1 /*-------------------------------------------------------------------------
4 * Routines copied from PostgreSQL core distribution.
6 * src/backend/optimizer/path/allpaths.c
7 * set_append_rel_pathlist()
8 * generate_mergeappend_paths()
9 * get_cheapest_parameterized_child_path()
10 * accumulate_append_subpath()
11 * standard_join_search()
13 * src/backend/optimizer/path/joinrels.c
14 * join_search_one_level()
15 * make_rels_by_clause_joins()
16 * make_rels_by_clauseless_joins()
18 * has_join_restriction()
21 * restriction_is_constant_false()
23 * Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
24 * Portions Copyright (c) 1994, Regents of the University of California
26 *-------------------------------------------------------------------------
30 * set_append_rel_pathlist
31 * Build access paths for an "append relation"
34 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
35 Index rti, RangeTblEntry *rte)
37 int parentRTindex = rti;
38 List *live_childrels = NIL;
40 bool subpaths_valid = true;
41 List *all_child_pathkeys = NIL;
42 List *all_child_outers = NIL;
46 * Generate access paths for each member relation, and remember the
47 * cheapest path for each one. Also, identify all pathkeys (orderings)
48 * and parameterizations (required_outer sets) available for the member
51 foreach(l, root->append_rel_list)
53 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
55 RangeTblEntry *childRTE;
59 /* append_rel_list contains all append rels; ignore others */
60 if (appinfo->parent_relid != parentRTindex)
63 /* Re-locate the child RTE and RelOptInfo */
64 childRTindex = appinfo->child_relid;
65 childRTE = root->simple_rte_array[childRTindex];
66 childrel = root->simple_rel_array[childRTindex];
69 * Compute the child's access paths.
71 set_rel_pathlist(root, childrel, childRTindex, childRTE);
74 * If child is dummy, ignore it.
76 if (IS_DUMMY_REL(childrel))
80 * Child is live, so add it to the live_childrels list for use below.
82 live_childrels = lappend(live_childrels, childrel);
85 * If child has an unparameterized cheapest-total path, add that to
86 * the unparameterized Append path we are constructing for the parent.
87 * If not, there's no workable unparameterized path.
89 if (childrel->cheapest_total_path->param_info == NULL)
90 subpaths = accumulate_append_subpath(subpaths,
91 childrel->cheapest_total_path);
93 subpaths_valid = false;
96 * Collect lists of all the available path orderings and
97 * parameterizations for all the children. We use these as a
98 * heuristic to indicate which sort orderings and parameterizations we
99 * should build Append and MergeAppend paths for.
101 foreach(lcp, childrel->pathlist)
103 Path *childpath = (Path *) lfirst(lcp);
104 List *childkeys = childpath->pathkeys;
105 Relids childouter = PATH_REQ_OUTER(childpath);
107 /* Unsorted paths don't contribute to pathkey list */
108 if (childkeys != NIL)
113 /* Have we already seen this ordering? */
114 foreach(lpk, all_child_pathkeys)
116 List *existing_pathkeys = (List *) lfirst(lpk);
118 if (compare_pathkeys(existing_pathkeys,
119 childkeys) == PATHKEYS_EQUAL)
127 /* No, so add it to all_child_pathkeys */
128 all_child_pathkeys = lappend(all_child_pathkeys,
133 /* Unparameterized paths don't contribute to param-set list */
139 /* Have we already seen this param set? */
140 foreach(lco, all_child_outers)
142 Relids existing_outers = (Relids) lfirst(lco);
144 if (bms_equal(existing_outers, childouter))
152 /* No, so add it to all_child_outers */
153 all_child_outers = lappend(all_child_outers,
161 * If we found unparameterized paths for all children, build an unordered,
162 * unparameterized Append path for the rel. (Note: this is correct even
163 * if we have zero or one live subpath due to constraint exclusion.)
166 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL));
169 * Also build unparameterized MergeAppend paths based on the collected
170 * list of child pathkeys.
173 generate_mergeappend_paths(root, rel, live_childrels,
177 * Build Append paths for each parameterization seen among the child rels.
178 * (This may look pretty expensive, but in most cases of practical
179 * interest, the child rels will expose mostly the same parameterizations,
180 * so that not that many cases actually get considered here.)
182 * The Append node itself cannot enforce quals, so all qual checking must
183 * be done in the child paths. This means that to have a parameterized
184 * Append path, we must have the exact same parameterization for each
185 * child path; otherwise some children might be failing to check the
186 * moved-down quals. To make them match up, we can try to increase the
187 * parameterization of lesser-parameterized paths.
189 foreach(l, all_child_outers)
191 Relids required_outer = (Relids) lfirst(l);
194 /* Select the child paths for an Append with this parameterization */
196 subpaths_valid = true;
197 foreach(lcr, live_childrels)
199 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
202 subpath = get_cheapest_parameterized_child_path(root,
207 /* failed to make a suitable path for this child */
208 subpaths_valid = false;
211 subpaths = accumulate_append_subpath(subpaths, subpath);
215 add_path(rel, (Path *)
216 create_append_path(rel, subpaths, required_outer));
221 * generate_mergeappend_paths
222 * Generate MergeAppend paths for an append relation
224 * Generate a path for each ordering (pathkey list) appearing in
225 * all_child_pathkeys.
227 * We consider both cheapest-startup and cheapest-total cases, ie, for each
228 * interesting ordering, collect all the cheapest startup subpaths and all the
229 * cheapest total paths, and build a MergeAppend path for each case.
231 * We don't currently generate any parameterized MergeAppend paths. While
232 * it would not take much more code here to do so, it's very unclear that it
233 * is worth the planning cycles to investigate such paths: there's little
234 * use for an ordered path on the inside of a nestloop. In fact, it's likely
235 * that the current coding of add_path would reject such paths out of hand,
236 * because add_path gives no credit for sort ordering of parameterized paths,
237 * and a parameterized MergeAppend is going to be more expensive than the
238 * corresponding parameterized Append path. If we ever try harder to support
239 * parameterized mergejoin plans, it might be worth adding support for
240 * parameterized MergeAppends to feed such joins. (See notes in
241 * optimizer/README for why that might not ever happen, though.)
244 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
245 List *live_childrels,
246 List *all_child_pathkeys)
250 foreach(lcp, all_child_pathkeys)
252 List *pathkeys = (List *) lfirst(lcp);
253 List *startup_subpaths = NIL;
254 List *total_subpaths = NIL;
255 bool startup_neq_total = false;
258 /* Select the child paths for this ordering... */
259 foreach(lcr, live_childrels)
261 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
262 Path *cheapest_startup,
265 /* Locate the right paths, if they are available. */
267 get_cheapest_path_for_pathkeys(childrel->pathlist,
272 get_cheapest_path_for_pathkeys(childrel->pathlist,
278 * If we can't find any paths with the right order just use the
279 * cheapest-total path; we'll have to sort it later.
281 if (cheapest_startup == NULL || cheapest_total == NULL)
283 cheapest_startup = cheapest_total =
284 childrel->cheapest_total_path;
285 /* Assert we do have an unparameterized path for this child */
286 Assert(cheapest_total->param_info == NULL);
290 * Notice whether we actually have different paths for the
291 * "cheapest" and "total" cases; frequently there will be no point
292 * in two create_merge_append_path() calls.
294 if (cheapest_startup != cheapest_total)
295 startup_neq_total = true;
298 accumulate_append_subpath(startup_subpaths, cheapest_startup);
300 accumulate_append_subpath(total_subpaths, cheapest_total);
303 /* ... and build the MergeAppend paths */
304 add_path(rel, (Path *) create_merge_append_path(root,
309 if (startup_neq_total)
310 add_path(rel, (Path *) create_merge_append_path(root,
319 * get_cheapest_parameterized_child_path
320 * Get cheapest path for this relation that has exactly the requested
323 * Returns NULL if unable to create such a path.
326 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
327 Relids required_outer)
333 * Look up the cheapest existing path with no more than the needed
334 * parameterization. If it has exactly the needed parameterization, we're
337 cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
341 Assert(cheapest != NULL);
342 if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
346 * Otherwise, we can "reparameterize" an existing path to match the given
347 * parameterization, which effectively means pushing down additional
348 * joinquals to be checked within the path's scan. However, some existing
349 * paths might check the available joinquals already while others don't;
350 * therefore, it's not clear which existing path will be cheapest after
351 * reparameterization. We have to go through them all and find out.
354 foreach(lc, rel->pathlist)
356 Path *path = (Path *) lfirst(lc);
358 /* Can't use it if it needs more than requested parameterization */
359 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
363 * Reparameterization can only increase the path's cost, so if it's
364 * already more expensive than the current cheapest, forget it.
366 if (cheapest != NULL &&
367 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
370 /* Reparameterize if needed, then recheck cost */
371 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
373 path = reparameterize_path(root, path, required_outer, 1.0);
375 continue; /* failed to reparameterize this one */
376 Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
378 if (cheapest != NULL &&
379 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
383 /* We have a new best path */
387 /* Return the best path, or NULL if we found no suitable candidate */
392 * accumulate_append_subpath
393 * Add a subpath to the list being built for an Append or MergeAppend
395 * It's possible that the child is itself an Append or MergeAppend path, in
396 * which case we can "cut out the middleman" and just add its child paths to
397 * our own list. (We don't try to do this earlier because we need to apply
398 * both levels of transformation to the quals.)
400 * Note that if we omit a child MergeAppend in this way, we are effectively
401 * omitting a sort step, which seems fine: if the parent is to be an Append,
402 * its result would be unsorted anyway, while if the parent is to be a
403 * MergeAppend, there's no point in a separate sort on a child.
406 accumulate_append_subpath(List *subpaths, Path *path)
408 if (IsA(path, AppendPath))
410 AppendPath *apath = (AppendPath *) path;
412 /* list_copy is important here to avoid sharing list substructure */
413 return list_concat(subpaths, list_copy(apath->subpaths));
415 else if (IsA(path, MergeAppendPath))
417 MergeAppendPath *mpath = (MergeAppendPath *) path;
419 /* list_copy is important here to avoid sharing list substructure */
420 return list_concat(subpaths, list_copy(mpath->subpaths));
423 return lappend(subpaths, path);
427 * standard_join_search
428 * Find possible joinpaths for a query by successively finding ways
429 * to join component relations into join relations.
431 * 'levels_needed' is the number of iterations needed, ie, the number of
432 * independent jointree items in the query. This is > 1.
434 * 'initial_rels' is a list of RelOptInfo nodes for each independent
435 * jointree item. These are the components to be joined together.
436 * Note that levels_needed == list_length(initial_rels).
438 * Returns the final level of join relations, i.e., the relation that is
439 * the result of joining all the original relations together.
440 * At least one implementation path must be provided for this relation and
441 * all required sub-relations.
443 * To support loadable plugins that modify planner behavior by changing the
444 * join searching algorithm, we provide a hook variable that lets a plugin
445 * replace or supplement this function. Any such hook must return the same
446 * final join relation as the standard code would, but it might have a
447 * different set of implementation paths attached, and only the sub-joinrels
448 * needed for these paths need have been instantiated.
450 * Note to plugin authors: the functions invoked during standard_join_search()
451 * modify root->join_rel_list and root->join_rel_hash. If you want to do more
452 * than one join-order search, you'll probably need to save and restore the
453 * original states of those data structures. See geqo_eval() for an example.
456 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
462 * This function cannot be invoked recursively within any one planning
463 * problem, so join_rel_level[] can't be in use already.
465 Assert(root->join_rel_level == NULL);
468 * We employ a simple "dynamic programming" algorithm: we first find all
469 * ways to build joins of two jointree items, then all ways to build joins
470 * of three items (from two-item joins and single items), then four-item
471 * joins, and so on until we have considered all ways to join all the
472 * items into one rel.
474 * root->join_rel_level[j] is a list of all the j-item rels. Initially we
475 * set root->join_rel_level[1] to represent all the single-jointree-item
478 root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
480 root->join_rel_level[1] = initial_rels;
482 for (lev = 2; lev <= levels_needed; lev++)
487 * Determine all possible pairs of relations to be joined at this
488 * level, and build paths for making each one from every available
489 * pair of lower-level relations.
491 join_search_one_level(root, lev);
494 * Do cleanup work on each just-processed rel.
496 foreach(lc, root->join_rel_level[lev])
498 rel = (RelOptInfo *) lfirst(lc);
500 /* Find and save the cheapest paths for this rel */
503 #ifdef OPTIMIZER_DEBUG
504 debug_print_rel(root, rel);
510 * We should have a single rel at the final level.
512 if (root->join_rel_level[levels_needed] == NIL)
513 elog(ERROR, "failed to build any %d-way joins", levels_needed);
514 Assert(list_length(root->join_rel_level[levels_needed]) == 1);
516 rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
518 root->join_rel_level = NULL;
524 * join_search_one_level
525 * Consider ways to produce join relations containing exactly 'level'
526 * jointree items. (This is one step of the dynamic-programming method
527 * embodied in standard_join_search.) Join rel nodes for each feasible
528 * combination of lower-level rels are created and returned in a list.
529 * Implementation paths are created for each such joinrel, too.
531 * level: level of rels we want to make this time
532 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
534 * The result is returned in root->join_rel_level[level].
537 join_search_one_level(PlannerInfo *root, int level)
539 List **joinrels = root->join_rel_level;
543 Assert(joinrels[level] == NIL);
545 /* Set join_cur_level so that new joinrels are added to proper list */
546 root->join_cur_level = level;
549 * First, consider left-sided and right-sided plans, in which rels of
550 * exactly level-1 member relations are joined against initial relations.
551 * We prefer to join using join clauses, but if we find a rel of level-1
552 * members that has no join clauses, we will generate Cartesian-product
553 * joins against all initial rels not already contained in it.
555 foreach(r, joinrels[level - 1])
557 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
559 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
560 has_join_restriction(root, old_rel))
563 * There are join clauses or join order restrictions relevant to
564 * this rel, so consider joins between this rel and (only) those
565 * initial rels it is linked to by a clause or restriction.
567 * At level 2 this condition is symmetric, so there is no need to
568 * look at initial rels before this one in the list; we already
569 * considered such joins when we were at the earlier rel. (The
570 * mirror-image joins are handled automatically by make_join_rel.)
571 * In later passes (level > 2), we join rels of the previous level
572 * to each initial rel they don't already include but have a join
573 * clause or restriction with.
575 ListCell *other_rels;
577 if (level == 2) /* consider remaining initial rels */
578 other_rels = lnext(r);
579 else /* consider all initial rels */
580 other_rels = list_head(joinrels[1]);
582 make_rels_by_clause_joins(root,
589 * Oops, we have a relation that is not joined to any other
590 * relation, either directly or by join-order restrictions.
591 * Cartesian product time.
593 * We consider a cartesian product with each not-already-included
594 * initial rel, whether it has other join clauses or not. At
595 * level 2, if there are two or more clauseless initial rels, we
596 * will redundantly consider joining them in both directions; but
597 * such cases aren't common enough to justify adding complexity to
598 * avoid the duplicated effort.
600 make_rels_by_clauseless_joins(root,
602 list_head(joinrels[1]));
607 * Now, consider "bushy plans" in which relations of k initial rels are
608 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
610 * We only consider bushy-plan joins for pairs of rels where there is a
611 * suitable join clause (or join order restriction), in order to avoid
612 * unreasonable growth of planning time.
616 int other_level = level - k;
619 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
620 * need to go as far as the halfway point.
625 foreach(r, joinrels[k])
627 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
628 ListCell *other_rels;
632 * We can ignore relations without join clauses here, unless they
633 * participate in join-order restrictions --- then we might have
634 * to force a bushy join plan.
636 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
637 !has_join_restriction(root, old_rel))
640 if (k == other_level)
641 other_rels = lnext(r); /* only consider remaining rels */
643 other_rels = list_head(joinrels[other_level]);
645 for_each_cell(r2, other_rels)
647 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
649 if (!bms_overlap(old_rel->relids, new_rel->relids))
652 * OK, we can build a rel of the right level from this
653 * pair of rels. Do so if there is at least one relevant
654 * join clause or join order restriction.
656 if (have_relevant_joinclause(root, old_rel, new_rel) ||
657 have_join_order_restriction(root, old_rel, new_rel))
659 (void) make_join_rel(root, old_rel, new_rel);
667 * Last-ditch effort: if we failed to find any usable joins so far, force
668 * a set of cartesian-product joins to be generated. This handles the
669 * special case where all the available rels have join clauses but we
670 * cannot use any of those clauses yet. This can only happen when we are
671 * considering a join sub-problem (a sub-joinlist) and all the rels in the
672 * sub-problem have only join clauses with rels outside the sub-problem.
675 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
676 * WHERE a.w = c.x and b.y = d.z;
678 * If the "a INNER JOIN b" sub-problem does not get flattened into the
679 * upper level, we must be willing to make a cartesian join of a and b;
680 * but the code above will not have done so, because it thought that both
681 * a and b have joinclauses. We consider only left-sided and right-sided
682 * cartesian joins in this case (no bushy).
685 if (joinrels[level] == NIL)
688 * This loop is just like the first one, except we always call
689 * make_rels_by_clauseless_joins().
691 foreach(r, joinrels[level - 1])
693 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
695 make_rels_by_clauseless_joins(root,
697 list_head(joinrels[1]));
701 * When special joins are involved, there may be no legal way
702 * to make an N-way join for some values of N. For example consider
704 * SELECT ... FROM t1 WHERE
705 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
706 * y IN (SELECT ... FROM t4,t5 WHERE ...)
708 * We will flatten this query to a 5-way join problem, but there are
709 * no 4-way joins that join_is_legal() will consider legal. We have
710 * to accept failure at level 4 and go on to discover a workable
711 * bushy plan at level 5.
713 * However, if there are no special joins and no lateral references
714 * then join_is_legal() should never fail, and so the following sanity
718 if (joinrels[level] == NIL &&
719 root->join_info_list == NIL &&
720 !root->hasLateralRTEs)
721 elog(ERROR, "failed to build any %d-way joins", level);
726 * make_rels_by_clause_joins
727 * Build joins between the given relation 'old_rel' and other relations
728 * that participate in join clauses that 'old_rel' also participates in
729 * (or participate in join-order restrictions with it).
730 * The join rels are returned in root->join_rel_level[join_cur_level].
732 * Note: at levels above 2 we will generate the same joined relation in
733 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
734 * (b join c) join a, though the second case will add a different set of Paths
735 * to it. This is the reason for using the join_rel_level mechanism, which
736 * automatically ensures that each new joinrel is only added to the list once.
738 * 'old_rel' is the relation entry for the relation to be joined
739 * 'other_rels': the first cell in a linked list containing the other
740 * rels to be considered for joining
742 * Currently, this is only used with initial rels in other_rels, but it
743 * will work for joining to joinrels too.
746 make_rels_by_clause_joins(PlannerInfo *root,
748 ListCell *other_rels)
752 for_each_cell(l, other_rels)
754 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
756 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
757 (have_relevant_joinclause(root, old_rel, other_rel) ||
758 have_join_order_restriction(root, old_rel, other_rel)))
760 (void) make_join_rel(root, old_rel, other_rel);
766 * make_rels_by_clauseless_joins
767 * Given a relation 'old_rel' and a list of other relations
768 * 'other_rels', create a join relation between 'old_rel' and each
769 * member of 'other_rels' that isn't already included in 'old_rel'.
770 * The join rels are returned in root->join_rel_level[join_cur_level].
772 * 'old_rel' is the relation entry for the relation to be joined
773 * 'other_rels': the first cell of a linked list containing the
774 * other rels to be considered for joining
776 * Currently, this is only used with initial rels in other_rels, but it would
777 * work for joining to joinrels too.
780 make_rels_by_clauseless_joins(PlannerInfo *root,
782 ListCell *other_rels)
786 for_each_cell(l, other_rels)
788 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
790 if (!bms_overlap(other_rel->relids, old_rel->relids))
792 (void) make_join_rel(root, old_rel, other_rel);
799 * Determine whether a proposed join is legal given the query's
800 * join order constraints; and if it is, determine the join type.
802 * Caller must supply not only the two rels, but the union of their relids.
803 * (We could simplify the API by computing joinrelids locally, but this
804 * would be redundant work in the normal path through make_join_rel.)
806 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
807 * else it's set to point to the associated SpecialJoinInfo node. Also,
808 * *reversed_p is set TRUE if the given relations need to be swapped to
809 * match the SpecialJoinInfo node.
812 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
814 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
816 SpecialJoinInfo *match_sjinfo;
819 bool must_be_leftjoin;
823 * Ensure output params are set on failure return. This is just to
824 * suppress uninitialized-variable warnings from overly anal compilers.
830 * If we have any special joins, the proposed join might be illegal; and
831 * in any case we have to determine its join type. Scan the join info
832 * list for matches and conflicts.
836 unique_ified = false;
837 must_be_leftjoin = false;
839 foreach(l, root->join_info_list)
841 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
844 * This special join is not relevant unless its RHS overlaps the
845 * proposed join. (Check this first as a fast path for dismissing
846 * most irrelevant SJs quickly.)
848 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
852 * Also, not relevant if proposed join is fully contained within RHS
853 * (ie, we're still building up the RHS).
855 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
859 * Also, not relevant if SJ is already done within either input.
861 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
862 bms_is_subset(sjinfo->min_righthand, rel1->relids))
864 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
865 bms_is_subset(sjinfo->min_righthand, rel2->relids))
869 * If it's a semijoin and we already joined the RHS to any other rels
870 * within either input, then we must have unique-ified the RHS at that
871 * point (see below). Therefore the semijoin is no longer relevant in
874 if (sjinfo->jointype == JOIN_SEMI)
876 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
877 !bms_equal(sjinfo->syn_righthand, rel1->relids))
879 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
880 !bms_equal(sjinfo->syn_righthand, rel2->relids))
885 * If one input contains min_lefthand and the other contains
886 * min_righthand, then we can perform the SJ at this join.
888 * Reject if we get matches to more than one SJ; that implies we're
889 * considering something that's not really valid.
891 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
892 bms_is_subset(sjinfo->min_righthand, rel2->relids))
895 return false; /* invalid join path */
896 match_sjinfo = sjinfo;
899 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
900 bms_is_subset(sjinfo->min_righthand, rel1->relids))
903 return false; /* invalid join path */
904 match_sjinfo = sjinfo;
907 else if (sjinfo->jointype == JOIN_SEMI &&
908 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
909 create_unique_path(root, rel2, rel2->cheapest_total_path,
913 * For a semijoin, we can join the RHS to anything else by
914 * unique-ifying the RHS (if the RHS can be unique-ified).
915 * We will only get here if we have the full RHS but less
916 * than min_lefthand on the LHS.
918 * The reason to consider such a join path is exemplified by
919 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
920 * If we insist on doing this as a semijoin we will first have
921 * to form the cartesian product of A*B. But if we unique-ify
922 * C then the semijoin becomes a plain innerjoin and we can join
923 * in any order, eg C to A and then to B. When C is much smaller
924 * than A and B this can be a huge win. So we allow C to be
925 * joined to just A or just B here, and then make_join_rel has
926 * to handle the case properly.
928 * Note that actually we'll allow unique-ified C to be joined to
929 * some other relation D here, too. That is legal, if usually not
930 * very sane, and this routine is only concerned with legality not
931 * with whether the join is good strategy.
935 return false; /* invalid join path */
936 match_sjinfo = sjinfo;
940 else if (sjinfo->jointype == JOIN_SEMI &&
941 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
942 create_unique_path(root, rel1, rel1->cheapest_total_path,
945 /* Reversed semijoin case */
947 return false; /* invalid join path */
948 match_sjinfo = sjinfo;
955 * Otherwise, the proposed join overlaps the RHS but isn't a valid
956 * implementation of this SJ. But don't panic quite yet: the RHS
957 * violation might have occurred previously, in one or both input
958 * relations, in which case we must have previously decided that
959 * it was OK to commute some other SJ with this one. If we need
960 * to perform this join to finish building up the RHS, rejecting
961 * it could lead to not finding any plan at all. (This can occur
962 * because of the heuristics elsewhere in this file that postpone
963 * clauseless joins: we might not consider doing a clauseless join
964 * within the RHS until after we've performed other, validly
965 * commutable SJs with one or both sides of the clauseless join.)
966 * This consideration boils down to the rule that if both inputs
967 * overlap the RHS, we can allow the join --- they are either
968 * fully within the RHS, or represent previously-allowed joins to
971 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
972 bms_overlap(rel2->relids, sjinfo->min_righthand))
973 continue; /* assume valid previous violation of RHS */
976 * The proposed join could still be legal, but only if we're
977 * allowed to associate it into the RHS of this SJ. That means
978 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
979 * not FULL) and the proposed join must not overlap the LHS.
981 if (sjinfo->jointype != JOIN_LEFT ||
982 bms_overlap(joinrelids, sjinfo->min_lefthand))
983 return false; /* invalid join path */
986 * To be valid, the proposed join must be a LEFT join; otherwise
987 * it can't associate into this SJ's RHS. But we may not yet have
988 * found the SpecialJoinInfo matching the proposed join, so we
989 * can't test that yet. Remember the requirement for later.
991 must_be_leftjoin = true;
996 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
997 * proposed join can't associate into an SJ's RHS.
999 * Also, fail if the proposed join's predicate isn't strict; we're
1000 * essentially checking to see if we can apply outer-join identity 3, and
1001 * that's a requirement. (This check may be redundant with checks in
1002 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1004 if (must_be_leftjoin &&
1005 (match_sjinfo == NULL ||
1006 match_sjinfo->jointype != JOIN_LEFT ||
1007 !match_sjinfo->lhs_strict))
1008 return false; /* invalid join path */
1011 * We also have to check for constraints imposed by LATERAL references.
1013 if (root->hasLateralRTEs)
1017 Relids join_lateral_rels;
1020 * The proposed rels could each contain lateral references to the
1021 * other, in which case the join is impossible. If there are lateral
1022 * references in just one direction, then the join has to be done with
1023 * a nestloop with the lateral referencer on the inside. If the join
1024 * matches an SJ that cannot be implemented by such a nestloop, the
1025 * join is impossible.
1027 * Also, if the lateral reference is only indirect, we should reject
1028 * the join; whatever rel(s) the reference chain goes through must be
1031 * Another case that might keep us from building a valid plan is the
1032 * implementation restriction described by have_dangerous_phv().
1034 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1035 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1036 if (lateral_fwd && lateral_rev)
1037 return false; /* have lateral refs in both directions */
1040 /* has to be implemented as nestloop with rel1 on left */
1044 match_sjinfo->jointype == JOIN_FULL))
1045 return false; /* not implementable as nestloop */
1046 /* check there is a direct reference from rel2 to rel1 */
1047 if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1048 return false; /* only indirect refs, so reject */
1049 /* check we won't have a dangerous PHV */
1050 if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1051 return false; /* might be unable to handle required PHV */
1053 else if (lateral_rev)
1055 /* has to be implemented as nestloop with rel2 on left */
1059 match_sjinfo->jointype == JOIN_FULL))
1060 return false; /* not implementable as nestloop */
1061 /* check there is a direct reference from rel1 to rel2 */
1062 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1063 return false; /* only indirect refs, so reject */
1064 /* check we won't have a dangerous PHV */
1065 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1066 return false; /* might be unable to handle required PHV */
1070 * LATERAL references could also cause problems later on if we accept
1071 * this join: if the join's minimum parameterization includes any rels
1072 * that would have to be on the inside of an outer join with this join
1073 * rel, then it's never going to be possible to build the complete
1074 * query using this join. We should reject this join not only because
1075 * it'll save work, but because if we don't, the clauseless-join
1076 * heuristics might think that legality of this join means that some
1077 * other join rel need not be formed, and that could lead to failure
1078 * to find any plan at all. We have to consider not only rels that
1079 * are directly on the inner side of an OJ with the joinrel, but also
1080 * ones that are indirectly so, so search to find all such rels.
1082 join_lateral_rels = min_join_parameterization(root, joinrelids,
1084 if (join_lateral_rels)
1086 Relids join_plus_rhs = bms_copy(joinrelids);
1092 foreach(l, root->join_info_list)
1094 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1096 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1097 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1099 join_plus_rhs = bms_add_members(join_plus_rhs,
1100 sjinfo->min_righthand);
1103 /* full joins constrain both sides symmetrically */
1104 if (sjinfo->jointype == JOIN_FULL &&
1105 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1106 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1108 join_plus_rhs = bms_add_members(join_plus_rhs,
1109 sjinfo->min_lefthand);
1114 if (bms_overlap(join_plus_rhs, join_lateral_rels))
1115 return false; /* will not be able to join to some RHS rel */
1119 /* Otherwise, it's a valid join */
1120 *sjinfo_p = match_sjinfo;
1121 *reversed_p = reversed;
1126 * has_join_restriction
1127 * Detect whether the specified relation has join-order restrictions,
1128 * due to being inside an outer join or an IN (sub-SELECT),
1129 * or participating in any LATERAL references or multi-rel PHVs.
1131 * Essentially, this tests whether have_join_order_restriction() could
1132 * succeed with this rel and some other one. It's OK if we sometimes
1133 * say "true" incorrectly. (Therefore, we don't bother with the relatively
1134 * expensive has_legal_joinclause test.)
1137 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1141 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1144 foreach(l, root->placeholder_list)
1146 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1148 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1149 !bms_equal(rel->relids, phinfo->ph_eval_at))
1153 foreach(l, root->join_info_list)
1155 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1157 /* ignore full joins --- other mechanisms preserve their ordering */
1158 if (sjinfo->jointype == JOIN_FULL)
1161 /* ignore if SJ is already contained in rel */
1162 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1163 bms_is_subset(sjinfo->min_righthand, rel->relids))
1166 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1167 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1168 bms_overlap(sjinfo->min_righthand, rel->relids))
1176 * is_dummy_rel --- has relation been proven empty?
1179 is_dummy_rel(RelOptInfo *rel)
1181 return IS_DUMMY_REL(rel);
1185 * Mark a relation as proven empty.
1187 * During GEQO planning, this can get invoked more than once on the same
1188 * baserel struct, so it's worth checking to see if the rel is already marked
1191 * Also, when called during GEQO join planning, we are in a short-lived
1192 * memory context. We must make sure that the dummy path attached to a
1193 * baserel survives the GEQO cycle, else the baserel is trashed for future
1194 * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1195 * we don't want the dummy path to clutter the main planning context. Upshot
1196 * is that the best solution is to explicitly make the dummy path in the same
1197 * context the given RelOptInfo is in.
1200 mark_dummy_rel(RelOptInfo *rel)
1202 MemoryContext oldcontext;
1204 /* Already marked? */
1205 if (is_dummy_rel(rel))
1208 /* No, so choose correct context to make the dummy path in */
1209 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1211 /* Set dummy size estimate */
1214 /* Evict any previously chosen paths */
1215 rel->pathlist = NIL;
1217 /* Set up the dummy path */
1218 add_path(rel, (Path *) create_append_path(rel, NIL, NULL));
1220 /* Set or update cheapest_total_path and related fields */
1223 MemoryContextSwitchTo(oldcontext);
1227 * restriction_is_constant_false --- is a restrictlist just FALSE?
1229 * In cases where a qual is provably constant FALSE, eval_const_expressions
1230 * will generally have thrown away anything that's ANDed with it. In outer
1231 * join situations this will leave us computing cartesian products only to
1232 * decide there's no match for an outer row, which is pretty stupid. So,
1233 * we need to detect the case.
1235 * If only_pushed_down is TRUE, then consider only pushed-down quals.
1238 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1243 * Despite the above comment, the restriction list we see here might
1244 * possibly have other members besides the FALSE constant, since other
1245 * quals could get "pushed down" to the outer join level. So we check
1246 * each member of the list.
1248 foreach(lc, restrictlist)
1250 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1252 Assert(IsA(rinfo, RestrictInfo));
1253 if (only_pushed_down && !rinfo->is_pushed_down)
1256 if (rinfo->clause && IsA(rinfo->clause, Const))
1258 Const *con = (Const *) rinfo->clause;
1260 /* constant NULL is as good as constant FALSE for our purposes */
1261 if (con->constisnull)
1263 if (!DatumGetBool(con->constvalue))