1 /*-------------------------------------------------------------------------
4 * Routines copied from PostgreSQL core distribution.
6 * The main purpose of this files is having access to static functions in core.
7 * Another purpose is tweaking functions behavior by replacing part of them by
8 * macro definitions. See at the end of pg_hint_plan.c for details. Anyway,
9 * this file *must* contain required functions without making any change.
11 * This file contains the following functions from corresponding files.
13 * src/backend/optimizer/path/allpaths.c
16 * set_plain_rel_pathlist()
17 * set_append_rel_pathlist()
18 * add_paths_to_append_rel()
19 * generate_mergeappend_paths()
20 * get_cheapest_parameterized_child_path()
21 * accumulate_append_subpath()
24 * standard_join_search(): This funcion is not static. The reason for
25 * including this function is make_rels_by_clause_joins. In order to
26 * avoid generating apparently unwanted join combination, we decided to
27 * change the behavior of make_join_rel, which is called under this
30 * src/backend/optimizer/path/joinrels.c
33 * join_search_one_level(): We have to modify this to call my definition of
34 * make_rels_by_clause_joins.
37 * make_rels_by_clause_joins()
38 * make_rels_by_clauseless_joins()
40 * has_join_restriction()
42 * restriction_is_constant_false()
45 * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
46 * Portions Copyright (c) 1994, Regents of the University of California
48 *-------------------------------------------------------------------------
53 * set_plain_rel_pathlist
54 * Build access paths for a plain relation (no subquery, no inheritance)
57 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
59 Relids required_outer;
62 * We don't support pushing join clauses into the quals of a seqscan, but
63 * it could still have required parameterization due to LATERAL refs in
66 required_outer = rel->lateral_relids;
68 /* Consider sequential scan */
69 add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
71 /* If appropriate, consider parallel sequential scan */
72 if (rel->consider_parallel && required_outer == NULL)
73 create_plain_partial_paths(root, rel);
75 /* Consider index scans */
76 create_index_paths(root, rel);
78 /* Consider TID scans */
79 create_tidscan_paths(root, rel);
84 * set_append_rel_pathlist
85 * Build access paths for an "append relation"
88 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
89 Index rti, RangeTblEntry *rte)
91 int parentRTindex = rti;
92 List *live_childrels = NIL;
96 * Generate access paths for each member relation, and remember the
99 foreach(l, root->append_rel_list)
101 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
103 RangeTblEntry *childRTE;
104 RelOptInfo *childrel;
106 /* append_rel_list contains all append rels; ignore others */
107 if (appinfo->parent_relid != parentRTindex)
110 /* Re-locate the child RTE and RelOptInfo */
111 childRTindex = appinfo->child_relid;
112 childRTE = root->simple_rte_array[childRTindex];
113 childrel = root->simple_rel_array[childRTindex];
116 * If set_append_rel_size() decided the parent appendrel was
117 * parallel-unsafe at some point after visiting this child rel, we
118 * need to propagate the unsafety marking down to the child, so that
119 * we don't generate useless partial paths for it.
121 if (!rel->consider_parallel)
122 childrel->consider_parallel = false;
125 * Compute the child's access paths.
127 set_rel_pathlist(root, childrel, childRTindex, childRTE);
130 * If child is dummy, ignore it.
132 if (IS_DUMMY_REL(childrel))
136 * Child is live, so add it to the live_childrels list for use below.
138 live_childrels = lappend(live_childrels, childrel);
141 /* Add paths to the "append" relation. */
142 add_paths_to_append_rel(root, rel, live_childrels);
146 * add_paths_to_append_rel
147 * Generate paths for given "append" relation given the set of non-dummy
150 * The function collects all parameterizations and orderings supported by the
151 * non-dummy children. For every such parameterization or ordering, it creates
152 * an append path collecting one path from each non-dummy child with given
153 * parameterization or ordering. Similarly it collects partial paths from
154 * non-dummy children to create partial append paths.
157 add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
158 List *live_childrels)
160 List *subpaths = NIL;
161 bool subpaths_valid = true;
162 List *partial_subpaths = NIL;
163 bool partial_subpaths_valid = true;
164 List *all_child_pathkeys = NIL;
165 List *all_child_outers = NIL;
167 List *partitioned_rels = NIL;
169 bool build_partitioned_rels = false;
172 * A plain relation will already have a PartitionedChildRelInfo if it is
173 * partitioned. For a subquery RTE, no PartitionedChildRelInfo exists; we
174 * collect all partitioned_rels associated with any child. (This assumes
175 * that we don't need to look through multiple levels of subquery RTEs; if
176 * we ever do, we could create a PartitionedChildRelInfo with the
177 * accumulated list of partitioned_rels which would then be found when
178 * populated our parent rel with paths. For the present, that appears to
181 rte = planner_rt_fetch(rel->relid, root);
182 switch (rte->rtekind)
185 if (rte->relkind == RELKIND_PARTITIONED_TABLE)
188 get_partitioned_child_rels(root, rel->relid);
189 Assert(list_length(partitioned_rels) >= 1);
193 build_partitioned_rels = true;
196 elog(ERROR, "unexpcted rtekind: %d", (int) rte->rtekind);
200 * For every non-dummy child, remember the cheapest path. Also, identify
201 * all pathkeys (orderings) and parameterizations (required_outer sets)
202 * available for the non-dummy member relations.
204 foreach(l, live_childrels)
206 RelOptInfo *childrel = lfirst(l);
210 * If we need to build partitioned_rels, accumulate the partitioned
211 * rels for this child.
213 if (build_partitioned_rels)
217 cprels = get_partitioned_child_rels(root, childrel->relid);
218 partitioned_rels = list_concat(partitioned_rels,
223 * If child has an unparameterized cheapest-total path, add that to
224 * the unparameterized Append path we are constructing for the parent.
225 * If not, there's no workable unparameterized path.
227 if (childrel->cheapest_total_path->param_info == NULL)
228 subpaths = accumulate_append_subpath(subpaths,
229 childrel->cheapest_total_path);
231 subpaths_valid = false;
233 /* Same idea, but for a partial plan. */
234 if (childrel->partial_pathlist != NIL)
235 partial_subpaths = accumulate_append_subpath(partial_subpaths,
236 linitial(childrel->partial_pathlist));
238 partial_subpaths_valid = false;
241 * Collect lists of all the available path orderings and
242 * parameterizations for all the children. We use these as a
243 * heuristic to indicate which sort orderings and parameterizations we
244 * should build Append and MergeAppend paths for.
246 foreach(lcp, childrel->pathlist)
248 Path *childpath = (Path *) lfirst(lcp);
249 List *childkeys = childpath->pathkeys;
250 Relids childouter = PATH_REQ_OUTER(childpath);
252 /* Unsorted paths don't contribute to pathkey list */
253 if (childkeys != NIL)
258 /* Have we already seen this ordering? */
259 foreach(lpk, all_child_pathkeys)
261 List *existing_pathkeys = (List *) lfirst(lpk);
263 if (compare_pathkeys(existing_pathkeys,
264 childkeys) == PATHKEYS_EQUAL)
272 /* No, so add it to all_child_pathkeys */
273 all_child_pathkeys = lappend(all_child_pathkeys,
278 /* Unparameterized paths don't contribute to param-set list */
284 /* Have we already seen this param set? */
285 foreach(lco, all_child_outers)
287 Relids existing_outers = (Relids) lfirst(lco);
289 if (bms_equal(existing_outers, childouter))
297 /* No, so add it to all_child_outers */
298 all_child_outers = lappend(all_child_outers,
306 * If we found unparameterized paths for all children, build an unordered,
307 * unparameterized Append path for the rel. (Note: this is correct even
308 * if we have zero or one live subpath due to constraint exclusion.)
311 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0,
315 * Consider an append of partial unordered, unparameterized partial paths.
317 if (partial_subpaths_valid)
319 AppendPath *appendpath;
321 int parallel_workers = 0;
324 * Decide on the number of workers to request for this append path.
325 * For now, we just use the maximum value from among the members. It
326 * might be useful to use a higher number if the Append node were
327 * smart enough to spread out the workers, but it currently isn't.
329 foreach(lc, partial_subpaths)
331 Path *path = lfirst(lc);
333 parallel_workers = Max(parallel_workers, path->parallel_workers);
335 Assert(parallel_workers > 0);
337 /* Generate a partial append path. */
338 appendpath = create_append_path(rel, partial_subpaths, NULL,
339 parallel_workers, partitioned_rels);
340 add_partial_path(rel, (Path *) appendpath);
344 * Also build unparameterized MergeAppend paths based on the collected
345 * list of child pathkeys.
348 generate_mergeappend_paths(root, rel, live_childrels,
353 * Build Append paths for each parameterization seen among the child rels.
354 * (This may look pretty expensive, but in most cases of practical
355 * interest, the child rels will expose mostly the same parameterizations,
356 * so that not that many cases actually get considered here.)
358 * The Append node itself cannot enforce quals, so all qual checking must
359 * be done in the child paths. This means that to have a parameterized
360 * Append path, we must have the exact same parameterization for each
361 * child path; otherwise some children might be failing to check the
362 * moved-down quals. To make them match up, we can try to increase the
363 * parameterization of lesser-parameterized paths.
365 foreach(l, all_child_outers)
367 Relids required_outer = (Relids) lfirst(l);
370 /* Select the child paths for an Append with this parameterization */
372 subpaths_valid = true;
373 foreach(lcr, live_childrels)
375 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
378 subpath = get_cheapest_parameterized_child_path(root,
383 /* failed to make a suitable path for this child */
384 subpaths_valid = false;
387 subpaths = accumulate_append_subpath(subpaths, subpath);
391 add_path(rel, (Path *)
392 create_append_path(rel, subpaths, required_outer, 0,
399 * generate_mergeappend_paths
400 * Generate MergeAppend paths for an append relation
402 * Generate a path for each ordering (pathkey list) appearing in
403 * all_child_pathkeys.
405 * We consider both cheapest-startup and cheapest-total cases, ie, for each
406 * interesting ordering, collect all the cheapest startup subpaths and all the
407 * cheapest total paths, and build a MergeAppend path for each case.
409 * We don't currently generate any parameterized MergeAppend paths. While
410 * it would not take much more code here to do so, it's very unclear that it
411 * is worth the planning cycles to investigate such paths: there's little
412 * use for an ordered path on the inside of a nestloop. In fact, it's likely
413 * that the current coding of add_path would reject such paths out of hand,
414 * because add_path gives no credit for sort ordering of parameterized paths,
415 * and a parameterized MergeAppend is going to be more expensive than the
416 * corresponding parameterized Append path. If we ever try harder to support
417 * parameterized mergejoin plans, it might be worth adding support for
418 * parameterized MergeAppends to feed such joins. (See notes in
419 * optimizer/README for why that might not ever happen, though.)
422 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
423 List *live_childrels,
424 List *all_child_pathkeys,
425 List *partitioned_rels)
429 foreach(lcp, all_child_pathkeys)
431 List *pathkeys = (List *) lfirst(lcp);
432 List *startup_subpaths = NIL;
433 List *total_subpaths = NIL;
434 bool startup_neq_total = false;
437 /* Select the child paths for this ordering... */
438 foreach(lcr, live_childrels)
440 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
441 Path *cheapest_startup,
444 /* Locate the right paths, if they are available. */
446 get_cheapest_path_for_pathkeys(childrel->pathlist,
452 get_cheapest_path_for_pathkeys(childrel->pathlist,
459 * If we can't find any paths with the right order just use the
460 * cheapest-total path; we'll have to sort it later.
462 if (cheapest_startup == NULL || cheapest_total == NULL)
464 cheapest_startup = cheapest_total =
465 childrel->cheapest_total_path;
466 /* Assert we do have an unparameterized path for this child */
467 Assert(cheapest_total->param_info == NULL);
471 * Notice whether we actually have different paths for the
472 * "cheapest" and "total" cases; frequently there will be no point
473 * in two create_merge_append_path() calls.
475 if (cheapest_startup != cheapest_total)
476 startup_neq_total = true;
479 accumulate_append_subpath(startup_subpaths, cheapest_startup);
481 accumulate_append_subpath(total_subpaths, cheapest_total);
484 /* ... and build the MergeAppend paths */
485 add_path(rel, (Path *) create_merge_append_path(root,
491 if (startup_neq_total)
492 add_path(rel, (Path *) create_merge_append_path(root,
503 * get_cheapest_parameterized_child_path
504 * Get cheapest path for this relation that has exactly the requested
507 * Returns NULL if unable to create such a path.
510 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
511 Relids required_outer)
517 * Look up the cheapest existing path with no more than the needed
518 * parameterization. If it has exactly the needed parameterization, we're
521 cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
526 Assert(cheapest != NULL);
527 if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
531 * Otherwise, we can "reparameterize" an existing path to match the given
532 * parameterization, which effectively means pushing down additional
533 * joinquals to be checked within the path's scan. However, some existing
534 * paths might check the available joinquals already while others don't;
535 * therefore, it's not clear which existing path will be cheapest after
536 * reparameterization. We have to go through them all and find out.
539 foreach(lc, rel->pathlist)
541 Path *path = (Path *) lfirst(lc);
543 /* Can't use it if it needs more than requested parameterization */
544 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
548 * Reparameterization can only increase the path's cost, so if it's
549 * already more expensive than the current cheapest, forget it.
551 if (cheapest != NULL &&
552 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
555 /* Reparameterize if needed, then recheck cost */
556 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
558 path = reparameterize_path(root, path, required_outer, 1.0);
560 continue; /* failed to reparameterize this one */
561 Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
563 if (cheapest != NULL &&
564 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
568 /* We have a new best path */
572 /* Return the best path, or NULL if we found no suitable candidate */
578 * accumulate_append_subpath
579 * Add a subpath to the list being built for an Append or MergeAppend
581 * It's possible that the child is itself an Append or MergeAppend path, in
582 * which case we can "cut out the middleman" and just add its child paths to
583 * our own list. (We don't try to do this earlier because we need to apply
584 * both levels of transformation to the quals.)
586 * Note that if we omit a child MergeAppend in this way, we are effectively
587 * omitting a sort step, which seems fine: if the parent is to be an Append,
588 * its result would be unsorted anyway, while if the parent is to be a
589 * MergeAppend, there's no point in a separate sort on a child.
592 accumulate_append_subpath(List *subpaths, Path *path)
594 if (IsA(path, AppendPath))
596 AppendPath *apath = (AppendPath *) path;
598 /* list_copy is important here to avoid sharing list substructure */
599 return list_concat(subpaths, list_copy(apath->subpaths));
601 else if (IsA(path, MergeAppendPath))
603 MergeAppendPath *mpath = (MergeAppendPath *) path;
605 /* list_copy is important here to avoid sharing list substructure */
606 return list_concat(subpaths, list_copy(mpath->subpaths));
609 return lappend(subpaths, path);
614 * standard_join_search
615 * Find possible joinpaths for a query by successively finding ways
616 * to join component relations into join relations.
618 * 'levels_needed' is the number of iterations needed, ie, the number of
619 * independent jointree items in the query. This is > 1.
621 * 'initial_rels' is a list of RelOptInfo nodes for each independent
622 * jointree item. These are the components to be joined together.
623 * Note that levels_needed == list_length(initial_rels).
625 * Returns the final level of join relations, i.e., the relation that is
626 * the result of joining all the original relations together.
627 * At least one implementation path must be provided for this relation and
628 * all required sub-relations.
630 * To support loadable plugins that modify planner behavior by changing the
631 * join searching algorithm, we provide a hook variable that lets a plugin
632 * replace or supplement this function. Any such hook must return the same
633 * final join relation as the standard code would, but it might have a
634 * different set of implementation paths attached, and only the sub-joinrels
635 * needed for these paths need have been instantiated.
637 * Note to plugin authors: the functions invoked during standard_join_search()
638 * modify root->join_rel_list and root->join_rel_hash. If you want to do more
639 * than one join-order search, you'll probably need to save and restore the
640 * original states of those data structures. See geqo_eval() for an example.
643 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
649 * This function cannot be invoked recursively within any one planning
650 * problem, so join_rel_level[] can't be in use already.
652 Assert(root->join_rel_level == NULL);
655 * We employ a simple "dynamic programming" algorithm: we first find all
656 * ways to build joins of two jointree items, then all ways to build joins
657 * of three items (from two-item joins and single items), then four-item
658 * joins, and so on until we have considered all ways to join all the
659 * items into one rel.
661 * root->join_rel_level[j] is a list of all the j-item rels. Initially we
662 * set root->join_rel_level[1] to represent all the single-jointree-item
665 root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
667 root->join_rel_level[1] = initial_rels;
669 for (lev = 2; lev <= levels_needed; lev++)
674 * Determine all possible pairs of relations to be joined at this
675 * level, and build paths for making each one from every available
676 * pair of lower-level relations.
678 join_search_one_level(root, lev);
681 * Run generate_gather_paths() for each just-processed joinrel. We
682 * could not do this earlier because both regular and partial paths
683 * can get added to a particular joinrel at multiple times within
684 * join_search_one_level. After that, we're done creating paths for
685 * the joinrel, so run set_cheapest().
687 foreach(lc, root->join_rel_level[lev])
689 rel = (RelOptInfo *) lfirst(lc);
691 /* Create GatherPaths for any useful partial paths for rel */
692 generate_gather_paths(root, rel);
694 /* Find and save the cheapest paths for this rel */
697 #ifdef OPTIMIZER_DEBUG
698 debug_print_rel(root, rel);
704 * We should have a single rel at the final level.
706 if (root->join_rel_level[levels_needed] == NIL)
707 elog(ERROR, "failed to build any %d-way joins", levels_needed);
708 Assert(list_length(root->join_rel_level[levels_needed]) == 1);
710 rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
712 root->join_rel_level = NULL;
718 * create_plain_partial_paths
719 * Build partial access paths for parallel scan of a plain relation
722 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
724 int parallel_workers;
726 parallel_workers = compute_parallel_worker(rel, rel->pages, -1);
728 /* If any limit was set to zero, the user doesn't want a parallel scan. */
729 if (parallel_workers <= 0)
732 /* Add an unordered partial path based on a parallel sequential scan. */
733 add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
738 * join_search_one_level
739 * Consider ways to produce join relations containing exactly 'level'
740 * jointree items. (This is one step of the dynamic-programming method
741 * embodied in standard_join_search.) Join rel nodes for each feasible
742 * combination of lower-level rels are created and returned in a list.
743 * Implementation paths are created for each such joinrel, too.
745 * level: level of rels we want to make this time
746 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
748 * The result is returned in root->join_rel_level[level].
751 join_search_one_level(PlannerInfo *root, int level)
753 List **joinrels = root->join_rel_level;
757 Assert(joinrels[level] == NIL);
759 /* Set join_cur_level so that new joinrels are added to proper list */
760 root->join_cur_level = level;
763 * First, consider left-sided and right-sided plans, in which rels of
764 * exactly level-1 member relations are joined against initial relations.
765 * We prefer to join using join clauses, but if we find a rel of level-1
766 * members that has no join clauses, we will generate Cartesian-product
767 * joins against all initial rels not already contained in it.
769 foreach(r, joinrels[level - 1])
771 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
773 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
774 has_join_restriction(root, old_rel))
777 * There are join clauses or join order restrictions relevant to
778 * this rel, so consider joins between this rel and (only) those
779 * initial rels it is linked to by a clause or restriction.
781 * At level 2 this condition is symmetric, so there is no need to
782 * look at initial rels before this one in the list; we already
783 * considered such joins when we were at the earlier rel. (The
784 * mirror-image joins are handled automatically by make_join_rel.)
785 * In later passes (level > 2), we join rels of the previous level
786 * to each initial rel they don't already include but have a join
787 * clause or restriction with.
789 ListCell *other_rels;
791 if (level == 2) /* consider remaining initial rels */
792 other_rels = lnext(r);
793 else /* consider all initial rels */
794 other_rels = list_head(joinrels[1]);
796 make_rels_by_clause_joins(root,
803 * Oops, we have a relation that is not joined to any other
804 * relation, either directly or by join-order restrictions.
805 * Cartesian product time.
807 * We consider a cartesian product with each not-already-included
808 * initial rel, whether it has other join clauses or not. At
809 * level 2, if there are two or more clauseless initial rels, we
810 * will redundantly consider joining them in both directions; but
811 * such cases aren't common enough to justify adding complexity to
812 * avoid the duplicated effort.
814 make_rels_by_clauseless_joins(root,
816 list_head(joinrels[1]));
821 * Now, consider "bushy plans" in which relations of k initial rels are
822 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
824 * We only consider bushy-plan joins for pairs of rels where there is a
825 * suitable join clause (or join order restriction), in order to avoid
826 * unreasonable growth of planning time.
830 int other_level = level - k;
833 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
834 * need to go as far as the halfway point.
839 foreach(r, joinrels[k])
841 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
842 ListCell *other_rels;
846 * We can ignore relations without join clauses here, unless they
847 * participate in join-order restrictions --- then we might have
848 * to force a bushy join plan.
850 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
851 !has_join_restriction(root, old_rel))
854 if (k == other_level)
855 other_rels = lnext(r); /* only consider remaining rels */
857 other_rels = list_head(joinrels[other_level]);
859 for_each_cell(r2, other_rels)
861 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
863 if (!bms_overlap(old_rel->relids, new_rel->relids))
866 * OK, we can build a rel of the right level from this
867 * pair of rels. Do so if there is at least one relevant
868 * join clause or join order restriction.
870 if (have_relevant_joinclause(root, old_rel, new_rel) ||
871 have_join_order_restriction(root, old_rel, new_rel))
873 (void) make_join_rel(root, old_rel, new_rel);
881 * Last-ditch effort: if we failed to find any usable joins so far, force
882 * a set of cartesian-product joins to be generated. This handles the
883 * special case where all the available rels have join clauses but we
884 * cannot use any of those clauses yet. This can only happen when we are
885 * considering a join sub-problem (a sub-joinlist) and all the rels in the
886 * sub-problem have only join clauses with rels outside the sub-problem.
889 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
890 * WHERE a.w = c.x and b.y = d.z;
892 * If the "a INNER JOIN b" sub-problem does not get flattened into the
893 * upper level, we must be willing to make a cartesian join of a and b;
894 * but the code above will not have done so, because it thought that both
895 * a and b have joinclauses. We consider only left-sided and right-sided
896 * cartesian joins in this case (no bushy).
899 if (joinrels[level] == NIL)
902 * This loop is just like the first one, except we always call
903 * make_rels_by_clauseless_joins().
905 foreach(r, joinrels[level - 1])
907 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
909 make_rels_by_clauseless_joins(root,
911 list_head(joinrels[1]));
915 * When special joins are involved, there may be no legal way
916 * to make an N-way join for some values of N. For example consider
918 * SELECT ... FROM t1 WHERE
919 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
920 * y IN (SELECT ... FROM t4,t5 WHERE ...)
922 * We will flatten this query to a 5-way join problem, but there are
923 * no 4-way joins that join_is_legal() will consider legal. We have
924 * to accept failure at level 4 and go on to discover a workable
925 * bushy plan at level 5.
927 * However, if there are no special joins and no lateral references
928 * then join_is_legal() should never fail, and so the following sanity
932 if (joinrels[level] == NIL &&
933 root->join_info_list == NIL &&
934 !root->hasLateralRTEs)
935 elog(ERROR, "failed to build any %d-way joins", level);
941 * make_rels_by_clause_joins
942 * Build joins between the given relation 'old_rel' and other relations
943 * that participate in join clauses that 'old_rel' also participates in
944 * (or participate in join-order restrictions with it).
945 * The join rels are returned in root->join_rel_level[join_cur_level].
947 * Note: at levels above 2 we will generate the same joined relation in
948 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
949 * (b join c) join a, though the second case will add a different set of Paths
950 * to it. This is the reason for using the join_rel_level mechanism, which
951 * automatically ensures that each new joinrel is only added to the list once.
953 * 'old_rel' is the relation entry for the relation to be joined
954 * 'other_rels': the first cell in a linked list containing the other
955 * rels to be considered for joining
957 * Currently, this is only used with initial rels in other_rels, but it
958 * will work for joining to joinrels too.
961 make_rels_by_clause_joins(PlannerInfo *root,
963 ListCell *other_rels)
967 for_each_cell(l, other_rels)
969 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
971 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
972 (have_relevant_joinclause(root, old_rel, other_rel) ||
973 have_join_order_restriction(root, old_rel, other_rel)))
975 (void) make_join_rel(root, old_rel, other_rel);
982 * make_rels_by_clauseless_joins
983 * Given a relation 'old_rel' and a list of other relations
984 * 'other_rels', create a join relation between 'old_rel' and each
985 * member of 'other_rels' that isn't already included in 'old_rel'.
986 * The join rels are returned in root->join_rel_level[join_cur_level].
988 * 'old_rel' is the relation entry for the relation to be joined
989 * 'other_rels': the first cell of a linked list containing the
990 * other rels to be considered for joining
992 * Currently, this is only used with initial rels in other_rels, but it would
993 * work for joining to joinrels too.
996 make_rels_by_clauseless_joins(PlannerInfo *root,
998 ListCell *other_rels)
1002 for_each_cell(l, other_rels)
1004 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
1006 if (!bms_overlap(other_rel->relids, old_rel->relids))
1008 (void) make_join_rel(root, old_rel, other_rel);
1016 * Determine whether a proposed join is legal given the query's
1017 * join order constraints; and if it is, determine the join type.
1019 * Caller must supply not only the two rels, but the union of their relids.
1020 * (We could simplify the API by computing joinrelids locally, but this
1021 * would be redundant work in the normal path through make_join_rel.)
1023 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
1024 * else it's set to point to the associated SpecialJoinInfo node. Also,
1025 * *reversed_p is set TRUE if the given relations need to be swapped to
1026 * match the SpecialJoinInfo node.
1029 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1031 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
1033 SpecialJoinInfo *match_sjinfo;
1036 bool must_be_leftjoin;
1040 * Ensure output params are set on failure return. This is just to
1041 * suppress uninitialized-variable warnings from overly anal compilers.
1044 *reversed_p = false;
1047 * If we have any special joins, the proposed join might be illegal; and
1048 * in any case we have to determine its join type. Scan the join info
1049 * list for matches and conflicts.
1051 match_sjinfo = NULL;
1053 unique_ified = false;
1054 must_be_leftjoin = false;
1056 foreach(l, root->join_info_list)
1058 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1061 * This special join is not relevant unless its RHS overlaps the
1062 * proposed join. (Check this first as a fast path for dismissing
1063 * most irrelevant SJs quickly.)
1065 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1069 * Also, not relevant if proposed join is fully contained within RHS
1070 * (ie, we're still building up the RHS).
1072 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1076 * Also, not relevant if SJ is already done within either input.
1078 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1079 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1081 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1082 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1086 * If it's a semijoin and we already joined the RHS to any other rels
1087 * within either input, then we must have unique-ified the RHS at that
1088 * point (see below). Therefore the semijoin is no longer relevant in
1091 if (sjinfo->jointype == JOIN_SEMI)
1093 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1094 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1096 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1097 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1102 * If one input contains min_lefthand and the other contains
1103 * min_righthand, then we can perform the SJ at this join.
1105 * Reject if we get matches to more than one SJ; that implies we're
1106 * considering something that's not really valid.
1108 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1109 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1112 return false; /* invalid join path */
1113 match_sjinfo = sjinfo;
1116 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1117 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1120 return false; /* invalid join path */
1121 match_sjinfo = sjinfo;
1124 else if (sjinfo->jointype == JOIN_SEMI &&
1125 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1126 create_unique_path(root, rel2, rel2->cheapest_total_path,
1130 * For a semijoin, we can join the RHS to anything else by
1131 * unique-ifying the RHS (if the RHS can be unique-ified).
1132 * We will only get here if we have the full RHS but less
1133 * than min_lefthand on the LHS.
1135 * The reason to consider such a join path is exemplified by
1136 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1137 * If we insist on doing this as a semijoin we will first have
1138 * to form the cartesian product of A*B. But if we unique-ify
1139 * C then the semijoin becomes a plain innerjoin and we can join
1140 * in any order, eg C to A and then to B. When C is much smaller
1141 * than A and B this can be a huge win. So we allow C to be
1142 * joined to just A or just B here, and then make_join_rel has
1143 * to handle the case properly.
1145 * Note that actually we'll allow unique-ified C to be joined to
1146 * some other relation D here, too. That is legal, if usually not
1147 * very sane, and this routine is only concerned with legality not
1148 * with whether the join is good strategy.
1152 return false; /* invalid join path */
1153 match_sjinfo = sjinfo;
1155 unique_ified = true;
1157 else if (sjinfo->jointype == JOIN_SEMI &&
1158 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1159 create_unique_path(root, rel1, rel1->cheapest_total_path,
1162 /* Reversed semijoin case */
1164 return false; /* invalid join path */
1165 match_sjinfo = sjinfo;
1167 unique_ified = true;
1172 * Otherwise, the proposed join overlaps the RHS but isn't a valid
1173 * implementation of this SJ. But don't panic quite yet: the RHS
1174 * violation might have occurred previously, in one or both input
1175 * relations, in which case we must have previously decided that
1176 * it was OK to commute some other SJ with this one. If we need
1177 * to perform this join to finish building up the RHS, rejecting
1178 * it could lead to not finding any plan at all. (This can occur
1179 * because of the heuristics elsewhere in this file that postpone
1180 * clauseless joins: we might not consider doing a clauseless join
1181 * within the RHS until after we've performed other, validly
1182 * commutable SJs with one or both sides of the clauseless join.)
1183 * This consideration boils down to the rule that if both inputs
1184 * overlap the RHS, we can allow the join --- they are either
1185 * fully within the RHS, or represent previously-allowed joins to
1188 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1189 bms_overlap(rel2->relids, sjinfo->min_righthand))
1190 continue; /* assume valid previous violation of RHS */
1193 * The proposed join could still be legal, but only if we're
1194 * allowed to associate it into the RHS of this SJ. That means
1195 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1196 * not FULL) and the proposed join must not overlap the LHS.
1198 if (sjinfo->jointype != JOIN_LEFT ||
1199 bms_overlap(joinrelids, sjinfo->min_lefthand))
1200 return false; /* invalid join path */
1203 * To be valid, the proposed join must be a LEFT join; otherwise
1204 * it can't associate into this SJ's RHS. But we may not yet have
1205 * found the SpecialJoinInfo matching the proposed join, so we
1206 * can't test that yet. Remember the requirement for later.
1208 must_be_leftjoin = true;
1213 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1214 * proposed join can't associate into an SJ's RHS.
1216 * Also, fail if the proposed join's predicate isn't strict; we're
1217 * essentially checking to see if we can apply outer-join identity 3, and
1218 * that's a requirement. (This check may be redundant with checks in
1219 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1221 if (must_be_leftjoin &&
1222 (match_sjinfo == NULL ||
1223 match_sjinfo->jointype != JOIN_LEFT ||
1224 !match_sjinfo->lhs_strict))
1225 return false; /* invalid join path */
1228 * We also have to check for constraints imposed by LATERAL references.
1230 if (root->hasLateralRTEs)
1234 Relids join_lateral_rels;
1237 * The proposed rels could each contain lateral references to the
1238 * other, in which case the join is impossible. If there are lateral
1239 * references in just one direction, then the join has to be done with
1240 * a nestloop with the lateral referencer on the inside. If the join
1241 * matches an SJ that cannot be implemented by such a nestloop, the
1242 * join is impossible.
1244 * Also, if the lateral reference is only indirect, we should reject
1245 * the join; whatever rel(s) the reference chain goes through must be
1248 * Another case that might keep us from building a valid plan is the
1249 * implementation restriction described by have_dangerous_phv().
1251 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1252 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1253 if (lateral_fwd && lateral_rev)
1254 return false; /* have lateral refs in both directions */
1257 /* has to be implemented as nestloop with rel1 on left */
1261 match_sjinfo->jointype == JOIN_FULL))
1262 return false; /* not implementable as nestloop */
1263 /* check there is a direct reference from rel2 to rel1 */
1264 if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1265 return false; /* only indirect refs, so reject */
1266 /* check we won't have a dangerous PHV */
1267 if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1268 return false; /* might be unable to handle required PHV */
1270 else if (lateral_rev)
1272 /* has to be implemented as nestloop with rel2 on left */
1276 match_sjinfo->jointype == JOIN_FULL))
1277 return false; /* not implementable as nestloop */
1278 /* check there is a direct reference from rel1 to rel2 */
1279 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1280 return false; /* only indirect refs, so reject */
1281 /* check we won't have a dangerous PHV */
1282 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1283 return false; /* might be unable to handle required PHV */
1287 * LATERAL references could also cause problems later on if we accept
1288 * this join: if the join's minimum parameterization includes any rels
1289 * that would have to be on the inside of an outer join with this join
1290 * rel, then it's never going to be possible to build the complete
1291 * query using this join. We should reject this join not only because
1292 * it'll save work, but because if we don't, the clauseless-join
1293 * heuristics might think that legality of this join means that some
1294 * other join rel need not be formed, and that could lead to failure
1295 * to find any plan at all. We have to consider not only rels that
1296 * are directly on the inner side of an OJ with the joinrel, but also
1297 * ones that are indirectly so, so search to find all such rels.
1299 join_lateral_rels = min_join_parameterization(root, joinrelids,
1301 if (join_lateral_rels)
1303 Relids join_plus_rhs = bms_copy(joinrelids);
1309 foreach(l, root->join_info_list)
1311 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1313 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1314 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1316 join_plus_rhs = bms_add_members(join_plus_rhs,
1317 sjinfo->min_righthand);
1320 /* full joins constrain both sides symmetrically */
1321 if (sjinfo->jointype == JOIN_FULL &&
1322 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1323 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1325 join_plus_rhs = bms_add_members(join_plus_rhs,
1326 sjinfo->min_lefthand);
1331 if (bms_overlap(join_plus_rhs, join_lateral_rels))
1332 return false; /* will not be able to join to some RHS rel */
1336 /* Otherwise, it's a valid join */
1337 *sjinfo_p = match_sjinfo;
1338 *reversed_p = reversed;
1344 * has_join_restriction
1345 * Detect whether the specified relation has join-order restrictions,
1346 * due to being inside an outer join or an IN (sub-SELECT),
1347 * or participating in any LATERAL references or multi-rel PHVs.
1349 * Essentially, this tests whether have_join_order_restriction() could
1350 * succeed with this rel and some other one. It's OK if we sometimes
1351 * say "true" incorrectly. (Therefore, we don't bother with the relatively
1352 * expensive has_legal_joinclause test.)
1355 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1359 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1362 foreach(l, root->placeholder_list)
1364 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1366 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1367 !bms_equal(rel->relids, phinfo->ph_eval_at))
1371 foreach(l, root->join_info_list)
1373 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1375 /* ignore full joins --- other mechanisms preserve their ordering */
1376 if (sjinfo->jointype == JOIN_FULL)
1379 /* ignore if SJ is already contained in rel */
1380 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1381 bms_is_subset(sjinfo->min_righthand, rel->relids))
1384 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1385 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1386 bms_overlap(sjinfo->min_righthand, rel->relids))
1395 * is_dummy_rel --- has relation been proven empty?
1398 is_dummy_rel(RelOptInfo *rel)
1400 return IS_DUMMY_REL(rel);
1405 * restriction_is_constant_false --- is a restrictlist just FALSE?
1407 * In cases where a qual is provably constant FALSE, eval_const_expressions
1408 * will generally have thrown away anything that's ANDed with it. In outer
1409 * join situations this will leave us computing cartesian products only to
1410 * decide there's no match for an outer row, which is pretty stupid. So,
1411 * we need to detect the case.
1413 * If only_pushed_down is TRUE, then consider only pushed-down quals.
1416 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1421 * Despite the above comment, the restriction list we see here might
1422 * possibly have other members besides the FALSE constant, since other
1423 * quals could get "pushed down" to the outer join level. So we check
1424 * each member of the list.
1426 foreach(lc, restrictlist)
1428 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1430 if (only_pushed_down && !rinfo->is_pushed_down)
1433 if (rinfo->clause && IsA(rinfo->clause, Const))
1435 Const *con = (Const *) rinfo->clause;
1437 /* constant NULL is as good as constant FALSE for our purposes */
1438 if (con->constisnull)
1440 if (!DatumGetBool(con->constvalue))