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/prep.h"
44 : #include "optimizer/tlist.h"
45 : #include "parser/parse_clause.h"
46 : #include "parser/parsetree.h"
47 : #include "partitioning/partbounds.h"
48 : #include "port/pg_bitutils.h"
49 : #include "rewrite/rewriteManip.h"
50 : #include "utils/lsyscache.h"
51 : #include "utils/selfuncs.h"
52 :
53 :
54 : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
55 : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
56 : #define UNSAFE_HAS_SET_FUNC (1 << 1)
57 : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
58 : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
59 : #define UNSAFE_TYPE_MISMATCH (1 << 4)
60 :
61 : /* results of subquery_is_pushdown_safe */
62 : typedef struct pushdown_safety_info
63 : {
64 : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
65 : * column is unsafe for qual pushdown, or 0 if
66 : * no reason. */
67 : bool unsafeVolatile; /* don't push down volatile quals */
68 : bool unsafeLeaky; /* don't push down leaky quals */
69 : } pushdown_safety_info;
70 :
71 : /* Return type for qual_is_pushdown_safe */
72 : typedef enum pushdown_safe_type
73 : {
74 : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
75 : PUSHDOWN_SAFE, /* safe to push qual into subquery */
76 : PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
77 : * run condition */
78 : } pushdown_safe_type;
79 :
80 : /* These parameters are set by GUC */
81 : bool enable_geqo = false; /* just in case GUC doesn't set it */
82 : bool enable_eager_aggregate = true;
83 : int geqo_threshold;
84 : double min_eager_agg_group_size;
85 : int min_parallel_table_scan_size;
86 : int min_parallel_index_scan_size;
87 :
88 : /* Hook for plugins to get control in set_rel_pathlist() */
89 : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
90 :
91 : /* Hook for plugins to replace standard_join_search() */
92 : join_search_hook_type join_search_hook = NULL;
93 :
94 :
95 : static void set_base_rel_consider_startup(PlannerInfo *root);
96 : static void set_base_rel_sizes(PlannerInfo *root);
97 : static void setup_simple_grouped_rels(PlannerInfo *root);
98 : static void set_base_rel_pathlists(PlannerInfo *root);
99 : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
100 : Index rti, RangeTblEntry *rte);
101 : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
102 : Index rti, RangeTblEntry *rte);
103 : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
104 : RangeTblEntry *rte);
105 : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
106 : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
107 : RangeTblEntry *rte);
108 : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
109 : RangeTblEntry *rte);
110 : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
111 : RangeTblEntry *rte);
112 : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
113 : RangeTblEntry *rte);
114 : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
115 : RangeTblEntry *rte);
116 : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
117 : RangeTblEntry *rte);
118 : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
119 : Index rti, RangeTblEntry *rte);
120 : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
121 : Index rti, RangeTblEntry *rte);
122 : static void set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel);
123 : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
124 : List *live_childrels,
125 : List *all_child_pathkeys);
126 : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
127 : RelOptInfo *rel,
128 : Relids required_outer);
129 : static void accumulate_append_subpath(Path *path,
130 : List **subpaths,
131 : List **special_subpaths);
132 : static Path *get_singleton_append_subpath(Path *path);
133 : static void set_dummy_rel_pathlist(RelOptInfo *rel);
134 : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
135 : Index rti, RangeTblEntry *rte);
136 : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : RangeTblEntry *rte);
138 : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : RangeTblEntry *rte);
140 : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : RangeTblEntry *rte);
142 : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : RangeTblEntry *rte);
144 : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
145 : RangeTblEntry *rte);
146 : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
147 : RangeTblEntry *rte);
148 : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
149 : RangeTblEntry *rte);
150 : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
151 : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
152 : pushdown_safety_info *safetyInfo);
153 : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
154 : pushdown_safety_info *safetyInfo);
155 : static void check_output_expressions(Query *subquery,
156 : pushdown_safety_info *safetyInfo);
157 : static void compare_tlist_datatypes(List *tlist, List *colTypes,
158 : pushdown_safety_info *safetyInfo);
159 : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
160 : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
161 : RestrictInfo *rinfo,
162 : pushdown_safety_info *safetyInfo);
163 : static void subquery_push_qual(Query *subquery,
164 : RangeTblEntry *rte, Index rti, Node *qual);
165 : static void recurse_push_qual(Node *setOp, Query *topquery,
166 : RangeTblEntry *rte, Index rti, Node *qual);
167 : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
168 : Bitmapset *extra_used_attrs);
169 :
170 :
171 : /*
172 : * make_one_rel
173 : * Finds all possible access paths for executing a query, returning a
174 : * single rel that represents the join of all base rels in the query.
175 : */
176 : RelOptInfo *
177 325366 : make_one_rel(PlannerInfo *root, List *joinlist)
178 : {
179 : RelOptInfo *rel;
180 : Index rti;
181 : double total_pages;
182 :
183 : /* Mark base rels as to whether we care about fast-start plans */
184 325366 : set_base_rel_consider_startup(root);
185 :
186 : /*
187 : * Compute size estimates and consider_parallel flags for each base rel.
188 : */
189 325366 : set_base_rel_sizes(root);
190 :
191 : /*
192 : * Build grouped relations for simple rels (i.e., base or "other" member
193 : * relations) where possible.
194 : */
195 325332 : setup_simple_grouped_rels(root);
196 :
197 : /*
198 : * We should now have size estimates for every actual table involved in
199 : * the query, and we also know which if any have been deleted from the
200 : * query by join removal, pruned by partition pruning, or eliminated by
201 : * constraint exclusion. So we can now compute total_table_pages.
202 : *
203 : * Note that appendrels are not double-counted here, even though we don't
204 : * bother to distinguish RelOptInfos for appendrel parents, because the
205 : * parents will have pages = 0.
206 : *
207 : * XXX if a table is self-joined, we will count it once per appearance,
208 : * which perhaps is the wrong thing ... but that's not completely clear,
209 : * and detecting self-joins here is difficult, so ignore it for now.
210 : */
211 325332 : total_pages = 0;
212 1007684 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
213 : {
214 682352 : RelOptInfo *brel = root->simple_rel_array[rti];
215 :
216 : /* there may be empty slots corresponding to non-baserel RTEs */
217 682352 : if (brel == NULL)
218 161690 : continue;
219 :
220 : Assert(brel->relid == rti); /* sanity check on array */
221 :
222 520662 : if (IS_DUMMY_REL(brel))
223 1352 : continue;
224 :
225 519310 : if (IS_SIMPLE_REL(brel))
226 519310 : total_pages += (double) brel->pages;
227 : }
228 325332 : root->total_table_pages = total_pages;
229 :
230 : /*
231 : * Generate access paths for each base rel.
232 : */
233 325332 : set_base_rel_pathlists(root);
234 :
235 : /*
236 : * Generate access paths for the entire join tree.
237 : */
238 325332 : rel = make_rel_from_joinlist(root, joinlist);
239 :
240 : /*
241 : * The result should join all and only the query's base + outer-join rels.
242 : */
243 : Assert(bms_equal(rel->relids, root->all_query_rels));
244 :
245 325332 : return rel;
246 : }
247 :
248 : /*
249 : * set_base_rel_consider_startup
250 : * Set the consider_[param_]startup flags for each base-relation entry.
251 : *
252 : * For the moment, we only deal with consider_param_startup here; because the
253 : * logic for consider_startup is pretty trivial and is the same for every base
254 : * relation, we just let build_simple_rel() initialize that flag correctly to
255 : * start with. If that logic ever gets more complicated it would probably
256 : * be better to move it here.
257 : */
258 : static void
259 325366 : set_base_rel_consider_startup(PlannerInfo *root)
260 : {
261 : /*
262 : * Since parameterized paths can only be used on the inside of a nestloop
263 : * join plan, there is usually little value in considering fast-start
264 : * plans for them. However, for relations that are on the RHS of a SEMI
265 : * or ANTI join, a fast-start plan can be useful because we're only going
266 : * to care about fetching one tuple anyway.
267 : *
268 : * To minimize growth of planning time, we currently restrict this to
269 : * cases where the RHS is a single base relation, not a join; there is no
270 : * provision for consider_param_startup to get set at all on joinrels.
271 : * Also we don't worry about appendrels. costsize.c's costing rules for
272 : * nestloop semi/antijoins don't consider such cases either.
273 : */
274 : ListCell *lc;
275 :
276 367840 : foreach(lc, root->join_info_list)
277 : {
278 42474 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
279 : int varno;
280 :
281 51000 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
282 8526 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
283 : {
284 8326 : RelOptInfo *rel = find_base_rel(root, varno);
285 :
286 8326 : rel->consider_param_startup = true;
287 : }
288 : }
289 325366 : }
290 :
291 : /*
292 : * set_base_rel_sizes
293 : * Set the size estimates (rows and widths) for each base-relation entry.
294 : * Also determine whether to consider parallel paths for base relations.
295 : *
296 : * We do this in a separate pass over the base rels so that rowcount
297 : * estimates are available for parameterized path generation, and also so
298 : * that each rel's consider_parallel flag is set correctly before we begin to
299 : * generate paths.
300 : */
301 : static void
302 325366 : set_base_rel_sizes(PlannerInfo *root)
303 : {
304 : Index rti;
305 :
306 1007720 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
307 : {
308 682388 : RelOptInfo *rel = root->simple_rel_array[rti];
309 : RangeTblEntry *rte;
310 :
311 : /* there may be empty slots corresponding to non-baserel RTEs */
312 682388 : if (rel == NULL)
313 161692 : continue;
314 :
315 : Assert(rel->relid == rti); /* sanity check on array */
316 :
317 : /* ignore RTEs that are "other rels" */
318 520696 : if (rel->reloptkind != RELOPT_BASEREL)
319 57410 : continue;
320 :
321 463286 : rte = root->simple_rte_array[rti];
322 :
323 : /*
324 : * If parallelism is allowable for this query in general, see whether
325 : * it's allowable for this rel in particular. We have to do this
326 : * before set_rel_size(), because (a) if this rel is an inheritance
327 : * parent, set_append_rel_size() will use and perhaps change the rel's
328 : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
329 : * goes ahead and makes paths immediately.
330 : */
331 463286 : if (root->glob->parallelModeOK)
332 369648 : set_rel_consider_parallel(root, rel, rte);
333 :
334 463286 : set_rel_size(root, rel, rti, rte);
335 : }
336 325332 : }
337 :
338 : /*
339 : * setup_simple_grouped_rels
340 : * For each simple relation, build a grouped simple relation if eager
341 : * aggregation is possible and if this relation can produce grouped paths.
342 : */
343 : static void
344 325332 : setup_simple_grouped_rels(PlannerInfo *root)
345 : {
346 : Index rti;
347 :
348 : /*
349 : * If there are no aggregate expressions or grouping expressions, eager
350 : * aggregation is not possible.
351 : */
352 325332 : if (root->agg_clause_list == NIL ||
353 658 : root->group_expr_list == NIL)
354 324740 : return;
355 :
356 5074 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
357 : {
358 4482 : RelOptInfo *rel = root->simple_rel_array[rti];
359 :
360 : /* there may be empty slots corresponding to non-baserel RTEs */
361 4482 : if (rel == NULL)
362 1414 : continue;
363 :
364 : Assert(rel->relid == rti); /* sanity check on array */
365 : Assert(IS_SIMPLE_REL(rel)); /* sanity check on rel */
366 :
367 3068 : (void) build_simple_grouped_rel(root, rel);
368 : }
369 : }
370 :
371 : /*
372 : * set_base_rel_pathlists
373 : * Finds all paths available for scanning each base-relation entry.
374 : * Sequential scan and any available indices are considered.
375 : * Each useful path is attached to its relation's 'pathlist' field.
376 : */
377 : static void
378 325332 : set_base_rel_pathlists(PlannerInfo *root)
379 : {
380 : Index rti;
381 :
382 1007684 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
383 : {
384 682352 : RelOptInfo *rel = root->simple_rel_array[rti];
385 :
386 : /* there may be empty slots corresponding to non-baserel RTEs */
387 682352 : if (rel == NULL)
388 161690 : continue;
389 :
390 : Assert(rel->relid == rti); /* sanity check on array */
391 :
392 : /* ignore RTEs that are "other rels" */
393 520662 : if (rel->reloptkind != RELOPT_BASEREL)
394 57410 : continue;
395 :
396 463252 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
397 : }
398 325332 : }
399 :
400 : /*
401 : * set_rel_size
402 : * Set size estimates for a base relation
403 : */
404 : static void
405 520386 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
406 : Index rti, RangeTblEntry *rte)
407 : {
408 983672 : if (rel->reloptkind == RELOPT_BASEREL &&
409 463286 : relation_excluded_by_constraints(root, rel, rte))
410 : {
411 : /*
412 : * We proved we don't need to scan the rel via constraint exclusion,
413 : * so set up a single dummy path for it. Here we only check this for
414 : * regular baserels; if it's an otherrel, CE was already checked in
415 : * set_append_rel_size().
416 : *
417 : * In this case, we go ahead and set up the relation's path right away
418 : * instead of leaving it for set_rel_pathlist to do. This is because
419 : * we don't have a convention for marking a rel as dummy except by
420 : * assigning a dummy path to it.
421 : */
422 616 : set_dummy_rel_pathlist(rel);
423 : }
424 519770 : else if (rte->inh)
425 : {
426 : /* It's an "append relation", process accordingly */
427 25374 : set_append_rel_size(root, rel, rti, rte);
428 : }
429 : else
430 : {
431 494396 : switch (rel->rtekind)
432 : {
433 407602 : case RTE_RELATION:
434 407602 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
435 : {
436 : /* Foreign table */
437 2438 : set_foreign_size(root, rel, rte);
438 : }
439 405164 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
440 : {
441 : /*
442 : * We could get here if asked to scan a partitioned table
443 : * with ONLY. In that case we shouldn't scan any of the
444 : * partitions, so mark it as a dummy rel.
445 : */
446 40 : set_dummy_rel_pathlist(rel);
447 : }
448 405124 : else if (rte->tablesample != NULL)
449 : {
450 : /* Sampled relation */
451 306 : set_tablesample_rel_size(root, rel, rte);
452 : }
453 : else
454 : {
455 : /* Plain relation */
456 404818 : set_plain_rel_size(root, rel, rte);
457 : }
458 407568 : break;
459 16760 : case RTE_SUBQUERY:
460 :
461 : /*
462 : * Subqueries don't support making a choice between
463 : * parameterized and unparameterized paths, so just go ahead
464 : * and build their paths immediately.
465 : */
466 16760 : set_subquery_pathlist(root, rel, rti, rte);
467 16760 : break;
468 51268 : case RTE_FUNCTION:
469 51268 : set_function_size_estimates(root, rel);
470 51268 : break;
471 626 : case RTE_TABLEFUNC:
472 626 : set_tablefunc_size_estimates(root, rel);
473 626 : break;
474 8264 : case RTE_VALUES:
475 8264 : set_values_size_estimates(root, rel);
476 8264 : break;
477 5190 : case RTE_CTE:
478 :
479 : /*
480 : * CTEs don't support making a choice between parameterized
481 : * and unparameterized paths, so just go ahead and build their
482 : * paths immediately.
483 : */
484 5190 : if (rte->self_reference)
485 932 : set_worktable_pathlist(root, rel, rte);
486 : else
487 4258 : set_cte_pathlist(root, rel, rte);
488 5190 : break;
489 478 : case RTE_NAMEDTUPLESTORE:
490 : /* Might as well just build the path immediately */
491 478 : set_namedtuplestore_pathlist(root, rel, rte);
492 478 : break;
493 4208 : case RTE_RESULT:
494 : /* Might as well just build the path immediately */
495 4208 : set_result_pathlist(root, rel, rte);
496 4208 : break;
497 0 : default:
498 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
499 : break;
500 : }
501 : }
502 :
503 : /*
504 : * We insist that all non-dummy rels have a nonzero rowcount estimate.
505 : */
506 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
507 520350 : }
508 :
509 : /*
510 : * set_rel_pathlist
511 : * Build access paths for a base relation
512 : */
513 : static void
514 520488 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
515 : Index rti, RangeTblEntry *rte)
516 : {
517 520488 : if (IS_DUMMY_REL(rel))
518 : {
519 : /* We already proved the relation empty, so nothing more to do */
520 : }
521 519274 : else if (rte->inh)
522 : {
523 : /* It's an "append relation", process accordingly */
524 25078 : set_append_rel_pathlist(root, rel, rti, rte);
525 : }
526 : else
527 : {
528 494196 : switch (rel->rtekind)
529 : {
530 407528 : case RTE_RELATION:
531 407528 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
532 : {
533 : /* Foreign table */
534 2434 : set_foreign_pathlist(root, rel, rte);
535 : }
536 405094 : else if (rte->tablesample != NULL)
537 : {
538 : /* Sampled relation */
539 306 : set_tablesample_rel_pathlist(root, rel, rte);
540 : }
541 : else
542 : {
543 : /* Plain relation */
544 404788 : set_plain_rel_pathlist(root, rel, rte);
545 : }
546 407528 : break;
547 16634 : case RTE_SUBQUERY:
548 : /* Subquery --- fully handled during set_rel_size */
549 16634 : break;
550 51268 : case RTE_FUNCTION:
551 : /* RangeFunction */
552 51268 : set_function_pathlist(root, rel, rte);
553 51268 : break;
554 626 : case RTE_TABLEFUNC:
555 : /* Table Function */
556 626 : set_tablefunc_pathlist(root, rel, rte);
557 626 : break;
558 8264 : case RTE_VALUES:
559 : /* Values list */
560 8264 : set_values_pathlist(root, rel, rte);
561 8264 : break;
562 5190 : case RTE_CTE:
563 : /* CTE reference --- fully handled during set_rel_size */
564 5190 : break;
565 478 : case RTE_NAMEDTUPLESTORE:
566 : /* tuplestore reference --- fully handled during set_rel_size */
567 478 : break;
568 4208 : case RTE_RESULT:
569 : /* simple Result --- fully handled during set_rel_size */
570 4208 : break;
571 0 : default:
572 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
573 : break;
574 : }
575 : }
576 :
577 : /*
578 : * Allow a plugin to editorialize on the set of Paths for this base
579 : * relation. It could add new paths (such as CustomPaths) by calling
580 : * add_path(), or add_partial_path() if parallel aware. It could also
581 : * delete or modify paths added by the core code.
582 : */
583 520488 : if (set_rel_pathlist_hook)
584 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
585 :
586 : /*
587 : * If this is a baserel, we should normally consider gathering any partial
588 : * paths we may have created for it. We have to do this after calling the
589 : * set_rel_pathlist_hook, else it cannot add partial paths to be included
590 : * here.
591 : *
592 : * However, if this is an inheritance child, skip it. Otherwise, we could
593 : * end up with a very large number of gather nodes, each trying to grab
594 : * its own pool of workers. Instead, we'll consider gathering partial
595 : * paths for the parent appendrel.
596 : *
597 : * Also, if this is the topmost scan/join rel, we postpone gathering until
598 : * the final scan/join targetlist is available (see grouping_planner).
599 : */
600 520488 : if (rel->reloptkind == RELOPT_BASEREL &&
601 463252 : !bms_equal(rel->relids, root->all_query_rels))
602 238398 : generate_useful_gather_paths(root, rel, false);
603 :
604 : /* Now find the cheapest of the paths for this rel */
605 520488 : set_cheapest(rel);
606 :
607 : /*
608 : * If a grouped relation for this rel exists, build partial aggregation
609 : * paths for it.
610 : *
611 : * Note that this can only happen after we've called set_cheapest() for
612 : * this base rel, because we need its cheapest paths.
613 : */
614 520488 : set_grouped_rel_pathlist(root, rel);
615 :
616 : #ifdef OPTIMIZER_DEBUG
617 : pprint(rel);
618 : #endif
619 520488 : }
620 :
621 : /*
622 : * set_plain_rel_size
623 : * Set size estimates for a plain relation (no subquery, no inheritance)
624 : */
625 : static void
626 404818 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
627 : {
628 : /*
629 : * Test any partial indexes of rel for applicability. We must do this
630 : * first since partial unique indexes can affect size estimates.
631 : */
632 404818 : check_index_predicates(root, rel);
633 :
634 : /* Mark rel with estimated output rows, width, etc */
635 404818 : set_baserel_size_estimates(root, rel);
636 404788 : }
637 :
638 : /*
639 : * If this relation could possibly be scanned from within a worker, then set
640 : * its consider_parallel flag.
641 : */
642 : static void
643 412206 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
644 : RangeTblEntry *rte)
645 : {
646 : /*
647 : * The flag has previously been initialized to false, so we can just
648 : * return if it becomes clear that we can't safely set it.
649 : */
650 : Assert(!rel->consider_parallel);
651 :
652 : /* Don't call this if parallelism is disallowed for the entire query. */
653 : Assert(root->glob->parallelModeOK);
654 :
655 : /* This should only be called for baserels and appendrel children. */
656 : Assert(IS_SIMPLE_REL(rel));
657 :
658 : /* Assorted checks based on rtekind. */
659 412206 : switch (rte->rtekind)
660 : {
661 354658 : case RTE_RELATION:
662 :
663 : /*
664 : * Currently, parallel workers can't access the leader's temporary
665 : * tables. We could possibly relax this if we wrote all of its
666 : * local buffers at the start of the query and made no changes
667 : * thereafter (maybe we could allow hint bit changes), and if we
668 : * taught the workers to read them. Writing a large number of
669 : * temporary buffers could be expensive, though, and we don't have
670 : * the rest of the necessary infrastructure right now anyway. So
671 : * for now, bail out if we see a temporary table.
672 : */
673 354658 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
674 8468 : return;
675 :
676 : /*
677 : * Table sampling can be pushed down to workers if the sample
678 : * function and its arguments are safe.
679 : */
680 346190 : if (rte->tablesample != NULL)
681 : {
682 330 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
683 :
684 330 : if (proparallel != PROPARALLEL_SAFE)
685 36 : return;
686 294 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
687 12 : return;
688 : }
689 :
690 : /*
691 : * Ask FDWs whether they can support performing a ForeignScan
692 : * within a worker. Most often, the answer will be no. For
693 : * example, if the nature of the FDW is such that it opens a TCP
694 : * connection with a remote server, each parallel worker would end
695 : * up with a separate connection, and these connections might not
696 : * be appropriately coordinated between workers and the leader.
697 : */
698 346142 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
699 : {
700 : Assert(rel->fdwroutine);
701 1552 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
702 1480 : return;
703 72 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
704 0 : return;
705 : }
706 :
707 : /*
708 : * There are additional considerations for appendrels, which we'll
709 : * deal with in set_append_rel_size and set_append_rel_pathlist.
710 : * For now, just set consider_parallel based on the rel's own
711 : * quals and targetlist.
712 : */
713 344662 : break;
714 :
715 18216 : case RTE_SUBQUERY:
716 :
717 : /*
718 : * There's no intrinsic problem with scanning a subquery-in-FROM
719 : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
720 : * If the subquery doesn't happen to have any parallel-safe paths,
721 : * then flagging it as consider_parallel won't change anything,
722 : * but that's true for plain tables, too. We must set
723 : * consider_parallel based on the rel's own quals and targetlist,
724 : * so that if a subquery path is parallel-safe but the quals and
725 : * projection we're sticking onto it are not, we correctly mark
726 : * the SubqueryScanPath as not parallel-safe. (Note that
727 : * set_subquery_pathlist() might push some of these quals down
728 : * into the subquery itself, but that doesn't change anything.)
729 : *
730 : * We can't push sub-select containing LIMIT/OFFSET to workers as
731 : * there is no guarantee that the row order will be fully
732 : * deterministic, and applying LIMIT/OFFSET will lead to
733 : * inconsistent results at the top-level. (In some cases, where
734 : * the result is ordered, we could relax this restriction. But it
735 : * doesn't currently seem worth expending extra effort to do so.)
736 : */
737 : {
738 18216 : Query *subquery = castNode(Query, rte->subquery);
739 :
740 18216 : if (limit_needed(subquery))
741 508 : return;
742 : }
743 17708 : break;
744 :
745 0 : case RTE_JOIN:
746 : /* Shouldn't happen; we're only considering baserels here. */
747 : Assert(false);
748 0 : return;
749 :
750 27416 : case RTE_FUNCTION:
751 : /* Check for parallel-restricted functions. */
752 27416 : if (!is_parallel_safe(root, (Node *) rte->functions))
753 13034 : return;
754 14382 : break;
755 :
756 626 : case RTE_TABLEFUNC:
757 : /* not parallel safe */
758 626 : return;
759 :
760 2834 : case RTE_VALUES:
761 : /* Check for parallel-restricted functions. */
762 2834 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
763 12 : return;
764 2822 : break;
765 :
766 4228 : case RTE_CTE:
767 :
768 : /*
769 : * CTE tuplestores aren't shared among parallel workers, so we
770 : * force all CTE scans to happen in the leader. Also, populating
771 : * the CTE would require executing a subplan that's not available
772 : * in the worker, might be parallel-restricted, and must get
773 : * executed only once.
774 : */
775 4228 : return;
776 :
777 450 : case RTE_NAMEDTUPLESTORE:
778 :
779 : /*
780 : * tuplestore cannot be shared, at least without more
781 : * infrastructure to support that.
782 : */
783 450 : return;
784 :
785 3778 : case RTE_RESULT:
786 : /* RESULT RTEs, in themselves, are no problem. */
787 3778 : break;
788 0 : case RTE_GROUP:
789 : /* Shouldn't happen; we're only considering baserels here. */
790 : Assert(false);
791 0 : return;
792 : }
793 :
794 : /*
795 : * If there's anything in baserestrictinfo that's parallel-restricted, we
796 : * give up on parallelizing access to this relation. We could consider
797 : * instead postponing application of the restricted quals until we're
798 : * above all the parallelism in the plan tree, but it's not clear that
799 : * that would be a win in very many cases, and it might be tricky to make
800 : * outer join clauses work correctly. It would likely break equivalence
801 : * classes, too.
802 : */
803 383352 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
804 27274 : return;
805 :
806 : /*
807 : * Likewise, if the relation's outputs are not parallel-safe, give up.
808 : * (Usually, they're just Vars, but sometimes they're not.)
809 : */
810 356078 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
811 18 : return;
812 :
813 : /* We have a winner. */
814 356060 : rel->consider_parallel = true;
815 : }
816 :
817 : /*
818 : * set_plain_rel_pathlist
819 : * Build access paths for a plain relation (no subquery, no inheritance)
820 : */
821 : static void
822 404788 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
823 : {
824 : Relids required_outer;
825 :
826 : /*
827 : * We don't support pushing join clauses into the quals of a seqscan, but
828 : * it could still have required parameterization due to LATERAL refs in
829 : * its tlist.
830 : */
831 404788 : required_outer = rel->lateral_relids;
832 :
833 : /*
834 : * Consider TID scans.
835 : *
836 : * If create_tidscan_paths returns true, then a TID scan path is forced.
837 : * This happens when rel->baserestrictinfo contains CurrentOfExpr, because
838 : * the executor can't handle any other type of path for such queries.
839 : * Hence, we return without adding any other paths.
840 : */
841 404788 : if (create_tidscan_paths(root, rel))
842 404 : return;
843 :
844 : /* Consider sequential scan */
845 404384 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
846 :
847 : /* If appropriate, consider parallel sequential scan */
848 404384 : if (rel->consider_parallel && required_outer == NULL)
849 303706 : create_plain_partial_paths(root, rel);
850 :
851 : /* Consider index scans */
852 404384 : create_index_paths(root, rel);
853 : }
854 :
855 : /*
856 : * create_plain_partial_paths
857 : * Build partial access paths for parallel scan of a plain relation
858 : */
859 : static void
860 303706 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
861 : {
862 : int parallel_workers;
863 :
864 303706 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
865 : max_parallel_workers_per_gather);
866 :
867 : /* If any limit was set to zero, the user doesn't want a parallel scan. */
868 303706 : if (parallel_workers <= 0)
869 276318 : return;
870 :
871 : /* Add an unordered partial path based on a parallel sequential scan. */
872 27388 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
873 : }
874 :
875 : /*
876 : * set_tablesample_rel_size
877 : * Set size estimates for a sampled relation
878 : */
879 : static void
880 306 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
881 : {
882 306 : TableSampleClause *tsc = rte->tablesample;
883 : TsmRoutine *tsm;
884 : BlockNumber pages;
885 : double tuples;
886 :
887 : /*
888 : * Test any partial indexes of rel for applicability. We must do this
889 : * first since partial unique indexes can affect size estimates.
890 : */
891 306 : check_index_predicates(root, rel);
892 :
893 : /*
894 : * Call the sampling method's estimation function to estimate the number
895 : * of pages it will read and the number of tuples it will return. (Note:
896 : * we assume the function returns sane values.)
897 : */
898 306 : tsm = GetTsmRoutine(tsc->tsmhandler);
899 306 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
900 : &pages, &tuples);
901 :
902 : /*
903 : * For the moment, because we will only consider a SampleScan path for the
904 : * rel, it's okay to just overwrite the pages and tuples estimates for the
905 : * whole relation. If we ever consider multiple path types for sampled
906 : * rels, we'll need more complication.
907 : */
908 306 : rel->pages = pages;
909 306 : rel->tuples = tuples;
910 :
911 : /* Mark rel with estimated output rows, width, etc */
912 306 : set_baserel_size_estimates(root, rel);
913 306 : }
914 :
915 : /*
916 : * set_tablesample_rel_pathlist
917 : * Build access paths for a sampled relation
918 : */
919 : static void
920 306 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
921 : {
922 : Relids required_outer;
923 : Path *path;
924 :
925 : /*
926 : * We don't support pushing join clauses into the quals of a samplescan,
927 : * but it could still have required parameterization due to LATERAL refs
928 : * in its tlist or TABLESAMPLE arguments.
929 : */
930 306 : required_outer = rel->lateral_relids;
931 :
932 : /* Consider sampled scan */
933 306 : path = create_samplescan_path(root, rel, required_outer);
934 :
935 : /*
936 : * If the sampling method does not support repeatable scans, we must avoid
937 : * plans that would scan the rel multiple times. Ideally, we'd simply
938 : * avoid putting the rel on the inside of a nestloop join; but adding such
939 : * a consideration to the planner seems like a great deal of complication
940 : * to support an uncommon usage of second-rate sampling methods. Instead,
941 : * if there is a risk that the query might perform an unsafe join, just
942 : * wrap the SampleScan in a Materialize node. We can check for joins by
943 : * counting the membership of all_query_rels (note that this correctly
944 : * counts inheritance trees as single rels). If we're inside a subquery,
945 : * we can't easily check whether a join might occur in the outer query, so
946 : * just assume one is possible.
947 : *
948 : * GetTsmRoutine is relatively expensive compared to the other tests here,
949 : * so check repeatable_across_scans last, even though that's a bit odd.
950 : */
951 586 : if ((root->query_level > 1 ||
952 280 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
953 98 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
954 : {
955 8 : path = (Path *) create_material_path(rel, path);
956 : }
957 :
958 306 : add_path(rel, path);
959 :
960 : /* For the moment, at least, there are no other paths to consider */
961 306 : }
962 :
963 : /*
964 : * set_foreign_size
965 : * Set size estimates for a foreign table RTE
966 : */
967 : static void
968 2438 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
969 : {
970 : /* Mark rel with estimated output rows, width, etc */
971 2438 : set_foreign_size_estimates(root, rel);
972 :
973 : /* Let FDW adjust the size estimates, if it can */
974 2438 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
975 :
976 : /* ... but do not let it set the rows estimate to zero */
977 2434 : rel->rows = clamp_row_est(rel->rows);
978 :
979 : /*
980 : * Also, make sure rel->tuples is not insane relative to rel->rows.
981 : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
982 : * FDW doesn't do anything to replace that.
983 : */
984 2434 : rel->tuples = Max(rel->tuples, rel->rows);
985 2434 : }
986 :
987 : /*
988 : * set_foreign_pathlist
989 : * Build access paths for a foreign table RTE
990 : */
991 : static void
992 2434 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
993 : {
994 : /* Call the FDW's GetForeignPaths function to generate path(s) */
995 2434 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
996 2434 : }
997 :
998 : /*
999 : * set_append_rel_size
1000 : * Set size estimates for a simple "append relation"
1001 : *
1002 : * The passed-in rel and RTE represent the entire append relation. The
1003 : * relation's contents are computed by appending together the output of the
1004 : * individual member relations. Note that in the non-partitioned inheritance
1005 : * case, the first member relation is actually the same table as is mentioned
1006 : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
1007 : * a good thing because their outputs are not the same size.
1008 : */
1009 : static void
1010 25374 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
1011 : Index rti, RangeTblEntry *rte)
1012 : {
1013 25374 : int parentRTindex = rti;
1014 : bool has_live_children;
1015 : double parent_tuples;
1016 : double parent_rows;
1017 : double parent_size;
1018 : double *parent_attrsizes;
1019 : int nattrs;
1020 : ListCell *l;
1021 :
1022 : /* Guard against stack overflow due to overly deep inheritance tree. */
1023 25374 : check_stack_depth();
1024 :
1025 : Assert(IS_SIMPLE_REL(rel));
1026 :
1027 : /*
1028 : * If this is a partitioned baserel, set the consider_partitionwise_join
1029 : * flag; currently, we only consider partitionwise joins with the baserel
1030 : * if its targetlist doesn't contain a whole-row Var.
1031 : */
1032 25374 : if (enable_partitionwise_join &&
1033 4982 : rel->reloptkind == RELOPT_BASEREL &&
1034 3962 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
1035 3962 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
1036 3886 : rel->consider_partitionwise_join = true;
1037 :
1038 : /*
1039 : * Initialize to compute size estimates for whole append relation.
1040 : *
1041 : * We handle tuples estimates by setting "tuples" to the total number of
1042 : * tuples accumulated from each live child, rather than using "rows".
1043 : * Although an appendrel itself doesn't directly enforce any quals, its
1044 : * child relations may. Therefore, setting "tuples" equal to "rows" for
1045 : * an appendrel isn't always appropriate, and can lead to inaccurate cost
1046 : * estimates. For example, when estimating the number of distinct values
1047 : * from an appendrel, we would be unable to adjust the estimate based on
1048 : * the restriction selectivity (see estimate_num_groups).
1049 : *
1050 : * We handle width estimates by weighting the widths of different child
1051 : * rels proportionally to their number of rows. This is sensible because
1052 : * the use of width estimates is mainly to compute the total relation
1053 : * "footprint" if we have to sort or hash it. To do this, we sum the
1054 : * total equivalent size (in "double" arithmetic) and then divide by the
1055 : * total rowcount estimate. This is done separately for the total rel
1056 : * width and each attribute.
1057 : *
1058 : * Note: if you consider changing this logic, beware that child rels could
1059 : * have zero rows and/or width, if they were excluded by constraints.
1060 : */
1061 25374 : has_live_children = false;
1062 25374 : parent_tuples = 0;
1063 25374 : parent_rows = 0;
1064 25374 : parent_size = 0;
1065 25374 : nattrs = rel->max_attr - rel->min_attr + 1;
1066 25374 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1067 :
1068 134196 : foreach(l, root->append_rel_list)
1069 : {
1070 108824 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1071 : int childRTindex;
1072 : RangeTblEntry *childRTE;
1073 : RelOptInfo *childrel;
1074 : List *childrinfos;
1075 : ListCell *parentvars;
1076 : ListCell *childvars;
1077 : ListCell *lc;
1078 :
1079 : /* append_rel_list contains all append rels; ignore others */
1080 108824 : if (appinfo->parent_relid != parentRTindex)
1081 51862 : continue;
1082 :
1083 57304 : childRTindex = appinfo->child_relid;
1084 57304 : childRTE = root->simple_rte_array[childRTindex];
1085 :
1086 : /*
1087 : * The child rel's RelOptInfo was already created during
1088 : * add_other_rels_to_query.
1089 : */
1090 57304 : childrel = find_base_rel(root, childRTindex);
1091 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1092 :
1093 : /* We may have already proven the child to be dummy. */
1094 57304 : if (IS_DUMMY_REL(childrel))
1095 18 : continue;
1096 :
1097 : /*
1098 : * We have to copy the parent's targetlist and quals to the child,
1099 : * with appropriate substitution of variables. However, the
1100 : * baserestrictinfo quals were already copied/substituted when the
1101 : * child RelOptInfo was built. So we don't need any additional setup
1102 : * before applying constraint exclusion.
1103 : */
1104 57286 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1105 : {
1106 : /*
1107 : * This child need not be scanned, so we can omit it from the
1108 : * appendrel.
1109 : */
1110 186 : set_dummy_rel_pathlist(childrel);
1111 186 : continue;
1112 : }
1113 :
1114 : /*
1115 : * Constraint exclusion failed, so copy the parent's join quals and
1116 : * targetlist to the child, with appropriate variable substitutions.
1117 : *
1118 : * We skip join quals that came from above outer joins that can null
1119 : * this rel, since they would be of no value while generating paths
1120 : * for the child. This saves some effort while processing the child
1121 : * rel, and it also avoids an implementation restriction in
1122 : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1123 : */
1124 57100 : childrinfos = NIL;
1125 70162 : foreach(lc, rel->joininfo)
1126 : {
1127 13062 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1128 :
1129 13062 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1130 10740 : childrinfos = lappend(childrinfos,
1131 10740 : adjust_appendrel_attrs(root,
1132 : (Node *) rinfo,
1133 : 1, &appinfo));
1134 : }
1135 57100 : childrel->joininfo = childrinfos;
1136 :
1137 : /*
1138 : * Now for the child's targetlist.
1139 : *
1140 : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1141 : * expressions, which otherwise would not occur in a rel's targetlist.
1142 : * Code that might be looking at an appendrel child must cope with
1143 : * such. (Normally, a rel's targetlist would only include Vars and
1144 : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1145 : * fields of childrel->reltarget; not clear if that would be useful.
1146 : */
1147 114200 : childrel->reltarget->exprs = (List *)
1148 57100 : adjust_appendrel_attrs(root,
1149 57100 : (Node *) rel->reltarget->exprs,
1150 : 1, &appinfo);
1151 :
1152 : /*
1153 : * We have to make child entries in the EquivalenceClass data
1154 : * structures as well. This is needed either if the parent
1155 : * participates in some eclass joins (because we will want to consider
1156 : * inner-indexscan joins on the individual children) or if the parent
1157 : * has useful pathkeys (because we should try to build MergeAppend
1158 : * paths that produce those sort orderings).
1159 : */
1160 57100 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1161 35326 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1162 57100 : childrel->has_eclass_joins = rel->has_eclass_joins;
1163 :
1164 : /*
1165 : * Note: we could compute appropriate attr_needed data for the child's
1166 : * variables, by transforming the parent's attr_needed through the
1167 : * translated_vars mapping. However, currently there's no need
1168 : * because attr_needed is only examined for base relations not
1169 : * otherrels. So we just leave the child's attr_needed empty.
1170 : */
1171 :
1172 : /*
1173 : * If we consider partitionwise joins with the parent rel, do the same
1174 : * for partitioned child rels.
1175 : *
1176 : * Note: here we abuse the consider_partitionwise_join flag by setting
1177 : * it for child rels that are not themselves partitioned. We do so to
1178 : * tell try_partitionwise_join() that the child rel is sufficiently
1179 : * valid to be used as a per-partition input, even if it later gets
1180 : * proven to be dummy. (It's not usable until we've set up the
1181 : * reltarget and EC entries, which we just did.)
1182 : */
1183 57100 : if (rel->consider_partitionwise_join)
1184 13148 : childrel->consider_partitionwise_join = true;
1185 :
1186 : /*
1187 : * If parallelism is allowable for this query in general, see whether
1188 : * it's allowable for this childrel in particular. But if we've
1189 : * already decided the appendrel is not parallel-safe as a whole,
1190 : * there's no point in considering parallelism for this child. For
1191 : * consistency, do this before calling set_rel_size() for the child.
1192 : */
1193 57100 : if (root->glob->parallelModeOK && rel->consider_parallel)
1194 42558 : set_rel_consider_parallel(root, childrel, childRTE);
1195 :
1196 : /*
1197 : * Compute the child's size.
1198 : */
1199 57100 : set_rel_size(root, childrel, childRTindex, childRTE);
1200 :
1201 : /*
1202 : * It is possible that constraint exclusion detected a contradiction
1203 : * within a child subquery, even though we didn't prove one above. If
1204 : * so, we can skip this child.
1205 : */
1206 57098 : if (IS_DUMMY_REL(childrel))
1207 138 : continue;
1208 :
1209 : /* We have at least one live child. */
1210 56960 : has_live_children = true;
1211 :
1212 : /*
1213 : * If any live child is not parallel-safe, treat the whole appendrel
1214 : * as not parallel-safe. In future we might be able to generate plans
1215 : * in which some children are farmed out to workers while others are
1216 : * not; but we don't have that today, so it's a waste to consider
1217 : * partial paths anywhere in the appendrel unless it's all safe.
1218 : * (Child rels visited before this one will be unmarked in
1219 : * set_append_rel_pathlist().)
1220 : */
1221 56960 : if (!childrel->consider_parallel)
1222 15234 : rel->consider_parallel = false;
1223 :
1224 : /*
1225 : * Accumulate size information from each live child.
1226 : */
1227 : Assert(childrel->rows > 0);
1228 :
1229 56960 : parent_tuples += childrel->tuples;
1230 56960 : parent_rows += childrel->rows;
1231 56960 : parent_size += childrel->reltarget->width * childrel->rows;
1232 :
1233 : /*
1234 : * Accumulate per-column estimates too. We need not do anything for
1235 : * PlaceHolderVars in the parent list. If child expression isn't a
1236 : * Var, or we didn't record a width estimate for it, we have to fall
1237 : * back on a datatype-based estimate.
1238 : *
1239 : * By construction, child's targetlist is 1-to-1 with parent's.
1240 : */
1241 185960 : forboth(parentvars, rel->reltarget->exprs,
1242 : childvars, childrel->reltarget->exprs)
1243 : {
1244 129000 : Var *parentvar = (Var *) lfirst(parentvars);
1245 129000 : Node *childvar = (Node *) lfirst(childvars);
1246 :
1247 129000 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1248 : {
1249 116112 : int pndx = parentvar->varattno - rel->min_attr;
1250 116112 : int32 child_width = 0;
1251 :
1252 116112 : if (IsA(childvar, Var) &&
1253 111320 : ((Var *) childvar)->varno == childrel->relid)
1254 : {
1255 111254 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1256 :
1257 111254 : child_width = childrel->attr_widths[cndx];
1258 : }
1259 116112 : if (child_width <= 0)
1260 4858 : child_width = get_typavgwidth(exprType(childvar),
1261 : exprTypmod(childvar));
1262 : Assert(child_width > 0);
1263 116112 : parent_attrsizes[pndx] += child_width * childrel->rows;
1264 : }
1265 : }
1266 : }
1267 :
1268 25372 : if (has_live_children)
1269 : {
1270 : /*
1271 : * Save the finished size estimates.
1272 : */
1273 : int i;
1274 :
1275 : Assert(parent_rows > 0);
1276 25078 : rel->tuples = parent_tuples;
1277 25078 : rel->rows = parent_rows;
1278 25078 : rel->reltarget->width = rint(parent_size / parent_rows);
1279 234358 : for (i = 0; i < nattrs; i++)
1280 209280 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1281 :
1282 : /*
1283 : * Note that we leave rel->pages as zero; this is important to avoid
1284 : * double-counting the appendrel tree in total_table_pages.
1285 : */
1286 : }
1287 : else
1288 : {
1289 : /*
1290 : * All children were excluded by constraints, so mark the whole
1291 : * appendrel dummy. We must do this in this phase so that the rel's
1292 : * dummy-ness is visible when we generate paths for other rels.
1293 : */
1294 294 : set_dummy_rel_pathlist(rel);
1295 : }
1296 :
1297 25372 : pfree(parent_attrsizes);
1298 25372 : }
1299 :
1300 : /*
1301 : * set_append_rel_pathlist
1302 : * Build access paths for an "append relation"
1303 : */
1304 : static void
1305 25078 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1306 : Index rti, RangeTblEntry *rte)
1307 : {
1308 25078 : int parentRTindex = rti;
1309 25078 : List *live_childrels = NIL;
1310 : ListCell *l;
1311 :
1312 : /*
1313 : * Generate access paths for each member relation, and remember the
1314 : * non-dummy children.
1315 : */
1316 133486 : foreach(l, root->append_rel_list)
1317 : {
1318 108408 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1319 : int childRTindex;
1320 : RangeTblEntry *childRTE;
1321 : RelOptInfo *childrel;
1322 :
1323 : /* append_rel_list contains all append rels; ignore others */
1324 108408 : if (appinfo->parent_relid != parentRTindex)
1325 51172 : continue;
1326 :
1327 : /* Re-locate the child RTE and RelOptInfo */
1328 57236 : childRTindex = appinfo->child_relid;
1329 57236 : childRTE = root->simple_rte_array[childRTindex];
1330 57236 : childrel = root->simple_rel_array[childRTindex];
1331 :
1332 : /*
1333 : * If set_append_rel_size() decided the parent appendrel was
1334 : * parallel-unsafe at some point after visiting this child rel, we
1335 : * need to propagate the unsafety marking down to the child, so that
1336 : * we don't generate useless partial paths for it.
1337 : */
1338 57236 : if (!rel->consider_parallel)
1339 15430 : childrel->consider_parallel = false;
1340 :
1341 : /*
1342 : * Compute the child's access paths.
1343 : */
1344 57236 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1345 :
1346 : /*
1347 : * If child is dummy, ignore it.
1348 : */
1349 57236 : if (IS_DUMMY_REL(childrel))
1350 276 : continue;
1351 :
1352 : /*
1353 : * Child is live, so add it to the live_childrels list for use below.
1354 : */
1355 56960 : live_childrels = lappend(live_childrels, childrel);
1356 : }
1357 :
1358 : /* Add paths to the append relation. */
1359 25078 : add_paths_to_append_rel(root, rel, live_childrels);
1360 25078 : }
1361 :
1362 : /*
1363 : * set_grouped_rel_pathlist
1364 : * If a grouped relation for the given 'rel' exists, build partial
1365 : * aggregation paths for it.
1366 : */
1367 : static void
1368 520488 : set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel)
1369 : {
1370 : RelOptInfo *grouped_rel;
1371 :
1372 : /*
1373 : * If there are no aggregate expressions or grouping expressions, eager
1374 : * aggregation is not possible.
1375 : */
1376 520488 : if (root->agg_clause_list == NIL ||
1377 3332 : root->group_expr_list == NIL)
1378 517420 : return;
1379 :
1380 : /* Add paths to the grouped base relation if one exists. */
1381 3068 : grouped_rel = rel->grouped_rel;
1382 3068 : if (grouped_rel)
1383 : {
1384 : Assert(IS_GROUPED_REL(grouped_rel));
1385 :
1386 586 : generate_grouped_paths(root, grouped_rel, rel);
1387 586 : set_cheapest(grouped_rel);
1388 : }
1389 : }
1390 :
1391 :
1392 : /*
1393 : * add_paths_to_append_rel
1394 : * Generate paths for the given append relation given the set of non-dummy
1395 : * child rels.
1396 : *
1397 : * The function collects all parameterizations and orderings supported by the
1398 : * non-dummy children. For every such parameterization or ordering, it creates
1399 : * an append path collecting one path from each non-dummy child with given
1400 : * parameterization or ordering. Similarly it collects partial paths from
1401 : * non-dummy children to create partial append paths.
1402 : */
1403 : void
1404 46178 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1405 : List *live_childrels)
1406 : {
1407 46178 : List *subpaths = NIL;
1408 46178 : bool subpaths_valid = true;
1409 46178 : List *startup_subpaths = NIL;
1410 46178 : bool startup_subpaths_valid = true;
1411 46178 : List *partial_subpaths = NIL;
1412 46178 : List *pa_partial_subpaths = NIL;
1413 46178 : List *pa_nonpartial_subpaths = NIL;
1414 46178 : bool partial_subpaths_valid = true;
1415 : bool pa_subpaths_valid;
1416 46178 : List *all_child_pathkeys = NIL;
1417 46178 : List *all_child_outers = NIL;
1418 : ListCell *l;
1419 46178 : double partial_rows = -1;
1420 :
1421 : /* If appropriate, consider parallel append */
1422 46178 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1423 :
1424 : /*
1425 : * For every non-dummy child, remember the cheapest path. Also, identify
1426 : * all pathkeys (orderings) and parameterizations (required_outer sets)
1427 : * available for the non-dummy member relations.
1428 : */
1429 147794 : foreach(l, live_childrels)
1430 : {
1431 101616 : RelOptInfo *childrel = lfirst(l);
1432 : ListCell *lcp;
1433 101616 : Path *cheapest_partial_path = NULL;
1434 :
1435 : /*
1436 : * If child has an unparameterized cheapest-total path, add that to
1437 : * the unparameterized Append path we are constructing for the parent.
1438 : * If not, there's no workable unparameterized path.
1439 : *
1440 : * With partitionwise aggregates, the child rel's pathlist may be
1441 : * empty, so don't assume that a path exists here.
1442 : */
1443 101616 : if (childrel->pathlist != NIL &&
1444 101616 : childrel->cheapest_total_path->param_info == NULL)
1445 100884 : accumulate_append_subpath(childrel->cheapest_total_path,
1446 : &subpaths, NULL);
1447 : else
1448 732 : subpaths_valid = false;
1449 :
1450 : /*
1451 : * When the planner is considering cheap startup plans, we'll also
1452 : * collect all the cheapest_startup_paths (if set) and build an
1453 : * AppendPath containing those as subpaths.
1454 : */
1455 101616 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1456 1622 : {
1457 : Path *cheapest_path;
1458 :
1459 : /*
1460 : * With an indication of how many tuples the query should provide,
1461 : * the optimizer tries to choose the path optimal for that
1462 : * specific number of tuples.
1463 : */
1464 1622 : if (root->tuple_fraction > 0.0)
1465 : cheapest_path =
1466 1622 : get_cheapest_fractional_path(childrel,
1467 : root->tuple_fraction);
1468 : else
1469 0 : cheapest_path = childrel->cheapest_startup_path;
1470 :
1471 : /* cheapest_startup_path must not be a parameterized path. */
1472 : Assert(cheapest_path->param_info == NULL);
1473 1622 : accumulate_append_subpath(cheapest_path,
1474 : &startup_subpaths,
1475 : NULL);
1476 : }
1477 : else
1478 99994 : startup_subpaths_valid = false;
1479 :
1480 :
1481 : /* Same idea, but for a partial plan. */
1482 101616 : if (childrel->partial_pathlist != NIL)
1483 : {
1484 63330 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1485 63330 : accumulate_append_subpath(cheapest_partial_path,
1486 : &partial_subpaths, NULL);
1487 : }
1488 : else
1489 38286 : partial_subpaths_valid = false;
1490 :
1491 : /*
1492 : * Same idea, but for a parallel append mixing partial and non-partial
1493 : * paths.
1494 : */
1495 101616 : if (pa_subpaths_valid)
1496 : {
1497 76532 : Path *nppath = NULL;
1498 :
1499 : nppath =
1500 76532 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1501 :
1502 76532 : if (cheapest_partial_path == NULL && nppath == NULL)
1503 : {
1504 : /* Neither a partial nor a parallel-safe path? Forget it. */
1505 546 : pa_subpaths_valid = false;
1506 : }
1507 75986 : else if (nppath == NULL ||
1508 62880 : (cheapest_partial_path != NULL &&
1509 62880 : cheapest_partial_path->total_cost < nppath->total_cost))
1510 : {
1511 : /* Partial path is cheaper or the only option. */
1512 : Assert(cheapest_partial_path != NULL);
1513 62740 : accumulate_append_subpath(cheapest_partial_path,
1514 : &pa_partial_subpaths,
1515 : &pa_nonpartial_subpaths);
1516 : }
1517 : else
1518 : {
1519 : /*
1520 : * Either we've got only a non-partial path, or we think that
1521 : * a single backend can execute the best non-partial path
1522 : * faster than all the parallel backends working together can
1523 : * execute the best partial path.
1524 : *
1525 : * It might make sense to be more aggressive here. Even if
1526 : * the best non-partial path is more expensive than the best
1527 : * partial path, it could still be better to choose the
1528 : * non-partial path if there are several such paths that can
1529 : * be given to different workers. For now, we don't try to
1530 : * figure that out.
1531 : */
1532 13246 : accumulate_append_subpath(nppath,
1533 : &pa_nonpartial_subpaths,
1534 : NULL);
1535 : }
1536 : }
1537 :
1538 : /*
1539 : * Collect lists of all the available path orderings and
1540 : * parameterizations for all the children. We use these as a
1541 : * heuristic to indicate which sort orderings and parameterizations we
1542 : * should build Append and MergeAppend paths for.
1543 : */
1544 240300 : foreach(lcp, childrel->pathlist)
1545 : {
1546 138684 : Path *childpath = (Path *) lfirst(lcp);
1547 138684 : List *childkeys = childpath->pathkeys;
1548 138684 : Relids childouter = PATH_REQ_OUTER(childpath);
1549 :
1550 : /* Unsorted paths don't contribute to pathkey list */
1551 138684 : if (childkeys != NIL)
1552 : {
1553 : ListCell *lpk;
1554 36794 : bool found = false;
1555 :
1556 : /* Have we already seen this ordering? */
1557 37000 : foreach(lpk, all_child_pathkeys)
1558 : {
1559 24856 : List *existing_pathkeys = (List *) lfirst(lpk);
1560 :
1561 24856 : if (compare_pathkeys(existing_pathkeys,
1562 : childkeys) == PATHKEYS_EQUAL)
1563 : {
1564 24650 : found = true;
1565 24650 : break;
1566 : }
1567 : }
1568 36794 : if (!found)
1569 : {
1570 : /* No, so add it to all_child_pathkeys */
1571 12144 : all_child_pathkeys = lappend(all_child_pathkeys,
1572 : childkeys);
1573 : }
1574 : }
1575 :
1576 : /* Unparameterized paths don't contribute to param-set list */
1577 138684 : if (childouter)
1578 : {
1579 : ListCell *lco;
1580 6596 : bool found = false;
1581 :
1582 : /* Have we already seen this param set? */
1583 7316 : foreach(lco, all_child_outers)
1584 : {
1585 4810 : Relids existing_outers = (Relids) lfirst(lco);
1586 :
1587 4810 : if (bms_equal(existing_outers, childouter))
1588 : {
1589 4090 : found = true;
1590 4090 : break;
1591 : }
1592 : }
1593 6596 : if (!found)
1594 : {
1595 : /* No, so add it to all_child_outers */
1596 2506 : all_child_outers = lappend(all_child_outers,
1597 : childouter);
1598 : }
1599 : }
1600 : }
1601 : }
1602 :
1603 : /*
1604 : * If we found unparameterized paths for all children, build an unordered,
1605 : * unparameterized Append path for the rel. (Note: this is correct even
1606 : * if we have zero or one live subpath due to constraint exclusion.)
1607 : */
1608 46178 : if (subpaths_valid)
1609 45866 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1610 : NIL, NULL, 0, false,
1611 : -1));
1612 :
1613 : /* build an AppendPath for the cheap startup paths, if valid */
1614 46178 : if (startup_subpaths_valid)
1615 664 : add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1616 : NIL, NIL, NULL, 0, false, -1));
1617 :
1618 : /*
1619 : * Consider an append of unordered, unparameterized partial paths. Make
1620 : * it parallel-aware if possible.
1621 : */
1622 46178 : if (partial_subpaths_valid && partial_subpaths != NIL)
1623 : {
1624 : AppendPath *appendpath;
1625 : ListCell *lc;
1626 27782 : int parallel_workers = 0;
1627 :
1628 : /* Find the highest number of workers requested for any subpath. */
1629 95586 : foreach(lc, partial_subpaths)
1630 : {
1631 67804 : Path *path = lfirst(lc);
1632 :
1633 67804 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1634 : }
1635 : Assert(parallel_workers > 0);
1636 :
1637 : /*
1638 : * If the use of parallel append is permitted, always request at least
1639 : * log2(# of children) workers. We assume it can be useful to have
1640 : * extra workers in this case because they will be spread out across
1641 : * the children. The precise formula is just a guess, but we don't
1642 : * want to end up with a radically different answer for a table with N
1643 : * partitions vs. an unpartitioned table with the same data, so the
1644 : * use of some kind of log-scaling here seems to make some sense.
1645 : */
1646 27782 : if (enable_parallel_append)
1647 : {
1648 27734 : parallel_workers = Max(parallel_workers,
1649 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1650 27734 : parallel_workers = Min(parallel_workers,
1651 : max_parallel_workers_per_gather);
1652 : }
1653 : Assert(parallel_workers > 0);
1654 :
1655 : /* Generate a partial append path. */
1656 27782 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1657 : NIL, NULL, parallel_workers,
1658 : enable_parallel_append,
1659 : -1);
1660 :
1661 : /*
1662 : * Make sure any subsequent partial paths use the same row count
1663 : * estimate.
1664 : */
1665 27782 : partial_rows = appendpath->path.rows;
1666 :
1667 : /* Add the path. */
1668 27782 : add_partial_path(rel, (Path *) appendpath);
1669 : }
1670 :
1671 : /*
1672 : * Consider a parallel-aware append using a mix of partial and non-partial
1673 : * paths. (This only makes sense if there's at least one child which has
1674 : * a non-partial path that is substantially cheaper than any partial path;
1675 : * otherwise, we should use the append path added in the previous step.)
1676 : */
1677 46178 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1678 : {
1679 : AppendPath *appendpath;
1680 : ListCell *lc;
1681 4664 : int parallel_workers = 0;
1682 :
1683 : /*
1684 : * Find the highest number of workers requested for any partial
1685 : * subpath.
1686 : */
1687 5566 : foreach(lc, pa_partial_subpaths)
1688 : {
1689 902 : Path *path = lfirst(lc);
1690 :
1691 902 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1692 : }
1693 :
1694 : /*
1695 : * Same formula here as above. It's even more important in this
1696 : * instance because the non-partial paths won't contribute anything to
1697 : * the planned number of parallel workers.
1698 : */
1699 4664 : parallel_workers = Max(parallel_workers,
1700 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1701 4664 : parallel_workers = Min(parallel_workers,
1702 : max_parallel_workers_per_gather);
1703 : Assert(parallel_workers > 0);
1704 :
1705 4664 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1706 : pa_partial_subpaths,
1707 : NIL, NULL, parallel_workers, true,
1708 : partial_rows);
1709 4664 : add_partial_path(rel, (Path *) appendpath);
1710 : }
1711 :
1712 : /*
1713 : * Also build unparameterized ordered append paths based on the collected
1714 : * list of child pathkeys.
1715 : */
1716 46178 : if (subpaths_valid)
1717 45866 : generate_orderedappend_paths(root, rel, live_childrels,
1718 : all_child_pathkeys);
1719 :
1720 : /*
1721 : * Build Append paths for each parameterization seen among the child rels.
1722 : * (This may look pretty expensive, but in most cases of practical
1723 : * interest, the child rels will expose mostly the same parameterizations,
1724 : * so that not that many cases actually get considered here.)
1725 : *
1726 : * The Append node itself cannot enforce quals, so all qual checking must
1727 : * be done in the child paths. This means that to have a parameterized
1728 : * Append path, we must have the exact same parameterization for each
1729 : * child path; otherwise some children might be failing to check the
1730 : * moved-down quals. To make them match up, we can try to increase the
1731 : * parameterization of lesser-parameterized paths.
1732 : */
1733 48684 : foreach(l, all_child_outers)
1734 : {
1735 2506 : Relids required_outer = (Relids) lfirst(l);
1736 : ListCell *lcr;
1737 :
1738 : /* Select the child paths for an Append with this parameterization */
1739 2506 : subpaths = NIL;
1740 2506 : subpaths_valid = true;
1741 9198 : foreach(lcr, live_childrels)
1742 : {
1743 6704 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1744 : Path *subpath;
1745 :
1746 6704 : if (childrel->pathlist == NIL)
1747 : {
1748 : /* failed to make a suitable path for this child */
1749 0 : subpaths_valid = false;
1750 0 : break;
1751 : }
1752 :
1753 6704 : subpath = get_cheapest_parameterized_child_path(root,
1754 : childrel,
1755 : required_outer);
1756 6704 : if (subpath == NULL)
1757 : {
1758 : /* failed to make a suitable path for this child */
1759 12 : subpaths_valid = false;
1760 12 : break;
1761 : }
1762 6692 : accumulate_append_subpath(subpath, &subpaths, NULL);
1763 : }
1764 :
1765 2506 : if (subpaths_valid)
1766 2494 : add_path(rel, (Path *)
1767 2494 : create_append_path(root, rel, subpaths, NIL,
1768 : NIL, required_outer, 0, false,
1769 : -1));
1770 : }
1771 :
1772 : /*
1773 : * When there is only a single child relation, the Append path can inherit
1774 : * any ordering available for the child rel's path, so that it's useful to
1775 : * consider ordered partial paths. Above we only considered the cheapest
1776 : * partial path for each child, but let's also make paths using any
1777 : * partial paths that have pathkeys.
1778 : */
1779 46178 : if (list_length(live_childrels) == 1)
1780 : {
1781 14370 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1782 :
1783 : /* skip the cheapest partial path, since we already used that above */
1784 14574 : for_each_from(l, childrel->partial_pathlist, 1)
1785 : {
1786 204 : Path *path = (Path *) lfirst(l);
1787 : AppendPath *appendpath;
1788 :
1789 : /* skip paths with no pathkeys. */
1790 204 : if (path->pathkeys == NIL)
1791 0 : continue;
1792 :
1793 204 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1794 : NIL, NULL,
1795 : path->parallel_workers, true,
1796 : partial_rows);
1797 204 : add_partial_path(rel, (Path *) appendpath);
1798 : }
1799 : }
1800 46178 : }
1801 :
1802 : /*
1803 : * generate_orderedappend_paths
1804 : * Generate ordered append paths for an append relation
1805 : *
1806 : * Usually we generate MergeAppend paths here, but there are some special
1807 : * cases where we can generate simple Append paths, because the subpaths
1808 : * can provide tuples in the required order already.
1809 : *
1810 : * We generate a path for each ordering (pathkey list) appearing in
1811 : * all_child_pathkeys.
1812 : *
1813 : * We consider both cheapest-startup and cheapest-total cases, ie, for each
1814 : * interesting ordering, collect all the cheapest startup subpaths and all the
1815 : * cheapest total paths, and build a suitable path for each case.
1816 : *
1817 : * We don't currently generate any parameterized ordered paths here. While
1818 : * it would not take much more code here to do so, it's very unclear that it
1819 : * is worth the planning cycles to investigate such paths: there's little
1820 : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1821 : * that the current coding of add_path would reject such paths out of hand,
1822 : * because add_path gives no credit for sort ordering of parameterized paths,
1823 : * and a parameterized MergeAppend is going to be more expensive than the
1824 : * corresponding parameterized Append path. If we ever try harder to support
1825 : * parameterized mergejoin plans, it might be worth adding support for
1826 : * parameterized paths here to feed such joins. (See notes in
1827 : * optimizer/README for why that might not ever happen, though.)
1828 : */
1829 : static void
1830 45866 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1831 : List *live_childrels,
1832 : List *all_child_pathkeys)
1833 : {
1834 : ListCell *lcp;
1835 45866 : List *partition_pathkeys = NIL;
1836 45866 : List *partition_pathkeys_desc = NIL;
1837 45866 : bool partition_pathkeys_partial = true;
1838 45866 : bool partition_pathkeys_desc_partial = true;
1839 :
1840 : /*
1841 : * Some partitioned table setups may allow us to use an Append node
1842 : * instead of a MergeAppend. This is possible in cases such as RANGE
1843 : * partitioned tables where it's guaranteed that an earlier partition must
1844 : * contain rows which come earlier in the sort order. To detect whether
1845 : * this is relevant, build pathkey descriptions of the partition ordering,
1846 : * for both forward and reverse scans.
1847 : */
1848 74224 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
1849 28358 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1850 : {
1851 23782 : partition_pathkeys = build_partition_pathkeys(root, rel,
1852 : ForwardScanDirection,
1853 : &partition_pathkeys_partial);
1854 :
1855 23782 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1856 : BackwardScanDirection,
1857 : &partition_pathkeys_desc_partial);
1858 :
1859 : /*
1860 : * You might think we should truncate_useless_pathkeys here, but
1861 : * allowing partition keys which are a subset of the query's pathkeys
1862 : * can often be useful. For example, consider a table partitioned by
1863 : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1864 : * paths that can produce the a, b, c ordering (perhaps via indexes on
1865 : * (a, b, c)) then it works to consider the appendrel output as
1866 : * ordered by a, b, c.
1867 : */
1868 : }
1869 :
1870 : /* Now consider each interesting sort ordering */
1871 57950 : foreach(lcp, all_child_pathkeys)
1872 : {
1873 12084 : List *pathkeys = (List *) lfirst(lcp);
1874 12084 : List *startup_subpaths = NIL;
1875 12084 : List *total_subpaths = NIL;
1876 12084 : List *fractional_subpaths = NIL;
1877 12084 : bool startup_neq_total = false;
1878 : bool match_partition_order;
1879 : bool match_partition_order_desc;
1880 : int end_index;
1881 : int first_index;
1882 : int direction;
1883 :
1884 : /*
1885 : * Determine if this sort ordering matches any partition pathkeys we
1886 : * have, for both ascending and descending partition order. If the
1887 : * partition pathkeys happen to be contained in pathkeys then it still
1888 : * works, as described above, providing that the partition pathkeys
1889 : * are complete and not just a prefix of the partition keys. (In such
1890 : * cases we'll be relying on the child paths to have sorted the
1891 : * lower-order columns of the required pathkeys.)
1892 : */
1893 12084 : match_partition_order =
1894 21900 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1895 10024 : (!partition_pathkeys_partial &&
1896 208 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1897 :
1898 31464 : match_partition_order_desc = !match_partition_order &&
1899 9708 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1900 9736 : (!partition_pathkeys_desc_partial &&
1901 64 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1902 :
1903 : /*
1904 : * When the required pathkeys match the reverse of the partition
1905 : * order, we must build the list of paths in reverse starting with the
1906 : * last matching partition first. We can get away without making any
1907 : * special cases for this in the loop below by just looping backward
1908 : * over the child relations in this case.
1909 : */
1910 12084 : if (match_partition_order_desc)
1911 : {
1912 : /* loop backward */
1913 48 : first_index = list_length(live_childrels) - 1;
1914 48 : end_index = -1;
1915 48 : direction = -1;
1916 :
1917 : /*
1918 : * Set this to true to save us having to check for
1919 : * match_partition_order_desc in the loop below.
1920 : */
1921 48 : match_partition_order = true;
1922 : }
1923 : else
1924 : {
1925 : /* for all other case, loop forward */
1926 12036 : first_index = 0;
1927 12036 : end_index = list_length(live_childrels);
1928 12036 : direction = 1;
1929 : }
1930 :
1931 : /* Select the child paths for this ordering... */
1932 43398 : for (int i = first_index; i != end_index; i += direction)
1933 : {
1934 31314 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1935 : Path *cheapest_startup,
1936 : *cheapest_total,
1937 31314 : *cheapest_fractional = NULL;
1938 :
1939 : /* Locate the right paths, if they are available. */
1940 : cheapest_startup =
1941 31314 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1942 : pathkeys,
1943 : NULL,
1944 : STARTUP_COST,
1945 : false);
1946 : cheapest_total =
1947 31314 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1948 : pathkeys,
1949 : NULL,
1950 : TOTAL_COST,
1951 : false);
1952 :
1953 : /*
1954 : * If we can't find any paths with the right order just use the
1955 : * cheapest-total path; we'll have to sort it later.
1956 : */
1957 31314 : if (cheapest_startup == NULL || cheapest_total == NULL)
1958 : {
1959 340 : cheapest_startup = cheapest_total =
1960 : childrel->cheapest_total_path;
1961 : /* Assert we do have an unparameterized path for this child */
1962 : Assert(cheapest_total->param_info == NULL);
1963 : }
1964 :
1965 : /*
1966 : * When building a fractional path, determine a cheapest
1967 : * fractional path for each child relation too. Looking at startup
1968 : * and total costs is not enough, because the cheapest fractional
1969 : * path may be dominated by two separate paths (one for startup,
1970 : * one for total).
1971 : *
1972 : * When needed (building fractional path), determine the cheapest
1973 : * fractional path too.
1974 : */
1975 31314 : if (root->tuple_fraction > 0)
1976 : {
1977 860 : double path_fraction = root->tuple_fraction;
1978 :
1979 : /*
1980 : * Merge Append considers only live children relations. Dummy
1981 : * relations must be filtered out before.
1982 : */
1983 : Assert(childrel->rows > 0);
1984 :
1985 : /* Convert absolute limit to a path fraction */
1986 860 : if (path_fraction >= 1.0)
1987 860 : path_fraction /= childrel->rows;
1988 :
1989 : cheapest_fractional =
1990 860 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1991 : pathkeys,
1992 : NULL,
1993 : path_fraction);
1994 :
1995 : /*
1996 : * If we found no path with matching pathkeys, use the
1997 : * cheapest total path instead.
1998 : *
1999 : * XXX We might consider partially sorted paths too (with an
2000 : * incremental sort on top). But we'd have to build all the
2001 : * incremental paths, do the costing etc.
2002 : */
2003 860 : if (!cheapest_fractional)
2004 44 : cheapest_fractional = cheapest_total;
2005 : }
2006 :
2007 : /*
2008 : * Notice whether we actually have different paths for the
2009 : * "cheapest" and "total" cases; frequently there will be no point
2010 : * in two create_merge_append_path() calls.
2011 : */
2012 31314 : if (cheapest_startup != cheapest_total)
2013 96 : startup_neq_total = true;
2014 :
2015 : /*
2016 : * Collect the appropriate child paths. The required logic varies
2017 : * for the Append and MergeAppend cases.
2018 : */
2019 31314 : if (match_partition_order)
2020 : {
2021 : /*
2022 : * We're going to make a plain Append path. We don't need
2023 : * most of what accumulate_append_subpath would do, but we do
2024 : * want to cut out child Appends or MergeAppends if they have
2025 : * just a single subpath (and hence aren't doing anything
2026 : * useful).
2027 : */
2028 6476 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
2029 6476 : cheapest_total = get_singleton_append_subpath(cheapest_total);
2030 :
2031 6476 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
2032 6476 : total_subpaths = lappend(total_subpaths, cheapest_total);
2033 :
2034 6476 : if (cheapest_fractional)
2035 : {
2036 144 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
2037 144 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
2038 : }
2039 : }
2040 : else
2041 : {
2042 : /*
2043 : * Otherwise, rely on accumulate_append_subpath to collect the
2044 : * child paths for the MergeAppend.
2045 : */
2046 24838 : accumulate_append_subpath(cheapest_startup,
2047 : &startup_subpaths, NULL);
2048 24838 : accumulate_append_subpath(cheapest_total,
2049 : &total_subpaths, NULL);
2050 :
2051 24838 : if (cheapest_fractional)
2052 716 : accumulate_append_subpath(cheapest_fractional,
2053 : &fractional_subpaths, NULL);
2054 : }
2055 : }
2056 :
2057 : /* ... and build the Append or MergeAppend paths */
2058 12084 : if (match_partition_order)
2059 : {
2060 : /* We only need Append */
2061 2424 : add_path(rel, (Path *) create_append_path(root,
2062 : rel,
2063 : startup_subpaths,
2064 : NIL,
2065 : pathkeys,
2066 : NULL,
2067 : 0,
2068 : false,
2069 : -1));
2070 2424 : if (startup_neq_total)
2071 0 : add_path(rel, (Path *) create_append_path(root,
2072 : rel,
2073 : total_subpaths,
2074 : NIL,
2075 : pathkeys,
2076 : NULL,
2077 : 0,
2078 : false,
2079 : -1));
2080 :
2081 2424 : if (fractional_subpaths)
2082 72 : add_path(rel, (Path *) create_append_path(root,
2083 : rel,
2084 : fractional_subpaths,
2085 : NIL,
2086 : pathkeys,
2087 : NULL,
2088 : 0,
2089 : false,
2090 : -1));
2091 : }
2092 : else
2093 : {
2094 : /* We need MergeAppend */
2095 9660 : add_path(rel, (Path *) create_merge_append_path(root,
2096 : rel,
2097 : startup_subpaths,
2098 : pathkeys,
2099 : NULL));
2100 9660 : if (startup_neq_total)
2101 60 : add_path(rel, (Path *) create_merge_append_path(root,
2102 : rel,
2103 : total_subpaths,
2104 : pathkeys,
2105 : NULL));
2106 :
2107 9660 : if (fractional_subpaths)
2108 256 : add_path(rel, (Path *) create_merge_append_path(root,
2109 : rel,
2110 : fractional_subpaths,
2111 : pathkeys,
2112 : NULL));
2113 : }
2114 : }
2115 45866 : }
2116 :
2117 : /*
2118 : * get_cheapest_parameterized_child_path
2119 : * Get cheapest path for this relation that has exactly the requested
2120 : * parameterization.
2121 : *
2122 : * Returns NULL if unable to create such a path.
2123 : */
2124 : static Path *
2125 6704 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2126 : Relids required_outer)
2127 : {
2128 : Path *cheapest;
2129 : ListCell *lc;
2130 :
2131 : /*
2132 : * Look up the cheapest existing path with no more than the needed
2133 : * parameterization. If it has exactly the needed parameterization, we're
2134 : * done.
2135 : */
2136 6704 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2137 : NIL,
2138 : required_outer,
2139 : TOTAL_COST,
2140 : false);
2141 : Assert(cheapest != NULL);
2142 6704 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2143 6364 : return cheapest;
2144 :
2145 : /*
2146 : * Otherwise, we can "reparameterize" an existing path to match the given
2147 : * parameterization, which effectively means pushing down additional
2148 : * joinquals to be checked within the path's scan. However, some existing
2149 : * paths might check the available joinquals already while others don't;
2150 : * therefore, it's not clear which existing path will be cheapest after
2151 : * reparameterization. We have to go through them all and find out.
2152 : */
2153 340 : cheapest = NULL;
2154 1180 : foreach(lc, rel->pathlist)
2155 : {
2156 840 : Path *path = (Path *) lfirst(lc);
2157 :
2158 : /* Can't use it if it needs more than requested parameterization */
2159 840 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2160 24 : continue;
2161 :
2162 : /*
2163 : * Reparameterization can only increase the path's cost, so if it's
2164 : * already more expensive than the current cheapest, forget it.
2165 : */
2166 1272 : if (cheapest != NULL &&
2167 456 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2168 384 : continue;
2169 :
2170 : /* Reparameterize if needed, then recheck cost */
2171 432 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2172 : {
2173 356 : path = reparameterize_path(root, path, required_outer, 1.0);
2174 356 : if (path == NULL)
2175 32 : continue; /* failed to reparameterize this one */
2176 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2177 :
2178 324 : if (cheapest != NULL &&
2179 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2180 0 : continue;
2181 : }
2182 :
2183 : /* We have a new best path */
2184 400 : cheapest = path;
2185 : }
2186 :
2187 : /* Return the best path, or NULL if we found no suitable candidate */
2188 340 : return cheapest;
2189 : }
2190 :
2191 : /*
2192 : * accumulate_append_subpath
2193 : * Add a subpath to the list being built for an Append or MergeAppend.
2194 : *
2195 : * It's possible that the child is itself an Append or MergeAppend path, in
2196 : * which case we can "cut out the middleman" and just add its child paths to
2197 : * our own list. (We don't try to do this earlier because we need to apply
2198 : * both levels of transformation to the quals.)
2199 : *
2200 : * Note that if we omit a child MergeAppend in this way, we are effectively
2201 : * omitting a sort step, which seems fine: if the parent is to be an Append,
2202 : * its result would be unsorted anyway, while if the parent is to be a
2203 : * MergeAppend, there's no point in a separate sort on a child.
2204 : *
2205 : * Normally, either path is a partial path and subpaths is a list of partial
2206 : * paths, or else path is a non-partial plan and subpaths is a list of those.
2207 : * However, if path is a parallel-aware Append, then we add its partial path
2208 : * children to subpaths and the rest to special_subpaths. If the latter is
2209 : * NULL, we don't flatten the path at all (unless it contains only partial
2210 : * paths).
2211 : */
2212 : static void
2213 298906 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2214 : {
2215 298906 : if (IsA(path, AppendPath))
2216 : {
2217 15872 : AppendPath *apath = (AppendPath *) path;
2218 :
2219 15872 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2220 : {
2221 15536 : *subpaths = list_concat(*subpaths, apath->subpaths);
2222 15536 : return;
2223 : }
2224 336 : else if (special_subpaths != NULL)
2225 : {
2226 : List *new_special_subpaths;
2227 :
2228 : /* Split Parallel Append into partial and non-partial subpaths */
2229 168 : *subpaths = list_concat(*subpaths,
2230 168 : list_copy_tail(apath->subpaths,
2231 : apath->first_partial_path));
2232 168 : new_special_subpaths = list_copy_head(apath->subpaths,
2233 : apath->first_partial_path);
2234 168 : *special_subpaths = list_concat(*special_subpaths,
2235 : new_special_subpaths);
2236 168 : return;
2237 : }
2238 : }
2239 283034 : else if (IsA(path, MergeAppendPath))
2240 : {
2241 1220 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2242 :
2243 1220 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2244 1220 : return;
2245 : }
2246 :
2247 281982 : *subpaths = lappend(*subpaths, path);
2248 : }
2249 :
2250 : /*
2251 : * get_singleton_append_subpath
2252 : * Returns the single subpath of an Append/MergeAppend, or just
2253 : * return 'path' if it's not a single sub-path Append/MergeAppend.
2254 : *
2255 : * Note: 'path' must not be a parallel-aware path.
2256 : */
2257 : static Path *
2258 13096 : get_singleton_append_subpath(Path *path)
2259 : {
2260 : Assert(!path->parallel_aware);
2261 :
2262 13096 : if (IsA(path, AppendPath))
2263 : {
2264 388 : AppendPath *apath = (AppendPath *) path;
2265 :
2266 388 : if (list_length(apath->subpaths) == 1)
2267 192 : return (Path *) linitial(apath->subpaths);
2268 : }
2269 12708 : else if (IsA(path, MergeAppendPath))
2270 : {
2271 348 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2272 :
2273 348 : if (list_length(mpath->subpaths) == 1)
2274 0 : return (Path *) linitial(mpath->subpaths);
2275 : }
2276 :
2277 12904 : return path;
2278 : }
2279 :
2280 : /*
2281 : * set_dummy_rel_pathlist
2282 : * Build a dummy path for a relation that's been excluded by constraints
2283 : *
2284 : * Rather than inventing a special "dummy" path type, we represent this as an
2285 : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2286 : *
2287 : * (See also mark_dummy_rel, which does basically the same thing, but is
2288 : * typically used to change a rel into dummy state after we already made
2289 : * paths for it.)
2290 : */
2291 : static void
2292 1262 : set_dummy_rel_pathlist(RelOptInfo *rel)
2293 : {
2294 : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2295 1262 : rel->rows = 0;
2296 1262 : rel->reltarget->width = 0;
2297 :
2298 : /* Discard any pre-existing paths; no further need for them */
2299 1262 : rel->pathlist = NIL;
2300 1262 : rel->partial_pathlist = NIL;
2301 :
2302 : /* Set up the dummy path */
2303 1262 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2304 : NIL, rel->lateral_relids,
2305 : 0, false, -1));
2306 :
2307 : /*
2308 : * We set the cheapest-path fields immediately, just in case they were
2309 : * pointing at some discarded path. This is redundant in current usage
2310 : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2311 : * for safety and consistency with mark_dummy_rel.
2312 : */
2313 1262 : set_cheapest(rel);
2314 1262 : }
2315 :
2316 : /*
2317 : * find_window_run_conditions
2318 : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2319 : * function to determine the function's monotonic properties. We then
2320 : * see if 'opexpr' can be used to short-circuit execution.
2321 : *
2322 : * For example row_number() over (order by ...) always produces a value one
2323 : * higher than the previous. If someone has a window function in a subquery
2324 : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2325 : * as well stop processing the windowagg once the row number reaches 11. Here
2326 : * we check if 'opexpr' might help us to stop doing needless extra processing
2327 : * in WindowAgg nodes.
2328 : *
2329 : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2330 : * its original purpose. This is set to false if the caller can assume that
2331 : * the run condition will handle all of the required filtering.
2332 : *
2333 : * Returns true if 'opexpr' was found to be useful and was added to the
2334 : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2335 : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2336 : * If the 'opexpr' cannot be used then we set *keep_original to true and
2337 : * return false.
2338 : */
2339 : static bool
2340 240 : find_window_run_conditions(Query *subquery, AttrNumber attno,
2341 : WindowFunc *wfunc, OpExpr *opexpr, bool wfunc_left,
2342 : bool *keep_original, Bitmapset **run_cond_attrs)
2343 : {
2344 : Oid prosupport;
2345 : Expr *otherexpr;
2346 : SupportRequestWFuncMonotonic req;
2347 : SupportRequestWFuncMonotonic *res;
2348 : WindowClause *wclause;
2349 : List *opinfos;
2350 : OpExpr *runopexpr;
2351 : Oid runoperator;
2352 : ListCell *lc;
2353 :
2354 240 : *keep_original = true;
2355 :
2356 240 : while (IsA(wfunc, RelabelType))
2357 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2358 :
2359 : /* we can only work with window functions */
2360 240 : if (!IsA(wfunc, WindowFunc))
2361 24 : return false;
2362 :
2363 : /* can't use it if there are subplans in the WindowFunc */
2364 216 : if (contain_subplans((Node *) wfunc))
2365 6 : return false;
2366 :
2367 210 : prosupport = get_func_support(wfunc->winfnoid);
2368 :
2369 : /* Check if there's a support function for 'wfunc' */
2370 210 : if (!OidIsValid(prosupport))
2371 18 : return false;
2372 :
2373 : /* get the Expr from the other side of the OpExpr */
2374 192 : if (wfunc_left)
2375 168 : otherexpr = lsecond(opexpr->args);
2376 : else
2377 24 : otherexpr = linitial(opexpr->args);
2378 :
2379 : /*
2380 : * The value being compared must not change during the evaluation of the
2381 : * window partition.
2382 : */
2383 192 : if (!is_pseudo_constant_clause((Node *) otherexpr))
2384 0 : return false;
2385 :
2386 : /* find the window clause belonging to the window function */
2387 192 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2388 192 : wfunc->winref - 1);
2389 :
2390 192 : req.type = T_SupportRequestWFuncMonotonic;
2391 192 : req.window_func = wfunc;
2392 192 : req.window_clause = wclause;
2393 :
2394 : /* call the support function */
2395 : res = (SupportRequestWFuncMonotonic *)
2396 192 : DatumGetPointer(OidFunctionCall1(prosupport,
2397 : PointerGetDatum(&req)));
2398 :
2399 : /*
2400 : * Nothing to do if the function is neither monotonically increasing nor
2401 : * monotonically decreasing.
2402 : */
2403 192 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2404 0 : return false;
2405 :
2406 192 : runopexpr = NULL;
2407 192 : runoperator = InvalidOid;
2408 192 : opinfos = get_op_index_interpretation(opexpr->opno);
2409 :
2410 192 : foreach(lc, opinfos)
2411 : {
2412 192 : OpIndexInterpretation *opinfo = (OpIndexInterpretation *) lfirst(lc);
2413 192 : CompareType cmptype = opinfo->cmptype;
2414 :
2415 : /* handle < / <= */
2416 192 : if (cmptype == COMPARE_LT || cmptype == COMPARE_LE)
2417 : {
2418 : /*
2419 : * < / <= is supported for monotonically increasing functions in
2420 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2421 : * for monotonically decreasing functions.
2422 : */
2423 138 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2424 18 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2425 : {
2426 126 : *keep_original = false;
2427 126 : runopexpr = opexpr;
2428 126 : runoperator = opexpr->opno;
2429 : }
2430 138 : break;
2431 : }
2432 : /* handle > / >= */
2433 54 : else if (cmptype == COMPARE_GT || cmptype == COMPARE_GE)
2434 : {
2435 : /*
2436 : * > / >= is supported for monotonically decreasing functions in
2437 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2438 : * for monotonically increasing functions.
2439 : */
2440 18 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2441 12 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2442 : {
2443 18 : *keep_original = false;
2444 18 : runopexpr = opexpr;
2445 18 : runoperator = opexpr->opno;
2446 : }
2447 18 : break;
2448 : }
2449 : /* handle = */
2450 36 : else if (cmptype == COMPARE_EQ)
2451 : {
2452 : CompareType newcmptype;
2453 :
2454 : /*
2455 : * When both monotonically increasing and decreasing then the
2456 : * return value of the window function will be the same each time.
2457 : * We can simply use 'opexpr' as the run condition without
2458 : * modifying it.
2459 : */
2460 36 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2461 : {
2462 6 : *keep_original = false;
2463 6 : runopexpr = opexpr;
2464 6 : runoperator = opexpr->opno;
2465 6 : break;
2466 : }
2467 :
2468 : /*
2469 : * When monotonically increasing we make a qual with <wfunc> <=
2470 : * <value> or <value> >= <wfunc> in order to filter out values
2471 : * which are above the value in the equality condition. For
2472 : * monotonically decreasing functions we want to filter values
2473 : * below the value in the equality condition.
2474 : */
2475 30 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2476 30 : newcmptype = wfunc_left ? COMPARE_LE : COMPARE_GE;
2477 : else
2478 0 : newcmptype = wfunc_left ? COMPARE_GE : COMPARE_LE;
2479 :
2480 : /* We must keep the original equality qual */
2481 30 : *keep_original = true;
2482 30 : runopexpr = opexpr;
2483 :
2484 : /* determine the operator to use for the WindowFuncRunCondition */
2485 30 : runoperator = get_opfamily_member_for_cmptype(opinfo->opfamily_id,
2486 : opinfo->oplefttype,
2487 : opinfo->oprighttype,
2488 : newcmptype);
2489 30 : break;
2490 : }
2491 : }
2492 :
2493 192 : if (runopexpr != NULL)
2494 : {
2495 : WindowFuncRunCondition *wfuncrc;
2496 :
2497 180 : wfuncrc = makeNode(WindowFuncRunCondition);
2498 180 : wfuncrc->opno = runoperator;
2499 180 : wfuncrc->inputcollid = runopexpr->inputcollid;
2500 180 : wfuncrc->wfunc_left = wfunc_left;
2501 180 : wfuncrc->arg = copyObject(otherexpr);
2502 :
2503 180 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2504 :
2505 : /* record that this attno was used in a run condition */
2506 180 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2507 : attno - FirstLowInvalidHeapAttributeNumber);
2508 180 : return true;
2509 : }
2510 :
2511 : /* unsupported OpExpr */
2512 12 : return false;
2513 : }
2514 :
2515 : /*
2516 : * check_and_push_window_quals
2517 : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2518 : * as a 'runCondition' qual. These, when present, allow some unnecessary
2519 : * work to be skipped during execution.
2520 : *
2521 : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2522 : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2523 : * window quals for.
2524 : *
2525 : * Returns true if the caller still must keep the original qual or false if
2526 : * the caller can safely ignore the original qual because the WindowAgg node
2527 : * will use the runCondition to stop returning tuples.
2528 : */
2529 : static bool
2530 252 : check_and_push_window_quals(Query *subquery, Node *clause,
2531 : Bitmapset **run_cond_attrs)
2532 : {
2533 252 : OpExpr *opexpr = (OpExpr *) clause;
2534 252 : bool keep_original = true;
2535 : Var *var1;
2536 : Var *var2;
2537 :
2538 : /* We're only able to use OpExprs with 2 operands */
2539 252 : if (!IsA(opexpr, OpExpr))
2540 18 : return true;
2541 :
2542 234 : if (list_length(opexpr->args) != 2)
2543 0 : return true;
2544 :
2545 : /*
2546 : * Currently, we restrict this optimization to strict OpExprs. The reason
2547 : * for this is that during execution, once the runcondition becomes false,
2548 : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2549 : * function result values, we set them to NULL. Having only strict
2550 : * OpExprs here ensures that we properly filter out the tuples with NULLs
2551 : * in the top-level WindowAgg.
2552 : */
2553 234 : set_opfuncid(opexpr);
2554 234 : if (!func_strict(opexpr->opfuncid))
2555 0 : return true;
2556 :
2557 : /*
2558 : * Check for plain Vars that reference window functions in the subquery.
2559 : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2560 : * be used as part of the run condition.
2561 : */
2562 :
2563 : /* Check the left side of the OpExpr */
2564 234 : var1 = linitial(opexpr->args);
2565 234 : if (IsA(var1, Var) && var1->varattno > 0)
2566 : {
2567 198 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2568 198 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2569 :
2570 198 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2571 : true, &keep_original, run_cond_attrs))
2572 162 : return keep_original;
2573 : }
2574 :
2575 : /* and check the right side */
2576 72 : var2 = lsecond(opexpr->args);
2577 72 : if (IsA(var2, Var) && var2->varattno > 0)
2578 : {
2579 42 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2580 42 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2581 :
2582 42 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2583 : false, &keep_original, run_cond_attrs))
2584 18 : return keep_original;
2585 : }
2586 :
2587 54 : return true;
2588 : }
2589 :
2590 : /*
2591 : * set_subquery_pathlist
2592 : * Generate SubqueryScan access paths for a subquery RTE
2593 : *
2594 : * We don't currently support generating parameterized paths for subqueries
2595 : * by pushing join clauses down into them; it seems too expensive to re-plan
2596 : * the subquery multiple times to consider different alternatives.
2597 : * (XXX that could stand to be reconsidered, now that we use Paths.)
2598 : * So the paths made here will be parameterized if the subquery contains
2599 : * LATERAL references, otherwise not. As long as that's true, there's no need
2600 : * for a separate set_subquery_size phase: just make the paths right away.
2601 : */
2602 : static void
2603 16760 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2604 : Index rti, RangeTblEntry *rte)
2605 : {
2606 16760 : Query *parse = root->parse;
2607 16760 : Query *subquery = rte->subquery;
2608 : bool trivial_pathtarget;
2609 : Relids required_outer;
2610 : pushdown_safety_info safetyInfo;
2611 : double tuple_fraction;
2612 : RelOptInfo *sub_final_rel;
2613 16760 : Bitmapset *run_cond_attrs = NULL;
2614 : ListCell *lc;
2615 : char *plan_name;
2616 :
2617 : /*
2618 : * Must copy the Query so that planning doesn't mess up the RTE contents
2619 : * (really really need to fix the planner to not scribble on its input,
2620 : * someday ... but see remove_unused_subquery_outputs to start with).
2621 : */
2622 16760 : subquery = copyObject(subquery);
2623 :
2624 : /*
2625 : * If it's a LATERAL subquery, it might contain some Vars of the current
2626 : * query level, requiring it to be treated as parameterized, even though
2627 : * we don't support pushing down join quals into subqueries.
2628 : */
2629 16760 : required_outer = rel->lateral_relids;
2630 :
2631 : /*
2632 : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2633 : * flags as needed while recursing. In particular, we need a workspace
2634 : * for keeping track of the reasons why columns are unsafe to reference.
2635 : * These reasons are stored in the bits inside unsafeFlags[i] when we
2636 : * discover reasons that column i of the subquery is unsafe to be used in
2637 : * a pushed-down qual.
2638 : */
2639 16760 : memset(&safetyInfo, 0, sizeof(safetyInfo));
2640 16760 : safetyInfo.unsafeFlags = (unsigned char *)
2641 16760 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2642 :
2643 : /*
2644 : * If the subquery has the "security_barrier" flag, it means the subquery
2645 : * originated from a view that must enforce row-level security. Then we
2646 : * must not push down quals that contain leaky functions. (Ideally this
2647 : * would be checked inside subquery_is_pushdown_safe, but since we don't
2648 : * currently pass the RTE to that function, we must do it here.)
2649 : */
2650 16760 : safetyInfo.unsafeLeaky = rte->security_barrier;
2651 :
2652 : /*
2653 : * If there are any restriction clauses that have been attached to the
2654 : * subquery relation, consider pushing them down to become WHERE or HAVING
2655 : * quals of the subquery itself. This transformation is useful because it
2656 : * may allow us to generate a better plan for the subquery than evaluating
2657 : * all the subquery output rows and then filtering them.
2658 : *
2659 : * There are several cases where we cannot push down clauses. Restrictions
2660 : * involving the subquery are checked by subquery_is_pushdown_safe().
2661 : * Restrictions on individual clauses are checked by
2662 : * qual_is_pushdown_safe(). Also, we don't want to push down
2663 : * pseudoconstant clauses; better to have the gating node above the
2664 : * subquery.
2665 : *
2666 : * Non-pushed-down clauses will get evaluated as qpquals of the
2667 : * SubqueryScan node.
2668 : *
2669 : * XXX Are there any cases where we want to make a policy decision not to
2670 : * push down a pushable qual, because it'd result in a worse plan?
2671 : */
2672 18628 : if (rel->baserestrictinfo != NIL &&
2673 1868 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2674 : {
2675 : /* OK to consider pushing down individual quals */
2676 1722 : List *upperrestrictlist = NIL;
2677 : ListCell *l;
2678 :
2679 4396 : foreach(l, rel->baserestrictinfo)
2680 : {
2681 2674 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2682 2674 : Node *clause = (Node *) rinfo->clause;
2683 :
2684 2674 : if (rinfo->pseudoconstant)
2685 : {
2686 4 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2687 4 : continue;
2688 : }
2689 :
2690 2670 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2691 : {
2692 1964 : case PUSHDOWN_SAFE:
2693 : /* Push it down */
2694 1964 : subquery_push_qual(subquery, rte, rti, clause);
2695 1964 : break;
2696 :
2697 252 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2698 :
2699 : /*
2700 : * Since we can't push the qual down into the subquery,
2701 : * check if it happens to reference a window function. If
2702 : * so then it might be useful to use for the WindowAgg's
2703 : * runCondition.
2704 : */
2705 504 : if (!subquery->hasWindowFuncs ||
2706 252 : check_and_push_window_quals(subquery, clause,
2707 : &run_cond_attrs))
2708 : {
2709 : /*
2710 : * subquery has no window funcs or the clause is not a
2711 : * suitable window run condition qual or it is, but
2712 : * the original must also be kept in the upper query.
2713 : */
2714 102 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2715 : }
2716 252 : break;
2717 :
2718 454 : case PUSHDOWN_UNSAFE:
2719 454 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2720 454 : break;
2721 : }
2722 : }
2723 1722 : rel->baserestrictinfo = upperrestrictlist;
2724 : /* We don't bother recomputing baserestrict_min_security */
2725 : }
2726 :
2727 16760 : pfree(safetyInfo.unsafeFlags);
2728 :
2729 : /*
2730 : * The upper query might not use all the subquery's output columns; if
2731 : * not, we can simplify. Pass the attributes that were pushed down into
2732 : * WindowAgg run conditions to ensure we don't accidentally think those
2733 : * are unused.
2734 : */
2735 16760 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2736 :
2737 : /*
2738 : * We can safely pass the outer tuple_fraction down to the subquery if the
2739 : * outer level has no joining, aggregation, or sorting to do. Otherwise
2740 : * we'd better tell the subquery to plan for full retrieval. (XXX This
2741 : * could probably be made more intelligent ...)
2742 : */
2743 16760 : if (parse->hasAggs ||
2744 15514 : parse->groupClause ||
2745 15496 : parse->groupingSets ||
2746 15496 : root->hasHavingQual ||
2747 15496 : parse->distinctClause ||
2748 20534 : parse->sortClause ||
2749 5546 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
2750 12536 : tuple_fraction = 0.0; /* default case */
2751 : else
2752 4224 : tuple_fraction = root->tuple_fraction;
2753 :
2754 : /* plan_params should not be in use in current query level */
2755 : Assert(root->plan_params == NIL);
2756 :
2757 : /* Generate a subroot and Paths for the subquery */
2758 16760 : plan_name = choose_plan_name(root->glob, rte->eref->aliasname, false);
2759 16760 : rel->subroot = subquery_planner(root->glob, subquery, plan_name,
2760 : root, false, tuple_fraction, NULL);
2761 :
2762 : /* Isolate the params needed by this specific subplan */
2763 16760 : rel->subplan_params = root->plan_params;
2764 16760 : root->plan_params = NIL;
2765 :
2766 : /*
2767 : * It's possible that constraint exclusion proved the subquery empty. If
2768 : * so, it's desirable to produce an unadorned dummy path so that we will
2769 : * recognize appropriate optimizations at this query level.
2770 : */
2771 16760 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2772 :
2773 16760 : if (IS_DUMMY_REL(sub_final_rel))
2774 : {
2775 126 : set_dummy_rel_pathlist(rel);
2776 126 : return;
2777 : }
2778 :
2779 : /*
2780 : * Mark rel with estimated output rows, width, etc. Note that we have to
2781 : * do this before generating outer-query paths, else cost_subqueryscan is
2782 : * not happy.
2783 : */
2784 16634 : set_subquery_size_estimates(root, rel);
2785 :
2786 : /*
2787 : * Also detect whether the reltarget is trivial, so that we can pass that
2788 : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2789 : * It's trivial if it fetches all the subplan output columns in order.
2790 : */
2791 16634 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2792 7578 : trivial_pathtarget = false;
2793 : else
2794 : {
2795 9056 : trivial_pathtarget = true;
2796 28668 : foreach(lc, rel->reltarget->exprs)
2797 : {
2798 19910 : Node *node = (Node *) lfirst(lc);
2799 : Var *var;
2800 :
2801 19910 : if (!IsA(node, Var))
2802 : {
2803 0 : trivial_pathtarget = false;
2804 0 : break;
2805 : }
2806 19910 : var = (Var *) node;
2807 19910 : if (var->varno != rti ||
2808 19910 : var->varattno != foreach_current_index(lc) + 1)
2809 : {
2810 298 : trivial_pathtarget = false;
2811 298 : break;
2812 : }
2813 : }
2814 : }
2815 :
2816 : /*
2817 : * For each Path that subquery_planner produced, make a SubqueryScanPath
2818 : * in the outer query.
2819 : */
2820 34044 : foreach(lc, sub_final_rel->pathlist)
2821 : {
2822 17410 : Path *subpath = (Path *) lfirst(lc);
2823 : List *pathkeys;
2824 :
2825 : /* Convert subpath's pathkeys to outer representation */
2826 17410 : pathkeys = convert_subquery_pathkeys(root,
2827 : rel,
2828 : subpath->pathkeys,
2829 : make_tlist_from_pathtarget(subpath->pathtarget));
2830 :
2831 : /* Generate outer path using this subpath */
2832 17410 : add_path(rel, (Path *)
2833 17410 : create_subqueryscan_path(root, rel, subpath,
2834 : trivial_pathtarget,
2835 : pathkeys, required_outer));
2836 : }
2837 :
2838 : /* If outer rel allows parallelism, do same for partial paths. */
2839 16634 : if (rel->consider_parallel && bms_is_empty(required_outer))
2840 : {
2841 : /* If consider_parallel is false, there should be no partial paths. */
2842 : Assert(sub_final_rel->consider_parallel ||
2843 : sub_final_rel->partial_pathlist == NIL);
2844 :
2845 : /* Same for partial paths. */
2846 12458 : foreach(lc, sub_final_rel->partial_pathlist)
2847 : {
2848 42 : Path *subpath = (Path *) lfirst(lc);
2849 : List *pathkeys;
2850 :
2851 : /* Convert subpath's pathkeys to outer representation */
2852 42 : pathkeys = convert_subquery_pathkeys(root,
2853 : rel,
2854 : subpath->pathkeys,
2855 : make_tlist_from_pathtarget(subpath->pathtarget));
2856 :
2857 : /* Generate outer path using this subpath */
2858 42 : add_partial_path(rel, (Path *)
2859 42 : create_subqueryscan_path(root, rel, subpath,
2860 : trivial_pathtarget,
2861 : pathkeys,
2862 : required_outer));
2863 : }
2864 : }
2865 : }
2866 :
2867 : /*
2868 : * set_function_pathlist
2869 : * Build the (single) access path for a function RTE
2870 : */
2871 : static void
2872 51268 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2873 : {
2874 : Relids required_outer;
2875 51268 : List *pathkeys = NIL;
2876 :
2877 : /*
2878 : * We don't support pushing join clauses into the quals of a function
2879 : * scan, but it could still have required parameterization due to LATERAL
2880 : * refs in the function expression.
2881 : */
2882 51268 : required_outer = rel->lateral_relids;
2883 :
2884 : /*
2885 : * The result is considered unordered unless ORDINALITY was used, in which
2886 : * case it is ordered by the ordinal column (the last one). See if we
2887 : * care, by checking for uses of that Var in equivalence classes.
2888 : */
2889 51268 : if (rte->funcordinality)
2890 : {
2891 932 : AttrNumber ordattno = rel->max_attr;
2892 932 : Var *var = NULL;
2893 : ListCell *lc;
2894 :
2895 : /*
2896 : * Is there a Var for it in rel's targetlist? If not, the query did
2897 : * not reference the ordinality column, or at least not in any way
2898 : * that would be interesting for sorting.
2899 : */
2900 2116 : foreach(lc, rel->reltarget->exprs)
2901 : {
2902 2110 : Var *node = (Var *) lfirst(lc);
2903 :
2904 : /* checking varno/varlevelsup is just paranoia */
2905 2110 : if (IsA(node, Var) &&
2906 2110 : node->varattno == ordattno &&
2907 926 : node->varno == rel->relid &&
2908 926 : node->varlevelsup == 0)
2909 : {
2910 926 : var = node;
2911 926 : break;
2912 : }
2913 : }
2914 :
2915 : /*
2916 : * Try to build pathkeys for this Var with int8 sorting. We tell
2917 : * build_expression_pathkey not to build any new equivalence class; if
2918 : * the Var isn't already mentioned in some EC, it means that nothing
2919 : * cares about the ordering.
2920 : */
2921 932 : if (var)
2922 926 : pathkeys = build_expression_pathkey(root,
2923 : (Expr *) var,
2924 : Int8LessOperator,
2925 : rel->relids,
2926 : false);
2927 : }
2928 :
2929 : /* Generate appropriate path */
2930 51268 : add_path(rel, create_functionscan_path(root, rel,
2931 : pathkeys, required_outer));
2932 51268 : }
2933 :
2934 : /*
2935 : * set_values_pathlist
2936 : * Build the (single) access path for a VALUES RTE
2937 : */
2938 : static void
2939 8264 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2940 : {
2941 : Relids required_outer;
2942 :
2943 : /*
2944 : * We don't support pushing join clauses into the quals of a values scan,
2945 : * but it could still have required parameterization due to LATERAL refs
2946 : * in the values expressions.
2947 : */
2948 8264 : required_outer = rel->lateral_relids;
2949 :
2950 : /* Generate appropriate path */
2951 8264 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
2952 8264 : }
2953 :
2954 : /*
2955 : * set_tablefunc_pathlist
2956 : * Build the (single) access path for a table func RTE
2957 : */
2958 : static void
2959 626 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2960 : {
2961 : Relids required_outer;
2962 :
2963 : /*
2964 : * We don't support pushing join clauses into the quals of a tablefunc
2965 : * scan, but it could still have required parameterization due to LATERAL
2966 : * refs in the function expression.
2967 : */
2968 626 : required_outer = rel->lateral_relids;
2969 :
2970 : /* Generate appropriate path */
2971 626 : add_path(rel, create_tablefuncscan_path(root, rel,
2972 : required_outer));
2973 626 : }
2974 :
2975 : /*
2976 : * set_cte_pathlist
2977 : * Build the (single) access path for a non-self-reference CTE RTE
2978 : *
2979 : * There's no need for a separate set_cte_size phase, since we don't
2980 : * support join-qual-parameterized paths for CTEs.
2981 : */
2982 : static void
2983 4258 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2984 : {
2985 : Path *ctepath;
2986 : Plan *cteplan;
2987 : PlannerInfo *cteroot;
2988 : Index levelsup;
2989 : List *pathkeys;
2990 : int ndx;
2991 : ListCell *lc;
2992 : int plan_id;
2993 : Relids required_outer;
2994 :
2995 : /*
2996 : * Find the referenced CTE, and locate the path and plan previously made
2997 : * for it.
2998 : */
2999 4258 : levelsup = rte->ctelevelsup;
3000 4258 : cteroot = root;
3001 7452 : while (levelsup-- > 0)
3002 : {
3003 3194 : cteroot = cteroot->parent_root;
3004 3194 : if (!cteroot) /* shouldn't happen */
3005 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3006 : }
3007 :
3008 : /*
3009 : * Note: cte_plan_ids can be shorter than cteList, if we are still working
3010 : * on planning the CTEs (ie, this is a side-reference from another CTE).
3011 : * So we mustn't use forboth here.
3012 : */
3013 4258 : ndx = 0;
3014 5910 : foreach(lc, cteroot->parse->cteList)
3015 : {
3016 5910 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
3017 :
3018 5910 : if (strcmp(cte->ctename, rte->ctename) == 0)
3019 4258 : break;
3020 1652 : ndx++;
3021 : }
3022 4258 : if (lc == NULL) /* shouldn't happen */
3023 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
3024 4258 : if (ndx >= list_length(cteroot->cte_plan_ids))
3025 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
3026 4258 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
3027 4258 : if (plan_id <= 0)
3028 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
3029 :
3030 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
3031 4258 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
3032 4258 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
3033 :
3034 : /* Mark rel with estimated output rows, width, etc */
3035 4258 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
3036 :
3037 : /* Convert the ctepath's pathkeys to outer query's representation */
3038 4258 : pathkeys = convert_subquery_pathkeys(root,
3039 : rel,
3040 : ctepath->pathkeys,
3041 : cteplan->targetlist);
3042 :
3043 : /*
3044 : * We don't support pushing join clauses into the quals of a CTE scan, but
3045 : * it could still have required parameterization due to LATERAL refs in
3046 : * its tlist.
3047 : */
3048 4258 : required_outer = rel->lateral_relids;
3049 :
3050 : /* Generate appropriate path */
3051 4258 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
3052 4258 : }
3053 :
3054 : /*
3055 : * set_namedtuplestore_pathlist
3056 : * Build the (single) access path for a named tuplestore RTE
3057 : *
3058 : * There's no need for a separate set_namedtuplestore_size phase, since we
3059 : * don't support join-qual-parameterized paths for tuplestores.
3060 : */
3061 : static void
3062 478 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
3063 : RangeTblEntry *rte)
3064 : {
3065 : Relids required_outer;
3066 :
3067 : /* Mark rel with estimated output rows, width, etc */
3068 478 : set_namedtuplestore_size_estimates(root, rel);
3069 :
3070 : /*
3071 : * We don't support pushing join clauses into the quals of a tuplestore
3072 : * scan, but it could still have required parameterization due to LATERAL
3073 : * refs in its tlist.
3074 : */
3075 478 : required_outer = rel->lateral_relids;
3076 :
3077 : /* Generate appropriate path */
3078 478 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
3079 478 : }
3080 :
3081 : /*
3082 : * set_result_pathlist
3083 : * Build the (single) access path for an RTE_RESULT RTE
3084 : *
3085 : * There's no need for a separate set_result_size phase, since we
3086 : * don't support join-qual-parameterized paths for these RTEs.
3087 : */
3088 : static void
3089 4208 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
3090 : RangeTblEntry *rte)
3091 : {
3092 : Relids required_outer;
3093 :
3094 : /* Mark rel with estimated output rows, width, etc */
3095 4208 : set_result_size_estimates(root, rel);
3096 :
3097 : /*
3098 : * We don't support pushing join clauses into the quals of a Result scan,
3099 : * but it could still have required parameterization due to LATERAL refs
3100 : * in its tlist.
3101 : */
3102 4208 : required_outer = rel->lateral_relids;
3103 :
3104 : /* Generate appropriate path */
3105 4208 : add_path(rel, create_resultscan_path(root, rel, required_outer));
3106 4208 : }
3107 :
3108 : /*
3109 : * set_worktable_pathlist
3110 : * Build the (single) access path for a self-reference CTE RTE
3111 : *
3112 : * There's no need for a separate set_worktable_size phase, since we don't
3113 : * support join-qual-parameterized paths for CTEs.
3114 : */
3115 : static void
3116 932 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3117 : {
3118 : Path *ctepath;
3119 : PlannerInfo *cteroot;
3120 : Index levelsup;
3121 : Relids required_outer;
3122 :
3123 : /*
3124 : * We need to find the non-recursive term's path, which is in the plan
3125 : * level that's processing the recursive UNION, which is one level *below*
3126 : * where the CTE comes from.
3127 : */
3128 932 : levelsup = rte->ctelevelsup;
3129 932 : if (levelsup == 0) /* shouldn't happen */
3130 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3131 932 : levelsup--;
3132 932 : cteroot = root;
3133 2276 : while (levelsup-- > 0)
3134 : {
3135 1344 : cteroot = cteroot->parent_root;
3136 1344 : if (!cteroot) /* shouldn't happen */
3137 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3138 : }
3139 932 : ctepath = cteroot->non_recursive_path;
3140 932 : if (!ctepath) /* shouldn't happen */
3141 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3142 :
3143 : /* Mark rel with estimated output rows, width, etc */
3144 932 : set_cte_size_estimates(root, rel, ctepath->rows);
3145 :
3146 : /*
3147 : * We don't support pushing join clauses into the quals of a worktable
3148 : * scan, but it could still have required parameterization due to LATERAL
3149 : * refs in its tlist. (I'm not sure this is actually possible given the
3150 : * restrictions on recursive references, but it's easy enough to support.)
3151 : */
3152 932 : required_outer = rel->lateral_relids;
3153 :
3154 : /* Generate appropriate path */
3155 932 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
3156 932 : }
3157 :
3158 : /*
3159 : * generate_gather_paths
3160 : * Generate parallel access paths for a relation by pushing a Gather or
3161 : * Gather Merge on top of a partial path.
3162 : *
3163 : * This must not be called until after we're done creating all partial paths
3164 : * for the specified relation. (Otherwise, add_partial_path might delete a
3165 : * path that some GatherPath or GatherMergePath has a reference to.)
3166 : *
3167 : * If we're generating paths for a scan or join relation, override_rows will
3168 : * be false, and we'll just use the relation's size estimate. When we're
3169 : * being called for a partially-grouped or partially-distinct path, though, we
3170 : * need to override the rowcount estimate. (It's not clear that the
3171 : * particular value we're using here is actually best, but the underlying rel
3172 : * has no estimate so we must do something.)
3173 : */
3174 : void
3175 24456 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3176 : {
3177 : Path *cheapest_partial_path;
3178 : Path *simple_gather_path;
3179 : ListCell *lc;
3180 : double rows;
3181 24456 : double *rowsp = NULL;
3182 :
3183 : /* If there are no partial paths, there's nothing to do here. */
3184 24456 : if (rel->partial_pathlist == NIL)
3185 0 : return;
3186 :
3187 : /* Should we override the rel's rowcount estimate? */
3188 24456 : if (override_rows)
3189 6084 : rowsp = &rows;
3190 :
3191 : /*
3192 : * The output of Gather is always unsorted, so there's only one partial
3193 : * path of interest: the cheapest one. That will be the one at the front
3194 : * of partial_pathlist because of the way add_partial_path works.
3195 : */
3196 24456 : cheapest_partial_path = linitial(rel->partial_pathlist);
3197 24456 : rows = compute_gather_rows(cheapest_partial_path);
3198 : simple_gather_path = (Path *)
3199 24456 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3200 : NULL, rowsp);
3201 24456 : add_path(rel, simple_gather_path);
3202 :
3203 : /*
3204 : * For each useful ordering, we can consider an order-preserving Gather
3205 : * Merge.
3206 : */
3207 54246 : foreach(lc, rel->partial_pathlist)
3208 : {
3209 29790 : Path *subpath = (Path *) lfirst(lc);
3210 : GatherMergePath *path;
3211 :
3212 29790 : if (subpath->pathkeys == NIL)
3213 23742 : continue;
3214 :
3215 6048 : rows = compute_gather_rows(subpath);
3216 6048 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3217 : subpath->pathkeys, NULL, rowsp);
3218 6048 : add_path(rel, &path->path);
3219 : }
3220 : }
3221 :
3222 : /*
3223 : * get_useful_pathkeys_for_relation
3224 : * Determine which orderings of a relation might be useful.
3225 : *
3226 : * Getting data in sorted order can be useful either because the requested
3227 : * order matches the final output ordering for the overall query we're
3228 : * planning, or because it enables an efficient merge join. Here, we try
3229 : * to figure out which pathkeys to consider.
3230 : *
3231 : * This allows us to do incremental sort on top of an index scan under a gather
3232 : * merge node, i.e. parallelized.
3233 : *
3234 : * If the require_parallel_safe is true, we also require the expressions to
3235 : * be parallel safe (which allows pushing the sort below Gather Merge).
3236 : *
3237 : * XXX At the moment this can only ever return a list with a single element,
3238 : * because it looks at query_pathkeys only. So we might return the pathkeys
3239 : * directly, but it seems plausible we'll want to consider other orderings
3240 : * in the future. For example, we might want to consider pathkeys useful for
3241 : * merge joins.
3242 : */
3243 : static List *
3244 24456 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3245 : bool require_parallel_safe)
3246 : {
3247 24456 : List *useful_pathkeys_list = NIL;
3248 :
3249 : /*
3250 : * Considering query_pathkeys is always worth it, because it might allow
3251 : * us to avoid a total sort when we have a partially presorted path
3252 : * available or to push the total sort into the parallel portion of the
3253 : * query.
3254 : */
3255 24456 : if (root->query_pathkeys)
3256 : {
3257 : ListCell *lc;
3258 14940 : int npathkeys = 0; /* useful pathkeys */
3259 :
3260 26078 : foreach(lc, root->query_pathkeys)
3261 : {
3262 18868 : PathKey *pathkey = (PathKey *) lfirst(lc);
3263 18868 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3264 :
3265 : /*
3266 : * We can only build a sort for pathkeys that contain a
3267 : * safe-to-compute-early EC member computable from the current
3268 : * relation's reltarget, so ignore the remainder of the list as
3269 : * soon as we find a pathkey without such a member.
3270 : *
3271 : * It's still worthwhile to return any prefix of the pathkeys list
3272 : * that meets this requirement, as we may be able to do an
3273 : * incremental sort.
3274 : *
3275 : * If requested, ensure the sort expression is parallel-safe too.
3276 : */
3277 18868 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3278 : require_parallel_safe))
3279 7730 : break;
3280 :
3281 11138 : npathkeys++;
3282 : }
3283 :
3284 : /*
3285 : * The whole query_pathkeys list matches, so append it directly, to
3286 : * allow comparing pathkeys easily by comparing list pointer. If we
3287 : * have to truncate the pathkeys, we gotta do a copy though.
3288 : */
3289 14940 : if (npathkeys == list_length(root->query_pathkeys))
3290 7210 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3291 7210 : root->query_pathkeys);
3292 7730 : else if (npathkeys > 0)
3293 474 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3294 474 : list_copy_head(root->query_pathkeys,
3295 : npathkeys));
3296 : }
3297 :
3298 24456 : return useful_pathkeys_list;
3299 : }
3300 :
3301 : /*
3302 : * generate_useful_gather_paths
3303 : * Generate parallel access paths for a relation by pushing a Gather or
3304 : * Gather Merge on top of a partial path.
3305 : *
3306 : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3307 : * paths (aiming to preserve the ordering), but also considers ordering that
3308 : * might be useful for nodes above the gather merge node, and tries to add
3309 : * a sort (regular or incremental) to provide that.
3310 : */
3311 : void
3312 612466 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3313 : {
3314 : ListCell *lc;
3315 : double rows;
3316 612466 : double *rowsp = NULL;
3317 612466 : List *useful_pathkeys_list = NIL;
3318 612466 : Path *cheapest_partial_path = NULL;
3319 :
3320 : /* If there are no partial paths, there's nothing to do here. */
3321 612466 : if (rel->partial_pathlist == NIL)
3322 588010 : return;
3323 :
3324 : /* Should we override the rel's rowcount estimate? */
3325 24456 : if (override_rows)
3326 6084 : rowsp = &rows;
3327 :
3328 : /* generate the regular gather (merge) paths */
3329 24456 : generate_gather_paths(root, rel, override_rows);
3330 :
3331 : /* consider incremental sort for interesting orderings */
3332 24456 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3333 :
3334 : /* used for explicit (full) sort paths */
3335 24456 : cheapest_partial_path = linitial(rel->partial_pathlist);
3336 :
3337 : /*
3338 : * Consider sorted paths for each interesting ordering. We generate both
3339 : * incremental and full sort.
3340 : */
3341 32140 : foreach(lc, useful_pathkeys_list)
3342 : {
3343 7684 : List *useful_pathkeys = lfirst(lc);
3344 : ListCell *lc2;
3345 : bool is_sorted;
3346 : int presorted_keys;
3347 :
3348 17500 : foreach(lc2, rel->partial_pathlist)
3349 : {
3350 9816 : Path *subpath = (Path *) lfirst(lc2);
3351 : GatherMergePath *path;
3352 :
3353 9816 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3354 : subpath->pathkeys,
3355 : &presorted_keys);
3356 :
3357 : /*
3358 : * We don't need to consider the case where a subpath is already
3359 : * fully sorted because generate_gather_paths already creates a
3360 : * gather merge path for every subpath that has pathkeys present.
3361 : *
3362 : * But since the subpath is already sorted, we know we don't need
3363 : * to consider adding a sort (full or incremental) on top of it,
3364 : * so we can continue here.
3365 : */
3366 9816 : if (is_sorted)
3367 2398 : continue;
3368 :
3369 : /*
3370 : * Try at least sorting the cheapest path and also try
3371 : * incrementally sorting any path which is partially sorted
3372 : * already (no need to deal with paths which have presorted keys
3373 : * when incremental sort is disabled unless it's the cheapest
3374 : * input path).
3375 : */
3376 7418 : if (subpath != cheapest_partial_path &&
3377 246 : (presorted_keys == 0 || !enable_incremental_sort))
3378 102 : continue;
3379 :
3380 : /*
3381 : * Consider regular sort for any path that's not presorted or if
3382 : * incremental sort is disabled. We've no need to consider both
3383 : * sort and incremental sort on the same path. We assume that
3384 : * incremental sort is always faster when there are presorted
3385 : * keys.
3386 : *
3387 : * This is not redundant with the gather paths created in
3388 : * generate_gather_paths, because that doesn't generate ordered
3389 : * output. Here we add an explicit sort to match the useful
3390 : * ordering.
3391 : */
3392 7316 : if (presorted_keys == 0 || !enable_incremental_sort)
3393 : {
3394 7160 : subpath = (Path *) create_sort_path(root,
3395 : rel,
3396 : subpath,
3397 : useful_pathkeys,
3398 : -1.0);
3399 : }
3400 : else
3401 156 : subpath = (Path *) create_incremental_sort_path(root,
3402 : rel,
3403 : subpath,
3404 : useful_pathkeys,
3405 : presorted_keys,
3406 : -1);
3407 7316 : rows = compute_gather_rows(subpath);
3408 7316 : path = create_gather_merge_path(root, rel,
3409 : subpath,
3410 7316 : rel->reltarget,
3411 : subpath->pathkeys,
3412 : NULL,
3413 : rowsp);
3414 :
3415 7316 : add_path(rel, &path->path);
3416 : }
3417 : }
3418 : }
3419 :
3420 : /*
3421 : * generate_grouped_paths
3422 : * Generate paths for a grouped relation by adding sorted and hashed
3423 : * partial aggregation paths on top of paths of the ungrouped relation.
3424 : *
3425 : * The information needed is provided by the RelAggInfo structure stored in
3426 : * "grouped_rel".
3427 : */
3428 : void
3429 898 : generate_grouped_paths(PlannerInfo *root, RelOptInfo *grouped_rel,
3430 : RelOptInfo *rel)
3431 : {
3432 898 : RelAggInfo *agg_info = grouped_rel->agg_info;
3433 : AggClauseCosts agg_costs;
3434 : bool can_hash;
3435 : bool can_sort;
3436 898 : Path *cheapest_total_path = NULL;
3437 898 : Path *cheapest_partial_path = NULL;
3438 898 : double dNumGroups = 0;
3439 898 : double dNumPartialGroups = 0;
3440 898 : List *group_pathkeys = NIL;
3441 :
3442 898 : if (IS_DUMMY_REL(rel))
3443 : {
3444 0 : mark_dummy_rel(grouped_rel);
3445 0 : return;
3446 : }
3447 :
3448 : /*
3449 : * We push partial aggregation only to the lowest possible level in the
3450 : * join tree that is deemed useful.
3451 : */
3452 898 : if (!bms_equal(agg_info->apply_at, rel->relids) ||
3453 898 : !agg_info->agg_useful)
3454 0 : return;
3455 :
3456 5388 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3457 898 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL, &agg_costs);
3458 :
3459 : /*
3460 : * Determine whether it's possible to perform sort-based implementations
3461 : * of grouping, and generate the pathkeys that represent the grouping
3462 : * requirements in that case.
3463 : */
3464 898 : can_sort = grouping_is_sortable(agg_info->group_clauses);
3465 898 : if (can_sort)
3466 : {
3467 : RelOptInfo *top_grouped_rel;
3468 : List *top_group_tlist;
3469 :
3470 502 : top_grouped_rel = IS_OTHER_REL(rel) ?
3471 1400 : rel->top_parent->grouped_rel : grouped_rel;
3472 : top_group_tlist =
3473 898 : make_tlist_from_pathtarget(top_grouped_rel->agg_info->target);
3474 :
3475 : group_pathkeys =
3476 898 : make_pathkeys_for_sortclauses(root, agg_info->group_clauses,
3477 : top_group_tlist);
3478 : }
3479 :
3480 : /*
3481 : * Determine whether we should consider hash-based implementations of
3482 : * grouping.
3483 : */
3484 : Assert(root->numOrderedAggs == 0);
3485 1796 : can_hash = (agg_info->group_clauses != NIL &&
3486 898 : grouping_is_hashable(agg_info->group_clauses));
3487 :
3488 : /*
3489 : * Consider whether we should generate partially aggregated non-partial
3490 : * paths. We can only do this if we have a non-partial path.
3491 : */
3492 898 : if (rel->pathlist != NIL)
3493 : {
3494 898 : cheapest_total_path = rel->cheapest_total_path;
3495 : Assert(cheapest_total_path != NULL);
3496 : }
3497 :
3498 : /*
3499 : * If parallelism is possible for grouped_rel, then we should consider
3500 : * generating partially-grouped partial paths. However, if the ungrouped
3501 : * rel has no partial paths, then we can't.
3502 : */
3503 898 : if (grouped_rel->consider_parallel && rel->partial_pathlist != NIL)
3504 : {
3505 732 : cheapest_partial_path = linitial(rel->partial_pathlist);
3506 : Assert(cheapest_partial_path != NULL);
3507 : }
3508 :
3509 : /* Estimate number of partial groups. */
3510 898 : if (cheapest_total_path != NULL)
3511 898 : dNumGroups = estimate_num_groups(root,
3512 : agg_info->group_exprs,
3513 : cheapest_total_path->rows,
3514 : NULL, NULL);
3515 898 : if (cheapest_partial_path != NULL)
3516 732 : dNumPartialGroups = estimate_num_groups(root,
3517 : agg_info->group_exprs,
3518 : cheapest_partial_path->rows,
3519 : NULL, NULL);
3520 :
3521 898 : if (can_sort && cheapest_total_path != NULL)
3522 : {
3523 : ListCell *lc;
3524 :
3525 : /*
3526 : * Use any available suitably-sorted path as input, and also consider
3527 : * sorting the cheapest-total path and incremental sort on any paths
3528 : * with presorted keys.
3529 : *
3530 : * To save planning time, we ignore parameterized input paths unless
3531 : * they are the cheapest-total path.
3532 : */
3533 2174 : foreach(lc, rel->pathlist)
3534 : {
3535 1276 : Path *input_path = (Path *) lfirst(lc);
3536 : Path *path;
3537 : bool is_sorted;
3538 : int presorted_keys;
3539 :
3540 : /*
3541 : * Ignore parameterized paths that are not the cheapest-total
3542 : * path.
3543 : */
3544 1276 : if (input_path->param_info &&
3545 : input_path != cheapest_total_path)
3546 30 : continue;
3547 :
3548 1270 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3549 : input_path->pathkeys,
3550 : &presorted_keys);
3551 :
3552 : /*
3553 : * Ignore paths that are not suitably or partially sorted, unless
3554 : * they are the cheapest total path (no need to deal with paths
3555 : * which have presorted keys when incremental sort is disabled).
3556 : */
3557 1270 : if (!is_sorted && input_path != cheapest_total_path &&
3558 168 : (presorted_keys == 0 || !enable_incremental_sort))
3559 24 : continue;
3560 :
3561 : /*
3562 : * Since the path originates from a non-grouped relation that is
3563 : * not aware of eager aggregation, we must ensure that it provides
3564 : * the correct input for partial aggregation.
3565 : */
3566 1246 : path = (Path *) create_projection_path(root,
3567 : grouped_rel,
3568 : input_path,
3569 1246 : agg_info->agg_input);
3570 :
3571 1246 : if (!is_sorted)
3572 : {
3573 : /*
3574 : * We've no need to consider both a sort and incremental sort.
3575 : * We'll just do a sort if there are no presorted keys and an
3576 : * incremental sort when there are presorted keys.
3577 : */
3578 1036 : if (presorted_keys == 0 || !enable_incremental_sort)
3579 892 : path = (Path *) create_sort_path(root,
3580 : grouped_rel,
3581 : path,
3582 : group_pathkeys,
3583 : -1.0);
3584 : else
3585 144 : path = (Path *) create_incremental_sort_path(root,
3586 : grouped_rel,
3587 : path,
3588 : group_pathkeys,
3589 : presorted_keys,
3590 : -1.0);
3591 : }
3592 :
3593 : /*
3594 : * qual is NIL because the HAVING clause cannot be evaluated until
3595 : * the final value of the aggregate is known.
3596 : */
3597 1246 : path = (Path *) create_agg_path(root,
3598 : grouped_rel,
3599 : path,
3600 1246 : agg_info->target,
3601 : AGG_SORTED,
3602 : AGGSPLIT_INITIAL_SERIAL,
3603 : agg_info->group_clauses,
3604 : NIL,
3605 : &agg_costs,
3606 : dNumGroups);
3607 :
3608 1246 : add_path(grouped_rel, path);
3609 : }
3610 : }
3611 :
3612 898 : if (can_sort && cheapest_partial_path != NULL)
3613 : {
3614 : ListCell *lc;
3615 :
3616 : /* Similar to above logic, but for partial paths. */
3617 1704 : foreach(lc, rel->partial_pathlist)
3618 : {
3619 972 : Path *input_path = (Path *) lfirst(lc);
3620 : Path *path;
3621 : bool is_sorted;
3622 : int presorted_keys;
3623 :
3624 972 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3625 : input_path->pathkeys,
3626 : &presorted_keys);
3627 :
3628 : /*
3629 : * Ignore paths that are not suitably or partially sorted, unless
3630 : * they are the cheapest partial path (no need to deal with paths
3631 : * which have presorted keys when incremental sort is disabled).
3632 : */
3633 972 : if (!is_sorted && input_path != cheapest_partial_path &&
3634 96 : (presorted_keys == 0 || !enable_incremental_sort))
3635 0 : continue;
3636 :
3637 : /*
3638 : * Since the path originates from a non-grouped relation that is
3639 : * not aware of eager aggregation, we must ensure that it provides
3640 : * the correct input for partial aggregation.
3641 : */
3642 972 : path = (Path *) create_projection_path(root,
3643 : grouped_rel,
3644 : input_path,
3645 972 : agg_info->agg_input);
3646 :
3647 972 : if (!is_sorted)
3648 : {
3649 : /*
3650 : * We've no need to consider both a sort and incremental sort.
3651 : * We'll just do a sort if there are no presorted keys and an
3652 : * incremental sort when there are presorted keys.
3653 : */
3654 828 : if (presorted_keys == 0 || !enable_incremental_sort)
3655 732 : path = (Path *) create_sort_path(root,
3656 : grouped_rel,
3657 : path,
3658 : group_pathkeys,
3659 : -1.0);
3660 : else
3661 96 : path = (Path *) create_incremental_sort_path(root,
3662 : grouped_rel,
3663 : path,
3664 : group_pathkeys,
3665 : presorted_keys,
3666 : -1.0);
3667 : }
3668 :
3669 : /*
3670 : * qual is NIL because the HAVING clause cannot be evaluated until
3671 : * the final value of the aggregate is known.
3672 : */
3673 972 : path = (Path *) create_agg_path(root,
3674 : grouped_rel,
3675 : path,
3676 972 : agg_info->target,
3677 : AGG_SORTED,
3678 : AGGSPLIT_INITIAL_SERIAL,
3679 : agg_info->group_clauses,
3680 : NIL,
3681 : &agg_costs,
3682 : dNumPartialGroups);
3683 :
3684 972 : add_partial_path(grouped_rel, path);
3685 : }
3686 : }
3687 :
3688 : /*
3689 : * Add a partially-grouped HashAgg Path where possible
3690 : */
3691 898 : if (can_hash && cheapest_total_path != NULL)
3692 : {
3693 : Path *path;
3694 :
3695 : /*
3696 : * Since the path originates from a non-grouped relation that is not
3697 : * aware of eager aggregation, we must ensure that it provides the
3698 : * correct input for partial aggregation.
3699 : */
3700 898 : path = (Path *) create_projection_path(root,
3701 : grouped_rel,
3702 : cheapest_total_path,
3703 898 : agg_info->agg_input);
3704 :
3705 : /*
3706 : * qual is NIL because the HAVING clause cannot be evaluated until the
3707 : * final value of the aggregate is known.
3708 : */
3709 898 : path = (Path *) create_agg_path(root,
3710 : grouped_rel,
3711 : path,
3712 898 : agg_info->target,
3713 : AGG_HASHED,
3714 : AGGSPLIT_INITIAL_SERIAL,
3715 : agg_info->group_clauses,
3716 : NIL,
3717 : &agg_costs,
3718 : dNumGroups);
3719 :
3720 898 : add_path(grouped_rel, path);
3721 : }
3722 :
3723 : /*
3724 : * Now add a partially-grouped HashAgg partial Path where possible
3725 : */
3726 898 : if (can_hash && cheapest_partial_path != NULL)
3727 : {
3728 : Path *path;
3729 :
3730 : /*
3731 : * Since the path originates from a non-grouped relation that is not
3732 : * aware of eager aggregation, we must ensure that it provides the
3733 : * correct input for partial aggregation.
3734 : */
3735 732 : path = (Path *) create_projection_path(root,
3736 : grouped_rel,
3737 : cheapest_partial_path,
3738 732 : agg_info->agg_input);
3739 :
3740 : /*
3741 : * qual is NIL because the HAVING clause cannot be evaluated until the
3742 : * final value of the aggregate is known.
3743 : */
3744 732 : path = (Path *) create_agg_path(root,
3745 : grouped_rel,
3746 : path,
3747 732 : agg_info->target,
3748 : AGG_HASHED,
3749 : AGGSPLIT_INITIAL_SERIAL,
3750 : agg_info->group_clauses,
3751 : NIL,
3752 : &agg_costs,
3753 : dNumPartialGroups);
3754 :
3755 732 : add_partial_path(grouped_rel, path);
3756 : }
3757 : }
3758 :
3759 : /*
3760 : * make_rel_from_joinlist
3761 : * Build access paths using a "joinlist" to guide the join path search.
3762 : *
3763 : * See comments for deconstruct_jointree() for definition of the joinlist
3764 : * data structure.
3765 : */
3766 : static RelOptInfo *
3767 328710 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3768 : {
3769 : int levels_needed;
3770 : List *initial_rels;
3771 : ListCell *jl;
3772 :
3773 : /*
3774 : * Count the number of child joinlist nodes. This is the depth of the
3775 : * dynamic-programming algorithm we must employ to consider all ways of
3776 : * joining the child nodes.
3777 : */
3778 328710 : levels_needed = list_length(joinlist);
3779 :
3780 328710 : if (levels_needed <= 0)
3781 0 : return NULL; /* nothing to do? */
3782 :
3783 : /*
3784 : * Construct a list of rels corresponding to the child joinlist nodes.
3785 : * This may contain both base rels and rels constructed according to
3786 : * sub-joinlists.
3787 : */
3788 328710 : initial_rels = NIL;
3789 795340 : foreach(jl, joinlist)
3790 : {
3791 466630 : Node *jlnode = (Node *) lfirst(jl);
3792 : RelOptInfo *thisrel;
3793 :
3794 466630 : if (IsA(jlnode, RangeTblRef))
3795 : {
3796 463252 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3797 :
3798 463252 : thisrel = find_base_rel(root, varno);
3799 : }
3800 3378 : else if (IsA(jlnode, List))
3801 : {
3802 : /* Recurse to handle subproblem */
3803 3378 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3804 : }
3805 : else
3806 : {
3807 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3808 : (int) nodeTag(jlnode));
3809 : thisrel = NULL; /* keep compiler quiet */
3810 : }
3811 :
3812 466630 : initial_rels = lappend(initial_rels, thisrel);
3813 : }
3814 :
3815 328710 : if (levels_needed == 1)
3816 : {
3817 : /*
3818 : * Single joinlist node, so we're done.
3819 : */
3820 227734 : return (RelOptInfo *) linitial(initial_rels);
3821 : }
3822 : else
3823 : {
3824 : /*
3825 : * Consider the different orders in which we could join the rels,
3826 : * using a plugin, GEQO, or the regular join search code.
3827 : *
3828 : * We put the initial_rels list into a PlannerInfo field because
3829 : * has_legal_joinclause() needs to look at it (ugly :-().
3830 : */
3831 100976 : root->initial_rels = initial_rels;
3832 :
3833 100976 : if (join_search_hook)
3834 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
3835 100976 : else if (enable_geqo && levels_needed >= geqo_threshold)
3836 42 : return geqo(root, levels_needed, initial_rels);
3837 : else
3838 100934 : return standard_join_search(root, levels_needed, initial_rels);
3839 : }
3840 : }
3841 :
3842 : /*
3843 : * standard_join_search
3844 : * Find possible joinpaths for a query by successively finding ways
3845 : * to join component relations into join relations.
3846 : *
3847 : * 'levels_needed' is the number of iterations needed, ie, the number of
3848 : * independent jointree items in the query. This is > 1.
3849 : *
3850 : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3851 : * jointree item. These are the components to be joined together.
3852 : * Note that levels_needed == list_length(initial_rels).
3853 : *
3854 : * Returns the final level of join relations, i.e., the relation that is
3855 : * the result of joining all the original relations together.
3856 : * At least one implementation path must be provided for this relation and
3857 : * all required sub-relations.
3858 : *
3859 : * To support loadable plugins that modify planner behavior by changing the
3860 : * join searching algorithm, we provide a hook variable that lets a plugin
3861 : * replace or supplement this function. Any such hook must return the same
3862 : * final join relation as the standard code would, but it might have a
3863 : * different set of implementation paths attached, and only the sub-joinrels
3864 : * needed for these paths need have been instantiated.
3865 : *
3866 : * Note to plugin authors: the functions invoked during standard_join_search()
3867 : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3868 : * than one join-order search, you'll probably need to save and restore the
3869 : * original states of those data structures. See geqo_eval() for an example.
3870 : */
3871 : RelOptInfo *
3872 100934 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3873 : {
3874 : int lev;
3875 : RelOptInfo *rel;
3876 :
3877 : /*
3878 : * This function cannot be invoked recursively within any one planning
3879 : * problem, so join_rel_level[] can't be in use already.
3880 : */
3881 : Assert(root->join_rel_level == NULL);
3882 :
3883 : /*
3884 : * We employ a simple "dynamic programming" algorithm: we first find all
3885 : * ways to build joins of two jointree items, then all ways to build joins
3886 : * of three items (from two-item joins and single items), then four-item
3887 : * joins, and so on until we have considered all ways to join all the
3888 : * items into one rel.
3889 : *
3890 : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3891 : * set root->join_rel_level[1] to represent all the single-jointree-item
3892 : * relations.
3893 : */
3894 100934 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3895 :
3896 100934 : root->join_rel_level[1] = initial_rels;
3897 :
3898 238794 : for (lev = 2; lev <= levels_needed; lev++)
3899 : {
3900 : ListCell *lc;
3901 :
3902 : /*
3903 : * Determine all possible pairs of relations to be joined at this
3904 : * level, and build paths for making each one from every available
3905 : * pair of lower-level relations.
3906 : */
3907 137860 : join_search_one_level(root, lev);
3908 :
3909 : /*
3910 : * Run generate_partitionwise_join_paths() and
3911 : * generate_useful_gather_paths() for each just-processed joinrel. We
3912 : * could not do this earlier because both regular and partial paths
3913 : * can get added to a particular joinrel at multiple times within
3914 : * join_search_one_level.
3915 : *
3916 : * After that, we're done creating paths for the joinrel, so run
3917 : * set_cheapest().
3918 : *
3919 : * In addition, we also run generate_grouped_paths() for the grouped
3920 : * relation of each just-processed joinrel, and run set_cheapest() for
3921 : * the grouped relation afterwards.
3922 : */
3923 348130 : foreach(lc, root->join_rel_level[lev])
3924 : {
3925 : bool is_top_rel;
3926 :
3927 210270 : rel = (RelOptInfo *) lfirst(lc);
3928 :
3929 210270 : is_top_rel = bms_equal(rel->relids, root->all_query_rels);
3930 :
3931 : /* Create paths for partitionwise joins. */
3932 210270 : generate_partitionwise_join_paths(root, rel);
3933 :
3934 : /*
3935 : * Except for the topmost scan/join rel, consider gathering
3936 : * partial paths. We'll do the same for the topmost scan/join rel
3937 : * once we know the final targetlist (see grouping_planner's and
3938 : * its call to apply_scanjoin_target_to_paths).
3939 : */
3940 210270 : if (!is_top_rel)
3941 109834 : generate_useful_gather_paths(root, rel, false);
3942 :
3943 : /* Find and save the cheapest paths for this rel */
3944 210270 : set_cheapest(rel);
3945 :
3946 : /*
3947 : * Except for the topmost scan/join rel, consider generating
3948 : * partial aggregation paths for the grouped relation on top of
3949 : * the paths of this rel. After that, we're done creating paths
3950 : * for the grouped relation, so run set_cheapest().
3951 : */
3952 210270 : if (rel->grouped_rel != NULL && !is_top_rel)
3953 : {
3954 72 : RelOptInfo *grouped_rel = rel->grouped_rel;
3955 :
3956 : Assert(IS_GROUPED_REL(grouped_rel));
3957 :
3958 72 : generate_grouped_paths(root, grouped_rel, rel);
3959 72 : set_cheapest(grouped_rel);
3960 : }
3961 :
3962 : #ifdef OPTIMIZER_DEBUG
3963 : pprint(rel);
3964 : #endif
3965 : }
3966 : }
3967 :
3968 : /*
3969 : * We should have a single rel at the final level.
3970 : */
3971 100934 : if (root->join_rel_level[levels_needed] == NIL)
3972 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
3973 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3974 :
3975 100934 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3976 :
3977 100934 : root->join_rel_level = NULL;
3978 :
3979 100934 : return rel;
3980 : }
3981 :
3982 : /*****************************************************************************
3983 : * PUSHING QUALS DOWN INTO SUBQUERIES
3984 : *****************************************************************************/
3985 :
3986 : /*
3987 : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3988 : *
3989 : * subquery is the particular component query being checked. topquery
3990 : * is the top component of a set-operations tree (the same Query if no
3991 : * set-op is involved).
3992 : *
3993 : * Conditions checked here:
3994 : *
3995 : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3996 : * since that could change the set of rows returned.
3997 : *
3998 : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3999 : * quals into it, because that could change the results.
4000 : *
4001 : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
4002 : * This is because upper-level quals should semantically be evaluated only
4003 : * once per distinct row, not once per original row, and if the qual is
4004 : * volatile then extra evaluations could change the results. (This issue
4005 : * does not apply to other forms of aggregation such as GROUP BY, because
4006 : * when those are present we push into HAVING not WHERE, so that the quals
4007 : * are still applied after aggregation.)
4008 : *
4009 : * 4. If the subquery contains window functions, we cannot push volatile quals
4010 : * into it. The issue here is a bit different from DISTINCT: a volatile qual
4011 : * might succeed for some rows of a window partition and fail for others,
4012 : * thereby changing the partition contents and thus the window functions'
4013 : * results for rows that remain.
4014 : *
4015 : * 5. If the subquery contains any set-returning functions in its targetlist,
4016 : * we cannot push volatile quals into it. That would push them below the SRFs
4017 : * and thereby change the number of times they are evaluated. Also, a
4018 : * volatile qual could succeed for some SRF output rows and fail for others,
4019 : * a behavior that cannot occur if it's evaluated before SRF expansion.
4020 : *
4021 : * 6. If the subquery has nonempty grouping sets, we cannot push down any
4022 : * quals. The concern here is that a qual referencing a "constant" grouping
4023 : * column could get constant-folded, which would be improper because the value
4024 : * is potentially nullable by grouping-set expansion. This restriction could
4025 : * be removed if we had a parsetree representation that shows that such
4026 : * grouping columns are not really constant. (There are other ideas that
4027 : * could be used to relax this restriction, but that's the approach most
4028 : * likely to get taken in the future. Note that there's not much to be gained
4029 : * so long as subquery_planner can't move HAVING clauses to WHERE within such
4030 : * a subquery.)
4031 : *
4032 : * In addition, we make several checks on the subquery's output columns to see
4033 : * if it is safe to reference them in pushed-down quals. If output column k
4034 : * is found to be unsafe to reference, we set the reason for that inside
4035 : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
4036 : * column k might not be referenced by some/all quals. The unsafeFlags[]
4037 : * array will be consulted later by qual_is_pushdown_safe(). It's better to
4038 : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
4039 : * because when the subquery involves set operations we have to check the
4040 : * output expressions in each arm of the set op.
4041 : *
4042 : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
4043 : * we're effectively assuming that the quals cannot distinguish values that
4044 : * the DISTINCT's equality operator sees as equal, yet there are many
4045 : * counterexamples to that assumption. However use of such a qual with a
4046 : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
4047 : * "equal" value will be chosen as the output value by the DISTINCT operation.
4048 : * So we don't worry too much about that. Another objection is that if the
4049 : * qual is expensive to evaluate, running it for each original row might cost
4050 : * more than we save by eliminating rows before the DISTINCT step. But it
4051 : * would be very hard to estimate that at this stage, and in practice pushdown
4052 : * seldom seems to make things worse, so we ignore that problem too.
4053 : *
4054 : * Note: likewise, pushing quals into a subquery with window functions is a
4055 : * bit dubious: the quals might remove some rows of a window partition while
4056 : * leaving others, causing changes in the window functions' results for the
4057 : * surviving rows. We insist that such a qual reference only partitioning
4058 : * columns, but again that only protects us if the qual does not distinguish
4059 : * values that the partitioning equality operator sees as equal. The risks
4060 : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
4061 : * occurs and thus there is no theoretical problem with such a qual. But
4062 : * we'll do this anyway because the potential performance benefits are very
4063 : * large, and we've seen no field complaints about the longstanding comparable
4064 : * behavior with DISTINCT.
4065 : */
4066 : static bool
4067 2180 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
4068 : pushdown_safety_info *safetyInfo)
4069 : {
4070 : SetOperationStmt *topop;
4071 :
4072 : /* Check point 1 */
4073 2180 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
4074 134 : return false;
4075 :
4076 : /* Check point 6 */
4077 2046 : if (subquery->groupClause && subquery->groupingSets)
4078 12 : return false;
4079 :
4080 : /* Check points 3, 4, and 5 */
4081 2034 : if (subquery->distinctClause ||
4082 1950 : subquery->hasWindowFuncs ||
4083 1684 : subquery->hasTargetSRFs)
4084 546 : safetyInfo->unsafeVolatile = true;
4085 :
4086 : /*
4087 : * If we're at a leaf query, check for unsafe expressions in its target
4088 : * list, and mark any reasons why they're unsafe in unsafeFlags[].
4089 : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
4090 : * so no need to check them.)
4091 : */
4092 2034 : if (subquery->setOperations == NULL)
4093 1878 : check_output_expressions(subquery, safetyInfo);
4094 :
4095 : /* Are we at top level, or looking at a setop component? */
4096 2034 : if (subquery == topquery)
4097 : {
4098 : /* Top level, so check any component queries */
4099 1722 : if (subquery->setOperations != NULL)
4100 156 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
4101 : safetyInfo))
4102 0 : return false;
4103 : }
4104 : else
4105 : {
4106 : /* Setop component must not have more components (too weird) */
4107 312 : if (subquery->setOperations != NULL)
4108 0 : return false;
4109 : /* Check whether setop component output types match top level */
4110 312 : topop = castNode(SetOperationStmt, topquery->setOperations);
4111 : Assert(topop);
4112 312 : compare_tlist_datatypes(subquery->targetList,
4113 : topop->colTypes,
4114 : safetyInfo);
4115 : }
4116 2034 : return true;
4117 : }
4118 :
4119 : /*
4120 : * Helper routine to recurse through setOperations tree
4121 : */
4122 : static bool
4123 468 : recurse_pushdown_safe(Node *setOp, Query *topquery,
4124 : pushdown_safety_info *safetyInfo)
4125 : {
4126 468 : if (IsA(setOp, RangeTblRef))
4127 : {
4128 312 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4129 312 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
4130 312 : Query *subquery = rte->subquery;
4131 :
4132 : Assert(subquery != NULL);
4133 312 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
4134 : }
4135 156 : else if (IsA(setOp, SetOperationStmt))
4136 : {
4137 156 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4138 :
4139 : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
4140 156 : if (op->op == SETOP_EXCEPT)
4141 0 : return false;
4142 : /* Else recurse */
4143 156 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
4144 0 : return false;
4145 156 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
4146 0 : return false;
4147 : }
4148 : else
4149 : {
4150 0 : elog(ERROR, "unrecognized node type: %d",
4151 : (int) nodeTag(setOp));
4152 : }
4153 156 : return true;
4154 : }
4155 :
4156 : /*
4157 : * check_output_expressions - check subquery's output expressions for safety
4158 : *
4159 : * There are several cases in which it's unsafe to push down an upper-level
4160 : * qual if it references a particular output column of a subquery. We check
4161 : * each output column of the subquery and set flags in unsafeFlags[k] when we
4162 : * see that column is unsafe for a pushed-down qual to reference. The
4163 : * conditions checked here are:
4164 : *
4165 : * 1. We must not push down any quals that refer to subselect outputs that
4166 : * return sets, else we'd introduce functions-returning-sets into the
4167 : * subquery's WHERE/HAVING quals.
4168 : *
4169 : * 2. We must not push down any quals that refer to subselect outputs that
4170 : * contain volatile functions, for fear of introducing strange results due
4171 : * to multiple evaluation of a volatile function.
4172 : *
4173 : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
4174 : * refer to non-DISTINCT output columns, because that could change the set
4175 : * of rows returned. (This condition is vacuous for DISTINCT, because then
4176 : * there are no non-DISTINCT output columns, so we needn't check. Note that
4177 : * subquery_is_pushdown_safe already reported that we can't use volatile
4178 : * quals if there's DISTINCT or DISTINCT ON.)
4179 : *
4180 : * 4. If the subquery has any window functions, we must not push down quals
4181 : * that reference any output columns that are not listed in all the subquery's
4182 : * window PARTITION BY clauses. We can push down quals that use only
4183 : * partitioning columns because they should succeed or fail identically for
4184 : * every row of any one window partition, and totally excluding some
4185 : * partitions will not change a window function's results for remaining
4186 : * partitions. (Again, this also requires nonvolatile quals, but
4187 : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
4188 : * unsafe for this reason can still have WindowClause run conditions pushed
4189 : * down.
4190 : */
4191 : static void
4192 1878 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
4193 : {
4194 : ListCell *lc;
4195 :
4196 18022 : foreach(lc, subquery->targetList)
4197 : {
4198 16144 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4199 :
4200 16144 : if (tle->resjunk)
4201 134 : continue; /* ignore resjunk columns */
4202 :
4203 : /* Functions returning sets are unsafe (point 1) */
4204 16010 : if (subquery->hasTargetSRFs &&
4205 668 : (safetyInfo->unsafeFlags[tle->resno] &
4206 668 : UNSAFE_HAS_SET_FUNC) == 0 &&
4207 668 : expression_returns_set((Node *) tle->expr))
4208 : {
4209 376 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
4210 376 : continue;
4211 : }
4212 :
4213 : /* Volatile functions are unsafe (point 2) */
4214 15634 : if ((safetyInfo->unsafeFlags[tle->resno] &
4215 15622 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
4216 15622 : contain_volatile_functions((Node *) tle->expr))
4217 : {
4218 78 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
4219 78 : continue;
4220 : }
4221 :
4222 : /* If subquery uses DISTINCT ON, check point 3 */
4223 15556 : if (subquery->hasDistinctOn &&
4224 0 : (safetyInfo->unsafeFlags[tle->resno] &
4225 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4226 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
4227 : {
4228 : /* non-DISTINCT column, so mark it unsafe */
4229 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
4230 0 : continue;
4231 : }
4232 :
4233 : /* If subquery uses window functions, check point 4 */
4234 15556 : if (subquery->hasWindowFuncs &&
4235 1090 : (safetyInfo->unsafeFlags[tle->resno] &
4236 2092 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4237 1090 : !targetIsInAllPartitionLists(tle, subquery))
4238 : {
4239 : /* not present in all PARTITION BY clauses, so mark it unsafe */
4240 1002 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
4241 1002 : continue;
4242 : }
4243 : }
4244 1878 : }
4245 :
4246 : /*
4247 : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
4248 : * push quals into each component query, but the quals can only reference
4249 : * subquery columns that suffer no type coercions in the set operation.
4250 : * Otherwise there are possible semantic gotchas. So, we check the
4251 : * component queries to see if any of them have output types different from
4252 : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
4253 : * unsafeFlags[k] if column k has different type in any component.
4254 : *
4255 : * We don't have to care about typmods here: the only allowed difference
4256 : * between set-op input and output typmods is input is a specific typmod
4257 : * and output is -1, and that does not require a coercion.
4258 : *
4259 : * tlist is a subquery tlist.
4260 : * colTypes is an OID list of the top-level setop's output column types.
4261 : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
4262 : */
4263 : static void
4264 312 : compare_tlist_datatypes(List *tlist, List *colTypes,
4265 : pushdown_safety_info *safetyInfo)
4266 : {
4267 : ListCell *l;
4268 312 : ListCell *colType = list_head(colTypes);
4269 :
4270 984 : foreach(l, tlist)
4271 : {
4272 672 : TargetEntry *tle = (TargetEntry *) lfirst(l);
4273 :
4274 672 : if (tle->resjunk)
4275 0 : continue; /* ignore resjunk columns */
4276 672 : if (colType == NULL)
4277 0 : elog(ERROR, "wrong number of tlist entries");
4278 672 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
4279 104 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
4280 672 : colType = lnext(colTypes, colType);
4281 : }
4282 312 : if (colType != NULL)
4283 0 : elog(ERROR, "wrong number of tlist entries");
4284 312 : }
4285 :
4286 : /*
4287 : * targetIsInAllPartitionLists
4288 : * True if the TargetEntry is listed in the PARTITION BY clause
4289 : * of every window defined in the query.
4290 : *
4291 : * It would be safe to ignore windows not actually used by any window
4292 : * function, but it's not easy to get that info at this stage; and it's
4293 : * unlikely to be useful to spend any extra cycles getting it, since
4294 : * unreferenced window definitions are probably infrequent in practice.
4295 : */
4296 : static bool
4297 1090 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
4298 : {
4299 : ListCell *lc;
4300 :
4301 1202 : foreach(lc, query->windowClause)
4302 : {
4303 1114 : WindowClause *wc = (WindowClause *) lfirst(lc);
4304 :
4305 1114 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
4306 1002 : return false;
4307 : }
4308 88 : return true;
4309 : }
4310 :
4311 : /*
4312 : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
4313 : *
4314 : * rinfo is a restriction clause applying to the given subquery (whose RTE
4315 : * has index rti in the parent query).
4316 : *
4317 : * Conditions checked here:
4318 : *
4319 : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
4320 : * unclear that it will work correctly: SubLinks will already have been
4321 : * transformed into SubPlans in the qual, but not in the subquery). Note that
4322 : * SubLinks that transform to initplans are safe, and will be accepted here
4323 : * because what we'll see in the qual is just a Param referencing the initplan
4324 : * output.
4325 : *
4326 : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
4327 : * functions.
4328 : *
4329 : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
4330 : * functions that are passed Var nodes, and therefore might reveal values from
4331 : * the subquery as side effects.
4332 : *
4333 : * 4. rinfo's clause must not refer to the whole-row output of the subquery
4334 : * (since there is no easy way to name that within the subquery itself).
4335 : *
4336 : * 5. rinfo's clause must not refer to any subquery output columns that were
4337 : * found to be unsafe to reference by subquery_is_pushdown_safe().
4338 : */
4339 : static pushdown_safe_type
4340 2670 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
4341 : pushdown_safety_info *safetyInfo)
4342 : {
4343 2670 : pushdown_safe_type safe = PUSHDOWN_SAFE;
4344 2670 : Node *qual = (Node *) rinfo->clause;
4345 : List *vars;
4346 : ListCell *vl;
4347 :
4348 : /* Refuse subselects (point 1) */
4349 2670 : if (contain_subplans(qual))
4350 66 : return PUSHDOWN_UNSAFE;
4351 :
4352 : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
4353 3256 : if (safetyInfo->unsafeVolatile &&
4354 652 : contain_volatile_functions((Node *) rinfo))
4355 18 : return PUSHDOWN_UNSAFE;
4356 :
4357 : /* Refuse leaky quals if told to (point 3) */
4358 3768 : if (safetyInfo->unsafeLeaky &&
4359 1182 : contain_leaked_vars(qual))
4360 162 : return PUSHDOWN_UNSAFE;
4361 :
4362 : /*
4363 : * Examine all Vars used in clause. Since it's a restriction clause, all
4364 : * such Vars must refer to subselect output columns ... unless this is
4365 : * part of a LATERAL subquery, in which case there could be lateral
4366 : * references.
4367 : *
4368 : * By omitting the relevant flags, this also gives us a cheap sanity check
4369 : * that no aggregates or window functions appear in the qual. Those would
4370 : * be unsafe to push down, but at least for the moment we could never see
4371 : * any in a qual anyhow.
4372 : */
4373 2424 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
4374 4748 : foreach(vl, vars)
4375 : {
4376 2532 : Var *var = (Var *) lfirst(vl);
4377 :
4378 : /*
4379 : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
4380 : * It's not clear whether a PHV could safely be pushed down, and even
4381 : * less clear whether such a situation could arise in any cases of
4382 : * practical interest anyway. So for the moment, just refuse to push
4383 : * down.
4384 : */
4385 2532 : if (!IsA(var, Var))
4386 : {
4387 0 : safe = PUSHDOWN_UNSAFE;
4388 0 : break;
4389 : }
4390 :
4391 : /*
4392 : * Punt if we find any lateral references. It would be safe to push
4393 : * these down, but we'd have to convert them into outer references,
4394 : * which subquery_push_qual lacks the infrastructure to do. The case
4395 : * arises so seldom that it doesn't seem worth working hard on.
4396 : */
4397 2532 : if (var->varno != rti)
4398 : {
4399 12 : safe = PUSHDOWN_UNSAFE;
4400 12 : break;
4401 : }
4402 :
4403 : /* Subqueries have no system columns */
4404 : Assert(var->varattno >= 0);
4405 :
4406 : /* Check point 4 */
4407 2520 : if (var->varattno == 0)
4408 : {
4409 0 : safe = PUSHDOWN_UNSAFE;
4410 0 : break;
4411 : }
4412 :
4413 : /* Check point 5 */
4414 2520 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
4415 : {
4416 526 : if (safetyInfo->unsafeFlags[var->varattno] &
4417 : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
4418 : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
4419 : {
4420 196 : safe = PUSHDOWN_UNSAFE;
4421 196 : break;
4422 : }
4423 : else
4424 : {
4425 : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
4426 330 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
4427 : /* don't break, we might find another Var that's unsafe */
4428 : }
4429 : }
4430 : }
4431 :
4432 2424 : list_free(vars);
4433 :
4434 2424 : return safe;
4435 : }
4436 :
4437 : /*
4438 : * subquery_push_qual - push down a qual that we have determined is safe
4439 : */
4440 : static void
4441 2228 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
4442 : {
4443 2228 : if (subquery->setOperations != NULL)
4444 : {
4445 : /* Recurse to push it separately to each component query */
4446 132 : recurse_push_qual(subquery->setOperations, subquery,
4447 : rte, rti, qual);
4448 : }
4449 : else
4450 : {
4451 : /*
4452 : * We need to replace Vars in the qual (which must refer to outputs of
4453 : * the subquery) with copies of the subquery's targetlist expressions.
4454 : * Note that at this point, any uplevel Vars in the qual should have
4455 : * been replaced with Params, so they need no work.
4456 : *
4457 : * This step also ensures that when we are pushing into a setop tree,
4458 : * each component query gets its own copy of the qual.
4459 : */
4460 2096 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
4461 : subquery->targetList,
4462 : subquery->resultRelation,
4463 : REPLACEVARS_REPORT_ERROR, 0,
4464 : &subquery->hasSubLinks);
4465 :
4466 : /*
4467 : * Now attach the qual to the proper place: normally WHERE, but if the
4468 : * subquery uses grouping or aggregation, put it in HAVING (since the
4469 : * qual really refers to the group-result rows).
4470 : */
4471 2096 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
4472 278 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
4473 : else
4474 1818 : subquery->jointree->quals =
4475 1818 : make_and_qual(subquery->jointree->quals, qual);
4476 :
4477 : /*
4478 : * We need not change the subquery's hasAggs or hasSubLinks flags,
4479 : * since we can't be pushing down any aggregates that weren't there
4480 : * before, and we don't push down subselects at all.
4481 : */
4482 : }
4483 2228 : }
4484 :
4485 : /*
4486 : * Helper routine to recurse through setOperations tree
4487 : */
4488 : static void
4489 396 : recurse_push_qual(Node *setOp, Query *topquery,
4490 : RangeTblEntry *rte, Index rti, Node *qual)
4491 : {
4492 396 : if (IsA(setOp, RangeTblRef))
4493 : {
4494 264 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4495 264 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4496 264 : Query *subquery = subrte->subquery;
4497 :
4498 : Assert(subquery != NULL);
4499 264 : subquery_push_qual(subquery, rte, rti, qual);
4500 : }
4501 132 : else if (IsA(setOp, SetOperationStmt))
4502 : {
4503 132 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4504 :
4505 132 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4506 132 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4507 : }
4508 : else
4509 : {
4510 0 : elog(ERROR, "unrecognized node type: %d",
4511 : (int) nodeTag(setOp));
4512 : }
4513 396 : }
4514 :
4515 : /*****************************************************************************
4516 : * SIMPLIFYING SUBQUERY TARGETLISTS
4517 : *****************************************************************************/
4518 :
4519 : /*
4520 : * remove_unused_subquery_outputs
4521 : * Remove subquery targetlist items we don't need
4522 : *
4523 : * It's possible, even likely, that the upper query does not read all the
4524 : * output columns of the subquery. We can remove any such outputs that are
4525 : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4526 : * affect semantics otherwise (e.g., volatile functions can't be removed).
4527 : * This is useful not only because we might be able to remove expensive-to-
4528 : * compute expressions, but because deletion of output columns might allow
4529 : * optimizations such as join removal to occur within the subquery.
4530 : *
4531 : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4532 : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4533 : * parameter is modified by the function, so callers must make a copy if they
4534 : * need to use the passed in Bitmapset after calling this function.
4535 : *
4536 : * To avoid affecting column numbering in the targetlist, we don't physically
4537 : * remove unused tlist entries, but rather replace their expressions with NULL
4538 : * constants. This is implemented by modifying subquery->targetList.
4539 : */
4540 : static void
4541 16760 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4542 : Bitmapset *extra_used_attrs)
4543 : {
4544 : Bitmapset *attrs_used;
4545 : ListCell *lc;
4546 :
4547 : /*
4548 : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4549 : * the current callers use the Bitmapset after calling this function.
4550 : */
4551 16760 : attrs_used = extra_used_attrs;
4552 :
4553 : /*
4554 : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4555 : * could update all the child SELECTs' tlists, but it seems not worth the
4556 : * trouble presently.
4557 : */
4558 16760 : if (subquery->setOperations)
4559 1386 : return;
4560 :
4561 : /*
4562 : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4563 : * time: all its output columns must be used in the distinctClause.
4564 : */
4565 15924 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4566 246 : return;
4567 :
4568 : /*
4569 : * Collect a bitmap of all the output column numbers used by the upper
4570 : * query.
4571 : *
4572 : * Add all the attributes needed for joins or final output. Note: we must
4573 : * look at rel's targetlist, not the attr_needed data, because attr_needed
4574 : * isn't computed for inheritance child rels, cf set_append_rel_size().
4575 : * (XXX might be worth changing that sometime.)
4576 : */
4577 15678 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4578 :
4579 : /* Add all the attributes used by un-pushed-down restriction clauses. */
4580 16414 : foreach(lc, rel->baserestrictinfo)
4581 : {
4582 736 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4583 :
4584 736 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4585 : }
4586 :
4587 : /*
4588 : * If there's a whole-row reference to the subquery, we can't remove
4589 : * anything.
4590 : */
4591 15678 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4592 304 : return;
4593 :
4594 : /*
4595 : * Run through the tlist and zap entries we don't need. It's okay to
4596 : * modify the tlist items in-place because set_subquery_pathlist made a
4597 : * copy of the subquery.
4598 : */
4599 87656 : foreach(lc, subquery->targetList)
4600 : {
4601 72282 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4602 72282 : Node *texpr = (Node *) tle->expr;
4603 :
4604 : /*
4605 : * If it has a sortgroupref number, it's used in some sort/group
4606 : * clause so we'd better not remove it. Also, don't remove any
4607 : * resjunk columns, since their reason for being has nothing to do
4608 : * with anybody reading the subquery's output. (It's likely that
4609 : * resjunk columns in a sub-SELECT would always have ressortgroupref
4610 : * set, but even if they don't, it seems imprudent to remove them.)
4611 : */
4612 72282 : if (tle->ressortgroupref || tle->resjunk)
4613 2622 : continue;
4614 :
4615 : /*
4616 : * If it's used by the upper query, we can't remove it.
4617 : */
4618 69660 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4619 : attrs_used))
4620 47150 : continue;
4621 :
4622 : /*
4623 : * If it contains a set-returning function, we can't remove it since
4624 : * that could change the number of rows returned by the subquery.
4625 : */
4626 23550 : if (subquery->hasTargetSRFs &&
4627 1040 : expression_returns_set(texpr))
4628 772 : continue;
4629 :
4630 : /*
4631 : * If it contains volatile functions, we daren't remove it for fear
4632 : * that the user is expecting their side-effects to happen.
4633 : */
4634 21738 : if (contain_volatile_functions(texpr))
4635 32 : continue;
4636 :
4637 : /*
4638 : * OK, we don't need it. Replace the expression with a NULL constant.
4639 : * Preserve the exposed type of the expression, in case something
4640 : * looks at the rowtype of the subquery's result.
4641 : */
4642 21706 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4643 : exprTypmod(texpr),
4644 : exprCollation(texpr));
4645 : }
4646 : }
4647 :
4648 : /*
4649 : * create_partial_bitmap_paths
4650 : * Build partial bitmap heap path for the relation
4651 : */
4652 : void
4653 144430 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4654 : Path *bitmapqual)
4655 : {
4656 : int parallel_workers;
4657 : double pages_fetched;
4658 :
4659 : /* Compute heap pages for bitmap heap scan */
4660 144430 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4661 : NULL, NULL);
4662 :
4663 144430 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4664 : max_parallel_workers_per_gather);
4665 :
4666 144430 : if (parallel_workers <= 0)
4667 140242 : return;
4668 :
4669 4188 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4670 : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4671 : }
4672 :
4673 : /*
4674 : * Compute the number of parallel workers that should be used to scan a
4675 : * relation. We compute the parallel workers based on the size of the heap to
4676 : * be scanned and the size of the index to be scanned, then choose a minimum
4677 : * of those.
4678 : *
4679 : * "heap_pages" is the number of pages from the table that we expect to scan, or
4680 : * -1 if we don't expect to scan any.
4681 : *
4682 : * "index_pages" is the number of pages from the index that we expect to scan, or
4683 : * -1 if we don't expect to scan any.
4684 : *
4685 : * "max_workers" is caller's limit on the number of workers. This typically
4686 : * comes from a GUC.
4687 : */
4688 : int
4689 757162 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4690 : int max_workers)
4691 : {
4692 757162 : int parallel_workers = 0;
4693 :
4694 : /*
4695 : * If the user has set the parallel_workers reloption, use that; otherwise
4696 : * select a default number of workers.
4697 : */
4698 757162 : if (rel->rel_parallel_workers != -1)
4699 1914 : parallel_workers = rel->rel_parallel_workers;
4700 : else
4701 : {
4702 : /*
4703 : * If the number of pages being scanned is insufficient to justify a
4704 : * parallel scan, just return zero ... unless it's an inheritance
4705 : * child. In that case, we want to generate a parallel path here
4706 : * anyway. It might not be worthwhile just for this relation, but
4707 : * when combined with all of its inheritance siblings it may well pay
4708 : * off.
4709 : */
4710 755248 : if (rel->reloptkind == RELOPT_BASEREL &&
4711 715734 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4712 23402 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4713 714950 : return 0;
4714 :
4715 40298 : if (heap_pages >= 0)
4716 : {
4717 : int heap_parallel_threshold;
4718 38164 : int heap_parallel_workers = 1;
4719 :
4720 : /*
4721 : * Select the number of workers based on the log of the size of
4722 : * the relation. This probably needs to be a good deal more
4723 : * sophisticated, but we need something here for now. Note that
4724 : * the upper limit of the min_parallel_table_scan_size GUC is
4725 : * chosen to prevent overflow here.
4726 : */
4727 38164 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4728 43028 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4729 : {
4730 4864 : heap_parallel_workers++;
4731 4864 : heap_parallel_threshold *= 3;
4732 4864 : if (heap_parallel_threshold > INT_MAX / 3)
4733 0 : break; /* avoid overflow */
4734 : }
4735 :
4736 38164 : parallel_workers = heap_parallel_workers;
4737 : }
4738 :
4739 40298 : if (index_pages >= 0)
4740 : {
4741 9824 : int index_parallel_workers = 1;
4742 : int index_parallel_threshold;
4743 :
4744 : /* same calculation as for heap_pages above */
4745 9824 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4746 10100 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4747 : {
4748 276 : index_parallel_workers++;
4749 276 : index_parallel_threshold *= 3;
4750 276 : if (index_parallel_threshold > INT_MAX / 3)
4751 0 : break; /* avoid overflow */
4752 : }
4753 :
4754 9824 : if (parallel_workers > 0)
4755 7690 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4756 : else
4757 2134 : parallel_workers = index_parallel_workers;
4758 : }
4759 : }
4760 :
4761 : /* In no case use more than caller supplied maximum number of workers */
4762 42212 : parallel_workers = Min(parallel_workers, max_workers);
4763 :
4764 42212 : return parallel_workers;
4765 : }
4766 :
4767 : /*
4768 : * generate_partitionwise_join_paths
4769 : * Create paths representing partitionwise join for given partitioned
4770 : * join relation.
4771 : *
4772 : * This must not be called until after we are done adding paths for all
4773 : * child-joins. Otherwise, add_path might delete a path to which some path
4774 : * generated here has a reference.
4775 : */
4776 : void
4777 235240 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4778 : {
4779 235240 : List *live_children = NIL;
4780 : int cnt_parts;
4781 : int num_parts;
4782 : RelOptInfo **part_rels;
4783 :
4784 : /* Handle only join relations here. */
4785 235240 : if (!IS_JOIN_REL(rel))
4786 0 : return;
4787 :
4788 : /* We've nothing to do if the relation is not partitioned. */
4789 235240 : if (!IS_PARTITIONED_REL(rel))
4790 228094 : return;
4791 :
4792 : /* The relation should have consider_partitionwise_join set. */
4793 : Assert(rel->consider_partitionwise_join);
4794 :
4795 : /* Guard against stack overflow due to overly deep partition hierarchy. */
4796 7146 : check_stack_depth();
4797 :
4798 7146 : num_parts = rel->nparts;
4799 7146 : part_rels = rel->part_rels;
4800 :
4801 : /* Collect non-dummy child-joins. */
4802 25448 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4803 : {
4804 18302 : RelOptInfo *child_rel = part_rels[cnt_parts];
4805 :
4806 : /* If it's been pruned entirely, it's certainly dummy. */
4807 18302 : if (child_rel == NULL)
4808 64 : continue;
4809 :
4810 : /* Make partitionwise join paths for this partitioned child-join. */
4811 18238 : generate_partitionwise_join_paths(root, child_rel);
4812 :
4813 : /* If we failed to make any path for this child, we must give up. */
4814 18238 : if (child_rel->pathlist == NIL)
4815 : {
4816 : /*
4817 : * Mark the parent joinrel as unpartitioned so that later
4818 : * functions treat it correctly.
4819 : */
4820 0 : rel->nparts = 0;
4821 0 : return;
4822 : }
4823 :
4824 : /* Else, identify the cheapest path for it. */
4825 18238 : set_cheapest(child_rel);
4826 :
4827 : /* Dummy children need not be scanned, so ignore those. */
4828 18238 : if (IS_DUMMY_REL(child_rel))
4829 0 : continue;
4830 :
4831 : /*
4832 : * Except for the topmost scan/join rel, consider generating partial
4833 : * aggregation paths for the grouped relation on top of the paths of
4834 : * this partitioned child-join. After that, we're done creating paths
4835 : * for the grouped relation, so run set_cheapest().
4836 : */
4837 18238 : if (child_rel->grouped_rel != NULL &&
4838 12876 : !bms_equal(IS_OTHER_REL(rel) ?
4839 : rel->top_parent_relids : rel->relids,
4840 12876 : root->all_query_rels))
4841 : {
4842 240 : RelOptInfo *grouped_rel = child_rel->grouped_rel;
4843 :
4844 : Assert(IS_GROUPED_REL(grouped_rel));
4845 :
4846 240 : generate_grouped_paths(root, grouped_rel, child_rel);
4847 240 : set_cheapest(grouped_rel);
4848 : }
4849 :
4850 : #ifdef OPTIMIZER_DEBUG
4851 : pprint(child_rel);
4852 : #endif
4853 :
4854 18238 : live_children = lappend(live_children, child_rel);
4855 : }
4856 :
4857 : /* If all child-joins are dummy, parent join is also dummy. */
4858 7146 : if (!live_children)
4859 : {
4860 0 : mark_dummy_rel(rel);
4861 0 : return;
4862 : }
4863 :
4864 : /* Build additional paths for this rel from child-join paths. */
4865 7146 : add_paths_to_append_rel(root, rel, live_children);
4866 7146 : list_free(live_children);
4867 : }
|
