Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * joinrels.c
4 : * Routines to determine which relations should be joined
5 : *
6 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/optimizer/path/joinrels.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include "miscadmin.h"
18 : #include "optimizer/appendinfo.h"
19 : #include "optimizer/joininfo.h"
20 : #include "optimizer/pathnode.h"
21 : #include "optimizer/paths.h"
22 : #include "partitioning/partbounds.h"
23 : #include "utils/memutils.h"
24 :
25 :
26 : static void make_rels_by_clause_joins(PlannerInfo *root,
27 : RelOptInfo *old_rel,
28 : List *other_rels,
29 : int first_rel_idx);
30 : static void make_rels_by_clauseless_joins(PlannerInfo *root,
31 : RelOptInfo *old_rel,
32 : List *other_rels);
33 : static bool has_join_restriction(PlannerInfo *root, RelOptInfo *rel);
34 : static bool has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel);
35 : static bool restriction_is_constant_false(List *restrictlist,
36 : RelOptInfo *joinrel,
37 : bool only_pushed_down);
38 : static void populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
39 : RelOptInfo *rel2, RelOptInfo *joinrel,
40 : SpecialJoinInfo *sjinfo, List *restrictlist);
41 : static void try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1,
42 : RelOptInfo *rel2, RelOptInfo *joinrel,
43 : SpecialJoinInfo *parent_sjinfo,
44 : List *parent_restrictlist);
45 : static SpecialJoinInfo *build_child_join_sjinfo(PlannerInfo *root,
46 : SpecialJoinInfo *parent_sjinfo,
47 : Relids left_relids, Relids right_relids);
48 : static void compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
49 : RelOptInfo *rel2, RelOptInfo *joinrel,
50 : SpecialJoinInfo *parent_sjinfo,
51 : List **parts1, List **parts2);
52 : static void get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
53 : RelOptInfo *rel1, RelOptInfo *rel2,
54 : List **parts1, List **parts2);
55 :
56 :
57 : /*
58 : * join_search_one_level
59 : * Consider ways to produce join relations containing exactly 'level'
60 : * jointree items. (This is one step of the dynamic-programming method
61 : * embodied in standard_join_search.) Join rel nodes for each feasible
62 : * combination of lower-level rels are created and returned in a list.
63 : * Implementation paths are created for each such joinrel, too.
64 : *
65 : * level: level of rels we want to make this time
66 : * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
67 : *
68 : * The result is returned in root->join_rel_level[level].
69 : */
70 : void
71 106664 : join_search_one_level(PlannerInfo *root, int level)
72 : {
73 106664 : List **joinrels = root->join_rel_level;
74 : ListCell *r;
75 : int k;
76 :
77 : Assert(joinrels[level] == NIL);
78 :
79 : /* Set join_cur_level so that new joinrels are added to proper list */
80 106664 : root->join_cur_level = level;
81 :
82 : /*
83 : * First, consider left-sided and right-sided plans, in which rels of
84 : * exactly level-1 member relations are joined against initial relations.
85 : * We prefer to join using join clauses, but if we find a rel of level-1
86 : * members that has no join clauses, we will generate Cartesian-product
87 : * joins against all initial rels not already contained in it.
88 : */
89 376412 : foreach(r, joinrels[level - 1])
90 : {
91 269748 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
92 :
93 296996 : if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
94 27248 : has_join_restriction(root, old_rel))
95 260912 : {
96 : int first_rel;
97 :
98 : /*
99 : * There are join clauses or join order restrictions relevant to
100 : * this rel, so consider joins between this rel and (only) those
101 : * initial rels it is linked to by a clause or restriction.
102 : *
103 : * At level 2 this condition is symmetric, so there is no need to
104 : * look at initial rels before this one in the list; we already
105 : * considered such joins when we were at the earlier rel. (The
106 : * mirror-image joins are handled automatically by make_join_rel.)
107 : * In later passes (level > 2), we join rels of the previous level
108 : * to each initial rel they don't already include but have a join
109 : * clause or restriction with.
110 : */
111 260912 : if (level == 2) /* consider remaining initial rels */
112 175778 : first_rel = foreach_current_index(r) + 1;
113 : else
114 85134 : first_rel = 0;
115 :
116 260912 : make_rels_by_clause_joins(root, old_rel, joinrels[1], first_rel);
117 : }
118 : else
119 : {
120 : /*
121 : * Oops, we have a relation that is not joined to any other
122 : * relation, either directly or by join-order restrictions.
123 : * Cartesian product time.
124 : *
125 : * We consider a cartesian product with each not-already-included
126 : * initial rel, whether it has other join clauses or not. At
127 : * level 2, if there are two or more clauseless initial rels, we
128 : * will redundantly consider joining them in both directions; but
129 : * such cases aren't common enough to justify adding complexity to
130 : * avoid the duplicated effort.
131 : */
132 8836 : make_rels_by_clauseless_joins(root,
133 : old_rel,
134 8836 : joinrels[1]);
135 : }
136 : }
137 :
138 : /*
139 : * Now, consider "bushy plans" in which relations of k initial rels are
140 : * joined to relations of level-k initial rels, for 2 <= k <= level-2.
141 : *
142 : * We only consider bushy-plan joins for pairs of rels where there is a
143 : * suitable join clause (or join order restriction), in order to avoid
144 : * unreasonable growth of planning time.
145 : */
146 106664 : for (k = 2;; k++)
147 10952 : {
148 117616 : int other_level = level - k;
149 :
150 : /*
151 : * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
152 : * need to go as far as the halfway point.
153 : */
154 117616 : if (k > other_level)
155 106664 : break;
156 :
157 54952 : foreach(r, joinrels[k])
158 : {
159 44000 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
160 : int first_rel;
161 : ListCell *r2;
162 :
163 : /*
164 : * We can ignore relations without join clauses here, unless they
165 : * participate in join-order restrictions --- then we might have
166 : * to force a bushy join plan.
167 : */
168 44000 : if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
169 276 : !has_join_restriction(root, old_rel))
170 180 : continue;
171 :
172 43820 : if (k == other_level) /* only consider remaining rels */
173 30752 : first_rel = foreach_current_index(r) + 1;
174 : else
175 13068 : first_rel = 0;
176 :
177 185422 : for_each_from(r2, joinrels[other_level], first_rel)
178 : {
179 141602 : RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
180 :
181 141602 : if (!bms_overlap(old_rel->relids, new_rel->relids))
182 : {
183 : /*
184 : * OK, we can build a rel of the right level from this
185 : * pair of rels. Do so if there is at least one relevant
186 : * join clause or join order restriction.
187 : */
188 18184 : if (have_relevant_joinclause(root, old_rel, new_rel) ||
189 1018 : have_join_order_restriction(root, old_rel, new_rel))
190 : {
191 16202 : (void) make_join_rel(root, old_rel, new_rel);
192 : }
193 : }
194 : }
195 : }
196 : }
197 :
198 : /*----------
199 : * Last-ditch effort: if we failed to find any usable joins so far, force
200 : * a set of cartesian-product joins to be generated. This handles the
201 : * special case where all the available rels have join clauses but we
202 : * cannot use any of those clauses yet. This can only happen when we are
203 : * considering a join sub-problem (a sub-joinlist) and all the rels in the
204 : * sub-problem have only join clauses with rels outside the sub-problem.
205 : * An example is
206 : *
207 : * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
208 : * WHERE a.w = c.x and b.y = d.z;
209 : *
210 : * If the "a INNER JOIN b" sub-problem does not get flattened into the
211 : * upper level, we must be willing to make a cartesian join of a and b;
212 : * but the code above will not have done so, because it thought that both
213 : * a and b have joinclauses. We consider only left-sided and right-sided
214 : * cartesian joins in this case (no bushy).
215 : *----------
216 : */
217 106664 : if (joinrels[level] == NIL)
218 : {
219 : /*
220 : * This loop is just like the first one, except we always call
221 : * make_rels_by_clauseless_joins().
222 : */
223 54 : foreach(r, joinrels[level - 1])
224 : {
225 36 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
226 :
227 36 : make_rels_by_clauseless_joins(root,
228 : old_rel,
229 36 : joinrels[1]);
230 : }
231 :
232 : /*----------
233 : * When special joins are involved, there may be no legal way
234 : * to make an N-way join for some values of N. For example consider
235 : *
236 : * SELECT ... FROM t1 WHERE
237 : * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
238 : * y IN (SELECT ... FROM t4,t5 WHERE ...)
239 : *
240 : * We will flatten this query to a 5-way join problem, but there are
241 : * no 4-way joins that join_is_legal() will consider legal. We have
242 : * to accept failure at level 4 and go on to discover a workable
243 : * bushy plan at level 5.
244 : *
245 : * However, if there are no special joins and no lateral references
246 : * then join_is_legal() should never fail, and so the following sanity
247 : * check is useful.
248 : *----------
249 : */
250 18 : if (joinrels[level] == NIL &&
251 6 : root->join_info_list == NIL &&
252 0 : !root->hasLateralRTEs)
253 0 : elog(ERROR, "failed to build any %d-way joins", level);
254 : }
255 106664 : }
256 :
257 : /*
258 : * make_rels_by_clause_joins
259 : * Build joins between the given relation 'old_rel' and other relations
260 : * that participate in join clauses that 'old_rel' also participates in
261 : * (or participate in join-order restrictions with it).
262 : * The join rels are returned in root->join_rel_level[join_cur_level].
263 : *
264 : * Note: at levels above 2 we will generate the same joined relation in
265 : * multiple ways --- for example (a join b) join c is the same RelOptInfo as
266 : * (b join c) join a, though the second case will add a different set of Paths
267 : * to it. This is the reason for using the join_rel_level mechanism, which
268 : * automatically ensures that each new joinrel is only added to the list once.
269 : *
270 : * 'old_rel' is the relation entry for the relation to be joined
271 : * 'other_rels': a list containing the other rels to be considered for joining
272 : * 'first_rel_idx': the first rel to be considered in 'other_rels'
273 : *
274 : * Currently, this is only used with initial rels in other_rels, but it
275 : * will work for joining to joinrels too.
276 : */
277 : static void
278 260912 : make_rels_by_clause_joins(PlannerInfo *root,
279 : RelOptInfo *old_rel,
280 : List *other_rels,
281 : int first_rel_idx)
282 : {
283 : ListCell *l;
284 :
285 769044 : for_each_from(l, other_rels, first_rel_idx)
286 : {
287 508132 : RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
288 :
289 797884 : if (!bms_overlap(old_rel->relids, other_rel->relids) &&
290 353230 : (have_relevant_joinclause(root, old_rel, other_rel) ||
291 63478 : have_join_order_restriction(root, old_rel, other_rel)))
292 : {
293 236790 : (void) make_join_rel(root, old_rel, other_rel);
294 : }
295 : }
296 260912 : }
297 :
298 : /*
299 : * make_rels_by_clauseless_joins
300 : * Given a relation 'old_rel' and a list of other relations
301 : * 'other_rels', create a join relation between 'old_rel' and each
302 : * member of 'other_rels' that isn't already included in 'old_rel'.
303 : * The join rels are returned in root->join_rel_level[join_cur_level].
304 : *
305 : * 'old_rel' is the relation entry for the relation to be joined
306 : * 'other_rels': a list containing the other rels to be considered for joining
307 : *
308 : * Currently, this is only used with initial rels in other_rels, but it would
309 : * work for joining to joinrels too.
310 : */
311 : static void
312 8872 : make_rels_by_clauseless_joins(PlannerInfo *root,
313 : RelOptInfo *old_rel,
314 : List *other_rels)
315 : {
316 : ListCell *l;
317 :
318 28128 : foreach(l, other_rels)
319 : {
320 19256 : RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
321 :
322 19256 : if (!bms_overlap(other_rel->relids, old_rel->relids))
323 : {
324 9510 : (void) make_join_rel(root, old_rel, other_rel);
325 : }
326 : }
327 8872 : }
328 :
329 :
330 : /*
331 : * join_is_legal
332 : * Determine whether a proposed join is legal given the query's
333 : * join order constraints; and if it is, determine the join type.
334 : *
335 : * Caller must supply not only the two rels, but the union of their relids.
336 : * (We could simplify the API by computing joinrelids locally, but this
337 : * would be redundant work in the normal path through make_join_rel.
338 : * Note that this value does NOT include the RT index of any outer join that
339 : * might need to be performed here, so it's not the canonical identifier
340 : * of the join relation.)
341 : *
342 : * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
343 : * else it's set to point to the associated SpecialJoinInfo node. Also,
344 : * *reversed_p is set true if the given relations need to be swapped to
345 : * match the SpecialJoinInfo node.
346 : */
347 : static bool
348 267542 : join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
349 : Relids joinrelids,
350 : SpecialJoinInfo **sjinfo_p, bool *reversed_p)
351 : {
352 : SpecialJoinInfo *match_sjinfo;
353 : bool reversed;
354 : bool unique_ified;
355 : bool must_be_leftjoin;
356 : ListCell *l;
357 :
358 : /*
359 : * Ensure output params are set on failure return. This is just to
360 : * suppress uninitialized-variable warnings from overly anal compilers.
361 : */
362 267542 : *sjinfo_p = NULL;
363 267542 : *reversed_p = false;
364 :
365 : /*
366 : * If we have any special joins, the proposed join might be illegal; and
367 : * in any case we have to determine its join type. Scan the join info
368 : * list for matches and conflicts.
369 : */
370 267542 : match_sjinfo = NULL;
371 267542 : reversed = false;
372 267542 : unique_ified = false;
373 267542 : must_be_leftjoin = false;
374 :
375 583828 : foreach(l, root->join_info_list)
376 : {
377 326200 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
378 :
379 : /*
380 : * This special join is not relevant unless its RHS overlaps the
381 : * proposed join. (Check this first as a fast path for dismissing
382 : * most irrelevant SJs quickly.)
383 : */
384 326200 : if (!bms_overlap(sjinfo->min_righthand, joinrelids))
385 111236 : continue;
386 :
387 : /*
388 : * Also, not relevant if proposed join is fully contained within RHS
389 : * (ie, we're still building up the RHS).
390 : */
391 214964 : if (bms_is_subset(joinrelids, sjinfo->min_righthand))
392 4888 : continue;
393 :
394 : /*
395 : * Also, not relevant if SJ is already done within either input.
396 : */
397 391616 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
398 181540 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
399 88926 : continue;
400 138382 : if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
401 17232 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
402 8710 : continue;
403 :
404 : /*
405 : * If it's a semijoin and we already joined the RHS to any other rels
406 : * within either input, then we must have unique-ified the RHS at that
407 : * point (see below). Therefore the semijoin is no longer relevant in
408 : * this join path.
409 : */
410 112440 : if (sjinfo->jointype == JOIN_SEMI)
411 : {
412 7440 : if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
413 1528 : !bms_equal(sjinfo->syn_righthand, rel1->relids))
414 600 : continue;
415 6840 : if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
416 3830 : !bms_equal(sjinfo->syn_righthand, rel2->relids))
417 214 : continue;
418 : }
419 :
420 : /*
421 : * If one input contains min_lefthand and the other contains
422 : * min_righthand, then we can perform the SJ at this join.
423 : *
424 : * Reject if we get matches to more than one SJ; that implies we're
425 : * considering something that's not really valid.
426 : */
427 204088 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
428 92462 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
429 : {
430 86458 : if (match_sjinfo)
431 9914 : return false; /* invalid join path */
432 86458 : match_sjinfo = sjinfo;
433 86458 : reversed = false;
434 : }
435 33358 : else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
436 8190 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
437 : {
438 7044 : if (match_sjinfo)
439 0 : return false; /* invalid join path */
440 7044 : match_sjinfo = sjinfo;
441 7044 : reversed = true;
442 : }
443 21024 : else if (sjinfo->jointype == JOIN_SEMI &&
444 3428 : bms_equal(sjinfo->syn_righthand, rel2->relids) &&
445 528 : create_unique_path(root, rel2, rel2->cheapest_total_path,
446 : sjinfo) != NULL)
447 : {
448 : /*----------
449 : * For a semijoin, we can join the RHS to anything else by
450 : * unique-ifying the RHS (if the RHS can be unique-ified).
451 : * We will only get here if we have the full RHS but less
452 : * than min_lefthand on the LHS.
453 : *
454 : * The reason to consider such a join path is exemplified by
455 : * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
456 : * If we insist on doing this as a semijoin we will first have
457 : * to form the cartesian product of A*B. But if we unique-ify
458 : * C then the semijoin becomes a plain innerjoin and we can join
459 : * in any order, eg C to A and then to B. When C is much smaller
460 : * than A and B this can be a huge win. So we allow C to be
461 : * joined to just A or just B here, and then make_join_rel has
462 : * to handle the case properly.
463 : *
464 : * Note that actually we'll allow unique-ified C to be joined to
465 : * some other relation D here, too. That is legal, if usually not
466 : * very sane, and this routine is only concerned with legality not
467 : * with whether the join is good strategy.
468 : *----------
469 : */
470 522 : if (match_sjinfo)
471 84 : return false; /* invalid join path */
472 438 : match_sjinfo = sjinfo;
473 438 : reversed = false;
474 438 : unique_ified = true;
475 : }
476 19980 : else if (sjinfo->jointype == JOIN_SEMI &&
477 2668 : bms_equal(sjinfo->syn_righthand, rel1->relids) &&
478 290 : create_unique_path(root, rel1, rel1->cheapest_total_path,
479 : sjinfo) != NULL)
480 : {
481 : /* Reversed semijoin case */
482 290 : if (match_sjinfo)
483 78 : return false; /* invalid join path */
484 212 : match_sjinfo = sjinfo;
485 212 : reversed = true;
486 212 : unique_ified = true;
487 : }
488 : else
489 : {
490 : /*
491 : * Otherwise, the proposed join overlaps the RHS but isn't a valid
492 : * implementation of this SJ. But don't panic quite yet: the RHS
493 : * violation might have occurred previously, in one or both input
494 : * relations, in which case we must have previously decided that
495 : * it was OK to commute some other SJ with this one. If we need
496 : * to perform this join to finish building up the RHS, rejecting
497 : * it could lead to not finding any plan at all. (This can occur
498 : * because of the heuristics elsewhere in this file that postpone
499 : * clauseless joins: we might not consider doing a clauseless join
500 : * within the RHS until after we've performed other, validly
501 : * commutable SJs with one or both sides of the clauseless join.)
502 : * This consideration boils down to the rule that if both inputs
503 : * overlap the RHS, we can allow the join --- they are either
504 : * fully within the RHS, or represent previously-allowed joins to
505 : * rels outside it.
506 : */
507 24726 : if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
508 7414 : bms_overlap(rel2->relids, sjinfo->min_righthand))
509 174 : continue; /* assume valid previous violation of RHS */
510 :
511 : /*
512 : * The proposed join could still be legal, but only if we're
513 : * allowed to associate it into the RHS of this SJ. That means
514 : * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
515 : * not FULL) and the proposed join must not overlap the LHS.
516 : */
517 32056 : if (sjinfo->jointype != JOIN_LEFT ||
518 14918 : bms_overlap(joinrelids, sjinfo->min_lefthand))
519 9752 : return false; /* invalid join path */
520 :
521 : /*
522 : * To be valid, the proposed join must be a LEFT join; otherwise
523 : * it can't associate into this SJ's RHS. But we may not yet have
524 : * found the SpecialJoinInfo matching the proposed join, so we
525 : * can't test that yet. Remember the requirement for later.
526 : */
527 7386 : must_be_leftjoin = true;
528 : }
529 : }
530 :
531 : /*
532 : * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
533 : * proposed join can't associate into an SJ's RHS.
534 : *
535 : * Also, fail if the proposed join's predicate isn't strict; we're
536 : * essentially checking to see if we can apply outer-join identity 3, and
537 : * that's a requirement. (This check may be redundant with checks in
538 : * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
539 : */
540 257628 : if (must_be_leftjoin &&
541 4622 : (match_sjinfo == NULL ||
542 4622 : match_sjinfo->jointype != JOIN_LEFT ||
543 4622 : !match_sjinfo->lhs_strict))
544 1418 : return false; /* invalid join path */
545 :
546 : /*
547 : * We also have to check for constraints imposed by LATERAL references.
548 : */
549 256210 : if (root->hasLateralRTEs)
550 : {
551 : bool lateral_fwd;
552 : bool lateral_rev;
553 : Relids join_lateral_rels;
554 :
555 : /*
556 : * The proposed rels could each contain lateral references to the
557 : * other, in which case the join is impossible. If there are lateral
558 : * references in just one direction, then the join has to be done with
559 : * a nestloop with the lateral referencer on the inside. If the join
560 : * matches an SJ that cannot be implemented by such a nestloop, the
561 : * join is impossible.
562 : *
563 : * Also, if the lateral reference is only indirect, we should reject
564 : * the join; whatever rel(s) the reference chain goes through must be
565 : * joined to first.
566 : *
567 : * Another case that might keep us from building a valid plan is the
568 : * implementation restriction described by have_dangerous_phv().
569 : */
570 14492 : lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
571 14492 : lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
572 14492 : if (lateral_fwd && lateral_rev)
573 18 : return false; /* have lateral refs in both directions */
574 14474 : if (lateral_fwd)
575 : {
576 : /* has to be implemented as nestloop with rel1 on left */
577 9408 : if (match_sjinfo &&
578 204 : (reversed ||
579 192 : unique_ified ||
580 192 : match_sjinfo->jointype == JOIN_FULL))
581 12 : return false; /* not implementable as nestloop */
582 : /* check there is a direct reference from rel2 to rel1 */
583 9396 : if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
584 42 : return false; /* only indirect refs, so reject */
585 : /* check we won't have a dangerous PHV */
586 9354 : if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
587 72 : return false; /* might be unable to handle required PHV */
588 : }
589 5066 : else if (lateral_rev)
590 : {
591 : /* has to be implemented as nestloop with rel2 on left */
592 1056 : if (match_sjinfo &&
593 72 : (!reversed ||
594 72 : unique_ified ||
595 72 : match_sjinfo->jointype == JOIN_FULL))
596 0 : return false; /* not implementable as nestloop */
597 : /* check there is a direct reference from rel1 to rel2 */
598 1056 : if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
599 0 : return false; /* only indirect refs, so reject */
600 : /* check we won't have a dangerous PHV */
601 1056 : if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
602 84 : return false; /* might be unable to handle required PHV */
603 : }
604 :
605 : /*
606 : * LATERAL references could also cause problems later on if we accept
607 : * this join: if the join's minimum parameterization includes any rels
608 : * that would have to be on the inside of an outer join with this join
609 : * rel, then it's never going to be possible to build the complete
610 : * query using this join. We should reject this join not only because
611 : * it'll save work, but because if we don't, the clauseless-join
612 : * heuristics might think that legality of this join means that some
613 : * other join rel need not be formed, and that could lead to failure
614 : * to find any plan at all. We have to consider not only rels that
615 : * are directly on the inner side of an OJ with the joinrel, but also
616 : * ones that are indirectly so, so search to find all such rels.
617 : */
618 14264 : join_lateral_rels = min_join_parameterization(root, joinrelids,
619 : rel1, rel2);
620 14264 : if (join_lateral_rels)
621 : {
622 1488 : Relids join_plus_rhs = bms_copy(joinrelids);
623 : bool more;
624 :
625 : do
626 : {
627 1800 : more = false;
628 3222 : foreach(l, root->join_info_list)
629 : {
630 1422 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
631 :
632 : /* ignore full joins --- their ordering is predetermined */
633 1422 : if (sjinfo->jointype == JOIN_FULL)
634 18 : continue;
635 :
636 1404 : if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
637 1182 : !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
638 : {
639 402 : join_plus_rhs = bms_add_members(join_plus_rhs,
640 402 : sjinfo->min_righthand);
641 402 : more = true;
642 : }
643 : }
644 1800 : } while (more);
645 1488 : if (bms_overlap(join_plus_rhs, join_lateral_rels))
646 228 : return false; /* will not be able to join to some RHS rel */
647 : }
648 : }
649 :
650 : /* Otherwise, it's a valid join */
651 255754 : *sjinfo_p = match_sjinfo;
652 255754 : *reversed_p = reversed;
653 255754 : return true;
654 : }
655 :
656 :
657 : /*
658 : * make_join_rel
659 : * Find or create a join RelOptInfo that represents the join of
660 : * the two given rels, and add to it path information for paths
661 : * created with the two rels as outer and inner rel.
662 : * (The join rel may already contain paths generated from other
663 : * pairs of rels that add up to the same set of base rels.)
664 : *
665 : * NB: will return NULL if attempted join is not valid. This can happen
666 : * when working with outer joins, or with IN or EXISTS clauses that have been
667 : * turned into joins.
668 : */
669 : RelOptInfo *
670 267254 : make_join_rel(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2)
671 : {
672 : Relids joinrelids;
673 : SpecialJoinInfo *sjinfo;
674 : bool reversed;
675 267254 : List *pushed_down_joins = NIL;
676 : SpecialJoinInfo sjinfo_data;
677 : RelOptInfo *joinrel;
678 : List *restrictlist;
679 :
680 : /* We should never try to join two overlapping sets of rels. */
681 : Assert(!bms_overlap(rel1->relids, rel2->relids));
682 :
683 : /* Construct Relids set that identifies the joinrel (without OJ as yet). */
684 267254 : joinrelids = bms_union(rel1->relids, rel2->relids);
685 :
686 : /* Check validity and determine join type. */
687 267254 : if (!join_is_legal(root, rel1, rel2, joinrelids,
688 : &sjinfo, &reversed))
689 : {
690 : /* invalid join path */
691 11650 : bms_free(joinrelids);
692 11650 : return NULL;
693 : }
694 :
695 : /*
696 : * Add outer join relid(s) to form the canonical relids. Any added outer
697 : * joins besides sjinfo itself are appended to pushed_down_joins.
698 : */
699 255604 : joinrelids = add_outer_joins_to_relids(root, joinrelids, sjinfo,
700 : &pushed_down_joins);
701 :
702 : /* Swap rels if needed to match the join info. */
703 255604 : if (reversed)
704 : {
705 7100 : RelOptInfo *trel = rel1;
706 :
707 7100 : rel1 = rel2;
708 7100 : rel2 = trel;
709 : }
710 :
711 : /*
712 : * If it's a plain inner join, then we won't have found anything in
713 : * join_info_list. Make up a SpecialJoinInfo so that selectivity
714 : * estimation functions will know what's being joined.
715 : */
716 255604 : if (sjinfo == NULL)
717 : {
718 161948 : sjinfo = &sjinfo_data;
719 161948 : sjinfo->type = T_SpecialJoinInfo;
720 161948 : sjinfo->min_lefthand = rel1->relids;
721 161948 : sjinfo->min_righthand = rel2->relids;
722 161948 : sjinfo->syn_lefthand = rel1->relids;
723 161948 : sjinfo->syn_righthand = rel2->relids;
724 161948 : sjinfo->jointype = JOIN_INNER;
725 161948 : sjinfo->ojrelid = 0;
726 161948 : sjinfo->commute_above_l = NULL;
727 161948 : sjinfo->commute_above_r = NULL;
728 161948 : sjinfo->commute_below_l = NULL;
729 161948 : sjinfo->commute_below_r = NULL;
730 : /* we don't bother trying to make the remaining fields valid */
731 161948 : sjinfo->lhs_strict = false;
732 161948 : sjinfo->semi_can_btree = false;
733 161948 : sjinfo->semi_can_hash = false;
734 161948 : sjinfo->semi_operators = NIL;
735 161948 : sjinfo->semi_rhs_exprs = NIL;
736 : }
737 :
738 : /*
739 : * Find or build the join RelOptInfo, and compute the restrictlist that
740 : * goes with this particular joining.
741 : */
742 255604 : joinrel = build_join_rel(root, joinrelids, rel1, rel2,
743 : sjinfo, pushed_down_joins,
744 : &restrictlist);
745 :
746 : /*
747 : * If we've already proven this join is empty, we needn't consider any
748 : * more paths for it.
749 : */
750 255604 : if (is_dummy_rel(joinrel))
751 : {
752 456 : bms_free(joinrelids);
753 456 : return joinrel;
754 : }
755 :
756 : /* Add paths to the join relation. */
757 255148 : populate_joinrel_with_paths(root, rel1, rel2, joinrel, sjinfo,
758 : restrictlist);
759 :
760 255148 : bms_free(joinrelids);
761 :
762 255148 : return joinrel;
763 : }
764 :
765 : /*
766 : * add_outer_joins_to_relids
767 : * Add relids to input_relids to represent any outer joins that will be
768 : * calculated at this join.
769 : *
770 : * input_relids is the union of the relid sets of the two input relations.
771 : * Note that we modify this in-place and return it; caller must bms_copy()
772 : * it first, if a separate value is desired.
773 : *
774 : * sjinfo represents the join being performed.
775 : *
776 : * If the current join completes the calculation of any outer joins that
777 : * have been pushed down per outer-join identity 3, those relids will be
778 : * added to the result along with sjinfo's own relid. If pushed_down_joins
779 : * is not NULL, then also the SpecialJoinInfos for such added outer joins will
780 : * be appended to *pushed_down_joins (so caller must initialize it to NIL).
781 : */
782 : Relids
783 262256 : add_outer_joins_to_relids(PlannerInfo *root, Relids input_relids,
784 : SpecialJoinInfo *sjinfo,
785 : List **pushed_down_joins)
786 : {
787 : /* Nothing to do if this isn't an outer join with an assigned relid. */
788 262256 : if (sjinfo == NULL || sjinfo->ojrelid == 0)
789 175668 : return input_relids;
790 :
791 : /*
792 : * If it's not a left join, we have no rules that would permit executing
793 : * it in non-syntactic order, so just form the syntactic relid set. (This
794 : * is just a quick-exit test; we'd come to the same conclusion anyway,
795 : * since its commute_below_l and commute_above_l sets must be empty.)
796 : */
797 86588 : if (sjinfo->jointype != JOIN_LEFT)
798 1892 : return bms_add_member(input_relids, sjinfo->ojrelid);
799 :
800 : /*
801 : * We cannot add the OJ relid if this join has been pushed into the RHS of
802 : * a syntactically-lower left join per OJ identity 3. (If it has, then we
803 : * cannot claim that its outputs represent the final state of its RHS.)
804 : * There will not be any other OJs that can be added either, so we're
805 : * done.
806 : */
807 84696 : if (!bms_is_subset(sjinfo->commute_below_l, input_relids))
808 4382 : return input_relids;
809 :
810 : /* OK to add OJ's own relid */
811 80314 : input_relids = bms_add_member(input_relids, sjinfo->ojrelid);
812 :
813 : /*
814 : * Contrariwise, if we are now forming the final result of such a commuted
815 : * pair of OJs, it's time to add the relid(s) of the pushed-down join(s).
816 : * We can skip this if this join was never a candidate to be pushed up.
817 : */
818 80314 : if (sjinfo->commute_above_l)
819 : {
820 15624 : Relids commute_above_rels = bms_copy(sjinfo->commute_above_l);
821 : ListCell *lc;
822 :
823 : /*
824 : * The current join could complete the nulling of more than one
825 : * pushed-down join, so we have to examine all the SpecialJoinInfos.
826 : * Because join_info_list was built in bottom-up order, it's
827 : * sufficient to traverse it once: an ojrelid we add in one loop
828 : * iteration would not have affected decisions of earlier iterations.
829 : */
830 57692 : foreach(lc, root->join_info_list)
831 : {
832 42068 : SpecialJoinInfo *othersj = (SpecialJoinInfo *) lfirst(lc);
833 :
834 42068 : if (othersj == sjinfo ||
835 26444 : othersj->ojrelid == 0 || othersj->jointype != JOIN_LEFT)
836 15636 : continue; /* definitely not interesting */
837 :
838 26432 : if (!bms_is_member(othersj->ojrelid, commute_above_rels))
839 10700 : continue;
840 :
841 : /* Add it if not already present but conditions now satisfied */
842 31464 : if (!bms_is_member(othersj->ojrelid, input_relids) &&
843 31440 : bms_is_subset(othersj->min_lefthand, input_relids) &&
844 23622 : bms_is_subset(othersj->min_righthand, input_relids) &&
845 7914 : bms_is_subset(othersj->commute_below_l, input_relids))
846 : {
847 7878 : input_relids = bms_add_member(input_relids, othersj->ojrelid);
848 : /* report such pushed down outer joins, if asked */
849 7878 : if (pushed_down_joins != NULL)
850 7878 : *pushed_down_joins = lappend(*pushed_down_joins, othersj);
851 :
852 : /*
853 : * We must also check any joins that othersj potentially
854 : * commutes with. They likewise must appear later in
855 : * join_info_list than othersj itself, so we can visit them
856 : * later in this loop.
857 : */
858 7878 : commute_above_rels = bms_add_members(commute_above_rels,
859 7878 : othersj->commute_above_l);
860 : }
861 : }
862 : }
863 :
864 80314 : return input_relids;
865 : }
866 :
867 : /*
868 : * populate_joinrel_with_paths
869 : * Add paths to the given joinrel for given pair of joining relations. The
870 : * SpecialJoinInfo provides details about the join and the restrictlist
871 : * contains the join clauses and the other clauses applicable for given pair
872 : * of the joining relations.
873 : */
874 : static void
875 260028 : populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
876 : RelOptInfo *rel2, RelOptInfo *joinrel,
877 : SpecialJoinInfo *sjinfo, List *restrictlist)
878 : {
879 : /*
880 : * Consider paths using each rel as both outer and inner. Depending on
881 : * the join type, a provably empty outer or inner rel might mean the join
882 : * is provably empty too; in which case throw away any previously computed
883 : * paths and mark the join as dummy. (We do it this way since it's
884 : * conceivable that dummy-ness of a multi-element join might only be
885 : * noticeable for certain construction paths.)
886 : *
887 : * Also, a provably constant-false join restriction typically means that
888 : * we can skip evaluating one or both sides of the join. We do this by
889 : * marking the appropriate rel as dummy. For outer joins, a
890 : * constant-false restriction that is pushed down still means the whole
891 : * join is dummy, while a non-pushed-down one means that no inner rows
892 : * will join so we can treat the inner rel as dummy.
893 : *
894 : * We need only consider the jointypes that appear in join_info_list, plus
895 : * JOIN_INNER.
896 : */
897 260028 : switch (sjinfo->jointype)
898 : {
899 163348 : case JOIN_INNER:
900 326678 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
901 163330 : restriction_is_constant_false(restrictlist, joinrel, false))
902 : {
903 168 : mark_dummy_rel(joinrel);
904 168 : break;
905 : }
906 163180 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
907 : JOIN_INNER, sjinfo,
908 : restrictlist);
909 163180 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
910 : JOIN_INNER, sjinfo,
911 : restrictlist);
912 163180 : break;
913 86362 : case JOIN_LEFT:
914 172670 : if (is_dummy_rel(rel1) ||
915 86308 : restriction_is_constant_false(restrictlist, joinrel, true))
916 : {
917 86 : mark_dummy_rel(joinrel);
918 86 : break;
919 : }
920 86432 : if (restriction_is_constant_false(restrictlist, joinrel, false) &&
921 156 : bms_is_subset(rel2->relids, sjinfo->syn_righthand))
922 132 : mark_dummy_rel(rel2);
923 86276 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
924 : JOIN_LEFT, sjinfo,
925 : restrictlist);
926 86276 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
927 : JOIN_RIGHT, sjinfo,
928 : restrictlist);
929 86276 : break;
930 1600 : case JOIN_FULL:
931 3200 : if ((is_dummy_rel(rel1) && is_dummy_rel(rel2)) ||
932 1600 : restriction_is_constant_false(restrictlist, joinrel, true))
933 : {
934 0 : mark_dummy_rel(joinrel);
935 0 : break;
936 : }
937 1600 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
938 : JOIN_FULL, sjinfo,
939 : restrictlist);
940 1600 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
941 : JOIN_FULL, sjinfo,
942 : restrictlist);
943 :
944 : /*
945 : * If there are join quals that aren't mergeable or hashable, we
946 : * may not be able to build any valid plan. Complain here so that
947 : * we can give a somewhat-useful error message. (Since we have no
948 : * flexibility of planning for a full join, there's no chance of
949 : * succeeding later with another pair of input rels.)
950 : */
951 1600 : if (joinrel->pathlist == NIL)
952 0 : ereport(ERROR,
953 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
954 : errmsg("FULL JOIN is only supported with merge-joinable or hash-joinable join conditions")));
955 1600 : break;
956 4408 : case JOIN_SEMI:
957 :
958 : /*
959 : * We might have a normal semijoin, or a case where we don't have
960 : * enough rels to do the semijoin but can unique-ify the RHS and
961 : * then do an innerjoin (see comments in join_is_legal). In the
962 : * latter case we can't apply JOIN_SEMI joining.
963 : */
964 8530 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
965 4122 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
966 : {
967 8238 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
968 4116 : restriction_is_constant_false(restrictlist, joinrel, false))
969 : {
970 12 : mark_dummy_rel(joinrel);
971 12 : break;
972 : }
973 4110 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
974 : JOIN_SEMI, sjinfo,
975 : restrictlist);
976 : }
977 :
978 : /*
979 : * If we know how to unique-ify the RHS and one input rel is
980 : * exactly the RHS (not a superset) we can consider unique-ifying
981 : * it and then doing a regular join. (The create_unique_path
982 : * check here is probably redundant with what join_is_legal did,
983 : * but if so the check is cheap because it's cached. So test
984 : * anyway to be sure.)
985 : */
986 8792 : if (bms_equal(sjinfo->syn_righthand, rel2->relids) &&
987 4396 : create_unique_path(root, rel2, rel2->cheapest_total_path,
988 : sjinfo) != NULL)
989 : {
990 8600 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
991 4300 : restriction_is_constant_false(restrictlist, joinrel, false))
992 : {
993 0 : mark_dummy_rel(joinrel);
994 0 : break;
995 : }
996 4300 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
997 : JOIN_UNIQUE_INNER, sjinfo,
998 : restrictlist);
999 4300 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1000 : JOIN_UNIQUE_OUTER, sjinfo,
1001 : restrictlist);
1002 : }
1003 4396 : break;
1004 4310 : case JOIN_ANTI:
1005 8620 : if (is_dummy_rel(rel1) ||
1006 4310 : restriction_is_constant_false(restrictlist, joinrel, true))
1007 : {
1008 0 : mark_dummy_rel(joinrel);
1009 0 : break;
1010 : }
1011 4310 : if (restriction_is_constant_false(restrictlist, joinrel, false) &&
1012 0 : bms_is_subset(rel2->relids, sjinfo->syn_righthand))
1013 0 : mark_dummy_rel(rel2);
1014 4310 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1015 : JOIN_ANTI, sjinfo,
1016 : restrictlist);
1017 4310 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1018 : JOIN_RIGHT_ANTI, sjinfo,
1019 : restrictlist);
1020 4310 : break;
1021 0 : default:
1022 : /* other values not expected here */
1023 0 : elog(ERROR, "unrecognized join type: %d", (int) sjinfo->jointype);
1024 : break;
1025 : }
1026 :
1027 : /* Apply partitionwise join technique, if possible. */
1028 260028 : try_partitionwise_join(root, rel1, rel2, joinrel, sjinfo, restrictlist);
1029 260028 : }
1030 :
1031 :
1032 : /*
1033 : * have_join_order_restriction
1034 : * Detect whether the two relations should be joined to satisfy
1035 : * a join-order restriction arising from special or lateral joins.
1036 : *
1037 : * In practice this is always used with have_relevant_joinclause(), and so
1038 : * could be merged with that function, but it seems clearer to separate the
1039 : * two concerns. We need this test because there are degenerate cases where
1040 : * a clauseless join must be performed to satisfy join-order restrictions.
1041 : * Also, if one rel has a lateral reference to the other, or both are needed
1042 : * to compute some PHV, we should consider joining them even if the join would
1043 : * be clauseless.
1044 : *
1045 : * Note: this is only a problem if one side of a degenerate outer join
1046 : * contains multiple rels, or a clauseless join is required within an
1047 : * IN/EXISTS RHS; else we will find a join path via the "last ditch" case in
1048 : * join_search_one_level(). We could dispense with this test if we were
1049 : * willing to try bushy plans in the "last ditch" case, but that seems much
1050 : * less efficient.
1051 : */
1052 : bool
1053 66776 : have_join_order_restriction(PlannerInfo *root,
1054 : RelOptInfo *rel1, RelOptInfo *rel2)
1055 : {
1056 66776 : bool result = false;
1057 : ListCell *l;
1058 :
1059 : /*
1060 : * If either side has a direct lateral reference to the other, attempt the
1061 : * join regardless of outer-join considerations.
1062 : */
1063 124648 : if (bms_overlap(rel1->relids, rel2->direct_lateral_relids) ||
1064 57872 : bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1065 9612 : return true;
1066 :
1067 : /*
1068 : * Likewise, if both rels are needed to compute some PlaceHolderVar,
1069 : * attempt the join regardless of outer-join considerations. (This is not
1070 : * very desirable, because a PHV with a large eval_at set will cause a lot
1071 : * of probably-useless joins to be considered, but failing to do this can
1072 : * cause us to fail to construct a plan at all.)
1073 : */
1074 58744 : foreach(l, root->placeholder_list)
1075 : {
1076 1634 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1077 :
1078 2006 : if (bms_is_subset(rel1->relids, phinfo->ph_eval_at) &&
1079 372 : bms_is_subset(rel2->relids, phinfo->ph_eval_at))
1080 54 : return true;
1081 : }
1082 :
1083 : /*
1084 : * It's possible that the rels correspond to the left and right sides of a
1085 : * degenerate outer join, that is, one with no joinclause mentioning the
1086 : * non-nullable side; in which case we should force the join to occur.
1087 : *
1088 : * Also, the two rels could represent a clauseless join that has to be
1089 : * completed to build up the LHS or RHS of an outer join.
1090 : */
1091 162486 : foreach(l, root->join_info_list)
1092 : {
1093 106430 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1094 :
1095 : /* ignore full joins --- other mechanisms handle them */
1096 106430 : if (sjinfo->jointype == JOIN_FULL)
1097 42 : continue;
1098 :
1099 : /* Can we perform the SJ with these rels? */
1100 133320 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1101 26932 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
1102 : {
1103 802 : result = true;
1104 802 : break;
1105 : }
1106 109868 : if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1107 4282 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
1108 : {
1109 138 : result = true;
1110 138 : break;
1111 : }
1112 :
1113 : /*
1114 : * Might we need to join these rels to complete the RHS? We have to
1115 : * use "overlap" tests since either rel might include a lower SJ that
1116 : * has been proven to commute with this one.
1117 : */
1118 130982 : if (bms_overlap(sjinfo->min_righthand, rel1->relids) &&
1119 25534 : bms_overlap(sjinfo->min_righthand, rel2->relids))
1120 : {
1121 90 : result = true;
1122 90 : break;
1123 : }
1124 :
1125 : /* Likewise for the LHS. */
1126 134780 : if (bms_overlap(sjinfo->min_lefthand, rel1->relids) &&
1127 29422 : bms_overlap(sjinfo->min_lefthand, rel2->relids))
1128 : {
1129 24 : result = true;
1130 24 : break;
1131 : }
1132 : }
1133 :
1134 : /*
1135 : * We do not force the join to occur if either input rel can legally be
1136 : * joined to anything else using joinclauses. This essentially means that
1137 : * clauseless bushy joins are put off as long as possible. The reason is
1138 : * that when there is a join order restriction high up in the join tree
1139 : * (that is, with many rels inside the LHS or RHS), we would otherwise
1140 : * expend lots of effort considering very stupid join combinations within
1141 : * its LHS or RHS.
1142 : */
1143 57110 : if (result)
1144 : {
1145 2048 : if (has_legal_joinclause(root, rel1) ||
1146 994 : has_legal_joinclause(root, rel2))
1147 150 : result = false;
1148 : }
1149 :
1150 57110 : return result;
1151 : }
1152 :
1153 :
1154 : /*
1155 : * has_join_restriction
1156 : * Detect whether the specified relation has join-order restrictions,
1157 : * due to being inside an outer join or an IN (sub-SELECT),
1158 : * or participating in any LATERAL references or multi-rel PHVs.
1159 : *
1160 : * Essentially, this tests whether have_join_order_restriction() could
1161 : * succeed with this rel and some other one. It's OK if we sometimes
1162 : * say "true" incorrectly. (Therefore, we don't bother with the relatively
1163 : * expensive has_legal_joinclause test.)
1164 : */
1165 : static bool
1166 27524 : has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1167 : {
1168 : ListCell *l;
1169 :
1170 27524 : if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1171 16874 : return true;
1172 :
1173 11280 : foreach(l, root->placeholder_list)
1174 : {
1175 672 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1176 :
1177 672 : if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1178 132 : !bms_equal(rel->relids, phinfo->ph_eval_at))
1179 42 : return true;
1180 : }
1181 :
1182 11342 : foreach(l, root->join_info_list)
1183 : {
1184 2326 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1185 :
1186 : /* ignore full joins --- other mechanisms preserve their ordering */
1187 2326 : if (sjinfo->jointype == JOIN_FULL)
1188 86 : continue;
1189 :
1190 : /* ignore if SJ is already contained in rel */
1191 3450 : if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1192 1210 : bms_is_subset(sjinfo->min_righthand, rel->relids))
1193 324 : continue;
1194 :
1195 : /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1196 2922 : if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1197 1006 : bms_overlap(sjinfo->min_righthand, rel->relids))
1198 1592 : return true;
1199 : }
1200 :
1201 9016 : return false;
1202 : }
1203 :
1204 :
1205 : /*
1206 : * has_legal_joinclause
1207 : * Detect whether the specified relation can legally be joined
1208 : * to any other rels using join clauses.
1209 : *
1210 : * We consider only joins to single other relations in the current
1211 : * initial_rels list. This is sufficient to get a "true" result in most real
1212 : * queries, and an occasional erroneous "false" will only cost a bit more
1213 : * planning time. The reason for this limitation is that considering joins to
1214 : * other joins would require proving that the other join rel can legally be
1215 : * formed, which seems like too much trouble for something that's only a
1216 : * heuristic to save planning time. (Note: we must look at initial_rels
1217 : * and not all of the query, since when we are planning a sub-joinlist we
1218 : * may be forced to make clauseless joins within initial_rels even though
1219 : * there are join clauses linking to other parts of the query.)
1220 : */
1221 : static bool
1222 2048 : has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel)
1223 : {
1224 : ListCell *lc;
1225 :
1226 7782 : foreach(lc, root->initial_rels)
1227 : {
1228 5884 : RelOptInfo *rel2 = (RelOptInfo *) lfirst(lc);
1229 :
1230 : /* ignore rels that are already in "rel" */
1231 5884 : if (bms_overlap(rel->relids, rel2->relids))
1232 2450 : continue;
1233 :
1234 3434 : if (have_relevant_joinclause(root, rel, rel2))
1235 : {
1236 : Relids joinrelids;
1237 : SpecialJoinInfo *sjinfo;
1238 : bool reversed;
1239 :
1240 : /* join_is_legal needs relids of the union */
1241 288 : joinrelids = bms_union(rel->relids, rel2->relids);
1242 :
1243 288 : if (join_is_legal(root, rel, rel2, joinrelids,
1244 : &sjinfo, &reversed))
1245 : {
1246 : /* Yes, this will work */
1247 150 : bms_free(joinrelids);
1248 150 : return true;
1249 : }
1250 :
1251 138 : bms_free(joinrelids);
1252 : }
1253 : }
1254 :
1255 1898 : return false;
1256 : }
1257 :
1258 :
1259 : /*
1260 : * There's a pitfall for creating parameterized nestloops: suppose the inner
1261 : * rel (call it A) has a parameter that is a PlaceHolderVar, and that PHV's
1262 : * minimum eval_at set includes the outer rel (B) and some third rel (C).
1263 : * We might think we could create a B/A nestloop join that's parameterized by
1264 : * C. But we would end up with a plan in which the PHV's expression has to be
1265 : * evaluated as a nestloop parameter at the B/A join; and the executor is only
1266 : * set up to handle simple Vars as NestLoopParams. Rather than add complexity
1267 : * and overhead to the executor for such corner cases, it seems better to
1268 : * forbid the join. (Note that we can still make use of A's parameterized
1269 : * path with pre-joined B+C as the outer rel. have_join_order_restriction()
1270 : * ensures that we will consider making such a join even if there are not
1271 : * other reasons to do so.)
1272 : *
1273 : * So we check whether any PHVs used in the query could pose such a hazard.
1274 : * We don't have any simple way of checking whether a risky PHV would actually
1275 : * be used in the inner plan, and the case is so unusual that it doesn't seem
1276 : * worth working very hard on it.
1277 : *
1278 : * This needs to be checked in two places. If the inner rel's minimum
1279 : * parameterization would trigger the restriction, then join_is_legal() should
1280 : * reject the join altogether, because there will be no workable paths for it.
1281 : * But joinpath.c has to check again for every proposed nestloop path, because
1282 : * the inner path might have more than the minimum parameterization, causing
1283 : * some PHV to be dangerous for it that otherwise wouldn't be.
1284 : */
1285 : bool
1286 42178 : have_dangerous_phv(PlannerInfo *root,
1287 : Relids outer_relids, Relids inner_params)
1288 : {
1289 : ListCell *lc;
1290 :
1291 44938 : foreach(lc, root->placeholder_list)
1292 : {
1293 3012 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
1294 :
1295 3012 : if (!bms_is_subset(phinfo->ph_eval_at, inner_params))
1296 2220 : continue; /* ignore, could not be a nestloop param */
1297 792 : if (!bms_overlap(phinfo->ph_eval_at, outer_relids))
1298 192 : continue; /* ignore, not relevant to this join */
1299 600 : if (bms_is_subset(phinfo->ph_eval_at, outer_relids))
1300 348 : continue; /* safe, it can be eval'd within outerrel */
1301 : /* Otherwise, it's potentially unsafe, so reject the join */
1302 252 : return true;
1303 : }
1304 :
1305 : /* OK to perform the join */
1306 41926 : return false;
1307 : }
1308 :
1309 :
1310 : /*
1311 : * is_dummy_rel --- has relation been proven empty?
1312 : */
1313 : bool
1314 2039718 : is_dummy_rel(RelOptInfo *rel)
1315 : {
1316 : Path *path;
1317 :
1318 : /*
1319 : * A rel that is known dummy will have just one path that is a childless
1320 : * Append. (Even if somehow it has more paths, a childless Append will
1321 : * have cost zero and hence should be at the front of the pathlist.)
1322 : */
1323 2039718 : if (rel->pathlist == NIL)
1324 1117692 : return false;
1325 922026 : path = (Path *) linitial(rel->pathlist);
1326 :
1327 : /*
1328 : * Initially, a dummy path will just be a childless Append. But in later
1329 : * planning stages we might stick a ProjectSetPath and/or ProjectionPath
1330 : * on top, since Append can't project. Rather than make assumptions about
1331 : * which combinations can occur, just descend through whatever we find.
1332 : */
1333 : for (;;)
1334 : {
1335 950616 : if (IsA(path, ProjectionPath))
1336 25190 : path = ((ProjectionPath *) path)->subpath;
1337 925426 : else if (IsA(path, ProjectSetPath))
1338 3400 : path = ((ProjectSetPath *) path)->subpath;
1339 : else
1340 922026 : break;
1341 : }
1342 922026 : if (IS_DUMMY_APPEND(path))
1343 3404 : return true;
1344 918622 : return false;
1345 : }
1346 :
1347 : /*
1348 : * Mark a relation as proven empty.
1349 : *
1350 : * During GEQO planning, this can get invoked more than once on the same
1351 : * baserel struct, so it's worth checking to see if the rel is already marked
1352 : * dummy.
1353 : *
1354 : * Also, when called during GEQO join planning, we are in a short-lived
1355 : * memory context. We must make sure that the dummy path attached to a
1356 : * baserel survives the GEQO cycle, else the baserel is trashed for future
1357 : * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1358 : * we don't want the dummy path to clutter the main planning context. Upshot
1359 : * is that the best solution is to explicitly make the dummy path in the same
1360 : * context the given RelOptInfo is in.
1361 : */
1362 : void
1363 476 : mark_dummy_rel(RelOptInfo *rel)
1364 : {
1365 : MemoryContext oldcontext;
1366 :
1367 : /* Already marked? */
1368 476 : if (is_dummy_rel(rel))
1369 12 : return;
1370 :
1371 : /* No, so choose correct context to make the dummy path in */
1372 464 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1373 :
1374 : /* Set dummy size estimate */
1375 464 : rel->rows = 0;
1376 :
1377 : /* Evict any previously chosen paths */
1378 464 : rel->pathlist = NIL;
1379 464 : rel->partial_pathlist = NIL;
1380 :
1381 : /* Set up the dummy path */
1382 464 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
1383 : NIL, rel->lateral_relids,
1384 : 0, false, -1));
1385 :
1386 : /* Set or update cheapest_total_path and related fields */
1387 464 : set_cheapest(rel);
1388 :
1389 464 : MemoryContextSwitchTo(oldcontext);
1390 : }
1391 :
1392 :
1393 : /*
1394 : * restriction_is_constant_false --- is a restrictlist just FALSE?
1395 : *
1396 : * In cases where a qual is provably constant FALSE, eval_const_expressions
1397 : * will generally have thrown away anything that's ANDed with it. In outer
1398 : * join situations this will leave us computing cartesian products only to
1399 : * decide there's no match for an outer row, which is pretty stupid. So,
1400 : * we need to detect the case.
1401 : *
1402 : * If only_pushed_down is true, then consider only quals that are pushed-down
1403 : * from the point of view of the joinrel.
1404 : */
1405 : static bool
1406 354550 : restriction_is_constant_false(List *restrictlist,
1407 : RelOptInfo *joinrel,
1408 : bool only_pushed_down)
1409 : {
1410 : ListCell *lc;
1411 :
1412 : /*
1413 : * Despite the above comment, the restriction list we see here might
1414 : * possibly have other members besides the FALSE constant, since other
1415 : * quals could get "pushed down" to the outer join level. So we check
1416 : * each member of the list.
1417 : */
1418 741442 : foreach(lc, restrictlist)
1419 : {
1420 387236 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1421 :
1422 387236 : if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1423 111922 : continue;
1424 :
1425 275314 : if (rinfo->clause && IsA(rinfo->clause, Const))
1426 : {
1427 4100 : Const *con = (Const *) rinfo->clause;
1428 :
1429 : /* constant NULL is as good as constant FALSE for our purposes */
1430 4100 : if (con->constisnull)
1431 344 : return true;
1432 3992 : if (!DatumGetBool(con->constvalue))
1433 236 : return true;
1434 : }
1435 : }
1436 354206 : return false;
1437 : }
1438 :
1439 : /*
1440 : * Assess whether join between given two partitioned relations can be broken
1441 : * down into joins between matching partitions; a technique called
1442 : * "partitionwise join"
1443 : *
1444 : * Partitionwise join is possible when a. Joining relations have same
1445 : * partitioning scheme b. There exists an equi-join between the partition keys
1446 : * of the two relations.
1447 : *
1448 : * Partitionwise join is planned as follows (details: optimizer/README.)
1449 : *
1450 : * 1. Create the RelOptInfos for joins between matching partitions i.e
1451 : * child-joins and add paths to them.
1452 : *
1453 : * 2. Construct Append or MergeAppend paths across the set of child joins.
1454 : * This second phase is implemented by generate_partitionwise_join_paths().
1455 : *
1456 : * The RelOptInfo, SpecialJoinInfo and restrictlist for each child join are
1457 : * obtained by translating the respective parent join structures.
1458 : */
1459 : static void
1460 260028 : try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1461 : RelOptInfo *joinrel, SpecialJoinInfo *parent_sjinfo,
1462 : List *parent_restrictlist)
1463 : {
1464 260028 : bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1465 260028 : bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1466 260028 : List *parts1 = NIL;
1467 260028 : List *parts2 = NIL;
1468 260028 : ListCell *lcr1 = NULL;
1469 260028 : ListCell *lcr2 = NULL;
1470 : int cnt_parts;
1471 :
1472 : /* Guard against stack overflow due to overly deep partition hierarchy. */
1473 260028 : check_stack_depth();
1474 :
1475 : /* Nothing to do, if the join relation is not partitioned. */
1476 260028 : if (joinrel->part_scheme == NULL || joinrel->nparts == 0)
1477 258178 : return;
1478 :
1479 : /* The join relation should have consider_partitionwise_join set. */
1480 : Assert(joinrel->consider_partitionwise_join);
1481 :
1482 : /*
1483 : * We can not perform partitionwise join if either of the joining
1484 : * relations is not partitioned.
1485 : */
1486 1988 : if (!IS_PARTITIONED_REL(rel1) || !IS_PARTITIONED_REL(rel2))
1487 18 : return;
1488 :
1489 : Assert(REL_HAS_ALL_PART_PROPS(rel1) && REL_HAS_ALL_PART_PROPS(rel2));
1490 :
1491 : /* The joining relations should have consider_partitionwise_join set. */
1492 : Assert(rel1->consider_partitionwise_join &&
1493 : rel2->consider_partitionwise_join);
1494 :
1495 : /*
1496 : * The partition scheme of the join relation should match that of the
1497 : * joining relations.
1498 : */
1499 : Assert(joinrel->part_scheme == rel1->part_scheme &&
1500 : joinrel->part_scheme == rel2->part_scheme);
1501 :
1502 : Assert(!(joinrel->partbounds_merged && (joinrel->nparts <= 0)));
1503 :
1504 1970 : compute_partition_bounds(root, rel1, rel2, joinrel, parent_sjinfo,
1505 : &parts1, &parts2);
1506 :
1507 1970 : if (joinrel->partbounds_merged)
1508 : {
1509 768 : lcr1 = list_head(parts1);
1510 768 : lcr2 = list_head(parts2);
1511 : }
1512 :
1513 : /*
1514 : * Create child-join relations for this partitioned join, if those don't
1515 : * exist. Add paths to child-joins for a pair of child relations
1516 : * corresponding to the given pair of parent relations.
1517 : */
1518 6902 : for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1519 : {
1520 : RelOptInfo *child_rel1;
1521 : RelOptInfo *child_rel2;
1522 : bool rel1_empty;
1523 : bool rel2_empty;
1524 : SpecialJoinInfo *child_sjinfo;
1525 : List *child_restrictlist;
1526 : RelOptInfo *child_joinrel;
1527 : AppendRelInfo **appinfos;
1528 : int nappinfos;
1529 :
1530 5052 : if (joinrel->partbounds_merged)
1531 : {
1532 2010 : child_rel1 = lfirst_node(RelOptInfo, lcr1);
1533 2010 : child_rel2 = lfirst_node(RelOptInfo, lcr2);
1534 2010 : lcr1 = lnext(parts1, lcr1);
1535 2010 : lcr2 = lnext(parts2, lcr2);
1536 : }
1537 : else
1538 : {
1539 3042 : child_rel1 = rel1->part_rels[cnt_parts];
1540 3042 : child_rel2 = rel2->part_rels[cnt_parts];
1541 : }
1542 :
1543 5052 : rel1_empty = (child_rel1 == NULL || IS_DUMMY_REL(child_rel1));
1544 5052 : rel2_empty = (child_rel2 == NULL || IS_DUMMY_REL(child_rel2));
1545 :
1546 : /*
1547 : * Check for cases where we can prove that this segment of the join
1548 : * returns no rows, due to one or both inputs being empty (including
1549 : * inputs that have been pruned away entirely). If so just ignore it.
1550 : * These rules are equivalent to populate_joinrel_with_paths's rules
1551 : * for dummy input relations.
1552 : */
1553 5052 : switch (parent_sjinfo->jointype)
1554 : {
1555 2246 : case JOIN_INNER:
1556 : case JOIN_SEMI:
1557 2246 : if (rel1_empty || rel2_empty)
1558 52 : continue; /* ignore this join segment */
1559 2222 : break;
1560 2084 : case JOIN_LEFT:
1561 : case JOIN_ANTI:
1562 2084 : if (rel1_empty)
1563 28 : continue; /* ignore this join segment */
1564 2056 : break;
1565 722 : case JOIN_FULL:
1566 722 : if (rel1_empty && rel2_empty)
1567 0 : continue; /* ignore this join segment */
1568 722 : break;
1569 0 : default:
1570 : /* other values not expected here */
1571 0 : elog(ERROR, "unrecognized join type: %d",
1572 : (int) parent_sjinfo->jointype);
1573 : break;
1574 : }
1575 :
1576 : /*
1577 : * If a child has been pruned entirely then we can't generate paths
1578 : * for it, so we have to reject partitionwise joining unless we were
1579 : * able to eliminate this partition above.
1580 : */
1581 5000 : if (child_rel1 == NULL || child_rel2 == NULL)
1582 : {
1583 : /*
1584 : * Mark the joinrel as unpartitioned so that later functions treat
1585 : * it correctly.
1586 : */
1587 120 : joinrel->nparts = 0;
1588 120 : return;
1589 : }
1590 :
1591 : /*
1592 : * If a leaf relation has consider_partitionwise_join=false, it means
1593 : * that it's a dummy relation for which we skipped setting up tlist
1594 : * expressions and adding EC members in set_append_rel_size(), so
1595 : * again we have to fail here.
1596 : */
1597 4880 : if (rel1_is_simple && !child_rel1->consider_partitionwise_join)
1598 : {
1599 : Assert(child_rel1->reloptkind == RELOPT_OTHER_MEMBER_REL);
1600 : Assert(IS_DUMMY_REL(child_rel1));
1601 0 : joinrel->nparts = 0;
1602 0 : return;
1603 : }
1604 4880 : if (rel2_is_simple && !child_rel2->consider_partitionwise_join)
1605 : {
1606 : Assert(child_rel2->reloptkind == RELOPT_OTHER_MEMBER_REL);
1607 : Assert(IS_DUMMY_REL(child_rel2));
1608 0 : joinrel->nparts = 0;
1609 0 : return;
1610 : }
1611 :
1612 : /* We should never try to join two overlapping sets of rels. */
1613 : Assert(!bms_overlap(child_rel1->relids, child_rel2->relids));
1614 :
1615 : /*
1616 : * Construct SpecialJoinInfo from parent join relations's
1617 : * SpecialJoinInfo.
1618 : */
1619 4880 : child_sjinfo = build_child_join_sjinfo(root, parent_sjinfo,
1620 : child_rel1->relids,
1621 : child_rel2->relids);
1622 :
1623 : /* Find the AppendRelInfo structures */
1624 4880 : appinfos = find_appinfos_by_relids(root,
1625 4880 : bms_union(child_rel1->relids,
1626 4880 : child_rel2->relids),
1627 : &nappinfos);
1628 :
1629 : /*
1630 : * Construct restrictions applicable to the child join from those
1631 : * applicable to the parent join.
1632 : */
1633 : child_restrictlist =
1634 4880 : (List *) adjust_appendrel_attrs(root,
1635 : (Node *) parent_restrictlist,
1636 : nappinfos, appinfos);
1637 :
1638 : /* Find or construct the child join's RelOptInfo */
1639 4880 : child_joinrel = joinrel->part_rels[cnt_parts];
1640 4880 : if (!child_joinrel)
1641 : {
1642 4408 : child_joinrel = build_child_join_rel(root, child_rel1, child_rel2,
1643 : joinrel, child_restrictlist,
1644 : child_sjinfo);
1645 4408 : joinrel->part_rels[cnt_parts] = child_joinrel;
1646 4408 : joinrel->live_parts = bms_add_member(joinrel->live_parts, cnt_parts);
1647 4408 : joinrel->all_partrels = bms_add_members(joinrel->all_partrels,
1648 4408 : child_joinrel->relids);
1649 : }
1650 :
1651 : /* Assert we got the right one */
1652 : Assert(bms_equal(child_joinrel->relids,
1653 : adjust_child_relids(joinrel->relids,
1654 : nappinfos, appinfos)));
1655 :
1656 : /* And make paths for the child join */
1657 4880 : populate_joinrel_with_paths(root, child_rel1, child_rel2,
1658 : child_joinrel, child_sjinfo,
1659 : child_restrictlist);
1660 :
1661 4880 : pfree(appinfos);
1662 : }
1663 : }
1664 :
1665 : /*
1666 : * Construct the SpecialJoinInfo for a child-join by translating
1667 : * SpecialJoinInfo for the join between parents. left_relids and right_relids
1668 : * are the relids of left and right side of the join respectively.
1669 : */
1670 : static SpecialJoinInfo *
1671 4880 : build_child_join_sjinfo(PlannerInfo *root, SpecialJoinInfo *parent_sjinfo,
1672 : Relids left_relids, Relids right_relids)
1673 : {
1674 4880 : SpecialJoinInfo *sjinfo = makeNode(SpecialJoinInfo);
1675 : AppendRelInfo **left_appinfos;
1676 : int left_nappinfos;
1677 : AppendRelInfo **right_appinfos;
1678 : int right_nappinfos;
1679 :
1680 4880 : memcpy(sjinfo, parent_sjinfo, sizeof(SpecialJoinInfo));
1681 4880 : left_appinfos = find_appinfos_by_relids(root, left_relids,
1682 : &left_nappinfos);
1683 4880 : right_appinfos = find_appinfos_by_relids(root, right_relids,
1684 : &right_nappinfos);
1685 :
1686 4880 : sjinfo->min_lefthand = adjust_child_relids(sjinfo->min_lefthand,
1687 : left_nappinfos, left_appinfos);
1688 4880 : sjinfo->min_righthand = adjust_child_relids(sjinfo->min_righthand,
1689 : right_nappinfos,
1690 : right_appinfos);
1691 4880 : sjinfo->syn_lefthand = adjust_child_relids(sjinfo->syn_lefthand,
1692 : left_nappinfos, left_appinfos);
1693 4880 : sjinfo->syn_righthand = adjust_child_relids(sjinfo->syn_righthand,
1694 : right_nappinfos,
1695 : right_appinfos);
1696 : /* outer-join relids need no adjustment */
1697 9760 : sjinfo->semi_rhs_exprs = (List *) adjust_appendrel_attrs(root,
1698 4880 : (Node *) sjinfo->semi_rhs_exprs,
1699 : right_nappinfos,
1700 : right_appinfos);
1701 :
1702 4880 : pfree(left_appinfos);
1703 4880 : pfree(right_appinfos);
1704 :
1705 4880 : return sjinfo;
1706 : }
1707 :
1708 : /*
1709 : * compute_partition_bounds
1710 : * Compute the partition bounds for a join rel from those for inputs
1711 : */
1712 : static void
1713 1970 : compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
1714 : RelOptInfo *rel2, RelOptInfo *joinrel,
1715 : SpecialJoinInfo *parent_sjinfo,
1716 : List **parts1, List **parts2)
1717 : {
1718 : /*
1719 : * If we don't have the partition bounds for the join rel yet, try to
1720 : * compute those along with pairs of partitions to be joined.
1721 : */
1722 1970 : if (joinrel->nparts == -1)
1723 : {
1724 1806 : PartitionScheme part_scheme = joinrel->part_scheme;
1725 1806 : PartitionBoundInfo boundinfo = NULL;
1726 1806 : int nparts = 0;
1727 :
1728 : Assert(joinrel->boundinfo == NULL);
1729 : Assert(joinrel->part_rels == NULL);
1730 :
1731 : /*
1732 : * See if the partition bounds for inputs are exactly the same, in
1733 : * which case we don't need to work hard: the join rel will have the
1734 : * same partition bounds as inputs, and the partitions with the same
1735 : * cardinal positions will form the pairs.
1736 : *
1737 : * Note: even in cases where one or both inputs have merged bounds, it
1738 : * would be possible for both the bounds to be exactly the same, but
1739 : * it seems unlikely to be worth the cycles to check.
1740 : */
1741 1806 : if (!rel1->partbounds_merged &&
1742 1746 : !rel2->partbounds_merged &&
1743 3234 : rel1->nparts == rel2->nparts &&
1744 1488 : partition_bounds_equal(part_scheme->partnatts,
1745 : part_scheme->parttyplen,
1746 : part_scheme->parttypbyval,
1747 : rel1->boundinfo, rel2->boundinfo))
1748 : {
1749 960 : boundinfo = rel1->boundinfo;
1750 960 : nparts = rel1->nparts;
1751 : }
1752 : else
1753 : {
1754 : /* Try merging the partition bounds for inputs. */
1755 846 : boundinfo = partition_bounds_merge(part_scheme->partnatts,
1756 846 : part_scheme->partsupfunc,
1757 : part_scheme->partcollation,
1758 : rel1, rel2,
1759 : parent_sjinfo->jointype,
1760 : parts1, parts2);
1761 846 : if (boundinfo == NULL)
1762 : {
1763 114 : joinrel->nparts = 0;
1764 114 : return;
1765 : }
1766 732 : nparts = list_length(*parts1);
1767 732 : joinrel->partbounds_merged = true;
1768 : }
1769 :
1770 : Assert(nparts > 0);
1771 1692 : joinrel->boundinfo = boundinfo;
1772 1692 : joinrel->nparts = nparts;
1773 1692 : joinrel->part_rels =
1774 1692 : (RelOptInfo **) palloc0(sizeof(RelOptInfo *) * nparts);
1775 : }
1776 : else
1777 : {
1778 : Assert(joinrel->nparts > 0);
1779 : Assert(joinrel->boundinfo);
1780 : Assert(joinrel->part_rels);
1781 :
1782 : /*
1783 : * If the join rel's partbounds_merged flag is true, it means inputs
1784 : * are not guaranteed to have the same partition bounds, therefore we
1785 : * can't assume that the partitions at the same cardinal positions
1786 : * form the pairs; let get_matching_part_pairs() generate the pairs.
1787 : * Otherwise, nothing to do since we can assume that.
1788 : */
1789 164 : if (joinrel->partbounds_merged)
1790 : {
1791 36 : get_matching_part_pairs(root, joinrel, rel1, rel2,
1792 : parts1, parts2);
1793 : Assert(list_length(*parts1) == joinrel->nparts);
1794 : Assert(list_length(*parts2) == joinrel->nparts);
1795 : }
1796 : }
1797 : }
1798 :
1799 : /*
1800 : * get_matching_part_pairs
1801 : * Generate pairs of partitions to be joined from inputs
1802 : */
1803 : static void
1804 36 : get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
1805 : RelOptInfo *rel1, RelOptInfo *rel2,
1806 : List **parts1, List **parts2)
1807 : {
1808 36 : bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1809 36 : bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1810 : int cnt_parts;
1811 :
1812 36 : *parts1 = NIL;
1813 36 : *parts2 = NIL;
1814 :
1815 132 : for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1816 : {
1817 96 : RelOptInfo *child_joinrel = joinrel->part_rels[cnt_parts];
1818 : RelOptInfo *child_rel1;
1819 : RelOptInfo *child_rel2;
1820 : Relids child_relids1;
1821 : Relids child_relids2;
1822 :
1823 : /*
1824 : * If this segment of the join is empty, it means that this segment
1825 : * was ignored when previously creating child-join paths for it in
1826 : * try_partitionwise_join() as it would not contribute to the join
1827 : * result, due to one or both inputs being empty; add NULL to each of
1828 : * the given lists so that this segment will be ignored again in that
1829 : * function.
1830 : */
1831 96 : if (!child_joinrel)
1832 : {
1833 0 : *parts1 = lappend(*parts1, NULL);
1834 0 : *parts2 = lappend(*parts2, NULL);
1835 0 : continue;
1836 : }
1837 :
1838 : /*
1839 : * Get a relids set of partition(s) involved in this join segment that
1840 : * are from the rel1 side.
1841 : */
1842 96 : child_relids1 = bms_intersect(child_joinrel->relids,
1843 96 : rel1->all_partrels);
1844 : Assert(bms_num_members(child_relids1) == bms_num_members(rel1->relids));
1845 :
1846 : /*
1847 : * Get a child rel for rel1 with the relids. Note that we should have
1848 : * the child rel even if rel1 is a join rel, because in that case the
1849 : * partitions specified in the relids would have matching/overlapping
1850 : * boundaries, so the specified partitions should be considered as
1851 : * ones to be joined when planning partitionwise joins of rel1,
1852 : * meaning that the child rel would have been built by the time we get
1853 : * here.
1854 : */
1855 96 : if (rel1_is_simple)
1856 : {
1857 0 : int varno = bms_singleton_member(child_relids1);
1858 :
1859 0 : child_rel1 = find_base_rel(root, varno);
1860 : }
1861 : else
1862 96 : child_rel1 = find_join_rel(root, child_relids1);
1863 : Assert(child_rel1);
1864 :
1865 : /*
1866 : * Get a relids set of partition(s) involved in this join segment that
1867 : * are from the rel2 side.
1868 : */
1869 96 : child_relids2 = bms_intersect(child_joinrel->relids,
1870 96 : rel2->all_partrels);
1871 : Assert(bms_num_members(child_relids2) == bms_num_members(rel2->relids));
1872 :
1873 : /*
1874 : * Get a child rel for rel2 with the relids. See above comments.
1875 : */
1876 96 : if (rel2_is_simple)
1877 : {
1878 96 : int varno = bms_singleton_member(child_relids2);
1879 :
1880 96 : child_rel2 = find_base_rel(root, varno);
1881 : }
1882 : else
1883 0 : child_rel2 = find_join_rel(root, child_relids2);
1884 : Assert(child_rel2);
1885 :
1886 : /*
1887 : * The join of rel1 and rel2 is legal, so is the join of the child
1888 : * rels obtained above; add them to the given lists as a join pair
1889 : * producing this join segment.
1890 : */
1891 96 : *parts1 = lappend(*parts1, child_rel1);
1892 96 : *parts2 = lappend(*parts2, child_rel2);
1893 : }
1894 36 : }
|