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 * set_append_rel_pathlist()
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()
43 * restriction_is_constant_false()
46 * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
47 * Portions Copyright (c) 1994, Regents of the University of California
49 *-------------------------------------------------------------------------
54 * set_plain_rel_pathlist
55 * Build access paths for a plain relation (no subquery, no inheritance)
58 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
60 Relids required_outer;
63 * We don't support pushing join clauses into the quals of a seqscan, but
64 * it could still have required parameterization due to LATERAL refs in
67 required_outer = rel->lateral_relids;
69 /* Consider sequential scan */
70 add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
72 /* If appropriate, consider parallel sequential scan */
73 if (rel->consider_parallel && required_outer == NULL)
74 create_plain_partial_paths(root, rel);
76 /* Consider index scans */
77 create_index_paths(root, rel);
79 /* Consider TID scans */
80 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;
94 bool subpaths_valid = true;
95 List *partial_subpaths = NIL;
96 bool partial_subpaths_valid = true;
97 List *all_child_pathkeys = NIL;
98 List *all_child_outers = NIL;
102 * Generate access paths for each member relation, and remember the
103 * cheapest path for each one. Also, identify all pathkeys (orderings)
104 * and parameterizations (required_outer sets) available for the member
107 foreach(l, root->append_rel_list)
109 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
111 RangeTblEntry *childRTE;
112 RelOptInfo *childrel;
115 /* append_rel_list contains all append rels; ignore others */
116 if (appinfo->parent_relid != parentRTindex)
119 /* Re-locate the child RTE and RelOptInfo */
120 childRTindex = appinfo->child_relid;
121 childRTE = root->simple_rte_array[childRTindex];
122 childrel = root->simple_rel_array[childRTindex];
125 * If set_append_rel_size() decided the parent appendrel was
126 * parallel-unsafe at some point after visiting this child rel, we
127 * need to propagate the unsafety marking down to the child, so that
128 * we don't generate useless partial paths for it.
130 if (!rel->consider_parallel)
131 childrel->consider_parallel = false;
134 * Compute the child's access paths.
136 set_rel_pathlist(root, childrel, childRTindex, childRTE);
139 * If child is dummy, ignore it.
141 if (IS_DUMMY_REL(childrel))
145 * Child is live, so add it to the live_childrels list for use below.
147 live_childrels = lappend(live_childrels, childrel);
150 * If child has an unparameterized cheapest-total path, add that to
151 * the unparameterized Append path we are constructing for the parent.
152 * If not, there's no workable unparameterized path.
154 if (childrel->cheapest_total_path->param_info == NULL)
155 subpaths = accumulate_append_subpath(subpaths,
156 childrel->cheapest_total_path);
158 subpaths_valid = false;
160 /* Same idea, but for a partial plan. */
161 if (childrel->partial_pathlist != NIL)
162 partial_subpaths = accumulate_append_subpath(partial_subpaths,
163 linitial(childrel->partial_pathlist));
165 partial_subpaths_valid = false;
168 * Collect lists of all the available path orderings and
169 * parameterizations for all the children. We use these as a
170 * heuristic to indicate which sort orderings and parameterizations we
171 * should build Append and MergeAppend paths for.
173 foreach(lcp, childrel->pathlist)
175 Path *childpath = (Path *) lfirst(lcp);
176 List *childkeys = childpath->pathkeys;
177 Relids childouter = PATH_REQ_OUTER(childpath);
179 /* Unsorted paths don't contribute to pathkey list */
180 if (childkeys != NIL)
185 /* Have we already seen this ordering? */
186 foreach(lpk, all_child_pathkeys)
188 List *existing_pathkeys = (List *) lfirst(lpk);
190 if (compare_pathkeys(existing_pathkeys,
191 childkeys) == PATHKEYS_EQUAL)
199 /* No, so add it to all_child_pathkeys */
200 all_child_pathkeys = lappend(all_child_pathkeys,
205 /* Unparameterized paths don't contribute to param-set list */
211 /* Have we already seen this param set? */
212 foreach(lco, all_child_outers)
214 Relids existing_outers = (Relids) lfirst(lco);
216 if (bms_equal(existing_outers, childouter))
224 /* No, so add it to all_child_outers */
225 all_child_outers = lappend(all_child_outers,
233 * If we found unparameterized paths for all children, build an unordered,
234 * unparameterized Append path for the rel. (Note: this is correct even
235 * if we have zero or one live subpath due to constraint exclusion.)
238 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0));
241 * Consider an append of partial unordered, unparameterized partial paths.
243 if (partial_subpaths_valid)
245 AppendPath *appendpath;
247 int parallel_workers = 0;
250 * Decide on the number of workers to request for this append path.
251 * For now, we just use the maximum value from among the members. It
252 * might be useful to use a higher number if the Append node were
253 * smart enough to spread out the workers, but it currently isn't.
255 foreach(lc, partial_subpaths)
257 Path *path = lfirst(lc);
259 parallel_workers = Max(parallel_workers, path->parallel_workers);
261 Assert(parallel_workers > 0);
263 /* Generate a partial append path. */
264 appendpath = create_append_path(rel, partial_subpaths, NULL,
266 add_partial_path(rel, (Path *) appendpath);
270 * Also build unparameterized MergeAppend paths based on the collected
271 * list of child pathkeys.
274 generate_mergeappend_paths(root, rel, live_childrels,
278 * Build Append paths for each parameterization seen among the child rels.
279 * (This may look pretty expensive, but in most cases of practical
280 * interest, the child rels will expose mostly the same parameterizations,
281 * so that not that many cases actually get considered here.)
283 * The Append node itself cannot enforce quals, so all qual checking must
284 * be done in the child paths. This means that to have a parameterized
285 * Append path, we must have the exact same parameterization for each
286 * child path; otherwise some children might be failing to check the
287 * moved-down quals. To make them match up, we can try to increase the
288 * parameterization of lesser-parameterized paths.
290 foreach(l, all_child_outers)
292 Relids required_outer = (Relids) lfirst(l);
295 /* Select the child paths for an Append with this parameterization */
297 subpaths_valid = true;
298 foreach(lcr, live_childrels)
300 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
303 subpath = get_cheapest_parameterized_child_path(root,
308 /* failed to make a suitable path for this child */
309 subpaths_valid = false;
312 subpaths = accumulate_append_subpath(subpaths, subpath);
316 add_path(rel, (Path *)
317 create_append_path(rel, subpaths, required_outer, 0));
322 * generate_mergeappend_paths
323 * Generate MergeAppend paths for an append relation
325 * Generate a path for each ordering (pathkey list) appearing in
326 * all_child_pathkeys.
328 * We consider both cheapest-startup and cheapest-total cases, ie, for each
329 * interesting ordering, collect all the cheapest startup subpaths and all the
330 * cheapest total paths, and build a MergeAppend path for each case.
332 * We don't currently generate any parameterized MergeAppend paths. While
333 * it would not take much more code here to do so, it's very unclear that it
334 * is worth the planning cycles to investigate such paths: there's little
335 * use for an ordered path on the inside of a nestloop. In fact, it's likely
336 * that the current coding of add_path would reject such paths out of hand,
337 * because add_path gives no credit for sort ordering of parameterized paths,
338 * and a parameterized MergeAppend is going to be more expensive than the
339 * corresponding parameterized Append path. If we ever try harder to support
340 * parameterized mergejoin plans, it might be worth adding support for
341 * parameterized MergeAppends to feed such joins. (See notes in
342 * optimizer/README for why that might not ever happen, though.)
345 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
346 List *live_childrels,
347 List *all_child_pathkeys)
351 foreach(lcp, all_child_pathkeys)
353 List *pathkeys = (List *) lfirst(lcp);
354 List *startup_subpaths = NIL;
355 List *total_subpaths = NIL;
356 bool startup_neq_total = false;
359 /* Select the child paths for this ordering... */
360 foreach(lcr, live_childrels)
362 RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
363 Path *cheapest_startup,
366 /* Locate the right paths, if they are available. */
368 get_cheapest_path_for_pathkeys(childrel->pathlist,
373 get_cheapest_path_for_pathkeys(childrel->pathlist,
379 * If we can't find any paths with the right order just use the
380 * cheapest-total path; we'll have to sort it later.
382 if (cheapest_startup == NULL || cheapest_total == NULL)
384 cheapest_startup = cheapest_total =
385 childrel->cheapest_total_path;
386 /* Assert we do have an unparameterized path for this child */
387 Assert(cheapest_total->param_info == NULL);
391 * Notice whether we actually have different paths for the
392 * "cheapest" and "total" cases; frequently there will be no point
393 * in two create_merge_append_path() calls.
395 if (cheapest_startup != cheapest_total)
396 startup_neq_total = true;
399 accumulate_append_subpath(startup_subpaths, cheapest_startup);
401 accumulate_append_subpath(total_subpaths, cheapest_total);
404 /* ... and build the MergeAppend paths */
405 add_path(rel, (Path *) create_merge_append_path(root,
410 if (startup_neq_total)
411 add_path(rel, (Path *) create_merge_append_path(root,
420 * get_cheapest_parameterized_child_path
421 * Get cheapest path for this relation that has exactly the requested
424 * Returns NULL if unable to create such a path.
427 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
428 Relids required_outer)
434 * Look up the cheapest existing path with no more than the needed
435 * parameterization. If it has exactly the needed parameterization, we're
438 cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
442 Assert(cheapest != NULL);
443 if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
447 * Otherwise, we can "reparameterize" an existing path to match the given
448 * parameterization, which effectively means pushing down additional
449 * joinquals to be checked within the path's scan. However, some existing
450 * paths might check the available joinquals already while others don't;
451 * therefore, it's not clear which existing path will be cheapest after
452 * reparameterization. We have to go through them all and find out.
455 foreach(lc, rel->pathlist)
457 Path *path = (Path *) lfirst(lc);
459 /* Can't use it if it needs more than requested parameterization */
460 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
464 * Reparameterization can only increase the path's cost, so if it's
465 * already more expensive than the current cheapest, forget it.
467 if (cheapest != NULL &&
468 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
471 /* Reparameterize if needed, then recheck cost */
472 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
474 path = reparameterize_path(root, path, required_outer, 1.0);
476 continue; /* failed to reparameterize this one */
477 Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
479 if (cheapest != NULL &&
480 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
484 /* We have a new best path */
488 /* Return the best path, or NULL if we found no suitable candidate */
493 * accumulate_append_subpath
494 * Add a subpath to the list being built for an Append or MergeAppend
496 * It's possible that the child is itself an Append or MergeAppend path, in
497 * which case we can "cut out the middleman" and just add its child paths to
498 * our own list. (We don't try to do this earlier because we need to apply
499 * both levels of transformation to the quals.)
501 * Note that if we omit a child MergeAppend in this way, we are effectively
502 * omitting a sort step, which seems fine: if the parent is to be an Append,
503 * its result would be unsorted anyway, while if the parent is to be a
504 * MergeAppend, there's no point in a separate sort on a child.
507 accumulate_append_subpath(List *subpaths, Path *path)
509 if (IsA(path, AppendPath))
511 AppendPath *apath = (AppendPath *) path;
513 /* list_copy is important here to avoid sharing list substructure */
514 return list_concat(subpaths, list_copy(apath->subpaths));
516 else if (IsA(path, MergeAppendPath))
518 MergeAppendPath *mpath = (MergeAppendPath *) path;
520 /* list_copy is important here to avoid sharing list substructure */
521 return list_concat(subpaths, list_copy(mpath->subpaths));
524 return lappend(subpaths, path);
528 * standard_join_search
529 * Find possible joinpaths for a query by successively finding ways
530 * to join component relations into join relations.
532 * 'levels_needed' is the number of iterations needed, ie, the number of
533 * independent jointree items in the query. This is > 1.
535 * 'initial_rels' is a list of RelOptInfo nodes for each independent
536 * jointree item. These are the components to be joined together.
537 * Note that levels_needed == list_length(initial_rels).
539 * Returns the final level of join relations, i.e., the relation that is
540 * the result of joining all the original relations together.
541 * At least one implementation path must be provided for this relation and
542 * all required sub-relations.
544 * To support loadable plugins that modify planner behavior by changing the
545 * join searching algorithm, we provide a hook variable that lets a plugin
546 * replace or supplement this function. Any such hook must return the same
547 * final join relation as the standard code would, but it might have a
548 * different set of implementation paths attached, and only the sub-joinrels
549 * needed for these paths need have been instantiated.
551 * Note to plugin authors: the functions invoked during standard_join_search()
552 * modify root->join_rel_list and root->join_rel_hash. If you want to do more
553 * than one join-order search, you'll probably need to save and restore the
554 * original states of those data structures. See geqo_eval() for an example.
557 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
563 * This function cannot be invoked recursively within any one planning
564 * problem, so join_rel_level[] can't be in use already.
566 Assert(root->join_rel_level == NULL);
569 * We employ a simple "dynamic programming" algorithm: we first find all
570 * ways to build joins of two jointree items, then all ways to build joins
571 * of three items (from two-item joins and single items), then four-item
572 * joins, and so on until we have considered all ways to join all the
573 * items into one rel.
575 * root->join_rel_level[j] is a list of all the j-item rels. Initially we
576 * set root->join_rel_level[1] to represent all the single-jointree-item
579 root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
581 root->join_rel_level[1] = initial_rels;
583 for (lev = 2; lev <= levels_needed; lev++)
588 * Determine all possible pairs of relations to be joined at this
589 * level, and build paths for making each one from every available
590 * pair of lower-level relations.
592 join_search_one_level(root, lev);
595 * Run generate_gather_paths() for each just-processed joinrel. We
596 * could not do this earlier because both regular and partial paths
597 * can get added to a particular joinrel at multiple times within
598 * join_search_one_level. After that, we're done creating paths for
599 * the joinrel, so run set_cheapest().
601 foreach(lc, root->join_rel_level[lev])
603 rel = (RelOptInfo *) lfirst(lc);
605 /* Create GatherPaths for any useful partial paths for rel */
606 generate_gather_paths(root, rel);
608 /* Find and save the cheapest paths for this rel */
611 #ifdef OPTIMIZER_DEBUG
612 debug_print_rel(root, rel);
618 * We should have a single rel at the final level.
620 if (root->join_rel_level[levels_needed] == NIL)
621 elog(ERROR, "failed to build any %d-way joins", levels_needed);
622 Assert(list_length(root->join_rel_level[levels_needed]) == 1);
624 rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
626 root->join_rel_level = NULL;
632 * create_plain_partial_paths
633 * Build partial access paths for parallel scan of a plain relation
636 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
638 int parallel_workers;
640 parallel_workers = compute_parallel_worker(rel, rel->pages);
642 /* If any limit was set to zero, the user doesn't want a parallel scan. */
643 if (parallel_workers <= 0)
646 /* Add an unordered partial path based on a parallel sequential scan. */
647 add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
651 * Compute the number of parallel workers that should be used to scan a
652 * relation. "pages" is the number of pages from the relation that we
656 compute_parallel_worker(RelOptInfo *rel, BlockNumber pages)
658 int parallel_workers;
661 * If the user has set the parallel_workers reloption, use that; otherwise
662 * select a default number of workers.
664 if (rel->rel_parallel_workers != -1)
665 parallel_workers = rel->rel_parallel_workers;
668 int parallel_threshold;
671 * If this relation is too small to be worth a parallel scan, just
672 * return without doing anything ... unless it's an inheritance child.
673 * In that case, we want to generate a parallel path here anyway. It
674 * might not be worthwhile just for this relation, but when combined
675 * with all of its inheritance siblings it may well pay off.
677 if (pages < (BlockNumber) min_parallel_relation_size &&
678 rel->reloptkind == RELOPT_BASEREL)
682 * Select the number of workers based on the log of the size of the
683 * relation. This probably needs to be a good deal more
684 * sophisticated, but we need something here for now. Note that the
685 * upper limit of the min_parallel_relation_size GUC is chosen to
686 * prevent overflow here.
688 parallel_workers = 1;
689 parallel_threshold = Max(min_parallel_relation_size, 1);
690 while (pages >= (BlockNumber) (parallel_threshold * 3))
693 parallel_threshold *= 3;
694 if (parallel_threshold > INT_MAX / 3)
695 break; /* avoid overflow */
700 * In no case use more than max_parallel_workers_per_gather workers.
702 parallel_workers = Min(parallel_workers, max_parallel_workers_per_gather);
704 return parallel_workers;
709 * join_search_one_level
710 * Consider ways to produce join relations containing exactly 'level'
711 * jointree items. (This is one step of the dynamic-programming method
712 * embodied in standard_join_search.) Join rel nodes for each feasible
713 * combination of lower-level rels are created and returned in a list.
714 * Implementation paths are created for each such joinrel, too.
716 * level: level of rels we want to make this time
717 * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
719 * The result is returned in root->join_rel_level[level].
722 join_search_one_level(PlannerInfo *root, int level)
724 List **joinrels = root->join_rel_level;
728 Assert(joinrels[level] == NIL);
730 /* Set join_cur_level so that new joinrels are added to proper list */
731 root->join_cur_level = level;
734 * First, consider left-sided and right-sided plans, in which rels of
735 * exactly level-1 member relations are joined against initial relations.
736 * We prefer to join using join clauses, but if we find a rel of level-1
737 * members that has no join clauses, we will generate Cartesian-product
738 * joins against all initial rels not already contained in it.
740 foreach(r, joinrels[level - 1])
742 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
744 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
745 has_join_restriction(root, old_rel))
748 * There are join clauses or join order restrictions relevant to
749 * this rel, so consider joins between this rel and (only) those
750 * initial rels it is linked to by a clause or restriction.
752 * At level 2 this condition is symmetric, so there is no need to
753 * look at initial rels before this one in the list; we already
754 * considered such joins when we were at the earlier rel. (The
755 * mirror-image joins are handled automatically by make_join_rel.)
756 * In later passes (level > 2), we join rels of the previous level
757 * to each initial rel they don't already include but have a join
758 * clause or restriction with.
760 ListCell *other_rels;
762 if (level == 2) /* consider remaining initial rels */
763 other_rels = lnext(r);
764 else /* consider all initial rels */
765 other_rels = list_head(joinrels[1]);
767 make_rels_by_clause_joins(root,
774 * Oops, we have a relation that is not joined to any other
775 * relation, either directly or by join-order restrictions.
776 * Cartesian product time.
778 * We consider a cartesian product with each not-already-included
779 * initial rel, whether it has other join clauses or not. At
780 * level 2, if there are two or more clauseless initial rels, we
781 * will redundantly consider joining them in both directions; but
782 * such cases aren't common enough to justify adding complexity to
783 * avoid the duplicated effort.
785 make_rels_by_clauseless_joins(root,
787 list_head(joinrels[1]));
792 * Now, consider "bushy plans" in which relations of k initial rels are
793 * joined to relations of level-k initial rels, for 2 <= k <= level-2.
795 * We only consider bushy-plan joins for pairs of rels where there is a
796 * suitable join clause (or join order restriction), in order to avoid
797 * unreasonable growth of planning time.
801 int other_level = level - k;
804 * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
805 * need to go as far as the halfway point.
810 foreach(r, joinrels[k])
812 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
813 ListCell *other_rels;
817 * We can ignore relations without join clauses here, unless they
818 * participate in join-order restrictions --- then we might have
819 * to force a bushy join plan.
821 if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
822 !has_join_restriction(root, old_rel))
825 if (k == other_level)
826 other_rels = lnext(r); /* only consider remaining rels */
828 other_rels = list_head(joinrels[other_level]);
830 for_each_cell(r2, other_rels)
832 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
834 if (!bms_overlap(old_rel->relids, new_rel->relids))
837 * OK, we can build a rel of the right level from this
838 * pair of rels. Do so if there is at least one relevant
839 * join clause or join order restriction.
841 if (have_relevant_joinclause(root, old_rel, new_rel) ||
842 have_join_order_restriction(root, old_rel, new_rel))
844 (void) make_join_rel(root, old_rel, new_rel);
852 * Last-ditch effort: if we failed to find any usable joins so far, force
853 * a set of cartesian-product joins to be generated. This handles the
854 * special case where all the available rels have join clauses but we
855 * cannot use any of those clauses yet. This can only happen when we are
856 * considering a join sub-problem (a sub-joinlist) and all the rels in the
857 * sub-problem have only join clauses with rels outside the sub-problem.
860 * SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
861 * WHERE a.w = c.x and b.y = d.z;
863 * If the "a INNER JOIN b" sub-problem does not get flattened into the
864 * upper level, we must be willing to make a cartesian join of a and b;
865 * but the code above will not have done so, because it thought that both
866 * a and b have joinclauses. We consider only left-sided and right-sided
867 * cartesian joins in this case (no bushy).
870 if (joinrels[level] == NIL)
873 * This loop is just like the first one, except we always call
874 * make_rels_by_clauseless_joins().
876 foreach(r, joinrels[level - 1])
878 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
880 make_rels_by_clauseless_joins(root,
882 list_head(joinrels[1]));
886 * When special joins are involved, there may be no legal way
887 * to make an N-way join for some values of N. For example consider
889 * SELECT ... FROM t1 WHERE
890 * x IN (SELECT ... FROM t2,t3 WHERE ...) AND
891 * y IN (SELECT ... FROM t4,t5 WHERE ...)
893 * We will flatten this query to a 5-way join problem, but there are
894 * no 4-way joins that join_is_legal() will consider legal. We have
895 * to accept failure at level 4 and go on to discover a workable
896 * bushy plan at level 5.
898 * However, if there are no special joins and no lateral references
899 * then join_is_legal() should never fail, and so the following sanity
903 if (joinrels[level] == NIL &&
904 root->join_info_list == NIL &&
905 !root->hasLateralRTEs)
906 elog(ERROR, "failed to build any %d-way joins", level);
911 * make_rels_by_clause_joins
912 * Build joins between the given relation 'old_rel' and other relations
913 * that participate in join clauses that 'old_rel' also participates in
914 * (or participate in join-order restrictions with it).
915 * The join rels are returned in root->join_rel_level[join_cur_level].
917 * Note: at levels above 2 we will generate the same joined relation in
918 * multiple ways --- for example (a join b) join c is the same RelOptInfo as
919 * (b join c) join a, though the second case will add a different set of Paths
920 * to it. This is the reason for using the join_rel_level mechanism, which
921 * automatically ensures that each new joinrel is only added to the list once.
923 * 'old_rel' is the relation entry for the relation to be joined
924 * 'other_rels': the first cell in a linked list containing the other
925 * rels to be considered for joining
927 * Currently, this is only used with initial rels in other_rels, but it
928 * will work for joining to joinrels too.
931 make_rels_by_clause_joins(PlannerInfo *root,
933 ListCell *other_rels)
937 for_each_cell(l, other_rels)
939 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
941 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
942 (have_relevant_joinclause(root, old_rel, other_rel) ||
943 have_join_order_restriction(root, old_rel, other_rel)))
945 (void) make_join_rel(root, old_rel, other_rel);
951 * make_rels_by_clauseless_joins
952 * Given a relation 'old_rel' and a list of other relations
953 * 'other_rels', create a join relation between 'old_rel' and each
954 * member of 'other_rels' that isn't already included in 'old_rel'.
955 * The join rels are returned in root->join_rel_level[join_cur_level].
957 * 'old_rel' is the relation entry for the relation to be joined
958 * 'other_rels': the first cell of a linked list containing the
959 * other rels to be considered for joining
961 * Currently, this is only used with initial rels in other_rels, but it would
962 * work for joining to joinrels too.
965 make_rels_by_clauseless_joins(PlannerInfo *root,
967 ListCell *other_rels)
971 for_each_cell(l, other_rels)
973 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
975 if (!bms_overlap(other_rel->relids, old_rel->relids))
977 (void) make_join_rel(root, old_rel, other_rel);
984 * Determine whether a proposed join is legal given the query's
985 * join order constraints; and if it is, determine the join type.
987 * Caller must supply not only the two rels, but the union of their relids.
988 * (We could simplify the API by computing joinrelids locally, but this
989 * would be redundant work in the normal path through make_join_rel.)
991 * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
992 * else it's set to point to the associated SpecialJoinInfo node. Also,
993 * *reversed_p is set TRUE if the given relations need to be swapped to
994 * match the SpecialJoinInfo node.
997 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
999 SpecialJoinInfo **sjinfo_p, bool *reversed_p)
1001 SpecialJoinInfo *match_sjinfo;
1004 bool must_be_leftjoin;
1008 * Ensure output params are set on failure return. This is just to
1009 * suppress uninitialized-variable warnings from overly anal compilers.
1012 *reversed_p = false;
1015 * If we have any special joins, the proposed join might be illegal; and
1016 * in any case we have to determine its join type. Scan the join info
1017 * list for matches and conflicts.
1019 match_sjinfo = NULL;
1021 unique_ified = false;
1022 must_be_leftjoin = false;
1024 foreach(l, root->join_info_list)
1026 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1029 * This special join is not relevant unless its RHS overlaps the
1030 * proposed join. (Check this first as a fast path for dismissing
1031 * most irrelevant SJs quickly.)
1033 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1037 * Also, not relevant if proposed join is fully contained within RHS
1038 * (ie, we're still building up the RHS).
1040 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1044 * Also, not relevant if SJ is already done within either input.
1046 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1047 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1049 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1050 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1054 * If it's a semijoin and we already joined the RHS to any other rels
1055 * within either input, then we must have unique-ified the RHS at that
1056 * point (see below). Therefore the semijoin is no longer relevant in
1059 if (sjinfo->jointype == JOIN_SEMI)
1061 if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1062 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1064 if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1065 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1070 * If one input contains min_lefthand and the other contains
1071 * min_righthand, then we can perform the SJ at this join.
1073 * Reject if we get matches to more than one SJ; that implies we're
1074 * considering something that's not really valid.
1076 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1077 bms_is_subset(sjinfo->min_righthand, rel2->relids))
1080 return false; /* invalid join path */
1081 match_sjinfo = sjinfo;
1084 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1085 bms_is_subset(sjinfo->min_righthand, rel1->relids))
1088 return false; /* invalid join path */
1089 match_sjinfo = sjinfo;
1092 else if (sjinfo->jointype == JOIN_SEMI &&
1093 bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1094 create_unique_path(root, rel2, rel2->cheapest_total_path,
1098 * For a semijoin, we can join the RHS to anything else by
1099 * unique-ifying the RHS (if the RHS can be unique-ified).
1100 * We will only get here if we have the full RHS but less
1101 * than min_lefthand on the LHS.
1103 * The reason to consider such a join path is exemplified by
1104 * SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1105 * If we insist on doing this as a semijoin we will first have
1106 * to form the cartesian product of A*B. But if we unique-ify
1107 * C then the semijoin becomes a plain innerjoin and we can join
1108 * in any order, eg C to A and then to B. When C is much smaller
1109 * than A and B this can be a huge win. So we allow C to be
1110 * joined to just A or just B here, and then make_join_rel has
1111 * to handle the case properly.
1113 * Note that actually we'll allow unique-ified C to be joined to
1114 * some other relation D here, too. That is legal, if usually not
1115 * very sane, and this routine is only concerned with legality not
1116 * with whether the join is good strategy.
1120 return false; /* invalid join path */
1121 match_sjinfo = sjinfo;
1123 unique_ified = true;
1125 else if (sjinfo->jointype == JOIN_SEMI &&
1126 bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1127 create_unique_path(root, rel1, rel1->cheapest_total_path,
1130 /* Reversed semijoin case */
1132 return false; /* invalid join path */
1133 match_sjinfo = sjinfo;
1135 unique_ified = true;
1140 * Otherwise, the proposed join overlaps the RHS but isn't a valid
1141 * implementation of this SJ. But don't panic quite yet: the RHS
1142 * violation might have occurred previously, in one or both input
1143 * relations, in which case we must have previously decided that
1144 * it was OK to commute some other SJ with this one. If we need
1145 * to perform this join to finish building up the RHS, rejecting
1146 * it could lead to not finding any plan at all. (This can occur
1147 * because of the heuristics elsewhere in this file that postpone
1148 * clauseless joins: we might not consider doing a clauseless join
1149 * within the RHS until after we've performed other, validly
1150 * commutable SJs with one or both sides of the clauseless join.)
1151 * This consideration boils down to the rule that if both inputs
1152 * overlap the RHS, we can allow the join --- they are either
1153 * fully within the RHS, or represent previously-allowed joins to
1156 if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1157 bms_overlap(rel2->relids, sjinfo->min_righthand))
1158 continue; /* assume valid previous violation of RHS */
1161 * The proposed join could still be legal, but only if we're
1162 * allowed to associate it into the RHS of this SJ. That means
1163 * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1164 * not FULL) and the proposed join must not overlap the LHS.
1166 if (sjinfo->jointype != JOIN_LEFT ||
1167 bms_overlap(joinrelids, sjinfo->min_lefthand))
1168 return false; /* invalid join path */
1171 * To be valid, the proposed join must be a LEFT join; otherwise
1172 * it can't associate into this SJ's RHS. But we may not yet have
1173 * found the SpecialJoinInfo matching the proposed join, so we
1174 * can't test that yet. Remember the requirement for later.
1176 must_be_leftjoin = true;
1181 * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1182 * proposed join can't associate into an SJ's RHS.
1184 * Also, fail if the proposed join's predicate isn't strict; we're
1185 * essentially checking to see if we can apply outer-join identity 3, and
1186 * that's a requirement. (This check may be redundant with checks in
1187 * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1189 if (must_be_leftjoin &&
1190 (match_sjinfo == NULL ||
1191 match_sjinfo->jointype != JOIN_LEFT ||
1192 !match_sjinfo->lhs_strict))
1193 return false; /* invalid join path */
1196 * We also have to check for constraints imposed by LATERAL references.
1198 if (root->hasLateralRTEs)
1202 Relids join_lateral_rels;
1205 * The proposed rels could each contain lateral references to the
1206 * other, in which case the join is impossible. If there are lateral
1207 * references in just one direction, then the join has to be done with
1208 * a nestloop with the lateral referencer on the inside. If the join
1209 * matches an SJ that cannot be implemented by such a nestloop, the
1210 * join is impossible.
1212 * Also, if the lateral reference is only indirect, we should reject
1213 * the join; whatever rel(s) the reference chain goes through must be
1216 * Another case that might keep us from building a valid plan is the
1217 * implementation restriction described by have_dangerous_phv().
1219 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1220 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1221 if (lateral_fwd && lateral_rev)
1222 return false; /* have lateral refs in both directions */
1225 /* has to be implemented as nestloop with rel1 on left */
1229 match_sjinfo->jointype == JOIN_FULL))
1230 return false; /* not implementable as nestloop */
1231 /* check there is a direct reference from rel2 to rel1 */
1232 if (!bms_overlap(rel1->relids, rel2->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, rel1->relids, rel2->lateral_relids))
1236 return false; /* might be unable to handle required PHV */
1238 else if (lateral_rev)
1240 /* has to be implemented as nestloop with rel2 on left */
1244 match_sjinfo->jointype == JOIN_FULL))
1245 return false; /* not implementable as nestloop */
1246 /* check there is a direct reference from rel1 to rel2 */
1247 if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1248 return false; /* only indirect refs, so reject */
1249 /* check we won't have a dangerous PHV */
1250 if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1251 return false; /* might be unable to handle required PHV */
1255 * LATERAL references could also cause problems later on if we accept
1256 * this join: if the join's minimum parameterization includes any rels
1257 * that would have to be on the inside of an outer join with this join
1258 * rel, then it's never going to be possible to build the complete
1259 * query using this join. We should reject this join not only because
1260 * it'll save work, but because if we don't, the clauseless-join
1261 * heuristics might think that legality of this join means that some
1262 * other join rel need not be formed, and that could lead to failure
1263 * to find any plan at all. We have to consider not only rels that
1264 * are directly on the inner side of an OJ with the joinrel, but also
1265 * ones that are indirectly so, so search to find all such rels.
1267 join_lateral_rels = min_join_parameterization(root, joinrelids,
1269 if (join_lateral_rels)
1271 Relids join_plus_rhs = bms_copy(joinrelids);
1277 foreach(l, root->join_info_list)
1279 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1281 if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1282 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1284 join_plus_rhs = bms_add_members(join_plus_rhs,
1285 sjinfo->min_righthand);
1288 /* full joins constrain both sides symmetrically */
1289 if (sjinfo->jointype == JOIN_FULL &&
1290 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1291 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1293 join_plus_rhs = bms_add_members(join_plus_rhs,
1294 sjinfo->min_lefthand);
1299 if (bms_overlap(join_plus_rhs, join_lateral_rels))
1300 return false; /* will not be able to join to some RHS rel */
1304 /* Otherwise, it's a valid join */
1305 *sjinfo_p = match_sjinfo;
1306 *reversed_p = reversed;
1311 * has_join_restriction
1312 * Detect whether the specified relation has join-order restrictions,
1313 * due to being inside an outer join or an IN (sub-SELECT),
1314 * or participating in any LATERAL references or multi-rel PHVs.
1316 * Essentially, this tests whether have_join_order_restriction() could
1317 * succeed with this rel and some other one. It's OK if we sometimes
1318 * say "true" incorrectly. (Therefore, we don't bother with the relatively
1319 * expensive has_legal_joinclause test.)
1322 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1326 if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1329 foreach(l, root->placeholder_list)
1331 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1333 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1334 !bms_equal(rel->relids, phinfo->ph_eval_at))
1338 foreach(l, root->join_info_list)
1340 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1342 /* ignore full joins --- other mechanisms preserve their ordering */
1343 if (sjinfo->jointype == JOIN_FULL)
1346 /* ignore if SJ is already contained in rel */
1347 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1348 bms_is_subset(sjinfo->min_righthand, rel->relids))
1351 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1352 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1353 bms_overlap(sjinfo->min_righthand, rel->relids))
1361 * is_dummy_rel --- has relation been proven empty?
1364 is_dummy_rel(RelOptInfo *rel)
1366 return IS_DUMMY_REL(rel);
1370 * Mark a relation as proven empty.
1372 * During GEQO planning, this can get invoked more than once on the same
1373 * baserel struct, so it's worth checking to see if the rel is already marked
1376 * Also, when called during GEQO join planning, we are in a short-lived
1377 * memory context. We must make sure that the dummy path attached to a
1378 * baserel survives the GEQO cycle, else the baserel is trashed for future
1379 * GEQO cycles. On the other hand, when we are marking a joinrel during GEQO,
1380 * we don't want the dummy path to clutter the main planning context. Upshot
1381 * is that the best solution is to explicitly make the dummy path in the same
1382 * context the given RelOptInfo is in.
1385 mark_dummy_rel(RelOptInfo *rel)
1387 MemoryContext oldcontext;
1389 /* Already marked? */
1390 if (is_dummy_rel(rel))
1393 /* No, so choose correct context to make the dummy path in */
1394 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1396 /* Set dummy size estimate */
1399 /* Evict any previously chosen paths */
1400 rel->pathlist = NIL;
1401 rel->partial_pathlist = NIL;
1403 /* Set up the dummy path */
1404 add_path(rel, (Path *) create_append_path(rel, NIL, NULL, 0));
1406 /* Set or update cheapest_total_path and related fields */
1409 MemoryContextSwitchTo(oldcontext);
1413 * restriction_is_constant_false --- is a restrictlist just FALSE?
1415 * In cases where a qual is provably constant FALSE, eval_const_expressions
1416 * will generally have thrown away anything that's ANDed with it. In outer
1417 * join situations this will leave us computing cartesian products only to
1418 * decide there's no match for an outer row, which is pretty stupid. So,
1419 * we need to detect the case.
1421 * If only_pushed_down is TRUE, then consider only pushed-down quals.
1424 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1429 * Despite the above comment, the restriction list we see here might
1430 * possibly have other members besides the FALSE constant, since other
1431 * quals could get "pushed down" to the outer join level. So we check
1432 * each member of the list.
1434 foreach(lc, restrictlist)
1436 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1438 Assert(IsA(rinfo, RestrictInfo));
1439 if (only_pushed_down && !rinfo->is_pushed_down)
1442 if (rinfo->clause && IsA(rinfo->clause, Const))
1444 Const *con = (Const *) rinfo->clause;
1446 /* constant NULL is as good as constant FALSE for our purposes */
1447 if (con->constisnull)
1449 if (!DatumGetBool(con->constvalue))