1 /*-------------------------------------------------------------------------
4 * Routines copied from PostgreSQL core distribution.
6 * The main purpose of this files is having access to static functions in core.
7 * Another purpose is tweaking functions behavior by replacing part of them by
8 * macro definitions. See at the end of pg_hint_plan.c for details. Anyway,
9 * this file *must* contain required functions without making any change.
11 * This file contains the following functions from corresponding files.
13 * src/backend/optimizer/path/allpaths.c
16 * standard_join_search(): This funcion is not static. The reason for
17 * including this function is make_rels_by_clause_joins. In order to
18 * avoid generating apparently unwanted join combination, we decided to
19 * change the behavior of make_join_rel, which is called under this
23 * set_plain_rel_pathlist()
24 * set_append_rel_pathlist()
25 * create_plain_partial_paths()
27 * src/backend/optimizer/path/joinrels.c
30 * join_search_one_level(): We have to modify this to call my definition of
31 * make_rels_by_clause_joins.
34 * make_rels_by_clause_joins()
35 * make_rels_by_clauseless_joins()
37 * has_join_restriction()
39 * restriction_is_constant_false()
40 * try_partitionwise_join()
42 * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
43 * Portions Copyright (c) 1994, Regents of the University of California
45 *-------------------------------------------------------------------------
48 static void populate_joinrel_with_paths(PlannerInfo *root, RelOptInfo *rel1,
49 RelOptInfo *rel2, RelOptInfo *joinrel,
50 SpecialJoinInfo *sjinfo, List *restrictlist);
53 * set_plain_rel_pathlist
54 * Build access paths for a plain relation (no subquery, no inheritance)
57 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
59 Relids required_outer;
62 * We don't support pushing join clauses into the quals of a seqscan, but
63 * it could still have required parameterization due to LATERAL refs in
66 required_outer = rel->lateral_relids;
68 /* Consider sequential scan */
69 add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
71 /* If appropriate, consider parallel sequential scan */
72 if (rel->consider_parallel && required_outer == NULL)
73 create_plain_partial_paths(root, rel);
75 /* Consider index scans */
76 create_index_paths(root, rel);
78 /* Consider TID scans */
79 create_tidscan_paths(root, rel);
84 * set_append_rel_pathlist
85 * Build access paths for an "append relation"
88 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
89 Index rti, RangeTblEntry *rte)
91 int parentRTindex = rti;
92 List *live_childrels = NIL;
96 * Generate access paths for each member relation, and remember the
99 foreach(l, root->append_rel_list)
101 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
103 RangeTblEntry *childRTE;
104 RelOptInfo *childrel;
106 /* append_rel_list contains all append rels; ignore others */
107 if (appinfo->parent_relid != parentRTindex)
110 /* Re-locate the child RTE and RelOptInfo */
111 childRTindex = appinfo->child_relid;
112 childRTE = root->simple_rte_array[childRTindex];
113 childrel = root->simple_rel_array[childRTindex];
116 * If set_append_rel_size() decided the parent appendrel was
117 * parallel-unsafe at some point after visiting this child rel, we
118 * need to propagate the unsafety marking down to the child, so that
119 * we don't generate useless partial paths for it.
121 if (!rel->consider_parallel)
122 childrel->consider_parallel = false;
125 * Compute the child's access paths.
127 set_rel_pathlist(root, childrel, childRTindex, childRTE);
130 * If child is dummy, ignore it.
132 if (IS_DUMMY_REL(childrel))
135 /* Bubble up childrel's partitioned children. */
136 if (rel->part_scheme)
137 rel->partitioned_child_rels =
138 list_concat(rel->partitioned_child_rels,
139 list_copy(childrel->partitioned_child_rels));
142 * Child is live, so add it to the live_childrels list for use below.
144 live_childrels = lappend(live_childrels, childrel);
147 /* Add paths to the append relation. */
148 add_paths_to_append_rel(root, rel, live_childrels);
153 * standard_join_search
154 * Find possible joinpaths for a query by successively finding ways
155 * to join component relations into join relations.
157 * 'levels_needed' is the number of iterations needed, ie, the number of
158 * independent jointree items in the query. This is > 1.
160 * 'initial_rels' is a list of RelOptInfo nodes for each independent
161 * jointree item. These are the components to be joined together.
162 * Note that levels_needed == list_length(initial_rels).
164 * Returns the final level of join relations, i.e., the relation that is
165 * the result of joining all the original relations together.
166 * At least one implementation path must be provided for this relation and
167 * all required sub-relations.
169 * To support loadable plugins that modify planner behavior by changing the
170 * join searching algorithm, we provide a hook variable that lets a plugin
171 * replace or supplement this function. Any such hook must return the same
172 * final join relation as the standard code would, but it might have a
173 * different set of implementation paths attached, and only the sub-joinrels
174 * needed for these paths need have been instantiated.
176 * Note to plugin authors: the functions invoked during standard_join_search()
177 * modify root->join_rel_list and root->join_rel_hash. If you want to do more
178 * than one join-order search, you'll probably need to save and restore the
179 * original states of those data structures. See geqo_eval() for an example.
182 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
188 * This function cannot be invoked recursively within any one planning
189 * problem, so join_rel_level[] can't be in use already.
191 Assert(root->join_rel_level == NULL);
194 * We employ a simple "dynamic programming" algorithm: we first find all
195 * ways to build joins of two jointree items, then all ways to build joins
196 * of three items (from two-item joins and single items), then four-item
197 * joins, and so on until we have considered all ways to join all the
198 * items into one rel.
200 * root->join_rel_level[j] is a list of all the j-item rels. Initially we
201 * set root->join_rel_level[1] to represent all the single-jointree-item
204 root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
206 root->join_rel_level[1] = initial_rels;
208 for (lev = 2; lev <= levels_needed; lev++)
213 * Determine all possible pairs of relations to be joined at this
214 * level, and build paths for making each one from every available
215 * pair of lower-level relations.
217 join_search_one_level(root, lev);
220 * Run generate_partitionwise_join_paths() and generate_gather_paths()
221 * for each just-processed joinrel. We could not do this earlier
222 * because both regular and partial paths can get added to a
223 * particular joinrel at multiple times within join_search_one_level.
225 * After that, we're done creating paths for the joinrel, so run
228 foreach(lc, root->join_rel_level[lev])
230 rel = (RelOptInfo *) lfirst(lc);
232 /* Create paths for partitionwise joins. */
233 generate_partitionwise_join_paths(root, rel);
236 * Except for the topmost scan/join rel, consider gathering
237 * partial paths. We'll do the same for the topmost scan/join rel
238 * once we know the final targetlist (see grouping_planner).
240 if (lev < levels_needed)
241 generate_gather_paths(root, rel, false);
243 /* Find and save the cheapest paths for this rel */
246 #ifdef OPTIMIZER_DEBUG
247 debug_print_rel(root, rel);
253 * We should have a single rel at the final level.
255 if (root->join_rel_level[levels_needed] == NIL)
256 elog(ERROR, "failed to build any %d-way joins", levels_needed);
257 Assert(list_length(root->join_rel_level[levels_needed]) == 1);
259 rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
261 root->join_rel_level = NULL;
268 * create_plain_partial_paths
269 * Build partial access paths for parallel scan of a plain relation
272 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
274 int parallel_workers;
276 parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
277 max_parallel_workers_per_gather);
279 /* If any limit was set to zero, the user doesn't want a parallel scan. */
280 if (parallel_workers <= 0)
283 /* Add an unordered partial path based on a parallel sequential scan. */
284 add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
289 * join_search_one_level
290 * Consider ways to produce join relations containing exactly 'level'
291 * jointree items. (This is one step of the dynamic-programming method
292 * embodied in standard_join_search.) Join rel nodes for each feasible
293 * combination of lower-level rels are created and returned in a list.
294 * Implementation paths are created for each such joinrel, too.
296 * level: level of rels we want to make this time
297 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
299 * The result is returned in root->join_rel_level[level].
302 join_search_one_level(PlannerInfo *root, int level)
304 List **joinrels = root->join_rel_level;
308 Assert(joinrels[level] == NIL);
310 /* Set join_cur_level so that new joinrels are added to proper list */
311 root->join_cur_level = level;
314 * First, consider left-sided and right-sided plans, in which rels of
315 * exactly level-1 member relations are joined against initial relations.
316 * We prefer to join using join clauses, but if we find a rel of level-1
317 * members that has no join clauses, we will generate Cartesian-product
318 * joins against all initial rels not already contained in it.
320 foreach(r, joinrels[level - 1])
322 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
324 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
325 has_join_restriction(root, old_rel))
328 * There are join clauses or join order restrictions relevant to
329 * this rel, so consider joins between this rel and (only) those
330 * initial rels it is linked to by a clause or restriction.
332 * At level 2 this condition is symmetric, so there is no need to
333 * look at initial rels before this one in the list; we already
334 * considered such joins when we were at the earlier rel. (The
335 * mirror-image joins are handled automatically by make_join_rel.)
336 * In later passes (level > 2), we join rels of the previous level
337 * to each initial rel they don't already include but have a join
338 * clause or restriction with.
340 ListCell *other_rels;
342 if (level == 2) /* consider remaining initial rels */
343 other_rels = lnext(r);
344 else /* consider all initial rels */
345 other_rels = list_head(joinrels[1]);
347 make_rels_by_clause_joins(root,
354 * Oops, we have a relation that is not joined to any other
355 * relation, either directly or by join-order restrictions.
356 * Cartesian product time.
358 * We consider a cartesian product with each not-already-included
359 * initial rel, whether it has other join clauses or not. At
360 * level 2, if there are two or more clauseless initial rels, we
361 * will redundantly consider joining them in both directions; but
362 * such cases aren't common enough to justify adding complexity to
363 * avoid the duplicated effort.
365 make_rels_by_clauseless_joins(root,
367 list_head(joinrels[1]));
372 * Now, consider "bushy plans" in which relations of k initial rels are
373 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
375 * We only consider bushy-plan joins for pairs of rels where there is a
376 * suitable join clause (or join order restriction), in order to avoid
377 * unreasonable growth of planning time.
381 int other_level = level - k;
384 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
385 * need to go as far as the halfway point.
390 foreach(r, joinrels[k])
392 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
393 ListCell *other_rels;
397 * We can ignore relations without join clauses here, unless they
398 * participate in join-order restrictions --- then we might have
399 * to force a bushy join plan.
401 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
402 !has_join_restriction(root, old_rel))
405 if (k == other_level)
406 other_rels = lnext(r); /* only consider remaining rels */
408 other_rels = list_head(joinrels[other_level]);
410 for_each_cell(r2, other_rels)
412 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
414 if (!bms_overlap(old_rel->relids, new_rel->relids))
417 * OK, we can build a rel of the right level from this
418 * pair of rels. Do so if there is at least one relevant
419 * join clause or join order restriction.
421 if (have_relevant_joinclause(root, old_rel, new_rel) ||
422 have_join_order_restriction(root, old_rel, new_rel))
424 (void) make_join_rel(root, old_rel, new_rel);
432 * Last-ditch effort: if we failed to find any usable joins so far, force
433 * a set of cartesian-product joins to be generated. This handles the
434 * special case where all the available rels have join clauses but we
435 * cannot use any of those clauses yet. This can only happen when we are
436 * considering a join sub-problem (a sub-joinlist) and all the rels in the
437 * sub-problem have only join clauses with rels outside the sub-problem.
440 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
441 * WHERE a.w = c.x and b.y = d.z;
443 * If the "a INNER JOIN b" sub-problem does not get flattened into the
444 * upper level, we must be willing to make a cartesian join of a and b;
445 * but the code above will not have done so, because it thought that both
446 * a and b have joinclauses. We consider only left-sided and right-sided
447 * cartesian joins in this case (no bushy).
450 if (joinrels[level] == NIL)
453 * This loop is just like the first one, except we always call
454 * make_rels_by_clauseless_joins().
456 foreach(r, joinrels[level - 1])
458 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
460 make_rels_by_clauseless_joins(root,
462 list_head(joinrels[1]));
466 * When special joins are involved, there may be no legal way
467 * to make an N-way join for some values of N. For example consider
469 * SELECT ... FROM t1 WHERE
470 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
471 * y IN (SELECT ... FROM t4,t5 WHERE ...)
473 * We will flatten this query to a 5-way join problem, but there are
474 * no 4-way joins that join_is_legal() will consider legal. We have
475 * to accept failure at level 4 and go on to discover a workable
476 * bushy plan at level 5.
478 * However, if there are no special joins and no lateral references
479 * then join_is_legal() should never fail, and so the following sanity
483 if (joinrels[level] == NIL &&
484 root->join_info_list == NIL &&
485 !root->hasLateralRTEs)
486 elog(ERROR, "failed to build any %d-way joins", level);
492 * make_rels_by_clause_joins
493 * Build joins between the given relation 'old_rel' and other relations
494 * that participate in join clauses that 'old_rel' also participates in
495 * (or participate in join-order restrictions with it).
496 * The join rels are returned in root->join_rel_level[join_cur_level].
498 * Note: at levels above 2 we will generate the same joined relation in
499 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
500 * (b join c) join a, though the second case will add a different set of Paths
501 * to it. This is the reason for using the join_rel_level mechanism, which
502 * automatically ensures that each new joinrel is only added to the list once.
504 * 'old_rel' is the relation entry for the relation to be joined
505 * 'other_rels': the first cell in a linked list containing the other
506 * rels to be considered for joining
508 * Currently, this is only used with initial rels in other_rels, but it
509 * will work for joining to joinrels too.
512 make_rels_by_clause_joins(PlannerInfo *root,
514 ListCell *other_rels)
518 for_each_cell(l, other_rels)
520 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
522 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
523 (have_relevant_joinclause(root, old_rel, other_rel) ||
524 have_join_order_restriction(root, old_rel, other_rel)))
526 (void) make_join_rel(root, old_rel, other_rel);
533 * make_rels_by_clauseless_joins
534 * Given a relation 'old_rel' and a list of other relations
535 * 'other_rels', create a join relation between 'old_rel' and each
536 * member of 'other_rels' that isn't already included in 'old_rel'.
537 * The join rels are returned in root->join_rel_level[join_cur_level].
539 * 'old_rel' is the relation entry for the relation to be joined
540 * 'other_rels': the first cell of a linked list containing the
541 * other rels to be considered for joining
543 * Currently, this is only used with initial rels in other_rels, but it would
544 * work for joining to joinrels too.
547 make_rels_by_clauseless_joins(PlannerInfo *root,
549 ListCell *other_rels)
553 for_each_cell(l, other_rels)
555 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
557 if (!bms_overlap(other_rel->relids, old_rel->relids))
559 (void) make_join_rel(root, old_rel, other_rel);
567 * Determine whether a proposed join is legal given the query's
568 * join order constraints; and if it is, determine the join type.
570 * Caller must supply not only the two rels, but the union of their relids.
571 * (We could simplify the API by computing joinrelids locally, but this
572 * would be redundant work in the normal path through make_join_rel.)
574 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
575 * else it's set to point to the associated SpecialJoinInfo node. Also,
576 * *reversed_p is set true if the given relations need to be swapped to
577 * match the SpecialJoinInfo node.
580 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
582 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
584 SpecialJoinInfo *match_sjinfo;
587 bool must_be_leftjoin;
591 * Ensure output params are set on failure return. This is just to
592 * suppress uninitialized-variable warnings from overly anal compilers.
598 * If we have any special joins, the proposed join might be illegal; and
599 * in any case we have to determine its join type. Scan the join info
600 * list for matches and conflicts.
604 unique_ified = false;
605 must_be_leftjoin = false;
607 foreach(l, root->join_info_list)
609 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
612 * This special join is not relevant unless its RHS overlaps the
613 * proposed join. (Check this first as a fast path for dismissing
614 * most irrelevant SJs quickly.)
616 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
620 * Also, not relevant if proposed join is fully contained within RHS
621 * (ie, we're still building up the RHS).
623 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
627 * Also, not relevant if SJ is already done within either input.
629 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
630 bms_is_subset(sjinfo->min_righthand, rel1->relids))
632 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
633 bms_is_subset(sjinfo->min_righthand, rel2->relids))
637 * If it's a semijoin and we already joined the RHS to any other rels
638 * within either input, then we must have unique-ified the RHS at that
639 * point (see below). Therefore the semijoin is no longer relevant in
642 if (sjinfo->jointype == JOIN_SEMI)
644 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
645 !bms_equal(sjinfo->syn_righthand, rel1->relids))
647 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
648 !bms_equal(sjinfo->syn_righthand, rel2->relids))
653 * If one input contains min_lefthand and the other contains
654 * min_righthand, then we can perform the SJ at this join.
656 * Reject if we get matches to more than one SJ; that implies we're
657 * considering something that's not really valid.
659 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
660 bms_is_subset(sjinfo->min_righthand, rel2->relids))
663 return false; /* invalid join path */
664 match_sjinfo = sjinfo;
667 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
668 bms_is_subset(sjinfo->min_righthand, rel1->relids))
671 return false; /* invalid join path */
672 match_sjinfo = sjinfo;
675 else if (sjinfo->jointype == JOIN_SEMI &&
676 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
677 create_unique_path(root, rel2, rel2->cheapest_total_path,
681 * For a semijoin, we can join the RHS to anything else by
682 * unique-ifying the RHS (if the RHS can be unique-ified).
683 * We will only get here if we have the full RHS but less
684 * than min_lefthand on the LHS.
686 * The reason to consider such a join path is exemplified by
687 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
688 * If we insist on doing this as a semijoin we will first have
689 * to form the cartesian product of A*B. But if we unique-ify
690 * C then the semijoin becomes a plain innerjoin and we can join
691 * in any order, eg C to A and then to B. When C is much smaller
692 * than A and B this can be a huge win. So we allow C to be
693 * joined to just A or just B here, and then make_join_rel has
694 * to handle the case properly.
696 * Note that actually we'll allow unique-ified C to be joined to
697 * some other relation D here, too. That is legal, if usually not
698 * very sane, and this routine is only concerned with legality not
699 * with whether the join is good strategy.
703 return false; /* invalid join path */
704 match_sjinfo = sjinfo;
708 else if (sjinfo->jointype == JOIN_SEMI &&
709 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
710 create_unique_path(root, rel1, rel1->cheapest_total_path,
713 /* Reversed semijoin case */
715 return false; /* invalid join path */
716 match_sjinfo = sjinfo;
723 * Otherwise, the proposed join overlaps the RHS but isn't a valid
724 * implementation of this SJ. But don't panic quite yet: the RHS
725 * violation might have occurred previously, in one or both input
726 * relations, in which case we must have previously decided that
727 * it was OK to commute some other SJ with this one. If we need
728 * to perform this join to finish building up the RHS, rejecting
729 * it could lead to not finding any plan at all. (This can occur
730 * because of the heuristics elsewhere in this file that postpone
731 * clauseless joins: we might not consider doing a clauseless join
732 * within the RHS until after we've performed other, validly
733 * commutable SJs with one or both sides of the clauseless join.)
734 * This consideration boils down to the rule that if both inputs
735 * overlap the RHS, we can allow the join --- they are either
736 * fully within the RHS, or represent previously-allowed joins to
739 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
740 bms_overlap(rel2->relids, sjinfo->min_righthand))
741 continue; /* assume valid previous violation of RHS */
744 * The proposed join could still be legal, but only if we're
745 * allowed to associate it into the RHS of this SJ. That means
746 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
747 * not FULL) and the proposed join must not overlap the LHS.
749 if (sjinfo->jointype != JOIN_LEFT ||
750 bms_overlap(joinrelids, sjinfo->min_lefthand))
751 return false; /* invalid join path */
754 * To be valid, the proposed join must be a LEFT join; otherwise
755 * it can't associate into this SJ's RHS. But we may not yet have
756 * found the SpecialJoinInfo matching the proposed join, so we
757 * can't test that yet. Remember the requirement for later.
759 must_be_leftjoin = true;
764 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
765 * proposed join can't associate into an SJ's RHS.
767 * Also, fail if the proposed join's predicate isn't strict; we're
768 * essentially checking to see if we can apply outer-join identity 3, and
769 * that's a requirement. (This check may be redundant with checks in
770 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
772 if (must_be_leftjoin &&
773 (match_sjinfo == NULL ||
774 match_sjinfo->jointype != JOIN_LEFT ||
775 !match_sjinfo->lhs_strict))
776 return false; /* invalid join path */
779 * We also have to check for constraints imposed by LATERAL references.
781 if (root->hasLateralRTEs)
785 Relids join_lateral_rels;
788 * The proposed rels could each contain lateral references to the
789 * other, in which case the join is impossible. If there are lateral
790 * references in just one direction, then the join has to be done with
791 * a nestloop with the lateral referencer on the inside. If the join
792 * matches an SJ that cannot be implemented by such a nestloop, the
793 * join is impossible.
795 * Also, if the lateral reference is only indirect, we should reject
796 * the join; whatever rel(s) the reference chain goes through must be
799 * Another case that might keep us from building a valid plan is the
800 * implementation restriction described by have_dangerous_phv().
802 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
803 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
804 if (lateral_fwd && lateral_rev)
805 return false; /* have lateral refs in both directions */
808 /* has to be implemented as nestloop with rel1 on left */
812 match_sjinfo->jointype == JOIN_FULL))
813 return false; /* not implementable as nestloop */
814 /* check there is a direct reference from rel2 to rel1 */
815 if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
816 return false; /* only indirect refs, so reject */
817 /* check we won't have a dangerous PHV */
818 if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
819 return false; /* might be unable to handle required PHV */
821 else if (lateral_rev)
823 /* has to be implemented as nestloop with rel2 on left */
827 match_sjinfo->jointype == JOIN_FULL))
828 return false; /* not implementable as nestloop */
829 /* check there is a direct reference from rel1 to rel2 */
830 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
831 return false; /* only indirect refs, so reject */
832 /* check we won't have a dangerous PHV */
833 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
834 return false; /* might be unable to handle required PHV */
838 * LATERAL references could also cause problems later on if we accept
839 * this join: if the join's minimum parameterization includes any rels
840 * that would have to be on the inside of an outer join with this join
841 * rel, then it's never going to be possible to build the complete
842 * query using this join. We should reject this join not only because
843 * it'll save work, but because if we don't, the clauseless-join
844 * heuristics might think that legality of this join means that some
845 * other join rel need not be formed, and that could lead to failure
846 * to find any plan at all. We have to consider not only rels that
847 * are directly on the inner side of an OJ with the joinrel, but also
848 * ones that are indirectly so, so search to find all such rels.
850 join_lateral_rels = min_join_parameterization(root, joinrelids,
852 if (join_lateral_rels)
854 Relids join_plus_rhs = bms_copy(joinrelids);
860 foreach(l, root->join_info_list)
862 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
864 /* ignore full joins --- their ordering is predetermined */
865 if (sjinfo->jointype == JOIN_FULL)
868 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
869 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
871 join_plus_rhs = bms_add_members(join_plus_rhs,
872 sjinfo->min_righthand);
877 if (bms_overlap(join_plus_rhs, join_lateral_rels))
878 return false; /* will not be able to join to some RHS rel */
882 /* Otherwise, it's a valid join */
883 *sjinfo_p = match_sjinfo;
884 *reversed_p = reversed;
890 * has_join_restriction
891 * Detect whether the specified relation has join-order restrictions,
892 * due to being inside an outer join or an IN (sub-SELECT),
893 * or participating in any LATERAL references or multi-rel PHVs.
895 * Essentially, this tests whether have_join_order_restriction() could
896 * succeed with this rel and some other one. It's OK if we sometimes
897 * say "true" incorrectly. (Therefore, we don't bother with the relatively
898 * expensive has_legal_joinclause test.)
901 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
905 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
908 foreach(l, root->placeholder_list)
910 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
912 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
913 !bms_equal(rel->relids, phinfo->ph_eval_at))
917 foreach(l, root->join_info_list)
919 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
921 /* ignore full joins --- other mechanisms preserve their ordering */
922 if (sjinfo->jointype == JOIN_FULL)
925 /* ignore if SJ is already contained in rel */
926 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
927 bms_is_subset(sjinfo->min_righthand, rel->relids))
930 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
931 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
932 bms_overlap(sjinfo->min_righthand, rel->relids))
941 * Mark a relation as proven empty.
943 * During GEQO planning, this can get invoked more than once on the same
944 * baserel struct, so it's worth checking to see if the rel is already marked
947 * Also, when called during GEQO join planning, we are in a short-lived
948 * memory context. We must make sure that the dummy path attached to a
949 * baserel survives the GEQO cycle, else the baserel is trashed for future
950 * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
951 * we don't want the dummy path to clutter the main planning context. Upshot
952 * is that the best solution is to explicitly make the dummy path in the same
953 * context the given RelOptInfo is in.
956 mark_dummy_rel(RelOptInfo *rel)
958 MemoryContext oldcontext;
960 /* Already marked? */
961 if (is_dummy_rel(rel))
964 /* No, so choose correct context to make the dummy path in */
965 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
967 /* Set dummy size estimate */
970 /* Evict any previously chosen paths */
972 rel->partial_pathlist = NIL;
974 /* Set up the dummy path */
975 add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
979 /* Set or update cheapest_total_path and related fields */
982 MemoryContextSwitchTo(oldcontext);
987 * restriction_is_constant_false --- is a restrictlist just FALSE?
989 * In cases where a qual is provably constant FALSE, eval_const_expressions
990 * will generally have thrown away anything that's ANDed with it. In outer
991 * join situations this will leave us computing cartesian products only to
992 * decide there's no match for an outer row, which is pretty stupid. So,
993 * we need to detect the case.
995 * If only_pushed_down is true, then consider only quals that are pushed-down
996 * from the point of view of the joinrel.
999 restriction_is_constant_false(List *restrictlist,
1000 RelOptInfo *joinrel,
1001 bool only_pushed_down)
1006 * Despite the above comment, the restriction list we see here might
1007 * possibly have other members besides the FALSE constant, since other
1008 * quals could get "pushed down" to the outer join level. So we check
1009 * each member of the list.
1011 foreach(lc, restrictlist)
1013 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1015 if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1018 if (rinfo->clause && IsA(rinfo->clause, Const))
1020 Const *con = (Const *) rinfo->clause;
1022 /* constant NULL is as good as constant FALSE for our purposes */
1023 if (con->constisnull)
1025 if (!DatumGetBool(con->constvalue))
1034 * Assess whether join between given two partitioned relations can be broken
1035 * down into joins between matching partitions; a technique called
1036 * "partitionwise join"
1038 * Partitionwise join is possible when a. Joining relations have same
1039 * partitioning scheme b. There exists an equi-join between the partition keys
1040 * of the two relations.
1042 * Partitionwise join is planned as follows (details: optimizer/README.)
1044 * 1. Create the RelOptInfos for joins between matching partitions i.e
1045 * child-joins and add paths to them.
1047 * 2. Construct Append or MergeAppend paths across the set of child joins.
1048 * This second phase is implemented by generate_partitionwise_join_paths().
1050 * The RelOptInfo, SpecialJoinInfo and restrictlist for each child join are
1051 * obtained by translating the respective parent join structures.
1054 try_partitionwise_join(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1055 RelOptInfo *joinrel, SpecialJoinInfo *parent_sjinfo,
1056 List *parent_restrictlist)
1058 bool rel1_is_simple = IS_SIMPLE_REL(rel1);
1059 bool rel2_is_simple = IS_SIMPLE_REL(rel2);
1063 /* Guard against stack overflow due to overly deep partition hierarchy. */
1064 check_stack_depth();
1066 /* Nothing to do, if the join relation is not partitioned. */
1067 if (!IS_PARTITIONED_REL(joinrel))
1070 /* The join relation should have consider_partitionwise_join set. */
1071 Assert(joinrel->consider_partitionwise_join);
1074 * Since this join relation is partitioned, all the base relations
1075 * participating in this join must be partitioned and so are all the
1076 * intermediate join relations.
1078 Assert(IS_PARTITIONED_REL(rel1) && IS_PARTITIONED_REL(rel2));
1079 Assert(REL_HAS_ALL_PART_PROPS(rel1) && REL_HAS_ALL_PART_PROPS(rel2));
1081 /* The joining relations should have consider_partitionwise_join set. */
1082 Assert(rel1->consider_partitionwise_join &&
1083 rel2->consider_partitionwise_join);
1086 * The partition scheme of the join relation should match that of the
1087 * joining relations.
1089 Assert(joinrel->part_scheme == rel1->part_scheme &&
1090 joinrel->part_scheme == rel2->part_scheme);
1093 * Since we allow partitionwise join only when the partition bounds of the
1094 * joining relations exactly match, the partition bounds of the join
1095 * should match those of the joining relations.
1097 Assert(partition_bounds_equal(joinrel->part_scheme->partnatts,
1098 joinrel->part_scheme->parttyplen,
1099 joinrel->part_scheme->parttypbyval,
1100 joinrel->boundinfo, rel1->boundinfo));
1101 Assert(partition_bounds_equal(joinrel->part_scheme->partnatts,
1102 joinrel->part_scheme->parttyplen,
1103 joinrel->part_scheme->parttypbyval,
1104 joinrel->boundinfo, rel2->boundinfo));
1106 nparts = joinrel->nparts;
1109 * Create child-join relations for this partitioned join, if those don't
1110 * exist. Add paths to child-joins for a pair of child relations
1111 * corresponding to the given pair of parent relations.
1113 for (cnt_parts = 0; cnt_parts < nparts; cnt_parts++)
1115 RelOptInfo *child_rel1 = rel1->part_rels[cnt_parts];
1116 RelOptInfo *child_rel2 = rel2->part_rels[cnt_parts];
1117 bool rel1_empty = (child_rel1 == NULL ||
1118 IS_DUMMY_REL(child_rel1));
1119 bool rel2_empty = (child_rel2 == NULL ||
1120 IS_DUMMY_REL(child_rel2));
1121 SpecialJoinInfo *child_sjinfo;
1122 List *child_restrictlist;
1123 RelOptInfo *child_joinrel;
1124 Relids child_joinrelids;
1125 AppendRelInfo **appinfos;
1129 * Check for cases where we can prove that this segment of the join
1130 * returns no rows, due to one or both inputs being empty (including
1131 * inputs that have been pruned away entirely). If so just ignore it.
1132 * These rules are equivalent to populate_joinrel_with_paths's rules
1133 * for dummy input relations.
1135 switch (parent_sjinfo->jointype)
1139 if (rel1_empty || rel2_empty)
1140 continue; /* ignore this join segment */
1145 continue; /* ignore this join segment */
1148 if (rel1_empty && rel2_empty)
1149 continue; /* ignore this join segment */
1152 /* other values not expected here */
1153 elog(ERROR, "unrecognized join type: %d",
1154 (int) parent_sjinfo->jointype);
1159 * If a child has been pruned entirely then we can't generate paths
1160 * for it, so we have to reject partitionwise joining unless we were
1161 * able to eliminate this partition above.
1163 if (child_rel1 == NULL || child_rel2 == NULL)
1166 * Mark the joinrel as unpartitioned so that later functions treat
1169 joinrel->nparts = 0;
1174 * If a leaf relation has consider_partitionwise_join=false, it means
1175 * that it's a dummy relation for which we skipped setting up tlist
1176 * expressions and adding EC members in set_append_rel_size(), so
1177 * again we have to fail here.
1179 if (rel1_is_simple && !child_rel1->consider_partitionwise_join)
1181 Assert(child_rel1->reloptkind == RELOPT_OTHER_MEMBER_REL);
1182 Assert(IS_DUMMY_REL(child_rel1));
1183 joinrel->nparts = 0;
1186 if (rel2_is_simple && !child_rel2->consider_partitionwise_join)
1188 Assert(child_rel2->reloptkind == RELOPT_OTHER_MEMBER_REL);
1189 Assert(IS_DUMMY_REL(child_rel2));
1190 joinrel->nparts = 0;
1194 /* We should never try to join two overlapping sets of rels. */
1195 Assert(!bms_overlap(child_rel1->relids, child_rel2->relids));
1196 child_joinrelids = bms_union(child_rel1->relids, child_rel2->relids);
1197 appinfos = find_appinfos_by_relids(root, child_joinrelids, &nappinfos);
1200 * Construct SpecialJoinInfo from parent join relations's
1203 child_sjinfo = build_child_join_sjinfo(root, parent_sjinfo,
1205 child_rel2->relids);
1208 * Construct restrictions applicable to the child join from those
1209 * applicable to the parent join.
1211 child_restrictlist =
1212 (List *) adjust_appendrel_attrs(root,
1213 (Node *) parent_restrictlist,
1214 nappinfos, appinfos);
1217 child_joinrel = joinrel->part_rels[cnt_parts];
1220 child_joinrel = build_child_join_rel(root, child_rel1, child_rel2,
1221 joinrel, child_restrictlist,
1223 child_sjinfo->jointype);
1224 joinrel->part_rels[cnt_parts] = child_joinrel;
1227 Assert(bms_equal(child_joinrel->relids, child_joinrelids));
1229 populate_joinrel_with_paths(root, child_rel1, child_rel2,
1230 child_joinrel, child_sjinfo,
1231 child_restrictlist);