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/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 free_child_join_sjinfo(SpecialJoinInfo *child_sjinfo,
49 : SpecialJoinInfo *parent_sjinfo);
50 : static void compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
51 : RelOptInfo *rel2, RelOptInfo *joinrel,
52 : SpecialJoinInfo *parent_sjinfo,
53 : List **parts1, List **parts2);
54 : static void get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
55 : RelOptInfo *rel1, RelOptInfo *rel2,
56 : List **parts1, List **parts2);
57 :
58 :
59 : /*
60 : * join_search_one_level
61 : * Consider ways to produce join relations containing exactly 'level'
62 : * jointree items. (This is one step of the dynamic-programming method
63 : * embodied in standard_join_search.) Join rel nodes for each feasible
64 : * combination of lower-level rels are created and returned in a list.
65 : * Implementation paths are created for each such joinrel, too.
66 : *
67 : * level: level of rels we want to make this time
68 : * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
69 : *
70 : * The result is returned in root->join_rel_level[level].
71 : */
72 : void
73 134642 : join_search_one_level(PlannerInfo *root, int level)
74 : {
75 134642 : List **joinrels = root->join_rel_level;
76 : ListCell *r;
77 : int k;
78 :
79 : Assert(joinrels[level] == NIL);
80 :
81 : /* Set join_cur_level so that new joinrels are added to proper list */
82 134642 : root->join_cur_level = level;
83 :
84 : /*
85 : * First, consider left-sided and right-sided plans, in which rels of
86 : * exactly level-1 member relations are joined against initial relations.
87 : * We prefer to join using join clauses, but if we find a rel of level-1
88 : * members that has no join clauses, we will generate Cartesian-product
89 : * joins against all initial rels not already contained in it.
90 : */
91 478370 : foreach(r, joinrels[level - 1])
92 : {
93 343728 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
94 :
95 374594 : if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
96 30866 : has_join_restriction(root, old_rel))
97 333354 : {
98 : int first_rel;
99 :
100 : /*
101 : * There are join clauses or join order restrictions relevant to
102 : * this rel, so consider joins between this rel and (only) those
103 : * initial rels it is linked to by a clause or restriction.
104 : *
105 : * At level 2 this condition is symmetric, so there is no need to
106 : * look at initial rels before this one in the list; we already
107 : * considered such joins when we were at the earlier rel. (The
108 : * mirror-image joins are handled automatically by make_join_rel.)
109 : * In later passes (level > 2), we join rels of the previous level
110 : * to each initial rel they don't already include but have a join
111 : * clause or restriction with.
112 : */
113 333354 : if (level == 2) /* consider remaining initial rels */
114 223122 : first_rel = foreach_current_index(r) + 1;
115 : else
116 110232 : first_rel = 0;
117 :
118 333354 : make_rels_by_clause_joins(root, old_rel, joinrels[1], first_rel);
119 : }
120 : else
121 : {
122 : /*
123 : * Oops, we have a relation that is not joined to any other
124 : * relation, either directly or by join-order restrictions.
125 : * Cartesian product time.
126 : *
127 : * We consider a cartesian product with each not-already-included
128 : * initial rel, whether it has other join clauses or not. At
129 : * level 2, if there are two or more clauseless initial rels, we
130 : * will redundantly consider joining them in both directions; but
131 : * such cases aren't common enough to justify adding complexity to
132 : * avoid the duplicated effort.
133 : */
134 10374 : make_rels_by_clauseless_joins(root,
135 : old_rel,
136 10374 : joinrels[1]);
137 : }
138 : }
139 :
140 : /*
141 : * Now, consider "bushy plans" in which relations of k initial rels are
142 : * joined to relations of level-k initial rels, for 2 <= k <= level-2.
143 : *
144 : * We only consider bushy-plan joins for pairs of rels where there is a
145 : * suitable join clause (or join order restriction), in order to avoid
146 : * unreasonable growth of planning time.
147 : */
148 134642 : for (k = 2;; k++)
149 13338 : {
150 147980 : int other_level = level - k;
151 :
152 : /*
153 : * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
154 : * need to go as far as the halfway point.
155 : */
156 147980 : if (k > other_level)
157 134642 : break;
158 :
159 70070 : foreach(r, joinrels[k])
160 : {
161 56732 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
162 : int first_rel;
163 : ListCell *r2;
164 :
165 : /*
166 : * We can ignore relations without join clauses here, unless they
167 : * participate in join-order restrictions --- then we might have
168 : * to force a bushy join plan.
169 : */
170 56732 : if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
171 312 : !has_join_restriction(root, old_rel))
172 216 : continue;
173 :
174 56516 : if (k == other_level) /* only consider remaining rels */
175 39696 : first_rel = foreach_current_index(r) + 1;
176 : else
177 16820 : first_rel = 0;
178 :
179 253474 : for_each_from(r2, joinrels[other_level], first_rel)
180 : {
181 196958 : RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
182 :
183 196958 : if (!bms_overlap(old_rel->relids, new_rel->relids))
184 : {
185 : /*
186 : * OK, we can build a rel of the right level from this
187 : * pair of rels. Do so if there is at least one relevant
188 : * join clause or join order restriction.
189 : */
190 22186 : if (have_relevant_joinclause(root, old_rel, new_rel) ||
191 1118 : have_join_order_restriction(root, old_rel, new_rel))
192 : {
193 20004 : (void) make_join_rel(root, old_rel, new_rel);
194 : }
195 : }
196 : }
197 : }
198 : }
199 :
200 : /*----------
201 : * Last-ditch effort: if we failed to find any usable joins so far, force
202 : * a set of cartesian-product joins to be generated. This handles the
203 : * special case where all the available rels have join clauses but we
204 : * cannot use any of those clauses yet. This can only happen when we are
205 : * considering a join sub-problem (a sub-joinlist) and all the rels in the
206 : * sub-problem have only join clauses with rels outside the sub-problem.
207 : * An example is
208 : *
209 : * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
210 : * WHERE a.w = c.x and b.y = d.z;
211 : *
212 : * If the "a INNER JOIN b" sub-problem does not get flattened into the
213 : * upper level, we must be willing to make a cartesian join of a and b;
214 : * but the code above will not have done so, because it thought that both
215 : * a and b have joinclauses. We consider only left-sided and right-sided
216 : * cartesian joins in this case (no bushy).
217 : *----------
218 : */
219 134642 : if (joinrels[level] == NIL)
220 : {
221 : /*
222 : * This loop is just like the first one, except we always call
223 : * make_rels_by_clauseless_joins().
224 : */
225 54 : foreach(r, joinrels[level - 1])
226 : {
227 36 : RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
228 :
229 36 : make_rels_by_clauseless_joins(root,
230 : old_rel,
231 36 : joinrels[1]);
232 : }
233 :
234 : /*----------
235 : * When special joins are involved, there may be no legal way
236 : * to make an N-way join for some values of N. For example consider
237 : *
238 : * SELECT ... FROM t1 WHERE
239 : * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
240 : * y IN (SELECT ... FROM t4,t5 WHERE ...)
241 : *
242 : * We will flatten this query to a 5-way join problem, but there are
243 : * no 4-way joins that join_is_legal() will consider legal. We have
244 : * to accept failure at level 4 and go on to discover a workable
245 : * bushy plan at level 5.
246 : *
247 : * However, if there are no special joins and no lateral references
248 : * then join_is_legal() should never fail, and so the following sanity
249 : * check is useful.
250 : *----------
251 : */
252 18 : if (joinrels[level] == NIL &&
253 6 : root->join_info_list == NIL &&
254 0 : !root->hasLateralRTEs)
255 0 : elog(ERROR, "failed to build any %d-way joins", level);
256 : }
257 134642 : }
258 :
259 : /*
260 : * make_rels_by_clause_joins
261 : * Build joins between the given relation 'old_rel' and other relations
262 : * that participate in join clauses that 'old_rel' also participates in
263 : * (or participate in join-order restrictions with it).
264 : * The join rels are returned in root->join_rel_level[join_cur_level].
265 : *
266 : * Note: at levels above 2 we will generate the same joined relation in
267 : * multiple ways --- for example (a join b) join c is the same RelOptInfo as
268 : * (b join c) join a, though the second case will add a different set of Paths
269 : * to it. This is the reason for using the join_rel_level mechanism, which
270 : * automatically ensures that each new joinrel is only added to the list once.
271 : *
272 : * 'old_rel' is the relation entry for the relation to be joined
273 : * 'other_rels': a list containing the other rels to be considered for joining
274 : * 'first_rel_idx': the first rel to be considered in 'other_rels'
275 : *
276 : * Currently, this is only used with initial rels in other_rels, but it
277 : * will work for joining to joinrels too.
278 : */
279 : static void
280 333354 : make_rels_by_clause_joins(PlannerInfo *root,
281 : RelOptInfo *old_rel,
282 : List *other_rels,
283 : int first_rel_idx)
284 : {
285 : ListCell *l;
286 :
287 988142 : for_each_from(l, other_rels, first_rel_idx)
288 : {
289 654788 : RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
290 :
291 1025006 : if (!bms_overlap(old_rel->relids, other_rel->relids) &&
292 445342 : (have_relevant_joinclause(root, old_rel, other_rel) ||
293 75124 : have_join_order_restriction(root, old_rel, other_rel)))
294 : {
295 306934 : (void) make_join_rel(root, old_rel, other_rel);
296 : }
297 : }
298 333354 : }
299 :
300 : /*
301 : * make_rels_by_clauseless_joins
302 : * Given a relation 'old_rel' and a list of other relations
303 : * 'other_rels', create a join relation between 'old_rel' and each
304 : * member of 'other_rels' that isn't already included in 'old_rel'.
305 : * The join rels are returned in root->join_rel_level[join_cur_level].
306 : *
307 : * 'old_rel' is the relation entry for the relation to be joined
308 : * 'other_rels': a list containing the other rels to be considered for joining
309 : *
310 : * Currently, this is only used with initial rels in other_rels, but it would
311 : * work for joining to joinrels too.
312 : */
313 : static void
314 10410 : make_rels_by_clauseless_joins(PlannerInfo *root,
315 : RelOptInfo *old_rel,
316 : List *other_rels)
317 : {
318 : ListCell *l;
319 :
320 32958 : foreach(l, other_rels)
321 : {
322 22548 : RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
323 :
324 22548 : if (!bms_overlap(other_rel->relids, old_rel->relids))
325 : {
326 11132 : (void) make_join_rel(root, old_rel, other_rel);
327 : }
328 : }
329 10410 : }
330 :
331 :
332 : /*
333 : * join_is_legal
334 : * Determine whether a proposed join is legal given the query's
335 : * join order constraints; and if it is, determine the join type.
336 : *
337 : * Caller must supply not only the two rels, but the union of their relids.
338 : * (We could simplify the API by computing joinrelids locally, but this
339 : * would be redundant work in the normal path through make_join_rel.
340 : * Note that this value does NOT include the RT index of any outer join that
341 : * might need to be performed here, so it's not the canonical identifier
342 : * of the join relation.)
343 : *
344 : * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
345 : * else it's set to point to the associated SpecialJoinInfo node. Also,
346 : * *reversed_p is set true if the given relations need to be swapped to
347 : * match the SpecialJoinInfo node.
348 : */
349 : static bool
350 343212 : join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
351 : Relids joinrelids,
352 : SpecialJoinInfo **sjinfo_p, bool *reversed_p)
353 : {
354 : SpecialJoinInfo *match_sjinfo;
355 : bool reversed;
356 : bool unique_ified;
357 : bool must_be_leftjoin;
358 : ListCell *l;
359 :
360 : /*
361 : * Ensure output params are set on failure return. This is just to
362 : * suppress uninitialized-variable warnings from overly anal compilers.
363 : */
364 343212 : *sjinfo_p = NULL;
365 343212 : *reversed_p = false;
366 :
367 : /*
368 : * If we have any special joins, the proposed join might be illegal; and
369 : * in any case we have to determine its join type. Scan the join info
370 : * list for matches and conflicts.
371 : */
372 343212 : match_sjinfo = NULL;
373 343212 : reversed = false;
374 343212 : unique_ified = false;
375 343212 : must_be_leftjoin = false;
376 :
377 717672 : foreach(l, root->join_info_list)
378 : {
379 383696 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
380 :
381 : /*
382 : * This special join is not relevant unless its RHS overlaps the
383 : * proposed join. (Check this first as a fast path for dismissing
384 : * most irrelevant SJs quickly.)
385 : */
386 383696 : if (!bms_overlap(sjinfo->min_righthand, joinrelids))
387 129852 : continue;
388 :
389 : /*
390 : * Also, not relevant if proposed join is fully contained within RHS
391 : * (ie, we're still building up the RHS).
392 : */
393 253844 : if (bms_is_subset(joinrelids, sjinfo->min_righthand))
394 5252 : continue;
395 :
396 : /*
397 : * Also, not relevant if SJ is already done within either input.
398 : */
399 466720 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
400 218128 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
401 107244 : continue;
402 161280 : if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
403 19932 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
404 10792 : continue;
405 :
406 : /*
407 : * If it's a semijoin and we already joined the RHS to any other rels
408 : * within either input, then we must have unique-ified the RHS at that
409 : * point (see below). Therefore the semijoin is no longer relevant in
410 : * this join path.
411 : */
412 130556 : if (sjinfo->jointype == JOIN_SEMI)
413 : {
414 10718 : if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
415 1552 : !bms_equal(sjinfo->syn_righthand, rel1->relids))
416 618 : continue;
417 10100 : if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
418 7030 : !bms_equal(sjinfo->syn_righthand, rel2->relids))
419 214 : continue;
420 : }
421 :
422 : /*
423 : * If one input contains min_lefthand and the other contains
424 : * min_righthand, then we can perform the SJ at this join.
425 : *
426 : * Reject if we get matches to more than one SJ; that implies we're
427 : * considering something that's not really valid.
428 : */
429 240456 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
430 110732 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
431 : {
432 104120 : if (match_sjinfo)
433 9236 : return false; /* invalid join path */
434 104120 : match_sjinfo = sjinfo;
435 104120 : reversed = false;
436 : }
437 34406 : else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
438 8802 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
439 : {
440 7578 : if (match_sjinfo)
441 0 : return false; /* invalid join path */
442 7578 : match_sjinfo = sjinfo;
443 7578 : reversed = true;
444 : }
445 21018 : else if (sjinfo->jointype == JOIN_SEMI &&
446 3558 : bms_equal(sjinfo->syn_righthand, rel2->relids) &&
447 566 : create_unique_path(root, rel2, rel2->cheapest_total_path,
448 : sjinfo) != NULL)
449 : {
450 : /*----------
451 : * For a semijoin, we can join the RHS to anything else by
452 : * unique-ifying the RHS (if the RHS can be unique-ified).
453 : * We will only get here if we have the full RHS but less
454 : * than min_lefthand on the LHS.
455 : *
456 : * The reason to consider such a join path is exemplified by
457 : * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
458 : * If we insist on doing this as a semijoin we will first have
459 : * to form the cartesian product of A*B. But if we unique-ify
460 : * C then the semijoin becomes a plain innerjoin and we can join
461 : * in any order, eg C to A and then to B. When C is much smaller
462 : * than A and B this can be a huge win. So we allow C to be
463 : * joined to just A or just B here, and then make_join_rel has
464 : * to handle the case properly.
465 : *
466 : * Note that actually we'll allow unique-ified C to be joined to
467 : * some other relation D here, too. That is legal, if usually not
468 : * very sane, and this routine is only concerned with legality not
469 : * with whether the join is good strategy.
470 : *----------
471 : */
472 560 : if (match_sjinfo)
473 84 : return false; /* invalid join path */
474 476 : match_sjinfo = sjinfo;
475 476 : reversed = false;
476 476 : unique_ified = true;
477 : }
478 19898 : else if (sjinfo->jointype == JOIN_SEMI &&
479 2722 : bms_equal(sjinfo->syn_righthand, rel1->relids) &&
480 290 : create_unique_path(root, rel1, rel1->cheapest_total_path,
481 : sjinfo) != NULL)
482 : {
483 : /* Reversed semijoin case */
484 290 : if (match_sjinfo)
485 78 : return false; /* invalid join path */
486 212 : match_sjinfo = sjinfo;
487 212 : reversed = true;
488 212 : unique_ified = true;
489 : }
490 : else
491 : {
492 : /*
493 : * Otherwise, the proposed join overlaps the RHS but isn't a valid
494 : * implementation of this SJ. But don't panic quite yet: the RHS
495 : * violation might have occurred previously, in one or both input
496 : * relations, in which case we must have previously decided that
497 : * it was OK to commute some other SJ with this one. If we need
498 : * to perform this join to finish building up the RHS, rejecting
499 : * it could lead to not finding any plan at all. (This can occur
500 : * because of the heuristics elsewhere in this file that postpone
501 : * clauseless joins: we might not consider doing a clauseless join
502 : * within the RHS until after we've performed other, validly
503 : * commutable SJs with one or both sides of the clauseless join.)
504 : * This consideration boils down to the rule that if both inputs
505 : * overlap the RHS, we can allow the join --- they are either
506 : * fully within the RHS, or represent previously-allowed joins to
507 : * rels outside it.
508 : */
509 25076 : if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
510 7900 : bms_overlap(rel2->relids, sjinfo->min_righthand))
511 174 : continue; /* assume valid previous violation of RHS */
512 :
513 : /*
514 : * The proposed join could still be legal, but only if we're
515 : * allowed to associate it into the RHS of this SJ. That means
516 : * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
517 : * not FULL) and the proposed join must not overlap the LHS.
518 : */
519 31730 : if (sjinfo->jointype != JOIN_LEFT ||
520 14728 : bms_overlap(joinrelids, sjinfo->min_lefthand))
521 9074 : return false; /* invalid join path */
522 :
523 : /*
524 : * To be valid, the proposed join must be a LEFT join; otherwise
525 : * it can't associate into this SJ's RHS. But we may not yet have
526 : * found the SpecialJoinInfo matching the proposed join, so we
527 : * can't test that yet. Remember the requirement for later.
528 : */
529 7928 : must_be_leftjoin = true;
530 : }
531 : }
532 :
533 : /*
534 : * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
535 : * proposed join can't associate into an SJ's RHS.
536 : *
537 : * Also, fail if the proposed join's predicate isn't strict; we're
538 : * essentially checking to see if we can apply outer-join identity 3, and
539 : * that's a requirement. (This check may be redundant with checks in
540 : * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
541 : */
542 333976 : if (must_be_leftjoin &&
543 4934 : (match_sjinfo == NULL ||
544 4934 : match_sjinfo->jointype != JOIN_LEFT ||
545 4934 : !match_sjinfo->lhs_strict))
546 1512 : return false; /* invalid join path */
547 :
548 : /*
549 : * We also have to check for constraints imposed by LATERAL references.
550 : */
551 332464 : if (root->hasLateralRTEs)
552 : {
553 : bool lateral_fwd;
554 : bool lateral_rev;
555 : Relids join_lateral_rels;
556 :
557 : /*
558 : * The proposed rels could each contain lateral references to the
559 : * other, in which case the join is impossible. If there are lateral
560 : * references in just one direction, then the join has to be done with
561 : * a nestloop with the lateral referencer on the inside. If the join
562 : * matches an SJ that cannot be implemented by such a nestloop, the
563 : * join is impossible.
564 : *
565 : * Also, if the lateral reference is only indirect, we should reject
566 : * the join; whatever rel(s) the reference chain goes through must be
567 : * joined to first.
568 : *
569 : * Another case that might keep us from building a valid plan is the
570 : * implementation restriction described by have_dangerous_phv().
571 : */
572 16850 : lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
573 16850 : lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
574 16850 : if (lateral_fwd && lateral_rev)
575 18 : return false; /* have lateral refs in both directions */
576 16832 : if (lateral_fwd)
577 : {
578 : /* has to be implemented as nestloop with rel1 on left */
579 10572 : if (match_sjinfo &&
580 408 : (reversed ||
581 390 : unique_ified ||
582 390 : 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 10554 : if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
586 42 : return false; /* only indirect refs, so reject */
587 : /* check we won't have a dangerous PHV */
588 10512 : if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
589 72 : return false; /* might be unable to handle required PHV */
590 : }
591 6260 : else if (lateral_rev)
592 : {
593 : /* has to be implemented as nestloop with rel2 on left */
594 1196 : if (match_sjinfo &&
595 72 : (!reversed ||
596 72 : unique_ified ||
597 72 : match_sjinfo->jointype == JOIN_FULL))
598 0 : return false; /* not implementable as nestloop */
599 : /* check there is a direct reference from rel1 to rel2 */
600 1196 : if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
601 0 : return false; /* only indirect refs, so reject */
602 : /* check we won't have a dangerous PHV */
603 1196 : if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
604 84 : return false; /* might be unable to handle required PHV */
605 : }
606 :
607 : /*
608 : * LATERAL references could also cause problems later on if we accept
609 : * this join: if the join's minimum parameterization includes any rels
610 : * that would have to be on the inside of an outer join with this join
611 : * rel, then it's never going to be possible to build the complete
612 : * query using this join. We should reject this join not only because
613 : * it'll save work, but because if we don't, the clauseless-join
614 : * heuristics might think that legality of this join means that some
615 : * other join rel need not be formed, and that could lead to failure
616 : * to find any plan at all. We have to consider not only rels that
617 : * are directly on the inner side of an OJ with the joinrel, but also
618 : * ones that are indirectly so, so search to find all such rels.
619 : */
620 16616 : join_lateral_rels = min_join_parameterization(root, joinrelids,
621 : rel1, rel2);
622 16616 : if (join_lateral_rels)
623 : {
624 1704 : Relids join_plus_rhs = bms_copy(joinrelids);
625 : bool more;
626 :
627 : do
628 : {
629 2100 : more = false;
630 4008 : foreach(l, root->join_info_list)
631 : {
632 1908 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
633 :
634 : /* ignore full joins --- their ordering is predetermined */
635 1908 : if (sjinfo->jointype == JOIN_FULL)
636 18 : continue;
637 :
638 1890 : if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
639 1596 : !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
640 : {
641 546 : join_plus_rhs = bms_add_members(join_plus_rhs,
642 546 : sjinfo->min_righthand);
643 546 : more = true;
644 : }
645 : }
646 2100 : } while (more);
647 1704 : if (bms_overlap(join_plus_rhs, join_lateral_rels))
648 312 : return false; /* will not be able to join to some RHS rel */
649 : }
650 : }
651 :
652 : /* Otherwise, it's a valid join */
653 331918 : *sjinfo_p = match_sjinfo;
654 331918 : *reversed_p = reversed;
655 331918 : return true;
656 : }
657 :
658 : /*
659 : * init_dummy_sjinfo
660 : * Populate the given SpecialJoinInfo for a plain inner join between the
661 : * left and right relations specified by left_relids and right_relids
662 : * respectively.
663 : *
664 : * Normally, an inner join does not have a SpecialJoinInfo node associated with
665 : * it. But some functions involved in join planning require one containing at
666 : * least the information of which relations are being joined. So we initialize
667 : * that information here.
668 : */
669 : void
670 935488 : init_dummy_sjinfo(SpecialJoinInfo *sjinfo, Relids left_relids,
671 : Relids right_relids)
672 : {
673 935488 : sjinfo->type = T_SpecialJoinInfo;
674 935488 : sjinfo->min_lefthand = left_relids;
675 935488 : sjinfo->min_righthand = right_relids;
676 935488 : sjinfo->syn_lefthand = left_relids;
677 935488 : sjinfo->syn_righthand = right_relids;
678 935488 : sjinfo->jointype = JOIN_INNER;
679 935488 : sjinfo->ojrelid = 0;
680 935488 : sjinfo->commute_above_l = NULL;
681 935488 : sjinfo->commute_above_r = NULL;
682 935488 : sjinfo->commute_below_l = NULL;
683 935488 : sjinfo->commute_below_r = NULL;
684 : /* we don't bother trying to make the remaining fields valid */
685 935488 : sjinfo->lhs_strict = false;
686 935488 : sjinfo->semi_can_btree = false;
687 935488 : sjinfo->semi_can_hash = false;
688 935488 : sjinfo->semi_operators = NIL;
689 935488 : sjinfo->semi_rhs_exprs = NIL;
690 935488 : }
691 :
692 : /*
693 : * make_join_rel
694 : * Find or create a join RelOptInfo that represents the join of
695 : * the two given rels, and add to it path information for paths
696 : * created with the two rels as outer and inner rel.
697 : * (The join rel may already contain paths generated from other
698 : * pairs of rels that add up to the same set of base rels.)
699 : *
700 : * NB: will return NULL if attempted join is not valid. This can happen
701 : * when working with outer joins, or with IN or EXISTS clauses that have been
702 : * turned into joins.
703 : */
704 : RelOptInfo *
705 342822 : make_join_rel(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2)
706 : {
707 : Relids joinrelids;
708 : SpecialJoinInfo *sjinfo;
709 : bool reversed;
710 342822 : List *pushed_down_joins = NIL;
711 : SpecialJoinInfo sjinfo_data;
712 : RelOptInfo *joinrel;
713 : List *restrictlist;
714 :
715 : /* We should never try to join two overlapping sets of rels. */
716 : Assert(!bms_overlap(rel1->relids, rel2->relids));
717 :
718 : /* Construct Relids set that identifies the joinrel (without OJ as yet). */
719 342822 : joinrelids = bms_union(rel1->relids, rel2->relids);
720 :
721 : /* Check validity and determine join type. */
722 342822 : if (!join_is_legal(root, rel1, rel2, joinrelids,
723 : &sjinfo, &reversed))
724 : {
725 : /* invalid join path */
726 11138 : bms_free(joinrelids);
727 11138 : return NULL;
728 : }
729 :
730 : /*
731 : * Add outer join relid(s) to form the canonical relids. Any added outer
732 : * joins besides sjinfo itself are appended to pushed_down_joins.
733 : */
734 331684 : joinrelids = add_outer_joins_to_relids(root, joinrelids, sjinfo,
735 : &pushed_down_joins);
736 :
737 : /* Swap rels if needed to match the join info. */
738 331684 : if (reversed)
739 : {
740 7634 : RelOptInfo *trel = rel1;
741 :
742 7634 : rel1 = rel2;
743 7634 : rel2 = trel;
744 : }
745 :
746 : /*
747 : * If it's a plain inner join, then we won't have found anything in
748 : * join_info_list. Make up a SpecialJoinInfo so that selectivity
749 : * estimation functions will know what's being joined.
750 : */
751 331684 : if (sjinfo == NULL)
752 : {
753 219862 : sjinfo = &sjinfo_data;
754 219862 : init_dummy_sjinfo(sjinfo, rel1->relids, rel2->relids);
755 : }
756 :
757 : /*
758 : * Find or build the join RelOptInfo, and compute the restrictlist that
759 : * goes with this particular joining.
760 : */
761 331684 : joinrel = build_join_rel(root, joinrelids, rel1, rel2,
762 : sjinfo, pushed_down_joins,
763 : &restrictlist);
764 :
765 : /*
766 : * If we've already proven this join is empty, we needn't consider any
767 : * more paths for it.
768 : */
769 331684 : if (is_dummy_rel(joinrel))
770 : {
771 456 : bms_free(joinrelids);
772 456 : return joinrel;
773 : }
774 :
775 : /* Add paths to the join relation. */
776 331228 : populate_joinrel_with_paths(root, rel1, rel2, joinrel, sjinfo,
777 : restrictlist);
778 :
779 331228 : bms_free(joinrelids);
780 :
781 331228 : return joinrel;
782 : }
783 :
784 : /*
785 : * add_outer_joins_to_relids
786 : * Add relids to input_relids to represent any outer joins that will be
787 : * calculated at this join.
788 : *
789 : * input_relids is the union of the relid sets of the two input relations.
790 : * Note that we modify this in-place and return it; caller must bms_copy()
791 : * it first, if a separate value is desired.
792 : *
793 : * sjinfo represents the join being performed.
794 : *
795 : * If the current join completes the calculation of any outer joins that
796 : * have been pushed down per outer-join identity 3, those relids will be
797 : * added to the result along with sjinfo's own relid. If pushed_down_joins
798 : * is not NULL, then also the SpecialJoinInfos for such added outer joins will
799 : * be appended to *pushed_down_joins (so caller must initialize it to NIL).
800 : */
801 : Relids
802 339018 : add_outer_joins_to_relids(PlannerInfo *root, Relids input_relids,
803 : SpecialJoinInfo *sjinfo,
804 : List **pushed_down_joins)
805 : {
806 : /* Nothing to do if this isn't an outer join with an assigned relid. */
807 339018 : if (sjinfo == NULL || sjinfo->ojrelid == 0)
808 238356 : return input_relids;
809 :
810 : /*
811 : * If it's not a left join, we have no rules that would permit executing
812 : * it in non-syntactic order, so just form the syntactic relid set. (This
813 : * is just a quick-exit test; we'd come to the same conclusion anyway,
814 : * since its commute_below_l and commute_above_l sets must be empty.)
815 : */
816 100662 : if (sjinfo->jointype != JOIN_LEFT)
817 2260 : return bms_add_member(input_relids, sjinfo->ojrelid);
818 :
819 : /*
820 : * We cannot add the OJ relid if this join has been pushed into the RHS of
821 : * a syntactically-lower left join per OJ identity 3. (If it has, then we
822 : * cannot claim that its outputs represent the final state of its RHS.)
823 : * There will not be any other OJs that can be added either, so we're
824 : * done.
825 : */
826 98402 : if (!bms_is_subset(sjinfo->commute_below_l, input_relids))
827 4634 : return input_relids;
828 :
829 : /* OK to add OJ's own relid */
830 93768 : input_relids = bms_add_member(input_relids, sjinfo->ojrelid);
831 :
832 : /*
833 : * Contrariwise, if we are now forming the final result of such a commuted
834 : * pair of OJs, it's time to add the relid(s) of the pushed-down join(s).
835 : * We can skip this if this join was never a candidate to be pushed up.
836 : */
837 93768 : if (sjinfo->commute_above_l)
838 : {
839 14896 : Relids commute_above_rels = bms_copy(sjinfo->commute_above_l);
840 : ListCell *lc;
841 :
842 : /*
843 : * The current join could complete the nulling of more than one
844 : * pushed-down join, so we have to examine all the SpecialJoinInfos.
845 : * Because join_info_list was built in bottom-up order, it's
846 : * sufficient to traverse it once: an ojrelid we add in one loop
847 : * iteration would not have affected decisions of earlier iterations.
848 : */
849 47964 : foreach(lc, root->join_info_list)
850 : {
851 33068 : SpecialJoinInfo *othersj = (SpecialJoinInfo *) lfirst(lc);
852 :
853 33068 : if (othersj == sjinfo ||
854 18172 : othersj->ojrelid == 0 || othersj->jointype != JOIN_LEFT)
855 14908 : continue; /* definitely not interesting */
856 :
857 18160 : if (!bms_is_member(othersj->ojrelid, commute_above_rels))
858 3116 : continue;
859 :
860 : /* Add it if not already present but conditions now satisfied */
861 30088 : if (!bms_is_member(othersj->ojrelid, input_relids) &&
862 30064 : bms_is_subset(othersj->min_lefthand, input_relids) &&
863 22618 : bms_is_subset(othersj->min_righthand, input_relids) &&
864 7598 : bms_is_subset(othersj->commute_below_l, input_relids))
865 : {
866 7562 : input_relids = bms_add_member(input_relids, othersj->ojrelid);
867 : /* report such pushed down outer joins, if asked */
868 7562 : if (pushed_down_joins != NULL)
869 7562 : *pushed_down_joins = lappend(*pushed_down_joins, othersj);
870 :
871 : /*
872 : * We must also check any joins that othersj potentially
873 : * commutes with. They likewise must appear later in
874 : * join_info_list than othersj itself, so we can visit them
875 : * later in this loop.
876 : */
877 7562 : commute_above_rels = bms_add_members(commute_above_rels,
878 7562 : othersj->commute_above_l);
879 : }
880 : }
881 : }
882 :
883 93768 : return input_relids;
884 : }
885 :
886 : /*
887 : * populate_joinrel_with_paths
888 : * Add paths to the given joinrel for given pair of joining relations. The
889 : * SpecialJoinInfo provides details about the join and the restrictlist
890 : * contains the join clauses and the other clauses applicable for given pair
891 : * of the joining relations.
892 : */
893 : static void
894 336654 : populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
895 : RelOptInfo *rel2, RelOptInfo *joinrel,
896 : SpecialJoinInfo *sjinfo, List *restrictlist)
897 : {
898 : /*
899 : * Consider paths using each rel as both outer and inner. Depending on
900 : * the join type, a provably empty outer or inner rel might mean the join
901 : * is provably empty too; in which case throw away any previously computed
902 : * paths and mark the join as dummy. (We do it this way since it's
903 : * conceivable that dummy-ness of a multi-element join might only be
904 : * noticeable for certain construction paths.)
905 : *
906 : * Also, a provably constant-false join restriction typically means that
907 : * we can skip evaluating one or both sides of the join. We do this by
908 : * marking the appropriate rel as dummy. For outer joins, a
909 : * constant-false restriction that is pushed down still means the whole
910 : * join is dummy, while a non-pushed-down one means that no inner rows
911 : * will join so we can treat the inner rel as dummy.
912 : *
913 : * We need only consider the jointypes that appear in join_info_list, plus
914 : * JOIN_INNER.
915 : */
916 336654 : switch (sjinfo->jointype)
917 : {
918 221754 : case JOIN_INNER:
919 443490 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
920 221736 : restriction_is_constant_false(restrictlist, joinrel, false))
921 : {
922 198 : mark_dummy_rel(joinrel);
923 198 : break;
924 : }
925 221556 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
926 : JOIN_INNER, sjinfo,
927 : restrictlist);
928 221556 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
929 : JOIN_INNER, sjinfo,
930 : restrictlist);
931 221556 : break;
932 100068 : case JOIN_LEFT:
933 200082 : if (is_dummy_rel(rel1) ||
934 100014 : restriction_is_constant_false(restrictlist, joinrel, true))
935 : {
936 86 : mark_dummy_rel(joinrel);
937 86 : break;
938 : }
939 100174 : if (restriction_is_constant_false(restrictlist, joinrel, false) &&
940 192 : bms_is_subset(rel2->relids, sjinfo->syn_righthand))
941 168 : mark_dummy_rel(rel2);
942 99982 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
943 : JOIN_LEFT, sjinfo,
944 : restrictlist);
945 99982 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
946 : JOIN_RIGHT, sjinfo,
947 : restrictlist);
948 99982 : break;
949 1732 : case JOIN_FULL:
950 3464 : if ((is_dummy_rel(rel1) && is_dummy_rel(rel2)) ||
951 1732 : restriction_is_constant_false(restrictlist, joinrel, true))
952 : {
953 12 : mark_dummy_rel(joinrel);
954 12 : break;
955 : }
956 1720 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
957 : JOIN_FULL, sjinfo,
958 : restrictlist);
959 1720 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
960 : JOIN_FULL, sjinfo,
961 : restrictlist);
962 :
963 : /*
964 : * If there are join quals that aren't mergeable or hashable, we
965 : * may not be able to build any valid plan. Complain here so that
966 : * we can give a somewhat-useful error message. (Since we have no
967 : * flexibility of planning for a full join, there's no chance of
968 : * succeeding later with another pair of input rels.)
969 : */
970 1720 : if (joinrel->pathlist == NIL)
971 0 : ereport(ERROR,
972 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
973 : errmsg("FULL JOIN is only supported with merge-joinable or hash-joinable join conditions")));
974 1720 : break;
975 7648 : case JOIN_SEMI:
976 :
977 : /*
978 : * We might have a normal semijoin, or a case where we don't have
979 : * enough rels to do the semijoin but can unique-ify the RHS and
980 : * then do an innerjoin (see comments in join_is_legal). In the
981 : * latter case we can't apply JOIN_SEMI joining.
982 : */
983 14974 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
984 7326 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
985 : {
986 14646 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
987 7320 : restriction_is_constant_false(restrictlist, joinrel, false))
988 : {
989 12 : mark_dummy_rel(joinrel);
990 12 : break;
991 : }
992 7314 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
993 : JOIN_SEMI, sjinfo,
994 : restrictlist);
995 7314 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
996 : JOIN_RIGHT_SEMI, sjinfo,
997 : restrictlist);
998 : }
999 :
1000 : /*
1001 : * If we know how to unique-ify the RHS and one input rel is
1002 : * exactly the RHS (not a superset) we can consider unique-ifying
1003 : * it and then doing a regular join. (The create_unique_path
1004 : * check here is probably redundant with what join_is_legal did,
1005 : * but if so the check is cheap because it's cached. So test
1006 : * anyway to be sure.)
1007 : */
1008 15272 : if (bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1009 7636 : create_unique_path(root, rel2, rel2->cheapest_total_path,
1010 : sjinfo) != NULL)
1011 : {
1012 15044 : if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
1013 7522 : restriction_is_constant_false(restrictlist, joinrel, false))
1014 : {
1015 0 : mark_dummy_rel(joinrel);
1016 0 : break;
1017 : }
1018 7522 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1019 : JOIN_UNIQUE_INNER, sjinfo,
1020 : restrictlist);
1021 7522 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1022 : JOIN_UNIQUE_OUTER, sjinfo,
1023 : restrictlist);
1024 : }
1025 7636 : break;
1026 5452 : case JOIN_ANTI:
1027 10904 : if (is_dummy_rel(rel1) ||
1028 5452 : restriction_is_constant_false(restrictlist, joinrel, true))
1029 : {
1030 0 : mark_dummy_rel(joinrel);
1031 0 : break;
1032 : }
1033 5452 : if (restriction_is_constant_false(restrictlist, joinrel, false) &&
1034 0 : bms_is_subset(rel2->relids, sjinfo->syn_righthand))
1035 0 : mark_dummy_rel(rel2);
1036 5452 : add_paths_to_joinrel(root, joinrel, rel1, rel2,
1037 : JOIN_ANTI, sjinfo,
1038 : restrictlist);
1039 5452 : add_paths_to_joinrel(root, joinrel, rel2, rel1,
1040 : JOIN_RIGHT_ANTI, sjinfo,
1041 : restrictlist);
1042 5452 : break;
1043 0 : default:
1044 : /* other values not expected here */
1045 0 : elog(ERROR, "unrecognized join type: %d", (int) sjinfo->jointype);
1046 : break;
1047 : }
1048 :
1049 : /* Apply partitionwise join technique, if possible. */
1050 336654 : try_partitionwise_join(root, rel1, rel2, joinrel, sjinfo, restrictlist);
1051 336654 : }
1052 :
1053 :
1054 : /*
1055 : * have_join_order_restriction
1056 : * Detect whether the two relations should be joined to satisfy
1057 : * a join-order restriction arising from special or lateral joins.
1058 : *
1059 : * In practice this is always used with have_relevant_joinclause(), and so
1060 : * could be merged with that function, but it seems clearer to separate the
1061 : * two concerns. We need this test because there are degenerate cases where
1062 : * a clauseless join must be performed to satisfy join-order restrictions.
1063 : * Also, if one rel has a lateral reference to the other, or both are needed
1064 : * to compute some PHV, we should consider joining them even if the join would
1065 : * be clauseless.
1066 : *
1067 : * Note: this is only a problem if one side of a degenerate outer join
1068 : * contains multiple rels, or a clauseless join is required within an
1069 : * IN/EXISTS RHS; else we will find a join path via the "last ditch" case in
1070 : * join_search_one_level(). We could dispense with this test if we were
1071 : * willing to try bushy plans in the "last ditch" case, but that seems much
1072 : * less efficient.
1073 : */
1074 : bool
1075 78522 : have_join_order_restriction(PlannerInfo *root,
1076 : RelOptInfo *rel1, RelOptInfo *rel2)
1077 : {
1078 78522 : bool result = false;
1079 : ListCell *l;
1080 :
1081 : /*
1082 : * If either side has a direct lateral reference to the other, attempt the
1083 : * join regardless of outer-join considerations.
1084 : */
1085 147228 : if (bms_overlap(rel1->relids, rel2->direct_lateral_relids) ||
1086 68706 : bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1087 10664 : return true;
1088 :
1089 : /*
1090 : * Likewise, if both rels are needed to compute some PlaceHolderVar,
1091 : * attempt the join regardless of outer-join considerations. (This is not
1092 : * very desirable, because a PHV with a large eval_at set will cause a lot
1093 : * of probably-useless joins to be considered, but failing to do this can
1094 : * cause us to fail to construct a plan at all.)
1095 : */
1096 69512 : foreach(l, root->placeholder_list)
1097 : {
1098 1708 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1099 :
1100 2080 : if (bms_is_subset(rel1->relids, phinfo->ph_eval_at) &&
1101 372 : bms_is_subset(rel2->relids, phinfo->ph_eval_at))
1102 54 : return true;
1103 : }
1104 :
1105 : /*
1106 : * It's possible that the rels correspond to the left and right sides of a
1107 : * degenerate outer join, that is, one with no joinclause mentioning the
1108 : * non-nullable side; in which case we should force the join to occur.
1109 : *
1110 : * Also, the two rels could represent a clauseless join that has to be
1111 : * completed to build up the LHS or RHS of an outer join.
1112 : */
1113 176070 : foreach(l, root->join_info_list)
1114 : {
1115 109676 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1116 :
1117 : /* ignore full joins --- other mechanisms handle them */
1118 109676 : if (sjinfo->jointype == JOIN_FULL)
1119 42 : continue;
1120 :
1121 : /* Can we perform the SJ with these rels? */
1122 131722 : if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1123 22088 : bms_is_subset(sjinfo->min_righthand, rel2->relids))
1124 : {
1125 1056 : result = true;
1126 1056 : break;
1127 : }
1128 114788 : if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1129 6210 : bms_is_subset(sjinfo->min_righthand, rel1->relids))
1130 : {
1131 204 : result = true;
1132 204 : break;
1133 : }
1134 :
1135 : /*
1136 : * Might we need to join these rels to complete the RHS? We have to
1137 : * use "overlap" tests since either rel might include a lower SJ that
1138 : * has been proven to commute with this one.
1139 : */
1140 129684 : if (bms_overlap(sjinfo->min_righthand, rel1->relids) &&
1141 21310 : bms_overlap(sjinfo->min_righthand, rel2->relids))
1142 : {
1143 120 : result = true;
1144 120 : break;
1145 : }
1146 :
1147 : /* Likewise for the LHS. */
1148 132222 : if (bms_overlap(sjinfo->min_lefthand, rel1->relids) &&
1149 23968 : bms_overlap(sjinfo->min_lefthand, rel2->relids))
1150 : {
1151 30 : result = true;
1152 30 : break;
1153 : }
1154 : }
1155 :
1156 : /*
1157 : * We do not force the join to occur if either input rel can legally be
1158 : * joined to anything else using joinclauses. This essentially means that
1159 : * clauseless bushy joins are put off as long as possible. The reason is
1160 : * that when there is a join order restriction high up in the join tree
1161 : * (that is, with many rels inside the LHS or RHS), we would otherwise
1162 : * expend lots of effort considering very stupid join combinations within
1163 : * its LHS or RHS.
1164 : */
1165 67804 : if (result)
1166 : {
1167 2700 : if (has_legal_joinclause(root, rel1) ||
1168 1290 : has_legal_joinclause(root, rel2))
1169 234 : result = false;
1170 : }
1171 :
1172 67804 : return result;
1173 : }
1174 :
1175 :
1176 : /*
1177 : * has_join_restriction
1178 : * Detect whether the specified relation has join-order restrictions,
1179 : * due to being inside an outer join or an IN (sub-SELECT),
1180 : * or participating in any LATERAL references or multi-rel PHVs.
1181 : *
1182 : * Essentially, this tests whether have_join_order_restriction() could
1183 : * succeed with this rel and some other one. It's OK if we sometimes
1184 : * say "true" incorrectly. (Therefore, we don't bother with the relatively
1185 : * expensive has_legal_joinclause test.)
1186 : */
1187 : static bool
1188 31178 : has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1189 : {
1190 : ListCell *l;
1191 :
1192 31178 : if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1193 18440 : return true;
1194 :
1195 13476 : foreach(l, root->placeholder_list)
1196 : {
1197 786 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1198 :
1199 786 : if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1200 174 : !bms_equal(rel->relids, phinfo->ph_eval_at))
1201 48 : return true;
1202 : }
1203 :
1204 13484 : foreach(l, root->join_info_list)
1205 : {
1206 2894 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1207 :
1208 : /* ignore full joins --- other mechanisms preserve their ordering */
1209 2894 : if (sjinfo->jointype == JOIN_FULL)
1210 86 : continue;
1211 :
1212 : /* ignore if SJ is already contained in rel */
1213 4332 : if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1214 1524 : bms_is_subset(sjinfo->min_righthand, rel->relids))
1215 372 : continue;
1216 :
1217 : /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1218 3696 : if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1219 1260 : bms_overlap(sjinfo->min_righthand, rel->relids))
1220 2100 : return true;
1221 : }
1222 :
1223 10590 : return false;
1224 : }
1225 :
1226 :
1227 : /*
1228 : * has_legal_joinclause
1229 : * Detect whether the specified relation can legally be joined
1230 : * to any other rels using join clauses.
1231 : *
1232 : * We consider only joins to single other relations in the current
1233 : * initial_rels list. This is sufficient to get a "true" result in most real
1234 : * queries, and an occasional erroneous "false" will only cost a bit more
1235 : * planning time. The reason for this limitation is that considering joins to
1236 : * other joins would require proving that the other join rel can legally be
1237 : * formed, which seems like too much trouble for something that's only a
1238 : * heuristic to save planning time. (Note: we must look at initial_rels
1239 : * and not all of the query, since when we are planning a sub-joinlist we
1240 : * may be forced to make clauseless joins within initial_rels even though
1241 : * there are join clauses linking to other parts of the query.)
1242 : */
1243 : static bool
1244 2700 : has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel)
1245 : {
1246 : ListCell *lc;
1247 :
1248 9996 : foreach(lc, root->initial_rels)
1249 : {
1250 7530 : RelOptInfo *rel2 = (RelOptInfo *) lfirst(lc);
1251 :
1252 : /* ignore rels that are already in "rel" */
1253 7530 : if (bms_overlap(rel->relids, rel2->relids))
1254 3198 : continue;
1255 :
1256 4332 : if (have_relevant_joinclause(root, rel, rel2))
1257 : {
1258 : Relids joinrelids;
1259 : SpecialJoinInfo *sjinfo;
1260 : bool reversed;
1261 :
1262 : /* join_is_legal needs relids of the union */
1263 390 : joinrelids = bms_union(rel->relids, rel2->relids);
1264 :
1265 390 : if (join_is_legal(root, rel, rel2, joinrelids,
1266 : &sjinfo, &reversed))
1267 : {
1268 : /* Yes, this will work */
1269 234 : bms_free(joinrelids);
1270 234 : return true;
1271 : }
1272 :
1273 156 : bms_free(joinrelids);
1274 : }
1275 : }
1276 :
1277 2466 : return false;
1278 : }
1279 :
1280 :
1281 : /*
1282 : * There's a pitfall for creating parameterized nestloops: suppose the inner
1283 : * rel (call it A) has a parameter that is a PlaceHolderVar, and that PHV's
1284 : * minimum eval_at set includes the outer rel (B) and some third rel (C).
1285 : * We might think we could create a B/A nestloop join that's parameterized by
1286 : * C. But we would end up with a plan in which the PHV's expression has to be
1287 : * evaluated as a nestloop parameter at the B/A join; and the executor is only
1288 : * set up to handle simple Vars as NestLoopParams. Rather than add complexity
1289 : * and overhead to the executor for such corner cases, it seems better to
1290 : * forbid the join. (Note that we can still make use of A's parameterized
1291 : * path with pre-joined B+C as the outer rel. have_join_order_restriction()
1292 : * ensures that we will consider making such a join even if there are not
1293 : * other reasons to do so.)
1294 : *
1295 : * So we check whether any PHVs used in the query could pose such a hazard.
1296 : * We don't have any simple way of checking whether a risky PHV would actually
1297 : * be used in the inner plan, and the case is so unusual that it doesn't seem
1298 : * worth working very hard on it.
1299 : *
1300 : * This needs to be checked in two places. If the inner rel's minimum
1301 : * parameterization would trigger the restriction, then join_is_legal() should
1302 : * reject the join altogether, because there will be no workable paths for it.
1303 : * But joinpath.c has to check again for every proposed nestloop path, because
1304 : * the inner path might have more than the minimum parameterization, causing
1305 : * some PHV to be dangerous for it that otherwise wouldn't be.
1306 : */
1307 : bool
1308 44922 : have_dangerous_phv(PlannerInfo *root,
1309 : Relids outer_relids, Relids inner_params)
1310 : {
1311 : ListCell *lc;
1312 :
1313 47820 : foreach(lc, root->placeholder_list)
1314 : {
1315 3150 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
1316 :
1317 3150 : if (!bms_is_subset(phinfo->ph_eval_at, inner_params))
1318 2298 : continue; /* ignore, could not be a nestloop param */
1319 852 : if (!bms_overlap(phinfo->ph_eval_at, outer_relids))
1320 192 : continue; /* ignore, not relevant to this join */
1321 660 : if (bms_is_subset(phinfo->ph_eval_at, outer_relids))
1322 408 : continue; /* safe, it can be eval'd within outerrel */
1323 : /* Otherwise, it's potentially unsafe, so reject the join */
1324 252 : return true;
1325 : }
1326 :
1327 : /* OK to perform the join */
1328 44670 : return false;
1329 : }
1330 :
1331 :
1332 : /*
1333 : * is_dummy_rel --- has relation been proven empty?
1334 : */
1335 : bool
1336 2631978 : is_dummy_rel(RelOptInfo *rel)
1337 : {
1338 : Path *path;
1339 :
1340 : /*
1341 : * A rel that is known dummy will have just one path that is a childless
1342 : * Append. (Even if somehow it has more paths, a childless Append will
1343 : * have cost zero and hence should be at the front of the pathlist.)
1344 : */
1345 2631978 : if (rel->pathlist == NIL)
1346 1413820 : return false;
1347 1218158 : path = (Path *) linitial(rel->pathlist);
1348 :
1349 : /*
1350 : * Initially, a dummy path will just be a childless Append. But in later
1351 : * planning stages we might stick a ProjectSetPath and/or ProjectionPath
1352 : * on top, since Append can't project. Rather than make assumptions about
1353 : * which combinations can occur, just descend through whatever we find.
1354 : */
1355 : for (;;)
1356 : {
1357 1262630 : if (IsA(path, ProjectionPath))
1358 35418 : path = ((ProjectionPath *) path)->subpath;
1359 1227212 : else if (IsA(path, ProjectSetPath))
1360 9054 : path = ((ProjectSetPath *) path)->subpath;
1361 : else
1362 1218158 : break;
1363 : }
1364 1218158 : if (IS_DUMMY_APPEND(path))
1365 3770 : return true;
1366 1214388 : return false;
1367 : }
1368 :
1369 : /*
1370 : * Mark a relation as proven empty.
1371 : *
1372 : * During GEQO planning, this can get invoked more than once on the same
1373 : * baserel struct, so it's worth checking to see if the rel is already marked
1374 : * dummy.
1375 : *
1376 : * Also, when called during GEQO join planning, we are in a short-lived
1377 : * memory context. We must make sure that the dummy path attached to a
1378 : * baserel survives the GEQO cycle, else the baserel is trashed for future
1379 : * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1380 : * we don't want the dummy path to clutter the main planning context. Upshot
1381 : * is that the best solution is to explicitly make the dummy path in the same
1382 : * context the given RelOptInfo is in.
1383 : */
1384 : void
1385 566 : mark_dummy_rel(RelOptInfo *rel)
1386 : {
1387 : MemoryContext oldcontext;
1388 :
1389 : /* Already marked? */
1390 566 : if (is_dummy_rel(rel))
1391 12 : return;
1392 :
1393 : /* No, so choose correct context to make the dummy path in */
1394 554 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1395 :
1396 : /* Set dummy size estimate */
1397 554 : rel->rows = 0;
1398 :
1399 : /* Evict any previously chosen paths */
1400 554 : rel->pathlist = NIL;
1401 554 : rel->partial_pathlist = NIL;
1402 :
1403 : /* Set up the dummy path */
1404 554 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
1405 : NIL, rel->lateral_relids,
1406 : 0, false, -1));
1407 :
1408 : /* Set or update cheapest_total_path and related fields */
1409 554 : set_cheapest(rel);
1410 :
1411 554 : MemoryContextSwitchTo(oldcontext);
1412 : }
1413 :
1414 :
1415 : /*
1416 : * restriction_is_constant_false --- is a restrictlist just FALSE?
1417 : *
1418 : * In cases where a qual is provably constant FALSE, eval_const_expressions
1419 : * will generally have thrown away anything that's ANDed with it. In outer
1420 : * join situations this will leave us computing cartesian products only to
1421 : * decide there's no match for an outer row, which is pretty stupid. So,
1422 : * we need to detect the case.
1423 : *
1424 : * If only_pushed_down is true, then consider only quals that are pushed-down
1425 : * from the point of view of the joinrel.
1426 : */
1427 : static bool
1428 449210 : restriction_is_constant_false(List *restrictlist,
1429 : RelOptInfo *joinrel,
1430 : bool only_pushed_down)
1431 : {
1432 : ListCell *lc;
1433 :
1434 : /*
1435 : * Despite the above comment, the restriction list we see here might
1436 : * possibly have other members besides the FALSE constant, since other
1437 : * quals could get "pushed down" to the outer join level. So we check
1438 : * each member of the list.
1439 : */
1440 960574 : foreach(lc, restrictlist)
1441 : {
1442 511786 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1443 :
1444 511786 : if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1445 133352 : continue;
1446 :
1447 378434 : if (rinfo->clause && IsA(rinfo->clause, Const))
1448 : {
1449 5364 : Const *con = (Const *) rinfo->clause;
1450 :
1451 : /* constant NULL is as good as constant FALSE for our purposes */
1452 5364 : if (con->constisnull)
1453 422 : return true;
1454 5256 : if (!DatumGetBool(con->constvalue))
1455 314 : return true;
1456 : }
1457 : }
1458 448788 : return false;
1459 : }
1460 :
1461 : /*
1462 : * Assess whether join between given two partitioned relations can be broken
1463 : * down into joins between matching partitions; a technique called
1464 : * "partitionwise join"
1465 : *
1466 : * Partitionwise join is possible when a. Joining relations have same
1467 : * partitioning scheme b. There exists an equi-join between the partition keys
1468 : * of the two relations.
1469 : *
1470 : * Partitionwise join is planned as follows (details: optimizer/README.)
1471 : *
1472 : * 1. Create the RelOptInfos for joins between matching partitions i.e
1473 : * child-joins and add paths to them.
1474 : *
1475 : * 2. Construct Append or MergeAppend paths across the set of child joins.
1476 : * This second phase is implemented by generate_partitionwise_join_paths().
1477 : *
1478 : * The RelOptInfo, SpecialJoinInfo and restrictlist for each child join are
1479 : * obtained by translating the respective parent join structures.
1480 : */
1481 : static void
1482 336654 : try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1483 : RelOptInfo *joinrel, SpecialJoinInfo *parent_sjinfo,
1484 : List *parent_restrictlist)
1485 : {
1486 336654 : bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1487 336654 : bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1488 336654 : List *parts1 = NIL;
1489 336654 : List *parts2 = NIL;
1490 336654 : ListCell *lcr1 = NULL;
1491 336654 : ListCell *lcr2 = NULL;
1492 : int cnt_parts;
1493 :
1494 : /* Guard against stack overflow due to overly deep partition hierarchy. */
1495 336654 : check_stack_depth();
1496 :
1497 : /* Nothing to do, if the join relation is not partitioned. */
1498 336654 : if (joinrel->part_scheme == NULL || joinrel->nparts == 0)
1499 334594 : return;
1500 :
1501 : /* The join relation should have consider_partitionwise_join set. */
1502 : Assert(joinrel->consider_partitionwise_join);
1503 :
1504 : /*
1505 : * We can not perform partitionwise join if either of the joining
1506 : * relations is not partitioned.
1507 : */
1508 2198 : if (!IS_PARTITIONED_REL(rel1) || !IS_PARTITIONED_REL(rel2))
1509 18 : return;
1510 :
1511 : Assert(REL_HAS_ALL_PART_PROPS(rel1) && REL_HAS_ALL_PART_PROPS(rel2));
1512 :
1513 : /* The joining relations should have consider_partitionwise_join set. */
1514 : Assert(rel1->consider_partitionwise_join &&
1515 : rel2->consider_partitionwise_join);
1516 :
1517 : /*
1518 : * The partition scheme of the join relation should match that of the
1519 : * joining relations.
1520 : */
1521 : Assert(joinrel->part_scheme == rel1->part_scheme &&
1522 : joinrel->part_scheme == rel2->part_scheme);
1523 :
1524 : Assert(!(joinrel->partbounds_merged && (joinrel->nparts <= 0)));
1525 :
1526 2180 : compute_partition_bounds(root, rel1, rel2, joinrel, parent_sjinfo,
1527 : &parts1, &parts2);
1528 :
1529 2180 : if (joinrel->partbounds_merged)
1530 : {
1531 768 : lcr1 = list_head(parts1);
1532 768 : lcr2 = list_head(parts2);
1533 : }
1534 :
1535 : /*
1536 : * Create child-join relations for this partitioned join, if those don't
1537 : * exist. Add paths to child-joins for a pair of child relations
1538 : * corresponding to the given pair of parent relations.
1539 : */
1540 7670 : for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1541 : {
1542 : RelOptInfo *child_rel1;
1543 : RelOptInfo *child_rel2;
1544 : bool rel1_empty;
1545 : bool rel2_empty;
1546 : SpecialJoinInfo *child_sjinfo;
1547 : List *child_restrictlist;
1548 : RelOptInfo *child_joinrel;
1549 : AppendRelInfo **appinfos;
1550 : int nappinfos;
1551 : Relids child_relids;
1552 :
1553 5610 : if (joinrel->partbounds_merged)
1554 : {
1555 2010 : child_rel1 = lfirst_node(RelOptInfo, lcr1);
1556 2010 : child_rel2 = lfirst_node(RelOptInfo, lcr2);
1557 2010 : lcr1 = lnext(parts1, lcr1);
1558 2010 : lcr2 = lnext(parts2, lcr2);
1559 : }
1560 : else
1561 : {
1562 3600 : child_rel1 = rel1->part_rels[cnt_parts];
1563 3600 : child_rel2 = rel2->part_rels[cnt_parts];
1564 : }
1565 :
1566 5610 : rel1_empty = (child_rel1 == NULL || IS_DUMMY_REL(child_rel1));
1567 5610 : rel2_empty = (child_rel2 == NULL || IS_DUMMY_REL(child_rel2));
1568 :
1569 : /*
1570 : * Check for cases where we can prove that this segment of the join
1571 : * returns no rows, due to one or both inputs being empty (including
1572 : * inputs that have been pruned away entirely). If so just ignore it.
1573 : * These rules are equivalent to populate_joinrel_with_paths's rules
1574 : * for dummy input relations.
1575 : */
1576 5610 : switch (parent_sjinfo->jointype)
1577 : {
1578 2804 : case JOIN_INNER:
1579 : case JOIN_SEMI:
1580 2804 : if (rel1_empty || rel2_empty)
1581 64 : continue; /* ignore this join segment */
1582 2768 : break;
1583 2084 : case JOIN_LEFT:
1584 : case JOIN_ANTI:
1585 2084 : if (rel1_empty)
1586 28 : continue; /* ignore this join segment */
1587 2056 : break;
1588 722 : case JOIN_FULL:
1589 722 : if (rel1_empty && rel2_empty)
1590 0 : continue; /* ignore this join segment */
1591 722 : break;
1592 0 : default:
1593 : /* other values not expected here */
1594 0 : elog(ERROR, "unrecognized join type: %d",
1595 : (int) parent_sjinfo->jointype);
1596 : break;
1597 : }
1598 :
1599 : /*
1600 : * If a child has been pruned entirely then we can't generate paths
1601 : * for it, so we have to reject partitionwise joining unless we were
1602 : * able to eliminate this partition above.
1603 : */
1604 5546 : if (child_rel1 == NULL || child_rel2 == NULL)
1605 : {
1606 : /*
1607 : * Mark the joinrel as unpartitioned so that later functions treat
1608 : * it correctly.
1609 : */
1610 120 : joinrel->nparts = 0;
1611 120 : return;
1612 : }
1613 :
1614 : /*
1615 : * If a leaf relation has consider_partitionwise_join=false, it means
1616 : * that it's a dummy relation for which we skipped setting up tlist
1617 : * expressions and adding EC members in set_append_rel_size(), so
1618 : * again we have to fail here.
1619 : */
1620 5426 : if (rel1_is_simple && !child_rel1->consider_partitionwise_join)
1621 : {
1622 : Assert(child_rel1->reloptkind == RELOPT_OTHER_MEMBER_REL);
1623 : Assert(IS_DUMMY_REL(child_rel1));
1624 0 : joinrel->nparts = 0;
1625 0 : return;
1626 : }
1627 5426 : if (rel2_is_simple && !child_rel2->consider_partitionwise_join)
1628 : {
1629 : Assert(child_rel2->reloptkind == RELOPT_OTHER_MEMBER_REL);
1630 : Assert(IS_DUMMY_REL(child_rel2));
1631 0 : joinrel->nparts = 0;
1632 0 : return;
1633 : }
1634 :
1635 : /* We should never try to join two overlapping sets of rels. */
1636 : Assert(!bms_overlap(child_rel1->relids, child_rel2->relids));
1637 :
1638 : /*
1639 : * Construct SpecialJoinInfo from parent join relations's
1640 : * SpecialJoinInfo.
1641 : */
1642 5426 : child_sjinfo = build_child_join_sjinfo(root, parent_sjinfo,
1643 : child_rel1->relids,
1644 : child_rel2->relids);
1645 :
1646 : /* Find the AppendRelInfo structures */
1647 5426 : child_relids = bms_union(child_rel1->relids, child_rel2->relids);
1648 5426 : appinfos = find_appinfos_by_relids(root, child_relids,
1649 : &nappinfos);
1650 :
1651 : /*
1652 : * Construct restrictions applicable to the child join from those
1653 : * applicable to the parent join.
1654 : */
1655 : child_restrictlist =
1656 5426 : (List *) adjust_appendrel_attrs(root,
1657 : (Node *) parent_restrictlist,
1658 : nappinfos, appinfos);
1659 :
1660 : /* Find or construct the child join's RelOptInfo */
1661 5426 : child_joinrel = joinrel->part_rels[cnt_parts];
1662 5426 : if (!child_joinrel)
1663 : {
1664 4918 : child_joinrel = build_child_join_rel(root, child_rel1, child_rel2,
1665 : joinrel, child_restrictlist,
1666 : child_sjinfo, nappinfos, appinfos);
1667 4918 : joinrel->part_rels[cnt_parts] = child_joinrel;
1668 4918 : joinrel->live_parts = bms_add_member(joinrel->live_parts, cnt_parts);
1669 4918 : joinrel->all_partrels = bms_add_members(joinrel->all_partrels,
1670 4918 : child_joinrel->relids);
1671 : }
1672 :
1673 : /* Assert we got the right one */
1674 : Assert(bms_equal(child_joinrel->relids,
1675 : adjust_child_relids(joinrel->relids,
1676 : nappinfos, appinfos)));
1677 :
1678 : /* And make paths for the child join */
1679 5426 : populate_joinrel_with_paths(root, child_rel1, child_rel2,
1680 : child_joinrel, child_sjinfo,
1681 : child_restrictlist);
1682 :
1683 : /*
1684 : * When there are thousands of partitions involved, this loop will
1685 : * accumulate a significant amount of memory usage from objects that
1686 : * are only needed within the loop. Free these local objects eagerly
1687 : * at the end of each iteration.
1688 : */
1689 5426 : pfree(appinfos);
1690 5426 : bms_free(child_relids);
1691 5426 : free_child_join_sjinfo(child_sjinfo, parent_sjinfo);
1692 : }
1693 : }
1694 :
1695 : /*
1696 : * Construct the SpecialJoinInfo for a child-join by translating
1697 : * SpecialJoinInfo for the join between parents. left_relids and right_relids
1698 : * are the relids of left and right side of the join respectively.
1699 : *
1700 : * If translations are added to or removed from this function, consider
1701 : * updating free_child_join_sjinfo() accordingly.
1702 : */
1703 : static SpecialJoinInfo *
1704 5426 : build_child_join_sjinfo(PlannerInfo *root, SpecialJoinInfo *parent_sjinfo,
1705 : Relids left_relids, Relids right_relids)
1706 : {
1707 5426 : SpecialJoinInfo *sjinfo = makeNode(SpecialJoinInfo);
1708 : AppendRelInfo **left_appinfos;
1709 : int left_nappinfos;
1710 : AppendRelInfo **right_appinfos;
1711 : int right_nappinfos;
1712 :
1713 : /* Dummy SpecialJoinInfos can be created without any translation. */
1714 5426 : if (parent_sjinfo->jointype == JOIN_INNER)
1715 : {
1716 : Assert(parent_sjinfo->ojrelid == 0);
1717 2318 : init_dummy_sjinfo(sjinfo, left_relids, right_relids);
1718 2318 : return sjinfo;
1719 : }
1720 :
1721 3108 : memcpy(sjinfo, parent_sjinfo, sizeof(SpecialJoinInfo));
1722 3108 : left_appinfos = find_appinfos_by_relids(root, left_relids,
1723 : &left_nappinfos);
1724 3108 : right_appinfos = find_appinfos_by_relids(root, right_relids,
1725 : &right_nappinfos);
1726 :
1727 3108 : sjinfo->min_lefthand = adjust_child_relids(sjinfo->min_lefthand,
1728 : left_nappinfos, left_appinfos);
1729 3108 : sjinfo->min_righthand = adjust_child_relids(sjinfo->min_righthand,
1730 : right_nappinfos,
1731 : right_appinfos);
1732 3108 : sjinfo->syn_lefthand = adjust_child_relids(sjinfo->syn_lefthand,
1733 : left_nappinfos, left_appinfos);
1734 3108 : sjinfo->syn_righthand = adjust_child_relids(sjinfo->syn_righthand,
1735 : right_nappinfos,
1736 : right_appinfos);
1737 : /* outer-join relids need no adjustment */
1738 6216 : sjinfo->semi_rhs_exprs = (List *) adjust_appendrel_attrs(root,
1739 3108 : (Node *) sjinfo->semi_rhs_exprs,
1740 : right_nappinfos,
1741 : right_appinfos);
1742 :
1743 3108 : pfree(left_appinfos);
1744 3108 : pfree(right_appinfos);
1745 :
1746 3108 : return sjinfo;
1747 : }
1748 :
1749 : /*
1750 : * free_child_join_sjinfo
1751 : * Free memory consumed by a SpecialJoinInfo created by
1752 : * build_child_join_sjinfo()
1753 : *
1754 : * Only members that are translated copies of their counterpart in the parent
1755 : * SpecialJoinInfo are freed here.
1756 : */
1757 : static void
1758 5426 : free_child_join_sjinfo(SpecialJoinInfo *child_sjinfo,
1759 : SpecialJoinInfo *parent_sjinfo)
1760 : {
1761 : /*
1762 : * Dummy SpecialJoinInfos of inner joins do not have any translated fields
1763 : * and hence no fields that to be freed.
1764 : */
1765 5426 : if (child_sjinfo->jointype != JOIN_INNER)
1766 : {
1767 3108 : if (child_sjinfo->min_lefthand != parent_sjinfo->min_lefthand)
1768 3090 : bms_free(child_sjinfo->min_lefthand);
1769 :
1770 3108 : if (child_sjinfo->min_righthand != parent_sjinfo->min_righthand)
1771 3108 : bms_free(child_sjinfo->min_righthand);
1772 :
1773 3108 : if (child_sjinfo->syn_lefthand != parent_sjinfo->syn_lefthand)
1774 3108 : bms_free(child_sjinfo->syn_lefthand);
1775 :
1776 3108 : if (child_sjinfo->syn_righthand != parent_sjinfo->syn_righthand)
1777 3108 : bms_free(child_sjinfo->syn_righthand);
1778 :
1779 : Assert(child_sjinfo->commute_above_l == parent_sjinfo->commute_above_l);
1780 : Assert(child_sjinfo->commute_above_r == parent_sjinfo->commute_above_r);
1781 : Assert(child_sjinfo->commute_below_l == parent_sjinfo->commute_below_l);
1782 : Assert(child_sjinfo->commute_below_r == parent_sjinfo->commute_below_r);
1783 :
1784 : Assert(child_sjinfo->semi_operators == parent_sjinfo->semi_operators);
1785 :
1786 : /*
1787 : * semi_rhs_exprs may in principle be freed, but a simple pfree() does
1788 : * not suffice, so we leave it alone.
1789 : */
1790 : }
1791 :
1792 5426 : pfree(child_sjinfo);
1793 5426 : }
1794 :
1795 : /*
1796 : * compute_partition_bounds
1797 : * Compute the partition bounds for a join rel from those for inputs
1798 : */
1799 : static void
1800 2180 : compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
1801 : RelOptInfo *rel2, RelOptInfo *joinrel,
1802 : SpecialJoinInfo *parent_sjinfo,
1803 : List **parts1, List **parts2)
1804 : {
1805 : /*
1806 : * If we don't have the partition bounds for the join rel yet, try to
1807 : * compute those along with pairs of partitions to be joined.
1808 : */
1809 2180 : if (joinrel->nparts == -1)
1810 : {
1811 2004 : PartitionScheme part_scheme = joinrel->part_scheme;
1812 2004 : PartitionBoundInfo boundinfo = NULL;
1813 2004 : int nparts = 0;
1814 :
1815 : Assert(joinrel->boundinfo == NULL);
1816 : Assert(joinrel->part_rels == NULL);
1817 :
1818 : /*
1819 : * See if the partition bounds for inputs are exactly the same, in
1820 : * which case we don't need to work hard: the join rel will have the
1821 : * same partition bounds as inputs, and the partitions with the same
1822 : * cardinal positions will form the pairs.
1823 : *
1824 : * Note: even in cases where one or both inputs have merged bounds, it
1825 : * would be possible for both the bounds to be exactly the same, but
1826 : * it seems unlikely to be worth the cycles to check.
1827 : */
1828 2004 : if (!rel1->partbounds_merged &&
1829 1944 : !rel2->partbounds_merged &&
1830 3630 : rel1->nparts == rel2->nparts &&
1831 1686 : partition_bounds_equal(part_scheme->partnatts,
1832 : part_scheme->parttyplen,
1833 : part_scheme->parttypbyval,
1834 : rel1->boundinfo, rel2->boundinfo))
1835 : {
1836 1158 : boundinfo = rel1->boundinfo;
1837 1158 : nparts = rel1->nparts;
1838 : }
1839 : else
1840 : {
1841 : /* Try merging the partition bounds for inputs. */
1842 846 : boundinfo = partition_bounds_merge(part_scheme->partnatts,
1843 846 : part_scheme->partsupfunc,
1844 : part_scheme->partcollation,
1845 : rel1, rel2,
1846 : parent_sjinfo->jointype,
1847 : parts1, parts2);
1848 846 : if (boundinfo == NULL)
1849 : {
1850 114 : joinrel->nparts = 0;
1851 114 : return;
1852 : }
1853 732 : nparts = list_length(*parts1);
1854 732 : joinrel->partbounds_merged = true;
1855 : }
1856 :
1857 : Assert(nparts > 0);
1858 1890 : joinrel->boundinfo = boundinfo;
1859 1890 : joinrel->nparts = nparts;
1860 1890 : joinrel->part_rels =
1861 1890 : (RelOptInfo **) palloc0(sizeof(RelOptInfo *) * nparts);
1862 : }
1863 : else
1864 : {
1865 : Assert(joinrel->nparts > 0);
1866 : Assert(joinrel->boundinfo);
1867 : Assert(joinrel->part_rels);
1868 :
1869 : /*
1870 : * If the join rel's partbounds_merged flag is true, it means inputs
1871 : * are not guaranteed to have the same partition bounds, therefore we
1872 : * can't assume that the partitions at the same cardinal positions
1873 : * form the pairs; let get_matching_part_pairs() generate the pairs.
1874 : * Otherwise, nothing to do since we can assume that.
1875 : */
1876 176 : if (joinrel->partbounds_merged)
1877 : {
1878 36 : get_matching_part_pairs(root, joinrel, rel1, rel2,
1879 : parts1, parts2);
1880 : Assert(list_length(*parts1) == joinrel->nparts);
1881 : Assert(list_length(*parts2) == joinrel->nparts);
1882 : }
1883 : }
1884 : }
1885 :
1886 : /*
1887 : * get_matching_part_pairs
1888 : * Generate pairs of partitions to be joined from inputs
1889 : */
1890 : static void
1891 36 : get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
1892 : RelOptInfo *rel1, RelOptInfo *rel2,
1893 : List **parts1, List **parts2)
1894 : {
1895 36 : bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1896 36 : bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1897 : int cnt_parts;
1898 :
1899 36 : *parts1 = NIL;
1900 36 : *parts2 = NIL;
1901 :
1902 132 : for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1903 : {
1904 96 : RelOptInfo *child_joinrel = joinrel->part_rels[cnt_parts];
1905 : RelOptInfo *child_rel1;
1906 : RelOptInfo *child_rel2;
1907 : Relids child_relids1;
1908 : Relids child_relids2;
1909 :
1910 : /*
1911 : * If this segment of the join is empty, it means that this segment
1912 : * was ignored when previously creating child-join paths for it in
1913 : * try_partitionwise_join() as it would not contribute to the join
1914 : * result, due to one or both inputs being empty; add NULL to each of
1915 : * the given lists so that this segment will be ignored again in that
1916 : * function.
1917 : */
1918 96 : if (!child_joinrel)
1919 : {
1920 0 : *parts1 = lappend(*parts1, NULL);
1921 0 : *parts2 = lappend(*parts2, NULL);
1922 0 : continue;
1923 : }
1924 :
1925 : /*
1926 : * Get a relids set of partition(s) involved in this join segment that
1927 : * are from the rel1 side.
1928 : */
1929 96 : child_relids1 = bms_intersect(child_joinrel->relids,
1930 96 : rel1->all_partrels);
1931 : Assert(bms_num_members(child_relids1) == bms_num_members(rel1->relids));
1932 :
1933 : /*
1934 : * Get a child rel for rel1 with the relids. Note that we should have
1935 : * the child rel even if rel1 is a join rel, because in that case the
1936 : * partitions specified in the relids would have matching/overlapping
1937 : * boundaries, so the specified partitions should be considered as
1938 : * ones to be joined when planning partitionwise joins of rel1,
1939 : * meaning that the child rel would have been built by the time we get
1940 : * here.
1941 : */
1942 96 : if (rel1_is_simple)
1943 : {
1944 0 : int varno = bms_singleton_member(child_relids1);
1945 :
1946 0 : child_rel1 = find_base_rel(root, varno);
1947 : }
1948 : else
1949 96 : child_rel1 = find_join_rel(root, child_relids1);
1950 : Assert(child_rel1);
1951 :
1952 : /*
1953 : * Get a relids set of partition(s) involved in this join segment that
1954 : * are from the rel2 side.
1955 : */
1956 96 : child_relids2 = bms_intersect(child_joinrel->relids,
1957 96 : rel2->all_partrels);
1958 : Assert(bms_num_members(child_relids2) == bms_num_members(rel2->relids));
1959 :
1960 : /*
1961 : * Get a child rel for rel2 with the relids. See above comments.
1962 : */
1963 96 : if (rel2_is_simple)
1964 : {
1965 96 : int varno = bms_singleton_member(child_relids2);
1966 :
1967 96 : child_rel2 = find_base_rel(root, varno);
1968 : }
1969 : else
1970 0 : child_rel2 = find_join_rel(root, child_relids2);
1971 : Assert(child_rel2);
1972 :
1973 : /*
1974 : * The join of rel1 and rel2 is legal, so is the join of the child
1975 : * rels obtained above; add them to the given lists as a join pair
1976 : * producing this join segment.
1977 : */
1978 96 : *parts1 = lappend(*parts1, child_rel1);
1979 96 : *parts2 = lappend(*parts2, child_rel2);
1980 : }
1981 36 : }
|
