1 /*-------------------------------------------------------------------------
2 *
3 * joinrels.c
4 * Routines to determine which relations should be joined
5 *
6 * Portions Copyright (c) 1996-2020, 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_list,
29 ListCell *other_rels);
30 static void make_rels_by_clauseless_joins(PlannerInfo *root,
31 RelOptInfo *old_rel,
32 List *other_rels);
33 static bool has_join_restriction(PlannerInfo *root, RelOptInfo *rel);
34 static bool has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel);
35 static bool restriction_is_constant_false(List *restrictlist,
36 RelOptInfo *joinrel,
37 bool only_pushed_down);
38 static void populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
39 RelOptInfo *rel2, RelOptInfo *joinrel,
40 SpecialJoinInfo *sjinfo, List *restrictlist);
41 static void try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1,
42 RelOptInfo *rel2, RelOptInfo *joinrel,
43 SpecialJoinInfo *parent_sjinfo,
44 List *parent_restrictlist);
45 static SpecialJoinInfo *build_child_join_sjinfo(PlannerInfo *root,
46 SpecialJoinInfo *parent_sjinfo,
47 Relids left_relids, Relids right_relids);
48 static void compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
49 RelOptInfo *rel2, RelOptInfo *joinrel,
50 SpecialJoinInfo *parent_sjinfo,
51 List **parts1, List **parts2);
52 static void get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
53 RelOptInfo *rel1, RelOptInfo *rel2,
54 List **parts1, List **parts2);
55
56
57 /*
58 * join_search_one_level
59 * Consider ways to produce join relations containing exactly 'level'
60 * jointree items. (This is one step of the dynamic-programming method
61 * embodied in standard_join_search.) Join rel nodes for each feasible
62 * combination of lower-level rels are created and returned in a list.
63 * Implementation paths are created for each such joinrel, too.
64 *
65 * level: level of rels we want to make this time
66 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
67 *
68 * The result is returned in root->join_rel_level[level].
69 */
70 void
join_search_one_level(PlannerInfo * root,int level)71 join_search_one_level(PlannerInfo *root, int level)
72 {
73 List **joinrels = root->join_rel_level;
74 ListCell *r;
75 int k;
76
77 Assert(joinrels[level] == NIL);
78
79 /* Set join_cur_level so that new joinrels are added to proper list */
80 root->join_cur_level = level;
81
82 /*
83 * First, consider left-sided and right-sided plans, in which rels of
84 * exactly level-1 member relations are joined against initial relations.
85 * We prefer to join using join clauses, but if we find a rel of level-1
86 * members that has no join clauses, we will generate Cartesian-product
87 * joins against all initial rels not already contained in it.
88 */
89 foreach(r, joinrels[level - 1])
90 {
91 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
92
93 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
94 has_join_restriction(root, old_rel))
95 {
96 /*
97 * There are join clauses or join order restrictions relevant to
98 * this rel, so consider joins between this rel and (only) those
99 * initial rels it is linked to by a clause or restriction.
100 *
101 * At level 2 this condition is symmetric, so there is no need to
102 * look at initial rels before this one in the list; we already
103 * considered such joins when we were at the earlier rel. (The
104 * mirror-image joins are handled automatically by make_join_rel.)
105 * In later passes (level > 2), we join rels of the previous level
106 * to each initial rel they don't already include but have a join
107 * clause or restriction with.
108 */
109 List *other_rels_list;
110 ListCell *other_rels;
111
112 if (level == 2) /* consider remaining initial rels */
113 {
114 other_rels_list = joinrels[level - 1];
115 other_rels = lnext(other_rels_list, r);
116 }
117 else /* consider all initial rels */
118 {
119 other_rels_list = joinrels[1];
120 other_rels = list_head(other_rels_list);
121 }
122
123 make_rels_by_clause_joins(root,
124 old_rel,
125 other_rels_list,
126 other_rels);
127 }
128 else
129 {
130 /*
131 * Oops, we have a relation that is not joined to any other
132 * relation, either directly or by join-order restrictions.
133 * Cartesian product time.
134 *
135 * We consider a cartesian product with each not-already-included
136 * initial rel, whether it has other join clauses or not. At
137 * level 2, if there are two or more clauseless initial rels, we
138 * will redundantly consider joining them in both directions; but
139 * such cases aren't common enough to justify adding complexity to
140 * avoid the duplicated effort.
141 */
142 make_rels_by_clauseless_joins(root,
143 old_rel,
144 joinrels[1]);
145 }
146 }
147
148 /*
149 * Now, consider "bushy plans" in which relations of k initial rels are
150 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
151 *
152 * We only consider bushy-plan joins for pairs of rels where there is a
153 * suitable join clause (or join order restriction), in order to avoid
154 * unreasonable growth of planning time.
155 */
156 for (k = 2;; k++)
157 {
158 int other_level = level - k;
159
160 /*
161 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
162 * need to go as far as the halfway point.
163 */
164 if (k > other_level)
165 break;
166
167 foreach(r, joinrels[k])
168 {
169 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
170 List *other_rels_list;
171 ListCell *other_rels;
172 ListCell *r2;
173
174 /*
175 * We can ignore relations without join clauses here, unless they
176 * participate in join-order restrictions --- then we might have
177 * to force a bushy join plan.
178 */
179 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
180 !has_join_restriction(root, old_rel))
181 continue;
182
183 if (k == other_level)
184 {
185 /* only consider remaining rels */
186 other_rels_list = joinrels[k];
187 other_rels = lnext(other_rels_list, r);
188 }
189 else
190 {
191 other_rels_list = joinrels[other_level];
192 other_rels = list_head(other_rels_list);
193 }
194
195 for_each_cell(r2, other_rels_list, other_rels)
196 {
197 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
198
199 if (!bms_overlap(old_rel->relids, new_rel->relids))
200 {
201 /*
202 * OK, we can build a rel of the right level from this
203 * pair of rels. Do so if there is at least one relevant
204 * join clause or join order restriction.
205 */
206 if (have_relevant_joinclause(root, old_rel, new_rel) ||
207 have_join_order_restriction(root, old_rel, new_rel))
208 {
209 (void) make_join_rel(root, old_rel, new_rel);
210 }
211 }
212 }
213 }
214 }
215
216 /*----------
217 * Last-ditch effort: if we failed to find any usable joins so far, force
218 * a set of cartesian-product joins to be generated. This handles the
219 * special case where all the available rels have join clauses but we
220 * cannot use any of those clauses yet. This can only happen when we are
221 * considering a join sub-problem (a sub-joinlist) and all the rels in the
222 * sub-problem have only join clauses with rels outside the sub-problem.
223 * An example is
224 *
225 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
226 * WHERE a.w = c.x and b.y = d.z;
227 *
228 * If the "a INNER JOIN b" sub-problem does not get flattened into the
229 * upper level, we must be willing to make a cartesian join of a and b;
230 * but the code above will not have done so, because it thought that both
231 * a and b have joinclauses. We consider only left-sided and right-sided
232 * cartesian joins in this case (no bushy).
233 *----------
234 */
235 if (joinrels[level] == NIL)
236 {
237 /*
238 * This loop is just like the first one, except we always call
239 * make_rels_by_clauseless_joins().
240 */
241 foreach(r, joinrels[level - 1])
242 {
243 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
244
245 make_rels_by_clauseless_joins(root,
246 old_rel,
247 joinrels[1]);
248 }
249
250 /*----------
251 * When special joins are involved, there may be no legal way
252 * to make an N-way join for some values of N. For example consider
253 *
254 * SELECT ... FROM t1 WHERE
255 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
256 * y IN (SELECT ... FROM t4,t5 WHERE ...)
257 *
258 * We will flatten this query to a 5-way join problem, but there are
259 * no 4-way joins that join_is_legal() will consider legal. We have
260 * to accept failure at level 4 and go on to discover a workable
261 * bushy plan at level 5.
262 *
263 * However, if there are no special joins and no lateral references
264 * then join_is_legal() should never fail, and so the following sanity
265 * check is useful.
266 *----------
267 */
268 if (joinrels[level] == NIL &&
269 root->join_info_list == NIL &&
270 !root->hasLateralRTEs)
271 elog(ERROR, "failed to build any %d-way joins", level);
272 }
273 }
274
275 /*
276 * make_rels_by_clause_joins
277 * Build joins between the given relation 'old_rel' and other relations
278 * that participate in join clauses that 'old_rel' also participates in
279 * (or participate in join-order restrictions with it).
280 * The join rels are returned in root->join_rel_level[join_cur_level].
281 *
282 * Note: at levels above 2 we will generate the same joined relation in
283 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
284 * (b join c) join a, though the second case will add a different set of Paths
285 * to it. This is the reason for using the join_rel_level mechanism, which
286 * automatically ensures that each new joinrel is only added to the list once.
287 *
288 * 'old_rel' is the relation entry for the relation to be joined
289 * 'other_rels_list': a list containing the other
290 * rels to be considered for joining
291 * 'other_rels': the first cell to be considered
292 *
293 * Currently, this is only used with initial rels in other_rels, but it
294 * will work for joining to joinrels too.
295 */
296 static void
make_rels_by_clause_joins(PlannerInfo * root,RelOptInfo * old_rel,List * other_rels_list,ListCell * other_rels)297 make_rels_by_clause_joins(PlannerInfo *root,
298 RelOptInfo *old_rel,
299 List *other_rels_list,
300 ListCell *other_rels)
301 {
302 ListCell *l;
303
304 for_each_cell(l, other_rels_list, other_rels)
305 {
306 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
307
308 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
309 (have_relevant_joinclause(root, old_rel, other_rel) ||
310 have_join_order_restriction(root, old_rel, other_rel)))
311 {
312 (void) make_join_rel(root, old_rel, other_rel);
313 }
314 }
315 }
316
317 /*
318 * make_rels_by_clauseless_joins
319 * Given a relation 'old_rel' and a list of other relations
320 * 'other_rels', create a join relation between 'old_rel' and each
321 * member of 'other_rels' that isn't already included in 'old_rel'.
322 * The join rels are returned in root->join_rel_level[join_cur_level].
323 *
324 * 'old_rel' is the relation entry for the relation to be joined
325 * 'other_rels': a list containing the other rels to be considered for joining
326 *
327 * Currently, this is only used with initial rels in other_rels, but it would
328 * work for joining to joinrels too.
329 */
330 static void
make_rels_by_clauseless_joins(PlannerInfo * root,RelOptInfo * old_rel,List * other_rels)331 make_rels_by_clauseless_joins(PlannerInfo *root,
332 RelOptInfo *old_rel,
333 List *other_rels)
334 {
335 ListCell *l;
336
337 foreach(l, other_rels)
338 {
339 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
340
341 if (!bms_overlap(other_rel->relids, old_rel->relids))
342 {
343 (void) make_join_rel(root, old_rel, other_rel);
344 }
345 }
346 }
347
348
349 /*
350 * join_is_legal
351 * Determine whether a proposed join is legal given the query's
352 * join order constraints; and if it is, determine the join type.
353 *
354 * Caller must supply not only the two rels, but the union of their relids.
355 * (We could simplify the API by computing joinrelids locally, but this
356 * would be redundant work in the normal path through make_join_rel.)
357 *
358 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
359 * else it's set to point to the associated SpecialJoinInfo node. Also,
360 * *reversed_p is set true if the given relations need to be swapped to
361 * match the SpecialJoinInfo node.
362 */
363 static bool
join_is_legal(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2,Relids joinrelids,SpecialJoinInfo ** sjinfo_p,bool * reversed_p)364 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
365 Relids joinrelids,
366 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
367 {
368 SpecialJoinInfo *match_sjinfo;
369 bool reversed;
370 bool unique_ified;
371 bool must_be_leftjoin;
372 ListCell *l;
373
374 /*
375 * Ensure output params are set on failure return. This is just to
376 * suppress uninitialized-variable warnings from overly anal compilers.
377 */
378 *sjinfo_p = NULL;
379 *reversed_p = false;
380
381 /*
382 * If we have any special joins, the proposed join might be illegal; and
383 * in any case we have to determine its join type. Scan the join info
384 * list for matches and conflicts.
385 */
386 match_sjinfo = NULL;
387 reversed = false;
388 unique_ified = false;
389 must_be_leftjoin = false;
390
391 foreach(l, root->join_info_list)
392 {
393 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
394
395 /*
396 * This special join is not relevant unless its RHS overlaps the
397 * proposed join. (Check this first as a fast path for dismissing
398 * most irrelevant SJs quickly.)
399 */
400 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
401 continue;
402
403 /*
404 * Also, not relevant if proposed join is fully contained within RHS
405 * (ie, we're still building up the RHS).
406 */
407 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
408 continue;
409
410 /*
411 * Also, not relevant if SJ is already done within either input.
412 */
413 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
414 bms_is_subset(sjinfo->min_righthand, rel1->relids))
415 continue;
416 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
417 bms_is_subset(sjinfo->min_righthand, rel2->relids))
418 continue;
419
420 /*
421 * If it's a semijoin and we already joined the RHS to any other rels
422 * within either input, then we must have unique-ified the RHS at that
423 * point (see below). Therefore the semijoin is no longer relevant in
424 * this join path.
425 */
426 if (sjinfo->jointype == JOIN_SEMI)
427 {
428 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
429 !bms_equal(sjinfo->syn_righthand, rel1->relids))
430 continue;
431 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
432 !bms_equal(sjinfo->syn_righthand, rel2->relids))
433 continue;
434 }
435
436 /*
437 * If one input contains min_lefthand and the other contains
438 * min_righthand, then we can perform the SJ at this join.
439 *
440 * Reject if we get matches to more than one SJ; that implies we're
441 * considering something that's not really valid.
442 */
443 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
444 bms_is_subset(sjinfo->min_righthand, rel2->relids))
445 {
446 if (match_sjinfo)
447 return false; /* invalid join path */
448 match_sjinfo = sjinfo;
449 reversed = false;
450 }
451 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
452 bms_is_subset(sjinfo->min_righthand, rel1->relids))
453 {
454 if (match_sjinfo)
455 return false; /* invalid join path */
456 match_sjinfo = sjinfo;
457 reversed = true;
458 }
459 else if (sjinfo->jointype == JOIN_SEMI &&
460 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
461 create_unique_path(root, rel2, rel2->cheapest_total_path,
462 sjinfo) != NULL)
463 {
464 /*----------
465 * For a semijoin, we can join the RHS to anything else by
466 * unique-ifying the RHS (if the RHS can be unique-ified).
467 * We will only get here if we have the full RHS but less
468 * than min_lefthand on the LHS.
469 *
470 * The reason to consider such a join path is exemplified by
471 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
472 * If we insist on doing this as a semijoin we will first have
473 * to form the cartesian product of A*B. But if we unique-ify
474 * C then the semijoin becomes a plain innerjoin and we can join
475 * in any order, eg C to A and then to B. When C is much smaller
476 * than A and B this can be a huge win. So we allow C to be
477 * joined to just A or just B here, and then make_join_rel has
478 * to handle the case properly.
479 *
480 * Note that actually we'll allow unique-ified C to be joined to
481 * some other relation D here, too. That is legal, if usually not
482 * very sane, and this routine is only concerned with legality not
483 * with whether the join is good strategy.
484 *----------
485 */
486 if (match_sjinfo)
487 return false; /* invalid join path */
488 match_sjinfo = sjinfo;
489 reversed = false;
490 unique_ified = true;
491 }
492 else if (sjinfo->jointype == JOIN_SEMI &&
493 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
494 create_unique_path(root, rel1, rel1->cheapest_total_path,
495 sjinfo) != NULL)
496 {
497 /* Reversed semijoin case */
498 if (match_sjinfo)
499 return false; /* invalid join path */
500 match_sjinfo = sjinfo;
501 reversed = true;
502 unique_ified = true;
503 }
504 else
505 {
506 /*
507 * Otherwise, the proposed join overlaps the RHS but isn't a valid
508 * implementation of this SJ. But don't panic quite yet: the RHS
509 * violation might have occurred previously, in one or both input
510 * relations, in which case we must have previously decided that
511 * it was OK to commute some other SJ with this one. If we need
512 * to perform this join to finish building up the RHS, rejecting
513 * it could lead to not finding any plan at all. (This can occur
514 * because of the heuristics elsewhere in this file that postpone
515 * clauseless joins: we might not consider doing a clauseless join
516 * within the RHS until after we've performed other, validly
517 * commutable SJs with one or both sides of the clauseless join.)
518 * This consideration boils down to the rule that if both inputs
519 * overlap the RHS, we can allow the join --- they are either
520 * fully within the RHS, or represent previously-allowed joins to
521 * rels outside it.
522 */
523 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
524 bms_overlap(rel2->relids, sjinfo->min_righthand))
525 continue; /* assume valid previous violation of RHS */
526
527 /*
528 * The proposed join could still be legal, but only if we're
529 * allowed to associate it into the RHS of this SJ. That means
530 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
531 * not FULL) and the proposed join must not overlap the LHS.
532 */
533 if (sjinfo->jointype != JOIN_LEFT ||
534 bms_overlap(joinrelids, sjinfo->min_lefthand))
535 return false; /* invalid join path */
536
537 /*
538 * To be valid, the proposed join must be a LEFT join; otherwise
539 * it can't associate into this SJ's RHS. But we may not yet have
540 * found the SpecialJoinInfo matching the proposed join, so we
541 * can't test that yet. Remember the requirement for later.
542 */
543 must_be_leftjoin = true;
544 }
545 }
546
547 /*
548 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
549 * proposed join can't associate into an SJ's RHS.
550 *
551 * Also, fail if the proposed join's predicate isn't strict; we're
552 * essentially checking to see if we can apply outer-join identity 3, and
553 * that's a requirement. (This check may be redundant with checks in
554 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
555 */
556 if (must_be_leftjoin &&
557 (match_sjinfo == NULL ||
558 match_sjinfo->jointype != JOIN_LEFT ||
559 !match_sjinfo->lhs_strict))
560 return false; /* invalid join path */
561
562 /*
563 * We also have to check for constraints imposed by LATERAL references.
564 */
565 if (root->hasLateralRTEs)
566 {
567 bool lateral_fwd;
568 bool lateral_rev;
569 Relids join_lateral_rels;
570
571 /*
572 * The proposed rels could each contain lateral references to the
573 * other, in which case the join is impossible. If there are lateral
574 * references in just one direction, then the join has to be done with
575 * a nestloop with the lateral referencer on the inside. If the join
576 * matches an SJ that cannot be implemented by such a nestloop, the
577 * join is impossible.
578 *
579 * Also, if the lateral reference is only indirect, we should reject
580 * the join; whatever rel(s) the reference chain goes through must be
581 * joined to first.
582 *
583 * Another case that might keep us from building a valid plan is the
584 * implementation restriction described by have_dangerous_phv().
585 */
586 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
587 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
588 if (lateral_fwd && lateral_rev)
589 return false; /* have lateral refs in both directions */
590 if (lateral_fwd)
591 {
592 /* has to be implemented as nestloop with rel1 on left */
593 if (match_sjinfo &&
594 (reversed ||
595 unique_ified ||
596 match_sjinfo->jointype == JOIN_FULL))
597 return false; /* not implementable as nestloop */
598 /* check there is a direct reference from rel2 to rel1 */
599 if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
600 return false; /* only indirect refs, so reject */
601 /* check we won't have a dangerous PHV */
602 if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
603 return false; /* might be unable to handle required PHV */
604 }
605 else if (lateral_rev)
606 {
607 /* has to be implemented as nestloop with rel2 on left */
608 if (match_sjinfo &&
609 (!reversed ||
610 unique_ified ||
611 match_sjinfo->jointype == JOIN_FULL))
612 return false; /* not implementable as nestloop */
613 /* check there is a direct reference from rel1 to rel2 */
614 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
615 return false; /* only indirect refs, so reject */
616 /* check we won't have a dangerous PHV */
617 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
618 return false; /* might be unable to handle required PHV */
619 }
620
621 /*
622 * LATERAL references could also cause problems later on if we accept
623 * this join: if the join's minimum parameterization includes any rels
624 * that would have to be on the inside of an outer join with this join
625 * rel, then it's never going to be possible to build the complete
626 * query using this join. We should reject this join not only because
627 * it'll save work, but because if we don't, the clauseless-join
628 * heuristics might think that legality of this join means that some
629 * other join rel need not be formed, and that could lead to failure
630 * to find any plan at all. We have to consider not only rels that
631 * are directly on the inner side of an OJ with the joinrel, but also
632 * ones that are indirectly so, so search to find all such rels.
633 */
634 join_lateral_rels = min_join_parameterization(root, joinrelids,
635 rel1, rel2);
636 if (join_lateral_rels)
637 {
638 Relids join_plus_rhs = bms_copy(joinrelids);
639 bool more;
640
641 do
642 {
643 more = false;
644 foreach(l, root->join_info_list)
645 {
646 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
647
648 /* ignore full joins --- their ordering is predetermined */
649 if (sjinfo->jointype == JOIN_FULL)
650 continue;
651
652 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
653 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
654 {
655 join_plus_rhs = bms_add_members(join_plus_rhs,
656 sjinfo->min_righthand);
657 more = true;
658 }
659 }
660 } while (more);
661 if (bms_overlap(join_plus_rhs, join_lateral_rels))
662 return false; /* will not be able to join to some RHS rel */
663 }
664 }
665
666 /* Otherwise, it's a valid join */
667 *sjinfo_p = match_sjinfo;
668 *reversed_p = reversed;
669 return true;
670 }
671
672
673 /*
674 * make_join_rel
675 * Find or create a join RelOptInfo that represents the join of
676 * the two given rels, and add to it path information for paths
677 * created with the two rels as outer and inner rel.
678 * (The join rel may already contain paths generated from other
679 * pairs of rels that add up to the same set of base rels.)
680 *
681 * NB: will return NULL if attempted join is not valid. This can happen
682 * when working with outer joins, or with IN or EXISTS clauses that have been
683 * turned into joins.
684 */
685 RelOptInfo *
make_join_rel(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2)686 make_join_rel(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2)
687 {
688 Relids joinrelids;
689 SpecialJoinInfo *sjinfo;
690 bool reversed;
691 SpecialJoinInfo sjinfo_data;
692 RelOptInfo *joinrel;
693 List *restrictlist;
694
695 /* We should never try to join two overlapping sets of rels. */
696 Assert(!bms_overlap(rel1->relids, rel2->relids));
697
698 /* Construct Relids set that identifies the joinrel. */
699 joinrelids = bms_union(rel1->relids, rel2->relids);
700
701 /* Check validity and determine join type. */
702 if (!join_is_legal(root, rel1, rel2, joinrelids,
703 &sjinfo, &reversed))
704 {
705 /* invalid join path */
706 bms_free(joinrelids);
707 return NULL;
708 }
709
710 /* Swap rels if needed to match the join info. */
711 if (reversed)
712 {
713 RelOptInfo *trel = rel1;
714
715 rel1 = rel2;
716 rel2 = trel;
717 }
718
719 /*
720 * If it's a plain inner join, then we won't have found anything in
721 * join_info_list. Make up a SpecialJoinInfo so that selectivity
722 * estimation functions will know what's being joined.
723 */
724 if (sjinfo == NULL)
725 {
726 sjinfo = &sjinfo_data;
727 sjinfo->type = T_SpecialJoinInfo;
728 sjinfo->min_lefthand = rel1->relids;
729 sjinfo->min_righthand = rel2->relids;
730 sjinfo->syn_lefthand = rel1->relids;
731 sjinfo->syn_righthand = rel2->relids;
732 sjinfo->jointype = JOIN_INNER;
733 /* we don't bother trying to make the remaining fields valid */
734 sjinfo->lhs_strict = false;
735 sjinfo->delay_upper_joins = false;
736 sjinfo->semi_can_btree = false;
737 sjinfo->semi_can_hash = false;
738 sjinfo->semi_operators = NIL;
739 sjinfo->semi_rhs_exprs = NIL;
740 }
741
742 /*
743 * Find or build the join RelOptInfo, and compute the restrictlist that
744 * goes with this particular joining.
745 */
746 joinrel = build_join_rel(root, joinrelids, rel1, rel2, sjinfo,
747 &restrictlist);
748
749 /*
750 * If we've already proven this join is empty, we needn't consider any
751 * more paths for it.
752 */
753 if (is_dummy_rel(joinrel))
754 {
755 bms_free(joinrelids);
756 return joinrel;
757 }
758
759 /* Add paths to the join relation. */
760 populate_joinrel_with_paths(root, rel1, rel2, joinrel, sjinfo,
761 restrictlist);
762
763 bms_free(joinrelids);
764
765 return joinrel;
766 }
767
768 /*
769 * populate_joinrel_with_paths
770 * Add paths to the given joinrel for given pair of joining relations. The
771 * SpecialJoinInfo provides details about the join and the restrictlist
772 * contains the join clauses and the other clauses applicable for given pair
773 * of the joining relations.
774 */
775 static void
populate_joinrel_with_paths(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2,RelOptInfo * joinrel,SpecialJoinInfo * sjinfo,List * restrictlist)776 populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
777 RelOptInfo *rel2, RelOptInfo *joinrel,
778 SpecialJoinInfo *sjinfo, List *restrictlist)
779 {
780 /*
781 * Consider paths using each rel as both outer and inner. Depending on
782 * the join type, a provably empty outer or inner rel might mean the join
783 * is provably empty too; in which case throw away any previously computed
784 * paths and mark the join as dummy. (We do it this way since it's
785 * conceivable that dummy-ness of a multi-element join might only be
786 * noticeable for certain construction paths.)
787 *
788 * Also, a provably constant-false join restriction typically means that
789 * we can skip evaluating one or both sides of the join. We do this by
790 * marking the appropriate rel as dummy. For outer joins, a
791 * constant-false restriction that is pushed down still means the whole
792 * join is dummy, while a non-pushed-down one means that no inner rows
793 * will join so we can treat the inner rel as dummy.
794 *
795 * We need only consider the jointypes that appear in join_info_list, plus
796 * JOIN_INNER.
797 */
798 switch (sjinfo->jointype)
799 {
800 case JOIN_INNER:
801 if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
802 restriction_is_constant_false(restrictlist, joinrel, false))
803 {
804 mark_dummy_rel(joinrel);
805 break;
806 }
807 add_paths_to_joinrel(root, joinrel, rel1, rel2,
808 JOIN_INNER, sjinfo,
809 restrictlist);
810 add_paths_to_joinrel(root, joinrel, rel2, rel1,
811 JOIN_INNER, sjinfo,
812 restrictlist);
813 break;
814 case JOIN_LEFT:
815 if (is_dummy_rel(rel1) ||
816 restriction_is_constant_false(restrictlist, joinrel, true))
817 {
818 mark_dummy_rel(joinrel);
819 break;
820 }
821 if (restriction_is_constant_false(restrictlist, joinrel, false) &&
822 bms_is_subset(rel2->relids, sjinfo->syn_righthand))
823 mark_dummy_rel(rel2);
824 add_paths_to_joinrel(root, joinrel, rel1, rel2,
825 JOIN_LEFT, sjinfo,
826 restrictlist);
827 add_paths_to_joinrel(root, joinrel, rel2, rel1,
828 JOIN_RIGHT, sjinfo,
829 restrictlist);
830 break;
831 case JOIN_FULL:
832 if ((is_dummy_rel(rel1) && is_dummy_rel(rel2)) ||
833 restriction_is_constant_false(restrictlist, joinrel, true))
834 {
835 mark_dummy_rel(joinrel);
836 break;
837 }
838 add_paths_to_joinrel(root, joinrel, rel1, rel2,
839 JOIN_FULL, sjinfo,
840 restrictlist);
841 add_paths_to_joinrel(root, joinrel, rel2, rel1,
842 JOIN_FULL, sjinfo,
843 restrictlist);
844
845 /*
846 * If there are join quals that aren't mergeable or hashable, we
847 * may not be able to build any valid plan. Complain here so that
848 * we can give a somewhat-useful error message. (Since we have no
849 * flexibility of planning for a full join, there's no chance of
850 * succeeding later with another pair of input rels.)
851 */
852 if (joinrel->pathlist == NIL)
853 ereport(ERROR,
854 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
855 errmsg("FULL JOIN is only supported with merge-joinable or hash-joinable join conditions")));
856 break;
857 case JOIN_SEMI:
858
859 /*
860 * We might have a normal semijoin, or a case where we don't have
861 * enough rels to do the semijoin but can unique-ify the RHS and
862 * then do an innerjoin (see comments in join_is_legal). In the
863 * latter case we can't apply JOIN_SEMI joining.
864 */
865 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
866 bms_is_subset(sjinfo->min_righthand, rel2->relids))
867 {
868 if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
869 restriction_is_constant_false(restrictlist, joinrel, false))
870 {
871 mark_dummy_rel(joinrel);
872 break;
873 }
874 add_paths_to_joinrel(root, joinrel, rel1, rel2,
875 JOIN_SEMI, sjinfo,
876 restrictlist);
877 }
878
879 /*
880 * If we know how to unique-ify the RHS and one input rel is
881 * exactly the RHS (not a superset) we can consider unique-ifying
882 * it and then doing a regular join. (The create_unique_path
883 * check here is probably redundant with what join_is_legal did,
884 * but if so the check is cheap because it's cached. So test
885 * anyway to be sure.)
886 */
887 if (bms_equal(sjinfo->syn_righthand, rel2->relids) &&
888 create_unique_path(root, rel2, rel2->cheapest_total_path,
889 sjinfo) != NULL)
890 {
891 if (is_dummy_rel(rel1) || is_dummy_rel(rel2) ||
892 restriction_is_constant_false(restrictlist, joinrel, false))
893 {
894 mark_dummy_rel(joinrel);
895 break;
896 }
897 add_paths_to_joinrel(root, joinrel, rel1, rel2,
898 JOIN_UNIQUE_INNER, sjinfo,
899 restrictlist);
900 add_paths_to_joinrel(root, joinrel, rel2, rel1,
901 JOIN_UNIQUE_OUTER, sjinfo,
902 restrictlist);
903 }
904 break;
905 case JOIN_ANTI:
906 if (is_dummy_rel(rel1) ||
907 restriction_is_constant_false(restrictlist, joinrel, true))
908 {
909 mark_dummy_rel(joinrel);
910 break;
911 }
912 if (restriction_is_constant_false(restrictlist, joinrel, false) &&
913 bms_is_subset(rel2->relids, sjinfo->syn_righthand))
914 mark_dummy_rel(rel2);
915 add_paths_to_joinrel(root, joinrel, rel1, rel2,
916 JOIN_ANTI, sjinfo,
917 restrictlist);
918 break;
919 default:
920 /* other values not expected here */
921 elog(ERROR, "unrecognized join type: %d", (int) sjinfo->jointype);
922 break;
923 }
924
925 /* Apply partitionwise join technique, if possible. */
926 try_partitionwise_join(root, rel1, rel2, joinrel, sjinfo, restrictlist);
927 }
928
929
930 /*
931 * have_join_order_restriction
932 * Detect whether the two relations should be joined to satisfy
933 * a join-order restriction arising from special or lateral joins.
934 *
935 * In practice this is always used with have_relevant_joinclause(), and so
936 * could be merged with that function, but it seems clearer to separate the
937 * two concerns. We need this test because there are degenerate cases where
938 * a clauseless join must be performed to satisfy join-order restrictions.
939 * Also, if one rel has a lateral reference to the other, or both are needed
940 * to compute some PHV, we should consider joining them even if the join would
941 * be clauseless.
942 *
943 * Note: this is only a problem if one side of a degenerate outer join
944 * contains multiple rels, or a clauseless join is required within an
945 * IN/EXISTS RHS; else we will find a join path via the "last ditch" case in
946 * join_search_one_level(). We could dispense with this test if we were
947 * willing to try bushy plans in the "last ditch" case, but that seems much
948 * less efficient.
949 */
950 bool
have_join_order_restriction(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2)951 have_join_order_restriction(PlannerInfo *root,
952 RelOptInfo *rel1, RelOptInfo *rel2)
953 {
954 bool result = false;
955 ListCell *l;
956
957 /*
958 * If either side has a direct lateral reference to the other, attempt the
959 * join regardless of outer-join considerations.
960 */
961 if (bms_overlap(rel1->relids, rel2->direct_lateral_relids) ||
962 bms_overlap(rel2->relids, rel1->direct_lateral_relids))
963 return true;
964
965 /*
966 * Likewise, if both rels are needed to compute some PlaceHolderVar,
967 * attempt the join regardless of outer-join considerations. (This is not
968 * very desirable, because a PHV with a large eval_at set will cause a lot
969 * of probably-useless joins to be considered, but failing to do this can
970 * cause us to fail to construct a plan at all.)
971 */
972 foreach(l, root->placeholder_list)
973 {
974 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
975
976 if (bms_is_subset(rel1->relids, phinfo->ph_eval_at) &&
977 bms_is_subset(rel2->relids, phinfo->ph_eval_at))
978 return true;
979 }
980
981 /*
982 * It's possible that the rels correspond to the left and right sides of a
983 * degenerate outer join, that is, one with no joinclause mentioning the
984 * non-nullable side; in which case we should force the join to occur.
985 *
986 * Also, the two rels could represent a clauseless join that has to be
987 * completed to build up the LHS or RHS of an outer join.
988 */
989 foreach(l, root->join_info_list)
990 {
991 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
992
993 /* ignore full joins --- other mechanisms handle them */
994 if (sjinfo->jointype == JOIN_FULL)
995 continue;
996
997 /* Can we perform the SJ with these rels? */
998 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
999 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1000 {
1001 result = true;
1002 break;
1003 }
1004 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1005 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1006 {
1007 result = true;
1008 break;
1009 }
1010
1011 /*
1012 * Might we need to join these rels to complete the RHS? We have to
1013 * use "overlap" tests since either rel might include a lower SJ that
1014 * has been proven to commute with this one.
1015 */
1016 if (bms_overlap(sjinfo->min_righthand, rel1->relids) &&
1017 bms_overlap(sjinfo->min_righthand, rel2->relids))
1018 {
1019 result = true;
1020 break;
1021 }
1022
1023 /* Likewise for the LHS. */
1024 if (bms_overlap(sjinfo->min_lefthand, rel1->relids) &&
1025 bms_overlap(sjinfo->min_lefthand, rel2->relids))
1026 {
1027 result = true;
1028 break;
1029 }
1030 }
1031
1032 /*
1033 * We do not force the join to occur if either input rel can legally be
1034 * joined to anything else using joinclauses. This essentially means that
1035 * clauseless bushy joins are put off as long as possible. The reason is
1036 * that when there is a join order restriction high up in the join tree
1037 * (that is, with many rels inside the LHS or RHS), we would otherwise
1038 * expend lots of effort considering very stupid join combinations within
1039 * its LHS or RHS.
1040 */
1041 if (result)
1042 {
1043 if (has_legal_joinclause(root, rel1) ||
1044 has_legal_joinclause(root, rel2))
1045 result = false;
1046 }
1047
1048 return result;
1049 }
1050
1051
1052 /*
1053 * has_join_restriction
1054 * Detect whether the specified relation has join-order restrictions,
1055 * due to being inside an outer join or an IN (sub-SELECT),
1056 * or participating in any LATERAL references or multi-rel PHVs.
1057 *
1058 * Essentially, this tests whether have_join_order_restriction() could
1059 * succeed with this rel and some other one. It's OK if we sometimes
1060 * say "true" incorrectly. (Therefore, we don't bother with the relatively
1061 * expensive has_legal_joinclause test.)
1062 */
1063 static bool
has_join_restriction(PlannerInfo * root,RelOptInfo * rel)1064 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1065 {
1066 ListCell *l;
1067
1068 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1069 return true;
1070
1071 foreach(l, root->placeholder_list)
1072 {
1073 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1074
1075 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1076 !bms_equal(rel->relids, phinfo->ph_eval_at))
1077 return true;
1078 }
1079
1080 foreach(l, root->join_info_list)
1081 {
1082 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1083
1084 /* ignore full joins --- other mechanisms preserve their ordering */
1085 if (sjinfo->jointype == JOIN_FULL)
1086 continue;
1087
1088 /* ignore if SJ is already contained in rel */
1089 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1090 bms_is_subset(sjinfo->min_righthand, rel->relids))
1091 continue;
1092
1093 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1094 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1095 bms_overlap(sjinfo->min_righthand, rel->relids))
1096 return true;
1097 }
1098
1099 return false;
1100 }
1101
1102
1103 /*
1104 * has_legal_joinclause
1105 * Detect whether the specified relation can legally be joined
1106 * to any other rels using join clauses.
1107 *
1108 * We consider only joins to single other relations in the current
1109 * initial_rels list. This is sufficient to get a "true" result in most real
1110 * queries, and an occasional erroneous "false" will only cost a bit more
1111 * planning time. The reason for this limitation is that considering joins to
1112 * other joins would require proving that the other join rel can legally be
1113 * formed, which seems like too much trouble for something that's only a
1114 * heuristic to save planning time. (Note: we must look at initial_rels
1115 * and not all of the query, since when we are planning a sub-joinlist we
1116 * may be forced to make clauseless joins within initial_rels even though
1117 * there are join clauses linking to other parts of the query.)
1118 */
1119 static bool
has_legal_joinclause(PlannerInfo * root,RelOptInfo * rel)1120 has_legal_joinclause(PlannerInfo *root, RelOptInfo *rel)
1121 {
1122 ListCell *lc;
1123
1124 foreach(lc, root->initial_rels)
1125 {
1126 RelOptInfo *rel2 = (RelOptInfo *) lfirst(lc);
1127
1128 /* ignore rels that are already in "rel" */
1129 if (bms_overlap(rel->relids, rel2->relids))
1130 continue;
1131
1132 if (have_relevant_joinclause(root, rel, rel2))
1133 {
1134 Relids joinrelids;
1135 SpecialJoinInfo *sjinfo;
1136 bool reversed;
1137
1138 /* join_is_legal needs relids of the union */
1139 joinrelids = bms_union(rel->relids, rel2->relids);
1140
1141 if (join_is_legal(root, rel, rel2, joinrelids,
1142 &sjinfo, &reversed))
1143 {
1144 /* Yes, this will work */
1145 bms_free(joinrelids);
1146 return true;
1147 }
1148
1149 bms_free(joinrelids);
1150 }
1151 }
1152
1153 return false;
1154 }
1155
1156
1157 /*
1158 * There's a pitfall for creating parameterized nestloops: suppose the inner
1159 * rel (call it A) has a parameter that is a PlaceHolderVar, and that PHV's
1160 * minimum eval_at set includes the outer rel (B) and some third rel (C).
1161 * We might think we could create a B/A nestloop join that's parameterized by
1162 * C. But we would end up with a plan in which the PHV's expression has to be
1163 * evaluated as a nestloop parameter at the B/A join; and the executor is only
1164 * set up to handle simple Vars as NestLoopParams. Rather than add complexity
1165 * and overhead to the executor for such corner cases, it seems better to
1166 * forbid the join. (Note that we can still make use of A's parameterized
1167 * path with pre-joined B+C as the outer rel. have_join_order_restriction()
1168 * ensures that we will consider making such a join even if there are not
1169 * other reasons to do so.)
1170 *
1171 * So we check whether any PHVs used in the query could pose such a hazard.
1172 * We don't have any simple way of checking whether a risky PHV would actually
1173 * be used in the inner plan, and the case is so unusual that it doesn't seem
1174 * worth working very hard on it.
1175 *
1176 * This needs to be checked in two places. If the inner rel's minimum
1177 * parameterization would trigger the restriction, then join_is_legal() should
1178 * reject the join altogether, because there will be no workable paths for it.
1179 * But joinpath.c has to check again for every proposed nestloop path, because
1180 * the inner path might have more than the minimum parameterization, causing
1181 * some PHV to be dangerous for it that otherwise wouldn't be.
1182 */
1183 bool
have_dangerous_phv(PlannerInfo * root,Relids outer_relids,Relids inner_params)1184 have_dangerous_phv(PlannerInfo *root,
1185 Relids outer_relids, Relids inner_params)
1186 {
1187 ListCell *lc;
1188
1189 foreach(lc, root->placeholder_list)
1190 {
1191 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
1192
1193 if (!bms_is_subset(phinfo->ph_eval_at, inner_params))
1194 continue; /* ignore, could not be a nestloop param */
1195 if (!bms_overlap(phinfo->ph_eval_at, outer_relids))
1196 continue; /* ignore, not relevant to this join */
1197 if (bms_is_subset(phinfo->ph_eval_at, outer_relids))
1198 continue; /* safe, it can be eval'd within outerrel */
1199 /* Otherwise, it's potentially unsafe, so reject the join */
1200 return true;
1201 }
1202
1203 /* OK to perform the join */
1204 return false;
1205 }
1206
1207
1208 /*
1209 * is_dummy_rel --- has relation been proven empty?
1210 */
1211 bool
is_dummy_rel(RelOptInfo * rel)1212 is_dummy_rel(RelOptInfo *rel)
1213 {
1214 Path *path;
1215
1216 /*
1217 * A rel that is known dummy will have just one path that is a childless
1218 * Append. (Even if somehow it has more paths, a childless Append will
1219 * have cost zero and hence should be at the front of the pathlist.)
1220 */
1221 if (rel->pathlist == NIL)
1222 return false;
1223 path = (Path *) linitial(rel->pathlist);
1224
1225 /*
1226 * Initially, a dummy path will just be a childless Append. But in later
1227 * planning stages we might stick a ProjectSetPath and/or ProjectionPath
1228 * on top, since Append can't project. Rather than make assumptions about
1229 * which combinations can occur, just descend through whatever we find.
1230 */
1231 for (;;)
1232 {
1233 if (IsA(path, ProjectionPath))
1234 path = ((ProjectionPath *) path)->subpath;
1235 else if (IsA(path, ProjectSetPath))
1236 path = ((ProjectSetPath *) path)->subpath;
1237 else
1238 break;
1239 }
1240 if (IS_DUMMY_APPEND(path))
1241 return true;
1242 return false;
1243 }
1244
1245 /*
1246 * Mark a relation as proven empty.
1247 *
1248 * During GEQO planning, this can get invoked more than once on the same
1249 * baserel struct, so it's worth checking to see if the rel is already marked
1250 * dummy.
1251 *
1252 * Also, when called during GEQO join planning, we are in a short-lived
1253 * memory context. We must make sure that the dummy path attached to a
1254 * baserel survives the GEQO cycle, else the baserel is trashed for future
1255 * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1256 * we don't want the dummy path to clutter the main planning context. Upshot
1257 * is that the best solution is to explicitly make the dummy path in the same
1258 * context the given RelOptInfo is in.
1259 */
1260 void
mark_dummy_rel(RelOptInfo * rel)1261 mark_dummy_rel(RelOptInfo *rel)
1262 {
1263 MemoryContext oldcontext;
1264
1265 /* Already marked? */
1266 if (is_dummy_rel(rel))
1267 return;
1268
1269 /* No, so choose correct context to make the dummy path in */
1270 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1271
1272 /* Set dummy size estimate */
1273 rel->rows = 0;
1274
1275 /* Evict any previously chosen paths */
1276 rel->pathlist = NIL;
1277 rel->partial_pathlist = NIL;
1278
1279 /* Set up the dummy path */
1280 add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
1281 NIL, rel->lateral_relids,
1282 0, false, NIL, -1));
1283
1284 /* Set or update cheapest_total_path and related fields */
1285 set_cheapest(rel);
1286
1287 MemoryContextSwitchTo(oldcontext);
1288 }
1289
1290
1291 /*
1292 * restriction_is_constant_false --- is a restrictlist just FALSE?
1293 *
1294 * In cases where a qual is provably constant FALSE, eval_const_expressions
1295 * will generally have thrown away anything that's ANDed with it. In outer
1296 * join situations this will leave us computing cartesian products only to
1297 * decide there's no match for an outer row, which is pretty stupid. So,
1298 * we need to detect the case.
1299 *
1300 * If only_pushed_down is true, then consider only quals that are pushed-down
1301 * from the point of view of the joinrel.
1302 */
1303 static bool
restriction_is_constant_false(List * restrictlist,RelOptInfo * joinrel,bool only_pushed_down)1304 restriction_is_constant_false(List *restrictlist,
1305 RelOptInfo *joinrel,
1306 bool only_pushed_down)
1307 {
1308 ListCell *lc;
1309
1310 /*
1311 * Despite the above comment, the restriction list we see here might
1312 * possibly have other members besides the FALSE constant, since other
1313 * quals could get "pushed down" to the outer join level. So we check
1314 * each member of the list.
1315 */
1316 foreach(lc, restrictlist)
1317 {
1318 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1319
1320 if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1321 continue;
1322
1323 if (rinfo->clause && IsA(rinfo->clause, Const))
1324 {
1325 Const *con = (Const *) rinfo->clause;
1326
1327 /* constant NULL is as good as constant FALSE for our purposes */
1328 if (con->constisnull)
1329 return true;
1330 if (!DatumGetBool(con->constvalue))
1331 return true;
1332 }
1333 }
1334 return false;
1335 }
1336
1337 /*
1338 * Assess whether join between given two partitioned relations can be broken
1339 * down into joins between matching partitions; a technique called
1340 * "partitionwise join"
1341 *
1342 * Partitionwise join is possible when a. Joining relations have same
1343 * partitioning scheme b. There exists an equi-join between the partition keys
1344 * of the two relations.
1345 *
1346 * Partitionwise join is planned as follows (details: optimizer/README.)
1347 *
1348 * 1. Create the RelOptInfos for joins between matching partitions i.e
1349 * child-joins and add paths to them.
1350 *
1351 * 2. Construct Append or MergeAppend paths across the set of child joins.
1352 * This second phase is implemented by generate_partitionwise_join_paths().
1353 *
1354 * The RelOptInfo, SpecialJoinInfo and restrictlist for each child join are
1355 * obtained by translating the respective parent join structures.
1356 */
1357 static void
try_partitionwise_join(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2,RelOptInfo * joinrel,SpecialJoinInfo * parent_sjinfo,List * parent_restrictlist)1358 try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1359 RelOptInfo *joinrel, SpecialJoinInfo *parent_sjinfo,
1360 List *parent_restrictlist)
1361 {
1362 bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1363 bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1364 List *parts1 = NIL;
1365 List *parts2 = NIL;
1366 ListCell *lcr1 = NULL;
1367 ListCell *lcr2 = NULL;
1368 int cnt_parts;
1369
1370 /* Guard against stack overflow due to overly deep partition hierarchy. */
1371 check_stack_depth();
1372
1373 /* Nothing to do, if the join relation is not partitioned. */
1374 if (joinrel->part_scheme == NULL || joinrel->nparts == 0)
1375 return;
1376
1377 /* The join relation should have consider_partitionwise_join set. */
1378 Assert(joinrel->consider_partitionwise_join);
1379
1380 /*
1381 * We can not perform partitionwise join if either of the joining
1382 * relations is not partitioned.
1383 */
1384 if (!IS_PARTITIONED_REL(rel1) || !IS_PARTITIONED_REL(rel2))
1385 return;
1386
1387 Assert(REL_HAS_ALL_PART_PROPS(rel1) && REL_HAS_ALL_PART_PROPS(rel2));
1388
1389 /* The joining relations should have consider_partitionwise_join set. */
1390 Assert(rel1->consider_partitionwise_join &&
1391 rel2->consider_partitionwise_join);
1392
1393 /*
1394 * The partition scheme of the join relation should match that of the
1395 * joining relations.
1396 */
1397 Assert(joinrel->part_scheme == rel1->part_scheme &&
1398 joinrel->part_scheme == rel2->part_scheme);
1399
1400 Assert(!(joinrel->partbounds_merged && (joinrel->nparts <= 0)));
1401
1402 compute_partition_bounds(root, rel1, rel2, joinrel, parent_sjinfo,
1403 &parts1, &parts2);
1404
1405 if (joinrel->partbounds_merged)
1406 {
1407 lcr1 = list_head(parts1);
1408 lcr2 = list_head(parts2);
1409 }
1410
1411 /*
1412 * Create child-join relations for this partitioned join, if those don't
1413 * exist. Add paths to child-joins for a pair of child relations
1414 * corresponding to the given pair of parent relations.
1415 */
1416 for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1417 {
1418 RelOptInfo *child_rel1;
1419 RelOptInfo *child_rel2;
1420 bool rel1_empty;
1421 bool rel2_empty;
1422 SpecialJoinInfo *child_sjinfo;
1423 List *child_restrictlist;
1424 RelOptInfo *child_joinrel;
1425 Relids child_joinrelids;
1426 AppendRelInfo **appinfos;
1427 int nappinfos;
1428
1429 if (joinrel->partbounds_merged)
1430 {
1431 child_rel1 = lfirst_node(RelOptInfo, lcr1);
1432 child_rel2 = lfirst_node(RelOptInfo, lcr2);
1433 lcr1 = lnext(parts1, lcr1);
1434 lcr2 = lnext(parts2, lcr2);
1435 }
1436 else
1437 {
1438 child_rel1 = rel1->part_rels[cnt_parts];
1439 child_rel2 = rel2->part_rels[cnt_parts];
1440 }
1441
1442 rel1_empty = (child_rel1 == NULL || IS_DUMMY_REL(child_rel1));
1443 rel2_empty = (child_rel2 == NULL || IS_DUMMY_REL(child_rel2));
1444
1445 /*
1446 * Check for cases where we can prove that this segment of the join
1447 * returns no rows, due to one or both inputs being empty (including
1448 * inputs that have been pruned away entirely). If so just ignore it.
1449 * These rules are equivalent to populate_joinrel_with_paths's rules
1450 * for dummy input relations.
1451 */
1452 switch (parent_sjinfo->jointype)
1453 {
1454 case JOIN_INNER:
1455 case JOIN_SEMI:
1456 if (rel1_empty || rel2_empty)
1457 continue; /* ignore this join segment */
1458 break;
1459 case JOIN_LEFT:
1460 case JOIN_ANTI:
1461 if (rel1_empty)
1462 continue; /* ignore this join segment */
1463 break;
1464 case JOIN_FULL:
1465 if (rel1_empty && rel2_empty)
1466 continue; /* ignore this join segment */
1467 break;
1468 default:
1469 /* other values not expected here */
1470 elog(ERROR, "unrecognized join type: %d",
1471 (int) parent_sjinfo->jointype);
1472 break;
1473 }
1474
1475 /*
1476 * If a child has been pruned entirely then we can't generate paths
1477 * for it, so we have to reject partitionwise joining unless we were
1478 * able to eliminate this partition above.
1479 */
1480 if (child_rel1 == NULL || child_rel2 == NULL)
1481 {
1482 /*
1483 * Mark the joinrel as unpartitioned so that later functions treat
1484 * it correctly.
1485 */
1486 joinrel->nparts = 0;
1487 return;
1488 }
1489
1490 /*
1491 * If a leaf relation has consider_partitionwise_join=false, it means
1492 * that it's a dummy relation for which we skipped setting up tlist
1493 * expressions and adding EC members in set_append_rel_size(), so
1494 * again we have to fail here.
1495 */
1496 if (rel1_is_simple && !child_rel1->consider_partitionwise_join)
1497 {
1498 Assert(child_rel1->reloptkind == RELOPT_OTHER_MEMBER_REL);
1499 Assert(IS_DUMMY_REL(child_rel1));
1500 joinrel->nparts = 0;
1501 return;
1502 }
1503 if (rel2_is_simple && !child_rel2->consider_partitionwise_join)
1504 {
1505 Assert(child_rel2->reloptkind == RELOPT_OTHER_MEMBER_REL);
1506 Assert(IS_DUMMY_REL(child_rel2));
1507 joinrel->nparts = 0;
1508 return;
1509 }
1510
1511 /* We should never try to join two overlapping sets of rels. */
1512 Assert(!bms_overlap(child_rel1->relids, child_rel2->relids));
1513 child_joinrelids = bms_union(child_rel1->relids, child_rel2->relids);
1514 appinfos = find_appinfos_by_relids(root, child_joinrelids, &nappinfos);
1515
1516 /*
1517 * Construct SpecialJoinInfo from parent join relations's
1518 * SpecialJoinInfo.
1519 */
1520 child_sjinfo = build_child_join_sjinfo(root, parent_sjinfo,
1521 child_rel1->relids,
1522 child_rel2->relids);
1523
1524 /*
1525 * Construct restrictions applicable to the child join from those
1526 * applicable to the parent join.
1527 */
1528 child_restrictlist =
1529 (List *) adjust_appendrel_attrs(root,
1530 (Node *) parent_restrictlist,
1531 nappinfos, appinfos);
1532 pfree(appinfos);
1533
1534 child_joinrel = joinrel->part_rels[cnt_parts];
1535 if (!child_joinrel)
1536 {
1537 child_joinrel = build_child_join_rel(root, child_rel1, child_rel2,
1538 joinrel, child_restrictlist,
1539 child_sjinfo,
1540 child_sjinfo->jointype);
1541 joinrel->part_rels[cnt_parts] = child_joinrel;
1542 joinrel->all_partrels = bms_add_members(joinrel->all_partrels,
1543 child_joinrel->relids);
1544 }
1545
1546 Assert(bms_equal(child_joinrel->relids, child_joinrelids));
1547
1548 populate_joinrel_with_paths(root, child_rel1, child_rel2,
1549 child_joinrel, child_sjinfo,
1550 child_restrictlist);
1551 }
1552 }
1553
1554 /*
1555 * Construct the SpecialJoinInfo for a child-join by translating
1556 * SpecialJoinInfo for the join between parents. left_relids and right_relids
1557 * are the relids of left and right side of the join respectively.
1558 */
1559 static SpecialJoinInfo *
build_child_join_sjinfo(PlannerInfo * root,SpecialJoinInfo * parent_sjinfo,Relids left_relids,Relids right_relids)1560 build_child_join_sjinfo(PlannerInfo *root, SpecialJoinInfo *parent_sjinfo,
1561 Relids left_relids, Relids right_relids)
1562 {
1563 SpecialJoinInfo *sjinfo = makeNode(SpecialJoinInfo);
1564 AppendRelInfo **left_appinfos;
1565 int left_nappinfos;
1566 AppendRelInfo **right_appinfos;
1567 int right_nappinfos;
1568
1569 memcpy(sjinfo, parent_sjinfo, sizeof(SpecialJoinInfo));
1570 left_appinfos = find_appinfos_by_relids(root, left_relids,
1571 &left_nappinfos);
1572 right_appinfos = find_appinfos_by_relids(root, right_relids,
1573 &right_nappinfos);
1574
1575 sjinfo->min_lefthand = adjust_child_relids(sjinfo->min_lefthand,
1576 left_nappinfos, left_appinfos);
1577 sjinfo->min_righthand = adjust_child_relids(sjinfo->min_righthand,
1578 right_nappinfos,
1579 right_appinfos);
1580 sjinfo->syn_lefthand = adjust_child_relids(sjinfo->syn_lefthand,
1581 left_nappinfos, left_appinfos);
1582 sjinfo->syn_righthand = adjust_child_relids(sjinfo->syn_righthand,
1583 right_nappinfos,
1584 right_appinfos);
1585 sjinfo->semi_rhs_exprs = (List *) adjust_appendrel_attrs(root,
1586 (Node *) sjinfo->semi_rhs_exprs,
1587 right_nappinfos,
1588 right_appinfos);
1589
1590 pfree(left_appinfos);
1591 pfree(right_appinfos);
1592
1593 return sjinfo;
1594 }
1595
1596 /*
1597 * compute_partition_bounds
1598 * Compute the partition bounds for a join rel from those for inputs
1599 */
1600 static void
compute_partition_bounds(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2,RelOptInfo * joinrel,SpecialJoinInfo * parent_sjinfo,List ** parts1,List ** parts2)1601 compute_partition_bounds(PlannerInfo *root, RelOptInfo *rel1,
1602 RelOptInfo *rel2, RelOptInfo *joinrel,
1603 SpecialJoinInfo *parent_sjinfo,
1604 List **parts1, List **parts2)
1605 {
1606 /*
1607 * If we don't have the partition bounds for the join rel yet, try to
1608 * compute those along with pairs of partitions to be joined.
1609 */
1610 if (joinrel->nparts == -1)
1611 {
1612 PartitionScheme part_scheme = joinrel->part_scheme;
1613 PartitionBoundInfo boundinfo = NULL;
1614 int nparts = 0;
1615
1616 Assert(joinrel->boundinfo == NULL);
1617 Assert(joinrel->part_rels == NULL);
1618
1619 /*
1620 * See if the partition bounds for inputs are exactly the same, in
1621 * which case we don't need to work hard: the join rel will have the
1622 * same partition bounds as inputs, and the partitions with the same
1623 * cardinal positions will form the pairs.
1624 *
1625 * Note: even in cases where one or both inputs have merged bounds, it
1626 * would be possible for both the bounds to be exactly the same, but
1627 * it seems unlikely to be worth the cycles to check.
1628 */
1629 if (!rel1->partbounds_merged &&
1630 !rel2->partbounds_merged &&
1631 rel1->nparts == rel2->nparts &&
1632 partition_bounds_equal(part_scheme->partnatts,
1633 part_scheme->parttyplen,
1634 part_scheme->parttypbyval,
1635 rel1->boundinfo, rel2->boundinfo))
1636 {
1637 boundinfo = rel1->boundinfo;
1638 nparts = rel1->nparts;
1639 }
1640 else
1641 {
1642 /* Try merging the partition bounds for inputs. */
1643 boundinfo = partition_bounds_merge(part_scheme->partnatts,
1644 part_scheme->partsupfunc,
1645 part_scheme->partcollation,
1646 rel1, rel2,
1647 parent_sjinfo->jointype,
1648 parts1, parts2);
1649 if (boundinfo == NULL)
1650 {
1651 joinrel->nparts = 0;
1652 return;
1653 }
1654 nparts = list_length(*parts1);
1655 joinrel->partbounds_merged = true;
1656 }
1657
1658 Assert(nparts > 0);
1659 joinrel->boundinfo = boundinfo;
1660 joinrel->nparts = nparts;
1661 joinrel->part_rels =
1662 (RelOptInfo **) palloc0(sizeof(RelOptInfo *) * nparts);
1663 }
1664 else
1665 {
1666 Assert(joinrel->nparts > 0);
1667 Assert(joinrel->boundinfo);
1668 Assert(joinrel->part_rels);
1669
1670 /*
1671 * If the join rel's partbounds_merged flag is true, it means inputs
1672 * are not guaranteed to have the same partition bounds, therefore we
1673 * can't assume that the partitions at the same cardinal positions
1674 * form the pairs; let get_matching_part_pairs() generate the pairs.
1675 * Otherwise, nothing to do since we can assume that.
1676 */
1677 if (joinrel->partbounds_merged)
1678 {
1679 get_matching_part_pairs(root, joinrel, rel1, rel2,
1680 parts1, parts2);
1681 Assert(list_length(*parts1) == joinrel->nparts);
1682 Assert(list_length(*parts2) == joinrel->nparts);
1683 }
1684 }
1685 }
1686
1687 /*
1688 * get_matching_part_pairs
1689 * Generate pairs of partitions to be joined from inputs
1690 */
1691 static void
get_matching_part_pairs(PlannerInfo * root,RelOptInfo * joinrel,RelOptInfo * rel1,RelOptInfo * rel2,List ** parts1,List ** parts2)1692 get_matching_part_pairs(PlannerInfo *root, RelOptInfo *joinrel,
1693 RelOptInfo *rel1, RelOptInfo *rel2,
1694 List **parts1, List **parts2)
1695 {
1696 bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1697 bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1698 int cnt_parts;
1699
1700 *parts1 = NIL;
1701 *parts2 = NIL;
1702
1703 for (cnt_parts = 0; cnt_parts < joinrel->nparts; cnt_parts++)
1704 {
1705 RelOptInfo *child_joinrel = joinrel->part_rels[cnt_parts];
1706 RelOptInfo *child_rel1;
1707 RelOptInfo *child_rel2;
1708 Relids child_relids1;
1709 Relids child_relids2;
1710
1711 /*
1712 * If this segment of the join is empty, it means that this segment
1713 * was ignored when previously creating child-join paths for it in
1714 * try_partitionwise_join() as it would not contribute to the join
1715 * result, due to one or both inputs being empty; add NULL to each of
1716 * the given lists so that this segment will be ignored again in that
1717 * function.
1718 */
1719 if (!child_joinrel)
1720 {
1721 *parts1 = lappend(*parts1, NULL);
1722 *parts2 = lappend(*parts2, NULL);
1723 continue;
1724 }
1725
1726 /*
1727 * Get a relids set of partition(s) involved in this join segment that
1728 * are from the rel1 side.
1729 */
1730 child_relids1 = bms_intersect(child_joinrel->relids,
1731 rel1->all_partrels);
1732 Assert(bms_num_members(child_relids1) == bms_num_members(rel1->relids));
1733
1734 /*
1735 * Get a child rel for rel1 with the relids. Note that we should have
1736 * the child rel even if rel1 is a join rel, because in that case the
1737 * partitions specified in the relids would have matching/overlapping
1738 * boundaries, so the specified partitions should be considered as
1739 * ones to be joined when planning partitionwise joins of rel1,
1740 * meaning that the child rel would have been built by the time we get
1741 * here.
1742 */
1743 if (rel1_is_simple)
1744 {
1745 int varno = bms_singleton_member(child_relids1);
1746
1747 child_rel1 = find_base_rel(root, varno);
1748 }
1749 else
1750 child_rel1 = find_join_rel(root, child_relids1);
1751 Assert(child_rel1);
1752
1753 /*
1754 * Get a relids set of partition(s) involved in this join segment that
1755 * are from the rel2 side.
1756 */
1757 child_relids2 = bms_intersect(child_joinrel->relids,
1758 rel2->all_partrels);
1759 Assert(bms_num_members(child_relids2) == bms_num_members(rel2->relids));
1760
1761 /*
1762 * Get a child rel for rel2 with the relids. See above comments.
1763 */
1764 if (rel2_is_simple)
1765 {
1766 int varno = bms_singleton_member(child_relids2);
1767
1768 child_rel2 = find_base_rel(root, varno);
1769 }
1770 else
1771 child_rel2 = find_join_rel(root, child_relids2);
1772 Assert(child_rel2);
1773
1774 /*
1775 * The join of rel1 and rel2 is legal, so is the join of the child
1776 * rels obtained above; add them to the given lists as a join pair
1777 * producing this join segment.
1778 */
1779 *parts1 = lappend(*parts1, child_rel1);
1780 *parts2 = lappend(*parts2, child_rel2);
1781 }
1782 }
1783