass clauselist_selectivity, and
4228  * simply multiply the independent clause selectivities together. Now
4229  * clauselist_selectivity often can't do any better than that anyhow, but
4230  * for some situations (such as range constraints) it is smarter. However,
4231  * we can't effectively cache the results of clauselist_selectivity, whereas
4232  * the individual clause selectivities can be and are cached.
4233  *
4234  * Since we are only using the results to estimate how many potential
4235  * output tuples are generated and passed through qpqual checking, it
4236  * seems OK to live with the approximation.
4237  */
4238 static double
4240 {
4241  double tuples;
4242  double outer_tuples = path->outerjoinpath->rows;
4243  double inner_tuples = path->innerjoinpath->rows;
4244  SpecialJoinInfo sjinfo;
4245  Selectivity selec = 1.0;
4246  ListCell *l;
4247 
4248  /*
4249  * Make up a SpecialJoinInfo for JOIN_INNER semantics.
4250  */
4251  sjinfo.type = T_SpecialJoinInfo;
4252  sjinfo.min_lefthand = path->outerjoinpath->parent->relids;
4253  sjinfo.min_righthand = path->innerjoinpath->parent->relids;
4254  sjinfo.syn_lefthand = path->outerjoinpath->parent->relids;
4255  sjinfo.syn_righthand = path->innerjoinpath->parent->relids;
4256  sjinfo.jointype = JOIN_INNER;
4257  /* we don't bother trying to make the remaining fields valid */
4258  sjinfo.lhs_strict = false;
4259  sjinfo.delay_upper_joins = false;
4260  sjinfo.semi_can_btree = false;
4261  sjinfo.semi_can_hash = false;
4262  sjinfo.semi_operators = NIL;
4263  sjinfo.semi_rhs_exprs = NIL;
4264 
4265  /* Get the approximate selectivity */
4266  foreach(l, quals)
4267  {
4268  Node *qual = (Node *) lfirst(l);
4269 
4270  /* Note that clause_selectivity will be able to cache its result */
4271  selec *= clause_selectivity(root, qual, 0, JOIN_INNER, &sjinfo);
4272  }
4273 
4274  /* Apply it to the input relation sizes */
4275  tuples = selec * outer_tuples * inner_tuples;
4276 
4277  return clamp_row_est(tuples);
4278 }
4279 
4280 
4281 /*
4282  * set_baserel_size_estimates
4283  * Set the size estimates for the given base relation.
4284  *
4285  * The rel's targetlist and restrictinfo list must have been constructed
4286  * already, and rel->tuples must be set.
4287  *
4288  * We set the following fields of the rel node:
4289  * rows: the estimated number of output tuples (after applying
4290  * restriction clauses).
4291  * width: the estimated average output tuple width in bytes.
4292  * baserestrictcost: estimated cost of evaluating baserestrictinfo clauses.
4293  */
4294 void
4296 {
4297  double nrows;
4298 
4299  /* Should only be applied to base relations */
4300  Assert(rel->relid > 0);
4301 
4302  nrows = rel->tuples *
4304  rel->baserestrictinfo,
4305  0,
4306  JOIN_INNER,
4307  NULL);
4308 
4309  rel->rows = clamp_row_est(nrows);
4310 
4312 
4313  set_rel_width(root, rel);
4314 }
4315 
4316 /*
4317  * get_parameterized_baserel_size
4318  * Make a size estimate for a parameterized scan of a base relation.
4319  *
4320  * 'param_clauses' lists the additional join clauses to be used.
4321  *
4322  * set_baserel_size_estimates must have been applied already.
4323  */
4324 double
4326  List *param_clauses)
4327 {
4328  List *allclauses;
4329  double nrows;
4330 
4331  /*
4332  * Estimate the number of rows returned by the parameterized scan, knowing
4333  * that it will apply all the extra join clauses as well as the rel's own
4334  * restriction clauses. Note that we force the clauses to be treated as
4335  * non-join clauses during selectivity estimation.
4336  */
4337  allclauses = list_concat(list_copy(param_clauses),
4338  rel->baserestrictinfo);
4339  nrows = rel->tuples *
4341  allclauses,
4342  rel->relid, /* do not use 0! */
4343  JOIN_INNER,
4344  NULL);
4345  nrows = clamp_row_est(nrows);
4346  /* For safety, make sure result is not more than the base estimate */
4347  if (nrows > rel->rows)
4348  nrows = rel->rows;
4349  return nrows;
4350 }
4351 
4352 /*
4353  * set_joinrel_size_estimates
4354  * Set the size estimates for the given join relation.
4355  *
4356  * The rel's targetlist must have been constructed already, and a
4357  * restriction clause list that matches the given component rels must
4358  * be provided.
4359  *
4360  * Since there is more than one way to make a joinrel for more than two
4361  * base relations, the results we get here could depend on which component
4362  * rel pair is provided. In theory we should get the same answers no matter
4363  * which pair is provided; in practice, since the selectivity estimation
4364  * routines don't handle all cases equally well, we might not. But there's
4365  * not much to be done about it. (Would it make sense to repeat the
4366  * calculations for each pair of input rels that's encountered, and somehow
4367  * average the results? Probably way more trouble than it's worth, and
4368  * anyway we must keep the rowcount estimate the same for all paths for the
4369  * joinrel.)
4370  *
4371  * We set only the rows field here. The reltarget field was already set by
4372  * build_joinrel_tlist, and baserestrictcost is not used for join rels.
4373  */
4374 void
4376  RelOptInfo *outer_rel,
4377  RelOptInfo *inner_rel,
4378  SpecialJoinInfo *sjinfo,
4379  List *restrictlist)
4380 {
4381  rel->rows = calc_joinrel_size_estimate(root,
4382  rel,
4383  outer_rel,
4384  inner_rel,
4385  outer_rel->rows,
4386  inner_rel->rows,
4387  sjinfo,
4388  restrictlist);
4389 }
4390 
4391 /*
4392  * get_parameterized_joinrel_size
4393  * Make a size estimate for a parameterized scan of a join relation.
4394  *
4395  * 'rel' is the joinrel under consideration.
4396  * 'outer_path', 'inner_path' are (probably also parameterized) Paths that
4397  * produce the relations being joined.
4398  * 'sjinfo' is any SpecialJoinInfo relevant to this join.
4399  * 'restrict_clauses' lists the join clauses that need to be applied at the
4400  * join node (including any movable clauses that were moved down to this join,
4401  * and not including any movable clauses that were pushed down into the
4402  * child paths).
4403  *
4404  * set_joinrel_size_estimates must have been applied already.
4405  */
4406 double
4408  Path *outer_path,
4409  Path *inner_path,
4410  SpecialJoinInfo *sjinfo,
4411  List *restrict_clauses)
4412 {
4413  double nrows;
4414 
4415  /*
4416  * Estimate the number of rows returned by the parameterized join as the
4417  * sizes of the input paths times the selectivity of the clauses that have
4418  * ended up at this join node.
4419  *
4420  * As with set_joinrel_size_estimates, the rowcount estimate could depend
4421  * on the pair of input paths provided, though ideally we'd get the same
4422  * estimate for any pair with the same parameterization.
4423  */
4424  nrows = calc_joinrel_size_estimate(root,
4425  rel,
4426  outer_path->parent,
4427  inner_path->parent,
4428  outer_path->rows,
4429  inner_path->rows,
4430  sjinfo,
4431  restrict_clauses);
4432  /* For safety, make sure result is not more than the base estimate */
4433  if (nrows > rel->rows)
4434  nrows = rel->rows;
4435  return nrows;
4436 }
4437 
4438 /*
4439  * calc_joinrel_size_estimate
4440  * Workhorse for set_joinrel_size_estimates and
4441  * get_parameterized_joinrel_size.
4442  *
4443  * outer_rel/inner_rel are the relations being joined, but they should be
4444  * assumed to have sizes outer_rows/inner_rows; those numbers might be less
4445  * than what rel->rows says, when we are considering parameterized paths.
4446  */
4447 static double
4449  RelOptInfo *joinrel,
4450  RelOptInfo *outer_rel,
4451  RelOptInfo *inner_rel,
4452  double outer_rows,
4453  double inner_rows,
4454  SpecialJoinInfo *sjinfo,
4455  List *restrictlist_in)
4456 {
4457  /* This apparently-useless variable dodges a compiler bug in VS2013: */
4458  List *restrictlist = restrictlist_in;
4459  JoinType jointype = sjinfo->jointype;
4460  Selectivity fkselec;
4461  Selectivity jselec;
4462  Selectivity pselec;
4463  double nrows;
4464 
4465  /*
4466  * Compute joinclause selectivity. Note that we are only considering
4467  * clauses that become restriction clauses at this join level; we are not
4468  * double-counting them because they were not considered in estimating the
4469  * sizes of the component rels.
4470  *
4471  * First, see whether any of the joinclauses can be matched to known FK
4472  * constraints. If so, drop those clauses from the restrictlist, and
4473  * instead estimate their selectivity using FK semantics. (We do this
4474  * without regard to whether said clauses are local or "pushed down".
4475  * Probably, an FK-matching clause could never be seen as pushed down at
4476  * an outer join, since it would be strict and hence would be grounds for
4477  * join strength reduction.) fkselec gets the net selectivity for
4478  * FK-matching clauses, or 1.0 if there are none.
4479  */
4480  fkselec = get_foreign_key_join_selectivity(root,
4481  outer_rel->relids,
4482  inner_rel->relids,
4483  sjinfo,
4484  &restrictlist);
4485 
4486  /*
4487  * For an outer join, we have to distinguish the selectivity of the join's
4488  * own clauses (JOIN/ON conditions) from any clauses that were "pushed
4489  * down". For inner joins we just count them all as joinclauses.
4490  */
4491  if (IS_OUTER_JOIN(jointype))
4492  {
4493  List *joinquals = NIL;
4494  List *pushedquals = NIL;
4495  ListCell *l;
4496 
4497  /* Grovel through the clauses to separate into two lists */
4498  foreach(l, restrictlist)
4499  {
4500  RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
4501 
4502  if (RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
4503  pushedquals = lappend(pushedquals, rinfo);
4504  else
4505  joinquals = lappend(joinquals, rinfo);
4506  }
4507 
4508  /* Get the separate selectivities */
4509  jselec = clauselist_selectivity(root,
4510  joinquals,
4511  0,
4512  jointype,
4513  sjinfo);
4514  pselec = clauselist_selectivity(root,
4515  pushedquals,
4516  0,
4517  jointype,
4518  sjinfo);
4519 
4520  /* Avoid leaking a lot of ListCells */
4521  list_free(joinquals);
4522  list_free(pushedquals);
4523  }
4524  else
4525  {
4526  jselec = clauselist_selectivity(root,
4527  restrictlist,
4528  0,
4529  jointype,
4530  sjinfo);
4531  pselec = 0.0; /* not used, keep compiler quiet */
4532  }
4533 
4534  /*
4535  * Basically, we multiply size of Cartesian product by selectivity.
4536  *
4537  * If we are doing an outer join, take that into account: the joinqual
4538  * selectivity has to be clamped using the knowledge that the output must
4539  * be at least as large as the non-nullable input. However, any
4540  * pushed-down quals are applied after the outer join, so their
4541  * selectivity applies fully.
4542  *
4543  * For JOIN_SEMI and JOIN_ANTI, the selectivity is defined as the fraction
4544  * of LHS rows that have matches, and we apply that straightforwardly.
4545  */
4546  switch (jointype)
4547  {
4548  case JOIN_INNER:
4549  nrows = outer_rows * inner_rows * fkselec * jselec;
4550  /* pselec not used */
4551  break;
4552  case JOIN_LEFT:
4553  nrows = outer_rows * inner_rows * fkselec * jselec;
4554  if (nrows < outer_rows)
4555  nrows = outer_rows;
4556  nrows *= pselec;
4557  break;
4558  case JOIN_FULL:
4559  nrows = outer_rows * inner_rows * fkselec * jselec;
4560  if (nrows < outer_rows)
4561  nrows = outer_rows;
4562  if (nrows < inner_rows)
4563  nrows = inner_rows;
4564  nrows *= pselec;
4565  break;
4566  case JOIN_SEMI:
4567  nrows = outer_rows * fkselec * jselec;
4568  /* pselec not used */
4569  break;
4570  case JOIN_ANTI:
4571  nrows = outer_rows * (1.0 - fkselec * jselec);
4572  nrows *= pselec;
4573  break;
4574  default:
4575  /* other values not expected here */
4576  elog(ERROR, "unrecognized join type: %d", (int) jointype);
4577  nrows = 0; /* keep compiler quiet */
4578  break;
4579  }
4580 
4581  return clamp_row_est(nrows);
4582 }
4583 
4584 /*
4585  * get_foreign_key_join_selectivity
4586  * Estimate join selectivity for foreign-key-related clauses.
4587  *
4588  * Remove any clauses that can be matched to FK constraints from *restrictlist,
4589  * and return a substitute estimate of their selectivity. 1.0 is returned
4590  * when there are no such clauses.
4591  *
4592  * The reason for treating such clauses specially is that we can get better
4593  * estimates this way than by relying on clauselist_selectivity(), especially
4594  * for multi-column FKs where that function's assumption that the clauses are
4595  * independent falls down badly. But even with single-column FKs, we may be
4596  * able to get a better answer when the pg_statistic stats are missing or out
4597  * of date.
4598  */
4599 static Selectivity
4601  Relids outer_relids,
4602  Relids inner_relids,
4603  SpecialJoinInfo *sjinfo,
4604  List **restrictlist)
4605 {
4606  Selectivity fkselec = 1.0;
4607  JoinType jointype = sjinfo->jointype;
4608  List *worklist = *restrictlist;
4609  ListCell *lc;
4610 
4611  /* Consider each FK constraint that is known to match the query */
4612  foreach(lc, root->fkey_list)
4613  {
4614  ForeignKeyOptInfo *fkinfo = (ForeignKeyOptInfo *) lfirst(lc);
4615  bool ref_is_outer;
4616  List *removedlist;
4617  ListCell *cell;
4618  ListCell *prev;
4619  ListCell *next;
4620 
4621  /*
4622  * This FK is not relevant unless it connects a baserel on one side of
4623  * this join to a baserel on the other side.
4624  */
4625  if (bms_is_member(fkinfo->con_relid, outer_relids) &&
4626  bms_is_member(fkinfo->ref_relid, inner_relids))
4627  ref_is_outer = false;
4628  else if (bms_is_member(fkinfo->ref_relid, outer_relids) &&
4629  bms_is_member(fkinfo->con_relid, inner_relids))
4630  ref_is_outer = true;
4631  else
4632  continue;
4633 
4634  /*
4635  * If we're dealing with a semi/anti join, and the FK's referenced
4636  * relation is on the outside, then knowledge of the FK doesn't help
4637  * us figure out what we need to know (which is the fraction of outer
4638  * rows that have matches). On the other hand, if the referenced rel
4639  * is on the inside, then all outer rows must have matches in the
4640  * referenced table (ignoring nulls). But any restriction or join
4641  * clauses that filter that table will reduce the fraction of matches.
4642  * We can account for restriction clauses, but it's too hard to guess
4643  * how many table rows would get through a join that's inside the RHS.
4644  * Hence, if either case applies, punt and ignore the FK.
4645  */
4646  if ((jointype == JOIN_SEMI || jointype == JOIN_ANTI) &&
4647  (ref_is_outer || bms_membership(inner_relids) != BMS_SINGLETON))
4648  continue;
4649 
4650  /*
4651  * Modify the restrictlist by removing clauses that match the FK (and
4652  * putting them into removedlist instead). It seems unsafe to modify
4653  * the originally-passed List structure, so we make a shallow copy the
4654  * first time through.
4655  */
4656  if (worklist == *restrictlist)
4657  worklist = list_copy(worklist);
4658 
4659  removedlist = NIL;
4660  prev = NULL;
4661  for (cell = list_head(worklist); cell; cell = next)
4662  {
4663  RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
4664  bool remove_it = false;
4665  int i;
4666 
4667  next = lnext(cell);
4668  /* Drop this clause if it matches any column of the FK */
4669  for (i = 0; i < fkinfo->nkeys; i++)
4670  {
4671  if (rinfo->parent_ec)
4672  {
4673  /*
4674  * EC-derived clauses can only match by EC. It is okay to
4675  * consider any clause derived from the same EC as
4676  * matching the FK: even if equivclass.c chose to generate
4677  * a clause equating some other pair of Vars, it could
4678  * have generated one equating the FK's Vars. So for
4679  * purposes of estimation, we can act as though it did so.
4680  *
4681  * Note: checking parent_ec is a bit of a cheat because
4682  * there are EC-derived clauses that don't have parent_ec
4683  * set; but such clauses must compare expressions that
4684  * aren't just Vars, so they cannot match the FK anyway.
4685  */
4686  if (fkinfo->eclass[i] == rinfo->parent_ec)
4687  {
4688  remove_it = true;
4689  break;
4690  }
4691  }
4692  else
4693  {
4694  /*
4695  * Otherwise, see if rinfo was previously matched to FK as
4696  * a "loose" clause.
4697  */
4698  if (list_member_ptr(fkinfo->rinfos[i], rinfo))
4699  {
4700  remove_it = true;
4701  break;
4702  }
4703  }
4704  }
4705  if (remove_it)
4706  {
4707  worklist = list_delete_cell(worklist, cell, prev);
4708  removedlist = lappend(removedlist, rinfo);
4709  }
4710  else
4711  prev = cell;
4712  }
4713 
4714  /*
4715  * If we failed to remove all the matching clauses we expected to
4716  * find, chicken out and ignore this FK; applying its selectivity
4717  * might result in double-counting. Put any clauses we did manage to
4718  * remove back into the worklist.
4719  *
4720  * Since the matching clauses are known not outerjoin-delayed, they
4721  * should certainly have appeared in the initial joinclause list. If
4722  * we didn't find them, they must have been matched to, and removed
4723  * by, some other FK in a previous iteration of this loop. (A likely
4724  * case is that two FKs are matched to the same EC; there will be only
4725  * one EC-derived clause in the initial list, so the first FK will
4726  * consume it.) Applying both FKs' selectivity independently risks
4727  * underestimating the join size; in particular, this would undo one
4728  * of the main things that ECs were invented for, namely to avoid
4729  * double-counting the selectivity of redundant equality conditions.
4730  * Later we might think of a reasonable way to combine the estimates,
4731  * but for now, just punt, since this is a fairly uncommon situation.
4732  */
4733  if (list_length(removedlist) !=
4734  (fkinfo->nmatched_ec + fkinfo->nmatched_ri))
4735  {
4736  worklist = list_concat(worklist, removedlist);
4737  continue;
4738  }
4739 
4740  /*
4741  * Finally we get to the payoff: estimate selectivity using the
4742  * knowledge that each referencing row will match exactly one row in
4743  * the referenced table.
4744  *
4745  * XXX that's not true in the presence of nulls in the referencing
4746  * column(s), so in principle we should derate the estimate for those.
4747  * However (1) if there are any strict restriction clauses for the
4748  * referencing column(s) elsewhere in the query, derating here would
4749  * be double-counting the null fraction, and (2) it's not very clear
4750  * how to combine null fractions for multiple referencing columns. So
4751  * we do nothing for now about correcting for nulls.
4752  *
4753  * XXX another point here is that if either side of an FK constraint
4754  * is an inheritance parent, we estimate as though the constraint
4755  * covers all its children as well. This is not an unreasonable
4756  * assumption for a referencing table, ie the user probably applied
4757  * identical constraints to all child tables (though perhaps we ought
4758  * to check that). But it's not possible to have done that for a
4759  * referenced table. Fortunately, precisely because that doesn't
4760  * work, it is uncommon in practice to have an FK referencing a parent
4761  * table. So, at least for now, disregard inheritance here.
4762  */
4763  if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
4764  {
4765  /*
4766  * For JOIN_SEMI and JOIN_ANTI, we only get here when the FK's
4767  * referenced table is exactly the inside of the join. The join
4768  * selectivity is defined as the fraction of LHS rows that have
4769  * matches. The FK implies that every LHS row has a match *in the
4770  * referenced table*; but any restriction clauses on it will
4771  * reduce the number of matches. Hence we take the join
4772  * selectivity as equal to the selectivity of the table's
4773  * restriction clauses, which is rows / tuples; but we must guard
4774  * against tuples == 0.
4775  */
4776  RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
4777  double ref_tuples = Max(ref_rel->tuples, 1.0);
4778 
4779  fkselec *= ref_rel->rows / ref_tuples;
4780  }
4781  else
4782  {
4783  /*
4784  * Otherwise, selectivity is exactly 1/referenced-table-size; but
4785  * guard against tuples == 0. Note we should use the raw table
4786  * tuple count, not any estimate of its filtered or joined size.
4787  */
4788  RelOptInfo *ref_rel = find_base_rel(root, fkinfo->ref_relid);
4789  double ref_tuples = Max(ref_rel->tuples, 1.0);
4790 
4791  fkselec *= 1.0 / ref_tuples;
4792  }
4793  }
4794 
4795  *restrictlist = worklist;
4796  return fkselec;
4797 }
4798 
4799 /*
4800  * set_subquery_size_estimates
4801  * Set the size estimates for a base relation that is a subquery.
4802  *
4803  * The rel's targetlist and restrictinfo list must have been constructed
4804  * already, and the Paths for the subquery must have been completed.
4805  * We look at the subquery's PlannerInfo to extract data.
4806  *
4807  * We set the same fields as set_baserel_size_estimates.
4808  */
4809 void
4811 {
4812  PlannerInfo *subroot = rel->subroot;
4813  RelOptInfo *sub_final_rel;
4814  ListCell *lc;
4815 
4816  /* Should only be applied to base relations that are subqueries */
4817  Assert(rel->relid > 0);
4818  Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_SUBQUERY);
4819 
4820  /*
4821  * Copy raw number of output rows from subquery. All of its paths should
4822  * have the same output rowcount, so just look at cheapest-total.
4823  */
4824  sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
4825  rel->tuples = sub_final_rel->cheapest_total_path->rows;
4826 
4827  /*
4828  * Compute per-output-column width estimates by examining the subquery's
4829  * targetlist. For any output that is a plain Var, get the width estimate
4830  * that was made while planning the subquery. Otherwise, we leave it to
4831  * set_rel_width to fill in a datatype-based default estimate.
4832  */
4833  foreach(lc, subroot->parse->targetList)
4834  {
4835  TargetEntry *te = lfirst_node(TargetEntry, lc);
4836  Node *texpr = (Node *) te->expr;
4837  int32 item_width = 0;
4838 
4839  /* junk columns aren't visible to upper query */
4840  if (te->resjunk)
4841  continue;
4842 
4843  /*
4844  * The subquery could be an expansion of a view that's had columns
4845  * added to it since the current query was parsed, so that there are
4846  * non-junk tlist columns in it that don't correspond to any column
4847  * visible at our query level. Ignore such columns.
4848  */
4849  if (te->resno < rel->min_attr || te->resno > rel->max_attr)
4850  continue;
4851 
4852  /*
4853  * XXX This currently doesn't work for subqueries containing set
4854  * operations, because the Vars in their tlists are bogus references
4855  * to the first leaf subquery, which wouldn't give the right answer
4856  * even if we could still get to its PlannerInfo.
4857  *
4858  * Also, the subquery could be an appendrel for which all branches are
4859  * known empty due to constraint exclusion, in which case
4860  * set_append_rel_pathlist will have left the attr_widths set to zero.
4861  *
4862  * In either case, we just leave the width estimate zero until
4863  * set_rel_width fixes it.
4864  */
4865  if (IsA(texpr, Var) &&
4866  subroot->parse->setOperations == NULL)
4867  {
4868  Var *var = (Var *) texpr;
4869  RelOptInfo *subrel = find_base_rel(subroot, var->varno);
4870 
4871  item_width = subrel->attr_widths[var->varattno - subrel->min_attr];
4872  }
4873  rel->attr_widths[te->resno - rel->min_attr] = item_width;
4874  }
4875 
4876  /* Now estimate number of output rows, etc */
4877  set_baserel_size_estimates(root, rel);
4878 }
4879 
4880 /*
4881  * set_function_size_estimates
4882  * Set the size estimates for a base relation that is a function call.
4883  *
4884  * The rel's targetlist and restrictinfo list must have been constructed
4885  * already.
4886  *
4887  * We set the same fields as set_baserel_size_estimates.
4888  */
4889 void
4891 {
4892  RangeTblEntry *rte;
4893  ListCell *lc;
4894 
4895  /* Should only be applied to base relations that are functions */
4896  Assert(rel->relid > 0);
4897  rte = planner_rt_fetch(rel->relid, root);
4898  Assert(rte->rtekind == RTE_FUNCTION);
4899 
4900  /*
4901  * Estimate number of rows the functions will return. The rowcount of the
4902  * node is that of the largest function result.
4903  */
4904  rel->tuples = 0;
4905  foreach(lc, rte->functions)
4906  {
4907  RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
4908  double ntup = expression_returns_set_rows(rtfunc->funcexpr);
4909 
4910  if (ntup > rel->tuples)
4911  rel->tuples = ntup;
4912  }
4913 
4914  /* Now estimate number of output rows, etc */
4915  set_baserel_size_estimates(root, rel);
4916 }
4917 
4918 /*
4919  * set_function_size_estimates
4920  * Set the size estimates for a base relation that is a function call.
4921  *
4922  * The rel's targetlist and restrictinfo list must have been constructed
4923  * already.
4924  *
4925  * We set the same fields as set_tablefunc_size_estimates.
4926  */
4927 void
4929 {
4930  /* Should only be applied to base relations that are functions */
4931  Assert(rel->relid > 0);
4932  Assert(planner_rt_fetch(rel->relid, root)->rtekind == RTE_TABLEFUNC);
4933 
4934  rel->tuples = 100;
4935 
4936  /* Now estimate number of output rows, etc */
4937  set_baserel_size_estimates(root, rel);
4938 }
4939 
4940 /*
4941  * set_values_size_estimates
4942  * Set the size estimates for a base relation that is a values list.
4943  *
4944  * The rel's targetlist and restrictinfo list must have been constructed
4945  * already.
4946  *
4947  * We set the same fields as set_baserel_size_estimates.
4948  */
4949 void
4951 {
4952  RangeTblEntry *rte;
4953 
4954  /* Should only be applied to base relations that are values lists */
4955  Assert(rel->relid > 0);
4956  rte = planner_rt_fetch(rel->relid, root);
4957  Assert(rte->rtekind == RTE_VALUES);
4958 
4959  /*
4960  * Estimate number of rows the values list will return. We know this
4961  * precisely based on the list length (well, barring set-returning
4962  * functions in list items, but that's a refinement not catered for
4963  * anywhere else either).
4964  */
4965  rel->tuples = list_length(rte->values_lists);
4966 
4967  /* Now estimate number of output rows, etc */
4968  set_baserel_size_estimates(root, rel);
4969 }
4970 
4971 /*
4972  * set_cte_size_estimates
4973  * Set the size estimates for a base relation that is a CTE reference.
4974  *
4975  * The rel's targetlist and restrictinfo list must have been constructed
4976  * already, and we need an estimate of the number of rows returned by the CTE
4977  * (if a regular CTE) or the non-recursive term (if a self-reference).
4978  *
4979  * We set the same fields as set_baserel_size_estimates.
4980  */
4981 void
4982 set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
4983 {
4984  RangeTblEntry *rte;
4985 
4986  /* Should only be applied to base relations that are CTE references */
4987  Assert(rel->relid > 0);
4988  rte = planner_rt_fetch(rel->relid, root);
4989  Assert(rte->rtekind == RTE_CTE);
4990 
4991  if (rte->self_reference)
4992  {
4993  /*
4994  * In a self-reference, arbitrarily assume the average worktable size
4995  * is about 10 times the nonrecursive term's size.
4996  */
4997  rel->tuples = 10 * cte_rows;
4998  }
4999  else
5000  {
5001  /* Otherwise just believe the CTE's rowcount estimate */
5002  rel->tuples = cte_rows;
5003  }
5004 
5005  /* Now estimate number of output rows, etc */
5006  set_baserel_size_estimates(root, rel);
5007 }
5008 
5009 /*
5010  * set_namedtuplestore_size_estimates
5011  * Set the size estimates for a base relation that is a tuplestore reference.
5012  *
5013  * The rel's targetlist and restrictinfo list must have been constructed
5014  * already.
5015  *
5016  * We set the same fields as set_baserel_size_estimates.
5017  */
5018 void
5020 {
5021  RangeTblEntry *rte;
5022 
5023  /* Should only be applied to base relations that are tuplestore references */
5024  Assert(rel->relid > 0);
5025  rte = planner_rt_fetch(rel->relid, root);
5027 
5028  /*
5029  * Use the estimate provided by the code which is generating the named
5030  * tuplestore. In some cases, the actual number might be available; in
5031  * others the same plan will be re-used, so a "typical" value might be
5032  * estimated and used.
5033  */
5034  rel->tuples = rte->enrtuples;
5035  if (rel->tuples < 0)
5036  rel->tuples = 1000;
5037 
5038  /* Now estimate number of output rows, etc */
5039  set_baserel_size_estimates(root, rel);
5040 }
5041 
5042 /*
5043  * set_foreign_size_estimates
5044  * Set the size estimates for a base relation that is a foreign table.
5045  *
5046  * There is not a whole lot that we can do here; the foreign-data wrapper
5047  * is responsible for producing useful estimates. We can do a decent job
5048  * of estimating baserestrictcost, so we set that, and we also set up width
5049  * using what will be purely datatype-driven estimates from the targetlist.
5050  * There is no way to do anything sane with the rows value, so we just put
5051  * a default estimate and hope that the wrapper can improve on it. The
5052  * wrapper's GetForeignRelSize function will be called momentarily.
5053  *
5054  * The rel's targetlist and restrictinfo list must have been constructed
5055  * already.
5056  */
5057 void
5059 {
5060  /* Should only be applied to base relations */
5061  Assert(rel->relid > 0);
5062 
5063  rel->rows = 1000; /* entirely bogus default estimate */
5064 
5066 
5067  set_rel_width(root, rel);
5068 }
5069 
5070 
5071 /*
5072  * set_rel_width
5073  * Set the estimated output width of a base relation.
5074  *
5075  * The estimated output width is the sum of the per-attribute width estimates
5076  * for the actually-referenced columns, plus any PHVs or other expressions
5077  * that have to be calculated at this relation. This is the amount of data
5078  * we'd need to pass upwards in case of a sort, hash, etc.
5079  *
5080  * This function also sets reltarget->cost, so it's a bit misnamed now.
5081  *
5082  * NB: this works best on plain relations because it prefers to look at
5083  * real Vars. For subqueries, set_subquery_size_estimates will already have
5084  * copied up whatever per-column estimates were made within the subquery,
5085  * and for other types of rels there isn't much we can do anyway. We fall
5086  * back on (fairly stupid) datatype-based width estimates if we can't get
5087  * any better number.
5088  *
5089  * The per-attribute width estimates are cached for possible re-use while
5090  * building join relations or post-scan/join pathtargets.
5091  */
5092 static void
5094 {
5095  Oid reloid = planner_rt_fetch(rel->relid, root)->relid;
5096  int32 tuple_width = 0;
5097  bool have_wholerow_var = false;
5098  ListCell *lc;
5099 
5100  /* Vars are assumed to have cost zero, but other exprs do not */
5101  rel->reltarget->cost.startup = 0;
5102  rel->reltarget->cost.per_tuple = 0;
5103 
5104  foreach(lc, rel->reltarget->exprs)
5105  {
5106  Node *node = (Node *) lfirst(lc);
5107 
5108  /*
5109  * Ordinarily, a Var in a rel's targetlist must belong to that rel;
5110  * but there are corner cases involving LATERAL references where that
5111  * isn't so. If the Var has the wrong varno, fall through to the
5112  * generic case (it doesn't seem worth the trouble to be any smarter).
5113  */
5114  if (IsA(node, Var) &&
5115  ((Var *) node)->varno == rel->relid)
5116  {
5117  Var *var = (Var *) node;
5118  int ndx;
5119  int32 item_width;
5120 
5121  Assert(var->varattno >= rel->min_attr);
5122  Assert(var->varattno <= rel->max_attr);
5123 
5124  ndx = var->varattno - rel->min_attr;
5125 
5126  /*
5127  * If it's a whole-row Var, we'll deal with it below after we have
5128  * already cached as many attr widths as possible.
5129  */
5130  if (var->varattno == 0)
5131  {
5132  have_wholerow_var = true;
5133  continue;
5134  }
5135 
5136  /*
5137  * The width may have been cached already (especially if it's a
5138  * subquery), so don't duplicate effort.
5139  */
5140  if (rel->attr_widths[ndx] > 0)
5141  {
5142  tuple_width += rel->attr_widths[ndx];
5143  continue;
5144  }
5145 
5146  /* Try to get column width from statistics */
5147  if (reloid != InvalidOid && var->varattno > 0)
5148  {
5149  item_width = get_attavgwidth(reloid, var->varattno);
5150  if (item_width > 0)
5151  {
5152  rel->attr_widths[ndx] = item_width;
5153  tuple_width += item_width;
5154  continue;
5155  }
5156  }
5157 
5158  /*
5159  * Not a plain relation, or can't find statistics for it. Estimate
5160  * using just the type info.
5161  */
5162  item_width = get_typavgwidth(var->vartype, var->vartypmod);
5163  Assert(item_width > 0);
5164  rel->attr_widths[ndx] = item_width;
5165  tuple_width += item_width;
5166  }
5167  else if (IsA(node, PlaceHolderVar))
5168  {
5169  /*
5170  * We will need to evaluate the PHV's contained expression while
5171  * scanning this rel, so be sure to include it in reltarget->cost.
5172  */
5173  PlaceHolderVar *phv = (PlaceHolderVar *) node;
5174  PlaceHolderInfo *phinfo = find_placeholder_info(root, phv, false);
5175  QualCost cost;
5176 
5177  tuple_width += phinfo->ph_width;
5178  cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
5179  rel->reltarget->cost.startup += cost.startup;
5180  rel->reltarget->cost.per_tuple += cost.per_tuple;
5181  }
5182  else
5183  {
5184  /*
5185  * We could be looking at an expression pulled up from a subquery,
5186  * or a ROW() representing a whole-row child Var, etc. Do what we
5187  * can using the expression type information.
5188  */
5189  int32 item_width;
5190  QualCost cost;
5191 
5192  item_width = get_typavgwidth(exprType(node), exprTypmod(node));
5193  Assert(item_width > 0);
5194  tuple_width += item_width;
5195  /* Not entirely clear if we need to account for cost, but do so */
5196  cost_qual_eval_node(&cost, node, root);
5197  rel->reltarget->cost.startup += cost.startup;
5198  rel->reltarget->cost.per_tuple += cost.per_tuple;
5199  }
5200  }
5201 
5202  /*
5203  * If we have a whole-row reference, estimate its width as the sum of
5204  * per-column widths plus heap tuple header overhead.
5205  */
5206  if (have_wholerow_var)
5207  {
5208  int32 wholerow_width = MAXALIGN(SizeofHeapTupleHeader);
5209 
5210  if (reloid != InvalidOid)
5211  {
5212  /* Real relation, so estimate true tuple width */
5213  wholerow_width += get_relation_data_width(reloid,
5214  rel->attr_widths - rel->min_attr);
5215  }
5216  else
5217  {
5218  /* Do what we can with info for a phony rel */
5219  AttrNumber i;
5220 
5221  for (i = 1; i <= rel->max_attr; i++)
5222  wholerow_width += rel->attr_widths[i - rel->min_attr];
5223  }
5224 
5225  rel->attr_widths[0 - rel->min_attr] = wholerow_width;
5226 
5227  /*
5228  * Include the whole-row Var as part of the output tuple. Yes, that
5229  * really is what happens at runtime.
5230  */
5231  tuple_width += wholerow_width;
5232  }
5233 
5234  Assert(tuple_width >= 0);
5235  rel->reltarget->width = tuple_width;
5236 }
5237 
5238 /*
5239  * set_pathtarget_cost_width
5240  * Set the estimated eval cost and output width of a PathTarget tlist.
5241  *
5242  * As a notational convenience, returns the same PathTarget pointer passed in.
5243  *
5244  * Most, though not quite all, uses of this function occur after we've run
5245  * set_rel_width() for base relations; so we can usually obtain cached width
5246  * estimates for Vars. If we can't, fall back on datatype-based width
5247  * estimates. Present early-planning uses of PathTargets don't need accurate
5248  * widths badly enough to justify going to the catalogs for better data.
5249  */
5250 PathTarget *
5252 {
5253  int32 tuple_width = 0;
5254  ListCell *lc;
5255 
5256  /* Vars are assumed to have cost zero, but other exprs do not */
5257  target->cost.startup = 0;
5258  target->cost.per_tuple = 0;
5259 
5260  foreach(lc, target->exprs)
5261  {
5262  Node *node = (Node *) lfirst(lc);
5263 
5264  if (IsA(node, Var))
5265  {
5266  Var *var = (Var *) node;
5267  int32 item_width;
5268 
5269  /* We should not see any upper-level Vars here */
5270  Assert(var->varlevelsup == 0);
5271 
5272  /* Try to get data from RelOptInfo cache */
5273  if (var->varno < root->simple_rel_array_size)
5274  {
5275  RelOptInfo *rel = root->simple_rel_array[var->varno];
5276 
5277  if (rel != NULL &&
5278  var->varattno >= rel->min_attr &&
5279  var->varattno <= rel->max_attr)
5280  {
5281  int ndx = var->varattno - rel->min_attr;
5282 
5283  if (rel->attr_widths[ndx] > 0)
5284  {
5285  tuple_width += rel->attr_widths[ndx];
5286  continue;
5287  }
5288  }
5289  }
5290 
5291  /*
5292  * No cached data available, so estimate using just the type info.
5293  */
5294  item_width = get_typavgwidth(var->vartype, var->vartypmod);
5295  Assert(item_width > 0);
5296  tuple_width += item_width;
5297  }
5298  else
5299  {
5300  /*
5301  * Handle general expressions using type info.
5302  */
5303  int32 item_width;
5304  QualCost cost;
5305 
5306  item_width = get_typavgwidth(exprType(node), exprTypmod(node));
5307  Assert(item_width > 0);
5308  tuple_width += item_width;
5309 
5310  /* Account for cost, too */
5311  cost_qual_eval_node(&cost, node, root);
5312  target->cost.startup += cost.startup;
5313  target->cost.per_tuple += cost.per_tuple;
5314  }
5315  }
5316 
5317  Assert(tuple_width >= 0);
5318  target->width = tuple_width;
5319 
5320  return target;
5321 }
5322 
5323 /*
5324  * relation_byte_size
5325  * Estimate the storage space in bytes for a given number of tuples
5326  * of a given width (size in bytes).
5327  */
5328 static double
5329 relation_byte_size(double tuples, int width)
5330 {
5331  return tuples * (MAXALIGN(width) + MAXALIGN(SizeofHeapTupleHeader));
5332 }
5333 
5334 /*
5335  * page_size
5336  * Returns an estimate of the number of pages covered by a given
5337  * number of tuples of a given width (size in bytes).
5338  */
5339 static double
5340 page_size(double tuples, int width)
5341 {
5342  return ceil(relation_byte_size(tuples, width) / BLCKSZ);
5343 }
5344 
5345 /*
5346  * Estimate the fraction of the work that each worker will do given the
5347  * number of workers budgeted for the path.
5348  */
5349 static double
5351 {
5352  double parallel_divisor = path->parallel_workers;
5353 
5354  /*
5355  * Early experience with parallel query suggests that when there is only
5356  * one worker, the leader often makes a very substantial contribution to
5357  * executing the parallel portion of the plan, but as more workers are
5358  * added, it does less and less, because it's busy reading tuples from the
5359  * workers and doing whatever non-parallel post-processing is needed. By
5360  * the time we reach 4 workers, the leader no longer makes a meaningful
5361  * contribution. Thus, for now, estimate that the leader spends 30% of
5362  * its time servicing each worker, and the remainder executing the
5363  * parallel plan.
5364  */
5366  {
5367  double leader_contribution;
5368 
5369  leader_contribution = 1.0 - (0.3 * path->parallel_workers);
5370  if (leader_contribution > 0)
5371  parallel_divisor += leader_contribution;
5372  }
5373 
5374  return parallel_divisor;
5375 }
5376 
5377 /*
5378  * compute_bitmap_pages
5379  *
5380  * compute number of pages fetched from heap in bitmap heap scan.
5381  */
5382 double
5383 compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel, Path *bitmapqual,
5384  int loop_count, Cost *cost, double *tuple)
5385 {
5386  Cost indexTotalCost;
5387  Selectivity indexSelectivity;
5388  double T;
5389  double pages_fetched;
5390  double tuples_fetched;
5391  double heap_pages;
5392  long maxentries;
5393 
5394  /*
5395  * Fetch total cost of obtaining the bitmap, as well as its total
5396  * selectivity.
5397  */
5398  cost_bitmap_tree_node(bitmapqual, &indexTotalCost, &indexSelectivity);
5399 
5400  /*
5401  * Estimate number of main-table pages fetched.
5402  */
5403  tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
5404 
5405  T = (baserel->pages > 1) ? (double) baserel->pages : 1.0;
5406 
5407  /*
5408  * For a single scan, the number of heap pages that need to be fetched is
5409  * the same as the Mackert and Lohman formula for the case T <= b (ie, no
5410  * re-reads needed).
5411  */
5412  pages_fetched = (2.0 * T * tuples_fetched) / (2.0 * T + tuples_fetched);
5413 
5414  /*
5415  * Calculate the number of pages fetched from the heap. Then based on
5416  * current work_mem estimate get the estimated maxentries in the bitmap.
5417  * (Note that we always do this calculation based on the number of pages
5418  * that would be fetched in a single iteration, even if loop_count > 1.
5419  * That's correct, because only that number of entries will be stored in
5420  * the bitmap at one time.)
5421  */
5422  heap_pages = Min(pages_fetched, baserel->pages);
5423  maxentries = tbm_calculate_entries(work_mem * 1024L);
5424 
5425  if (loop_count > 1)
5426  {
5427  /*
5428  * For repeated bitmap scans, scale up the number of tuples fetched in
5429  * the Mackert and Lohman formula by the number of scans, so that we
5430  * estimate the number of pages fetched by all the scans. Then
5431  * pro-rate for one scan.
5432  */
5433  pages_fetched = index_pages_fetched(tuples_fetched * loop_count,
5434  baserel->pages,
5435  get_indexpath_pages(bitmapqual),
5436  root);
5437  pages_fetched /= loop_count;
5438  }
5439 
5440  if (pages_fetched >= T)
5441  pages_fetched = T;
5442  else
5443  pages_fetched = ceil(pages_fetched);
5444 
5445  if (maxentries < heap_pages)
5446  {
5447  double exact_pages;
5448  double lossy_pages;
5449 
5450  /*
5451  * Crude approximation of the number of lossy pages. Because of the
5452  * way tbm_lossify() is coded, the number of lossy pages increases
5453  * very sharply as soon as we run short of memory; this formula has
5454  * that property and seems to perform adequately in testing, but it's
5455  * possible we could do better somehow.
5456  */
5457  lossy_pages = Max(0, heap_pages - maxentries / 2);
5458  exact_pages = heap_pages - lossy_pages;
5459 
5460  /*
5461  * If there are lossy pages then recompute the number of tuples
5462  * processed by the bitmap heap node. We assume here that the chance
5463  * of a given tuple coming from an exact page is the same as the
5464  * chance that a given page is exact. This might not be true, but
5465  * it's not clear how we can do any better.
5466  */
5467  if (lossy_pages > 0)
5468  tuples_fetched =
5469  clamp_row_est(indexSelectivity *
5470  (exact_pages / heap_pages) * baserel->tuples +
5471  (lossy_pages / heap_pages) * baserel->tuples);
5472  }
5473 
5474  if (cost)
5475  *cost = indexTotalCost;
5476  if (tuple)
5477  *tuple = tuples_fetched;
5478 
5479  return pages_fetched;
5480 }
QualCost eval_cost
Definition: relation.h:1917
void set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:4810
static void cost_rescan(PlannerInfo *root, Path *path, Cost *rescan_startup_cost, Cost *rescan_total_cost)
Definition: costsize.c:3609
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:3742
#define NIL
Definition: pg_list.h:69
void final_cost_hashjoin(PlannerInfo *root, HashPath *path, JoinCostWorkspace *workspace, JoinPathExtraData *extra)
Definition: costsize.c:3261
bool semi_can_btree
Definition: relation.h:2074
Selectivity leftendsel
Definition: relation.h:1977
#define SizeofHeapTupleHeader
Definition: htup_details.h:183
List * path_mergeclauses
Definition: relation.h:1482
double plan_rows
Definition: plannodes.h:134
#define IsA(nodeptr, _type_)
Definition: nodes.h:567
JoinPath jpath
Definition: relation.h:1500
PlannerInfo * root
Definition: costsize.c:145
PathTarget * pathtarget
Definition: relation.h:1079
Query * parse
Definition: relation.h:169
bool enable_tidscan
Definition: costsize.c:129
bool ExecSupportsMarkRestore(Path *pathnode)
Definition: execAmi.c:405
void cost_bitmap_heap_scan(Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info, Path *bitmapqual, double loop_count)
Definition: costsize.c:945
Index varlevelsup
Definition: primnodes.h:173
void getTypeOutputInfo(Oid type, Oid *typOutput, bool *typIsVarlena)
Definition: lsyscache.c:2650
Path path
Definition: relation.h:1154
bool is_redundant_derived_clause(RestrictInfo *rinfo, List *clauselist)
Definition: equivclass.c:2495
IndexOptInfo * indexinfo
Definition: relation.h:1155
SemiAntiJoinFactors semifactors
Definition: relation.h:2314
void cost_tidscan(Path *path, PlannerInfo *root, RelOptInfo *baserel, List *tidquals, ParamPathInfo *param_info)
Definition: costsize.c:1181
static int32 next
Definition: blutils.c:211
void cost_windowagg(Path *path, PlannerInfo *root, List *windowFuncs, int numPartCols, int numOrderCols, Cost input_startup_cost, Cost input_total_cost, double input_tuples)
Definition: costsize.c:2178
bool enable_nestloop
Definition: costsize.c:132
int num_batches
Definition: relation.h:1502
PathTarget * set_pathtarget_cost_width(PlannerInfo *root, PathTarget *target)
Definition: costsize.c:5251
List * args
Definition: primnodes.h:359
Relids min_righthand
Definition: relation.h:2067
bool materialize_inner
Definition: relation.h:1486
#define DEFAULT_CPU_TUPLE_COST
Definition: cost.h:26
void cost_gather_merge(GatherMergePath *path, PlannerInfo *root, RelOptInfo *rel, ParamPathInfo *param_info, Cost input_startup_cost, Cost input_total_cost, double *rows)
Definition: costsize.c:402
void mergejoinscansel(PlannerInfo *root, Node *clause, Oid opfamily, int strategy, bool nulls_first, Selectivity *leftstart, Selectivity *leftend, Selectivity *rightstart, Selectivity *rightend)
Definition: selfuncs.c:2995
Selectivity right_mcvfreq
Definition: relation.h:1944
double expression_returns_set_rows(Node *clause)
Definition: clauses.c:803
Expr * orclause
Definition: relation.h:1911
void cost_agg(Path *path, PlannerInfo *root, AggStrategy aggstrategy, const AggClauseCosts *aggcosts, int numGroupCols, double numGroups, List *quals, Cost input_startup_cost, Cost input_total_cost, double input_tuples)
Definition: costsize.c:2060
int compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages, int max_workers)
Definition: allpaths.c:3443
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:276
Selectivity outer_match_frac
Definition: relation.h:2291
Path * innerjoinpath
Definition: relation.h:1427
static MergeScanSelCache * cached_scansel(PlannerInfo *root, RestrictInfo *rinfo, PathKey *pathkey)
Definition: costsize.c:3074
static double approx_tuple_count(PlannerInfo *root, JoinPath *path, List *quals)
Definition: costsize.c:4239
void set_joinrel_size_estimates(PlannerInfo *root, RelOptInfo *rel, RelOptInfo *outer_rel, RelOptInfo *inner_rel, SpecialJoinInfo *sjinfo, List *restrictlist)
Definition: costsize.c:4375
void set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
Definition: costsize.c:4982
void cost_namedtuplestorescan(Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
Definition: costsize.c:1541
double tuples
Definition: relation.h:652
Oid reltablespace
Definition: relation.h:641
List * baserestrictinfo
Definition: relation.h:672
Oid resulttype
Definition: primnodes.h:815
NodeTag type
Definition: relation.h:2065
#define DEFAULT_PARALLEL_SETUP_COST
Definition: cost.h:30
#define Min(x, y)
Definition: c.h:857
int parallel_workers
Definition: relation.h:1085
bool pseudoconstant
Definition: relation.h:1888
void cost_bitmap_and_node(BitmapAndPath *path, PlannerInfo *root)
Definition: costsize.c:1089
int effective_cache_size
Definition: costsize.c:119
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
SubLinkType subLinkType
Definition: primnodes.h:684
#define IS_OUTER_JOIN(jointype)
Definition: nodes.h:729
ParamPathInfo * param_info
Definition: relation.h:1081
bool enable_parallel_hash
Definition: costsize.c:140
List * list_copy(const List *oldlist)
Definition: list.c:1160
Definition: nodes.h:516
#define MemSet(start, val, len)
Definition: c.h:908
Relids left_relids
Definition: relation.h:1907
static Cost append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
Definition: costsize.c:1767
AttrNumber varattno
Definition: primnodes.h:168
void initial_cost_hashjoin(PlannerInfo *root, JoinCostWorkspace *workspace, JoinType jointype, List *hashclauses, Path *outer_path, Path *inner_path, JoinPathExtraData *extra, bool parallel_hash)
Definition: costsize.c:3153
void cost_valuesscan(Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
Definition: costsize.c:1450
static double page_size(double tuples, int width)
Definition: costsize.c:5340
void cost_ctescan(Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
Definition: costsize.c:1500
List * list_concat(List *list1, List *list2)
Definition: list.c:321
bool enable_parallel_append
Definition: costsize.c:139
double parallel_setup_cost
Definition: costsize.c:117
uint32 BlockNumber
Definition: block.h:31
static List * extract_nonindex_conditions(List *qual_clauses, List *indexquals)
Definition: costsize.c:770
Definition: nodes.h:76
double Selectivity
Definition: nodes.h:646
QualCost transCost
Definition: relation.h:63
float4 get_func_cost(Oid funcid)
Definition: lsyscache.c:1645
Selectivity bitmapselectivity
Definition: relation.h:1199
unsigned int Oid
Definition: postgres_ext.h:31
Definition: primnodes.h:163
bool enable_seqscan
Definition: costsize.c:125
List * fkey_list
Definition: relation.h:272
#define DEFAULT_EFFECTIVE_CACHE_SIZE
Definition: cost.h:32
double inner_rows_total
Definition: relation.h:1503
#define DEFAULT_CPU_OPERATOR_COST
Definition: cost.h:28
void initial_cost_mergejoin(PlannerInfo *root, JoinCostWorkspace *workspace, JoinType jointype, List *mergeclauses, Path *outer_path, Path *inner_path, List *outersortkeys, List *innersortkeys, JoinPathExtraData *extra)
Definition: costsize.c:2600
bool enable_partitionwise_aggregate
Definition: costsize.c:138
Cost inner_rescan_run_cost
Definition: relation.h:2406
List * values_lists
Definition: parsenodes.h:1027
#define lsecond(l)
Definition: pg_list.h:116
Relids syn_lefthand
Definition: relation.h:2068
int pk_strategy
Definition: relation.h:977
Cost startup
Definition: relation.h:46
double allvisfrac
Definition: relation.h:653
signed int int32
Definition: c.h:313
List * bitmapquals
Definition: relation.h:1198
Path path
Definition: relation.h:1304
JoinType
Definition: nodes.h:680
List * targetList
Definition: parsenodes.h:140
int first_partial_path
Definition: relation.h:1310
struct RelOptInfo ** simple_rel_array
Definition: relation.h:193
List * bitmapquals
Definition: relation.h:1211
Definition: type.h:89
BlockNumber pages
Definition: relation.h:758
NodeTag pathtype
Definition: relation.h:1076
Relids syn_righthand
Definition: relation.h:2069
List * subpaths
Definition: relation.h:1307
PlannerInfo * subroot
Definition: relation.h:654
bool enable_sort
Definition: costsize.c:130
void final_cost_nestloop(PlannerInfo *root, NestPath *path, JoinCostWorkspace *workspace, JoinPathExtraData *extra)
Definition: costsize.c:2393
double random_page_cost
Definition: costsize.c:112
void cost_seqscan(Path *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
Definition: costsize.c:212
static Selectivity get_foreign_key_join_selectivity(PlannerInfo *root, Relids outer_relids, Relids inner_relids, SpecialJoinInfo *sjinfo, List **restrictlist)
Definition: costsize.c:4600
Cost per_tuple
Definition: relation.h:47
List * indexquals
Definition: relation.h:1157
int estimate_array_length(Node *arrayexpr)
Definition: selfuncs.c:2179
bool skip_mark_restore
Definition: relation.h:1485
static void set_rel_width(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5093
RelOptInfo * rel
Definition: relation.h:755
#define RINFO_IS_PUSHED_DOWN(rinfo, joinrelids)
Definition: relation.h:1957
bool resjunk
Definition: primnodes.h:1382
#define linitial(l)
Definition: pg_list.h:111
#define planner_rt_fetch(rti, root)
Definition: relation.h:344
List * make_ands_implicit(Expr *clause)
Definition: clauses.c:379
bool pk_nulls_first
Definition: relation.h:978
#define ERROR
Definition: elog.h:43
Expr * phexpr
Definition: relation.h:1999
TableFunc * tablefunc
Definition: parsenodes.h:1022
Oid vartype
Definition: primnodes.h:170
void cost_qual_eval(QualCost *cost, List *quals, PlannerInfo *root)
Definition: costsize.c:3716
Node * get_leftop(const Expr *clause)
Definition: clauses.c:200
Cost indextotalcost
Definition: relation.h:1162
Cost startup_cost
Definition: relation.h:1089
Cost disable_cost
Definition: costsize.c:121
List * semi_rhs_exprs
Definition: relation.h:2077
void cost_subqueryscan(SubqueryScanPath *path, PlannerInfo *root, RelOptInfo *baserel, ParamPathInfo *param_info)
Definition: costsize.c:1284
void cost_group(Path *path, PlannerInfo *root, int numGroupCols, double numGroups, List *quals, Cost input_startup_cost, Cost input_total_cost, double input_tuples)
Definition: costsize.c:2248
#define DEFAULT_RANDOM_PAGE_COST
Definition: cost.h:25
bool semi_can_hash
Definition: relation.h:2075
List * joinrestrictinfo
Definition: relation.h:1429
EquivalenceClass * parent_ec
Definition: relation.h:1914
RelOptInfo * parent
Definition: relation.h:1078
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:374
RelOptInfo * fetch_upper_rel(PlannerInfo *root, UpperRelationKind kind, Relids relids)
Definition: relnode.c:1145
#define lfirst_node(type, lc)
Definition: pg_list.h:109
static const uint32 T[65]
Definition: md5.c:101
double get_parameterized_joinrel_size(PlannerInfo *root, RelOptInfo *rel, Path *outer_path, Path *inner_path, SpecialJoinInfo *sjinfo, List *restrict_clauses)
Definition: costsize.c:4407
struct Path * cheapest_total_path
Definition: relation.h:630
Selectivity rightstartsel
Definition: relation.h:1978
Selectivity indexselectivity
Definition: relation.h:1163
Selectivity clause_selectivity(PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:574
Cost startup_cost
Definition: plannodes.h:128
PlaceHolderInfo * find_placeholder_info(PlannerInfo *root, PlaceHolderVar *phv, bool create_new_ph)
Definition: placeholder.c:70
AttrNumber resno
Definition: primnodes.h:1376
static ListCell * list_head(const List *l)
Definition: pg_list.h:77
static double get_parallel_divisor(Path *path)
Definition: costsize.c:5350
Relids relids
Definition: relation.h:612
double cpu_operator_cost
Definition: costsize.c:115
Path * subpath
Definition: relation.h:1395
Oid winfnoid
Definition: primnodes.h:355
Expr * arg
Definition: primnodes.h:814
double total_table_pages
Definition: relation.h:304
Selectivity bitmapselectivity
Definition: relation.h:1212
Expr * elemexpr
Definition: primnodes.h:839
NextSampleBlock_function NextSampleBlock
Definition: tsmapi.h:72