1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2006, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.206 2006/08/02 01:59:46 joe Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "executor/executor.h"
22 #include "executor/nodeAgg.h"
23 #include "miscadmin.h"
24 #include "nodes/makefuncs.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/cost.h"
27 #include "optimizer/pathnode.h"
28 #include "optimizer/paths.h"
29 #include "optimizer/planmain.h"
30 #include "optimizer/planner.h"
31 #include "optimizer/prep.h"
32 #include "optimizer/subselect.h"
33 #include "optimizer/tlist.h"
34 #include "optimizer/var.h"
35 #ifdef OPTIMIZER_DEBUG
36 #include "nodes/print.h"
38 #include "parser/parse_expr.h"
39 #include "parser/parse_oper.h"
40 #include "parser/parsetree.h"
41 #include "utils/syscache.h"
44 ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
47 /* Expression kind codes for preprocess_expression */
48 #define EXPRKIND_QUAL 0
49 #define EXPRKIND_TARGET 1
50 #define EXPRKIND_RTFUNC 2
51 #define EXPRKIND_VALUES 3
52 #define EXPRKIND_LIMIT 4
53 #define EXPRKIND_ININFO 5
54 #define EXPRKIND_APPINFO 6
57 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
58 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
59 static Plan *inheritance_planner(PlannerInfo *root);
60 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
61 static double preprocess_limit(PlannerInfo *root,
62 double tuple_fraction,
63 int64 *offset_est, int64 *count_est);
64 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
65 Path *cheapest_path, Path *sorted_path,
66 double dNumGroups, AggClauseCounts *agg_counts);
67 static bool hash_safe_grouping(PlannerInfo *root);
68 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
69 AttrNumber **groupColIdx, bool *need_tlist_eval);
70 static void locate_grouping_columns(PlannerInfo *root,
73 AttrNumber *groupColIdx);
74 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
77 /*****************************************************************************
79 * Query optimizer entry point
81 *****************************************************************************/
83 planner(Query *parse, bool isCursor, int cursorOptions,
84 ParamListInfo boundParams)
86 double tuple_fraction;
88 Index save_PlannerQueryLevel;
89 List *save_PlannerParamList;
90 ParamListInfo save_PlannerBoundParamList;
93 * The planner can be called recursively (an example is when
94 * eval_const_expressions tries to pre-evaluate an SQL function). So,
95 * these global state variables must be saved and restored.
97 * Query level and the param list cannot be moved into the per-query
98 * PlannerInfo structure since their whole purpose is communication across
99 * multiple sub-queries. Also, boundParams is explicitly info from outside
100 * the query, and so is likewise better handled as a global variable.
102 * Note we do NOT save and restore PlannerPlanId: it exists to assign
103 * unique IDs to SubPlan nodes, and we want those IDs to be unique for the
104 * life of a backend. Also, PlannerInitPlan is saved/restored in
105 * subquery_planner, not here.
107 save_PlannerQueryLevel = PlannerQueryLevel;
108 save_PlannerParamList = PlannerParamList;
109 save_PlannerBoundParamList = PlannerBoundParamList;
111 /* Initialize state for handling outer-level references and params */
112 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
113 PlannerParamList = NIL;
114 PlannerBoundParamList = boundParams;
116 /* Determine what fraction of the plan is likely to be scanned */
120 * We have no real idea how many tuples the user will ultimately FETCH
121 * from a cursor, but it seems a good bet that he doesn't want 'em
122 * all. Optimize for 10% retrieval (you gotta better number? Should
123 * this be a SETtable parameter?)
125 tuple_fraction = 0.10;
129 /* Default assumption is we need all the tuples */
130 tuple_fraction = 0.0;
133 /* primary planning entry point (may recurse for subqueries) */
134 result_plan = subquery_planner(parse, tuple_fraction, NULL);
136 /* check we popped out the right number of levels */
137 Assert(PlannerQueryLevel == 0);
140 * If creating a plan for a scrollable cursor, make sure it can run
141 * backwards on demand. Add a Material node at the top at need.
143 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
145 if (!ExecSupportsBackwardScan(result_plan))
146 result_plan = materialize_finished_plan(result_plan);
149 /* final cleanup of the plan */
150 result_plan = set_plan_references(result_plan, parse->rtable);
152 /* executor wants to know total number of Params used overall */
153 result_plan->nParamExec = list_length(PlannerParamList);
155 /* restore state for outer planner, if any */
156 PlannerQueryLevel = save_PlannerQueryLevel;
157 PlannerParamList = save_PlannerParamList;
158 PlannerBoundParamList = save_PlannerBoundParamList;
164 /*--------------------
166 * Invokes the planner on a subquery. We recurse to here for each
167 * sub-SELECT found in the query tree.
169 * parse is the querytree produced by the parser & rewriter.
170 * tuple_fraction is the fraction of tuples we expect will be retrieved.
171 * tuple_fraction is interpreted as explained for grouping_planner, below.
173 * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
174 * the output sort ordering of the completed plan.
176 * Basically, this routine does the stuff that should only be done once
177 * per Query object. It then calls grouping_planner. At one time,
178 * grouping_planner could be invoked recursively on the same Query object;
179 * that's not currently true, but we keep the separation between the two
180 * routines anyway, in case we need it again someday.
182 * subquery_planner will be called recursively to handle sub-Query nodes
183 * found within the query's expressions and rangetable.
185 * Returns a query plan.
186 *--------------------
189 subquery_planner(Query *parse, double tuple_fraction,
190 List **subquery_pathkeys)
192 List *saved_initplan = PlannerInitPlan;
193 int saved_planid = PlannerPlanId;
199 /* Set up for a new level of subquery */
201 PlannerInitPlan = NIL;
203 /* Create a PlannerInfo data structure for this subquery */
204 root = makeNode(PlannerInfo);
206 root->in_info_list = NIL;
207 root->append_rel_list = NIL;
210 * Look for IN clauses at the top level of WHERE, and transform them into
211 * joins. Note that this step only handles IN clauses originally at top
212 * level of WHERE; if we pull up any subqueries in the next step, their
213 * INs are processed just before pulling them up.
215 if (parse->hasSubLinks)
216 parse->jointree->quals = pull_up_IN_clauses(root,
217 parse->jointree->quals);
220 * Check to see if any subqueries in the rangetable can be merged into
223 parse->jointree = (FromExpr *)
224 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
227 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
228 * avoid the expense of doing flatten_join_alias_vars(). Also check for
229 * outer joins --- if none, we can skip reduce_outer_joins() and some
230 * other processing. This must be done after we have done
231 * pull_up_subqueries, of course.
233 * Note: if reduce_outer_joins manages to eliminate all outer joins,
234 * root->hasOuterJoins is not reset currently. This is OK since its
235 * purpose is merely to suppress unnecessary processing in simple cases.
237 root->hasJoinRTEs = false;
238 root->hasOuterJoins = false;
239 foreach(l, parse->rtable)
241 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
243 if (rte->rtekind == RTE_JOIN)
245 root->hasJoinRTEs = true;
246 if (IS_OUTER_JOIN(rte->jointype))
248 root->hasOuterJoins = true;
249 /* Can quit scanning once we find an outer join */
256 * Expand any rangetable entries that are inheritance sets into "append
257 * relations". This can add entries to the rangetable, but they must be
258 * plain base relations not joins, so it's OK (and marginally more
259 * efficient) to do it after checking for join RTEs. We must do it after
260 * pulling up subqueries, else we'd fail to handle inherited tables in
263 expand_inherited_tables(root);
266 * Set hasHavingQual to remember if HAVING clause is present. Needed
267 * because preprocess_expression will reduce a constant-true condition to
268 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
270 root->hasHavingQual = (parse->havingQual != NULL);
272 /* Clear this flag; might get set in distribute_qual_to_rels */
273 root->hasPseudoConstantQuals = false;
276 * Do expression preprocessing on targetlist and quals.
278 parse->targetList = (List *)
279 preprocess_expression(root, (Node *) parse->targetList,
282 preprocess_qual_conditions(root, (Node *) parse->jointree);
284 parse->havingQual = preprocess_expression(root, parse->havingQual,
287 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
289 parse->limitCount = preprocess_expression(root, parse->limitCount,
292 root->in_info_list = (List *)
293 preprocess_expression(root, (Node *) root->in_info_list,
295 root->append_rel_list = (List *)
296 preprocess_expression(root, (Node *) root->append_rel_list,
299 /* Also need to preprocess expressions for function and values RTEs */
300 foreach(l, parse->rtable)
302 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
304 if (rte->rtekind == RTE_FUNCTION)
305 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
307 else if (rte->rtekind == RTE_VALUES)
308 rte->values_lists = (List *)
309 preprocess_expression(root, (Node *) rte->values_lists,
314 * In some cases we may want to transfer a HAVING clause into WHERE. We
315 * cannot do so if the HAVING clause contains aggregates (obviously) or
316 * volatile functions (since a HAVING clause is supposed to be executed
317 * only once per group). Also, it may be that the clause is so expensive
318 * to execute that we're better off doing it only once per group, despite
319 * the loss of selectivity. This is hard to estimate short of doing the
320 * entire planning process twice, so we use a heuristic: clauses
321 * containing subplans are left in HAVING. Otherwise, we move or copy the
322 * HAVING clause into WHERE, in hopes of eliminating tuples before
323 * aggregation instead of after.
325 * If the query has explicit grouping then we can simply move such a
326 * clause into WHERE; any group that fails the clause will not be in the
327 * output because none of its tuples will reach the grouping or
328 * aggregation stage. Otherwise we must have a degenerate (variable-free)
329 * HAVING clause, which we put in WHERE so that query_planner() can use it
330 * in a gating Result node, but also keep in HAVING to ensure that we
331 * don't emit a bogus aggregated row. (This could be done better, but it
332 * seems not worth optimizing.)
334 * Note that both havingQual and parse->jointree->quals are in
335 * implicitly-ANDed-list form at this point, even though they are declared
339 foreach(l, (List *) parse->havingQual)
341 Node *havingclause = (Node *) lfirst(l);
343 if (contain_agg_clause(havingclause) ||
344 contain_volatile_functions(havingclause) ||
345 contain_subplans(havingclause))
347 /* keep it in HAVING */
348 newHaving = lappend(newHaving, havingclause);
350 else if (parse->groupClause)
352 /* move it to WHERE */
353 parse->jointree->quals = (Node *)
354 lappend((List *) parse->jointree->quals, havingclause);
358 /* put a copy in WHERE, keep it in HAVING */
359 parse->jointree->quals = (Node *)
360 lappend((List *) parse->jointree->quals,
361 copyObject(havingclause));
362 newHaving = lappend(newHaving, havingclause);
365 parse->havingQual = (Node *) newHaving;
368 * If we have any outer joins, try to reduce them to plain inner joins.
369 * This step is most easily done after we've done expression
372 if (root->hasOuterJoins)
373 reduce_outer_joins(root);
376 * Do the main planning. If we have an inherited target relation, that
377 * needs special processing, else go straight to grouping_planner.
379 if (parse->resultRelation &&
380 rt_fetch(parse->resultRelation, parse->rtable)->inh)
381 plan = inheritance_planner(root);
383 plan = grouping_planner(root, tuple_fraction);
386 * If any subplans were generated, or if we're inside a subplan, build
387 * initPlan list and extParam/allParam sets for plan nodes, and attach the
388 * initPlans to the top plan node.
390 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
391 SS_finalize_plan(plan, parse->rtable);
393 /* Return sort ordering info if caller wants it */
394 if (subquery_pathkeys)
395 *subquery_pathkeys = root->query_pathkeys;
397 /* Return to outer subquery context */
399 PlannerInitPlan = saved_initplan;
400 /* we do NOT restore PlannerPlanId; that's not an oversight! */
406 * preprocess_expression
407 * Do subquery_planner's preprocessing work for an expression,
408 * which can be a targetlist, a WHERE clause (including JOIN/ON
409 * conditions), or a HAVING clause.
412 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
415 * Fall out quickly if expression is empty. This occurs often enough to
416 * be worth checking. Note that null->null is the correct conversion for
417 * implicit-AND result format, too.
423 * If the query has any join RTEs, replace join alias variables with
424 * base-relation variables. We must do this before sublink processing,
425 * else sublinks expanded out from join aliases wouldn't get processed.
426 * We can skip it in VALUES lists, however, since they can't contain
429 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
430 expr = flatten_join_alias_vars(root, expr);
433 * Simplify constant expressions.
435 * Note: this also flattens nested AND and OR expressions into N-argument
436 * form. All processing of a qual expression after this point must be
437 * careful to maintain AND/OR flatness --- that is, do not generate a tree
438 * with AND directly under AND, nor OR directly under OR.
440 * Because this is a relatively expensive process, we skip it when the
441 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
442 * expression will only be evaluated once anyway, so no point in
443 * pre-simplifying; we can't execute it any faster than the executor can,
444 * and we will waste cycles copying the tree. Notice however that we
445 * still must do it for quals (to get AND/OR flatness); and if we are in a
446 * subquery we should not assume it will be done only once.
448 * For VALUES lists we never do this at all, again on the grounds that
449 * we should optimize for one-time evaluation.
451 if (kind != EXPRKIND_VALUES &&
452 (root->parse->jointree->fromlist != NIL ||
453 kind == EXPRKIND_QUAL ||
454 PlannerQueryLevel > 1))
455 expr = eval_const_expressions(expr);
458 * If it's a qual or havingQual, canonicalize it.
460 if (kind == EXPRKIND_QUAL)
462 expr = (Node *) canonicalize_qual((Expr *) expr);
464 #ifdef OPTIMIZER_DEBUG
465 printf("After canonicalize_qual()\n");
470 /* Expand SubLinks to SubPlans */
471 if (root->parse->hasSubLinks)
472 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
475 * XXX do not insert anything here unless you have grokked the comments in
476 * SS_replace_correlation_vars ...
479 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
480 if (PlannerQueryLevel > 1)
481 expr = SS_replace_correlation_vars(expr);
484 * If it's a qual or havingQual, convert it to implicit-AND format. (We
485 * don't want to do this before eval_const_expressions, since the latter
486 * would be unable to simplify a top-level AND correctly. Also,
487 * SS_process_sublinks expects explicit-AND format.)
489 if (kind == EXPRKIND_QUAL)
490 expr = (Node *) make_ands_implicit((Expr *) expr);
496 * preprocess_qual_conditions
497 * Recursively scan the query's jointree and do subquery_planner's
498 * preprocessing work on each qual condition found therein.
501 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
505 if (IsA(jtnode, RangeTblRef))
507 /* nothing to do here */
509 else if (IsA(jtnode, FromExpr))
511 FromExpr *f = (FromExpr *) jtnode;
514 foreach(l, f->fromlist)
515 preprocess_qual_conditions(root, lfirst(l));
517 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
519 else if (IsA(jtnode, JoinExpr))
521 JoinExpr *j = (JoinExpr *) jtnode;
523 preprocess_qual_conditions(root, j->larg);
524 preprocess_qual_conditions(root, j->rarg);
526 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
529 elog(ERROR, "unrecognized node type: %d",
530 (int) nodeTag(jtnode));
534 * inheritance_planner
535 * Generate a plan in the case where the result relation is an
538 * We have to handle this case differently from cases where a source relation
539 * is an inheritance set. Source inheritance is expanded at the bottom of the
540 * plan tree (see allpaths.c), but target inheritance has to be expanded at
541 * the top. The reason is that for UPDATE, each target relation needs a
542 * different targetlist matching its own column set. Also, for both UPDATE
543 * and DELETE, the executor needs the Append plan node at the top, else it
544 * can't keep track of which table is the current target table. Fortunately,
545 * the UPDATE/DELETE target can never be the nullable side of an outer join,
546 * so it's OK to generate the plan this way.
548 * Returns a query plan.
551 inheritance_planner(PlannerInfo *root)
553 Query *parse = root->parse;
554 int parentRTindex = parse->resultRelation;
555 List *subplans = NIL;
560 subroot.parse = NULL; /* catch it if no matches in loop */
562 parse->resultRelations = NIL;
564 foreach(l, root->append_rel_list)
566 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
569 /* append_rel_list contains all append rels; ignore others */
570 if (appinfo->parent_relid != parentRTindex)
573 /* Build target-relations list for the executor */
574 parse->resultRelations = lappend_int(parse->resultRelations,
575 appinfo->child_relid);
578 * Generate modified query with this rel as target. We have to be
579 * prepared to translate varnos in in_info_list as well as in the
582 memcpy(&subroot, root, sizeof(PlannerInfo));
583 subroot.parse = (Query *)
584 adjust_appendrel_attrs((Node *) parse,
586 subroot.in_info_list = (List *)
587 adjust_appendrel_attrs((Node *) root->in_info_list,
589 /* There shouldn't be any OJ info to translate, as yet */
590 Assert(subroot.oj_info_list == NIL);
593 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
595 subplans = lappend(subplans, subplan);
597 /* Save preprocessed tlist from first rel for use in Append */
599 tlist = subplan->targetlist;
603 * Planning might have modified the rangetable, due to changes of the
604 * Query structures inside subquery RTEs. We have to ensure that this
605 * gets propagated back to the master copy. But can't do this until we
606 * are done planning, because all the calls to grouping_planner need
607 * virgin sub-Queries to work from. (We are effectively assuming that
608 * sub-Queries will get planned identically each time, or at least that
609 * the impacts on their rangetables will be the same each time.)
611 * XXX should clean this up someday
613 parse->rtable = subroot.parse->rtable;
615 /* Mark result as unordered (probably unnecessary) */
616 root->query_pathkeys = NIL;
618 return (Plan *) make_append(subplans, true, tlist);
621 /*--------------------
623 * Perform planning steps related to grouping, aggregation, etc.
624 * This primarily means adding top-level processing to the basic
625 * query plan produced by query_planner.
627 * tuple_fraction is the fraction of tuples we expect will be retrieved
629 * tuple_fraction is interpreted as follows:
630 * 0: expect all tuples to be retrieved (normal case)
631 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
632 * from the plan to be retrieved
633 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
634 * expected to be retrieved (ie, a LIMIT specification)
636 * Returns a query plan. Also, root->query_pathkeys is returned as the
637 * actual output ordering of the plan (in pathkey format).
638 *--------------------
641 grouping_planner(PlannerInfo *root, double tuple_fraction)
643 Query *parse = root->parse;
644 List *tlist = parse->targetList;
645 int64 offset_est = 0;
648 List *current_pathkeys;
650 double dNumGroups = 0;
652 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
653 if (parse->limitCount || parse->limitOffset)
654 tuple_fraction = preprocess_limit(root, tuple_fraction,
655 &offset_est, &count_est);
657 if (parse->setOperations)
659 List *set_sortclauses;
662 * If there's a top-level ORDER BY, assume we have to fetch all the
663 * tuples. This might seem too simplistic given all the hackery below
664 * to possibly avoid the sort ... but a nonzero tuple_fraction is only
665 * of use to plan_set_operations() when the setop is UNION ALL, and
666 * the result of UNION ALL is always unsorted.
668 if (parse->sortClause)
669 tuple_fraction = 0.0;
672 * Construct the plan for set operations. The result will not need
673 * any work except perhaps a top-level sort and/or LIMIT.
675 result_plan = plan_set_operations(root, tuple_fraction,
679 * Calculate pathkeys representing the sort order (if any) of the set
680 * operation's result. We have to do this before overwriting the sort
683 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
684 result_plan->targetlist);
685 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
688 * We should not need to call preprocess_targetlist, since we must be
689 * in a SELECT query node. Instead, use the targetlist returned by
690 * plan_set_operations (since this tells whether it returned any
691 * resjunk columns!), and transfer any sort key information from the
694 Assert(parse->commandType == CMD_SELECT);
696 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
699 * Can't handle FOR UPDATE/SHARE here (parser should have checked
700 * already, but let's make sure).
704 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
705 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
708 * Calculate pathkeys that represent result ordering requirements
710 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
712 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
716 /* No set operations, do regular planning */
718 List *group_pathkeys;
719 AttrNumber *groupColIdx = NULL;
720 bool need_tlist_eval = true;
726 AggClauseCounts agg_counts;
727 int numGroupCols = list_length(parse->groupClause);
728 bool use_hashed_grouping = false;
730 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
732 /* Preprocess targetlist */
733 tlist = preprocess_targetlist(root, tlist);
736 * Generate appropriate target list for subplan; may be different from
737 * tlist if grouping or aggregation is needed.
739 sub_tlist = make_subplanTargetList(root, tlist,
740 &groupColIdx, &need_tlist_eval);
743 * Calculate pathkeys that represent grouping/ordering requirements.
744 * Stash them in PlannerInfo so that query_planner can canonicalize
747 root->group_pathkeys =
748 make_pathkeys_for_sortclauses(parse->groupClause, tlist);
749 root->sort_pathkeys =
750 make_pathkeys_for_sortclauses(parse->sortClause, tlist);
753 * Will need actual number of aggregates for estimating costs.
755 * Note: we do not attempt to detect duplicate aggregates here; a
756 * somewhat-overestimated count is okay for our present purposes.
758 * Note: think not that we can turn off hasAggs if we find no aggs. It
759 * is possible for constant-expression simplification to remove all
760 * explicit references to aggs, but we still have to follow the
761 * aggregate semantics (eg, producing only one output row).
765 count_agg_clauses((Node *) tlist, &agg_counts);
766 count_agg_clauses(parse->havingQual, &agg_counts);
770 * Figure out whether we need a sorted result from query_planner.
772 * If we have a GROUP BY clause, then we want a result sorted properly
773 * for grouping. Otherwise, if there is an ORDER BY clause, we want
774 * to sort by the ORDER BY clause. (Note: if we have both, and ORDER
775 * BY is a superset of GROUP BY, it would be tempting to request sort
776 * by ORDER BY --- but that might just leave us failing to exploit an
777 * available sort order at all. Needs more thought...)
779 if (parse->groupClause)
780 root->query_pathkeys = root->group_pathkeys;
781 else if (parse->sortClause)
782 root->query_pathkeys = root->sort_pathkeys;
784 root->query_pathkeys = NIL;
787 * Generate the best unsorted and presorted paths for this Query (but
788 * note there may not be any presorted path). query_planner will also
789 * estimate the number of groups in the query, and canonicalize all
792 query_planner(root, sub_tlist, tuple_fraction,
793 &cheapest_path, &sorted_path, &dNumGroups);
795 group_pathkeys = root->group_pathkeys;
796 sort_pathkeys = root->sort_pathkeys;
799 * If grouping, decide whether we want to use hashed grouping.
801 if (parse->groupClause)
803 use_hashed_grouping =
804 choose_hashed_grouping(root, tuple_fraction,
805 cheapest_path, sorted_path,
806 dNumGroups, &agg_counts);
808 /* Also convert # groups to long int --- but 'ware overflow! */
809 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
813 * Select the best path. If we are doing hashed grouping, we will
814 * always read all the input tuples, so use the cheapest-total path.
815 * Otherwise, trust query_planner's decision about which to use.
817 if (use_hashed_grouping || !sorted_path)
818 best_path = cheapest_path;
820 best_path = sorted_path;
823 * Check to see if it's possible to optimize MIN/MAX aggregates. If
824 * so, we will forget all the work we did so far to choose a "regular"
825 * path ... but we had to do it anyway to be able to tell which way is
828 result_plan = optimize_minmax_aggregates(root,
831 if (result_plan != NULL)
834 * optimize_minmax_aggregates generated the full plan, with the
835 * right tlist, and it has no sort order.
837 current_pathkeys = NIL;
842 * Normal case --- create a plan according to query_planner's
845 result_plan = create_plan(root, best_path);
846 current_pathkeys = best_path->pathkeys;
849 * create_plan() returns a plan with just a "flat" tlist of
850 * required Vars. Usually we need to insert the sub_tlist as the
851 * tlist of the top plan node. However, we can skip that if we
852 * determined that whatever query_planner chose to return will be
858 * If the top-level plan node is one that cannot do expression
859 * evaluation, we must insert a Result node to project the
862 if (!is_projection_capable_plan(result_plan))
864 result_plan = (Plan *) make_result(sub_tlist, NULL,
870 * Otherwise, just replace the subplan's flat tlist with
873 result_plan->targetlist = sub_tlist;
877 * Also, account for the cost of evaluation of the sub_tlist.
879 * Up to now, we have only been dealing with "flat" tlists,
880 * containing just Vars. So their evaluation cost is zero
881 * according to the model used by cost_qual_eval() (or if you
882 * prefer, the cost is factored into cpu_tuple_cost). Thus we
883 * can avoid accounting for tlist cost throughout
884 * query_planner() and subroutines. But now we've inserted a
885 * tlist that might contain actual operators, sub-selects, etc
886 * --- so we'd better account for its cost.
888 * Below this point, any tlist eval cost for added-on nodes
889 * should be accounted for as we create those nodes.
890 * Presently, of the node types we can add on, only Agg and
891 * Group project new tlists (the rest just copy their input
892 * tuples) --- so make_agg() and make_group() are responsible
893 * for computing the added cost.
895 cost_qual_eval(&tlist_cost, sub_tlist);
896 result_plan->startup_cost += tlist_cost.startup;
897 result_plan->total_cost += tlist_cost.startup +
898 tlist_cost.per_tuple * result_plan->plan_rows;
903 * Since we're using query_planner's tlist and not the one
904 * make_subplanTargetList calculated, we have to refigure any
905 * grouping-column indexes make_subplanTargetList computed.
907 locate_grouping_columns(root, tlist, result_plan->targetlist,
912 * Insert AGG or GROUP node if needed, plus an explicit sort step
915 * HAVING clause, if any, becomes qual of the Agg or Group node.
917 if (use_hashed_grouping)
919 /* Hashed aggregate plan --- no sort needed */
920 result_plan = (Plan *) make_agg(root,
922 (List *) parse->havingQual,
929 /* Hashed aggregation produces randomly-ordered results */
930 current_pathkeys = NIL;
932 else if (parse->hasAggs)
934 /* Plain aggregate plan --- sort if needed */
935 AggStrategy aggstrategy;
937 if (parse->groupClause)
939 if (!pathkeys_contained_in(group_pathkeys,
942 result_plan = (Plan *)
943 make_sort_from_groupcols(root,
947 current_pathkeys = group_pathkeys;
949 aggstrategy = AGG_SORTED;
952 * The AGG node will not change the sort ordering of its
953 * groups, so current_pathkeys describes the result too.
958 aggstrategy = AGG_PLAIN;
959 /* Result will be only one row anyway; no sort order */
960 current_pathkeys = NIL;
963 result_plan = (Plan *) make_agg(root,
965 (List *) parse->havingQual,
973 else if (parse->groupClause)
976 * GROUP BY without aggregation, so insert a group node (plus
977 * the appropriate sort node, if necessary).
979 * Add an explicit sort if we couldn't make the path come out
980 * the way the GROUP node needs it.
982 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
984 result_plan = (Plan *)
985 make_sort_from_groupcols(root,
989 current_pathkeys = group_pathkeys;
992 result_plan = (Plan *) make_group(root,
994 (List *) parse->havingQual,
999 /* The Group node won't change sort ordering */
1001 else if (root->hasHavingQual)
1004 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1005 * This is a degenerate case in which we are supposed to emit
1006 * either 0 or 1 row depending on whether HAVING succeeds.
1007 * Furthermore, there cannot be any variables in either HAVING
1008 * or the targetlist, so we actually do not need the FROM
1009 * table at all! We can just throw away the plan-so-far and
1010 * generate a Result node. This is a sufficiently unusual
1011 * corner case that it's not worth contorting the structure of
1012 * this routine to avoid having to generate the plan in the
1015 result_plan = (Plan *) make_result(tlist,
1019 } /* end of non-minmax-aggregate case */
1020 } /* end of if (setOperations) */
1023 * If we were not able to make the plan come out in the right order, add
1024 * an explicit sort step.
1026 if (parse->sortClause)
1028 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1030 result_plan = (Plan *)
1031 make_sort_from_sortclauses(root,
1034 current_pathkeys = sort_pathkeys;
1039 * If there is a DISTINCT clause, add the UNIQUE node.
1041 if (parse->distinctClause)
1043 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1046 * If there was grouping or aggregation, leave plan_rows as-is (ie,
1047 * assume the result was already mostly unique). If not, use the
1048 * number of distinct-groups calculated by query_planner.
1050 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1051 result_plan->plan_rows = dNumGroups;
1055 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1057 if (parse->limitCount || parse->limitOffset)
1059 result_plan = (Plan *) make_limit(result_plan,
1067 * Return the actual output ordering in query_pathkeys for possible use by
1068 * an outer query level.
1070 root->query_pathkeys = current_pathkeys;
1076 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1078 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1079 * results back in *count_est and *offset_est. These variables are set to
1080 * 0 if the corresponding clause is not present, and -1 if it's present
1081 * but we couldn't estimate the value for it. (The "0" convention is OK
1082 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1083 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1084 * usual practice of never estimating less than one row.) These values will
1085 * be passed to make_limit, which see if you change this code.
1087 * The return value is the suitably adjusted tuple_fraction to use for
1088 * planning the query. This adjustment is not overridable, since it reflects
1089 * plan actions that grouping_planner() will certainly take, not assumptions
1093 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1094 int64 *offset_est, int64 *count_est)
1096 Query *parse = root->parse;
1098 double limit_fraction;
1100 /* Should not be called unless LIMIT or OFFSET */
1101 Assert(parse->limitCount || parse->limitOffset);
1104 * Try to obtain the clause values. We use estimate_expression_value
1105 * primarily because it can sometimes do something useful with Params.
1107 if (parse->limitCount)
1109 est = estimate_expression_value(parse->limitCount);
1110 if (est && IsA(est, Const))
1112 if (((Const *) est)->constisnull)
1114 /* NULL indicates LIMIT ALL, ie, no limit */
1115 *count_est = 0; /* treat as not present */
1119 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1120 if (*count_est <= 0)
1121 *count_est = 1; /* force to at least 1 */
1125 *count_est = -1; /* can't estimate */
1128 *count_est = 0; /* not present */
1130 if (parse->limitOffset)
1132 est = estimate_expression_value(parse->limitOffset);
1133 if (est && IsA(est, Const))
1135 if (((Const *) est)->constisnull)
1137 /* Treat NULL as no offset; the executor will too */
1138 *offset_est = 0; /* treat as not present */
1142 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1143 if (*offset_est < 0)
1144 *offset_est = 0; /* less than 0 is same as 0 */
1148 *offset_est = -1; /* can't estimate */
1151 *offset_est = 0; /* not present */
1153 if (*count_est != 0)
1156 * A LIMIT clause limits the absolute number of tuples returned.
1157 * However, if it's not a constant LIMIT then we have to guess; for
1158 * lack of a better idea, assume 10% of the plan's result is wanted.
1160 if (*count_est < 0 || *offset_est < 0)
1162 /* LIMIT or OFFSET is an expression ... punt ... */
1163 limit_fraction = 0.10;
1167 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1168 limit_fraction = (double) *count_est + (double) *offset_est;
1172 * If we have absolute limits from both caller and LIMIT, use the
1173 * smaller value; likewise if they are both fractional. If one is
1174 * fractional and the other absolute, we can't easily determine which
1175 * is smaller, but we use the heuristic that the absolute will usually
1178 if (tuple_fraction >= 1.0)
1180 if (limit_fraction >= 1.0)
1183 tuple_fraction = Min(tuple_fraction, limit_fraction);
1187 /* caller absolute, limit fractional; use caller's value */
1190 else if (tuple_fraction > 0.0)
1192 if (limit_fraction >= 1.0)
1194 /* caller fractional, limit absolute; use limit */
1195 tuple_fraction = limit_fraction;
1199 /* both fractional */
1200 tuple_fraction = Min(tuple_fraction, limit_fraction);
1205 /* no info from caller, just use limit */
1206 tuple_fraction = limit_fraction;
1209 else if (*offset_est != 0 && tuple_fraction > 0.0)
1212 * We have an OFFSET but no LIMIT. This acts entirely differently
1213 * from the LIMIT case: here, we need to increase rather than decrease
1214 * the caller's tuple_fraction, because the OFFSET acts to cause more
1215 * tuples to be fetched instead of fewer. This only matters if we got
1216 * a tuple_fraction > 0, however.
1218 * As above, use 10% if OFFSET is present but unestimatable.
1220 if (*offset_est < 0)
1221 limit_fraction = 0.10;
1223 limit_fraction = (double) *offset_est;
1226 * If we have absolute counts from both caller and OFFSET, add them
1227 * together; likewise if they are both fractional. If one is
1228 * fractional and the other absolute, we want to take the larger, and
1229 * we heuristically assume that's the fractional one.
1231 if (tuple_fraction >= 1.0)
1233 if (limit_fraction >= 1.0)
1235 /* both absolute, so add them together */
1236 tuple_fraction += limit_fraction;
1240 /* caller absolute, limit fractional; use limit */
1241 tuple_fraction = limit_fraction;
1246 if (limit_fraction >= 1.0)
1248 /* caller fractional, limit absolute; use caller's value */
1252 /* both fractional, so add them together */
1253 tuple_fraction += limit_fraction;
1254 if (tuple_fraction >= 1.0)
1255 tuple_fraction = 0.0; /* assume fetch all */
1260 return tuple_fraction;
1264 * choose_hashed_grouping - should we use hashed grouping?
1267 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1268 Path *cheapest_path, Path *sorted_path,
1269 double dNumGroups, AggClauseCounts *agg_counts)
1271 int numGroupCols = list_length(root->parse->groupClause);
1272 double cheapest_path_rows;
1273 int cheapest_path_width;
1275 List *current_pathkeys;
1280 * Check can't-do-it conditions, including whether the grouping operators
1283 * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1284 * (Doing so would imply storing *all* the input values in the hash table,
1285 * which seems like a certain loser.)
1287 if (!enable_hashagg)
1289 if (agg_counts->numDistinctAggs != 0)
1291 if (!hash_safe_grouping(root))
1295 * Don't do it if it doesn't look like the hashtable will fit into
1298 * Beware here of the possibility that cheapest_path->parent is NULL. This
1299 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1301 if (cheapest_path->parent)
1303 cheapest_path_rows = cheapest_path->parent->rows;
1304 cheapest_path_width = cheapest_path->parent->width;
1308 cheapest_path_rows = 1; /* assume non-set result */
1309 cheapest_path_width = 100; /* arbitrary */
1312 /* Estimate per-hash-entry space at tuple width... */
1313 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1314 /* plus space for pass-by-ref transition values... */
1315 hashentrysize += agg_counts->transitionSpace;
1316 /* plus the per-hash-entry overhead */
1317 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1319 if (hashentrysize * dNumGroups > work_mem * 1024L)
1323 * See if the estimated cost is no more than doing it the other way. While
1324 * avoiding the need for sorted input is usually a win, the fact that the
1325 * output won't be sorted may be a loss; so we need to do an actual cost
1328 * We need to consider cheapest_path + hashagg [+ final sort] versus
1329 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1330 * presorted_path + group or agg [+ final sort] where brackets indicate a
1331 * step that may not be needed. We assume query_planner() will have
1332 * returned a presorted path only if it's a winner compared to
1333 * cheapest_path for this purpose.
1335 * These path variables are dummies that just hold cost fields; we don't
1336 * make actual Paths for these steps.
1338 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1339 numGroupCols, dNumGroups,
1340 cheapest_path->startup_cost, cheapest_path->total_cost,
1341 cheapest_path_rows);
1342 /* Result of hashed agg is always unsorted */
1343 if (root->sort_pathkeys)
1344 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1345 dNumGroups, cheapest_path_width);
1349 sorted_p.startup_cost = sorted_path->startup_cost;
1350 sorted_p.total_cost = sorted_path->total_cost;
1351 current_pathkeys = sorted_path->pathkeys;
1355 sorted_p.startup_cost = cheapest_path->startup_cost;
1356 sorted_p.total_cost = cheapest_path->total_cost;
1357 current_pathkeys = cheapest_path->pathkeys;
1359 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1361 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1362 cheapest_path_rows, cheapest_path_width);
1363 current_pathkeys = root->group_pathkeys;
1366 if (root->parse->hasAggs)
1367 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1368 numGroupCols, dNumGroups,
1369 sorted_p.startup_cost, sorted_p.total_cost,
1370 cheapest_path_rows);
1372 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1373 sorted_p.startup_cost, sorted_p.total_cost,
1374 cheapest_path_rows);
1375 /* The Agg or Group node will preserve ordering */
1376 if (root->sort_pathkeys &&
1377 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1378 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1379 dNumGroups, cheapest_path_width);
1382 * Now make the decision using the top-level tuple fraction. First we
1383 * have to convert an absolute count (LIMIT) into fractional form.
1385 if (tuple_fraction >= 1.0)
1386 tuple_fraction /= dNumGroups;
1388 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1389 tuple_fraction) < 0)
1391 /* Hashed is cheaper, so use it */
1398 * hash_safe_grouping - are grouping operators hashable?
1400 * We assume hashed aggregation will work if the datatype's equality operator
1401 * is marked hashjoinable.
1404 hash_safe_grouping(PlannerInfo *root)
1408 foreach(gl, root->parse->groupClause)
1410 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1411 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1412 root->parse->targetList);
1416 optup = equality_oper(exprType((Node *) tle->expr), true);
1419 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1420 ReleaseSysCache(optup);
1428 * make_subplanTargetList
1429 * Generate appropriate target list when grouping is required.
1431 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1432 * above the result of query_planner, we typically want to pass a different
1433 * target list to query_planner than the outer plan nodes should have.
1434 * This routine generates the correct target list for the subplan.
1436 * The initial target list passed from the parser already contains entries
1437 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1438 * for variables used only in HAVING clauses; so we need to add those
1439 * variables to the subplan target list. Also, we flatten all expressions
1440 * except GROUP BY items into their component variables; the other expressions
1441 * will be computed by the inserted nodes rather than by the subplan.
1442 * For example, given a query like
1443 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1444 * we want to pass this targetlist to the subplan:
1446 * where the a+b target will be used by the Sort/Group steps, and the
1447 * other targets will be used for computing the final results. (In the
1448 * above example we could theoretically suppress the a and b targets and
1449 * pass down only c,d,a+b, but it's not really worth the trouble to
1450 * eliminate simple var references from the subplan. We will avoid doing
1451 * the extra computation to recompute a+b at the outer level; see
1452 * replace_vars_with_subplan_refs() in setrefs.c.)
1454 * If we are grouping or aggregating, *and* there are no non-Var grouping
1455 * expressions, then the returned tlist is effectively dummy; we do not
1456 * need to force it to be evaluated, because all the Vars it contains
1457 * should be present in the output of query_planner anyway.
1459 * 'tlist' is the query's target list.
1460 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1461 * expressions (if there are any) in the subplan's target list.
1462 * 'need_tlist_eval' is set true if we really need to evaluate the
1465 * The result is the targetlist to be passed to the subplan.
1469 make_subplanTargetList(PlannerInfo *root,
1471 AttrNumber **groupColIdx,
1472 bool *need_tlist_eval)
1474 Query *parse = root->parse;
1479 *groupColIdx = NULL;
1482 * If we're not grouping or aggregating, there's nothing to do here;
1483 * query_planner should receive the unmodified target list.
1485 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1487 *need_tlist_eval = true;
1492 * Otherwise, start with a "flattened" tlist (having just the vars
1493 * mentioned in the targetlist and HAVING qual --- but not upper- level
1494 * Vars; they will be replaced by Params later on).
1496 sub_tlist = flatten_tlist(tlist);
1497 extravars = pull_var_clause(parse->havingQual, false);
1498 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1499 list_free(extravars);
1500 *need_tlist_eval = false; /* only eval if not flat tlist */
1503 * If grouping, create sub_tlist entries for all GROUP BY expressions
1504 * (GROUP BY items that are simple Vars should be in the list already),
1505 * and make an array showing where the group columns are in the sub_tlist.
1507 numCols = list_length(parse->groupClause);
1511 AttrNumber *grpColIdx;
1514 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1515 *groupColIdx = grpColIdx;
1517 foreach(gl, parse->groupClause)
1519 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1520 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1521 TargetEntry *te = NULL;
1524 /* Find or make a matching sub_tlist entry */
1525 foreach(sl, sub_tlist)
1527 te = (TargetEntry *) lfirst(sl);
1528 if (equal(groupexpr, te->expr))
1533 te = makeTargetEntry((Expr *) groupexpr,
1534 list_length(sub_tlist) + 1,
1537 sub_tlist = lappend(sub_tlist, te);
1538 *need_tlist_eval = true; /* it's not flat anymore */
1541 /* and save its resno */
1542 grpColIdx[keyno++] = te->resno;
1550 * locate_grouping_columns
1551 * Locate grouping columns in the tlist chosen by query_planner.
1553 * This is only needed if we don't use the sub_tlist chosen by
1554 * make_subplanTargetList. We have to forget the column indexes found
1555 * by that routine and re-locate the grouping vars in the real sub_tlist.
1558 locate_grouping_columns(PlannerInfo *root,
1561 AttrNumber *groupColIdx)
1567 * No work unless grouping.
1569 if (!root->parse->groupClause)
1571 Assert(groupColIdx == NULL);
1574 Assert(groupColIdx != NULL);
1576 foreach(gl, root->parse->groupClause)
1578 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1579 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1580 TargetEntry *te = NULL;
1583 foreach(sl, sub_tlist)
1585 te = (TargetEntry *) lfirst(sl);
1586 if (equal(groupexpr, te->expr))
1590 elog(ERROR, "failed to locate grouping columns");
1592 groupColIdx[keyno++] = te->resno;
1597 * postprocess_setop_tlist
1598 * Fix up targetlist returned by plan_set_operations().
1600 * We need to transpose sort key info from the orig_tlist into new_tlist.
1601 * NOTE: this would not be good enough if we supported resjunk sort keys
1602 * for results of set operations --- then, we'd need to project a whole
1603 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1604 * find any resjunk columns in orig_tlist.
1607 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1610 ListCell *orig_tlist_item = list_head(orig_tlist);
1612 foreach(l, new_tlist)
1614 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1615 TargetEntry *orig_tle;
1617 /* ignore resjunk columns in setop result */
1618 if (new_tle->resjunk)
1621 Assert(orig_tlist_item != NULL);
1622 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1623 orig_tlist_item = lnext(orig_tlist_item);
1624 if (orig_tle->resjunk) /* should not happen */
1625 elog(ERROR, "resjunk output columns are not implemented");
1626 Assert(new_tle->resno == orig_tle->resno);
1627 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1629 if (orig_tlist_item != NULL)
1630 elog(ERROR, "resjunk output columns are not implemented");