Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * joinrels.c
4 : * Routines to determine which relations should be joined
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/optimizer/path/joinrels.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include "miscadmin.h"
18 : #include "optimizer/appendinfo.h"
19 : #include "optimizer/cost.h"
20 : #include "optimizer/joininfo.h"
21 : #include "optimizer/pathnode.h"
22 : #include "optimizer/paths.h"
23 : #include "optimizer/planner.h"
24 : #include "partitioning/partbounds.h"
25 : #include "utils/memutils.h"
26 :
27 :
28 : static void make_rels_by_clause_joins(PlannerInfo *root,
29 : RelOptInfo *old_rel,
30 : List *other_rels,
31 : int first_rel_idx);
32 : static void make_rels_by_clauseless_joins(PlannerInfo *root,
33 : RelOptInfo *old_rel,
34 : List *other_rels);
35 : static bool has_join_restriction(PlannerInfo *root, RelOptInfo *rel);
36 : static bool has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel);
37 : static bool restriction_is_constant_false(List *restrictlist,
38 : RelOptInfo *joinrel,
39 : bool only_pushed_down);
40 : static void make_grouped_join_rel(PlannerInfo *root, RelOptInfo *rel1,
41 : RelOptInfo *rel2, RelOptInfo *joinrel,
42 : SpecialJoinInfo *sjinfo, List *restrictlist);
43 : static void populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
44 : RelOptInfo *rel2, RelOptInfo *joinrel,
45 : SpecialJoinInfo *sjinfo, List *restrictlist);
46 : static void try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1,
47 : RelOptInfo *rel2, RelOptInfo *joinrel,
48 : SpecialJoinInfo *parent_sjinfo,
49 : List *parent_restrictlist);
50 : static SpecialJoinInfo *build_child_join_sjinfo(PlannerInfo *root,
51 : SpecialJoinInfo *parent_sjinfo,
52 : Relids left_relids, Relids right_relids);
53 : static void free_child_join_sjinfo(SpecialJoinInfo *child_sjinfo,
54 : SpecialJoinInfo *parent_sjinfo);
55 : static void compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
56 : RelOptInfo *rel2, RelOptInfo *joinrel,
57 : SpecialJoinInfo *parent_sjinfo,
58 : List **parts1, List **parts2);
59 : static void get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
60 : RelOptInfo *rel1, RelOptInfo *rel2,
61 : List **parts1, List **parts2);
62 :
63 :
64 : /*
65 : * join_search_one_level
66 : * Consider ways to produce join relations containing exactly 'level'
67 : * jointree items. (This is one step of the dynamic-programming method
68 : * embodied in standard_join_search.) Join rel nodes for each feasible
69 : * combination of lower-level rels are created and returned in a list.
70 : * Implementation paths are created for each such joinrel, too.
71 : *
72 : * level: level of rels we want to make this time
73 : * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
74 : *
75 : * The result is returned in root->join_rel_level[level].
76 : */
77 : void
78 136722 : join_search_one_level(PlannerInfo *root, int level)
79 : {
80 136722 : List **joinrels = root->join_rel_level;
81 : ListCell *r;
82 : int k;
83 :
84 : Assert(joinrels[level] == NIL);
85 :
86 : /* Set join_cur_level so that new joinrels are added to proper list */
87 136722 : root->join_cur_level = level;
88 :
89 : /*
90 : * First, consider left-sided and right-sided plans, in which rels of
91 : * exactly level-1 member relations are joined against initial relations.
92 : * We prefer to join using join clauses, but if we find a rel of level-1
93 : * members that has no join clauses, we will generate Cartesian-product
94 : * joins against all initial rels not already contained in it.
95 : */
96 482206 : foreach(r, joinrels[level - 1])
97 : {
98 345484 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
99 :
100 375850 : if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
101 30366 : has_join_restriction(root, old_rel))
102 334796 : {
103 : int first_rel;
104 :
105 : /*
106 : * There are join clauses or join order restrictions relevant to
107 : * this rel, so consider joins between this rel and (only) those
108 : * initial rels it is linked to by a clause or restriction.
109 : *
110 : * At level 2 this condition is symmetric, so there is no need to
111 : * look at initial rels before this one in the list; we already
112 : * considered such joins when we were at the earlier rel. (The
113 : * mirror-image joins are handled automatically by make_join_rel.)
114 : * In later passes (level > 2), we join rels of the previous level
115 : * to each initial rel they don't already include but have a join
116 : * clause or restriction with.
117 : */
118 334796 : if (level == 2) /* consider remaining initial rels */
119 226852 : first_rel = foreach_current_index(r) + 1;
120 : else
121 107944 : first_rel = 0;
122 :
123 334796 : make_rels_by_clause_joins(root, old_rel, joinrels[1], first_rel);
124 : }
125 : else
126 : {
127 : /*
128 : * Oops, we have a relation that is not joined to any other
129 : * relation, either directly or by join-order restrictions.
130 : * Cartesian product time.
131 : *
132 : * We consider a cartesian product with each not-already-included
133 : * initial rel, whether it has other join clauses or not. At
134 : * level 2, if there are two or more clauseless initial rels, we
135 : * will redundantly consider joining them in both directions; but
136 : * such cases aren't common enough to justify adding complexity to
137 : * avoid the duplicated effort.
138 : */
139 10688 : make_rels_by_clauseless_joins(root,
140 : old_rel,
141 10688 : joinrels[1]);
142 : }
143 : }
144 :
145 : /*
146 : * Now, consider "bushy plans" in which relations of k initial rels are
147 : * joined to relations of level-k initial rels, for 2 <= k <= level-2.
148 : *
149 : * We only consider bushy-plan joins for pairs of rels where there is a
150 : * suitable join clause (or join order restriction), in order to avoid
151 : * unreasonable growth of planning time.
152 : */
153 136722 : for (k = 2;; k++)
154 12920 : {
155 149642 : int other_level = level - k;
156 :
157 : /*
158 : * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
159 : * need to go as far as the halfway point.
160 : */
161 149642 : if (k > other_level)
162 136722 : break;
163 :
164 67038 : foreach(r, joinrels[k])
165 : {
166 54118 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
167 : int first_rel;
168 : ListCell *r2;
169 :
170 : /*
171 : * We can ignore relations without join clauses here, unless they
172 : * participate in join-order restrictions --- then we might have
173 : * to force a bushy join plan.
174 : */
175 54118 : if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
176 432 : !has_join_restriction(root, old_rel))
177 300 : continue;
178 :
179 53818 : if (k == other_level) /* only consider remaining rels */
180 37520 : first_rel = foreach_current_index(r) + 1;
181 : else
182 16298 : first_rel = 0;
183 :
184 233618 : for_each_from(r2, joinrels[other_level], first_rel)
185 : {
186 179800 : RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
187 :
188 179800 : if (!bms_overlap(old_rel->relids, new_rel->relids))
189 : {
190 : /*
191 : * OK, we can build a rel of the right level from this
192 : * pair of rels. Do so if there is at least one relevant
193 : * join clause or join order restriction.
194 : */
195 22098 : if (have_relevant_joinclause(root, old_rel, new_rel) ||
196 1154 : have_join_order_restriction(root, old_rel, new_rel))
197 : {
198 19856 : (void) make_join_rel(root, old_rel, new_rel);
199 : }
200 : }
201 : }
202 : }
203 : }
204 :
205 : /*----------
206 : * Last-ditch effort: if we failed to find any usable joins so far, force
207 : * a set of cartesian-product joins to be generated. This handles the
208 : * special case where all the available rels have join clauses but we
209 : * cannot use any of those clauses yet. This can only happen when we are
210 : * considering a join sub-problem (a sub-joinlist) and all the rels in the
211 : * sub-problem have only join clauses with rels outside the sub-problem.
212 : * An example is
213 : *
214 : * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
215 : * WHERE a.w = c.x and b.y = d.z;
216 : *
217 : * If the "a INNER JOIN b" sub-problem does not get flattened into the
218 : * upper level, we must be willing to make a cartesian join of a and b;
219 : * but the code above will not have done so, because it thought that both
220 : * a and b have joinclauses. We consider only left-sided and right-sided
221 : * cartesian joins in this case (no bushy).
222 : *----------
223 : */
224 136722 : if (joinrels[level] == NIL)
225 : {
226 : /*
227 : * This loop is just like the first one, except we always call
228 : * make_rels_by_clauseless_joins().
229 : */
230 54 : foreach(r, joinrels[level - 1])
231 : {
232 36 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
233 :
234 36 : make_rels_by_clauseless_joins(root,
235 : old_rel,
236 36 : joinrels[1]);
237 : }
238 :
239 : /*----------
240 : * When special joins are involved, there may be no legal way
241 : * to make an N-way join for some values of N. For example consider
242 : *
243 : * SELECT ... FROM t1 WHERE
244 : * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
245 : * y IN (SELECT ... FROM t4,t5 WHERE ...)
246 : *
247 : * We will flatten this query to a 5-way join problem, but there are
248 : * no 4-way joins that join_is_legal() will consider legal. We have
249 : * to accept failure at level 4 and go on to discover a workable
250 : * bushy plan at level 5.
251 : *
252 : * However, if there are no special joins and no lateral references
253 : * then join_is_legal() should never fail, and so the following sanity
254 : * check is useful.
255 : *----------
256 : */
257 18 : if (joinrels[level] == NIL &&
258 6 : root->join_info_list == NIL &&
259 0 : !root->hasLateralRTEs)
260 0 : elog(ERROR, "failed to build any %d-way joins", level);
261 : }
262 136722 : }
263 :
264 : /*
265 : * make_rels_by_clause_joins
266 : * Build joins between the given relation 'old_rel' and other relations
267 : * that participate in join clauses that 'old_rel' also participates in
268 : * (or participate in join-order restrictions with it).
269 : * The join rels are returned in root->join_rel_level[join_cur_level].
270 : *
271 : * Note: at levels above 2 we will generate the same joined relation in
272 : * multiple ways --- for example (a join b) join c is the same RelOptInfo as
273 : * (b join c) join a, though the second case will add a different set of Paths
274 : * to it. This is the reason for using the join_rel_level mechanism, which
275 : * automatically ensures that each new joinrel is only added to the list once.
276 : *
277 : * 'old_rel' is the relation entry for the relation to be joined
278 : * 'other_rels': a list containing the other rels to be considered for joining
279 : * 'first_rel_idx': the first rel to be considered in 'other_rels'
280 : *
281 : * Currently, this is only used with initial rels in other_rels, but it
282 : * will work for joining to joinrels too.
283 : */
284 : static void
285 334796 : make_rels_by_clause_joins(PlannerInfo *root,
286 : RelOptInfo *old_rel,
287 : List *other_rels,
288 : int first_rel_idx)
289 : {
290 : ListCell *l;
291 :
292 975906 : for_each_from(l, other_rels, first_rel_idx)
293 : {
294 641110 : RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
295 :
296 1006622 : if (!bms_overlap(old_rel->relids, other_rel->relids) &&
297 441702 : (have_relevant_joinclause(root, old_rel, other_rel) ||
298 76190 : have_join_order_restriction(root, old_rel, other_rel)))
299 : {
300 301036 : (void) make_join_rel(root, old_rel, other_rel);
301 : }
302 : }
303 334796 : }
304 :
305 : /*
306 : * make_rels_by_clauseless_joins
307 : * Given a relation 'old_rel' and a list of other relations
308 : * 'other_rels', create a join relation between 'old_rel' and each
309 : * member of 'other_rels' that isn't already included in 'old_rel'.
310 : * The join rels are returned in root->join_rel_level[join_cur_level].
311 : *
312 : * 'old_rel' is the relation entry for the relation to be joined
313 : * 'other_rels': a list containing the other rels to be considered for joining
314 : *
315 : * Currently, this is only used with initial rels in other_rels, but it would
316 : * work for joining to joinrels too.
317 : */
318 : static void
319 10724 : make_rels_by_clauseless_joins(PlannerInfo *root,
320 : RelOptInfo *old_rel,
321 : List *other_rels)
322 : {
323 : ListCell *l;
324 :
325 34290 : foreach(l, other_rels)
326 : {
327 23566 : RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
328 :
329 23566 : if (!bms_overlap(other_rel->relids, old_rel->relids))
330 : {
331 11626 : (void) make_join_rel(root, old_rel, other_rel);
332 : }
333 : }
334 10724 : }
335 :
336 :
337 : /*
338 : * join_is_legal
339 : * Determine whether a proposed join is legal given the query's
340 : * join order constraints; and if it is, determine the join type.
341 : *
342 : * Caller must supply not only the two rels, but the union of their relids.
343 : * (We could simplify the API by computing joinrelids locally, but this
344 : * would be redundant work in the normal path through make_join_rel.
345 : * Note that this value does NOT include the RT index of any outer join that
346 : * might need to be performed here, so it's not the canonical identifier
347 : * of the join relation.)
348 : *
349 : * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
350 : * else it's set to point to the associated SpecialJoinInfo node. Also,
351 : * *reversed_p is set true if the given relations need to be swapped to
352 : * match the SpecialJoinInfo node.
353 : */
354 : static bool
355 341362 : join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
356 : Relids joinrelids,
357 : SpecialJoinInfo **sjinfo_p, bool *reversed_p)
358 : {
359 : SpecialJoinInfo *match_sjinfo;
360 : bool reversed;
361 : bool unique_ified;
362 : bool must_be_leftjoin;
363 : ListCell *l;
364 :
365 : /*
366 : * Ensure output params are set on failure return. This is just to
367 : * suppress uninitialized-variable warnings from overly anal compilers.
368 : */
369 341362 : *sjinfo_p = NULL;
370 341362 : *reversed_p = false;
371 :
372 : /*
373 : * If we have any special joins, the proposed join might be illegal; and
374 : * in any case we have to determine its join type. Scan the join info
375 : * list for matches and conflicts.
376 : */
377 341362 : match_sjinfo = NULL;
378 341362 : reversed = false;
379 341362 : unique_ified = false;
380 341362 : must_be_leftjoin = false;
381 :
382 695630 : foreach(l, root->join_info_list)
383 : {
384 363708 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
385 :
386 : /*
387 : * This special join is not relevant unless its RHS overlaps the
388 : * proposed join. (Check this first as a fast path for dismissing
389 : * most irrelevant SJs quickly.)
390 : */
391 363708 : if (!bms_overlap(sjinfo->min_righthand, joinrelids))
392 121676 : continue;
393 :
394 : /*
395 : * Also, not relevant if proposed join is fully contained within RHS
396 : * (ie, we're still building up the RHS).
397 : */
398 242032 : if (bms_is_subset(joinrelids, sjinfo->min_righthand))
399 5344 : continue;
400 :
401 : /*
402 : * Also, not relevant if SJ is already done within either input.
403 : */
404 442018 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
405 205330 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
406 100376 : continue;
407 156902 : if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
408 20590 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
409 10812 : continue;
410 :
411 : /*
412 : * If it's a semijoin and we already joined the RHS to any other rels
413 : * within either input, then we must have unique-ified the RHS at that
414 : * point (see below). Therefore the semijoin is no longer relevant in
415 : * this join path.
416 : */
417 125500 : if (sjinfo->jointype == JOIN_SEMI)
418 : {
419 10966 : if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
420 1576 : !bms_equal(sjinfo->syn_righthand, rel1->relids))
421 690 : continue;
422 10276 : if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
423 7224 : !bms_equal(sjinfo->syn_righthand, rel2->relids))
424 214 : continue;
425 : }
426 :
427 : /*
428 : * If one input contains min_lefthand and the other contains
429 : * min_righthand, then we can perform the SJ at this join.
430 : *
431 : * Reject if we get matches to more than one SJ; that implies we're
432 : * considering something that's not really valid.
433 : */
434 229398 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
435 104802 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
436 : {
437 98114 : if (match_sjinfo)
438 9440 : return false; /* invalid join path */
439 98114 : match_sjinfo = sjinfo;
440 98114 : reversed = false;
441 : }
442 35922 : else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
443 9440 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
444 : {
445 8216 : if (match_sjinfo)
446 0 : return false; /* invalid join path */
447 8216 : match_sjinfo = sjinfo;
448 8216 : reversed = true;
449 : }
450 21276 : else if (sjinfo->jointype == JOIN_SEMI &&
451 3618 : bms_equal(sjinfo->syn_righthand, rel2->relids) &&
452 608 : create_unique_paths(root, rel2, sjinfo) != NULL)
453 : {
454 : /*----------
455 : * For a semijoin, we can join the RHS to anything else by
456 : * unique-ifying the RHS (if the RHS can be unique-ified).
457 : * We will only get here if we have the full RHS but less
458 : * than min_lefthand on the LHS.
459 : *
460 : * The reason to consider such a join path is exemplified by
461 : * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
462 : * If we insist on doing this as a semijoin we will first have
463 : * to form the cartesian product of A*B. But if we unique-ify
464 : * C then the semijoin becomes a plain innerjoin and we can join
465 : * in any order, eg C to A and then to B. When C is much smaller
466 : * than A and B this can be a huge win. So we allow C to be
467 : * joined to just A or just B here, and then make_join_rel has
468 : * to handle the case properly.
469 : *
470 : * Note that actually we'll allow unique-ified C to be joined to
471 : * some other relation D here, too. That is legal, if usually not
472 : * very sane, and this routine is only concerned with legality not
473 : * with whether the join is good strategy.
474 : *----------
475 : */
476 392 : if (match_sjinfo)
477 6 : return false; /* invalid join path */
478 386 : match_sjinfo = sjinfo;
479 386 : reversed = false;
480 386 : unique_ified = true;
481 : }
482 20492 : else if (sjinfo->jointype == JOIN_SEMI &&
483 2854 : bms_equal(sjinfo->syn_righthand, rel1->relids) &&
484 236 : create_unique_paths(root, rel1, sjinfo) != NULL)
485 : {
486 : /* Reversed semijoin case */
487 104 : if (match_sjinfo)
488 0 : return false; /* invalid join path */
489 104 : match_sjinfo = sjinfo;
490 104 : reversed = true;
491 104 : unique_ified = true;
492 : }
493 : else
494 : {
495 : /*
496 : * Otherwise, the proposed join overlaps the RHS but isn't a valid
497 : * implementation of this SJ. But don't panic quite yet: the RHS
498 : * violation might have occurred previously, in one or both input
499 : * relations, in which case we must have previously decided that
500 : * it was OK to commute some other SJ with this one. If we need
501 : * to perform this join to finish building up the RHS, rejecting
502 : * it could lead to not finding any plan at all. (This can occur
503 : * because of the heuristics elsewhere in this file that postpone
504 : * clauseless joins: we might not consider doing a clauseless join
505 : * within the RHS until after we've performed other, validly
506 : * commutable SJs with one or both sides of the clauseless join.)
507 : * This consideration boils down to the rule that if both inputs
508 : * overlap the RHS, we can allow the join --- they are either
509 : * fully within the RHS, or represent previously-allowed joins to
510 : * rels outside it.
511 : */
512 26014 : if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
513 8244 : bms_overlap(rel2->relids, sjinfo->min_righthand))
514 174 : continue; /* assume valid previous violation of RHS */
515 :
516 : /*
517 : * The proposed join could still be legal, but only if we're
518 : * allowed to associate it into the RHS of this SJ. That means
519 : * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
520 : * not FULL) and the proposed join must not overlap the LHS.
521 : */
522 32546 : if (sjinfo->jointype != JOIN_LEFT ||
523 14950 : bms_overlap(joinrelids, sjinfo->min_lefthand))
524 9434 : return false; /* invalid join path */
525 :
526 : /*
527 : * To be valid, the proposed join must be a LEFT join; otherwise
528 : * it can't associate into this SJ's RHS. But we may not yet have
529 : * found the SpecialJoinInfo matching the proposed join, so we
530 : * can't test that yet. Remember the requirement for later.
531 : */
532 8162 : must_be_leftjoin = true;
533 : }
534 : }
535 :
536 : /*
537 : * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
538 : * proposed join can't associate into an SJ's RHS.
539 : *
540 : * Also, fail if the proposed join's predicate isn't strict; we're
541 : * essentially checking to see if we can apply outer-join identity 3, and
542 : * that's a requirement. (This check may be redundant with checks in
543 : * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
544 : */
545 331922 : if (must_be_leftjoin &&
546 5190 : (match_sjinfo == NULL ||
547 5190 : match_sjinfo->jointype != JOIN_LEFT ||
548 5190 : !match_sjinfo->lhs_strict))
549 1498 : return false; /* invalid join path */
550 :
551 : /*
552 : * We also have to check for constraints imposed by LATERAL references.
553 : */
554 330424 : if (root->hasLateralRTEs)
555 : {
556 : bool lateral_fwd;
557 : bool lateral_rev;
558 : Relids join_lateral_rels;
559 :
560 : /*
561 : * The proposed rels could each contain lateral references to the
562 : * other, in which case the join is impossible. If there are lateral
563 : * references in just one direction, then the join has to be done with
564 : * a nestloop with the lateral referencer on the inside. If the join
565 : * matches an SJ that cannot be implemented by such a nestloop, the
566 : * join is impossible.
567 : *
568 : * Also, if the lateral reference is only indirect, we should reject
569 : * the join; whatever rel(s) the reference chain goes through must be
570 : * joined to first.
571 : */
572 17650 : lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
573 17650 : lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
574 17650 : if (lateral_fwd && lateral_rev)
575 30 : return false; /* have lateral refs in both directions */
576 17620 : if (lateral_fwd)
577 : {
578 : /* has to be implemented as nestloop with rel1 on left */
579 10232 : if (match_sjinfo &&
580 426 : (reversed ||
581 408 : unique_ified ||
582 408 : match_sjinfo->jointype == JOIN_FULL))
583 18 : return false; /* not implementable as nestloop */
584 : /* check there is a direct reference from rel2 to rel1 */
585 10214 : if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
586 42 : return false; /* only indirect refs, so reject */
587 : }
588 7388 : else if (lateral_rev)
589 : {
590 : /* has to be implemented as nestloop with rel2 on left */
591 1660 : if (match_sjinfo &&
592 78 : (!reversed ||
593 78 : unique_ified ||
594 78 : match_sjinfo->jointype == JOIN_FULL))
595 0 : return false; /* not implementable as nestloop */
596 : /* check there is a direct reference from rel1 to rel2 */
597 1660 : if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
598 0 : return false; /* only indirect refs, so reject */
599 : }
600 :
601 : /*
602 : * LATERAL references could also cause problems later on if we accept
603 : * this join: if the join's minimum parameterization includes any rels
604 : * that would have to be on the inside of an outer join with this join
605 : * rel, then it's never going to be possible to build the complete
606 : * query using this join. We should reject this join not only because
607 : * it'll save work, but because if we don't, the clauseless-join
608 : * heuristics might think that legality of this join means that some
609 : * other join rel need not be formed, and that could lead to failure
610 : * to find any plan at all. We have to consider not only rels that
611 : * are directly on the inner side of an OJ with the joinrel, but also
612 : * ones that are indirectly so, so search to find all such rels.
613 : */
614 17560 : join_lateral_rels = min_join_parameterization(root, joinrelids,
615 : rel1, rel2);
616 17560 : if (join_lateral_rels)
617 : {
618 2390 : Relids join_plus_rhs = bms_copy(joinrelids);
619 : bool more;
620 :
621 : do
622 : {
623 2786 : more = false;
624 4712 : foreach(l, root->join_info_list)
625 : {
626 1926 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
627 :
628 : /* ignore full joins --- their ordering is predetermined */
629 1926 : if (sjinfo->jointype == JOIN_FULL)
630 18 : continue;
631 :
632 1908 : if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
633 1602 : !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
634 : {
635 546 : join_plus_rhs = bms_add_members(join_plus_rhs,
636 546 : sjinfo->min_righthand);
637 546 : more = true;
638 : }
639 : }
640 2786 : } while (more);
641 2390 : if (bms_overlap(join_plus_rhs, join_lateral_rels))
642 312 : return false; /* will not be able to join to some RHS rel */
643 : }
644 : }
645 :
646 : /* Otherwise, it's a valid join */
647 330022 : *sjinfo_p = match_sjinfo;
648 330022 : *reversed_p = reversed;
649 330022 : return true;
650 : }
651 :
652 : /*
653 : * init_dummy_sjinfo
654 : * Populate the given SpecialJoinInfo for a plain inner join between the
655 : * left and right relations specified by left_relids and right_relids
656 : * respectively.
657 : *
658 : * Normally, an inner join does not have a SpecialJoinInfo node associated with
659 : * it. But some functions involved in join planning require one containing at
660 : * least the information of which relations are being joined. So we initialize
661 : * that information here.
662 : */
663 : void
664 1233092 : init_dummy_sjinfo(SpecialJoinInfo *sjinfo, Relids left_relids,
665 : Relids right_relids)
666 : {
667 1233092 : sjinfo->type = T_SpecialJoinInfo;
668 1233092 : sjinfo->min_lefthand = left_relids;
669 1233092 : sjinfo->min_righthand = right_relids;
670 1233092 : sjinfo->syn_lefthand = left_relids;
671 1233092 : sjinfo->syn_righthand = right_relids;
672 1233092 : sjinfo->jointype = JOIN_INNER;
673 1233092 : sjinfo->ojrelid = 0;
674 1233092 : sjinfo->commute_above_l = NULL;
675 1233092 : sjinfo->commute_above_r = NULL;
676 1233092 : sjinfo->commute_below_l = NULL;
677 1233092 : sjinfo->commute_below_r = NULL;
678 : /* we don't bother trying to make the remaining fields valid */
679 1233092 : sjinfo->lhs_strict = false;
680 1233092 : sjinfo->semi_can_btree = false;
681 1233092 : sjinfo->semi_can_hash = false;
682 1233092 : sjinfo->semi_operators = NIL;
683 1233092 : sjinfo->semi_rhs_exprs = NIL;
684 1233092 : }
685 :
686 : /*
687 : * make_join_rel
688 : * Find or create a join RelOptInfo that represents the join of
689 : * the two given rels, and add to it path information for paths
690 : * created with the two rels as outer and inner rel.
691 : * (The join rel may already contain paths generated from other
692 : * pairs of rels that add up to the same set of base rels.)
693 : *
694 : * NB: will return NULL if attempted join is not valid. This can happen
695 : * when working with outer joins, or with IN or EXISTS clauses that have been
696 : * turned into joins.
697 : */
698 : RelOptInfo *
699 340906 : make_join_rel(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2)
700 : {
701 : Relids joinrelids;
702 : SpecialJoinInfo *sjinfo;
703 : bool reversed;
704 340906 : List *pushed_down_joins = NIL;
705 : SpecialJoinInfo sjinfo_data;
706 : RelOptInfo *joinrel;
707 : List *restrictlist;
708 :
709 : /* We should never try to join two overlapping sets of rels. */
710 : Assert(!bms_overlap(rel1->relids, rel2->relids));
711 :
712 : /* Construct Relids set that identifies the joinrel (without OJ as yet). */
713 340906 : joinrelids = bms_union(rel1->relids, rel2->relids);
714 :
715 : /* Check validity and determine join type. */
716 340906 : if (!join_is_legal(root, rel1, rel2, joinrelids,
717 : &sjinfo, &reversed))
718 : {
719 : /* invalid join path */
720 11136 : bms_free(joinrelids);
721 11136 : return NULL;
722 : }
723 :
724 : /*
725 : * Add outer join relid(s) to form the canonical relids. Any added outer
726 : * joins besides sjinfo itself are appended to pushed_down_joins.
727 : */
728 329770 : joinrelids = add_outer_joins_to_relids(root, joinrelids, sjinfo,
729 : &pushed_down_joins);
730 :
731 : /* Swap rels if needed to match the join info. */
732 329770 : if (reversed)
733 : {
734 8296 : RelOptInfo *trel = rel1;
735 :
736 8296 : rel1 = rel2;
737 8296 : rel2 = trel;
738 : }
739 :
740 : /*
741 : * If it's a plain inner join, then we won't have found anything in
742 : * join_info_list. Make up a SpecialJoinInfo so that selectivity
743 : * estimation functions will know what's being joined.
744 : */
745 329770 : if (sjinfo == NULL)
746 : {
747 223244 : sjinfo = &sjinfo_data;
748 223244 : init_dummy_sjinfo(sjinfo, rel1->relids, rel2->relids);
749 : }
750 :
751 : /*
752 : * Find or build the join RelOptInfo, and compute the restrictlist that
753 : * goes with this particular joining.
754 : */
755 329770 : joinrel = build_join_rel(root, joinrelids, rel1, rel2,
756 : sjinfo, pushed_down_joins,
757 : &restrictlist);
758 :
759 : /*
760 : * If we've already proven this join is empty, we needn't consider any
761 : * more paths for it.
762 : */
763 329770 : if (is_dummy_rel(joinrel))
764 : {
765 504 : bms_free(joinrelids);
766 504 : return joinrel;
767 : }
768 :
769 : /* Build a grouped join relation for 'joinrel' if possible. */
770 329266 : make_grouped_join_rel(root, rel1, rel2, joinrel, sjinfo,
771 : restrictlist);
772 :
773 : /* Add paths to the join relation. */
774 329266 : populate_joinrel_with_paths(root, rel1, rel2, joinrel, sjinfo,
775 : restrictlist);
776 :
777 329266 : bms_free(joinrelids);
778 :
779 329266 : return joinrel;
780 : }
781 :
782 : /*
783 : * add_outer_joins_to_relids
784 : * Add relids to input_relids to represent any outer joins that will be
785 : * calculated at this join.
786 : *
787 : * input_relids is the union of the relid sets of the two input relations.
788 : * Note that we modify this in-place and return it; caller must bms_copy()
789 : * it first, if a separate value is desired.
790 : *
791 : * sjinfo represents the join being performed.
792 : *
793 : * If the current join completes the calculation of any outer joins that
794 : * have been pushed down per outer-join identity 3, those relids will be
795 : * added to the result along with sjinfo's own relid. If pushed_down_joins
796 : * is not NULL, then also the SpecialJoinInfos for such added outer joins will
797 : * be appended to *pushed_down_joins (so caller must initialize it to NIL).
798 : */
799 : Relids
800 337288 : add_outer_joins_to_relids(PlannerInfo *root, Relids input_relids,
801 : SpecialJoinInfo *sjinfo,
802 : List **pushed_down_joins)
803 : {
804 : /* Nothing to do if this isn't an outer join with an assigned relid. */
805 337288 : if (sjinfo == NULL || sjinfo->ojrelid == 0)
806 241882 : return input_relids;
807 :
808 : /*
809 : * If it's not a left join, we have no rules that would permit executing
810 : * it in non-syntactic order, so just form the syntactic relid set. (This
811 : * is just a quick-exit test; we'd come to the same conclusion anyway,
812 : * since its commute_below_l and commute_above_l sets must be empty.)
813 : */
814 95406 : if (sjinfo->jointype != JOIN_LEFT)
815 2236 : return bms_add_member(input_relids, sjinfo->ojrelid);
816 :
817 : /*
818 : * We cannot add the OJ relid if this join has been pushed into the RHS of
819 : * a syntactically-lower left join per OJ identity 3. (If it has, then we
820 : * cannot claim that its outputs represent the final state of its RHS.)
821 : * There will not be any other OJs that can be added either, so we're
822 : * done.
823 : */
824 93170 : if (!bms_is_subset(sjinfo->commute_below_l, input_relids))
825 4890 : return input_relids;
826 :
827 : /* OK to add OJ's own relid */
828 88280 : input_relids = bms_add_member(input_relids, sjinfo->ojrelid);
829 :
830 : /*
831 : * Contrariwise, if we are now forming the final result of such a commuted
832 : * pair of OJs, it's time to add the relid(s) of the pushed-down join(s).
833 : * We can skip this if this join was never a candidate to be pushed up.
834 : */
835 88280 : if (sjinfo->commute_above_l)
836 : {
837 17464 : Relids commute_above_rels = bms_copy(sjinfo->commute_above_l);
838 : ListCell *lc;
839 :
840 : /*
841 : * The current join could complete the nulling of more than one
842 : * pushed-down join, so we have to examine all the SpecialJoinInfos.
843 : * Because join_info_list was built in bottom-up order, it's
844 : * sufficient to traverse it once: an ojrelid we add in one loop
845 : * iteration would not have affected decisions of earlier iterations.
846 : */
847 58720 : foreach(lc, root->join_info_list)
848 : {
849 41256 : SpecialJoinInfo *othersj = (SpecialJoinInfo *) lfirst(lc);
850 :
851 41256 : if (othersj == sjinfo ||
852 23792 : othersj->ojrelid == 0 || othersj->jointype != JOIN_LEFT)
853 17476 : continue; /* definitely not interesting */
854 :
855 23780 : if (!bms_is_member(othersj->ojrelid, commute_above_rels))
856 6168 : continue;
857 :
858 : /* Add it if not already present but conditions now satisfied */
859 35224 : if (!bms_is_member(othersj->ojrelid, input_relids) &&
860 35200 : bms_is_subset(othersj->min_lefthand, input_relids) &&
861 26470 : bms_is_subset(othersj->min_righthand, input_relids) &&
862 8882 : bms_is_subset(othersj->commute_below_l, input_relids))
863 : {
864 8846 : input_relids = bms_add_member(input_relids, othersj->ojrelid);
865 : /* report such pushed down outer joins, if asked */
866 8846 : if (pushed_down_joins != NULL)
867 8846 : *pushed_down_joins = lappend(*pushed_down_joins, othersj);
868 :
869 : /*
870 : * We must also check any joins that othersj potentially
871 : * commutes with. They likewise must appear later in
872 : * join_info_list than othersj itself, so we can visit them
873 : * later in this loop.
874 : */
875 8846 : commute_above_rels = bms_add_members(commute_above_rels,
876 8846 : othersj->commute_above_l);
877 : }
878 : }
879 : }
880 :
881 88280 : return input_relids;
882 : }
883 :
884 : /*
885 : * make_grouped_join_rel
886 : * Build a grouped join relation for the given "joinrel" if eager
887 : * aggregation is applicable and the resulting grouped paths are considered
888 : * useful.
889 : *
890 : * There are two strategies for generating grouped paths for a join relation:
891 : *
892 : * 1. Join a grouped (partially aggregated) input relation with a non-grouped
893 : * input (e.g., AGG(B) JOIN A).
894 : *
895 : * 2. Apply partial aggregation (sorted or hashed) on top of existing
896 : * non-grouped join paths (e.g., AGG(A JOIN B)).
897 : *
898 : * To limit planning effort and avoid an explosion of alternatives, we adopt a
899 : * strategy where partial aggregation is only pushed to the lowest possible
900 : * level in the join tree that is deemed useful. That is, if grouped paths can
901 : * be built using the first strategy, we skip consideration of the second
902 : * strategy for the same join level.
903 : *
904 : * Additionally, if there are multiple lowest useful levels where partial
905 : * aggregation could be applied, such as in a join tree with relations A, B,
906 : * and C where both "AGG(A JOIN B) JOIN C" and "A JOIN AGG(B JOIN C)" are valid
907 : * placements, we choose only the first one encountered during join search.
908 : * This avoids generating multiple versions of the same grouped relation based
909 : * on different aggregation placements.
910 : *
911 : * These heuristics also ensure that all grouped paths for the same grouped
912 : * relation produce the same set of rows, which is a basic assumption in the
913 : * planner.
914 : */
915 : static void
916 348660 : make_grouped_join_rel(PlannerInfo *root, RelOptInfo *rel1,
917 : RelOptInfo *rel2, RelOptInfo *joinrel,
918 : SpecialJoinInfo *sjinfo, List *restrictlist)
919 : {
920 : RelOptInfo *grouped_rel;
921 : RelOptInfo *grouped_rel1;
922 : RelOptInfo *grouped_rel2;
923 : bool rel1_empty;
924 : bool rel2_empty;
925 : Relids apply_agg_at;
926 :
927 : /*
928 : * If there are no aggregate expressions or grouping expressions, eager
929 : * aggregation is not possible.
930 : */
931 348660 : if (root->agg_clause_list == NIL ||
932 18780 : root->group_expr_list == NIL)
933 330114 : return;
934 :
935 : /* Retrieve the grouped relations for the two input rels */
936 18546 : grouped_rel1 = rel1->grouped_rel;
937 18546 : grouped_rel2 = rel2->grouped_rel;
938 :
939 18546 : rel1_empty = (grouped_rel1 == NULL || IS_DUMMY_REL(grouped_rel1));
940 18546 : rel2_empty = (grouped_rel2 == NULL || IS_DUMMY_REL(grouped_rel2));
941 :
942 : /* Find or construct a grouped joinrel for this joinrel */
943 18546 : grouped_rel = joinrel->grouped_rel;
944 18546 : if (grouped_rel == NULL)
945 : {
946 17970 : RelAggInfo *agg_info = NULL;
947 :
948 : /*
949 : * Prepare the information needed to create grouped paths for this
950 : * join relation.
951 : */
952 17970 : agg_info = create_rel_agg_info(root, joinrel, rel1_empty == rel2_empty);
953 17970 : if (agg_info == NULL)
954 840 : return;
955 :
956 : /*
957 : * If grouped paths for the given join relation are not considered
958 : * useful, and no grouped paths can be built by joining grouped input
959 : * relations, skip building the grouped join relation.
960 : */
961 17130 : if (!agg_info->agg_useful &&
962 : (rel1_empty == rel2_empty))
963 260 : return;
964 :
965 : /* build the grouped relation */
966 16870 : grouped_rel = build_grouped_rel(root, joinrel);
967 16870 : grouped_rel->reltarget = agg_info->target;
968 :
969 16870 : if (rel1_empty != rel2_empty)
970 : {
971 : /*
972 : * If there is exactly one grouped input relation, then we can
973 : * build grouped paths by joining the input relations. Set size
974 : * estimates for the grouped join relation based on the input
975 : * relations, and update the set of relids where partial
976 : * aggregation is applied to that of the grouped input relation.
977 : */
978 16210 : set_joinrel_size_estimates(root, grouped_rel,
979 : rel1_empty ? rel1 : grouped_rel1,
980 : rel2_empty ? rel2 : grouped_rel2,
981 : sjinfo, restrictlist);
982 16210 : agg_info->apply_agg_at = rel1_empty ?
983 16210 : grouped_rel2->agg_info->apply_agg_at :
984 8032 : grouped_rel1->agg_info->apply_agg_at;
985 : }
986 : else
987 : {
988 : /*
989 : * Otherwise, grouped paths can be built by applying partial
990 : * aggregation on top of existing non-grouped join paths. Set
991 : * size estimates for the grouped join relation based on the
992 : * estimated number of groups, and track the set of relids where
993 : * partial aggregation is applied. Note that these values may be
994 : * updated later if it is determined that grouped paths can be
995 : * constructed by joining other input relations.
996 : */
997 660 : grouped_rel->rows = agg_info->grouped_rows;
998 660 : agg_info->apply_agg_at = bms_copy(joinrel->relids);
999 : }
1000 :
1001 16870 : grouped_rel->agg_info = agg_info;
1002 16870 : joinrel->grouped_rel = grouped_rel;
1003 : }
1004 :
1005 : Assert(IS_GROUPED_REL(grouped_rel));
1006 :
1007 : /* We may have already proven this grouped join relation to be dummy. */
1008 17446 : if (IS_DUMMY_REL(grouped_rel))
1009 0 : return;
1010 :
1011 : /*
1012 : * Nothing to do if there's no grouped input relation. Also, joining two
1013 : * grouped relations is not currently supported.
1014 : */
1015 17446 : if (rel1_empty == rel2_empty)
1016 948 : return;
1017 :
1018 : /*
1019 : * Get the set of relids where partial aggregation is applied among the
1020 : * given input relations.
1021 : */
1022 16498 : apply_agg_at = rel1_empty ?
1023 16498 : grouped_rel2->agg_info->apply_agg_at :
1024 8320 : grouped_rel1->agg_info->apply_agg_at;
1025 :
1026 : /*
1027 : * If it's not the designated level, skip building grouped paths.
1028 : *
1029 : * One exception is when it is a subset of the previously recorded level.
1030 : * In that case, we need to update the designated level to this one, and
1031 : * adjust the size estimates for the grouped join relation accordingly.
1032 : * For example, suppose partial aggregation can be applied on top of (B
1033 : * JOIN C). If we first construct the join as ((A JOIN B) JOIN C), we'd
1034 : * record the designated level as including all three relations (A B C).
1035 : * Later, when we consider (A JOIN (B JOIN C)), we encounter the smaller
1036 : * (B C) join level directly. Since this is a subset of the previous
1037 : * level and still valid for partial aggregation, we update the designated
1038 : * level to (B C), and adjust the size estimates accordingly.
1039 : */
1040 16498 : if (!bms_equal(apply_agg_at, grouped_rel->agg_info->apply_agg_at))
1041 : {
1042 288 : if (bms_is_subset(apply_agg_at, grouped_rel->agg_info->apply_agg_at))
1043 : {
1044 : /* Adjust the size estimates for the grouped join relation. */
1045 288 : set_joinrel_size_estimates(root, grouped_rel,
1046 : rel1_empty ? rel1 : grouped_rel1,
1047 : rel2_empty ? rel2 : grouped_rel2,
1048 : sjinfo, restrictlist);
1049 288 : grouped_rel->agg_info->apply_agg_at = apply_agg_at;
1050 : }
1051 : else
1052 0 : return;
1053 : }
1054 :
1055 : /* Make paths for the grouped join relation. */
1056 16498 : populate_joinrel_with_paths(root,
1057 : rel1_empty ? rel1 : grouped_rel1,
1058 : rel2_empty ? rel2 : grouped_rel2,
1059 : grouped_rel,
1060 : sjinfo,
1061 : restrictlist);
1062 : }
1063 :
1064 : /*
1065 : * populate_joinrel_with_paths
1066 : * Add paths to the given joinrel for given pair of joining relations. The
1067 : * SpecialJoinInfo provides details about the join and the restrictlist
1068 : * contains the join clauses and the other clauses applicable for given pair
1069 : * of the joining relations.
1070 : */
1071 : static void
1072 365158 : populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
1073 : RelOptInfo *rel2, RelOptInfo *joinrel,
1074 : SpecialJoinInfo *sjinfo, List *restrictlist)
1075 : {
1076 : RelOptInfo *unique_rel2;
1077 :
1078 : /*
1079 : * Consider paths using each rel as both outer and inner. Depending on
1080 : * the join type, a provably empty outer or inner rel might mean the join
1081 : * is provably empty too; in which case throw away any previously computed
1082 : * paths and mark the join as dummy. (We do it this way since it's
1083 : * conceivable that dummy-ness of a multi-element join might only be
1084 : * noticeable for certain construction paths.)
1085 : *
1086 : * Also, a provably constant-false join restriction typically means that
1087 : * we can skip evaluating one or both sides of the join. We do this by
1088 : * marking the appropriate rel as dummy. For outer joins, a
1089 : * constant-false restriction that is pushed down still means the whole
1090 : * join is dummy, while a non-pushed-down one means that no inner rows
1091 : * will join so we can treat the inner rel as dummy.
1092 : *
1093 : * We need only consider the jointypes that appear in join_info_list, plus
1094 : * JOIN_INNER.
1095 : */
1096 365158 : switch (sjinfo->jointype)
1097 : {
1098 255464 : case JOIN_INNER:
1099 510898 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
1100 255434 : restriction_is_constant_false(restrictlist, joinrel, false))
1101 : {
1102 216 : mark_dummy_rel(joinrel);
1103 216 : break;
1104 : }
1105 255248 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1106 : JOIN_INNER, sjinfo,
1107 : restrictlist);
1108 255248 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1109 : JOIN_INNER, sjinfo,
1110 : restrictlist);
1111 255248 : break;
1112 94860 : case JOIN_LEFT:
1113 189666 : if (is_dummy_rel(rel1) ||
1114 94806 : restriction_is_constant_false(restrictlist, joinrel, true))
1115 : {
1116 86 : mark_dummy_rel(joinrel);
1117 86 : break;
1118 : }
1119 94990 : if (restriction_is_constant_false(restrictlist, joinrel, false) &&
1120 216 : bms_is_subset(rel2->relids, sjinfo->syn_righthand))
1121 192 : mark_dummy_rel(rel2);
1122 94774 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1123 : JOIN_LEFT, sjinfo,
1124 : restrictlist);
1125 94774 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1126 : JOIN_RIGHT, sjinfo,
1127 : restrictlist);
1128 94774 : break;
1129 1732 : case JOIN_FULL:
1130 3464 : if ((is_dummy_rel(rel1) && is_dummy_rel(rel2)) ||
1131 1732 : restriction_is_constant_false(restrictlist, joinrel, true))
1132 : {
1133 12 : mark_dummy_rel(joinrel);
1134 12 : break;
1135 : }
1136 1720 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1137 : JOIN_FULL, sjinfo,
1138 : restrictlist);
1139 1720 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1140 : JOIN_FULL, sjinfo,
1141 : restrictlist);
1142 :
1143 : /*
1144 : * If there are join quals that aren't mergeable or hashable, we
1145 : * may not be able to build any valid plan. Complain here so that
1146 : * we can give a somewhat-useful error message. (Since we have no
1147 : * flexibility of planning for a full join, there's no chance of
1148 : * succeeding later with another pair of input rels.)
1149 : */
1150 1720 : if (joinrel->pathlist == NIL)
1151 0 : ereport(ERROR,
1152 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1153 : errmsg("FULL JOIN is only supported with merge-joinable or hash-joinable join conditions")));
1154 1720 : break;
1155 7944 : case JOIN_SEMI:
1156 :
1157 : /*
1158 : * We might have a normal semijoin, or a case where we don't have
1159 : * enough rels to do the semijoin but can unique-ify the RHS and
1160 : * then do an innerjoin (see comments in join_is_legal). In the
1161 : * latter case we can't apply JOIN_SEMI joining.
1162 : */
1163 15458 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1164 7514 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
1165 : {
1166 15022 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
1167 7508 : restriction_is_constant_false(restrictlist, joinrel, false))
1168 : {
1169 12 : mark_dummy_rel(joinrel);
1170 12 : break;
1171 : }
1172 7502 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1173 : JOIN_SEMI, sjinfo,
1174 : restrictlist);
1175 7502 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1176 : JOIN_RIGHT_SEMI, sjinfo,
1177 : restrictlist);
1178 : }
1179 :
1180 : /*
1181 : * If we know how to unique-ify the RHS and one input rel is
1182 : * exactly the RHS (not a superset) we can consider unique-ifying
1183 : * it and then doing a regular join. (The create_unique_paths
1184 : * check here is probably redundant with what join_is_legal did,
1185 : * but if so the check is cheap because it's cached. So test
1186 : * anyway to be sure.)
1187 : */
1188 15864 : if (bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1189 7932 : (unique_rel2 = create_unique_paths(root, rel2, sjinfo)) != NULL)
1190 : {
1191 12372 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
1192 6186 : restriction_is_constant_false(restrictlist, joinrel, false))
1193 : {
1194 0 : mark_dummy_rel(joinrel);
1195 0 : break;
1196 : }
1197 6186 : add_paths_to_joinrel(root, joinrel, rel1, unique_rel2,
1198 : JOIN_UNIQUE_INNER, sjinfo,
1199 : restrictlist);
1200 6186 : add_paths_to_joinrel(root, joinrel, unique_rel2, rel1,
1201 : JOIN_UNIQUE_OUTER, sjinfo,
1202 : restrictlist);
1203 : }
1204 7932 : break;
1205 5158 : case JOIN_ANTI:
1206 10316 : if (is_dummy_rel(rel1) ||
1207 5158 : restriction_is_constant_false(restrictlist, joinrel, true))
1208 : {
1209 0 : mark_dummy_rel(joinrel);
1210 0 : break;
1211 : }
1212 5158 : if (restriction_is_constant_false(restrictlist, joinrel, false) &&
1213 0 : bms_is_subset(rel2->relids, sjinfo->syn_righthand))
1214 0 : mark_dummy_rel(rel2);
1215 5158 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1216 : JOIN_ANTI, sjinfo,
1217 : restrictlist);
1218 5158 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1219 : JOIN_RIGHT_ANTI, sjinfo,
1220 : restrictlist);
1221 5158 : break;
1222 0 : default:
1223 : /* other values not expected here */
1224 0 : elog(ERROR, "unrecognized join type: %d", (int) sjinfo->jointype);
1225 : break;
1226 : }
1227 :
1228 : /* Apply partitionwise join technique, if possible. */
1229 365158 : try_partitionwise_join(root, rel1, rel2, joinrel, sjinfo, restrictlist);
1230 365158 : }
1231 :
1232 :
1233 : /*
1234 : * have_join_order_restriction
1235 : * Detect whether the two relations should be joined to satisfy
1236 : * a join-order restriction arising from special or lateral joins.
1237 : *
1238 : * In practice this is always used with have_relevant_joinclause(), and so
1239 : * could be merged with that function, but it seems clearer to separate the
1240 : * two concerns. We need this test because there are degenerate cases where
1241 : * a clauseless join must be performed to satisfy join-order restrictions.
1242 : * Also, if one rel has a lateral reference to the other, or both are needed
1243 : * to compute some PHV, we should consider joining them even if the join would
1244 : * be clauseless.
1245 : *
1246 : * Note: this is only a problem if one side of a degenerate outer join
1247 : * contains multiple rels, or a clauseless join is required within an
1248 : * IN/EXISTS RHS; else we will find a join path via the "last ditch" case in
1249 : * join_search_one_level(). We could dispense with this test if we were
1250 : * willing to try bushy plans in the "last ditch" case, but that seems much
1251 : * less efficient.
1252 : */
1253 : bool
1254 79624 : have_join_order_restriction(PlannerInfo *root,
1255 : RelOptInfo *rel1, RelOptInfo *rel2)
1256 : {
1257 79624 : bool result = false;
1258 : ListCell *l;
1259 :
1260 : /*
1261 : * If either side has a direct lateral reference to the other, attempt the
1262 : * join regardless of outer-join considerations.
1263 : */
1264 149868 : if (bms_overlap(rel1->relids, rel2->direct_lateral_relids) ||
1265 70244 : bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1266 10494 : return true;
1267 :
1268 : /*
1269 : * Likewise, if both rels are needed to compute some PlaceHolderVar,
1270 : * attempt the join regardless of outer-join considerations. (This is not
1271 : * very desirable, because a PHV with a large eval_at set will cause a lot
1272 : * of probably-useless joins to be considered, but failing to do this can
1273 : * cause us to fail to construct a plan at all.)
1274 : */
1275 71012 : foreach(l, root->placeholder_list)
1276 : {
1277 1942 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1278 :
1279 2332 : if (bms_is_subset(rel1->relids, phinfo->ph_eval_at) &&
1280 390 : bms_is_subset(rel2->relids, phinfo->ph_eval_at))
1281 60 : return true;
1282 : }
1283 :
1284 : /*
1285 : * It's possible that the rels correspond to the left and right sides of a
1286 : * degenerate outer join, that is, one with no joinclause mentioning the
1287 : * non-nullable side; in which case we should force the join to occur.
1288 : *
1289 : * Also, the two rels could represent a clauseless join that has to be
1290 : * completed to build up the LHS or RHS of an outer join.
1291 : */
1292 180466 : foreach(l, root->join_info_list)
1293 : {
1294 112874 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1295 :
1296 : /* ignore full joins --- other mechanisms handle them */
1297 112874 : if (sjinfo->jointype == JOIN_FULL)
1298 42 : continue;
1299 :
1300 : /* Can we perform the SJ with these rels? */
1301 138470 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1302 25638 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
1303 : {
1304 1076 : result = true;
1305 1076 : break;
1306 : }
1307 118650 : if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1308 6894 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
1309 : {
1310 210 : result = true;
1311 210 : break;
1312 : }
1313 :
1314 : /*
1315 : * Might we need to join these rels to complete the RHS? We have to
1316 : * use "overlap" tests since either rel might include a lower SJ that
1317 : * has been proven to commute with this one.
1318 : */
1319 134538 : if (bms_overlap(sjinfo->min_righthand, rel1->relids) &&
1320 22992 : bms_overlap(sjinfo->min_righthand, rel2->relids))
1321 : {
1322 120 : result = true;
1323 120 : break;
1324 : }
1325 :
1326 : /* Likewise for the LHS. */
1327 139020 : if (bms_overlap(sjinfo->min_lefthand, rel1->relids) &&
1328 27594 : bms_overlap(sjinfo->min_lefthand, rel2->relids))
1329 : {
1330 72 : result = true;
1331 72 : break;
1332 : }
1333 : }
1334 :
1335 : /*
1336 : * We do not force the join to occur if either input rel can legally be
1337 : * joined to anything else using joinclauses. This essentially means that
1338 : * clauseless bushy joins are put off as long as possible. The reason is
1339 : * that when there is a join order restriction high up in the join tree
1340 : * (that is, with many rels inside the LHS or RHS), we would otherwise
1341 : * expend lots of effort considering very stupid join combinations within
1342 : * its LHS or RHS.
1343 : */
1344 69070 : if (result)
1345 : {
1346 2818 : if (has_legal_joinclause(root, rel1) ||
1347 1340 : has_legal_joinclause(root, rel2))
1348 252 : result = false;
1349 : }
1350 :
1351 69070 : return result;
1352 : }
1353 :
1354 :
1355 : /*
1356 : * has_join_restriction
1357 : * Detect whether the specified relation has join-order restrictions,
1358 : * due to being inside an outer join or an IN (sub-SELECT),
1359 : * or participating in any LATERAL references or multi-rel PHVs.
1360 : *
1361 : * Essentially, this tests whether have_join_order_restriction() could
1362 : * succeed with this rel and some other one. It's OK if we sometimes
1363 : * say "true" incorrectly. (Therefore, we don't bother with the relatively
1364 : * expensive has_legal_joinclause test.)
1365 : */
1366 : static bool
1367 30798 : has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1368 : {
1369 : ListCell *l;
1370 :
1371 30798 : if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1372 17610 : return true;
1373 :
1374 13980 : foreach(l, root->placeholder_list)
1375 : {
1376 840 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1377 :
1378 840 : if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1379 204 : !bms_equal(rel->relids, phinfo->ph_eval_at))
1380 48 : return true;
1381 : }
1382 :
1383 13934 : foreach(l, root->join_info_list)
1384 : {
1385 2946 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1386 :
1387 : /* ignore full joins --- other mechanisms preserve their ordering */
1388 2946 : if (sjinfo->jointype == JOIN_FULL)
1389 86 : continue;
1390 :
1391 : /* ignore if SJ is already contained in rel */
1392 4410 : if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1393 1550 : bms_is_subset(sjinfo->min_righthand, rel->relids))
1394 372 : continue;
1395 :
1396 : /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1397 3774 : if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1398 1286 : bms_overlap(sjinfo->min_righthand, rel->relids))
1399 2152 : return true;
1400 : }
1401 :
1402 10988 : return false;
1403 : }
1404 :
1405 :
1406 : /*
1407 : * has_legal_joinclause
1408 : * Detect whether the specified relation can legally be joined
1409 : * to any other rels using join clauses.
1410 : *
1411 : * We consider only joins to single other relations in the current
1412 : * initial_rels list. This is sufficient to get a "true" result in most real
1413 : * queries, and an occasional erroneous "false" will only cost a bit more
1414 : * planning time. The reason for this limitation is that considering joins to
1415 : * other joins would require proving that the other join rel can legally be
1416 : * formed, which seems like too much trouble for something that's only a
1417 : * heuristic to save planning time. (Note: we must look at initial_rels
1418 : * and not all of the query, since when we are planning a sub-joinlist we
1419 : * may be forced to make clauseless joins within initial_rels even though
1420 : * there are join clauses linking to other parts of the query.)
1421 : */
1422 : static bool
1423 2818 : has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel)
1424 : {
1425 : ListCell *lc;
1426 :
1427 10458 : foreach(lc, root->initial_rels)
1428 : {
1429 7892 : RelOptInfo *rel2 = (RelOptInfo *) lfirst(lc);
1430 :
1431 : /* ignore rels that are already in "rel" */
1432 7892 : if (bms_overlap(rel->relids, rel2->relids))
1433 3322 : continue;
1434 :
1435 4570 : if (have_relevant_joinclause(root, rel, rel2))
1436 : {
1437 : Relids joinrelids;
1438 : SpecialJoinInfo *sjinfo;
1439 : bool reversed;
1440 :
1441 : /* join_is_legal needs relids of the union */
1442 456 : joinrelids = bms_union(rel->relids, rel2->relids);
1443 :
1444 456 : if (join_is_legal(root, rel, rel2, joinrelids,
1445 : &sjinfo, &reversed))
1446 : {
1447 : /* Yes, this will work */
1448 252 : bms_free(joinrelids);
1449 252 : return true;
1450 : }
1451 :
1452 204 : bms_free(joinrelids);
1453 : }
1454 : }
1455 :
1456 2566 : return false;
1457 : }
1458 :
1459 :
1460 : /*
1461 : * is_dummy_rel --- has relation been proven empty?
1462 : */
1463 : bool
1464 2876694 : is_dummy_rel(RelOptInfo *rel)
1465 : {
1466 : Path *path;
1467 :
1468 : /*
1469 : * A rel that is known dummy will have just one path that is a childless
1470 : * Append. (Even if somehow it has more paths, a childless Append will
1471 : * have cost zero and hence should be at the front of the pathlist.)
1472 : */
1473 2876694 : if (rel->pathlist == NIL)
1474 1449572 : return false;
1475 1427122 : path = (Path *) linitial(rel->pathlist);
1476 :
1477 : /*
1478 : * Initially, a dummy path will just be a childless Append. But in later
1479 : * planning stages we might stick a ProjectSetPath and/or ProjectionPath
1480 : * on top, since Append can't project. Rather than make assumptions about
1481 : * which combinations can occur, just descend through whatever we find.
1482 : */
1483 : for (;;)
1484 : {
1485 1502416 : if (IsA(path, ProjectionPath))
1486 66146 : path = ((ProjectionPath *) path)->subpath;
1487 1436270 : else if (IsA(path, ProjectSetPath))
1488 9148 : path = ((ProjectSetPath *) path)->subpath;
1489 : else
1490 1427122 : break;
1491 : }
1492 1427122 : if (IS_DUMMY_APPEND(path))
1493 4672 : return true;
1494 1422450 : return false;
1495 : }
1496 :
1497 : /*
1498 : * Mark a relation as proven empty.
1499 : *
1500 : * During GEQO planning, this can get invoked more than once on the same
1501 : * baserel struct, so it's worth checking to see if the rel is already marked
1502 : * dummy.
1503 : *
1504 : * Also, when called during GEQO join planning, we are in a short-lived
1505 : * memory context. We must make sure that the dummy path attached to a
1506 : * baserel survives the GEQO cycle, else the baserel is trashed for future
1507 : * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1508 : * we don't want the dummy path to clutter the main planning context. Upshot
1509 : * is that the best solution is to explicitly make the dummy path in the same
1510 : * context the given RelOptInfo is in.
1511 : */
1512 : void
1513 720 : mark_dummy_rel(RelOptInfo *rel)
1514 : {
1515 : MemoryContext oldcontext;
1516 :
1517 : /* Already marked? */
1518 720 : if (is_dummy_rel(rel))
1519 18 : return;
1520 :
1521 : /* No, so choose correct context to make the dummy path in */
1522 702 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1523 :
1524 : /* Set dummy size estimate */
1525 702 : rel->rows = 0;
1526 :
1527 : /* Evict any previously chosen paths */
1528 702 : rel->pathlist = NIL;
1529 702 : rel->partial_pathlist = NIL;
1530 :
1531 : /* Set up the dummy path */
1532 702 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
1533 : NIL, rel->lateral_relids,
1534 : 0, false, -1));
1535 :
1536 : /* Set or update cheapest_total_path and related fields */
1537 702 : set_cheapest(rel);
1538 :
1539 702 : MemoryContextSwitchTo(oldcontext);
1540 : }
1541 :
1542 :
1543 : /*
1544 : * restriction_is_constant_false --- is a restrictlist just FALSE?
1545 : *
1546 : * In cases where a qual is provably constant FALSE, eval_const_expressions
1547 : * will generally have thrown away anything that's ANDed with it. In outer
1548 : * join situations this will leave us computing cartesian products only to
1549 : * decide there's no match for an outer row, which is pretty stupid. So,
1550 : * we need to detect the case.
1551 : *
1552 : * If only_pushed_down is true, then consider only quals that are pushed-down
1553 : * from the point of view of the joinrel.
1554 : */
1555 : static bool
1556 470756 : restriction_is_constant_false(List *restrictlist,
1557 : RelOptInfo *joinrel,
1558 : bool only_pushed_down)
1559 : {
1560 : ListCell *lc;
1561 :
1562 : /*
1563 : * Despite the above comment, the restriction list we see here might
1564 : * possibly have other members besides the FALSE constant, since other
1565 : * quals could get "pushed down" to the outer join level. So we check
1566 : * each member of the list.
1567 : */
1568 1003812 : foreach(lc, restrictlist)
1569 : {
1570 533508 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1571 :
1572 533508 : if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1573 127324 : continue;
1574 :
1575 406184 : if (rinfo->clause && IsA(rinfo->clause, Const))
1576 : {
1577 5252 : Const *con = (Const *) rinfo->clause;
1578 :
1579 : /* constant NULL is as good as constant FALSE for our purposes */
1580 5252 : if (con->constisnull)
1581 452 : return true;
1582 5144 : if (!DatumGetBool(con->constvalue))
1583 344 : return true;
1584 : }
1585 : }
1586 470304 : return false;
1587 : }
1588 :
1589 : /*
1590 : * Assess whether join between given two partitioned relations can be broken
1591 : * down into joins between matching partitions; a technique called
1592 : * "partitionwise join"
1593 : *
1594 : * Partitionwise join is possible when a. Joining relations have same
1595 : * partitioning scheme b. There exists an equi-join between the partition keys
1596 : * of the two relations.
1597 : *
1598 : * Partitionwise join is planned as follows (details: optimizer/README.)
1599 : *
1600 : * 1. Create the RelOptInfos for joins between matching partitions i.e
1601 : * child-joins and add paths to them.
1602 : *
1603 : * 2. Construct Append or MergeAppend paths across the set of child joins.
1604 : * This second phase is implemented by generate_partitionwise_join_paths().
1605 : *
1606 : * The RelOptInfo, SpecialJoinInfo and restrictlist for each child join are
1607 : * obtained by translating the respective parent join structures.
1608 : */
1609 : static void
1610 365158 : try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1611 : RelOptInfo *joinrel, SpecialJoinInfo *parent_sjinfo,
1612 : List *parent_restrictlist)
1613 : {
1614 365158 : bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1615 365158 : bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1616 365158 : List *parts1 = NIL;
1617 365158 : List *parts2 = NIL;
1618 365158 : ListCell *lcr1 = NULL;
1619 365158 : ListCell *lcr2 = NULL;
1620 : int cnt_parts;
1621 :
1622 : /* Guard against stack overflow due to overly deep partition hierarchy. */
1623 365158 : check_stack_depth();
1624 :
1625 : /* Nothing to do, if the join relation is not partitioned. */
1626 365158 : if (joinrel->part_scheme == NULL || joinrel->nparts == 0)
1627 357512 : return;
1628 :
1629 : /* The join relation should have consider_partitionwise_join set. */
1630 : Assert(joinrel->consider_partitionwise_join);
1631 :
1632 : /*
1633 : * We can not perform partitionwise join if either of the joining
1634 : * relations is not partitioned.
1635 : */
1636 7784 : if (!IS_PARTITIONED_REL(rel1) || !IS_PARTITIONED_REL(rel2))
1637 18 : return;
1638 :
1639 : Assert(REL_HAS_ALL_PART_PROPS(rel1) && REL_HAS_ALL_PART_PROPS(rel2));
1640 :
1641 : /* The joining relations should have consider_partitionwise_join set. */
1642 : Assert(rel1->consider_partitionwise_join &&
1643 : rel2->consider_partitionwise_join);
1644 :
1645 : /*
1646 : * The partition scheme of the join relation should match that of the
1647 : * joining relations.
1648 : */
1649 : Assert(joinrel->part_scheme == rel1->part_scheme &&
1650 : joinrel->part_scheme == rel2->part_scheme);
1651 :
1652 : Assert(!(joinrel->partbounds_merged && (joinrel->nparts <= 0)));
1653 :
1654 7766 : compute_partition_bounds(root, rel1, rel2, joinrel, parent_sjinfo,
1655 : &parts1, &parts2);
1656 :
1657 7766 : if (joinrel->partbounds_merged)
1658 : {
1659 768 : lcr1 = list_head(parts1);
1660 768 : lcr2 = list_head(parts2);
1661 : }
1662 :
1663 : /*
1664 : * Create child-join relations for this partitioned join, if those don't
1665 : * exist. Add paths to child-joins for a pair of child relations
1666 : * corresponding to the given pair of parent relations.
1667 : */
1668 27236 : for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1669 : {
1670 : RelOptInfo *child_rel1;
1671 : RelOptInfo *child_rel2;
1672 : bool rel1_empty;
1673 : bool rel2_empty;
1674 : SpecialJoinInfo *child_sjinfo;
1675 : List *child_restrictlist;
1676 : RelOptInfo *child_joinrel;
1677 : AppendRelInfo **appinfos;
1678 : int nappinfos;
1679 : Relids child_relids;
1680 :
1681 19590 : if (joinrel->partbounds_merged)
1682 : {
1683 2010 : child_rel1 = lfirst_node(RelOptInfo, lcr1);
1684 2010 : child_rel2 = lfirst_node(RelOptInfo, lcr2);
1685 2010 : lcr1 = lnext(parts1, lcr1);
1686 2010 : lcr2 = lnext(parts2, lcr2);
1687 : }
1688 : else
1689 : {
1690 17580 : child_rel1 = rel1->part_rels[cnt_parts];
1691 17580 : child_rel2 = rel2->part_rels[cnt_parts];
1692 : }
1693 :
1694 19590 : rel1_empty = (child_rel1 == NULL || IS_DUMMY_REL(child_rel1));
1695 19590 : rel2_empty = (child_rel2 == NULL || IS_DUMMY_REL(child_rel2));
1696 :
1697 : /*
1698 : * Check for cases where we can prove that this segment of the join
1699 : * returns no rows, due to one or both inputs being empty (including
1700 : * inputs that have been pruned away entirely). If so just ignore it.
1701 : * These rules are equivalent to populate_joinrel_with_paths's rules
1702 : * for dummy input relations.
1703 : */
1704 19590 : switch (parent_sjinfo->jointype)
1705 : {
1706 16772 : case JOIN_INNER:
1707 : case JOIN_SEMI:
1708 16772 : if (rel1_empty || rel2_empty)
1709 76 : continue; /* ignore this join segment */
1710 16724 : break;
1711 2096 : case JOIN_LEFT:
1712 : case JOIN_ANTI:
1713 2096 : if (rel1_empty)
1714 28 : continue; /* ignore this join segment */
1715 2068 : break;
1716 722 : case JOIN_FULL:
1717 722 : if (rel1_empty && rel2_empty)
1718 0 : continue; /* ignore this join segment */
1719 722 : break;
1720 0 : default:
1721 : /* other values not expected here */
1722 0 : elog(ERROR, "unrecognized join type: %d",
1723 : (int) parent_sjinfo->jointype);
1724 : break;
1725 : }
1726 :
1727 : /*
1728 : * If a child has been pruned entirely then we can't generate paths
1729 : * for it, so we have to reject partitionwise joining unless we were
1730 : * able to eliminate this partition above.
1731 : */
1732 19514 : if (child_rel1 == NULL || child_rel2 == NULL)
1733 : {
1734 : /*
1735 : * Mark the joinrel as unpartitioned so that later functions treat
1736 : * it correctly.
1737 : */
1738 120 : joinrel->nparts = 0;
1739 120 : return;
1740 : }
1741 :
1742 : /*
1743 : * If a leaf relation has consider_partitionwise_join=false, it means
1744 : * that it's a dummy relation for which we skipped setting up tlist
1745 : * expressions and adding EC members in set_append_rel_size(), so
1746 : * again we have to fail here.
1747 : */
1748 19394 : if (rel1_is_simple && !child_rel1->consider_partitionwise_join)
1749 : {
1750 : Assert(child_rel1->reloptkind == RELOPT_OTHER_MEMBER_REL);
1751 : Assert(IS_DUMMY_REL(child_rel1));
1752 0 : joinrel->nparts = 0;
1753 0 : return;
1754 : }
1755 19394 : if (rel2_is_simple && !child_rel2->consider_partitionwise_join)
1756 : {
1757 : Assert(child_rel2->reloptkind == RELOPT_OTHER_MEMBER_REL);
1758 : Assert(IS_DUMMY_REL(child_rel2));
1759 0 : joinrel->nparts = 0;
1760 0 : return;
1761 : }
1762 :
1763 : /* We should never try to join two overlapping sets of rels. */
1764 : Assert(!bms_overlap(child_rel1->relids, child_rel2->relids));
1765 :
1766 : /*
1767 : * Construct SpecialJoinInfo from parent join relations's
1768 : * SpecialJoinInfo.
1769 : */
1770 19394 : child_sjinfo = build_child_join_sjinfo(root, parent_sjinfo,
1771 : child_rel1->relids,
1772 : child_rel2->relids);
1773 :
1774 : /* Find the AppendRelInfo structures */
1775 19394 : child_relids = bms_union(child_rel1->relids, child_rel2->relids);
1776 19394 : appinfos = find_appinfos_by_relids(root, child_relids,
1777 : &nappinfos);
1778 :
1779 : /*
1780 : * Construct restrictions applicable to the child join from those
1781 : * applicable to the parent join.
1782 : */
1783 : child_restrictlist =
1784 19394 : (List *) adjust_appendrel_attrs(root,
1785 : (Node *) parent_restrictlist,
1786 : nappinfos, appinfos);
1787 :
1788 : /* Find or construct the child join's RelOptInfo */
1789 19394 : child_joinrel = joinrel->part_rels[cnt_parts];
1790 19394 : if (!child_joinrel)
1791 : {
1792 18370 : child_joinrel = build_child_join_rel(root, child_rel1, child_rel2,
1793 : joinrel, child_restrictlist,
1794 : child_sjinfo, nappinfos, appinfos);
1795 18370 : joinrel->part_rels[cnt_parts] = child_joinrel;
1796 18370 : joinrel->live_parts = bms_add_member(joinrel->live_parts, cnt_parts);
1797 18370 : joinrel->all_partrels = bms_add_members(joinrel->all_partrels,
1798 18370 : child_joinrel->relids);
1799 : }
1800 :
1801 : /* Assert we got the right one */
1802 : Assert(bms_equal(child_joinrel->relids,
1803 : adjust_child_relids(joinrel->relids,
1804 : nappinfos, appinfos)));
1805 :
1806 : /* Build a grouped join relation for 'child_joinrel' if possible */
1807 19394 : make_grouped_join_rel(root, child_rel1, child_rel2,
1808 : child_joinrel, child_sjinfo,
1809 : child_restrictlist);
1810 :
1811 : /* And make paths for the child join */
1812 19394 : populate_joinrel_with_paths(root, child_rel1, child_rel2,
1813 : child_joinrel, child_sjinfo,
1814 : child_restrictlist);
1815 :
1816 : /*
1817 : * When there are thousands of partitions involved, this loop will
1818 : * accumulate a significant amount of memory usage from objects that
1819 : * are only needed within the loop. Free these local objects eagerly
1820 : * at the end of each iteration.
1821 : */
1822 19394 : pfree(appinfos);
1823 19394 : bms_free(child_relids);
1824 19394 : free_child_join_sjinfo(child_sjinfo, parent_sjinfo);
1825 : }
1826 : }
1827 :
1828 : /*
1829 : * Construct the SpecialJoinInfo for a child-join by translating
1830 : * SpecialJoinInfo for the join between parents. left_relids and right_relids
1831 : * are the relids of left and right side of the join respectively.
1832 : *
1833 : * If translations are added to or removed from this function, consider
1834 : * updating free_child_join_sjinfo() accordingly.
1835 : */
1836 : static SpecialJoinInfo *
1837 19394 : build_child_join_sjinfo(PlannerInfo *root, SpecialJoinInfo *parent_sjinfo,
1838 : Relids left_relids, Relids right_relids)
1839 : {
1840 19394 : SpecialJoinInfo *sjinfo = makeNode(SpecialJoinInfo);
1841 : AppendRelInfo **left_appinfos;
1842 : int left_nappinfos;
1843 : AppendRelInfo **right_appinfos;
1844 : int right_nappinfos;
1845 :
1846 : /* Dummy SpecialJoinInfos can be created without any translation. */
1847 19394 : if (parent_sjinfo->jointype == JOIN_INNER)
1848 : {
1849 : Assert(parent_sjinfo->ojrelid == 0);
1850 16208 : init_dummy_sjinfo(sjinfo, left_relids, right_relids);
1851 16208 : return sjinfo;
1852 : }
1853 :
1854 3186 : memcpy(sjinfo, parent_sjinfo, sizeof(SpecialJoinInfo));
1855 3186 : left_appinfos = find_appinfos_by_relids(root, left_relids,
1856 : &left_nappinfos);
1857 3186 : right_appinfos = find_appinfos_by_relids(root, right_relids,
1858 : &right_nappinfos);
1859 :
1860 3186 : sjinfo->min_lefthand = adjust_child_relids(sjinfo->min_lefthand,
1861 : left_nappinfos, left_appinfos);
1862 3186 : sjinfo->min_righthand = adjust_child_relids(sjinfo->min_righthand,
1863 : right_nappinfos,
1864 : right_appinfos);
1865 3186 : sjinfo->syn_lefthand = adjust_child_relids(sjinfo->syn_lefthand,
1866 : left_nappinfos, left_appinfos);
1867 3186 : sjinfo->syn_righthand = adjust_child_relids(sjinfo->syn_righthand,
1868 : right_nappinfos,
1869 : right_appinfos);
1870 : /* outer-join relids need no adjustment */
1871 6372 : sjinfo->semi_rhs_exprs = (List *) adjust_appendrel_attrs(root,
1872 3186 : (Node *) sjinfo->semi_rhs_exprs,
1873 : right_nappinfos,
1874 : right_appinfos);
1875 :
1876 3186 : pfree(left_appinfos);
1877 3186 : pfree(right_appinfos);
1878 :
1879 3186 : return sjinfo;
1880 : }
1881 :
1882 : /*
1883 : * free_child_join_sjinfo
1884 : * Free memory consumed by a SpecialJoinInfo created by
1885 : * build_child_join_sjinfo()
1886 : *
1887 : * Only members that are translated copies of their counterpart in the parent
1888 : * SpecialJoinInfo are freed here.
1889 : */
1890 : static void
1891 19394 : free_child_join_sjinfo(SpecialJoinInfo *child_sjinfo,
1892 : SpecialJoinInfo *parent_sjinfo)
1893 : {
1894 : /*
1895 : * Dummy SpecialJoinInfos of inner joins do not have any translated fields
1896 : * and hence no fields that to be freed.
1897 : */
1898 19394 : if (child_sjinfo->jointype != JOIN_INNER)
1899 : {
1900 3186 : if (child_sjinfo->min_lefthand != parent_sjinfo->min_lefthand)
1901 3168 : bms_free(child_sjinfo->min_lefthand);
1902 :
1903 3186 : if (child_sjinfo->min_righthand != parent_sjinfo->min_righthand)
1904 3186 : bms_free(child_sjinfo->min_righthand);
1905 :
1906 3186 : if (child_sjinfo->syn_lefthand != parent_sjinfo->syn_lefthand)
1907 3186 : bms_free(child_sjinfo->syn_lefthand);
1908 :
1909 3186 : if (child_sjinfo->syn_righthand != parent_sjinfo->syn_righthand)
1910 3186 : bms_free(child_sjinfo->syn_righthand);
1911 :
1912 : Assert(child_sjinfo->commute_above_l == parent_sjinfo->commute_above_l);
1913 : Assert(child_sjinfo->commute_above_r == parent_sjinfo->commute_above_r);
1914 : Assert(child_sjinfo->commute_below_l == parent_sjinfo->commute_below_l);
1915 : Assert(child_sjinfo->commute_below_r == parent_sjinfo->commute_below_r);
1916 :
1917 : Assert(child_sjinfo->semi_operators == parent_sjinfo->semi_operators);
1918 :
1919 : /*
1920 : * semi_rhs_exprs may in principle be freed, but a simple pfree() does
1921 : * not suffice, so we leave it alone.
1922 : */
1923 : }
1924 :
1925 19394 : pfree(child_sjinfo);
1926 19394 : }
1927 :
1928 : /*
1929 : * compute_partition_bounds
1930 : * Compute the partition bounds for a join rel from those for inputs
1931 : */
1932 : static void
1933 7766 : compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
1934 : RelOptInfo *rel2, RelOptInfo *joinrel,
1935 : SpecialJoinInfo *parent_sjinfo,
1936 : List **parts1, List **parts2)
1937 : {
1938 : /*
1939 : * If we don't have the partition bounds for the join rel yet, try to
1940 : * compute those along with pairs of partitions to be joined.
1941 : */
1942 7766 : if (joinrel->nparts == -1)
1943 : {
1944 7386 : PartitionScheme part_scheme = joinrel->part_scheme;
1945 7386 : PartitionBoundInfo boundinfo = NULL;
1946 7386 : int nparts = 0;
1947 :
1948 : Assert(joinrel->boundinfo == NULL);
1949 : Assert(joinrel->part_rels == NULL);
1950 :
1951 : /*
1952 : * See if the partition bounds for inputs are exactly the same, in
1953 : * which case we don't need to work hard: the join rel will have the
1954 : * same partition bounds as inputs, and the partitions with the same
1955 : * cardinal positions will form the pairs.
1956 : *
1957 : * Note: even in cases where one or both inputs have merged bounds, it
1958 : * would be possible for both the bounds to be exactly the same, but
1959 : * it seems unlikely to be worth the cycles to check.
1960 : */
1961 7386 : if (!rel1->partbounds_merged &&
1962 7326 : !rel2->partbounds_merged &&
1963 14394 : rel1->nparts == rel2->nparts &&
1964 7068 : partition_bounds_equal(part_scheme->partnatts,
1965 : part_scheme->parttyplen,
1966 : part_scheme->parttypbyval,
1967 : rel1->boundinfo, rel2->boundinfo))
1968 : {
1969 6540 : boundinfo = rel1->boundinfo;
1970 6540 : nparts = rel1->nparts;
1971 : }
1972 : else
1973 : {
1974 : /* Try merging the partition bounds for inputs. */
1975 846 : boundinfo = partition_bounds_merge(part_scheme->partnatts,
1976 846 : part_scheme->partsupfunc,
1977 : part_scheme->partcollation,
1978 : rel1, rel2,
1979 : parent_sjinfo->jointype,
1980 : parts1, parts2);
1981 846 : if (boundinfo == NULL)
1982 : {
1983 114 : joinrel->nparts = 0;
1984 114 : return;
1985 : }
1986 732 : nparts = list_length(*parts1);
1987 732 : joinrel->partbounds_merged = true;
1988 : }
1989 :
1990 : Assert(nparts > 0);
1991 7272 : joinrel->boundinfo = boundinfo;
1992 7272 : joinrel->nparts = nparts;
1993 7272 : joinrel->part_rels =
1994 7272 : (RelOptInfo **) palloc0(sizeof(RelOptInfo *) * nparts);
1995 : }
1996 : else
1997 : {
1998 : Assert(joinrel->nparts > 0);
1999 : Assert(joinrel->boundinfo);
2000 : Assert(joinrel->part_rels);
2001 :
2002 : /*
2003 : * If the join rel's partbounds_merged flag is true, it means inputs
2004 : * are not guaranteed to have the same partition bounds, therefore we
2005 : * can't assume that the partitions at the same cardinal positions
2006 : * form the pairs; let get_matching_part_pairs() generate the pairs.
2007 : * Otherwise, nothing to do since we can assume that.
2008 : */
2009 380 : if (joinrel->partbounds_merged)
2010 : {
2011 36 : get_matching_part_pairs(root, joinrel, rel1, rel2,
2012 : parts1, parts2);
2013 : Assert(list_length(*parts1) == joinrel->nparts);
2014 : Assert(list_length(*parts2) == joinrel->nparts);
2015 : }
2016 : }
2017 : }
2018 :
2019 : /*
2020 : * get_matching_part_pairs
2021 : * Generate pairs of partitions to be joined from inputs
2022 : */
2023 : static void
2024 36 : get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
2025 : RelOptInfo *rel1, RelOptInfo *rel2,
2026 : List **parts1, List **parts2)
2027 : {
2028 36 : bool rel1_is_simple = IS_SIMPLE_REL(rel1);
2029 36 : bool rel2_is_simple = IS_SIMPLE_REL(rel2);
2030 : int cnt_parts;
2031 :
2032 36 : *parts1 = NIL;
2033 36 : *parts2 = NIL;
2034 :
2035 132 : for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
2036 : {
2037 96 : RelOptInfo *child_joinrel = joinrel->part_rels[cnt_parts];
2038 : RelOptInfo *child_rel1;
2039 : RelOptInfo *child_rel2;
2040 : Relids child_relids1;
2041 : Relids child_relids2;
2042 :
2043 : /*
2044 : * If this segment of the join is empty, it means that this segment
2045 : * was ignored when previously creating child-join paths for it in
2046 : * try_partitionwise_join() as it would not contribute to the join
2047 : * result, due to one or both inputs being empty; add NULL to each of
2048 : * the given lists so that this segment will be ignored again in that
2049 : * function.
2050 : */
2051 96 : if (!child_joinrel)
2052 : {
2053 0 : *parts1 = lappend(*parts1, NULL);
2054 0 : *parts2 = lappend(*parts2, NULL);
2055 0 : continue;
2056 : }
2057 :
2058 : /*
2059 : * Get a relids set of partition(s) involved in this join segment that
2060 : * are from the rel1 side.
2061 : */
2062 96 : child_relids1 = bms_intersect(child_joinrel->relids,
2063 96 : rel1->all_partrels);
2064 : Assert(bms_num_members(child_relids1) == bms_num_members(rel1->relids));
2065 :
2066 : /*
2067 : * Get a child rel for rel1 with the relids. Note that we should have
2068 : * the child rel even if rel1 is a join rel, because in that case the
2069 : * partitions specified in the relids would have matching/overlapping
2070 : * boundaries, so the specified partitions should be considered as
2071 : * ones to be joined when planning partitionwise joins of rel1,
2072 : * meaning that the child rel would have been built by the time we get
2073 : * here.
2074 : */
2075 96 : if (rel1_is_simple)
2076 : {
2077 0 : int varno = bms_singleton_member(child_relids1);
2078 :
2079 0 : child_rel1 = find_base_rel(root, varno);
2080 : }
2081 : else
2082 96 : child_rel1 = find_join_rel(root, child_relids1);
2083 : Assert(child_rel1);
2084 :
2085 : /*
2086 : * Get a relids set of partition(s) involved in this join segment that
2087 : * are from the rel2 side.
2088 : */
2089 96 : child_relids2 = bms_intersect(child_joinrel->relids,
2090 96 : rel2->all_partrels);
2091 : Assert(bms_num_members(child_relids2) == bms_num_members(rel2->relids));
2092 :
2093 : /*
2094 : * Get a child rel for rel2 with the relids. See above comments.
2095 : */
2096 96 : if (rel2_is_simple)
2097 : {
2098 96 : int varno = bms_singleton_member(child_relids2);
2099 :
2100 96 : child_rel2 = find_base_rel(root, varno);
2101 : }
2102 : else
2103 0 : child_rel2 = find_join_rel(root, child_relids2);
2104 : Assert(child_rel2);
2105 :
2106 : /*
2107 : * The join of rel1 and rel2 is legal, so is the join of the child
2108 : * rels obtained above; add them to the given lists as a join pair
2109 : * producing this join segment.
2110 : */
2111 96 : *parts1 = lappend(*parts1, child_rel1);
2112 96 : *parts2 = lappend(*parts2, child_rel2);
2113 : }
2114 36 : }
|
