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 326154 : 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 326154 : set_base_rel_consider_startup(root);
179 :
180 : /*
181 : * Compute size estimates and consider_parallel flags for each base rel.
182 : */
183 326154 : 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 326120 : total_pages = 0;
200 995160 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
201 : {
202 669040 : RelOptInfo *brel = root->simple_rel_array[rti];
203 :
204 : /* there may be empty slots corresponding to non-baserel RTEs */
205 669040 : if (brel == NULL)
206 158146 : continue;
207 :
208 : Assert(brel->relid == rti); /* sanity check on array */
209 :
210 510894 : if (IS_DUMMY_REL(brel))
211 1268 : continue;
212 :
213 509626 : if (IS_SIMPLE_REL(brel))
214 509626 : total_pages += (double) brel->pages;
215 : }
216 326120 : root->total_table_pages = total_pages;
217 :
218 : /*
219 : * Generate access paths for each base rel.
220 : */
221 326120 : set_base_rel_pathlists(root);
222 :
223 : /*
224 : * Generate access paths for the entire join tree.
225 : */
226 326120 : 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 326120 : 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 326154 : 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 369918 : foreach(lc, root->join_info_list)
265 : {
266 43764 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
267 : int varno;
268 :
269 52384 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
270 8620 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
271 : {
272 8438 : RelOptInfo *rel = find_base_rel(root, varno);
273 :
274 8438 : rel->consider_param_startup = true;
275 : }
276 : }
277 326154 : }
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 326154 : set_base_rel_sizes(PlannerInfo *root)
291 : {
292 : Index rti;
293 :
294 995196 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
295 : {
296 669076 : RelOptInfo *rel = root->simple_rel_array[rti];
297 : RangeTblEntry *rte;
298 :
299 : /* there may be empty slots corresponding to non-baserel RTEs */
300 669076 : if (rel == NULL)
301 158148 : continue;
302 :
303 : Assert(rel->relid == rti); /* sanity check on array */
304 :
305 : /* ignore RTEs that are "other rels" */
306 510928 : if (rel->reloptkind != RELOPT_BASEREL)
307 50590 : continue;
308 :
309 460338 : 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 460338 : if (root->glob->parallelModeOK)
320 367164 : set_rel_consider_parallel(root, rel, rte);
321 :
322 460338 : set_rel_size(root, rel, rti, rte);
323 : }
324 326120 : }
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 326120 : set_base_rel_pathlists(PlannerInfo *root)
334 : {
335 : Index rti;
336 :
337 995160 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
338 : {
339 669040 : RelOptInfo *rel = root->simple_rel_array[rti];
340 :
341 : /* there may be empty slots corresponding to non-baserel RTEs */
342 669040 : if (rel == NULL)
343 158146 : continue;
344 :
345 : Assert(rel->relid == rti); /* sanity check on array */
346 :
347 : /* ignore RTEs that are "other rels" */
348 510894 : if (rel->reloptkind != RELOPT_BASEREL)
349 50590 : continue;
350 :
351 460304 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
352 : }
353 326120 : }
354 :
355 : /*
356 : * set_rel_size
357 : * Set size estimates for a base relation
358 : */
359 : static void
360 510618 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
361 : Index rti, RangeTblEntry *rte)
362 : {
363 970956 : if (rel->reloptkind == RELOPT_BASEREL &&
364 460338 : 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 510080 : else if (rte->inh)
380 : {
381 : /* It's an "append relation", process accordingly */
382 24158 : set_append_rel_size(root, rel, rti, rte);
383 : }
384 : else
385 : {
386 485922 : switch (rel->rtekind)
387 : {
388 403374 : case RTE_RELATION:
389 403374 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
390 : {
391 : /* Foreign table */
392 2422 : set_foreign_size(root, rel, rte);
393 : }
394 400952 : 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 400912 : 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 400606 : set_plain_rel_size(root, rel, rte);
412 : }
413 403340 : break;
414 11588 : 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 11588 : set_subquery_pathlist(root, rel, rti, rte);
422 11588 : break;
423 52316 : case RTE_FUNCTION:
424 52316 : set_function_size_estimates(root, rel);
425 52316 : break;
426 626 : case RTE_TABLEFUNC:
427 626 : set_tablefunc_size_estimates(root, rel);
428 626 : break;
429 8240 : case RTE_VALUES:
430 8240 : set_values_size_estimates(root, rel);
431 8240 : break;
432 5092 : 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 5092 : if (rte->self_reference)
440 1006 : set_worktable_pathlist(root, rel, rte);
441 : else
442 4086 : set_cte_pathlist(root, rel, rte);
443 5092 : break;
444 466 : case RTE_NAMEDTUPLESTORE:
445 : /* Might as well just build the path immediately */
446 466 : set_namedtuplestore_pathlist(root, rel, rte);
447 466 : break;
448 4220 : case RTE_RESULT:
449 : /* Might as well just build the path immediately */
450 4220 : set_result_pathlist(root, rel, rte);
451 4220 : 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 510582 : }
463 :
464 : /*
465 : * set_rel_pathlist
466 : * Build access paths for a base relation
467 : */
468 : static void
469 510720 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
470 : Index rti, RangeTblEntry *rte)
471 : {
472 510720 : if (IS_DUMMY_REL(rel))
473 : {
474 : /* We already proved the relation empty, so nothing more to do */
475 : }
476 509590 : else if (rte->inh)
477 : {
478 : /* It's an "append relation", process accordingly */
479 23862 : set_append_rel_pathlist(root, rel, rti, rte);
480 : }
481 : else
482 : {
483 485728 : switch (rel->rtekind)
484 : {
485 403300 : case RTE_RELATION:
486 403300 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
487 : {
488 : /* Foreign table */
489 2418 : set_foreign_pathlist(root, rel, rte);
490 : }
491 400882 : 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 400576 : set_plain_rel_pathlist(root, rel, rte);
500 : }
501 403300 : break;
502 11468 : case RTE_SUBQUERY:
503 : /* Subquery --- fully handled during set_rel_size */
504 11468 : break;
505 52316 : case RTE_FUNCTION:
506 : /* RangeFunction */
507 52316 : set_function_pathlist(root, rel, rte);
508 52316 : break;
509 626 : case RTE_TABLEFUNC:
510 : /* Table Function */
511 626 : set_tablefunc_pathlist(root, rel, rte);
512 626 : break;
513 8240 : case RTE_VALUES:
514 : /* Values list */
515 8240 : set_values_pathlist(root, rel, rte);
516 8240 : break;
517 5092 : case RTE_CTE:
518 : /* CTE reference --- fully handled during set_rel_size */
519 5092 : break;
520 466 : case RTE_NAMEDTUPLESTORE:
521 : /* tuplestore reference --- fully handled during set_rel_size */
522 466 : break;
523 4220 : case RTE_RESULT:
524 : /* simple Result --- fully handled during set_rel_size */
525 4220 : 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 510720 : 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 510720 : if (rel->reloptkind == RELOPT_BASEREL &&
556 460304 : !bms_equal(rel->relids, root->all_query_rels))
557 231506 : generate_useful_gather_paths(root, rel, false);
558 :
559 : /* Now find the cheapest of the paths for this rel */
560 510720 : set_cheapest(rel);
561 :
562 : #ifdef OPTIMIZER_DEBUG
563 : pprint(rel);
564 : #endif
565 510720 : }
566 :
567 : /*
568 : * set_plain_rel_size
569 : * Set size estimates for a plain relation (no subquery, no inheritance)
570 : */
571 : static void
572 400606 : 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 400606 : check_index_predicates(root, rel);
579 :
580 : /* Mark rel with estimated output rows, width, etc */
581 400606 : set_baserel_size_estimates(root, rel);
582 400576 : }
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 403276 : 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 403276 : switch (rte->rtekind)
606 : {
607 350112 : 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 350112 : 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 342100 : 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 342052 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
645 : {
646 : Assert(rel->fdwroutine);
647 1552 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
648 1480 : 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 340572 : break;
660 :
661 12690 : 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 12690 : Query *subquery = castNode(Query, rte->subquery);
685 :
686 12690 : if (limit_needed(subquery))
687 466 : return;
688 : }
689 12224 : 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 28606 : case RTE_FUNCTION:
697 : /* Check for parallel-restricted functions. */
698 28606 : if (!is_parallel_safe(root, (Node *) rte->functions))
699 13868 : return;
700 14738 : 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 4186 : 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 4186 : return;
722 :
723 438 : case RTE_NAMEDTUPLESTORE:
724 :
725 : /*
726 : * tuplestore cannot be shared, at least without more
727 : * infrastructure to support that.
728 : */
729 438 : 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 374140 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
750 28474 : 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 345666 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
757 18 : return;
758 :
759 : /* We have a winner. */
760 345648 : 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 400576 : 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 400576 : 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 400576 : if (create_tidscan_paths(root, rel))
788 404 : return;
789 :
790 : /* Consider sequential scan */
791 400172 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
792 :
793 : /* If appropriate, consider parallel sequential scan */
794 400172 : if (rel->consider_parallel && required_outer == NULL)
795 299128 : create_plain_partial_paths(root, rel);
796 :
797 : /* Consider index scans */
798 400172 : 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 299128 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
807 : {
808 : int parallel_workers;
809 :
810 299128 : 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 299128 : if (parallel_workers <= 0)
815 272998 : return;
816 :
817 : /* Add an unordered partial path based on a parallel sequential scan. */
818 26130 : 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 2422 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
915 : {
916 : /* Mark rel with estimated output rows, width, etc */
917 2422 : set_foreign_size_estimates(root, rel);
918 :
919 : /* Let FDW adjust the size estimates, if it can */
920 2422 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
921 :
922 : /* ... but do not let it set the rows estimate to zero */
923 2418 : 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 2418 : rel->tuples = Max(rel->tuples, rel->rows);
931 2418 : }
932 :
933 : /*
934 : * set_foreign_pathlist
935 : * Build access paths for a foreign table RTE
936 : */
937 : static void
938 2418 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
939 : {
940 : /* Call the FDW's GetForeignPaths function to generate path(s) */
941 2418 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
942 2418 : }
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 24158 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
957 : Index rti, RangeTblEntry *rte)
958 : {
959 24158 : 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 24158 : 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 24158 : if (enable_partitionwise_join &&
979 4340 : rel->reloptkind == RELOPT_BASEREL &&
980 3608 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
981 3608 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
982 3532 : 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 24158 : has_live_children = false;
1008 24158 : parent_tuples = 0;
1009 24158 : parent_rows = 0;
1010 24158 : parent_size = 0;
1011 24158 : nattrs = rel->max_attr - rel->min_attr + 1;
1012 24158 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1013 :
1014 118774 : foreach(l, root->append_rel_list)
1015 : {
1016 94618 : 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 94618 : if (appinfo->parent_relid != parentRTindex)
1027 44476 : continue;
1028 :
1029 50484 : childRTindex = appinfo->child_relid;
1030 50484 : 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 50484 : 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 50484 : 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 50466 : 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 186 : set_dummy_rel_pathlist(childrel);
1057 186 : 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 50280 : childrinfos = NIL;
1071 63468 : foreach(lc, rel->joininfo)
1072 : {
1073 13188 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1074 :
1075 13188 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1076 10866 : childrinfos = lappend(childrinfos,
1077 10866 : adjust_appendrel_attrs(root,
1078 : (Node *) rinfo,
1079 : 1, &appinfo));
1080 : }
1081 50280 : 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 100560 : childrel->reltarget->exprs = (List *)
1094 50280 : adjust_appendrel_attrs(root,
1095 50280 : (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 50280 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1107 28924 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1108 50280 : 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 50280 : if (rel->consider_partitionwise_join)
1130 11558 : 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 50280 : if (root->glob->parallelModeOK && rel->consider_parallel)
1140 36112 : set_rel_consider_parallel(root, childrel, childRTE);
1141 :
1142 : /*
1143 : * Compute the child's size.
1144 : */
1145 50280 : 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 50278 : if (IS_DUMMY_REL(childrel))
1153 138 : continue;
1154 :
1155 : /* We have at least one live child. */
1156 50140 : 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 50140 : if (!childrel->consider_parallel)
1168 14802 : rel->consider_parallel = false;
1169 :
1170 : /*
1171 : * Accumulate size information from each live child.
1172 : */
1173 : Assert(childrel->rows > 0);
1174 :
1175 50140 : parent_tuples += childrel->tuples;
1176 50140 : parent_rows += childrel->rows;
1177 50140 : 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 151768 : forboth(parentvars, rel->reltarget->exprs,
1188 : childvars, childrel->reltarget->exprs)
1189 : {
1190 101628 : Var *parentvar = (Var *) lfirst(parentvars);
1191 101628 : Node *childvar = (Node *) lfirst(childvars);
1192 :
1193 101628 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1194 : {
1195 88924 : int pndx = parentvar->varattno - rel->min_attr;
1196 88924 : int32 child_width = 0;
1197 :
1198 88924 : if (IsA(childvar, Var) &&
1199 84132 : ((Var *) childvar)->varno == childrel->relid)
1200 : {
1201 84066 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1202 :
1203 84066 : child_width = childrel->attr_widths[cndx];
1204 : }
1205 88924 : if (child_width <= 0)
1206 4858 : child_width = get_typavgwidth(exprType(childvar),
1207 : exprTypmod(childvar));
1208 : Assert(child_width > 0);
1209 88924 : parent_attrsizes[pndx] += child_width * childrel->rows;
1210 : }
1211 : }
1212 : }
1213 :
1214 24156 : if (has_live_children)
1215 : {
1216 : /*
1217 : * Save the finished size estimates.
1218 : */
1219 : int i;
1220 :
1221 : Assert(parent_rows > 0);
1222 23862 : rel->tuples = parent_tuples;
1223 23862 : rel->rows = parent_rows;
1224 23862 : rel->reltarget->width = rint(parent_size / parent_rows);
1225 222872 : for (i = 0; i < nattrs; i++)
1226 199010 : 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 24156 : pfree(parent_attrsizes);
1244 24156 : }
1245 :
1246 : /*
1247 : * set_append_rel_pathlist
1248 : * Build access paths for an "append relation"
1249 : */
1250 : static void
1251 23862 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1252 : Index rti, RangeTblEntry *rte)
1253 : {
1254 23862 : int parentRTindex = rti;
1255 23862 : 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 118064 : foreach(l, root->append_rel_list)
1263 : {
1264 94202 : 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 94202 : if (appinfo->parent_relid != parentRTindex)
1271 43786 : continue;
1272 :
1273 : /* Re-locate the child RTE and RelOptInfo */
1274 50416 : childRTindex = appinfo->child_relid;
1275 50416 : childRTE = root->simple_rte_array[childRTindex];
1276 50416 : 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 50416 : if (!rel->consider_parallel)
1285 14998 : childrel->consider_parallel = false;
1286 :
1287 : /*
1288 : * Compute the child's access paths.
1289 : */
1290 50416 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1291 :
1292 : /*
1293 : * If child is dummy, ignore it.
1294 : */
1295 50416 : if (IS_DUMMY_REL(childrel))
1296 276 : continue;
1297 :
1298 : /*
1299 : * Child is live, so add it to the live_childrels list for use below.
1300 : */
1301 50140 : live_childrels = lappend(live_childrels, childrel);
1302 : }
1303 :
1304 : /* Add paths to the append relation. */
1305 23862 : add_paths_to_append_rel(root, rel, live_childrels);
1306 23862 : }
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 38860 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1322 : List *live_childrels)
1323 : {
1324 38860 : List *subpaths = NIL;
1325 38860 : bool subpaths_valid = true;
1326 38860 : List *startup_subpaths = NIL;
1327 38860 : bool startup_subpaths_valid = true;
1328 38860 : List *partial_subpaths = NIL;
1329 38860 : List *pa_partial_subpaths = NIL;
1330 38860 : List *pa_nonpartial_subpaths = NIL;
1331 38860 : bool partial_subpaths_valid = true;
1332 : bool pa_subpaths_valid;
1333 38860 : List *all_child_pathkeys = NIL;
1334 38860 : List *all_child_outers = NIL;
1335 : ListCell *l;
1336 38860 : double partial_rows = -1;
1337 :
1338 : /* If appropriate, consider parallel append */
1339 38860 : 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 118444 : foreach(l, live_childrels)
1347 : {
1348 79584 : RelOptInfo *childrel = lfirst(l);
1349 : ListCell *lcp;
1350 79584 : 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 79584 : if (childrel->pathlist != NIL &&
1361 79584 : childrel->cheapest_total_path->param_info == NULL)
1362 78852 : 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 79584 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1373 1622 : {
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 1622 : if (root->tuple_fraction > 0.0)
1382 : cheapest_path =
1383 1622 : 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 1622 : accumulate_append_subpath(cheapest_path,
1391 : &startup_subpaths,
1392 : NULL);
1393 : }
1394 : else
1395 77962 : startup_subpaths_valid = false;
1396 :
1397 :
1398 : /* Same idea, but for a partial plan. */
1399 79584 : if (childrel->partial_pathlist != NIL)
1400 : {
1401 47230 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1402 47230 : accumulate_append_subpath(cheapest_partial_path,
1403 : &partial_subpaths, NULL);
1404 : }
1405 : else
1406 32354 : partial_subpaths_valid = false;
1407 :
1408 : /*
1409 : * Same idea, but for a parallel append mixing partial and non-partial
1410 : * paths.
1411 : */
1412 79584 : if (pa_subpaths_valid)
1413 : {
1414 55308 : Path *nppath = NULL;
1415 :
1416 : nppath =
1417 55308 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1418 :
1419 55308 : 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 54790 : else if (nppath == NULL ||
1425 46780 : (cheapest_partial_path != NULL &&
1426 46780 : 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 46640 : 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 181452 : foreach(lcp, childrel->pathlist)
1462 : {
1463 101868 : Path *childpath = (Path *) lfirst(lcp);
1464 101868 : List *childkeys = childpath->pathkeys;
1465 101868 : Relids childouter = PATH_REQ_OUTER(childpath);
1466 :
1467 : /* Unsorted paths don't contribute to pathkey list */
1468 101868 : if (childkeys != NIL)
1469 : {
1470 : ListCell *lpk;
1471 22116 : bool found = false;
1472 :
1473 : /* Have we already seen this ordering? */
1474 22304 : foreach(lpk, all_child_pathkeys)
1475 : {
1476 16008 : List *existing_pathkeys = (List *) lfirst(lpk);
1477 :
1478 16008 : if (compare_pathkeys(existing_pathkeys,
1479 : childkeys) == PATHKEYS_EQUAL)
1480 : {
1481 15820 : found = true;
1482 15820 : break;
1483 : }
1484 : }
1485 22116 : if (!found)
1486 : {
1487 : /* No, so add it to all_child_pathkeys */
1488 6296 : all_child_pathkeys = lappend(all_child_pathkeys,
1489 : childkeys);
1490 : }
1491 : }
1492 :
1493 : /* Unparameterized paths don't contribute to param-set list */
1494 101868 : if (childouter)
1495 : {
1496 : ListCell *lco;
1497 6608 : bool found = false;
1498 :
1499 : /* Have we already seen this param set? */
1500 7274 : foreach(lco, all_child_outers)
1501 : {
1502 4760 : Relids existing_outers = (Relids) lfirst(lco);
1503 :
1504 4760 : if (bms_equal(existing_outers, childouter))
1505 : {
1506 4094 : found = true;
1507 4094 : break;
1508 : }
1509 : }
1510 6608 : if (!found)
1511 : {
1512 : /* No, so add it to all_child_outers */
1513 2514 : 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 38860 : if (subpaths_valid)
1526 38548 : 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 38860 : if (startup_subpaths_valid)
1532 664 : 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 38860 : if (partial_subpaths_valid && partial_subpaths != NIL)
1540 : {
1541 : AppendPath *appendpath;
1542 : ListCell *lc;
1543 21358 : int parallel_workers = 0;
1544 :
1545 : /* Find the highest number of workers requested for any subpath. */
1546 72726 : foreach(lc, partial_subpaths)
1547 : {
1548 51368 : Path *path = lfirst(lc);
1549 :
1550 51368 : 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 21358 : if (enable_parallel_append)
1564 : {
1565 21310 : parallel_workers = Max(parallel_workers,
1566 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1567 21310 : 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 21358 : 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 21358 : partial_rows = appendpath->path.rows;
1583 :
1584 : /* Add the path. */
1585 21358 : 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 38860 : 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 38860 : if (subpaths_valid)
1634 38548 : 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 41374 : foreach(l, all_child_outers)
1651 : {
1652 2514 : Relids required_outer = (Relids) lfirst(l);
1653 : ListCell *lcr;
1654 :
1655 : /* Select the child paths for an Append with this parameterization */
1656 2514 : subpaths = NIL;
1657 2514 : subpaths_valid = true;
1658 9212 : foreach(lcr, live_childrels)
1659 : {
1660 6710 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1661 : Path *subpath;
1662 :
1663 6710 : 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 6710 : subpath = get_cheapest_parameterized_child_path(root,
1671 : childrel,
1672 : required_outer);
1673 6710 : if (subpath == NULL)
1674 : {
1675 : /* failed to make a suitable path for this child */
1676 12 : subpaths_valid = false;
1677 12 : break;
1678 : }
1679 6698 : accumulate_append_subpath(subpath, &subpaths, NULL);
1680 : }
1681 :
1682 2514 : if (subpaths_valid)
1683 2502 : add_path(rel, (Path *)
1684 2502 : 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 38860 : if (list_length(live_childrels) == 1)
1697 : {
1698 14286 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1699 :
1700 : /* skip the cheapest partial path, since we already used that above */
1701 14490 : 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 38860 : }
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 38548 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1748 : List *live_childrels,
1749 : List *all_child_pathkeys)
1750 : {
1751 : ListCell *lcp;
1752 38548 : List *partition_pathkeys = NIL;
1753 38548 : List *partition_pathkeys_desc = NIL;
1754 38548 : bool partition_pathkeys_partial = true;
1755 38548 : 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 66218 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
1766 27670 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1767 : {
1768 23090 : partition_pathkeys = build_partition_pathkeys(root, rel,
1769 : ForwardScanDirection,
1770 : &partition_pathkeys_partial);
1771 :
1772 23090 : 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 44784 : foreach(lcp, all_child_pathkeys)
1789 : {
1790 6236 : List *pathkeys = (List *) lfirst(lcp);
1791 6236 : List *startup_subpaths = NIL;
1792 6236 : List *total_subpaths = NIL;
1793 6236 : List *fractional_subpaths = NIL;
1794 6236 : 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 6236 : match_partition_order =
1811 10188 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1812 4148 : (!partition_pathkeys_partial &&
1813 196 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1814 :
1815 13912 : match_partition_order_desc = !match_partition_order &&
1816 3856 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1817 3884 : (!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 6236 : if (match_partition_order_desc)
1828 : {
1829 : /* loop backward */
1830 48 : first_index = list_length(live_childrels) - 1;
1831 48 : end_index = -1;
1832 48 : 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 48 : match_partition_order = true;
1839 : }
1840 : else
1841 : {
1842 : /* for all other case, loop forward */
1843 6188 : first_index = 0;
1844 6188 : end_index = list_length(live_childrels);
1845 6188 : direction = 1;
1846 : }
1847 :
1848 : /* Select the child paths for this ordering... */
1849 22904 : for (int i = first_index; i != end_index; i += direction)
1850 : {
1851 16668 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1852 : Path *cheapest_startup,
1853 : *cheapest_total,
1854 16668 : *cheapest_fractional = NULL;
1855 :
1856 : /* Locate the right paths, if they are available. */
1857 : cheapest_startup =
1858 16668 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1859 : pathkeys,
1860 : NULL,
1861 : STARTUP_COST,
1862 : false);
1863 : cheapest_total =
1864 16668 : 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 16668 : 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 16668 : if (root->tuple_fraction > 0)
1893 : {
1894 860 : double path_fraction = root->tuple_fraction;
1895 :
1896 : /*
1897 : * Merge Append considers only live children relations. Dummy
1898 : * relations must be filtered out before.
1899 : */
1900 : Assert(childrel->rows > 0);
1901 :
1902 : /* Convert absolute limit to a path fraction */
1903 860 : if (path_fraction >= 1.0)
1904 860 : path_fraction /= childrel->rows;
1905 :
1906 : cheapest_fractional =
1907 860 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1908 : pathkeys,
1909 : NULL,
1910 : path_fraction);
1911 :
1912 : /*
1913 : * If we found no path with matching pathkeys, use the
1914 : * cheapest total path instead.
1915 : *
1916 : * XXX We might consider partially sorted paths too (with an
1917 : * incremental sort on top). But we'd have to build all the
1918 : * incremental paths, do the costing etc.
1919 : */
1920 860 : if (!cheapest_fractional)
1921 44 : cheapest_fractional = cheapest_total;
1922 : }
1923 :
1924 : /*
1925 : * Notice whether we actually have different paths for the
1926 : * "cheapest" and "total" cases; frequently there will be no point
1927 : * in two create_merge_append_path() calls.
1928 : */
1929 16668 : if (cheapest_startup != cheapest_total)
1930 96 : startup_neq_total = true;
1931 :
1932 : /*
1933 : * Collect the appropriate child paths. The required logic varies
1934 : * for the Append and MergeAppend cases.
1935 : */
1936 16668 : if (match_partition_order)
1937 : {
1938 : /*
1939 : * We're going to make a plain Append path. We don't need
1940 : * most of what accumulate_append_subpath would do, but we do
1941 : * want to cut out child Appends or MergeAppends if they have
1942 : * just a single subpath (and hence aren't doing anything
1943 : * useful).
1944 : */
1945 6492 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1946 6492 : cheapest_total = get_singleton_append_subpath(cheapest_total);
1947 :
1948 6492 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1949 6492 : total_subpaths = lappend(total_subpaths, cheapest_total);
1950 :
1951 6492 : if (cheapest_fractional)
1952 : {
1953 144 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1954 144 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1955 : }
1956 : }
1957 : else
1958 : {
1959 : /*
1960 : * Otherwise, rely on accumulate_append_subpath to collect the
1961 : * child paths for the MergeAppend.
1962 : */
1963 10176 : accumulate_append_subpath(cheapest_startup,
1964 : &startup_subpaths, NULL);
1965 10176 : accumulate_append_subpath(cheapest_total,
1966 : &total_subpaths, NULL);
1967 :
1968 10176 : if (cheapest_fractional)
1969 716 : accumulate_append_subpath(cheapest_fractional,
1970 : &fractional_subpaths, NULL);
1971 : }
1972 : }
1973 :
1974 : /* ... and build the Append or MergeAppend paths */
1975 6236 : if (match_partition_order)
1976 : {
1977 : /* We only need Append */
1978 2428 : add_path(rel, (Path *) create_append_path(root,
1979 : rel,
1980 : startup_subpaths,
1981 : NIL,
1982 : pathkeys,
1983 : NULL,
1984 : 0,
1985 : false,
1986 : -1));
1987 2428 : if (startup_neq_total)
1988 0 : add_path(rel, (Path *) create_append_path(root,
1989 : rel,
1990 : total_subpaths,
1991 : NIL,
1992 : pathkeys,
1993 : NULL,
1994 : 0,
1995 : false,
1996 : -1));
1997 :
1998 2428 : if (fractional_subpaths)
1999 72 : add_path(rel, (Path *) create_append_path(root,
2000 : rel,
2001 : fractional_subpaths,
2002 : NIL,
2003 : pathkeys,
2004 : NULL,
2005 : 0,
2006 : false,
2007 : -1));
2008 : }
2009 : else
2010 : {
2011 : /* We need MergeAppend */
2012 3808 : add_path(rel, (Path *) create_merge_append_path(root,
2013 : rel,
2014 : startup_subpaths,
2015 : pathkeys,
2016 : NULL));
2017 3808 : if (startup_neq_total)
2018 60 : add_path(rel, (Path *) create_merge_append_path(root,
2019 : rel,
2020 : total_subpaths,
2021 : pathkeys,
2022 : NULL));
2023 :
2024 3808 : if (fractional_subpaths)
2025 256 : add_path(rel, (Path *) create_merge_append_path(root,
2026 : rel,
2027 : fractional_subpaths,
2028 : pathkeys,
2029 : NULL));
2030 : }
2031 : }
2032 38548 : }
2033 :
2034 : /*
2035 : * get_cheapest_parameterized_child_path
2036 : * Get cheapest path for this relation that has exactly the requested
2037 : * parameterization.
2038 : *
2039 : * Returns NULL if unable to create such a path.
2040 : */
2041 : static Path *
2042 6710 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2043 : Relids required_outer)
2044 : {
2045 : Path *cheapest;
2046 : ListCell *lc;
2047 :
2048 : /*
2049 : * Look up the cheapest existing path with no more than the needed
2050 : * parameterization. If it has exactly the needed parameterization, we're
2051 : * done.
2052 : */
2053 6710 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2054 : NIL,
2055 : required_outer,
2056 : TOTAL_COST,
2057 : false);
2058 : Assert(cheapest != NULL);
2059 6710 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2060 6376 : return cheapest;
2061 :
2062 : /*
2063 : * Otherwise, we can "reparameterize" an existing path to match the given
2064 : * parameterization, which effectively means pushing down additional
2065 : * joinquals to be checked within the path's scan. However, some existing
2066 : * paths might check the available joinquals already while others don't;
2067 : * therefore, it's not clear which existing path will be cheapest after
2068 : * reparameterization. We have to go through them all and find out.
2069 : */
2070 334 : cheapest = NULL;
2071 1168 : foreach(lc, rel->pathlist)
2072 : {
2073 834 : Path *path = (Path *) lfirst(lc);
2074 :
2075 : /* Can't use it if it needs more than requested parameterization */
2076 834 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2077 24 : continue;
2078 :
2079 : /*
2080 : * Reparameterization can only increase the path's cost, so if it's
2081 : * already more expensive than the current cheapest, forget it.
2082 : */
2083 1266 : if (cheapest != NULL &&
2084 456 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2085 384 : continue;
2086 :
2087 : /* Reparameterize if needed, then recheck cost */
2088 426 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2089 : {
2090 350 : path = reparameterize_path(root, path, required_outer, 1.0);
2091 350 : if (path == NULL)
2092 32 : continue; /* failed to reparameterize this one */
2093 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2094 :
2095 318 : if (cheapest != NULL &&
2096 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2097 0 : continue;
2098 : }
2099 :
2100 : /* We have a new best path */
2101 394 : cheapest = path;
2102 : }
2103 :
2104 : /* Return the best path, or NULL if we found no suitable candidate */
2105 334 : return cheapest;
2106 : }
2107 :
2108 : /*
2109 : * accumulate_append_subpath
2110 : * Add a subpath to the list being built for an Append or MergeAppend.
2111 : *
2112 : * It's possible that the child is itself an Append or MergeAppend path, in
2113 : * which case we can "cut out the middleman" and just add its child paths to
2114 : * our own list. (We don't try to do this earlier because we need to apply
2115 : * both levels of transformation to the quals.)
2116 : *
2117 : * Note that if we omit a child MergeAppend in this way, we are effectively
2118 : * omitting a sort step, which seems fine: if the parent is to be an Append,
2119 : * its result would be unsorted anyway, while if the parent is to be a
2120 : * MergeAppend, there's no point in a separate sort on a child.
2121 : *
2122 : * Normally, either path is a partial path and subpaths is a list of partial
2123 : * paths, or else path is a non-partial plan and subpaths is a list of those.
2124 : * However, if path is a parallel-aware Append, then we add its partial path
2125 : * children to subpaths and the rest to special_subpaths. If the latter is
2126 : * NULL, we don't flatten the path at all (unless it contains only partial
2127 : * paths).
2128 : */
2129 : static void
2130 210260 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2131 : {
2132 210260 : if (IsA(path, AppendPath))
2133 : {
2134 14344 : AppendPath *apath = (AppendPath *) path;
2135 :
2136 14344 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2137 : {
2138 14152 : *subpaths = list_concat(*subpaths, apath->subpaths);
2139 14152 : return;
2140 : }
2141 192 : else if (special_subpaths != NULL)
2142 : {
2143 : List *new_special_subpaths;
2144 :
2145 : /* Split Parallel Append into partial and non-partial subpaths */
2146 96 : *subpaths = list_concat(*subpaths,
2147 96 : list_copy_tail(apath->subpaths,
2148 : apath->first_partial_path));
2149 96 : new_special_subpaths = list_copy_head(apath->subpaths,
2150 : apath->first_partial_path);
2151 96 : *special_subpaths = list_concat(*special_subpaths,
2152 : new_special_subpaths);
2153 96 : return;
2154 : }
2155 : }
2156 195916 : else if (IsA(path, MergeAppendPath))
2157 : {
2158 696 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2159 :
2160 696 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2161 696 : return;
2162 : }
2163 :
2164 195316 : *subpaths = lappend(*subpaths, path);
2165 : }
2166 :
2167 : /*
2168 : * get_singleton_append_subpath
2169 : * Returns the single subpath of an Append/MergeAppend, or just
2170 : * return 'path' if it's not a single sub-path Append/MergeAppend.
2171 : *
2172 : * Note: 'path' must not be a parallel-aware path.
2173 : */
2174 : static Path *
2175 13128 : get_singleton_append_subpath(Path *path)
2176 : {
2177 : Assert(!path->parallel_aware);
2178 :
2179 13128 : if (IsA(path, AppendPath))
2180 : {
2181 440 : AppendPath *apath = (AppendPath *) path;
2182 :
2183 440 : if (list_length(apath->subpaths) == 1)
2184 204 : return (Path *) linitial(apath->subpaths);
2185 : }
2186 12688 : else if (IsA(path, MergeAppendPath))
2187 : {
2188 348 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2189 :
2190 348 : if (list_length(mpath->subpaths) == 1)
2191 0 : return (Path *) linitial(mpath->subpaths);
2192 : }
2193 :
2194 12924 : return path;
2195 : }
2196 :
2197 : /*
2198 : * set_dummy_rel_pathlist
2199 : * Build a dummy path for a relation that's been excluded by constraints
2200 : *
2201 : * Rather than inventing a special "dummy" path type, we represent this as an
2202 : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2203 : *
2204 : * (See also mark_dummy_rel, which does basically the same thing, but is
2205 : * typically used to change a rel into dummy state after we already made
2206 : * paths for it.)
2207 : */
2208 : static void
2209 1178 : set_dummy_rel_pathlist(RelOptInfo *rel)
2210 : {
2211 : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2212 1178 : rel->rows = 0;
2213 1178 : rel->reltarget->width = 0;
2214 :
2215 : /* Discard any pre-existing paths; no further need for them */
2216 1178 : rel->pathlist = NIL;
2217 1178 : rel->partial_pathlist = NIL;
2218 :
2219 : /* Set up the dummy path */
2220 1178 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2221 : NIL, rel->lateral_relids,
2222 : 0, false, -1));
2223 :
2224 : /*
2225 : * We set the cheapest-path fields immediately, just in case they were
2226 : * pointing at some discarded path. This is redundant in current usage
2227 : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2228 : * for safety and consistency with mark_dummy_rel.
2229 : */
2230 1178 : set_cheapest(rel);
2231 1178 : }
2232 :
2233 : /*
2234 : * find_window_run_conditions
2235 : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2236 : * function to determine the function's monotonic properties. We then
2237 : * see if 'opexpr' can be used to short-circuit execution.
2238 : *
2239 : * For example row_number() over (order by ...) always produces a value one
2240 : * higher than the previous. If someone has a window function in a subquery
2241 : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2242 : * as well stop processing the windowagg once the row number reaches 11. Here
2243 : * we check if 'opexpr' might help us to stop doing needless extra processing
2244 : * in WindowAgg nodes.
2245 : *
2246 : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2247 : * its original purpose. This is set to false if the caller can assume that
2248 : * the run condition will handle all of the required filtering.
2249 : *
2250 : * Returns true if 'opexpr' was found to be useful and was added to the
2251 : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2252 : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2253 : * If the 'opexpr' cannot be used then we set *keep_original to true and
2254 : * return false.
2255 : */
2256 : static bool
2257 240 : find_window_run_conditions(Query *subquery, RangeTblEntry *rte, Index rti,
2258 : AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2259 : bool wfunc_left, bool *keep_original,
2260 : Bitmapset **run_cond_attrs)
2261 : {
2262 : Oid prosupport;
2263 : Expr *otherexpr;
2264 : SupportRequestWFuncMonotonic req;
2265 : SupportRequestWFuncMonotonic *res;
2266 : WindowClause *wclause;
2267 : List *opinfos;
2268 : OpExpr *runopexpr;
2269 : Oid runoperator;
2270 : ListCell *lc;
2271 :
2272 240 : *keep_original = true;
2273 :
2274 240 : while (IsA(wfunc, RelabelType))
2275 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2276 :
2277 : /* we can only work with window functions */
2278 240 : if (!IsA(wfunc, WindowFunc))
2279 24 : return false;
2280 :
2281 : /* can't use it if there are subplans in the WindowFunc */
2282 216 : if (contain_subplans((Node *) wfunc))
2283 6 : return false;
2284 :
2285 210 : prosupport = get_func_support(wfunc->winfnoid);
2286 :
2287 : /* Check if there's a support function for 'wfunc' */
2288 210 : if (!OidIsValid(prosupport))
2289 18 : return false;
2290 :
2291 : /* get the Expr from the other side of the OpExpr */
2292 192 : if (wfunc_left)
2293 168 : otherexpr = lsecond(opexpr->args);
2294 : else
2295 24 : otherexpr = linitial(opexpr->args);
2296 :
2297 : /*
2298 : * The value being compared must not change during the evaluation of the
2299 : * window partition.
2300 : */
2301 192 : if (!is_pseudo_constant_clause((Node *) otherexpr))
2302 0 : return false;
2303 :
2304 : /* find the window clause belonging to the window function */
2305 192 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2306 192 : wfunc->winref - 1);
2307 :
2308 192 : req.type = T_SupportRequestWFuncMonotonic;
2309 192 : req.window_func = wfunc;
2310 192 : req.window_clause = wclause;
2311 :
2312 : /* call the support function */
2313 : res = (SupportRequestWFuncMonotonic *)
2314 192 : DatumGetPointer(OidFunctionCall1(prosupport,
2315 : PointerGetDatum(&req)));
2316 :
2317 : /*
2318 : * Nothing to do if the function is neither monotonically increasing nor
2319 : * monotonically decreasing.
2320 : */
2321 192 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2322 0 : return false;
2323 :
2324 192 : runopexpr = NULL;
2325 192 : runoperator = InvalidOid;
2326 192 : opinfos = get_op_index_interpretation(opexpr->opno);
2327 :
2328 192 : foreach(lc, opinfos)
2329 : {
2330 192 : OpIndexInterpretation *opinfo = (OpIndexInterpretation *) lfirst(lc);
2331 192 : CompareType cmptype = opinfo->cmptype;
2332 :
2333 : /* handle < / <= */
2334 192 : if (cmptype == COMPARE_LT || cmptype == COMPARE_LE)
2335 : {
2336 : /*
2337 : * < / <= is supported for monotonically increasing functions in
2338 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2339 : * for monotonically decreasing functions.
2340 : */
2341 138 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2342 18 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2343 : {
2344 126 : *keep_original = false;
2345 126 : runopexpr = opexpr;
2346 126 : runoperator = opexpr->opno;
2347 : }
2348 138 : break;
2349 : }
2350 : /* handle > / >= */
2351 54 : else if (cmptype == COMPARE_GT || cmptype == COMPARE_GE)
2352 : {
2353 : /*
2354 : * > / >= is supported for monotonically decreasing functions in
2355 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2356 : * for monotonically increasing functions.
2357 : */
2358 18 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2359 12 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2360 : {
2361 18 : *keep_original = false;
2362 18 : runopexpr = opexpr;
2363 18 : runoperator = opexpr->opno;
2364 : }
2365 18 : break;
2366 : }
2367 : /* handle = */
2368 36 : else if (cmptype == COMPARE_EQ)
2369 : {
2370 : CompareType newcmptype;
2371 :
2372 : /*
2373 : * When both monotonically increasing and decreasing then the
2374 : * return value of the window function will be the same each time.
2375 : * We can simply use 'opexpr' as the run condition without
2376 : * modifying it.
2377 : */
2378 36 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2379 : {
2380 6 : *keep_original = false;
2381 6 : runopexpr = opexpr;
2382 6 : runoperator = opexpr->opno;
2383 6 : break;
2384 : }
2385 :
2386 : /*
2387 : * When monotonically increasing we make a qual with <wfunc> <=
2388 : * <value> or <value> >= <wfunc> in order to filter out values
2389 : * which are above the value in the equality condition. For
2390 : * monotonically decreasing functions we want to filter values
2391 : * below the value in the equality condition.
2392 : */
2393 30 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2394 30 : newcmptype = wfunc_left ? COMPARE_LE : COMPARE_GE;
2395 : else
2396 0 : newcmptype = wfunc_left ? COMPARE_GE : COMPARE_LE;
2397 :
2398 : /* We must keep the original equality qual */
2399 30 : *keep_original = true;
2400 30 : runopexpr = opexpr;
2401 :
2402 : /* determine the operator to use for the WindowFuncRunCondition */
2403 30 : runoperator = get_opfamily_member_for_cmptype(opinfo->opfamily_id,
2404 : opinfo->oplefttype,
2405 : opinfo->oprighttype,
2406 : newcmptype);
2407 30 : break;
2408 : }
2409 : }
2410 :
2411 192 : if (runopexpr != NULL)
2412 : {
2413 : WindowFuncRunCondition *wfuncrc;
2414 :
2415 180 : wfuncrc = makeNode(WindowFuncRunCondition);
2416 180 : wfuncrc->opno = runoperator;
2417 180 : wfuncrc->inputcollid = runopexpr->inputcollid;
2418 180 : wfuncrc->wfunc_left = wfunc_left;
2419 180 : wfuncrc->arg = copyObject(otherexpr);
2420 :
2421 180 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2422 :
2423 : /* record that this attno was used in a run condition */
2424 180 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2425 : attno - FirstLowInvalidHeapAttributeNumber);
2426 180 : return true;
2427 : }
2428 :
2429 : /* unsupported OpExpr */
2430 12 : return false;
2431 : }
2432 :
2433 : /*
2434 : * check_and_push_window_quals
2435 : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2436 : * as a 'runCondition' qual. These, when present, allow some unnecessary
2437 : * work to be skipped during execution.
2438 : *
2439 : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2440 : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2441 : * window quals for.
2442 : *
2443 : * Returns true if the caller still must keep the original qual or false if
2444 : * the caller can safely ignore the original qual because the WindowAgg node
2445 : * will use the runCondition to stop returning tuples.
2446 : */
2447 : static bool
2448 252 : check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti,
2449 : Node *clause, Bitmapset **run_cond_attrs)
2450 : {
2451 252 : OpExpr *opexpr = (OpExpr *) clause;
2452 252 : bool keep_original = true;
2453 : Var *var1;
2454 : Var *var2;
2455 :
2456 : /* We're only able to use OpExprs with 2 operands */
2457 252 : if (!IsA(opexpr, OpExpr))
2458 18 : return true;
2459 :
2460 234 : if (list_length(opexpr->args) != 2)
2461 0 : return true;
2462 :
2463 : /*
2464 : * Currently, we restrict this optimization to strict OpExprs. The reason
2465 : * for this is that during execution, once the runcondition becomes false,
2466 : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2467 : * function result values, we set them to NULL. Having only strict
2468 : * OpExprs here ensures that we properly filter out the tuples with NULLs
2469 : * in the top-level WindowAgg.
2470 : */
2471 234 : set_opfuncid(opexpr);
2472 234 : if (!func_strict(opexpr->opfuncid))
2473 0 : return true;
2474 :
2475 : /*
2476 : * Check for plain Vars that reference window functions in the subquery.
2477 : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2478 : * be used as part of the run condition.
2479 : */
2480 :
2481 : /* Check the left side of the OpExpr */
2482 234 : var1 = linitial(opexpr->args);
2483 234 : if (IsA(var1, Var) && var1->varattno > 0)
2484 : {
2485 198 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2486 198 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2487 :
2488 198 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2489 : opexpr, true, &keep_original,
2490 : run_cond_attrs))
2491 162 : return keep_original;
2492 : }
2493 :
2494 : /* and check the right side */
2495 72 : var2 = lsecond(opexpr->args);
2496 72 : if (IsA(var2, Var) && var2->varattno > 0)
2497 : {
2498 42 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2499 42 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2500 :
2501 42 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2502 : opexpr, false, &keep_original,
2503 : run_cond_attrs))
2504 18 : return keep_original;
2505 : }
2506 :
2507 54 : return true;
2508 : }
2509 :
2510 : /*
2511 : * set_subquery_pathlist
2512 : * Generate SubqueryScan access paths for a subquery RTE
2513 : *
2514 : * We don't currently support generating parameterized paths for subqueries
2515 : * by pushing join clauses down into them; it seems too expensive to re-plan
2516 : * the subquery multiple times to consider different alternatives.
2517 : * (XXX that could stand to be reconsidered, now that we use Paths.)
2518 : * So the paths made here will be parameterized if the subquery contains
2519 : * LATERAL references, otherwise not. As long as that's true, there's no need
2520 : * for a separate set_subquery_size phase: just make the paths right away.
2521 : */
2522 : static void
2523 11588 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2524 : Index rti, RangeTblEntry *rte)
2525 : {
2526 11588 : Query *parse = root->parse;
2527 11588 : Query *subquery = rte->subquery;
2528 : bool trivial_pathtarget;
2529 : Relids required_outer;
2530 : pushdown_safety_info safetyInfo;
2531 : double tuple_fraction;
2532 : RelOptInfo *sub_final_rel;
2533 11588 : Bitmapset *run_cond_attrs = NULL;
2534 : ListCell *lc;
2535 :
2536 : /*
2537 : * Must copy the Query so that planning doesn't mess up the RTE contents
2538 : * (really really need to fix the planner to not scribble on its input,
2539 : * someday ... but see remove_unused_subquery_outputs to start with).
2540 : */
2541 11588 : subquery = copyObject(subquery);
2542 :
2543 : /*
2544 : * If it's a LATERAL subquery, it might contain some Vars of the current
2545 : * query level, requiring it to be treated as parameterized, even though
2546 : * we don't support pushing down join quals into subqueries.
2547 : */
2548 11588 : required_outer = rel->lateral_relids;
2549 :
2550 : /*
2551 : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2552 : * flags as needed while recursing. In particular, we need a workspace
2553 : * for keeping track of the reasons why columns are unsafe to reference.
2554 : * These reasons are stored in the bits inside unsafeFlags[i] when we
2555 : * discover reasons that column i of the subquery is unsafe to be used in
2556 : * a pushed-down qual.
2557 : */
2558 11588 : memset(&safetyInfo, 0, sizeof(safetyInfo));
2559 11588 : safetyInfo.unsafeFlags = (unsigned char *)
2560 11588 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2561 :
2562 : /*
2563 : * If the subquery has the "security_barrier" flag, it means the subquery
2564 : * originated from a view that must enforce row-level security. Then we
2565 : * must not push down quals that contain leaky functions. (Ideally this
2566 : * would be checked inside subquery_is_pushdown_safe, but since we don't
2567 : * currently pass the RTE to that function, we must do it here.)
2568 : */
2569 11588 : safetyInfo.unsafeLeaky = rte->security_barrier;
2570 :
2571 : /*
2572 : * If there are any restriction clauses that have been attached to the
2573 : * subquery relation, consider pushing them down to become WHERE or HAVING
2574 : * quals of the subquery itself. This transformation is useful because it
2575 : * may allow us to generate a better plan for the subquery than evaluating
2576 : * all the subquery output rows and then filtering them.
2577 : *
2578 : * There are several cases where we cannot push down clauses. Restrictions
2579 : * involving the subquery are checked by subquery_is_pushdown_safe().
2580 : * Restrictions on individual clauses are checked by
2581 : * qual_is_pushdown_safe(). Also, we don't want to push down
2582 : * pseudoconstant clauses; better to have the gating node above the
2583 : * subquery.
2584 : *
2585 : * Non-pushed-down clauses will get evaluated as qpquals of the
2586 : * SubqueryScan node.
2587 : *
2588 : * XXX Are there any cases where we want to make a policy decision not to
2589 : * push down a pushable qual, because it'd result in a worse plan?
2590 : */
2591 13426 : if (rel->baserestrictinfo != NIL &&
2592 1838 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2593 : {
2594 : /* OK to consider pushing down individual quals */
2595 1692 : List *upperrestrictlist = NIL;
2596 : ListCell *l;
2597 :
2598 4330 : foreach(l, rel->baserestrictinfo)
2599 : {
2600 2638 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2601 2638 : Node *clause = (Node *) rinfo->clause;
2602 :
2603 2638 : if (rinfo->pseudoconstant)
2604 : {
2605 4 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2606 4 : continue;
2607 : }
2608 :
2609 2634 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2610 : {
2611 1940 : case PUSHDOWN_SAFE:
2612 : /* Push it down */
2613 1940 : subquery_push_qual(subquery, rte, rti, clause);
2614 1940 : break;
2615 :
2616 252 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2617 :
2618 : /*
2619 : * Since we can't push the qual down into the subquery,
2620 : * check if it happens to reference a window function. If
2621 : * so then it might be useful to use for the WindowAgg's
2622 : * runCondition.
2623 : */
2624 504 : if (!subquery->hasWindowFuncs ||
2625 252 : check_and_push_window_quals(subquery, rte, rti, clause,
2626 : &run_cond_attrs))
2627 : {
2628 : /*
2629 : * subquery has no window funcs or the clause is not a
2630 : * suitable window run condition qual or it is, but
2631 : * the original must also be kept in the upper query.
2632 : */
2633 102 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2634 : }
2635 252 : break;
2636 :
2637 442 : case PUSHDOWN_UNSAFE:
2638 442 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2639 442 : break;
2640 : }
2641 : }
2642 1692 : rel->baserestrictinfo = upperrestrictlist;
2643 : /* We don't bother recomputing baserestrict_min_security */
2644 : }
2645 :
2646 11588 : pfree(safetyInfo.unsafeFlags);
2647 :
2648 : /*
2649 : * The upper query might not use all the subquery's output columns; if
2650 : * not, we can simplify. Pass the attributes that were pushed down into
2651 : * WindowAgg run conditions to ensure we don't accidentally think those
2652 : * are unused.
2653 : */
2654 11588 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2655 :
2656 : /*
2657 : * We can safely pass the outer tuple_fraction down to the subquery if the
2658 : * outer level has no joining, aggregation, or sorting to do. Otherwise
2659 : * we'd better tell the subquery to plan for full retrieval. (XXX This
2660 : * could probably be made more intelligent ...)
2661 : */
2662 11588 : if (parse->hasAggs ||
2663 10420 : parse->groupClause ||
2664 10402 : parse->groupingSets ||
2665 10402 : root->hasHavingQual ||
2666 10402 : parse->distinctClause ||
2667 15186 : parse->sortClause ||
2668 5282 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
2669 7514 : tuple_fraction = 0.0; /* default case */
2670 : else
2671 4074 : tuple_fraction = root->tuple_fraction;
2672 :
2673 : /* plan_params should not be in use in current query level */
2674 : Assert(root->plan_params == NIL);
2675 :
2676 : /* Generate a subroot and Paths for the subquery */
2677 11588 : rel->subroot = subquery_planner(root->glob, subquery, root, false,
2678 : tuple_fraction, NULL);
2679 :
2680 : /* Isolate the params needed by this specific subplan */
2681 11588 : rel->subplan_params = root->plan_params;
2682 11588 : root->plan_params = NIL;
2683 :
2684 : /*
2685 : * It's possible that constraint exclusion proved the subquery empty. If
2686 : * so, it's desirable to produce an unadorned dummy path so that we will
2687 : * recognize appropriate optimizations at this query level.
2688 : */
2689 11588 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2690 :
2691 11588 : if (IS_DUMMY_REL(sub_final_rel))
2692 : {
2693 120 : set_dummy_rel_pathlist(rel);
2694 120 : return;
2695 : }
2696 :
2697 : /*
2698 : * Mark rel with estimated output rows, width, etc. Note that we have to
2699 : * do this before generating outer-query paths, else cost_subqueryscan is
2700 : * not happy.
2701 : */
2702 11468 : set_subquery_size_estimates(root, rel);
2703 :
2704 : /*
2705 : * Also detect whether the reltarget is trivial, so that we can pass that
2706 : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2707 : * It's trivial if it fetches all the subplan output columns in order.
2708 : */
2709 11468 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2710 2688 : trivial_pathtarget = false;
2711 : else
2712 : {
2713 8780 : trivial_pathtarget = true;
2714 27778 : foreach(lc, rel->reltarget->exprs)
2715 : {
2716 19296 : Node *node = (Node *) lfirst(lc);
2717 : Var *var;
2718 :
2719 19296 : if (!IsA(node, Var))
2720 : {
2721 0 : trivial_pathtarget = false;
2722 0 : break;
2723 : }
2724 19296 : var = (Var *) node;
2725 19296 : if (var->varno != rti ||
2726 19296 : var->varattno != foreach_current_index(lc) + 1)
2727 : {
2728 298 : trivial_pathtarget = false;
2729 298 : break;
2730 : }
2731 : }
2732 : }
2733 :
2734 : /*
2735 : * For each Path that subquery_planner produced, make a SubqueryScanPath
2736 : * in the outer query.
2737 : */
2738 23716 : foreach(lc, sub_final_rel->pathlist)
2739 : {
2740 12248 : Path *subpath = (Path *) lfirst(lc);
2741 : List *pathkeys;
2742 :
2743 : /* Convert subpath's pathkeys to outer representation */
2744 12248 : pathkeys = convert_subquery_pathkeys(root,
2745 : rel,
2746 : subpath->pathkeys,
2747 : make_tlist_from_pathtarget(subpath->pathtarget));
2748 :
2749 : /* Generate outer path using this subpath */
2750 12248 : add_path(rel, (Path *)
2751 12248 : create_subqueryscan_path(root, rel, subpath,
2752 : trivial_pathtarget,
2753 : pathkeys, required_outer));
2754 : }
2755 :
2756 : /* If outer rel allows parallelism, do same for partial paths. */
2757 11468 : if (rel->consider_parallel && bms_is_empty(required_outer))
2758 : {
2759 : /* If consider_parallel is false, there should be no partial paths. */
2760 : Assert(sub_final_rel->consider_parallel ||
2761 : sub_final_rel->partial_pathlist == NIL);
2762 :
2763 : /* Same for partial paths. */
2764 7440 : foreach(lc, sub_final_rel->partial_pathlist)
2765 : {
2766 42 : Path *subpath = (Path *) lfirst(lc);
2767 : List *pathkeys;
2768 :
2769 : /* Convert subpath's pathkeys to outer representation */
2770 42 : pathkeys = convert_subquery_pathkeys(root,
2771 : rel,
2772 : subpath->pathkeys,
2773 : make_tlist_from_pathtarget(subpath->pathtarget));
2774 :
2775 : /* Generate outer path using this subpath */
2776 42 : add_partial_path(rel, (Path *)
2777 42 : create_subqueryscan_path(root, rel, subpath,
2778 : trivial_pathtarget,
2779 : pathkeys,
2780 : required_outer));
2781 : }
2782 : }
2783 : }
2784 :
2785 : /*
2786 : * set_function_pathlist
2787 : * Build the (single) access path for a function RTE
2788 : */
2789 : static void
2790 52316 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2791 : {
2792 : Relids required_outer;
2793 52316 : List *pathkeys = NIL;
2794 :
2795 : /*
2796 : * We don't support pushing join clauses into the quals of a function
2797 : * scan, but it could still have required parameterization due to LATERAL
2798 : * refs in the function expression.
2799 : */
2800 52316 : required_outer = rel->lateral_relids;
2801 :
2802 : /*
2803 : * The result is considered unordered unless ORDINALITY was used, in which
2804 : * case it is ordered by the ordinal column (the last one). See if we
2805 : * care, by checking for uses of that Var in equivalence classes.
2806 : */
2807 52316 : if (rte->funcordinality)
2808 : {
2809 914 : AttrNumber ordattno = rel->max_attr;
2810 914 : Var *var = NULL;
2811 : ListCell *lc;
2812 :
2813 : /*
2814 : * Is there a Var for it in rel's targetlist? If not, the query did
2815 : * not reference the ordinality column, or at least not in any way
2816 : * that would be interesting for sorting.
2817 : */
2818 2098 : foreach(lc, rel->reltarget->exprs)
2819 : {
2820 2092 : Var *node = (Var *) lfirst(lc);
2821 :
2822 : /* checking varno/varlevelsup is just paranoia */
2823 2092 : if (IsA(node, Var) &&
2824 2092 : node->varattno == ordattno &&
2825 908 : node->varno == rel->relid &&
2826 908 : node->varlevelsup == 0)
2827 : {
2828 908 : var = node;
2829 908 : break;
2830 : }
2831 : }
2832 :
2833 : /*
2834 : * Try to build pathkeys for this Var with int8 sorting. We tell
2835 : * build_expression_pathkey not to build any new equivalence class; if
2836 : * the Var isn't already mentioned in some EC, it means that nothing
2837 : * cares about the ordering.
2838 : */
2839 914 : if (var)
2840 908 : pathkeys = build_expression_pathkey(root,
2841 : (Expr *) var,
2842 : Int8LessOperator,
2843 : rel->relids,
2844 : false);
2845 : }
2846 :
2847 : /* Generate appropriate path */
2848 52316 : add_path(rel, create_functionscan_path(root, rel,
2849 : pathkeys, required_outer));
2850 52316 : }
2851 :
2852 : /*
2853 : * set_values_pathlist
2854 : * Build the (single) access path for a VALUES RTE
2855 : */
2856 : static void
2857 8240 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2858 : {
2859 : Relids required_outer;
2860 :
2861 : /*
2862 : * We don't support pushing join clauses into the quals of a values scan,
2863 : * but it could still have required parameterization due to LATERAL refs
2864 : * in the values expressions.
2865 : */
2866 8240 : required_outer = rel->lateral_relids;
2867 :
2868 : /* Generate appropriate path */
2869 8240 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
2870 8240 : }
2871 :
2872 : /*
2873 : * set_tablefunc_pathlist
2874 : * Build the (single) access path for a table func RTE
2875 : */
2876 : static void
2877 626 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2878 : {
2879 : Relids required_outer;
2880 :
2881 : /*
2882 : * We don't support pushing join clauses into the quals of a tablefunc
2883 : * scan, but it could still have required parameterization due to LATERAL
2884 : * refs in the function expression.
2885 : */
2886 626 : required_outer = rel->lateral_relids;
2887 :
2888 : /* Generate appropriate path */
2889 626 : add_path(rel, create_tablefuncscan_path(root, rel,
2890 : required_outer));
2891 626 : }
2892 :
2893 : /*
2894 : * set_cte_pathlist
2895 : * Build the (single) access path for a non-self-reference CTE RTE
2896 : *
2897 : * There's no need for a separate set_cte_size phase, since we don't
2898 : * support join-qual-parameterized paths for CTEs.
2899 : */
2900 : static void
2901 4086 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2902 : {
2903 : Path *ctepath;
2904 : Plan *cteplan;
2905 : PlannerInfo *cteroot;
2906 : Index levelsup;
2907 : List *pathkeys;
2908 : int ndx;
2909 : ListCell *lc;
2910 : int plan_id;
2911 : Relids required_outer;
2912 :
2913 : /*
2914 : * Find the referenced CTE, and locate the path and plan previously made
2915 : * for it.
2916 : */
2917 4086 : levelsup = rte->ctelevelsup;
2918 4086 : cteroot = root;
2919 7090 : while (levelsup-- > 0)
2920 : {
2921 3004 : cteroot = cteroot->parent_root;
2922 3004 : if (!cteroot) /* shouldn't happen */
2923 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2924 : }
2925 :
2926 : /*
2927 : * Note: cte_plan_ids can be shorter than cteList, if we are still working
2928 : * on planning the CTEs (ie, this is a side-reference from another CTE).
2929 : * So we mustn't use forboth here.
2930 : */
2931 4086 : ndx = 0;
2932 5686 : foreach(lc, cteroot->parse->cteList)
2933 : {
2934 5686 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
2935 :
2936 5686 : if (strcmp(cte->ctename, rte->ctename) == 0)
2937 4086 : break;
2938 1600 : ndx++;
2939 : }
2940 4086 : if (lc == NULL) /* shouldn't happen */
2941 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
2942 4086 : if (ndx >= list_length(cteroot->cte_plan_ids))
2943 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
2944 4086 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
2945 4086 : if (plan_id <= 0)
2946 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
2947 :
2948 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
2949 4086 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
2950 4086 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2951 :
2952 : /* Mark rel with estimated output rows, width, etc */
2953 4086 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
2954 :
2955 : /* Convert the ctepath's pathkeys to outer query's representation */
2956 4086 : pathkeys = convert_subquery_pathkeys(root,
2957 : rel,
2958 : ctepath->pathkeys,
2959 : cteplan->targetlist);
2960 :
2961 : /*
2962 : * We don't support pushing join clauses into the quals of a CTE scan, but
2963 : * it could still have required parameterization due to LATERAL refs in
2964 : * its tlist.
2965 : */
2966 4086 : required_outer = rel->lateral_relids;
2967 :
2968 : /* Generate appropriate path */
2969 4086 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
2970 4086 : }
2971 :
2972 : /*
2973 : * set_namedtuplestore_pathlist
2974 : * Build the (single) access path for a named tuplestore RTE
2975 : *
2976 : * There's no need for a separate set_namedtuplestore_size phase, since we
2977 : * don't support join-qual-parameterized paths for tuplestores.
2978 : */
2979 : static void
2980 466 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
2981 : RangeTblEntry *rte)
2982 : {
2983 : Relids required_outer;
2984 :
2985 : /* Mark rel with estimated output rows, width, etc */
2986 466 : set_namedtuplestore_size_estimates(root, rel);
2987 :
2988 : /*
2989 : * We don't support pushing join clauses into the quals of a tuplestore
2990 : * scan, but it could still have required parameterization due to LATERAL
2991 : * refs in its tlist.
2992 : */
2993 466 : required_outer = rel->lateral_relids;
2994 :
2995 : /* Generate appropriate path */
2996 466 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2997 466 : }
2998 :
2999 : /*
3000 : * set_result_pathlist
3001 : * Build the (single) access path for an RTE_RESULT RTE
3002 : *
3003 : * There's no need for a separate set_result_size phase, since we
3004 : * don't support join-qual-parameterized paths for these RTEs.
3005 : */
3006 : static void
3007 4220 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
3008 : RangeTblEntry *rte)
3009 : {
3010 : Relids required_outer;
3011 :
3012 : /* Mark rel with estimated output rows, width, etc */
3013 4220 : set_result_size_estimates(root, rel);
3014 :
3015 : /*
3016 : * We don't support pushing join clauses into the quals of a Result scan,
3017 : * but it could still have required parameterization due to LATERAL refs
3018 : * in its tlist.
3019 : */
3020 4220 : required_outer = rel->lateral_relids;
3021 :
3022 : /* Generate appropriate path */
3023 4220 : add_path(rel, create_resultscan_path(root, rel, required_outer));
3024 4220 : }
3025 :
3026 : /*
3027 : * set_worktable_pathlist
3028 : * Build the (single) access path for a self-reference CTE RTE
3029 : *
3030 : * There's no need for a separate set_worktable_size phase, since we don't
3031 : * support join-qual-parameterized paths for CTEs.
3032 : */
3033 : static void
3034 1006 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3035 : {
3036 : Path *ctepath;
3037 : PlannerInfo *cteroot;
3038 : Index levelsup;
3039 : Relids required_outer;
3040 :
3041 : /*
3042 : * We need to find the non-recursive term's path, which is in the plan
3043 : * level that's processing the recursive UNION, which is one level *below*
3044 : * where the CTE comes from.
3045 : */
3046 1006 : levelsup = rte->ctelevelsup;
3047 1006 : if (levelsup == 0) /* shouldn't happen */
3048 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3049 1006 : levelsup--;
3050 1006 : cteroot = root;
3051 2368 : while (levelsup-- > 0)
3052 : {
3053 1362 : cteroot = cteroot->parent_root;
3054 1362 : if (!cteroot) /* shouldn't happen */
3055 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3056 : }
3057 1006 : ctepath = cteroot->non_recursive_path;
3058 1006 : if (!ctepath) /* shouldn't happen */
3059 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3060 :
3061 : /* Mark rel with estimated output rows, width, etc */
3062 1006 : set_cte_size_estimates(root, rel, ctepath->rows);
3063 :
3064 : /*
3065 : * We don't support pushing join clauses into the quals of a worktable
3066 : * scan, but it could still have required parameterization due to LATERAL
3067 : * refs in its tlist. (I'm not sure this is actually possible given the
3068 : * restrictions on recursive references, but it's easy enough to support.)
3069 : */
3070 1006 : required_outer = rel->lateral_relids;
3071 :
3072 : /* Generate appropriate path */
3073 1006 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
3074 1006 : }
3075 :
3076 : /*
3077 : * generate_gather_paths
3078 : * Generate parallel access paths for a relation by pushing a Gather or
3079 : * Gather Merge on top of a partial path.
3080 : *
3081 : * This must not be called until after we're done creating all partial paths
3082 : * for the specified relation. (Otherwise, add_partial_path might delete a
3083 : * path that some GatherPath or GatherMergePath has a reference to.)
3084 : *
3085 : * If we're generating paths for a scan or join relation, override_rows will
3086 : * be false, and we'll just use the relation's size estimate. When we're
3087 : * being called for a partially-grouped or partially-distinct path, though, we
3088 : * need to override the rowcount estimate. (It's not clear that the
3089 : * particular value we're using here is actually best, but the underlying rel
3090 : * has no estimate so we must do something.)
3091 : */
3092 : void
3093 19106 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3094 : {
3095 : Path *cheapest_partial_path;
3096 : Path *simple_gather_path;
3097 : ListCell *lc;
3098 : double rows;
3099 19106 : double *rowsp = NULL;
3100 :
3101 : /* If there are no partial paths, there's nothing to do here. */
3102 19106 : if (rel->partial_pathlist == NIL)
3103 0 : return;
3104 :
3105 : /* Should we override the rel's rowcount estimate? */
3106 19106 : if (override_rows)
3107 1754 : rowsp = &rows;
3108 :
3109 : /*
3110 : * The output of Gather is always unsorted, so there's only one partial
3111 : * path of interest: the cheapest one. That will be the one at the front
3112 : * of partial_pathlist because of the way add_partial_path works.
3113 : */
3114 19106 : cheapest_partial_path = linitial(rel->partial_pathlist);
3115 19106 : rows = compute_gather_rows(cheapest_partial_path);
3116 : simple_gather_path = (Path *)
3117 19106 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3118 : NULL, rowsp);
3119 19106 : add_path(rel, simple_gather_path);
3120 :
3121 : /*
3122 : * For each useful ordering, we can consider an order-preserving Gather
3123 : * Merge.
3124 : */
3125 39534 : foreach(lc, rel->partial_pathlist)
3126 : {
3127 20428 : Path *subpath = (Path *) lfirst(lc);
3128 : GatherMergePath *path;
3129 :
3130 20428 : if (subpath->pathkeys == NIL)
3131 18712 : continue;
3132 :
3133 1716 : rows = compute_gather_rows(subpath);
3134 1716 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3135 : subpath->pathkeys, NULL, rowsp);
3136 1716 : add_path(rel, &path->path);
3137 : }
3138 : }
3139 :
3140 : /*
3141 : * get_useful_pathkeys_for_relation
3142 : * Determine which orderings of a relation might be useful.
3143 : *
3144 : * Getting data in sorted order can be useful either because the requested
3145 : * order matches the final output ordering for the overall query we're
3146 : * planning, or because it enables an efficient merge join. Here, we try
3147 : * to figure out which pathkeys to consider.
3148 : *
3149 : * This allows us to do incremental sort on top of an index scan under a gather
3150 : * merge node, i.e. parallelized.
3151 : *
3152 : * If the require_parallel_safe is true, we also require the expressions to
3153 : * be parallel safe (which allows pushing the sort below Gather Merge).
3154 : *
3155 : * XXX At the moment this can only ever return a list with a single element,
3156 : * because it looks at query_pathkeys only. So we might return the pathkeys
3157 : * directly, but it seems plausible we'll want to consider other orderings
3158 : * in the future. For example, we might want to consider pathkeys useful for
3159 : * merge joins.
3160 : */
3161 : static List *
3162 19106 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3163 : bool require_parallel_safe)
3164 : {
3165 19106 : List *useful_pathkeys_list = NIL;
3166 :
3167 : /*
3168 : * Considering query_pathkeys is always worth it, because it might allow
3169 : * us to avoid a total sort when we have a partially presorted path
3170 : * available or to push the total sort into the parallel portion of the
3171 : * query.
3172 : */
3173 19106 : if (root->query_pathkeys)
3174 : {
3175 : ListCell *lc;
3176 9666 : int npathkeys = 0; /* useful pathkeys */
3177 :
3178 17870 : foreach(lc, root->query_pathkeys)
3179 : {
3180 12810 : PathKey *pathkey = (PathKey *) lfirst(lc);
3181 12810 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3182 :
3183 : /*
3184 : * We can only build a sort for pathkeys that contain a
3185 : * safe-to-compute-early EC member computable from the current
3186 : * relation's reltarget, so ignore the remainder of the list as
3187 : * soon as we find a pathkey without such a member.
3188 : *
3189 : * It's still worthwhile to return any prefix of the pathkeys list
3190 : * that meets this requirement, as we may be able to do an
3191 : * incremental sort.
3192 : *
3193 : * If requested, ensure the sort expression is parallel-safe too.
3194 : */
3195 12810 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3196 : require_parallel_safe))
3197 4606 : break;
3198 :
3199 8204 : npathkeys++;
3200 : }
3201 :
3202 : /*
3203 : * The whole query_pathkeys list matches, so append it directly, to
3204 : * allow comparing pathkeys easily by comparing list pointer. If we
3205 : * have to truncate the pathkeys, we gotta do a copy though.
3206 : */
3207 9666 : if (npathkeys == list_length(root->query_pathkeys))
3208 5060 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3209 5060 : root->query_pathkeys);
3210 4606 : else if (npathkeys > 0)
3211 450 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3212 450 : list_copy_head(root->query_pathkeys,
3213 : npathkeys));
3214 : }
3215 :
3216 19106 : return useful_pathkeys_list;
3217 : }
3218 :
3219 : /*
3220 : * generate_useful_gather_paths
3221 : * Generate parallel access paths for a relation by pushing a Gather or
3222 : * Gather Merge on top of a partial path.
3223 : *
3224 : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3225 : * paths (aiming to preserve the ordering), but also considers ordering that
3226 : * might be useful for nodes above the gather merge node, and tries to add
3227 : * a sort (regular or incremental) to provide that.
3228 : */
3229 : void
3230 602604 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3231 : {
3232 : ListCell *lc;
3233 : double rows;
3234 602604 : double *rowsp = NULL;
3235 602604 : List *useful_pathkeys_list = NIL;
3236 602604 : Path *cheapest_partial_path = NULL;
3237 :
3238 : /* If there are no partial paths, there's nothing to do here. */
3239 602604 : if (rel->partial_pathlist == NIL)
3240 583498 : return;
3241 :
3242 : /* Should we override the rel's rowcount estimate? */
3243 19106 : if (override_rows)
3244 1754 : rowsp = &rows;
3245 :
3246 : /* generate the regular gather (merge) paths */
3247 19106 : generate_gather_paths(root, rel, override_rows);
3248 :
3249 : /* consider incremental sort for interesting orderings */
3250 19106 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3251 :
3252 : /* used for explicit (full) sort paths */
3253 19106 : cheapest_partial_path = linitial(rel->partial_pathlist);
3254 :
3255 : /*
3256 : * Consider sorted paths for each interesting ordering. We generate both
3257 : * incremental and full sort.
3258 : */
3259 24616 : foreach(lc, useful_pathkeys_list)
3260 : {
3261 5510 : List *useful_pathkeys = lfirst(lc);
3262 : ListCell *lc2;
3263 : bool is_sorted;
3264 : int presorted_keys;
3265 :
3266 12210 : foreach(lc2, rel->partial_pathlist)
3267 : {
3268 6700 : Path *subpath = (Path *) lfirst(lc2);
3269 : GatherMergePath *path;
3270 :
3271 6700 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3272 : subpath->pathkeys,
3273 : &presorted_keys);
3274 :
3275 : /*
3276 : * We don't need to consider the case where a subpath is already
3277 : * fully sorted because generate_gather_paths already creates a
3278 : * gather merge path for every subpath that has pathkeys present.
3279 : *
3280 : * But since the subpath is already sorted, we know we don't need
3281 : * to consider adding a sort (full or incremental) on top of it,
3282 : * so we can continue here.
3283 : */
3284 6700 : if (is_sorted)
3285 1192 : continue;
3286 :
3287 : /*
3288 : * Try at least sorting the cheapest path and also try
3289 : * incrementally sorting any path which is partially sorted
3290 : * already (no need to deal with paths which have presorted keys
3291 : * when incremental sort is disabled unless it's the cheapest
3292 : * input path).
3293 : */
3294 5508 : if (subpath != cheapest_partial_path &&
3295 222 : (presorted_keys == 0 || !enable_incremental_sort))
3296 78 : continue;
3297 :
3298 : /*
3299 : * Consider regular sort for any path that's not presorted or if
3300 : * incremental sort is disabled. We've no need to consider both
3301 : * sort and incremental sort on the same path. We assume that
3302 : * incremental sort is always faster when there are presorted
3303 : * keys.
3304 : *
3305 : * This is not redundant with the gather paths created in
3306 : * generate_gather_paths, because that doesn't generate ordered
3307 : * output. Here we add an explicit sort to match the useful
3308 : * ordering.
3309 : */
3310 5430 : if (presorted_keys == 0 || !enable_incremental_sort)
3311 : {
3312 5274 : subpath = (Path *) create_sort_path(root,
3313 : rel,
3314 : subpath,
3315 : useful_pathkeys,
3316 : -1.0);
3317 : }
3318 : else
3319 156 : subpath = (Path *) create_incremental_sort_path(root,
3320 : rel,
3321 : subpath,
3322 : useful_pathkeys,
3323 : presorted_keys,
3324 : -1);
3325 5430 : rows = compute_gather_rows(subpath);
3326 5430 : path = create_gather_merge_path(root, rel,
3327 : subpath,
3328 5430 : rel->reltarget,
3329 : subpath->pathkeys,
3330 : NULL,
3331 : rowsp);
3332 :
3333 5430 : add_path(rel, &path->path);
3334 : }
3335 : }
3336 : }
3337 :
3338 : /*
3339 : * make_rel_from_joinlist
3340 : * Build access paths using a "joinlist" to guide the join path search.
3341 : *
3342 : * See comments for deconstruct_jointree() for definition of the joinlist
3343 : * data structure.
3344 : */
3345 : static RelOptInfo *
3346 329498 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3347 : {
3348 : int levels_needed;
3349 : List *initial_rels;
3350 : ListCell *jl;
3351 :
3352 : /*
3353 : * Count the number of child joinlist nodes. This is the depth of the
3354 : * dynamic-programming algorithm we must employ to consider all ways of
3355 : * joining the child nodes.
3356 : */
3357 329498 : levels_needed = list_length(joinlist);
3358 :
3359 329498 : if (levels_needed <= 0)
3360 0 : return NULL; /* nothing to do? */
3361 :
3362 : /*
3363 : * Construct a list of rels corresponding to the child joinlist nodes.
3364 : * This may contain both base rels and rels constructed according to
3365 : * sub-joinlists.
3366 : */
3367 329498 : initial_rels = NIL;
3368 793180 : foreach(jl, joinlist)
3369 : {
3370 463682 : Node *jlnode = (Node *) lfirst(jl);
3371 : RelOptInfo *thisrel;
3372 :
3373 463682 : if (IsA(jlnode, RangeTblRef))
3374 : {
3375 460304 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3376 :
3377 460304 : thisrel = find_base_rel(root, varno);
3378 : }
3379 3378 : else if (IsA(jlnode, List))
3380 : {
3381 : /* Recurse to handle subproblem */
3382 3378 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3383 : }
3384 : else
3385 : {
3386 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3387 : (int) nodeTag(jlnode));
3388 : thisrel = NULL; /* keep compiler quiet */
3389 : }
3390 :
3391 463682 : initial_rels = lappend(initial_rels, thisrel);
3392 : }
3393 :
3394 329498 : if (levels_needed == 1)
3395 : {
3396 : /*
3397 : * Single joinlist node, so we're done.
3398 : */
3399 231678 : return (RelOptInfo *) linitial(initial_rels);
3400 : }
3401 : else
3402 : {
3403 : /*
3404 : * Consider the different orders in which we could join the rels,
3405 : * using a plugin, GEQO, or the regular join search code.
3406 : *
3407 : * We put the initial_rels list into a PlannerInfo field because
3408 : * has_legal_joinclause() needs to look at it (ugly :-().
3409 : */
3410 97820 : root->initial_rels = initial_rels;
3411 :
3412 97820 : if (join_search_hook)
3413 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
3414 97820 : else if (enable_geqo && levels_needed >= geqo_threshold)
3415 6 : return geqo(root, levels_needed, initial_rels);
3416 : else
3417 97814 : return standard_join_search(root, levels_needed, initial_rels);
3418 : }
3419 : }
3420 :
3421 : /*
3422 : * standard_join_search
3423 : * Find possible joinpaths for a query by successively finding ways
3424 : * to join component relations into join relations.
3425 : *
3426 : * 'levels_needed' is the number of iterations needed, ie, the number of
3427 : * independent jointree items in the query. This is > 1.
3428 : *
3429 : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3430 : * jointree item. These are the components to be joined together.
3431 : * Note that levels_needed == list_length(initial_rels).
3432 : *
3433 : * Returns the final level of join relations, i.e., the relation that is
3434 : * the result of joining all the original relations together.
3435 : * At least one implementation path must be provided for this relation and
3436 : * all required sub-relations.
3437 : *
3438 : * To support loadable plugins that modify planner behavior by changing the
3439 : * join searching algorithm, we provide a hook variable that lets a plugin
3440 : * replace or supplement this function. Any such hook must return the same
3441 : * final join relation as the standard code would, but it might have a
3442 : * different set of implementation paths attached, and only the sub-joinrels
3443 : * needed for these paths need have been instantiated.
3444 : *
3445 : * Note to plugin authors: the functions invoked during standard_join_search()
3446 : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3447 : * than one join-order search, you'll probably need to save and restore the
3448 : * original states of those data structures. See geqo_eval() for an example.
3449 : */
3450 : RelOptInfo *
3451 97814 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3452 : {
3453 : int lev;
3454 : RelOptInfo *rel;
3455 :
3456 : /*
3457 : * This function cannot be invoked recursively within any one planning
3458 : * problem, so join_rel_level[] can't be in use already.
3459 : */
3460 : Assert(root->join_rel_level == NULL);
3461 :
3462 : /*
3463 : * We employ a simple "dynamic programming" algorithm: we first find all
3464 : * ways to build joins of two jointree items, then all ways to build joins
3465 : * of three items (from two-item joins and single items), then four-item
3466 : * joins, and so on until we have considered all ways to join all the
3467 : * items into one rel.
3468 : *
3469 : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3470 : * set root->join_rel_level[1] to represent all the single-jointree-item
3471 : * relations.
3472 : */
3473 97814 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3474 :
3475 97814 : root->join_rel_level[1] = initial_rels;
3476 :
3477 231974 : for (lev = 2; lev <= levels_needed; lev++)
3478 : {
3479 : ListCell *lc;
3480 :
3481 : /*
3482 : * Determine all possible pairs of relations to be joined at this
3483 : * level, and build paths for making each one from every available
3484 : * pair of lower-level relations.
3485 : */
3486 134160 : join_search_one_level(root, lev);
3487 :
3488 : /*
3489 : * Run generate_partitionwise_join_paths() and
3490 : * generate_useful_gather_paths() for each just-processed joinrel. We
3491 : * could not do this earlier because both regular and partial paths
3492 : * can get added to a particular joinrel at multiple times within
3493 : * join_search_one_level.
3494 : *
3495 : * After that, we're done creating paths for the joinrel, so run
3496 : * set_cheapest().
3497 : */
3498 343448 : foreach(lc, root->join_rel_level[lev])
3499 : {
3500 209288 : rel = (RelOptInfo *) lfirst(lc);
3501 :
3502 : /* Create paths for partitionwise joins. */
3503 209288 : generate_partitionwise_join_paths(root, rel);
3504 :
3505 : /*
3506 : * Except for the topmost scan/join rel, consider gathering
3507 : * partial paths. We'll do the same for the topmost scan/join rel
3508 : * once we know the final targetlist (see grouping_planner's and
3509 : * its call to apply_scanjoin_target_to_paths).
3510 : */
3511 209288 : if (!bms_equal(rel->relids, root->all_query_rels))
3512 111972 : generate_useful_gather_paths(root, rel, false);
3513 :
3514 : /* Find and save the cheapest paths for this rel */
3515 209288 : set_cheapest(rel);
3516 :
3517 : #ifdef OPTIMIZER_DEBUG
3518 : pprint(rel);
3519 : #endif
3520 : }
3521 : }
3522 :
3523 : /*
3524 : * We should have a single rel at the final level.
3525 : */
3526 97814 : if (root->join_rel_level[levels_needed] == NIL)
3527 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
3528 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3529 :
3530 97814 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3531 :
3532 97814 : root->join_rel_level = NULL;
3533 :
3534 97814 : return rel;
3535 : }
3536 :
3537 : /*****************************************************************************
3538 : * PUSHING QUALS DOWN INTO SUBQUERIES
3539 : *****************************************************************************/
3540 :
3541 : /*
3542 : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3543 : *
3544 : * subquery is the particular component query being checked. topquery
3545 : * is the top component of a set-operations tree (the same Query if no
3546 : * set-op is involved).
3547 : *
3548 : * Conditions checked here:
3549 : *
3550 : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3551 : * since that could change the set of rows returned.
3552 : *
3553 : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3554 : * quals into it, because that could change the results.
3555 : *
3556 : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3557 : * This is because upper-level quals should semantically be evaluated only
3558 : * once per distinct row, not once per original row, and if the qual is
3559 : * volatile then extra evaluations could change the results. (This issue
3560 : * does not apply to other forms of aggregation such as GROUP BY, because
3561 : * when those are present we push into HAVING not WHERE, so that the quals
3562 : * are still applied after aggregation.)
3563 : *
3564 : * 4. If the subquery contains window functions, we cannot push volatile quals
3565 : * into it. The issue here is a bit different from DISTINCT: a volatile qual
3566 : * might succeed for some rows of a window partition and fail for others,
3567 : * thereby changing the partition contents and thus the window functions'
3568 : * results for rows that remain.
3569 : *
3570 : * 5. If the subquery contains any set-returning functions in its targetlist,
3571 : * we cannot push volatile quals into it. That would push them below the SRFs
3572 : * and thereby change the number of times they are evaluated. Also, a
3573 : * volatile qual could succeed for some SRF output rows and fail for others,
3574 : * a behavior that cannot occur if it's evaluated before SRF expansion.
3575 : *
3576 : * 6. If the subquery has nonempty grouping sets, we cannot push down any
3577 : * quals. The concern here is that a qual referencing a "constant" grouping
3578 : * column could get constant-folded, which would be improper because the value
3579 : * is potentially nullable by grouping-set expansion. This restriction could
3580 : * be removed if we had a parsetree representation that shows that such
3581 : * grouping columns are not really constant. (There are other ideas that
3582 : * could be used to relax this restriction, but that's the approach most
3583 : * likely to get taken in the future. Note that there's not much to be gained
3584 : * so long as subquery_planner can't move HAVING clauses to WHERE within such
3585 : * a subquery.)
3586 : *
3587 : * In addition, we make several checks on the subquery's output columns to see
3588 : * if it is safe to reference them in pushed-down quals. If output column k
3589 : * is found to be unsafe to reference, we set the reason for that inside
3590 : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3591 : * column k might not be referenced by some/all quals. The unsafeFlags[]
3592 : * array will be consulted later by qual_is_pushdown_safe(). It's better to
3593 : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3594 : * because when the subquery involves set operations we have to check the
3595 : * output expressions in each arm of the set op.
3596 : *
3597 : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3598 : * we're effectively assuming that the quals cannot distinguish values that
3599 : * the DISTINCT's equality operator sees as equal, yet there are many
3600 : * counterexamples to that assumption. However use of such a qual with a
3601 : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3602 : * "equal" value will be chosen as the output value by the DISTINCT operation.
3603 : * So we don't worry too much about that. Another objection is that if the
3604 : * qual is expensive to evaluate, running it for each original row might cost
3605 : * more than we save by eliminating rows before the DISTINCT step. But it
3606 : * would be very hard to estimate that at this stage, and in practice pushdown
3607 : * seldom seems to make things worse, so we ignore that problem too.
3608 : *
3609 : * Note: likewise, pushing quals into a subquery with window functions is a
3610 : * bit dubious: the quals might remove some rows of a window partition while
3611 : * leaving others, causing changes in the window functions' results for the
3612 : * surviving rows. We insist that such a qual reference only partitioning
3613 : * columns, but again that only protects us if the qual does not distinguish
3614 : * values that the partitioning equality operator sees as equal. The risks
3615 : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3616 : * occurs and thus there is no theoretical problem with such a qual. But
3617 : * we'll do this anyway because the potential performance benefits are very
3618 : * large, and we've seen no field complaints about the longstanding comparable
3619 : * behavior with DISTINCT.
3620 : */
3621 : static bool
3622 2150 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
3623 : pushdown_safety_info *safetyInfo)
3624 : {
3625 : SetOperationStmt *topop;
3626 :
3627 : /* Check point 1 */
3628 2150 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
3629 134 : return false;
3630 :
3631 : /* Check point 6 */
3632 2016 : if (subquery->groupClause && subquery->groupingSets)
3633 12 : return false;
3634 :
3635 : /* Check points 3, 4, and 5 */
3636 2004 : if (subquery->distinctClause ||
3637 1920 : subquery->hasWindowFuncs ||
3638 1654 : subquery->hasTargetSRFs)
3639 546 : safetyInfo->unsafeVolatile = true;
3640 :
3641 : /*
3642 : * If we're at a leaf query, check for unsafe expressions in its target
3643 : * list, and mark any reasons why they're unsafe in unsafeFlags[].
3644 : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3645 : * so no need to check them.)
3646 : */
3647 2004 : if (subquery->setOperations == NULL)
3648 1848 : check_output_expressions(subquery, safetyInfo);
3649 :
3650 : /* Are we at top level, or looking at a setop component? */
3651 2004 : if (subquery == topquery)
3652 : {
3653 : /* Top level, so check any component queries */
3654 1692 : if (subquery->setOperations != NULL)
3655 156 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3656 : safetyInfo))
3657 0 : return false;
3658 : }
3659 : else
3660 : {
3661 : /* Setop component must not have more components (too weird) */
3662 312 : if (subquery->setOperations != NULL)
3663 0 : return false;
3664 : /* Check whether setop component output types match top level */
3665 312 : topop = castNode(SetOperationStmt, topquery->setOperations);
3666 : Assert(topop);
3667 312 : compare_tlist_datatypes(subquery->targetList,
3668 : topop->colTypes,
3669 : safetyInfo);
3670 : }
3671 2004 : return true;
3672 : }
3673 :
3674 : /*
3675 : * Helper routine to recurse through setOperations tree
3676 : */
3677 : static bool
3678 468 : recurse_pushdown_safe(Node *setOp, Query *topquery,
3679 : pushdown_safety_info *safetyInfo)
3680 : {
3681 468 : if (IsA(setOp, RangeTblRef))
3682 : {
3683 312 : RangeTblRef *rtr = (RangeTblRef *) setOp;
3684 312 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3685 312 : Query *subquery = rte->subquery;
3686 :
3687 : Assert(subquery != NULL);
3688 312 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3689 : }
3690 156 : else if (IsA(setOp, SetOperationStmt))
3691 : {
3692 156 : SetOperationStmt *op = (SetOperationStmt *) setOp;
3693 :
3694 : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
3695 156 : if (op->op == SETOP_EXCEPT)
3696 0 : return false;
3697 : /* Else recurse */
3698 156 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
3699 0 : return false;
3700 156 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
3701 0 : return false;
3702 : }
3703 : else
3704 : {
3705 0 : elog(ERROR, "unrecognized node type: %d",
3706 : (int) nodeTag(setOp));
3707 : }
3708 156 : return true;
3709 : }
3710 :
3711 : /*
3712 : * check_output_expressions - check subquery's output expressions for safety
3713 : *
3714 : * There are several cases in which it's unsafe to push down an upper-level
3715 : * qual if it references a particular output column of a subquery. We check
3716 : * each output column of the subquery and set flags in unsafeFlags[k] when we
3717 : * see that column is unsafe for a pushed-down qual to reference. The
3718 : * conditions checked here are:
3719 : *
3720 : * 1. We must not push down any quals that refer to subselect outputs that
3721 : * return sets, else we'd introduce functions-returning-sets into the
3722 : * subquery's WHERE/HAVING quals.
3723 : *
3724 : * 2. We must not push down any quals that refer to subselect outputs that
3725 : * contain volatile functions, for fear of introducing strange results due
3726 : * to multiple evaluation of a volatile function.
3727 : *
3728 : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3729 : * refer to non-DISTINCT output columns, because that could change the set
3730 : * of rows returned. (This condition is vacuous for DISTINCT, because then
3731 : * there are no non-DISTINCT output columns, so we needn't check. Note that
3732 : * subquery_is_pushdown_safe already reported that we can't use volatile
3733 : * quals if there's DISTINCT or DISTINCT ON.)
3734 : *
3735 : * 4. If the subquery has any window functions, we must not push down quals
3736 : * that reference any output columns that are not listed in all the subquery's
3737 : * window PARTITION BY clauses. We can push down quals that use only
3738 : * partitioning columns because they should succeed or fail identically for
3739 : * every row of any one window partition, and totally excluding some
3740 : * partitions will not change a window function's results for remaining
3741 : * partitions. (Again, this also requires nonvolatile quals, but
3742 : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3743 : * unsafe for this reason can still have WindowClause run conditions pushed
3744 : * down.
3745 : */
3746 : static void
3747 1848 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
3748 : {
3749 : ListCell *lc;
3750 :
3751 17372 : foreach(lc, subquery->targetList)
3752 : {
3753 15524 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
3754 :
3755 15524 : if (tle->resjunk)
3756 134 : continue; /* ignore resjunk columns */
3757 :
3758 : /* Functions returning sets are unsafe (point 1) */
3759 15390 : if (subquery->hasTargetSRFs &&
3760 668 : (safetyInfo->unsafeFlags[tle->resno] &
3761 668 : UNSAFE_HAS_SET_FUNC) == 0 &&
3762 668 : expression_returns_set((Node *) tle->expr))
3763 : {
3764 376 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3765 376 : continue;
3766 : }
3767 :
3768 : /* Volatile functions are unsafe (point 2) */
3769 15014 : if ((safetyInfo->unsafeFlags[tle->resno] &
3770 15002 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3771 15002 : contain_volatile_functions((Node *) tle->expr))
3772 : {
3773 78 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3774 78 : continue;
3775 : }
3776 :
3777 : /* If subquery uses DISTINCT ON, check point 3 */
3778 14936 : if (subquery->hasDistinctOn &&
3779 0 : (safetyInfo->unsafeFlags[tle->resno] &
3780 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3781 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3782 : {
3783 : /* non-DISTINCT column, so mark it unsafe */
3784 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3785 0 : continue;
3786 : }
3787 :
3788 : /* If subquery uses window functions, check point 4 */
3789 14936 : if (subquery->hasWindowFuncs &&
3790 1090 : (safetyInfo->unsafeFlags[tle->resno] &
3791 2092 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3792 1090 : !targetIsInAllPartitionLists(tle, subquery))
3793 : {
3794 : /* not present in all PARTITION BY clauses, so mark it unsafe */
3795 1002 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3796 1002 : continue;
3797 : }
3798 : }
3799 1848 : }
3800 :
3801 : /*
3802 : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3803 : * push quals into each component query, but the quals can only reference
3804 : * subquery columns that suffer no type coercions in the set operation.
3805 : * Otherwise there are possible semantic gotchas. So, we check the
3806 : * component queries to see if any of them have output types different from
3807 : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3808 : * unsafeFlags[k] if column k has different type in any component.
3809 : *
3810 : * We don't have to care about typmods here: the only allowed difference
3811 : * between set-op input and output typmods is input is a specific typmod
3812 : * and output is -1, and that does not require a coercion.
3813 : *
3814 : * tlist is a subquery tlist.
3815 : * colTypes is an OID list of the top-level setop's output column types.
3816 : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
3817 : */
3818 : static void
3819 312 : compare_tlist_datatypes(List *tlist, List *colTypes,
3820 : pushdown_safety_info *safetyInfo)
3821 : {
3822 : ListCell *l;
3823 312 : ListCell *colType = list_head(colTypes);
3824 :
3825 984 : foreach(l, tlist)
3826 : {
3827 672 : TargetEntry *tle = (TargetEntry *) lfirst(l);
3828 :
3829 672 : if (tle->resjunk)
3830 0 : continue; /* ignore resjunk columns */
3831 672 : if (colType == NULL)
3832 0 : elog(ERROR, "wrong number of tlist entries");
3833 672 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
3834 104 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
3835 672 : colType = lnext(colTypes, colType);
3836 : }
3837 312 : if (colType != NULL)
3838 0 : elog(ERROR, "wrong number of tlist entries");
3839 312 : }
3840 :
3841 : /*
3842 : * targetIsInAllPartitionLists
3843 : * True if the TargetEntry is listed in the PARTITION BY clause
3844 : * of every window defined in the query.
3845 : *
3846 : * It would be safe to ignore windows not actually used by any window
3847 : * function, but it's not easy to get that info at this stage; and it's
3848 : * unlikely to be useful to spend any extra cycles getting it, since
3849 : * unreferenced window definitions are probably infrequent in practice.
3850 : */
3851 : static bool
3852 1090 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
3853 : {
3854 : ListCell *lc;
3855 :
3856 1202 : foreach(lc, query->windowClause)
3857 : {
3858 1114 : WindowClause *wc = (WindowClause *) lfirst(lc);
3859 :
3860 1114 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
3861 1002 : return false;
3862 : }
3863 88 : return true;
3864 : }
3865 :
3866 : /*
3867 : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3868 : *
3869 : * rinfo is a restriction clause applying to the given subquery (whose RTE
3870 : * has index rti in the parent query).
3871 : *
3872 : * Conditions checked here:
3873 : *
3874 : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3875 : * unclear that it will work correctly: SubLinks will already have been
3876 : * transformed into SubPlans in the qual, but not in the subquery). Note that
3877 : * SubLinks that transform to initplans are safe, and will be accepted here
3878 : * because what we'll see in the qual is just a Param referencing the initplan
3879 : * output.
3880 : *
3881 : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3882 : * functions.
3883 : *
3884 : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
3885 : * functions that are passed Var nodes, and therefore might reveal values from
3886 : * the subquery as side effects.
3887 : *
3888 : * 4. rinfo's clause must not refer to the whole-row output of the subquery
3889 : * (since there is no easy way to name that within the subquery itself).
3890 : *
3891 : * 5. rinfo's clause must not refer to any subquery output columns that were
3892 : * found to be unsafe to reference by subquery_is_pushdown_safe().
3893 : */
3894 : static pushdown_safe_type
3895 2634 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
3896 : pushdown_safety_info *safetyInfo)
3897 : {
3898 2634 : pushdown_safe_type safe = PUSHDOWN_SAFE;
3899 2634 : Node *qual = (Node *) rinfo->clause;
3900 : List *vars;
3901 : ListCell *vl;
3902 :
3903 : /* Refuse subselects (point 1) */
3904 2634 : if (contain_subplans(qual))
3905 66 : return PUSHDOWN_UNSAFE;
3906 :
3907 : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3908 3220 : if (safetyInfo->unsafeVolatile &&
3909 652 : contain_volatile_functions((Node *) rinfo))
3910 18 : return PUSHDOWN_UNSAFE;
3911 :
3912 : /* Refuse leaky quals if told to (point 3) */
3913 3702 : if (safetyInfo->unsafeLeaky &&
3914 1152 : contain_leaked_vars(qual))
3915 150 : return PUSHDOWN_UNSAFE;
3916 :
3917 : /*
3918 : * Examine all Vars used in clause. Since it's a restriction clause, all
3919 : * such Vars must refer to subselect output columns ... unless this is
3920 : * part of a LATERAL subquery, in which case there could be lateral
3921 : * references.
3922 : *
3923 : * By omitting the relevant flags, this also gives us a cheap sanity check
3924 : * that no aggregates or window functions appear in the qual. Those would
3925 : * be unsafe to push down, but at least for the moment we could never see
3926 : * any in a qual anyhow.
3927 : */
3928 2400 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
3929 4700 : foreach(vl, vars)
3930 : {
3931 2508 : Var *var = (Var *) lfirst(vl);
3932 :
3933 : /*
3934 : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3935 : * It's not clear whether a PHV could safely be pushed down, and even
3936 : * less clear whether such a situation could arise in any cases of
3937 : * practical interest anyway. So for the moment, just refuse to push
3938 : * down.
3939 : */
3940 2508 : if (!IsA(var, Var))
3941 : {
3942 0 : safe = PUSHDOWN_UNSAFE;
3943 0 : break;
3944 : }
3945 :
3946 : /*
3947 : * Punt if we find any lateral references. It would be safe to push
3948 : * these down, but we'd have to convert them into outer references,
3949 : * which subquery_push_qual lacks the infrastructure to do. The case
3950 : * arises so seldom that it doesn't seem worth working hard on.
3951 : */
3952 2508 : if (var->varno != rti)
3953 : {
3954 12 : safe = PUSHDOWN_UNSAFE;
3955 12 : break;
3956 : }
3957 :
3958 : /* Subqueries have no system columns */
3959 : Assert(var->varattno >= 0);
3960 :
3961 : /* Check point 4 */
3962 2496 : if (var->varattno == 0)
3963 : {
3964 0 : safe = PUSHDOWN_UNSAFE;
3965 0 : break;
3966 : }
3967 :
3968 : /* Check point 5 */
3969 2496 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
3970 : {
3971 526 : if (safetyInfo->unsafeFlags[var->varattno] &
3972 : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
3973 : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
3974 : {
3975 196 : safe = PUSHDOWN_UNSAFE;
3976 196 : break;
3977 : }
3978 : else
3979 : {
3980 : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
3981 330 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
3982 : /* don't break, we might find another Var that's unsafe */
3983 : }
3984 : }
3985 : }
3986 :
3987 2400 : list_free(vars);
3988 :
3989 2400 : return safe;
3990 : }
3991 :
3992 : /*
3993 : * subquery_push_qual - push down a qual that we have determined is safe
3994 : */
3995 : static void
3996 2204 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3997 : {
3998 2204 : if (subquery->setOperations != NULL)
3999 : {
4000 : /* Recurse to push it separately to each component query */
4001 132 : recurse_push_qual(subquery->setOperations, subquery,
4002 : rte, rti, qual);
4003 : }
4004 : else
4005 : {
4006 : /*
4007 : * We need to replace Vars in the qual (which must refer to outputs of
4008 : * the subquery) with copies of the subquery's targetlist expressions.
4009 : * Note that at this point, any uplevel Vars in the qual should have
4010 : * been replaced with Params, so they need no work.
4011 : *
4012 : * This step also ensures that when we are pushing into a setop tree,
4013 : * each component query gets its own copy of the qual.
4014 : */
4015 2072 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
4016 : subquery->targetList,
4017 : subquery->resultRelation,
4018 : REPLACEVARS_REPORT_ERROR, 0,
4019 : &subquery->hasSubLinks);
4020 :
4021 : /*
4022 : * Now attach the qual to the proper place: normally WHERE, but if the
4023 : * subquery uses grouping or aggregation, put it in HAVING (since the
4024 : * qual really refers to the group-result rows).
4025 : */
4026 2072 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
4027 272 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
4028 : else
4029 1800 : subquery->jointree->quals =
4030 1800 : make_and_qual(subquery->jointree->quals, qual);
4031 :
4032 : /*
4033 : * We need not change the subquery's hasAggs or hasSubLinks flags,
4034 : * since we can't be pushing down any aggregates that weren't there
4035 : * before, and we don't push down subselects at all.
4036 : */
4037 : }
4038 2204 : }
4039 :
4040 : /*
4041 : * Helper routine to recurse through setOperations tree
4042 : */
4043 : static void
4044 396 : recurse_push_qual(Node *setOp, Query *topquery,
4045 : RangeTblEntry *rte, Index rti, Node *qual)
4046 : {
4047 396 : if (IsA(setOp, RangeTblRef))
4048 : {
4049 264 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4050 264 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4051 264 : Query *subquery = subrte->subquery;
4052 :
4053 : Assert(subquery != NULL);
4054 264 : subquery_push_qual(subquery, rte, rti, qual);
4055 : }
4056 132 : else if (IsA(setOp, SetOperationStmt))
4057 : {
4058 132 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4059 :
4060 132 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4061 132 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4062 : }
4063 : else
4064 : {
4065 0 : elog(ERROR, "unrecognized node type: %d",
4066 : (int) nodeTag(setOp));
4067 : }
4068 396 : }
4069 :
4070 : /*****************************************************************************
4071 : * SIMPLIFYING SUBQUERY TARGETLISTS
4072 : *****************************************************************************/
4073 :
4074 : /*
4075 : * remove_unused_subquery_outputs
4076 : * Remove subquery targetlist items we don't need
4077 : *
4078 : * It's possible, even likely, that the upper query does not read all the
4079 : * output columns of the subquery. We can remove any such outputs that are
4080 : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4081 : * affect semantics otherwise (e.g., volatile functions can't be removed).
4082 : * This is useful not only because we might be able to remove expensive-to-
4083 : * compute expressions, but because deletion of output columns might allow
4084 : * optimizations such as join removal to occur within the subquery.
4085 : *
4086 : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4087 : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4088 : * parameter is modified by the function, so callers must make a copy if they
4089 : * need to use the passed in Bitmapset after calling this function.
4090 : *
4091 : * To avoid affecting column numbering in the targetlist, we don't physically
4092 : * remove unused tlist entries, but rather replace their expressions with NULL
4093 : * constants. This is implemented by modifying subquery->targetList.
4094 : */
4095 : static void
4096 11588 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4097 : Bitmapset *extra_used_attrs)
4098 : {
4099 : Bitmapset *attrs_used;
4100 : ListCell *lc;
4101 :
4102 : /*
4103 : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4104 : * the current callers use the Bitmapset after calling this function.
4105 : */
4106 11588 : attrs_used = extra_used_attrs;
4107 :
4108 : /*
4109 : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4110 : * could update all the child SELECTs' tlists, but it seems not worth the
4111 : * trouble presently.
4112 : */
4113 11588 : if (subquery->setOperations)
4114 1280 : return;
4115 :
4116 : /*
4117 : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4118 : * time: all its output columns must be used in the distinctClause.
4119 : */
4120 10852 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4121 246 : return;
4122 :
4123 : /*
4124 : * Collect a bitmap of all the output column numbers used by the upper
4125 : * query.
4126 : *
4127 : * Add all the attributes needed for joins or final output. Note: we must
4128 : * look at rel's targetlist, not the attr_needed data, because attr_needed
4129 : * isn't computed for inheritance child rels, cf set_append_rel_size().
4130 : * (XXX might be worth changing that sometime.)
4131 : */
4132 10606 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4133 :
4134 : /* Add all the attributes used by un-pushed-down restriction clauses. */
4135 11330 : foreach(lc, rel->baserestrictinfo)
4136 : {
4137 724 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4138 :
4139 724 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4140 : }
4141 :
4142 : /*
4143 : * If there's a whole-row reference to the subquery, we can't remove
4144 : * anything.
4145 : */
4146 10606 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4147 298 : return;
4148 :
4149 : /*
4150 : * Run through the tlist and zap entries we don't need. It's okay to
4151 : * modify the tlist items in-place because set_subquery_pathlist made a
4152 : * copy of the subquery.
4153 : */
4154 43326 : foreach(lc, subquery->targetList)
4155 : {
4156 33018 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4157 33018 : Node *texpr = (Node *) tle->expr;
4158 :
4159 : /*
4160 : * If it has a sortgroupref number, it's used in some sort/group
4161 : * clause so we'd better not remove it. Also, don't remove any
4162 : * resjunk columns, since their reason for being has nothing to do
4163 : * with anybody reading the subquery's output. (It's likely that
4164 : * resjunk columns in a sub-SELECT would always have ressortgroupref
4165 : * set, but even if they don't, it seems imprudent to remove them.)
4166 : */
4167 33018 : if (tle->ressortgroupref || tle->resjunk)
4168 2596 : continue;
4169 :
4170 : /*
4171 : * If it's used by the upper query, we can't remove it.
4172 : */
4173 30422 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4174 : attrs_used))
4175 22616 : continue;
4176 :
4177 : /*
4178 : * If it contains a set-returning function, we can't remove it since
4179 : * that could change the number of rows returned by the subquery.
4180 : */
4181 8826 : if (subquery->hasTargetSRFs &&
4182 1020 : expression_returns_set(texpr))
4183 752 : continue;
4184 :
4185 : /*
4186 : * If it contains volatile functions, we daren't remove it for fear
4187 : * that the user is expecting their side-effects to happen.
4188 : */
4189 7054 : if (contain_volatile_functions(texpr))
4190 32 : continue;
4191 :
4192 : /*
4193 : * OK, we don't need it. Replace the expression with a NULL constant.
4194 : * Preserve the exposed type of the expression, in case something
4195 : * looks at the rowtype of the subquery's result.
4196 : */
4197 7022 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4198 : exprTypmod(texpr),
4199 : exprCollation(texpr));
4200 : }
4201 : }
4202 :
4203 : /*
4204 : * create_partial_bitmap_paths
4205 : * Build partial bitmap heap path for the relation
4206 : */
4207 : void
4208 144400 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4209 : Path *bitmapqual)
4210 : {
4211 : int parallel_workers;
4212 : double pages_fetched;
4213 :
4214 : /* Compute heap pages for bitmap heap scan */
4215 144400 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4216 : NULL, NULL);
4217 :
4218 144400 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4219 : max_parallel_workers_per_gather);
4220 :
4221 144400 : if (parallel_workers <= 0)
4222 140224 : return;
4223 :
4224 4176 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4225 : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4226 : }
4227 :
4228 : /*
4229 : * Compute the number of parallel workers that should be used to scan a
4230 : * relation. We compute the parallel workers based on the size of the heap to
4231 : * be scanned and the size of the index to be scanned, then choose a minimum
4232 : * of those.
4233 : *
4234 : * "heap_pages" is the number of pages from the table that we expect to scan, or
4235 : * -1 if we don't expect to scan any.
4236 : *
4237 : * "index_pages" is the number of pages from the index that we expect to scan, or
4238 : * -1 if we don't expect to scan any.
4239 : *
4240 : * "max_workers" is caller's limit on the number of workers. This typically
4241 : * comes from a GUC.
4242 : */
4243 : int
4244 749292 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4245 : int max_workers)
4246 : {
4247 749292 : int parallel_workers = 0;
4248 :
4249 : /*
4250 : * If the user has set the parallel_workers reloption, use that; otherwise
4251 : * select a default number of workers.
4252 : */
4253 749292 : if (rel->rel_parallel_workers != -1)
4254 1914 : parallel_workers = rel->rel_parallel_workers;
4255 : else
4256 : {
4257 : /*
4258 : * If the number of pages being scanned is insufficient to justify a
4259 : * parallel scan, just return zero ... unless it's an inheritance
4260 : * child. In that case, we want to generate a parallel path here
4261 : * anyway. It might not be worthwhile just for this relation, but
4262 : * when combined with all of its inheritance siblings it may well pay
4263 : * off.
4264 : */
4265 747378 : if (rel->reloptkind == RELOPT_BASEREL &&
4266 709182 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4267 21058 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4268 708448 : return 0;
4269 :
4270 38930 : if (heap_pages >= 0)
4271 : {
4272 : int heap_parallel_threshold;
4273 36840 : int heap_parallel_workers = 1;
4274 :
4275 : /*
4276 : * Select the number of workers based on the log of the size of
4277 : * the relation. This probably needs to be a good deal more
4278 : * sophisticated, but we need something here for now. Note that
4279 : * the upper limit of the min_parallel_table_scan_size GUC is
4280 : * chosen to prevent overflow here.
4281 : */
4282 36840 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4283 41660 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4284 : {
4285 4820 : heap_parallel_workers++;
4286 4820 : heap_parallel_threshold *= 3;
4287 4820 : if (heap_parallel_threshold > INT_MAX / 3)
4288 0 : break; /* avoid overflow */
4289 : }
4290 :
4291 36840 : parallel_workers = heap_parallel_workers;
4292 : }
4293 :
4294 38930 : if (index_pages >= 0)
4295 : {
4296 9726 : int index_parallel_workers = 1;
4297 : int index_parallel_threshold;
4298 :
4299 : /* same calculation as for heap_pages above */
4300 9726 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4301 10002 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4302 : {
4303 276 : index_parallel_workers++;
4304 276 : index_parallel_threshold *= 3;
4305 276 : if (index_parallel_threshold > INT_MAX / 3)
4306 0 : break; /* avoid overflow */
4307 : }
4308 :
4309 9726 : if (parallel_workers > 0)
4310 7636 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4311 : else
4312 2090 : parallel_workers = index_parallel_workers;
4313 : }
4314 : }
4315 :
4316 : /* In no case use more than caller supplied maximum number of workers */
4317 40844 : parallel_workers = Min(parallel_workers, max_workers);
4318 :
4319 40844 : return parallel_workers;
4320 : }
4321 :
4322 : /*
4323 : * generate_partitionwise_join_paths
4324 : * Create paths representing partitionwise join for given partitioned
4325 : * join relation.
4326 : *
4327 : * This must not be called until after we are done adding paths for all
4328 : * child-joins. Otherwise, add_path might delete a path to which some path
4329 : * generated here has a reference.
4330 : */
4331 : void
4332 217182 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4333 : {
4334 217182 : List *live_children = NIL;
4335 : int cnt_parts;
4336 : int num_parts;
4337 : RelOptInfo **part_rels;
4338 :
4339 : /* Handle only join relations here. */
4340 217182 : if (!IS_JOIN_REL(rel))
4341 0 : return;
4342 :
4343 : /* We've nothing to do if the relation is not partitioned. */
4344 217182 : if (!IS_PARTITIONED_REL(rel))
4345 215412 : return;
4346 :
4347 : /* The relation should have consider_partitionwise_join set. */
4348 : Assert(rel->consider_partitionwise_join);
4349 :
4350 : /* Guard against stack overflow due to overly deep partition hierarchy. */
4351 1770 : check_stack_depth();
4352 :
4353 1770 : num_parts = rel->nparts;
4354 1770 : part_rels = rel->part_rels;
4355 :
4356 : /* Collect non-dummy child-joins. */
4357 6620 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4358 : {
4359 4850 : RelOptInfo *child_rel = part_rels[cnt_parts];
4360 :
4361 : /* If it's been pruned entirely, it's certainly dummy. */
4362 4850 : if (child_rel == NULL)
4363 52 : continue;
4364 :
4365 : /* Make partitionwise join paths for this partitioned child-join. */
4366 4798 : generate_partitionwise_join_paths(root, child_rel);
4367 :
4368 : /* If we failed to make any path for this child, we must give up. */
4369 4798 : if (child_rel->pathlist == NIL)
4370 : {
4371 : /*
4372 : * Mark the parent joinrel as unpartitioned so that later
4373 : * functions treat it correctly.
4374 : */
4375 0 : rel->nparts = 0;
4376 0 : return;
4377 : }
4378 :
4379 : /* Else, identify the cheapest path for it. */
4380 4798 : set_cheapest(child_rel);
4381 :
4382 : /* Dummy children need not be scanned, so ignore those. */
4383 4798 : if (IS_DUMMY_REL(child_rel))
4384 0 : continue;
4385 :
4386 : #ifdef OPTIMIZER_DEBUG
4387 : pprint(child_rel);
4388 : #endif
4389 :
4390 4798 : live_children = lappend(live_children, child_rel);
4391 : }
4392 :
4393 : /* If all child-joins are dummy, parent join is also dummy. */
4394 1770 : if (!live_children)
4395 : {
4396 0 : mark_dummy_rel(rel);
4397 0 : return;
4398 : }
4399 :
4400 : /* Build additional paths for this rel from child-join paths. */
4401 1770 : add_paths_to_append_rel(root, rel, live_children);
4402 1770 : list_free(live_children);
4403 : }
|
