Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * allpaths.c
4 : * Routines to find possible search paths for processing a query
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/optimizer/path/allpaths.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 :
16 : #include "postgres.h"
17 :
18 : #include <limits.h>
19 : #include <math.h>
20 :
21 : #include "access/sysattr.h"
22 : #include "access/tsmapi.h"
23 : #include "catalog/pg_class.h"
24 : #include "catalog/pg_operator.h"
25 : #include "catalog/pg_proc.h"
26 : #include "foreign/fdwapi.h"
27 : #include "miscadmin.h"
28 : #include "nodes/makefuncs.h"
29 : #include "nodes/nodeFuncs.h"
30 : #include "nodes/supportnodes.h"
31 : #ifdef OPTIMIZER_DEBUG
32 : #include "nodes/print.h"
33 : #endif
34 : #include "optimizer/appendinfo.h"
35 : #include "optimizer/clauses.h"
36 : #include "optimizer/cost.h"
37 : #include "optimizer/geqo.h"
38 : #include "optimizer/optimizer.h"
39 : #include "optimizer/pathnode.h"
40 : #include "optimizer/paths.h"
41 : #include "optimizer/plancat.h"
42 : #include "optimizer/planner.h"
43 : #include "optimizer/tlist.h"
44 : #include "parser/parse_clause.h"
45 : #include "parser/parsetree.h"
46 : #include "partitioning/partbounds.h"
47 : #include "port/pg_bitutils.h"
48 : #include "rewrite/rewriteManip.h"
49 : #include "utils/lsyscache.h"
50 :
51 :
52 : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
53 : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
54 : #define UNSAFE_HAS_SET_FUNC (1 << 1)
55 : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
56 : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
57 : #define UNSAFE_TYPE_MISMATCH (1 << 4)
58 :
59 : /* results of subquery_is_pushdown_safe */
60 : typedef struct pushdown_safety_info
61 : {
62 : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
63 : * column is unsafe for qual pushdown, or 0 if
64 : * no reason. */
65 : bool unsafeVolatile; /* don't push down volatile quals */
66 : bool unsafeLeaky; /* don't push down leaky quals */
67 : } pushdown_safety_info;
68 :
69 : /* Return type for qual_is_pushdown_safe */
70 : typedef enum pushdown_safe_type
71 : {
72 : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
73 : PUSHDOWN_SAFE, /* safe to push qual into subquery */
74 : PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
75 : * run condition */
76 : } pushdown_safe_type;
77 :
78 : /* These parameters are set by GUC */
79 : bool enable_geqo = false; /* just in case GUC doesn't set it */
80 : int geqo_threshold;
81 : int min_parallel_table_scan_size;
82 : int min_parallel_index_scan_size;
83 :
84 : /* Hook for plugins to get control in set_rel_pathlist() */
85 : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
86 :
87 : /* Hook for plugins to replace standard_join_search() */
88 : join_search_hook_type join_search_hook = NULL;
89 :
90 :
91 : static void set_base_rel_consider_startup(PlannerInfo *root);
92 : static void set_base_rel_sizes(PlannerInfo *root);
93 : static void set_base_rel_pathlists(PlannerInfo *root);
94 : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
95 : Index rti, RangeTblEntry *rte);
96 : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
97 : Index rti, RangeTblEntry *rte);
98 : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
99 : RangeTblEntry *rte);
100 : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
101 : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
102 : RangeTblEntry *rte);
103 : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
104 : RangeTblEntry *rte);
105 : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
106 : RangeTblEntry *rte);
107 : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
108 : RangeTblEntry *rte);
109 : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
110 : RangeTblEntry *rte);
111 : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
112 : RangeTblEntry *rte);
113 : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
114 : Index rti, RangeTblEntry *rte);
115 : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
116 : Index rti, RangeTblEntry *rte);
117 : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
118 : List *live_childrels,
119 : List *all_child_pathkeys);
120 : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
121 : RelOptInfo *rel,
122 : Relids required_outer);
123 : static void accumulate_append_subpath(Path *path,
124 : List **subpaths,
125 : List **special_subpaths);
126 : static Path *get_singleton_append_subpath(Path *path);
127 : static void set_dummy_rel_pathlist(RelOptInfo *rel);
128 : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
129 : Index rti, RangeTblEntry *rte);
130 : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
131 : RangeTblEntry *rte);
132 : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
133 : RangeTblEntry *rte);
134 : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
135 : RangeTblEntry *rte);
136 : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : RangeTblEntry *rte);
138 : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : RangeTblEntry *rte);
140 : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : RangeTblEntry *rte);
142 : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : RangeTblEntry *rte);
144 : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
145 : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
146 : pushdown_safety_info *safetyInfo);
147 : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
148 : pushdown_safety_info *safetyInfo);
149 : static void check_output_expressions(Query *subquery,
150 : pushdown_safety_info *safetyInfo);
151 : static void compare_tlist_datatypes(List *tlist, List *colTypes,
152 : pushdown_safety_info *safetyInfo);
153 : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
154 : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
155 : RestrictInfo *rinfo,
156 : pushdown_safety_info *safetyInfo);
157 : static void subquery_push_qual(Query *subquery,
158 : RangeTblEntry *rte, Index rti, Node *qual);
159 : static void recurse_push_qual(Node *setOp, Query *topquery,
160 : RangeTblEntry *rte, Index rti, Node *qual);
161 : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
162 : Bitmapset *extra_used_attrs);
163 :
164 :
165 : /*
166 : * make_one_rel
167 : * Finds all possible access paths for executing a query, returning a
168 : * single rel that represents the join of all base rels in the query.
169 : */
170 : RelOptInfo *
171 328966 : make_one_rel(PlannerInfo *root, List *joinlist)
172 : {
173 : RelOptInfo *rel;
174 : Index rti;
175 : double total_pages;
176 :
177 : /* Mark base rels as to whether we care about fast-start plans */
178 328966 : set_base_rel_consider_startup(root);
179 :
180 : /*
181 : * Compute size estimates and consider_parallel flags for each base rel.
182 : */
183 328966 : set_base_rel_sizes(root);
184 :
185 : /*
186 : * We should now have size estimates for every actual table involved in
187 : * the query, and we also know which if any have been deleted from the
188 : * query by join removal, pruned by partition pruning, or eliminated by
189 : * constraint exclusion. So we can now compute total_table_pages.
190 : *
191 : * Note that appendrels are not double-counted here, even though we don't
192 : * bother to distinguish RelOptInfos for appendrel parents, because the
193 : * parents will have pages = 0.
194 : *
195 : * XXX if a table is self-joined, we will count it once per appearance,
196 : * which perhaps is the wrong thing ... but that's not completely clear,
197 : * and detecting self-joins here is difficult, so ignore it for now.
198 : */
199 328932 : total_pages = 0;
200 999040 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
201 : {
202 670108 : RelOptInfo *brel = root->simple_rel_array[rti];
203 :
204 : /* there may be empty slots corresponding to non-baserel RTEs */
205 670108 : if (brel == NULL)
206 156920 : continue;
207 :
208 : Assert(brel->relid == rti); /* sanity check on array */
209 :
210 513188 : if (IS_DUMMY_REL(brel))
211 1178 : continue;
212 :
213 512010 : if (IS_SIMPLE_REL(brel))
214 512010 : total_pages += (double) brel->pages;
215 : }
216 328932 : root->total_table_pages = total_pages;
217 :
218 : /*
219 : * Generate access paths for each base rel.
220 : */
221 328932 : set_base_rel_pathlists(root);
222 :
223 : /*
224 : * Generate access paths for the entire join tree.
225 : */
226 328932 : rel = make_rel_from_joinlist(root, joinlist);
227 :
228 : /*
229 : * The result should join all and only the query's base + outer-join rels.
230 : */
231 : Assert(bms_equal(rel->relids, root->all_query_rels));
232 :
233 328932 : return rel;
234 : }
235 :
236 : /*
237 : * set_base_rel_consider_startup
238 : * Set the consider_[param_]startup flags for each base-relation entry.
239 : *
240 : * For the moment, we only deal with consider_param_startup here; because the
241 : * logic for consider_startup is pretty trivial and is the same for every base
242 : * relation, we just let build_simple_rel() initialize that flag correctly to
243 : * start with. If that logic ever gets more complicated it would probably
244 : * be better to move it here.
245 : */
246 : static void
247 328966 : set_base_rel_consider_startup(PlannerInfo *root)
248 : {
249 : /*
250 : * Since parameterized paths can only be used on the inside of a nestloop
251 : * join plan, there is usually little value in considering fast-start
252 : * plans for them. However, for relations that are on the RHS of a SEMI
253 : * or ANTI join, a fast-start plan can be useful because we're only going
254 : * to care about fetching one tuple anyway.
255 : *
256 : * To minimize growth of planning time, we currently restrict this to
257 : * cases where the RHS is a single base relation, not a join; there is no
258 : * provision for consider_param_startup to get set at all on joinrels.
259 : * Also we don't worry about appendrels. costsize.c's costing rules for
260 : * nestloop semi/antijoins don't consider such cases either.
261 : */
262 : ListCell *lc;
263 :
264 372372 : foreach(lc, root->join_info_list)
265 : {
266 43406 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
267 : int varno;
268 :
269 52008 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
270 8602 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
271 : {
272 8420 : RelOptInfo *rel = find_base_rel(root, varno);
273 :
274 8420 : rel->consider_param_startup = true;
275 : }
276 : }
277 328966 : }
278 :
279 : /*
280 : * set_base_rel_sizes
281 : * Set the size estimates (rows and widths) for each base-relation entry.
282 : * Also determine whether to consider parallel paths for base relations.
283 : *
284 : * We do this in a separate pass over the base rels so that rowcount
285 : * estimates are available for parameterized path generation, and also so
286 : * that each rel's consider_parallel flag is set correctly before we begin to
287 : * generate paths.
288 : */
289 : static void
290 328966 : set_base_rel_sizes(PlannerInfo *root)
291 : {
292 : Index rti;
293 :
294 999076 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
295 : {
296 670144 : RelOptInfo *rel = root->simple_rel_array[rti];
297 : RangeTblEntry *rte;
298 :
299 : /* there may be empty slots corresponding to non-baserel RTEs */
300 670144 : if (rel == NULL)
301 156922 : continue;
302 :
303 : Assert(rel->relid == rti); /* sanity check on array */
304 :
305 : /* ignore RTEs that are "other rels" */
306 513222 : if (rel->reloptkind != RELOPT_BASEREL)
307 50068 : continue;
308 :
309 463154 : rte = root->simple_rte_array[rti];
310 :
311 : /*
312 : * If parallelism is allowable for this query in general, see whether
313 : * it's allowable for this rel in particular. We have to do this
314 : * before set_rel_size(), because (a) if this rel is an inheritance
315 : * parent, set_append_rel_size() will use and perhaps change the rel's
316 : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
317 : * goes ahead and makes paths immediately.
318 : */
319 463154 : if (root->glob->parallelModeOK)
320 367218 : set_rel_consider_parallel(root, rel, rte);
321 :
322 463154 : set_rel_size(root, rel, rti, rte);
323 : }
324 328932 : }
325 :
326 : /*
327 : * set_base_rel_pathlists
328 : * Finds all paths available for scanning each base-relation entry.
329 : * Sequential scan and any available indices are considered.
330 : * Each useful path is attached to its relation's 'pathlist' field.
331 : */
332 : static void
333 328932 : set_base_rel_pathlists(PlannerInfo *root)
334 : {
335 : Index rti;
336 :
337 999040 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
338 : {
339 670108 : RelOptInfo *rel = root->simple_rel_array[rti];
340 :
341 : /* there may be empty slots corresponding to non-baserel RTEs */
342 670108 : if (rel == NULL)
343 156920 : continue;
344 :
345 : Assert(rel->relid == rti); /* sanity check on array */
346 :
347 : /* ignore RTEs that are "other rels" */
348 513188 : if (rel->reloptkind != RELOPT_BASEREL)
349 50068 : continue;
350 :
351 463120 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
352 : }
353 328932 : }
354 :
355 : /*
356 : * set_rel_size
357 : * Set size estimates for a base relation
358 : */
359 : static void
360 513002 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
361 : Index rti, RangeTblEntry *rte)
362 : {
363 976156 : if (rel->reloptkind == RELOPT_BASEREL &&
364 463154 : relation_excluded_by_constraints(root, rel, rte))
365 : {
366 : /*
367 : * We proved we don't need to scan the rel via constraint exclusion,
368 : * so set up a single dummy path for it. Here we only check this for
369 : * regular baserels; if it's an otherrel, CE was already checked in
370 : * set_append_rel_size().
371 : *
372 : * In this case, we go ahead and set up the relation's path right away
373 : * instead of leaving it for set_rel_pathlist to do. This is because
374 : * we don't have a convention for marking a rel as dummy except by
375 : * assigning a dummy path to it.
376 : */
377 538 : set_dummy_rel_pathlist(rel);
378 : }
379 512464 : else if (rte->inh)
380 : {
381 : /* It's an "append relation", process accordingly */
382 23928 : set_append_rel_size(root, rel, rti, rte);
383 : }
384 : else
385 : {
386 488536 : switch (rel->rtekind)
387 : {
388 403050 : case RTE_RELATION:
389 403050 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
390 : {
391 : /* Foreign table */
392 2412 : set_foreign_size(root, rel, rte);
393 : }
394 400638 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
395 : {
396 : /*
397 : * We could get here if asked to scan a partitioned table
398 : * with ONLY. In that case we shouldn't scan any of the
399 : * partitions, so mark it as a dummy rel.
400 : */
401 40 : set_dummy_rel_pathlist(rel);
402 : }
403 400598 : else if (rte->tablesample != NULL)
404 : {
405 : /* Sampled relation */
406 306 : set_tablesample_rel_size(root, rel, rte);
407 : }
408 : else
409 : {
410 : /* Plain relation */
411 400292 : set_plain_rel_size(root, rel, rte);
412 : }
413 403016 : break;
414 11762 : case RTE_SUBQUERY:
415 :
416 : /*
417 : * Subqueries don't support making a choice between
418 : * parameterized and unparameterized paths, so just go ahead
419 : * and build their paths immediately.
420 : */
421 11762 : set_subquery_pathlist(root, rel, rti, rte);
422 11762 : break;
423 55114 : case RTE_FUNCTION:
424 55114 : set_function_size_estimates(root, rel);
425 55114 : break;
426 626 : case RTE_TABLEFUNC:
427 626 : set_tablefunc_size_estimates(root, rel);
428 626 : break;
429 8232 : case RTE_VALUES:
430 8232 : set_values_size_estimates(root, rel);
431 8232 : break;
432 5100 : case RTE_CTE:
433 :
434 : /*
435 : * CTEs don't support making a choice between parameterized
436 : * and unparameterized paths, so just go ahead and build their
437 : * paths immediately.
438 : */
439 5100 : if (rte->self_reference)
440 1010 : set_worktable_pathlist(root, rel, rte);
441 : else
442 4090 : set_cte_pathlist(root, rel, rte);
443 5100 : break;
444 462 : case RTE_NAMEDTUPLESTORE:
445 : /* Might as well just build the path immediately */
446 462 : set_namedtuplestore_pathlist(root, rel, rte);
447 462 : break;
448 4190 : case RTE_RESULT:
449 : /* Might as well just build the path immediately */
450 4190 : set_result_pathlist(root, rel, rte);
451 4190 : break;
452 0 : default:
453 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
454 : break;
455 : }
456 : }
457 :
458 : /*
459 : * We insist that all non-dummy rels have a nonzero rowcount estimate.
460 : */
461 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
462 512966 : }
463 :
464 : /*
465 : * set_rel_pathlist
466 : * Build access paths for a base relation
467 : */
468 : static void
469 513014 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
470 : Index rti, RangeTblEntry *rte)
471 : {
472 513014 : if (IS_DUMMY_REL(rel))
473 : {
474 : /* We already proved the relation empty, so nothing more to do */
475 : }
476 511974 : else if (rte->inh)
477 : {
478 : /* It's an "append relation", process accordingly */
479 23632 : set_append_rel_pathlist(root, rel, rti, rte);
480 : }
481 : else
482 : {
483 488342 : switch (rel->rtekind)
484 : {
485 402976 : case RTE_RELATION:
486 402976 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
487 : {
488 : /* Foreign table */
489 2408 : set_foreign_pathlist(root, rel, rte);
490 : }
491 400568 : else if (rte->tablesample != NULL)
492 : {
493 : /* Sampled relation */
494 306 : set_tablesample_rel_pathlist(root, rel, rte);
495 : }
496 : else
497 : {
498 : /* Plain relation */
499 400262 : set_plain_rel_pathlist(root, rel, rte);
500 : }
501 402976 : break;
502 11642 : case RTE_SUBQUERY:
503 : /* Subquery --- fully handled during set_rel_size */
504 11642 : break;
505 55114 : case RTE_FUNCTION:
506 : /* RangeFunction */
507 55114 : set_function_pathlist(root, rel, rte);
508 55114 : break;
509 626 : case RTE_TABLEFUNC:
510 : /* Table Function */
511 626 : set_tablefunc_pathlist(root, rel, rte);
512 626 : break;
513 8232 : case RTE_VALUES:
514 : /* Values list */
515 8232 : set_values_pathlist(root, rel, rte);
516 8232 : break;
517 5100 : case RTE_CTE:
518 : /* CTE reference --- fully handled during set_rel_size */
519 5100 : break;
520 462 : case RTE_NAMEDTUPLESTORE:
521 : /* tuplestore reference --- fully handled during set_rel_size */
522 462 : break;
523 4190 : case RTE_RESULT:
524 : /* simple Result --- fully handled during set_rel_size */
525 4190 : break;
526 0 : default:
527 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
528 : break;
529 : }
530 : }
531 :
532 : /*
533 : * Allow a plugin to editorialize on the set of Paths for this base
534 : * relation. It could add new paths (such as CustomPaths) by calling
535 : * add_path(), or add_partial_path() if parallel aware. It could also
536 : * delete or modify paths added by the core code.
537 : */
538 513014 : if (set_rel_pathlist_hook)
539 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
540 :
541 : /*
542 : * If this is a baserel, we should normally consider gathering any partial
543 : * paths we may have created for it. We have to do this after calling the
544 : * set_rel_pathlist_hook, else it cannot add partial paths to be included
545 : * here.
546 : *
547 : * However, if this is an inheritance child, skip it. Otherwise, we could
548 : * end up with a very large number of gather nodes, each trying to grab
549 : * its own pool of workers. Instead, we'll consider gathering partial
550 : * paths for the parent appendrel.
551 : *
552 : * Also, if this is the topmost scan/join rel, we postpone gathering until
553 : * the final scan/join targetlist is available (see grouping_planner).
554 : */
555 513014 : if (rel->reloptkind == RELOPT_BASEREL &&
556 463120 : !bms_equal(rel->relids, root->all_query_rels))
557 231614 : generate_useful_gather_paths(root, rel, false);
558 :
559 : /* Now find the cheapest of the paths for this rel */
560 513014 : set_cheapest(rel);
561 :
562 : #ifdef OPTIMIZER_DEBUG
563 : pprint(rel);
564 : #endif
565 513014 : }
566 :
567 : /*
568 : * set_plain_rel_size
569 : * Set size estimates for a plain relation (no subquery, no inheritance)
570 : */
571 : static void
572 400292 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
573 : {
574 : /*
575 : * Test any partial indexes of rel for applicability. We must do this
576 : * first since partial unique indexes can affect size estimates.
577 : */
578 400292 : check_index_predicates(root, rel);
579 :
580 : /* Mark rel with estimated output rows, width, etc */
581 400292 : set_baserel_size_estimates(root, rel);
582 400262 : }
583 :
584 : /*
585 : * If this relation could possibly be scanned from within a worker, then set
586 : * its consider_parallel flag.
587 : */
588 : static void
589 403312 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
590 : RangeTblEntry *rte)
591 : {
592 : /*
593 : * The flag has previously been initialized to false, so we can just
594 : * return if it becomes clear that we can't safely set it.
595 : */
596 : Assert(!rel->consider_parallel);
597 :
598 : /* Don't call this if parallelism is disallowed for the entire query. */
599 : Assert(root->glob->parallelModeOK);
600 :
601 : /* This should only be called for baserels and appendrel children. */
602 : Assert(IS_SIMPLE_REL(rel));
603 :
604 : /* Assorted checks based on rtekind. */
605 403312 : switch (rte->rtekind)
606 : {
607 350388 : case RTE_RELATION:
608 :
609 : /*
610 : * Currently, parallel workers can't access the leader's temporary
611 : * tables. We could possibly relax this if we wrote all of its
612 : * local buffers at the start of the query and made no changes
613 : * thereafter (maybe we could allow hint bit changes), and if we
614 : * taught the workers to read them. Writing a large number of
615 : * temporary buffers could be expensive, though, and we don't have
616 : * the rest of the necessary infrastructure right now anyway. So
617 : * for now, bail out if we see a temporary table.
618 : */
619 350388 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
620 8012 : return;
621 :
622 : /*
623 : * Table sampling can be pushed down to workers if the sample
624 : * function and its arguments are safe.
625 : */
626 342376 : if (rte->tablesample != NULL)
627 : {
628 330 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
629 :
630 330 : if (proparallel != PROPARALLEL_SAFE)
631 36 : return;
632 294 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
633 12 : return;
634 : }
635 :
636 : /*
637 : * Ask FDWs whether they can support performing a ForeignScan
638 : * within a worker. Most often, the answer will be no. For
639 : * example, if the nature of the FDW is such that it opens a TCP
640 : * connection with a remote server, each parallel worker would end
641 : * up with a separate connection, and these connections might not
642 : * be appropriately coordinated between workers and the leader.
643 : */
644 342328 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
645 : {
646 : Assert(rel->fdwroutine);
647 1554 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
648 1482 : return;
649 72 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
650 0 : return;
651 : }
652 :
653 : /*
654 : * There are additional considerations for appendrels, which we'll
655 : * deal with in set_append_rel_size and set_append_rel_pathlist.
656 : * For now, just set consider_parallel based on the rel's own
657 : * quals and targetlist.
658 : */
659 340846 : break;
660 :
661 12872 : case RTE_SUBQUERY:
662 :
663 : /*
664 : * There's no intrinsic problem with scanning a subquery-in-FROM
665 : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
666 : * If the subquery doesn't happen to have any parallel-safe paths,
667 : * then flagging it as consider_parallel won't change anything,
668 : * but that's true for plain tables, too. We must set
669 : * consider_parallel based on the rel's own quals and targetlist,
670 : * so that if a subquery path is parallel-safe but the quals and
671 : * projection we're sticking onto it are not, we correctly mark
672 : * the SubqueryScanPath as not parallel-safe. (Note that
673 : * set_subquery_pathlist() might push some of these quals down
674 : * into the subquery itself, but that doesn't change anything.)
675 : *
676 : * We can't push sub-select containing LIMIT/OFFSET to workers as
677 : * there is no guarantee that the row order will be fully
678 : * deterministic, and applying LIMIT/OFFSET will lead to
679 : * inconsistent results at the top-level. (In some cases, where
680 : * the result is ordered, we could relax this restriction. But it
681 : * doesn't currently seem worth expending extra effort to do so.)
682 : */
683 : {
684 12872 : Query *subquery = castNode(Query, rte->subquery);
685 :
686 12872 : if (limit_needed(subquery))
687 460 : return;
688 : }
689 12412 : break;
690 :
691 0 : case RTE_JOIN:
692 : /* Shouldn't happen; we're only considering baserels here. */
693 : Assert(false);
694 0 : return;
695 :
696 28180 : case RTE_FUNCTION:
697 : /* Check for parallel-restricted functions. */
698 28180 : if (!is_parallel_safe(root, (Node *) rte->functions))
699 13356 : return;
700 14824 : break;
701 :
702 626 : case RTE_TABLEFUNC:
703 : /* not parallel safe */
704 626 : return;
705 :
706 2834 : case RTE_VALUES:
707 : /* Check for parallel-restricted functions. */
708 2834 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
709 12 : return;
710 2822 : break;
711 :
712 4194 : case RTE_CTE:
713 :
714 : /*
715 : * CTE tuplestores aren't shared among parallel workers, so we
716 : * force all CTE scans to happen in the leader. Also, populating
717 : * the CTE would require executing a subplan that's not available
718 : * in the worker, might be parallel-restricted, and must get
719 : * executed only once.
720 : */
721 4194 : return;
722 :
723 434 : case RTE_NAMEDTUPLESTORE:
724 :
725 : /*
726 : * tuplestore cannot be shared, at least without more
727 : * infrastructure to support that.
728 : */
729 434 : return;
730 :
731 3784 : case RTE_RESULT:
732 : /* RESULT RTEs, in themselves, are no problem. */
733 3784 : break;
734 0 : case RTE_GROUP:
735 : /* Shouldn't happen; we're only considering baserels here. */
736 : Assert(false);
737 0 : return;
738 : }
739 :
740 : /*
741 : * If there's anything in baserestrictinfo that's parallel-restricted, we
742 : * give up on parallelizing access to this relation. We could consider
743 : * instead postponing application of the restricted quals until we're
744 : * above all the parallelism in the plan tree, but it's not clear that
745 : * that would be a win in very many cases, and it might be tricky to make
746 : * outer join clauses work correctly. It would likely break equivalence
747 : * classes, too.
748 : */
749 374688 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
750 28456 : return;
751 :
752 : /*
753 : * Likewise, if the relation's outputs are not parallel-safe, give up.
754 : * (Usually, they're just Vars, but sometimes they're not.)
755 : */
756 346232 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
757 18 : return;
758 :
759 : /* We have a winner. */
760 346214 : rel->consider_parallel = true;
761 : }
762 :
763 : /*
764 : * set_plain_rel_pathlist
765 : * Build access paths for a plain relation (no subquery, no inheritance)
766 : */
767 : static void
768 400262 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
769 : {
770 : Relids required_outer;
771 :
772 : /*
773 : * We don't support pushing join clauses into the quals of a seqscan, but
774 : * it could still have required parameterization due to LATERAL refs in
775 : * its tlist.
776 : */
777 400262 : required_outer = rel->lateral_relids;
778 :
779 : /*
780 : * Consider TID scans.
781 : *
782 : * If create_tidscan_paths returns true, then a TID scan path is forced.
783 : * This happens when rel->baserestrictinfo contains CurrentOfExpr, because
784 : * the executor can't handle any other type of path for such queries.
785 : * Hence, we return without adding any other paths.
786 : */
787 400262 : if (create_tidscan_paths(root, rel))
788 404 : return;
789 :
790 : /* Consider sequential scan */
791 399858 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
792 :
793 : /* If appropriate, consider parallel sequential scan */
794 399858 : if (rel->consider_parallel && required_outer == NULL)
795 299422 : create_plain_partial_paths(root, rel);
796 :
797 : /* Consider index scans */
798 399858 : create_index_paths(root, rel);
799 : }
800 :
801 : /*
802 : * create_plain_partial_paths
803 : * Build partial access paths for parallel scan of a plain relation
804 : */
805 : static void
806 299422 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
807 : {
808 : int parallel_workers;
809 :
810 299422 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
811 : max_parallel_workers_per_gather);
812 :
813 : /* If any limit was set to zero, the user doesn't want a parallel scan. */
814 299422 : if (parallel_workers <= 0)
815 273304 : return;
816 :
817 : /* Add an unordered partial path based on a parallel sequential scan. */
818 26118 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
819 : }
820 :
821 : /*
822 : * set_tablesample_rel_size
823 : * Set size estimates for a sampled relation
824 : */
825 : static void
826 306 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
827 : {
828 306 : TableSampleClause *tsc = rte->tablesample;
829 : TsmRoutine *tsm;
830 : BlockNumber pages;
831 : double tuples;
832 :
833 : /*
834 : * Test any partial indexes of rel for applicability. We must do this
835 : * first since partial unique indexes can affect size estimates.
836 : */
837 306 : check_index_predicates(root, rel);
838 :
839 : /*
840 : * Call the sampling method's estimation function to estimate the number
841 : * of pages it will read and the number of tuples it will return. (Note:
842 : * we assume the function returns sane values.)
843 : */
844 306 : tsm = GetTsmRoutine(tsc->tsmhandler);
845 306 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
846 : &pages, &tuples);
847 :
848 : /*
849 : * For the moment, because we will only consider a SampleScan path for the
850 : * rel, it's okay to just overwrite the pages and tuples estimates for the
851 : * whole relation. If we ever consider multiple path types for sampled
852 : * rels, we'll need more complication.
853 : */
854 306 : rel->pages = pages;
855 306 : rel->tuples = tuples;
856 :
857 : /* Mark rel with estimated output rows, width, etc */
858 306 : set_baserel_size_estimates(root, rel);
859 306 : }
860 :
861 : /*
862 : * set_tablesample_rel_pathlist
863 : * Build access paths for a sampled relation
864 : */
865 : static void
866 306 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
867 : {
868 : Relids required_outer;
869 : Path *path;
870 :
871 : /*
872 : * We don't support pushing join clauses into the quals of a samplescan,
873 : * but it could still have required parameterization due to LATERAL refs
874 : * in its tlist or TABLESAMPLE arguments.
875 : */
876 306 : required_outer = rel->lateral_relids;
877 :
878 : /* Consider sampled scan */
879 306 : path = create_samplescan_path(root, rel, required_outer);
880 :
881 : /*
882 : * If the sampling method does not support repeatable scans, we must avoid
883 : * plans that would scan the rel multiple times. Ideally, we'd simply
884 : * avoid putting the rel on the inside of a nestloop join; but adding such
885 : * a consideration to the planner seems like a great deal of complication
886 : * to support an uncommon usage of second-rate sampling methods. Instead,
887 : * if there is a risk that the query might perform an unsafe join, just
888 : * wrap the SampleScan in a Materialize node. We can check for joins by
889 : * counting the membership of all_query_rels (note that this correctly
890 : * counts inheritance trees as single rels). If we're inside a subquery,
891 : * we can't easily check whether a join might occur in the outer query, so
892 : * just assume one is possible.
893 : *
894 : * GetTsmRoutine is relatively expensive compared to the other tests here,
895 : * so check repeatable_across_scans last, even though that's a bit odd.
896 : */
897 586 : if ((root->query_level > 1 ||
898 280 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
899 98 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
900 : {
901 8 : path = (Path *) create_material_path(rel, path);
902 : }
903 :
904 306 : add_path(rel, path);
905 :
906 : /* For the moment, at least, there are no other paths to consider */
907 306 : }
908 :
909 : /*
910 : * set_foreign_size
911 : * Set size estimates for a foreign table RTE
912 : */
913 : static void
914 2412 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
915 : {
916 : /* Mark rel with estimated output rows, width, etc */
917 2412 : set_foreign_size_estimates(root, rel);
918 :
919 : /* Let FDW adjust the size estimates, if it can */
920 2412 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
921 :
922 : /* ... but do not let it set the rows estimate to zero */
923 2408 : rel->rows = clamp_row_est(rel->rows);
924 :
925 : /*
926 : * Also, make sure rel->tuples is not insane relative to rel->rows.
927 : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
928 : * FDW doesn't do anything to replace that.
929 : */
930 2408 : rel->tuples = Max(rel->tuples, rel->rows);
931 2408 : }
932 :
933 : /*
934 : * set_foreign_pathlist
935 : * Build access paths for a foreign table RTE
936 : */
937 : static void
938 2408 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
939 : {
940 : /* Call the FDW's GetForeignPaths function to generate path(s) */
941 2408 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
942 2408 : }
943 :
944 : /*
945 : * set_append_rel_size
946 : * Set size estimates for a simple "append relation"
947 : *
948 : * The passed-in rel and RTE represent the entire append relation. The
949 : * relation's contents are computed by appending together the output of the
950 : * individual member relations. Note that in the non-partitioned inheritance
951 : * case, the first member relation is actually the same table as is mentioned
952 : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
953 : * a good thing because their outputs are not the same size.
954 : */
955 : static void
956 23928 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
957 : Index rti, RangeTblEntry *rte)
958 : {
959 23928 : int parentRTindex = rti;
960 : bool has_live_children;
961 : double parent_tuples;
962 : double parent_rows;
963 : double parent_size;
964 : double *parent_attrsizes;
965 : int nattrs;
966 : ListCell *l;
967 :
968 : /* Guard against stack overflow due to overly deep inheritance tree. */
969 23928 : check_stack_depth();
970 :
971 : Assert(IS_SIMPLE_REL(rel));
972 :
973 : /*
974 : * If this is a partitioned baserel, set the consider_partitionwise_join
975 : * flag; currently, we only consider partitionwise joins with the baserel
976 : * if its targetlist doesn't contain a whole-row Var.
977 : */
978 23928 : if (enable_partitionwise_join &&
979 4328 : rel->reloptkind == RELOPT_BASEREL &&
980 3596 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
981 3596 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
982 3520 : rel->consider_partitionwise_join = true;
983 :
984 : /*
985 : * Initialize to compute size estimates for whole append relation.
986 : *
987 : * We handle tuples estimates by setting "tuples" to the total number of
988 : * tuples accumulated from each live child, rather than using "rows".
989 : * Although an appendrel itself doesn't directly enforce any quals, its
990 : * child relations may. Therefore, setting "tuples" equal to "rows" for
991 : * an appendrel isn't always appropriate, and can lead to inaccurate cost
992 : * estimates. For example, when estimating the number of distinct values
993 : * from an appendrel, we would be unable to adjust the estimate based on
994 : * the restriction selectivity (see estimate_num_groups).
995 : *
996 : * We handle width estimates by weighting the widths of different child
997 : * rels proportionally to their number of rows. This is sensible because
998 : * the use of width estimates is mainly to compute the total relation
999 : * "footprint" if we have to sort or hash it. To do this, we sum the
1000 : * total equivalent size (in "double" arithmetic) and then divide by the
1001 : * total rowcount estimate. This is done separately for the total rel
1002 : * width and each attribute.
1003 : *
1004 : * Note: if you consider changing this logic, beware that child rels could
1005 : * have zero rows and/or width, if they were excluded by constraints.
1006 : */
1007 23928 : has_live_children = false;
1008 23928 : parent_tuples = 0;
1009 23928 : parent_rows = 0;
1010 23928 : parent_size = 0;
1011 23928 : nattrs = rel->max_attr - rel->min_attr + 1;
1012 23928 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1013 :
1014 117596 : foreach(l, root->append_rel_list)
1015 : {
1016 93670 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1017 : int childRTindex;
1018 : RangeTblEntry *childRTE;
1019 : RelOptInfo *childrel;
1020 : List *childrinfos;
1021 : ListCell *parentvars;
1022 : ListCell *childvars;
1023 : ListCell *lc;
1024 :
1025 : /* append_rel_list contains all append rels; ignore others */
1026 93670 : if (appinfo->parent_relid != parentRTindex)
1027 43960 : continue;
1028 :
1029 49962 : childRTindex = appinfo->child_relid;
1030 49962 : childRTE = root->simple_rte_array[childRTindex];
1031 :
1032 : /*
1033 : * The child rel's RelOptInfo was already created during
1034 : * add_other_rels_to_query.
1035 : */
1036 49962 : childrel = find_base_rel(root, childRTindex);
1037 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1038 :
1039 : /* We may have already proven the child to be dummy. */
1040 49962 : if (IS_DUMMY_REL(childrel))
1041 18 : continue;
1042 :
1043 : /*
1044 : * We have to copy the parent's targetlist and quals to the child,
1045 : * with appropriate substitution of variables. However, the
1046 : * baserestrictinfo quals were already copied/substituted when the
1047 : * child RelOptInfo was built. So we don't need any additional setup
1048 : * before applying constraint exclusion.
1049 : */
1050 49944 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1051 : {
1052 : /*
1053 : * This child need not be scanned, so we can omit it from the
1054 : * appendrel.
1055 : */
1056 96 : set_dummy_rel_pathlist(childrel);
1057 96 : continue;
1058 : }
1059 :
1060 : /*
1061 : * Constraint exclusion failed, so copy the parent's join quals and
1062 : * targetlist to the child, with appropriate variable substitutions.
1063 : *
1064 : * We skip join quals that came from above outer joins that can null
1065 : * this rel, since they would be of no value while generating paths
1066 : * for the child. This saves some effort while processing the child
1067 : * rel, and it also avoids an implementation restriction in
1068 : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1069 : */
1070 49848 : childrinfos = NIL;
1071 62832 : foreach(lc, rel->joininfo)
1072 : {
1073 12984 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1074 :
1075 12984 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1076 10662 : childrinfos = lappend(childrinfos,
1077 10662 : adjust_appendrel_attrs(root,
1078 : (Node *) rinfo,
1079 : 1, &appinfo));
1080 : }
1081 49848 : childrel->joininfo = childrinfos;
1082 :
1083 : /*
1084 : * Now for the child's targetlist.
1085 : *
1086 : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1087 : * expressions, which otherwise would not occur in a rel's targetlist.
1088 : * Code that might be looking at an appendrel child must cope with
1089 : * such. (Normally, a rel's targetlist would only include Vars and
1090 : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1091 : * fields of childrel->reltarget; not clear if that would be useful.
1092 : */
1093 99696 : childrel->reltarget->exprs = (List *)
1094 49848 : adjust_appendrel_attrs(root,
1095 49848 : (Node *) rel->reltarget->exprs,
1096 : 1, &appinfo);
1097 :
1098 : /*
1099 : * We have to make child entries in the EquivalenceClass data
1100 : * structures as well. This is needed either if the parent
1101 : * participates in some eclass joins (because we will want to consider
1102 : * inner-indexscan joins on the individual children) or if the parent
1103 : * has useful pathkeys (because we should try to build MergeAppend
1104 : * paths that produce those sort orderings).
1105 : */
1106 49848 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1107 28678 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1108 49848 : childrel->has_eclass_joins = rel->has_eclass_joins;
1109 :
1110 : /*
1111 : * Note: we could compute appropriate attr_needed data for the child's
1112 : * variables, by transforming the parent's attr_needed through the
1113 : * translated_vars mapping. However, currently there's no need
1114 : * because attr_needed is only examined for base relations not
1115 : * otherrels. So we just leave the child's attr_needed empty.
1116 : */
1117 :
1118 : /*
1119 : * If we consider partitionwise joins with the parent rel, do the same
1120 : * for partitioned child rels.
1121 : *
1122 : * Note: here we abuse the consider_partitionwise_join flag by setting
1123 : * it for child rels that are not themselves partitioned. We do so to
1124 : * tell try_partitionwise_join() that the child rel is sufficiently
1125 : * valid to be used as a per-partition input, even if it later gets
1126 : * proven to be dummy. (It's not usable until we've set up the
1127 : * reltarget and EC entries, which we just did.)
1128 : */
1129 49848 : if (rel->consider_partitionwise_join)
1130 11534 : childrel->consider_partitionwise_join = true;
1131 :
1132 : /*
1133 : * If parallelism is allowable for this query in general, see whether
1134 : * it's allowable for this childrel in particular. But if we've
1135 : * already decided the appendrel is not parallel-safe as a whole,
1136 : * there's no point in considering parallelism for this child. For
1137 : * consistency, do this before calling set_rel_size() for the child.
1138 : */
1139 49848 : if (root->glob->parallelModeOK && rel->consider_parallel)
1140 36094 : set_rel_consider_parallel(root, childrel, childRTE);
1141 :
1142 : /*
1143 : * Compute the child's size.
1144 : */
1145 49848 : set_rel_size(root, childrel, childRTindex, childRTE);
1146 :
1147 : /*
1148 : * It is possible that constraint exclusion detected a contradiction
1149 : * within a child subquery, even though we didn't prove one above. If
1150 : * so, we can skip this child.
1151 : */
1152 49846 : if (IS_DUMMY_REL(childrel))
1153 138 : continue;
1154 :
1155 : /* We have at least one live child. */
1156 49708 : has_live_children = true;
1157 :
1158 : /*
1159 : * If any live child is not parallel-safe, treat the whole appendrel
1160 : * as not parallel-safe. In future we might be able to generate plans
1161 : * in which some children are farmed out to workers while others are
1162 : * not; but we don't have that today, so it's a waste to consider
1163 : * partial paths anywhere in the appendrel unless it's all safe.
1164 : * (Child rels visited before this one will be unmarked in
1165 : * set_append_rel_pathlist().)
1166 : */
1167 49708 : if (!childrel->consider_parallel)
1168 14388 : rel->consider_parallel = false;
1169 :
1170 : /*
1171 : * Accumulate size information from each live child.
1172 : */
1173 : Assert(childrel->rows > 0);
1174 :
1175 49708 : parent_tuples += childrel->tuples;
1176 49708 : parent_rows += childrel->rows;
1177 49708 : parent_size += childrel->reltarget->width * childrel->rows;
1178 :
1179 : /*
1180 : * Accumulate per-column estimates too. We need not do anything for
1181 : * PlaceHolderVars in the parent list. If child expression isn't a
1182 : * Var, or we didn't record a width estimate for it, we have to fall
1183 : * back on a datatype-based estimate.
1184 : *
1185 : * By construction, child's targetlist is 1-to-1 with parent's.
1186 : */
1187 150646 : forboth(parentvars, rel->reltarget->exprs,
1188 : childvars, childrel->reltarget->exprs)
1189 : {
1190 100938 : Var *parentvar = (Var *) lfirst(parentvars);
1191 100938 : Node *childvar = (Node *) lfirst(childvars);
1192 :
1193 100938 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1194 : {
1195 88330 : int pndx = parentvar->varattno - rel->min_attr;
1196 88330 : int32 child_width = 0;
1197 :
1198 88330 : if (IsA(childvar, Var) &&
1199 83538 : ((Var *) childvar)->varno == childrel->relid)
1200 : {
1201 83472 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1202 :
1203 83472 : child_width = childrel->attr_widths[cndx];
1204 : }
1205 88330 : if (child_width <= 0)
1206 4858 : child_width = get_typavgwidth(exprType(childvar),
1207 : exprTypmod(childvar));
1208 : Assert(child_width > 0);
1209 88330 : parent_attrsizes[pndx] += child_width * childrel->rows;
1210 : }
1211 : }
1212 : }
1213 :
1214 23926 : if (has_live_children)
1215 : {
1216 : /*
1217 : * Save the finished size estimates.
1218 : */
1219 : int i;
1220 :
1221 : Assert(parent_rows > 0);
1222 23632 : rel->tuples = parent_tuples;
1223 23632 : rel->rows = parent_rows;
1224 23632 : rel->reltarget->width = rint(parent_size / parent_rows);
1225 220536 : for (i = 0; i < nattrs; i++)
1226 196904 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1227 :
1228 : /*
1229 : * Note that we leave rel->pages as zero; this is important to avoid
1230 : * double-counting the appendrel tree in total_table_pages.
1231 : */
1232 : }
1233 : else
1234 : {
1235 : /*
1236 : * All children were excluded by constraints, so mark the whole
1237 : * appendrel dummy. We must do this in this phase so that the rel's
1238 : * dummy-ness is visible when we generate paths for other rels.
1239 : */
1240 294 : set_dummy_rel_pathlist(rel);
1241 : }
1242 :
1243 23926 : pfree(parent_attrsizes);
1244 23926 : }
1245 :
1246 : /*
1247 : * set_append_rel_pathlist
1248 : * Build access paths for an "append relation"
1249 : */
1250 : static void
1251 23632 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1252 : Index rti, RangeTblEntry *rte)
1253 : {
1254 23632 : int parentRTindex = rti;
1255 23632 : List *live_childrels = NIL;
1256 : ListCell *l;
1257 :
1258 : /*
1259 : * Generate access paths for each member relation, and remember the
1260 : * non-dummy children.
1261 : */
1262 116886 : foreach(l, root->append_rel_list)
1263 : {
1264 93254 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1265 : int childRTindex;
1266 : RangeTblEntry *childRTE;
1267 : RelOptInfo *childrel;
1268 :
1269 : /* append_rel_list contains all append rels; ignore others */
1270 93254 : if (appinfo->parent_relid != parentRTindex)
1271 43360 : continue;
1272 :
1273 : /* Re-locate the child RTE and RelOptInfo */
1274 49894 : childRTindex = appinfo->child_relid;
1275 49894 : childRTE = root->simple_rte_array[childRTindex];
1276 49894 : childrel = root->simple_rel_array[childRTindex];
1277 :
1278 : /*
1279 : * If set_append_rel_size() decided the parent appendrel was
1280 : * parallel-unsafe at some point after visiting this child rel, we
1281 : * need to propagate the unsafety marking down to the child, so that
1282 : * we don't generate useless partial paths for it.
1283 : */
1284 49894 : if (!rel->consider_parallel)
1285 14494 : childrel->consider_parallel = false;
1286 :
1287 : /*
1288 : * Compute the child's access paths.
1289 : */
1290 49894 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1291 :
1292 : /*
1293 : * If child is dummy, ignore it.
1294 : */
1295 49894 : if (IS_DUMMY_REL(childrel))
1296 186 : continue;
1297 :
1298 : /*
1299 : * Child is live, so add it to the live_childrels list for use below.
1300 : */
1301 49708 : live_childrels = lappend(live_childrels, childrel);
1302 : }
1303 :
1304 : /* Add paths to the append relation. */
1305 23632 : add_paths_to_append_rel(root, rel, live_childrels);
1306 23632 : }
1307 :
1308 :
1309 : /*
1310 : * add_paths_to_append_rel
1311 : * Generate paths for the given append relation given the set of non-dummy
1312 : * child rels.
1313 : *
1314 : * The function collects all parameterizations and orderings supported by the
1315 : * non-dummy children. For every such parameterization or ordering, it creates
1316 : * an append path collecting one path from each non-dummy child with given
1317 : * parameterization or ordering. Similarly it collects partial paths from
1318 : * non-dummy children to create partial append paths.
1319 : */
1320 : void
1321 38526 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1322 : List *live_childrels)
1323 : {
1324 38526 : List *subpaths = NIL;
1325 38526 : bool subpaths_valid = true;
1326 38526 : List *startup_subpaths = NIL;
1327 38526 : bool startup_subpaths_valid = true;
1328 38526 : List *partial_subpaths = NIL;
1329 38526 : List *pa_partial_subpaths = NIL;
1330 38526 : List *pa_nonpartial_subpaths = NIL;
1331 38526 : bool partial_subpaths_valid = true;
1332 : bool pa_subpaths_valid;
1333 38526 : List *all_child_pathkeys = NIL;
1334 38526 : List *all_child_outers = NIL;
1335 : ListCell *l;
1336 38526 : double partial_rows = -1;
1337 :
1338 : /* If appropriate, consider parallel append */
1339 38526 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1340 :
1341 : /*
1342 : * For every non-dummy child, remember the cheapest path. Also, identify
1343 : * all pathkeys (orderings) and parameterizations (required_outer sets)
1344 : * available for the non-dummy member relations.
1345 : */
1346 117498 : foreach(l, live_childrels)
1347 : {
1348 78972 : RelOptInfo *childrel = lfirst(l);
1349 : ListCell *lcp;
1350 78972 : Path *cheapest_partial_path = NULL;
1351 :
1352 : /*
1353 : * If child has an unparameterized cheapest-total path, add that to
1354 : * the unparameterized Append path we are constructing for the parent.
1355 : * If not, there's no workable unparameterized path.
1356 : *
1357 : * With partitionwise aggregates, the child rel's pathlist may be
1358 : * empty, so don't assume that a path exists here.
1359 : */
1360 78972 : if (childrel->pathlist != NIL &&
1361 78972 : childrel->cheapest_total_path->param_info == NULL)
1362 78240 : accumulate_append_subpath(childrel->cheapest_total_path,
1363 : &subpaths, NULL);
1364 : else
1365 732 : subpaths_valid = false;
1366 :
1367 : /*
1368 : * When the planner is considering cheap startup plans, we'll also
1369 : * collect all the cheapest_startup_paths (if set) and build an
1370 : * AppendPath containing those as subpaths.
1371 : */
1372 78972 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1373 1574 : {
1374 : Path *cheapest_path;
1375 :
1376 : /*
1377 : * With an indication of how many tuples the query should provide,
1378 : * the optimizer tries to choose the path optimal for that
1379 : * specific number of tuples.
1380 : */
1381 1574 : if (root->tuple_fraction > 0.0)
1382 : cheapest_path =
1383 1574 : get_cheapest_fractional_path(childrel,
1384 : root->tuple_fraction);
1385 : else
1386 0 : cheapest_path = childrel->cheapest_startup_path;
1387 :
1388 : /* cheapest_startup_path must not be a parameterized path. */
1389 : Assert(cheapest_path->param_info == NULL);
1390 1574 : accumulate_append_subpath(cheapest_path,
1391 : &startup_subpaths,
1392 : NULL);
1393 : }
1394 : else
1395 77398 : startup_subpaths_valid = false;
1396 :
1397 :
1398 : /* Same idea, but for a partial plan. */
1399 78972 : if (childrel->partial_pathlist != NIL)
1400 : {
1401 47194 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1402 47194 : accumulate_append_subpath(cheapest_partial_path,
1403 : &partial_subpaths, NULL);
1404 : }
1405 : else
1406 31778 : partial_subpaths_valid = false;
1407 :
1408 : /*
1409 : * Same idea, but for a parallel append mixing partial and non-partial
1410 : * paths.
1411 : */
1412 78972 : if (pa_subpaths_valid)
1413 : {
1414 55272 : Path *nppath = NULL;
1415 :
1416 : nppath =
1417 55272 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1418 :
1419 55272 : if (cheapest_partial_path == NULL && nppath == NULL)
1420 : {
1421 : /* Neither a partial nor a parallel-safe path? Forget it. */
1422 518 : pa_subpaths_valid = false;
1423 : }
1424 54754 : else if (nppath == NULL ||
1425 46744 : (cheapest_partial_path != NULL &&
1426 46744 : cheapest_partial_path->total_cost < nppath->total_cost))
1427 : {
1428 : /* Partial path is cheaper or the only option. */
1429 : Assert(cheapest_partial_path != NULL);
1430 46604 : accumulate_append_subpath(cheapest_partial_path,
1431 : &pa_partial_subpaths,
1432 : &pa_nonpartial_subpaths);
1433 : }
1434 : else
1435 : {
1436 : /*
1437 : * Either we've got only a non-partial path, or we think that
1438 : * a single backend can execute the best non-partial path
1439 : * faster than all the parallel backends working together can
1440 : * execute the best partial path.
1441 : *
1442 : * It might make sense to be more aggressive here. Even if
1443 : * the best non-partial path is more expensive than the best
1444 : * partial path, it could still be better to choose the
1445 : * non-partial path if there are several such paths that can
1446 : * be given to different workers. For now, we don't try to
1447 : * figure that out.
1448 : */
1449 8150 : accumulate_append_subpath(nppath,
1450 : &pa_nonpartial_subpaths,
1451 : NULL);
1452 : }
1453 : }
1454 :
1455 : /*
1456 : * Collect lists of all the available path orderings and
1457 : * parameterizations for all the children. We use these as a
1458 : * heuristic to indicate which sort orderings and parameterizations we
1459 : * should build Append and MergeAppend paths for.
1460 : */
1461 179868 : foreach(lcp, childrel->pathlist)
1462 : {
1463 100896 : Path *childpath = (Path *) lfirst(lcp);
1464 100896 : List *childkeys = childpath->pathkeys;
1465 100896 : Relids childouter = PATH_REQ_OUTER(childpath);
1466 :
1467 : /* Unsorted paths don't contribute to pathkey list */
1468 100896 : if (childkeys != NIL)
1469 : {
1470 : ListCell *lpk;
1471 21804 : bool found = false;
1472 :
1473 : /* Have we already seen this ordering? */
1474 21992 : foreach(lpk, all_child_pathkeys)
1475 : {
1476 15768 : List *existing_pathkeys = (List *) lfirst(lpk);
1477 :
1478 15768 : if (compare_pathkeys(existing_pathkeys,
1479 : childkeys) == PATHKEYS_EQUAL)
1480 : {
1481 15580 : found = true;
1482 15580 : break;
1483 : }
1484 : }
1485 21804 : if (!found)
1486 : {
1487 : /* No, so add it to all_child_pathkeys */
1488 6224 : all_child_pathkeys = lappend(all_child_pathkeys,
1489 : childkeys);
1490 : }
1491 : }
1492 :
1493 : /* Unparameterized paths don't contribute to param-set list */
1494 100896 : if (childouter)
1495 : {
1496 : ListCell *lco;
1497 6494 : bool found = false;
1498 :
1499 : /* Have we already seen this param set? */
1500 7160 : foreach(lco, all_child_outers)
1501 : {
1502 4700 : Relids existing_outers = (Relids) lfirst(lco);
1503 :
1504 4700 : if (bms_equal(existing_outers, childouter))
1505 : {
1506 4034 : found = true;
1507 4034 : break;
1508 : }
1509 : }
1510 6494 : if (!found)
1511 : {
1512 : /* No, so add it to all_child_outers */
1513 2460 : all_child_outers = lappend(all_child_outers,
1514 : childouter);
1515 : }
1516 : }
1517 : }
1518 : }
1519 :
1520 : /*
1521 : * If we found unparameterized paths for all children, build an unordered,
1522 : * unparameterized Append path for the rel. (Note: this is correct even
1523 : * if we have zero or one live subpath due to constraint exclusion.)
1524 : */
1525 38526 : if (subpaths_valid)
1526 38214 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1527 : NIL, NULL, 0, false,
1528 : -1));
1529 :
1530 : /* build an AppendPath for the cheap startup paths, if valid */
1531 38526 : if (startup_subpaths_valid)
1532 640 : add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1533 : NIL, NIL, NULL, 0, false, -1));
1534 :
1535 : /*
1536 : * Consider an append of unordered, unparameterized partial paths. Make
1537 : * it parallel-aware if possible.
1538 : */
1539 38526 : if (partial_subpaths_valid && partial_subpaths != NIL)
1540 : {
1541 : AppendPath *appendpath;
1542 : ListCell *lc;
1543 21354 : int parallel_workers = 0;
1544 :
1545 : /* Find the highest number of workers requested for any subpath. */
1546 72662 : foreach(lc, partial_subpaths)
1547 : {
1548 51308 : Path *path = lfirst(lc);
1549 :
1550 51308 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1551 : }
1552 : Assert(parallel_workers > 0);
1553 :
1554 : /*
1555 : * If the use of parallel append is permitted, always request at least
1556 : * log2(# of children) workers. We assume it can be useful to have
1557 : * extra workers in this case because they will be spread out across
1558 : * the children. The precise formula is just a guess, but we don't
1559 : * want to end up with a radically different answer for a table with N
1560 : * partitions vs. an unpartitioned table with the same data, so the
1561 : * use of some kind of log-scaling here seems to make some sense.
1562 : */
1563 21354 : if (enable_parallel_append)
1564 : {
1565 21306 : parallel_workers = Max(parallel_workers,
1566 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1567 21306 : parallel_workers = Min(parallel_workers,
1568 : max_parallel_workers_per_gather);
1569 : }
1570 : Assert(parallel_workers > 0);
1571 :
1572 : /* Generate a partial append path. */
1573 21354 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1574 : NIL, NULL, parallel_workers,
1575 : enable_parallel_append,
1576 : -1);
1577 :
1578 : /*
1579 : * Make sure any subsequent partial paths use the same row count
1580 : * estimate.
1581 : */
1582 21354 : partial_rows = appendpath->path.rows;
1583 :
1584 : /* Add the path. */
1585 21354 : add_partial_path(rel, (Path *) appendpath);
1586 : }
1587 :
1588 : /*
1589 : * Consider a parallel-aware append using a mix of partial and non-partial
1590 : * paths. (This only makes sense if there's at least one child which has
1591 : * a non-partial path that is substantially cheaper than any partial path;
1592 : * otherwise, we should use the append path added in the previous step.)
1593 : */
1594 38526 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1595 : {
1596 : AppendPath *appendpath;
1597 : ListCell *lc;
1598 4146 : int parallel_workers = 0;
1599 :
1600 : /*
1601 : * Find the highest number of workers requested for any partial
1602 : * subpath.
1603 : */
1604 5048 : foreach(lc, pa_partial_subpaths)
1605 : {
1606 902 : Path *path = lfirst(lc);
1607 :
1608 902 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1609 : }
1610 :
1611 : /*
1612 : * Same formula here as above. It's even more important in this
1613 : * instance because the non-partial paths won't contribute anything to
1614 : * the planned number of parallel workers.
1615 : */
1616 4146 : parallel_workers = Max(parallel_workers,
1617 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1618 4146 : parallel_workers = Min(parallel_workers,
1619 : max_parallel_workers_per_gather);
1620 : Assert(parallel_workers > 0);
1621 :
1622 4146 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1623 : pa_partial_subpaths,
1624 : NIL, NULL, parallel_workers, true,
1625 : partial_rows);
1626 4146 : add_partial_path(rel, (Path *) appendpath);
1627 : }
1628 :
1629 : /*
1630 : * Also build unparameterized ordered append paths based on the collected
1631 : * list of child pathkeys.
1632 : */
1633 38526 : if (subpaths_valid)
1634 38214 : generate_orderedappend_paths(root, rel, live_childrels,
1635 : all_child_pathkeys);
1636 :
1637 : /*
1638 : * Build Append paths for each parameterization seen among the child rels.
1639 : * (This may look pretty expensive, but in most cases of practical
1640 : * interest, the child rels will expose mostly the same parameterizations,
1641 : * so that not that many cases actually get considered here.)
1642 : *
1643 : * The Append node itself cannot enforce quals, so all qual checking must
1644 : * be done in the child paths. This means that to have a parameterized
1645 : * Append path, we must have the exact same parameterization for each
1646 : * child path; otherwise some children might be failing to check the
1647 : * moved-down quals. To make them match up, we can try to increase the
1648 : * parameterization of lesser-parameterized paths.
1649 : */
1650 40986 : foreach(l, all_child_outers)
1651 : {
1652 2460 : Relids required_outer = (Relids) lfirst(l);
1653 : ListCell *lcr;
1654 :
1655 : /* Select the child paths for an Append with this parameterization */
1656 2460 : subpaths = NIL;
1657 2460 : subpaths_valid = true;
1658 9044 : foreach(lcr, live_childrels)
1659 : {
1660 6596 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1661 : Path *subpath;
1662 :
1663 6596 : if (childrel->pathlist == NIL)
1664 : {
1665 : /* failed to make a suitable path for this child */
1666 0 : subpaths_valid = false;
1667 0 : break;
1668 : }
1669 :
1670 6596 : subpath = get_cheapest_parameterized_child_path(root,
1671 : childrel,
1672 : required_outer);
1673 6596 : if (subpath == NULL)
1674 : {
1675 : /* failed to make a suitable path for this child */
1676 12 : subpaths_valid = false;
1677 12 : break;
1678 : }
1679 6584 : accumulate_append_subpath(subpath, &subpaths, NULL);
1680 : }
1681 :
1682 2460 : if (subpaths_valid)
1683 2448 : add_path(rel, (Path *)
1684 2448 : create_append_path(root, rel, subpaths, NIL,
1685 : NIL, required_outer, 0, false,
1686 : -1));
1687 : }
1688 :
1689 : /*
1690 : * When there is only a single child relation, the Append path can inherit
1691 : * any ordering available for the child rel's path, so that it's useful to
1692 : * consider ordered partial paths. Above we only considered the cheapest
1693 : * partial path for each child, but let's also make paths using any
1694 : * partial paths that have pathkeys.
1695 : */
1696 38526 : if (list_length(live_childrels) == 1)
1697 : {
1698 14148 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1699 :
1700 : /* skip the cheapest partial path, since we already used that above */
1701 14352 : for_each_from(l, childrel->partial_pathlist, 1)
1702 : {
1703 204 : Path *path = (Path *) lfirst(l);
1704 : AppendPath *appendpath;
1705 :
1706 : /* skip paths with no pathkeys. */
1707 204 : if (path->pathkeys == NIL)
1708 0 : continue;
1709 :
1710 204 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1711 : NIL, NULL,
1712 : path->parallel_workers, true,
1713 : partial_rows);
1714 204 : add_partial_path(rel, (Path *) appendpath);
1715 : }
1716 : }
1717 38526 : }
1718 :
1719 : /*
1720 : * generate_orderedappend_paths
1721 : * Generate ordered append paths for an append relation
1722 : *
1723 : * Usually we generate MergeAppend paths here, but there are some special
1724 : * cases where we can generate simple Append paths, because the subpaths
1725 : * can provide tuples in the required order already.
1726 : *
1727 : * We generate a path for each ordering (pathkey list) appearing in
1728 : * all_child_pathkeys.
1729 : *
1730 : * We consider both cheapest-startup and cheapest-total cases, ie, for each
1731 : * interesting ordering, collect all the cheapest startup subpaths and all the
1732 : * cheapest total paths, and build a suitable path for each case.
1733 : *
1734 : * We don't currently generate any parameterized ordered paths here. While
1735 : * it would not take much more code here to do so, it's very unclear that it
1736 : * is worth the planning cycles to investigate such paths: there's little
1737 : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1738 : * that the current coding of add_path would reject such paths out of hand,
1739 : * because add_path gives no credit for sort ordering of parameterized paths,
1740 : * and a parameterized MergeAppend is going to be more expensive than the
1741 : * corresponding parameterized Append path. If we ever try harder to support
1742 : * parameterized mergejoin plans, it might be worth adding support for
1743 : * parameterized paths here to feed such joins. (See notes in
1744 : * optimizer/README for why that might not ever happen, though.)
1745 : */
1746 : static void
1747 38214 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1748 : List *live_childrels,
1749 : List *all_child_pathkeys)
1750 : {
1751 : ListCell *lcp;
1752 38214 : List *partition_pathkeys = NIL;
1753 38214 : List *partition_pathkeys_desc = NIL;
1754 38214 : bool partition_pathkeys_partial = true;
1755 38214 : bool partition_pathkeys_desc_partial = true;
1756 :
1757 : /*
1758 : * Some partitioned table setups may allow us to use an Append node
1759 : * instead of a MergeAppend. This is possible in cases such as RANGE
1760 : * partitioned tables where it's guaranteed that an earlier partition must
1761 : * contain rows which come earlier in the sort order. To detect whether
1762 : * this is relevant, build pathkey descriptions of the partition ordering,
1763 : * for both forward and reverse scans.
1764 : */
1765 65592 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
1766 27378 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1767 : {
1768 22798 : partition_pathkeys = build_partition_pathkeys(root, rel,
1769 : ForwardScanDirection,
1770 : &partition_pathkeys_partial);
1771 :
1772 22798 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1773 : BackwardScanDirection,
1774 : &partition_pathkeys_desc_partial);
1775 :
1776 : /*
1777 : * You might think we should truncate_useless_pathkeys here, but
1778 : * allowing partition keys which are a subset of the query's pathkeys
1779 : * can often be useful. For example, consider a table partitioned by
1780 : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1781 : * paths that can produce the a, b, c ordering (perhaps via indexes on
1782 : * (a, b, c)) then it works to consider the appendrel output as
1783 : * ordered by a, b, c.
1784 : */
1785 : }
1786 :
1787 : /* Now consider each interesting sort ordering */
1788 44378 : foreach(lcp, all_child_pathkeys)
1789 : {
1790 6164 : List *pathkeys = (List *) lfirst(lcp);
1791 6164 : List *startup_subpaths = NIL;
1792 6164 : List *total_subpaths = NIL;
1793 6164 : List *fractional_subpaths = NIL;
1794 6164 : bool startup_neq_total = false;
1795 : bool match_partition_order;
1796 : bool match_partition_order_desc;
1797 : int end_index;
1798 : int first_index;
1799 : int direction;
1800 :
1801 : /*
1802 : * Determine if this sort ordering matches any partition pathkeys we
1803 : * have, for both ascending and descending partition order. If the
1804 : * partition pathkeys happen to be contained in pathkeys then it still
1805 : * works, as described above, providing that the partition pathkeys
1806 : * are complete and not just a prefix of the partition keys. (In such
1807 : * cases we'll be relying on the child paths to have sorted the
1808 : * lower-order columns of the required pathkeys.)
1809 : */
1810 6164 : match_partition_order =
1811 10080 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1812 4088 : (!partition_pathkeys_partial &&
1813 172 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1814 :
1815 13810 : match_partition_order_desc = !match_partition_order &&
1816 3838 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1817 3872 : (!partition_pathkeys_desc_partial &&
1818 64 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1819 :
1820 : /*
1821 : * When the required pathkeys match the reverse of the partition
1822 : * order, we must build the list of paths in reverse starting with the
1823 : * last matching partition first. We can get away without making any
1824 : * special cases for this in the loop below by just looping backward
1825 : * over the child relations in this case.
1826 : */
1827 6164 : if (match_partition_order_desc)
1828 : {
1829 : /* loop backward */
1830 42 : first_index = list_length(live_childrels) - 1;
1831 42 : end_index = -1;
1832 42 : direction = -1;
1833 :
1834 : /*
1835 : * Set this to true to save us having to check for
1836 : * match_partition_order_desc in the loop below.
1837 : */
1838 42 : match_partition_order = true;
1839 : }
1840 : else
1841 : {
1842 : /* for all other case, loop forward */
1843 6122 : first_index = 0;
1844 6122 : end_index = list_length(live_childrels);
1845 6122 : direction = 1;
1846 : }
1847 :
1848 : /* Select the child paths for this ordering... */
1849 22682 : for (int i = first_index; i != end_index; i += direction)
1850 : {
1851 16518 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1852 : Path *cheapest_startup,
1853 : *cheapest_total,
1854 16518 : *cheapest_fractional = NULL;
1855 :
1856 : /* Locate the right paths, if they are available. */
1857 : cheapest_startup =
1858 16518 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1859 : pathkeys,
1860 : NULL,
1861 : STARTUP_COST,
1862 : false);
1863 : cheapest_total =
1864 16518 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1865 : pathkeys,
1866 : NULL,
1867 : TOTAL_COST,
1868 : false);
1869 :
1870 : /*
1871 : * If we can't find any paths with the right order just use the
1872 : * cheapest-total path; we'll have to sort it later.
1873 : */
1874 16518 : if (cheapest_startup == NULL || cheapest_total == NULL)
1875 : {
1876 316 : cheapest_startup = cheapest_total =
1877 : childrel->cheapest_total_path;
1878 : /* Assert we do have an unparameterized path for this child */
1879 : Assert(cheapest_total->param_info == NULL);
1880 : }
1881 :
1882 : /*
1883 : * When building a fractional path, determine a cheapest
1884 : * fractional path for each child relation too. Looking at startup
1885 : * and total costs is not enough, because the cheapest fractional
1886 : * path may be dominated by two separate paths (one for startup,
1887 : * one for total).
1888 : *
1889 : * When needed (building fractional path), determine the cheapest
1890 : * fractional path too.
1891 : */
1892 16518 : if (root->tuple_fraction > 0)
1893 : {
1894 812 : double path_fraction = (1.0 / root->tuple_fraction);
1895 :
1896 : cheapest_fractional =
1897 812 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1898 : pathkeys,
1899 : NULL,
1900 : path_fraction);
1901 :
1902 : /*
1903 : * If we found no path with matching pathkeys, use the
1904 : * cheapest total path instead.
1905 : *
1906 : * XXX We might consider partially sorted paths too (with an
1907 : * incremental sort on top). But we'd have to build all the
1908 : * incremental paths, do the costing etc.
1909 : */
1910 812 : if (!cheapest_fractional)
1911 44 : cheapest_fractional = cheapest_total;
1912 : }
1913 :
1914 : /*
1915 : * Notice whether we actually have different paths for the
1916 : * "cheapest" and "total" cases; frequently there will be no point
1917 : * in two create_merge_append_path() calls.
1918 : */
1919 16518 : if (cheapest_startup != cheapest_total)
1920 72 : startup_neq_total = true;
1921 :
1922 : /*
1923 : * Collect the appropriate child paths. The required logic varies
1924 : * for the Append and MergeAppend cases.
1925 : */
1926 16518 : if (match_partition_order)
1927 : {
1928 : /*
1929 : * We're going to make a plain Append path. We don't need
1930 : * most of what accumulate_append_subpath would do, but we do
1931 : * want to cut out child Appends or MergeAppends if they have
1932 : * just a single subpath (and hence aren't doing anything
1933 : * useful).
1934 : */
1935 6366 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1936 6366 : cheapest_total = get_singleton_append_subpath(cheapest_total);
1937 :
1938 6366 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1939 6366 : total_subpaths = lappend(total_subpaths, cheapest_total);
1940 :
1941 6366 : if (cheapest_fractional)
1942 : {
1943 120 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1944 120 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1945 : }
1946 : }
1947 : else
1948 : {
1949 : /*
1950 : * Otherwise, rely on accumulate_append_subpath to collect the
1951 : * child paths for the MergeAppend.
1952 : */
1953 10152 : accumulate_append_subpath(cheapest_startup,
1954 : &startup_subpaths, NULL);
1955 10152 : accumulate_append_subpath(cheapest_total,
1956 : &total_subpaths, NULL);
1957 :
1958 10152 : if (cheapest_fractional)
1959 692 : accumulate_append_subpath(cheapest_fractional,
1960 : &fractional_subpaths, NULL);
1961 : }
1962 : }
1963 :
1964 : /* ... and build the Append or MergeAppend paths */
1965 6164 : if (match_partition_order)
1966 : {
1967 : /* We only need Append */
1968 2368 : add_path(rel, (Path *) create_append_path(root,
1969 : rel,
1970 : startup_subpaths,
1971 : NIL,
1972 : pathkeys,
1973 : NULL,
1974 : 0,
1975 : false,
1976 : -1));
1977 2368 : if (startup_neq_total)
1978 0 : add_path(rel, (Path *) create_append_path(root,
1979 : rel,
1980 : total_subpaths,
1981 : NIL,
1982 : pathkeys,
1983 : NULL,
1984 : 0,
1985 : false,
1986 : -1));
1987 :
1988 2368 : if (fractional_subpaths)
1989 60 : add_path(rel, (Path *) create_append_path(root,
1990 : rel,
1991 : fractional_subpaths,
1992 : NIL,
1993 : pathkeys,
1994 : NULL,
1995 : 0,
1996 : false,
1997 : -1));
1998 : }
1999 : else
2000 : {
2001 : /* We need MergeAppend */
2002 3796 : add_path(rel, (Path *) create_merge_append_path(root,
2003 : rel,
2004 : startup_subpaths,
2005 : pathkeys,
2006 : NULL));
2007 3796 : if (startup_neq_total)
2008 48 : add_path(rel, (Path *) create_merge_append_path(root,
2009 : rel,
2010 : total_subpaths,
2011 : pathkeys,
2012 : NULL));
2013 :
2014 3796 : if (fractional_subpaths)
2015 244 : add_path(rel, (Path *) create_merge_append_path(root,
2016 : rel,
2017 : fractional_subpaths,
2018 : pathkeys,
2019 : NULL));
2020 : }
2021 : }
2022 38214 : }
2023 :
2024 : /*
2025 : * get_cheapest_parameterized_child_path
2026 : * Get cheapest path for this relation that has exactly the requested
2027 : * parameterization.
2028 : *
2029 : * Returns NULL if unable to create such a path.
2030 : */
2031 : static Path *
2032 6596 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2033 : Relids required_outer)
2034 : {
2035 : Path *cheapest;
2036 : ListCell *lc;
2037 :
2038 : /*
2039 : * Look up the cheapest existing path with no more than the needed
2040 : * parameterization. If it has exactly the needed parameterization, we're
2041 : * done.
2042 : */
2043 6596 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2044 : NIL,
2045 : required_outer,
2046 : TOTAL_COST,
2047 : false);
2048 : Assert(cheapest != NULL);
2049 6596 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2050 6262 : return cheapest;
2051 :
2052 : /*
2053 : * Otherwise, we can "reparameterize" an existing path to match the given
2054 : * parameterization, which effectively means pushing down additional
2055 : * joinquals to be checked within the path's scan. However, some existing
2056 : * paths might check the available joinquals already while others don't;
2057 : * therefore, it's not clear which existing path will be cheapest after
2058 : * reparameterization. We have to go through them all and find out.
2059 : */
2060 334 : cheapest = NULL;
2061 1168 : foreach(lc, rel->pathlist)
2062 : {
2063 834 : Path *path = (Path *) lfirst(lc);
2064 :
2065 : /* Can't use it if it needs more than requested parameterization */
2066 834 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2067 24 : continue;
2068 :
2069 : /*
2070 : * Reparameterization can only increase the path's cost, so if it's
2071 : * already more expensive than the current cheapest, forget it.
2072 : */
2073 1266 : if (cheapest != NULL &&
2074 456 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2075 384 : continue;
2076 :
2077 : /* Reparameterize if needed, then recheck cost */
2078 426 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2079 : {
2080 350 : path = reparameterize_path(root, path, required_outer, 1.0);
2081 350 : if (path == NULL)
2082 32 : continue; /* failed to reparameterize this one */
2083 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2084 :
2085 318 : if (cheapest != NULL &&
2086 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2087 0 : continue;
2088 : }
2089 :
2090 : /* We have a new best path */
2091 394 : cheapest = path;
2092 : }
2093 :
2094 : /* Return the best path, or NULL if we found no suitable candidate */
2095 334 : return cheapest;
2096 : }
2097 :
2098 : /*
2099 : * accumulate_append_subpath
2100 : * Add a subpath to the list being built for an Append or MergeAppend.
2101 : *
2102 : * It's possible that the child is itself an Append or MergeAppend path, in
2103 : * which case we can "cut out the middleman" and just add its child paths to
2104 : * our own list. (We don't try to do this earlier because we need to apply
2105 : * both levels of transformation to the quals.)
2106 : *
2107 : * Note that if we omit a child MergeAppend in this way, we are effectively
2108 : * omitting a sort step, which seems fine: if the parent is to be an Append,
2109 : * its result would be unsorted anyway, while if the parent is to be a
2110 : * MergeAppend, there's no point in a separate sort on a child.
2111 : *
2112 : * Normally, either path is a partial path and subpaths is a list of partial
2113 : * paths, or else path is a non-partial plan and subpaths is a list of those.
2114 : * However, if path is a parallel-aware Append, then we add its partial path
2115 : * children to subpaths and the rest to special_subpaths. If the latter is
2116 : * NULL, we don't flatten the path at all (unless it contains only partial
2117 : * paths).
2118 : */
2119 : static void
2120 209342 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2121 : {
2122 209342 : if (IsA(path, AppendPath))
2123 : {
2124 14212 : AppendPath *apath = (AppendPath *) path;
2125 :
2126 14212 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2127 : {
2128 14020 : *subpaths = list_concat(*subpaths, apath->subpaths);
2129 14020 : return;
2130 : }
2131 192 : else if (special_subpaths != NULL)
2132 : {
2133 : List *new_special_subpaths;
2134 :
2135 : /* Split Parallel Append into partial and non-partial subpaths */
2136 96 : *subpaths = list_concat(*subpaths,
2137 96 : list_copy_tail(apath->subpaths,
2138 : apath->first_partial_path));
2139 96 : new_special_subpaths = list_copy_head(apath->subpaths,
2140 : apath->first_partial_path);
2141 96 : *special_subpaths = list_concat(*special_subpaths,
2142 : new_special_subpaths);
2143 96 : return;
2144 : }
2145 : }
2146 195130 : else if (IsA(path, MergeAppendPath))
2147 : {
2148 696 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2149 :
2150 696 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2151 696 : return;
2152 : }
2153 :
2154 194530 : *subpaths = lappend(*subpaths, path);
2155 : }
2156 :
2157 : /*
2158 : * get_singleton_append_subpath
2159 : * Returns the single subpath of an Append/MergeAppend, or just
2160 : * return 'path' if it's not a single sub-path Append/MergeAppend.
2161 : *
2162 : * Note: 'path' must not be a parallel-aware path.
2163 : */
2164 : static Path *
2165 12852 : get_singleton_append_subpath(Path *path)
2166 : {
2167 : Assert(!path->parallel_aware);
2168 :
2169 12852 : if (IsA(path, AppendPath))
2170 : {
2171 404 : AppendPath *apath = (AppendPath *) path;
2172 :
2173 404 : if (list_length(apath->subpaths) == 1)
2174 168 : return (Path *) linitial(apath->subpaths);
2175 : }
2176 12448 : else if (IsA(path, MergeAppendPath))
2177 : {
2178 348 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2179 :
2180 348 : if (list_length(mpath->subpaths) == 1)
2181 0 : return (Path *) linitial(mpath->subpaths);
2182 : }
2183 :
2184 12684 : return path;
2185 : }
2186 :
2187 : /*
2188 : * set_dummy_rel_pathlist
2189 : * Build a dummy path for a relation that's been excluded by constraints
2190 : *
2191 : * Rather than inventing a special "dummy" path type, we represent this as an
2192 : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2193 : *
2194 : * (See also mark_dummy_rel, which does basically the same thing, but is
2195 : * typically used to change a rel into dummy state after we already made
2196 : * paths for it.)
2197 : */
2198 : static void
2199 1088 : set_dummy_rel_pathlist(RelOptInfo *rel)
2200 : {
2201 : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2202 1088 : rel->rows = 0;
2203 1088 : rel->reltarget->width = 0;
2204 :
2205 : /* Discard any pre-existing paths; no further need for them */
2206 1088 : rel->pathlist = NIL;
2207 1088 : rel->partial_pathlist = NIL;
2208 :
2209 : /* Set up the dummy path */
2210 1088 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2211 : NIL, rel->lateral_relids,
2212 : 0, false, -1));
2213 :
2214 : /*
2215 : * We set the cheapest-path fields immediately, just in case they were
2216 : * pointing at some discarded path. This is redundant in current usage
2217 : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2218 : * for safety and consistency with mark_dummy_rel.
2219 : */
2220 1088 : set_cheapest(rel);
2221 1088 : }
2222 :
2223 : /*
2224 : * find_window_run_conditions
2225 : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2226 : * function to determine the function's monotonic properties. We then
2227 : * see if 'opexpr' can be used to short-circuit execution.
2228 : *
2229 : * For example row_number() over (order by ...) always produces a value one
2230 : * higher than the previous. If someone has a window function in a subquery
2231 : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2232 : * as well stop processing the windowagg once the row number reaches 11. Here
2233 : * we check if 'opexpr' might help us to stop doing needless extra processing
2234 : * in WindowAgg nodes.
2235 : *
2236 : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2237 : * its original purpose. This is set to false if the caller can assume that
2238 : * the run condition will handle all of the required filtering.
2239 : *
2240 : * Returns true if 'opexpr' was found to be useful and was added to the
2241 : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2242 : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2243 : * If the 'opexpr' cannot be used then we set *keep_original to true and
2244 : * return false.
2245 : */
2246 : static bool
2247 240 : find_window_run_conditions(Query *subquery, RangeTblEntry *rte, Index rti,
2248 : AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2249 : bool wfunc_left, bool *keep_original,
2250 : Bitmapset **run_cond_attrs)
2251 : {
2252 : Oid prosupport;
2253 : Expr *otherexpr;
2254 : SupportRequestWFuncMonotonic req;
2255 : SupportRequestWFuncMonotonic *res;
2256 : WindowClause *wclause;
2257 : List *opinfos;
2258 : OpExpr *runopexpr;
2259 : Oid runoperator;
2260 : ListCell *lc;
2261 :
2262 240 : *keep_original = true;
2263 :
2264 240 : while (IsA(wfunc, RelabelType))
2265 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2266 :
2267 : /* we can only work with window functions */
2268 240 : if (!IsA(wfunc, WindowFunc))
2269 24 : return false;
2270 :
2271 : /* can't use it if there are subplans in the WindowFunc */
2272 216 : if (contain_subplans((Node *) wfunc))
2273 6 : return false;
2274 :
2275 210 : prosupport = get_func_support(wfunc->winfnoid);
2276 :
2277 : /* Check if there's a support function for 'wfunc' */
2278 210 : if (!OidIsValid(prosupport))
2279 18 : return false;
2280 :
2281 : /* get the Expr from the other side of the OpExpr */
2282 192 : if (wfunc_left)
2283 168 : otherexpr = lsecond(opexpr->args);
2284 : else
2285 24 : otherexpr = linitial(opexpr->args);
2286 :
2287 : /*
2288 : * The value being compared must not change during the evaluation of the
2289 : * window partition.
2290 : */
2291 192 : if (!is_pseudo_constant_clause((Node *) otherexpr))
2292 0 : return false;
2293 :
2294 : /* find the window clause belonging to the window function */
2295 192 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2296 192 : wfunc->winref - 1);
2297 :
2298 192 : req.type = T_SupportRequestWFuncMonotonic;
2299 192 : req.window_func = wfunc;
2300 192 : req.window_clause = wclause;
2301 :
2302 : /* call the support function */
2303 : res = (SupportRequestWFuncMonotonic *)
2304 192 : DatumGetPointer(OidFunctionCall1(prosupport,
2305 : PointerGetDatum(&req)));
2306 :
2307 : /*
2308 : * Nothing to do if the function is neither monotonically increasing nor
2309 : * monotonically decreasing.
2310 : */
2311 192 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2312 0 : return false;
2313 :
2314 192 : runopexpr = NULL;
2315 192 : runoperator = InvalidOid;
2316 192 : opinfos = get_op_index_interpretation(opexpr->opno);
2317 :
2318 192 : foreach(lc, opinfos)
2319 : {
2320 192 : OpIndexInterpretation *opinfo = (OpIndexInterpretation *) lfirst(lc);
2321 192 : CompareType cmptype = opinfo->cmptype;
2322 :
2323 : /* handle < / <= */
2324 192 : if (cmptype == COMPARE_LT || cmptype == COMPARE_LE)
2325 : {
2326 : /*
2327 : * < / <= is supported for monotonically increasing functions in
2328 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2329 : * for monotonically decreasing functions.
2330 : */
2331 138 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2332 18 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2333 : {
2334 126 : *keep_original = false;
2335 126 : runopexpr = opexpr;
2336 126 : runoperator = opexpr->opno;
2337 : }
2338 138 : break;
2339 : }
2340 : /* handle > / >= */
2341 54 : else if (cmptype == COMPARE_GT || cmptype == COMPARE_GE)
2342 : {
2343 : /*
2344 : * > / >= is supported for monotonically decreasing functions in
2345 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2346 : * for monotonically increasing functions.
2347 : */
2348 18 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2349 12 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2350 : {
2351 18 : *keep_original = false;
2352 18 : runopexpr = opexpr;
2353 18 : runoperator = opexpr->opno;
2354 : }
2355 18 : break;
2356 : }
2357 : /* handle = */
2358 36 : else if (cmptype == COMPARE_EQ)
2359 : {
2360 : CompareType newcmptype;
2361 :
2362 : /*
2363 : * When both monotonically increasing and decreasing then the
2364 : * return value of the window function will be the same each time.
2365 : * We can simply use 'opexpr' as the run condition without
2366 : * modifying it.
2367 : */
2368 36 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2369 : {
2370 6 : *keep_original = false;
2371 6 : runopexpr = opexpr;
2372 6 : runoperator = opexpr->opno;
2373 6 : break;
2374 : }
2375 :
2376 : /*
2377 : * When monotonically increasing we make a qual with <wfunc> <=
2378 : * <value> or <value> >= <wfunc> in order to filter out values
2379 : * which are above the value in the equality condition. For
2380 : * monotonically decreasing functions we want to filter values
2381 : * below the value in the equality condition.
2382 : */
2383 30 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2384 30 : newcmptype = wfunc_left ? COMPARE_LE : COMPARE_GE;
2385 : else
2386 0 : newcmptype = wfunc_left ? COMPARE_GE : COMPARE_LE;
2387 :
2388 : /* We must keep the original equality qual */
2389 30 : *keep_original = true;
2390 30 : runopexpr = opexpr;
2391 :
2392 : /* determine the operator to use for the WindowFuncRunCondition */
2393 30 : runoperator = get_opfamily_member_for_cmptype(opinfo->opfamily_id,
2394 : opinfo->oplefttype,
2395 : opinfo->oprighttype,
2396 : newcmptype);
2397 30 : break;
2398 : }
2399 : }
2400 :
2401 192 : if (runopexpr != NULL)
2402 : {
2403 : WindowFuncRunCondition *wfuncrc;
2404 :
2405 180 : wfuncrc = makeNode(WindowFuncRunCondition);
2406 180 : wfuncrc->opno = runoperator;
2407 180 : wfuncrc->inputcollid = runopexpr->inputcollid;
2408 180 : wfuncrc->wfunc_left = wfunc_left;
2409 180 : wfuncrc->arg = copyObject(otherexpr);
2410 :
2411 180 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2412 :
2413 : /* record that this attno was used in a run condition */
2414 180 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2415 : attno - FirstLowInvalidHeapAttributeNumber);
2416 180 : return true;
2417 : }
2418 :
2419 : /* unsupported OpExpr */
2420 12 : return false;
2421 : }
2422 :
2423 : /*
2424 : * check_and_push_window_quals
2425 : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2426 : * as a 'runCondition' qual. These, when present, allow some unnecessary
2427 : * work to be skipped during execution.
2428 : *
2429 : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2430 : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2431 : * window quals for.
2432 : *
2433 : * Returns true if the caller still must keep the original qual or false if
2434 : * the caller can safely ignore the original qual because the WindowAgg node
2435 : * will use the runCondition to stop returning tuples.
2436 : */
2437 : static bool
2438 252 : check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti,
2439 : Node *clause, Bitmapset **run_cond_attrs)
2440 : {
2441 252 : OpExpr *opexpr = (OpExpr *) clause;
2442 252 : bool keep_original = true;
2443 : Var *var1;
2444 : Var *var2;
2445 :
2446 : /* We're only able to use OpExprs with 2 operands */
2447 252 : if (!IsA(opexpr, OpExpr))
2448 18 : return true;
2449 :
2450 234 : if (list_length(opexpr->args) != 2)
2451 0 : return true;
2452 :
2453 : /*
2454 : * Currently, we restrict this optimization to strict OpExprs. The reason
2455 : * for this is that during execution, once the runcondition becomes false,
2456 : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2457 : * function result values, we set them to NULL. Having only strict
2458 : * OpExprs here ensures that we properly filter out the tuples with NULLs
2459 : * in the top-level WindowAgg.
2460 : */
2461 234 : set_opfuncid(opexpr);
2462 234 : if (!func_strict(opexpr->opfuncid))
2463 0 : return true;
2464 :
2465 : /*
2466 : * Check for plain Vars that reference window functions in the subquery.
2467 : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2468 : * be used as part of the run condition.
2469 : */
2470 :
2471 : /* Check the left side of the OpExpr */
2472 234 : var1 = linitial(opexpr->args);
2473 234 : if (IsA(var1, Var) && var1->varattno > 0)
2474 : {
2475 198 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2476 198 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2477 :
2478 198 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2479 : opexpr, true, &keep_original,
2480 : run_cond_attrs))
2481 162 : return keep_original;
2482 : }
2483 :
2484 : /* and check the right side */
2485 72 : var2 = lsecond(opexpr->args);
2486 72 : if (IsA(var2, Var) && var2->varattno > 0)
2487 : {
2488 42 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2489 42 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2490 :
2491 42 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2492 : opexpr, false, &keep_original,
2493 : run_cond_attrs))
2494 18 : return keep_original;
2495 : }
2496 :
2497 54 : return true;
2498 : }
2499 :
2500 : /*
2501 : * set_subquery_pathlist
2502 : * Generate SubqueryScan access paths for a subquery RTE
2503 : *
2504 : * We don't currently support generating parameterized paths for subqueries
2505 : * by pushing join clauses down into them; it seems too expensive to re-plan
2506 : * the subquery multiple times to consider different alternatives.
2507 : * (XXX that could stand to be reconsidered, now that we use Paths.)
2508 : * So the paths made here will be parameterized if the subquery contains
2509 : * LATERAL references, otherwise not. As long as that's true, there's no need
2510 : * for a separate set_subquery_size phase: just make the paths right away.
2511 : */
2512 : static void
2513 11762 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2514 : Index rti, RangeTblEntry *rte)
2515 : {
2516 11762 : Query *parse = root->parse;
2517 11762 : Query *subquery = rte->subquery;
2518 : bool trivial_pathtarget;
2519 : Relids required_outer;
2520 : pushdown_safety_info safetyInfo;
2521 : double tuple_fraction;
2522 : RelOptInfo *sub_final_rel;
2523 11762 : Bitmapset *run_cond_attrs = NULL;
2524 : ListCell *lc;
2525 :
2526 : /*
2527 : * Must copy the Query so that planning doesn't mess up the RTE contents
2528 : * (really really need to fix the planner to not scribble on its input,
2529 : * someday ... but see remove_unused_subquery_outputs to start with).
2530 : */
2531 11762 : subquery = copyObject(subquery);
2532 :
2533 : /*
2534 : * If it's a LATERAL subquery, it might contain some Vars of the current
2535 : * query level, requiring it to be treated as parameterized, even though
2536 : * we don't support pushing down join quals into subqueries.
2537 : */
2538 11762 : required_outer = rel->lateral_relids;
2539 :
2540 : /*
2541 : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2542 : * flags as needed while recursing. In particular, we need a workspace
2543 : * for keeping track of the reasons why columns are unsafe to reference.
2544 : * These reasons are stored in the bits inside unsafeFlags[i] when we
2545 : * discover reasons that column i of the subquery is unsafe to be used in
2546 : * a pushed-down qual.
2547 : */
2548 11762 : memset(&safetyInfo, 0, sizeof(safetyInfo));
2549 11762 : safetyInfo.unsafeFlags = (unsigned char *)
2550 11762 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2551 :
2552 : /*
2553 : * If the subquery has the "security_barrier" flag, it means the subquery
2554 : * originated from a view that must enforce row-level security. Then we
2555 : * must not push down quals that contain leaky functions. (Ideally this
2556 : * would be checked inside subquery_is_pushdown_safe, but since we don't
2557 : * currently pass the RTE to that function, we must do it here.)
2558 : */
2559 11762 : safetyInfo.unsafeLeaky = rte->security_barrier;
2560 :
2561 : /*
2562 : * If there are any restriction clauses that have been attached to the
2563 : * subquery relation, consider pushing them down to become WHERE or HAVING
2564 : * quals of the subquery itself. This transformation is useful because it
2565 : * may allow us to generate a better plan for the subquery than evaluating
2566 : * all the subquery output rows and then filtering them.
2567 : *
2568 : * There are several cases where we cannot push down clauses. Restrictions
2569 : * involving the subquery are checked by subquery_is_pushdown_safe().
2570 : * Restrictions on individual clauses are checked by
2571 : * qual_is_pushdown_safe(). Also, we don't want to push down
2572 : * pseudoconstant clauses; better to have the gating node above the
2573 : * subquery.
2574 : *
2575 : * Non-pushed-down clauses will get evaluated as qpquals of the
2576 : * SubqueryScan node.
2577 : *
2578 : * XXX Are there any cases where we want to make a policy decision not to
2579 : * push down a pushable qual, because it'd result in a worse plan?
2580 : */
2581 13794 : if (rel->baserestrictinfo != NIL &&
2582 2032 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2583 : {
2584 : /* OK to consider pushing down individual quals */
2585 1886 : List *upperrestrictlist = NIL;
2586 : ListCell *l;
2587 :
2588 4718 : foreach(l, rel->baserestrictinfo)
2589 : {
2590 2832 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2591 2832 : Node *clause = (Node *) rinfo->clause;
2592 :
2593 2832 : if (rinfo->pseudoconstant)
2594 : {
2595 4 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2596 4 : continue;
2597 : }
2598 :
2599 2828 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2600 : {
2601 2134 : case PUSHDOWN_SAFE:
2602 : /* Push it down */
2603 2134 : subquery_push_qual(subquery, rte, rti, clause);
2604 2134 : break;
2605 :
2606 252 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2607 :
2608 : /*
2609 : * Since we can't push the qual down into the subquery,
2610 : * check if it happens to reference a window function. If
2611 : * so then it might be useful to use for the WindowAgg's
2612 : * runCondition.
2613 : */
2614 504 : if (!subquery->hasWindowFuncs ||
2615 252 : check_and_push_window_quals(subquery, rte, rti, clause,
2616 : &run_cond_attrs))
2617 : {
2618 : /*
2619 : * subquery has no window funcs or the clause is not a
2620 : * suitable window run condition qual or it is, but
2621 : * the original must also be kept in the upper query.
2622 : */
2623 102 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2624 : }
2625 252 : break;
2626 :
2627 442 : case PUSHDOWN_UNSAFE:
2628 442 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2629 442 : break;
2630 : }
2631 2832 : }
2632 1886 : rel->baserestrictinfo = upperrestrictlist;
2633 : /* We don't bother recomputing baserestrict_min_security */
2634 : }
2635 :
2636 11762 : pfree(safetyInfo.unsafeFlags);
2637 :
2638 : /*
2639 : * The upper query might not use all the subquery's output columns; if
2640 : * not, we can simplify. Pass the attributes that were pushed down into
2641 : * WindowAgg run conditions to ensure we don't accidentally think those
2642 : * are unused.
2643 : */
2644 11762 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2645 :
2646 : /*
2647 : * We can safely pass the outer tuple_fraction down to the subquery if the
2648 : * outer level has no joining, aggregation, or sorting to do. Otherwise
2649 : * we'd better tell the subquery to plan for full retrieval. (XXX This
2650 : * could probably be made more intelligent ...)
2651 : */
2652 11762 : if (parse->hasAggs ||
2653 10598 : parse->groupClause ||
2654 10580 : parse->groupingSets ||
2655 10580 : root->hasHavingQual ||
2656 10580 : parse->distinctClause ||
2657 15346 : parse->sortClause ||
2658 5264 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
2659 7694 : tuple_fraction = 0.0; /* default case */
2660 : else
2661 4068 : tuple_fraction = root->tuple_fraction;
2662 :
2663 : /* plan_params should not be in use in current query level */
2664 : Assert(root->plan_params == NIL);
2665 :
2666 : /* Generate a subroot and Paths for the subquery */
2667 11762 : rel->subroot = subquery_planner(root->glob, subquery, root, false,
2668 : tuple_fraction, NULL);
2669 :
2670 : /* Isolate the params needed by this specific subplan */
2671 11762 : rel->subplan_params = root->plan_params;
2672 11762 : root->plan_params = NIL;
2673 :
2674 : /*
2675 : * It's possible that constraint exclusion proved the subquery empty. If
2676 : * so, it's desirable to produce an unadorned dummy path so that we will
2677 : * recognize appropriate optimizations at this query level.
2678 : */
2679 11762 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2680 :
2681 11762 : if (IS_DUMMY_REL(sub_final_rel))
2682 : {
2683 120 : set_dummy_rel_pathlist(rel);
2684 120 : return;
2685 : }
2686 :
2687 : /*
2688 : * Mark rel with estimated output rows, width, etc. Note that we have to
2689 : * do this before generating outer-query paths, else cost_subqueryscan is
2690 : * not happy.
2691 : */
2692 11642 : set_subquery_size_estimates(root, rel);
2693 :
2694 : /*
2695 : * Also detect whether the reltarget is trivial, so that we can pass that
2696 : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2697 : * It's trivial if it fetches all the subplan output columns in order.
2698 : */
2699 11642 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2700 2670 : trivial_pathtarget = false;
2701 : else
2702 : {
2703 8972 : trivial_pathtarget = true;
2704 31298 : foreach(lc, rel->reltarget->exprs)
2705 : {
2706 22624 : Node *node = (Node *) lfirst(lc);
2707 : Var *var;
2708 :
2709 22624 : if (!IsA(node, Var))
2710 : {
2711 0 : trivial_pathtarget = false;
2712 0 : break;
2713 : }
2714 22624 : var = (Var *) node;
2715 22624 : if (var->varno != rti ||
2716 22624 : var->varattno != foreach_current_index(lc) + 1)
2717 : {
2718 298 : trivial_pathtarget = false;
2719 298 : break;
2720 : }
2721 : }
2722 : }
2723 :
2724 : /*
2725 : * For each Path that subquery_planner produced, make a SubqueryScanPath
2726 : * in the outer query.
2727 : */
2728 24062 : foreach(lc, sub_final_rel->pathlist)
2729 : {
2730 12420 : Path *subpath = (Path *) lfirst(lc);
2731 : List *pathkeys;
2732 :
2733 : /* Convert subpath's pathkeys to outer representation */
2734 12420 : pathkeys = convert_subquery_pathkeys(root,
2735 : rel,
2736 : subpath->pathkeys,
2737 : make_tlist_from_pathtarget(subpath->pathtarget));
2738 :
2739 : /* Generate outer path using this subpath */
2740 12420 : add_path(rel, (Path *)
2741 12420 : create_subqueryscan_path(root, rel, subpath,
2742 : trivial_pathtarget,
2743 : pathkeys, required_outer));
2744 : }
2745 :
2746 : /* If outer rel allows parallelism, do same for partial paths. */
2747 11642 : if (rel->consider_parallel && bms_is_empty(required_outer))
2748 : {
2749 : /* If consider_parallel is false, there should be no partial paths. */
2750 : Assert(sub_final_rel->consider_parallel ||
2751 : sub_final_rel->partial_pathlist == NIL);
2752 :
2753 : /* Same for partial paths. */
2754 7634 : foreach(lc, sub_final_rel->partial_pathlist)
2755 : {
2756 42 : Path *subpath = (Path *) lfirst(lc);
2757 : List *pathkeys;
2758 :
2759 : /* Convert subpath's pathkeys to outer representation */
2760 42 : pathkeys = convert_subquery_pathkeys(root,
2761 : rel,
2762 : subpath->pathkeys,
2763 : make_tlist_from_pathtarget(subpath->pathtarget));
2764 :
2765 : /* Generate outer path using this subpath */
2766 42 : add_partial_path(rel, (Path *)
2767 42 : create_subqueryscan_path(root, rel, subpath,
2768 : trivial_pathtarget,
2769 : pathkeys,
2770 : required_outer));
2771 : }
2772 : }
2773 : }
2774 :
2775 : /*
2776 : * set_function_pathlist
2777 : * Build the (single) access path for a function RTE
2778 : */
2779 : static void
2780 55114 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2781 : {
2782 : Relids required_outer;
2783 55114 : List *pathkeys = NIL;
2784 :
2785 : /*
2786 : * We don't support pushing join clauses into the quals of a function
2787 : * scan, but it could still have required parameterization due to LATERAL
2788 : * refs in the function expression.
2789 : */
2790 55114 : required_outer = rel->lateral_relids;
2791 :
2792 : /*
2793 : * The result is considered unordered unless ORDINALITY was used, in which
2794 : * case it is ordered by the ordinal column (the last one). See if we
2795 : * care, by checking for uses of that Var in equivalence classes.
2796 : */
2797 55114 : if (rte->funcordinality)
2798 : {
2799 1110 : AttrNumber ordattno = rel->max_attr;
2800 1110 : Var *var = NULL;
2801 : ListCell *lc;
2802 :
2803 : /*
2804 : * Is there a Var for it in rel's targetlist? If not, the query did
2805 : * not reference the ordinality column, or at least not in any way
2806 : * that would be interesting for sorting.
2807 : */
2808 2294 : foreach(lc, rel->reltarget->exprs)
2809 : {
2810 2288 : Var *node = (Var *) lfirst(lc);
2811 :
2812 : /* checking varno/varlevelsup is just paranoia */
2813 2288 : if (IsA(node, Var) &&
2814 2288 : node->varattno == ordattno &&
2815 1104 : node->varno == rel->relid &&
2816 1104 : node->varlevelsup == 0)
2817 : {
2818 1104 : var = node;
2819 1104 : break;
2820 : }
2821 : }
2822 :
2823 : /*
2824 : * Try to build pathkeys for this Var with int8 sorting. We tell
2825 : * build_expression_pathkey not to build any new equivalence class; if
2826 : * the Var isn't already mentioned in some EC, it means that nothing
2827 : * cares about the ordering.
2828 : */
2829 1110 : if (var)
2830 1104 : pathkeys = build_expression_pathkey(root,
2831 : (Expr *) var,
2832 : Int8LessOperator,
2833 : rel->relids,
2834 : false);
2835 : }
2836 :
2837 : /* Generate appropriate path */
2838 55114 : add_path(rel, create_functionscan_path(root, rel,
2839 : pathkeys, required_outer));
2840 55114 : }
2841 :
2842 : /*
2843 : * set_values_pathlist
2844 : * Build the (single) access path for a VALUES RTE
2845 : */
2846 : static void
2847 8232 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2848 : {
2849 : Relids required_outer;
2850 :
2851 : /*
2852 : * We don't support pushing join clauses into the quals of a values scan,
2853 : * but it could still have required parameterization due to LATERAL refs
2854 : * in the values expressions.
2855 : */
2856 8232 : required_outer = rel->lateral_relids;
2857 :
2858 : /* Generate appropriate path */
2859 8232 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
2860 8232 : }
2861 :
2862 : /*
2863 : * set_tablefunc_pathlist
2864 : * Build the (single) access path for a table func RTE
2865 : */
2866 : static void
2867 626 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2868 : {
2869 : Relids required_outer;
2870 :
2871 : /*
2872 : * We don't support pushing join clauses into the quals of a tablefunc
2873 : * scan, but it could still have required parameterization due to LATERAL
2874 : * refs in the function expression.
2875 : */
2876 626 : required_outer = rel->lateral_relids;
2877 :
2878 : /* Generate appropriate path */
2879 626 : add_path(rel, create_tablefuncscan_path(root, rel,
2880 : required_outer));
2881 626 : }
2882 :
2883 : /*
2884 : * set_cte_pathlist
2885 : * Build the (single) access path for a non-self-reference CTE RTE
2886 : *
2887 : * There's no need for a separate set_cte_size phase, since we don't
2888 : * support join-qual-parameterized paths for CTEs.
2889 : */
2890 : static void
2891 4090 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2892 : {
2893 : Path *ctepath;
2894 : Plan *cteplan;
2895 : PlannerInfo *cteroot;
2896 : Index levelsup;
2897 : List *pathkeys;
2898 : int ndx;
2899 : ListCell *lc;
2900 : int plan_id;
2901 : Relids required_outer;
2902 :
2903 : /*
2904 : * Find the referenced CTE, and locate the path and plan previously made
2905 : * for it.
2906 : */
2907 4090 : levelsup = rte->ctelevelsup;
2908 4090 : cteroot = root;
2909 7094 : while (levelsup-- > 0)
2910 : {
2911 3004 : cteroot = cteroot->parent_root;
2912 3004 : if (!cteroot) /* shouldn't happen */
2913 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2914 : }
2915 :
2916 : /*
2917 : * Note: cte_plan_ids can be shorter than cteList, if we are still working
2918 : * on planning the CTEs (ie, this is a side-reference from another CTE).
2919 : * So we mustn't use forboth here.
2920 : */
2921 4090 : ndx = 0;
2922 5690 : foreach(lc, cteroot->parse->cteList)
2923 : {
2924 5690 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
2925 :
2926 5690 : if (strcmp(cte->ctename, rte->ctename) == 0)
2927 4090 : break;
2928 1600 : ndx++;
2929 : }
2930 4090 : if (lc == NULL) /* shouldn't happen */
2931 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
2932 4090 : if (ndx >= list_length(cteroot->cte_plan_ids))
2933 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
2934 4090 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
2935 4090 : if (plan_id <= 0)
2936 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
2937 :
2938 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
2939 4090 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
2940 4090 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2941 :
2942 : /* Mark rel with estimated output rows, width, etc */
2943 4090 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
2944 :
2945 : /* Convert the ctepath's pathkeys to outer query's representation */
2946 4090 : pathkeys = convert_subquery_pathkeys(root,
2947 : rel,
2948 : ctepath->pathkeys,
2949 : cteplan->targetlist);
2950 :
2951 : /*
2952 : * We don't support pushing join clauses into the quals of a CTE scan, but
2953 : * it could still have required parameterization due to LATERAL refs in
2954 : * its tlist.
2955 : */
2956 4090 : required_outer = rel->lateral_relids;
2957 :
2958 : /* Generate appropriate path */
2959 4090 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
2960 4090 : }
2961 :
2962 : /*
2963 : * set_namedtuplestore_pathlist
2964 : * Build the (single) access path for a named tuplestore RTE
2965 : *
2966 : * There's no need for a separate set_namedtuplestore_size phase, since we
2967 : * don't support join-qual-parameterized paths for tuplestores.
2968 : */
2969 : static void
2970 462 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
2971 : RangeTblEntry *rte)
2972 : {
2973 : Relids required_outer;
2974 :
2975 : /* Mark rel with estimated output rows, width, etc */
2976 462 : set_namedtuplestore_size_estimates(root, rel);
2977 :
2978 : /*
2979 : * We don't support pushing join clauses into the quals of a tuplestore
2980 : * scan, but it could still have required parameterization due to LATERAL
2981 : * refs in its tlist.
2982 : */
2983 462 : required_outer = rel->lateral_relids;
2984 :
2985 : /* Generate appropriate path */
2986 462 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2987 462 : }
2988 :
2989 : /*
2990 : * set_result_pathlist
2991 : * Build the (single) access path for an RTE_RESULT RTE
2992 : *
2993 : * There's no need for a separate set_result_size phase, since we
2994 : * don't support join-qual-parameterized paths for these RTEs.
2995 : */
2996 : static void
2997 4190 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
2998 : RangeTblEntry *rte)
2999 : {
3000 : Relids required_outer;
3001 :
3002 : /* Mark rel with estimated output rows, width, etc */
3003 4190 : set_result_size_estimates(root, rel);
3004 :
3005 : /*
3006 : * We don't support pushing join clauses into the quals of a Result scan,
3007 : * but it could still have required parameterization due to LATERAL refs
3008 : * in its tlist.
3009 : */
3010 4190 : required_outer = rel->lateral_relids;
3011 :
3012 : /* Generate appropriate path */
3013 4190 : add_path(rel, create_resultscan_path(root, rel, required_outer));
3014 4190 : }
3015 :
3016 : /*
3017 : * set_worktable_pathlist
3018 : * Build the (single) access path for a self-reference CTE RTE
3019 : *
3020 : * There's no need for a separate set_worktable_size phase, since we don't
3021 : * support join-qual-parameterized paths for CTEs.
3022 : */
3023 : static void
3024 1010 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3025 : {
3026 : Path *ctepath;
3027 : PlannerInfo *cteroot;
3028 : Index levelsup;
3029 : Relids required_outer;
3030 :
3031 : /*
3032 : * We need to find the non-recursive term's path, which is in the plan
3033 : * level that's processing the recursive UNION, which is one level *below*
3034 : * where the CTE comes from.
3035 : */
3036 1010 : levelsup = rte->ctelevelsup;
3037 1010 : if (levelsup == 0) /* shouldn't happen */
3038 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3039 1010 : levelsup--;
3040 1010 : cteroot = root;
3041 2376 : while (levelsup-- > 0)
3042 : {
3043 1366 : cteroot = cteroot->parent_root;
3044 1366 : if (!cteroot) /* shouldn't happen */
3045 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3046 : }
3047 1010 : ctepath = cteroot->non_recursive_path;
3048 1010 : if (!ctepath) /* shouldn't happen */
3049 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3050 :
3051 : /* Mark rel with estimated output rows, width, etc */
3052 1010 : set_cte_size_estimates(root, rel, ctepath->rows);
3053 :
3054 : /*
3055 : * We don't support pushing join clauses into the quals of a worktable
3056 : * scan, but it could still have required parameterization due to LATERAL
3057 : * refs in its tlist. (I'm not sure this is actually possible given the
3058 : * restrictions on recursive references, but it's easy enough to support.)
3059 : */
3060 1010 : required_outer = rel->lateral_relids;
3061 :
3062 : /* Generate appropriate path */
3063 1010 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
3064 1010 : }
3065 :
3066 : /*
3067 : * generate_gather_paths
3068 : * Generate parallel access paths for a relation by pushing a Gather or
3069 : * Gather Merge on top of a partial path.
3070 : *
3071 : * This must not be called until after we're done creating all partial paths
3072 : * for the specified relation. (Otherwise, add_partial_path might delete a
3073 : * path that some GatherPath or GatherMergePath has a reference to.)
3074 : *
3075 : * If we're generating paths for a scan or join relation, override_rows will
3076 : * be false, and we'll just use the relation's size estimate. When we're
3077 : * being called for a partially-grouped or partially-distinct path, though, we
3078 : * need to override the rowcount estimate. (It's not clear that the
3079 : * particular value we're using here is actually best, but the underlying rel
3080 : * has no estimate so we must do something.)
3081 : */
3082 : void
3083 19076 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3084 : {
3085 : Path *cheapest_partial_path;
3086 : Path *simple_gather_path;
3087 : ListCell *lc;
3088 : double rows;
3089 19076 : double *rowsp = NULL;
3090 :
3091 : /* If there are no partial paths, there's nothing to do here. */
3092 19076 : if (rel->partial_pathlist == NIL)
3093 0 : return;
3094 :
3095 : /* Should we override the rel's rowcount estimate? */
3096 19076 : if (override_rows)
3097 1752 : rowsp = &rows;
3098 :
3099 : /*
3100 : * The output of Gather is always unsorted, so there's only one partial
3101 : * path of interest: the cheapest one. That will be the one at the front
3102 : * of partial_pathlist because of the way add_partial_path works.
3103 : */
3104 19076 : cheapest_partial_path = linitial(rel->partial_pathlist);
3105 19076 : rows = compute_gather_rows(cheapest_partial_path);
3106 : simple_gather_path = (Path *)
3107 19076 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3108 : NULL, rowsp);
3109 19076 : add_path(rel, simple_gather_path);
3110 :
3111 : /*
3112 : * For each useful ordering, we can consider an order-preserving Gather
3113 : * Merge.
3114 : */
3115 39474 : foreach(lc, rel->partial_pathlist)
3116 : {
3117 20398 : Path *subpath = (Path *) lfirst(lc);
3118 : GatherMergePath *path;
3119 :
3120 20398 : if (subpath->pathkeys == NIL)
3121 18714 : continue;
3122 :
3123 1684 : rows = compute_gather_rows(subpath);
3124 1684 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3125 : subpath->pathkeys, NULL, rowsp);
3126 1684 : add_path(rel, &path->path);
3127 : }
3128 : }
3129 :
3130 : /*
3131 : * get_useful_pathkeys_for_relation
3132 : * Determine which orderings of a relation might be useful.
3133 : *
3134 : * Getting data in sorted order can be useful either because the requested
3135 : * order matches the final output ordering for the overall query we're
3136 : * planning, or because it enables an efficient merge join. Here, we try
3137 : * to figure out which pathkeys to consider.
3138 : *
3139 : * This allows us to do incremental sort on top of an index scan under a gather
3140 : * merge node, i.e. parallelized.
3141 : *
3142 : * If the require_parallel_safe is true, we also require the expressions to
3143 : * be parallel safe (which allows pushing the sort below Gather Merge).
3144 : *
3145 : * XXX At the moment this can only ever return a list with a single element,
3146 : * because it looks at query_pathkeys only. So we might return the pathkeys
3147 : * directly, but it seems plausible we'll want to consider other orderings
3148 : * in the future. For example, we might want to consider pathkeys useful for
3149 : * merge joins.
3150 : */
3151 : static List *
3152 19076 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3153 : bool require_parallel_safe)
3154 : {
3155 19076 : List *useful_pathkeys_list = NIL;
3156 :
3157 : /*
3158 : * Considering query_pathkeys is always worth it, because it might allow
3159 : * us to avoid a total sort when we have a partially presorted path
3160 : * available or to push the total sort into the parallel portion of the
3161 : * query.
3162 : */
3163 19076 : if (root->query_pathkeys)
3164 : {
3165 : ListCell *lc;
3166 9666 : int npathkeys = 0; /* useful pathkeys */
3167 :
3168 17870 : foreach(lc, root->query_pathkeys)
3169 : {
3170 12810 : PathKey *pathkey = (PathKey *) lfirst(lc);
3171 12810 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3172 :
3173 : /*
3174 : * We can only build a sort for pathkeys that contain a
3175 : * safe-to-compute-early EC member computable from the current
3176 : * relation's reltarget, so ignore the remainder of the list as
3177 : * soon as we find a pathkey without such a member.
3178 : *
3179 : * It's still worthwhile to return any prefix of the pathkeys list
3180 : * that meets this requirement, as we may be able to do an
3181 : * incremental sort.
3182 : *
3183 : * If requested, ensure the sort expression is parallel-safe too.
3184 : */
3185 12810 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3186 : require_parallel_safe))
3187 4606 : break;
3188 :
3189 8204 : npathkeys++;
3190 : }
3191 :
3192 : /*
3193 : * The whole query_pathkeys list matches, so append it directly, to
3194 : * allow comparing pathkeys easily by comparing list pointer. If we
3195 : * have to truncate the pathkeys, we gotta do a copy though.
3196 : */
3197 9666 : if (npathkeys == list_length(root->query_pathkeys))
3198 5060 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3199 5060 : root->query_pathkeys);
3200 4606 : else if (npathkeys > 0)
3201 450 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3202 450 : list_copy_head(root->query_pathkeys,
3203 : npathkeys));
3204 : }
3205 :
3206 19076 : return useful_pathkeys_list;
3207 : }
3208 :
3209 : /*
3210 : * generate_useful_gather_paths
3211 : * Generate parallel access paths for a relation by pushing a Gather or
3212 : * Gather Merge on top of a partial path.
3213 : *
3214 : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3215 : * paths (aiming to preserve the ordering), but also considers ordering that
3216 : * might be useful for nodes above the gather merge node, and tries to add
3217 : * a sort (regular or incremental) to provide that.
3218 : */
3219 : void
3220 601534 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3221 : {
3222 : ListCell *lc;
3223 : double rows;
3224 601534 : double *rowsp = NULL;
3225 601534 : List *useful_pathkeys_list = NIL;
3226 601534 : Path *cheapest_partial_path = NULL;
3227 :
3228 : /* If there are no partial paths, there's nothing to do here. */
3229 601534 : if (rel->partial_pathlist == NIL)
3230 582458 : return;
3231 :
3232 : /* Should we override the rel's rowcount estimate? */
3233 19076 : if (override_rows)
3234 1752 : rowsp = &rows;
3235 :
3236 : /* generate the regular gather (merge) paths */
3237 19076 : generate_gather_paths(root, rel, override_rows);
3238 :
3239 : /* consider incremental sort for interesting orderings */
3240 19076 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3241 :
3242 : /* used for explicit (full) sort paths */
3243 19076 : cheapest_partial_path = linitial(rel->partial_pathlist);
3244 :
3245 : /*
3246 : * Consider sorted paths for each interesting ordering. We generate both
3247 : * incremental and full sort.
3248 : */
3249 24586 : foreach(lc, useful_pathkeys_list)
3250 : {
3251 5510 : List *useful_pathkeys = lfirst(lc);
3252 : ListCell *lc2;
3253 : bool is_sorted;
3254 : int presorted_keys;
3255 :
3256 12210 : foreach(lc2, rel->partial_pathlist)
3257 : {
3258 6700 : Path *subpath = (Path *) lfirst(lc2);
3259 : GatherMergePath *path;
3260 :
3261 6700 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3262 : subpath->pathkeys,
3263 : &presorted_keys);
3264 :
3265 : /*
3266 : * We don't need to consider the case where a subpath is already
3267 : * fully sorted because generate_gather_paths already creates a
3268 : * gather merge path for every subpath that has pathkeys present.
3269 : *
3270 : * But since the subpath is already sorted, we know we don't need
3271 : * to consider adding a sort (full or incremental) on top of it,
3272 : * so we can continue here.
3273 : */
3274 6700 : if (is_sorted)
3275 1192 : continue;
3276 :
3277 : /*
3278 : * Try at least sorting the cheapest path and also try
3279 : * incrementally sorting any path which is partially sorted
3280 : * already (no need to deal with paths which have presorted keys
3281 : * when incremental sort is disabled unless it's the cheapest
3282 : * input path).
3283 : */
3284 5508 : if (subpath != cheapest_partial_path &&
3285 222 : (presorted_keys == 0 || !enable_incremental_sort))
3286 78 : continue;
3287 :
3288 : /*
3289 : * Consider regular sort for any path that's not presorted or if
3290 : * incremental sort is disabled. We've no need to consider both
3291 : * sort and incremental sort on the same path. We assume that
3292 : * incremental sort is always faster when there are presorted
3293 : * keys.
3294 : *
3295 : * This is not redundant with the gather paths created in
3296 : * generate_gather_paths, because that doesn't generate ordered
3297 : * output. Here we add an explicit sort to match the useful
3298 : * ordering.
3299 : */
3300 5430 : if (presorted_keys == 0 || !enable_incremental_sort)
3301 : {
3302 5274 : subpath = (Path *) create_sort_path(root,
3303 : rel,
3304 : subpath,
3305 : useful_pathkeys,
3306 : -1.0);
3307 : }
3308 : else
3309 156 : subpath = (Path *) create_incremental_sort_path(root,
3310 : rel,
3311 : subpath,
3312 : useful_pathkeys,
3313 : presorted_keys,
3314 : -1);
3315 5430 : rows = compute_gather_rows(subpath);
3316 5430 : path = create_gather_merge_path(root, rel,
3317 : subpath,
3318 5430 : rel->reltarget,
3319 : subpath->pathkeys,
3320 : NULL,
3321 : rowsp);
3322 :
3323 5430 : add_path(rel, &path->path);
3324 : }
3325 : }
3326 : }
3327 :
3328 : /*
3329 : * make_rel_from_joinlist
3330 : * Build access paths using a "joinlist" to guide the join path search.
3331 : *
3332 : * See comments for deconstruct_jointree() for definition of the joinlist
3333 : * data structure.
3334 : */
3335 : static RelOptInfo *
3336 332298 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3337 : {
3338 : int levels_needed;
3339 : List *initial_rels;
3340 : ListCell *jl;
3341 :
3342 : /*
3343 : * Count the number of child joinlist nodes. This is the depth of the
3344 : * dynamic-programming algorithm we must employ to consider all ways of
3345 : * joining the child nodes.
3346 : */
3347 332298 : levels_needed = list_length(joinlist);
3348 :
3349 332298 : if (levels_needed <= 0)
3350 0 : return NULL; /* nothing to do? */
3351 :
3352 : /*
3353 : * Construct a list of rels corresponding to the child joinlist nodes.
3354 : * This may contain both base rels and rels constructed according to
3355 : * sub-joinlists.
3356 : */
3357 332298 : initial_rels = NIL;
3358 798784 : foreach(jl, joinlist)
3359 : {
3360 466486 : Node *jlnode = (Node *) lfirst(jl);
3361 : RelOptInfo *thisrel;
3362 :
3363 466486 : if (IsA(jlnode, RangeTblRef))
3364 : {
3365 463120 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3366 :
3367 463120 : thisrel = find_base_rel(root, varno);
3368 : }
3369 3366 : else if (IsA(jlnode, List))
3370 : {
3371 : /* Recurse to handle subproblem */
3372 3366 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3373 : }
3374 : else
3375 : {
3376 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3377 : (int) nodeTag(jlnode));
3378 : thisrel = NULL; /* keep compiler quiet */
3379 : }
3380 :
3381 466486 : initial_rels = lappend(initial_rels, thisrel);
3382 : }
3383 :
3384 332298 : if (levels_needed == 1)
3385 : {
3386 : /*
3387 : * Single joinlist node, so we're done.
3388 : */
3389 234380 : return (RelOptInfo *) linitial(initial_rels);
3390 : }
3391 : else
3392 : {
3393 : /*
3394 : * Consider the different orders in which we could join the rels,
3395 : * using a plugin, GEQO, or the regular join search code.
3396 : *
3397 : * We put the initial_rels list into a PlannerInfo field because
3398 : * has_legal_joinclause() needs to look at it (ugly :-().
3399 : */
3400 97918 : root->initial_rels = initial_rels;
3401 :
3402 97918 : if (join_search_hook)
3403 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
3404 97918 : else if (enable_geqo && levels_needed >= geqo_threshold)
3405 6 : return geqo(root, levels_needed, initial_rels);
3406 : else
3407 97912 : return standard_join_search(root, levels_needed, initial_rels);
3408 : }
3409 : }
3410 :
3411 : /*
3412 : * standard_join_search
3413 : * Find possible joinpaths for a query by successively finding ways
3414 : * to join component relations into join relations.
3415 : *
3416 : * 'levels_needed' is the number of iterations needed, ie, the number of
3417 : * independent jointree items in the query. This is > 1.
3418 : *
3419 : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3420 : * jointree item. These are the components to be joined together.
3421 : * Note that levels_needed == list_length(initial_rels).
3422 : *
3423 : * Returns the final level of join relations, i.e., the relation that is
3424 : * the result of joining all the original relations together.
3425 : * At least one implementation path must be provided for this relation and
3426 : * all required sub-relations.
3427 : *
3428 : * To support loadable plugins that modify planner behavior by changing the
3429 : * join searching algorithm, we provide a hook variable that lets a plugin
3430 : * replace or supplement this function. Any such hook must return the same
3431 : * final join relation as the standard code would, but it might have a
3432 : * different set of implementation paths attached, and only the sub-joinrels
3433 : * needed for these paths need have been instantiated.
3434 : *
3435 : * Note to plugin authors: the functions invoked during standard_join_search()
3436 : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3437 : * than one join-order search, you'll probably need to save and restore the
3438 : * original states of those data structures. See geqo_eval() for an example.
3439 : */
3440 : RelOptInfo *
3441 97912 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3442 : {
3443 : int lev;
3444 : RelOptInfo *rel;
3445 :
3446 : /*
3447 : * This function cannot be invoked recursively within any one planning
3448 : * problem, so join_rel_level[] can't be in use already.
3449 : */
3450 : Assert(root->join_rel_level == NULL);
3451 :
3452 : /*
3453 : * We employ a simple "dynamic programming" algorithm: we first find all
3454 : * ways to build joins of two jointree items, then all ways to build joins
3455 : * of three items (from two-item joins and single items), then four-item
3456 : * joins, and so on until we have considered all ways to join all the
3457 : * items into one rel.
3458 : *
3459 : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3460 : * set root->join_rel_level[1] to represent all the single-jointree-item
3461 : * relations.
3462 : */
3463 97912 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3464 :
3465 97912 : root->join_rel_level[1] = initial_rels;
3466 :
3467 232076 : for (lev = 2; lev <= levels_needed; lev++)
3468 : {
3469 : ListCell *lc;
3470 :
3471 : /*
3472 : * Determine all possible pairs of relations to be joined at this
3473 : * level, and build paths for making each one from every available
3474 : * pair of lower-level relations.
3475 : */
3476 134164 : join_search_one_level(root, lev);
3477 :
3478 : /*
3479 : * Run generate_partitionwise_join_paths() and
3480 : * generate_useful_gather_paths() for each just-processed joinrel. We
3481 : * could not do this earlier because both regular and partial paths
3482 : * can get added to a particular joinrel at multiple times within
3483 : * join_search_one_level.
3484 : *
3485 : * After that, we're done creating paths for the joinrel, so run
3486 : * set_cheapest().
3487 : */
3488 342420 : foreach(lc, root->join_rel_level[lev])
3489 : {
3490 208256 : rel = (RelOptInfo *) lfirst(lc);
3491 :
3492 : /* Create paths for partitionwise joins. */
3493 208256 : generate_partitionwise_join_paths(root, rel);
3494 :
3495 : /*
3496 : * Except for the topmost scan/join rel, consider gathering
3497 : * partial paths. We'll do the same for the topmost scan/join rel
3498 : * once we know the final targetlist (see grouping_planner's and
3499 : * its call to apply_scanjoin_target_to_paths).
3500 : */
3501 208256 : if (!bms_equal(rel->relids, root->all_query_rels))
3502 110836 : generate_useful_gather_paths(root, rel, false);
3503 :
3504 : /* Find and save the cheapest paths for this rel */
3505 208256 : set_cheapest(rel);
3506 :
3507 : #ifdef OPTIMIZER_DEBUG
3508 : pprint(rel);
3509 : #endif
3510 : }
3511 : }
3512 :
3513 : /*
3514 : * We should have a single rel at the final level.
3515 : */
3516 97912 : if (root->join_rel_level[levels_needed] == NIL)
3517 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
3518 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3519 :
3520 97912 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3521 :
3522 97912 : root->join_rel_level = NULL;
3523 :
3524 97912 : return rel;
3525 : }
3526 :
3527 : /*****************************************************************************
3528 : * PUSHING QUALS DOWN INTO SUBQUERIES
3529 : *****************************************************************************/
3530 :
3531 : /*
3532 : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3533 : *
3534 : * subquery is the particular component query being checked. topquery
3535 : * is the top component of a set-operations tree (the same Query if no
3536 : * set-op is involved).
3537 : *
3538 : * Conditions checked here:
3539 : *
3540 : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3541 : * since that could change the set of rows returned.
3542 : *
3543 : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3544 : * quals into it, because that could change the results.
3545 : *
3546 : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3547 : * This is because upper-level quals should semantically be evaluated only
3548 : * once per distinct row, not once per original row, and if the qual is
3549 : * volatile then extra evaluations could change the results. (This issue
3550 : * does not apply to other forms of aggregation such as GROUP BY, because
3551 : * when those are present we push into HAVING not WHERE, so that the quals
3552 : * are still applied after aggregation.)
3553 : *
3554 : * 4. If the subquery contains window functions, we cannot push volatile quals
3555 : * into it. The issue here is a bit different from DISTINCT: a volatile qual
3556 : * might succeed for some rows of a window partition and fail for others,
3557 : * thereby changing the partition contents and thus the window functions'
3558 : * results for rows that remain.
3559 : *
3560 : * 5. If the subquery contains any set-returning functions in its targetlist,
3561 : * we cannot push volatile quals into it. That would push them below the SRFs
3562 : * and thereby change the number of times they are evaluated. Also, a
3563 : * volatile qual could succeed for some SRF output rows and fail for others,
3564 : * a behavior that cannot occur if it's evaluated before SRF expansion.
3565 : *
3566 : * 6. If the subquery has nonempty grouping sets, we cannot push down any
3567 : * quals. The concern here is that a qual referencing a "constant" grouping
3568 : * column could get constant-folded, which would be improper because the value
3569 : * is potentially nullable by grouping-set expansion. This restriction could
3570 : * be removed if we had a parsetree representation that shows that such
3571 : * grouping columns are not really constant. (There are other ideas that
3572 : * could be used to relax this restriction, but that's the approach most
3573 : * likely to get taken in the future. Note that there's not much to be gained
3574 : * so long as subquery_planner can't move HAVING clauses to WHERE within such
3575 : * a subquery.)
3576 : *
3577 : * In addition, we make several checks on the subquery's output columns to see
3578 : * if it is safe to reference them in pushed-down quals. If output column k
3579 : * is found to be unsafe to reference, we set the reason for that inside
3580 : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3581 : * column k might not be referenced by some/all quals. The unsafeFlags[]
3582 : * array will be consulted later by qual_is_pushdown_safe(). It's better to
3583 : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3584 : * because when the subquery involves set operations we have to check the
3585 : * output expressions in each arm of the set op.
3586 : *
3587 : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3588 : * we're effectively assuming that the quals cannot distinguish values that
3589 : * the DISTINCT's equality operator sees as equal, yet there are many
3590 : * counterexamples to that assumption. However use of such a qual with a
3591 : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3592 : * "equal" value will be chosen as the output value by the DISTINCT operation.
3593 : * So we don't worry too much about that. Another objection is that if the
3594 : * qual is expensive to evaluate, running it for each original row might cost
3595 : * more than we save by eliminating rows before the DISTINCT step. But it
3596 : * would be very hard to estimate that at this stage, and in practice pushdown
3597 : * seldom seems to make things worse, so we ignore that problem too.
3598 : *
3599 : * Note: likewise, pushing quals into a subquery with window functions is a
3600 : * bit dubious: the quals might remove some rows of a window partition while
3601 : * leaving others, causing changes in the window functions' results for the
3602 : * surviving rows. We insist that such a qual reference only partitioning
3603 : * columns, but again that only protects us if the qual does not distinguish
3604 : * values that the partitioning equality operator sees as equal. The risks
3605 : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3606 : * occurs and thus there is no theoretical problem with such a qual. But
3607 : * we'll do this anyway because the potential performance benefits are very
3608 : * large, and we've seen no field complaints about the longstanding comparable
3609 : * behavior with DISTINCT.
3610 : */
3611 : static bool
3612 2344 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
3613 : pushdown_safety_info *safetyInfo)
3614 : {
3615 : SetOperationStmt *topop;
3616 :
3617 : /* Check point 1 */
3618 2344 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
3619 134 : return false;
3620 :
3621 : /* Check point 6 */
3622 2210 : if (subquery->groupClause && subquery->groupingSets)
3623 12 : return false;
3624 :
3625 : /* Check points 3, 4, and 5 */
3626 2198 : if (subquery->distinctClause ||
3627 2114 : subquery->hasWindowFuncs ||
3628 1848 : subquery->hasTargetSRFs)
3629 546 : safetyInfo->unsafeVolatile = true;
3630 :
3631 : /*
3632 : * If we're at a leaf query, check for unsafe expressions in its target
3633 : * list, and mark any reasons why they're unsafe in unsafeFlags[].
3634 : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3635 : * so no need to check them.)
3636 : */
3637 2198 : if (subquery->setOperations == NULL)
3638 2042 : check_output_expressions(subquery, safetyInfo);
3639 :
3640 : /* Are we at top level, or looking at a setop component? */
3641 2198 : if (subquery == topquery)
3642 : {
3643 : /* Top level, so check any component queries */
3644 1886 : if (subquery->setOperations != NULL)
3645 156 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3646 : safetyInfo))
3647 0 : return false;
3648 : }
3649 : else
3650 : {
3651 : /* Setop component must not have more components (too weird) */
3652 312 : if (subquery->setOperations != NULL)
3653 0 : return false;
3654 : /* Check whether setop component output types match top level */
3655 312 : topop = castNode(SetOperationStmt, topquery->setOperations);
3656 : Assert(topop);
3657 312 : compare_tlist_datatypes(subquery->targetList,
3658 : topop->colTypes,
3659 : safetyInfo);
3660 : }
3661 2198 : return true;
3662 : }
3663 :
3664 : /*
3665 : * Helper routine to recurse through setOperations tree
3666 : */
3667 : static bool
3668 468 : recurse_pushdown_safe(Node *setOp, Query *topquery,
3669 : pushdown_safety_info *safetyInfo)
3670 : {
3671 468 : if (IsA(setOp, RangeTblRef))
3672 : {
3673 312 : RangeTblRef *rtr = (RangeTblRef *) setOp;
3674 312 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3675 312 : Query *subquery = rte->subquery;
3676 :
3677 : Assert(subquery != NULL);
3678 312 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3679 : }
3680 156 : else if (IsA(setOp, SetOperationStmt))
3681 : {
3682 156 : SetOperationStmt *op = (SetOperationStmt *) setOp;
3683 :
3684 : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
3685 156 : if (op->op == SETOP_EXCEPT)
3686 0 : return false;
3687 : /* Else recurse */
3688 156 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
3689 0 : return false;
3690 156 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
3691 0 : return false;
3692 : }
3693 : else
3694 : {
3695 0 : elog(ERROR, "unrecognized node type: %d",
3696 : (int) nodeTag(setOp));
3697 : }
3698 156 : return true;
3699 : }
3700 :
3701 : /*
3702 : * check_output_expressions - check subquery's output expressions for safety
3703 : *
3704 : * There are several cases in which it's unsafe to push down an upper-level
3705 : * qual if it references a particular output column of a subquery. We check
3706 : * each output column of the subquery and set flags in unsafeFlags[k] when we
3707 : * see that column is unsafe for a pushed-down qual to reference. The
3708 : * conditions checked here are:
3709 : *
3710 : * 1. We must not push down any quals that refer to subselect outputs that
3711 : * return sets, else we'd introduce functions-returning-sets into the
3712 : * subquery's WHERE/HAVING quals.
3713 : *
3714 : * 2. We must not push down any quals that refer to subselect outputs that
3715 : * contain volatile functions, for fear of introducing strange results due
3716 : * to multiple evaluation of a volatile function.
3717 : *
3718 : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3719 : * refer to non-DISTINCT output columns, because that could change the set
3720 : * of rows returned. (This condition is vacuous for DISTINCT, because then
3721 : * there are no non-DISTINCT output columns, so we needn't check. Note that
3722 : * subquery_is_pushdown_safe already reported that we can't use volatile
3723 : * quals if there's DISTINCT or DISTINCT ON.)
3724 : *
3725 : * 4. If the subquery has any window functions, we must not push down quals
3726 : * that reference any output columns that are not listed in all the subquery's
3727 : * window PARTITION BY clauses. We can push down quals that use only
3728 : * partitioning columns because they should succeed or fail identically for
3729 : * every row of any one window partition, and totally excluding some
3730 : * partitions will not change a window function's results for remaining
3731 : * partitions. (Again, this also requires nonvolatile quals, but
3732 : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3733 : * unsafe for this reason can still have WindowClause run conditions pushed
3734 : * down.
3735 : */
3736 : static void
3737 2042 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
3738 : {
3739 : ListCell *lc;
3740 :
3741 20838 : foreach(lc, subquery->targetList)
3742 : {
3743 18796 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
3744 :
3745 18796 : if (tle->resjunk)
3746 134 : continue; /* ignore resjunk columns */
3747 :
3748 : /* Functions returning sets are unsafe (point 1) */
3749 18662 : if (subquery->hasTargetSRFs &&
3750 668 : (safetyInfo->unsafeFlags[tle->resno] &
3751 668 : UNSAFE_HAS_SET_FUNC) == 0 &&
3752 668 : expression_returns_set((Node *) tle->expr))
3753 : {
3754 376 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3755 376 : continue;
3756 : }
3757 :
3758 : /* Volatile functions are unsafe (point 2) */
3759 18286 : if ((safetyInfo->unsafeFlags[tle->resno] &
3760 18274 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3761 18274 : contain_volatile_functions((Node *) tle->expr))
3762 : {
3763 78 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3764 78 : continue;
3765 : }
3766 :
3767 : /* If subquery uses DISTINCT ON, check point 3 */
3768 18208 : if (subquery->hasDistinctOn &&
3769 0 : (safetyInfo->unsafeFlags[tle->resno] &
3770 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3771 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3772 : {
3773 : /* non-DISTINCT column, so mark it unsafe */
3774 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3775 0 : continue;
3776 : }
3777 :
3778 : /* If subquery uses window functions, check point 4 */
3779 18208 : if (subquery->hasWindowFuncs &&
3780 1090 : (safetyInfo->unsafeFlags[tle->resno] &
3781 1090 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3782 1090 : !targetIsInAllPartitionLists(tle, subquery))
3783 : {
3784 : /* not present in all PARTITION BY clauses, so mark it unsafe */
3785 1002 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3786 1002 : continue;
3787 : }
3788 : }
3789 2042 : }
3790 :
3791 : /*
3792 : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3793 : * push quals into each component query, but the quals can only reference
3794 : * subquery columns that suffer no type coercions in the set operation.
3795 : * Otherwise there are possible semantic gotchas. So, we check the
3796 : * component queries to see if any of them have output types different from
3797 : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3798 : * unsafeFlags[k] if column k has different type in any component.
3799 : *
3800 : * We don't have to care about typmods here: the only allowed difference
3801 : * between set-op input and output typmods is input is a specific typmod
3802 : * and output is -1, and that does not require a coercion.
3803 : *
3804 : * tlist is a subquery tlist.
3805 : * colTypes is an OID list of the top-level setop's output column types.
3806 : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
3807 : */
3808 : static void
3809 312 : compare_tlist_datatypes(List *tlist, List *colTypes,
3810 : pushdown_safety_info *safetyInfo)
3811 : {
3812 : ListCell *l;
3813 312 : ListCell *colType = list_head(colTypes);
3814 :
3815 984 : foreach(l, tlist)
3816 : {
3817 672 : TargetEntry *tle = (TargetEntry *) lfirst(l);
3818 :
3819 672 : if (tle->resjunk)
3820 0 : continue; /* ignore resjunk columns */
3821 672 : if (colType == NULL)
3822 0 : elog(ERROR, "wrong number of tlist entries");
3823 672 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
3824 104 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
3825 672 : colType = lnext(colTypes, colType);
3826 : }
3827 312 : if (colType != NULL)
3828 0 : elog(ERROR, "wrong number of tlist entries");
3829 312 : }
3830 :
3831 : /*
3832 : * targetIsInAllPartitionLists
3833 : * True if the TargetEntry is listed in the PARTITION BY clause
3834 : * of every window defined in the query.
3835 : *
3836 : * It would be safe to ignore windows not actually used by any window
3837 : * function, but it's not easy to get that info at this stage; and it's
3838 : * unlikely to be useful to spend any extra cycles getting it, since
3839 : * unreferenced window definitions are probably infrequent in practice.
3840 : */
3841 : static bool
3842 1090 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
3843 : {
3844 : ListCell *lc;
3845 :
3846 1202 : foreach(lc, query->windowClause)
3847 : {
3848 1114 : WindowClause *wc = (WindowClause *) lfirst(lc);
3849 :
3850 1114 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
3851 1002 : return false;
3852 : }
3853 88 : return true;
3854 : }
3855 :
3856 : /*
3857 : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3858 : *
3859 : * rinfo is a restriction clause applying to the given subquery (whose RTE
3860 : * has index rti in the parent query).
3861 : *
3862 : * Conditions checked here:
3863 : *
3864 : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3865 : * unclear that it will work correctly: SubLinks will already have been
3866 : * transformed into SubPlans in the qual, but not in the subquery). Note that
3867 : * SubLinks that transform to initplans are safe, and will be accepted here
3868 : * because what we'll see in the qual is just a Param referencing the initplan
3869 : * output.
3870 : *
3871 : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3872 : * functions.
3873 : *
3874 : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
3875 : * functions that are passed Var nodes, and therefore might reveal values from
3876 : * the subquery as side effects.
3877 : *
3878 : * 4. rinfo's clause must not refer to the whole-row output of the subquery
3879 : * (since there is no easy way to name that within the subquery itself).
3880 : *
3881 : * 5. rinfo's clause must not refer to any subquery output columns that were
3882 : * found to be unsafe to reference by subquery_is_pushdown_safe().
3883 : */
3884 : static pushdown_safe_type
3885 2828 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
3886 : pushdown_safety_info *safetyInfo)
3887 : {
3888 2828 : pushdown_safe_type safe = PUSHDOWN_SAFE;
3889 2828 : Node *qual = (Node *) rinfo->clause;
3890 : List *vars;
3891 : ListCell *vl;
3892 :
3893 : /* Refuse subselects (point 1) */
3894 2828 : if (contain_subplans(qual))
3895 66 : return PUSHDOWN_UNSAFE;
3896 :
3897 : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3898 3414 : if (safetyInfo->unsafeVolatile &&
3899 652 : contain_volatile_functions((Node *) rinfo))
3900 18 : return PUSHDOWN_UNSAFE;
3901 :
3902 : /* Refuse leaky quals if told to (point 3) */
3903 4092 : if (safetyInfo->unsafeLeaky &&
3904 1348 : contain_leaked_vars(qual))
3905 150 : return PUSHDOWN_UNSAFE;
3906 :
3907 : /*
3908 : * Examine all Vars used in clause. Since it's a restriction clause, all
3909 : * such Vars must refer to subselect output columns ... unless this is
3910 : * part of a LATERAL subquery, in which case there could be lateral
3911 : * references.
3912 : *
3913 : * By omitting the relevant flags, this also gives us a cheap sanity check
3914 : * that no aggregates or window functions appear in the qual. Those would
3915 : * be unsafe to push down, but at least for the moment we could never see
3916 : * any in a qual anyhow.
3917 : */
3918 2594 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
3919 5088 : foreach(vl, vars)
3920 : {
3921 2702 : Var *var = (Var *) lfirst(vl);
3922 :
3923 : /*
3924 : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3925 : * It's not clear whether a PHV could safely be pushed down, and even
3926 : * less clear whether such a situation could arise in any cases of
3927 : * practical interest anyway. So for the moment, just refuse to push
3928 : * down.
3929 : */
3930 2702 : if (!IsA(var, Var))
3931 : {
3932 0 : safe = PUSHDOWN_UNSAFE;
3933 0 : break;
3934 : }
3935 :
3936 : /*
3937 : * Punt if we find any lateral references. It would be safe to push
3938 : * these down, but we'd have to convert them into outer references,
3939 : * which subquery_push_qual lacks the infrastructure to do. The case
3940 : * arises so seldom that it doesn't seem worth working hard on.
3941 : */
3942 2702 : if (var->varno != rti)
3943 : {
3944 12 : safe = PUSHDOWN_UNSAFE;
3945 12 : break;
3946 : }
3947 :
3948 : /* Subqueries have no system columns */
3949 : Assert(var->varattno >= 0);
3950 :
3951 : /* Check point 4 */
3952 2690 : if (var->varattno == 0)
3953 : {
3954 0 : safe = PUSHDOWN_UNSAFE;
3955 0 : break;
3956 : }
3957 :
3958 : /* Check point 5 */
3959 2690 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
3960 : {
3961 526 : if (safetyInfo->unsafeFlags[var->varattno] &
3962 : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
3963 : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
3964 : {
3965 196 : safe = PUSHDOWN_UNSAFE;
3966 196 : break;
3967 : }
3968 : else
3969 : {
3970 : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
3971 330 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
3972 : /* don't break, we might find another Var that's unsafe */
3973 : }
3974 : }
3975 : }
3976 :
3977 2594 : list_free(vars);
3978 :
3979 2594 : return safe;
3980 : }
3981 :
3982 : /*
3983 : * subquery_push_qual - push down a qual that we have determined is safe
3984 : */
3985 : static void
3986 2398 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3987 : {
3988 2398 : if (subquery->setOperations != NULL)
3989 : {
3990 : /* Recurse to push it separately to each component query */
3991 132 : recurse_push_qual(subquery->setOperations, subquery,
3992 : rte, rti, qual);
3993 : }
3994 : else
3995 : {
3996 : /*
3997 : * We need to replace Vars in the qual (which must refer to outputs of
3998 : * the subquery) with copies of the subquery's targetlist expressions.
3999 : * Note that at this point, any uplevel Vars in the qual should have
4000 : * been replaced with Params, so they need no work.
4001 : *
4002 : * This step also ensures that when we are pushing into a setop tree,
4003 : * each component query gets its own copy of the qual.
4004 : */
4005 2266 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
4006 : subquery->targetList,
4007 : subquery->resultRelation,
4008 : REPLACEVARS_REPORT_ERROR, 0,
4009 : &subquery->hasSubLinks);
4010 :
4011 : /*
4012 : * Now attach the qual to the proper place: normally WHERE, but if the
4013 : * subquery uses grouping or aggregation, put it in HAVING (since the
4014 : * qual really refers to the group-result rows).
4015 : */
4016 2266 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
4017 270 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
4018 : else
4019 1996 : subquery->jointree->quals =
4020 1996 : make_and_qual(subquery->jointree->quals, qual);
4021 :
4022 : /*
4023 : * We need not change the subquery's hasAggs or hasSubLinks flags,
4024 : * since we can't be pushing down any aggregates that weren't there
4025 : * before, and we don't push down subselects at all.
4026 : */
4027 : }
4028 2398 : }
4029 :
4030 : /*
4031 : * Helper routine to recurse through setOperations tree
4032 : */
4033 : static void
4034 396 : recurse_push_qual(Node *setOp, Query *topquery,
4035 : RangeTblEntry *rte, Index rti, Node *qual)
4036 : {
4037 396 : if (IsA(setOp, RangeTblRef))
4038 : {
4039 264 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4040 264 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4041 264 : Query *subquery = subrte->subquery;
4042 :
4043 : Assert(subquery != NULL);
4044 264 : subquery_push_qual(subquery, rte, rti, qual);
4045 : }
4046 132 : else if (IsA(setOp, SetOperationStmt))
4047 : {
4048 132 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4049 :
4050 132 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4051 132 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4052 : }
4053 : else
4054 : {
4055 0 : elog(ERROR, "unrecognized node type: %d",
4056 : (int) nodeTag(setOp));
4057 : }
4058 396 : }
4059 :
4060 : /*****************************************************************************
4061 : * SIMPLIFYING SUBQUERY TARGETLISTS
4062 : *****************************************************************************/
4063 :
4064 : /*
4065 : * remove_unused_subquery_outputs
4066 : * Remove subquery targetlist items we don't need
4067 : *
4068 : * It's possible, even likely, that the upper query does not read all the
4069 : * output columns of the subquery. We can remove any such outputs that are
4070 : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4071 : * affect semantics otherwise (e.g., volatile functions can't be removed).
4072 : * This is useful not only because we might be able to remove expensive-to-
4073 : * compute expressions, but because deletion of output columns might allow
4074 : * optimizations such as join removal to occur within the subquery.
4075 : *
4076 : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4077 : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4078 : * parameter is modified by the function, so callers must make a copy if they
4079 : * need to use the passed in Bitmapset after calling this function.
4080 : *
4081 : * To avoid affecting column numbering in the targetlist, we don't physically
4082 : * remove unused tlist entries, but rather replace their expressions with NULL
4083 : * constants. This is implemented by modifying subquery->targetList.
4084 : */
4085 : static void
4086 11762 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4087 : Bitmapset *extra_used_attrs)
4088 : {
4089 : Bitmapset *attrs_used;
4090 : ListCell *lc;
4091 :
4092 : /*
4093 : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4094 : * the current callers use the Bitmapset after calling this function.
4095 : */
4096 11762 : attrs_used = extra_used_attrs;
4097 :
4098 : /*
4099 : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4100 : * could update all the child SELECTs' tlists, but it seems not worth the
4101 : * trouble presently.
4102 : */
4103 11762 : if (subquery->setOperations)
4104 1278 : return;
4105 :
4106 : /*
4107 : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4108 : * time: all its output columns must be used in the distinctClause.
4109 : */
4110 11028 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4111 246 : return;
4112 :
4113 : /*
4114 : * Collect a bitmap of all the output column numbers used by the upper
4115 : * query.
4116 : *
4117 : * Add all the attributes needed for joins or final output. Note: we must
4118 : * look at rel's targetlist, not the attr_needed data, because attr_needed
4119 : * isn't computed for inheritance child rels, cf set_append_rel_size().
4120 : * (XXX might be worth changing that sometime.)
4121 : */
4122 10782 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4123 :
4124 : /* Add all the attributes used by un-pushed-down restriction clauses. */
4125 11506 : foreach(lc, rel->baserestrictinfo)
4126 : {
4127 724 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4128 :
4129 724 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4130 : }
4131 :
4132 : /*
4133 : * If there's a whole-row reference to the subquery, we can't remove
4134 : * anything.
4135 : */
4136 10782 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4137 298 : return;
4138 :
4139 : /*
4140 : * Run through the tlist and zap entries we don't need. It's okay to
4141 : * modify the tlist items in-place because set_subquery_pathlist made a
4142 : * copy of the subquery.
4143 : */
4144 46732 : foreach(lc, subquery->targetList)
4145 : {
4146 36248 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4147 36248 : Node *texpr = (Node *) tle->expr;
4148 :
4149 : /*
4150 : * If it has a sortgroupref number, it's used in some sort/group
4151 : * clause so we'd better not remove it. Also, don't remove any
4152 : * resjunk columns, since their reason for being has nothing to do
4153 : * with anybody reading the subquery's output. (It's likely that
4154 : * resjunk columns in a sub-SELECT would always have ressortgroupref
4155 : * set, but even if they don't, it seems imprudent to remove them.)
4156 : */
4157 36248 : if (tle->ressortgroupref || tle->resjunk)
4158 2586 : continue;
4159 :
4160 : /*
4161 : * If it's used by the upper query, we can't remove it.
4162 : */
4163 33662 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4164 : attrs_used))
4165 25940 : continue;
4166 :
4167 : /*
4168 : * If it contains a set-returning function, we can't remove it since
4169 : * that could change the number of rows returned by the subquery.
4170 : */
4171 8742 : if (subquery->hasTargetSRFs &&
4172 1020 : expression_returns_set(texpr))
4173 752 : continue;
4174 :
4175 : /*
4176 : * If it contains volatile functions, we daren't remove it for fear
4177 : * that the user is expecting their side-effects to happen.
4178 : */
4179 6970 : if (contain_volatile_functions(texpr))
4180 26 : continue;
4181 :
4182 : /*
4183 : * OK, we don't need it. Replace the expression with a NULL constant.
4184 : * Preserve the exposed type of the expression, in case something
4185 : * looks at the rowtype of the subquery's result.
4186 : */
4187 6944 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4188 : exprTypmod(texpr),
4189 : exprCollation(texpr));
4190 : }
4191 : }
4192 :
4193 : /*
4194 : * create_partial_bitmap_paths
4195 : * Build partial bitmap heap path for the relation
4196 : */
4197 : void
4198 144632 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4199 : Path *bitmapqual)
4200 : {
4201 : int parallel_workers;
4202 : double pages_fetched;
4203 :
4204 : /* Compute heap pages for bitmap heap scan */
4205 144632 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4206 : NULL, NULL);
4207 :
4208 144632 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4209 : max_parallel_workers_per_gather);
4210 :
4211 144632 : if (parallel_workers <= 0)
4212 140450 : return;
4213 :
4214 4182 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4215 : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4216 : }
4217 :
4218 : /*
4219 : * Compute the number of parallel workers that should be used to scan a
4220 : * relation. We compute the parallel workers based on the size of the heap to
4221 : * be scanned and the size of the index to be scanned, then choose a minimum
4222 : * of those.
4223 : *
4224 : * "heap_pages" is the number of pages from the table that we expect to scan, or
4225 : * -1 if we don't expect to scan any.
4226 : *
4227 : * "index_pages" is the number of pages from the index that we expect to scan, or
4228 : * -1 if we don't expect to scan any.
4229 : *
4230 : * "max_workers" is caller's limit on the number of workers. This typically
4231 : * comes from a GUC.
4232 : */
4233 : int
4234 750118 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4235 : int max_workers)
4236 : {
4237 750118 : int parallel_workers = 0;
4238 :
4239 : /*
4240 : * If the user has set the parallel_workers reloption, use that; otherwise
4241 : * select a default number of workers.
4242 : */
4243 750118 : if (rel->rel_parallel_workers != -1)
4244 1914 : parallel_workers = rel->rel_parallel_workers;
4245 : else
4246 : {
4247 : /*
4248 : * If the number of pages being scanned is insufficient to justify a
4249 : * parallel scan, just return zero ... unless it's an inheritance
4250 : * child. In that case, we want to generate a parallel path here
4251 : * anyway. It might not be worthwhile just for this relation, but
4252 : * when combined with all of its inheritance siblings it may well pay
4253 : * off.
4254 : */
4255 748204 : if (rel->reloptkind == RELOPT_BASEREL &&
4256 710008 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4257 21100 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4258 709274 : return 0;
4259 :
4260 38930 : if (heap_pages >= 0)
4261 : {
4262 : int heap_parallel_threshold;
4263 36840 : int heap_parallel_workers = 1;
4264 :
4265 : /*
4266 : * Select the number of workers based on the log of the size of
4267 : * the relation. This probably needs to be a good deal more
4268 : * sophisticated, but we need something here for now. Note that
4269 : * the upper limit of the min_parallel_table_scan_size GUC is
4270 : * chosen to prevent overflow here.
4271 : */
4272 36840 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4273 41624 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4274 : {
4275 4784 : heap_parallel_workers++;
4276 4784 : heap_parallel_threshold *= 3;
4277 4784 : if (heap_parallel_threshold > INT_MAX / 3)
4278 0 : break; /* avoid overflow */
4279 : }
4280 :
4281 36840 : parallel_workers = heap_parallel_workers;
4282 : }
4283 :
4284 38930 : if (index_pages >= 0)
4285 : {
4286 9732 : int index_parallel_workers = 1;
4287 : int index_parallel_threshold;
4288 :
4289 : /* same calculation as for heap_pages above */
4290 9732 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4291 10008 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4292 : {
4293 276 : index_parallel_workers++;
4294 276 : index_parallel_threshold *= 3;
4295 276 : if (index_parallel_threshold > INT_MAX / 3)
4296 0 : break; /* avoid overflow */
4297 : }
4298 :
4299 9732 : if (parallel_workers > 0)
4300 7642 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4301 : else
4302 2090 : parallel_workers = index_parallel_workers;
4303 : }
4304 : }
4305 :
4306 : /* In no case use more than caller supplied maximum number of workers */
4307 40844 : parallel_workers = Min(parallel_workers, max_workers);
4308 :
4309 40844 : return parallel_workers;
4310 : }
4311 :
4312 : /*
4313 : * generate_partitionwise_join_paths
4314 : * Create paths representing partitionwise join for given partitioned
4315 : * join relation.
4316 : *
4317 : * This must not be called until after we are done adding paths for all
4318 : * child-joins. Otherwise, add_path might delete a path to which some path
4319 : * generated here has a reference.
4320 : */
4321 : void
4322 216138 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4323 : {
4324 216138 : List *live_children = NIL;
4325 : int cnt_parts;
4326 : int num_parts;
4327 : RelOptInfo **part_rels;
4328 :
4329 : /* Handle only join relations here. */
4330 216138 : if (!IS_JOIN_REL(rel))
4331 0 : return;
4332 :
4333 : /* We've nothing to do if the relation is not partitioned. */
4334 216138 : if (!IS_PARTITIONED_REL(rel))
4335 214374 : return;
4336 :
4337 : /* The relation should have consider_partitionwise_join set. */
4338 : Assert(rel->consider_partitionwise_join);
4339 :
4340 : /* Guard against stack overflow due to overly deep partition hierarchy. */
4341 1764 : check_stack_depth();
4342 :
4343 1764 : num_parts = rel->nparts;
4344 1764 : part_rels = rel->part_rels;
4345 :
4346 : /* Collect non-dummy child-joins. */
4347 6602 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4348 : {
4349 4838 : RelOptInfo *child_rel = part_rels[cnt_parts];
4350 :
4351 : /* If it's been pruned entirely, it's certainly dummy. */
4352 4838 : if (child_rel == NULL)
4353 52 : continue;
4354 :
4355 : /* Make partitionwise join paths for this partitioned child-join. */
4356 4786 : generate_partitionwise_join_paths(root, child_rel);
4357 :
4358 : /* If we failed to make any path for this child, we must give up. */
4359 4786 : if (child_rel->pathlist == NIL)
4360 : {
4361 : /*
4362 : * Mark the parent joinrel as unpartitioned so that later
4363 : * functions treat it correctly.
4364 : */
4365 0 : rel->nparts = 0;
4366 0 : return;
4367 : }
4368 :
4369 : /* Else, identify the cheapest path for it. */
4370 4786 : set_cheapest(child_rel);
4371 :
4372 : /* Dummy children need not be scanned, so ignore those. */
4373 4786 : if (IS_DUMMY_REL(child_rel))
4374 0 : continue;
4375 :
4376 : #ifdef OPTIMIZER_DEBUG
4377 : pprint(child_rel);
4378 : #endif
4379 :
4380 4786 : live_children = lappend(live_children, child_rel);
4381 : }
4382 :
4383 : /* If all child-joins are dummy, parent join is also dummy. */
4384 1764 : if (!live_children)
4385 : {
4386 0 : mark_dummy_rel(rel);
4387 0 : return;
4388 : }
4389 :
4390 : /* Build additional paths for this rel from child-join paths. */
4391 1764 : add_paths_to_append_rel(root, rel, live_children);
4392 1764 : list_free(live_children);
4393 : }
|