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 *partial_subpaths = NIL;
42 bool partial_subpaths_valid = true;
43 List *all_child_pathkeys = NIL;
44 List *all_child_outers = NIL;
48 * Generate access paths for each member relation, and remember the
49 * cheapest path for each one. Also, identify all pathkeys (orderings)
50 * and parameterizations (required_outer sets) available for the member
53 foreach(l, root->append_rel_list)
55 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
57 RangeTblEntry *childRTE;
61 /* append_rel_list contains all append rels; ignore others */
62 if (appinfo->parent_relid != parentRTindex)
65 /* Re-locate the child RTE and RelOptInfo */
66 childRTindex = appinfo->child_relid;
67 childRTE = root->simple_rte_array[childRTindex];
68 childrel = root->simple_rel_array[childRTindex];
71 * If set_append_rel_size() decided the parent appendrel was
72 * parallel-unsafe at some point after visiting this child rel, we
73 * need to propagate the unsafety marking down to the child, so that
74 * we don't generate useless partial paths for it.
76 if (!rel->consider_parallel)
77 childrel->consider_parallel = false;
80 * Compute the child's access paths.
82 set_rel_pathlist(root, childrel, childRTindex, childRTE);
85 * If child is dummy, ignore it.
87 if (IS_DUMMY_REL(childrel))
91 * Child is live, so add it to the live_childrels list for use below.
93 live_childrels = lappend(live_childrels, childrel);
96 * If child has an unparameterized cheapest-total path, add that to
97 * the unparameterized Append path we are constructing for the parent.
98 * If not, there's no workable unparameterized path.
100 if (childrel->cheapest_total_path->param_info == NULL)
101 subpaths = accumulate_append_subpath(subpaths,
102 childrel->cheapest_total_path);
104 subpaths_valid = false;
106 /* Same idea, but for a partial plan. */
107 if (childrel->partial_pathlist != NIL)
108 partial_subpaths = accumulate_append_subpath(partial_subpaths,
109 linitial(childrel->partial_pathlist));
111 partial_subpaths_valid = false;
114 * Collect lists of all the available path orderings and
115 * parameterizations for all the children. We use these as a
116 * heuristic to indicate which sort orderings and parameterizations we
117 * should build Append and MergeAppend paths for.
119 foreach(lcp, childrel->pathlist)
121 Path *childpath = (Path *) lfirst(lcp);
122 List *childkeys = childpath->pathkeys;
123 Relids childouter = PATH_REQ_OUTER(childpath);
125 /* Unsorted paths don't contribute to pathkey list */
126 if (childkeys != NIL)
131 /* Have we already seen this ordering? */
132 foreach(lpk, all_child_pathkeys)
134 List *existing_pathkeys = (List *) lfirst(lpk);
136 if (compare_pathkeys(existing_pathkeys,
137 childkeys) == PATHKEYS_EQUAL)
145 /* No, so add it to all_child_pathkeys */
146 all_child_pathkeys = lappend(all_child_pathkeys,
151 /* Unparameterized paths don't contribute to param-set list */
157 /* Have we already seen this param set? */
158 foreach(lco, all_child_outers)
160 Relids existing_outers = (Relids) lfirst(lco);
162 if (bms_equal(existing_outers, childouter))
170 /* No, so add it to all_child_outers */
171 all_child_outers = lappend(all_child_outers,
179 * If we found unparameterized paths for all children, build an unordered,
180 * unparameterized Append path for the rel. (Note: this is correct even
181 * if we have zero or one live subpath due to constraint exclusion.)
184 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0));
187 * Consider an append of partial unordered, unparameterized partial paths.
189 if (partial_subpaths_valid)
191 AppendPath *appendpath;
193 int parallel_workers = 0;
196 * Decide on the number of workers to request for this append path.
197 * For now, we just use the maximum value from among the members. It
198 * might be useful to use a higher number if the Append node were
199 * smart enough to spread out the workers, but it currently isn't.
201 foreach(lc, partial_subpaths)
203 Path *path = lfirst(lc);
205 parallel_workers = Max(parallel_workers, path->parallel_workers);
207 Assert(parallel_workers > 0);
209 /* Generate a partial append path. */
210 appendpath = create_append_path(rel, partial_subpaths, NULL,
212 add_partial_path(rel, (Path *) appendpath);
216 * Also build unparameterized MergeAppend paths based on the collected
217 * list of child pathkeys.
220 generate_mergeappend_paths(root, rel, live_childrels,
224 * Build Append paths for each parameterization seen among the child rels.
225 * (This may look pretty expensive, but in most cases of practical
226 * interest, the child rels will expose mostly the same parameterizations,
227 * so that not that many cases actually get considered here.)
229 * The Append node itself cannot enforce quals, so all qual checking must
230 * be done in the child paths. This means that to have a parameterized
231 * Append path, we must have the exact same parameterization for each
232 * child path; otherwise some children might be failing to check the
233 * moved-down quals. To make them match up, we can try to increase the
234 * parameterization of lesser-parameterized paths.
236 foreach(l, all_child_outers)
238 Relids required_outer = (Relids) lfirst(l);
241 /* Select the child paths for an Append with this parameterization */
243 subpaths_valid = true;
244 foreach(lcr, live_childrels)
246 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
249 subpath = get_cheapest_parameterized_child_path(root,
254 /* failed to make a suitable path for this child */
255 subpaths_valid = false;
258 subpaths = accumulate_append_subpath(subpaths, subpath);
262 add_path(rel, (Path *)
263 create_append_path(rel, subpaths, required_outer, 0));
268 * generate_mergeappend_paths
269 * Generate MergeAppend paths for an append relation
271 * Generate a path for each ordering (pathkey list) appearing in
272 * all_child_pathkeys.
274 * We consider both cheapest-startup and cheapest-total cases, ie, for each
275 * interesting ordering, collect all the cheapest startup subpaths and all the
276 * cheapest total paths, and build a MergeAppend path for each case.
278 * We don't currently generate any parameterized MergeAppend paths. While
279 * it would not take much more code here to do so, it's very unclear that it
280 * is worth the planning cycles to investigate such paths: there's little
281 * use for an ordered path on the inside of a nestloop. In fact, it's likely
282 * that the current coding of add_path would reject such paths out of hand,
283 * because add_path gives no credit for sort ordering of parameterized paths,
284 * and a parameterized MergeAppend is going to be more expensive than the
285 * corresponding parameterized Append path. If we ever try harder to support
286 * parameterized mergejoin plans, it might be worth adding support for
287 * parameterized MergeAppends to feed such joins. (See notes in
288 * optimizer/README for why that might not ever happen, though.)
291 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
292 List *live_childrels,
293 List *all_child_pathkeys)
297 foreach(lcp, all_child_pathkeys)
299 List *pathkeys = (List *) lfirst(lcp);
300 List *startup_subpaths = NIL;
301 List *total_subpaths = NIL;
302 bool startup_neq_total = false;
305 /* Select the child paths for this ordering... */
306 foreach(lcr, live_childrels)
308 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
309 Path *cheapest_startup,
312 /* Locate the right paths, if they are available. */
314 get_cheapest_path_for_pathkeys(childrel->pathlist,
319 get_cheapest_path_for_pathkeys(childrel->pathlist,
325 * If we can't find any paths with the right order just use the
326 * cheapest-total path; we'll have to sort it later.
328 if (cheapest_startup == NULL || cheapest_total == NULL)
330 cheapest_startup = cheapest_total =
331 childrel->cheapest_total_path;
332 /* Assert we do have an unparameterized path for this child */
333 Assert(cheapest_total->param_info == NULL);
337 * Notice whether we actually have different paths for the
338 * "cheapest" and "total" cases; frequently there will be no point
339 * in two create_merge_append_path() calls.
341 if (cheapest_startup != cheapest_total)
342 startup_neq_total = true;
345 accumulate_append_subpath(startup_subpaths, cheapest_startup);
347 accumulate_append_subpath(total_subpaths, cheapest_total);
350 /* ... and build the MergeAppend paths */
351 add_path(rel, (Path *) create_merge_append_path(root,
356 if (startup_neq_total)
357 add_path(rel, (Path *) create_merge_append_path(root,
366 * get_cheapest_parameterized_child_path
367 * Get cheapest path for this relation that has exactly the requested
370 * Returns NULL if unable to create such a path.
373 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
374 Relids required_outer)
380 * Look up the cheapest existing path with no more than the needed
381 * parameterization. If it has exactly the needed parameterization, we're
384 cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
388 Assert(cheapest != NULL);
389 if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
393 * Otherwise, we can "reparameterize" an existing path to match the given
394 * parameterization, which effectively means pushing down additional
395 * joinquals to be checked within the path's scan. However, some existing
396 * paths might check the available joinquals already while others don't;
397 * therefore, it's not clear which existing path will be cheapest after
398 * reparameterization. We have to go through them all and find out.
401 foreach(lc, rel->pathlist)
403 Path *path = (Path *) lfirst(lc);
405 /* Can't use it if it needs more than requested parameterization */
406 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
410 * Reparameterization can only increase the path's cost, so if it's
411 * already more expensive than the current cheapest, forget it.
413 if (cheapest != NULL &&
414 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
417 /* Reparameterize if needed, then recheck cost */
418 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
420 path = reparameterize_path(root, path, required_outer, 1.0);
422 continue; /* failed to reparameterize this one */
423 Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
425 if (cheapest != NULL &&
426 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
430 /* We have a new best path */
434 /* Return the best path, or NULL if we found no suitable candidate */
439 * accumulate_append_subpath
440 * Add a subpath to the list being built for an Append or MergeAppend
442 * It's possible that the child is itself an Append or MergeAppend path, in
443 * which case we can "cut out the middleman" and just add its child paths to
444 * our own list. (We don't try to do this earlier because we need to apply
445 * both levels of transformation to the quals.)
447 * Note that if we omit a child MergeAppend in this way, we are effectively
448 * omitting a sort step, which seems fine: if the parent is to be an Append,
449 * its result would be unsorted anyway, while if the parent is to be a
450 * MergeAppend, there's no point in a separate sort on a child.
453 accumulate_append_subpath(List *subpaths, Path *path)
455 if (IsA(path, AppendPath))
457 AppendPath *apath = (AppendPath *) path;
459 /* list_copy is important here to avoid sharing list substructure */
460 return list_concat(subpaths, list_copy(apath->subpaths));
462 else if (IsA(path, MergeAppendPath))
464 MergeAppendPath *mpath = (MergeAppendPath *) path;
466 /* list_copy is important here to avoid sharing list substructure */
467 return list_concat(subpaths, list_copy(mpath->subpaths));
470 return lappend(subpaths, path);
474 * standard_join_search
475 * Find possible joinpaths for a query by successively finding ways
476 * to join component relations into join relations.
478 * 'levels_needed' is the number of iterations needed, ie, the number of
479 * independent jointree items in the query. This is > 1.
481 * 'initial_rels' is a list of RelOptInfo nodes for each independent
482 * jointree item. These are the components to be joined together.
483 * Note that levels_needed == list_length(initial_rels).
485 * Returns the final level of join relations, i.e., the relation that is
486 * the result of joining all the original relations together.
487 * At least one implementation path must be provided for this relation and
488 * all required sub-relations.
490 * To support loadable plugins that modify planner behavior by changing the
491 * join searching algorithm, we provide a hook variable that lets a plugin
492 * replace or supplement this function. Any such hook must return the same
493 * final join relation as the standard code would, but it might have a
494 * different set of implementation paths attached, and only the sub-joinrels
495 * needed for these paths need have been instantiated.
497 * Note to plugin authors: the functions invoked during standard_join_search()
498 * modify root->join_rel_list and root->join_rel_hash. If you want to do more
499 * than one join-order search, you'll probably need to save and restore the
500 * original states of those data structures. See geqo_eval() for an example.
503 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
509 * This function cannot be invoked recursively within any one planning
510 * problem, so join_rel_level[] can't be in use already.
512 Assert(root->join_rel_level == NULL);
515 * We employ a simple "dynamic programming" algorithm: we first find all
516 * ways to build joins of two jointree items, then all ways to build joins
517 * of three items (from two-item joins and single items), then four-item
518 * joins, and so on until we have considered all ways to join all the
519 * items into one rel.
521 * root->join_rel_level[j] is a list of all the j-item rels. Initially we
522 * set root->join_rel_level[1] to represent all the single-jointree-item
525 root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
527 root->join_rel_level[1] = initial_rels;
529 for (lev = 2; lev <= levels_needed; lev++)
534 * Determine all possible pairs of relations to be joined at this
535 * level, and build paths for making each one from every available
536 * pair of lower-level relations.
538 join_search_one_level(root, lev);
541 * Run generate_gather_paths() for each just-processed joinrel. We
542 * could not do this earlier because both regular and partial paths
543 * can get added to a particular joinrel at multiple times within
544 * join_search_one_level. After that, we're done creating paths for
545 * the joinrel, so run set_cheapest().
547 foreach(lc, root->join_rel_level[lev])
549 rel = (RelOptInfo *) lfirst(lc);
551 /* Create GatherPaths for any useful partial paths for rel */
552 generate_gather_paths(root, rel);
554 /* Find and save the cheapest paths for this rel */
557 #ifdef OPTIMIZER_DEBUG
558 debug_print_rel(root, rel);
564 * We should have a single rel at the final level.
566 if (root->join_rel_level[levels_needed] == NIL)
567 elog(ERROR, "failed to build any %d-way joins", levels_needed);
568 Assert(list_length(root->join_rel_level[levels_needed]) == 1);
570 rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
572 root->join_rel_level = NULL;
578 * create_plain_partial_paths
579 * Build partial access paths for parallel scan of a plain relation
582 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
584 int parallel_workers;
586 parallel_workers = compute_parallel_worker(rel, rel->pages);
588 /* If any limit was set to zero, the user doesn't want a parallel scan. */
589 if (parallel_workers <= 0)
592 /* Add an unordered partial path based on a parallel sequential scan. */
593 add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
597 * Compute the number of parallel workers that should be used to scan a
598 * relation. "pages" is the number of pages from the relation that we
602 compute_parallel_worker(RelOptInfo *rel, BlockNumber pages)
604 int parallel_workers;
607 * If the user has set the parallel_workers reloption, use that; otherwise
608 * select a default number of workers.
610 if (rel->rel_parallel_workers != -1)
611 parallel_workers = rel->rel_parallel_workers;
614 int parallel_threshold;
617 * If this relation is too small to be worth a parallel scan, just
618 * return without doing anything ... unless it's an inheritance child.
619 * In that case, we want to generate a parallel path here anyway. It
620 * might not be worthwhile just for this relation, but when combined
621 * with all of its inheritance siblings it may well pay off.
623 if (pages < (BlockNumber) min_parallel_relation_size &&
624 rel->reloptkind == RELOPT_BASEREL)
628 * Select the number of workers based on the log of the size of the
629 * relation. This probably needs to be a good deal more
630 * sophisticated, but we need something here for now. Note that the
631 * upper limit of the min_parallel_relation_size GUC is chosen to
632 * prevent overflow here.
634 parallel_workers = 1;
635 parallel_threshold = Max(min_parallel_relation_size, 1);
636 while (pages >= (BlockNumber) (parallel_threshold * 3))
639 parallel_threshold *= 3;
640 if (parallel_threshold > INT_MAX / 3)
641 break; /* avoid overflow */
646 * In no case use more than max_parallel_workers_per_gather workers.
648 parallel_workers = Min(parallel_workers, max_parallel_workers_per_gather);
650 return parallel_workers;
655 * join_search_one_level
656 * Consider ways to produce join relations containing exactly 'level'
657 * jointree items. (This is one step of the dynamic-programming method
658 * embodied in standard_join_search.) Join rel nodes for each feasible
659 * combination of lower-level rels are created and returned in a list.
660 * Implementation paths are created for each such joinrel, too.
662 * level: level of rels we want to make this time
663 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
665 * The result is returned in root->join_rel_level[level].
668 join_search_one_level(PlannerInfo *root, int level)
670 List **joinrels = root->join_rel_level;
674 Assert(joinrels[level] == NIL);
676 /* Set join_cur_level so that new joinrels are added to proper list */
677 root->join_cur_level = level;
680 * First, consider left-sided and right-sided plans, in which rels of
681 * exactly level-1 member relations are joined against initial relations.
682 * We prefer to join using join clauses, but if we find a rel of level-1
683 * members that has no join clauses, we will generate Cartesian-product
684 * joins against all initial rels not already contained in it.
686 foreach(r, joinrels[level - 1])
688 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
690 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
691 has_join_restriction(root, old_rel))
694 * There are join clauses or join order restrictions relevant to
695 * this rel, so consider joins between this rel and (only) those
696 * initial rels it is linked to by a clause or restriction.
698 * At level 2 this condition is symmetric, so there is no need to
699 * look at initial rels before this one in the list; we already
700 * considered such joins when we were at the earlier rel. (The
701 * mirror-image joins are handled automatically by make_join_rel.)
702 * In later passes (level > 2), we join rels of the previous level
703 * to each initial rel they don't already include but have a join
704 * clause or restriction with.
706 ListCell *other_rels;
708 if (level == 2) /* consider remaining initial rels */
709 other_rels = lnext(r);
710 else /* consider all initial rels */
711 other_rels = list_head(joinrels[1]);
713 make_rels_by_clause_joins(root,
720 * Oops, we have a relation that is not joined to any other
721 * relation, either directly or by join-order restrictions.
722 * Cartesian product time.
724 * We consider a cartesian product with each not-already-included
725 * initial rel, whether it has other join clauses or not. At
726 * level 2, if there are two or more clauseless initial rels, we
727 * will redundantly consider joining them in both directions; but
728 * such cases aren't common enough to justify adding complexity to
729 * avoid the duplicated effort.
731 make_rels_by_clauseless_joins(root,
733 list_head(joinrels[1]));
738 * Now, consider "bushy plans" in which relations of k initial rels are
739 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
741 * We only consider bushy-plan joins for pairs of rels where there is a
742 * suitable join clause (or join order restriction), in order to avoid
743 * unreasonable growth of planning time.
747 int other_level = level - k;
750 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
751 * need to go as far as the halfway point.
756 foreach(r, joinrels[k])
758 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
759 ListCell *other_rels;
763 * We can ignore relations without join clauses here, unless they
764 * participate in join-order restrictions --- then we might have
765 * to force a bushy join plan.
767 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
768 !has_join_restriction(root, old_rel))
771 if (k == other_level)
772 other_rels = lnext(r); /* only consider remaining rels */
774 other_rels = list_head(joinrels[other_level]);
776 for_each_cell(r2, other_rels)
778 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
780 if (!bms_overlap(old_rel->relids, new_rel->relids))
783 * OK, we can build a rel of the right level from this
784 * pair of rels. Do so if there is at least one relevant
785 * join clause or join order restriction.
787 if (have_relevant_joinclause(root, old_rel, new_rel) ||
788 have_join_order_restriction(root, old_rel, new_rel))
790 (void) make_join_rel(root, old_rel, new_rel);
798 * Last-ditch effort: if we failed to find any usable joins so far, force
799 * a set of cartesian-product joins to be generated. This handles the
800 * special case where all the available rels have join clauses but we
801 * cannot use any of those clauses yet. This can only happen when we are
802 * considering a join sub-problem (a sub-joinlist) and all the rels in the
803 * sub-problem have only join clauses with rels outside the sub-problem.
806 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
807 * WHERE a.w = c.x and b.y = d.z;
809 * If the "a INNER JOIN b" sub-problem does not get flattened into the
810 * upper level, we must be willing to make a cartesian join of a and b;
811 * but the code above will not have done so, because it thought that both
812 * a and b have joinclauses. We consider only left-sided and right-sided
813 * cartesian joins in this case (no bushy).
816 if (joinrels[level] == NIL)
819 * This loop is just like the first one, except we always call
820 * make_rels_by_clauseless_joins().
822 foreach(r, joinrels[level - 1])
824 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
826 make_rels_by_clauseless_joins(root,
828 list_head(joinrels[1]));
832 * When special joins are involved, there may be no legal way
833 * to make an N-way join for some values of N. For example consider
835 * SELECT ... FROM t1 WHERE
836 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
837 * y IN (SELECT ... FROM t4,t5 WHERE ...)
839 * We will flatten this query to a 5-way join problem, but there are
840 * no 4-way joins that join_is_legal() will consider legal. We have
841 * to accept failure at level 4 and go on to discover a workable
842 * bushy plan at level 5.
844 * However, if there are no special joins and no lateral references
845 * then join_is_legal() should never fail, and so the following sanity
849 if (joinrels[level] == NIL &&
850 root->join_info_list == NIL &&
851 !root->hasLateralRTEs)
852 elog(ERROR, "failed to build any %d-way joins", level);
857 * make_rels_by_clause_joins
858 * Build joins between the given relation 'old_rel' and other relations
859 * that participate in join clauses that 'old_rel' also participates in
860 * (or participate in join-order restrictions with it).
861 * The join rels are returned in root->join_rel_level[join_cur_level].
863 * Note: at levels above 2 we will generate the same joined relation in
864 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
865 * (b join c) join a, though the second case will add a different set of Paths
866 * to it. This is the reason for using the join_rel_level mechanism, which
867 * automatically ensures that each new joinrel is only added to the list once.
869 * 'old_rel' is the relation entry for the relation to be joined
870 * 'other_rels': the first cell in a linked list containing the other
871 * rels to be considered for joining
873 * Currently, this is only used with initial rels in other_rels, but it
874 * will work for joining to joinrels too.
877 make_rels_by_clause_joins(PlannerInfo *root,
879 ListCell *other_rels)
883 for_each_cell(l, other_rels)
885 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
887 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
888 (have_relevant_joinclause(root, old_rel, other_rel) ||
889 have_join_order_restriction(root, old_rel, other_rel)))
891 (void) make_join_rel(root, old_rel, other_rel);
897 * make_rels_by_clauseless_joins
898 * Given a relation 'old_rel' and a list of other relations
899 * 'other_rels', create a join relation between 'old_rel' and each
900 * member of 'other_rels' that isn't already included in 'old_rel'.
901 * The join rels are returned in root->join_rel_level[join_cur_level].
903 * 'old_rel' is the relation entry for the relation to be joined
904 * 'other_rels': the first cell of a linked list containing the
905 * other rels to be considered for joining
907 * Currently, this is only used with initial rels in other_rels, but it would
908 * work for joining to joinrels too.
911 make_rels_by_clauseless_joins(PlannerInfo *root,
913 ListCell *other_rels)
917 for_each_cell(l, other_rels)
919 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
921 if (!bms_overlap(other_rel->relids, old_rel->relids))
923 (void) make_join_rel(root, old_rel, other_rel);
930 * Determine whether a proposed join is legal given the query's
931 * join order constraints; and if it is, determine the join type.
933 * Caller must supply not only the two rels, but the union of their relids.
934 * (We could simplify the API by computing joinrelids locally, but this
935 * would be redundant work in the normal path through make_join_rel.)
937 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
938 * else it's set to point to the associated SpecialJoinInfo node. Also,
939 * *reversed_p is set TRUE if the given relations need to be swapped to
940 * match the SpecialJoinInfo node.
943 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
945 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
947 SpecialJoinInfo *match_sjinfo;
950 bool must_be_leftjoin;
954 * Ensure output params are set on failure return. This is just to
955 * suppress uninitialized-variable warnings from overly anal compilers.
961 * If we have any special joins, the proposed join might be illegal; and
962 * in any case we have to determine its join type. Scan the join info
963 * list for matches and conflicts.
967 unique_ified = false;
968 must_be_leftjoin = false;
970 foreach(l, root->join_info_list)
972 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
975 * This special join is not relevant unless its RHS overlaps the
976 * proposed join. (Check this first as a fast path for dismissing
977 * most irrelevant SJs quickly.)
979 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
983 * Also, not relevant if proposed join is fully contained within RHS
984 * (ie, we're still building up the RHS).
986 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
990 * Also, not relevant if SJ is already done within either input.
992 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
993 bms_is_subset(sjinfo->min_righthand, rel1->relids))
995 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
996 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1000 * If it's a semijoin and we already joined the RHS to any other rels
1001 * within either input, then we must have unique-ified the RHS at that
1002 * point (see below). Therefore the semijoin is no longer relevant in
1005 if (sjinfo->jointype == JOIN_SEMI)
1007 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1008 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1010 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1011 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1016 * If one input contains min_lefthand and the other contains
1017 * min_righthand, then we can perform the SJ at this join.
1019 * Reject if we get matches to more than one SJ; that implies we're
1020 * considering something that's not really valid.
1022 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1023 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1026 return false; /* invalid join path */
1027 match_sjinfo = sjinfo;
1030 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1031 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1034 return false; /* invalid join path */
1035 match_sjinfo = sjinfo;
1038 else if (sjinfo->jointype == JOIN_SEMI &&
1039 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1040 create_unique_path(root, rel2, rel2->cheapest_total_path,
1044 * For a semijoin, we can join the RHS to anything else by
1045 * unique-ifying the RHS (if the RHS can be unique-ified).
1046 * We will only get here if we have the full RHS but less
1047 * than min_lefthand on the LHS.
1049 * The reason to consider such a join path is exemplified by
1050 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1051 * If we insist on doing this as a semijoin we will first have
1052 * to form the cartesian product of A*B. But if we unique-ify
1053 * C then the semijoin becomes a plain innerjoin and we can join
1054 * in any order, eg C to A and then to B. When C is much smaller
1055 * than A and B this can be a huge win. So we allow C to be
1056 * joined to just A or just B here, and then make_join_rel has
1057 * to handle the case properly.
1059 * Note that actually we'll allow unique-ified C to be joined to
1060 * some other relation D here, too. That is legal, if usually not
1061 * very sane, and this routine is only concerned with legality not
1062 * with whether the join is good strategy.
1066 return false; /* invalid join path */
1067 match_sjinfo = sjinfo;
1069 unique_ified = true;
1071 else if (sjinfo->jointype == JOIN_SEMI &&
1072 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1073 create_unique_path(root, rel1, rel1->cheapest_total_path,
1076 /* Reversed semijoin case */
1078 return false; /* invalid join path */
1079 match_sjinfo = sjinfo;
1081 unique_ified = true;
1086 * Otherwise, the proposed join overlaps the RHS but isn't a valid
1087 * implementation of this SJ. But don't panic quite yet: the RHS
1088 * violation might have occurred previously, in one or both input
1089 * relations, in which case we must have previously decided that
1090 * it was OK to commute some other SJ with this one. If we need
1091 * to perform this join to finish building up the RHS, rejecting
1092 * it could lead to not finding any plan at all. (This can occur
1093 * because of the heuristics elsewhere in this file that postpone
1094 * clauseless joins: we might not consider doing a clauseless join
1095 * within the RHS until after we've performed other, validly
1096 * commutable SJs with one or both sides of the clauseless join.)
1097 * This consideration boils down to the rule that if both inputs
1098 * overlap the RHS, we can allow the join --- they are either
1099 * fully within the RHS, or represent previously-allowed joins to
1102 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1103 bms_overlap(rel2->relids, sjinfo->min_righthand))
1104 continue; /* assume valid previous violation of RHS */
1107 * The proposed join could still be legal, but only if we're
1108 * allowed to associate it into the RHS of this SJ. That means
1109 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1110 * not FULL) and the proposed join must not overlap the LHS.
1112 if (sjinfo->jointype != JOIN_LEFT ||
1113 bms_overlap(joinrelids, sjinfo->min_lefthand))
1114 return false; /* invalid join path */
1117 * To be valid, the proposed join must be a LEFT join; otherwise
1118 * it can't associate into this SJ's RHS. But we may not yet have
1119 * found the SpecialJoinInfo matching the proposed join, so we
1120 * can't test that yet. Remember the requirement for later.
1122 must_be_leftjoin = true;
1127 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1128 * proposed join can't associate into an SJ's RHS.
1130 * Also, fail if the proposed join's predicate isn't strict; we're
1131 * essentially checking to see if we can apply outer-join identity 3, and
1132 * that's a requirement. (This check may be redundant with checks in
1133 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1135 if (must_be_leftjoin &&
1136 (match_sjinfo == NULL ||
1137 match_sjinfo->jointype != JOIN_LEFT ||
1138 !match_sjinfo->lhs_strict))
1139 return false; /* invalid join path */
1142 * We also have to check for constraints imposed by LATERAL references.
1144 if (root->hasLateralRTEs)
1148 Relids join_lateral_rels;
1151 * The proposed rels could each contain lateral references to the
1152 * other, in which case the join is impossible. If there are lateral
1153 * references in just one direction, then the join has to be done with
1154 * a nestloop with the lateral referencer on the inside. If the join
1155 * matches an SJ that cannot be implemented by such a nestloop, the
1156 * join is impossible.
1158 * Also, if the lateral reference is only indirect, we should reject
1159 * the join; whatever rel(s) the reference chain goes through must be
1162 * Another case that might keep us from building a valid plan is the
1163 * implementation restriction described by have_dangerous_phv().
1165 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1166 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1167 if (lateral_fwd && lateral_rev)
1168 return false; /* have lateral refs in both directions */
1171 /* has to be implemented as nestloop with rel1 on left */
1175 match_sjinfo->jointype == JOIN_FULL))
1176 return false; /* not implementable as nestloop */
1177 /* check there is a direct reference from rel2 to rel1 */
1178 if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1179 return false; /* only indirect refs, so reject */
1180 /* check we won't have a dangerous PHV */
1181 if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1182 return false; /* might be unable to handle required PHV */
1184 else if (lateral_rev)
1186 /* has to be implemented as nestloop with rel2 on left */
1190 match_sjinfo->jointype == JOIN_FULL))
1191 return false; /* not implementable as nestloop */
1192 /* check there is a direct reference from rel1 to rel2 */
1193 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1194 return false; /* only indirect refs, so reject */
1195 /* check we won't have a dangerous PHV */
1196 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1197 return false; /* might be unable to handle required PHV */
1201 * LATERAL references could also cause problems later on if we accept
1202 * this join: if the join's minimum parameterization includes any rels
1203 * that would have to be on the inside of an outer join with this join
1204 * rel, then it's never going to be possible to build the complete
1205 * query using this join. We should reject this join not only because
1206 * it'll save work, but because if we don't, the clauseless-join
1207 * heuristics might think that legality of this join means that some
1208 * other join rel need not be formed, and that could lead to failure
1209 * to find any plan at all. We have to consider not only rels that
1210 * are directly on the inner side of an OJ with the joinrel, but also
1211 * ones that are indirectly so, so search to find all such rels.
1213 join_lateral_rels = min_join_parameterization(root, joinrelids,
1215 if (join_lateral_rels)
1217 Relids join_plus_rhs = bms_copy(joinrelids);
1223 foreach(l, root->join_info_list)
1225 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1227 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1228 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1230 join_plus_rhs = bms_add_members(join_plus_rhs,
1231 sjinfo->min_righthand);
1234 /* full joins constrain both sides symmetrically */
1235 if (sjinfo->jointype == JOIN_FULL &&
1236 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1237 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1239 join_plus_rhs = bms_add_members(join_plus_rhs,
1240 sjinfo->min_lefthand);
1245 if (bms_overlap(join_plus_rhs, join_lateral_rels))
1246 return false; /* will not be able to join to some RHS rel */
1250 /* Otherwise, it's a valid join */
1251 *sjinfo_p = match_sjinfo;
1252 *reversed_p = reversed;
1257 * has_join_restriction
1258 * Detect whether the specified relation has join-order restrictions,
1259 * due to being inside an outer join or an IN (sub-SELECT),
1260 * or participating in any LATERAL references or multi-rel PHVs.
1262 * Essentially, this tests whether have_join_order_restriction() could
1263 * succeed with this rel and some other one. It's OK if we sometimes
1264 * say "true" incorrectly. (Therefore, we don't bother with the relatively
1265 * expensive has_legal_joinclause test.)
1268 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1272 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1275 foreach(l, root->placeholder_list)
1277 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1279 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1280 !bms_equal(rel->relids, phinfo->ph_eval_at))
1284 foreach(l, root->join_info_list)
1286 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1288 /* ignore full joins --- other mechanisms preserve their ordering */
1289 if (sjinfo->jointype == JOIN_FULL)
1292 /* ignore if SJ is already contained in rel */
1293 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1294 bms_is_subset(sjinfo->min_righthand, rel->relids))
1297 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1298 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1299 bms_overlap(sjinfo->min_righthand, rel->relids))
1307 * is_dummy_rel --- has relation been proven empty?
1310 is_dummy_rel(RelOptInfo *rel)
1312 return IS_DUMMY_REL(rel);
1316 * Mark a relation as proven empty.
1318 * During GEQO planning, this can get invoked more than once on the same
1319 * baserel struct, so it's worth checking to see if the rel is already marked
1322 * Also, when called during GEQO join planning, we are in a short-lived
1323 * memory context. We must make sure that the dummy path attached to a
1324 * baserel survives the GEQO cycle, else the baserel is trashed for future
1325 * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1326 * we don't want the dummy path to clutter the main planning context. Upshot
1327 * is that the best solution is to explicitly make the dummy path in the same
1328 * context the given RelOptInfo is in.
1331 mark_dummy_rel(RelOptInfo *rel)
1333 MemoryContext oldcontext;
1335 /* Already marked? */
1336 if (is_dummy_rel(rel))
1339 /* No, so choose correct context to make the dummy path in */
1340 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1342 /* Set dummy size estimate */
1345 /* Evict any previously chosen paths */
1346 rel->pathlist = NIL;
1347 rel->partial_pathlist = NIL;
1349 /* Set up the dummy path */
1350 add_path(rel, (Path *) create_append_path(rel, NIL, NULL, 0));
1352 /* Set or update cheapest_total_path and related fields */
1355 MemoryContextSwitchTo(oldcontext);
1359 * restriction_is_constant_false --- is a restrictlist just FALSE?
1361 * In cases where a qual is provably constant FALSE, eval_const_expressions
1362 * will generally have thrown away anything that's ANDed with it. In outer
1363 * join situations this will leave us computing cartesian products only to
1364 * decide there's no match for an outer row, which is pretty stupid. So,
1365 * we need to detect the case.
1367 * If only_pushed_down is TRUE, then consider only pushed-down quals.
1370 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1375 * Despite the above comment, the restriction list we see here might
1376 * possibly have other members besides the FALSE constant, since other
1377 * quals could get "pushed down" to the outer join level. So we check
1378 * each member of the list.
1380 foreach(lc, restrictlist)
1382 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1384 Assert(IsA(rinfo, RestrictInfo));
1385 if (only_pushed_down && !rinfo->is_pushed_down)
1388 if (rinfo->clause && IsA(rinfo->clause, Const))
1390 Const *con = (Const *) rinfo->clause;
1392 /* constant NULL is as good as constant FALSE for our purposes */
1393 if (con->constisnull)
1395 if (!DatumGetBool(con->constvalue))