PostgreSQL Source Code  git master
allpaths.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * allpaths.c
4  * Routines to find possible search paths for processing a query
5  *
6  * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7  * Portions Copyright (c) 1994, Regents of the University of California
8  *
9  *
10  * IDENTIFICATION
11  * src/backend/optimizer/path/allpaths.c
12  *
13  *-------------------------------------------------------------------------
14  */
15 
16 #include "postgres.h"
17 
18 #include <limits.h>
19 #include <math.h>
20 
21 #include "access/sysattr.h"
22 #include "access/tsmapi.h"
23 #include "catalog/pg_class.h"
24 #include "catalog/pg_operator.h"
25 #include "catalog/pg_proc.h"
26 #include "foreign/fdwapi.h"
27 #include "miscadmin.h"
28 #include "nodes/makefuncs.h"
29 #include "nodes/nodeFuncs.h"
30 #include "nodes/supportnodes.h"
31 #ifdef OPTIMIZER_DEBUG
32 #include "nodes/print.h"
33 #endif
34 #include "optimizer/appendinfo.h"
35 #include "optimizer/clauses.h"
36 #include "optimizer/cost.h"
37 #include "optimizer/geqo.h"
38 #include "optimizer/optimizer.h"
39 #include "optimizer/pathnode.h"
40 #include "optimizer/paths.h"
41 #include "optimizer/plancat.h"
42 #include "optimizer/planner.h"
43 #include "optimizer/tlist.h"
44 #include "parser/parse_clause.h"
45 #include "parser/parsetree.h"
47 #include "port/pg_bitutils.h"
48 #include "rewrite/rewriteManip.h"
49 #include "utils/lsyscache.h"
50 
51 
52 /* Bitmask flags for pushdown_safety_info.unsafeFlags */
53 #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
54 #define UNSAFE_HAS_SET_FUNC (1 << 1)
55 #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
56 #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
57 #define UNSAFE_TYPE_MISMATCH (1 << 4)
58 
59 /* results of subquery_is_pushdown_safe */
60 typedef struct pushdown_safety_info
61 {
62  unsigned char *unsafeFlags; /* bitmask of reasons why this target list
63  * column is unsafe for qual pushdown, or 0 if
64  * no reason. */
65  bool unsafeVolatile; /* don't push down volatile quals */
66  bool unsafeLeaky; /* don't push down leaky quals */
68 
69 /* Return type for qual_is_pushdown_safe */
70 typedef enum pushdown_safe_type
71 {
72  PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
73  PUSHDOWN_SAFE, /* safe to push qual into subquery */
74  PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
75  * run condition */
77 
78 /* These parameters are set by GUC */
79 bool enable_geqo = false; /* just in case GUC doesn't set it */
83 
84 /* Hook for plugins to get control in set_rel_pathlist() */
86 
87 /* Hook for plugins to replace standard_join_search() */
89 
90 
92 static void set_base_rel_sizes(PlannerInfo *root);
93 static void set_base_rel_pathlists(PlannerInfo *root);
94 static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
95  Index rti, RangeTblEntry *rte);
96 static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
97  Index rti, RangeTblEntry *rte);
98 static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
99  RangeTblEntry *rte);
100 static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
101 static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
102  RangeTblEntry *rte);
103 static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
104  RangeTblEntry *rte);
105 static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
106  RangeTblEntry *rte);
107 static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
108  RangeTblEntry *rte);
109 static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
110  RangeTblEntry *rte);
111 static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
112  RangeTblEntry *rte);
113 static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
114  Index rti, RangeTblEntry *rte);
115 static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
116  Index rti, RangeTblEntry *rte);
117 static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
118  List *live_childrels,
119  List *all_child_pathkeys);
121  RelOptInfo *rel,
122  Relids required_outer);
123 static void accumulate_append_subpath(Path *path,
124  List **subpaths,
125  List **special_subpaths);
126 static Path *get_singleton_append_subpath(Path *path);
127 static void set_dummy_rel_pathlist(RelOptInfo *rel);
128 static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
129  Index rti, RangeTblEntry *rte);
130 static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
131  RangeTblEntry *rte);
132 static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
133  RangeTblEntry *rte);
134 static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
135  RangeTblEntry *rte);
136 static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
137  RangeTblEntry *rte);
138 static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
139  RangeTblEntry *rte);
140 static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
141  RangeTblEntry *rte);
142 static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
143  RangeTblEntry *rte);
144 static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
145 static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
146  pushdown_safety_info *safetyInfo);
147 static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
148  pushdown_safety_info *safetyInfo);
149 static void check_output_expressions(Query *subquery,
150  pushdown_safety_info *safetyInfo);
151 static void compare_tlist_datatypes(List *tlist, List *colTypes,
152  pushdown_safety_info *safetyInfo);
153 static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
155  RestrictInfo *rinfo,
156  pushdown_safety_info *safetyInfo);
157 static void subquery_push_qual(Query *subquery,
158  RangeTblEntry *rte, Index rti, Node *qual);
159 static void recurse_push_qual(Node *setOp, Query *topquery,
160  RangeTblEntry *rte, Index rti, Node *qual);
161 static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
162  Bitmapset *extra_used_attrs);
163 
164 
165 /*
166  * make_one_rel
167  * Finds all possible access paths for executing a query, returning a
168  * single rel that represents the join of all base rels in the query.
169  */
170 RelOptInfo *
171 make_one_rel(PlannerInfo *root, List *joinlist)
172 {
173  RelOptInfo *rel;
174  Index rti;
175  double total_pages;
176 
177  /* Mark base rels as to whether we care about fast-start plans */
179 
180  /*
181  * Compute size estimates and consider_parallel flags for each base rel.
182  */
183  set_base_rel_sizes(root);
184 
185  /*
186  * We should now have size estimates for every actual table involved in
187  * the query, and we also know which if any have been deleted from the
188  * query by join removal, pruned by partition pruning, or eliminated by
189  * constraint exclusion. So we can now compute total_table_pages.
190  *
191  * Note that appendrels are not double-counted here, even though we don't
192  * bother to distinguish RelOptInfos for appendrel parents, because the
193  * parents will have pages = 0.
194  *
195  * XXX if a table is self-joined, we will count it once per appearance,
196  * which perhaps is the wrong thing ... but that's not completely clear,
197  * and detecting self-joins here is difficult, so ignore it for now.
198  */
199  total_pages = 0;
200  for (rti = 1; rti < root->simple_rel_array_size; rti++)
201  {
202  RelOptInfo *brel = root->simple_rel_array[rti];
203 
204  /* there may be empty slots corresponding to non-baserel RTEs */
205  if (brel == NULL)
206  continue;
207 
208  Assert(brel->relid == rti); /* sanity check on array */
209 
210  if (IS_DUMMY_REL(brel))
211  continue;
212 
213  if (IS_SIMPLE_REL(brel))
214  total_pages += (double) brel->pages;
215  }
216  root->total_table_pages = total_pages;
217 
218  /*
219  * Generate access paths for each base rel.
220  */
222 
223  /*
224  * Generate access paths for the entire join tree.
225  */
226  rel = make_rel_from_joinlist(root, joinlist);
227 
228  /*
229  * The result should join all and only the query's base + outer-join rels.
230  */
231  Assert(bms_equal(rel->relids, root->all_query_rels));
232 
233  return rel;
234 }
235 
236 /*
237  * set_base_rel_consider_startup
238  * Set the consider_[param_]startup flags for each base-relation entry.
239  *
240  * For the moment, we only deal with consider_param_startup here; because the
241  * logic for consider_startup is pretty trivial and is the same for every base
242  * relation, we just let build_simple_rel() initialize that flag correctly to
243  * start with. If that logic ever gets more complicated it would probably
244  * be better to move it here.
245  */
246 static void
248 {
249  /*
250  * Since parameterized paths can only be used on the inside of a nestloop
251  * join plan, there is usually little value in considering fast-start
252  * plans for them. However, for relations that are on the RHS of a SEMI
253  * or ANTI join, a fast-start plan can be useful because we're only going
254  * to care about fetching one tuple anyway.
255  *
256  * To minimize growth of planning time, we currently restrict this to
257  * cases where the RHS is a single base relation, not a join; there is no
258  * provision for consider_param_startup to get set at all on joinrels.
259  * Also we don't worry about appendrels. costsize.c's costing rules for
260  * nestloop semi/antijoins don't consider such cases either.
261  */
262  ListCell *lc;
263 
264  foreach(lc, root->join_info_list)
265  {
266  SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
267  int varno;
268 
269  if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
270  bms_get_singleton_member(sjinfo->syn_righthand, &varno))
271  {
272  RelOptInfo *rel = find_base_rel(root, varno);
273 
274  rel->consider_param_startup = true;
275  }
276  }
277 }
278 
279 /*
280  * set_base_rel_sizes
281  * Set the size estimates (rows and widths) for each base-relation entry.
282  * Also determine whether to consider parallel paths for base relations.
283  *
284  * We do this in a separate pass over the base rels so that rowcount
285  * estimates are available for parameterized path generation, and also so
286  * that each rel's consider_parallel flag is set correctly before we begin to
287  * generate paths.
288  */
289 static void
291 {
292  Index rti;
293 
294  for (rti = 1; rti < root->simple_rel_array_size; rti++)
295  {
296  RelOptInfo *rel = root->simple_rel_array[rti];
297  RangeTblEntry *rte;
298 
299  /* there may be empty slots corresponding to non-baserel RTEs */
300  if (rel == NULL)
301  continue;
302 
303  Assert(rel->relid == rti); /* sanity check on array */
304 
305  /* ignore RTEs that are "other rels" */
306  if (rel->reloptkind != RELOPT_BASEREL)
307  continue;
308 
309  rte = root->simple_rte_array[rti];
310 
311  /*
312  * If parallelism is allowable for this query in general, see whether
313  * it's allowable for this rel in particular. We have to do this
314  * before set_rel_size(), because (a) if this rel is an inheritance
315  * parent, set_append_rel_size() will use and perhaps change the rel's
316  * consider_parallel flag, and (b) for some RTE types, set_rel_size()
317  * goes ahead and makes paths immediately.
318  */
319  if (root->glob->parallelModeOK)
320  set_rel_consider_parallel(root, rel, rte);
321 
322  set_rel_size(root, rel, rti, rte);
323  }
324 }
325 
326 /*
327  * set_base_rel_pathlists
328  * Finds all paths available for scanning each base-relation entry.
329  * Sequential scan and any available indices are considered.
330  * Each useful path is attached to its relation's 'pathlist' field.
331  */
332 static void
334 {
335  Index rti;
336 
337  for (rti = 1; rti < root->simple_rel_array_size; rti++)
338  {
339  RelOptInfo *rel = root->simple_rel_array[rti];
340 
341  /* there may be empty slots corresponding to non-baserel RTEs */
342  if (rel == NULL)
343  continue;
344 
345  Assert(rel->relid == rti); /* sanity check on array */
346 
347  /* ignore RTEs that are "other rels" */
348  if (rel->reloptkind != RELOPT_BASEREL)
349  continue;
350 
351  set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
352  }
353 }
354 
355 /*
356  * set_rel_size
357  * Set size estimates for a base relation
358  */
359 static void
361  Index rti, RangeTblEntry *rte)
362 {
363  if (rel->reloptkind == RELOPT_BASEREL &&
364  relation_excluded_by_constraints(root, rel, rte))
365  {
366  /*
367  * We proved we don't need to scan the rel via constraint exclusion,
368  * so set up a single dummy path for it. Here we only check this for
369  * regular baserels; if it's an otherrel, CE was already checked in
370  * set_append_rel_size().
371  *
372  * In this case, we go ahead and set up the relation's path right away
373  * instead of leaving it for set_rel_pathlist to do. This is because
374  * we don't have a convention for marking a rel as dummy except by
375  * assigning a dummy path to it.
376  */
378  }
379  else if (rte->inh)
380  {
381  /* It's an "append relation", process accordingly */
382  set_append_rel_size(root, rel, rti, rte);
383  }
384  else
385  {
386  switch (rel->rtekind)
387  {
388  case RTE_RELATION:
389  if (rte->relkind == RELKIND_FOREIGN_TABLE)
390  {
391  /* Foreign table */
392  set_foreign_size(root, rel, rte);
393  }
394  else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
395  {
396  /*
397  * We could get here if asked to scan a partitioned table
398  * with ONLY. In that case we shouldn't scan any of the
399  * partitions, so mark it as a dummy rel.
400  */
402  }
403  else if (rte->tablesample != NULL)
404  {
405  /* Sampled relation */
406  set_tablesample_rel_size(root, rel, rte);
407  }
408  else
409  {
410  /* Plain relation */
411  set_plain_rel_size(root, rel, rte);
412  }
413  break;
414  case RTE_SUBQUERY:
415 
416  /*
417  * Subqueries don't support making a choice between
418  * parameterized and unparameterized paths, so just go ahead
419  * and build their paths immediately.
420  */
421  set_subquery_pathlist(root, rel, rti, rte);
422  break;
423  case RTE_FUNCTION:
424  set_function_size_estimates(root, rel);
425  break;
426  case RTE_TABLEFUNC:
427  set_tablefunc_size_estimates(root, rel);
428  break;
429  case RTE_VALUES:
430  set_values_size_estimates(root, rel);
431  break;
432  case RTE_CTE:
433 
434  /*
435  * CTEs don't support making a choice between parameterized
436  * and unparameterized paths, so just go ahead and build their
437  * paths immediately.
438  */
439  if (rte->self_reference)
440  set_worktable_pathlist(root, rel, rte);
441  else
442  set_cte_pathlist(root, rel, rte);
443  break;
444  case RTE_NAMEDTUPLESTORE:
445  /* Might as well just build the path immediately */
446  set_namedtuplestore_pathlist(root, rel, rte);
447  break;
448  case RTE_RESULT:
449  /* Might as well just build the path immediately */
450  set_result_pathlist(root, rel, rte);
451  break;
452  default:
453  elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
454  break;
455  }
456  }
457 
458  /*
459  * We insist that all non-dummy rels have a nonzero rowcount estimate.
460  */
461  Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
462 }
463 
464 /*
465  * set_rel_pathlist
466  * Build access paths for a base relation
467  */
468 static void
470  Index rti, RangeTblEntry *rte)
471 {
472  if (IS_DUMMY_REL(rel))
473  {
474  /* We already proved the relation empty, so nothing more to do */
475  }
476  else if (rte->inh)
477  {
478  /* It's an "append relation", process accordingly */
479  set_append_rel_pathlist(root, rel, rti, rte);
480  }
481  else
482  {
483  switch (rel->rtekind)
484  {
485  case RTE_RELATION:
486  if (rte->relkind == RELKIND_FOREIGN_TABLE)
487  {
488  /* Foreign table */
489  set_foreign_pathlist(root, rel, rte);
490  }
491  else if (rte->tablesample != NULL)
492  {
493  /* Sampled relation */
494  set_tablesample_rel_pathlist(root, rel, rte);
495  }
496  else
497  {
498  /* Plain relation */
499  set_plain_rel_pathlist(root, rel, rte);
500  }
501  break;
502  case RTE_SUBQUERY:
503  /* Subquery --- fully handled during set_rel_size */
504  break;
505  case RTE_FUNCTION:
506  /* RangeFunction */
507  set_function_pathlist(root, rel, rte);
508  break;
509  case RTE_TABLEFUNC:
510  /* Table Function */
511  set_tablefunc_pathlist(root, rel, rte);
512  break;
513  case RTE_VALUES:
514  /* Values list */
515  set_values_pathlist(root, rel, rte);
516  break;
517  case RTE_CTE:
518  /* CTE reference --- fully handled during set_rel_size */
519  break;
520  case RTE_NAMEDTUPLESTORE:
521  /* tuplestore reference --- fully handled during set_rel_size */
522  break;
523  case RTE_RESULT:
524  /* simple Result --- fully handled during set_rel_size */
525  break;
526  default:
527  elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
528  break;
529  }
530  }
531 
532  /*
533  * Allow a plugin to editorialize on the set of Paths for this base
534  * relation. It could add new paths (such as CustomPaths) by calling
535  * add_path(), or add_partial_path() if parallel aware. It could also
536  * delete or modify paths added by the core code.
537  */
539  (*set_rel_pathlist_hook) (root, rel, rti, rte);
540 
541  /*
542  * If this is a baserel, we should normally consider gathering any partial
543  * paths we may have created for it. We have to do this after calling the
544  * set_rel_pathlist_hook, else it cannot add partial paths to be included
545  * here.
546  *
547  * However, if this is an inheritance child, skip it. Otherwise, we could
548  * end up with a very large number of gather nodes, each trying to grab
549  * its own pool of workers. Instead, we'll consider gathering partial
550  * paths for the parent appendrel.
551  *
552  * Also, if this is the topmost scan/join rel, we postpone gathering until
553  * the final scan/join targetlist is available (see grouping_planner).
554  */
555  if (rel->reloptkind == RELOPT_BASEREL &&
556  !bms_equal(rel->relids, root->all_query_rels))
557  generate_useful_gather_paths(root, rel, false);
558 
559  /* Now find the cheapest of the paths for this rel */
560  set_cheapest(rel);
561 
562 #ifdef OPTIMIZER_DEBUG
563  pprint(rel);
564 #endif
565 }
566 
567 /*
568  * set_plain_rel_size
569  * Set size estimates for a plain relation (no subquery, no inheritance)
570  */
571 static void
573 {
574  /*
575  * Test any partial indexes of rel for applicability. We must do this
576  * first since partial unique indexes can affect size estimates.
577  */
578  check_index_predicates(root, rel);
579 
580  /* Mark rel with estimated output rows, width, etc */
581  set_baserel_size_estimates(root, rel);
582 }
583 
584 /*
585  * If this relation could possibly be scanned from within a worker, then set
586  * its consider_parallel flag.
587  */
588 static void
590  RangeTblEntry *rte)
591 {
592  /*
593  * The flag has previously been initialized to false, so we can just
594  * return if it becomes clear that we can't safely set it.
595  */
596  Assert(!rel->consider_parallel);
597 
598  /* Don't call this if parallelism is disallowed for the entire query. */
599  Assert(root->glob->parallelModeOK);
600 
601  /* This should only be called for baserels and appendrel children. */
602  Assert(IS_SIMPLE_REL(rel));
603 
604  /* Assorted checks based on rtekind. */
605  switch (rte->rtekind)
606  {
607  case RTE_RELATION:
608 
609  /*
610  * Currently, parallel workers can't access the leader's temporary
611  * tables. We could possibly relax this if we wrote all of its
612  * local buffers at the start of the query and made no changes
613  * thereafter (maybe we could allow hint bit changes), and if we
614  * taught the workers to read them. Writing a large number of
615  * temporary buffers could be expensive, though, and we don't have
616  * the rest of the necessary infrastructure right now anyway. So
617  * for now, bail out if we see a temporary table.
618  */
619  if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
620  return;
621 
622  /*
623  * Table sampling can be pushed down to workers if the sample
624  * function and its arguments are safe.
625  */
626  if (rte->tablesample != NULL)
627  {
628  char proparallel = func_parallel(rte->tablesample->tsmhandler);
629 
630  if (proparallel != PROPARALLEL_SAFE)
631  return;
632  if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
633  return;
634  }
635 
636  /*
637  * Ask FDWs whether they can support performing a ForeignScan
638  * within a worker. Most often, the answer will be no. For
639  * example, if the nature of the FDW is such that it opens a TCP
640  * connection with a remote server, each parallel worker would end
641  * up with a separate connection, and these connections might not
642  * be appropriately coordinated between workers and the leader.
643  */
644  if (rte->relkind == RELKIND_FOREIGN_TABLE)
645  {
646  Assert(rel->fdwroutine);
647  if (!rel->fdwroutine->IsForeignScanParallelSafe)
648  return;
649  if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
650  return;
651  }
652 
653  /*
654  * There are additional considerations for appendrels, which we'll
655  * deal with in set_append_rel_size and set_append_rel_pathlist.
656  * For now, just set consider_parallel based on the rel's own
657  * quals and targetlist.
658  */
659  break;
660 
661  case RTE_SUBQUERY:
662 
663  /*
664  * There's no intrinsic problem with scanning a subquery-in-FROM
665  * (as distinct from a SubPlan or InitPlan) in a parallel worker.
666  * If the subquery doesn't happen to have any parallel-safe paths,
667  * then flagging it as consider_parallel won't change anything,
668  * but that's true for plain tables, too. We must set
669  * consider_parallel based on the rel's own quals and targetlist,
670  * so that if a subquery path is parallel-safe but the quals and
671  * projection we're sticking onto it are not, we correctly mark
672  * the SubqueryScanPath as not parallel-safe. (Note that
673  * set_subquery_pathlist() might push some of these quals down
674  * into the subquery itself, but that doesn't change anything.)
675  *
676  * We can't push sub-select containing LIMIT/OFFSET to workers as
677  * there is no guarantee that the row order will be fully
678  * deterministic, and applying LIMIT/OFFSET will lead to
679  * inconsistent results at the top-level. (In some cases, where
680  * the result is ordered, we could relax this restriction. But it
681  * doesn't currently seem worth expending extra effort to do so.)
682  */
683  {
684  Query *subquery = castNode(Query, rte->subquery);
685 
686  if (limit_needed(subquery))
687  return;
688  }
689  break;
690 
691  case RTE_JOIN:
692  /* Shouldn't happen; we're only considering baserels here. */
693  Assert(false);
694  return;
695 
696  case RTE_FUNCTION:
697  /* Check for parallel-restricted functions. */
698  if (!is_parallel_safe(root, (Node *) rte->functions))
699  return;
700  break;
701 
702  case RTE_TABLEFUNC:
703  /* not parallel safe */
704  return;
705 
706  case RTE_VALUES:
707  /* Check for parallel-restricted functions. */
708  if (!is_parallel_safe(root, (Node *) rte->values_lists))
709  return;
710  break;
711 
712  case RTE_CTE:
713 
714  /*
715  * CTE tuplestores aren't shared among parallel workers, so we
716  * force all CTE scans to happen in the leader. Also, populating
717  * the CTE would require executing a subplan that's not available
718  * in the worker, might be parallel-restricted, and must get
719  * executed only once.
720  */
721  return;
722 
723  case RTE_NAMEDTUPLESTORE:
724 
725  /*
726  * tuplestore cannot be shared, at least without more
727  * infrastructure to support that.
728  */
729  return;
730 
731  case RTE_RESULT:
732  /* RESULT RTEs, in themselves, are no problem. */
733  break;
734  }
735 
736  /*
737  * If there's anything in baserestrictinfo that's parallel-restricted, we
738  * give up on parallelizing access to this relation. We could consider
739  * instead postponing application of the restricted quals until we're
740  * above all the parallelism in the plan tree, but it's not clear that
741  * that would be a win in very many cases, and it might be tricky to make
742  * outer join clauses work correctly. It would likely break equivalence
743  * classes, too.
744  */
745  if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
746  return;
747 
748  /*
749  * Likewise, if the relation's outputs are not parallel-safe, give up.
750  * (Usually, they're just Vars, but sometimes they're not.)
751  */
752  if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
753  return;
754 
755  /* We have a winner. */
756  rel->consider_parallel = true;
757 }
758 
759 /*
760  * set_plain_rel_pathlist
761  * Build access paths for a plain relation (no subquery, no inheritance)
762  */
763 static void
765 {
766  Relids required_outer;
767 
768  /*
769  * We don't support pushing join clauses into the quals of a seqscan, but
770  * it could still have required parameterization due to LATERAL refs in
771  * its tlist.
772  */
773  required_outer = rel->lateral_relids;
774 
775  /* Consider sequential scan */
776  add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
777 
778  /* If appropriate, consider parallel sequential scan */
779  if (rel->consider_parallel && required_outer == NULL)
780  create_plain_partial_paths(root, rel);
781 
782  /* Consider index scans */
783  create_index_paths(root, rel);
784 
785  /* Consider TID scans */
786  create_tidscan_paths(root, rel);
787 }
788 
789 /*
790  * create_plain_partial_paths
791  * Build partial access paths for parallel scan of a plain relation
792  */
793 static void
795 {
796  int parallel_workers;
797 
798  parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
800 
801  /* If any limit was set to zero, the user doesn't want a parallel scan. */
802  if (parallel_workers <= 0)
803  return;
804 
805  /* Add an unordered partial path based on a parallel sequential scan. */
806  add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
807 }
808 
809 /*
810  * set_tablesample_rel_size
811  * Set size estimates for a sampled relation
812  */
813 static void
815 {
816  TableSampleClause *tsc = rte->tablesample;
817  TsmRoutine *tsm;
818  BlockNumber pages;
819  double tuples;
820 
821  /*
822  * Test any partial indexes of rel for applicability. We must do this
823  * first since partial unique indexes can affect size estimates.
824  */
825  check_index_predicates(root, rel);
826 
827  /*
828  * Call the sampling method's estimation function to estimate the number
829  * of pages it will read and the number of tuples it will return. (Note:
830  * we assume the function returns sane values.)
831  */
832  tsm = GetTsmRoutine(tsc->tsmhandler);
833  tsm->SampleScanGetSampleSize(root, rel, tsc->args,
834  &pages, &tuples);
835 
836  /*
837  * For the moment, because we will only consider a SampleScan path for the
838  * rel, it's okay to just overwrite the pages and tuples estimates for the
839  * whole relation. If we ever consider multiple path types for sampled
840  * rels, we'll need more complication.
841  */
842  rel->pages = pages;
843  rel->tuples = tuples;
844 
845  /* Mark rel with estimated output rows, width, etc */
846  set_baserel_size_estimates(root, rel);
847 }
848 
849 /*
850  * set_tablesample_rel_pathlist
851  * Build access paths for a sampled relation
852  */
853 static void
855 {
856  Relids required_outer;
857  Path *path;
858 
859  /*
860  * We don't support pushing join clauses into the quals of a samplescan,
861  * but it could still have required parameterization due to LATERAL refs
862  * in its tlist or TABLESAMPLE arguments.
863  */
864  required_outer = rel->lateral_relids;
865 
866  /* Consider sampled scan */
867  path = create_samplescan_path(root, rel, required_outer);
868 
869  /*
870  * If the sampling method does not support repeatable scans, we must avoid
871  * plans that would scan the rel multiple times. Ideally, we'd simply
872  * avoid putting the rel on the inside of a nestloop join; but adding such
873  * a consideration to the planner seems like a great deal of complication
874  * to support an uncommon usage of second-rate sampling methods. Instead,
875  * if there is a risk that the query might perform an unsafe join, just
876  * wrap the SampleScan in a Materialize node. We can check for joins by
877  * counting the membership of all_query_rels (note that this correctly
878  * counts inheritance trees as single rels). If we're inside a subquery,
879  * we can't easily check whether a join might occur in the outer query, so
880  * just assume one is possible.
881  *
882  * GetTsmRoutine is relatively expensive compared to the other tests here,
883  * so check repeatable_across_scans last, even though that's a bit odd.
884  */
885  if ((root->query_level > 1 ||
888  {
889  path = (Path *) create_material_path(rel, path);
890  }
891 
892  add_path(rel, path);
893 
894  /* For the moment, at least, there are no other paths to consider */
895 }
896 
897 /*
898  * set_foreign_size
899  * Set size estimates for a foreign table RTE
900  */
901 static void
903 {
904  /* Mark rel with estimated output rows, width, etc */
905  set_foreign_size_estimates(root, rel);
906 
907  /* Let FDW adjust the size estimates, if it can */
908  rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
909 
910  /* ... but do not let it set the rows estimate to zero */
911  rel->rows = clamp_row_est(rel->rows);
912 
913  /*
914  * Also, make sure rel->tuples is not insane relative to rel->rows.
915  * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
916  * FDW doesn't do anything to replace that.
917  */
918  rel->tuples = Max(rel->tuples, rel->rows);
919 }
920 
921 /*
922  * set_foreign_pathlist
923  * Build access paths for a foreign table RTE
924  */
925 static void
927 {
928  /* Call the FDW's GetForeignPaths function to generate path(s) */
929  rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
930 }
931 
932 /*
933  * set_append_rel_size
934  * Set size estimates for a simple "append relation"
935  *
936  * The passed-in rel and RTE represent the entire append relation. The
937  * relation's contents are computed by appending together the output of the
938  * individual member relations. Note that in the non-partitioned inheritance
939  * case, the first member relation is actually the same table as is mentioned
940  * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
941  * a good thing because their outputs are not the same size.
942  */
943 static void
945  Index rti, RangeTblEntry *rte)
946 {
947  int parentRTindex = rti;
948  bool has_live_children;
949  double parent_rows;
950  double parent_size;
951  double *parent_attrsizes;
952  int nattrs;
953  ListCell *l;
954 
955  /* Guard against stack overflow due to overly deep inheritance tree. */
957 
958  Assert(IS_SIMPLE_REL(rel));
959 
960  /*
961  * If this is a partitioned baserel, set the consider_partitionwise_join
962  * flag; currently, we only consider partitionwise joins with the baserel
963  * if its targetlist doesn't contain a whole-row Var.
964  */
966  rel->reloptkind == RELOPT_BASEREL &&
967  rte->relkind == RELKIND_PARTITIONED_TABLE &&
968  bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
969  rel->consider_partitionwise_join = true;
970 
971  /*
972  * Initialize to compute size estimates for whole append relation.
973  *
974  * We handle width estimates by weighting the widths of different child
975  * rels proportionally to their number of rows. This is sensible because
976  * the use of width estimates is mainly to compute the total relation
977  * "footprint" if we have to sort or hash it. To do this, we sum the
978  * total equivalent size (in "double" arithmetic) and then divide by the
979  * total rowcount estimate. This is done separately for the total rel
980  * width and each attribute.
981  *
982  * Note: if you consider changing this logic, beware that child rels could
983  * have zero rows and/or width, if they were excluded by constraints.
984  */
985  has_live_children = false;
986  parent_rows = 0;
987  parent_size = 0;
988  nattrs = rel->max_attr - rel->min_attr + 1;
989  parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
990 
991  foreach(l, root->append_rel_list)
992  {
993  AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
994  int childRTindex;
995  RangeTblEntry *childRTE;
996  RelOptInfo *childrel;
997  List *childrinfos;
998  ListCell *parentvars;
999  ListCell *childvars;
1000  ListCell *lc;
1001 
1002  /* append_rel_list contains all append rels; ignore others */
1003  if (appinfo->parent_relid != parentRTindex)
1004  continue;
1005 
1006  childRTindex = appinfo->child_relid;
1007  childRTE = root->simple_rte_array[childRTindex];
1008 
1009  /*
1010  * The child rel's RelOptInfo was already created during
1011  * add_other_rels_to_query.
1012  */
1013  childrel = find_base_rel(root, childRTindex);
1015 
1016  /* We may have already proven the child to be dummy. */
1017  if (IS_DUMMY_REL(childrel))
1018  continue;
1019 
1020  /*
1021  * We have to copy the parent's targetlist and quals to the child,
1022  * with appropriate substitution of variables. However, the
1023  * baserestrictinfo quals were already copied/substituted when the
1024  * child RelOptInfo was built. So we don't need any additional setup
1025  * before applying constraint exclusion.
1026  */
1027  if (relation_excluded_by_constraints(root, childrel, childRTE))
1028  {
1029  /*
1030  * This child need not be scanned, so we can omit it from the
1031  * appendrel.
1032  */
1033  set_dummy_rel_pathlist(childrel);
1034  continue;
1035  }
1036 
1037  /*
1038  * Constraint exclusion failed, so copy the parent's join quals and
1039  * targetlist to the child, with appropriate variable substitutions.
1040  *
1041  * We skip join quals that came from above outer joins that can null
1042  * this rel, since they would be of no value while generating paths
1043  * for the child. This saves some effort while processing the child
1044  * rel, and it also avoids an implementation restriction in
1045  * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1046  */
1047  childrinfos = NIL;
1048  foreach(lc, rel->joininfo)
1049  {
1050  RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1051 
1052  if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1053  childrinfos = lappend(childrinfos,
1055  (Node *) rinfo,
1056  1, &appinfo));
1057  }
1058  childrel->joininfo = childrinfos;
1059 
1060  /*
1061  * Now for the child's targetlist.
1062  *
1063  * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1064  * expressions, which otherwise would not occur in a rel's targetlist.
1065  * Code that might be looking at an appendrel child must cope with
1066  * such. (Normally, a rel's targetlist would only include Vars and
1067  * PlaceHolderVars.) XXX we do not bother to update the cost or width
1068  * fields of childrel->reltarget; not clear if that would be useful.
1069  */
1070  childrel->reltarget->exprs = (List *)
1072  (Node *) rel->reltarget->exprs,
1073  1, &appinfo);
1074 
1075  /*
1076  * We have to make child entries in the EquivalenceClass data
1077  * structures as well. This is needed either if the parent
1078  * participates in some eclass joins (because we will want to consider
1079  * inner-indexscan joins on the individual children) or if the parent
1080  * has useful pathkeys (because we should try to build MergeAppend
1081  * paths that produce those sort orderings).
1082  */
1083  if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1084  add_child_rel_equivalences(root, appinfo, rel, childrel);
1085  childrel->has_eclass_joins = rel->has_eclass_joins;
1086 
1087  /*
1088  * Note: we could compute appropriate attr_needed data for the child's
1089  * variables, by transforming the parent's attr_needed through the
1090  * translated_vars mapping. However, currently there's no need
1091  * because attr_needed is only examined for base relations not
1092  * otherrels. So we just leave the child's attr_needed empty.
1093  */
1094 
1095  /*
1096  * If we consider partitionwise joins with the parent rel, do the same
1097  * for partitioned child rels.
1098  *
1099  * Note: here we abuse the consider_partitionwise_join flag by setting
1100  * it for child rels that are not themselves partitioned. We do so to
1101  * tell try_partitionwise_join() that the child rel is sufficiently
1102  * valid to be used as a per-partition input, even if it later gets
1103  * proven to be dummy. (It's not usable until we've set up the
1104  * reltarget and EC entries, which we just did.)
1105  */
1106  if (rel->consider_partitionwise_join)
1107  childrel->consider_partitionwise_join = true;
1108 
1109  /*
1110  * If parallelism is allowable for this query in general, see whether
1111  * it's allowable for this childrel in particular. But if we've
1112  * already decided the appendrel is not parallel-safe as a whole,
1113  * there's no point in considering parallelism for this child. For
1114  * consistency, do this before calling set_rel_size() for the child.
1115  */
1116  if (root->glob->parallelModeOK && rel->consider_parallel)
1117  set_rel_consider_parallel(root, childrel, childRTE);
1118 
1119  /*
1120  * Compute the child's size.
1121  */
1122  set_rel_size(root, childrel, childRTindex, childRTE);
1123 
1124  /*
1125  * It is possible that constraint exclusion detected a contradiction
1126  * within a child subquery, even though we didn't prove one above. If
1127  * so, we can skip this child.
1128  */
1129  if (IS_DUMMY_REL(childrel))
1130  continue;
1131 
1132  /* We have at least one live child. */
1133  has_live_children = true;
1134 
1135  /*
1136  * If any live child is not parallel-safe, treat the whole appendrel
1137  * as not parallel-safe. In future we might be able to generate plans
1138  * in which some children are farmed out to workers while others are
1139  * not; but we don't have that today, so it's a waste to consider
1140  * partial paths anywhere in the appendrel unless it's all safe.
1141  * (Child rels visited before this one will be unmarked in
1142  * set_append_rel_pathlist().)
1143  */
1144  if (!childrel->consider_parallel)
1145  rel->consider_parallel = false;
1146 
1147  /*
1148  * Accumulate size information from each live child.
1149  */
1150  Assert(childrel->rows > 0);
1151 
1152  parent_rows += childrel->rows;
1153  parent_size += childrel->reltarget->width * childrel->rows;
1154 
1155  /*
1156  * Accumulate per-column estimates too. We need not do anything for
1157  * PlaceHolderVars in the parent list. If child expression isn't a
1158  * Var, or we didn't record a width estimate for it, we have to fall
1159  * back on a datatype-based estimate.
1160  *
1161  * By construction, child's targetlist is 1-to-1 with parent's.
1162  */
1163  forboth(parentvars, rel->reltarget->exprs,
1164  childvars, childrel->reltarget->exprs)
1165  {
1166  Var *parentvar = (Var *) lfirst(parentvars);
1167  Node *childvar = (Node *) lfirst(childvars);
1168 
1169  if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1170  {
1171  int pndx = parentvar->varattno - rel->min_attr;
1172  int32 child_width = 0;
1173 
1174  if (IsA(childvar, Var) &&
1175  ((Var *) childvar)->varno == childrel->relid)
1176  {
1177  int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1178 
1179  child_width = childrel->attr_widths[cndx];
1180  }
1181  if (child_width <= 0)
1182  child_width = get_typavgwidth(exprType(childvar),
1183  exprTypmod(childvar));
1184  Assert(child_width > 0);
1185  parent_attrsizes[pndx] += child_width * childrel->rows;
1186  }
1187  }
1188  }
1189 
1190  if (has_live_children)
1191  {
1192  /*
1193  * Save the finished size estimates.
1194  */
1195  int i;
1196 
1197  Assert(parent_rows > 0);
1198  rel->rows = parent_rows;
1199  rel->reltarget->width = rint(parent_size / parent_rows);
1200  for (i = 0; i < nattrs; i++)
1201  rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1202 
1203  /*
1204  * Set "raw tuples" count equal to "rows" for the appendrel; needed
1205  * because some places assume rel->tuples is valid for any baserel.
1206  */
1207  rel->tuples = parent_rows;
1208 
1209  /*
1210  * Note that we leave rel->pages as zero; this is important to avoid
1211  * double-counting the appendrel tree in total_table_pages.
1212  */
1213  }
1214  else
1215  {
1216  /*
1217  * All children were excluded by constraints, so mark the whole
1218  * appendrel dummy. We must do this in this phase so that the rel's
1219  * dummy-ness is visible when we generate paths for other rels.
1220  */
1222  }
1223 
1224  pfree(parent_attrsizes);
1225 }
1226 
1227 /*
1228  * set_append_rel_pathlist
1229  * Build access paths for an "append relation"
1230  */
1231 static void
1233  Index rti, RangeTblEntry *rte)
1234 {
1235  int parentRTindex = rti;
1236  List *live_childrels = NIL;
1237  ListCell *l;
1238 
1239  /*
1240  * Generate access paths for each member relation, and remember the
1241  * non-dummy children.
1242  */
1243  foreach(l, root->append_rel_list)
1244  {
1245  AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1246  int childRTindex;
1247  RangeTblEntry *childRTE;
1248  RelOptInfo *childrel;
1249 
1250  /* append_rel_list contains all append rels; ignore others */
1251  if (appinfo->parent_relid != parentRTindex)
1252  continue;
1253 
1254  /* Re-locate the child RTE and RelOptInfo */
1255  childRTindex = appinfo->child_relid;
1256  childRTE = root->simple_rte_array[childRTindex];
1257  childrel = root->simple_rel_array[childRTindex];
1258 
1259  /*
1260  * If set_append_rel_size() decided the parent appendrel was
1261  * parallel-unsafe at some point after visiting this child rel, we
1262  * need to propagate the unsafety marking down to the child, so that
1263  * we don't generate useless partial paths for it.
1264  */
1265  if (!rel->consider_parallel)
1266  childrel->consider_parallel = false;
1267 
1268  /*
1269  * Compute the child's access paths.
1270  */
1271  set_rel_pathlist(root, childrel, childRTindex, childRTE);
1272 
1273  /*
1274  * If child is dummy, ignore it.
1275  */
1276  if (IS_DUMMY_REL(childrel))
1277  continue;
1278 
1279  /*
1280  * Child is live, so add it to the live_childrels list for use below.
1281  */
1282  live_childrels = lappend(live_childrels, childrel);
1283  }
1284 
1285  /* Add paths to the append relation. */
1286  add_paths_to_append_rel(root, rel, live_childrels);
1287 }
1288 
1289 
1290 /*
1291  * add_paths_to_append_rel
1292  * Generate paths for the given append relation given the set of non-dummy
1293  * child rels.
1294  *
1295  * The function collects all parameterizations and orderings supported by the
1296  * non-dummy children. For every such parameterization or ordering, it creates
1297  * an append path collecting one path from each non-dummy child with given
1298  * parameterization or ordering. Similarly it collects partial paths from
1299  * non-dummy children to create partial append paths.
1300  */
1301 void
1303  List *live_childrels)
1304 {
1305  List *subpaths = NIL;
1306  bool subpaths_valid = true;
1307  List *startup_subpaths = NIL;
1308  bool startup_subpaths_valid = true;
1309  List *partial_subpaths = NIL;
1310  List *pa_partial_subpaths = NIL;
1311  List *pa_nonpartial_subpaths = NIL;
1312  bool partial_subpaths_valid = true;
1313  bool pa_subpaths_valid;
1314  List *all_child_pathkeys = NIL;
1315  List *all_child_outers = NIL;
1316  ListCell *l;
1317  double partial_rows = -1;
1318 
1319  /* If appropriate, consider parallel append */
1320  pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1321 
1322  /*
1323  * For every non-dummy child, remember the cheapest path. Also, identify
1324  * all pathkeys (orderings) and parameterizations (required_outer sets)
1325  * available for the non-dummy member relations.
1326  */
1327  foreach(l, live_childrels)
1328  {
1329  RelOptInfo *childrel = lfirst(l);
1330  ListCell *lcp;
1331  Path *cheapest_partial_path = NULL;
1332 
1333  /*
1334  * If child has an unparameterized cheapest-total path, add that to
1335  * the unparameterized Append path we are constructing for the parent.
1336  * If not, there's no workable unparameterized path.
1337  *
1338  * With partitionwise aggregates, the child rel's pathlist may be
1339  * empty, so don't assume that a path exists here.
1340  */
1341  if (childrel->pathlist != NIL &&
1342  childrel->cheapest_total_path->param_info == NULL)
1344  &subpaths, NULL);
1345  else
1346  subpaths_valid = false;
1347 
1348  /*
1349  * When the planner is considering cheap startup plans, we'll also
1350  * collect all the cheapest_startup_paths (if set) and build an
1351  * AppendPath containing those as subpaths.
1352  */
1353  if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1354  {
1355  /* cheapest_startup_path must not be a parameterized path. */
1356  Assert(childrel->cheapest_startup_path->param_info == NULL);
1358  &startup_subpaths,
1359  NULL);
1360  }
1361  else
1362  startup_subpaths_valid = false;
1363 
1364 
1365  /* Same idea, but for a partial plan. */
1366  if (childrel->partial_pathlist != NIL)
1367  {
1368  cheapest_partial_path = linitial(childrel->partial_pathlist);
1369  accumulate_append_subpath(cheapest_partial_path,
1370  &partial_subpaths, NULL);
1371  }
1372  else
1373  partial_subpaths_valid = false;
1374 
1375  /*
1376  * Same idea, but for a parallel append mixing partial and non-partial
1377  * paths.
1378  */
1379  if (pa_subpaths_valid)
1380  {
1381  Path *nppath = NULL;
1382 
1383  nppath =
1385 
1386  if (cheapest_partial_path == NULL && nppath == NULL)
1387  {
1388  /* Neither a partial nor a parallel-safe path? Forget it. */
1389  pa_subpaths_valid = false;
1390  }
1391  else if (nppath == NULL ||
1392  (cheapest_partial_path != NULL &&
1393  cheapest_partial_path->total_cost < nppath->total_cost))
1394  {
1395  /* Partial path is cheaper or the only option. */
1396  Assert(cheapest_partial_path != NULL);
1397  accumulate_append_subpath(cheapest_partial_path,
1398  &pa_partial_subpaths,
1399  &pa_nonpartial_subpaths);
1400  }
1401  else
1402  {
1403  /*
1404  * Either we've got only a non-partial path, or we think that
1405  * a single backend can execute the best non-partial path
1406  * faster than all the parallel backends working together can
1407  * execute the best partial path.
1408  *
1409  * It might make sense to be more aggressive here. Even if
1410  * the best non-partial path is more expensive than the best
1411  * partial path, it could still be better to choose the
1412  * non-partial path if there are several such paths that can
1413  * be given to different workers. For now, we don't try to
1414  * figure that out.
1415  */
1417  &pa_nonpartial_subpaths,
1418  NULL);
1419  }
1420  }
1421 
1422  /*
1423  * Collect lists of all the available path orderings and
1424  * parameterizations for all the children. We use these as a
1425  * heuristic to indicate which sort orderings and parameterizations we
1426  * should build Append and MergeAppend paths for.
1427  */
1428  foreach(lcp, childrel->pathlist)
1429  {
1430  Path *childpath = (Path *) lfirst(lcp);
1431  List *childkeys = childpath->pathkeys;
1432  Relids childouter = PATH_REQ_OUTER(childpath);
1433 
1434  /* Unsorted paths don't contribute to pathkey list */
1435  if (childkeys != NIL)
1436  {
1437  ListCell *lpk;
1438  bool found = false;
1439 
1440  /* Have we already seen this ordering? */
1441  foreach(lpk, all_child_pathkeys)
1442  {
1443  List *existing_pathkeys = (List *) lfirst(lpk);
1444 
1445  if (compare_pathkeys(existing_pathkeys,
1446  childkeys) == PATHKEYS_EQUAL)
1447  {
1448  found = true;
1449  break;
1450  }
1451  }
1452  if (!found)
1453  {
1454  /* No, so add it to all_child_pathkeys */
1455  all_child_pathkeys = lappend(all_child_pathkeys,
1456  childkeys);
1457  }
1458  }
1459 
1460  /* Unparameterized paths don't contribute to param-set list */
1461  if (childouter)
1462  {
1463  ListCell *lco;
1464  bool found = false;
1465 
1466  /* Have we already seen this param set? */
1467  foreach(lco, all_child_outers)
1468  {
1469  Relids existing_outers = (Relids) lfirst(lco);
1470 
1471  if (bms_equal(existing_outers, childouter))
1472  {
1473  found = true;
1474  break;
1475  }
1476  }
1477  if (!found)
1478  {
1479  /* No, so add it to all_child_outers */
1480  all_child_outers = lappend(all_child_outers,
1481  childouter);
1482  }
1483  }
1484  }
1485  }
1486 
1487  /*
1488  * If we found unparameterized paths for all children, build an unordered,
1489  * unparameterized Append path for the rel. (Note: this is correct even
1490  * if we have zero or one live subpath due to constraint exclusion.)
1491  */
1492  if (subpaths_valid)
1493  add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1494  NIL, NULL, 0, false,
1495  -1));
1496 
1497  /* build an AppendPath for the cheap startup paths, if valid */
1498  if (startup_subpaths_valid)
1499  add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1500  NIL, NIL, NULL, 0, false, -1));
1501 
1502  /*
1503  * Consider an append of unordered, unparameterized partial paths. Make
1504  * it parallel-aware if possible.
1505  */
1506  if (partial_subpaths_valid && partial_subpaths != NIL)
1507  {
1508  AppendPath *appendpath;
1509  ListCell *lc;
1510  int parallel_workers = 0;
1511 
1512  /* Find the highest number of workers requested for any subpath. */
1513  foreach(lc, partial_subpaths)
1514  {
1515  Path *path = lfirst(lc);
1516 
1517  parallel_workers = Max(parallel_workers, path->parallel_workers);
1518  }
1519  Assert(parallel_workers > 0);
1520 
1521  /*
1522  * If the use of parallel append is permitted, always request at least
1523  * log2(# of children) workers. We assume it can be useful to have
1524  * extra workers in this case because they will be spread out across
1525  * the children. The precise formula is just a guess, but we don't
1526  * want to end up with a radically different answer for a table with N
1527  * partitions vs. an unpartitioned table with the same data, so the
1528  * use of some kind of log-scaling here seems to make some sense.
1529  */
1531  {
1532  parallel_workers = Max(parallel_workers,
1533  pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1534  parallel_workers = Min(parallel_workers,
1536  }
1537  Assert(parallel_workers > 0);
1538 
1539  /* Generate a partial append path. */
1540  appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1541  NIL, NULL, parallel_workers,
1543  -1);
1544 
1545  /*
1546  * Make sure any subsequent partial paths use the same row count
1547  * estimate.
1548  */
1549  partial_rows = appendpath->path.rows;
1550 
1551  /* Add the path. */
1552  add_partial_path(rel, (Path *) appendpath);
1553  }
1554 
1555  /*
1556  * Consider a parallel-aware append using a mix of partial and non-partial
1557  * paths. (This only makes sense if there's at least one child which has
1558  * a non-partial path that is substantially cheaper than any partial path;
1559  * otherwise, we should use the append path added in the previous step.)
1560  */
1561  if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1562  {
1563  AppendPath *appendpath;
1564  ListCell *lc;
1565  int parallel_workers = 0;
1566 
1567  /*
1568  * Find the highest number of workers requested for any partial
1569  * subpath.
1570  */
1571  foreach(lc, pa_partial_subpaths)
1572  {
1573  Path *path = lfirst(lc);
1574 
1575  parallel_workers = Max(parallel_workers, path->parallel_workers);
1576  }
1577 
1578  /*
1579  * Same formula here as above. It's even more important in this
1580  * instance because the non-partial paths won't contribute anything to
1581  * the planned number of parallel workers.
1582  */
1583  parallel_workers = Max(parallel_workers,
1584  pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1585  parallel_workers = Min(parallel_workers,
1587  Assert(parallel_workers > 0);
1588 
1589  appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1590  pa_partial_subpaths,
1591  NIL, NULL, parallel_workers, true,
1592  partial_rows);
1593  add_partial_path(rel, (Path *) appendpath);
1594  }
1595 
1596  /*
1597  * Also build unparameterized ordered append paths based on the collected
1598  * list of child pathkeys.
1599  */
1600  if (subpaths_valid)
1601  generate_orderedappend_paths(root, rel, live_childrels,
1602  all_child_pathkeys);
1603 
1604  /*
1605  * Build Append paths for each parameterization seen among the child rels.
1606  * (This may look pretty expensive, but in most cases of practical
1607  * interest, the child rels will expose mostly the same parameterizations,
1608  * so that not that many cases actually get considered here.)
1609  *
1610  * The Append node itself cannot enforce quals, so all qual checking must
1611  * be done in the child paths. This means that to have a parameterized
1612  * Append path, we must have the exact same parameterization for each
1613  * child path; otherwise some children might be failing to check the
1614  * moved-down quals. To make them match up, we can try to increase the
1615  * parameterization of lesser-parameterized paths.
1616  */
1617  foreach(l, all_child_outers)
1618  {
1619  Relids required_outer = (Relids) lfirst(l);
1620  ListCell *lcr;
1621 
1622  /* Select the child paths for an Append with this parameterization */
1623  subpaths = NIL;
1624  subpaths_valid = true;
1625  foreach(lcr, live_childrels)
1626  {
1627  RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1628  Path *subpath;
1629 
1630  if (childrel->pathlist == NIL)
1631  {
1632  /* failed to make a suitable path for this child */
1633  subpaths_valid = false;
1634  break;
1635  }
1636 
1638  childrel,
1639  required_outer);
1640  if (subpath == NULL)
1641  {
1642  /* failed to make a suitable path for this child */
1643  subpaths_valid = false;
1644  break;
1645  }
1646  accumulate_append_subpath(subpath, &subpaths, NULL);
1647  }
1648 
1649  if (subpaths_valid)
1650  add_path(rel, (Path *)
1651  create_append_path(root, rel, subpaths, NIL,
1652  NIL, required_outer, 0, false,
1653  -1));
1654  }
1655 
1656  /*
1657  * When there is only a single child relation, the Append path can inherit
1658  * any ordering available for the child rel's path, so that it's useful to
1659  * consider ordered partial paths. Above we only considered the cheapest
1660  * partial path for each child, but let's also make paths using any
1661  * partial paths that have pathkeys.
1662  */
1663  if (list_length(live_childrels) == 1)
1664  {
1665  RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1666 
1667  /* skip the cheapest partial path, since we already used that above */
1668  for_each_from(l, childrel->partial_pathlist, 1)
1669  {
1670  Path *path = (Path *) lfirst(l);
1671  AppendPath *appendpath;
1672 
1673  /* skip paths with no pathkeys. */
1674  if (path->pathkeys == NIL)
1675  continue;
1676 
1677  appendpath = create_append_path(root, rel, NIL, list_make1(path),
1678  NIL, NULL,
1679  path->parallel_workers, true,
1680  partial_rows);
1681  add_partial_path(rel, (Path *) appendpath);
1682  }
1683  }
1684 }
1685 
1686 /*
1687  * generate_orderedappend_paths
1688  * Generate ordered append paths for an append relation
1689  *
1690  * Usually we generate MergeAppend paths here, but there are some special
1691  * cases where we can generate simple Append paths, because the subpaths
1692  * can provide tuples in the required order already.
1693  *
1694  * We generate a path for each ordering (pathkey list) appearing in
1695  * all_child_pathkeys.
1696  *
1697  * We consider both cheapest-startup and cheapest-total cases, ie, for each
1698  * interesting ordering, collect all the cheapest startup subpaths and all the
1699  * cheapest total paths, and build a suitable path for each case.
1700  *
1701  * We don't currently generate any parameterized ordered paths here. While
1702  * it would not take much more code here to do so, it's very unclear that it
1703  * is worth the planning cycles to investigate such paths: there's little
1704  * use for an ordered path on the inside of a nestloop. In fact, it's likely
1705  * that the current coding of add_path would reject such paths out of hand,
1706  * because add_path gives no credit for sort ordering of parameterized paths,
1707  * and a parameterized MergeAppend is going to be more expensive than the
1708  * corresponding parameterized Append path. If we ever try harder to support
1709  * parameterized mergejoin plans, it might be worth adding support for
1710  * parameterized paths here to feed such joins. (See notes in
1711  * optimizer/README for why that might not ever happen, though.)
1712  */
1713 static void
1715  List *live_childrels,
1716  List *all_child_pathkeys)
1717 {
1718  ListCell *lcp;
1719  List *partition_pathkeys = NIL;
1720  List *partition_pathkeys_desc = NIL;
1721  bool partition_pathkeys_partial = true;
1722  bool partition_pathkeys_desc_partial = true;
1723 
1724  /*
1725  * Some partitioned table setups may allow us to use an Append node
1726  * instead of a MergeAppend. This is possible in cases such as RANGE
1727  * partitioned tables where it's guaranteed that an earlier partition must
1728  * contain rows which come earlier in the sort order. To detect whether
1729  * this is relevant, build pathkey descriptions of the partition ordering,
1730  * for both forward and reverse scans.
1731  */
1732  if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
1733  partitions_are_ordered(rel->boundinfo, rel->live_parts))
1734  {
1735  partition_pathkeys = build_partition_pathkeys(root, rel,
1737  &partition_pathkeys_partial);
1738 
1739  partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1741  &partition_pathkeys_desc_partial);
1742 
1743  /*
1744  * You might think we should truncate_useless_pathkeys here, but
1745  * allowing partition keys which are a subset of the query's pathkeys
1746  * can often be useful. For example, consider a table partitioned by
1747  * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1748  * paths that can produce the a, b, c ordering (perhaps via indexes on
1749  * (a, b, c)) then it works to consider the appendrel output as
1750  * ordered by a, b, c.
1751  */
1752  }
1753 
1754  /* Now consider each interesting sort ordering */
1755  foreach(lcp, all_child_pathkeys)
1756  {
1757  List *pathkeys = (List *) lfirst(lcp);
1758  List *startup_subpaths = NIL;
1759  List *total_subpaths = NIL;
1760  List *fractional_subpaths = NIL;
1761  bool startup_neq_total = false;
1762  bool match_partition_order;
1763  bool match_partition_order_desc;
1764  int end_index;
1765  int first_index;
1766  int direction;
1767 
1768  /*
1769  * Determine if this sort ordering matches any partition pathkeys we
1770  * have, for both ascending and descending partition order. If the
1771  * partition pathkeys happen to be contained in pathkeys then it still
1772  * works, as described above, providing that the partition pathkeys
1773  * are complete and not just a prefix of the partition keys. (In such
1774  * cases we'll be relying on the child paths to have sorted the
1775  * lower-order columns of the required pathkeys.)
1776  */
1777  match_partition_order =
1778  pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1779  (!partition_pathkeys_partial &&
1780  pathkeys_contained_in(partition_pathkeys, pathkeys));
1781 
1782  match_partition_order_desc = !match_partition_order &&
1783  (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1784  (!partition_pathkeys_desc_partial &&
1785  pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1786 
1787  /*
1788  * When the required pathkeys match the reverse of the partition
1789  * order, we must build the list of paths in reverse starting with the
1790  * last matching partition first. We can get away without making any
1791  * special cases for this in the loop below by just looping backward
1792  * over the child relations in this case.
1793  */
1794  if (match_partition_order_desc)
1795  {
1796  /* loop backward */
1797  first_index = list_length(live_childrels) - 1;
1798  end_index = -1;
1799  direction = -1;
1800 
1801  /*
1802  * Set this to true to save us having to check for
1803  * match_partition_order_desc in the loop below.
1804  */
1805  match_partition_order = true;
1806  }
1807  else
1808  {
1809  /* for all other case, loop forward */
1810  first_index = 0;
1811  end_index = list_length(live_childrels);
1812  direction = 1;
1813  }
1814 
1815  /* Select the child paths for this ordering... */
1816  for (int i = first_index; i != end_index; i += direction)
1817  {
1818  RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1819  Path *cheapest_startup,
1820  *cheapest_total,
1821  *cheapest_fractional = NULL;
1822 
1823  /* Locate the right paths, if they are available. */
1824  cheapest_startup =
1826  pathkeys,
1827  NULL,
1828  STARTUP_COST,
1829  false);
1830  cheapest_total =
1832  pathkeys,
1833  NULL,
1834  TOTAL_COST,
1835  false);
1836 
1837  /*
1838  * If we can't find any paths with the right order just use the
1839  * cheapest-total path; we'll have to sort it later.
1840  */
1841  if (cheapest_startup == NULL || cheapest_total == NULL)
1842  {
1843  cheapest_startup = cheapest_total =
1844  childrel->cheapest_total_path;
1845  /* Assert we do have an unparameterized path for this child */
1846  Assert(cheapest_total->param_info == NULL);
1847  }
1848 
1849  /*
1850  * When building a fractional path, determine a cheapest
1851  * fractional path for each child relation too. Looking at startup
1852  * and total costs is not enough, because the cheapest fractional
1853  * path may be dominated by two separate paths (one for startup,
1854  * one for total).
1855  *
1856  * When needed (building fractional path), determine the cheapest
1857  * fractional path too.
1858  */
1859  if (root->tuple_fraction > 0)
1860  {
1861  double path_fraction = (1.0 / root->tuple_fraction);
1862 
1863  cheapest_fractional =
1865  pathkeys,
1866  NULL,
1867  path_fraction);
1868 
1869  /*
1870  * If we found no path with matching pathkeys, use the
1871  * cheapest total path instead.
1872  *
1873  * XXX We might consider partially sorted paths too (with an
1874  * incremental sort on top). But we'd have to build all the
1875  * incremental paths, do the costing etc.
1876  */
1877  if (!cheapest_fractional)
1878  cheapest_fractional = cheapest_total;
1879  }
1880 
1881  /*
1882  * Notice whether we actually have different paths for the
1883  * "cheapest" and "total" cases; frequently there will be no point
1884  * in two create_merge_append_path() calls.
1885  */
1886  if (cheapest_startup != cheapest_total)
1887  startup_neq_total = true;
1888 
1889  /*
1890  * Collect the appropriate child paths. The required logic varies
1891  * for the Append and MergeAppend cases.
1892  */
1893  if (match_partition_order)
1894  {
1895  /*
1896  * We're going to make a plain Append path. We don't need
1897  * most of what accumulate_append_subpath would do, but we do
1898  * want to cut out child Appends or MergeAppends if they have
1899  * just a single subpath (and hence aren't doing anything
1900  * useful).
1901  */
1902  cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1903  cheapest_total = get_singleton_append_subpath(cheapest_total);
1904 
1905  startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1906  total_subpaths = lappend(total_subpaths, cheapest_total);
1907 
1908  if (cheapest_fractional)
1909  {
1910  cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1911  fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1912  }
1913  }
1914  else
1915  {
1916  /*
1917  * Otherwise, rely on accumulate_append_subpath to collect the
1918  * child paths for the MergeAppend.
1919  */
1920  accumulate_append_subpath(cheapest_startup,
1921  &startup_subpaths, NULL);
1922  accumulate_append_subpath(cheapest_total,
1923  &total_subpaths, NULL);
1924 
1925  if (cheapest_fractional)
1926  accumulate_append_subpath(cheapest_fractional,
1927  &fractional_subpaths, NULL);
1928  }
1929  }
1930 
1931  /* ... and build the Append or MergeAppend paths */
1932  if (match_partition_order)
1933  {
1934  /* We only need Append */
1935  add_path(rel, (Path *) create_append_path(root,
1936  rel,
1937  startup_subpaths,
1938  NIL,
1939  pathkeys,
1940  NULL,
1941  0,
1942  false,
1943  -1));
1944  if (startup_neq_total)
1945  add_path(rel, (Path *) create_append_path(root,
1946  rel,
1947  total_subpaths,
1948  NIL,
1949  pathkeys,
1950  NULL,
1951  0,
1952  false,
1953  -1));
1954 
1955  if (fractional_subpaths)
1956  add_path(rel, (Path *) create_append_path(root,
1957  rel,
1958  fractional_subpaths,
1959  NIL,
1960  pathkeys,
1961  NULL,
1962  0,
1963  false,
1964  -1));
1965  }
1966  else
1967  {
1968  /* We need MergeAppend */
1969  add_path(rel, (Path *) create_merge_append_path(root,
1970  rel,
1971  startup_subpaths,
1972  pathkeys,
1973  NULL));
1974  if (startup_neq_total)
1975  add_path(rel, (Path *) create_merge_append_path(root,
1976  rel,
1977  total_subpaths,
1978  pathkeys,
1979  NULL));
1980 
1981  if (fractional_subpaths)
1982  add_path(rel, (Path *) create_merge_append_path(root,
1983  rel,
1984  fractional_subpaths,
1985  pathkeys,
1986  NULL));
1987  }
1988  }
1989 }
1990 
1991 /*
1992  * get_cheapest_parameterized_child_path
1993  * Get cheapest path for this relation that has exactly the requested
1994  * parameterization.
1995  *
1996  * Returns NULL if unable to create such a path.
1997  */
1998 static Path *
2000  Relids required_outer)
2001 {
2002  Path *cheapest;
2003  ListCell *lc;
2004 
2005  /*
2006  * Look up the cheapest existing path with no more than the needed
2007  * parameterization. If it has exactly the needed parameterization, we're
2008  * done.
2009  */
2010  cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2011  NIL,
2012  required_outer,
2013  TOTAL_COST,
2014  false);
2015  Assert(cheapest != NULL);
2016  if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2017  return cheapest;
2018 
2019  /*
2020  * Otherwise, we can "reparameterize" an existing path to match the given
2021  * parameterization, which effectively means pushing down additional
2022  * joinquals to be checked within the path's scan. However, some existing
2023  * paths might check the available joinquals already while others don't;
2024  * therefore, it's not clear which existing path will be cheapest after
2025  * reparameterization. We have to go through them all and find out.
2026  */
2027  cheapest = NULL;
2028  foreach(lc, rel->pathlist)
2029  {
2030  Path *path = (Path *) lfirst(lc);
2031 
2032  /* Can't use it if it needs more than requested parameterization */
2033  if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2034  continue;
2035 
2036  /*
2037  * Reparameterization can only increase the path's cost, so if it's
2038  * already more expensive than the current cheapest, forget it.
2039  */
2040  if (cheapest != NULL &&
2041  compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2042  continue;
2043 
2044  /* Reparameterize if needed, then recheck cost */
2045  if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2046  {
2047  path = reparameterize_path(root, path, required_outer, 1.0);
2048  if (path == NULL)
2049  continue; /* failed to reparameterize this one */
2050  Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2051 
2052  if (cheapest != NULL &&
2053  compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2054  continue;
2055  }
2056 
2057  /* We have a new best path */
2058  cheapest = path;
2059  }
2060 
2061  /* Return the best path, or NULL if we found no suitable candidate */
2062  return cheapest;
2063 }
2064 
2065 /*
2066  * accumulate_append_subpath
2067  * Add a subpath to the list being built for an Append or MergeAppend.
2068  *
2069  * It's possible that the child is itself an Append or MergeAppend path, in
2070  * which case we can "cut out the middleman" and just add its child paths to
2071  * our own list. (We don't try to do this earlier because we need to apply
2072  * both levels of transformation to the quals.)
2073  *
2074  * Note that if we omit a child MergeAppend in this way, we are effectively
2075  * omitting a sort step, which seems fine: if the parent is to be an Append,
2076  * its result would be unsorted anyway, while if the parent is to be a
2077  * MergeAppend, there's no point in a separate sort on a child.
2078  *
2079  * Normally, either path is a partial path and subpaths is a list of partial
2080  * paths, or else path is a non-partial plan and subpaths is a list of those.
2081  * However, if path is a parallel-aware Append, then we add its partial path
2082  * children to subpaths and the rest to special_subpaths. If the latter is
2083  * NULL, we don't flatten the path at all (unless it contains only partial
2084  * paths).
2085  */
2086 static void
2087 accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2088 {
2089  if (IsA(path, AppendPath))
2090  {
2091  AppendPath *apath = (AppendPath *) path;
2092 
2093  if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2094  {
2095  *subpaths = list_concat(*subpaths, apath->subpaths);
2096  return;
2097  }
2098  else if (special_subpaths != NULL)
2099  {
2100  List *new_special_subpaths;
2101 
2102  /* Split Parallel Append into partial and non-partial subpaths */
2103  *subpaths = list_concat(*subpaths,
2104  list_copy_tail(apath->subpaths,
2105  apath->first_partial_path));
2106  new_special_subpaths = list_copy_head(apath->subpaths,
2107  apath->first_partial_path);
2108  *special_subpaths = list_concat(*special_subpaths,
2109  new_special_subpaths);
2110  return;
2111  }
2112  }
2113  else if (IsA(path, MergeAppendPath))
2114  {
2115  MergeAppendPath *mpath = (MergeAppendPath *) path;
2116 
2117  *subpaths = list_concat(*subpaths, mpath->subpaths);
2118  return;
2119  }
2120 
2121  *subpaths = lappend(*subpaths, path);
2122 }
2123 
2124 /*
2125  * get_singleton_append_subpath
2126  * Returns the single subpath of an Append/MergeAppend, or just
2127  * return 'path' if it's not a single sub-path Append/MergeAppend.
2128  *
2129  * Note: 'path' must not be a parallel-aware path.
2130  */
2131 static Path *
2133 {
2134  Assert(!path->parallel_aware);
2135 
2136  if (IsA(path, AppendPath))
2137  {
2138  AppendPath *apath = (AppendPath *) path;
2139 
2140  if (list_length(apath->subpaths) == 1)
2141  return (Path *) linitial(apath->subpaths);
2142  }
2143  else if (IsA(path, MergeAppendPath))
2144  {
2145  MergeAppendPath *mpath = (MergeAppendPath *) path;
2146 
2147  if (list_length(mpath->subpaths) == 1)
2148  return (Path *) linitial(mpath->subpaths);
2149  }
2150 
2151  return path;
2152 }
2153 
2154 /*
2155  * set_dummy_rel_pathlist
2156  * Build a dummy path for a relation that's been excluded by constraints
2157  *
2158  * Rather than inventing a special "dummy" path type, we represent this as an
2159  * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2160  *
2161  * (See also mark_dummy_rel, which does basically the same thing, but is
2162  * typically used to change a rel into dummy state after we already made
2163  * paths for it.)
2164  */
2165 static void
2167 {
2168  /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2169  rel->rows = 0;
2170  rel->reltarget->width = 0;
2171 
2172  /* Discard any pre-existing paths; no further need for them */
2173  rel->pathlist = NIL;
2174  rel->partial_pathlist = NIL;
2175 
2176  /* Set up the dummy path */
2177  add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2178  NIL, rel->lateral_relids,
2179  0, false, -1));
2180 
2181  /*
2182  * We set the cheapest-path fields immediately, just in case they were
2183  * pointing at some discarded path. This is redundant when we're called
2184  * from set_rel_size(), but not when called from elsewhere, and doing it
2185  * twice is harmless anyway.
2186  */
2187  set_cheapest(rel);
2188 }
2189 
2190 /*
2191  * find_window_run_conditions
2192  * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2193  * function to determine the function's monotonic properties. We then
2194  * see if 'opexpr' can be used to short-circuit execution.
2195  *
2196  * For example row_number() over (order by ...) always produces a value one
2197  * higher than the previous. If someone has a window function in a subquery
2198  * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2199  * as well stop processing the windowagg once the row number reaches 11. Here
2200  * we check if 'opexpr' might help us to stop doing needless extra processing
2201  * in WindowAgg nodes.
2202  *
2203  * '*keep_original' is set to true if the caller should also use 'opexpr' for
2204  * its original purpose. This is set to false if the caller can assume that
2205  * the run condition will handle all of the required filtering.
2206  *
2207  * Returns true if 'opexpr' was found to be useful and was added to the
2208  * WindowClauses runCondition. We also set *keep_original accordingly and add
2209  * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2210  * If the 'opexpr' cannot be used then we set *keep_original to true and
2211  * return false.
2212  */
2213 static bool
2215  AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2216  bool wfunc_left, bool *keep_original,
2217  Bitmapset **run_cond_attrs)
2218 {
2219  Oid prosupport;
2220  Expr *otherexpr;
2223  WindowClause *wclause;
2224  List *opinfos;
2225  OpExpr *runopexpr;
2226  Oid runoperator;
2227  ListCell *lc;
2228 
2229  *keep_original = true;
2230 
2231  while (IsA(wfunc, RelabelType))
2232  wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2233 
2234  /* we can only work with window functions */
2235  if (!IsA(wfunc, WindowFunc))
2236  return false;
2237 
2238  /* can't use it if there are subplans in the WindowFunc */
2239  if (contain_subplans((Node *) wfunc))
2240  return false;
2241 
2242  prosupport = get_func_support(wfunc->winfnoid);
2243 
2244  /* Check if there's a support function for 'wfunc' */
2245  if (!OidIsValid(prosupport))
2246  return false;
2247 
2248  /* get the Expr from the other side of the OpExpr */
2249  if (wfunc_left)
2250  otherexpr = lsecond(opexpr->args);
2251  else
2252  otherexpr = linitial(opexpr->args);
2253 
2254  /*
2255  * The value being compared must not change during the evaluation of the
2256  * window partition.
2257  */
2258  if (!is_pseudo_constant_clause((Node *) otherexpr))
2259  return false;
2260 
2261  /* find the window clause belonging to the window function */
2262  wclause = (WindowClause *) list_nth(subquery->windowClause,
2263  wfunc->winref - 1);
2264 
2265  req.type = T_SupportRequestWFuncMonotonic;
2266  req.window_func = wfunc;
2267  req.window_clause = wclause;
2268 
2269  /* call the support function */
2271  DatumGetPointer(OidFunctionCall1(prosupport,
2272  PointerGetDatum(&req)));
2273 
2274  /*
2275  * Nothing to do if the function is neither monotonically increasing nor
2276  * monotonically decreasing.
2277  */
2278  if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2279  return false;
2280 
2281  runopexpr = NULL;
2282  runoperator = InvalidOid;
2283  opinfos = get_op_btree_interpretation(opexpr->opno);
2284 
2285  foreach(lc, opinfos)
2286  {
2288  int strategy = opinfo->strategy;
2289 
2290  /* handle < / <= */
2291  if (strategy == BTLessStrategyNumber ||
2292  strategy == BTLessEqualStrategyNumber)
2293  {
2294  /*
2295  * < / <= is supported for monotonically increasing functions in
2296  * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2297  * for monotonically decreasing functions.
2298  */
2299  if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2300  (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2301  {
2302  *keep_original = false;
2303  runopexpr = opexpr;
2304  runoperator = opexpr->opno;
2305  }
2306  break;
2307  }
2308  /* handle > / >= */
2309  else if (strategy == BTGreaterStrategyNumber ||
2310  strategy == BTGreaterEqualStrategyNumber)
2311  {
2312  /*
2313  * > / >= is supported for monotonically decreasing functions in
2314  * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2315  * for monotonically increasing functions.
2316  */
2317  if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2318  (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2319  {
2320  *keep_original = false;
2321  runopexpr = opexpr;
2322  runoperator = opexpr->opno;
2323  }
2324  break;
2325  }
2326  /* handle = */
2327  else if (strategy == BTEqualStrategyNumber)
2328  {
2329  int16 newstrategy;
2330 
2331  /*
2332  * When both monotonically increasing and decreasing then the
2333  * return value of the window function will be the same each time.
2334  * We can simply use 'opexpr' as the run condition without
2335  * modifying it.
2336  */
2337  if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2338  {
2339  *keep_original = false;
2340  runopexpr = opexpr;
2341  runoperator = opexpr->opno;
2342  break;
2343  }
2344 
2345  /*
2346  * When monotonically increasing we make a qual with <wfunc> <=
2347  * <value> or <value> >= <wfunc> in order to filter out values
2348  * which are above the value in the equality condition. For
2349  * monotonically decreasing functions we want to filter values
2350  * below the value in the equality condition.
2351  */
2352  if (res->monotonic & MONOTONICFUNC_INCREASING)
2353  newstrategy = wfunc_left ? BTLessEqualStrategyNumber : BTGreaterEqualStrategyNumber;
2354  else
2355  newstrategy = wfunc_left ? BTGreaterEqualStrategyNumber : BTLessEqualStrategyNumber;
2356 
2357  /* We must keep the original equality qual */
2358  *keep_original = true;
2359  runopexpr = opexpr;
2360 
2361  /* determine the operator to use for the runCondition qual */
2362  runoperator = get_opfamily_member(opinfo->opfamily_id,
2363  opinfo->oplefttype,
2364  opinfo->oprighttype,
2365  newstrategy);
2366  break;
2367  }
2368  }
2369 
2370  if (runopexpr != NULL)
2371  {
2372  Expr *newexpr;
2373 
2374  /*
2375  * Build the qual required for the run condition keeping the
2376  * WindowFunc on the same side as it was originally.
2377  */
2378  if (wfunc_left)
2379  newexpr = make_opclause(runoperator,
2380  runopexpr->opresulttype,
2381  runopexpr->opretset, (Expr *) wfunc,
2382  otherexpr, runopexpr->opcollid,
2383  runopexpr->inputcollid);
2384  else
2385  newexpr = make_opclause(runoperator,
2386  runopexpr->opresulttype,
2387  runopexpr->opretset,
2388  otherexpr, (Expr *) wfunc,
2389  runopexpr->opcollid,
2390  runopexpr->inputcollid);
2391 
2392  wclause->runCondition = lappend(wclause->runCondition, newexpr);
2393 
2394  /* record that this attno was used in a run condition */
2395  *run_cond_attrs = bms_add_member(*run_cond_attrs,
2397  return true;
2398  }
2399 
2400  /* unsupported OpExpr */
2401  return false;
2402 }
2403 
2404 /*
2405  * check_and_push_window_quals
2406  * Check if 'clause' is a qual that can be pushed into a WindowFunc's
2407  * WindowClause as a 'runCondition' qual. These, when present, allow
2408  * some unnecessary work to be skipped during execution.
2409  *
2410  * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2411  * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2412  * window quals for.
2413  *
2414  * Returns true if the caller still must keep the original qual or false if
2415  * the caller can safely ignore the original qual because the WindowAgg node
2416  * will use the runCondition to stop returning tuples.
2417  */
2418 static bool
2420  Node *clause, Bitmapset **run_cond_attrs)
2421 {
2422  OpExpr *opexpr = (OpExpr *) clause;
2423  bool keep_original = true;
2424  Var *var1;
2425  Var *var2;
2426 
2427  /* We're only able to use OpExprs with 2 operands */
2428  if (!IsA(opexpr, OpExpr))
2429  return true;
2430 
2431  if (list_length(opexpr->args) != 2)
2432  return true;
2433 
2434  /*
2435  * Currently, we restrict this optimization to strict OpExprs. The reason
2436  * for this is that during execution, once the runcondition becomes false,
2437  * we stop evaluating WindowFuncs. To avoid leaving around stale window
2438  * function result values, we set them to NULL. Having only strict
2439  * OpExprs here ensures that we properly filter out the tuples with NULLs
2440  * in the top-level WindowAgg.
2441  */
2442  set_opfuncid(opexpr);
2443  if (!func_strict(opexpr->opfuncid))
2444  return true;
2445 
2446  /*
2447  * Check for plain Vars that reference window functions in the subquery.
2448  * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2449  * be used as part of the run condition.
2450  */
2451 
2452  /* Check the left side of the OpExpr */
2453  var1 = linitial(opexpr->args);
2454  if (IsA(var1, Var) && var1->varattno > 0)
2455  {
2456  TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2457  WindowFunc *wfunc = (WindowFunc *) tle->expr;
2458 
2459  if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2460  opexpr, true, &keep_original,
2461  run_cond_attrs))
2462  return keep_original;
2463  }
2464 
2465  /* and check the right side */
2466  var2 = lsecond(opexpr->args);
2467  if (IsA(var2, Var) && var2->varattno > 0)
2468  {
2469  TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2470  WindowFunc *wfunc = (WindowFunc *) tle->expr;
2471 
2472  if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2473  opexpr, false, &keep_original,
2474  run_cond_attrs))
2475  return keep_original;
2476  }
2477 
2478  return true;
2479 }
2480 
2481 /*
2482  * set_subquery_pathlist
2483  * Generate SubqueryScan access paths for a subquery RTE
2484  *
2485  * We don't currently support generating parameterized paths for subqueries
2486  * by pushing join clauses down into them; it seems too expensive to re-plan
2487  * the subquery multiple times to consider different alternatives.
2488  * (XXX that could stand to be reconsidered, now that we use Paths.)
2489  * So the paths made here will be parameterized if the subquery contains
2490  * LATERAL references, otherwise not. As long as that's true, there's no need
2491  * for a separate set_subquery_size phase: just make the paths right away.
2492  */
2493 static void
2495  Index rti, RangeTblEntry *rte)
2496 {
2497  Query *parse = root->parse;
2498  Query *subquery = rte->subquery;
2499  bool trivial_pathtarget;
2500  Relids required_outer;
2501  pushdown_safety_info safetyInfo;
2502  double tuple_fraction;
2503  RelOptInfo *sub_final_rel;
2504  Bitmapset *run_cond_attrs = NULL;
2505  ListCell *lc;
2506 
2507  /*
2508  * Must copy the Query so that planning doesn't mess up the RTE contents
2509  * (really really need to fix the planner to not scribble on its input,
2510  * someday ... but see remove_unused_subquery_outputs to start with).
2511  */
2512  subquery = copyObject(subquery);
2513 
2514  /*
2515  * If it's a LATERAL subquery, it might contain some Vars of the current
2516  * query level, requiring it to be treated as parameterized, even though
2517  * we don't support pushing down join quals into subqueries.
2518  */
2519  required_outer = rel->lateral_relids;
2520 
2521  /*
2522  * Zero out result area for subquery_is_pushdown_safe, so that it can set
2523  * flags as needed while recursing. In particular, we need a workspace
2524  * for keeping track of the reasons why columns are unsafe to reference.
2525  * These reasons are stored in the bits inside unsafeFlags[i] when we
2526  * discover reasons that column i of the subquery is unsafe to be used in
2527  * a pushed-down qual.
2528  */
2529  memset(&safetyInfo, 0, sizeof(safetyInfo));
2530  safetyInfo.unsafeFlags = (unsigned char *)
2531  palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2532 
2533  /*
2534  * If the subquery has the "security_barrier" flag, it means the subquery
2535  * originated from a view that must enforce row-level security. Then we
2536  * must not push down quals that contain leaky functions. (Ideally this
2537  * would be checked inside subquery_is_pushdown_safe, but since we don't
2538  * currently pass the RTE to that function, we must do it here.)
2539  */
2540  safetyInfo.unsafeLeaky = rte->security_barrier;
2541 
2542  /*
2543  * If there are any restriction clauses that have been attached to the
2544  * subquery relation, consider pushing them down to become WHERE or HAVING
2545  * quals of the subquery itself. This transformation is useful because it
2546  * may allow us to generate a better plan for the subquery than evaluating
2547  * all the subquery output rows and then filtering them.
2548  *
2549  * There are several cases where we cannot push down clauses. Restrictions
2550  * involving the subquery are checked by subquery_is_pushdown_safe().
2551  * Restrictions on individual clauses are checked by
2552  * qual_is_pushdown_safe(). Also, we don't want to push down
2553  * pseudoconstant clauses; better to have the gating node above the
2554  * subquery.
2555  *
2556  * Non-pushed-down clauses will get evaluated as qpquals of the
2557  * SubqueryScan node.
2558  *
2559  * XXX Are there any cases where we want to make a policy decision not to
2560  * push down a pushable qual, because it'd result in a worse plan?
2561  */
2562  if (rel->baserestrictinfo != NIL &&
2563  subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2564  {
2565  /* OK to consider pushing down individual quals */
2566  List *upperrestrictlist = NIL;
2567  ListCell *l;
2568 
2569  foreach(l, rel->baserestrictinfo)
2570  {
2571  RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2572  Node *clause = (Node *) rinfo->clause;
2573 
2574  if (rinfo->pseudoconstant)
2575  {
2576  upperrestrictlist = lappend(upperrestrictlist, rinfo);
2577  continue;
2578  }
2579 
2580  switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2581  {
2582  case PUSHDOWN_SAFE:
2583  /* Push it down */
2584  subquery_push_qual(subquery, rte, rti, clause);
2585  break;
2586 
2588 
2589  /*
2590  * Since we can't push the qual down into the subquery,
2591  * check if it happens to reference a window function. If
2592  * so then it might be useful to use for the WindowAgg's
2593  * runCondition.
2594  */
2595  if (!subquery->hasWindowFuncs ||
2596  check_and_push_window_quals(subquery, rte, rti, clause,
2597  &run_cond_attrs))
2598  {
2599  /*
2600  * subquery has no window funcs or the clause is not a
2601  * suitable window run condition qual or it is, but
2602  * the original must also be kept in the upper query.
2603  */
2604  upperrestrictlist = lappend(upperrestrictlist, rinfo);
2605  }
2606  break;
2607 
2608  case PUSHDOWN_UNSAFE:
2609  upperrestrictlist = lappend(upperrestrictlist, rinfo);
2610  break;
2611  }
2612  }
2613  rel->baserestrictinfo = upperrestrictlist;
2614  /* We don't bother recomputing baserestrict_min_security */
2615  }
2616 
2617  pfree(safetyInfo.unsafeFlags);
2618 
2619  /*
2620  * The upper query might not use all the subquery's output columns; if
2621  * not, we can simplify. Pass the attributes that were pushed down into
2622  * WindowAgg run conditions to ensure we don't accidentally think those
2623  * are unused.
2624  */
2625  remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2626 
2627  /*
2628  * We can safely pass the outer tuple_fraction down to the subquery if the
2629  * outer level has no joining, aggregation, or sorting to do. Otherwise
2630  * we'd better tell the subquery to plan for full retrieval. (XXX This
2631  * could probably be made more intelligent ...)
2632  */
2633  if (parse->hasAggs ||
2634  parse->groupClause ||
2635  parse->groupingSets ||
2636  root->hasHavingQual ||
2637  parse->distinctClause ||
2638  parse->sortClause ||
2640  tuple_fraction = 0.0; /* default case */
2641  else
2642  tuple_fraction = root->tuple_fraction;
2643 
2644  /* plan_params should not be in use in current query level */
2645  Assert(root->plan_params == NIL);
2646 
2647  /* Generate a subroot and Paths for the subquery */
2648  rel->subroot = subquery_planner(root->glob, subquery,
2649  root,
2650  false, tuple_fraction);
2651 
2652  /* Isolate the params needed by this specific subplan */
2653  rel->subplan_params = root->plan_params;
2654  root->plan_params = NIL;
2655 
2656  /*
2657  * It's possible that constraint exclusion proved the subquery empty. If
2658  * so, it's desirable to produce an unadorned dummy path so that we will
2659  * recognize appropriate optimizations at this query level.
2660  */
2661  sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2662 
2663  if (IS_DUMMY_REL(sub_final_rel))
2664  {
2666  return;
2667  }
2668 
2669  /*
2670  * Mark rel with estimated output rows, width, etc. Note that we have to
2671  * do this before generating outer-query paths, else cost_subqueryscan is
2672  * not happy.
2673  */
2674  set_subquery_size_estimates(root, rel);
2675 
2676  /*
2677  * Also detect whether the reltarget is trivial, so that we can pass that
2678  * info to cost_subqueryscan (rather than re-deriving it multiple times).
2679  * It's trivial if it fetches all the subplan output columns in order.
2680  */
2681  if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2682  trivial_pathtarget = false;
2683  else
2684  {
2685  trivial_pathtarget = true;
2686  foreach(lc, rel->reltarget->exprs)
2687  {
2688  Node *node = (Node *) lfirst(lc);
2689  Var *var;
2690 
2691  if (!IsA(node, Var))
2692  {
2693  trivial_pathtarget = false;
2694  break;
2695  }
2696  var = (Var *) node;
2697  if (var->varno != rti ||
2698  var->varattno != foreach_current_index(lc) + 1)
2699  {
2700  trivial_pathtarget = false;
2701  break;
2702  }
2703  }
2704  }
2705 
2706  /*
2707  * For each Path that subquery_planner produced, make a SubqueryScanPath
2708  * in the outer query.
2709  */
2710  foreach(lc, sub_final_rel->pathlist)
2711  {
2712  Path *subpath = (Path *) lfirst(lc);
2713  List *pathkeys;
2714 
2715  /* Convert subpath's pathkeys to outer representation */
2716  pathkeys = convert_subquery_pathkeys(root,
2717  rel,
2718  subpath->pathkeys,
2719  make_tlist_from_pathtarget(subpath->pathtarget));
2720 
2721  /* Generate outer path using this subpath */
2722  add_path(rel, (Path *)
2723  create_subqueryscan_path(root, rel, subpath,
2724  trivial_pathtarget,
2725  pathkeys, required_outer));
2726  }
2727 
2728  /* If outer rel allows parallelism, do same for partial paths. */
2729  if (rel->consider_parallel && bms_is_empty(required_outer))
2730  {
2731  /* If consider_parallel is false, there should be no partial paths. */
2732  Assert(sub_final_rel->consider_parallel ||
2733  sub_final_rel->partial_pathlist == NIL);
2734 
2735  /* Same for partial paths. */
2736  foreach(lc, sub_final_rel->partial_pathlist)
2737  {
2738  Path *subpath = (Path *) lfirst(lc);
2739  List *pathkeys;
2740 
2741  /* Convert subpath's pathkeys to outer representation */
2742  pathkeys = convert_subquery_pathkeys(root,
2743  rel,
2744  subpath->pathkeys,
2745  make_tlist_from_pathtarget(subpath->pathtarget));
2746 
2747  /* Generate outer path using this subpath */
2748  add_partial_path(rel, (Path *)
2749  create_subqueryscan_path(root, rel, subpath,
2750  trivial_pathtarget,
2751  pathkeys,
2752  required_outer));
2753  }
2754  }
2755 }
2756 
2757 /*
2758  * set_function_pathlist
2759  * Build the (single) access path for a function RTE
2760  */
2761 static void
2763 {
2764  Relids required_outer;
2765  List *pathkeys = NIL;
2766 
2767  /*
2768  * We don't support pushing join clauses into the quals of a function
2769  * scan, but it could still have required parameterization due to LATERAL
2770  * refs in the function expression.
2771  */
2772  required_outer = rel->lateral_relids;
2773 
2774  /*
2775  * The result is considered unordered unless ORDINALITY was used, in which
2776  * case it is ordered by the ordinal column (the last one). See if we
2777  * care, by checking for uses of that Var in equivalence classes.
2778  */
2779  if (rte->funcordinality)
2780  {
2781  AttrNumber ordattno = rel->max_attr;
2782  Var *var = NULL;
2783  ListCell *lc;
2784 
2785  /*
2786  * Is there a Var for it in rel's targetlist? If not, the query did
2787  * not reference the ordinality column, or at least not in any way
2788  * that would be interesting for sorting.
2789  */
2790  foreach(lc, rel->reltarget->exprs)
2791  {
2792  Var *node = (Var *) lfirst(lc);
2793 
2794  /* checking varno/varlevelsup is just paranoia */
2795  if (IsA(node, Var) &&
2796  node->varattno == ordattno &&
2797  node->varno == rel->relid &&
2798  node->varlevelsup == 0)
2799  {
2800  var = node;
2801  break;
2802  }
2803  }
2804 
2805  /*
2806  * Try to build pathkeys for this Var with int8 sorting. We tell
2807  * build_expression_pathkey not to build any new equivalence class; if
2808  * the Var isn't already mentioned in some EC, it means that nothing
2809  * cares about the ordering.
2810  */
2811  if (var)
2812  pathkeys = build_expression_pathkey(root,
2813  (Expr *) var,
2814  Int8LessOperator,
2815  rel->relids,
2816  false);
2817  }
2818 
2819  /* Generate appropriate path */
2820  add_path(rel, create_functionscan_path(root, rel,
2821  pathkeys, required_outer));
2822 }
2823 
2824 /*
2825  * set_values_pathlist
2826  * Build the (single) access path for a VALUES RTE
2827  */
2828 static void
2830 {
2831  Relids required_outer;
2832 
2833  /*
2834  * We don't support pushing join clauses into the quals of a values scan,
2835  * but it could still have required parameterization due to LATERAL refs
2836  * in the values expressions.
2837  */
2838  required_outer = rel->lateral_relids;
2839 
2840  /* Generate appropriate path */
2841  add_path(rel, create_valuesscan_path(root, rel, required_outer));
2842 }
2843 
2844 /*
2845  * set_tablefunc_pathlist
2846  * Build the (single) access path for a table func RTE
2847  */
2848 static void
2850 {
2851  Relids required_outer;
2852 
2853  /*
2854  * We don't support pushing join clauses into the quals of a tablefunc
2855  * scan, but it could still have required parameterization due to LATERAL
2856  * refs in the function expression.
2857  */
2858  required_outer = rel->lateral_relids;
2859 
2860  /* Generate appropriate path */
2861  add_path(rel, create_tablefuncscan_path(root, rel,
2862  required_outer));
2863 }
2864 
2865 /*
2866  * set_cte_pathlist
2867  * Build the (single) access path for a non-self-reference CTE RTE
2868  *
2869  * There's no need for a separate set_cte_size phase, since we don't
2870  * support join-qual-parameterized paths for CTEs.
2871  */
2872 static void
2874 {
2875  Plan *cteplan;
2876  PlannerInfo *cteroot;
2877  Index levelsup;
2878  int ndx;
2879  ListCell *lc;
2880  int plan_id;
2881  Relids required_outer;
2882 
2883  /*
2884  * Find the referenced CTE, and locate the plan previously made for it.
2885  */
2886  levelsup = rte->ctelevelsup;
2887  cteroot = root;
2888  while (levelsup-- > 0)
2889  {
2890  cteroot = cteroot->parent_root;
2891  if (!cteroot) /* shouldn't happen */
2892  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2893  }
2894 
2895  /*
2896  * Note: cte_plan_ids can be shorter than cteList, if we are still working
2897  * on planning the CTEs (ie, this is a side-reference from another CTE).
2898  * So we mustn't use forboth here.
2899  */
2900  ndx = 0;
2901  foreach(lc, cteroot->parse->cteList)
2902  {
2903  CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
2904 
2905  if (strcmp(cte->ctename, rte->ctename) == 0)
2906  break;
2907  ndx++;
2908  }
2909  if (lc == NULL) /* shouldn't happen */
2910  elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
2911  if (ndx >= list_length(cteroot->cte_plan_ids))
2912  elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
2913  plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
2914  if (plan_id <= 0)
2915  elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
2916  cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2917 
2918  /* Mark rel with estimated output rows, width, etc */
2919  set_cte_size_estimates(root, rel, cteplan->plan_rows);
2920 
2921  /*
2922  * We don't support pushing join clauses into the quals of a CTE scan, but
2923  * it could still have required parameterization due to LATERAL refs in
2924  * its tlist.
2925  */
2926  required_outer = rel->lateral_relids;
2927 
2928  /* Generate appropriate path */
2929  add_path(rel, create_ctescan_path(root, rel, required_outer));
2930 }
2931 
2932 /*
2933  * set_namedtuplestore_pathlist
2934  * Build the (single) access path for a named tuplestore RTE
2935  *
2936  * There's no need for a separate set_namedtuplestore_size phase, since we
2937  * don't support join-qual-parameterized paths for tuplestores.
2938  */
2939 static void
2941  RangeTblEntry *rte)
2942 {
2943  Relids required_outer;
2944 
2945  /* Mark rel with estimated output rows, width, etc */
2947 
2948  /*
2949  * We don't support pushing join clauses into the quals of a tuplestore
2950  * scan, but it could still have required parameterization due to LATERAL
2951  * refs in its tlist.
2952  */
2953  required_outer = rel->lateral_relids;
2954 
2955  /* Generate appropriate path */
2956  add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2957 
2958  /* Select cheapest path (pretty easy in this case...) */
2959  set_cheapest(rel);
2960 }
2961 
2962 /*
2963  * set_result_pathlist
2964  * Build the (single) access path for an RTE_RESULT RTE
2965  *
2966  * There's no need for a separate set_result_size phase, since we
2967  * don't support join-qual-parameterized paths for these RTEs.
2968  */
2969 static void
2971  RangeTblEntry *rte)
2972 {
2973  Relids required_outer;
2974 
2975  /* Mark rel with estimated output rows, width, etc */
2976  set_result_size_estimates(root, rel);
2977 
2978  /*
2979  * We don't support pushing join clauses into the quals of a Result scan,
2980  * but it could still have required parameterization due to LATERAL refs
2981  * in its tlist.
2982  */
2983  required_outer = rel->lateral_relids;
2984 
2985  /* Generate appropriate path */
2986  add_path(rel, create_resultscan_path(root, rel, required_outer));
2987 
2988  /* Select cheapest path (pretty easy in this case...) */
2989  set_cheapest(rel);
2990 }
2991 
2992 /*
2993  * set_worktable_pathlist
2994  * Build the (single) access path for a self-reference CTE RTE
2995  *
2996  * There's no need for a separate set_worktable_size phase, since we don't
2997  * support join-qual-parameterized paths for CTEs.
2998  */
2999 static void
3001 {
3002  Path *ctepath;
3003  PlannerInfo *cteroot;
3004  Index levelsup;
3005  Relids required_outer;
3006 
3007  /*
3008  * We need to find the non-recursive term's path, which is in the plan
3009  * level that's processing the recursive UNION, which is one level *below*
3010  * where the CTE comes from.
3011  */
3012  levelsup = rte->ctelevelsup;
3013  if (levelsup == 0) /* shouldn't happen */
3014  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3015  levelsup--;
3016  cteroot = root;
3017  while (levelsup-- > 0)
3018  {
3019  cteroot = cteroot->parent_root;
3020  if (!cteroot) /* shouldn't happen */
3021  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3022  }
3023  ctepath = cteroot->non_recursive_path;
3024  if (!ctepath) /* shouldn't happen */
3025  elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3026 
3027  /* Mark rel with estimated output rows, width, etc */
3028  set_cte_size_estimates(root, rel, ctepath->rows);
3029 
3030  /*
3031  * We don't support pushing join clauses into the quals of a worktable
3032  * scan, but it could still have required parameterization due to LATERAL
3033  * refs in its tlist. (I'm not sure this is actually possible given the
3034  * restrictions on recursive references, but it's easy enough to support.)
3035  */
3036  required_outer = rel->lateral_relids;
3037 
3038  /* Generate appropriate path */
3039  add_path(rel, create_worktablescan_path(root, rel, required_outer));
3040 }
3041 
3042 /*
3043  * generate_gather_paths
3044  * Generate parallel access paths for a relation by pushing a Gather or
3045  * Gather Merge on top of a partial path.
3046  *
3047  * This must not be called until after we're done creating all partial paths
3048  * for the specified relation. (Otherwise, add_partial_path might delete a
3049  * path that some GatherPath or GatherMergePath has a reference to.)
3050  *
3051  * If we're generating paths for a scan or join relation, override_rows will
3052  * be false, and we'll just use the relation's size estimate. When we're
3053  * being called for a partially-grouped or partially-distinct path, though, we
3054  * need to override the rowcount estimate. (It's not clear that the
3055  * particular value we're using here is actually best, but the underlying rel
3056  * has no estimate so we must do something.)
3057  */
3058 void
3059 generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3060 {
3061  Path *cheapest_partial_path;
3062  Path *simple_gather_path;
3063  ListCell *lc;
3064  double rows;
3065  double *rowsp = NULL;
3066 
3067  /* If there are no partial paths, there's nothing to do here. */
3068  if (rel->partial_pathlist == NIL)
3069  return;
3070 
3071  /* Should we override the rel's rowcount estimate? */
3072  if (override_rows)
3073  rowsp = &rows;
3074 
3075  /*
3076  * The output of Gather is always unsorted, so there's only one partial
3077  * path of interest: the cheapest one. That will be the one at the front
3078  * of partial_pathlist because of the way add_partial_path works.
3079  */
3080  cheapest_partial_path = linitial(rel->partial_pathlist);
3081  rows =
3082  cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
3083  simple_gather_path = (Path *)
3084  create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3085  NULL, rowsp);
3086  add_path(rel, simple_gather_path);
3087 
3088  /*
3089  * For each useful ordering, we can consider an order-preserving Gather
3090  * Merge.
3091  */
3092  foreach(lc, rel->partial_pathlist)
3093  {
3094  Path *subpath = (Path *) lfirst(lc);
3095  GatherMergePath *path;
3096 
3097  if (subpath->pathkeys == NIL)
3098  continue;
3099 
3100  rows = subpath->rows * subpath->parallel_workers;
3101  path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3102  subpath->pathkeys, NULL, rowsp);
3103  add_path(rel, &path->path);
3104  }
3105 }
3106 
3107 /*
3108  * get_useful_pathkeys_for_relation
3109  * Determine which orderings of a relation might be useful.
3110  *
3111  * Getting data in sorted order can be useful either because the requested
3112  * order matches the final output ordering for the overall query we're
3113  * planning, or because it enables an efficient merge join. Here, we try
3114  * to figure out which pathkeys to consider.
3115  *
3116  * This allows us to do incremental sort on top of an index scan under a gather
3117  * merge node, i.e. parallelized.
3118  *
3119  * If the require_parallel_safe is true, we also require the expressions to
3120  * be parallel safe (which allows pushing the sort below Gather Merge).
3121  *
3122  * XXX At the moment this can only ever return a list with a single element,
3123  * because it looks at query_pathkeys only. So we might return the pathkeys
3124  * directly, but it seems plausible we'll want to consider other orderings
3125  * in the future. For example, we might want to consider pathkeys useful for
3126  * merge joins.
3127  */
3128 static List *
3130  bool require_parallel_safe)
3131 {
3132  List *useful_pathkeys_list = NIL;
3133 
3134  /*
3135  * Considering query_pathkeys is always worth it, because it might allow
3136  * us to avoid a total sort when we have a partially presorted path
3137  * available or to push the total sort into the parallel portion of the
3138  * query.
3139  */
3140  if (root->query_pathkeys)
3141  {
3142  ListCell *lc;
3143  int npathkeys = 0; /* useful pathkeys */
3144 
3145  foreach(lc, root->query_pathkeys)
3146  {
3147  PathKey *pathkey = (PathKey *) lfirst(lc);
3148  EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3149 
3150  /*
3151  * We can only build a sort for pathkeys that contain a
3152  * safe-to-compute-early EC member computable from the current
3153  * relation's reltarget, so ignore the remainder of the list as
3154  * soon as we find a pathkey without such a member.
3155  *
3156  * It's still worthwhile to return any prefix of the pathkeys list
3157  * that meets this requirement, as we may be able to do an
3158  * incremental sort.
3159  *
3160  * If requested, ensure the sort expression is parallel-safe too.
3161  */
3162  if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3163  require_parallel_safe))
3164  break;
3165 
3166  npathkeys++;
3167  }
3168 
3169  /*
3170  * The whole query_pathkeys list matches, so append it directly, to
3171  * allow comparing pathkeys easily by comparing list pointer. If we
3172  * have to truncate the pathkeys, we gotta do a copy though.
3173  */
3174  if (npathkeys == list_length(root->query_pathkeys))
3175  useful_pathkeys_list = lappend(useful_pathkeys_list,
3176  root->query_pathkeys);
3177  else if (npathkeys > 0)
3178  useful_pathkeys_list = lappend(useful_pathkeys_list,
3180  npathkeys));
3181  }
3182 
3183  return useful_pathkeys_list;
3184 }
3185 
3186 /*
3187  * generate_useful_gather_paths
3188  * Generate parallel access paths for a relation by pushing a Gather or
3189  * Gather Merge on top of a partial path.
3190  *
3191  * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3192  * paths (aiming to preserve the ordering), but also considers ordering that
3193  * might be useful for nodes above the gather merge node, and tries to add
3194  * a sort (regular or incremental) to provide that.
3195  */
3196 void
3197 generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3198 {
3199  ListCell *lc;
3200  double rows;
3201  double *rowsp = NULL;
3202  List *useful_pathkeys_list = NIL;
3203  Path *cheapest_partial_path = NULL;
3204 
3205  /* If there are no partial paths, there's nothing to do here. */
3206  if (rel->partial_pathlist == NIL)
3207  return;
3208 
3209  /* Should we override the rel's rowcount estimate? */
3210  if (override_rows)
3211  rowsp = &rows;
3212 
3213  /* generate the regular gather (merge) paths */
3214  generate_gather_paths(root, rel, override_rows);
3215 
3216  /* consider incremental sort for interesting orderings */
3217  useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3218 
3219  /* used for explicit (full) sort paths */
3220  cheapest_partial_path = linitial(rel->partial_pathlist);
3221 
3222  /*
3223  * Consider sorted paths for each interesting ordering. We generate both
3224  * incremental and full sort.
3225  */
3226  foreach(lc, useful_pathkeys_list)
3227  {
3228  List *useful_pathkeys = lfirst(lc);
3229  ListCell *lc2;
3230  bool is_sorted;
3231  int presorted_keys;
3232 
3233  foreach(lc2, rel->partial_pathlist)
3234  {
3235  Path *subpath = (Path *) lfirst(lc2);
3236  GatherMergePath *path;
3237 
3238  is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3239  subpath->pathkeys,
3240  &presorted_keys);
3241 
3242  /*
3243  * We don't need to consider the case where a subpath is already
3244  * fully sorted because generate_gather_paths already creates a
3245  * gather merge path for every subpath that has pathkeys present.
3246  *
3247  * But since the subpath is already sorted, we know we don't need
3248  * to consider adding a sort (full or incremental) on top of it,
3249  * so we can continue here.
3250  */
3251  if (is_sorted)
3252  continue;
3253 
3254  /*
3255  * Try at least sorting the cheapest path and also try
3256  * incrementally sorting any path which is partially sorted
3257  * already (no need to deal with paths which have presorted keys
3258  * when incremental sort is disabled unless it's the cheapest
3259  * input path).
3260  */
3261  if (subpath != cheapest_partial_path &&
3262  (presorted_keys == 0 || !enable_incremental_sort))
3263  continue;
3264 
3265  /*
3266  * Consider regular sort for any path that's not presorted or if
3267  * incremental sort is disabled. We've no need to consider both
3268  * sort and incremental sort on the same path. We assume that
3269  * incremental sort is always faster when there are presorted
3270  * keys.
3271  *
3272  * This is not redundant with the gather paths created in
3273  * generate_gather_paths, because that doesn't generate ordered
3274  * output. Here we add an explicit sort to match the useful
3275  * ordering.
3276  */
3277  if (presorted_keys == 0 || !enable_incremental_sort)
3278  {
3279  subpath = (Path *) create_sort_path(root,
3280  rel,
3281  subpath,
3282  useful_pathkeys,
3283  -1.0);
3284  rows = subpath->rows * subpath->parallel_workers;
3285  }
3286  else
3288  rel,
3289  subpath,
3290  useful_pathkeys,
3291  presorted_keys,
3292  -1);
3293  path = create_gather_merge_path(root, rel,
3294  subpath,
3295  rel->reltarget,
3296  subpath->pathkeys,
3297  NULL,
3298  rowsp);
3299 
3300  add_path(rel, &path->path);
3301  }
3302  }
3303 }
3304 
3305 /*
3306  * make_rel_from_joinlist
3307  * Build access paths using a "joinlist" to guide the join path search.
3308  *
3309  * See comments for deconstruct_jointree() for definition of the joinlist
3310  * data structure.
3311  */
3312 static RelOptInfo *
3314 {
3315  int levels_needed;
3316  List *initial_rels;
3317  ListCell *jl;
3318 
3319  /*
3320  * Count the number of child joinlist nodes. This is the depth of the
3321  * dynamic-programming algorithm we must employ to consider all ways of
3322  * joining the child nodes.
3323  */
3324  levels_needed = list_length(joinlist);
3325 
3326  if (levels_needed <= 0)
3327  return NULL; /* nothing to do? */
3328 
3329  /*
3330  * Construct a list of rels corresponding to the child joinlist nodes.
3331  * This may contain both base rels and rels constructed according to
3332  * sub-joinlists.
3333  */
3334  initial_rels = NIL;
3335  foreach(jl, joinlist)
3336  {
3337  Node *jlnode = (Node *) lfirst(jl);
3338  RelOptInfo *thisrel;
3339 
3340  if (IsA(jlnode, RangeTblRef))
3341  {
3342  int varno = ((RangeTblRef *) jlnode)->rtindex;
3343 
3344  thisrel = find_base_rel(root, varno);
3345  }
3346  else if (IsA(jlnode, List))
3347  {
3348  /* Recurse to handle subproblem */
3349  thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3350  }
3351  else
3352  {
3353  elog(ERROR, "unrecognized joinlist node type: %d",
3354  (int) nodeTag(jlnode));
3355  thisrel = NULL; /* keep compiler quiet */
3356  }
3357 
3358  initial_rels = lappend(initial_rels, thisrel);
3359  }
3360 
3361  if (levels_needed == 1)
3362  {
3363  /*
3364  * Single joinlist node, so we're done.
3365  */
3366  return (RelOptInfo *) linitial(initial_rels);
3367  }
3368  else
3369  {
3370  /*
3371  * Consider the different orders in which we could join the rels,
3372  * using a plugin, GEQO, or the regular join search code.
3373  *
3374  * We put the initial_rels list into a PlannerInfo field because
3375  * has_legal_joinclause() needs to look at it (ugly :-().
3376  */
3377  root->initial_rels = initial_rels;
3378 
3379  if (join_search_hook)
3380  return (*join_search_hook) (root, levels_needed, initial_rels);
3381  else if (enable_geqo && levels_needed >= geqo_threshold)
3382  return geqo(root, levels_needed, initial_rels);
3383  else
3384  return standard_join_search(root, levels_needed, initial_rels);
3385  }
3386 }
3387 
3388 /*
3389  * standard_join_search
3390  * Find possible joinpaths for a query by successively finding ways
3391  * to join component relations into join relations.
3392  *
3393  * 'levels_needed' is the number of iterations needed, ie, the number of
3394  * independent jointree items in the query. This is > 1.
3395  *
3396  * 'initial_rels' is a list of RelOptInfo nodes for each independent
3397  * jointree item. These are the components to be joined together.
3398  * Note that levels_needed == list_length(initial_rels).
3399  *
3400  * Returns the final level of join relations, i.e., the relation that is
3401  * the result of joining all the original relations together.
3402  * At least one implementation path must be provided for this relation and
3403  * all required sub-relations.
3404  *
3405  * To support loadable plugins that modify planner behavior by changing the
3406  * join searching algorithm, we provide a hook variable that lets a plugin
3407  * replace or supplement this function. Any such hook must return the same
3408  * final join relation as the standard code would, but it might have a
3409  * different set of implementation paths attached, and only the sub-joinrels
3410  * needed for these paths need have been instantiated.
3411  *
3412  * Note to plugin authors: the functions invoked during standard_join_search()
3413  * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3414  * than one join-order search, you'll probably need to save and restore the
3415  * original states of those data structures. See geqo_eval() for an example.
3416  */
3417 RelOptInfo *
3418 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3419 {
3420  int lev;
3421  RelOptInfo *rel;
3422 
3423  /*
3424  * This function cannot be invoked recursively within any one planning
3425  * problem, so join_rel_level[] can't be in use already.
3426  */
3427  Assert(root->join_rel_level == NULL);
3428 
3429  /*
3430  * We employ a simple "dynamic programming" algorithm: we first find all
3431  * ways to build joins of two jointree items, then all ways to build joins
3432  * of three items (from two-item joins and single items), then four-item
3433  * joins, and so on until we have considered all ways to join all the
3434  * items into one rel.
3435  *
3436  * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3437  * set root->join_rel_level[1] to represent all the single-jointree-item
3438  * relations.
3439  */
3440  root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3441 
3442  root->join_rel_level[1] = initial_rels;
3443 
3444  for (lev = 2; lev <= levels_needed; lev++)
3445  {
3446  ListCell *lc;
3447 
3448  /*
3449  * Determine all possible pairs of relations to be joined at this
3450  * level, and build paths for making each one from every available
3451  * pair of lower-level relations.
3452  */
3453  join_search_one_level(root, lev);
3454 
3455  /*
3456  * Run generate_partitionwise_join_paths() and
3457  * generate_useful_gather_paths() for each just-processed joinrel. We
3458  * could not do this earlier because both regular and partial paths
3459  * can get added to a particular joinrel at multiple times within
3460  * join_search_one_level.
3461  *
3462  * After that, we're done creating paths for the joinrel, so run
3463  * set_cheapest().
3464  */
3465  foreach(lc, root->join_rel_level[lev])
3466  {
3467  rel = (RelOptInfo *) lfirst(lc);
3468 
3469  /* Create paths for partitionwise joins. */
3471 
3472  /*
3473  * Except for the topmost scan/join rel, consider gathering
3474  * partial paths. We'll do the same for the topmost scan/join rel
3475  * once we know the final targetlist (see grouping_planner's and
3476  * its call to apply_scanjoin_target_to_paths).
3477  */
3478  if (!bms_equal(rel->relids, root->all_query_rels))
3479  generate_useful_gather_paths(root, rel, false);
3480 
3481  /* Find and save the cheapest paths for this rel */
3482  set_cheapest(rel);
3483 
3484 #ifdef OPTIMIZER_DEBUG
3485  pprint(rel);
3486 #endif
3487  }
3488  }
3489 
3490  /*
3491  * We should have a single rel at the final level.
3492  */
3493  if (root->join_rel_level[levels_needed] == NIL)
3494  elog(ERROR, "failed to build any %d-way joins", levels_needed);
3495  Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3496 
3497  rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3498 
3499  root->join_rel_level = NULL;
3500 
3501  return rel;
3502 }
3503 
3504 /*****************************************************************************
3505  * PUSHING QUALS DOWN INTO SUBQUERIES
3506  *****************************************************************************/
3507 
3508 /*
3509  * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3510  *
3511  * subquery is the particular component query being checked. topquery
3512  * is the top component of a set-operations tree (the same Query if no
3513  * set-op is involved).
3514  *
3515  * Conditions checked here:
3516  *
3517  * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3518  * since that could change the set of rows returned.
3519  *
3520  * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3521  * quals into it, because that could change the results.
3522  *
3523  * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3524  * This is because upper-level quals should semantically be evaluated only
3525  * once per distinct row, not once per original row, and if the qual is
3526  * volatile then extra evaluations could change the results. (This issue
3527  * does not apply to other forms of aggregation such as GROUP BY, because
3528  * when those are present we push into HAVING not WHERE, so that the quals
3529  * are still applied after aggregation.)
3530  *
3531  * 4. If the subquery contains window functions, we cannot push volatile quals
3532  * into it. The issue here is a bit different from DISTINCT: a volatile qual
3533  * might succeed for some rows of a window partition and fail for others,
3534  * thereby changing the partition contents and thus the window functions'
3535  * results for rows that remain.
3536  *
3537  * 5. If the subquery contains any set-returning functions in its targetlist,
3538  * we cannot push volatile quals into it. That would push them below the SRFs
3539  * and thereby change the number of times they are evaluated. Also, a
3540  * volatile qual could succeed for some SRF output rows and fail for others,
3541  * a behavior that cannot occur if it's evaluated before SRF expansion.
3542  *
3543  * 6. If the subquery has nonempty grouping sets, we cannot push down any
3544  * quals. The concern here is that a qual referencing a "constant" grouping
3545  * column could get constant-folded, which would be improper because the value
3546  * is potentially nullable by grouping-set expansion. This restriction could
3547  * be removed if we had a parsetree representation that shows that such
3548  * grouping columns are not really constant. (There are other ideas that
3549  * could be used to relax this restriction, but that's the approach most
3550  * likely to get taken in the future. Note that there's not much to be gained
3551  * so long as subquery_planner can't move HAVING clauses to WHERE within such
3552  * a subquery.)
3553  *
3554  * In addition, we make several checks on the subquery's output columns to see
3555  * if it is safe to reference them in pushed-down quals. If output column k
3556  * is found to be unsafe to reference, we set the reason for that inside
3557  * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3558  * column k might not be referenced by some/all quals. The unsafeFlags[]
3559  * array will be consulted later by qual_is_pushdown_safe(). It's better to
3560  * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3561  * because when the subquery involves set operations we have to check the
3562  * output expressions in each arm of the set op.
3563  *
3564  * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3565  * we're effectively assuming that the quals cannot distinguish values that
3566  * the DISTINCT's equality operator sees as equal, yet there are many
3567  * counterexamples to that assumption. However use of such a qual with a
3568  * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3569  * "equal" value will be chosen as the output value by the DISTINCT operation.
3570  * So we don't worry too much about that. Another objection is that if the
3571  * qual is expensive to evaluate, running it for each original row might cost
3572  * more than we save by eliminating rows before the DISTINCT step. But it
3573  * would be very hard to estimate that at this stage, and in practice pushdown
3574  * seldom seems to make things worse, so we ignore that problem too.
3575  *
3576  * Note: likewise, pushing quals into a subquery with window functions is a
3577  * bit dubious: the quals might remove some rows of a window partition while
3578  * leaving others, causing changes in the window functions' results for the
3579  * surviving rows. We insist that such a qual reference only partitioning
3580  * columns, but again that only protects us if the qual does not distinguish
3581  * values that the partitioning equality operator sees as equal. The risks
3582  * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3583  * occurs and thus there is no theoretical problem with such a qual. But
3584  * we'll do this anyway because the potential performance benefits are very
3585  * large, and we've seen no field complaints about the longstanding comparable
3586  * behavior with DISTINCT.
3587  */
3588 static bool
3590  pushdown_safety_info *safetyInfo)
3591 {
3592  SetOperationStmt *topop;
3593 
3594  /* Check point 1 */
3595  if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
3596  return false;
3597 
3598  /* Check point 6 */
3599  if (subquery->groupClause && subquery->groupingSets)
3600  return false;
3601 
3602  /* Check points 3, 4, and 5 */
3603  if (subquery->distinctClause ||
3604  subquery->hasWindowFuncs ||
3605  subquery->hasTargetSRFs)
3606  safetyInfo->unsafeVolatile = true;
3607 
3608  /*
3609  * If we're at a leaf query, check for unsafe expressions in its target
3610  * list, and mark any reasons why they're unsafe in unsafeFlags[].
3611  * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3612  * so no need to check them.)
3613  */
3614  if (subquery->setOperations == NULL)
3615  check_output_expressions(subquery, safetyInfo);
3616 
3617  /* Are we at top level, or looking at a setop component? */
3618  if (subquery == topquery)
3619  {
3620  /* Top level, so check any component queries */
3621  if (subquery->setOperations != NULL)
3622  if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3623  safetyInfo))
3624  return false;
3625  }
3626  else
3627  {
3628  /* Setop component must not have more components (too weird) */
3629  if (subquery->setOperations != NULL)
3630  return false;
3631  /* Check whether setop component output types match top level */
3632  topop = castNode(SetOperationStmt, topquery->setOperations);
3633  Assert(topop);
3635  topop->colTypes,
3636  safetyInfo);
3637  }
3638  return true;
3639 }
3640 
3641 /*
3642  * Helper routine to recurse through setOperations tree
3643  */
3644 static bool
3646  pushdown_safety_info *safetyInfo)
3647 {
3648  if (IsA(setOp, RangeTblRef))
3649  {
3650  RangeTblRef *rtr = (RangeTblRef *) setOp;
3651  RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3652  Query *subquery = rte->subquery;
3653 
3654  Assert(subquery != NULL);
3655  return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3656  }
3657  else if (IsA(setOp, SetOperationStmt))
3658  {
3659  SetOperationStmt *op = (SetOperationStmt *) setOp;
3660 
3661  /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
3662  if (op->op == SETOP_EXCEPT)
3663  return false;
3664  /* Else recurse */
3665  if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
3666  return false;
3667  if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
3668  return false;
3669  }
3670  else
3671  {
3672  elog(ERROR, "unrecognized node type: %d",
3673  (int) nodeTag(setOp));
3674  }
3675  return true;
3676 }
3677 
3678 /*
3679  * check_output_expressions - check subquery's output expressions for safety
3680  *
3681  * There are several cases in which it's unsafe to push down an upper-level
3682  * qual if it references a particular output column of a subquery. We check
3683  * each output column of the subquery and set flags in unsafeFlags[k] when we
3684  * see that column is unsafe for a pushed-down qual to reference. The
3685  * conditions checked here are:
3686  *
3687  * 1. We must not push down any quals that refer to subselect outputs that
3688  * return sets, else we'd introduce functions-returning-sets into the
3689  * subquery's WHERE/HAVING quals.
3690  *
3691  * 2. We must not push down any quals that refer to subselect outputs that
3692  * contain volatile functions, for fear of introducing strange results due
3693  * to multiple evaluation of a volatile function.
3694  *
3695  * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3696  * refer to non-DISTINCT output columns, because that could change the set
3697  * of rows returned. (This condition is vacuous for DISTINCT, because then
3698  * there are no non-DISTINCT output columns, so we needn't check. Note that
3699  * subquery_is_pushdown_safe already reported that we can't use volatile
3700  * quals if there's DISTINCT or DISTINCT ON.)
3701  *
3702  * 4. If the subquery has any window functions, we must not push down quals
3703  * that reference any output columns that are not listed in all the subquery's
3704  * window PARTITION BY clauses. We can push down quals that use only
3705  * partitioning columns because they should succeed or fail identically for
3706  * every row of any one window partition, and totally excluding some
3707  * partitions will not change a window function's results for remaining
3708  * partitions. (Again, this also requires nonvolatile quals, but
3709  * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3710  * unsafe for this reason can still have WindowClause run conditions pushed
3711  * down.
3712  */
3713 static void
3715 {
3716  ListCell *lc;
3717 
3718  foreach(lc, subquery->targetList)
3719  {
3720  TargetEntry *tle = (TargetEntry *) lfirst(lc);
3721 
3722  if (tle->resjunk)
3723  continue; /* ignore resjunk columns */
3724 
3725  /* Functions returning sets are unsafe (point 1) */
3726  if (subquery->hasTargetSRFs &&
3727  (safetyInfo->unsafeFlags[tle->resno] &
3728  UNSAFE_HAS_SET_FUNC) == 0 &&
3729  expression_returns_set((Node *) tle->expr))
3730  {
3731  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3732  continue;
3733  }
3734 
3735  /* Volatile functions are unsafe (point 2) */
3736  if ((safetyInfo->unsafeFlags[tle->resno] &
3737  UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3739  {
3740  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3741  continue;
3742  }
3743 
3744  /* If subquery uses DISTINCT ON, check point 3 */
3745  if (subquery->hasDistinctOn &&
3746  (safetyInfo->unsafeFlags[tle->resno] &
3748  !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3749  {
3750  /* non-DISTINCT column, so mark it unsafe */
3751  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3752  continue;
3753  }
3754 
3755  /* If subquery uses window functions, check point 4 */
3756  if (subquery->hasWindowFuncs &&
3757  (safetyInfo->unsafeFlags[tle->resno] &
3759  !targetIsInAllPartitionLists(tle, subquery))
3760  {
3761  /* not present in all PARTITION BY clauses, so mark it unsafe */
3762  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3763  continue;
3764  }
3765  }
3766 }
3767 
3768 /*
3769  * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3770  * push quals into each component query, but the quals can only reference
3771  * subquery columns that suffer no type coercions in the set operation.
3772  * Otherwise there are possible semantic gotchas. So, we check the
3773  * component queries to see if any of them have output types different from
3774  * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3775  * unsafeFlags[k] if column k has different type in any component.
3776  *
3777  * We don't have to care about typmods here: the only allowed difference
3778  * between set-op input and output typmods is input is a specific typmod
3779  * and output is -1, and that does not require a coercion.
3780  *
3781  * tlist is a subquery tlist.
3782  * colTypes is an OID list of the top-level setop's output column types.
3783  * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
3784  */
3785 static void
3787  pushdown_safety_info *safetyInfo)
3788 {
3789  ListCell *l;
3790  ListCell *colType = list_head(colTypes);
3791 
3792  foreach(l, tlist)
3793  {
3794  TargetEntry *tle = (TargetEntry *) lfirst(l);
3795 
3796  if (tle->resjunk)
3797  continue; /* ignore resjunk columns */
3798  if (colType == NULL)
3799  elog(ERROR, "wrong number of tlist entries");
3800  if (exprType((Node *) tle->expr) != lfirst_oid(colType))
3801  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
3802  colType = lnext(colTypes, colType);
3803  }
3804  if (colType != NULL)
3805  elog(ERROR, "wrong number of tlist entries");
3806 }
3807 
3808 /*
3809  * targetIsInAllPartitionLists
3810  * True if the TargetEntry is listed in the PARTITION BY clause
3811  * of every window defined in the query.
3812  *
3813  * It would be safe to ignore windows not actually used by any window
3814  * function, but it's not easy to get that info at this stage; and it's
3815  * unlikely to be useful to spend any extra cycles getting it, since
3816  * unreferenced window definitions are probably infrequent in practice.
3817  */
3818 static bool
3820 {
3821  ListCell *lc;
3822 
3823  foreach(lc, query->windowClause)
3824  {
3825  WindowClause *wc = (WindowClause *) lfirst(lc);
3826 
3828  return false;
3829  }
3830  return true;
3831 }
3832 
3833 /*
3834  * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3835  *
3836  * rinfo is a restriction clause applying to the given subquery (whose RTE
3837  * has index rti in the parent query).
3838  *
3839  * Conditions checked here:
3840  *
3841  * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3842  * unclear that it will work correctly: SubLinks will already have been
3843  * transformed into SubPlans in the qual, but not in the subquery). Note that
3844  * SubLinks that transform to initplans are safe, and will be accepted here
3845  * because what we'll see in the qual is just a Param referencing the initplan
3846  * output.
3847  *
3848  * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3849  * functions.
3850  *
3851  * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
3852  * functions that are passed Var nodes, and therefore might reveal values from
3853  * the subquery as side effects.
3854  *
3855  * 4. rinfo's clause must not refer to the whole-row output of the subquery
3856  * (since there is no easy way to name that within the subquery itself).
3857  *
3858  * 5. rinfo's clause must not refer to any subquery output columns that were
3859  * found to be unsafe to reference by subquery_is_pushdown_safe().
3860  */
3861 static pushdown_safe_type
3863  pushdown_safety_info *safetyInfo)
3864 {
3866  Node *qual = (Node *) rinfo->clause;
3867  List *vars;
3868  ListCell *vl;
3869 
3870  /* Refuse subselects (point 1) */
3871  if (contain_subplans(qual))
3872  return PUSHDOWN_UNSAFE;
3873 
3874  /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3875  if (safetyInfo->unsafeVolatile &&
3876  contain_volatile_functions((Node *) rinfo))
3877  return PUSHDOWN_UNSAFE;
3878 
3879  /* Refuse leaky quals if told to (point 3) */
3880  if (safetyInfo->unsafeLeaky &&
3881  contain_leaked_vars(qual))
3882  return PUSHDOWN_UNSAFE;
3883 
3884  /*
3885  * Examine all Vars used in clause. Since it's a restriction clause, all
3886  * such Vars must refer to subselect output columns ... unless this is
3887  * part of a LATERAL subquery, in which case there could be lateral
3888  * references.
3889  *
3890  * By omitting the relevant flags, this also gives us a cheap sanity check
3891  * that no aggregates or window functions appear in the qual. Those would
3892  * be unsafe to push down, but at least for the moment we could never see
3893  * any in a qual anyhow.
3894  */
3896  foreach(vl, vars)
3897  {
3898  Var *var = (Var *) lfirst(vl);
3899 
3900  /*
3901  * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3902  * It's not clear whether a PHV could safely be pushed down, and even
3903  * less clear whether such a situation could arise in any cases of
3904  * practical interest anyway. So for the moment, just refuse to push
3905  * down.
3906  */
3907  if (!IsA(var, Var))
3908  {
3909  safe = PUSHDOWN_UNSAFE;
3910  break;
3911  }
3912 
3913  /*
3914  * Punt if we find any lateral references. It would be safe to push
3915  * these down, but we'd have to convert them into outer references,
3916  * which subquery_push_qual lacks the infrastructure to do. The case
3917  * arises so seldom that it doesn't seem worth working hard on.
3918  */
3919  if (var->varno != rti)
3920  {
3921  safe = PUSHDOWN_UNSAFE;
3922  break;
3923  }
3924 
3925  /* Subqueries have no system columns */
3926  Assert(var->varattno >= 0);
3927 
3928  /* Check point 4 */
3929  if (var->varattno == 0)
3930  {
3931  safe = PUSHDOWN_UNSAFE;
3932  break;
3933  }
3934 
3935  /* Check point 5 */
3936  if (safetyInfo->unsafeFlags[var->varattno] != 0)
3937  {
3938  if (safetyInfo->unsafeFlags[var->varattno] &
3941  {
3942  safe = PUSHDOWN_UNSAFE;
3943  break;
3944  }
3945  else
3946  {
3947  /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
3949  /* don't break, we might find another Var that's unsafe */
3950  }
3951  }
3952  }
3953 
3954  list_free(vars);
3955 
3956  return safe;
3957 }
3958 
3959 /*
3960  * subquery_push_qual - push down a qual that we have determined is safe
3961  */
3962 static void
3963 subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3964 {
3965  if (subquery->setOperations != NULL)
3966  {
3967  /* Recurse to push it separately to each component query */
3968  recurse_push_qual(subquery->setOperations, subquery,
3969  rte, rti, qual);
3970  }
3971  else
3972  {
3973  /*
3974  * We need to replace Vars in the qual (which must refer to outputs of
3975  * the subquery) with copies of the subquery's targetlist expressions.
3976  * Note that at this point, any uplevel Vars in the qual should have
3977  * been replaced with Params, so they need no work.
3978  *
3979  * This step also ensures that when we are pushing into a setop tree,
3980  * each component query gets its own copy of the qual.
3981  */
3982  qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
3983  subquery->targetList,
3985  &subquery->hasSubLinks);
3986 
3987  /*
3988  * Now attach the qual to the proper place: normally WHERE, but if the
3989  * subquery uses grouping or aggregation, put it in HAVING (since the
3990  * qual really refers to the group-result rows).
3991  */
3992  if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
3993  subquery->havingQual = make_and_qual(subquery->havingQual, qual);
3994  else
3995  subquery->jointree->quals =
3996  make_and_qual(subquery->jointree->quals, qual);
3997 
3998  /*
3999  * We need not change the subquery's hasAggs or hasSubLinks flags,
4000  * since we can't be pushing down any aggregates that weren't there
4001  * before, and we don't push down subselects at all.
4002  */
4003  }
4004 }
4005 
4006 /*
4007  * Helper routine to recurse through setOperations tree
4008  */
4009 static void
4010 recurse_push_qual(Node *setOp, Query *topquery,
4011  RangeTblEntry *rte, Index rti, Node *qual)
4012 {
4013  if (IsA(setOp, RangeTblRef))
4014  {
4015  RangeTblRef *rtr = (RangeTblRef *) setOp;
4016  RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4017  Query *subquery = subrte->subquery;
4018 
4019  Assert(subquery != NULL);
4020  subquery_push_qual(subquery, rte, rti, qual);
4021  }
4022  else if (IsA(setOp, SetOperationStmt))
4023  {
4024  SetOperationStmt *op = (SetOperationStmt *) setOp;
4025 
4026  recurse_push_qual(op->larg, topquery, rte, rti, qual);
4027  recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4028  }
4029  else
4030  {
4031  elog(ERROR, "unrecognized node type: %d",
4032  (int) nodeTag(setOp));
4033  }
4034 }
4035 
4036 /*****************************************************************************
4037  * SIMPLIFYING SUBQUERY TARGETLISTS
4038  *****************************************************************************/
4039 
4040 /*
4041  * remove_unused_subquery_outputs
4042  * Remove subquery targetlist items we don't need
4043  *
4044  * It's possible, even likely, that the upper query does not read all the
4045  * output columns of the subquery. We can remove any such outputs that are
4046  * not needed by the subquery itself (e.g., as sort/group columns) and do not
4047  * affect semantics otherwise (e.g., volatile functions can't be removed).
4048  * This is useful not only because we might be able to remove expensive-to-
4049  * compute expressions, but because deletion of output columns might allow
4050  * optimizations such as join removal to occur within the subquery.
4051  *
4052  * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4053  * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4054  * parameter is modified by the function, so callers must make a copy if they
4055  * need to use the passed in Bitmapset after calling this function.
4056  *
4057  * To avoid affecting column numbering in the targetlist, we don't physically
4058  * remove unused tlist entries, but rather replace their expressions with NULL
4059  * constants. This is implemented by modifying subquery->targetList.
4060  */
4061 static void
4063  Bitmapset *extra_used_attrs)
4064 {
4065  Bitmapset *attrs_used;
4066  ListCell *lc;
4067 
4068  /*
4069  * Just point directly to extra_used_attrs. No need to bms_copy as none of
4070  * the current callers use the Bitmapset after calling this function.
4071  */
4072  attrs_used = extra_used_attrs;
4073 
4074  /*
4075  * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4076  * could update all the child SELECTs' tlists, but it seems not worth the
4077  * trouble presently.
4078  */
4079  if (subquery->setOperations)
4080  return;
4081 
4082  /*
4083  * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4084  * time: all its output columns must be used in the distinctClause.
4085  */
4086  if (subquery->distinctClause && !subquery->hasDistinctOn)
4087  return;
4088 
4089  /*
4090  * Collect a bitmap of all the output column numbers used by the upper
4091  * query.
4092  *
4093  * Add all the attributes needed for joins or final output. Note: we must
4094  * look at rel's targetlist, not the attr_needed data, because attr_needed
4095  * isn't computed for inheritance child rels, cf set_append_rel_size().
4096  * (XXX might be worth changing that sometime.)
4097  */
4098  pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4099 
4100  /* Add all the attributes used by un-pushed-down restriction clauses. */
4101  foreach(lc, rel->baserestrictinfo)
4102  {
4103  RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4104 
4105  pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4106  }
4107 
4108  /*
4109  * If there's a whole-row reference to the subquery, we can't remove
4110  * anything.
4111  */
4113  return;
4114 
4115  /*
4116  * Run through the tlist and zap entries we don't need. It's okay to
4117  * modify the tlist items in-place because set_subquery_pathlist made a
4118  * copy of the subquery.
4119  */
4120  foreach(lc, subquery->targetList)
4121  {
4122  TargetEntry *tle = (TargetEntry *) lfirst(lc);
4123  Node *texpr = (Node *) tle->expr;
4124 
4125  /*
4126  * If it has a sortgroupref number, it's used in some sort/group
4127  * clause so we'd better not remove it. Also, don't remove any
4128  * resjunk columns, since their reason for being has nothing to do
4129  * with anybody reading the subquery's output. (It's likely that
4130  * resjunk columns in a sub-SELECT would always have ressortgroupref
4131  * set, but even if they don't, it seems imprudent to remove them.)
4132  */
4133  if (tle->ressortgroupref || tle->resjunk)
4134  continue;
4135 
4136  /*
4137  * If it's used by the upper query, we can't remove it.
4138  */
4140  attrs_used))
4141  continue;
4142 
4143  /*
4144  * If it contains a set-returning function, we can't remove it since
4145  * that could change the number of rows returned by the subquery.
4146  */
4147  if (subquery->hasTargetSRFs &&
4148  expression_returns_set(texpr))
4149  continue;
4150 
4151  /*
4152  * If it contains volatile functions, we daren't remove it for fear
4153  * that the user is expecting their side-effects to happen.
4154  */
4155  if (contain_volatile_functions(texpr))
4156  continue;
4157 
4158  /*
4159  * OK, we don't need it. Replace the expression with a NULL constant.
4160  * Preserve the exposed type of the expression, in case something
4161  * looks at the rowtype of the subquery's result.
4162  */
4163  tle->expr = (Expr *) makeNullConst(exprType(texpr),
4164  exprTypmod(texpr),
4165  exprCollation(texpr));
4166  }
4167 }
4168 
4169 /*
4170  * create_partial_bitmap_paths
4171  * Build partial bitmap heap path for the relation
4172  */
4173 void
4175  Path *bitmapqual)
4176 {
4177  int parallel_workers;
4178  double pages_fetched;
4179 
4180  /* Compute heap pages for bitmap heap scan */
4181  pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4182  NULL, NULL);
4183 
4184  parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4186 
4187  if (parallel_workers <= 0)
4188  return;
4189 
4190  add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4191  bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4192 }
4193 
4194 /*
4195  * Compute the number of parallel workers that should be used to scan a
4196  * relation. We compute the parallel workers based on the size of the heap to
4197  * be scanned and the size of the index to be scanned, then choose a minimum
4198  * of those.
4199  *
4200  * "heap_pages" is the number of pages from the table that we expect to scan, or
4201  * -1 if we don't expect to scan any.
4202  *
4203  * "index_pages" is the number of pages from the index that we expect to scan, or
4204  * -1 if we don't expect to scan any.
4205  *
4206  * "max_workers" is caller's limit on the number of workers. This typically
4207  * comes from a GUC.
4208  */
4209 int
4210 compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4211  int max_workers)
4212 {
4213  int parallel_workers = 0;
4214 
4215  /*
4216  * If the user has set the parallel_workers reloption, use that; otherwise
4217  * select a default number of workers.
4218  */
4219  if (rel->rel_parallel_workers != -1)
4220  parallel_workers = rel->rel_parallel_workers;
4221  else
4222  {
4223  /*
4224  * If the number of pages being scanned is insufficient to justify a
4225  * parallel scan, just return zero ... unless it's an inheritance
4226  * child. In that case, we want to generate a parallel path here
4227  * anyway. It might not be worthwhile just for this relation, but
4228  * when combined with all of its inheritance siblings it may well pay
4229  * off.
4230  */
4231  if (rel->reloptkind == RELOPT_BASEREL &&
4232  ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4233  (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4234  return 0;
4235 
4236  if (heap_pages >= 0)
4237  {
4238  int heap_parallel_threshold;
4239  int heap_parallel_workers = 1;
4240 
4241  /*
4242  * Select the number of workers based on the log of the size of
4243  * the relation. This probably needs to be a good deal more
4244  * sophisticated, but we need something here for now. Note that
4245  * the upper limit of the min_parallel_table_scan_size GUC is
4246  * chosen to prevent overflow here.
4247  */
4248  heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4249  while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4250  {
4251  heap_parallel_workers++;
4252  heap_parallel_threshold *= 3;
4253  if (heap_parallel_threshold > INT_MAX / 3)
4254  break; /* avoid overflow */
4255  }
4256 
4257  parallel_workers = heap_parallel_workers;
4258  }
4259 
4260  if (index_pages >= 0)
4261  {
4262  int index_parallel_workers = 1;
4263  int index_parallel_threshold;
4264 
4265  /* same calculation as for heap_pages above */
4266  index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4267  while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4268  {
4269  index_parallel_workers++;
4270  index_parallel_threshold *= 3;
4271  if (index_parallel_threshold > INT_MAX / 3)
4272  break; /* avoid overflow */
4273  }
4274 
4275  if (parallel_workers > 0)
4276  parallel_workers = Min(parallel_workers, index_parallel_workers);
4277  else
4278  parallel_workers = index_parallel_workers;
4279  }
4280  }
4281 
4282  /* In no case use more than caller supplied maximum number of workers */
4283  parallel_workers = Min(parallel_workers, max_workers);
4284 
4285  return parallel_workers;
4286 }
4287 
4288 /*
4289  * generate_partitionwise_join_paths
4290  * Create paths representing partitionwise join for given partitioned
4291  * join relation.
4292  *
4293  * This must not be called until after we are done adding paths for all
4294  * child-joins. Otherwise, add_path might delete a path to which some path
4295  * generated here has a reference.
4296  */
4297 void
4299 {
4300  List *live_children = NIL;
4301  int cnt_parts;
4302  int num_parts;
4303  RelOptInfo **part_rels;
4304 
4305  /* Handle only join relations here. */
4306  if (!IS_JOIN_REL(rel))
4307  return;
4308 
4309  /* We've nothing to do if the relation is not partitioned. */
4310  if (!IS_PARTITIONED_REL(rel))
4311  return;
4312 
4313  /* The relation should have consider_partitionwise_join set. */
4315 
4316  /* Guard against stack overflow due to overly deep partition hierarchy. */
4318 
4319  num_parts = rel->nparts;
4320  part_rels = rel->part_rels;
4321 
4322  /* Collect non-dummy child-joins. */
4323  for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4324  {
4325  RelOptInfo *child_rel = part_rels[cnt_parts];
4326 
4327  /* If it's been pruned entirely, it's certainly dummy. */
4328  if (child_rel == NULL)
4329  continue;
4330 
4331  /* Make partitionwise join paths for this partitioned child-join. */
4332  generate_partitionwise_join_paths(root, child_rel);
4333 
4334  /* If we failed to make any path for this child, we must give up. */
4335  if (child_rel->pathlist == NIL)
4336  {
4337  /*
4338  * Mark the parent joinrel as unpartitioned so that later
4339  * functions treat it correctly.
4340  */
4341  rel->nparts = 0;
4342  return;
4343  }
4344 
4345  /* Else, identify the cheapest path for it. */
4346  set_cheapest(child_rel);
4347 
4348  /* Dummy children need not be scanned, so ignore those. */
4349  if (IS_DUMMY_REL(child_rel))
4350  continue;
4351 
4352 #ifdef OPTIMIZER_DEBUG
4353  pprint(child_rel);
4354 #endif
4355 
4356  live_children = lappend(live_children, child_rel);
4357  }
4358 
4359  /* If all child-joins are dummy, parent join is also dummy. */
4360  if (!live_children)
4361  {
4362  mark_dummy_rel(rel);
4363  return;
4364  }
4365 
4366  /* Build additional paths for this rel from child-join paths. */
4367  add_paths_to_append_rel(root, rel, live_children);
4368  list_free(live_children);
4369 }
static void set_base_rel_sizes(PlannerInfo *root)
Definition: allpaths.c:290
static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:2494
static bool find_window_run_conditions(Query *subquery, RangeTblEntry *rte, Index rti, AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr, bool wfunc_left, bool *keep_original, Bitmapset **run_cond_attrs)
Definition: allpaths.c:2214
#define UNSAFE_TYPE_MISMATCH
Definition: allpaths.c:57
static Path * get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: allpaths.c:1999
static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2940
static void subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
Definition: allpaths.c:3963
void generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
Definition: allpaths.c:4298
RelOptInfo * standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
Definition: allpaths.c:3418
static void set_base_rel_consider_startup(PlannerInfo *root)
Definition: allpaths.c:247
#define UNSAFE_HAS_VOLATILE_FUNC
Definition: allpaths.c:53
#define UNSAFE_NOTIN_DISTINCTON_CLAUSE
Definition: allpaths.c:55
static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:469
static Path * get_singleton_append_subpath(Path *path)
Definition: allpaths.c:2132
static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:814
static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:854
static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:926
static void set_base_rel_pathlists(PlannerInfo *root)
Definition: allpaths.c:333
int geqo_threshold
Definition: allpaths.c:80
static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:1232
static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3862
static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2970
int compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages, int max_workers)
Definition: allpaths.c:4210
void generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
Definition: allpaths.c:3059
static void set_dummy_rel_pathlist(RelOptInfo *rel)
Definition: allpaths.c:2166
static void compare_tlist_datatypes(List *tlist, List *colTypes, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3786
static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:3000
static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
Definition: allpaths.c:3819
static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
Definition: allpaths.c:794
static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3589
join_search_hook_type join_search_hook
Definition: allpaths.c:88
bool enable_geqo
Definition: allpaths.c:79
static bool check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti, Node *clause, Bitmapset **run_cond_attrs)
Definition: allpaths.c:2419
void generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
Definition: allpaths.c:3197
static RelOptInfo * make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
Definition: allpaths.c:3313
static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2762
static void recurse_push_qual(Node *setOp, Query *topquery, RangeTblEntry *rte, Index rti, Node *qual)
Definition: allpaths.c:4010
static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:764
static List * get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel, bool require_parallel_safe)
Definition: allpaths.c:3129
static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel, Bitmapset *extra_used_attrs)
Definition: allpaths.c:4062
static void check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3714
static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2849
static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:902
set_rel_pathlist_hook_type set_rel_pathlist_hook
Definition: allpaths.c:85
static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2829
struct pushdown_safety_info pushdown_safety_info
#define UNSAFE_HAS_SET_FUNC
Definition: allpaths.c:54
static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2873
static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:589
static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:944
void create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel, Path *bitmapqual)
Definition: allpaths.c:4174
static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:572
void add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel, List *live_childrels)
Definition: allpaths.c:1302
static void set_rel_size(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:360
static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel, List *live_childrels, List *all_child_pathkeys)
Definition: allpaths.c:1714
#define UNSAFE_NOTIN_PARTITIONBY_CLAUSE
Definition: allpaths.c:56
RelOptInfo * make_one_rel(PlannerInfo *root, List *joinlist)
Definition: allpaths.c:171
static bool recurse_pushdown_safe(Node *setOp, Query *topquery, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3645
pushdown_safe_type
Definition: allpaths.c:71
@ PUSHDOWN_WINDOWCLAUSE_RUNCOND
Definition: allpaths.c:74
@ PUSHDOWN_UNSAFE
Definition: allpaths.c:72
@ PUSHDOWN_SAFE
Definition: allpaths.c:73
static void accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
Definition: allpaths.c:2087
int min_parallel_index_scan_size
Definition: allpaths.c:82
int min_parallel_table_scan_size
Definition: allpaths.c:81
Node * adjust_appendrel_attrs(PlannerInfo *root, Node *node, int nappinfos, AppendRelInfo **appinfos)
Definition: appendinfo.c:196
int16 AttrNumber
Definition: attnum.h:21
#define InvalidAttrNumber
Definition: attnum.h:23
void pprint(const void *obj)
Definition: print.c:54
bool bms_equal(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:142
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:412
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:510
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:815
BMS_Membership bms_membership(const Bitmapset *a)
Definition: bitmapset.c:781
bool bms_overlap(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:582
bool bms_get_singleton_member(const Bitmapset *a, int *member)
Definition: bitmapset.c:715
#define bms_is_empty(a)
Definition: bitmapset.h:118
@ BMS_SINGLETON
Definition: bitmapset.h:72
@ BMS_MULTIPLE
Definition: bitmapset.h:73
uint32 BlockNumber
Definition: block.h:31
#define Min(x, y)
Definition: c.h:991
signed short int16
Definition: c.h:480
signed int int32
Definition: c.h:481
#define Max(x, y)
Definition: c.h:985
unsigned int Index
Definition: c.h:601
#define OidIsValid(objectId)
Definition: c.h:762
bool is_pseudo_constant_clause(Node *clause)
Definition: clauses.c:2068
bool contain_leaked_vars(Node *clause)
Definition: clauses.c:1243
bool is_parallel_safe(PlannerInfo *root, Node *node)
Definition: clauses.c:733
bool contain_subplans(Node *clause)
Definition: clauses.c:329
bool contain_volatile_functions(Node *clause)
Definition: clauses.c:518
void set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:6032
int max_parallel_workers_per_gather
Definition: costsize.c:132
void set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5269
void set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5902
void set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
Definition: costsize.c:5994
void set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:6065
bool enable_partitionwise_join
Definition: costsize.c:148
double compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel, Path *bitmapqual, double loop_count, Cost *cost_p, double *tuples_p)
Definition: costsize.c:6433
void set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5822
bool enable_parallel_append
Definition: costsize.c:150
void set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:6094
double clamp_row_est(double nrows)
Definition: costsize.c:202
void set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5940
void set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5962
bool enable_incremental_sort
Definition: costsize.c:140
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:224
void add_child_rel_equivalences(PlannerInfo *root, AppendRelInfo *appinfo, RelOptInfo *parent_rel, RelOptInfo *child_rel)
Definition: equivclass.c:2632
bool relation_can_be_sorted_early(PlannerInfo *root, RelOptInfo *rel, EquivalenceClass *ec, bool require_parallel_safe)
Definition: equivclass.c:933
#define OidFunctionCall1(functionId, arg1)
Definition: fmgr.h:680
RelOptInfo * geqo(PlannerInfo *root, int number_of_rels, List *initial_rels)
Definition: geqo_main.c:72
void check_index_predicates(PlannerInfo *root, RelOptInfo *rel)
Definition: indxpath.c:3298
void create_index_paths(PlannerInfo *root, RelOptInfo *rel)
Definition: indxpath.c:235
int i
Definition: isn.c:73
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:77
void join_search_one_level(PlannerInfo *root, int level)
Definition: joinrels.c:71
void mark_dummy_rel(RelOptInfo *rel)
Definition: joinrels.c:1363
Assert(fmt[strlen(fmt) - 1] !='\n')
List * lappend(List *list, void *datum)
Definition: list.c:339
List * list_copy_head(const List *oldlist, int len)
Definition: list.c:1593
void list_free(List *list)
Definition: list.c:1546
List * list_copy_tail(const List *oldlist, int nskip)
Definition: list.c:1613
List * list_concat(List *list1, const List *list2)
Definition: list.c:561
char get_rel_persistence(Oid relid)
Definition: lsyscache.c:2056
char func_parallel(Oid funcid)
Definition: lsyscache.c:1777
RegProcedure get_func_support(Oid funcid)
Definition: lsyscache.c:1836
Oid get_opfamily_member(Oid opfamily, Oid lefttype, Oid righttype, int16 strategy)
Definition: lsyscache.c:166
bool func_strict(Oid funcid)
Definition: lsyscache.c:1739
List * get_op_btree_interpretation(Oid opno)
Definition: lsyscache.c:601
int32 get_typavgwidth(Oid typid, int32 typmod)
Definition: lsyscache.c:2534
Datum subpath(PG_FUNCTION_ARGS)
Definition: ltree_op.c:241
Expr * make_opclause(Oid opno, Oid opresulttype, bool opretset, Expr *leftop, Expr *rightop, Oid opcollid, Oid inputcollid)
Definition: makefuncs.c:612
Const * makeNullConst(Oid consttype, int32 consttypmod, Oid constcollid)
Definition: makefuncs.c:339
Node * make_and_qual(Node *qual1, Node *qual2)
Definition: makefuncs.c:691
void pfree(void *pointer)
Definition: mcxt.c:1508
void * palloc0(Size size)
Definition: mcxt.c:1334
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:42
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:284
Oid exprCollation(const Node *expr)
Definition: nodeFuncs.c:788
bool expression_returns_set(Node *clause)
Definition: nodeFuncs.c:730
void set_opfuncid(OpExpr *opexpr)
Definition: nodeFuncs.c:1790
#define IsA(nodeptr, _type_)
Definition: nodes.h:158
#define copyObject(obj)
Definition: nodes.h:223
#define nodeTag(nodeptr)
Definition: nodes.h:133
#define castNode(_type_, nodeptr)
Definition: nodes.h:176
@ JOIN_SEMI
Definition: nodes.h:297
@ JOIN_ANTI
Definition: nodes.h:298
#define PVC_INCLUDE_PLACEHOLDERS
Definition: optimizer.h:190
bool targetIsInSortList(TargetEntry *tle, Oid sortop, List *sortList)
@ SETOP_EXCEPT
Definition: parsenodes.h:1964
@ RTE_JOIN
Definition: parsenodes.h:1013
@ RTE_CTE
Definition: parsenodes.h:1017
@ RTE_NAMEDTUPLESTORE
Definition: parsenodes.h:1018
@ RTE_VALUES
Definition: parsenodes.h:1016
@ RTE_SUBQUERY
Definition: parsenodes.h:1012
@ RTE_RESULT
Definition: parsenodes.h:1019
@ RTE_FUNCTION
Definition: parsenodes.h:1014
@ RTE_TABLEFUNC
Definition: parsenodes.h:1015
@ RTE_RELATION
Definition: parsenodes.h:1011
#define rt_fetch(rangetable_index, rangetable)
Definition: parsetree.h:31
bool partitions_are_ordered(PartitionBoundInfo boundinfo, Bitmapset *live_parts)
Definition: partbounds.c:2852
List * build_expression_pathkey(PlannerInfo *root, Expr *expr, Oid opno, Relids rel, bool create_it)
Definition: pathkeys.c:1015
Path * get_cheapest_path_for_pathkeys(List *paths, List *pathkeys, Relids required_outer, CostSelector cost_criterion, bool require_parallel_safe)
Definition: pathkeys.c:635
bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common)
Definition: pathkeys.c:573
bool has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
Definition: pathkeys.c:2242
Path * get_cheapest_parallel_safe_total_inner(List *paths)
Definition: pathkeys.c:714
List * convert_subquery_pathkeys(PlannerInfo *root, RelOptInfo *rel, List *subquery_pathkeys, List *subquery_tlist)
Definition: pathkeys.c:1069
Path * get_cheapest_fractional_path_for_pathkeys(List *paths, List *pathkeys, Relids required_outer, double fraction)
Definition: pathkeys.c:681
bool pathkeys_contained_in(List *keys1, List *keys2)
Definition: pathkeys.c:341
PathKeysComparison compare_pathkeys(List *keys1, List *keys2)
Definition: pathkeys.c:302
List * build_partition_pathkeys(PlannerInfo *root, RelOptInfo *partrel, ScanDirection scandir, bool *partialkeys)
Definition: pathkeys.c:934
AppendPath * create_append_path(PlannerInfo *root, RelOptInfo *rel, List *subpaths, List *partial_subpaths, List *pathkeys, Relids required_outer, int parallel_workers, bool parallel_aware, double rows)
Definition: pathnode.c:1242
SortPath * create_sort_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, List *pathkeys, double limit_tuples)
Definition: pathnode.c:2986
GatherMergePath * create_gather_merge_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, PathTarget *target, List *pathkeys, Relids required_outer, double *rows)
Definition: pathnode.c:1878
Path * create_tablefuncscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2069
Path * create_namedtuplestorescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2146
GatherPath * create_gather_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, PathTarget *target, Relids required_outer, double *rows)
Definition: pathnode.c:1969
Path * create_resultscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2172
Path * create_samplescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:950
MergeAppendPath * create_merge_append_path(PlannerInfo *root, RelOptInfo *rel, List *subpaths, List *pathkeys, Relids required_outer)
Definition: pathnode.c:1413
void set_cheapest(RelOptInfo *parent_rel)
Definition: pathnode.c:240
Path * reparameterize_path(PlannerInfo *root, Path *path, Relids required_outer, double loop_count)
Definition: pathnode.c:3929
SubqueryScanPath * create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, bool trivial_pathtarget, List *pathkeys, Relids required_outer)
Definition: pathnode.c:2013
Path * create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2121
void add_partial_path(RelOptInfo *parent_rel, Path *new_path)
Definition: pathnode.c:745
MaterialPath * create_material_path(RelOptInfo *rel, Path *subpath)
Definition: pathnode.c:1564
Path * create_functionscan_path(PlannerInfo *root, RelOptInfo *rel, List *pathkeys, Relids required_outer)
Definition: pathnode.c:2043
IncrementalSortPath * create_incremental_sort_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, List *pathkeys, int presorted_keys, double limit_tuples)
Definition: pathnode.c:2937
Path * create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2198
void add_path(RelOptInfo *parent_rel, Path *new_path)
Definition: pathnode.c:418
Path * create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2095
int compare_path_costs(Path *path1, Path *path2, CostSelector criterion)
Definition: pathnode.c:67
Path * create_seqscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer, int parallel_workers)
Definition: pathnode.c:925
BitmapHeapPath * create_bitmap_heap_path(PlannerInfo *root, RelOptInfo *rel, Path *bitmapqual, Relids required_outer, double loop_count, int parallel_degree)
Definition: pathnode.c:1040
#define IS_SIMPLE_REL(rel)
Definition: pathnodes.h:824
#define IS_DUMMY_REL(r)
Definition: pathnodes.h:1926
#define IS_JOIN_REL(rel)
Definition: pathnodes.h:829
@ TOTAL_COST
Definition: pathnodes.h:38
@ STARTUP_COST
Definition: pathnodes.h:38
#define IS_PARTITIONED_REL(rel)
Definition: pathnodes.h:1043
#define PATH_REQ_OUTER(path)
Definition: pathnodes.h:1649
Bitmapset * Relids
Definition: pathnodes.h:30
@ UPPERREL_FINAL
Definition: pathnodes.h:79
@ RELOPT_BASEREL
Definition: pathnodes.h:812
@ RELOPT_OTHER_MEMBER_REL
Definition: pathnodes.h:814
RelOptInfo *(* join_search_hook_type)(PlannerInfo *root, int levels_needed, List *initial_rels)
Definition: paths.h:46
void(* set_rel_pathlist_hook_type)(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: paths.h:30
@ PATHKEYS_EQUAL
Definition: paths.h:199
void * arg
static int pg_leftmost_one_pos32(uint32 word)
Definition: pg_bitutils.h:41
#define lfirst(lc)
Definition: pg_list.h:172
static int list_length(const List *l)
Definition: pg_list.h:152
#define NIL
Definition: pg_list.h:68
#define forboth(cell1, list1, cell2, list2)
Definition: pg_list.h:518
#define foreach_current_index(var_or_cell)
Definition: pg_list.h:403
#define list_make1(x1)
Definition: pg_list.h:212
static ListCell * list_head(const List *l)
Definition: pg_list.h:128
#define for_each_from(cell, lst, N)
Definition: pg_list.h:414
#define linitial(l)
Definition: pg_list.h:178
#define lsecond(l)
Definition: pg_list.h:183
static void * list_nth(const List *list, int n)
Definition: pg_list.h:299
#define list_nth_node(type, list, n)
Definition: pg_list.h:327
static ListCell * lnext(const List *l, const ListCell *c)
Definition: pg_list.h:343
#define lfirst_oid(lc)
Definition: pg_list.h:174
static int list_nth_int(const List *list, int n)
Definition: pg_list.h:310
bool relation_excluded_by_constraints(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: plancat.c:1546
PlannerInfo * subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root, bool hasRecursion, double tuple_fraction)
Definition: planner.c:618
bool limit_needed(Query *parse)
Definition: planner.c:2596
@ MONOTONICFUNC_NONE
Definition: plannodes.h:1585
@ MONOTONICFUNC_DECREASING
Definition: plannodes.h:1587
@ MONOTONICFUNC_INCREASING
Definition: plannodes.h:1586
@ MONOTONICFUNC_BOTH
Definition: plannodes.h:1588
void check_stack_depth(void)
Definition: postgres.c:3531
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:322
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:312
#define InvalidOid
Definition: postgres_ext.h:36
unsigned int Oid
Definition: postgres_ext.h:31
static struct subre * parse(struct vars *v, int stopper, int type, struct state *init, struct state *final)
Definition: regcomp.c:715
RelOptInfo * find_base_rel(PlannerInfo *root, int relid)
Definition: relnode.c:407
RelOptInfo * fetch_upper_rel(PlannerInfo *root, UpperRelationKind kind, Relids relids)
Definition: relnode.c:1463
Node * ReplaceVarsFromTargetList(Node *node, int target_varno, int sublevels_up, RangeTblEntry *target_rte, List *targetlist, ReplaceVarsNoMatchOption nomatch_option, int nomatch_varno, bool *outer_hasSubLinks)
@ REPLACEVARS_REPORT_ERROR
Definition: rewriteManip.h:38
@ BackwardScanDirection
Definition: sdir.h:26
@ ForwardScanDirection
Definition: sdir.h:28
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
#define BTLessStrategyNumber
Definition: stratnum.h:29
#define BTEqualStrategyNumber
Definition: stratnum.h:31
#define BTLessEqualStrategyNumber
Definition: stratnum.h:30
#define BTGreaterEqualStrategyNumber
Definition: stratnum.h:32
int first_partial_path
Definition: pathnodes.h:1914
List * subpaths
Definition: pathnodes.h:1912
Index child_relid
Definition: pathnodes.h:2944
Index parent_relid
Definition: pathnodes.h:2943
Node * quals
Definition: primnodes.h:2062
Definition: pg_list.h:54
Definition: nodes.h:129
Oid opno
Definition: primnodes.h:774
List * args
Definition: primnodes.h:792
List * exprs
Definition: pathnodes.h:1513
List * pathkeys
Definition: pathnodes.h:1645
Cardinality rows
Definition: pathnodes.h:1640
int parallel_workers
Definition: pathnodes.h:1637
Cost total_cost
Definition: pathnodes.h:1642
bool parallel_aware
Definition: pathnodes.h:1633
Cardinality plan_rows
Definition: plannodes.h:134
List * subplans
Definition: pathnodes.h:105
bool parallelModeOK
Definition: pathnodes.h:153
int simple_rel_array_size
Definition: pathnodes.h:229
Relids all_query_rels
Definition: pathnodes.h:266
List * cte_plan_ids
Definition: pathnodes.h:302
bool hasHavingQual
Definition: pathnodes.h:493
Index query_level
Definition: pathnodes.h:205
List * append_rel_list
Definition: pathnodes.h:362
struct Path * non_recursive_path
Definition: pathnodes.h:523
PlannerGlobal * glob
Definition: pathnodes.h:202
List * plan_params
Definition: pathnodes.h:217
Cardinality total_table_pages
Definition: pathnodes.h:475
Query * parse
Definition: pathnodes.h:199
List * query_pathkeys
Definition: pathnodes.h:382
Selectivity tuple_fraction
Definition: pathnodes.h:478
List * join_info_list
Definition: pathnodes.h:337
Relids all_baserels
Definition: pathnodes.h:252
Node * limitCount
Definition: parsenodes.h:213
FromExpr * jointree
Definition: parsenodes.h:175
Node * setOperations
Definition: parsenodes.h:218
List * cteList
Definition: parsenodes.h:166
List * groupClause
Definition: parsenodes.h:199
Node * havingQual
Definition: parsenodes.h:204
List * rtable
Definition: parsenodes.h:168
Node * limitOffset
Definition: parsenodes.h:212
List * windowClause
Definition: parsenodes.h:206
List * targetList
Definition: parsenodes.h:190
List * groupingSets
Definition: parsenodes.h:202
List * distinctClause
Definition: parsenodes.h:208
char * ctename
Definition: parsenodes.h:1169
bool self_reference
Definition: parsenodes.h:1171
bool security_barrier
Definition: parsenodes.h:1087
Index ctelevelsup
Definition: parsenodes.h:1170
bool funcordinality
Definition: parsenodes.h:1154
struct TableSampleClause * tablesample
Definition: parsenodes.h:1081
Query * subquery
Definition: parsenodes.h:1086
List * values_lists
Definition: parsenodes.h:1164
List * functions
Definition: parsenodes.h:1153
RTEKind rtekind
Definition: parsenodes.h:1030
List * baserestrictinfo
Definition: pathnodes.h:966
bool consider_param_startup
Definition: pathnodes.h:870
List * subplan_params
Definition: pathnodes.h:935
List * joininfo
Definition: pathnodes.h:972
Relids relids
Definition: pathnodes.h:856
struct PathTarget * reltarget
Definition: pathnodes.h:878
Index relid
Definition: pathnodes.h:903
Cardinality tuples
Definition: pathnodes.h:930
bool consider_parallel
Definition: pathnodes.h:872
BlockNumber pages
Definition: pathnodes.h:929
Relids lateral_relids
Definition: pathnodes.h:898
List * pathlist
Definition: pathnodes.h:883
RelOptKind reloptkind
Definition: pathnodes.h:850
struct Path * cheapest_startup_path
Definition: pathnodes.h:886
struct Path * cheapest_total_path
Definition: pathnodes.h:887
bool has_eclass_joins
Definition: pathnodes.h:974
bool consider_startup
Definition: pathnodes.h:868
Bitmapset * live_parts
Definition: pathnodes.h:1020
int rel_parallel_workers
Definition: pathnodes.h:937
bool consider_partitionwise_join
Definition: pathnodes.h:980
List * partial_pathlist
Definition: pathnodes.h:885
PlannerInfo * subroot
Definition: pathnodes.h:934
AttrNumber max_attr
Definition: pathnodes.h:911
Relids nulling_relids
Definition: pathnodes.h:919
Cardinality rows
Definition: pathnodes.h:862
AttrNumber min_attr
Definition: pathnodes.h:909
RTEKind rtekind
Definition: pathnodes.h:907
Expr * clause
Definition: pathnodes.h:2541
SetOperation op
Definition: parsenodes.h:2039
JoinType jointype
Definition: pathnodes.h:2872
Relids syn_righthand
Definition: pathnodes.h:2871
struct WindowClause * window_clause
Definition: supportnodes.h:296
Expr * expr
Definition: primnodes.h:1943
AttrNumber resno
Definition: primnodes.h:1945
Index ressortgroupref
Definition: primnodes.h:1949
bool repeatable_across_scans
Definition: tsmapi.h:65
SampleScanGetSampleSize_function SampleScanGetSampleSize
Definition: tsmapi.h:68
Definition: primnodes.h:234
AttrNumber varattno
Definition: primnodes.h:246
int varno
Definition: primnodes.h:241
Index varlevelsup
Definition: primnodes.h:266
List * partitionClause
Definition: parsenodes.h:1501
Index winref
Definition: primnodes.h:565
Oid winfnoid
Definition: primnodes.h:553
unsigned char * unsafeFlags
Definition: allpaths.c:62
Definition: regcomp.c:281
#define FirstLowInvalidHeapAttributeNumber
Definition: sysattr.h:27
TsmRoutine * GetTsmRoutine(Oid tsmhandler)
Definition: tablesample.c:27
void create_tidscan_paths(PlannerInfo *root, RelOptInfo *rel)
Definition: tidpath.c:459
List * make_tlist_from_pathtarget(PathTarget *target)
Definition: tlist.c:624
List * pull_var_clause(Node *node, int flags)
Definition: var.c:607
void pull_varattnos(Node *node, Index varno, Bitmapset **varattnos)
Definition: var.c:291