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) 1996-2023, 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/inherit.h"
39 #include "optimizer/optimizer.h"
40 #include "optimizer/pathnode.h"
41 #include "optimizer/paths.h"
42 #include "optimizer/plancat.h"
43 #include "optimizer/planner.h"
44 #include "optimizer/restrictinfo.h"
45 #include "optimizer/tlist.h"
46 #include "parser/parse_clause.h"
47 #include "parser/parsetree.h"
49 #include "partitioning/partprune.h"
50 #include "port/pg_bitutils.h"
51 #include "rewrite/rewriteManip.h"
52 #include "utils/lsyscache.h"
53 
54 
55 /* Bitmask flags for pushdown_safety_info.unsafeFlags */
56 #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
57 #define UNSAFE_HAS_SET_FUNC (1 << 1)
58 #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
59 #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
60 #define UNSAFE_TYPE_MISMATCH (1 << 4)
61 
62 /* results of subquery_is_pushdown_safe */
63 typedef struct pushdown_safety_info
64 {
65  unsigned char *unsafeFlags; /* bitmask of reasons why this target list
66  * column is unsafe for qual pushdown, or 0 if
67  * no reason. */
68  bool unsafeVolatile; /* don't push down volatile quals */
69  bool unsafeLeaky; /* don't push down leaky quals */
71 
72 /* Return type for qual_is_pushdown_safe */
73 typedef enum pushdown_safe_type
74 {
75  PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
76  PUSHDOWN_SAFE, /* safe to push qual into subquery */
77  PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
78  * run condition */
80 
81 /* These parameters are set by GUC */
82 bool enable_geqo = false; /* just in case GUC doesn't set it */
86 
87 /* Hook for plugins to get control in set_rel_pathlist() */
89 
90 /* Hook for plugins to replace standard_join_search() */
92 
93 
95 static void set_base_rel_sizes(PlannerInfo *root);
96 static void set_base_rel_pathlists(PlannerInfo *root);
97 static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
98  Index rti, RangeTblEntry *rte);
99 static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
100  Index rti, RangeTblEntry *rte);
101 static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
102  RangeTblEntry *rte);
103 static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
104 static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
105  RangeTblEntry *rte);
106 static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
107  RangeTblEntry *rte);
108 static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
109  RangeTblEntry *rte);
110 static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
111  RangeTblEntry *rte);
112 static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
113  RangeTblEntry *rte);
114 static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
115  RangeTblEntry *rte);
116 static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
117  Index rti, RangeTblEntry *rte);
118 static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
119  Index rti, RangeTblEntry *rte);
120 static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
121  List *live_childrels,
122  List *all_child_pathkeys);
124  RelOptInfo *rel,
125  Relids required_outer);
126 static void accumulate_append_subpath(Path *path,
127  List **subpaths,
128  List **special_subpaths);
129 static Path *get_singleton_append_subpath(Path *path);
130 static void set_dummy_rel_pathlist(RelOptInfo *rel);
131 static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
132  Index rti, RangeTblEntry *rte);
133 static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
134  RangeTblEntry *rte);
135 static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
136  RangeTblEntry *rte);
137 static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
138  RangeTblEntry *rte);
139 static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
140  RangeTblEntry *rte);
141 static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
142  RangeTblEntry *rte);
143 static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
144  RangeTblEntry *rte);
145 static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
146  RangeTblEntry *rte);
147 static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
148 static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
149  pushdown_safety_info *safetyInfo);
150 static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
151  pushdown_safety_info *safetyInfo);
152 static void check_output_expressions(Query *subquery,
153  pushdown_safety_info *safetyInfo);
154 static void compare_tlist_datatypes(List *tlist, List *colTypes,
155  pushdown_safety_info *safetyInfo);
156 static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
158  RestrictInfo *rinfo,
159  pushdown_safety_info *safetyInfo);
160 static void subquery_push_qual(Query *subquery,
161  RangeTblEntry *rte, Index rti, Node *qual);
162 static void recurse_push_qual(Node *setOp, Query *topquery,
163  RangeTblEntry *rte, Index rti, Node *qual);
164 static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
165  Bitmapset *extra_used_attrs);
166 
167 
168 /*
169  * make_one_rel
170  * Finds all possible access paths for executing a query, returning a
171  * single rel that represents the join of all base rels in the query.
172  */
173 RelOptInfo *
174 make_one_rel(PlannerInfo *root, List *joinlist)
175 {
176  RelOptInfo *rel;
177  Index rti;
178  double total_pages;
179 
180  /* Mark base rels as to whether we care about fast-start plans */
182 
183  /*
184  * Compute size estimates and consider_parallel flags for each base rel.
185  */
186  set_base_rel_sizes(root);
187 
188  /*
189  * We should now have size estimates for every actual table involved in
190  * the query, and we also know which if any have been deleted from the
191  * query by join removal, pruned by partition pruning, or eliminated by
192  * constraint exclusion. So we can now compute total_table_pages.
193  *
194  * Note that appendrels are not double-counted here, even though we don't
195  * bother to distinguish RelOptInfos for appendrel parents, because the
196  * parents will have pages = 0.
197  *
198  * XXX if a table is self-joined, we will count it once per appearance,
199  * which perhaps is the wrong thing ... but that's not completely clear,
200  * and detecting self-joins here is difficult, so ignore it for now.
201  */
202  total_pages = 0;
203  for (rti = 1; rti < root->simple_rel_array_size; rti++)
204  {
205  RelOptInfo *brel = root->simple_rel_array[rti];
206 
207  /* there may be empty slots corresponding to non-baserel RTEs */
208  if (brel == NULL)
209  continue;
210 
211  Assert(brel->relid == rti); /* sanity check on array */
212 
213  if (IS_DUMMY_REL(brel))
214  continue;
215 
216  if (IS_SIMPLE_REL(brel))
217  total_pages += (double) brel->pages;
218  }
219  root->total_table_pages = total_pages;
220 
221  /*
222  * Generate access paths for each base rel.
223  */
225 
226  /*
227  * Generate access paths for the entire join tree.
228  */
229  rel = make_rel_from_joinlist(root, joinlist);
230 
231  /*
232  * The result should join all and only the query's base + outer-join rels.
233  */
234  Assert(bms_equal(rel->relids, root->all_query_rels));
235 
236  return rel;
237 }
238 
239 /*
240  * set_base_rel_consider_startup
241  * Set the consider_[param_]startup flags for each base-relation entry.
242  *
243  * For the moment, we only deal with consider_param_startup here; because the
244  * logic for consider_startup is pretty trivial and is the same for every base
245  * relation, we just let build_simple_rel() initialize that flag correctly to
246  * start with. If that logic ever gets more complicated it would probably
247  * be better to move it here.
248  */
249 static void
251 {
252  /*
253  * Since parameterized paths can only be used on the inside of a nestloop
254  * join plan, there is usually little value in considering fast-start
255  * plans for them. However, for relations that are on the RHS of a SEMI
256  * or ANTI join, a fast-start plan can be useful because we're only going
257  * to care about fetching one tuple anyway.
258  *
259  * To minimize growth of planning time, we currently restrict this to
260  * cases where the RHS is a single base relation, not a join; there is no
261  * provision for consider_param_startup to get set at all on joinrels.
262  * Also we don't worry about appendrels. costsize.c's costing rules for
263  * nestloop semi/antijoins don't consider such cases either.
264  */
265  ListCell *lc;
266 
267  foreach(lc, root->join_info_list)
268  {
269  SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
270  int varno;
271 
272  if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
273  bms_get_singleton_member(sjinfo->syn_righthand, &varno))
274  {
275  RelOptInfo *rel = find_base_rel(root, varno);
276 
277  rel->consider_param_startup = true;
278  }
279  }
280 }
281 
282 /*
283  * set_base_rel_sizes
284  * Set the size estimates (rows and widths) for each base-relation entry.
285  * Also determine whether to consider parallel paths for base relations.
286  *
287  * We do this in a separate pass over the base rels so that rowcount
288  * estimates are available for parameterized path generation, and also so
289  * that each rel's consider_parallel flag is set correctly before we begin to
290  * generate paths.
291  */
292 static void
294 {
295  Index rti;
296 
297  for (rti = 1; rti < root->simple_rel_array_size; rti++)
298  {
299  RelOptInfo *rel = root->simple_rel_array[rti];
300  RangeTblEntry *rte;
301 
302  /* there may be empty slots corresponding to non-baserel RTEs */
303  if (rel == NULL)
304  continue;
305 
306  Assert(rel->relid == rti); /* sanity check on array */
307 
308  /* ignore RTEs that are "other rels" */
309  if (rel->reloptkind != RELOPT_BASEREL)
310  continue;
311 
312  rte = root->simple_rte_array[rti];
313 
314  /*
315  * If parallelism is allowable for this query in general, see whether
316  * it's allowable for this rel in particular. We have to do this
317  * before set_rel_size(), because (a) if this rel is an inheritance
318  * parent, set_append_rel_size() will use and perhaps change the rel's
319  * consider_parallel flag, and (b) for some RTE types, set_rel_size()
320  * goes ahead and makes paths immediately.
321  */
322  if (root->glob->parallelModeOK)
323  set_rel_consider_parallel(root, rel, rte);
324 
325  set_rel_size(root, rel, rti, rte);
326  }
327 }
328 
329 /*
330  * set_base_rel_pathlists
331  * Finds all paths available for scanning each base-relation entry.
332  * Sequential scan and any available indices are considered.
333  * Each useful path is attached to its relation's 'pathlist' field.
334  */
335 static void
337 {
338  Index rti;
339 
340  for (rti = 1; rti < root->simple_rel_array_size; rti++)
341  {
342  RelOptInfo *rel = root->simple_rel_array[rti];
343 
344  /* there may be empty slots corresponding to non-baserel RTEs */
345  if (rel == NULL)
346  continue;
347 
348  Assert(rel->relid == rti); /* sanity check on array */
349 
350  /* ignore RTEs that are "other rels" */
351  if (rel->reloptkind != RELOPT_BASEREL)
352  continue;
353 
354  set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
355  }
356 }
357 
358 /*
359  * set_rel_size
360  * Set size estimates for a base relation
361  */
362 static void
364  Index rti, RangeTblEntry *rte)
365 {
366  if (rel->reloptkind == RELOPT_BASEREL &&
367  relation_excluded_by_constraints(root, rel, rte))
368  {
369  /*
370  * We proved we don't need to scan the rel via constraint exclusion,
371  * so set up a single dummy path for it. Here we only check this for
372  * regular baserels; if it's an otherrel, CE was already checked in
373  * set_append_rel_size().
374  *
375  * In this case, we go ahead and set up the relation's path right away
376  * instead of leaving it for set_rel_pathlist to do. This is because
377  * we don't have a convention for marking a rel as dummy except by
378  * assigning a dummy path to it.
379  */
381  }
382  else if (rte->inh)
383  {
384  /* It's an "append relation", process accordingly */
385  set_append_rel_size(root, rel, rti, rte);
386  }
387  else
388  {
389  switch (rel->rtekind)
390  {
391  case RTE_RELATION:
392  if (rte->relkind == RELKIND_FOREIGN_TABLE)
393  {
394  /* Foreign table */
395  set_foreign_size(root, rel, rte);
396  }
397  else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
398  {
399  /*
400  * We could get here if asked to scan a partitioned table
401  * with ONLY. In that case we shouldn't scan any of the
402  * partitions, so mark it as a dummy rel.
403  */
405  }
406  else if (rte->tablesample != NULL)
407  {
408  /* Sampled relation */
409  set_tablesample_rel_size(root, rel, rte);
410  }
411  else
412  {
413  /* Plain relation */
414  set_plain_rel_size(root, rel, rte);
415  }
416  break;
417  case RTE_SUBQUERY:
418 
419  /*
420  * Subqueries don't support making a choice between
421  * parameterized and unparameterized paths, so just go ahead
422  * and build their paths immediately.
423  */
424  set_subquery_pathlist(root, rel, rti, rte);
425  break;
426  case RTE_FUNCTION:
427  set_function_size_estimates(root, rel);
428  break;
429  case RTE_TABLEFUNC:
430  set_tablefunc_size_estimates(root, rel);
431  break;
432  case RTE_VALUES:
433  set_values_size_estimates(root, rel);
434  break;
435  case RTE_CTE:
436 
437  /*
438  * CTEs don't support making a choice between parameterized
439  * and unparameterized paths, so just go ahead and build their
440  * paths immediately.
441  */
442  if (rte->self_reference)
443  set_worktable_pathlist(root, rel, rte);
444  else
445  set_cte_pathlist(root, rel, rte);
446  break;
447  case RTE_NAMEDTUPLESTORE:
448  /* Might as well just build the path immediately */
449  set_namedtuplestore_pathlist(root, rel, rte);
450  break;
451  case RTE_RESULT:
452  /* Might as well just build the path immediately */
453  set_result_pathlist(root, rel, rte);
454  break;
455  default:
456  elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
457  break;
458  }
459  }
460 
461  /*
462  * We insist that all non-dummy rels have a nonzero rowcount estimate.
463  */
464  Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
465 }
466 
467 /*
468  * set_rel_pathlist
469  * Build access paths for a base relation
470  */
471 static void
473  Index rti, RangeTblEntry *rte)
474 {
475  if (IS_DUMMY_REL(rel))
476  {
477  /* We already proved the relation empty, so nothing more to do */
478  }
479  else if (rte->inh)
480  {
481  /* It's an "append relation", process accordingly */
482  set_append_rel_pathlist(root, rel, rti, rte);
483  }
484  else
485  {
486  switch (rel->rtekind)
487  {
488  case RTE_RELATION:
489  if (rte->relkind == RELKIND_FOREIGN_TABLE)
490  {
491  /* Foreign table */
492  set_foreign_pathlist(root, rel, rte);
493  }
494  else if (rte->tablesample != NULL)
495  {
496  /* Sampled relation */
497  set_tablesample_rel_pathlist(root, rel, rte);
498  }
499  else
500  {
501  /* Plain relation */
502  set_plain_rel_pathlist(root, rel, rte);
503  }
504  break;
505  case RTE_SUBQUERY:
506  /* Subquery --- fully handled during set_rel_size */
507  break;
508  case RTE_FUNCTION:
509  /* RangeFunction */
510  set_function_pathlist(root, rel, rte);
511  break;
512  case RTE_TABLEFUNC:
513  /* Table Function */
514  set_tablefunc_pathlist(root, rel, rte);
515  break;
516  case RTE_VALUES:
517  /* Values list */
518  set_values_pathlist(root, rel, rte);
519  break;
520  case RTE_CTE:
521  /* CTE reference --- fully handled during set_rel_size */
522  break;
523  case RTE_NAMEDTUPLESTORE:
524  /* tuplestore reference --- fully handled during set_rel_size */
525  break;
526  case RTE_RESULT:
527  /* simple Result --- fully handled during set_rel_size */
528  break;
529  default:
530  elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
531  break;
532  }
533  }
534 
535  /*
536  * Allow a plugin to editorialize on the set of Paths for this base
537  * relation. It could add new paths (such as CustomPaths) by calling
538  * add_path(), or add_partial_path() if parallel aware. It could also
539  * delete or modify paths added by the core code.
540  */
542  (*set_rel_pathlist_hook) (root, rel, rti, rte);
543 
544  /*
545  * If this is a baserel, we should normally consider gathering any partial
546  * paths we may have created for it. We have to do this after calling the
547  * set_rel_pathlist_hook, else it cannot add partial paths to be included
548  * here.
549  *
550  * However, if this is an inheritance child, skip it. Otherwise, we could
551  * end up with a very large number of gather nodes, each trying to grab
552  * its own pool of workers. Instead, we'll consider gathering partial
553  * paths for the parent appendrel.
554  *
555  * Also, if this is the topmost scan/join rel, we postpone gathering until
556  * the final scan/join targetlist is available (see grouping_planner).
557  */
558  if (rel->reloptkind == RELOPT_BASEREL &&
559  !bms_equal(rel->relids, root->all_query_rels))
560  generate_useful_gather_paths(root, rel, false);
561 
562  /* Now find the cheapest of the paths for this rel */
563  set_cheapest(rel);
564 
565 #ifdef OPTIMIZER_DEBUG
566  pprint(rel);
567 #endif
568 }
569 
570 /*
571  * set_plain_rel_size
572  * Set size estimates for a plain relation (no subquery, no inheritance)
573  */
574 static void
576 {
577  /*
578  * Test any partial indexes of rel for applicability. We must do this
579  * first since partial unique indexes can affect size estimates.
580  */
581  check_index_predicates(root, rel);
582 
583  /* Mark rel with estimated output rows, width, etc */
584  set_baserel_size_estimates(root, rel);
585 }
586 
587 /*
588  * If this relation could possibly be scanned from within a worker, then set
589  * its consider_parallel flag.
590  */
591 static void
593  RangeTblEntry *rte)
594 {
595  /*
596  * The flag has previously been initialized to false, so we can just
597  * return if it becomes clear that we can't safely set it.
598  */
599  Assert(!rel->consider_parallel);
600 
601  /* Don't call this if parallelism is disallowed for the entire query. */
602  Assert(root->glob->parallelModeOK);
603 
604  /* This should only be called for baserels and appendrel children. */
605  Assert(IS_SIMPLE_REL(rel));
606 
607  /* Assorted checks based on rtekind. */
608  switch (rte->rtekind)
609  {
610  case RTE_RELATION:
611 
612  /*
613  * Currently, parallel workers can't access the leader's temporary
614  * tables. We could possibly relax this if we wrote all of its
615  * local buffers at the start of the query and made no changes
616  * thereafter (maybe we could allow hint bit changes), and if we
617  * taught the workers to read them. Writing a large number of
618  * temporary buffers could be expensive, though, and we don't have
619  * the rest of the necessary infrastructure right now anyway. So
620  * for now, bail out if we see a temporary table.
621  */
622  if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
623  return;
624 
625  /*
626  * Table sampling can be pushed down to workers if the sample
627  * function and its arguments are safe.
628  */
629  if (rte->tablesample != NULL)
630  {
631  char proparallel = func_parallel(rte->tablesample->tsmhandler);
632 
633  if (proparallel != PROPARALLEL_SAFE)
634  return;
635  if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
636  return;
637  }
638 
639  /*
640  * Ask FDWs whether they can support performing a ForeignScan
641  * within a worker. Most often, the answer will be no. For
642  * example, if the nature of the FDW is such that it opens a TCP
643  * connection with a remote server, each parallel worker would end
644  * up with a separate connection, and these connections might not
645  * be appropriately coordinated between workers and the leader.
646  */
647  if (rte->relkind == RELKIND_FOREIGN_TABLE)
648  {
649  Assert(rel->fdwroutine);
650  if (!rel->fdwroutine->IsForeignScanParallelSafe)
651  return;
652  if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
653  return;
654  }
655 
656  /*
657  * There are additional considerations for appendrels, which we'll
658  * deal with in set_append_rel_size and set_append_rel_pathlist.
659  * For now, just set consider_parallel based on the rel's own
660  * quals and targetlist.
661  */
662  break;
663 
664  case RTE_SUBQUERY:
665 
666  /*
667  * There's no intrinsic problem with scanning a subquery-in-FROM
668  * (as distinct from a SubPlan or InitPlan) in a parallel worker.
669  * If the subquery doesn't happen to have any parallel-safe paths,
670  * then flagging it as consider_parallel won't change anything,
671  * but that's true for plain tables, too. We must set
672  * consider_parallel based on the rel's own quals and targetlist,
673  * so that if a subquery path is parallel-safe but the quals and
674  * projection we're sticking onto it are not, we correctly mark
675  * the SubqueryScanPath as not parallel-safe. (Note that
676  * set_subquery_pathlist() might push some of these quals down
677  * into the subquery itself, but that doesn't change anything.)
678  *
679  * We can't push sub-select containing LIMIT/OFFSET to workers as
680  * there is no guarantee that the row order will be fully
681  * deterministic, and applying LIMIT/OFFSET will lead to
682  * inconsistent results at the top-level. (In some cases, where
683  * the result is ordered, we could relax this restriction. But it
684  * doesn't currently seem worth expending extra effort to do so.)
685  */
686  {
687  Query *subquery = castNode(Query, rte->subquery);
688 
689  if (limit_needed(subquery))
690  return;
691  }
692  break;
693 
694  case RTE_JOIN:
695  /* Shouldn't happen; we're only considering baserels here. */
696  Assert(false);
697  return;
698 
699  case RTE_FUNCTION:
700  /* Check for parallel-restricted functions. */
701  if (!is_parallel_safe(root, (Node *) rte->functions))
702  return;
703  break;
704 
705  case RTE_TABLEFUNC:
706  /* not parallel safe */
707  return;
708 
709  case RTE_VALUES:
710  /* Check for parallel-restricted functions. */
711  if (!is_parallel_safe(root, (Node *) rte->values_lists))
712  return;
713  break;
714 
715  case RTE_CTE:
716 
717  /*
718  * CTE tuplestores aren't shared among parallel workers, so we
719  * force all CTE scans to happen in the leader. Also, populating
720  * the CTE would require executing a subplan that's not available
721  * in the worker, might be parallel-restricted, and must get
722  * executed only once.
723  */
724  return;
725 
726  case RTE_NAMEDTUPLESTORE:
727 
728  /*
729  * tuplestore cannot be shared, at least without more
730  * infrastructure to support that.
731  */
732  return;
733 
734  case RTE_RESULT:
735  /* RESULT RTEs, in themselves, are no problem. */
736  break;
737  }
738 
739  /*
740  * If there's anything in baserestrictinfo that's parallel-restricted, we
741  * give up on parallelizing access to this relation. We could consider
742  * instead postponing application of the restricted quals until we're
743  * above all the parallelism in the plan tree, but it's not clear that
744  * that would be a win in very many cases, and it might be tricky to make
745  * outer join clauses work correctly. It would likely break equivalence
746  * classes, too.
747  */
748  if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
749  return;
750 
751  /*
752  * Likewise, if the relation's outputs are not parallel-safe, give up.
753  * (Usually, they're just Vars, but sometimes they're not.)
754  */
755  if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
756  return;
757 
758  /* We have a winner. */
759  rel->consider_parallel = true;
760 }
761 
762 /*
763  * set_plain_rel_pathlist
764  * Build access paths for a plain relation (no subquery, no inheritance)
765  */
766 static void
768 {
769  Relids required_outer;
770 
771  /*
772  * We don't support pushing join clauses into the quals of a seqscan, but
773  * it could still have required parameterization due to LATERAL refs in
774  * its tlist.
775  */
776  required_outer = rel->lateral_relids;
777 
778  /* Consider sequential scan */
779  add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
780 
781  /* If appropriate, consider parallel sequential scan */
782  if (rel->consider_parallel && required_outer == NULL)
783  create_plain_partial_paths(root, rel);
784 
785  /* Consider index scans */
786  create_index_paths(root, rel);
787 
788  /* Consider TID scans */
789  create_tidscan_paths(root, rel);
790 }
791 
792 /*
793  * create_plain_partial_paths
794  * Build partial access paths for parallel scan of a plain relation
795  */
796 static void
798 {
799  int parallel_workers;
800 
801  parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
803 
804  /* If any limit was set to zero, the user doesn't want a parallel scan. */
805  if (parallel_workers <= 0)
806  return;
807 
808  /* Add an unordered partial path based on a parallel sequential scan. */
809  add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
810 }
811 
812 /*
813  * set_tablesample_rel_size
814  * Set size estimates for a sampled relation
815  */
816 static void
818 {
819  TableSampleClause *tsc = rte->tablesample;
820  TsmRoutine *tsm;
821  BlockNumber pages;
822  double tuples;
823 
824  /*
825  * Test any partial indexes of rel for applicability. We must do this
826  * first since partial unique indexes can affect size estimates.
827  */
828  check_index_predicates(root, rel);
829 
830  /*
831  * Call the sampling method's estimation function to estimate the number
832  * of pages it will read and the number of tuples it will return. (Note:
833  * we assume the function returns sane values.)
834  */
835  tsm = GetTsmRoutine(tsc->tsmhandler);
836  tsm->SampleScanGetSampleSize(root, rel, tsc->args,
837  &pages, &tuples);
838 
839  /*
840  * For the moment, because we will only consider a SampleScan path for the
841  * rel, it's okay to just overwrite the pages and tuples estimates for the
842  * whole relation. If we ever consider multiple path types for sampled
843  * rels, we'll need more complication.
844  */
845  rel->pages = pages;
846  rel->tuples = tuples;
847 
848  /* Mark rel with estimated output rows, width, etc */
849  set_baserel_size_estimates(root, rel);
850 }
851 
852 /*
853  * set_tablesample_rel_pathlist
854  * Build access paths for a sampled relation
855  */
856 static void
858 {
859  Relids required_outer;
860  Path *path;
861 
862  /*
863  * We don't support pushing join clauses into the quals of a samplescan,
864  * but it could still have required parameterization due to LATERAL refs
865  * in its tlist or TABLESAMPLE arguments.
866  */
867  required_outer = rel->lateral_relids;
868 
869  /* Consider sampled scan */
870  path = create_samplescan_path(root, rel, required_outer);
871 
872  /*
873  * If the sampling method does not support repeatable scans, we must avoid
874  * plans that would scan the rel multiple times. Ideally, we'd simply
875  * avoid putting the rel on the inside of a nestloop join; but adding such
876  * a consideration to the planner seems like a great deal of complication
877  * to support an uncommon usage of second-rate sampling methods. Instead,
878  * if there is a risk that the query might perform an unsafe join, just
879  * wrap the SampleScan in a Materialize node. We can check for joins by
880  * counting the membership of all_query_rels (note that this correctly
881  * counts inheritance trees as single rels). If we're inside a subquery,
882  * we can't easily check whether a join might occur in the outer query, so
883  * just assume one is possible.
884  *
885  * GetTsmRoutine is relatively expensive compared to the other tests here,
886  * so check repeatable_across_scans last, even though that's a bit odd.
887  */
888  if ((root->query_level > 1 ||
891  {
892  path = (Path *) create_material_path(rel, path);
893  }
894 
895  add_path(rel, path);
896 
897  /* For the moment, at least, there are no other paths to consider */
898 }
899 
900 /*
901  * set_foreign_size
902  * Set size estimates for a foreign table RTE
903  */
904 static void
906 {
907  /* Mark rel with estimated output rows, width, etc */
908  set_foreign_size_estimates(root, rel);
909 
910  /* Let FDW adjust the size estimates, if it can */
911  rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
912 
913  /* ... but do not let it set the rows estimate to zero */
914  rel->rows = clamp_row_est(rel->rows);
915 
916  /*
917  * Also, make sure rel->tuples is not insane relative to rel->rows.
918  * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
919  * FDW doesn't do anything to replace that.
920  */
921  rel->tuples = Max(rel->tuples, rel->rows);
922 }
923 
924 /*
925  * set_foreign_pathlist
926  * Build access paths for a foreign table RTE
927  */
928 static void
930 {
931  /* Call the FDW's GetForeignPaths function to generate path(s) */
932  rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
933 }
934 
935 /*
936  * set_append_rel_size
937  * Set size estimates for a simple "append relation"
938  *
939  * The passed-in rel and RTE represent the entire append relation. The
940  * relation's contents are computed by appending together the output of the
941  * individual member relations. Note that in the non-partitioned inheritance
942  * case, the first member relation is actually the same table as is mentioned
943  * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
944  * a good thing because their outputs are not the same size.
945  */
946 static void
948  Index rti, RangeTblEntry *rte)
949 {
950  int parentRTindex = rti;
951  bool has_live_children;
952  double parent_rows;
953  double parent_size;
954  double *parent_attrsizes;
955  int nattrs;
956  ListCell *l;
957 
958  /* Guard against stack overflow due to overly deep inheritance tree. */
960 
961  Assert(IS_SIMPLE_REL(rel));
962 
963  /*
964  * If this is a partitioned baserel, set the consider_partitionwise_join
965  * flag; currently, we only consider partitionwise joins with the baserel
966  * if its targetlist doesn't contain a whole-row Var.
967  */
969  rel->reloptkind == RELOPT_BASEREL &&
970  rte->relkind == RELKIND_PARTITIONED_TABLE &&
971  bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
972  rel->consider_partitionwise_join = true;
973 
974  /*
975  * Initialize to compute size estimates for whole append relation.
976  *
977  * We handle width estimates by weighting the widths of different child
978  * rels proportionally to their number of rows. This is sensible because
979  * the use of width estimates is mainly to compute the total relation
980  * "footprint" if we have to sort or hash it. To do this, we sum the
981  * total equivalent size (in "double" arithmetic) and then divide by the
982  * total rowcount estimate. This is done separately for the total rel
983  * width and each attribute.
984  *
985  * Note: if you consider changing this logic, beware that child rels could
986  * have zero rows and/or width, if they were excluded by constraints.
987  */
988  has_live_children = false;
989  parent_rows = 0;
990  parent_size = 0;
991  nattrs = rel->max_attr - rel->min_attr + 1;
992  parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
993 
994  foreach(l, root->append_rel_list)
995  {
996  AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
997  int childRTindex;
998  RangeTblEntry *childRTE;
999  RelOptInfo *childrel;
1000  List *childrinfos;
1001  ListCell *parentvars;
1002  ListCell *childvars;
1003  ListCell *lc;
1004 
1005  /* append_rel_list contains all append rels; ignore others */
1006  if (appinfo->parent_relid != parentRTindex)
1007  continue;
1008 
1009  childRTindex = appinfo->child_relid;
1010  childRTE = root->simple_rte_array[childRTindex];
1011 
1012  /*
1013  * The child rel's RelOptInfo was already created during
1014  * add_other_rels_to_query.
1015  */
1016  childrel = find_base_rel(root, childRTindex);
1018 
1019  /* We may have already proven the child to be dummy. */
1020  if (IS_DUMMY_REL(childrel))
1021  continue;
1022 
1023  /*
1024  * We have to copy the parent's targetlist and quals to the child,
1025  * with appropriate substitution of variables. However, the
1026  * baserestrictinfo quals were already copied/substituted when the
1027  * child RelOptInfo was built. So we don't need any additional setup
1028  * before applying constraint exclusion.
1029  */
1030  if (relation_excluded_by_constraints(root, childrel, childRTE))
1031  {
1032  /*
1033  * This child need not be scanned, so we can omit it from the
1034  * appendrel.
1035  */
1036  set_dummy_rel_pathlist(childrel);
1037  continue;
1038  }
1039 
1040  /*
1041  * Constraint exclusion failed, so copy the parent's join quals and
1042  * targetlist to the child, with appropriate variable substitutions.
1043  *
1044  * We skip join quals that came from above outer joins that can null
1045  * this rel, since they would be of no value while generating paths
1046  * for the child. This saves some effort while processing the child
1047  * rel, and it also avoids an implementation restriction in
1048  * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1049  */
1050  childrinfos = NIL;
1051  foreach(lc, rel->joininfo)
1052  {
1053  RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1054 
1055  if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1056  childrinfos = lappend(childrinfos,
1058  (Node *) rinfo,
1059  1, &appinfo));
1060  }
1061  childrel->joininfo = childrinfos;
1062 
1063  /*
1064  * Now for the child's targetlist.
1065  *
1066  * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1067  * expressions, which otherwise would not occur in a rel's targetlist.
1068  * Code that might be looking at an appendrel child must cope with
1069  * such. (Normally, a rel's targetlist would only include Vars and
1070  * PlaceHolderVars.) XXX we do not bother to update the cost or width
1071  * fields of childrel->reltarget; not clear if that would be useful.
1072  */
1073  childrel->reltarget->exprs = (List *)
1075  (Node *) rel->reltarget->exprs,
1076  1, &appinfo);
1077 
1078  /*
1079  * We have to make child entries in the EquivalenceClass data
1080  * structures as well. This is needed either if the parent
1081  * participates in some eclass joins (because we will want to consider
1082  * inner-indexscan joins on the individual children) or if the parent
1083  * has useful pathkeys (because we should try to build MergeAppend
1084  * paths that produce those sort orderings).
1085  */
1086  if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1087  add_child_rel_equivalences(root, appinfo, rel, childrel);
1088  childrel->has_eclass_joins = rel->has_eclass_joins;
1089 
1090  /*
1091  * Note: we could compute appropriate attr_needed data for the child's
1092  * variables, by transforming the parent's attr_needed through the
1093  * translated_vars mapping. However, currently there's no need
1094  * because attr_needed is only examined for base relations not
1095  * otherrels. So we just leave the child's attr_needed empty.
1096  */
1097 
1098  /*
1099  * If we consider partitionwise joins with the parent rel, do the same
1100  * for partitioned child rels.
1101  *
1102  * Note: here we abuse the consider_partitionwise_join flag by setting
1103  * it for child rels that are not themselves partitioned. We do so to
1104  * tell try_partitionwise_join() that the child rel is sufficiently
1105  * valid to be used as a per-partition input, even if it later gets
1106  * proven to be dummy. (It's not usable until we've set up the
1107  * reltarget and EC entries, which we just did.)
1108  */
1109  if (rel->consider_partitionwise_join)
1110  childrel->consider_partitionwise_join = true;
1111 
1112  /*
1113  * If parallelism is allowable for this query in general, see whether
1114  * it's allowable for this childrel in particular. But if we've
1115  * already decided the appendrel is not parallel-safe as a whole,
1116  * there's no point in considering parallelism for this child. For
1117  * consistency, do this before calling set_rel_size() for the child.
1118  */
1119  if (root->glob->parallelModeOK && rel->consider_parallel)
1120  set_rel_consider_parallel(root, childrel, childRTE);
1121 
1122  /*
1123  * Compute the child's size.
1124  */
1125  set_rel_size(root, childrel, childRTindex, childRTE);
1126 
1127  /*
1128  * It is possible that constraint exclusion detected a contradiction
1129  * within a child subquery, even though we didn't prove one above. If
1130  * so, we can skip this child.
1131  */
1132  if (IS_DUMMY_REL(childrel))
1133  continue;
1134 
1135  /* We have at least one live child. */
1136  has_live_children = true;
1137 
1138  /*
1139  * If any live child is not parallel-safe, treat the whole appendrel
1140  * as not parallel-safe. In future we might be able to generate plans
1141  * in which some children are farmed out to workers while others are
1142  * not; but we don't have that today, so it's a waste to consider
1143  * partial paths anywhere in the appendrel unless it's all safe.
1144  * (Child rels visited before this one will be unmarked in
1145  * set_append_rel_pathlist().)
1146  */
1147  if (!childrel->consider_parallel)
1148  rel->consider_parallel = false;
1149 
1150  /*
1151  * Accumulate size information from each live child.
1152  */
1153  Assert(childrel->rows > 0);
1154 
1155  parent_rows += childrel->rows;
1156  parent_size += childrel->reltarget->width * childrel->rows;
1157 
1158  /*
1159  * Accumulate per-column estimates too. We need not do anything for
1160  * PlaceHolderVars in the parent list. If child expression isn't a
1161  * Var, or we didn't record a width estimate for it, we have to fall
1162  * back on a datatype-based estimate.
1163  *
1164  * By construction, child's targetlist is 1-to-1 with parent's.
1165  */
1166  forboth(parentvars, rel->reltarget->exprs,
1167  childvars, childrel->reltarget->exprs)
1168  {
1169  Var *parentvar = (Var *) lfirst(parentvars);
1170  Node *childvar = (Node *) lfirst(childvars);
1171 
1172  if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1173  {
1174  int pndx = parentvar->varattno - rel->min_attr;
1175  int32 child_width = 0;
1176 
1177  if (IsA(childvar, Var) &&
1178  ((Var *) childvar)->varno == childrel->relid)
1179  {
1180  int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1181 
1182  child_width = childrel->attr_widths[cndx];
1183  }
1184  if (child_width <= 0)
1185  child_width = get_typavgwidth(exprType(childvar),
1186  exprTypmod(childvar));
1187  Assert(child_width > 0);
1188  parent_attrsizes[pndx] += child_width * childrel->rows;
1189  }
1190  }
1191  }
1192 
1193  if (has_live_children)
1194  {
1195  /*
1196  * Save the finished size estimates.
1197  */
1198  int i;
1199 
1200  Assert(parent_rows > 0);
1201  rel->rows = parent_rows;
1202  rel->reltarget->width = rint(parent_size / parent_rows);
1203  for (i = 0; i < nattrs; i++)
1204  rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1205 
1206  /*
1207  * Set "raw tuples" count equal to "rows" for the appendrel; needed
1208  * because some places assume rel->tuples is valid for any baserel.
1209  */
1210  rel->tuples = parent_rows;
1211 
1212  /*
1213  * Note that we leave rel->pages as zero; this is important to avoid
1214  * double-counting the appendrel tree in total_table_pages.
1215  */
1216  }
1217  else
1218  {
1219  /*
1220  * All children were excluded by constraints, so mark the whole
1221  * appendrel dummy. We must do this in this phase so that the rel's
1222  * dummy-ness is visible when we generate paths for other rels.
1223  */
1225  }
1226 
1227  pfree(parent_attrsizes);
1228 }
1229 
1230 /*
1231  * set_append_rel_pathlist
1232  * Build access paths for an "append relation"
1233  */
1234 static void
1236  Index rti, RangeTblEntry *rte)
1237 {
1238  int parentRTindex = rti;
1239  List *live_childrels = NIL;
1240  ListCell *l;
1241 
1242  /*
1243  * Generate access paths for each member relation, and remember the
1244  * non-dummy children.
1245  */
1246  foreach(l, root->append_rel_list)
1247  {
1248  AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1249  int childRTindex;
1250  RangeTblEntry *childRTE;
1251  RelOptInfo *childrel;
1252 
1253  /* append_rel_list contains all append rels; ignore others */
1254  if (appinfo->parent_relid != parentRTindex)
1255  continue;
1256 
1257  /* Re-locate the child RTE and RelOptInfo */
1258  childRTindex = appinfo->child_relid;
1259  childRTE = root->simple_rte_array[childRTindex];
1260  childrel = root->simple_rel_array[childRTindex];
1261 
1262  /*
1263  * If set_append_rel_size() decided the parent appendrel was
1264  * parallel-unsafe at some point after visiting this child rel, we
1265  * need to propagate the unsafety marking down to the child, so that
1266  * we don't generate useless partial paths for it.
1267  */
1268  if (!rel->consider_parallel)
1269  childrel->consider_parallel = false;
1270 
1271  /*
1272  * Compute the child's access paths.
1273  */
1274  set_rel_pathlist(root, childrel, childRTindex, childRTE);
1275 
1276  /*
1277  * If child is dummy, ignore it.
1278  */
1279  if (IS_DUMMY_REL(childrel))
1280  continue;
1281 
1282  /*
1283  * Child is live, so add it to the live_childrels list for use below.
1284  */
1285  live_childrels = lappend(live_childrels, childrel);
1286  }
1287 
1288  /* Add paths to the append relation. */
1289  add_paths_to_append_rel(root, rel, live_childrels);
1290 }
1291 
1292 
1293 /*
1294  * add_paths_to_append_rel
1295  * Generate paths for the given append relation given the set of non-dummy
1296  * child rels.
1297  *
1298  * The function collects all parameterizations and orderings supported by the
1299  * non-dummy children. For every such parameterization or ordering, it creates
1300  * an append path collecting one path from each non-dummy child with given
1301  * parameterization or ordering. Similarly it collects partial paths from
1302  * non-dummy children to create partial append paths.
1303  */
1304 void
1306  List *live_childrels)
1307 {
1308  List *subpaths = NIL;
1309  bool subpaths_valid = true;
1310  List *startup_subpaths = NIL;
1311  bool startup_subpaths_valid = true;
1312  List *partial_subpaths = NIL;
1313  List *pa_partial_subpaths = NIL;
1314  List *pa_nonpartial_subpaths = NIL;
1315  bool partial_subpaths_valid = true;
1316  bool pa_subpaths_valid;
1317  List *all_child_pathkeys = NIL;
1318  List *all_child_outers = NIL;
1319  ListCell *l;
1320  double partial_rows = -1;
1321 
1322  /* If appropriate, consider parallel append */
1323  pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1324 
1325  /*
1326  * For every non-dummy child, remember the cheapest path. Also, identify
1327  * all pathkeys (orderings) and parameterizations (required_outer sets)
1328  * available for the non-dummy member relations.
1329  */
1330  foreach(l, live_childrels)
1331  {
1332  RelOptInfo *childrel = lfirst(l);
1333  ListCell *lcp;
1334  Path *cheapest_partial_path = NULL;
1335 
1336  /*
1337  * If child has an unparameterized cheapest-total path, add that to
1338  * the unparameterized Append path we are constructing for the parent.
1339  * If not, there's no workable unparameterized path.
1340  *
1341  * With partitionwise aggregates, the child rel's pathlist may be
1342  * empty, so don't assume that a path exists here.
1343  */
1344  if (childrel->pathlist != NIL &&
1345  childrel->cheapest_total_path->param_info == NULL)
1347  &subpaths, NULL);
1348  else
1349  subpaths_valid = false;
1350 
1351  /*
1352  * When the planner is considering cheap startup plans, we'll also
1353  * collect all the cheapest_startup_paths (if set) and build an
1354  * AppendPath containing those as subpaths.
1355  */
1356  if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1357  {
1358  /* cheapest_startup_path must not be a parameterized path. */
1359  Assert(childrel->cheapest_startup_path->param_info == NULL);
1361  &startup_subpaths,
1362  NULL);
1363  }
1364  else
1365  startup_subpaths_valid = false;
1366 
1367 
1368  /* Same idea, but for a partial plan. */
1369  if (childrel->partial_pathlist != NIL)
1370  {
1371  cheapest_partial_path = linitial(childrel->partial_pathlist);
1372  accumulate_append_subpath(cheapest_partial_path,
1373  &partial_subpaths, NULL);
1374  }
1375  else
1376  partial_subpaths_valid = false;
1377 
1378  /*
1379  * Same idea, but for a parallel append mixing partial and non-partial
1380  * paths.
1381  */
1382  if (pa_subpaths_valid)
1383  {
1384  Path *nppath = NULL;
1385 
1386  nppath =
1388 
1389  if (cheapest_partial_path == NULL && nppath == NULL)
1390  {
1391  /* Neither a partial nor a parallel-safe path? Forget it. */
1392  pa_subpaths_valid = false;
1393  }
1394  else if (nppath == NULL ||
1395  (cheapest_partial_path != NULL &&
1396  cheapest_partial_path->total_cost < nppath->total_cost))
1397  {
1398  /* Partial path is cheaper or the only option. */
1399  Assert(cheapest_partial_path != NULL);
1400  accumulate_append_subpath(cheapest_partial_path,
1401  &pa_partial_subpaths,
1402  &pa_nonpartial_subpaths);
1403  }
1404  else
1405  {
1406  /*
1407  * Either we've got only a non-partial path, or we think that
1408  * a single backend can execute the best non-partial path
1409  * faster than all the parallel backends working together can
1410  * execute the best partial path.
1411  *
1412  * It might make sense to be more aggressive here. Even if
1413  * the best non-partial path is more expensive than the best
1414  * partial path, it could still be better to choose the
1415  * non-partial path if there are several such paths that can
1416  * be given to different workers. For now, we don't try to
1417  * figure that out.
1418  */
1420  &pa_nonpartial_subpaths,
1421  NULL);
1422  }
1423  }
1424 
1425  /*
1426  * Collect lists of all the available path orderings and
1427  * parameterizations for all the children. We use these as a
1428  * heuristic to indicate which sort orderings and parameterizations we
1429  * should build Append and MergeAppend paths for.
1430  */
1431  foreach(lcp, childrel->pathlist)
1432  {
1433  Path *childpath = (Path *) lfirst(lcp);
1434  List *childkeys = childpath->pathkeys;
1435  Relids childouter = PATH_REQ_OUTER(childpath);
1436 
1437  /* Unsorted paths don't contribute to pathkey list */
1438  if (childkeys != NIL)
1439  {
1440  ListCell *lpk;
1441  bool found = false;
1442 
1443  /* Have we already seen this ordering? */
1444  foreach(lpk, all_child_pathkeys)
1445  {
1446  List *existing_pathkeys = (List *) lfirst(lpk);
1447 
1448  if (compare_pathkeys(existing_pathkeys,
1449  childkeys) == PATHKEYS_EQUAL)
1450  {
1451  found = true;
1452  break;
1453  }
1454  }
1455  if (!found)
1456  {
1457  /* No, so add it to all_child_pathkeys */
1458  all_child_pathkeys = lappend(all_child_pathkeys,
1459  childkeys);
1460  }
1461  }
1462 
1463  /* Unparameterized paths don't contribute to param-set list */
1464  if (childouter)
1465  {
1466  ListCell *lco;
1467  bool found = false;
1468 
1469  /* Have we already seen this param set? */
1470  foreach(lco, all_child_outers)
1471  {
1472  Relids existing_outers = (Relids) lfirst(lco);
1473 
1474  if (bms_equal(existing_outers, childouter))
1475  {
1476  found = true;
1477  break;
1478  }
1479  }
1480  if (!found)
1481  {
1482  /* No, so add it to all_child_outers */
1483  all_child_outers = lappend(all_child_outers,
1484  childouter);
1485  }
1486  }
1487  }
1488  }
1489 
1490  /*
1491  * If we found unparameterized paths for all children, build an unordered,
1492  * unparameterized Append path for the rel. (Note: this is correct even
1493  * if we have zero or one live subpath due to constraint exclusion.)
1494  */
1495  if (subpaths_valid)
1496  add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1497  NIL, NULL, 0, false,
1498  -1));
1499 
1500  /* build an AppendPath for the cheap startup paths, if valid */
1501  if (startup_subpaths_valid)
1502  add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1503  NIL, NIL, NULL, 0, false, -1));
1504 
1505  /*
1506  * Consider an append of unordered, unparameterized partial paths. Make
1507  * it parallel-aware if possible.
1508  */
1509  if (partial_subpaths_valid && partial_subpaths != NIL)
1510  {
1511  AppendPath *appendpath;
1512  ListCell *lc;
1513  int parallel_workers = 0;
1514 
1515  /* Find the highest number of workers requested for any subpath. */
1516  foreach(lc, partial_subpaths)
1517  {
1518  Path *path = lfirst(lc);
1519 
1520  parallel_workers = Max(parallel_workers, path->parallel_workers);
1521  }
1522  Assert(parallel_workers > 0);
1523 
1524  /*
1525  * If the use of parallel append is permitted, always request at least
1526  * log2(# of children) workers. We assume it can be useful to have
1527  * extra workers in this case because they will be spread out across
1528  * the children. The precise formula is just a guess, but we don't
1529  * want to end up with a radically different answer for a table with N
1530  * partitions vs. an unpartitioned table with the same data, so the
1531  * use of some kind of log-scaling here seems to make some sense.
1532  */
1534  {
1535  parallel_workers = Max(parallel_workers,
1536  pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1537  parallel_workers = Min(parallel_workers,
1539  }
1540  Assert(parallel_workers > 0);
1541 
1542  /* Generate a partial append path. */
1543  appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1544  NIL, NULL, parallel_workers,
1546  -1);
1547 
1548  /*
1549  * Make sure any subsequent partial paths use the same row count
1550  * estimate.
1551  */
1552  partial_rows = appendpath->path.rows;
1553 
1554  /* Add the path. */
1555  add_partial_path(rel, (Path *) appendpath);
1556  }
1557 
1558  /*
1559  * Consider a parallel-aware append using a mix of partial and non-partial
1560  * paths. (This only makes sense if there's at least one child which has
1561  * a non-partial path that is substantially cheaper than any partial path;
1562  * otherwise, we should use the append path added in the previous step.)
1563  */
1564  if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1565  {
1566  AppendPath *appendpath;
1567  ListCell *lc;
1568  int parallel_workers = 0;
1569 
1570  /*
1571  * Find the highest number of workers requested for any partial
1572  * subpath.
1573  */
1574  foreach(lc, pa_partial_subpaths)
1575  {
1576  Path *path = lfirst(lc);
1577 
1578  parallel_workers = Max(parallel_workers, path->parallel_workers);
1579  }
1580 
1581  /*
1582  * Same formula here as above. It's even more important in this
1583  * instance because the non-partial paths won't contribute anything to
1584  * the planned number of parallel workers.
1585  */
1586  parallel_workers = Max(parallel_workers,
1587  pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1588  parallel_workers = Min(parallel_workers,
1590  Assert(parallel_workers > 0);
1591 
1592  appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1593  pa_partial_subpaths,
1594  NIL, NULL, parallel_workers, true,
1595  partial_rows);
1596  add_partial_path(rel, (Path *) appendpath);
1597  }
1598 
1599  /*
1600  * Also build unparameterized ordered append paths based on the collected
1601  * list of child pathkeys.
1602  */
1603  if (subpaths_valid)
1604  generate_orderedappend_paths(root, rel, live_childrels,
1605  all_child_pathkeys);
1606 
1607  /*
1608  * Build Append paths for each parameterization seen among the child rels.
1609  * (This may look pretty expensive, but in most cases of practical
1610  * interest, the child rels will expose mostly the same parameterizations,
1611  * so that not that many cases actually get considered here.)
1612  *
1613  * The Append node itself cannot enforce quals, so all qual checking must
1614  * be done in the child paths. This means that to have a parameterized
1615  * Append path, we must have the exact same parameterization for each
1616  * child path; otherwise some children might be failing to check the
1617  * moved-down quals. To make them match up, we can try to increase the
1618  * parameterization of lesser-parameterized paths.
1619  */
1620  foreach(l, all_child_outers)
1621  {
1622  Relids required_outer = (Relids) lfirst(l);
1623  ListCell *lcr;
1624 
1625  /* Select the child paths for an Append with this parameterization */
1626  subpaths = NIL;
1627  subpaths_valid = true;
1628  foreach(lcr, live_childrels)
1629  {
1630  RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1631  Path *subpath;
1632 
1633  if (childrel->pathlist == NIL)
1634  {
1635  /* failed to make a suitable path for this child */
1636  subpaths_valid = false;
1637  break;
1638  }
1639 
1641  childrel,
1642  required_outer);
1643  if (subpath == NULL)
1644  {
1645  /* failed to make a suitable path for this child */
1646  subpaths_valid = false;
1647  break;
1648  }
1649  accumulate_append_subpath(subpath, &subpaths, NULL);
1650  }
1651 
1652  if (subpaths_valid)
1653  add_path(rel, (Path *)
1654  create_append_path(root, rel, subpaths, NIL,
1655  NIL, required_outer, 0, false,
1656  -1));
1657  }
1658 
1659  /*
1660  * When there is only a single child relation, the Append path can inherit
1661  * any ordering available for the child rel's path, so that it's useful to
1662  * consider ordered partial paths. Above we only considered the cheapest
1663  * partial path for each child, but let's also make paths using any
1664  * partial paths that have pathkeys.
1665  */
1666  if (list_length(live_childrels) == 1)
1667  {
1668  RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1669 
1670  /* skip the cheapest partial path, since we already used that above */
1671  for_each_from(l, childrel->partial_pathlist, 1)
1672  {
1673  Path *path = (Path *) lfirst(l);
1674  AppendPath *appendpath;
1675 
1676  /* skip paths with no pathkeys. */
1677  if (path->pathkeys == NIL)
1678  continue;
1679 
1680  appendpath = create_append_path(root, rel, NIL, list_make1(path),
1681  NIL, NULL,
1682  path->parallel_workers, true,
1683  partial_rows);
1684  add_partial_path(rel, (Path *) appendpath);
1685  }
1686  }
1687 }
1688 
1689 /*
1690  * generate_orderedappend_paths
1691  * Generate ordered append paths for an append relation
1692  *
1693  * Usually we generate MergeAppend paths here, but there are some special
1694  * cases where we can generate simple Append paths, because the subpaths
1695  * can provide tuples in the required order already.
1696  *
1697  * We generate a path for each ordering (pathkey list) appearing in
1698  * all_child_pathkeys.
1699  *
1700  * We consider both cheapest-startup and cheapest-total cases, ie, for each
1701  * interesting ordering, collect all the cheapest startup subpaths and all the
1702  * cheapest total paths, and build a suitable path for each case.
1703  *
1704  * We don't currently generate any parameterized ordered paths here. While
1705  * it would not take much more code here to do so, it's very unclear that it
1706  * is worth the planning cycles to investigate such paths: there's little
1707  * use for an ordered path on the inside of a nestloop. In fact, it's likely
1708  * that the current coding of add_path would reject such paths out of hand,
1709  * because add_path gives no credit for sort ordering of parameterized paths,
1710  * and a parameterized MergeAppend is going to be more expensive than the
1711  * corresponding parameterized Append path. If we ever try harder to support
1712  * parameterized mergejoin plans, it might be worth adding support for
1713  * parameterized paths here to feed such joins. (See notes in
1714  * optimizer/README for why that might not ever happen, though.)
1715  */
1716 static void
1718  List *live_childrels,
1719  List *all_child_pathkeys)
1720 {
1721  ListCell *lcp;
1722  List *partition_pathkeys = NIL;
1723  List *partition_pathkeys_desc = NIL;
1724  bool partition_pathkeys_partial = true;
1725  bool partition_pathkeys_desc_partial = true;
1726 
1727  /*
1728  * Some partitioned table setups may allow us to use an Append node
1729  * instead of a MergeAppend. This is possible in cases such as RANGE
1730  * partitioned tables where it's guaranteed that an earlier partition must
1731  * contain rows which come earlier in the sort order. To detect whether
1732  * this is relevant, build pathkey descriptions of the partition ordering,
1733  * for both forward and reverse scans.
1734  */
1735  if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
1736  partitions_are_ordered(rel->boundinfo, rel->live_parts))
1737  {
1738  partition_pathkeys = build_partition_pathkeys(root, rel,
1740  &partition_pathkeys_partial);
1741 
1742  partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1744  &partition_pathkeys_desc_partial);
1745 
1746  /*
1747  * You might think we should truncate_useless_pathkeys here, but
1748  * allowing partition keys which are a subset of the query's pathkeys
1749  * can often be useful. For example, consider a table partitioned by
1750  * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1751  * paths that can produce the a, b, c ordering (perhaps via indexes on
1752  * (a, b, c)) then it works to consider the appendrel output as
1753  * ordered by a, b, c.
1754  */
1755  }
1756 
1757  /* Now consider each interesting sort ordering */
1758  foreach(lcp, all_child_pathkeys)
1759  {
1760  List *pathkeys = (List *) lfirst(lcp);
1761  List *startup_subpaths = NIL;
1762  List *total_subpaths = NIL;
1763  List *fractional_subpaths = NIL;
1764  bool startup_neq_total = false;
1765  bool match_partition_order;
1766  bool match_partition_order_desc;
1767  int end_index;
1768  int first_index;
1769  int direction;
1770 
1771  /*
1772  * Determine if this sort ordering matches any partition pathkeys we
1773  * have, for both ascending and descending partition order. If the
1774  * partition pathkeys happen to be contained in pathkeys then it still
1775  * works, as described above, providing that the partition pathkeys
1776  * are complete and not just a prefix of the partition keys. (In such
1777  * cases we'll be relying on the child paths to have sorted the
1778  * lower-order columns of the required pathkeys.)
1779  */
1780  match_partition_order =
1781  pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1782  (!partition_pathkeys_partial &&
1783  pathkeys_contained_in(partition_pathkeys, pathkeys));
1784 
1785  match_partition_order_desc = !match_partition_order &&
1786  (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1787  (!partition_pathkeys_desc_partial &&
1788  pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1789 
1790  /*
1791  * When the required pathkeys match the reverse of the partition
1792  * order, we must build the list of paths in reverse starting with the
1793  * last matching partition first. We can get away without making any
1794  * special cases for this in the loop below by just looping backward
1795  * over the child relations in this case.
1796  */
1797  if (match_partition_order_desc)
1798  {
1799  /* loop backward */
1800  first_index = list_length(live_childrels) - 1;
1801  end_index = -1;
1802  direction = -1;
1803 
1804  /*
1805  * Set this to true to save us having to check for
1806  * match_partition_order_desc in the loop below.
1807  */
1808  match_partition_order = true;
1809  }
1810  else
1811  {
1812  /* for all other case, loop forward */
1813  first_index = 0;
1814  end_index = list_length(live_childrels);
1815  direction = 1;
1816  }
1817 
1818  /* Select the child paths for this ordering... */
1819  for (int i = first_index; i != end_index; i += direction)
1820  {
1821  RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1822  Path *cheapest_startup,
1823  *cheapest_total,
1824  *cheapest_fractional = NULL;
1825 
1826  /* Locate the right paths, if they are available. */
1827  cheapest_startup =
1829  pathkeys,
1830  NULL,
1831  STARTUP_COST,
1832  false);
1833  cheapest_total =
1835  pathkeys,
1836  NULL,
1837  TOTAL_COST,
1838  false);
1839 
1840  /*
1841  * If we can't find any paths with the right order just use the
1842  * cheapest-total path; we'll have to sort it later.
1843  */
1844  if (cheapest_startup == NULL || cheapest_total == NULL)
1845  {
1846  cheapest_startup = cheapest_total =
1847  childrel->cheapest_total_path;
1848  /* Assert we do have an unparameterized path for this child */
1849  Assert(cheapest_total->param_info == NULL);
1850  }
1851 
1852  /*
1853  * When building a fractional path, determine a cheapest
1854  * fractional path for each child relation too. Looking at startup
1855  * and total costs is not enough, because the cheapest fractional
1856  * path may be dominated by two separate paths (one for startup,
1857  * one for total).
1858  *
1859  * When needed (building fractional path), determine the cheapest
1860  * fractional path too.
1861  */
1862  if (root->tuple_fraction > 0)
1863  {
1864  double path_fraction = (1.0 / root->tuple_fraction);
1865 
1866  cheapest_fractional =
1868  pathkeys,
1869  NULL,
1870  path_fraction);
1871 
1872  /*
1873  * If we found no path with matching pathkeys, use the
1874  * cheapest total path instead.
1875  *
1876  * XXX We might consider partially sorted paths too (with an
1877  * incremental sort on top). But we'd have to build all the
1878  * incremental paths, do the costing etc.
1879  */
1880  if (!cheapest_fractional)
1881  cheapest_fractional = cheapest_total;
1882  }
1883 
1884  /*
1885  * Notice whether we actually have different paths for the
1886  * "cheapest" and "total" cases; frequently there will be no point
1887  * in two create_merge_append_path() calls.
1888  */
1889  if (cheapest_startup != cheapest_total)
1890  startup_neq_total = true;
1891 
1892  /*
1893  * Collect the appropriate child paths. The required logic varies
1894  * for the Append and MergeAppend cases.
1895  */
1896  if (match_partition_order)
1897  {
1898  /*
1899  * We're going to make a plain Append path. We don't need
1900  * most of what accumulate_append_subpath would do, but we do
1901  * want to cut out child Appends or MergeAppends if they have
1902  * just a single subpath (and hence aren't doing anything
1903  * useful).
1904  */
1905  cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1906  cheapest_total = get_singleton_append_subpath(cheapest_total);
1907 
1908  startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1909  total_subpaths = lappend(total_subpaths, cheapest_total);
1910 
1911  if (cheapest_fractional)
1912  {
1913  cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1914  fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1915  }
1916  }
1917  else
1918  {
1919  /*
1920  * Otherwise, rely on accumulate_append_subpath to collect the
1921  * child paths for the MergeAppend.
1922  */
1923  accumulate_append_subpath(cheapest_startup,
1924  &startup_subpaths, NULL);
1925  accumulate_append_subpath(cheapest_total,
1926  &total_subpaths, NULL);
1927 
1928  if (cheapest_fractional)
1929  accumulate_append_subpath(cheapest_fractional,
1930  &fractional_subpaths, NULL);
1931  }
1932  }
1933 
1934  /* ... and build the Append or MergeAppend paths */
1935  if (match_partition_order)
1936  {
1937  /* We only need Append */
1938  add_path(rel, (Path *) create_append_path(root,
1939  rel,
1940  startup_subpaths,
1941  NIL,
1942  pathkeys,
1943  NULL,
1944  0,
1945  false,
1946  -1));
1947  if (startup_neq_total)
1948  add_path(rel, (Path *) create_append_path(root,
1949  rel,
1950  total_subpaths,
1951  NIL,
1952  pathkeys,
1953  NULL,
1954  0,
1955  false,
1956  -1));
1957 
1958  if (fractional_subpaths)
1959  add_path(rel, (Path *) create_append_path(root,
1960  rel,
1961  fractional_subpaths,
1962  NIL,
1963  pathkeys,
1964  NULL,
1965  0,
1966  false,
1967  -1));
1968  }
1969  else
1970  {
1971  /* We need MergeAppend */
1972  add_path(rel, (Path *) create_merge_append_path(root,
1973  rel,
1974  startup_subpaths,
1975  pathkeys,
1976  NULL));
1977  if (startup_neq_total)
1978  add_path(rel, (Path *) create_merge_append_path(root,
1979  rel,
1980  total_subpaths,
1981  pathkeys,
1982  NULL));
1983 
1984  if (fractional_subpaths)
1985  add_path(rel, (Path *) create_merge_append_path(root,
1986  rel,
1987  fractional_subpaths,
1988  pathkeys,
1989  NULL));
1990  }
1991  }
1992 }
1993 
1994 /*
1995  * get_cheapest_parameterized_child_path
1996  * Get cheapest path for this relation that has exactly the requested
1997  * parameterization.
1998  *
1999  * Returns NULL if unable to create such a path.
2000  */
2001 static Path *
2003  Relids required_outer)
2004 {
2005  Path *cheapest;
2006  ListCell *lc;
2007 
2008  /*
2009  * Look up the cheapest existing path with no more than the needed
2010  * parameterization. If it has exactly the needed parameterization, we're
2011  * done.
2012  */
2013  cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2014  NIL,
2015  required_outer,
2016  TOTAL_COST,
2017  false);
2018  Assert(cheapest != NULL);
2019  if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2020  return cheapest;
2021 
2022  /*
2023  * Otherwise, we can "reparameterize" an existing path to match the given
2024  * parameterization, which effectively means pushing down additional
2025  * joinquals to be checked within the path's scan. However, some existing
2026  * paths might check the available joinquals already while others don't;
2027  * therefore, it's not clear which existing path will be cheapest after
2028  * reparameterization. We have to go through them all and find out.
2029  */
2030  cheapest = NULL;
2031  foreach(lc, rel->pathlist)
2032  {
2033  Path *path = (Path *) lfirst(lc);
2034 
2035  /* Can't use it if it needs more than requested parameterization */
2036  if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2037  continue;
2038 
2039  /*
2040  * Reparameterization can only increase the path's cost, so if it's
2041  * already more expensive than the current cheapest, forget it.
2042  */
2043  if (cheapest != NULL &&
2044  compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2045  continue;
2046 
2047  /* Reparameterize if needed, then recheck cost */
2048  if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2049  {
2050  path = reparameterize_path(root, path, required_outer, 1.0);
2051  if (path == NULL)
2052  continue; /* failed to reparameterize this one */
2053  Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2054 
2055  if (cheapest != NULL &&
2056  compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2057  continue;
2058  }
2059 
2060  /* We have a new best path */
2061  cheapest = path;
2062  }
2063 
2064  /* Return the best path, or NULL if we found no suitable candidate */
2065  return cheapest;
2066 }
2067 
2068 /*
2069  * accumulate_append_subpath
2070  * Add a subpath to the list being built for an Append or MergeAppend.
2071  *
2072  * It's possible that the child is itself an Append or MergeAppend path, in
2073  * which case we can "cut out the middleman" and just add its child paths to
2074  * our own list. (We don't try to do this earlier because we need to apply
2075  * both levels of transformation to the quals.)
2076  *
2077  * Note that if we omit a child MergeAppend in this way, we are effectively
2078  * omitting a sort step, which seems fine: if the parent is to be an Append,
2079  * its result would be unsorted anyway, while if the parent is to be a
2080  * MergeAppend, there's no point in a separate sort on a child.
2081  *
2082  * Normally, either path is a partial path and subpaths is a list of partial
2083  * paths, or else path is a non-partial plan and subpaths is a list of those.
2084  * However, if path is a parallel-aware Append, then we add its partial path
2085  * children to subpaths and the rest to special_subpaths. If the latter is
2086  * NULL, we don't flatten the path at all (unless it contains only partial
2087  * paths).
2088  */
2089 static void
2090 accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2091 {
2092  if (IsA(path, AppendPath))
2093  {
2094  AppendPath *apath = (AppendPath *) path;
2095 
2096  if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2097  {
2098  *subpaths = list_concat(*subpaths, apath->subpaths);
2099  return;
2100  }
2101  else if (special_subpaths != NULL)
2102  {
2103  List *new_special_subpaths;
2104 
2105  /* Split Parallel Append into partial and non-partial subpaths */
2106  *subpaths = list_concat(*subpaths,
2107  list_copy_tail(apath->subpaths,
2108  apath->first_partial_path));
2109  new_special_subpaths = list_copy_head(apath->subpaths,
2110  apath->first_partial_path);
2111  *special_subpaths = list_concat(*special_subpaths,
2112  new_special_subpaths);
2113  return;
2114  }
2115  }
2116  else if (IsA(path, MergeAppendPath))
2117  {
2118  MergeAppendPath *mpath = (MergeAppendPath *) path;
2119 
2120  *subpaths = list_concat(*subpaths, mpath->subpaths);
2121  return;
2122  }
2123 
2124  *subpaths = lappend(*subpaths, path);
2125 }
2126 
2127 /*
2128  * get_singleton_append_subpath
2129  * Returns the single subpath of an Append/MergeAppend, or just
2130  * return 'path' if it's not a single sub-path Append/MergeAppend.
2131  *
2132  * Note: 'path' must not be a parallel-aware path.
2133  */
2134 static Path *
2136 {
2137  Assert(!path->parallel_aware);
2138 
2139  if (IsA(path, AppendPath))
2140  {
2141  AppendPath *apath = (AppendPath *) path;
2142 
2143  if (list_length(apath->subpaths) == 1)
2144  return (Path *) linitial(apath->subpaths);
2145  }
2146  else if (IsA(path, MergeAppendPath))
2147  {
2148  MergeAppendPath *mpath = (MergeAppendPath *) path;
2149 
2150  if (list_length(mpath->subpaths) == 1)
2151  return (Path *) linitial(mpath->subpaths);
2152  }
2153 
2154  return path;
2155 }
2156 
2157 /*
2158  * set_dummy_rel_pathlist
2159  * Build a dummy path for a relation that's been excluded by constraints
2160  *
2161  * Rather than inventing a special "dummy" path type, we represent this as an
2162  * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2163  *
2164  * (See also mark_dummy_rel, which does basically the same thing, but is
2165  * typically used to change a rel into dummy state after we already made
2166  * paths for it.)
2167  */
2168 static void
2170 {
2171  /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2172  rel->rows = 0;
2173  rel->reltarget->width = 0;
2174 
2175  /* Discard any pre-existing paths; no further need for them */
2176  rel->pathlist = NIL;
2177  rel->partial_pathlist = NIL;
2178 
2179  /* Set up the dummy path */
2180  add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2181  NIL, rel->lateral_relids,
2182  0, false, -1));
2183 
2184  /*
2185  * We set the cheapest-path fields immediately, just in case they were
2186  * pointing at some discarded path. This is redundant when we're called
2187  * from set_rel_size(), but not when called from elsewhere, and doing it
2188  * twice is harmless anyway.
2189  */
2190  set_cheapest(rel);
2191 }
2192 
2193 /*
2194  * find_window_run_conditions
2195  * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2196  * function to determine the function's monotonic properties. We then
2197  * see if 'opexpr' can be used to short-circuit execution.
2198  *
2199  * For example row_number() over (order by ...) always produces a value one
2200  * higher than the previous. If someone has a window function in a subquery
2201  * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2202  * as well stop processing the windowagg once the row number reaches 11. Here
2203  * we check if 'opexpr' might help us to stop doing needless extra processing
2204  * in WindowAgg nodes.
2205  *
2206  * '*keep_original' is set to true if the caller should also use 'opexpr' for
2207  * its original purpose. This is set to false if the caller can assume that
2208  * the run condition will handle all of the required filtering.
2209  *
2210  * Returns true if 'opexpr' was found to be useful and was added to the
2211  * WindowClauses runCondition. We also set *keep_original accordingly and add
2212  * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2213  * If the 'opexpr' cannot be used then we set *keep_original to true and
2214  * return false.
2215  */
2216 static bool
2218  AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2219  bool wfunc_left, bool *keep_original,
2220  Bitmapset **run_cond_attrs)
2221 {
2222  Oid prosupport;
2223  Expr *otherexpr;
2226  WindowClause *wclause;
2227  List *opinfos;
2228  OpExpr *runopexpr;
2229  Oid runoperator;
2230  ListCell *lc;
2231 
2232  *keep_original = true;
2233 
2234  while (IsA(wfunc, RelabelType))
2235  wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2236 
2237  /* we can only work with window functions */
2238  if (!IsA(wfunc, WindowFunc))
2239  return false;
2240 
2241  /* can't use it if there are subplans in the WindowFunc */
2242  if (contain_subplans((Node *) wfunc))
2243  return false;
2244 
2245  prosupport = get_func_support(wfunc->winfnoid);
2246 
2247  /* Check if there's a support function for 'wfunc' */
2248  if (!OidIsValid(prosupport))
2249  return false;
2250 
2251  /* get the Expr from the other side of the OpExpr */
2252  if (wfunc_left)
2253  otherexpr = lsecond(opexpr->args);
2254  else
2255  otherexpr = linitial(opexpr->args);
2256 
2257  /*
2258  * The value being compared must not change during the evaluation of the
2259  * window partition.
2260  */
2261  if (!is_pseudo_constant_clause((Node *) otherexpr))
2262  return false;
2263 
2264  /* find the window clause belonging to the window function */
2265  wclause = (WindowClause *) list_nth(subquery->windowClause,
2266  wfunc->winref - 1);
2267 
2268  req.type = T_SupportRequestWFuncMonotonic;
2269  req.window_func = wfunc;
2270  req.window_clause = wclause;
2271 
2272  /* call the support function */
2274  DatumGetPointer(OidFunctionCall1(prosupport,
2275  PointerGetDatum(&req)));
2276 
2277  /*
2278  * Nothing to do if the function is neither monotonically increasing nor
2279  * monotonically decreasing.
2280  */
2281  if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2282  return false;
2283 
2284  runopexpr = NULL;
2285  runoperator = InvalidOid;
2286  opinfos = get_op_btree_interpretation(opexpr->opno);
2287 
2288  foreach(lc, opinfos)
2289  {
2291  int strategy = opinfo->strategy;
2292 
2293  /* handle < / <= */
2294  if (strategy == BTLessStrategyNumber ||
2295  strategy == BTLessEqualStrategyNumber)
2296  {
2297  /*
2298  * < / <= is supported for monotonically increasing functions in
2299  * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2300  * for monotonically decreasing functions.
2301  */
2302  if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2303  (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2304  {
2305  *keep_original = false;
2306  runopexpr = opexpr;
2307  runoperator = opexpr->opno;
2308  }
2309  break;
2310  }
2311  /* handle > / >= */
2312  else if (strategy == BTGreaterStrategyNumber ||
2313  strategy == BTGreaterEqualStrategyNumber)
2314  {
2315  /*
2316  * > / >= is supported for monotonically decreasing functions in
2317  * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2318  * for monotonically increasing functions.
2319  */
2320  if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2321  (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2322  {
2323  *keep_original = false;
2324  runopexpr = opexpr;
2325  runoperator = opexpr->opno;
2326  }
2327  break;
2328  }
2329  /* handle = */
2330  else if (strategy == BTEqualStrategyNumber)
2331  {
2332  int16 newstrategy;
2333 
2334  /*
2335  * When both monotonically increasing and decreasing then the
2336  * return value of the window function will be the same each time.
2337  * We can simply use 'opexpr' as the run condition without
2338  * modifying it.
2339  */
2340  if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2341  {
2342  *keep_original = false;
2343  runopexpr = opexpr;
2344  runoperator = opexpr->opno;
2345  break;
2346  }
2347 
2348  /*
2349  * When monotonically increasing we make a qual with <wfunc> <=
2350  * <value> or <value> >= <wfunc> in order to filter out values
2351  * which are above the value in the equality condition. For
2352  * monotonically decreasing functions we want to filter values
2353  * below the value in the equality condition.
2354  */
2355  if (res->monotonic & MONOTONICFUNC_INCREASING)
2356  newstrategy = wfunc_left ? BTLessEqualStrategyNumber : BTGreaterEqualStrategyNumber;
2357  else
2358  newstrategy = wfunc_left ? BTGreaterEqualStrategyNumber : BTLessEqualStrategyNumber;
2359 
2360  /* We must keep the original equality qual */
2361  *keep_original = true;
2362  runopexpr = opexpr;
2363 
2364  /* determine the operator to use for the runCondition qual */
2365  runoperator = get_opfamily_member(opinfo->opfamily_id,
2366  opinfo->oplefttype,
2367  opinfo->oprighttype,
2368  newstrategy);
2369  break;
2370  }
2371  }
2372 
2373  if (runopexpr != NULL)
2374  {
2375  Expr *newexpr;
2376 
2377  /*
2378  * Build the qual required for the run condition keeping the
2379  * WindowFunc on the same side as it was originally.
2380  */
2381  if (wfunc_left)
2382  newexpr = make_opclause(runoperator,
2383  runopexpr->opresulttype,
2384  runopexpr->opretset, (Expr *) wfunc,
2385  otherexpr, runopexpr->opcollid,
2386  runopexpr->inputcollid);
2387  else
2388  newexpr = make_opclause(runoperator,
2389  runopexpr->opresulttype,
2390  runopexpr->opretset,
2391  otherexpr, (Expr *) wfunc,
2392  runopexpr->opcollid,
2393  runopexpr->inputcollid);
2394 
2395  wclause->runCondition = lappend(wclause->runCondition, newexpr);
2396 
2397  /* record that this attno was used in a run condition */
2398  *run_cond_attrs = bms_add_member(*run_cond_attrs,
2400  return true;
2401  }
2402 
2403  /* unsupported OpExpr */
2404  return false;
2405 }
2406 
2407 /*
2408  * check_and_push_window_quals
2409  * Check if 'clause' is a qual that can be pushed into a WindowFunc's
2410  * WindowClause as a 'runCondition' qual. These, when present, allow
2411  * some unnecessary work to be skipped during execution.
2412  *
2413  * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2414  * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2415  * window quals for.
2416  *
2417  * Returns true if the caller still must keep the original qual or false if
2418  * the caller can safely ignore the original qual because the WindowAgg node
2419  * will use the runCondition to stop returning tuples.
2420  */
2421 static bool
2423  Node *clause, Bitmapset **run_cond_attrs)
2424 {
2425  OpExpr *opexpr = (OpExpr *) clause;
2426  bool keep_original = true;
2427  Var *var1;
2428  Var *var2;
2429 
2430  /* We're only able to use OpExprs with 2 operands */
2431  if (!IsA(opexpr, OpExpr))
2432  return true;
2433 
2434  if (list_length(opexpr->args) != 2)
2435  return true;
2436 
2437  /*
2438  * Currently, we restrict this optimization to strict OpExprs. The reason
2439  * for this is that during execution, once the runcondition becomes false,
2440  * we stop evaluating WindowFuncs. To avoid leaving around stale window
2441  * function result values, we set them to NULL. Having only strict
2442  * OpExprs here ensures that we properly filter out the tuples with NULLs
2443  * in the top-level WindowAgg.
2444  */
2445  set_opfuncid(opexpr);
2446  if (!func_strict(opexpr->opfuncid))
2447  return true;
2448 
2449  /*
2450  * Check for plain Vars that reference window functions in the subquery.
2451  * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2452  * be used as part of the run condition.
2453  */
2454 
2455  /* Check the left side of the OpExpr */
2456  var1 = linitial(opexpr->args);
2457  if (IsA(var1, Var) && var1->varattno > 0)
2458  {
2459  TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2460  WindowFunc *wfunc = (WindowFunc *) tle->expr;
2461 
2462  if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2463  opexpr, true, &keep_original,
2464  run_cond_attrs))
2465  return keep_original;
2466  }
2467 
2468  /* and check the right side */
2469  var2 = lsecond(opexpr->args);
2470  if (IsA(var2, Var) && var2->varattno > 0)
2471  {
2472  TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2473  WindowFunc *wfunc = (WindowFunc *) tle->expr;
2474 
2475  if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2476  opexpr, false, &keep_original,
2477  run_cond_attrs))
2478  return keep_original;
2479  }
2480 
2481  return true;
2482 }
2483 
2484 /*
2485  * set_subquery_pathlist
2486  * Generate SubqueryScan access paths for a subquery RTE
2487  *
2488  * We don't currently support generating parameterized paths for subqueries
2489  * by pushing join clauses down into them; it seems too expensive to re-plan
2490  * the subquery multiple times to consider different alternatives.
2491  * (XXX that could stand to be reconsidered, now that we use Paths.)
2492  * So the paths made here will be parameterized if the subquery contains
2493  * LATERAL references, otherwise not. As long as that's true, there's no need
2494  * for a separate set_subquery_size phase: just make the paths right away.
2495  */
2496 static void
2498  Index rti, RangeTblEntry *rte)
2499 {
2500  Query *parse = root->parse;
2501  Query *subquery = rte->subquery;
2502  bool trivial_pathtarget;
2503  Relids required_outer;
2504  pushdown_safety_info safetyInfo;
2505  double tuple_fraction;
2506  RelOptInfo *sub_final_rel;
2507  Bitmapset *run_cond_attrs = NULL;
2508  ListCell *lc;
2509 
2510  /*
2511  * Must copy the Query so that planning doesn't mess up the RTE contents
2512  * (really really need to fix the planner to not scribble on its input,
2513  * someday ... but see remove_unused_subquery_outputs to start with).
2514  */
2515  subquery = copyObject(subquery);
2516 
2517  /*
2518  * If it's a LATERAL subquery, it might contain some Vars of the current
2519  * query level, requiring it to be treated as parameterized, even though
2520  * we don't support pushing down join quals into subqueries.
2521  */
2522  required_outer = rel->lateral_relids;
2523 
2524  /*
2525  * Zero out result area for subquery_is_pushdown_safe, so that it can set
2526  * flags as needed while recursing. In particular, we need a workspace
2527  * for keeping track of the reasons why columns are unsafe to reference.
2528  * These reasons are stored in the bits inside unsafeFlags[i] when we
2529  * discover reasons that column i of the subquery is unsafe to be used in
2530  * a pushed-down qual.
2531  */
2532  memset(&safetyInfo, 0, sizeof(safetyInfo));
2533  safetyInfo.unsafeFlags = (unsigned char *)
2534  palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2535 
2536  /*
2537  * If the subquery has the "security_barrier" flag, it means the subquery
2538  * originated from a view that must enforce row-level security. Then we
2539  * must not push down quals that contain leaky functions. (Ideally this
2540  * would be checked inside subquery_is_pushdown_safe, but since we don't
2541  * currently pass the RTE to that function, we must do it here.)
2542  */
2543  safetyInfo.unsafeLeaky = rte->security_barrier;
2544 
2545  /*
2546  * If there are any restriction clauses that have been attached to the
2547  * subquery relation, consider pushing them down to become WHERE or HAVING
2548  * quals of the subquery itself. This transformation is useful because it
2549  * may allow us to generate a better plan for the subquery than evaluating
2550  * all the subquery output rows and then filtering them.
2551  *
2552  * There are several cases where we cannot push down clauses. Restrictions
2553  * involving the subquery are checked by subquery_is_pushdown_safe().
2554  * Restrictions on individual clauses are checked by
2555  * qual_is_pushdown_safe(). Also, we don't want to push down
2556  * pseudoconstant clauses; better to have the gating node above the
2557  * subquery.
2558  *
2559  * Non-pushed-down clauses will get evaluated as qpquals of the
2560  * SubqueryScan node.
2561  *
2562  * XXX Are there any cases where we want to make a policy decision not to
2563  * push down a pushable qual, because it'd result in a worse plan?
2564  */
2565  if (rel->baserestrictinfo != NIL &&
2566  subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2567  {
2568  /* OK to consider pushing down individual quals */
2569  List *upperrestrictlist = NIL;
2570  ListCell *l;
2571 
2572  foreach(l, rel->baserestrictinfo)
2573  {
2574  RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2575  Node *clause = (Node *) rinfo->clause;
2576 
2577  if (rinfo->pseudoconstant)
2578  {
2579  upperrestrictlist = lappend(upperrestrictlist, rinfo);
2580  continue;
2581  }
2582 
2583  switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2584  {
2585  case PUSHDOWN_SAFE:
2586  /* Push it down */
2587  subquery_push_qual(subquery, rte, rti, clause);
2588  break;
2589 
2591 
2592  /*
2593  * Since we can't push the qual down into the subquery,
2594  * check if it happens to reference a window function. If
2595  * so then it might be useful to use for the WindowAgg's
2596  * runCondition.
2597  */
2598  if (!subquery->hasWindowFuncs ||
2599  check_and_push_window_quals(subquery, rte, rti, clause,
2600  &run_cond_attrs))
2601  {
2602  /*
2603  * subquery has no window funcs or the clause is not a
2604  * suitable window run condition qual or it is, but
2605  * the original must also be kept in the upper query.
2606  */
2607  upperrestrictlist = lappend(upperrestrictlist, rinfo);
2608  }
2609  break;
2610 
2611  case PUSHDOWN_UNSAFE:
2612  upperrestrictlist = lappend(upperrestrictlist, rinfo);
2613  break;
2614  }
2615  }
2616  rel->baserestrictinfo = upperrestrictlist;
2617  /* We don't bother recomputing baserestrict_min_security */
2618  }
2619 
2620  pfree(safetyInfo.unsafeFlags);
2621 
2622  /*
2623  * The upper query might not use all the subquery's output columns; if
2624  * not, we can simplify. Pass the attributes that were pushed down into
2625  * WindowAgg run conditions to ensure we don't accidentally think those
2626  * are unused.
2627  */
2628  remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2629 
2630  /*
2631  * We can safely pass the outer tuple_fraction down to the subquery if the
2632  * outer level has no joining, aggregation, or sorting to do. Otherwise
2633  * we'd better tell the subquery to plan for full retrieval. (XXX This
2634  * could probably be made more intelligent ...)
2635  */
2636  if (parse->hasAggs ||
2637  parse->groupClause ||
2638  parse->groupingSets ||
2639  root->hasHavingQual ||
2640  parse->distinctClause ||
2641  parse->sortClause ||
2643  tuple_fraction = 0.0; /* default case */
2644  else
2645  tuple_fraction = root->tuple_fraction;
2646 
2647  /* plan_params should not be in use in current query level */
2648  Assert(root->plan_params == NIL);
2649 
2650  /* Generate a subroot and Paths for the subquery */
2651  rel->subroot = subquery_planner(root->glob, subquery,
2652  root,
2653  false, tuple_fraction);
2654 
2655  /* Isolate the params needed by this specific subplan */
2656  rel->subplan_params = root->plan_params;
2657  root->plan_params = NIL;
2658 
2659  /*
2660  * It's possible that constraint exclusion proved the subquery empty. If
2661  * so, it's desirable to produce an unadorned dummy path so that we will
2662  * recognize appropriate optimizations at this query level.
2663  */
2664  sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2665 
2666  if (IS_DUMMY_REL(sub_final_rel))
2667  {
2669  return;
2670  }
2671 
2672  /*
2673  * Mark rel with estimated output rows, width, etc. Note that we have to
2674  * do this before generating outer-query paths, else cost_subqueryscan is
2675  * not happy.
2676  */
2677  set_subquery_size_estimates(root, rel);
2678 
2679  /*
2680  * Also detect whether the reltarget is trivial, so that we can pass that
2681  * info to cost_subqueryscan (rather than re-deriving it multiple times).
2682  * It's trivial if it fetches all the subplan output columns in order.
2683  */
2684  if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2685  trivial_pathtarget = false;
2686  else
2687  {
2688  trivial_pathtarget = true;
2689  foreach(lc, rel->reltarget->exprs)
2690  {
2691  Node *node = (Node *) lfirst(lc);
2692  Var *var;
2693 
2694  if (!IsA(node, Var))
2695  {
2696  trivial_pathtarget = false;
2697  break;
2698  }
2699  var = (Var *) node;
2700  if (var->varno != rti ||
2701  var->varattno != foreach_current_index(lc) + 1)
2702  {
2703  trivial_pathtarget = false;
2704  break;
2705  }
2706  }
2707  }
2708 
2709  /*
2710  * For each Path that subquery_planner produced, make a SubqueryScanPath
2711  * in the outer query.
2712  */
2713  foreach(lc, sub_final_rel->pathlist)
2714  {
2715  Path *subpath = (Path *) lfirst(lc);
2716  List *pathkeys;
2717 
2718  /* Convert subpath's pathkeys to outer representation */
2719  pathkeys = convert_subquery_pathkeys(root,
2720  rel,
2721  subpath->pathkeys,
2722  make_tlist_from_pathtarget(subpath->pathtarget));
2723 
2724  /* Generate outer path using this subpath */
2725  add_path(rel, (Path *)
2726  create_subqueryscan_path(root, rel, subpath,
2727  trivial_pathtarget,
2728  pathkeys, required_outer));
2729  }
2730 
2731  /* If outer rel allows parallelism, do same for partial paths. */
2732  if (rel->consider_parallel && bms_is_empty(required_outer))
2733  {
2734  /* If consider_parallel is false, there should be no partial paths. */
2735  Assert(sub_final_rel->consider_parallel ||
2736  sub_final_rel->partial_pathlist == NIL);
2737 
2738  /* Same for partial paths. */
2739  foreach(lc, sub_final_rel->partial_pathlist)
2740  {
2741  Path *subpath = (Path *) lfirst(lc);
2742  List *pathkeys;
2743 
2744  /* Convert subpath's pathkeys to outer representation */
2745  pathkeys = convert_subquery_pathkeys(root,
2746  rel,
2747  subpath->pathkeys,
2748  make_tlist_from_pathtarget(subpath->pathtarget));
2749 
2750  /* Generate outer path using this subpath */
2751  add_partial_path(rel, (Path *)
2752  create_subqueryscan_path(root, rel, subpath,
2753  trivial_pathtarget,
2754  pathkeys,
2755  required_outer));
2756  }
2757  }
2758 }
2759 
2760 /*
2761  * set_function_pathlist
2762  * Build the (single) access path for a function RTE
2763  */
2764 static void
2766 {
2767  Relids required_outer;
2768  List *pathkeys = NIL;
2769 
2770  /*
2771  * We don't support pushing join clauses into the quals of a function
2772  * scan, but it could still have required parameterization due to LATERAL
2773  * refs in the function expression.
2774  */
2775  required_outer = rel->lateral_relids;
2776 
2777  /*
2778  * The result is considered unordered unless ORDINALITY was used, in which
2779  * case it is ordered by the ordinal column (the last one). See if we
2780  * care, by checking for uses of that Var in equivalence classes.
2781  */
2782  if (rte->funcordinality)
2783  {
2784  AttrNumber ordattno = rel->max_attr;
2785  Var *var = NULL;
2786  ListCell *lc;
2787 
2788  /*
2789  * Is there a Var for it in rel's targetlist? If not, the query did
2790  * not reference the ordinality column, or at least not in any way
2791  * that would be interesting for sorting.
2792  */
2793  foreach(lc, rel->reltarget->exprs)
2794  {
2795  Var *node = (Var *) lfirst(lc);
2796 
2797  /* checking varno/varlevelsup is just paranoia */
2798  if (IsA(node, Var) &&
2799  node->varattno == ordattno &&
2800  node->varno == rel->relid &&
2801  node->varlevelsup == 0)
2802  {
2803  var = node;
2804  break;
2805  }
2806  }
2807 
2808  /*
2809  * Try to build pathkeys for this Var with int8 sorting. We tell
2810  * build_expression_pathkey not to build any new equivalence class; if
2811  * the Var isn't already mentioned in some EC, it means that nothing
2812  * cares about the ordering.
2813  */
2814  if (var)
2815  pathkeys = build_expression_pathkey(root,
2816  (Expr *) var,
2817  Int8LessOperator,
2818  rel->relids,
2819  false);
2820  }
2821 
2822  /* Generate appropriate path */
2823  add_path(rel, create_functionscan_path(root, rel,
2824  pathkeys, required_outer));
2825 }
2826 
2827 /*
2828  * set_values_pathlist
2829  * Build the (single) access path for a VALUES RTE
2830  */
2831 static void
2833 {
2834  Relids required_outer;
2835 
2836  /*
2837  * We don't support pushing join clauses into the quals of a values scan,
2838  * but it could still have required parameterization due to LATERAL refs
2839  * in the values expressions.
2840  */
2841  required_outer = rel->lateral_relids;
2842 
2843  /* Generate appropriate path */
2844  add_path(rel, create_valuesscan_path(root, rel, required_outer));
2845 }
2846 
2847 /*
2848  * set_tablefunc_pathlist
2849  * Build the (single) access path for a table func RTE
2850  */
2851 static void
2853 {
2854  Relids required_outer;
2855 
2856  /*
2857  * We don't support pushing join clauses into the quals of a tablefunc
2858  * scan, but it could still have required parameterization due to LATERAL
2859  * refs in the function expression.
2860  */
2861  required_outer = rel->lateral_relids;
2862 
2863  /* Generate appropriate path */
2864  add_path(rel, create_tablefuncscan_path(root, rel,
2865  required_outer));
2866 }
2867 
2868 /*
2869  * set_cte_pathlist
2870  * Build the (single) access path for a non-self-reference CTE RTE
2871  *
2872  * There's no need for a separate set_cte_size phase, since we don't
2873  * support join-qual-parameterized paths for CTEs.
2874  */
2875 static void
2877 {
2878  Plan *cteplan;
2879  PlannerInfo *cteroot;
2880  Index levelsup;
2881  int ndx;
2882  ListCell *lc;
2883  int plan_id;
2884  Relids required_outer;
2885 
2886  /*
2887  * Find the referenced CTE, and locate the plan previously made for it.
2888  */
2889  levelsup = rte->ctelevelsup;
2890  cteroot = root;
2891  while (levelsup-- > 0)
2892  {
2893  cteroot = cteroot->parent_root;
2894  if (!cteroot) /* shouldn't happen */
2895  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2896  }
2897 
2898  /*
2899  * Note: cte_plan_ids can be shorter than cteList, if we are still working
2900  * on planning the CTEs (ie, this is a side-reference from another CTE).
2901  * So we mustn't use forboth here.
2902  */
2903  ndx = 0;
2904  foreach(lc, cteroot->parse->cteList)
2905  {
2906  CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
2907 
2908  if (strcmp(cte->ctename, rte->ctename) == 0)
2909  break;
2910  ndx++;
2911  }
2912  if (lc == NULL) /* shouldn't happen */
2913  elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
2914  if (ndx >= list_length(cteroot->cte_plan_ids))
2915  elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
2916  plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
2917  if (plan_id <= 0)
2918  elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
2919  cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2920 
2921  /* Mark rel with estimated output rows, width, etc */
2922  set_cte_size_estimates(root, rel, cteplan->plan_rows);
2923 
2924  /*
2925  * We don't support pushing join clauses into the quals of a CTE scan, but
2926  * it could still have required parameterization due to LATERAL refs in
2927  * its tlist.
2928  */
2929  required_outer = rel->lateral_relids;
2930 
2931  /* Generate appropriate path */
2932  add_path(rel, create_ctescan_path(root, rel, required_outer));
2933 }
2934 
2935 /*
2936  * set_namedtuplestore_pathlist
2937  * Build the (single) access path for a named tuplestore RTE
2938  *
2939  * There's no need for a separate set_namedtuplestore_size phase, since we
2940  * don't support join-qual-parameterized paths for tuplestores.
2941  */
2942 static void
2944  RangeTblEntry *rte)
2945 {
2946  Relids required_outer;
2947 
2948  /* Mark rel with estimated output rows, width, etc */
2950 
2951  /*
2952  * We don't support pushing join clauses into the quals of a tuplestore
2953  * scan, but it could still have required parameterization due to LATERAL
2954  * refs in its tlist.
2955  */
2956  required_outer = rel->lateral_relids;
2957 
2958  /* Generate appropriate path */
2959  add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2960 
2961  /* Select cheapest path (pretty easy in this case...) */
2962  set_cheapest(rel);
2963 }
2964 
2965 /*
2966  * set_result_pathlist
2967  * Build the (single) access path for an RTE_RESULT RTE
2968  *
2969  * There's no need for a separate set_result_size phase, since we
2970  * don't support join-qual-parameterized paths for these RTEs.
2971  */
2972 static void
2974  RangeTblEntry *rte)
2975 {
2976  Relids required_outer;
2977 
2978  /* Mark rel with estimated output rows, width, etc */
2979  set_result_size_estimates(root, rel);
2980 
2981  /*
2982  * We don't support pushing join clauses into the quals of a Result scan,
2983  * but it could still have required parameterization due to LATERAL refs
2984  * in its tlist.
2985  */
2986  required_outer = rel->lateral_relids;
2987 
2988  /* Generate appropriate path */
2989  add_path(rel, create_resultscan_path(root, rel, required_outer));
2990 
2991  /* Select cheapest path (pretty easy in this case...) */
2992  set_cheapest(rel);
2993 }
2994 
2995 /*
2996  * set_worktable_pathlist
2997  * Build the (single) access path for a self-reference CTE RTE
2998  *
2999  * There's no need for a separate set_worktable_size phase, since we don't
3000  * support join-qual-parameterized paths for CTEs.
3001  */
3002 static void
3004 {
3005  Path *ctepath;
3006  PlannerInfo *cteroot;
3007  Index levelsup;
3008  Relids required_outer;
3009 
3010  /*
3011  * We need to find the non-recursive term's path, which is in the plan
3012  * level that's processing the recursive UNION, which is one level *below*
3013  * where the CTE comes from.
3014  */
3015  levelsup = rte->ctelevelsup;
3016  if (levelsup == 0) /* shouldn't happen */
3017  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3018  levelsup--;
3019  cteroot = root;
3020  while (levelsup-- > 0)
3021  {
3022  cteroot = cteroot->parent_root;
3023  if (!cteroot) /* shouldn't happen */
3024  elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3025  }
3026  ctepath = cteroot->non_recursive_path;
3027  if (!ctepath) /* shouldn't happen */
3028  elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3029 
3030  /* Mark rel with estimated output rows, width, etc */
3031  set_cte_size_estimates(root, rel, ctepath->rows);
3032 
3033  /*
3034  * We don't support pushing join clauses into the quals of a worktable
3035  * scan, but it could still have required parameterization due to LATERAL
3036  * refs in its tlist. (I'm not sure this is actually possible given the
3037  * restrictions on recursive references, but it's easy enough to support.)
3038  */
3039  required_outer = rel->lateral_relids;
3040 
3041  /* Generate appropriate path */
3042  add_path(rel, create_worktablescan_path(root, rel, required_outer));
3043 }
3044 
3045 /*
3046  * generate_gather_paths
3047  * Generate parallel access paths for a relation by pushing a Gather or
3048  * Gather Merge on top of a partial path.
3049  *
3050  * This must not be called until after we're done creating all partial paths
3051  * for the specified relation. (Otherwise, add_partial_path might delete a
3052  * path that some GatherPath or GatherMergePath has a reference to.)
3053  *
3054  * If we're generating paths for a scan or join relation, override_rows will
3055  * be false, and we'll just use the relation's size estimate. When we're
3056  * being called for a partially-grouped path, though, we need to override
3057  * the rowcount estimate. (It's not clear that the particular value we're
3058  * using here is actually best, but the underlying rel has no estimate so
3059  * we must do something.)
3060  */
3061 void
3062 generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3063 {
3064  Path *cheapest_partial_path;
3065  Path *simple_gather_path;
3066  ListCell *lc;
3067  double rows;
3068  double *rowsp = NULL;
3069 
3070  /* If there are no partial paths, there's nothing to do here. */
3071  if (rel->partial_pathlist == NIL)
3072  return;
3073 
3074  /* Should we override the rel's rowcount estimate? */
3075  if (override_rows)
3076  rowsp = &rows;
3077 
3078  /*
3079  * The output of Gather is always unsorted, so there's only one partial
3080  * path of interest: the cheapest one. That will be the one at the front
3081  * of partial_pathlist because of the way add_partial_path works.
3082  */
3083  cheapest_partial_path = linitial(rel->partial_pathlist);
3084  rows =
3085  cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
3086  simple_gather_path = (Path *)
3087  create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3088  NULL, rowsp);
3089  add_path(rel, simple_gather_path);
3090 
3091  /*
3092  * For each useful ordering, we can consider an order-preserving Gather
3093  * Merge.
3094  */
3095  foreach(lc, rel->partial_pathlist)
3096  {
3097  Path *subpath = (Path *) lfirst(lc);
3098  GatherMergePath *path;
3099 
3100  if (subpath->pathkeys == NIL)
3101  continue;
3102 
3103  rows = subpath->rows * subpath->parallel_workers;
3104  path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3105  subpath->pathkeys, NULL, rowsp);
3106  add_path(rel, &path->path);
3107  }
3108 }
3109 
3110 /*
3111  * get_useful_pathkeys_for_relation
3112  * Determine which orderings of a relation might be useful.
3113  *
3114  * Getting data in sorted order can be useful either because the requested
3115  * order matches the final output ordering for the overall query we're
3116  * planning, or because it enables an efficient merge join. Here, we try
3117  * to figure out which pathkeys to consider.
3118  *
3119  * This allows us to do incremental sort on top of an index scan under a gather
3120  * merge node, i.e. parallelized.
3121  *
3122  * If the require_parallel_safe is true, we also require the expressions to
3123  * be parallel safe (which allows pushing the sort below Gather Merge).
3124  *
3125  * XXX At the moment this can only ever return a list with a single element,
3126  * because it looks at query_pathkeys only. So we might return the pathkeys
3127  * directly, but it seems plausible we'll want to consider other orderings
3128  * in the future. For example, we might want to consider pathkeys useful for
3129  * merge joins.
3130  */
3131 static List *
3133  bool require_parallel_safe)
3134 {
3135  List *useful_pathkeys_list = NIL;
3136 
3137  /*
3138  * Considering query_pathkeys is always worth it, because it might allow
3139  * us to avoid a total sort when we have a partially presorted path
3140  * available or to push the total sort into the parallel portion of the
3141  * query.
3142  */
3143  if (root->query_pathkeys)
3144  {
3145  ListCell *lc;
3146  int npathkeys = 0; /* useful pathkeys */
3147 
3148  foreach(lc, root->query_pathkeys)
3149  {
3150  PathKey *pathkey = (PathKey *) lfirst(lc);
3151  EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3152 
3153  /*
3154  * We can only build a sort for pathkeys that contain a
3155  * safe-to-compute-early EC member computable from the current
3156  * relation's reltarget, so ignore the remainder of the list as
3157  * soon as we find a pathkey without such a member.
3158  *
3159  * It's still worthwhile to return any prefix of the pathkeys list
3160  * that meets this requirement, as we may be able to do an
3161  * incremental sort.
3162  *
3163  * If requested, ensure the sort expression is parallel-safe too.
3164  */
3165  if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3166  require_parallel_safe))
3167  break;
3168 
3169  npathkeys++;
3170  }
3171 
3172  /*
3173  * The whole query_pathkeys list matches, so append it directly, to
3174  * allow comparing pathkeys easily by comparing list pointer. If we
3175  * have to truncate the pathkeys, we gotta do a copy though.
3176  */
3177  if (npathkeys == list_length(root->query_pathkeys))
3178  useful_pathkeys_list = lappend(useful_pathkeys_list,
3179  root->query_pathkeys);
3180  else if (npathkeys > 0)
3181  useful_pathkeys_list = lappend(useful_pathkeys_list,
3183  npathkeys));
3184  }
3185 
3186  return useful_pathkeys_list;
3187 }
3188 
3189 /*
3190  * generate_useful_gather_paths
3191  * Generate parallel access paths for a relation by pushing a Gather or
3192  * Gather Merge on top of a partial path.
3193  *
3194  * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3195  * paths (aiming to preserve the ordering), but also considers ordering that
3196  * might be useful for nodes above the gather merge node, and tries to add
3197  * a sort (regular or incremental) to provide that.
3198  */
3199 void
3200 generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3201 {
3202  ListCell *lc;
3203  double rows;
3204  double *rowsp = NULL;
3205  List *useful_pathkeys_list = NIL;
3206  Path *cheapest_partial_path = NULL;
3207 
3208  /* If there are no partial paths, there's nothing to do here. */
3209  if (rel->partial_pathlist == NIL)
3210  return;
3211 
3212  /* Should we override the rel's rowcount estimate? */
3213  if (override_rows)
3214  rowsp = &rows;
3215 
3216  /* generate the regular gather (merge) paths */
3217  generate_gather_paths(root, rel, override_rows);
3218 
3219  /* consider incremental sort for interesting orderings */
3220  useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3221 
3222  /* used for explicit (full) sort paths */
3223  cheapest_partial_path = linitial(rel->partial_pathlist);
3224 
3225  /*
3226  * Consider sorted paths for each interesting ordering. We generate both
3227  * incremental and full sort.
3228  */
3229  foreach(lc, useful_pathkeys_list)
3230  {
3231  List *useful_pathkeys = lfirst(lc);
3232  ListCell *lc2;
3233  bool is_sorted;
3234  int presorted_keys;
3235 
3236  foreach(lc2, rel->partial_pathlist)
3237  {
3238  Path *subpath = (Path *) lfirst(lc2);
3239  GatherMergePath *path;
3240 
3241  is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3242  subpath->pathkeys,
3243  &presorted_keys);
3244 
3245  /*
3246  * We don't need to consider the case where a subpath is already
3247  * fully sorted because generate_gather_paths already creates a
3248  * gather merge path for every subpath that has pathkeys present.
3249  *
3250  * But since the subpath is already sorted, we know we don't need
3251  * to consider adding a sort (full or incremental) on top of it,
3252  * so we can continue here.
3253  */
3254  if (is_sorted)
3255  continue;
3256 
3257  /*
3258  * Try at least sorting the cheapest path and also try
3259  * incrementally sorting any path which is partially sorted
3260  * already (no need to deal with paths which have presorted keys
3261  * when incremental sort is disabled unless it's the cheapest
3262  * input path).
3263  */
3264  if (subpath != cheapest_partial_path &&
3265  (presorted_keys == 0 || !enable_incremental_sort))
3266  continue;
3267 
3268  /*
3269  * Consider regular sort for any path that's not presorted or if
3270  * incremental sort is disabled. We've no need to consider both
3271  * sort and incremental sort on the same path. We assume that
3272  * incremental sort is always faster when there are presorted
3273  * keys.
3274  *
3275  * This is not redundant with the gather paths created in
3276  * generate_gather_paths, because that doesn't generate ordered
3277  * output. Here we add an explicit sort to match the useful
3278  * ordering.
3279  */
3280  if (presorted_keys == 0 || !enable_incremental_sort)
3281  {
3282  subpath = (Path *) create_sort_path(root,
3283  rel,
3284  subpath,
3285  useful_pathkeys,
3286  -1.0);
3287  rows = subpath->rows * subpath->parallel_workers;
3288  }
3289  else
3291  rel,
3292  subpath,
3293  useful_pathkeys,
3294  presorted_keys,
3295  -1);
3296  path = create_gather_merge_path(root, rel,
3297  subpath,
3298  rel->reltarget,
3299  subpath->pathkeys,
3300  NULL,
3301  rowsp);
3302 
3303  add_path(rel, &path->path);
3304  }
3305  }
3306 }
3307 
3308 /*
3309  * make_rel_from_joinlist
3310  * Build access paths using a "joinlist" to guide the join path search.
3311  *
3312  * See comments for deconstruct_jointree() for definition of the joinlist
3313  * data structure.
3314  */
3315 static RelOptInfo *
3317 {
3318  int levels_needed;
3319  List *initial_rels;
3320  ListCell *jl;
3321 
3322  /*
3323  * Count the number of child joinlist nodes. This is the depth of the
3324  * dynamic-programming algorithm we must employ to consider all ways of
3325  * joining the child nodes.
3326  */
3327  levels_needed = list_length(joinlist);
3328 
3329  if (levels_needed <= 0)
3330  return NULL; /* nothing to do? */
3331 
3332  /*
3333  * Construct a list of rels corresponding to the child joinlist nodes.
3334  * This may contain both base rels and rels constructed according to
3335  * sub-joinlists.
3336  */
3337  initial_rels = NIL;
3338  foreach(jl, joinlist)
3339  {
3340  Node *jlnode = (Node *) lfirst(jl);
3341  RelOptInfo *thisrel;
3342 
3343  if (IsA(jlnode, RangeTblRef))
3344  {
3345  int varno = ((RangeTblRef *) jlnode)->rtindex;
3346 
3347  thisrel = find_base_rel(root, varno);
3348  }
3349  else if (IsA(jlnode, List))
3350  {
3351  /* Recurse to handle subproblem */
3352  thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3353  }
3354  else
3355  {
3356  elog(ERROR, "unrecognized joinlist node type: %d",
3357  (int) nodeTag(jlnode));
3358  thisrel = NULL; /* keep compiler quiet */
3359  }
3360 
3361  initial_rels = lappend(initial_rels, thisrel);
3362  }
3363 
3364  if (levels_needed == 1)
3365  {
3366  /*
3367  * Single joinlist node, so we're done.
3368  */
3369  return (RelOptInfo *) linitial(initial_rels);
3370  }
3371  else
3372  {
3373  /*
3374  * Consider the different orders in which we could join the rels,
3375  * using a plugin, GEQO, or the regular join search code.
3376  *
3377  * We put the initial_rels list into a PlannerInfo field because
3378  * has_legal_joinclause() needs to look at it (ugly :-().
3379  */
3380  root->initial_rels = initial_rels;
3381 
3382  if (join_search_hook)
3383  return (*join_search_hook) (root, levels_needed, initial_rels);
3384  else if (enable_geqo && levels_needed >= geqo_threshold)
3385  return geqo(root, levels_needed, initial_rels);
3386  else
3387  return standard_join_search(root, levels_needed, initial_rels);
3388  }
3389 }
3390 
3391 /*
3392  * standard_join_search
3393  * Find possible joinpaths for a query by successively finding ways
3394  * to join component relations into join relations.
3395  *
3396  * 'levels_needed' is the number of iterations needed, ie, the number of
3397  * independent jointree items in the query. This is > 1.
3398  *
3399  * 'initial_rels' is a list of RelOptInfo nodes for each independent
3400  * jointree item. These are the components to be joined together.
3401  * Note that levels_needed == list_length(initial_rels).
3402  *
3403  * Returns the final level of join relations, i.e., the relation that is
3404  * the result of joining all the original relations together.
3405  * At least one implementation path must be provided for this relation and
3406  * all required sub-relations.
3407  *
3408  * To support loadable plugins that modify planner behavior by changing the
3409  * join searching algorithm, we provide a hook variable that lets a plugin
3410  * replace or supplement this function. Any such hook must return the same
3411  * final join relation as the standard code would, but it might have a
3412  * different set of implementation paths attached, and only the sub-joinrels
3413  * needed for these paths need have been instantiated.
3414  *
3415  * Note to plugin authors: the functions invoked during standard_join_search()
3416  * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3417  * than one join-order search, you'll probably need to save and restore the
3418  * original states of those data structures. See geqo_eval() for an example.
3419  */
3420 RelOptInfo *
3421 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3422 {
3423  int lev;
3424  RelOptInfo *rel;
3425 
3426  /*
3427  * This function cannot be invoked recursively within any one planning
3428  * problem, so join_rel_level[] can't be in use already.
3429  */
3430  Assert(root->join_rel_level == NULL);
3431 
3432  /*
3433  * We employ a simple "dynamic programming" algorithm: we first find all
3434  * ways to build joins of two jointree items, then all ways to build joins
3435  * of three items (from two-item joins and single items), then four-item
3436  * joins, and so on until we have considered all ways to join all the
3437  * items into one rel.
3438  *
3439  * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3440  * set root->join_rel_level[1] to represent all the single-jointree-item
3441  * relations.
3442  */
3443  root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3444 
3445  root->join_rel_level[1] = initial_rels;
3446 
3447  for (lev = 2; lev <= levels_needed; lev++)
3448  {
3449  ListCell *lc;
3450 
3451  /*
3452  * Determine all possible pairs of relations to be joined at this
3453  * level, and build paths for making each one from every available
3454  * pair of lower-level relations.
3455  */
3456  join_search_one_level(root, lev);
3457 
3458  /*
3459  * Run generate_partitionwise_join_paths() and
3460  * generate_useful_gather_paths() for each just-processed joinrel. We
3461  * could not do this earlier because both regular and partial paths
3462  * can get added to a particular joinrel at multiple times within
3463  * join_search_one_level.
3464  *
3465  * After that, we're done creating paths for the joinrel, so run
3466  * set_cheapest().
3467  */
3468  foreach(lc, root->join_rel_level[lev])
3469  {
3470  rel = (RelOptInfo *) lfirst(lc);
3471 
3472  /* Create paths for partitionwise joins. */
3474 
3475  /*
3476  * Except for the topmost scan/join rel, consider gathering
3477  * partial paths. We'll do the same for the topmost scan/join rel
3478  * once we know the final targetlist (see grouping_planner's and
3479  * its call to apply_scanjoin_target_to_paths).
3480  */
3481  if (!bms_equal(rel->relids, root->all_query_rels))
3482  generate_useful_gather_paths(root, rel, false);
3483 
3484  /* Find and save the cheapest paths for this rel */
3485  set_cheapest(rel);
3486 
3487 #ifdef OPTIMIZER_DEBUG
3488  pprint(rel);
3489 #endif
3490  }
3491  }
3492 
3493  /*
3494  * We should have a single rel at the final level.
3495  */
3496  if (root->join_rel_level[levels_needed] == NIL)
3497  elog(ERROR, "failed to build any %d-way joins", levels_needed);
3498  Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3499 
3500  rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3501 
3502  root->join_rel_level = NULL;
3503 
3504  return rel;
3505 }
3506 
3507 /*****************************************************************************
3508  * PUSHING QUALS DOWN INTO SUBQUERIES
3509  *****************************************************************************/
3510 
3511 /*
3512  * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3513  *
3514  * subquery is the particular component query being checked. topquery
3515  * is the top component of a set-operations tree (the same Query if no
3516  * set-op is involved).
3517  *
3518  * Conditions checked here:
3519  *
3520  * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3521  * since that could change the set of rows returned.
3522  *
3523  * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3524  * quals into it, because that could change the results.
3525  *
3526  * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3527  * This is because upper-level quals should semantically be evaluated only
3528  * once per distinct row, not once per original row, and if the qual is
3529  * volatile then extra evaluations could change the results. (This issue
3530  * does not apply to other forms of aggregation such as GROUP BY, because
3531  * when those are present we push into HAVING not WHERE, so that the quals
3532  * are still applied after aggregation.)
3533  *
3534  * 4. If the subquery contains window functions, we cannot push volatile quals
3535  * into it. The issue here is a bit different from DISTINCT: a volatile qual
3536  * might succeed for some rows of a window partition and fail for others,
3537  * thereby changing the partition contents and thus the window functions'
3538  * results for rows that remain.
3539  *
3540  * 5. If the subquery contains any set-returning functions in its targetlist,
3541  * we cannot push volatile quals into it. That would push them below the SRFs
3542  * and thereby change the number of times they are evaluated. Also, a
3543  * volatile qual could succeed for some SRF output rows and fail for others,
3544  * a behavior that cannot occur if it's evaluated before SRF expansion.
3545  *
3546  * 6. If the subquery has nonempty grouping sets, we cannot push down any
3547  * quals. The concern here is that a qual referencing a "constant" grouping
3548  * column could get constant-folded, which would be improper because the value
3549  * is potentially nullable by grouping-set expansion. This restriction could
3550  * be removed if we had a parsetree representation that shows that such
3551  * grouping columns are not really constant. (There are other ideas that
3552  * could be used to relax this restriction, but that's the approach most
3553  * likely to get taken in the future. Note that there's not much to be gained
3554  * so long as subquery_planner can't move HAVING clauses to WHERE within such
3555  * a subquery.)
3556  *
3557  * In addition, we make several checks on the subquery's output columns to see
3558  * if it is safe to reference them in pushed-down quals. If output column k
3559  * is found to be unsafe to reference, we set the reason for that inside
3560  * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3561  * column k might not be referenced by some/all quals. The unsafeFlags[]
3562  * array will be consulted later by qual_is_pushdown_safe(). It's better to
3563  * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3564  * because when the subquery involves set operations we have to check the
3565  * output expressions in each arm of the set op.
3566  *
3567  * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3568  * we're effectively assuming that the quals cannot distinguish values that
3569  * the DISTINCT's equality operator sees as equal, yet there are many
3570  * counterexamples to that assumption. However use of such a qual with a
3571  * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3572  * "equal" value will be chosen as the output value by the DISTINCT operation.
3573  * So we don't worry too much about that. Another objection is that if the
3574  * qual is expensive to evaluate, running it for each original row might cost
3575  * more than we save by eliminating rows before the DISTINCT step. But it
3576  * would be very hard to estimate that at this stage, and in practice pushdown
3577  * seldom seems to make things worse, so we ignore that problem too.
3578  *
3579  * Note: likewise, pushing quals into a subquery with window functions is a
3580  * bit dubious: the quals might remove some rows of a window partition while
3581  * leaving others, causing changes in the window functions' results for the
3582  * surviving rows. We insist that such a qual reference only partitioning
3583  * columns, but again that only protects us if the qual does not distinguish
3584  * values that the partitioning equality operator sees as equal. The risks
3585  * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3586  * occurs and thus there is no theoretical problem with such a qual. But
3587  * we'll do this anyway because the potential performance benefits are very
3588  * large, and we've seen no field complaints about the longstanding comparable
3589  * behavior with DISTINCT.
3590  */
3591 static bool
3593  pushdown_safety_info *safetyInfo)
3594 {
3595  SetOperationStmt *topop;
3596 
3597  /* Check point 1 */
3598  if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
3599  return false;
3600 
3601  /* Check point 6 */
3602  if (subquery->groupClause && subquery->groupingSets)
3603  return false;
3604 
3605  /* Check points 3, 4, and 5 */
3606  if (subquery->distinctClause ||
3607  subquery->hasWindowFuncs ||
3608  subquery->hasTargetSRFs)
3609  safetyInfo->unsafeVolatile = true;
3610 
3611  /*
3612  * If we're at a leaf query, check for unsafe expressions in its target
3613  * list, and mark any reasons why they're unsafe in unsafeFlags[].
3614  * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3615  * so no need to check them.)
3616  */
3617  if (subquery->setOperations == NULL)
3618  check_output_expressions(subquery, safetyInfo);
3619 
3620  /* Are we at top level, or looking at a setop component? */
3621  if (subquery == topquery)
3622  {
3623  /* Top level, so check any component queries */
3624  if (subquery->setOperations != NULL)
3625  if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3626  safetyInfo))
3627  return false;
3628  }
3629  else
3630  {
3631  /* Setop component must not have more components (too weird) */
3632  if (subquery->setOperations != NULL)
3633  return false;
3634  /* Check whether setop component output types match top level */
3635  topop = castNode(SetOperationStmt, topquery->setOperations);
3636  Assert(topop);
3638  topop->colTypes,
3639  safetyInfo);
3640  }
3641  return true;
3642 }
3643 
3644 /*
3645  * Helper routine to recurse through setOperations tree
3646  */
3647 static bool
3649  pushdown_safety_info *safetyInfo)
3650 {
3651  if (IsA(setOp, RangeTblRef))
3652  {
3653  RangeTblRef *rtr = (RangeTblRef *) setOp;
3654  RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3655  Query *subquery = rte->subquery;
3656 
3657  Assert(subquery != NULL);
3658  return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3659  }
3660  else if (IsA(setOp, SetOperationStmt))
3661  {
3662  SetOperationStmt *op = (SetOperationStmt *) setOp;
3663 
3664  /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
3665  if (op->op == SETOP_EXCEPT)
3666  return false;
3667  /* Else recurse */
3668  if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
3669  return false;
3670  if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
3671  return false;
3672  }
3673  else
3674  {
3675  elog(ERROR, "unrecognized node type: %d",
3676  (int) nodeTag(setOp));
3677  }
3678  return true;
3679 }
3680 
3681 /*
3682  * check_output_expressions - check subquery's output expressions for safety
3683  *
3684  * There are several cases in which it's unsafe to push down an upper-level
3685  * qual if it references a particular output column of a subquery. We check
3686  * each output column of the subquery and set flags in unsafeFlags[k] when we
3687  * see that column is unsafe for a pushed-down qual to reference. The
3688  * conditions checked here are:
3689  *
3690  * 1. We must not push down any quals that refer to subselect outputs that
3691  * return sets, else we'd introduce functions-returning-sets into the
3692  * subquery's WHERE/HAVING quals.
3693  *
3694  * 2. We must not push down any quals that refer to subselect outputs that
3695  * contain volatile functions, for fear of introducing strange results due
3696  * to multiple evaluation of a volatile function.
3697  *
3698  * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3699  * refer to non-DISTINCT output columns, because that could change the set
3700  * of rows returned. (This condition is vacuous for DISTINCT, because then
3701  * there are no non-DISTINCT output columns, so we needn't check. Note that
3702  * subquery_is_pushdown_safe already reported that we can't use volatile
3703  * quals if there's DISTINCT or DISTINCT ON.)
3704  *
3705  * 4. If the subquery has any window functions, we must not push down quals
3706  * that reference any output columns that are not listed in all the subquery's
3707  * window PARTITION BY clauses. We can push down quals that use only
3708  * partitioning columns because they should succeed or fail identically for
3709  * every row of any one window partition, and totally excluding some
3710  * partitions will not change a window function's results for remaining
3711  * partitions. (Again, this also requires nonvolatile quals, but
3712  * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3713  * unsafe for this reason can still have WindowClause run conditions pushed
3714  * down.
3715  */
3716 static void
3718 {
3719  ListCell *lc;
3720 
3721  foreach(lc, subquery->targetList)
3722  {
3723  TargetEntry *tle = (TargetEntry *) lfirst(lc);
3724 
3725  if (tle->resjunk)
3726  continue; /* ignore resjunk columns */
3727 
3728  /* Functions returning sets are unsafe (point 1) */
3729  if (subquery->hasTargetSRFs &&
3730  (safetyInfo->unsafeFlags[tle->resno] &
3731  UNSAFE_HAS_SET_FUNC) == 0 &&
3732  expression_returns_set((Node *) tle->expr))
3733  {
3734  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3735  continue;
3736  }
3737 
3738  /* Volatile functions are unsafe (point 2) */
3739  if ((safetyInfo->unsafeFlags[tle->resno] &
3740  UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3742  {
3743  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3744  continue;
3745  }
3746 
3747  /* If subquery uses DISTINCT ON, check point 3 */
3748  if (subquery->hasDistinctOn &&
3749  (safetyInfo->unsafeFlags[tle->resno] &
3751  !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3752  {
3753  /* non-DISTINCT column, so mark it unsafe */
3754  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3755  continue;
3756  }
3757 
3758  /* If subquery uses window functions, check point 4 */
3759  if (subquery->hasWindowFuncs &&
3760  (safetyInfo->unsafeFlags[tle->resno] &
3762  !targetIsInAllPartitionLists(tle, subquery))
3763  {
3764  /* not present in all PARTITION BY clauses, so mark it unsafe */
3765  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3766  continue;
3767  }
3768  }
3769 }
3770 
3771 /*
3772  * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3773  * push quals into each component query, but the quals can only reference
3774  * subquery columns that suffer no type coercions in the set operation.
3775  * Otherwise there are possible semantic gotchas. So, we check the
3776  * component queries to see if any of them have output types different from
3777  * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3778  * unsafeFlags[k] if column k has different type in any component.
3779  *
3780  * We don't have to care about typmods here: the only allowed difference
3781  * between set-op input and output typmods is input is a specific typmod
3782  * and output is -1, and that does not require a coercion.
3783  *
3784  * tlist is a subquery tlist.
3785  * colTypes is an OID list of the top-level setop's output column types.
3786  * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
3787  */
3788 static void
3790  pushdown_safety_info *safetyInfo)
3791 {
3792  ListCell *l;
3793  ListCell *colType = list_head(colTypes);
3794 
3795  foreach(l, tlist)
3796  {
3797  TargetEntry *tle = (TargetEntry *) lfirst(l);
3798 
3799  if (tle->resjunk)
3800  continue; /* ignore resjunk columns */
3801  if (colType == NULL)
3802  elog(ERROR, "wrong number of tlist entries");
3803  if (exprType((Node *) tle->expr) != lfirst_oid(colType))
3804  safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
3805  colType = lnext(colTypes, colType);
3806  }
3807  if (colType != NULL)
3808  elog(ERROR, "wrong number of tlist entries");
3809 }
3810 
3811 /*
3812  * targetIsInAllPartitionLists
3813  * True if the TargetEntry is listed in the PARTITION BY clause
3814  * of every window defined in the query.
3815  *
3816  * It would be safe to ignore windows not actually used by any window
3817  * function, but it's not easy to get that info at this stage; and it's
3818  * unlikely to be useful to spend any extra cycles getting it, since
3819  * unreferenced window definitions are probably infrequent in practice.
3820  */
3821 static bool
3823 {
3824  ListCell *lc;
3825 
3826  foreach(lc, query->windowClause)
3827  {
3828  WindowClause *wc = (WindowClause *) lfirst(lc);
3829 
3831  return false;
3832  }
3833  return true;
3834 }
3835 
3836 /*
3837  * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3838  *
3839  * rinfo is a restriction clause applying to the given subquery (whose RTE
3840  * has index rti in the parent query).
3841  *
3842  * Conditions checked here:
3843  *
3844  * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3845  * unclear that it will work correctly: SubLinks will already have been
3846  * transformed into SubPlans in the qual, but not in the subquery). Note that
3847  * SubLinks that transform to initplans are safe, and will be accepted here
3848  * because what we'll see in the qual is just a Param referencing the initplan
3849  * output.
3850  *
3851  * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3852  * functions.
3853  *
3854  * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
3855  * functions that are passed Var nodes, and therefore might reveal values from
3856  * the subquery as side effects.
3857  *
3858  * 4. rinfo's clause must not refer to the whole-row output of the subquery
3859  * (since there is no easy way to name that within the subquery itself).
3860  *
3861  * 5. rinfo's clause must not refer to any subquery output columns that were
3862  * found to be unsafe to reference by subquery_is_pushdown_safe().
3863  */
3864 static pushdown_safe_type
3866  pushdown_safety_info *safetyInfo)
3867 {
3869  Node *qual = (Node *) rinfo->clause;
3870  List *vars;
3871  ListCell *vl;
3872 
3873  /* Refuse subselects (point 1) */
3874  if (contain_subplans(qual))
3875  return PUSHDOWN_UNSAFE;
3876 
3877  /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3878  if (safetyInfo->unsafeVolatile &&
3879  contain_volatile_functions((Node *) rinfo))
3880  return PUSHDOWN_UNSAFE;
3881 
3882  /* Refuse leaky quals if told to (point 3) */
3883  if (safetyInfo->unsafeLeaky &&
3884  contain_leaked_vars(qual))
3885  return PUSHDOWN_UNSAFE;
3886 
3887  /*
3888  * Examine all Vars used in clause. Since it's a restriction clause, all
3889  * such Vars must refer to subselect output columns ... unless this is
3890  * part of a LATERAL subquery, in which case there could be lateral
3891  * references.
3892  *
3893  * By omitting the relevant flags, this also gives us a cheap sanity check
3894  * that no aggregates or window functions appear in the qual. Those would
3895  * be unsafe to push down, but at least for the moment we could never see
3896  * any in a qual anyhow.
3897  */
3899  foreach(vl, vars)
3900  {
3901  Var *var = (Var *) lfirst(vl);
3902 
3903  /*
3904  * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3905  * It's not clear whether a PHV could safely be pushed down, and even
3906  * less clear whether such a situation could arise in any cases of
3907  * practical interest anyway. So for the moment, just refuse to push
3908  * down.
3909  */
3910  if (!IsA(var, Var))
3911  {
3912  safe = PUSHDOWN_UNSAFE;
3913  break;
3914  }
3915 
3916  /*
3917  * Punt if we find any lateral references. It would be safe to push
3918  * these down, but we'd have to convert them into outer references,
3919  * which subquery_push_qual lacks the infrastructure to do. The case
3920  * arises so seldom that it doesn't seem worth working hard on.
3921  */
3922  if (var->varno != rti)
3923  {
3924  safe = PUSHDOWN_UNSAFE;
3925  break;
3926  }
3927 
3928  /* Subqueries have no system columns */
3929  Assert(var->varattno >= 0);
3930 
3931  /* Check point 4 */
3932  if (var->varattno == 0)
3933  {
3934  safe = PUSHDOWN_UNSAFE;
3935  break;
3936  }
3937 
3938  /* Check point 5 */
3939  if (safetyInfo->unsafeFlags[var->varattno] != 0)
3940  {
3941  if (safetyInfo->unsafeFlags[var->varattno] &
3944  {
3945  safe = PUSHDOWN_UNSAFE;
3946  break;
3947  }
3948  else
3949  {
3950  /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
3952  /* don't break, we might find another Var that's unsafe */
3953  }
3954  }
3955  }
3956 
3957  list_free(vars);
3958 
3959  return safe;
3960 }
3961 
3962 /*
3963  * subquery_push_qual - push down a qual that we have determined is safe
3964  */
3965 static void
3966 subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3967 {
3968  if (subquery->setOperations != NULL)
3969  {
3970  /* Recurse to push it separately to each component query */
3971  recurse_push_qual(subquery->setOperations, subquery,
3972  rte, rti, qual);
3973  }
3974  else
3975  {
3976  /*
3977  * We need to replace Vars in the qual (which must refer to outputs of
3978  * the subquery) with copies of the subquery's targetlist expressions.
3979  * Note that at this point, any uplevel Vars in the qual should have
3980  * been replaced with Params, so they need no work.
3981  *
3982  * This step also ensures that when we are pushing into a setop tree,
3983  * each component query gets its own copy of the qual.
3984  */
3985  qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
3986  subquery->targetList,
3988  &subquery->hasSubLinks);
3989 
3990  /*
3991  * Now attach the qual to the proper place: normally WHERE, but if the
3992  * subquery uses grouping or aggregation, put it in HAVING (since the
3993  * qual really refers to the group-result rows).
3994  */
3995  if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
3996  subquery->havingQual = make_and_qual(subquery->havingQual, qual);
3997  else
3998  subquery->jointree->quals =
3999  make_and_qual(subquery->jointree->quals, qual);
4000 
4001  /*
4002  * We need not change the subquery's hasAggs or hasSubLinks flags,
4003  * since we can't be pushing down any aggregates that weren't there
4004  * before, and we don't push down subselects at all.
4005  */
4006  }
4007 }
4008 
4009 /*
4010  * Helper routine to recurse through setOperations tree
4011  */
4012 static void
4013 recurse_push_qual(Node *setOp, Query *topquery,
4014  RangeTblEntry *rte, Index rti, Node *qual)
4015 {
4016  if (IsA(setOp, RangeTblRef))
4017  {
4018  RangeTblRef *rtr = (RangeTblRef *) setOp;
4019  RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4020  Query *subquery = subrte->subquery;
4021 
4022  Assert(subquery != NULL);
4023  subquery_push_qual(subquery, rte, rti, qual);
4024  }
4025  else if (IsA(setOp, SetOperationStmt))
4026  {
4027  SetOperationStmt *op = (SetOperationStmt *) setOp;
4028 
4029  recurse_push_qual(op->larg, topquery, rte, rti, qual);
4030  recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4031  }
4032  else
4033  {
4034  elog(ERROR, "unrecognized node type: %d",
4035  (int) nodeTag(setOp));
4036  }
4037 }
4038 
4039 /*****************************************************************************
4040  * SIMPLIFYING SUBQUERY TARGETLISTS
4041  *****************************************************************************/
4042 
4043 /*
4044  * remove_unused_subquery_outputs
4045  * Remove subquery targetlist items we don't need
4046  *
4047  * It's possible, even likely, that the upper query does not read all the
4048  * output columns of the subquery. We can remove any such outputs that are
4049  * not needed by the subquery itself (e.g., as sort/group columns) and do not
4050  * affect semantics otherwise (e.g., volatile functions can't be removed).
4051  * This is useful not only because we might be able to remove expensive-to-
4052  * compute expressions, but because deletion of output columns might allow
4053  * optimizations such as join removal to occur within the subquery.
4054  *
4055  * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4056  * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4057  * parameter is modified by the function, so callers must make a copy if they
4058  * need to use the passed in Bitmapset after calling this function.
4059  *
4060  * To avoid affecting column numbering in the targetlist, we don't physically
4061  * remove unused tlist entries, but rather replace their expressions with NULL
4062  * constants. This is implemented by modifying subquery->targetList.
4063  */
4064 static void
4066  Bitmapset *extra_used_attrs)
4067 {
4068  Bitmapset *attrs_used;
4069  ListCell *lc;
4070 
4071  /*
4072  * Just point directly to extra_used_attrs. No need to bms_copy as none of
4073  * the current callers use the Bitmapset after calling this function.
4074  */
4075  attrs_used = extra_used_attrs;
4076 
4077  /*
4078  * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4079  * could update all the child SELECTs' tlists, but it seems not worth the
4080  * trouble presently.
4081  */
4082  if (subquery->setOperations)
4083  return;
4084 
4085  /*
4086  * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4087  * time: all its output columns must be used in the distinctClause.
4088  */
4089  if (subquery->distinctClause && !subquery->hasDistinctOn)
4090  return;
4091 
4092  /*
4093  * Collect a bitmap of all the output column numbers used by the upper
4094  * query.
4095  *
4096  * Add all the attributes needed for joins or final output. Note: we must
4097  * look at rel's targetlist, not the attr_needed data, because attr_needed
4098  * isn't computed for inheritance child rels, cf set_append_rel_size().
4099  * (XXX might be worth changing that sometime.)
4100  */
4101  pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4102 
4103  /* Add all the attributes used by un-pushed-down restriction clauses. */
4104  foreach(lc, rel->baserestrictinfo)
4105  {
4106  RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4107 
4108  pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4109  }
4110 
4111  /*
4112  * If there's a whole-row reference to the subquery, we can't remove
4113  * anything.
4114  */
4116  return;
4117 
4118  /*
4119  * Run through the tlist and zap entries we don't need. It's okay to
4120  * modify the tlist items in-place because set_subquery_pathlist made a
4121  * copy of the subquery.
4122  */
4123  foreach(lc, subquery->targetList)
4124  {
4125  TargetEntry *tle = (TargetEntry *) lfirst(lc);
4126  Node *texpr = (Node *) tle->expr;
4127 
4128  /*
4129  * If it has a sortgroupref number, it's used in some sort/group
4130  * clause so we'd better not remove it. Also, don't remove any
4131  * resjunk columns, since their reason for being has nothing to do
4132  * with anybody reading the subquery's output. (It's likely that
4133  * resjunk columns in a sub-SELECT would always have ressortgroupref
4134  * set, but even if they don't, it seems imprudent to remove them.)
4135  */
4136  if (tle->ressortgroupref || tle->resjunk)
4137  continue;
4138 
4139  /*
4140  * If it's used by the upper query, we can't remove it.
4141  */
4143  attrs_used))
4144  continue;
4145 
4146  /*
4147  * If it contains a set-returning function, we can't remove it since
4148  * that could change the number of rows returned by the subquery.
4149  */
4150  if (subquery->hasTargetSRFs &&
4151  expression_returns_set(texpr))
4152  continue;
4153 
4154  /*
4155  * If it contains volatile functions, we daren't remove it for fear
4156  * that the user is expecting their side-effects to happen.
4157  */
4158  if (contain_volatile_functions(texpr))
4159  continue;
4160 
4161  /*
4162  * OK, we don't need it. Replace the expression with a NULL constant.
4163  * Preserve the exposed type of the expression, in case something
4164  * looks at the rowtype of the subquery's result.
4165  */
4166  tle->expr = (Expr *) makeNullConst(exprType(texpr),
4167  exprTypmod(texpr),
4168  exprCollation(texpr));
4169  }
4170 }
4171 
4172 /*
4173  * create_partial_bitmap_paths
4174  * Build partial bitmap heap path for the relation
4175  */
4176 void
4178  Path *bitmapqual)
4179 {
4180  int parallel_workers;
4181  double pages_fetched;
4182 
4183  /* Compute heap pages for bitmap heap scan */
4184  pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4185  NULL, NULL);
4186 
4187  parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4189 
4190  if (parallel_workers <= 0)
4191  return;
4192 
4193  add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4194  bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4195 }
4196 
4197 /*
4198  * Compute the number of parallel workers that should be used to scan a
4199  * relation. We compute the parallel workers based on the size of the heap to
4200  * be scanned and the size of the index to be scanned, then choose a minimum
4201  * of those.
4202  *
4203  * "heap_pages" is the number of pages from the table that we expect to scan, or
4204  * -1 if we don't expect to scan any.
4205  *
4206  * "index_pages" is the number of pages from the index that we expect to scan, or
4207  * -1 if we don't expect to scan any.
4208  *
4209  * "max_workers" is caller's limit on the number of workers. This typically
4210  * comes from a GUC.
4211  */
4212 int
4213 compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4214  int max_workers)
4215 {
4216  int parallel_workers = 0;
4217 
4218  /*
4219  * If the user has set the parallel_workers reloption, use that; otherwise
4220  * select a default number of workers.
4221  */
4222  if (rel->rel_parallel_workers != -1)
4223  parallel_workers = rel->rel_parallel_workers;
4224  else
4225  {
4226  /*
4227  * If the number of pages being scanned is insufficient to justify a
4228  * parallel scan, just return zero ... unless it's an inheritance
4229  * child. In that case, we want to generate a parallel path here
4230  * anyway. It might not be worthwhile just for this relation, but
4231  * when combined with all of its inheritance siblings it may well pay
4232  * off.
4233  */
4234  if (rel->reloptkind == RELOPT_BASEREL &&
4235  ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4236  (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4237  return 0;
4238 
4239  if (heap_pages >= 0)
4240  {
4241  int heap_parallel_threshold;
4242  int heap_parallel_workers = 1;
4243 
4244  /*
4245  * Select the number of workers based on the log of the size of
4246  * the relation. This probably needs to be a good deal more
4247  * sophisticated, but we need something here for now. Note that
4248  * the upper limit of the min_parallel_table_scan_size GUC is
4249  * chosen to prevent overflow here.
4250  */
4251  heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4252  while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4253  {
4254  heap_parallel_workers++;
4255  heap_parallel_threshold *= 3;
4256  if (heap_parallel_threshold > INT_MAX / 3)
4257  break; /* avoid overflow */
4258  }
4259 
4260  parallel_workers = heap_parallel_workers;
4261  }
4262 
4263  if (index_pages >= 0)
4264  {
4265  int index_parallel_workers = 1;
4266  int index_parallel_threshold;
4267 
4268  /* same calculation as for heap_pages above */
4269  index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4270  while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4271  {
4272  index_parallel_workers++;
4273  index_parallel_threshold *= 3;
4274  if (index_parallel_threshold > INT_MAX / 3)
4275  break; /* avoid overflow */
4276  }
4277 
4278  if (parallel_workers > 0)
4279  parallel_workers = Min(parallel_workers, index_parallel_workers);
4280  else
4281  parallel_workers = index_parallel_workers;
4282  }
4283  }
4284 
4285  /* In no case use more than caller supplied maximum number of workers */
4286  parallel_workers = Min(parallel_workers, max_workers);
4287 
4288  return parallel_workers;
4289 }
4290 
4291 /*
4292  * generate_partitionwise_join_paths
4293  * Create paths representing partitionwise join for given partitioned
4294  * join relation.
4295  *
4296  * This must not be called until after we are done adding paths for all
4297  * child-joins. Otherwise, add_path might delete a path to which some path
4298  * generated here has a reference.
4299  */
4300 void
4302 {
4303  List *live_children = NIL;
4304  int cnt_parts;
4305  int num_parts;
4306  RelOptInfo **part_rels;
4307 
4308  /* Handle only join relations here. */
4309  if (!IS_JOIN_REL(rel))
4310  return;
4311 
4312  /* We've nothing to do if the relation is not partitioned. */
4313  if (!IS_PARTITIONED_REL(rel))
4314  return;
4315 
4316  /* The relation should have consider_partitionwise_join set. */
4318 
4319  /* Guard against stack overflow due to overly deep partition hierarchy. */
4321 
4322  num_parts = rel->nparts;
4323  part_rels = rel->part_rels;
4324 
4325  /* Collect non-dummy child-joins. */
4326  for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4327  {
4328  RelOptInfo *child_rel = part_rels[cnt_parts];
4329 
4330  /* If it's been pruned entirely, it's certainly dummy. */
4331  if (child_rel == NULL)
4332  continue;
4333 
4334  /* Make partitionwise join paths for this partitioned child-join. */
4335  generate_partitionwise_join_paths(root, child_rel);
4336 
4337  /* If we failed to make any path for this child, we must give up. */
4338  if (child_rel->pathlist == NIL)
4339  {
4340  /*
4341  * Mark the parent joinrel as unpartitioned so that later
4342  * functions treat it correctly.
4343  */
4344  rel->nparts = 0;
4345  return;
4346  }
4347 
4348  /* Else, identify the cheapest path for it. */
4349  set_cheapest(child_rel);
4350 
4351  /* Dummy children need not be scanned, so ignore those. */
4352  if (IS_DUMMY_REL(child_rel))
4353  continue;
4354 
4355 #ifdef OPTIMIZER_DEBUG
4356  pprint(child_rel);
4357 #endif
4358 
4359  live_children = lappend(live_children, child_rel);
4360  }
4361 
4362  /* If all child-joins are dummy, parent join is also dummy. */
4363  if (!live_children)
4364  {
4365  mark_dummy_rel(rel);
4366  return;
4367  }
4368 
4369  /* Build additional paths for this rel from child-join paths. */
4370  add_paths_to_append_rel(root, rel, live_children);
4371  list_free(live_children);
4372 }
static void set_base_rel_sizes(PlannerInfo *root)
Definition: allpaths.c:293
static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:2497
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:2217
#define UNSAFE_TYPE_MISMATCH
Definition: allpaths.c:60
static Path * get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: allpaths.c:2002
static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2943
static void subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
Definition: allpaths.c:3966
void generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
Definition: allpaths.c:4301
RelOptInfo * standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
Definition: allpaths.c:3421
static void set_base_rel_consider_startup(PlannerInfo *root)
Definition: allpaths.c:250
#define UNSAFE_HAS_VOLATILE_FUNC
Definition: allpaths.c:56
#define UNSAFE_NOTIN_DISTINCTON_CLAUSE
Definition: allpaths.c:58
static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:472
static Path * get_singleton_append_subpath(Path *path)
Definition: allpaths.c:2135
static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:817
static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:857
static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:929
static void set_base_rel_pathlists(PlannerInfo *root)
Definition: allpaths.c:336
int geqo_threshold
Definition: allpaths.c:83
static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:1235
static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3865
static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2973
int compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages, int max_workers)
Definition: allpaths.c:4213
void generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
Definition: allpaths.c:3062
static void set_dummy_rel_pathlist(RelOptInfo *rel)
Definition: allpaths.c:2169
static void compare_tlist_datatypes(List *tlist, List *colTypes, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3789
static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:3003
static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
Definition: allpaths.c:3822
static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
Definition: allpaths.c:797
static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3592
join_search_hook_type join_search_hook
Definition: allpaths.c:91
bool enable_geqo
Definition: allpaths.c:82
static bool check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti, Node *clause, Bitmapset **run_cond_attrs)
Definition: allpaths.c:2422
void generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
Definition: allpaths.c:3200
static RelOptInfo * make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
Definition: allpaths.c:3316
static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2765
static void recurse_push_qual(Node *setOp, Query *topquery, RangeTblEntry *rte, Index rti, Node *qual)
Definition: allpaths.c:4013
static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:767
static List * get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel, bool require_parallel_safe)
Definition: allpaths.c:3132
static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel, Bitmapset *extra_used_attrs)
Definition: allpaths.c:4065
static void check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3717
static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2852
static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:905
set_rel_pathlist_hook_type set_rel_pathlist_hook
Definition: allpaths.c:88
static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2832
struct pushdown_safety_info pushdown_safety_info
#define UNSAFE_HAS_SET_FUNC
Definition: allpaths.c:57
static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:2876
static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:592
static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:947
void create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel, Path *bitmapqual)
Definition: allpaths.c:4177
static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: allpaths.c:575
void add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel, List *live_childrels)
Definition: allpaths.c:1305
static void set_rel_size(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: allpaths.c:363
static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel, List *live_childrels, List *all_child_pathkeys)
Definition: allpaths.c:1717
#define UNSAFE_NOTIN_PARTITIONBY_CLAUSE
Definition: allpaths.c:59
RelOptInfo * make_one_rel(PlannerInfo *root, List *joinlist)
Definition: allpaths.c:174
static bool recurse_pushdown_safe(Node *setOp, Query *topquery, pushdown_safety_info *safetyInfo)
Definition: allpaths.c:3648
pushdown_safe_type
Definition: allpaths.c:74
@ PUSHDOWN_WINDOWCLAUSE_RUNCOND
Definition: allpaths.c:77
@ PUSHDOWN_UNSAFE
Definition: allpaths.c:75
@ PUSHDOWN_SAFE
Definition: allpaths.c:76
static void accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
Definition: allpaths.c:2090
int min_parallel_index_scan_size
Definition: allpaths.c:85
int min_parallel_table_scan_size
Definition: allpaths.c:84
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:97
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:363
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:460
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:753
BMS_Membership bms_membership(const Bitmapset *a)
Definition: bitmapset.c:712
bool bms_overlap(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:527
bool bms_get_singleton_member(const Bitmapset *a, int *member)
Definition: bitmapset.c:652
#define bms_is_empty(a)
Definition: bitmapset.h:105
@ 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:993
signed short int16
Definition: c.h:482
signed int int32
Definition: c.h:483
#define Max(x, y)
Definition: c.h:987
unsigned int Index
Definition: c.h:603
#define OidIsValid(objectId)
Definition: c.h:764
bool is_pseudo_constant_clause(Node *clause)
Definition: clauses.c:2071
bool contain_leaked_vars(Node *clause)
Definition: clauses.c:1246
bool is_parallel_safe(PlannerInfo *root, Node *node)
Definition: clauses.c:736
bool contain_subplans(Node *clause)
Definition: clauses.c:332
bool contain_volatile_functions(Node *clause)
Definition: clauses.c:521
void set_namedtuplestore_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:6004
int max_parallel_workers_per_gather
Definition: costsize.c:133
void set_baserel_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5241
void set_function_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5874
double compute_bitmap_pages(PlannerInfo *root, RelOptInfo *baserel, Path *bitmapqual, int loop_count, Cost *cost, double *tuple)
Definition: costsize.c:6400
void set_cte_size_estimates(PlannerInfo *root, RelOptInfo *rel, double cte_rows)
Definition: costsize.c:5966
void set_result_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:6037
bool enable_partitionwise_join
Definition: costsize.c:149
void set_subquery_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5794
bool enable_parallel_append
Definition: costsize.c:151
void set_foreign_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:6066
double clamp_row_est(double nrows)
Definition: costsize.c:203
void set_tablefunc_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5912
void set_values_size_estimates(PlannerInfo *root, RelOptInfo *rel)
Definition: costsize.c:5934
bool enable_incremental_sort
Definition: costsize.c:141
#define ERROR
Definition: elog.h:39
void add_child_rel_equivalences(PlannerInfo *root, AppendRelInfo *appinfo, RelOptInfo *parent_rel, RelOptInfo *child_rel)
Definition: equivclass.c:2621
bool relation_can_be_sorted_early(PlannerInfo *root, RelOptInfo *rel, EquivalenceClass *ec, bool require_parallel_safe)
Definition: equivclass.c:922
#define OidFunctionCall1(functionId, arg1)
Definition: fmgr.h:680
RelOptInfo * geqo(PlannerInfo *root, int number_of_rels, List *initial_rels)
Definition: geqo_main.c:67
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:338
List * list_copy_head(const List *oldlist, int len)
Definition: list.c:1592
void list_free(List *list)
Definition: list.c:1545
List * list_copy_tail(const List *oldlist, int nskip)
Definition: list.c:1612
List * list_concat(List *list1, const List *list2)
Definition: list.c:560
char get_rel_persistence(Oid relid)
Definition: lsyscache.c:2082
char func_parallel(Oid funcid)
Definition: lsyscache.c:1803
RegProcedure get_func_support(Oid funcid)
Definition: lsyscache.c:1862
Oid get_opfamily_member(Oid opfamily, Oid lefttype, Oid righttype, int16 strategy)
Definition: lsyscache.c:165
bool func_strict(Oid funcid)
Definition: lsyscache.c:1765
List * get_op_btree_interpretation(Oid opno)
Definition: lsyscache.c:600
int32 get_typavgwidth(Oid typid, int32 typmod)
Definition: lsyscache.c:2560
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:613
Const * makeNullConst(Oid consttype, int32 consttypmod, Oid constcollid)
Definition: makefuncs.c:340
Node * make_and_qual(Node *qual1, Node *qual2)
Definition: makefuncs.c:692
void pfree(void *pointer)
Definition: mcxt.c:1456
void * palloc0(Size size)
Definition: mcxt.c:1257
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:43
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:282
Oid exprCollation(const Node *expr)
Definition: nodeFuncs.c:786
bool expression_returns_set(Node *clause)
Definition: nodeFuncs.c:728
void set_opfuncid(OpExpr *opexpr)
Definition: nodeFuncs.c:1779
#define IsA(nodeptr, _type_)
Definition: nodes.h:179
#define copyObject(obj)
Definition: nodes.h:244
#define nodeTag(nodeptr)
Definition: nodes.h:133
#define castNode(_type_, nodeptr)
Definition: nodes.h:197
@ JOIN_SEMI
Definition: nodes.h:318
@ JOIN_ANTI
Definition: nodes.h:319
#define PVC_INCLUDE_PLACEHOLDERS
Definition: optimizer.h:189
bool targetIsInSortList(TargetEntry *tle, Oid sortop, List *sortList)
@ SETOP_EXCEPT
Definition: parsenodes.h:1951
@ RTE_JOIN
Definition: parsenodes.h:1008
@ RTE_CTE
Definition: parsenodes.h:1012
@ RTE_NAMEDTUPLESTORE
Definition: parsenodes.h:1013
@ RTE_VALUES
Definition: parsenodes.h:1011
@ RTE_SUBQUERY
Definition: parsenodes.h:1007
@ RTE_RESULT
Definition: parsenodes.h:1014
@ RTE_FUNCTION
Definition: parsenodes.h:1009
@ RTE_TABLEFUNC
Definition: parsenodes.h:1010
@ RTE_RELATION
Definition: parsenodes.h:1006
#define rt_fetch(rangetable_index, rangetable)
Definition: parsetree.h:31
bool partitions_are_ordered(PartitionBoundInfo boundinfo, Bitmapset *live_parts)
Definition: partbounds.c:2853
List * build_expression_pathkey(PlannerInfo *root, Expr *expr, Oid opno, Relids rel, bool create_it)
Definition: pathkeys.c:801
Path * get_cheapest_path_for_pathkeys(List *paths, List *pathkeys, Relids required_outer, CostSelector cost_criterion, bool require_parallel_safe)
Definition: pathkeys.c:421
bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common)
Definition: pathkeys.c:359
bool has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
Definition: pathkeys.c:1987
Path * get_cheapest_parallel_safe_total_inner(List *paths)
Definition: pathkeys.c:500
List * convert_subquery_pathkeys(PlannerInfo *root, RelOptInfo *rel, List *subquery_pathkeys, List *subquery_tlist)
Definition: pathkeys.c:855
Path * get_cheapest_fractional_path_for_pathkeys(List *paths, List *pathkeys, Relids required_outer, double fraction)
Definition: pathkeys.c:467
bool pathkeys_contained_in(List *keys1, List *keys2)
Definition: pathkeys.c:340
PathKeysComparison compare_pathkeys(List *keys1, List *keys2)
Definition: pathkeys.c:301
List * build_partition_pathkeys(PlannerInfo *root, RelOptInfo *partrel, ScanDirection scandir, bool *partialkeys)
Definition: pathkeys.c:720
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:2959
GatherMergePath * create_gather_merge_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, PathTarget *target, List *pathkeys, Relids required_outer, double *rows)
Definition: pathnode.c:1873
Path * create_tablefuncscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2064
Path * create_namedtuplestorescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2141
GatherPath * create_gather_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, PathTarget *target, Relids required_outer, double *rows)
Definition: pathnode.c:1964
Path * create_resultscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2167
Path * create_samplescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:954
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:244
Path * reparameterize_path(PlannerInfo *root, Path *path, Relids required_outer, double loop_count)
Definition: pathnode.c:3902
SubqueryScanPath * create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, bool trivial_pathtarget, List *pathkeys, Relids required_outer)
Definition: pathnode.c:2008
Path * create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2116
void add_partial_path(RelOptInfo *parent_rel, Path *new_path)
Definition: pathnode.c:749
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:2038
IncrementalSortPath * create_incremental_sort_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, List *pathkeys, int presorted_keys, double limit_tuples)
Definition: pathnode.c:2910
Path * create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2193
void add_path(RelOptInfo *parent_rel, Path *new_path)
Definition: pathnode.c:422
Path * create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
Definition: pathnode.c:2090
int compare_path_costs(Path *path1, Path *path2, CostSelector criterion)
Definition: pathnode.c:71
Path * create_seqscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer, int parallel_workers)
Definition: pathnode.c:929
BitmapHeapPath * create_bitmap_heap_path(PlannerInfo *root, RelOptInfo *rel, Path *bitmapqual, Relids required_outer, double loop_count, int parallel_degree)
Definition: pathnode.c:1044
#define IS_SIMPLE_REL(rel)
Definition: pathnodes.h:824
#define IS_DUMMY_REL(r)
Definition: pathnodes.h:1914
#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:1041
#define PATH_REQ_OUTER(path)
Definition: pathnodes.h:1637
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:45
void(* set_rel_pathlist_hook_type)(PlannerInfo *root, RelOptInfo *rel, Index rti, RangeTblEntry *rte)
Definition: paths.h:29
@ PATHKEYS_EQUAL
Definition: paths.h:198
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:467
#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
#define foreach_current_index(cell)
Definition: pg_list.h:403
bool relation_excluded_by_constraints(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
Definition: plancat.c:1529
PlannerInfo * subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root, bool hasRecursion, double tuple_fraction)
Definition: planner.c:623
bool limit_needed(Query *parse)
Definition: planner.c:2600
@ 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:3520
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:405
RelOptInfo * fetch_upper_rel(PlannerInfo *root, UpperRelationKind kind, Relids relids)
Definition: relnode.c:1443
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:1902
List * subpaths
Definition: pathnodes.h:1900
Index child_relid
Definition: pathnodes.h:2929
Index parent_relid
Definition: pathnodes.h:2928
Node * quals
Definition: primnodes.h:2041
Definition: pg_list.h:54
Definition: nodes.h:129
Oid opno
Definition: primnodes.h:753
List * args
Definition: primnodes.h:771
List * exprs
Definition: pathnodes.h:1501
List * pathkeys
Definition: pathnodes.h:1633
Cardinality rows
Definition: pathnodes.h:1628
int parallel_workers
Definition: pathnodes.h:1625
Cost total_cost
Definition: pathnodes.h:1630
bool parallel_aware
Definition: pathnodes.h:1621
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:204
FromExpr * jointree
Definition: parsenodes.h:174
Node * setOperations
Definition: parsenodes.h:209
List * cteList
Definition: parsenodes.h:165
List * groupClause
Definition: parsenodes.h:190
Node * havingQual
Definition: parsenodes.h:195
List * rtable
Definition: parsenodes.h:167
Node * limitOffset
Definition: parsenodes.h:203
List * windowClause
Definition: parsenodes.h:197
List * targetList
Definition: parsenodes.h:181
List * groupingSets
Definition: parsenodes.h:193
List * distinctClause
Definition: parsenodes.h:199
char * ctename
Definition: parsenodes.h:1156
bool self_reference
Definition: parsenodes.h:1158
bool security_barrier
Definition: parsenodes.h:1074
Index ctelevelsup
Definition: parsenodes.h:1157
bool funcordinality
Definition: parsenodes.h:1141
struct TableSampleClause * tablesample
Definition: parsenodes.h:1067
Query * subquery
Definition: parsenodes.h:1073
List * values_lists
Definition: parsenodes.h:1151
List * functions
Definition: parsenodes.h:1140
RTEKind rtekind
Definition: parsenodes.h:1025
List * baserestrictinfo
Definition: pathnodes.h:964
bool consider_param_startup
Definition: pathnodes.h:870
List * subplan_params
Definition: pathnodes.h:933
List * joininfo
Definition: pathnodes.h:970
Relids relids
Definition: pathnodes.h:856
struct PathTarget * reltarget
Definition: pathnodes.h:878
Index relid
Definition: pathnodes.h:903
Cardinality tuples
Definition: pathnodes.h:928
bool consider_parallel
Definition: pathnodes.h:872
BlockNumber pages
Definition: pathnodes.h:927
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:972
bool consider_startup
Definition: pathnodes.h:868
Bitmapset * live_parts
Definition: pathnodes.h:1018
int rel_parallel_workers
Definition: pathnodes.h:935
bool consider_partitionwise_join
Definition: pathnodes.h:978
List * partial_pathlist
Definition: pathnodes.h:885
PlannerInfo * subroot
Definition: pathnodes.h:932
AttrNumber max_attr
Definition: pathnodes.h:911
Relids nulling_relids
Definition: pathnodes.h:917
Cardinality rows
Definition: pathnodes.h:862
AttrNumber min_attr
Definition: pathnodes.h:909
RTEKind rtekind
Definition: pathnodes.h:907
Expr * clause
Definition: pathnodes.h:2529
SetOperation op
Definition: parsenodes.h:2026
JoinType jointype
Definition: pathnodes.h:2857
Relids syn_righthand
Definition: pathnodes.h:2856
struct WindowClause * window_clause
Definition: supportnodes.h:296
Expr * expr
Definition: primnodes.h:1922
AttrNumber resno
Definition: primnodes.h:1924
Index ressortgroupref
Definition: primnodes.h:1928
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:1489
Index winref
Definition: primnodes.h:565
Oid winfnoid
Definition: primnodes.h:553
unsigned char * unsafeFlags
Definition: allpaths.c:65
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:460
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