1 /*-------------------------------------------------------------------------
4 * Routines copied from PostgreSQL core distribution.
7 * The main purpose of this files is having access to static functions in core.
8 * Another purpose is tweaking functions behavior by replacing part of them by
9 * macro definitions. See at the end of pg_hint_plan.c for details. Anyway,
10 * this file *must* contain required functions without making any change.
12 * This file contains the following functions from corresponding files.
14 * src/backend/optimizer/path/allpaths.c
17 * set_plain_rel_pathlist()
18 * create_plain_partial_paths()
19 * set_append_rel_pathlist()
20 * generate_mergeappend_paths()
21 * get_cheapest_parameterized_child_path()
22 * accumulate_append_subpath()
25 * standard_join_search(): This funcion is not static. The reason for
26 * including this function is make_rels_by_clause_joins. In order to
27 * avoid generating apparently unwanted join combination, we decided to
28 * change the behavior of make_join_rel, which is called under this
31 * src/backend/optimizer/path/joinrels.c
34 * join_search_one_level(): We have to modify this to call my definition of
35 * make_rels_by_clause_joins.
38 * make_rels_by_clause_joins()
39 * make_rels_by_clauseless_joins()
41 * has_join_restriction()
44 * restriction_is_constant_false()
47 * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
48 * Portions Copyright (c) 1994, Regents of the University of California
50 *-------------------------------------------------------------------------
55 * set_plain_rel_pathlist
56 * Build access paths for a plain relation (no subquery, no inheritance)
59 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
61 Relids required_outer;
64 * We don't support pushing join clauses into the quals of a seqscan, but
65 * it could still have required parameterization due to LATERAL refs in
68 required_outer = rel->lateral_relids;
70 /* Consider sequential scan */
71 add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
73 /* If appropriate, consider parallel sequential scan */
74 if (rel->consider_parallel && required_outer == NULL)
75 create_plain_partial_paths(root, rel);
77 /* Consider index scans */
78 create_index_paths(root, rel);
80 /* Consider TID scans */
81 create_tidscan_paths(root, rel);
85 * create_plain_partial_paths
86 * Build partial access paths for parallel scan of a plain relation
89 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
94 * If the user has set the parallel_workers reloption, use that; otherwise
95 * select a default number of workers.
97 if (rel->rel_parallel_workers != -1)
98 parallel_workers = rel->rel_parallel_workers;
101 int parallel_threshold;
104 * If this relation is too small to be worth a parallel scan, just
105 * return without doing anything ... unless it's an inheritance child.
106 * In that case, we want to generate a parallel path here anyway. It
107 * might not be worthwhile just for this relation, but when combined
108 * with all of its inheritance siblings it may well pay off.
110 if (rel->pages < (BlockNumber) min_parallel_relation_size &&
111 rel->reloptkind == RELOPT_BASEREL)
115 * Select the number of workers based on the log of the size of the
116 * relation. This probably needs to be a good deal more
117 * sophisticated, but we need something here for now. Note that the
118 * upper limit of the min_parallel_relation_size GUC is chosen to
119 * prevent overflow here.
121 parallel_workers = 1;
122 parallel_threshold = Max(min_parallel_relation_size, 1);
123 while (rel->pages >= (BlockNumber) (parallel_threshold * 3))
126 parallel_threshold *= 3;
127 if (parallel_threshold > INT_MAX / 3)
128 break; /* avoid overflow */
133 * In no case use more than max_parallel_workers_per_gather workers.
135 parallel_workers = Min(parallel_workers, max_parallel_workers_per_gather);
137 /* If any limit was set to zero, the user doesn't want a parallel scan. */
138 if (parallel_workers <= 0)
141 /* Add an unordered partial path based on a parallel sequential scan. */
142 add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
146 * set_append_rel_pathlist
147 * Build access paths for an "append relation"
150 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
151 Index rti, RangeTblEntry *rte)
153 int parentRTindex = rti;
154 List *live_childrels = NIL;
155 List *subpaths = NIL;
156 bool subpaths_valid = true;
157 List *partial_subpaths = NIL;
158 bool partial_subpaths_valid = true;
159 List *all_child_pathkeys = NIL;
160 List *all_child_outers = NIL;
164 * Generate access paths for each member relation, and remember the
165 * cheapest path for each one. Also, identify all pathkeys (orderings)
166 * and parameterizations (required_outer sets) available for the member
169 foreach(l, root->append_rel_list)
171 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
173 RangeTblEntry *childRTE;
174 RelOptInfo *childrel;
177 /* append_rel_list contains all append rels; ignore others */
178 if (appinfo->parent_relid != parentRTindex)
181 /* Re-locate the child RTE and RelOptInfo */
182 childRTindex = appinfo->child_relid;
183 childRTE = root->simple_rte_array[childRTindex];
184 childrel = root->simple_rel_array[childRTindex];
187 * If set_append_rel_size() decided the parent appendrel was
188 * parallel-unsafe at some point after visiting this child rel, we
189 * need to propagate the unsafety marking down to the child, so that
190 * we don't generate useless partial paths for it.
192 if (!rel->consider_parallel)
193 childrel->consider_parallel = false;
196 * Compute the child's access paths.
198 set_rel_pathlist(root, childrel, childRTindex, childRTE);
201 * If child is dummy, ignore it.
203 if (IS_DUMMY_REL(childrel))
207 * Child is live, so add it to the live_childrels list for use below.
209 live_childrels = lappend(live_childrels, childrel);
212 * If child has an unparameterized cheapest-total path, add that to
213 * the unparameterized Append path we are constructing for the parent.
214 * If not, there's no workable unparameterized path.
216 if (childrel->cheapest_total_path->param_info == NULL)
217 subpaths = accumulate_append_subpath(subpaths,
218 childrel->cheapest_total_path);
220 subpaths_valid = false;
222 /* Same idea, but for a partial plan. */
223 if (childrel->partial_pathlist != NIL)
224 partial_subpaths = accumulate_append_subpath(partial_subpaths,
225 linitial(childrel->partial_pathlist));
227 partial_subpaths_valid = false;
230 * Collect lists of all the available path orderings and
231 * parameterizations for all the children. We use these as a
232 * heuristic to indicate which sort orderings and parameterizations we
233 * should build Append and MergeAppend paths for.
235 foreach(lcp, childrel->pathlist)
237 Path *childpath = (Path *) lfirst(lcp);
238 List *childkeys = childpath->pathkeys;
239 Relids childouter = PATH_REQ_OUTER(childpath);
241 /* Unsorted paths don't contribute to pathkey list */
242 if (childkeys != NIL)
247 /* Have we already seen this ordering? */
248 foreach(lpk, all_child_pathkeys)
250 List *existing_pathkeys = (List *) lfirst(lpk);
252 if (compare_pathkeys(existing_pathkeys,
253 childkeys) == PATHKEYS_EQUAL)
261 /* No, so add it to all_child_pathkeys */
262 all_child_pathkeys = lappend(all_child_pathkeys,
267 /* Unparameterized paths don't contribute to param-set list */
273 /* Have we already seen this param set? */
274 foreach(lco, all_child_outers)
276 Relids existing_outers = (Relids) lfirst(lco);
278 if (bms_equal(existing_outers, childouter))
286 /* No, so add it to all_child_outers */
287 all_child_outers = lappend(all_child_outers,
295 * If we found unparameterized paths for all children, build an unordered,
296 * unparameterized Append path for the rel. (Note: this is correct even
297 * if we have zero or one live subpath due to constraint exclusion.)
300 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0));
303 * Consider an append of partial unordered, unparameterized partial paths.
305 if (partial_subpaths_valid)
307 AppendPath *appendpath;
309 int parallel_workers = 0;
312 * Decide on the number of workers to request for this append path.
313 * For now, we just use the maximum value from among the members. It
314 * might be useful to use a higher number if the Append node were
315 * smart enough to spread out the workers, but it currently isn't.
317 foreach(lc, partial_subpaths)
319 Path *path = lfirst(lc);
321 parallel_workers = Max(parallel_workers, path->parallel_workers);
323 Assert(parallel_workers > 0);
325 /* Generate a partial append path. */
326 appendpath = create_append_path(rel, partial_subpaths, NULL,
328 add_partial_path(rel, (Path *) appendpath);
332 * Also build unparameterized MergeAppend paths based on the collected
333 * list of child pathkeys.
336 generate_mergeappend_paths(root, rel, live_childrels,
340 * Build Append paths for each parameterization seen among the child rels.
341 * (This may look pretty expensive, but in most cases of practical
342 * interest, the child rels will expose mostly the same parameterizations,
343 * so that not that many cases actually get considered here.)
345 * The Append node itself cannot enforce quals, so all qual checking must
346 * be done in the child paths. This means that to have a parameterized
347 * Append path, we must have the exact same parameterization for each
348 * child path; otherwise some children might be failing to check the
349 * moved-down quals. To make them match up, we can try to increase the
350 * parameterization of lesser-parameterized paths.
352 foreach(l, all_child_outers)
354 Relids required_outer = (Relids) lfirst(l);
357 /* Select the child paths for an Append with this parameterization */
359 subpaths_valid = true;
360 foreach(lcr, live_childrels)
362 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
365 subpath = get_cheapest_parameterized_child_path(root,
370 /* failed to make a suitable path for this child */
371 subpaths_valid = false;
374 subpaths = accumulate_append_subpath(subpaths, subpath);
378 add_path(rel, (Path *)
379 create_append_path(rel, subpaths, required_outer, 0));
384 * generate_mergeappend_paths
385 * Generate MergeAppend paths for an append relation
387 * Generate a path for each ordering (pathkey list) appearing in
388 * all_child_pathkeys.
390 * We consider both cheapest-startup and cheapest-total cases, ie, for each
391 * interesting ordering, collect all the cheapest startup subpaths and all the
392 * cheapest total paths, and build a MergeAppend path for each case.
394 * We don't currently generate any parameterized MergeAppend paths. While
395 * it would not take much more code here to do so, it's very unclear that it
396 * is worth the planning cycles to investigate such paths: there's little
397 * use for an ordered path on the inside of a nestloop. In fact, it's likely
398 * that the current coding of add_path would reject such paths out of hand,
399 * because add_path gives no credit for sort ordering of parameterized paths,
400 * and a parameterized MergeAppend is going to be more expensive than the
401 * corresponding parameterized Append path. If we ever try harder to support
402 * parameterized mergejoin plans, it might be worth adding support for
403 * parameterized MergeAppends to feed such joins. (See notes in
404 * optimizer/README for why that might not ever happen, though.)
407 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
408 List *live_childrels,
409 List *all_child_pathkeys)
413 foreach(lcp, all_child_pathkeys)
415 List *pathkeys = (List *) lfirst(lcp);
416 List *startup_subpaths = NIL;
417 List *total_subpaths = NIL;
418 bool startup_neq_total = false;
421 /* Select the child paths for this ordering... */
422 foreach(lcr, live_childrels)
424 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
425 Path *cheapest_startup,
428 /* Locate the right paths, if they are available. */
430 get_cheapest_path_for_pathkeys(childrel->pathlist,
435 get_cheapest_path_for_pathkeys(childrel->pathlist,
441 * If we can't find any paths with the right order just use the
442 * cheapest-total path; we'll have to sort it later.
444 if (cheapest_startup == NULL || cheapest_total == NULL)
446 cheapest_startup = cheapest_total =
447 childrel->cheapest_total_path;
448 /* Assert we do have an unparameterized path for this child */
449 Assert(cheapest_total->param_info == NULL);
453 * Notice whether we actually have different paths for the
454 * "cheapest" and "total" cases; frequently there will be no point
455 * in two create_merge_append_path() calls.
457 if (cheapest_startup != cheapest_total)
458 startup_neq_total = true;
461 accumulate_append_subpath(startup_subpaths, cheapest_startup);
463 accumulate_append_subpath(total_subpaths, cheapest_total);
466 /* ... and build the MergeAppend paths */
467 add_path(rel, (Path *) create_merge_append_path(root,
472 if (startup_neq_total)
473 add_path(rel, (Path *) create_merge_append_path(root,
482 * get_cheapest_parameterized_child_path
483 * Get cheapest path for this relation that has exactly the requested
486 * Returns NULL if unable to create such a path.
489 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
490 Relids required_outer)
496 * Look up the cheapest existing path with no more than the needed
497 * parameterization. If it has exactly the needed parameterization, we're
500 cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
504 Assert(cheapest != NULL);
505 if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
509 * Otherwise, we can "reparameterize" an existing path to match the given
510 * parameterization, which effectively means pushing down additional
511 * joinquals to be checked within the path's scan. However, some existing
512 * paths might check the available joinquals already while others don't;
513 * therefore, it's not clear which existing path will be cheapest after
514 * reparameterization. We have to go through them all and find out.
517 foreach(lc, rel->pathlist)
519 Path *path = (Path *) lfirst(lc);
521 /* Can't use it if it needs more than requested parameterization */
522 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
526 * Reparameterization can only increase the path's cost, so if it's
527 * already more expensive than the current cheapest, forget it.
529 if (cheapest != NULL &&
530 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
533 /* Reparameterize if needed, then recheck cost */
534 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
536 path = reparameterize_path(root, path, required_outer, 1.0);
538 continue; /* failed to reparameterize this one */
539 Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
541 if (cheapest != NULL &&
542 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
546 /* We have a new best path */
550 /* Return the best path, or NULL if we found no suitable candidate */
555 * accumulate_append_subpath
556 * Add a subpath to the list being built for an Append or MergeAppend
558 * It's possible that the child is itself an Append or MergeAppend path, in
559 * which case we can "cut out the middleman" and just add its child paths to
560 * our own list. (We don't try to do this earlier because we need to apply
561 * both levels of transformation to the quals.)
563 * Note that if we omit a child MergeAppend in this way, we are effectively
564 * omitting a sort step, which seems fine: if the parent is to be an Append,
565 * its result would be unsorted anyway, while if the parent is to be a
566 * MergeAppend, there's no point in a separate sort on a child.
569 accumulate_append_subpath(List *subpaths, Path *path)
571 if (IsA(path, AppendPath))
573 AppendPath *apath = (AppendPath *) path;
575 /* list_copy is important here to avoid sharing list substructure */
576 return list_concat(subpaths, list_copy(apath->subpaths));
578 else if (IsA(path, MergeAppendPath))
580 MergeAppendPath *mpath = (MergeAppendPath *) path;
582 /* list_copy is important here to avoid sharing list substructure */
583 return list_concat(subpaths, list_copy(mpath->subpaths));
586 return lappend(subpaths, path);
590 * standard_join_search
591 * Find possible joinpaths for a query by successively finding ways
592 * to join component relations into join relations.
594 * 'levels_needed' is the number of iterations needed, ie, the number of
595 * independent jointree items in the query. This is > 1.
597 * 'initial_rels' is a list of RelOptInfo nodes for each independent
598 * jointree item. These are the components to be joined together.
599 * Note that levels_needed == list_length(initial_rels).
601 * Returns the final level of join relations, i.e., the relation that is
602 * the result of joining all the original relations together.
603 * At least one implementation path must be provided for this relation and
604 * all required sub-relations.
606 * To support loadable plugins that modify planner behavior by changing the
607 * join searching algorithm, we provide a hook variable that lets a plugin
608 * replace or supplement this function. Any such hook must return the same
609 * final join relation as the standard code would, but it might have a
610 * different set of implementation paths attached, and only the sub-joinrels
611 * needed for these paths need have been instantiated.
613 * Note to plugin authors: the functions invoked during standard_join_search()
614 * modify root->join_rel_list and root->join_rel_hash. If you want to do more
615 * than one join-order search, you'll probably need to save and restore the
616 * original states of those data structures. See geqo_eval() for an example.
619 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
625 * This function cannot be invoked recursively within any one planning
626 * problem, so join_rel_level[] can't be in use already.
628 Assert(root->join_rel_level == NULL);
631 * We employ a simple "dynamic programming" algorithm: we first find all
632 * ways to build joins of two jointree items, then all ways to build joins
633 * of three items (from two-item joins and single items), then four-item
634 * joins, and so on until we have considered all ways to join all the
635 * items into one rel.
637 * root->join_rel_level[j] is a list of all the j-item rels. Initially we
638 * set root->join_rel_level[1] to represent all the single-jointree-item
641 root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
643 root->join_rel_level[1] = initial_rels;
645 for (lev = 2; lev <= levels_needed; lev++)
650 * Determine all possible pairs of relations to be joined at this
651 * level, and build paths for making each one from every available
652 * pair of lower-level relations.
654 join_search_one_level(root, lev);
657 * Run generate_gather_paths() for each just-processed joinrel. We
658 * could not do this earlier because both regular and partial paths
659 * can get added to a particular joinrel at multiple times within
660 * join_search_one_level. After that, we're done creating paths for
661 * the joinrel, so run set_cheapest().
663 foreach(lc, root->join_rel_level[lev])
665 rel = (RelOptInfo *) lfirst(lc);
667 /* Create GatherPaths for any useful partial paths for rel */
668 generate_gather_paths(root, rel);
670 /* Find and save the cheapest paths for this rel */
673 #ifdef OPTIMIZER_DEBUG
674 debug_print_rel(root, rel);
680 * We should have a single rel at the final level.
682 if (root->join_rel_level[levels_needed] == NIL)
683 elog(ERROR, "failed to build any %d-way joins", levels_needed);
684 Assert(list_length(root->join_rel_level[levels_needed]) == 1);
686 rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
688 root->join_rel_level = NULL;
694 * join_search_one_level
695 * Consider ways to produce join relations containing exactly 'level'
696 * jointree items. (This is one step of the dynamic-programming method
697 * embodied in standard_join_search.) Join rel nodes for each feasible
698 * combination of lower-level rels are created and returned in a list.
699 * Implementation paths are created for each such joinrel, too.
701 * level: level of rels we want to make this time
702 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
704 * The result is returned in root->join_rel_level[level].
707 join_search_one_level(PlannerInfo *root, int level)
709 List **joinrels = root->join_rel_level;
713 Assert(joinrels[level] == NIL);
715 /* Set join_cur_level so that new joinrels are added to proper list */
716 root->join_cur_level = level;
719 * First, consider left-sided and right-sided plans, in which rels of
720 * exactly level-1 member relations are joined against initial relations.
721 * We prefer to join using join clauses, but if we find a rel of level-1
722 * members that has no join clauses, we will generate Cartesian-product
723 * joins against all initial rels not already contained in it.
725 foreach(r, joinrels[level - 1])
727 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
729 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
730 has_join_restriction(root, old_rel))
733 * There are join clauses or join order restrictions relevant to
734 * this rel, so consider joins between this rel and (only) those
735 * initial rels it is linked to by a clause or restriction.
737 * At level 2 this condition is symmetric, so there is no need to
738 * look at initial rels before this one in the list; we already
739 * considered such joins when we were at the earlier rel. (The
740 * mirror-image joins are handled automatically by make_join_rel.)
741 * In later passes (level > 2), we join rels of the previous level
742 * to each initial rel they don't already include but have a join
743 * clause or restriction with.
745 ListCell *other_rels;
747 if (level == 2) /* consider remaining initial rels */
748 other_rels = lnext(r);
749 else /* consider all initial rels */
750 other_rels = list_head(joinrels[1]);
752 make_rels_by_clause_joins(root,
759 * Oops, we have a relation that is not joined to any other
760 * relation, either directly or by join-order restrictions.
761 * Cartesian product time.
763 * We consider a cartesian product with each not-already-included
764 * initial rel, whether it has other join clauses or not. At
765 * level 2, if there are two or more clauseless initial rels, we
766 * will redundantly consider joining them in both directions; but
767 * such cases aren't common enough to justify adding complexity to
768 * avoid the duplicated effort.
770 make_rels_by_clauseless_joins(root,
772 list_head(joinrels[1]));
777 * Now, consider "bushy plans" in which relations of k initial rels are
778 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
780 * We only consider bushy-plan joins for pairs of rels where there is a
781 * suitable join clause (or join order restriction), in order to avoid
782 * unreasonable growth of planning time.
786 int other_level = level - k;
789 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
790 * need to go as far as the halfway point.
795 foreach(r, joinrels[k])
797 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
798 ListCell *other_rels;
802 * We can ignore relations without join clauses here, unless they
803 * participate in join-order restrictions --- then we might have
804 * to force a bushy join plan.
806 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
807 !has_join_restriction(root, old_rel))
810 if (k == other_level)
811 other_rels = lnext(r); /* only consider remaining rels */
813 other_rels = list_head(joinrels[other_level]);
815 for_each_cell(r2, other_rels)
817 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
819 if (!bms_overlap(old_rel->relids, new_rel->relids))
822 * OK, we can build a rel of the right level from this
823 * pair of rels. Do so if there is at least one relevant
824 * join clause or join order restriction.
826 if (have_relevant_joinclause(root, old_rel, new_rel) ||
827 have_join_order_restriction(root, old_rel, new_rel))
829 (void) make_join_rel(root, old_rel, new_rel);
837 * Last-ditch effort: if we failed to find any usable joins so far, force
838 * a set of cartesian-product joins to be generated. This handles the
839 * special case where all the available rels have join clauses but we
840 * cannot use any of those clauses yet. This can only happen when we are
841 * considering a join sub-problem (a sub-joinlist) and all the rels in the
842 * sub-problem have only join clauses with rels outside the sub-problem.
845 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
846 * WHERE a.w = c.x and b.y = d.z;
848 * If the "a INNER JOIN b" sub-problem does not get flattened into the
849 * upper level, we must be willing to make a cartesian join of a and b;
850 * but the code above will not have done so, because it thought that both
851 * a and b have joinclauses. We consider only left-sided and right-sided
852 * cartesian joins in this case (no bushy).
855 if (joinrels[level] == NIL)
858 * This loop is just like the first one, except we always call
859 * make_rels_by_clauseless_joins().
861 foreach(r, joinrels[level - 1])
863 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
865 make_rels_by_clauseless_joins(root,
867 list_head(joinrels[1]));
871 * When special joins are involved, there may be no legal way
872 * to make an N-way join for some values of N. For example consider
874 * SELECT ... FROM t1 WHERE
875 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
876 * y IN (SELECT ... FROM t4,t5 WHERE ...)
878 * We will flatten this query to a 5-way join problem, but there are
879 * no 4-way joins that join_is_legal() will consider legal. We have
880 * to accept failure at level 4 and go on to discover a workable
881 * bushy plan at level 5.
883 * However, if there are no special joins and no lateral references
884 * then join_is_legal() should never fail, and so the following sanity
888 if (joinrels[level] == NIL &&
889 root->join_info_list == NIL &&
890 !root->hasLateralRTEs)
891 elog(ERROR, "failed to build any %d-way joins", level);
896 * make_rels_by_clause_joins
897 * Build joins between the given relation 'old_rel' and other relations
898 * that participate in join clauses that 'old_rel' also participates in
899 * (or participate in join-order restrictions with it).
900 * The join rels are returned in root->join_rel_level[join_cur_level].
902 * Note: at levels above 2 we will generate the same joined relation in
903 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
904 * (b join c) join a, though the second case will add a different set of Paths
905 * to it. This is the reason for using the join_rel_level mechanism, which
906 * automatically ensures that each new joinrel is only added to the list once.
908 * 'old_rel' is the relation entry for the relation to be joined
909 * 'other_rels': the first cell in a linked list containing the other
910 * rels to be considered for joining
912 * Currently, this is only used with initial rels in other_rels, but it
913 * will work for joining to joinrels too.
916 make_rels_by_clause_joins(PlannerInfo *root,
918 ListCell *other_rels)
922 for_each_cell(l, other_rels)
924 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
926 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
927 (have_relevant_joinclause(root, old_rel, other_rel) ||
928 have_join_order_restriction(root, old_rel, other_rel)))
930 (void) make_join_rel(root, old_rel, other_rel);
936 * make_rels_by_clauseless_joins
937 * Given a relation 'old_rel' and a list of other relations
938 * 'other_rels', create a join relation between 'old_rel' and each
939 * member of 'other_rels' that isn't already included in 'old_rel'.
940 * The join rels are returned in root->join_rel_level[join_cur_level].
942 * 'old_rel' is the relation entry for the relation to be joined
943 * 'other_rels': the first cell of a linked list containing the
944 * other rels to be considered for joining
946 * Currently, this is only used with initial rels in other_rels, but it would
947 * work for joining to joinrels too.
950 make_rels_by_clauseless_joins(PlannerInfo *root,
952 ListCell *other_rels)
956 for_each_cell(l, other_rels)
958 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
960 if (!bms_overlap(other_rel->relids, old_rel->relids))
962 (void) make_join_rel(root, old_rel, other_rel);
969 * Determine whether a proposed join is legal given the query's
970 * join order constraints; and if it is, determine the join type.
972 * Caller must supply not only the two rels, but the union of their relids.
973 * (We could simplify the API by computing joinrelids locally, but this
974 * would be redundant work in the normal path through make_join_rel.)
976 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
977 * else it's set to point to the associated SpecialJoinInfo node. Also,
978 * *reversed_p is set TRUE if the given relations need to be swapped to
979 * match the SpecialJoinInfo node.
982 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
984 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
986 SpecialJoinInfo *match_sjinfo;
989 bool must_be_leftjoin;
993 * Ensure output params are set on failure return. This is just to
994 * suppress uninitialized-variable warnings from overly anal compilers.
1000 * If we have any special joins, the proposed join might be illegal; and
1001 * in any case we have to determine its join type. Scan the join info
1002 * list for matches and conflicts.
1004 match_sjinfo = NULL;
1006 unique_ified = false;
1007 must_be_leftjoin = false;
1009 foreach(l, root->join_info_list)
1011 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1014 * This special join is not relevant unless its RHS overlaps the
1015 * proposed join. (Check this first as a fast path for dismissing
1016 * most irrelevant SJs quickly.)
1018 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1022 * Also, not relevant if proposed join is fully contained within RHS
1023 * (ie, we're still building up the RHS).
1025 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1029 * Also, not relevant if SJ is already done within either input.
1031 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1032 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1034 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1035 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1039 * If it's a semijoin and we already joined the RHS to any other rels
1040 * within either input, then we must have unique-ified the RHS at that
1041 * point (see below). Therefore the semijoin is no longer relevant in
1044 if (sjinfo->jointype == JOIN_SEMI)
1046 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1047 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1049 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1050 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1055 * If one input contains min_lefthand and the other contains
1056 * min_righthand, then we can perform the SJ at this join.
1058 * Reject if we get matches to more than one SJ; that implies we're
1059 * considering something that's not really valid.
1061 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1062 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1065 return false; /* invalid join path */
1066 match_sjinfo = sjinfo;
1069 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1070 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1073 return false; /* invalid join path */
1074 match_sjinfo = sjinfo;
1077 else if (sjinfo->jointype == JOIN_SEMI &&
1078 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1079 create_unique_path(root, rel2, rel2->cheapest_total_path,
1083 * For a semijoin, we can join the RHS to anything else by
1084 * unique-ifying the RHS (if the RHS can be unique-ified).
1085 * We will only get here if we have the full RHS but less
1086 * than min_lefthand on the LHS.
1088 * The reason to consider such a join path is exemplified by
1089 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1090 * If we insist on doing this as a semijoin we will first have
1091 * to form the cartesian product of A*B. But if we unique-ify
1092 * C then the semijoin becomes a plain innerjoin and we can join
1093 * in any order, eg C to A and then to B. When C is much smaller
1094 * than A and B this can be a huge win. So we allow C to be
1095 * joined to just A or just B here, and then make_join_rel has
1096 * to handle the case properly.
1098 * Note that actually we'll allow unique-ified C to be joined to
1099 * some other relation D here, too. That is legal, if usually not
1100 * very sane, and this routine is only concerned with legality not
1101 * with whether the join is good strategy.
1105 return false; /* invalid join path */
1106 match_sjinfo = sjinfo;
1108 unique_ified = true;
1110 else if (sjinfo->jointype == JOIN_SEMI &&
1111 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1112 create_unique_path(root, rel1, rel1->cheapest_total_path,
1115 /* Reversed semijoin case */
1117 return false; /* invalid join path */
1118 match_sjinfo = sjinfo;
1120 unique_ified = true;
1125 * Otherwise, the proposed join overlaps the RHS but isn't a valid
1126 * implementation of this SJ. But don't panic quite yet: the RHS
1127 * violation might have occurred previously, in one or both input
1128 * relations, in which case we must have previously decided that
1129 * it was OK to commute some other SJ with this one. If we need
1130 * to perform this join to finish building up the RHS, rejecting
1131 * it could lead to not finding any plan at all. (This can occur
1132 * because of the heuristics elsewhere in this file that postpone
1133 * clauseless joins: we might not consider doing a clauseless join
1134 * within the RHS until after we've performed other, validly
1135 * commutable SJs with one or both sides of the clauseless join.)
1136 * This consideration boils down to the rule that if both inputs
1137 * overlap the RHS, we can allow the join --- they are either
1138 * fully within the RHS, or represent previously-allowed joins to
1141 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1142 bms_overlap(rel2->relids, sjinfo->min_righthand))
1143 continue; /* assume valid previous violation of RHS */
1146 * The proposed join could still be legal, but only if we're
1147 * allowed to associate it into the RHS of this SJ. That means
1148 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1149 * not FULL) and the proposed join must not overlap the LHS.
1151 if (sjinfo->jointype != JOIN_LEFT ||
1152 bms_overlap(joinrelids, sjinfo->min_lefthand))
1153 return false; /* invalid join path */
1156 * To be valid, the proposed join must be a LEFT join; otherwise
1157 * it can't associate into this SJ's RHS. But we may not yet have
1158 * found the SpecialJoinInfo matching the proposed join, so we
1159 * can't test that yet. Remember the requirement for later.
1161 must_be_leftjoin = true;
1166 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1167 * proposed join can't associate into an SJ's RHS.
1169 * Also, fail if the proposed join's predicate isn't strict; we're
1170 * essentially checking to see if we can apply outer-join identity 3, and
1171 * that's a requirement. (This check may be redundant with checks in
1172 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1174 if (must_be_leftjoin &&
1175 (match_sjinfo == NULL ||
1176 match_sjinfo->jointype != JOIN_LEFT ||
1177 !match_sjinfo->lhs_strict))
1178 return false; /* invalid join path */
1181 * We also have to check for constraints imposed by LATERAL references.
1183 if (root->hasLateralRTEs)
1187 Relids join_lateral_rels;
1190 * The proposed rels could each contain lateral references to the
1191 * other, in which case the join is impossible. If there are lateral
1192 * references in just one direction, then the join has to be done with
1193 * a nestloop with the lateral referencer on the inside. If the join
1194 * matches an SJ that cannot be implemented by such a nestloop, the
1195 * join is impossible.
1197 * Also, if the lateral reference is only indirect, we should reject
1198 * the join; whatever rel(s) the reference chain goes through must be
1201 * Another case that might keep us from building a valid plan is the
1202 * implementation restriction described by have_dangerous_phv().
1204 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1205 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1206 if (lateral_fwd && lateral_rev)
1207 return false; /* have lateral refs in both directions */
1210 /* has to be implemented as nestloop with rel1 on left */
1214 match_sjinfo->jointype == JOIN_FULL))
1215 return false; /* not implementable as nestloop */
1216 /* check there is a direct reference from rel2 to rel1 */
1217 if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1218 return false; /* only indirect refs, so reject */
1219 /* check we won't have a dangerous PHV */
1220 if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1221 return false; /* might be unable to handle required PHV */
1223 else if (lateral_rev)
1225 /* has to be implemented as nestloop with rel2 on left */
1229 match_sjinfo->jointype == JOIN_FULL))
1230 return false; /* not implementable as nestloop */
1231 /* check there is a direct reference from rel1 to rel2 */
1232 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1233 return false; /* only indirect refs, so reject */
1234 /* check we won't have a dangerous PHV */
1235 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1236 return false; /* might be unable to handle required PHV */
1240 * LATERAL references could also cause problems later on if we accept
1241 * this join: if the join's minimum parameterization includes any rels
1242 * that would have to be on the inside of an outer join with this join
1243 * rel, then it's never going to be possible to build the complete
1244 * query using this join. We should reject this join not only because
1245 * it'll save work, but because if we don't, the clauseless-join
1246 * heuristics might think that legality of this join means that some
1247 * other join rel need not be formed, and that could lead to failure
1248 * to find any plan at all. We have to consider not only rels that
1249 * are directly on the inner side of an OJ with the joinrel, but also
1250 * ones that are indirectly so, so search to find all such rels.
1252 join_lateral_rels = min_join_parameterization(root, joinrelids,
1254 if (join_lateral_rels)
1256 Relids join_plus_rhs = bms_copy(joinrelids);
1262 foreach(l, root->join_info_list)
1264 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1266 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1267 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1269 join_plus_rhs = bms_add_members(join_plus_rhs,
1270 sjinfo->min_righthand);
1273 /* full joins constrain both sides symmetrically */
1274 if (sjinfo->jointype == JOIN_FULL &&
1275 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1276 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1278 join_plus_rhs = bms_add_members(join_plus_rhs,
1279 sjinfo->min_lefthand);
1284 if (bms_overlap(join_plus_rhs, join_lateral_rels))
1285 return false; /* will not be able to join to some RHS rel */
1289 /* Otherwise, it's a valid join */
1290 *sjinfo_p = match_sjinfo;
1291 *reversed_p = reversed;
1296 * has_join_restriction
1297 * Detect whether the specified relation has join-order restrictions,
1298 * due to being inside an outer join or an IN (sub-SELECT),
1299 * or participating in any LATERAL references or multi-rel PHVs.
1301 * Essentially, this tests whether have_join_order_restriction() could
1302 * succeed with this rel and some other one. It's OK if we sometimes
1303 * say "true" incorrectly. (Therefore, we don't bother with the relatively
1304 * expensive has_legal_joinclause test.)
1307 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1311 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1314 foreach(l, root->placeholder_list)
1316 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1318 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1319 !bms_equal(rel->relids, phinfo->ph_eval_at))
1323 foreach(l, root->join_info_list)
1325 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1327 /* ignore full joins --- other mechanisms preserve their ordering */
1328 if (sjinfo->jointype == JOIN_FULL)
1331 /* ignore if SJ is already contained in rel */
1332 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1333 bms_is_subset(sjinfo->min_righthand, rel->relids))
1336 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1337 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1338 bms_overlap(sjinfo->min_righthand, rel->relids))
1346 * is_dummy_rel --- has relation been proven empty?
1349 is_dummy_rel(RelOptInfo *rel)
1351 return IS_DUMMY_REL(rel);
1355 * Mark a relation as proven empty.
1357 * During GEQO planning, this can get invoked more than once on the same
1358 * baserel struct, so it's worth checking to see if the rel is already marked
1361 * Also, when called during GEQO join planning, we are in a short-lived
1362 * memory context. We must make sure that the dummy path attached to a
1363 * baserel survives the GEQO cycle, else the baserel is trashed for future
1364 * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1365 * we don't want the dummy path to clutter the main planning context. Upshot
1366 * is that the best solution is to explicitly make the dummy path in the same
1367 * context the given RelOptInfo is in.
1370 mark_dummy_rel(RelOptInfo *rel)
1372 MemoryContext oldcontext;
1374 /* Already marked? */
1375 if (is_dummy_rel(rel))
1378 /* No, so choose correct context to make the dummy path in */
1379 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1381 /* Set dummy size estimate */
1384 /* Evict any previously chosen paths */
1385 rel->pathlist = NIL;
1386 rel->partial_pathlist = NIL;
1388 /* Set up the dummy path */
1389 add_path(rel, (Path *) create_append_path(rel, NIL, NULL, 0));
1391 /* Set or update cheapest_total_path and related fields */
1394 MemoryContextSwitchTo(oldcontext);
1398 * restriction_is_constant_false --- is a restrictlist just false?
1400 * In cases where a qual is provably constant false, eval_const_expressions
1401 * will generally have thrown away anything that's ANDed with it. In outer
1402 * join situations this will leave us computing cartesian products only to
1403 * decide there's no match for an outer row, which is pretty stupid. So,
1404 * we need to detect the case.
1406 * If only_pushed_down is true, then consider only quals that are pushed-down
1407 * from the point of view of the joinrel.
1410 restriction_is_constant_false(List *restrictlist,
1411 RelOptInfo *joinrel,
1412 bool only_pushed_down)
1417 * Despite the above comment, the restriction list we see here might
1418 * possibly have other members besides the FALSE constant, since other
1419 * quals could get "pushed down" to the outer join level. So we check
1420 * each member of the list.
1422 foreach(lc, restrictlist)
1424 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1426 Assert(IsA(rinfo, RestrictInfo));
1427 if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1430 if (rinfo->clause && IsA(rinfo->clause, Const))
1432 Const *con = (Const *) rinfo->clause;
1434 /* constant NULL is as good as constant FALSE for our purposes */
1435 if (con->constisnull)
1437 if (!DatumGetBool(con->constvalue))