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) 1996-2024, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/optimizer/path/allpaths.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 :
16 : #include "postgres.h"
17 :
18 : #include <limits.h>
19 : #include <math.h>
20 :
21 : #include "access/sysattr.h"
22 : #include "access/tsmapi.h"
23 : #include "catalog/pg_class.h"
24 : #include "catalog/pg_operator.h"
25 : #include "catalog/pg_proc.h"
26 : #include "foreign/fdwapi.h"
27 : #include "miscadmin.h"
28 : #include "nodes/makefuncs.h"
29 : #include "nodes/nodeFuncs.h"
30 : #include "nodes/supportnodes.h"
31 : #ifdef OPTIMIZER_DEBUG
32 : #include "nodes/print.h"
33 : #endif
34 : #include "optimizer/appendinfo.h"
35 : #include "optimizer/clauses.h"
36 : #include "optimizer/cost.h"
37 : #include "optimizer/geqo.h"
38 : #include "optimizer/optimizer.h"
39 : #include "optimizer/pathnode.h"
40 : #include "optimizer/paths.h"
41 : #include "optimizer/plancat.h"
42 : #include "optimizer/planner.h"
43 : #include "optimizer/tlist.h"
44 : #include "parser/parse_clause.h"
45 : #include "parser/parsetree.h"
46 : #include "partitioning/partbounds.h"
47 : #include "port/pg_bitutils.h"
48 : #include "rewrite/rewriteManip.h"
49 : #include "utils/lsyscache.h"
50 :
51 :
52 : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
53 : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
54 : #define UNSAFE_HAS_SET_FUNC (1 << 1)
55 : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
56 : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
57 : #define UNSAFE_TYPE_MISMATCH (1 << 4)
58 :
59 : /* results of subquery_is_pushdown_safe */
60 : typedef struct pushdown_safety_info
61 : {
62 : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
63 : * column is unsafe for qual pushdown, or 0 if
64 : * no reason. */
65 : bool unsafeVolatile; /* don't push down volatile quals */
66 : bool unsafeLeaky; /* don't push down leaky quals */
67 : } pushdown_safety_info;
68 :
69 : /* Return type for qual_is_pushdown_safe */
70 : typedef enum pushdown_safe_type
71 : {
72 : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
73 : PUSHDOWN_SAFE, /* safe to push qual into subquery */
74 : PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
75 : * run condition */
76 : } pushdown_safe_type;
77 :
78 : /* These parameters are set by GUC */
79 : bool enable_geqo = false; /* just in case GUC doesn't set it */
80 : int geqo_threshold;
81 : int min_parallel_table_scan_size;
82 : int min_parallel_index_scan_size;
83 :
84 : /* Hook for plugins to get control in set_rel_pathlist() */
85 : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
86 :
87 : /* Hook for plugins to replace standard_join_search() */
88 : join_search_hook_type join_search_hook = NULL;
89 :
90 :
91 : static void set_base_rel_consider_startup(PlannerInfo *root);
92 : static void set_base_rel_sizes(PlannerInfo *root);
93 : static void set_base_rel_pathlists(PlannerInfo *root);
94 : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
95 : Index rti, RangeTblEntry *rte);
96 : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
97 : Index rti, RangeTblEntry *rte);
98 : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
99 : RangeTblEntry *rte);
100 : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
101 : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
102 : RangeTblEntry *rte);
103 : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
104 : RangeTblEntry *rte);
105 : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
106 : RangeTblEntry *rte);
107 : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
108 : RangeTblEntry *rte);
109 : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
110 : RangeTblEntry *rte);
111 : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
112 : RangeTblEntry *rte);
113 : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
114 : Index rti, RangeTblEntry *rte);
115 : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
116 : Index rti, RangeTblEntry *rte);
117 : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
118 : List *live_childrels,
119 : List *all_child_pathkeys);
120 : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
121 : RelOptInfo *rel,
122 : Relids required_outer);
123 : static void accumulate_append_subpath(Path *path,
124 : List **subpaths,
125 : List **special_subpaths);
126 : static Path *get_singleton_append_subpath(Path *path);
127 : static void set_dummy_rel_pathlist(RelOptInfo *rel);
128 : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
129 : Index rti, RangeTblEntry *rte);
130 : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
131 : RangeTblEntry *rte);
132 : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
133 : RangeTblEntry *rte);
134 : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
135 : RangeTblEntry *rte);
136 : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : RangeTblEntry *rte);
138 : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : RangeTblEntry *rte);
140 : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : RangeTblEntry *rte);
142 : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : RangeTblEntry *rte);
144 : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
145 : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
146 : pushdown_safety_info *safetyInfo);
147 : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
148 : pushdown_safety_info *safetyInfo);
149 : static void check_output_expressions(Query *subquery,
150 : pushdown_safety_info *safetyInfo);
151 : static void compare_tlist_datatypes(List *tlist, List *colTypes,
152 : pushdown_safety_info *safetyInfo);
153 : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
154 : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
155 : RestrictInfo *rinfo,
156 : pushdown_safety_info *safetyInfo);
157 : static void subquery_push_qual(Query *subquery,
158 : RangeTblEntry *rte, Index rti, Node *qual);
159 : static void recurse_push_qual(Node *setOp, Query *topquery,
160 : RangeTblEntry *rte, Index rti, Node *qual);
161 : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
162 : Bitmapset *extra_used_attrs);
163 :
164 :
165 : /*
166 : * make_one_rel
167 : * Finds all possible access paths for executing a query, returning a
168 : * single rel that represents the join of all base rels in the query.
169 : */
170 : RelOptInfo *
171 291858 : make_one_rel(PlannerInfo *root, List *joinlist)
172 : {
173 : RelOptInfo *rel;
174 : Index rti;
175 : double total_pages;
176 :
177 : /* Mark base rels as to whether we care about fast-start plans */
178 291858 : set_base_rel_consider_startup(root);
179 :
180 : /*
181 : * Compute size estimates and consider_parallel flags for each base rel.
182 : */
183 291858 : set_base_rel_sizes(root);
184 :
185 : /*
186 : * We should now have size estimates for every actual table involved in
187 : * the query, and we also know which if any have been deleted from the
188 : * query by join removal, pruned by partition pruning, or eliminated by
189 : * constraint exclusion. So we can now compute total_table_pages.
190 : *
191 : * Note that appendrels are not double-counted here, even though we don't
192 : * bother to distinguish RelOptInfos for appendrel parents, because the
193 : * parents will have pages = 0.
194 : *
195 : * XXX if a table is self-joined, we will count it once per appearance,
196 : * which perhaps is the wrong thing ... but that's not completely clear,
197 : * and detecting self-joins here is difficult, so ignore it for now.
198 : */
199 291830 : total_pages = 0;
200 880126 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
201 : {
202 588296 : RelOptInfo *brel = root->simple_rel_array[rti];
203 :
204 : /* there may be empty slots corresponding to non-baserel RTEs */
205 588296 : if (brel == NULL)
206 135494 : continue;
207 :
208 : Assert(brel->relid == rti); /* sanity check on array */
209 :
210 452802 : if (IS_DUMMY_REL(brel))
211 1126 : continue;
212 :
213 451676 : if (IS_SIMPLE_REL(brel))
214 451676 : total_pages += (double) brel->pages;
215 : }
216 291830 : root->total_table_pages = total_pages;
217 :
218 : /*
219 : * Generate access paths for each base rel.
220 : */
221 291830 : set_base_rel_pathlists(root);
222 :
223 : /*
224 : * Generate access paths for the entire join tree.
225 : */
226 291830 : rel = make_rel_from_joinlist(root, joinlist);
227 :
228 : /*
229 : * The result should join all and only the query's base + outer-join rels.
230 : */
231 : Assert(bms_equal(rel->relids, root->all_query_rels));
232 :
233 291830 : return rel;
234 : }
235 :
236 : /*
237 : * set_base_rel_consider_startup
238 : * Set the consider_[param_]startup flags for each base-relation entry.
239 : *
240 : * For the moment, we only deal with consider_param_startup here; because the
241 : * logic for consider_startup is pretty trivial and is the same for every base
242 : * relation, we just let build_simple_rel() initialize that flag correctly to
243 : * start with. If that logic ever gets more complicated it would probably
244 : * be better to move it here.
245 : */
246 : static void
247 291858 : set_base_rel_consider_startup(PlannerInfo *root)
248 : {
249 : /*
250 : * Since parameterized paths can only be used on the inside of a nestloop
251 : * join plan, there is usually little value in considering fast-start
252 : * plans for them. However, for relations that are on the RHS of a SEMI
253 : * or ANTI join, a fast-start plan can be useful because we're only going
254 : * to care about fetching one tuple anyway.
255 : *
256 : * To minimize growth of planning time, we currently restrict this to
257 : * cases where the RHS is a single base relation, not a join; there is no
258 : * provision for consider_param_startup to get set at all on joinrels.
259 : * Also we don't worry about appendrels. costsize.c's costing rules for
260 : * nestloop semi/antijoins don't consider such cases either.
261 : */
262 : ListCell *lc;
263 :
264 327482 : foreach(lc, root->join_info_list)
265 : {
266 35624 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
267 : int varno;
268 :
269 40806 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
270 5182 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
271 : {
272 5012 : RelOptInfo *rel = find_base_rel(root, varno);
273 :
274 5012 : rel->consider_param_startup = true;
275 : }
276 : }
277 291858 : }
278 :
279 : /*
280 : * set_base_rel_sizes
281 : * Set the size estimates (rows and widths) for each base-relation entry.
282 : * Also determine whether to consider parallel paths for base relations.
283 : *
284 : * We do this in a separate pass over the base rels so that rowcount
285 : * estimates are available for parameterized path generation, and also so
286 : * that each rel's consider_parallel flag is set correctly before we begin to
287 : * generate paths.
288 : */
289 : static void
290 291858 : set_base_rel_sizes(PlannerInfo *root)
291 : {
292 : Index rti;
293 :
294 880156 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
295 : {
296 588326 : RelOptInfo *rel = root->simple_rel_array[rti];
297 : RangeTblEntry *rte;
298 :
299 : /* there may be empty slots corresponding to non-baserel RTEs */
300 588326 : if (rel == NULL)
301 135496 : continue;
302 :
303 : Assert(rel->relid == rti); /* sanity check on array */
304 :
305 : /* ignore RTEs that are "other rels" */
306 452830 : if (rel->reloptkind != RELOPT_BASEREL)
307 43222 : continue;
308 :
309 409608 : rte = root->simple_rte_array[rti];
310 :
311 : /*
312 : * If parallelism is allowable for this query in general, see whether
313 : * it's allowable for this rel in particular. We have to do this
314 : * before set_rel_size(), because (a) if this rel is an inheritance
315 : * parent, set_append_rel_size() will use and perhaps change the rel's
316 : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
317 : * goes ahead and makes paths immediately.
318 : */
319 409608 : if (root->glob->parallelModeOK)
320 323504 : set_rel_consider_parallel(root, rel, rte);
321 :
322 409608 : set_rel_size(root, rel, rti, rte);
323 : }
324 291830 : }
325 :
326 : /*
327 : * set_base_rel_pathlists
328 : * Finds all paths available for scanning each base-relation entry.
329 : * Sequential scan and any available indices are considered.
330 : * Each useful path is attached to its relation's 'pathlist' field.
331 : */
332 : static void
333 291830 : set_base_rel_pathlists(PlannerInfo *root)
334 : {
335 : Index rti;
336 :
337 880126 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
338 : {
339 588296 : RelOptInfo *rel = root->simple_rel_array[rti];
340 :
341 : /* there may be empty slots corresponding to non-baserel RTEs */
342 588296 : if (rel == NULL)
343 135494 : continue;
344 :
345 : Assert(rel->relid == rti); /* sanity check on array */
346 :
347 : /* ignore RTEs that are "other rels" */
348 452802 : if (rel->reloptkind != RELOPT_BASEREL)
349 43222 : continue;
350 :
351 409580 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
352 : }
353 291830 : }
354 :
355 : /*
356 : * set_rel_size
357 : * Set size estimates for a base relation
358 : */
359 : static void
360 452610 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
361 : Index rti, RangeTblEntry *rte)
362 : {
363 862218 : if (rel->reloptkind == RELOPT_BASEREL &&
364 409608 : relation_excluded_by_constraints(root, rel, rte))
365 : {
366 : /*
367 : * We proved we don't need to scan the rel via constraint exclusion,
368 : * so set up a single dummy path for it. Here we only check this for
369 : * regular baserels; if it's an otherrel, CE was already checked in
370 : * set_append_rel_size().
371 : *
372 : * In this case, we go ahead and set up the relation's path right away
373 : * instead of leaving it for set_rel_pathlist to do. This is because
374 : * we don't have a convention for marking a rel as dummy except by
375 : * assigning a dummy path to it.
376 : */
377 492 : set_dummy_rel_pathlist(rel);
378 : }
379 452118 : else if (rte->inh)
380 : {
381 : /* It's an "append relation", process accordingly */
382 20656 : set_append_rel_size(root, rel, rti, rte);
383 : }
384 : else
385 : {
386 431462 : switch (rel->rtekind)
387 : {
388 364854 : case RTE_RELATION:
389 364854 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
390 : {
391 : /* Foreign table */
392 2368 : set_foreign_size(root, rel, rte);
393 : }
394 362486 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
395 : {
396 : /*
397 : * We could get here if asked to scan a partitioned table
398 : * with ONLY. In that case we shouldn't scan any of the
399 : * partitions, so mark it as a dummy rel.
400 : */
401 40 : set_dummy_rel_pathlist(rel);
402 : }
403 362446 : else if (rte->tablesample != NULL)
404 : {
405 : /* Sampled relation */
406 300 : set_tablesample_rel_size(root, rel, rte);
407 : }
408 : else
409 : {
410 : /* Plain relation */
411 362146 : set_plain_rel_size(root, rel, rte);
412 : }
413 364826 : break;
414 8056 : case RTE_SUBQUERY:
415 :
416 : /*
417 : * Subqueries don't support making a choice between
418 : * parameterized and unparameterized paths, so just go ahead
419 : * and build their paths immediately.
420 : */
421 8056 : set_subquery_pathlist(root, rel, rti, rte);
422 8056 : break;
423 43978 : case RTE_FUNCTION:
424 43978 : set_function_size_estimates(root, rel);
425 43978 : break;
426 626 : case RTE_TABLEFUNC:
427 626 : set_tablefunc_size_estimates(root, rel);
428 626 : break;
429 7896 : case RTE_VALUES:
430 7896 : set_values_size_estimates(root, rel);
431 7896 : break;
432 4008 : case RTE_CTE:
433 :
434 : /*
435 : * CTEs don't support making a choice between parameterized
436 : * and unparameterized paths, so just go ahead and build their
437 : * paths immediately.
438 : */
439 4008 : if (rte->self_reference)
440 812 : set_worktable_pathlist(root, rel, rte);
441 : else
442 3196 : set_cte_pathlist(root, rel, rte);
443 4008 : break;
444 446 : case RTE_NAMEDTUPLESTORE:
445 : /* Might as well just build the path immediately */
446 446 : set_namedtuplestore_pathlist(root, rel, rte);
447 446 : break;
448 1598 : case RTE_RESULT:
449 : /* Might as well just build the path immediately */
450 1598 : set_result_pathlist(root, rel, rte);
451 1598 : break;
452 0 : default:
453 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
454 : break;
455 : }
456 : }
457 :
458 : /*
459 : * We insist that all non-dummy rels have a nonzero rowcount estimate.
460 : */
461 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
462 452580 : }
463 :
464 : /*
465 : * set_rel_pathlist
466 : * Build access paths for a base relation
467 : */
468 : static void
469 452628 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
470 : Index rti, RangeTblEntry *rte)
471 : {
472 452628 : if (IS_DUMMY_REL(rel))
473 : {
474 : /* We already proved the relation empty, so nothing more to do */
475 : }
476 451640 : else if (rte->inh)
477 : {
478 : /* It's an "append relation", process accordingly */
479 20360 : set_append_rel_pathlist(root, rel, rti, rte);
480 : }
481 : else
482 : {
483 431280 : switch (rel->rtekind)
484 : {
485 364786 : case RTE_RELATION:
486 364786 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
487 : {
488 : /* Foreign table */
489 2364 : set_foreign_pathlist(root, rel, rte);
490 : }
491 362422 : else if (rte->tablesample != NULL)
492 : {
493 : /* Sampled relation */
494 300 : set_tablesample_rel_pathlist(root, rel, rte);
495 : }
496 : else
497 : {
498 : /* Plain relation */
499 362122 : set_plain_rel_pathlist(root, rel, rte);
500 : }
501 364786 : break;
502 7942 : case RTE_SUBQUERY:
503 : /* Subquery --- fully handled during set_rel_size */
504 7942 : break;
505 43978 : case RTE_FUNCTION:
506 : /* RangeFunction */
507 43978 : set_function_pathlist(root, rel, rte);
508 43978 : break;
509 626 : case RTE_TABLEFUNC:
510 : /* Table Function */
511 626 : set_tablefunc_pathlist(root, rel, rte);
512 626 : break;
513 7896 : case RTE_VALUES:
514 : /* Values list */
515 7896 : set_values_pathlist(root, rel, rte);
516 7896 : break;
517 4008 : case RTE_CTE:
518 : /* CTE reference --- fully handled during set_rel_size */
519 4008 : break;
520 446 : case RTE_NAMEDTUPLESTORE:
521 : /* tuplestore reference --- fully handled during set_rel_size */
522 446 : break;
523 1598 : case RTE_RESULT:
524 : /* simple Result --- fully handled during set_rel_size */
525 1598 : break;
526 0 : default:
527 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
528 : break;
529 : }
530 : }
531 :
532 : /*
533 : * Allow a plugin to editorialize on the set of Paths for this base
534 : * relation. It could add new paths (such as CustomPaths) by calling
535 : * add_path(), or add_partial_path() if parallel aware. It could also
536 : * delete or modify paths added by the core code.
537 : */
538 452628 : if (set_rel_pathlist_hook)
539 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
540 :
541 : /*
542 : * If this is a baserel, we should normally consider gathering any partial
543 : * paths we may have created for it. We have to do this after calling the
544 : * set_rel_pathlist_hook, else it cannot add partial paths to be included
545 : * here.
546 : *
547 : * However, if this is an inheritance child, skip it. Otherwise, we could
548 : * end up with a very large number of gather nodes, each trying to grab
549 : * its own pool of workers. Instead, we'll consider gathering partial
550 : * paths for the parent appendrel.
551 : *
552 : * Also, if this is the topmost scan/join rel, we postpone gathering until
553 : * the final scan/join targetlist is available (see grouping_planner).
554 : */
555 452628 : if (rel->reloptkind == RELOPT_BASEREL &&
556 409580 : !bms_equal(rel->relids, root->all_query_rels))
557 204084 : generate_useful_gather_paths(root, rel, false);
558 :
559 : /* Now find the cheapest of the paths for this rel */
560 452628 : set_cheapest(rel);
561 :
562 : #ifdef OPTIMIZER_DEBUG
563 : pprint(rel);
564 : #endif
565 452628 : }
566 :
567 : /*
568 : * set_plain_rel_size
569 : * Set size estimates for a plain relation (no subquery, no inheritance)
570 : */
571 : static void
572 362146 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
573 : {
574 : /*
575 : * Test any partial indexes of rel for applicability. We must do this
576 : * first since partial unique indexes can affect size estimates.
577 : */
578 362146 : check_index_predicates(root, rel);
579 :
580 : /* Mark rel with estimated output rows, width, etc */
581 362146 : set_baserel_size_estimates(root, rel);
582 362122 : }
583 :
584 : /*
585 : * If this relation could possibly be scanned from within a worker, then set
586 : * its consider_parallel flag.
587 : */
588 : static void
589 353608 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
590 : RangeTblEntry *rte)
591 : {
592 : /*
593 : * The flag has previously been initialized to false, so we can just
594 : * return if it becomes clear that we can't safely set it.
595 : */
596 : Assert(!rel->consider_parallel);
597 :
598 : /* Don't call this if parallelism is disallowed for the entire query. */
599 : Assert(root->glob->parallelModeOK);
600 :
601 : /* This should only be called for baserels and appendrel children. */
602 : Assert(IS_SIMPLE_REL(rel));
603 :
604 : /* Assorted checks based on rtekind. */
605 353608 : switch (rte->rtekind)
606 : {
607 315090 : case RTE_RELATION:
608 :
609 : /*
610 : * Currently, parallel workers can't access the leader's temporary
611 : * tables. We could possibly relax this if we wrote all of its
612 : * local buffers at the start of the query and made no changes
613 : * thereafter (maybe we could allow hint bit changes), and if we
614 : * taught the workers to read them. Writing a large number of
615 : * temporary buffers could be expensive, though, and we don't have
616 : * the rest of the necessary infrastructure right now anyway. So
617 : * for now, bail out if we see a temporary table.
618 : */
619 315090 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
620 7792 : return;
621 :
622 : /*
623 : * Table sampling can be pushed down to workers if the sample
624 : * function and its arguments are safe.
625 : */
626 307298 : if (rte->tablesample != NULL)
627 : {
628 324 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
629 :
630 324 : if (proparallel != PROPARALLEL_SAFE)
631 36 : return;
632 288 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
633 12 : return;
634 : }
635 :
636 : /*
637 : * Ask FDWs whether they can support performing a ForeignScan
638 : * within a worker. Most often, the answer will be no. For
639 : * example, if the nature of the FDW is such that it opens a TCP
640 : * connection with a remote server, each parallel worker would end
641 : * up with a separate connection, and these connections might not
642 : * be appropriately coordinated between workers and the leader.
643 : */
644 307250 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
645 : {
646 : Assert(rel->fdwroutine);
647 1530 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
648 1458 : return;
649 72 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
650 0 : return;
651 : }
652 :
653 : /*
654 : * There are additional considerations for appendrels, which we'll
655 : * deal with in set_append_rel_size and set_append_rel_pathlist.
656 : * For now, just set consider_parallel based on the rel's own
657 : * quals and targetlist.
658 : */
659 305792 : break;
660 :
661 6490 : case RTE_SUBQUERY:
662 :
663 : /*
664 : * There's no intrinsic problem with scanning a subquery-in-FROM
665 : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
666 : * If the subquery doesn't happen to have any parallel-safe paths,
667 : * then flagging it as consider_parallel won't change anything,
668 : * but that's true for plain tables, too. We must set
669 : * consider_parallel based on the rel's own quals and targetlist,
670 : * so that if a subquery path is parallel-safe but the quals and
671 : * projection we're sticking onto it are not, we correctly mark
672 : * the SubqueryScanPath as not parallel-safe. (Note that
673 : * set_subquery_pathlist() might push some of these quals down
674 : * into the subquery itself, but that doesn't change anything.)
675 : *
676 : * We can't push sub-select containing LIMIT/OFFSET to workers as
677 : * there is no guarantee that the row order will be fully
678 : * deterministic, and applying LIMIT/OFFSET will lead to
679 : * inconsistent results at the top-level. (In some cases, where
680 : * the result is ordered, we could relax this restriction. But it
681 : * doesn't currently seem worth expending extra effort to do so.)
682 : */
683 : {
684 6490 : Query *subquery = castNode(Query, rte->subquery);
685 :
686 6490 : if (limit_needed(subquery))
687 424 : return;
688 : }
689 6066 : break;
690 :
691 0 : case RTE_JOIN:
692 : /* Shouldn't happen; we're only considering baserels here. */
693 : Assert(false);
694 0 : return;
695 :
696 23682 : case RTE_FUNCTION:
697 : /* Check for parallel-restricted functions. */
698 23682 : if (!is_parallel_safe(root, (Node *) rte->functions))
699 11252 : return;
700 12430 : break;
701 :
702 626 : case RTE_TABLEFUNC:
703 : /* not parallel safe */
704 626 : return;
705 :
706 2806 : case RTE_VALUES:
707 : /* Check for parallel-restricted functions. */
708 2806 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
709 6 : return;
710 2800 : break;
711 :
712 3286 : case RTE_CTE:
713 :
714 : /*
715 : * CTE tuplestores aren't shared among parallel workers, so we
716 : * force all CTE scans to happen in the leader. Also, populating
717 : * the CTE would require executing a subplan that's not available
718 : * in the worker, might be parallel-restricted, and must get
719 : * executed only once.
720 : */
721 3286 : return;
722 :
723 418 : case RTE_NAMEDTUPLESTORE:
724 :
725 : /*
726 : * tuplestore cannot be shared, at least without more
727 : * infrastructure to support that.
728 : */
729 418 : return;
730 :
731 1210 : case RTE_RESULT:
732 : /* RESULT RTEs, in themselves, are no problem. */
733 1210 : break;
734 0 : case RTE_GROUP:
735 : /* Shouldn't happen; we're only considering baserels here. */
736 : Assert(false);
737 0 : return;
738 : }
739 :
740 : /*
741 : * If there's anything in baserestrictinfo that's parallel-restricted, we
742 : * give up on parallelizing access to this relation. We could consider
743 : * instead postponing application of the restricted quals until we're
744 : * above all the parallelism in the plan tree, but it's not clear that
745 : * that would be a win in very many cases, and it might be tricky to make
746 : * outer join clauses work correctly. It would likely break equivalence
747 : * classes, too.
748 : */
749 328298 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
750 24318 : return;
751 :
752 : /*
753 : * Likewise, if the relation's outputs are not parallel-safe, give up.
754 : * (Usually, they're just Vars, but sometimes they're not.)
755 : */
756 303980 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
757 18 : return;
758 :
759 : /* We have a winner. */
760 303962 : rel->consider_parallel = true;
761 : }
762 :
763 : /*
764 : * set_plain_rel_pathlist
765 : * Build access paths for a plain relation (no subquery, no inheritance)
766 : */
767 : static void
768 362122 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
769 : {
770 : Relids required_outer;
771 :
772 : /*
773 : * We don't support pushing join clauses into the quals of a seqscan, but
774 : * it could still have required parameterization due to LATERAL refs in
775 : * its tlist.
776 : */
777 362122 : required_outer = rel->lateral_relids;
778 :
779 : /*
780 : * Consider TID scans.
781 : *
782 : * If create_tidscan_paths returns true, then a TID scan path is forced.
783 : * This happens when rel->baserestrictinfo contains CurrentOfExpr, because
784 : * the executor can't handle any other type of path for such queries.
785 : * Hence, we return without adding any other paths.
786 : */
787 362122 : if (create_tidscan_paths(root, rel))
788 404 : return;
789 :
790 : /* Consider sequential scan */
791 361718 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
792 :
793 : /* If appropriate, consider parallel sequential scan */
794 361718 : if (rel->consider_parallel && required_outer == NULL)
795 268664 : create_plain_partial_paths(root, rel);
796 :
797 : /* Consider index scans */
798 361718 : create_index_paths(root, rel);
799 : }
800 :
801 : /*
802 : * create_plain_partial_paths
803 : * Build partial access paths for parallel scan of a plain relation
804 : */
805 : static void
806 268664 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
807 : {
808 : int parallel_workers;
809 :
810 268664 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
811 : max_parallel_workers_per_gather);
812 :
813 : /* If any limit was set to zero, the user doesn't want a parallel scan. */
814 268664 : if (parallel_workers <= 0)
815 242956 : return;
816 :
817 : /* Add an unordered partial path based on a parallel sequential scan. */
818 25708 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
819 : }
820 :
821 : /*
822 : * set_tablesample_rel_size
823 : * Set size estimates for a sampled relation
824 : */
825 : static void
826 300 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
827 : {
828 300 : TableSampleClause *tsc = rte->tablesample;
829 : TsmRoutine *tsm;
830 : BlockNumber pages;
831 : double tuples;
832 :
833 : /*
834 : * Test any partial indexes of rel for applicability. We must do this
835 : * first since partial unique indexes can affect size estimates.
836 : */
837 300 : check_index_predicates(root, rel);
838 :
839 : /*
840 : * Call the sampling method's estimation function to estimate the number
841 : * of pages it will read and the number of tuples it will return. (Note:
842 : * we assume the function returns sane values.)
843 : */
844 300 : tsm = GetTsmRoutine(tsc->tsmhandler);
845 300 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
846 : &pages, &tuples);
847 :
848 : /*
849 : * For the moment, because we will only consider a SampleScan path for the
850 : * rel, it's okay to just overwrite the pages and tuples estimates for the
851 : * whole relation. If we ever consider multiple path types for sampled
852 : * rels, we'll need more complication.
853 : */
854 300 : rel->pages = pages;
855 300 : rel->tuples = tuples;
856 :
857 : /* Mark rel with estimated output rows, width, etc */
858 300 : set_baserel_size_estimates(root, rel);
859 300 : }
860 :
861 : /*
862 : * set_tablesample_rel_pathlist
863 : * Build access paths for a sampled relation
864 : */
865 : static void
866 300 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
867 : {
868 : Relids required_outer;
869 : Path *path;
870 :
871 : /*
872 : * We don't support pushing join clauses into the quals of a samplescan,
873 : * but it could still have required parameterization due to LATERAL refs
874 : * in its tlist or TABLESAMPLE arguments.
875 : */
876 300 : required_outer = rel->lateral_relids;
877 :
878 : /* Consider sampled scan */
879 300 : path = create_samplescan_path(root, rel, required_outer);
880 :
881 : /*
882 : * If the sampling method does not support repeatable scans, we must avoid
883 : * plans that would scan the rel multiple times. Ideally, we'd simply
884 : * avoid putting the rel on the inside of a nestloop join; but adding such
885 : * a consideration to the planner seems like a great deal of complication
886 : * to support an uncommon usage of second-rate sampling methods. Instead,
887 : * if there is a risk that the query might perform an unsafe join, just
888 : * wrap the SampleScan in a Materialize node. We can check for joins by
889 : * counting the membership of all_query_rels (note that this correctly
890 : * counts inheritance trees as single rels). If we're inside a subquery,
891 : * we can't easily check whether a join might occur in the outer query, so
892 : * just assume one is possible.
893 : *
894 : * GetTsmRoutine is relatively expensive compared to the other tests here,
895 : * so check repeatable_across_scans last, even though that's a bit odd.
896 : */
897 574 : if ((root->query_level > 1 ||
898 274 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
899 92 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
900 : {
901 8 : path = (Path *) create_material_path(rel, path);
902 : }
903 :
904 300 : add_path(rel, path);
905 :
906 : /* For the moment, at least, there are no other paths to consider */
907 300 : }
908 :
909 : /*
910 : * set_foreign_size
911 : * Set size estimates for a foreign table RTE
912 : */
913 : static void
914 2368 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
915 : {
916 : /* Mark rel with estimated output rows, width, etc */
917 2368 : set_foreign_size_estimates(root, rel);
918 :
919 : /* Let FDW adjust the size estimates, if it can */
920 2368 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
921 :
922 : /* ... but do not let it set the rows estimate to zero */
923 2364 : rel->rows = clamp_row_est(rel->rows);
924 :
925 : /*
926 : * Also, make sure rel->tuples is not insane relative to rel->rows.
927 : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
928 : * FDW doesn't do anything to replace that.
929 : */
930 2364 : rel->tuples = Max(rel->tuples, rel->rows);
931 2364 : }
932 :
933 : /*
934 : * set_foreign_pathlist
935 : * Build access paths for a foreign table RTE
936 : */
937 : static void
938 2364 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
939 : {
940 : /* Call the FDW's GetForeignPaths function to generate path(s) */
941 2364 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
942 2364 : }
943 :
944 : /*
945 : * set_append_rel_size
946 : * Set size estimates for a simple "append relation"
947 : *
948 : * The passed-in rel and RTE represent the entire append relation. The
949 : * relation's contents are computed by appending together the output of the
950 : * individual member relations. Note that in the non-partitioned inheritance
951 : * case, the first member relation is actually the same table as is mentioned
952 : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
953 : * a good thing because their outputs are not the same size.
954 : */
955 : static void
956 20656 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
957 : Index rti, RangeTblEntry *rte)
958 : {
959 20656 : int parentRTindex = rti;
960 : bool has_live_children;
961 : double parent_rows;
962 : double parent_size;
963 : double *parent_attrsizes;
964 : int nattrs;
965 : ListCell *l;
966 :
967 : /* Guard against stack overflow due to overly deep inheritance tree. */
968 20656 : check_stack_depth();
969 :
970 : Assert(IS_SIMPLE_REL(rel));
971 :
972 : /*
973 : * If this is a partitioned baserel, set the consider_partitionwise_join
974 : * flag; currently, we only consider partitionwise joins with the baserel
975 : * if its targetlist doesn't contain a whole-row Var.
976 : */
977 20656 : if (enable_partitionwise_join &&
978 4262 : rel->reloptkind == RELOPT_BASEREL &&
979 3530 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
980 3530 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
981 3454 : rel->consider_partitionwise_join = true;
982 :
983 : /*
984 : * Initialize to compute size estimates for whole append relation.
985 : *
986 : * We handle width estimates by weighting the widths of different child
987 : * rels proportionally to their number of rows. This is sensible because
988 : * the use of width estimates is mainly to compute the total relation
989 : * "footprint" if we have to sort or hash it. To do this, we sum the
990 : * total equivalent size (in "double" arithmetic) and then divide by the
991 : * total rowcount estimate. This is done separately for the total rel
992 : * width and each attribute.
993 : *
994 : * Note: if you consider changing this logic, beware that child rels could
995 : * have zero rows and/or width, if they were excluded by constraints.
996 : */
997 20656 : has_live_children = false;
998 20656 : parent_rows = 0;
999 20656 : parent_size = 0;
1000 20656 : nattrs = rel->max_attr - rel->min_attr + 1;
1001 20656 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1002 :
1003 106894 : foreach(l, root->append_rel_list)
1004 : {
1005 86240 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1006 : int childRTindex;
1007 : RangeTblEntry *childRTE;
1008 : RelOptInfo *childrel;
1009 : List *childrinfos;
1010 : ListCell *parentvars;
1011 : ListCell *childvars;
1012 : ListCell *lc;
1013 :
1014 : /* append_rel_list contains all append rels; ignore others */
1015 86240 : if (appinfo->parent_relid != parentRTindex)
1016 43376 : continue;
1017 :
1018 43116 : childRTindex = appinfo->child_relid;
1019 43116 : childRTE = root->simple_rte_array[childRTindex];
1020 :
1021 : /*
1022 : * The child rel's RelOptInfo was already created during
1023 : * add_other_rels_to_query.
1024 : */
1025 43116 : childrel = find_base_rel(root, childRTindex);
1026 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1027 :
1028 : /* We may have already proven the child to be dummy. */
1029 43116 : if (IS_DUMMY_REL(childrel))
1030 18 : continue;
1031 :
1032 : /*
1033 : * We have to copy the parent's targetlist and quals to the child,
1034 : * with appropriate substitution of variables. However, the
1035 : * baserestrictinfo quals were already copied/substituted when the
1036 : * child RelOptInfo was built. So we don't need any additional setup
1037 : * before applying constraint exclusion.
1038 : */
1039 43098 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1040 : {
1041 : /*
1042 : * This child need not be scanned, so we can omit it from the
1043 : * appendrel.
1044 : */
1045 96 : set_dummy_rel_pathlist(childrel);
1046 96 : continue;
1047 : }
1048 :
1049 : /*
1050 : * Constraint exclusion failed, so copy the parent's join quals and
1051 : * targetlist to the child, with appropriate variable substitutions.
1052 : *
1053 : * We skip join quals that came from above outer joins that can null
1054 : * this rel, since they would be of no value while generating paths
1055 : * for the child. This saves some effort while processing the child
1056 : * rel, and it also avoids an implementation restriction in
1057 : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1058 : */
1059 43002 : childrinfos = NIL;
1060 55782 : foreach(lc, rel->joininfo)
1061 : {
1062 12780 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1063 :
1064 12780 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1065 10458 : childrinfos = lappend(childrinfos,
1066 10458 : adjust_appendrel_attrs(root,
1067 : (Node *) rinfo,
1068 : 1, &appinfo));
1069 : }
1070 43002 : childrel->joininfo = childrinfos;
1071 :
1072 : /*
1073 : * Now for the child's targetlist.
1074 : *
1075 : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1076 : * expressions, which otherwise would not occur in a rel's targetlist.
1077 : * Code that might be looking at an appendrel child must cope with
1078 : * such. (Normally, a rel's targetlist would only include Vars and
1079 : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1080 : * fields of childrel->reltarget; not clear if that would be useful.
1081 : */
1082 86004 : childrel->reltarget->exprs = (List *)
1083 43002 : adjust_appendrel_attrs(root,
1084 43002 : (Node *) rel->reltarget->exprs,
1085 : 1, &appinfo);
1086 :
1087 : /*
1088 : * We have to make child entries in the EquivalenceClass data
1089 : * structures as well. This is needed either if the parent
1090 : * participates in some eclass joins (because we will want to consider
1091 : * inner-indexscan joins on the individual children) or if the parent
1092 : * has useful pathkeys (because we should try to build MergeAppend
1093 : * paths that produce those sort orderings).
1094 : */
1095 43002 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1096 22330 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1097 43002 : childrel->has_eclass_joins = rel->has_eclass_joins;
1098 :
1099 : /*
1100 : * Note: we could compute appropriate attr_needed data for the child's
1101 : * variables, by transforming the parent's attr_needed through the
1102 : * translated_vars mapping. However, currently there's no need
1103 : * because attr_needed is only examined for base relations not
1104 : * otherrels. So we just leave the child's attr_needed empty.
1105 : */
1106 :
1107 : /*
1108 : * If we consider partitionwise joins with the parent rel, do the same
1109 : * for partitioned child rels.
1110 : *
1111 : * Note: here we abuse the consider_partitionwise_join flag by setting
1112 : * it for child rels that are not themselves partitioned. We do so to
1113 : * tell try_partitionwise_join() that the child rel is sufficiently
1114 : * valid to be used as a per-partition input, even if it later gets
1115 : * proven to be dummy. (It's not usable until we've set up the
1116 : * reltarget and EC entries, which we just did.)
1117 : */
1118 43002 : if (rel->consider_partitionwise_join)
1119 11336 : childrel->consider_partitionwise_join = true;
1120 :
1121 : /*
1122 : * If parallelism is allowable for this query in general, see whether
1123 : * it's allowable for this childrel in particular. But if we've
1124 : * already decided the appendrel is not parallel-safe as a whole,
1125 : * there's no point in considering parallelism for this child. For
1126 : * consistency, do this before calling set_rel_size() for the child.
1127 : */
1128 43002 : if (root->glob->parallelModeOK && rel->consider_parallel)
1129 30104 : set_rel_consider_parallel(root, childrel, childRTE);
1130 :
1131 : /*
1132 : * Compute the child's size.
1133 : */
1134 43002 : set_rel_size(root, childrel, childRTindex, childRTE);
1135 :
1136 : /*
1137 : * It is possible that constraint exclusion detected a contradiction
1138 : * within a child subquery, even though we didn't prove one above. If
1139 : * so, we can skip this child.
1140 : */
1141 43000 : if (IS_DUMMY_REL(childrel))
1142 138 : continue;
1143 :
1144 : /* We have at least one live child. */
1145 42862 : has_live_children = true;
1146 :
1147 : /*
1148 : * If any live child is not parallel-safe, treat the whole appendrel
1149 : * as not parallel-safe. In future we might be able to generate plans
1150 : * in which some children are farmed out to workers while others are
1151 : * not; but we don't have that today, so it's a waste to consider
1152 : * partial paths anywhere in the appendrel unless it's all safe.
1153 : * (Child rels visited before this one will be unmarked in
1154 : * set_append_rel_pathlist().)
1155 : */
1156 42862 : if (!childrel->consider_parallel)
1157 13360 : rel->consider_parallel = false;
1158 :
1159 : /*
1160 : * Accumulate size information from each live child.
1161 : */
1162 : Assert(childrel->rows > 0);
1163 :
1164 42862 : parent_rows += childrel->rows;
1165 42862 : parent_size += childrel->reltarget->width * childrel->rows;
1166 :
1167 : /*
1168 : * Accumulate per-column estimates too. We need not do anything for
1169 : * PlaceHolderVars in the parent list. If child expression isn't a
1170 : * Var, or we didn't record a width estimate for it, we have to fall
1171 : * back on a datatype-based estimate.
1172 : *
1173 : * By construction, child's targetlist is 1-to-1 with parent's.
1174 : */
1175 135040 : forboth(parentvars, rel->reltarget->exprs,
1176 : childvars, childrel->reltarget->exprs)
1177 : {
1178 92178 : Var *parentvar = (Var *) lfirst(parentvars);
1179 92178 : Node *childvar = (Node *) lfirst(childvars);
1180 :
1181 92178 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1182 : {
1183 80134 : int pndx = parentvar->varattno - rel->min_attr;
1184 80134 : int32 child_width = 0;
1185 :
1186 80134 : if (IsA(childvar, Var) &&
1187 77916 : ((Var *) childvar)->varno == childrel->relid)
1188 : {
1189 77850 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1190 :
1191 77850 : child_width = childrel->attr_widths[cndx];
1192 : }
1193 80134 : if (child_width <= 0)
1194 2284 : child_width = get_typavgwidth(exprType(childvar),
1195 : exprTypmod(childvar));
1196 : Assert(child_width > 0);
1197 80134 : parent_attrsizes[pndx] += child_width * childrel->rows;
1198 : }
1199 : }
1200 : }
1201 :
1202 20654 : if (has_live_children)
1203 : {
1204 : /*
1205 : * Save the finished size estimates.
1206 : */
1207 : int i;
1208 :
1209 : Assert(parent_rows > 0);
1210 20360 : rel->rows = parent_rows;
1211 20360 : rel->reltarget->width = rint(parent_size / parent_rows);
1212 207026 : for (i = 0; i < nattrs; i++)
1213 186666 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1214 :
1215 : /*
1216 : * Set "raw tuples" count equal to "rows" for the appendrel; needed
1217 : * because some places assume rel->tuples is valid for any baserel.
1218 : */
1219 20360 : rel->tuples = parent_rows;
1220 :
1221 : /*
1222 : * Note that we leave rel->pages as zero; this is important to avoid
1223 : * double-counting the appendrel tree in total_table_pages.
1224 : */
1225 : }
1226 : else
1227 : {
1228 : /*
1229 : * All children were excluded by constraints, so mark the whole
1230 : * appendrel dummy. We must do this in this phase so that the rel's
1231 : * dummy-ness is visible when we generate paths for other rels.
1232 : */
1233 294 : set_dummy_rel_pathlist(rel);
1234 : }
1235 :
1236 20654 : pfree(parent_attrsizes);
1237 20654 : }
1238 :
1239 : /*
1240 : * set_append_rel_pathlist
1241 : * Build access paths for an "append relation"
1242 : */
1243 : static void
1244 20360 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1245 : Index rti, RangeTblEntry *rte)
1246 : {
1247 20360 : int parentRTindex = rti;
1248 20360 : List *live_childrels = NIL;
1249 : ListCell *l;
1250 :
1251 : /*
1252 : * Generate access paths for each member relation, and remember the
1253 : * non-dummy children.
1254 : */
1255 106184 : foreach(l, root->append_rel_list)
1256 : {
1257 85824 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1258 : int childRTindex;
1259 : RangeTblEntry *childRTE;
1260 : RelOptInfo *childrel;
1261 :
1262 : /* append_rel_list contains all append rels; ignore others */
1263 85824 : if (appinfo->parent_relid != parentRTindex)
1264 42776 : continue;
1265 :
1266 : /* Re-locate the child RTE and RelOptInfo */
1267 43048 : childRTindex = appinfo->child_relid;
1268 43048 : childRTE = root->simple_rte_array[childRTindex];
1269 43048 : childrel = root->simple_rel_array[childRTindex];
1270 :
1271 : /*
1272 : * If set_append_rel_size() decided the parent appendrel was
1273 : * parallel-unsafe at some point after visiting this child rel, we
1274 : * need to propagate the unsafety marking down to the child, so that
1275 : * we don't generate useless partial paths for it.
1276 : */
1277 43048 : if (!rel->consider_parallel)
1278 13466 : childrel->consider_parallel = false;
1279 :
1280 : /*
1281 : * Compute the child's access paths.
1282 : */
1283 43048 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1284 :
1285 : /*
1286 : * If child is dummy, ignore it.
1287 : */
1288 43048 : if (IS_DUMMY_REL(childrel))
1289 186 : continue;
1290 :
1291 : /*
1292 : * Child is live, so add it to the live_childrels list for use below.
1293 : */
1294 42862 : live_childrels = lappend(live_childrels, childrel);
1295 : }
1296 :
1297 : /* Add paths to the append relation. */
1298 20360 : add_paths_to_append_rel(root, rel, live_childrels);
1299 20360 : }
1300 :
1301 :
1302 : /*
1303 : * add_paths_to_append_rel
1304 : * Generate paths for the given append relation given the set of non-dummy
1305 : * child rels.
1306 : *
1307 : * The function collects all parameterizations and orderings supported by the
1308 : * non-dummy children. For every such parameterization or ordering, it creates
1309 : * an append path collecting one path from each non-dummy child with given
1310 : * parameterization or ordering. Similarly it collects partial paths from
1311 : * non-dummy children to create partial append paths.
1312 : */
1313 : void
1314 35158 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1315 : List *live_childrels)
1316 : {
1317 35158 : List *subpaths = NIL;
1318 35158 : bool subpaths_valid = true;
1319 35158 : List *startup_subpaths = NIL;
1320 35158 : bool startup_subpaths_valid = true;
1321 35158 : List *partial_subpaths = NIL;
1322 35158 : List *pa_partial_subpaths = NIL;
1323 35158 : List *pa_nonpartial_subpaths = NIL;
1324 35158 : bool partial_subpaths_valid = true;
1325 : bool pa_subpaths_valid;
1326 35158 : List *all_child_pathkeys = NIL;
1327 35158 : List *all_child_outers = NIL;
1328 : ListCell *l;
1329 35158 : double partial_rows = -1;
1330 :
1331 : /* If appropriate, consider parallel append */
1332 35158 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1333 :
1334 : /*
1335 : * For every non-dummy child, remember the cheapest path. Also, identify
1336 : * all pathkeys (orderings) and parameterizations (required_outer sets)
1337 : * available for the non-dummy member relations.
1338 : */
1339 106894 : foreach(l, live_childrels)
1340 : {
1341 71736 : RelOptInfo *childrel = lfirst(l);
1342 : ListCell *lcp;
1343 71736 : Path *cheapest_partial_path = NULL;
1344 :
1345 : /*
1346 : * If child has an unparameterized cheapest-total path, add that to
1347 : * the unparameterized Append path we are constructing for the parent.
1348 : * If not, there's no workable unparameterized path.
1349 : *
1350 : * With partitionwise aggregates, the child rel's pathlist may be
1351 : * empty, so don't assume that a path exists here.
1352 : */
1353 71736 : if (childrel->pathlist != NIL &&
1354 71736 : childrel->cheapest_total_path->param_info == NULL)
1355 71004 : accumulate_append_subpath(childrel->cheapest_total_path,
1356 : &subpaths, NULL);
1357 : else
1358 732 : subpaths_valid = false;
1359 :
1360 : /*
1361 : * When the planner is considering cheap startup plans, we'll also
1362 : * collect all the cheapest_startup_paths (if set) and build an
1363 : * AppendPath containing those as subpaths.
1364 : */
1365 71736 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1366 : {
1367 : /* cheapest_startup_path must not be a parameterized path. */
1368 : Assert(childrel->cheapest_startup_path->param_info == NULL);
1369 1286 : accumulate_append_subpath(childrel->cheapest_startup_path,
1370 : &startup_subpaths,
1371 : NULL);
1372 : }
1373 : else
1374 70450 : startup_subpaths_valid = false;
1375 :
1376 :
1377 : /* Same idea, but for a partial plan. */
1378 71736 : if (childrel->partial_pathlist != NIL)
1379 : {
1380 46508 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1381 46508 : accumulate_append_subpath(cheapest_partial_path,
1382 : &partial_subpaths, NULL);
1383 : }
1384 : else
1385 25228 : partial_subpaths_valid = false;
1386 :
1387 : /*
1388 : * Same idea, but for a parallel append mixing partial and non-partial
1389 : * paths.
1390 : */
1391 71736 : if (pa_subpaths_valid)
1392 : {
1393 49352 : Path *nppath = NULL;
1394 :
1395 : nppath =
1396 49352 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1397 :
1398 49352 : if (cheapest_partial_path == NULL && nppath == NULL)
1399 : {
1400 : /* Neither a partial nor a parallel-safe path? Forget it. */
1401 450 : pa_subpaths_valid = false;
1402 : }
1403 48902 : else if (nppath == NULL ||
1404 46058 : (cheapest_partial_path != NULL &&
1405 46058 : cheapest_partial_path->total_cost < nppath->total_cost))
1406 : {
1407 : /* Partial path is cheaper or the only option. */
1408 : Assert(cheapest_partial_path != NULL);
1409 45918 : accumulate_append_subpath(cheapest_partial_path,
1410 : &pa_partial_subpaths,
1411 : &pa_nonpartial_subpaths);
1412 : }
1413 : else
1414 : {
1415 : /*
1416 : * Either we've got only a non-partial path, or we think that
1417 : * a single backend can execute the best non-partial path
1418 : * faster than all the parallel backends working together can
1419 : * execute the best partial path.
1420 : *
1421 : * It might make sense to be more aggressive here. Even if
1422 : * the best non-partial path is more expensive than the best
1423 : * partial path, it could still be better to choose the
1424 : * non-partial path if there are several such paths that can
1425 : * be given to different workers. For now, we don't try to
1426 : * figure that out.
1427 : */
1428 2984 : accumulate_append_subpath(nppath,
1429 : &pa_nonpartial_subpaths,
1430 : NULL);
1431 : }
1432 : }
1433 :
1434 : /*
1435 : * Collect lists of all the available path orderings and
1436 : * parameterizations for all the children. We use these as a
1437 : * heuristic to indicate which sort orderings and parameterizations we
1438 : * should build Append and MergeAppend paths for.
1439 : */
1440 164056 : foreach(lcp, childrel->pathlist)
1441 : {
1442 92320 : Path *childpath = (Path *) lfirst(lcp);
1443 92320 : List *childkeys = childpath->pathkeys;
1444 92320 : Relids childouter = PATH_REQ_OUTER(childpath);
1445 :
1446 : /* Unsorted paths don't contribute to pathkey list */
1447 92320 : if (childkeys != NIL)
1448 : {
1449 : ListCell *lpk;
1450 20782 : bool found = false;
1451 :
1452 : /* Have we already seen this ordering? */
1453 20970 : foreach(lpk, all_child_pathkeys)
1454 : {
1455 14962 : List *existing_pathkeys = (List *) lfirst(lpk);
1456 :
1457 14962 : if (compare_pathkeys(existing_pathkeys,
1458 : childkeys) == PATHKEYS_EQUAL)
1459 : {
1460 14774 : found = true;
1461 14774 : break;
1462 : }
1463 : }
1464 20782 : if (!found)
1465 : {
1466 : /* No, so add it to all_child_pathkeys */
1467 6008 : all_child_pathkeys = lappend(all_child_pathkeys,
1468 : childkeys);
1469 : }
1470 : }
1471 :
1472 : /* Unparameterized paths don't contribute to param-set list */
1473 92320 : if (childouter)
1474 : {
1475 : ListCell *lco;
1476 6024 : bool found = false;
1477 :
1478 : /* Have we already seen this param set? */
1479 6636 : foreach(lco, all_child_outers)
1480 : {
1481 4350 : Relids existing_outers = (Relids) lfirst(lco);
1482 :
1483 4350 : if (bms_equal(existing_outers, childouter))
1484 : {
1485 3738 : found = true;
1486 3738 : break;
1487 : }
1488 : }
1489 6024 : if (!found)
1490 : {
1491 : /* No, so add it to all_child_outers */
1492 2286 : all_child_outers = lappend(all_child_outers,
1493 : childouter);
1494 : }
1495 : }
1496 : }
1497 : }
1498 :
1499 : /*
1500 : * If we found unparameterized paths for all children, build an unordered,
1501 : * unparameterized Append path for the rel. (Note: this is correct even
1502 : * if we have zero or one live subpath due to constraint exclusion.)
1503 : */
1504 35158 : if (subpaths_valid)
1505 34846 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1506 : NIL, NULL, 0, false,
1507 : -1));
1508 :
1509 : /* build an AppendPath for the cheap startup paths, if valid */
1510 35158 : if (startup_subpaths_valid)
1511 544 : add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1512 : NIL, NIL, NULL, 0, false, -1));
1513 :
1514 : /*
1515 : * Consider an append of unordered, unparameterized partial paths. Make
1516 : * it parallel-aware if possible.
1517 : */
1518 35158 : if (partial_subpaths_valid && partial_subpaths != NIL)
1519 : {
1520 : AppendPath *appendpath;
1521 : ListCell *lc;
1522 21104 : int parallel_workers = 0;
1523 :
1524 : /* Find the highest number of workers requested for any subpath. */
1525 71732 : foreach(lc, partial_subpaths)
1526 : {
1527 50628 : Path *path = lfirst(lc);
1528 :
1529 50628 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1530 : }
1531 : Assert(parallel_workers > 0);
1532 :
1533 : /*
1534 : * If the use of parallel append is permitted, always request at least
1535 : * log2(# of children) workers. We assume it can be useful to have
1536 : * extra workers in this case because they will be spread out across
1537 : * the children. The precise formula is just a guess, but we don't
1538 : * want to end up with a radically different answer for a table with N
1539 : * partitions vs. an unpartitioned table with the same data, so the
1540 : * use of some kind of log-scaling here seems to make some sense.
1541 : */
1542 21104 : if (enable_parallel_append)
1543 : {
1544 21056 : parallel_workers = Max(parallel_workers,
1545 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1546 21056 : parallel_workers = Min(parallel_workers,
1547 : max_parallel_workers_per_gather);
1548 : }
1549 : Assert(parallel_workers > 0);
1550 :
1551 : /* Generate a partial append path. */
1552 21104 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1553 : NIL, NULL, parallel_workers,
1554 : enable_parallel_append,
1555 : -1);
1556 :
1557 : /*
1558 : * Make sure any subsequent partial paths use the same row count
1559 : * estimate.
1560 : */
1561 21104 : partial_rows = appendpath->path.rows;
1562 :
1563 : /* Add the path. */
1564 21104 : add_partial_path(rel, (Path *) appendpath);
1565 : }
1566 :
1567 : /*
1568 : * Consider a parallel-aware append using a mix of partial and non-partial
1569 : * paths. (This only makes sense if there's at least one child which has
1570 : * a non-partial path that is substantially cheaper than any partial path;
1571 : * otherwise, we should use the append path added in the previous step.)
1572 : */
1573 35158 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1574 : {
1575 : AppendPath *appendpath;
1576 : ListCell *lc;
1577 1560 : int parallel_workers = 0;
1578 :
1579 : /*
1580 : * Find the highest number of workers requested for any partial
1581 : * subpath.
1582 : */
1583 2456 : foreach(lc, pa_partial_subpaths)
1584 : {
1585 896 : Path *path = lfirst(lc);
1586 :
1587 896 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1588 : }
1589 :
1590 : /*
1591 : * Same formula here as above. It's even more important in this
1592 : * instance because the non-partial paths won't contribute anything to
1593 : * the planned number of parallel workers.
1594 : */
1595 1560 : parallel_workers = Max(parallel_workers,
1596 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1597 1560 : parallel_workers = Min(parallel_workers,
1598 : max_parallel_workers_per_gather);
1599 : Assert(parallel_workers > 0);
1600 :
1601 1560 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1602 : pa_partial_subpaths,
1603 : NIL, NULL, parallel_workers, true,
1604 : partial_rows);
1605 1560 : add_partial_path(rel, (Path *) appendpath);
1606 : }
1607 :
1608 : /*
1609 : * Also build unparameterized ordered append paths based on the collected
1610 : * list of child pathkeys.
1611 : */
1612 35158 : if (subpaths_valid)
1613 34846 : generate_orderedappend_paths(root, rel, live_childrels,
1614 : all_child_pathkeys);
1615 :
1616 : /*
1617 : * Build Append paths for each parameterization seen among the child rels.
1618 : * (This may look pretty expensive, but in most cases of practical
1619 : * interest, the child rels will expose mostly the same parameterizations,
1620 : * so that not that many cases actually get considered here.)
1621 : *
1622 : * The Append node itself cannot enforce quals, so all qual checking must
1623 : * be done in the child paths. This means that to have a parameterized
1624 : * Append path, we must have the exact same parameterization for each
1625 : * child path; otherwise some children might be failing to check the
1626 : * moved-down quals. To make them match up, we can try to increase the
1627 : * parameterization of lesser-parameterized paths.
1628 : */
1629 37444 : foreach(l, all_child_outers)
1630 : {
1631 2286 : Relids required_outer = (Relids) lfirst(l);
1632 : ListCell *lcr;
1633 :
1634 : /* Select the child paths for an Append with this parameterization */
1635 2286 : subpaths = NIL;
1636 2286 : subpaths_valid = true;
1637 8394 : foreach(lcr, live_childrels)
1638 : {
1639 6120 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1640 : Path *subpath;
1641 :
1642 6120 : if (childrel->pathlist == NIL)
1643 : {
1644 : /* failed to make a suitable path for this child */
1645 0 : subpaths_valid = false;
1646 0 : break;
1647 : }
1648 :
1649 6120 : subpath = get_cheapest_parameterized_child_path(root,
1650 : childrel,
1651 : required_outer);
1652 6120 : if (subpath == NULL)
1653 : {
1654 : /* failed to make a suitable path for this child */
1655 12 : subpaths_valid = false;
1656 12 : break;
1657 : }
1658 6108 : accumulate_append_subpath(subpath, &subpaths, NULL);
1659 : }
1660 :
1661 2286 : if (subpaths_valid)
1662 2274 : add_path(rel, (Path *)
1663 2274 : create_append_path(root, rel, subpaths, NIL,
1664 : NIL, required_outer, 0, false,
1665 : -1));
1666 : }
1667 :
1668 : /*
1669 : * When there is only a single child relation, the Append path can inherit
1670 : * any ordering available for the child rel's path, so that it's useful to
1671 : * consider ordered partial paths. Above we only considered the cheapest
1672 : * partial path for each child, but let's also make paths using any
1673 : * partial paths that have pathkeys.
1674 : */
1675 35158 : if (list_length(live_childrels) == 1)
1676 : {
1677 14032 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1678 :
1679 : /* skip the cheapest partial path, since we already used that above */
1680 14236 : for_each_from(l, childrel->partial_pathlist, 1)
1681 : {
1682 204 : Path *path = (Path *) lfirst(l);
1683 : AppendPath *appendpath;
1684 :
1685 : /* skip paths with no pathkeys. */
1686 204 : if (path->pathkeys == NIL)
1687 0 : continue;
1688 :
1689 204 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1690 : NIL, NULL,
1691 : path->parallel_workers, true,
1692 : partial_rows);
1693 204 : add_partial_path(rel, (Path *) appendpath);
1694 : }
1695 : }
1696 35158 : }
1697 :
1698 : /*
1699 : * generate_orderedappend_paths
1700 : * Generate ordered append paths for an append relation
1701 : *
1702 : * Usually we generate MergeAppend paths here, but there are some special
1703 : * cases where we can generate simple Append paths, because the subpaths
1704 : * can provide tuples in the required order already.
1705 : *
1706 : * We generate a path for each ordering (pathkey list) appearing in
1707 : * all_child_pathkeys.
1708 : *
1709 : * We consider both cheapest-startup and cheapest-total cases, ie, for each
1710 : * interesting ordering, collect all the cheapest startup subpaths and all the
1711 : * cheapest total paths, and build a suitable path for each case.
1712 : *
1713 : * We don't currently generate any parameterized ordered paths here. While
1714 : * it would not take much more code here to do so, it's very unclear that it
1715 : * is worth the planning cycles to investigate such paths: there's little
1716 : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1717 : * that the current coding of add_path would reject such paths out of hand,
1718 : * because add_path gives no credit for sort ordering of parameterized paths,
1719 : * and a parameterized MergeAppend is going to be more expensive than the
1720 : * corresponding parameterized Append path. If we ever try harder to support
1721 : * parameterized mergejoin plans, it might be worth adding support for
1722 : * parameterized paths here to feed such joins. (See notes in
1723 : * optimizer/README for why that might not ever happen, though.)
1724 : */
1725 : static void
1726 34846 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1727 : List *live_childrels,
1728 : List *all_child_pathkeys)
1729 : {
1730 : ListCell *lcp;
1731 34846 : List *partition_pathkeys = NIL;
1732 34846 : List *partition_pathkeys_desc = NIL;
1733 34846 : bool partition_pathkeys_partial = true;
1734 34846 : bool partition_pathkeys_desc_partial = true;
1735 :
1736 : /*
1737 : * Some partitioned table setups may allow us to use an Append node
1738 : * instead of a MergeAppend. This is possible in cases such as RANGE
1739 : * partitioned tables where it's guaranteed that an earlier partition must
1740 : * contain rows which come earlier in the sort order. To detect whether
1741 : * this is relevant, build pathkey descriptions of the partition ordering,
1742 : * for both forward and reverse scans.
1743 : */
1744 61990 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
1745 27144 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1746 : {
1747 22442 : partition_pathkeys = build_partition_pathkeys(root, rel,
1748 : ForwardScanDirection,
1749 : &partition_pathkeys_partial);
1750 :
1751 22442 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1752 : BackwardScanDirection,
1753 : &partition_pathkeys_desc_partial);
1754 :
1755 : /*
1756 : * You might think we should truncate_useless_pathkeys here, but
1757 : * allowing partition keys which are a subset of the query's pathkeys
1758 : * can often be useful. For example, consider a table partitioned by
1759 : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1760 : * paths that can produce the a, b, c ordering (perhaps via indexes on
1761 : * (a, b, c)) then it works to consider the appendrel output as
1762 : * ordered by a, b, c.
1763 : */
1764 : }
1765 :
1766 : /* Now consider each interesting sort ordering */
1767 40794 : foreach(lcp, all_child_pathkeys)
1768 : {
1769 5948 : List *pathkeys = (List *) lfirst(lcp);
1770 5948 : List *startup_subpaths = NIL;
1771 5948 : List *total_subpaths = NIL;
1772 5948 : List *fractional_subpaths = NIL;
1773 5948 : bool startup_neq_total = false;
1774 : bool match_partition_order;
1775 : bool match_partition_order_desc;
1776 : int end_index;
1777 : int first_index;
1778 : int direction;
1779 :
1780 : /*
1781 : * Determine if this sort ordering matches any partition pathkeys we
1782 : * have, for both ascending and descending partition order. If the
1783 : * partition pathkeys happen to be contained in pathkeys then it still
1784 : * works, as described above, providing that the partition pathkeys
1785 : * are complete and not just a prefix of the partition keys. (In such
1786 : * cases we'll be relying on the child paths to have sorted the
1787 : * lower-order columns of the required pathkeys.)
1788 : */
1789 5948 : match_partition_order =
1790 9754 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1791 3978 : (!partition_pathkeys_partial &&
1792 172 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1793 :
1794 13374 : match_partition_order_desc = !match_partition_order &&
1795 3728 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1796 3762 : (!partition_pathkeys_desc_partial &&
1797 64 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1798 :
1799 : /*
1800 : * When the required pathkeys match the reverse of the partition
1801 : * order, we must build the list of paths in reverse starting with the
1802 : * last matching partition first. We can get away without making any
1803 : * special cases for this in the loop below by just looping backward
1804 : * over the child relations in this case.
1805 : */
1806 5948 : if (match_partition_order_desc)
1807 : {
1808 : /* loop backward */
1809 42 : first_index = list_length(live_childrels) - 1;
1810 42 : end_index = -1;
1811 42 : direction = -1;
1812 :
1813 : /*
1814 : * Set this to true to save us having to check for
1815 : * match_partition_order_desc in the loop below.
1816 : */
1817 42 : match_partition_order = true;
1818 : }
1819 : else
1820 : {
1821 : /* for all other case, loop forward */
1822 5906 : first_index = 0;
1823 5906 : end_index = list_length(live_childrels);
1824 5906 : direction = 1;
1825 : }
1826 :
1827 : /* Select the child paths for this ordering... */
1828 21840 : for (int i = first_index; i != end_index; i += direction)
1829 : {
1830 15892 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1831 : Path *cheapest_startup,
1832 : *cheapest_total,
1833 15892 : *cheapest_fractional = NULL;
1834 :
1835 : /* Locate the right paths, if they are available. */
1836 : cheapest_startup =
1837 15892 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1838 : pathkeys,
1839 : NULL,
1840 : STARTUP_COST,
1841 : false);
1842 : cheapest_total =
1843 15892 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1844 : pathkeys,
1845 : NULL,
1846 : TOTAL_COST,
1847 : false);
1848 :
1849 : /*
1850 : * If we can't find any paths with the right order just use the
1851 : * cheapest-total path; we'll have to sort it later.
1852 : */
1853 15892 : if (cheapest_startup == NULL || cheapest_total == NULL)
1854 : {
1855 280 : cheapest_startup = cheapest_total =
1856 : childrel->cheapest_total_path;
1857 : /* Assert we do have an unparameterized path for this child */
1858 : Assert(cheapest_total->param_info == NULL);
1859 : }
1860 :
1861 : /*
1862 : * When building a fractional path, determine a cheapest
1863 : * fractional path for each child relation too. Looking at startup
1864 : * and total costs is not enough, because the cheapest fractional
1865 : * path may be dominated by two separate paths (one for startup,
1866 : * one for total).
1867 : *
1868 : * When needed (building fractional path), determine the cheapest
1869 : * fractional path too.
1870 : */
1871 15892 : if (root->tuple_fraction > 0)
1872 : {
1873 668 : double path_fraction = (1.0 / root->tuple_fraction);
1874 :
1875 : cheapest_fractional =
1876 668 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1877 : pathkeys,
1878 : NULL,
1879 : path_fraction);
1880 :
1881 : /*
1882 : * If we found no path with matching pathkeys, use the
1883 : * cheapest total path instead.
1884 : *
1885 : * XXX We might consider partially sorted paths too (with an
1886 : * incremental sort on top). But we'd have to build all the
1887 : * incremental paths, do the costing etc.
1888 : */
1889 668 : if (!cheapest_fractional)
1890 44 : cheapest_fractional = cheapest_total;
1891 : }
1892 :
1893 : /*
1894 : * Notice whether we actually have different paths for the
1895 : * "cheapest" and "total" cases; frequently there will be no point
1896 : * in two create_merge_append_path() calls.
1897 : */
1898 15892 : if (cheapest_startup != cheapest_total)
1899 72 : startup_neq_total = true;
1900 :
1901 : /*
1902 : * Collect the appropriate child paths. The required logic varies
1903 : * for the Append and MergeAppend cases.
1904 : */
1905 15892 : if (match_partition_order)
1906 : {
1907 : /*
1908 : * We're going to make a plain Append path. We don't need
1909 : * most of what accumulate_append_subpath would do, but we do
1910 : * want to cut out child Appends or MergeAppends if they have
1911 : * just a single subpath (and hence aren't doing anything
1912 : * useful).
1913 : */
1914 6074 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1915 6074 : cheapest_total = get_singleton_append_subpath(cheapest_total);
1916 :
1917 6074 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1918 6074 : total_subpaths = lappend(total_subpaths, cheapest_total);
1919 :
1920 6074 : if (cheapest_fractional)
1921 : {
1922 120 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1923 120 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1924 : }
1925 : }
1926 : else
1927 : {
1928 : /*
1929 : * Otherwise, rely on accumulate_append_subpath to collect the
1930 : * child paths for the MergeAppend.
1931 : */
1932 9818 : accumulate_append_subpath(cheapest_startup,
1933 : &startup_subpaths, NULL);
1934 9818 : accumulate_append_subpath(cheapest_total,
1935 : &total_subpaths, NULL);
1936 :
1937 9818 : if (cheapest_fractional)
1938 548 : accumulate_append_subpath(cheapest_fractional,
1939 : &fractional_subpaths, NULL);
1940 : }
1941 : }
1942 :
1943 : /* ... and build the Append or MergeAppend paths */
1944 5948 : if (match_partition_order)
1945 : {
1946 : /* We only need Append */
1947 2262 : add_path(rel, (Path *) create_append_path(root,
1948 : rel,
1949 : startup_subpaths,
1950 : NIL,
1951 : pathkeys,
1952 : NULL,
1953 : 0,
1954 : false,
1955 : -1));
1956 2262 : if (startup_neq_total)
1957 0 : add_path(rel, (Path *) create_append_path(root,
1958 : rel,
1959 : total_subpaths,
1960 : NIL,
1961 : pathkeys,
1962 : NULL,
1963 : 0,
1964 : false,
1965 : -1));
1966 :
1967 2262 : if (fractional_subpaths)
1968 60 : add_path(rel, (Path *) create_append_path(root,
1969 : rel,
1970 : fractional_subpaths,
1971 : NIL,
1972 : pathkeys,
1973 : NULL,
1974 : 0,
1975 : false,
1976 : -1));
1977 : }
1978 : else
1979 : {
1980 : /* We need MergeAppend */
1981 3686 : add_path(rel, (Path *) create_merge_append_path(root,
1982 : rel,
1983 : startup_subpaths,
1984 : pathkeys,
1985 : NULL));
1986 3686 : if (startup_neq_total)
1987 48 : add_path(rel, (Path *) create_merge_append_path(root,
1988 : rel,
1989 : total_subpaths,
1990 : pathkeys,
1991 : NULL));
1992 :
1993 3686 : if (fractional_subpaths)
1994 196 : add_path(rel, (Path *) create_merge_append_path(root,
1995 : rel,
1996 : fractional_subpaths,
1997 : pathkeys,
1998 : NULL));
1999 : }
2000 : }
2001 34846 : }
2002 :
2003 : /*
2004 : * get_cheapest_parameterized_child_path
2005 : * Get cheapest path for this relation that has exactly the requested
2006 : * parameterization.
2007 : *
2008 : * Returns NULL if unable to create such a path.
2009 : */
2010 : static Path *
2011 6120 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2012 : Relids required_outer)
2013 : {
2014 : Path *cheapest;
2015 : ListCell *lc;
2016 :
2017 : /*
2018 : * Look up the cheapest existing path with no more than the needed
2019 : * parameterization. If it has exactly the needed parameterization, we're
2020 : * done.
2021 : */
2022 6120 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2023 : NIL,
2024 : required_outer,
2025 : TOTAL_COST,
2026 : false);
2027 : Assert(cheapest != NULL);
2028 6120 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2029 5792 : return cheapest;
2030 :
2031 : /*
2032 : * Otherwise, we can "reparameterize" an existing path to match the given
2033 : * parameterization, which effectively means pushing down additional
2034 : * joinquals to be checked within the path's scan. However, some existing
2035 : * paths might check the available joinquals already while others don't;
2036 : * therefore, it's not clear which existing path will be cheapest after
2037 : * reparameterization. We have to go through them all and find out.
2038 : */
2039 328 : cheapest = NULL;
2040 1156 : foreach(lc, rel->pathlist)
2041 : {
2042 828 : Path *path = (Path *) lfirst(lc);
2043 :
2044 : /* Can't use it if it needs more than requested parameterization */
2045 828 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2046 24 : continue;
2047 :
2048 : /*
2049 : * Reparameterization can only increase the path's cost, so if it's
2050 : * already more expensive than the current cheapest, forget it.
2051 : */
2052 1260 : if (cheapest != NULL &&
2053 456 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2054 384 : continue;
2055 :
2056 : /* Reparameterize if needed, then recheck cost */
2057 420 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2058 : {
2059 344 : path = reparameterize_path(root, path, required_outer, 1.0);
2060 344 : if (path == NULL)
2061 32 : continue; /* failed to reparameterize this one */
2062 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2063 :
2064 312 : if (cheapest != NULL &&
2065 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2066 0 : continue;
2067 : }
2068 :
2069 : /* We have a new best path */
2070 388 : cheapest = path;
2071 : }
2072 :
2073 : /* Return the best path, or NULL if we found no suitable candidate */
2074 328 : return cheapest;
2075 : }
2076 :
2077 : /*
2078 : * accumulate_append_subpath
2079 : * Add a subpath to the list being built for an Append or MergeAppend.
2080 : *
2081 : * It's possible that the child is itself an Append or MergeAppend path, in
2082 : * which case we can "cut out the middleman" and just add its child paths to
2083 : * our own list. (We don't try to do this earlier because we need to apply
2084 : * both levels of transformation to the quals.)
2085 : *
2086 : * Note that if we omit a child MergeAppend in this way, we are effectively
2087 : * omitting a sort step, which seems fine: if the parent is to be an Append,
2088 : * its result would be unsorted anyway, while if the parent is to be a
2089 : * MergeAppend, there's no point in a separate sort on a child.
2090 : *
2091 : * Normally, either path is a partial path and subpaths is a list of partial
2092 : * paths, or else path is a non-partial plan and subpaths is a list of those.
2093 : * However, if path is a parallel-aware Append, then we add its partial path
2094 : * children to subpaths and the rest to special_subpaths. If the latter is
2095 : * NULL, we don't flatten the path at all (unless it contains only partial
2096 : * paths).
2097 : */
2098 : static void
2099 193992 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2100 : {
2101 193992 : if (IsA(path, AppendPath))
2102 : {
2103 14120 : AppendPath *apath = (AppendPath *) path;
2104 :
2105 14120 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2106 : {
2107 13928 : *subpaths = list_concat(*subpaths, apath->subpaths);
2108 13928 : return;
2109 : }
2110 192 : else if (special_subpaths != NULL)
2111 : {
2112 : List *new_special_subpaths;
2113 :
2114 : /* Split Parallel Append into partial and non-partial subpaths */
2115 96 : *subpaths = list_concat(*subpaths,
2116 96 : list_copy_tail(apath->subpaths,
2117 : apath->first_partial_path));
2118 96 : new_special_subpaths = list_copy_head(apath->subpaths,
2119 : apath->first_partial_path);
2120 96 : *special_subpaths = list_concat(*special_subpaths,
2121 : new_special_subpaths);
2122 96 : return;
2123 : }
2124 : }
2125 179872 : else if (IsA(path, MergeAppendPath))
2126 : {
2127 692 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2128 :
2129 692 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2130 692 : return;
2131 : }
2132 :
2133 179276 : *subpaths = lappend(*subpaths, path);
2134 : }
2135 :
2136 : /*
2137 : * get_singleton_append_subpath
2138 : * Returns the single subpath of an Append/MergeAppend, or just
2139 : * return 'path' if it's not a single sub-path Append/MergeAppend.
2140 : *
2141 : * Note: 'path' must not be a parallel-aware path.
2142 : */
2143 : static Path *
2144 12268 : get_singleton_append_subpath(Path *path)
2145 : {
2146 : Assert(!path->parallel_aware);
2147 :
2148 12268 : if (IsA(path, AppendPath))
2149 : {
2150 340 : AppendPath *apath = (AppendPath *) path;
2151 :
2152 340 : if (list_length(apath->subpaths) == 1)
2153 156 : return (Path *) linitial(apath->subpaths);
2154 : }
2155 11928 : else if (IsA(path, MergeAppendPath))
2156 : {
2157 300 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2158 :
2159 300 : if (list_length(mpath->subpaths) == 1)
2160 0 : return (Path *) linitial(mpath->subpaths);
2161 : }
2162 :
2163 12112 : return path;
2164 : }
2165 :
2166 : /*
2167 : * set_dummy_rel_pathlist
2168 : * Build a dummy path for a relation that's been excluded by constraints
2169 : *
2170 : * Rather than inventing a special "dummy" path type, we represent this as an
2171 : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2172 : *
2173 : * (See also mark_dummy_rel, which does basically the same thing, but is
2174 : * typically used to change a rel into dummy state after we already made
2175 : * paths for it.)
2176 : */
2177 : static void
2178 1036 : set_dummy_rel_pathlist(RelOptInfo *rel)
2179 : {
2180 : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2181 1036 : rel->rows = 0;
2182 1036 : rel->reltarget->width = 0;
2183 :
2184 : /* Discard any pre-existing paths; no further need for them */
2185 1036 : rel->pathlist = NIL;
2186 1036 : rel->partial_pathlist = NIL;
2187 :
2188 : /* Set up the dummy path */
2189 1036 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2190 : NIL, rel->lateral_relids,
2191 : 0, false, -1));
2192 :
2193 : /*
2194 : * We set the cheapest-path fields immediately, just in case they were
2195 : * pointing at some discarded path. This is redundant in current usage
2196 : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2197 : * for safety and consistency with mark_dummy_rel.
2198 : */
2199 1036 : set_cheapest(rel);
2200 1036 : }
2201 :
2202 : /*
2203 : * find_window_run_conditions
2204 : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2205 : * function to determine the function's monotonic properties. We then
2206 : * see if 'opexpr' can be used to short-circuit execution.
2207 : *
2208 : * For example row_number() over (order by ...) always produces a value one
2209 : * higher than the previous. If someone has a window function in a subquery
2210 : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2211 : * as well stop processing the windowagg once the row number reaches 11. Here
2212 : * we check if 'opexpr' might help us to stop doing needless extra processing
2213 : * in WindowAgg nodes.
2214 : *
2215 : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2216 : * its original purpose. This is set to false if the caller can assume that
2217 : * the run condition will handle all of the required filtering.
2218 : *
2219 : * Returns true if 'opexpr' was found to be useful and was added to the
2220 : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2221 : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2222 : * If the 'opexpr' cannot be used then we set *keep_original to true and
2223 : * return false.
2224 : */
2225 : static bool
2226 234 : find_window_run_conditions(Query *subquery, RangeTblEntry *rte, Index rti,
2227 : AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2228 : bool wfunc_left, bool *keep_original,
2229 : Bitmapset **run_cond_attrs)
2230 : {
2231 : Oid prosupport;
2232 : Expr *otherexpr;
2233 : SupportRequestWFuncMonotonic req;
2234 : SupportRequestWFuncMonotonic *res;
2235 : WindowClause *wclause;
2236 : List *opinfos;
2237 : OpExpr *runopexpr;
2238 : Oid runoperator;
2239 : ListCell *lc;
2240 :
2241 234 : *keep_original = true;
2242 :
2243 234 : while (IsA(wfunc, RelabelType))
2244 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2245 :
2246 : /* we can only work with window functions */
2247 234 : if (!IsA(wfunc, WindowFunc))
2248 24 : return false;
2249 :
2250 : /* can't use it if there are subplans in the WindowFunc */
2251 210 : if (contain_subplans((Node *) wfunc))
2252 6 : return false;
2253 :
2254 204 : prosupport = get_func_support(wfunc->winfnoid);
2255 :
2256 : /* Check if there's a support function for 'wfunc' */
2257 204 : if (!OidIsValid(prosupport))
2258 18 : return false;
2259 :
2260 : /* get the Expr from the other side of the OpExpr */
2261 186 : if (wfunc_left)
2262 162 : otherexpr = lsecond(opexpr->args);
2263 : else
2264 24 : otherexpr = linitial(opexpr->args);
2265 :
2266 : /*
2267 : * The value being compared must not change during the evaluation of the
2268 : * window partition.
2269 : */
2270 186 : if (!is_pseudo_constant_clause((Node *) otherexpr))
2271 0 : return false;
2272 :
2273 : /* find the window clause belonging to the window function */
2274 186 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2275 186 : wfunc->winref - 1);
2276 :
2277 186 : req.type = T_SupportRequestWFuncMonotonic;
2278 186 : req.window_func = wfunc;
2279 186 : req.window_clause = wclause;
2280 :
2281 : /* call the support function */
2282 : res = (SupportRequestWFuncMonotonic *)
2283 186 : DatumGetPointer(OidFunctionCall1(prosupport,
2284 : PointerGetDatum(&req)));
2285 :
2286 : /*
2287 : * Nothing to do if the function is neither monotonically increasing nor
2288 : * monotonically decreasing.
2289 : */
2290 186 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2291 0 : return false;
2292 :
2293 186 : runopexpr = NULL;
2294 186 : runoperator = InvalidOid;
2295 186 : opinfos = get_op_btree_interpretation(opexpr->opno);
2296 :
2297 186 : foreach(lc, opinfos)
2298 : {
2299 186 : OpBtreeInterpretation *opinfo = (OpBtreeInterpretation *) lfirst(lc);
2300 186 : int strategy = opinfo->strategy;
2301 :
2302 : /* handle < / <= */
2303 186 : if (strategy == BTLessStrategyNumber ||
2304 : strategy == BTLessEqualStrategyNumber)
2305 : {
2306 : /*
2307 : * < / <= is supported for monotonically increasing functions in
2308 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2309 : * for monotonically decreasing functions.
2310 : */
2311 132 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2312 18 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2313 : {
2314 120 : *keep_original = false;
2315 120 : runopexpr = opexpr;
2316 120 : runoperator = opexpr->opno;
2317 : }
2318 132 : break;
2319 : }
2320 : /* handle > / >= */
2321 54 : else if (strategy == BTGreaterStrategyNumber ||
2322 : strategy == BTGreaterEqualStrategyNumber)
2323 : {
2324 : /*
2325 : * > / >= is supported for monotonically decreasing functions in
2326 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2327 : * for monotonically increasing functions.
2328 : */
2329 18 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2330 12 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2331 : {
2332 18 : *keep_original = false;
2333 18 : runopexpr = opexpr;
2334 18 : runoperator = opexpr->opno;
2335 : }
2336 18 : break;
2337 : }
2338 : /* handle = */
2339 36 : else if (strategy == BTEqualStrategyNumber)
2340 : {
2341 : int16 newstrategy;
2342 :
2343 : /*
2344 : * When both monotonically increasing and decreasing then the
2345 : * return value of the window function will be the same each time.
2346 : * We can simply use 'opexpr' as the run condition without
2347 : * modifying it.
2348 : */
2349 36 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2350 : {
2351 6 : *keep_original = false;
2352 6 : runopexpr = opexpr;
2353 6 : runoperator = opexpr->opno;
2354 6 : break;
2355 : }
2356 :
2357 : /*
2358 : * When monotonically increasing we make a qual with <wfunc> <=
2359 : * <value> or <value> >= <wfunc> in order to filter out values
2360 : * which are above the value in the equality condition. For
2361 : * monotonically decreasing functions we want to filter values
2362 : * below the value in the equality condition.
2363 : */
2364 30 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2365 30 : newstrategy = wfunc_left ? BTLessEqualStrategyNumber : BTGreaterEqualStrategyNumber;
2366 : else
2367 0 : newstrategy = wfunc_left ? BTGreaterEqualStrategyNumber : BTLessEqualStrategyNumber;
2368 :
2369 : /* We must keep the original equality qual */
2370 30 : *keep_original = true;
2371 30 : runopexpr = opexpr;
2372 :
2373 : /* determine the operator to use for the WindowFuncRunCondition */
2374 30 : runoperator = get_opfamily_member(opinfo->opfamily_id,
2375 : opinfo->oplefttype,
2376 : opinfo->oprighttype,
2377 : newstrategy);
2378 30 : break;
2379 : }
2380 : }
2381 :
2382 186 : if (runopexpr != NULL)
2383 : {
2384 : WindowFuncRunCondition *wfuncrc;
2385 :
2386 174 : wfuncrc = makeNode(WindowFuncRunCondition);
2387 174 : wfuncrc->opno = runoperator;
2388 174 : wfuncrc->inputcollid = runopexpr->inputcollid;
2389 174 : wfuncrc->wfunc_left = wfunc_left;
2390 174 : wfuncrc->arg = copyObject(otherexpr);
2391 :
2392 174 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2393 :
2394 : /* record that this attno was used in a run condition */
2395 174 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2396 : attno - FirstLowInvalidHeapAttributeNumber);
2397 174 : return true;
2398 : }
2399 :
2400 : /* unsupported OpExpr */
2401 12 : return false;
2402 : }
2403 :
2404 : /*
2405 : * check_and_push_window_quals
2406 : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2407 : * as a 'runCondition' qual. These, when present, allow some unnecessary
2408 : * work to be skipped during execution.
2409 : *
2410 : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2411 : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2412 : * window quals for.
2413 : *
2414 : * Returns true if the caller still must keep the original qual or false if
2415 : * the caller can safely ignore the original qual because the WindowAgg node
2416 : * will use the runCondition to stop returning tuples.
2417 : */
2418 : static bool
2419 246 : check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti,
2420 : Node *clause, Bitmapset **run_cond_attrs)
2421 : {
2422 246 : OpExpr *opexpr = (OpExpr *) clause;
2423 246 : bool keep_original = true;
2424 : Var *var1;
2425 : Var *var2;
2426 :
2427 : /* We're only able to use OpExprs with 2 operands */
2428 246 : if (!IsA(opexpr, OpExpr))
2429 18 : return true;
2430 :
2431 228 : if (list_length(opexpr->args) != 2)
2432 0 : return true;
2433 :
2434 : /*
2435 : * Currently, we restrict this optimization to strict OpExprs. The reason
2436 : * for this is that during execution, once the runcondition becomes false,
2437 : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2438 : * function result values, we set them to NULL. Having only strict
2439 : * OpExprs here ensures that we properly filter out the tuples with NULLs
2440 : * in the top-level WindowAgg.
2441 : */
2442 228 : set_opfuncid(opexpr);
2443 228 : if (!func_strict(opexpr->opfuncid))
2444 0 : return true;
2445 :
2446 : /*
2447 : * Check for plain Vars that reference window functions in the subquery.
2448 : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2449 : * be used as part of the run condition.
2450 : */
2451 :
2452 : /* Check the left side of the OpExpr */
2453 228 : var1 = linitial(opexpr->args);
2454 228 : if (IsA(var1, Var) && var1->varattno > 0)
2455 : {
2456 192 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2457 192 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2458 :
2459 192 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2460 : opexpr, true, &keep_original,
2461 : run_cond_attrs))
2462 156 : return keep_original;
2463 : }
2464 :
2465 : /* and check the right side */
2466 72 : var2 = lsecond(opexpr->args);
2467 72 : if (IsA(var2, Var) && var2->varattno > 0)
2468 : {
2469 42 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2470 42 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2471 :
2472 42 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2473 : opexpr, false, &keep_original,
2474 : run_cond_attrs))
2475 18 : return keep_original;
2476 : }
2477 :
2478 54 : return true;
2479 : }
2480 :
2481 : /*
2482 : * set_subquery_pathlist
2483 : * Generate SubqueryScan access paths for a subquery RTE
2484 : *
2485 : * We don't currently support generating parameterized paths for subqueries
2486 : * by pushing join clauses down into them; it seems too expensive to re-plan
2487 : * the subquery multiple times to consider different alternatives.
2488 : * (XXX that could stand to be reconsidered, now that we use Paths.)
2489 : * So the paths made here will be parameterized if the subquery contains
2490 : * LATERAL references, otherwise not. As long as that's true, there's no need
2491 : * for a separate set_subquery_size phase: just make the paths right away.
2492 : */
2493 : static void
2494 8056 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2495 : Index rti, RangeTblEntry *rte)
2496 : {
2497 8056 : Query *parse = root->parse;
2498 8056 : Query *subquery = rte->subquery;
2499 : bool trivial_pathtarget;
2500 : Relids required_outer;
2501 : pushdown_safety_info safetyInfo;
2502 : double tuple_fraction;
2503 : RelOptInfo *sub_final_rel;
2504 8056 : Bitmapset *run_cond_attrs = NULL;
2505 : ListCell *lc;
2506 :
2507 : /*
2508 : * Must copy the Query so that planning doesn't mess up the RTE contents
2509 : * (really really need to fix the planner to not scribble on its input,
2510 : * someday ... but see remove_unused_subquery_outputs to start with).
2511 : */
2512 8056 : subquery = copyObject(subquery);
2513 :
2514 : /*
2515 : * If it's a LATERAL subquery, it might contain some Vars of the current
2516 : * query level, requiring it to be treated as parameterized, even though
2517 : * we don't support pushing down join quals into subqueries.
2518 : */
2519 8056 : required_outer = rel->lateral_relids;
2520 :
2521 : /*
2522 : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2523 : * flags as needed while recursing. In particular, we need a workspace
2524 : * for keeping track of the reasons why columns are unsafe to reference.
2525 : * These reasons are stored in the bits inside unsafeFlags[i] when we
2526 : * discover reasons that column i of the subquery is unsafe to be used in
2527 : * a pushed-down qual.
2528 : */
2529 8056 : memset(&safetyInfo, 0, sizeof(safetyInfo));
2530 8056 : safetyInfo.unsafeFlags = (unsigned char *)
2531 8056 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2532 :
2533 : /*
2534 : * If the subquery has the "security_barrier" flag, it means the subquery
2535 : * originated from a view that must enforce row-level security. Then we
2536 : * must not push down quals that contain leaky functions. (Ideally this
2537 : * would be checked inside subquery_is_pushdown_safe, but since we don't
2538 : * currently pass the RTE to that function, we must do it here.)
2539 : */
2540 8056 : safetyInfo.unsafeLeaky = rte->security_barrier;
2541 :
2542 : /*
2543 : * If there are any restriction clauses that have been attached to the
2544 : * subquery relation, consider pushing them down to become WHERE or HAVING
2545 : * quals of the subquery itself. This transformation is useful because it
2546 : * may allow us to generate a better plan for the subquery than evaluating
2547 : * all the subquery output rows and then filtering them.
2548 : *
2549 : * There are several cases where we cannot push down clauses. Restrictions
2550 : * involving the subquery are checked by subquery_is_pushdown_safe().
2551 : * Restrictions on individual clauses are checked by
2552 : * qual_is_pushdown_safe(). Also, we don't want to push down
2553 : * pseudoconstant clauses; better to have the gating node above the
2554 : * subquery.
2555 : *
2556 : * Non-pushed-down clauses will get evaluated as qpquals of the
2557 : * SubqueryScan node.
2558 : *
2559 : * XXX Are there any cases where we want to make a policy decision not to
2560 : * push down a pushable qual, because it'd result in a worse plan?
2561 : */
2562 9616 : if (rel->baserestrictinfo != NIL &&
2563 1560 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2564 : {
2565 : /* OK to consider pushing down individual quals */
2566 1420 : List *upperrestrictlist = NIL;
2567 : ListCell *l;
2568 :
2569 3744 : foreach(l, rel->baserestrictinfo)
2570 : {
2571 2324 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2572 2324 : Node *clause = (Node *) rinfo->clause;
2573 :
2574 2324 : if (rinfo->pseudoconstant)
2575 : {
2576 4 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2577 4 : continue;
2578 : }
2579 :
2580 2320 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2581 : {
2582 1650 : case PUSHDOWN_SAFE:
2583 : /* Push it down */
2584 1650 : subquery_push_qual(subquery, rte, rti, clause);
2585 1650 : break;
2586 :
2587 246 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2588 :
2589 : /*
2590 : * Since we can't push the qual down into the subquery,
2591 : * check if it happens to reference a window function. If
2592 : * so then it might be useful to use for the WindowAgg's
2593 : * runCondition.
2594 : */
2595 492 : if (!subquery->hasWindowFuncs ||
2596 246 : check_and_push_window_quals(subquery, rte, rti, clause,
2597 : &run_cond_attrs))
2598 : {
2599 : /*
2600 : * subquery has no window funcs or the clause is not a
2601 : * suitable window run condition qual or it is, but
2602 : * the original must also be kept in the upper query.
2603 : */
2604 102 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2605 : }
2606 246 : break;
2607 :
2608 424 : case PUSHDOWN_UNSAFE:
2609 424 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2610 424 : break;
2611 : }
2612 2324 : }
2613 1420 : rel->baserestrictinfo = upperrestrictlist;
2614 : /* We don't bother recomputing baserestrict_min_security */
2615 : }
2616 :
2617 8056 : pfree(safetyInfo.unsafeFlags);
2618 :
2619 : /*
2620 : * The upper query might not use all the subquery's output columns; if
2621 : * not, we can simplify. Pass the attributes that were pushed down into
2622 : * WindowAgg run conditions to ensure we don't accidentally think those
2623 : * are unused.
2624 : */
2625 8056 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2626 :
2627 : /*
2628 : * We can safely pass the outer tuple_fraction down to the subquery if the
2629 : * outer level has no joining, aggregation, or sorting to do. Otherwise
2630 : * we'd better tell the subquery to plan for full retrieval. (XXX This
2631 : * could probably be made more intelligent ...)
2632 : */
2633 8056 : if (parse->hasAggs ||
2634 6864 : parse->groupClause ||
2635 6846 : parse->groupingSets ||
2636 6846 : root->hasHavingQual ||
2637 6846 : parse->distinctClause ||
2638 11022 : parse->sortClause ||
2639 4608 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
2640 4560 : tuple_fraction = 0.0; /* default case */
2641 : else
2642 3496 : tuple_fraction = root->tuple_fraction;
2643 :
2644 : /* plan_params should not be in use in current query level */
2645 : Assert(root->plan_params == NIL);
2646 :
2647 : /* Generate a subroot and Paths for the subquery */
2648 8056 : rel->subroot = subquery_planner(root->glob, subquery, root, false,
2649 : tuple_fraction, NULL);
2650 :
2651 : /* Isolate the params needed by this specific subplan */
2652 8056 : rel->subplan_params = root->plan_params;
2653 8056 : root->plan_params = NIL;
2654 :
2655 : /*
2656 : * It's possible that constraint exclusion proved the subquery empty. If
2657 : * so, it's desirable to produce an unadorned dummy path so that we will
2658 : * recognize appropriate optimizations at this query level.
2659 : */
2660 8056 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2661 :
2662 8056 : if (IS_DUMMY_REL(sub_final_rel))
2663 : {
2664 114 : set_dummy_rel_pathlist(rel);
2665 114 : return;
2666 : }
2667 :
2668 : /*
2669 : * Mark rel with estimated output rows, width, etc. Note that we have to
2670 : * do this before generating outer-query paths, else cost_subqueryscan is
2671 : * not happy.
2672 : */
2673 7942 : set_subquery_size_estimates(root, rel);
2674 :
2675 : /*
2676 : * Also detect whether the reltarget is trivial, so that we can pass that
2677 : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2678 : * It's trivial if it fetches all the subplan output columns in order.
2679 : */
2680 7942 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2681 2518 : trivial_pathtarget = false;
2682 : else
2683 : {
2684 5424 : trivial_pathtarget = true;
2685 17966 : foreach(lc, rel->reltarget->exprs)
2686 : {
2687 12852 : Node *node = (Node *) lfirst(lc);
2688 : Var *var;
2689 :
2690 12852 : if (!IsA(node, Var))
2691 : {
2692 0 : trivial_pathtarget = false;
2693 0 : break;
2694 : }
2695 12852 : var = (Var *) node;
2696 12852 : if (var->varno != rti ||
2697 12852 : var->varattno != foreach_current_index(lc) + 1)
2698 : {
2699 310 : trivial_pathtarget = false;
2700 310 : break;
2701 : }
2702 : }
2703 : }
2704 :
2705 : /*
2706 : * For each Path that subquery_planner produced, make a SubqueryScanPath
2707 : * in the outer query.
2708 : */
2709 16556 : foreach(lc, sub_final_rel->pathlist)
2710 : {
2711 8614 : Path *subpath = (Path *) lfirst(lc);
2712 : List *pathkeys;
2713 :
2714 : /* Convert subpath's pathkeys to outer representation */
2715 8614 : pathkeys = convert_subquery_pathkeys(root,
2716 : rel,
2717 : subpath->pathkeys,
2718 : make_tlist_from_pathtarget(subpath->pathtarget));
2719 :
2720 : /* Generate outer path using this subpath */
2721 8614 : add_path(rel, (Path *)
2722 8614 : create_subqueryscan_path(root, rel, subpath,
2723 : trivial_pathtarget,
2724 : pathkeys, required_outer));
2725 : }
2726 :
2727 : /* If outer rel allows parallelism, do same for partial paths. */
2728 7942 : if (rel->consider_parallel && bms_is_empty(required_outer))
2729 : {
2730 : /* If consider_parallel is false, there should be no partial paths. */
2731 : Assert(sub_final_rel->consider_parallel ||
2732 : sub_final_rel->partial_pathlist == NIL);
2733 :
2734 : /* Same for partial paths. */
2735 4138 : foreach(lc, sub_final_rel->partial_pathlist)
2736 : {
2737 42 : Path *subpath = (Path *) lfirst(lc);
2738 : List *pathkeys;
2739 :
2740 : /* Convert subpath's pathkeys to outer representation */
2741 42 : pathkeys = convert_subquery_pathkeys(root,
2742 : rel,
2743 : subpath->pathkeys,
2744 : make_tlist_from_pathtarget(subpath->pathtarget));
2745 :
2746 : /* Generate outer path using this subpath */
2747 42 : add_partial_path(rel, (Path *)
2748 42 : create_subqueryscan_path(root, rel, subpath,
2749 : trivial_pathtarget,
2750 : pathkeys,
2751 : required_outer));
2752 : }
2753 : }
2754 : }
2755 :
2756 : /*
2757 : * set_function_pathlist
2758 : * Build the (single) access path for a function RTE
2759 : */
2760 : static void
2761 43978 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2762 : {
2763 : Relids required_outer;
2764 43978 : List *pathkeys = NIL;
2765 :
2766 : /*
2767 : * We don't support pushing join clauses into the quals of a function
2768 : * scan, but it could still have required parameterization due to LATERAL
2769 : * refs in the function expression.
2770 : */
2771 43978 : required_outer = rel->lateral_relids;
2772 :
2773 : /*
2774 : * The result is considered unordered unless ORDINALITY was used, in which
2775 : * case it is ordered by the ordinal column (the last one). See if we
2776 : * care, by checking for uses of that Var in equivalence classes.
2777 : */
2778 43978 : if (rte->funcordinality)
2779 : {
2780 698 : AttrNumber ordattno = rel->max_attr;
2781 698 : Var *var = NULL;
2782 : ListCell *lc;
2783 :
2784 : /*
2785 : * Is there a Var for it in rel's targetlist? If not, the query did
2786 : * not reference the ordinality column, or at least not in any way
2787 : * that would be interesting for sorting.
2788 : */
2789 1852 : foreach(lc, rel->reltarget->exprs)
2790 : {
2791 1846 : Var *node = (Var *) lfirst(lc);
2792 :
2793 : /* checking varno/varlevelsup is just paranoia */
2794 1846 : if (IsA(node, Var) &&
2795 1846 : node->varattno == ordattno &&
2796 692 : node->varno == rel->relid &&
2797 692 : node->varlevelsup == 0)
2798 : {
2799 692 : var = node;
2800 692 : break;
2801 : }
2802 : }
2803 :
2804 : /*
2805 : * Try to build pathkeys for this Var with int8 sorting. We tell
2806 : * build_expression_pathkey not to build any new equivalence class; if
2807 : * the Var isn't already mentioned in some EC, it means that nothing
2808 : * cares about the ordering.
2809 : */
2810 698 : if (var)
2811 692 : pathkeys = build_expression_pathkey(root,
2812 : (Expr *) var,
2813 : Int8LessOperator,
2814 : rel->relids,
2815 : false);
2816 : }
2817 :
2818 : /* Generate appropriate path */
2819 43978 : add_path(rel, create_functionscan_path(root, rel,
2820 : pathkeys, required_outer));
2821 43978 : }
2822 :
2823 : /*
2824 : * set_values_pathlist
2825 : * Build the (single) access path for a VALUES RTE
2826 : */
2827 : static void
2828 7896 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2829 : {
2830 : Relids required_outer;
2831 :
2832 : /*
2833 : * We don't support pushing join clauses into the quals of a values scan,
2834 : * but it could still have required parameterization due to LATERAL refs
2835 : * in the values expressions.
2836 : */
2837 7896 : required_outer = rel->lateral_relids;
2838 :
2839 : /* Generate appropriate path */
2840 7896 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
2841 7896 : }
2842 :
2843 : /*
2844 : * set_tablefunc_pathlist
2845 : * Build the (single) access path for a table func RTE
2846 : */
2847 : static void
2848 626 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2849 : {
2850 : Relids required_outer;
2851 :
2852 : /*
2853 : * We don't support pushing join clauses into the quals of a tablefunc
2854 : * scan, but it could still have required parameterization due to LATERAL
2855 : * refs in the function expression.
2856 : */
2857 626 : required_outer = rel->lateral_relids;
2858 :
2859 : /* Generate appropriate path */
2860 626 : add_path(rel, create_tablefuncscan_path(root, rel,
2861 : required_outer));
2862 626 : }
2863 :
2864 : /*
2865 : * set_cte_pathlist
2866 : * Build the (single) access path for a non-self-reference CTE RTE
2867 : *
2868 : * There's no need for a separate set_cte_size phase, since we don't
2869 : * support join-qual-parameterized paths for CTEs.
2870 : */
2871 : static void
2872 3196 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2873 : {
2874 : Path *ctepath;
2875 : Plan *cteplan;
2876 : PlannerInfo *cteroot;
2877 : Index levelsup;
2878 : List *pathkeys;
2879 : int ndx;
2880 : ListCell *lc;
2881 : int plan_id;
2882 : Relids required_outer;
2883 :
2884 : /*
2885 : * Find the referenced CTE, and locate the path and plan previously made
2886 : * for it.
2887 : */
2888 3196 : levelsup = rte->ctelevelsup;
2889 3196 : cteroot = root;
2890 5714 : while (levelsup-- > 0)
2891 : {
2892 2518 : cteroot = cteroot->parent_root;
2893 2518 : if (!cteroot) /* shouldn't happen */
2894 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2895 : }
2896 :
2897 : /*
2898 : * Note: cte_plan_ids can be shorter than cteList, if we are still working
2899 : * on planning the CTEs (ie, this is a side-reference from another CTE).
2900 : * So we mustn't use forboth here.
2901 : */
2902 3196 : ndx = 0;
2903 4632 : foreach(lc, cteroot->parse->cteList)
2904 : {
2905 4632 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
2906 :
2907 4632 : if (strcmp(cte->ctename, rte->ctename) == 0)
2908 3196 : break;
2909 1436 : ndx++;
2910 : }
2911 3196 : if (lc == NULL) /* shouldn't happen */
2912 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
2913 3196 : if (ndx >= list_length(cteroot->cte_plan_ids))
2914 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
2915 3196 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
2916 3196 : if (plan_id <= 0)
2917 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
2918 :
2919 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
2920 3196 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
2921 3196 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2922 :
2923 : /* Mark rel with estimated output rows, width, etc */
2924 3196 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
2925 :
2926 : /* Convert the ctepath's pathkeys to outer query's representation */
2927 3196 : pathkeys = convert_subquery_pathkeys(root,
2928 : rel,
2929 : ctepath->pathkeys,
2930 : cteplan->targetlist);
2931 :
2932 : /*
2933 : * We don't support pushing join clauses into the quals of a CTE scan, but
2934 : * it could still have required parameterization due to LATERAL refs in
2935 : * its tlist.
2936 : */
2937 3196 : required_outer = rel->lateral_relids;
2938 :
2939 : /* Generate appropriate path */
2940 3196 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
2941 3196 : }
2942 :
2943 : /*
2944 : * set_namedtuplestore_pathlist
2945 : * Build the (single) access path for a named tuplestore RTE
2946 : *
2947 : * There's no need for a separate set_namedtuplestore_size phase, since we
2948 : * don't support join-qual-parameterized paths for tuplestores.
2949 : */
2950 : static void
2951 446 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
2952 : RangeTblEntry *rte)
2953 : {
2954 : Relids required_outer;
2955 :
2956 : /* Mark rel with estimated output rows, width, etc */
2957 446 : set_namedtuplestore_size_estimates(root, rel);
2958 :
2959 : /*
2960 : * We don't support pushing join clauses into the quals of a tuplestore
2961 : * scan, but it could still have required parameterization due to LATERAL
2962 : * refs in its tlist.
2963 : */
2964 446 : required_outer = rel->lateral_relids;
2965 :
2966 : /* Generate appropriate path */
2967 446 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2968 446 : }
2969 :
2970 : /*
2971 : * set_result_pathlist
2972 : * Build the (single) access path for an RTE_RESULT RTE
2973 : *
2974 : * There's no need for a separate set_result_size phase, since we
2975 : * don't support join-qual-parameterized paths for these RTEs.
2976 : */
2977 : static void
2978 1598 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
2979 : RangeTblEntry *rte)
2980 : {
2981 : Relids required_outer;
2982 :
2983 : /* Mark rel with estimated output rows, width, etc */
2984 1598 : set_result_size_estimates(root, rel);
2985 :
2986 : /*
2987 : * We don't support pushing join clauses into the quals of a Result scan,
2988 : * but it could still have required parameterization due to LATERAL refs
2989 : * in its tlist.
2990 : */
2991 1598 : required_outer = rel->lateral_relids;
2992 :
2993 : /* Generate appropriate path */
2994 1598 : add_path(rel, create_resultscan_path(root, rel, required_outer));
2995 1598 : }
2996 :
2997 : /*
2998 : * set_worktable_pathlist
2999 : * Build the (single) access path for a self-reference CTE RTE
3000 : *
3001 : * There's no need for a separate set_worktable_size phase, since we don't
3002 : * support join-qual-parameterized paths for CTEs.
3003 : */
3004 : static void
3005 812 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3006 : {
3007 : Path *ctepath;
3008 : PlannerInfo *cteroot;
3009 : Index levelsup;
3010 : Relids required_outer;
3011 :
3012 : /*
3013 : * We need to find the non-recursive term's path, which is in the plan
3014 : * level that's processing the recursive UNION, which is one level *below*
3015 : * where the CTE comes from.
3016 : */
3017 812 : levelsup = rte->ctelevelsup;
3018 812 : if (levelsup == 0) /* shouldn't happen */
3019 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3020 812 : levelsup--;
3021 812 : cteroot = root;
3022 1844 : while (levelsup-- > 0)
3023 : {
3024 1032 : cteroot = cteroot->parent_root;
3025 1032 : if (!cteroot) /* shouldn't happen */
3026 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3027 : }
3028 812 : ctepath = cteroot->non_recursive_path;
3029 812 : if (!ctepath) /* shouldn't happen */
3030 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3031 :
3032 : /* Mark rel with estimated output rows, width, etc */
3033 812 : set_cte_size_estimates(root, rel, ctepath->rows);
3034 :
3035 : /*
3036 : * We don't support pushing join clauses into the quals of a worktable
3037 : * scan, but it could still have required parameterization due to LATERAL
3038 : * refs in its tlist. (I'm not sure this is actually possible given the
3039 : * restrictions on recursive references, but it's easy enough to support.)
3040 : */
3041 812 : required_outer = rel->lateral_relids;
3042 :
3043 : /* Generate appropriate path */
3044 812 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
3045 812 : }
3046 :
3047 : /*
3048 : * generate_gather_paths
3049 : * Generate parallel access paths for a relation by pushing a Gather or
3050 : * Gather Merge on top of a partial path.
3051 : *
3052 : * This must not be called until after we're done creating all partial paths
3053 : * for the specified relation. (Otherwise, add_partial_path might delete a
3054 : * path that some GatherPath or GatherMergePath has a reference to.)
3055 : *
3056 : * If we're generating paths for a scan or join relation, override_rows will
3057 : * be false, and we'll just use the relation's size estimate. When we're
3058 : * being called for a partially-grouped or partially-distinct path, though, we
3059 : * need to override the rowcount estimate. (It's not clear that the
3060 : * particular value we're using here is actually best, but the underlying rel
3061 : * has no estimate so we must do something.)
3062 : */
3063 : void
3064 16226 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3065 : {
3066 : Path *cheapest_partial_path;
3067 : Path *simple_gather_path;
3068 : ListCell *lc;
3069 : double rows;
3070 16226 : double *rowsp = NULL;
3071 :
3072 : /* If there are no partial paths, there's nothing to do here. */
3073 16226 : if (rel->partial_pathlist == NIL)
3074 0 : return;
3075 :
3076 : /* Should we override the rel's rowcount estimate? */
3077 16226 : if (override_rows)
3078 1728 : rowsp = &rows;
3079 :
3080 : /*
3081 : * The output of Gather is always unsorted, so there's only one partial
3082 : * path of interest: the cheapest one. That will be the one at the front
3083 : * of partial_pathlist because of the way add_partial_path works.
3084 : */
3085 16226 : cheapest_partial_path = linitial(rel->partial_pathlist);
3086 16226 : rows = compute_gather_rows(cheapest_partial_path);
3087 : simple_gather_path = (Path *)
3088 16226 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3089 : NULL, rowsp);
3090 16226 : add_path(rel, simple_gather_path);
3091 :
3092 : /*
3093 : * For each useful ordering, we can consider an order-preserving Gather
3094 : * Merge.
3095 : */
3096 33726 : foreach(lc, rel->partial_pathlist)
3097 : {
3098 17500 : Path *subpath = (Path *) lfirst(lc);
3099 : GatherMergePath *path;
3100 :
3101 17500 : if (subpath->pathkeys == NIL)
3102 15888 : continue;
3103 :
3104 1612 : rows = compute_gather_rows(subpath);
3105 1612 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3106 : subpath->pathkeys, NULL, rowsp);
3107 1612 : add_path(rel, &path->path);
3108 : }
3109 : }
3110 :
3111 : /*
3112 : * get_useful_pathkeys_for_relation
3113 : * Determine which orderings of a relation might be useful.
3114 : *
3115 : * Getting data in sorted order can be useful either because the requested
3116 : * order matches the final output ordering for the overall query we're
3117 : * planning, or because it enables an efficient merge join. Here, we try
3118 : * to figure out which pathkeys to consider.
3119 : *
3120 : * This allows us to do incremental sort on top of an index scan under a gather
3121 : * merge node, i.e. parallelized.
3122 : *
3123 : * If the require_parallel_safe is true, we also require the expressions to
3124 : * be parallel safe (which allows pushing the sort below Gather Merge).
3125 : *
3126 : * XXX At the moment this can only ever return a list with a single element,
3127 : * because it looks at query_pathkeys only. So we might return the pathkeys
3128 : * directly, but it seems plausible we'll want to consider other orderings
3129 : * in the future. For example, we might want to consider pathkeys useful for
3130 : * merge joins.
3131 : */
3132 : static List *
3133 16226 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3134 : bool require_parallel_safe)
3135 : {
3136 16226 : List *useful_pathkeys_list = NIL;
3137 :
3138 : /*
3139 : * Considering query_pathkeys is always worth it, because it might allow
3140 : * us to avoid a total sort when we have a partially presorted path
3141 : * available or to push the total sort into the parallel portion of the
3142 : * query.
3143 : */
3144 16226 : if (root->query_pathkeys)
3145 : {
3146 : ListCell *lc;
3147 7002 : int npathkeys = 0; /* useful pathkeys */
3148 :
3149 15044 : foreach(lc, root->query_pathkeys)
3150 : {
3151 10082 : PathKey *pathkey = (PathKey *) lfirst(lc);
3152 10082 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3153 :
3154 : /*
3155 : * We can only build a sort for pathkeys that contain a
3156 : * safe-to-compute-early EC member computable from the current
3157 : * relation's reltarget, so ignore the remainder of the list as
3158 : * soon as we find a pathkey without such a member.
3159 : *
3160 : * It's still worthwhile to return any prefix of the pathkeys list
3161 : * that meets this requirement, as we may be able to do an
3162 : * incremental sort.
3163 : *
3164 : * If requested, ensure the sort expression is parallel-safe too.
3165 : */
3166 10082 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3167 : require_parallel_safe))
3168 2040 : break;
3169 :
3170 8042 : npathkeys++;
3171 : }
3172 :
3173 : /*
3174 : * The whole query_pathkeys list matches, so append it directly, to
3175 : * allow comparing pathkeys easily by comparing list pointer. If we
3176 : * have to truncate the pathkeys, we gotta do a copy though.
3177 : */
3178 7002 : if (npathkeys == list_length(root->query_pathkeys))
3179 4962 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3180 4962 : root->query_pathkeys);
3181 2040 : else if (npathkeys > 0)
3182 452 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3183 452 : list_copy_head(root->query_pathkeys,
3184 : npathkeys));
3185 : }
3186 :
3187 16226 : return useful_pathkeys_list;
3188 : }
3189 :
3190 : /*
3191 : * generate_useful_gather_paths
3192 : * Generate parallel access paths for a relation by pushing a Gather or
3193 : * Gather Merge on top of a partial path.
3194 : *
3195 : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3196 : * paths (aiming to preserve the ordering), but also considers ordering that
3197 : * might be useful for nodes above the gather merge node, and tries to add
3198 : * a sort (regular or incremental) to provide that.
3199 : */
3200 : void
3201 533684 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3202 : {
3203 : ListCell *lc;
3204 : double rows;
3205 533684 : double *rowsp = NULL;
3206 533684 : List *useful_pathkeys_list = NIL;
3207 533684 : Path *cheapest_partial_path = NULL;
3208 :
3209 : /* If there are no partial paths, there's nothing to do here. */
3210 533684 : if (rel->partial_pathlist == NIL)
3211 517458 : return;
3212 :
3213 : /* Should we override the rel's rowcount estimate? */
3214 16226 : if (override_rows)
3215 1728 : rowsp = &rows;
3216 :
3217 : /* generate the regular gather (merge) paths */
3218 16226 : generate_gather_paths(root, rel, override_rows);
3219 :
3220 : /* consider incremental sort for interesting orderings */
3221 16226 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3222 :
3223 : /* used for explicit (full) sort paths */
3224 16226 : cheapest_partial_path = linitial(rel->partial_pathlist);
3225 :
3226 : /*
3227 : * Consider sorted paths for each interesting ordering. We generate both
3228 : * incremental and full sort.
3229 : */
3230 21640 : foreach(lc, useful_pathkeys_list)
3231 : {
3232 5414 : List *useful_pathkeys = lfirst(lc);
3233 : ListCell *lc2;
3234 : bool is_sorted;
3235 : int presorted_keys;
3236 :
3237 11988 : foreach(lc2, rel->partial_pathlist)
3238 : {
3239 6574 : Path *subpath = (Path *) lfirst(lc2);
3240 : GatherMergePath *path;
3241 :
3242 6574 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3243 : subpath->pathkeys,
3244 : &presorted_keys);
3245 :
3246 : /*
3247 : * We don't need to consider the case where a subpath is already
3248 : * fully sorted because generate_gather_paths already creates a
3249 : * gather merge path for every subpath that has pathkeys present.
3250 : *
3251 : * But since the subpath is already sorted, we know we don't need
3252 : * to consider adding a sort (full or incremental) on top of it,
3253 : * so we can continue here.
3254 : */
3255 6574 : if (is_sorted)
3256 1184 : continue;
3257 :
3258 : /*
3259 : * Try at least sorting the cheapest path and also try
3260 : * incrementally sorting any path which is partially sorted
3261 : * already (no need to deal with paths which have presorted keys
3262 : * when incremental sort is disabled unless it's the cheapest
3263 : * input path).
3264 : */
3265 5390 : if (subpath != cheapest_partial_path &&
3266 198 : (presorted_keys == 0 || !enable_incremental_sort))
3267 54 : continue;
3268 :
3269 : /*
3270 : * Consider regular sort for any path that's not presorted or if
3271 : * incremental sort is disabled. We've no need to consider both
3272 : * sort and incremental sort on the same path. We assume that
3273 : * incremental sort is always faster when there are presorted
3274 : * keys.
3275 : *
3276 : * This is not redundant with the gather paths created in
3277 : * generate_gather_paths, because that doesn't generate ordered
3278 : * output. Here we add an explicit sort to match the useful
3279 : * ordering.
3280 : */
3281 5336 : if (presorted_keys == 0 || !enable_incremental_sort)
3282 : {
3283 5160 : subpath = (Path *) create_sort_path(root,
3284 : rel,
3285 : subpath,
3286 : useful_pathkeys,
3287 : -1.0);
3288 : }
3289 : else
3290 176 : subpath = (Path *) create_incremental_sort_path(root,
3291 : rel,
3292 : subpath,
3293 : useful_pathkeys,
3294 : presorted_keys,
3295 : -1);
3296 5336 : rows = compute_gather_rows(subpath);
3297 5336 : path = create_gather_merge_path(root, rel,
3298 : subpath,
3299 5336 : rel->reltarget,
3300 : subpath->pathkeys,
3301 : NULL,
3302 : rowsp);
3303 :
3304 5336 : add_path(rel, &path->path);
3305 : }
3306 : }
3307 : }
3308 :
3309 : /*
3310 : * make_rel_from_joinlist
3311 : * Build access paths using a "joinlist" to guide the join path search.
3312 : *
3313 : * See comments for deconstruct_jointree() for definition of the joinlist
3314 : * data structure.
3315 : */
3316 : static RelOptInfo *
3317 295088 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3318 : {
3319 : int levels_needed;
3320 : List *initial_rels;
3321 : ListCell *jl;
3322 :
3323 : /*
3324 : * Count the number of child joinlist nodes. This is the depth of the
3325 : * dynamic-programming algorithm we must employ to consider all ways of
3326 : * joining the child nodes.
3327 : */
3328 295088 : levels_needed = list_length(joinlist);
3329 :
3330 295088 : if (levels_needed <= 0)
3331 0 : return NULL; /* nothing to do? */
3332 :
3333 : /*
3334 : * Construct a list of rels corresponding to the child joinlist nodes.
3335 : * This may contain both base rels and rels constructed according to
3336 : * sub-joinlists.
3337 : */
3338 295088 : initial_rels = NIL;
3339 707926 : foreach(jl, joinlist)
3340 : {
3341 412838 : Node *jlnode = (Node *) lfirst(jl);
3342 : RelOptInfo *thisrel;
3343 :
3344 412838 : if (IsA(jlnode, RangeTblRef))
3345 : {
3346 409580 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3347 :
3348 409580 : thisrel = find_base_rel(root, varno);
3349 : }
3350 3258 : else if (IsA(jlnode, List))
3351 : {
3352 : /* Recurse to handle subproblem */
3353 3258 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3354 : }
3355 : else
3356 : {
3357 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3358 : (int) nodeTag(jlnode));
3359 : thisrel = NULL; /* keep compiler quiet */
3360 : }
3361 :
3362 412838 : initial_rels = lappend(initial_rels, thisrel);
3363 : }
3364 :
3365 295088 : if (levels_needed == 1)
3366 : {
3367 : /*
3368 : * Single joinlist node, so we're done.
3369 : */
3370 208310 : return (RelOptInfo *) linitial(initial_rels);
3371 : }
3372 : else
3373 : {
3374 : /*
3375 : * Consider the different orders in which we could join the rels,
3376 : * using a plugin, GEQO, or the regular join search code.
3377 : *
3378 : * We put the initial_rels list into a PlannerInfo field because
3379 : * has_legal_joinclause() needs to look at it (ugly :-().
3380 : */
3381 86778 : root->initial_rels = initial_rels;
3382 :
3383 86778 : if (join_search_hook)
3384 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
3385 86778 : else if (enable_geqo && levels_needed >= geqo_threshold)
3386 6 : return geqo(root, levels_needed, initial_rels);
3387 : else
3388 86772 : return standard_join_search(root, levels_needed, initial_rels);
3389 : }
3390 : }
3391 :
3392 : /*
3393 : * standard_join_search
3394 : * Find possible joinpaths for a query by successively finding ways
3395 : * to join component relations into join relations.
3396 : *
3397 : * 'levels_needed' is the number of iterations needed, ie, the number of
3398 : * independent jointree items in the query. This is > 1.
3399 : *
3400 : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3401 : * jointree item. These are the components to be joined together.
3402 : * Note that levels_needed == list_length(initial_rels).
3403 : *
3404 : * Returns the final level of join relations, i.e., the relation that is
3405 : * the result of joining all the original relations together.
3406 : * At least one implementation path must be provided for this relation and
3407 : * all required sub-relations.
3408 : *
3409 : * To support loadable plugins that modify planner behavior by changing the
3410 : * join searching algorithm, we provide a hook variable that lets a plugin
3411 : * replace or supplement this function. Any such hook must return the same
3412 : * final join relation as the standard code would, but it might have a
3413 : * different set of implementation paths attached, and only the sub-joinrels
3414 : * needed for these paths need have been instantiated.
3415 : *
3416 : * Note to plugin authors: the functions invoked during standard_join_search()
3417 : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3418 : * than one join-order search, you'll probably need to save and restore the
3419 : * original states of those data structures. See geqo_eval() for an example.
3420 : */
3421 : RelOptInfo *
3422 86772 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3423 : {
3424 : int lev;
3425 : RelOptInfo *rel;
3426 :
3427 : /*
3428 : * This function cannot be invoked recursively within any one planning
3429 : * problem, so join_rel_level[] can't be in use already.
3430 : */
3431 : Assert(root->join_rel_level == NULL);
3432 :
3433 : /*
3434 : * We employ a simple "dynamic programming" algorithm: we first find all
3435 : * ways to build joins of two jointree items, then all ways to build joins
3436 : * of three items (from two-item joins and single items), then four-item
3437 : * joins, and so on until we have considered all ways to join all the
3438 : * items into one rel.
3439 : *
3440 : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3441 : * set root->join_rel_level[1] to represent all the single-jointree-item
3442 : * relations.
3443 : */
3444 86772 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3445 :
3446 86772 : root->join_rel_level[1] = initial_rels;
3447 :
3448 204498 : for (lev = 2; lev <= levels_needed; lev++)
3449 : {
3450 : ListCell *lc;
3451 :
3452 : /*
3453 : * Determine all possible pairs of relations to be joined at this
3454 : * level, and build paths for making each one from every available
3455 : * pair of lower-level relations.
3456 : */
3457 117726 : join_search_one_level(root, lev);
3458 :
3459 : /*
3460 : * Run generate_partitionwise_join_paths() and
3461 : * generate_useful_gather_paths() for each just-processed joinrel. We
3462 : * could not do this earlier because both regular and partial paths
3463 : * can get added to a particular joinrel at multiple times within
3464 : * join_search_one_level.
3465 : *
3466 : * After that, we're done creating paths for the joinrel, so run
3467 : * set_cheapest().
3468 : */
3469 294974 : foreach(lc, root->join_rel_level[lev])
3470 : {
3471 177248 : rel = (RelOptInfo *) lfirst(lc);
3472 :
3473 : /* Create paths for partitionwise joins. */
3474 177248 : generate_partitionwise_join_paths(root, rel);
3475 :
3476 : /*
3477 : * Except for the topmost scan/join rel, consider gathering
3478 : * partial paths. We'll do the same for the topmost scan/join rel
3479 : * once we know the final targetlist (see grouping_planner's and
3480 : * its call to apply_scanjoin_target_to_paths).
3481 : */
3482 177248 : if (!bms_equal(rel->relids, root->all_query_rels))
3483 90920 : generate_useful_gather_paths(root, rel, false);
3484 :
3485 : /* Find and save the cheapest paths for this rel */
3486 177248 : set_cheapest(rel);
3487 :
3488 : #ifdef OPTIMIZER_DEBUG
3489 : pprint(rel);
3490 : #endif
3491 : }
3492 : }
3493 :
3494 : /*
3495 : * We should have a single rel at the final level.
3496 : */
3497 86772 : if (root->join_rel_level[levels_needed] == NIL)
3498 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
3499 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3500 :
3501 86772 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3502 :
3503 86772 : root->join_rel_level = NULL;
3504 :
3505 86772 : return rel;
3506 : }
3507 :
3508 : /*****************************************************************************
3509 : * PUSHING QUALS DOWN INTO SUBQUERIES
3510 : *****************************************************************************/
3511 :
3512 : /*
3513 : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3514 : *
3515 : * subquery is the particular component query being checked. topquery
3516 : * is the top component of a set-operations tree (the same Query if no
3517 : * set-op is involved).
3518 : *
3519 : * Conditions checked here:
3520 : *
3521 : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3522 : * since that could change the set of rows returned.
3523 : *
3524 : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3525 : * quals into it, because that could change the results.
3526 : *
3527 : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3528 : * This is because upper-level quals should semantically be evaluated only
3529 : * once per distinct row, not once per original row, and if the qual is
3530 : * volatile then extra evaluations could change the results. (This issue
3531 : * does not apply to other forms of aggregation such as GROUP BY, because
3532 : * when those are present we push into HAVING not WHERE, so that the quals
3533 : * are still applied after aggregation.)
3534 : *
3535 : * 4. If the subquery contains window functions, we cannot push volatile quals
3536 : * into it. The issue here is a bit different from DISTINCT: a volatile qual
3537 : * might succeed for some rows of a window partition and fail for others,
3538 : * thereby changing the partition contents and thus the window functions'
3539 : * results for rows that remain.
3540 : *
3541 : * 5. If the subquery contains any set-returning functions in its targetlist,
3542 : * we cannot push volatile quals into it. That would push them below the SRFs
3543 : * and thereby change the number of times they are evaluated. Also, a
3544 : * volatile qual could succeed for some SRF output rows and fail for others,
3545 : * a behavior that cannot occur if it's evaluated before SRF expansion.
3546 : *
3547 : * 6. If the subquery has nonempty grouping sets, we cannot push down any
3548 : * quals. The concern here is that a qual referencing a "constant" grouping
3549 : * column could get constant-folded, which would be improper because the value
3550 : * is potentially nullable by grouping-set expansion. This restriction could
3551 : * be removed if we had a parsetree representation that shows that such
3552 : * grouping columns are not really constant. (There are other ideas that
3553 : * could be used to relax this restriction, but that's the approach most
3554 : * likely to get taken in the future. Note that there's not much to be gained
3555 : * so long as subquery_planner can't move HAVING clauses to WHERE within such
3556 : * a subquery.)
3557 : *
3558 : * In addition, we make several checks on the subquery's output columns to see
3559 : * if it is safe to reference them in pushed-down quals. If output column k
3560 : * is found to be unsafe to reference, we set the reason for that inside
3561 : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3562 : * column k might not be referenced by some/all quals. The unsafeFlags[]
3563 : * array will be consulted later by qual_is_pushdown_safe(). It's better to
3564 : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3565 : * because when the subquery involves set operations we have to check the
3566 : * output expressions in each arm of the set op.
3567 : *
3568 : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3569 : * we're effectively assuming that the quals cannot distinguish values that
3570 : * the DISTINCT's equality operator sees as equal, yet there are many
3571 : * counterexamples to that assumption. However use of such a qual with a
3572 : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3573 : * "equal" value will be chosen as the output value by the DISTINCT operation.
3574 : * So we don't worry too much about that. Another objection is that if the
3575 : * qual is expensive to evaluate, running it for each original row might cost
3576 : * more than we save by eliminating rows before the DISTINCT step. But it
3577 : * would be very hard to estimate that at this stage, and in practice pushdown
3578 : * seldom seems to make things worse, so we ignore that problem too.
3579 : *
3580 : * Note: likewise, pushing quals into a subquery with window functions is a
3581 : * bit dubious: the quals might remove some rows of a window partition while
3582 : * leaving others, causing changes in the window functions' results for the
3583 : * surviving rows. We insist that such a qual reference only partitioning
3584 : * columns, but again that only protects us if the qual does not distinguish
3585 : * values that the partitioning equality operator sees as equal. The risks
3586 : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3587 : * occurs and thus there is no theoretical problem with such a qual. But
3588 : * we'll do this anyway because the potential performance benefits are very
3589 : * large, and we've seen no field complaints about the longstanding comparable
3590 : * behavior with DISTINCT.
3591 : */
3592 : static bool
3593 1732 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
3594 : pushdown_safety_info *safetyInfo)
3595 : {
3596 : SetOperationStmt *topop;
3597 :
3598 : /* Check point 1 */
3599 1732 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
3600 128 : return false;
3601 :
3602 : /* Check point 6 */
3603 1604 : if (subquery->groupClause && subquery->groupingSets)
3604 12 : return false;
3605 :
3606 : /* Check points 3, 4, and 5 */
3607 1592 : if (subquery->distinctClause ||
3608 1514 : subquery->hasWindowFuncs ||
3609 1250 : subquery->hasTargetSRFs)
3610 532 : safetyInfo->unsafeVolatile = true;
3611 :
3612 : /*
3613 : * If we're at a leaf query, check for unsafe expressions in its target
3614 : * list, and mark any reasons why they're unsafe in unsafeFlags[].
3615 : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3616 : * so no need to check them.)
3617 : */
3618 1592 : if (subquery->setOperations == NULL)
3619 1506 : check_output_expressions(subquery, safetyInfo);
3620 :
3621 : /* Are we at top level, or looking at a setop component? */
3622 1592 : if (subquery == topquery)
3623 : {
3624 : /* Top level, so check any component queries */
3625 1420 : if (subquery->setOperations != NULL)
3626 86 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3627 : safetyInfo))
3628 0 : return false;
3629 : }
3630 : else
3631 : {
3632 : /* Setop component must not have more components (too weird) */
3633 172 : if (subquery->setOperations != NULL)
3634 0 : return false;
3635 : /* Check whether setop component output types match top level */
3636 172 : topop = castNode(SetOperationStmt, topquery->setOperations);
3637 : Assert(topop);
3638 172 : compare_tlist_datatypes(subquery->targetList,
3639 : topop->colTypes,
3640 : safetyInfo);
3641 : }
3642 1592 : return true;
3643 : }
3644 :
3645 : /*
3646 : * Helper routine to recurse through setOperations tree
3647 : */
3648 : static bool
3649 258 : recurse_pushdown_safe(Node *setOp, Query *topquery,
3650 : pushdown_safety_info *safetyInfo)
3651 : {
3652 258 : if (IsA(setOp, RangeTblRef))
3653 : {
3654 172 : RangeTblRef *rtr = (RangeTblRef *) setOp;
3655 172 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3656 172 : Query *subquery = rte->subquery;
3657 :
3658 : Assert(subquery != NULL);
3659 172 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3660 : }
3661 86 : else if (IsA(setOp, SetOperationStmt))
3662 : {
3663 86 : SetOperationStmt *op = (SetOperationStmt *) setOp;
3664 :
3665 : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
3666 86 : if (op->op == SETOP_EXCEPT)
3667 0 : return false;
3668 : /* Else recurse */
3669 86 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
3670 0 : return false;
3671 86 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
3672 0 : return false;
3673 : }
3674 : else
3675 : {
3676 0 : elog(ERROR, "unrecognized node type: %d",
3677 : (int) nodeTag(setOp));
3678 : }
3679 86 : return true;
3680 : }
3681 :
3682 : /*
3683 : * check_output_expressions - check subquery's output expressions for safety
3684 : *
3685 : * There are several cases in which it's unsafe to push down an upper-level
3686 : * qual if it references a particular output column of a subquery. We check
3687 : * each output column of the subquery and set flags in unsafeFlags[k] when we
3688 : * see that column is unsafe for a pushed-down qual to reference. The
3689 : * conditions checked here are:
3690 : *
3691 : * 1. We must not push down any quals that refer to subselect outputs that
3692 : * return sets, else we'd introduce functions-returning-sets into the
3693 : * subquery's WHERE/HAVING quals.
3694 : *
3695 : * 2. We must not push down any quals that refer to subselect outputs that
3696 : * contain volatile functions, for fear of introducing strange results due
3697 : * to multiple evaluation of a volatile function.
3698 : *
3699 : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3700 : * refer to non-DISTINCT output columns, because that could change the set
3701 : * of rows returned. (This condition is vacuous for DISTINCT, because then
3702 : * there are no non-DISTINCT output columns, so we needn't check. Note that
3703 : * subquery_is_pushdown_safe already reported that we can't use volatile
3704 : * quals if there's DISTINCT or DISTINCT ON.)
3705 : *
3706 : * 4. If the subquery has any window functions, we must not push down quals
3707 : * that reference any output columns that are not listed in all the subquery's
3708 : * window PARTITION BY clauses. We can push down quals that use only
3709 : * partitioning columns because they should succeed or fail identically for
3710 : * every row of any one window partition, and totally excluding some
3711 : * partitions will not change a window function's results for remaining
3712 : * partitions. (Again, this also requires nonvolatile quals, but
3713 : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3714 : * unsafe for this reason can still have WindowClause run conditions pushed
3715 : * down.
3716 : */
3717 : static void
3718 1506 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
3719 : {
3720 : ListCell *lc;
3721 :
3722 13132 : foreach(lc, subquery->targetList)
3723 : {
3724 11626 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
3725 :
3726 11626 : if (tle->resjunk)
3727 144 : continue; /* ignore resjunk columns */
3728 :
3729 : /* Functions returning sets are unsafe (point 1) */
3730 11482 : if (subquery->hasTargetSRFs &&
3731 614 : (safetyInfo->unsafeFlags[tle->resno] &
3732 614 : UNSAFE_HAS_SET_FUNC) == 0 &&
3733 614 : expression_returns_set((Node *) tle->expr))
3734 : {
3735 352 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3736 352 : continue;
3737 : }
3738 :
3739 : /* Volatile functions are unsafe (point 2) */
3740 11130 : if ((safetyInfo->unsafeFlags[tle->resno] &
3741 11118 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3742 11118 : contain_volatile_functions((Node *) tle->expr))
3743 : {
3744 78 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3745 78 : continue;
3746 : }
3747 :
3748 : /* If subquery uses DISTINCT ON, check point 3 */
3749 11052 : if (subquery->hasDistinctOn &&
3750 0 : (safetyInfo->unsafeFlags[tle->resno] &
3751 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3752 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3753 : {
3754 : /* non-DISTINCT column, so mark it unsafe */
3755 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3756 0 : continue;
3757 : }
3758 :
3759 : /* If subquery uses window functions, check point 4 */
3760 11052 : if (subquery->hasWindowFuncs &&
3761 1146 : (safetyInfo->unsafeFlags[tle->resno] &
3762 1146 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3763 1146 : !targetIsInAllPartitionLists(tle, subquery))
3764 : {
3765 : /* not present in all PARTITION BY clauses, so mark it unsafe */
3766 1050 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3767 1050 : continue;
3768 : }
3769 : }
3770 1506 : }
3771 :
3772 : /*
3773 : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3774 : * push quals into each component query, but the quals can only reference
3775 : * subquery columns that suffer no type coercions in the set operation.
3776 : * Otherwise there are possible semantic gotchas. So, we check the
3777 : * component queries to see if any of them have output types different from
3778 : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3779 : * unsafeFlags[k] if column k has different type in any component.
3780 : *
3781 : * We don't have to care about typmods here: the only allowed difference
3782 : * between set-op input and output typmods is input is a specific typmod
3783 : * and output is -1, and that does not require a coercion.
3784 : *
3785 : * tlist is a subquery tlist.
3786 : * colTypes is an OID list of the top-level setop's output column types.
3787 : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
3788 : */
3789 : static void
3790 172 : compare_tlist_datatypes(List *tlist, List *colTypes,
3791 : pushdown_safety_info *safetyInfo)
3792 : {
3793 : ListCell *l;
3794 172 : ListCell *colType = list_head(colTypes);
3795 :
3796 552 : foreach(l, tlist)
3797 : {
3798 380 : TargetEntry *tle = (TargetEntry *) lfirst(l);
3799 :
3800 380 : if (tle->resjunk)
3801 0 : continue; /* ignore resjunk columns */
3802 380 : if (colType == NULL)
3803 0 : elog(ERROR, "wrong number of tlist entries");
3804 380 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
3805 40 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
3806 380 : colType = lnext(colTypes, colType);
3807 : }
3808 172 : if (colType != NULL)
3809 0 : elog(ERROR, "wrong number of tlist entries");
3810 172 : }
3811 :
3812 : /*
3813 : * targetIsInAllPartitionLists
3814 : * True if the TargetEntry is listed in the PARTITION BY clause
3815 : * of every window defined in the query.
3816 : *
3817 : * It would be safe to ignore windows not actually used by any window
3818 : * function, but it's not easy to get that info at this stage; and it's
3819 : * unlikely to be useful to spend any extra cycles getting it, since
3820 : * unreferenced window definitions are probably infrequent in practice.
3821 : */
3822 : static bool
3823 1146 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
3824 : {
3825 : ListCell *lc;
3826 :
3827 1266 : foreach(lc, query->windowClause)
3828 : {
3829 1170 : WindowClause *wc = (WindowClause *) lfirst(lc);
3830 :
3831 1170 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
3832 1050 : return false;
3833 : }
3834 96 : return true;
3835 : }
3836 :
3837 : /*
3838 : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3839 : *
3840 : * rinfo is a restriction clause applying to the given subquery (whose RTE
3841 : * has index rti in the parent query).
3842 : *
3843 : * Conditions checked here:
3844 : *
3845 : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3846 : * unclear that it will work correctly: SubLinks will already have been
3847 : * transformed into SubPlans in the qual, but not in the subquery). Note that
3848 : * SubLinks that transform to initplans are safe, and will be accepted here
3849 : * because what we'll see in the qual is just a Param referencing the initplan
3850 : * output.
3851 : *
3852 : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3853 : * functions.
3854 : *
3855 : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
3856 : * functions that are passed Var nodes, and therefore might reveal values from
3857 : * the subquery as side effects.
3858 : *
3859 : * 4. rinfo's clause must not refer to the whole-row output of the subquery
3860 : * (since there is no easy way to name that within the subquery itself).
3861 : *
3862 : * 5. rinfo's clause must not refer to any subquery output columns that were
3863 : * found to be unsafe to reference by subquery_is_pushdown_safe().
3864 : */
3865 : static pushdown_safe_type
3866 2320 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
3867 : pushdown_safety_info *safetyInfo)
3868 : {
3869 2320 : pushdown_safe_type safe = PUSHDOWN_SAFE;
3870 2320 : Node *qual = (Node *) rinfo->clause;
3871 : List *vars;
3872 : ListCell *vl;
3873 :
3874 : /* Refuse subselects (point 1) */
3875 2320 : if (contain_subplans(qual))
3876 66 : return PUSHDOWN_UNSAFE;
3877 :
3878 : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3879 2886 : if (safetyInfo->unsafeVolatile &&
3880 632 : contain_volatile_functions((Node *) rinfo))
3881 18 : return PUSHDOWN_UNSAFE;
3882 :
3883 : /* Refuse leaky quals if told to (point 3) */
3884 3166 : if (safetyInfo->unsafeLeaky &&
3885 930 : contain_leaked_vars(qual))
3886 138 : return PUSHDOWN_UNSAFE;
3887 :
3888 : /*
3889 : * Examine all Vars used in clause. Since it's a restriction clause, all
3890 : * such Vars must refer to subselect output columns ... unless this is
3891 : * part of a LATERAL subquery, in which case there could be lateral
3892 : * references.
3893 : *
3894 : * By omitting the relevant flags, this also gives us a cheap sanity check
3895 : * that no aggregates or window functions appear in the qual. Those would
3896 : * be unsafe to push down, but at least for the moment we could never see
3897 : * any in a qual anyhow.
3898 : */
3899 2098 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
3900 4102 : foreach(vl, vars)
3901 : {
3902 2206 : Var *var = (Var *) lfirst(vl);
3903 :
3904 : /*
3905 : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3906 : * It's not clear whether a PHV could safely be pushed down, and even
3907 : * less clear whether such a situation could arise in any cases of
3908 : * practical interest anyway. So for the moment, just refuse to push
3909 : * down.
3910 : */
3911 2206 : if (!IsA(var, Var))
3912 : {
3913 0 : safe = PUSHDOWN_UNSAFE;
3914 0 : break;
3915 : }
3916 :
3917 : /*
3918 : * Punt if we find any lateral references. It would be safe to push
3919 : * these down, but we'd have to convert them into outer references,
3920 : * which subquery_push_qual lacks the infrastructure to do. The case
3921 : * arises so seldom that it doesn't seem worth working hard on.
3922 : */
3923 2206 : if (var->varno != rti)
3924 : {
3925 12 : safe = PUSHDOWN_UNSAFE;
3926 12 : break;
3927 : }
3928 :
3929 : /* Subqueries have no system columns */
3930 : Assert(var->varattno >= 0);
3931 :
3932 : /* Check point 4 */
3933 2194 : if (var->varattno == 0)
3934 : {
3935 0 : safe = PUSHDOWN_UNSAFE;
3936 0 : break;
3937 : }
3938 :
3939 : /* Check point 5 */
3940 2194 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
3941 : {
3942 514 : if (safetyInfo->unsafeFlags[var->varattno] &
3943 : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
3944 : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
3945 : {
3946 190 : safe = PUSHDOWN_UNSAFE;
3947 190 : break;
3948 : }
3949 : else
3950 : {
3951 : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
3952 324 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
3953 : /* don't break, we might find another Var that's unsafe */
3954 : }
3955 : }
3956 : }
3957 :
3958 2098 : list_free(vars);
3959 :
3960 2098 : return safe;
3961 : }
3962 :
3963 : /*
3964 : * subquery_push_qual - push down a qual that we have determined is safe
3965 : */
3966 : static void
3967 1774 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3968 : {
3969 1774 : if (subquery->setOperations != NULL)
3970 : {
3971 : /* Recurse to push it separately to each component query */
3972 62 : recurse_push_qual(subquery->setOperations, subquery,
3973 : rte, rti, qual);
3974 : }
3975 : else
3976 : {
3977 : /*
3978 : * We need to replace Vars in the qual (which must refer to outputs of
3979 : * the subquery) with copies of the subquery's targetlist expressions.
3980 : * Note that at this point, any uplevel Vars in the qual should have
3981 : * been replaced with Params, so they need no work.
3982 : *
3983 : * This step also ensures that when we are pushing into a setop tree,
3984 : * each component query gets its own copy of the qual.
3985 : */
3986 1712 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
3987 : subquery->targetList,
3988 : REPLACEVARS_REPORT_ERROR, 0,
3989 : &subquery->hasSubLinks);
3990 :
3991 : /*
3992 : * Now attach the qual to the proper place: normally WHERE, but if the
3993 : * subquery uses grouping or aggregation, put it in HAVING (since the
3994 : * qual really refers to the group-result rows).
3995 : */
3996 1712 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
3997 270 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
3998 : else
3999 1442 : subquery->jointree->quals =
4000 1442 : make_and_qual(subquery->jointree->quals, qual);
4001 :
4002 : /*
4003 : * We need not change the subquery's hasAggs or hasSubLinks flags,
4004 : * since we can't be pushing down any aggregates that weren't there
4005 : * before, and we don't push down subselects at all.
4006 : */
4007 : }
4008 1774 : }
4009 :
4010 : /*
4011 : * Helper routine to recurse through setOperations tree
4012 : */
4013 : static void
4014 186 : recurse_push_qual(Node *setOp, Query *topquery,
4015 : RangeTblEntry *rte, Index rti, Node *qual)
4016 : {
4017 186 : if (IsA(setOp, RangeTblRef))
4018 : {
4019 124 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4020 124 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4021 124 : Query *subquery = subrte->subquery;
4022 :
4023 : Assert(subquery != NULL);
4024 124 : subquery_push_qual(subquery, rte, rti, qual);
4025 : }
4026 62 : else if (IsA(setOp, SetOperationStmt))
4027 : {
4028 62 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4029 :
4030 62 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4031 62 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4032 : }
4033 : else
4034 : {
4035 0 : elog(ERROR, "unrecognized node type: %d",
4036 : (int) nodeTag(setOp));
4037 : }
4038 186 : }
4039 :
4040 : /*****************************************************************************
4041 : * SIMPLIFYING SUBQUERY TARGETLISTS
4042 : *****************************************************************************/
4043 :
4044 : /*
4045 : * remove_unused_subquery_outputs
4046 : * Remove subquery targetlist items we don't need
4047 : *
4048 : * It's possible, even likely, that the upper query does not read all the
4049 : * output columns of the subquery. We can remove any such outputs that are
4050 : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4051 : * affect semantics otherwise (e.g., volatile functions can't be removed).
4052 : * This is useful not only because we might be able to remove expensive-to-
4053 : * compute expressions, but because deletion of output columns might allow
4054 : * optimizations such as join removal to occur within the subquery.
4055 : *
4056 : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4057 : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4058 : * parameter is modified by the function, so callers must make a copy if they
4059 : * need to use the passed in Bitmapset after calling this function.
4060 : *
4061 : * To avoid affecting column numbering in the targetlist, we don't physically
4062 : * remove unused tlist entries, but rather replace their expressions with NULL
4063 : * constants. This is implemented by modifying subquery->targetList.
4064 : */
4065 : static void
4066 8056 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4067 : Bitmapset *extra_used_attrs)
4068 : {
4069 : Bitmapset *attrs_used;
4070 : ListCell *lc;
4071 :
4072 : /*
4073 : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4074 : * the current callers use the Bitmapset after calling this function.
4075 : */
4076 8056 : attrs_used = extra_used_attrs;
4077 :
4078 : /*
4079 : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4080 : * could update all the child SELECTs' tlists, but it seems not worth the
4081 : * trouble presently.
4082 : */
4083 8056 : if (subquery->setOperations)
4084 1074 : return;
4085 :
4086 : /*
4087 : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4088 : * time: all its output columns must be used in the distinctClause.
4089 : */
4090 7494 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4091 238 : return;
4092 :
4093 : /*
4094 : * Collect a bitmap of all the output column numbers used by the upper
4095 : * query.
4096 : *
4097 : * Add all the attributes needed for joins or final output. Note: we must
4098 : * look at rel's targetlist, not the attr_needed data, because attr_needed
4099 : * isn't computed for inheritance child rels, cf set_append_rel_size().
4100 : * (XXX might be worth changing that sometime.)
4101 : */
4102 7256 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4103 :
4104 : /* Add all the attributes used by un-pushed-down restriction clauses. */
4105 7950 : foreach(lc, rel->baserestrictinfo)
4106 : {
4107 694 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4108 :
4109 694 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4110 : }
4111 :
4112 : /*
4113 : * If there's a whole-row reference to the subquery, we can't remove
4114 : * anything.
4115 : */
4116 7256 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4117 274 : return;
4118 :
4119 : /*
4120 : * Run through the tlist and zap entries we don't need. It's okay to
4121 : * modify the tlist items in-place because set_subquery_pathlist made a
4122 : * copy of the subquery.
4123 : */
4124 32626 : foreach(lc, subquery->targetList)
4125 : {
4126 25644 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4127 25644 : Node *texpr = (Node *) tle->expr;
4128 :
4129 : /*
4130 : * If it has a sortgroupref number, it's used in some sort/group
4131 : * clause so we'd better not remove it. Also, don't remove any
4132 : * resjunk columns, since their reason for being has nothing to do
4133 : * with anybody reading the subquery's output. (It's likely that
4134 : * resjunk columns in a sub-SELECT would always have ressortgroupref
4135 : * set, but even if they don't, it seems imprudent to remove them.)
4136 : */
4137 25644 : if (tle->ressortgroupref || tle->resjunk)
4138 2482 : continue;
4139 :
4140 : /*
4141 : * If it's used by the upper query, we can't remove it.
4142 : */
4143 23162 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4144 : attrs_used))
4145 16230 : continue;
4146 :
4147 : /*
4148 : * If it contains a set-returning function, we can't remove it since
4149 : * that could change the number of rows returned by the subquery.
4150 : */
4151 7880 : if (subquery->hasTargetSRFs &&
4152 948 : expression_returns_set(texpr))
4153 692 : continue;
4154 :
4155 : /*
4156 : * If it contains volatile functions, we daren't remove it for fear
4157 : * that the user is expecting their side-effects to happen.
4158 : */
4159 6240 : if (contain_volatile_functions(texpr))
4160 26 : continue;
4161 :
4162 : /*
4163 : * OK, we don't need it. Replace the expression with a NULL constant.
4164 : * Preserve the exposed type of the expression, in case something
4165 : * looks at the rowtype of the subquery's result.
4166 : */
4167 6214 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4168 : exprTypmod(texpr),
4169 : exprCollation(texpr));
4170 : }
4171 : }
4172 :
4173 : /*
4174 : * create_partial_bitmap_paths
4175 : * Build partial bitmap heap path for the relation
4176 : */
4177 : void
4178 135540 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4179 : Path *bitmapqual)
4180 : {
4181 : int parallel_workers;
4182 : double pages_fetched;
4183 :
4184 : /* Compute heap pages for bitmap heap scan */
4185 135540 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4186 : NULL, NULL);
4187 :
4188 135540 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4189 : max_parallel_workers_per_gather);
4190 :
4191 135540 : if (parallel_workers <= 0)
4192 131380 : return;
4193 :
4194 4160 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4195 : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4196 : }
4197 :
4198 : /*
4199 : * Compute the number of parallel workers that should be used to scan a
4200 : * relation. We compute the parallel workers based on the size of the heap to
4201 : * be scanned and the size of the index to be scanned, then choose a minimum
4202 : * of those.
4203 : *
4204 : * "heap_pages" is the number of pages from the table that we expect to scan, or
4205 : * -1 if we don't expect to scan any.
4206 : *
4207 : * "index_pages" is the number of pages from the index that we expect to scan, or
4208 : * -1 if we don't expect to scan any.
4209 : *
4210 : * "max_workers" is caller's limit on the number of workers. This typically
4211 : * comes from a GUC.
4212 : */
4213 : int
4214 671546 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4215 : int max_workers)
4216 : {
4217 671546 : int parallel_workers = 0;
4218 :
4219 : /*
4220 : * If the user has set the parallel_workers reloption, use that; otherwise
4221 : * select a default number of workers.
4222 : */
4223 671546 : if (rel->rel_parallel_workers != -1)
4224 1914 : parallel_workers = rel->rel_parallel_workers;
4225 : else
4226 : {
4227 : /*
4228 : * If the number of pages being scanned is insufficient to justify a
4229 : * parallel scan, just return zero ... unless it's an inheritance
4230 : * child. In that case, we want to generate a parallel path here
4231 : * anyway. It might not be worthwhile just for this relation, but
4232 : * when combined with all of its inheritance siblings it may well pay
4233 : * off.
4234 : */
4235 669632 : if (rel->reloptkind == RELOPT_BASEREL &&
4236 631970 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4237 19076 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4238 631282 : return 0;
4239 :
4240 38350 : if (heap_pages >= 0)
4241 : {
4242 : int heap_parallel_threshold;
4243 36324 : int heap_parallel_workers = 1;
4244 :
4245 : /*
4246 : * Select the number of workers based on the log of the size of
4247 : * the relation. This probably needs to be a good deal more
4248 : * sophisticated, but we need something here for now. Note that
4249 : * the upper limit of the min_parallel_table_scan_size GUC is
4250 : * chosen to prevent overflow here.
4251 : */
4252 36324 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4253 41074 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4254 : {
4255 4750 : heap_parallel_workers++;
4256 4750 : heap_parallel_threshold *= 3;
4257 4750 : if (heap_parallel_threshold > INT_MAX / 3)
4258 0 : break; /* avoid overflow */
4259 : }
4260 :
4261 36324 : parallel_workers = heap_parallel_workers;
4262 : }
4263 :
4264 38350 : if (index_pages >= 0)
4265 : {
4266 9584 : int index_parallel_workers = 1;
4267 : int index_parallel_threshold;
4268 :
4269 : /* same calculation as for heap_pages above */
4270 9584 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4271 9860 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4272 : {
4273 276 : index_parallel_workers++;
4274 276 : index_parallel_threshold *= 3;
4275 276 : if (index_parallel_threshold > INT_MAX / 3)
4276 0 : break; /* avoid overflow */
4277 : }
4278 :
4279 9584 : if (parallel_workers > 0)
4280 7558 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4281 : else
4282 2026 : parallel_workers = index_parallel_workers;
4283 : }
4284 : }
4285 :
4286 : /* In no case use more than caller supplied maximum number of workers */
4287 40264 : parallel_workers = Min(parallel_workers, max_workers);
4288 :
4289 40264 : return parallel_workers;
4290 : }
4291 :
4292 : /*
4293 : * generate_partitionwise_join_paths
4294 : * Create paths representing partitionwise join for given partitioned
4295 : * join relation.
4296 : *
4297 : * This must not be called until after we are done adding paths for all
4298 : * child-joins. Otherwise, add_path might delete a path to which some path
4299 : * generated here has a reference.
4300 : */
4301 : void
4302 184986 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4303 : {
4304 184986 : List *live_children = NIL;
4305 : int cnt_parts;
4306 : int num_parts;
4307 : RelOptInfo **part_rels;
4308 :
4309 : /* Handle only join relations here. */
4310 184986 : if (!IS_JOIN_REL(rel))
4311 0 : return;
4312 :
4313 : /* We've nothing to do if the relation is not partitioned. */
4314 184986 : if (!IS_PARTITIONED_REL(rel))
4315 183270 : return;
4316 :
4317 : /* The relation should have consider_partitionwise_join set. */
4318 : Assert(rel->consider_partitionwise_join);
4319 :
4320 : /* Guard against stack overflow due to overly deep partition hierarchy. */
4321 1716 : check_stack_depth();
4322 :
4323 1716 : num_parts = rel->nparts;
4324 1716 : part_rels = rel->part_rels;
4325 :
4326 : /* Collect non-dummy child-joins. */
4327 6410 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4328 : {
4329 4694 : RelOptInfo *child_rel = part_rels[cnt_parts];
4330 :
4331 : /* If it's been pruned entirely, it's certainly dummy. */
4332 4694 : if (child_rel == NULL)
4333 52 : continue;
4334 :
4335 : /* Make partitionwise join paths for this partitioned child-join. */
4336 4642 : generate_partitionwise_join_paths(root, child_rel);
4337 :
4338 : /* If we failed to make any path for this child, we must give up. */
4339 4642 : if (child_rel->pathlist == NIL)
4340 : {
4341 : /*
4342 : * Mark the parent joinrel as unpartitioned so that later
4343 : * functions treat it correctly.
4344 : */
4345 0 : rel->nparts = 0;
4346 0 : return;
4347 : }
4348 :
4349 : /* Else, identify the cheapest path for it. */
4350 4642 : set_cheapest(child_rel);
4351 :
4352 : /* Dummy children need not be scanned, so ignore those. */
4353 4642 : if (IS_DUMMY_REL(child_rel))
4354 0 : continue;
4355 :
4356 : #ifdef OPTIMIZER_DEBUG
4357 : pprint(child_rel);
4358 : #endif
4359 :
4360 4642 : live_children = lappend(live_children, child_rel);
4361 : }
4362 :
4363 : /* If all child-joins are dummy, parent join is also dummy. */
4364 1716 : if (!live_children)
4365 : {
4366 0 : mark_dummy_rel(rel);
4367 0 : return;
4368 : }
4369 :
4370 : /* Build additional paths for this rel from child-join paths. */
4371 1716 : add_paths_to_append_rel(root, rel, live_children);
4372 1716 : list_free(live_children);
4373 : }
|
