1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2005, 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 1.189 2005/06/10 02:21:04 tgl Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "catalog/pg_type.h"
22 #include "executor/executor.h"
23 #include "executor/nodeAgg.h"
24 #include "miscadmin.h"
25 #include "nodes/makefuncs.h"
26 #ifdef OPTIMIZER_DEBUG
27 #include "nodes/print.h"
29 #include "optimizer/clauses.h"
30 #include "optimizer/cost.h"
31 #include "optimizer/pathnode.h"
32 #include "optimizer/paths.h"
33 #include "optimizer/planmain.h"
34 #include "optimizer/planner.h"
35 #include "optimizer/prep.h"
36 #include "optimizer/subselect.h"
37 #include "optimizer/tlist.h"
38 #include "optimizer/var.h"
39 #include "parser/parsetree.h"
40 #include "parser/parse_expr.h"
41 #include "parser/parse_oper.h"
42 #include "utils/selfuncs.h"
43 #include "utils/syscache.h"
46 ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
49 /* Expression kind codes for preprocess_expression */
50 #define EXPRKIND_QUAL 0
51 #define EXPRKIND_TARGET 1
52 #define EXPRKIND_RTFUNC 2
53 #define EXPRKIND_LIMIT 3
54 #define EXPRKIND_ININFO 4
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, List *inheritlist);
60 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
61 static double adjust_tuple_fraction_for_limit(PlannerInfo *root,
62 double tuple_fraction);
63 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
64 Path *cheapest_path, Path *sorted_path,
65 List *sort_pathkeys, List *group_pathkeys,
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
99 * across multiple sub-queries. Also, boundParams is explicitly info
100 * from outside the query, and so is likewise better handled as a global
103 * Note we do NOT save and restore PlannerPlanId: it exists to assign
104 * unique IDs to SubPlan nodes, and we want those IDs to be unique for
105 * the life of a backend. Also, PlannerInitPlan is saved/restored in
106 * subquery_planner, not here.
108 save_PlannerQueryLevel = PlannerQueryLevel;
109 save_PlannerParamList = PlannerParamList;
110 save_PlannerBoundParamList = PlannerBoundParamList;
112 /* Initialize state for handling outer-level references and params */
113 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
114 PlannerParamList = NIL;
115 PlannerBoundParamList = boundParams;
117 /* Determine what fraction of the plan is likely to be scanned */
121 * We have no real idea how many tuples the user will ultimately
122 * FETCH from a cursor, but it seems a good bet that he doesn't
123 * want 'em all. Optimize for 10% retrieval (you gotta better
124 * number? Should this be a SETtable parameter?)
126 tuple_fraction = 0.10;
130 /* Default assumption is we need all the tuples */
131 tuple_fraction = 0.0;
134 /* primary planning entry point (may recurse for subqueries) */
135 result_plan = subquery_planner(parse, tuple_fraction, NULL);
137 /* check we popped out the right number of levels */
138 Assert(PlannerQueryLevel == 0);
141 * If creating a plan for a scrollable cursor, make sure it can run
142 * backwards on demand. Add a Material node at the top at need.
144 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
146 if (!ExecSupportsBackwardScan(result_plan))
147 result_plan = materialize_finished_plan(result_plan);
150 /* final cleanup of the plan */
151 result_plan = set_plan_references(result_plan, parse->rtable);
153 /* executor wants to know total number of Params used overall */
154 result_plan->nParamExec = list_length(PlannerParamList);
156 /* restore state for outer planner, if any */
157 PlannerQueryLevel = save_PlannerQueryLevel;
158 PlannerParamList = save_PlannerParamList;
159 PlannerBoundParamList = save_PlannerBoundParamList;
165 /*--------------------
167 * Invokes the planner on a subquery. We recurse to here for each
168 * sub-SELECT found in the query tree.
170 * parse is the querytree produced by the parser & rewriter.
171 * tuple_fraction is the fraction of tuples we expect will be retrieved.
172 * tuple_fraction is interpreted as explained for grouping_planner, below.
174 * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
175 * the output sort ordering of the completed plan.
177 * Basically, this routine does the stuff that should only be done once
178 * per Query object. It then calls grouping_planner. At one time,
179 * grouping_planner could be invoked recursively on the same Query object;
180 * that's not currently true, but we keep the separation between the two
181 * routines anyway, in case we need it again someday.
183 * subquery_planner will be called recursively to handle sub-Query nodes
184 * found within the query's expressions and rangetable.
186 * Returns a query plan.
187 *--------------------
190 subquery_planner(Query *parse, double tuple_fraction,
191 List **subquery_pathkeys)
193 List *saved_initplan = PlannerInitPlan;
194 int saved_planid = PlannerPlanId;
202 /* Set up for a new level of subquery */
204 PlannerInitPlan = NIL;
206 /* Create a PlannerInfo data structure for this subquery */
207 root = makeNode(PlannerInfo);
211 * Look for IN clauses at the top level of WHERE, and transform them
212 * into joins. Note that this step only handles IN clauses originally
213 * at top level of WHERE; if we pull up any subqueries in the next
214 * step, their INs are processed just before pulling them up.
216 root->in_info_list = NIL;
217 if (parse->hasSubLinks)
218 parse->jointree->quals = pull_up_IN_clauses(root,
219 parse->jointree->quals);
222 * Check to see if any subqueries in the rangetable can be merged into
225 parse->jointree = (FromExpr *)
226 pull_up_subqueries(root, (Node *) parse->jointree, false);
229 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
230 * can avoid the expense of doing flatten_join_alias_vars(). Also
231 * check for outer joins --- if none, we can skip
232 * reduce_outer_joins(). This must be done after we have done
233 * pull_up_subqueries, of course.
235 root->hasJoinRTEs = false;
236 hasOuterJoins = false;
237 foreach(l, parse->rtable)
239 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
241 if (rte->rtekind == RTE_JOIN)
243 root->hasJoinRTEs = true;
244 if (IS_OUTER_JOIN(rte->jointype))
246 hasOuterJoins = true;
247 /* Can quit scanning once we find an outer join */
254 * Set hasHavingQual to remember if HAVING clause is present. Needed
255 * because preprocess_expression will reduce a constant-true condition
256 * to an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
258 root->hasHavingQual = (parse->havingQual != NULL);
261 * Do expression preprocessing on targetlist and quals.
263 parse->targetList = (List *)
264 preprocess_expression(root, (Node *) parse->targetList,
267 preprocess_qual_conditions(root, (Node *) parse->jointree);
269 parse->havingQual = preprocess_expression(root, parse->havingQual,
272 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
274 parse->limitCount = preprocess_expression(root, parse->limitCount,
277 root->in_info_list = (List *)
278 preprocess_expression(root, (Node *) root->in_info_list,
281 /* Also need to preprocess expressions for function RTEs */
282 foreach(l, parse->rtable)
284 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
286 if (rte->rtekind == RTE_FUNCTION)
287 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
292 * In some cases we may want to transfer a HAVING clause into WHERE.
293 * We cannot do so if the HAVING clause contains aggregates (obviously)
294 * or volatile functions (since a HAVING clause is supposed to be executed
295 * only once per group). Also, it may be that the clause is so expensive
296 * to execute that we're better off doing it only once per group, despite
297 * the loss of selectivity. This is hard to estimate short of doing the
298 * entire planning process twice, so we use a heuristic: clauses
299 * containing subplans are left in HAVING. Otherwise, we move or copy
300 * the HAVING clause into WHERE, in hopes of eliminating tuples before
301 * aggregation instead of after.
303 * If the query has explicit grouping then we can simply move such a
304 * clause into WHERE; any group that fails the clause will not be
305 * in the output because none of its tuples will reach the grouping
306 * or aggregation stage. Otherwise we must have a degenerate
307 * (variable-free) HAVING clause, which we put in WHERE so that
308 * query_planner() can use it in a gating Result node, but also keep
309 * in HAVING to ensure that we don't emit a bogus aggregated row.
310 * (This could be done better, but it seems not worth optimizing.)
312 * Note that both havingQual and parse->jointree->quals are in
313 * implicitly-ANDed-list form at this point, even though they are
314 * declared as Node *.
317 foreach(l, (List *) parse->havingQual)
319 Node *havingclause = (Node *) lfirst(l);
321 if (contain_agg_clause(havingclause) ||
322 contain_volatile_functions(havingclause) ||
323 contain_subplans(havingclause))
325 /* keep it in HAVING */
326 newHaving = lappend(newHaving, havingclause);
328 else if (parse->groupClause)
330 /* move it to WHERE */
331 parse->jointree->quals = (Node *)
332 lappend((List *) parse->jointree->quals, havingclause);
336 /* put a copy in WHERE, keep it in HAVING */
337 parse->jointree->quals = (Node *)
338 lappend((List *) parse->jointree->quals,
339 copyObject(havingclause));
340 newHaving = lappend(newHaving, havingclause);
343 parse->havingQual = (Node *) newHaving;
346 * If we have any outer joins, try to reduce them to plain inner
347 * joins. This step is most easily done after we've done expression
351 reduce_outer_joins(root);
354 * See if we can simplify the jointree; opportunities for this may
355 * come from having pulled up subqueries, or from flattening explicit
356 * JOIN syntax. We must do this after flattening JOIN alias
357 * variables, since eliminating explicit JOIN nodes from the jointree
358 * will cause get_relids_for_join() to fail. But it should happen
359 * after reduce_outer_joins, anyway.
361 parse->jointree = (FromExpr *)
362 simplify_jointree(root, (Node *) parse->jointree);
365 * Do the main planning. If we have an inherited target relation,
366 * that needs special processing, else go straight to
369 if (parse->resultRelation &&
370 (lst = expand_inherited_rtentry(root, parse->resultRelation)) != NIL)
371 plan = inheritance_planner(root, lst);
373 plan = grouping_planner(root, tuple_fraction);
376 * If any subplans were generated, or if we're inside a subplan, build
377 * initPlan list and extParam/allParam sets for plan nodes, and attach
378 * the initPlans to the top plan node.
380 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
381 SS_finalize_plan(plan, parse->rtable);
383 /* Return sort ordering info if caller wants it */
384 if (subquery_pathkeys)
385 *subquery_pathkeys = root->query_pathkeys;
387 /* Return to outer subquery context */
389 PlannerInitPlan = saved_initplan;
390 /* we do NOT restore PlannerPlanId; that's not an oversight! */
396 * preprocess_expression
397 * Do subquery_planner's preprocessing work for an expression,
398 * which can be a targetlist, a WHERE clause (including JOIN/ON
399 * conditions), or a HAVING clause.
402 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
405 * Fall out quickly if expression is empty. This occurs often enough
406 * to be worth checking. Note that null->null is the correct conversion
407 * for implicit-AND result format, too.
413 * If the query has any join RTEs, replace join alias variables with
414 * base-relation variables. We must do this before sublink processing,
415 * else sublinks expanded out from join aliases wouldn't get
418 if (root->hasJoinRTEs)
419 expr = flatten_join_alias_vars(root, expr);
422 * Simplify constant expressions.
424 * Note: this also flattens nested AND and OR expressions into N-argument
425 * form. All processing of a qual expression after this point must be
426 * careful to maintain AND/OR flatness --- that is, do not generate a tree
427 * with AND directly under AND, nor OR directly under OR.
429 * Because this is a relatively expensive process, we skip it when the
430 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()".
431 * The expression will only be evaluated once anyway, so no point in
432 * pre-simplifying; we can't execute it any faster than the executor can,
433 * and we will waste cycles copying the tree. Notice however that we
434 * still must do it for quals (to get AND/OR flatness); and if we are
435 * in a subquery we should not assume it will be done only once.
437 if (root->parse->jointree->fromlist != NIL ||
438 kind == EXPRKIND_QUAL ||
439 PlannerQueryLevel > 1)
440 expr = eval_const_expressions(expr);
443 * If it's a qual or havingQual, canonicalize it.
445 if (kind == EXPRKIND_QUAL)
447 expr = (Node *) canonicalize_qual((Expr *) expr);
449 #ifdef OPTIMIZER_DEBUG
450 printf("After canonicalize_qual()\n");
455 /* Expand SubLinks to SubPlans */
456 if (root->parse->hasSubLinks)
457 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
460 * XXX do not insert anything here unless you have grokked the
461 * comments in SS_replace_correlation_vars ...
464 /* Replace uplevel vars with Param nodes */
465 if (PlannerQueryLevel > 1)
466 expr = SS_replace_correlation_vars(expr);
469 * If it's a qual or havingQual, convert it to implicit-AND format.
470 * (We don't want to do this before eval_const_expressions, since the
471 * latter would be unable to simplify a top-level AND correctly. Also,
472 * SS_process_sublinks expects explicit-AND format.)
474 if (kind == EXPRKIND_QUAL)
475 expr = (Node *) make_ands_implicit((Expr *) expr);
481 * preprocess_qual_conditions
482 * Recursively scan the query's jointree and do subquery_planner's
483 * preprocessing work on each qual condition found therein.
486 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
490 if (IsA(jtnode, RangeTblRef))
492 /* nothing to do here */
494 else if (IsA(jtnode, FromExpr))
496 FromExpr *f = (FromExpr *) jtnode;
499 foreach(l, f->fromlist)
500 preprocess_qual_conditions(root, lfirst(l));
502 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
504 else if (IsA(jtnode, JoinExpr))
506 JoinExpr *j = (JoinExpr *) jtnode;
508 preprocess_qual_conditions(root, j->larg);
509 preprocess_qual_conditions(root, j->rarg);
511 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
514 elog(ERROR, "unrecognized node type: %d",
515 (int) nodeTag(jtnode));
518 /*--------------------
519 * inheritance_planner
520 * Generate a plan in the case where the result relation is an
523 * We have to handle this case differently from cases where a source
524 * relation is an inheritance set. Source inheritance is expanded at
525 * the bottom of the plan tree (see allpaths.c), but target inheritance
526 * has to be expanded at the top. The reason is that for UPDATE, each
527 * target relation needs a different targetlist matching its own column
528 * set. (This is not so critical for DELETE, but for simplicity we treat
529 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
530 * can never be the nullable side of an outer join, so it's OK to generate
533 * inheritlist is an integer list of RT indexes for the result relation set.
535 * Returns a query plan.
536 *--------------------
539 inheritance_planner(PlannerInfo *root, List *inheritlist)
541 Query *parse = root->parse;
542 int parentRTindex = parse->resultRelation;
543 Oid parentOID = getrelid(parentRTindex, parse->rtable);
544 int mainrtlength = list_length(parse->rtable);
545 List *subplans = NIL;
549 foreach(l, inheritlist)
551 int childRTindex = lfirst_int(l);
552 Oid childOID = getrelid(childRTindex, parse->rtable);
557 * Generate modified query with this rel as target. We have to
558 * be prepared to translate varnos in in_info_list as well as in
561 memcpy(&subroot, root, sizeof(PlannerInfo));
562 subroot.parse = (Query *)
563 adjust_inherited_attrs((Node *) parse,
564 parentRTindex, parentOID,
565 childRTindex, childOID);
566 subroot.in_info_list = (List *)
567 adjust_inherited_attrs((Node *) root->in_info_list,
568 parentRTindex, parentOID,
569 childRTindex, childOID);
572 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
574 subplans = lappend(subplans, subplan);
577 * XXX my goodness this next bit is ugly. Really need to think about
578 * ways to rein in planner's habit of scribbling on its input.
580 * Planning of the subquery might have modified the rangetable,
581 * either by addition of RTEs due to expansion of inherited source
582 * tables, or by changes of the Query structures inside subquery
583 * RTEs. We have to ensure that this gets propagated back to the
584 * master copy. However, if we aren't done planning yet, we also
585 * need to ensure that subsequent calls to grouping_planner have
586 * virgin sub-Queries to work from. So, if we are at the last
587 * list entry, just copy the subquery rangetable back to the master
588 * copy; if we are not, then extend the master copy by adding
589 * whatever the subquery added. (We assume these added entries
590 * will go untouched by the future grouping_planner calls. We are
591 * also effectively assuming that sub-Queries will get planned
592 * identically each time, or at least that the impacts on their
593 * rangetables will be the same each time. Did I say this is ugly?)
595 if (lnext(l) == NULL)
596 parse->rtable = subroot.parse->rtable;
599 int subrtlength = list_length(subroot.parse->rtable);
601 if (subrtlength > mainrtlength)
605 subrt = list_copy_tail(subroot.parse->rtable, mainrtlength);
606 parse->rtable = list_concat(parse->rtable, subrt);
607 mainrtlength = subrtlength;
611 /* Save preprocessed tlist from first rel for use in Append */
613 tlist = subplan->targetlist;
616 /* Save the target-relations list for the executor, too */
617 parse->resultRelations = inheritlist;
619 /* Mark result as unordered (probably unnecessary) */
620 root->query_pathkeys = NIL;
622 return (Plan *) make_append(subplans, true, tlist);
625 /*--------------------
627 * Perform planning steps related to grouping, aggregation, etc.
628 * This primarily means adding top-level processing to the basic
629 * query plan produced by query_planner.
631 * tuple_fraction is the fraction of tuples we expect will be retrieved
633 * tuple_fraction is interpreted as follows:
634 * 0: expect all tuples to be retrieved (normal case)
635 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
636 * from the plan to be retrieved
637 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
638 * expected to be retrieved (ie, a LIMIT specification)
640 * Returns a query plan. Also, root->query_pathkeys is returned as the
641 * actual output ordering of the plan (in pathkey format).
642 *--------------------
645 grouping_planner(PlannerInfo *root, double tuple_fraction)
647 Query *parse = root->parse;
648 List *tlist = parse->targetList;
650 List *current_pathkeys;
653 /* Tweak caller-supplied tuple_fraction if have LIMIT */
654 if (parse->limitCount != NULL)
655 tuple_fraction = adjust_tuple_fraction_for_limit(root, tuple_fraction);
657 if (parse->setOperations)
659 List *set_sortclauses;
662 * If there's a top-level ORDER BY, assume we have to fetch all
663 * the tuples. This might seem too simplistic given all the
664 * hackery below to possibly avoid the sort ... but a nonzero
665 * tuple_fraction is only of use to plan_set_operations() when
666 * the setop is UNION ALL, and the result of UNION ALL is always
669 if (parse->sortClause)
670 tuple_fraction = 0.0;
673 * Construct the plan for set operations. The result will not
674 * need any work except perhaps a top-level sort and/or LIMIT.
676 result_plan = plan_set_operations(root, tuple_fraction,
680 * Calculate pathkeys representing the sort order (if any) of the
681 * set operation's result. We have to do this before overwriting
682 * the sort key information...
684 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
685 result_plan->targetlist);
686 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
689 * We should not need to call preprocess_targetlist, since we must
690 * be in a SELECT query node. Instead, use the targetlist
691 * returned by plan_set_operations (since this tells whether it
692 * returned any resjunk columns!), and transfer any sort key
693 * information from the original tlist.
695 Assert(parse->commandType == CMD_SELECT);
697 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
700 * Can't handle FOR UPDATE/SHARE here (parser should have checked
701 * already, but let's make sure).
705 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
706 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
709 * Calculate pathkeys that represent result ordering requirements
711 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
713 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
717 /* No set operations, do regular planning */
719 List *group_pathkeys;
720 AttrNumber *groupColIdx = NULL;
721 bool need_tlist_eval = true;
723 double sub_tuple_fraction;
727 double dNumGroups = 0;
729 AggClauseCounts agg_counts;
730 int numGroupCols = list_length(parse->groupClause);
731 bool use_hashed_grouping = false;
733 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
735 /* Preprocess targetlist */
736 tlist = preprocess_targetlist(root, tlist);
739 * Generate appropriate target list for subplan; may be different
740 * from tlist if grouping or aggregation is needed.
742 sub_tlist = make_subplanTargetList(root, tlist,
743 &groupColIdx, &need_tlist_eval);
746 * Calculate pathkeys that represent grouping/ordering
749 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
751 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
755 * Will need actual number of aggregates for estimating costs.
757 * Note: we do not attempt to detect duplicate aggregates here; a
758 * somewhat-overestimated count is okay for our present purposes.
760 * Note: think not that we can turn off hasAggs if we find no aggs.
761 * It is possible for constant-expression simplification to remove
762 * all explicit references to aggs, but we still have to follow
763 * the aggregate semantics (eg, producing only one output row).
767 count_agg_clauses((Node *) tlist, &agg_counts);
768 count_agg_clauses(parse->havingQual, &agg_counts);
772 * Figure out whether we need a sorted result from query_planner.
774 * If we have a GROUP BY clause, then we want a result sorted
775 * properly for grouping. Otherwise, if there is an ORDER BY
776 * clause, we want to sort by the ORDER BY clause. (Note: if we
777 * have both, and ORDER BY is a superset of GROUP BY, it would be
778 * tempting to request sort by ORDER BY --- but that might just
779 * leave us failing to exploit an available sort order at all.
780 * Needs more thought...)
782 if (parse->groupClause)
783 root->query_pathkeys = group_pathkeys;
784 else if (parse->sortClause)
785 root->query_pathkeys = sort_pathkeys;
787 root->query_pathkeys = NIL;
790 * With grouping or aggregation, the tuple fraction to pass to
791 * query_planner() may be different from what it is at top level.
793 sub_tuple_fraction = tuple_fraction;
795 if (parse->groupClause)
798 * In GROUP BY mode, we have the little problem that we don't
799 * really know how many input tuples will be needed to make a
800 * group, so we can't translate an output LIMIT count into an
801 * input count. For lack of a better idea, assume 25% of the
802 * input data will be processed if there is any output limit.
803 * However, if the caller gave us a fraction rather than an
804 * absolute count, we can keep using that fraction (which
805 * amounts to assuming that all the groups are about the same
808 if (sub_tuple_fraction >= 1.0)
809 sub_tuple_fraction = 0.25;
812 * If both GROUP BY and ORDER BY are specified, we will need
813 * two levels of sort --- and, therefore, certainly need to
814 * read all the input tuples --- unless ORDER BY is a subset
815 * of GROUP BY. (We have not yet canonicalized the pathkeys,
816 * so must use the slower noncanonical comparison method.)
818 if (parse->groupClause && parse->sortClause &&
819 !noncanonical_pathkeys_contained_in(sort_pathkeys,
821 sub_tuple_fraction = 0.0;
823 else if (parse->hasAggs)
826 * Ungrouped aggregate will certainly want all the input
829 sub_tuple_fraction = 0.0;
831 else if (parse->distinctClause)
834 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
835 * number of input tuples per output tuple. Handle the same
838 if (sub_tuple_fraction >= 1.0)
839 sub_tuple_fraction = 0.25;
843 * Generate the best unsorted and presorted paths for this Query
844 * (but note there may not be any presorted path).
846 query_planner(root, sub_tlist, sub_tuple_fraction,
847 &cheapest_path, &sorted_path);
850 * We couldn't canonicalize group_pathkeys and sort_pathkeys
851 * before running query_planner(), so do it now.
853 group_pathkeys = canonicalize_pathkeys(root, group_pathkeys);
854 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
857 * If grouping, estimate the number of groups. (We can't do this
858 * until after running query_planner(), either.) Then decide
859 * whether we want to use hashed grouping.
861 if (parse->groupClause)
864 double cheapest_path_rows;
867 * Beware of the possibility that cheapest_path->parent is NULL.
868 * This could happen if user does something silly like
869 * SELECT 'foo' GROUP BY 1;
871 if (cheapest_path->parent)
872 cheapest_path_rows = cheapest_path->parent->rows;
874 cheapest_path_rows = 1; /* assume non-set result */
876 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
878 dNumGroups = estimate_num_groups(root,
881 /* Also want it as a long int --- but 'ware overflow! */
882 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
884 use_hashed_grouping =
885 choose_hashed_grouping(root, tuple_fraction,
886 cheapest_path, sorted_path,
887 sort_pathkeys, group_pathkeys,
888 dNumGroups, &agg_counts);
892 * Select the best path. If we are doing hashed grouping, we will
893 * always read all the input tuples, so use the cheapest-total
894 * path. Otherwise, trust query_planner's decision about which to use.
896 if (use_hashed_grouping || !sorted_path)
897 best_path = cheapest_path;
899 best_path = sorted_path;
902 * Check to see if it's possible to optimize MIN/MAX aggregates.
903 * If so, we will forget all the work we did so far to choose a
904 * "regular" path ... but we had to do it anyway to be able to
905 * tell which way is cheaper.
907 result_plan = optimize_minmax_aggregates(root,
910 if (result_plan != NULL)
913 * optimize_minmax_aggregates generated the full plan, with
914 * the right tlist, and it has no sort order.
916 current_pathkeys = NIL;
921 * Normal case --- create a plan according to query_planner's
924 result_plan = create_plan(root, best_path);
925 current_pathkeys = best_path->pathkeys;
928 * create_plan() returns a plan with just a "flat" tlist of
929 * required Vars. Usually we need to insert the sub_tlist as the
930 * tlist of the top plan node. However, we can skip that if we
931 * determined that whatever query_planner chose to return will be
937 * If the top-level plan node is one that cannot do expression
938 * evaluation, we must insert a Result node to project the
941 if (!is_projection_capable_plan(result_plan))
943 result_plan = (Plan *) make_result(sub_tlist, NULL,
949 * Otherwise, just replace the subplan's flat tlist with
952 result_plan->targetlist = sub_tlist;
956 * Also, account for the cost of evaluation of the sub_tlist.
958 * Up to now, we have only been dealing with "flat" tlists,
959 * containing just Vars. So their evaluation cost is zero
960 * according to the model used by cost_qual_eval() (or if you
961 * prefer, the cost is factored into cpu_tuple_cost). Thus we
962 * can avoid accounting for tlist cost throughout
963 * query_planner() and subroutines. But now we've inserted a
964 * tlist that might contain actual operators, sub-selects, etc
965 * --- so we'd better account for its cost.
967 * Below this point, any tlist eval cost for added-on nodes
968 * should be accounted for as we create those nodes.
969 * Presently, of the node types we can add on, only Agg and
970 * Group project new tlists (the rest just copy their input
971 * tuples) --- so make_agg() and make_group() are responsible
972 * for computing the added cost.
974 cost_qual_eval(&tlist_cost, sub_tlist);
975 result_plan->startup_cost += tlist_cost.startup;
976 result_plan->total_cost += tlist_cost.startup +
977 tlist_cost.per_tuple * result_plan->plan_rows;
982 * Since we're using query_planner's tlist and not the one
983 * make_subplanTargetList calculated, we have to refigure any
984 * grouping-column indexes make_subplanTargetList computed.
986 locate_grouping_columns(root, tlist, result_plan->targetlist,
991 * Insert AGG or GROUP node if needed, plus an explicit sort step
994 * HAVING clause, if any, becomes qual of the Agg or Group node.
996 if (use_hashed_grouping)
998 /* Hashed aggregate plan --- no sort needed */
999 result_plan = (Plan *) make_agg(root,
1001 (List *) parse->havingQual,
1008 /* Hashed aggregation produces randomly-ordered results */
1009 current_pathkeys = NIL;
1011 else if (parse->hasAggs)
1013 /* Plain aggregate plan --- sort if needed */
1014 AggStrategy aggstrategy;
1016 if (parse->groupClause)
1018 if (!pathkeys_contained_in(group_pathkeys,
1021 result_plan = (Plan *)
1022 make_sort_from_groupcols(root,
1026 current_pathkeys = group_pathkeys;
1028 aggstrategy = AGG_SORTED;
1031 * The AGG node will not change the sort ordering of its
1032 * groups, so current_pathkeys describes the result too.
1037 aggstrategy = AGG_PLAIN;
1038 /* Result will be only one row anyway; no sort order */
1039 current_pathkeys = NIL;
1042 result_plan = (Plan *) make_agg(root,
1044 (List *) parse->havingQual,
1052 else if (parse->groupClause)
1055 * GROUP BY without aggregation, so insert a group node (plus
1056 * the appropriate sort node, if necessary).
1058 * Add an explicit sort if we couldn't make the path come
1059 * out the way the GROUP node needs it.
1061 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1063 result_plan = (Plan *)
1064 make_sort_from_groupcols(root,
1068 current_pathkeys = group_pathkeys;
1071 result_plan = (Plan *) make_group(root,
1073 (List *) parse->havingQual,
1078 /* The Group node won't change sort ordering */
1080 else if (root->hasHavingQual)
1083 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1084 * This is a degenerate case in which we are supposed to emit
1085 * either 0 or 1 row depending on whether HAVING succeeds.
1086 * Furthermore, there cannot be any variables in either HAVING
1087 * or the targetlist, so we actually do not need the FROM table
1088 * at all! We can just throw away the plan-so-far and generate
1089 * a Result node. This is a sufficiently unusual corner case
1090 * that it's not worth contorting the structure of this routine
1091 * to avoid having to generate the plan in the first place.
1093 result_plan = (Plan *) make_result(tlist,
1097 } /* end of non-minmax-aggregate case */
1098 } /* end of if (setOperations) */
1101 * If we were not able to make the plan come out in the right order,
1102 * add an explicit sort step.
1104 if (parse->sortClause)
1106 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1108 result_plan = (Plan *)
1109 make_sort_from_sortclauses(root,
1112 current_pathkeys = sort_pathkeys;
1117 * If there is a DISTINCT clause, add the UNIQUE node.
1119 if (parse->distinctClause)
1121 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1124 * If there was grouping or aggregation, leave plan_rows as-is
1125 * (ie, assume the result was already mostly unique). If not,
1126 * it's reasonable to assume the UNIQUE filter has effects
1127 * comparable to GROUP BY.
1129 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1131 List *distinctExprs;
1133 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1135 result_plan->plan_rows = estimate_num_groups(root,
1137 result_plan->plan_rows);
1142 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1144 if (parse->limitOffset || parse->limitCount)
1146 result_plan = (Plan *) make_limit(result_plan,
1152 * Return the actual output ordering in query_pathkeys for possible
1153 * use by an outer query level.
1155 root->query_pathkeys = current_pathkeys;
1161 * adjust_tuple_fraction_for_limit - adjust tuple fraction for LIMIT
1163 * If the query contains LIMIT, we adjust the caller-supplied tuple_fraction
1164 * accordingly. This is not overridable by the caller, since it reflects plan
1165 * actions that grouping_planner() will certainly take, not assumptions about
1169 adjust_tuple_fraction_for_limit(PlannerInfo *root, double tuple_fraction)
1171 Query *parse = root->parse;
1172 double limit_fraction = 0.0;
1174 /* Should not be called unless LIMIT */
1175 Assert(parse->limitCount != NULL);
1178 * A LIMIT clause limits the absolute number of tuples returned. However,
1179 * if it's not a constant LIMIT then we have to punt; for lack of a better
1180 * idea, assume 10% of the plan's result is wanted.
1182 if (IsA(parse->limitCount, Const))
1184 Const *limitc = (Const *) parse->limitCount;
1185 int32 count = DatumGetInt32(limitc->constvalue);
1188 * A NULL-constant LIMIT represents "LIMIT ALL", which we treat the
1189 * same as no limit (ie, expect to retrieve all the tuples).
1191 if (!limitc->constisnull && count > 0)
1193 limit_fraction = (double) count;
1194 /* We must also consider the OFFSET, if present */
1195 if (parse->limitOffset != NULL)
1197 if (IsA(parse->limitOffset, Const))
1201 limitc = (Const *) parse->limitOffset;
1202 offset = DatumGetInt32(limitc->constvalue);
1203 if (!limitc->constisnull && offset > 0)
1204 limit_fraction += (double) offset;
1208 /* OFFSET is an expression ... punt ... */
1209 limit_fraction = 0.10;
1216 /* LIMIT is an expression ... punt ... */
1217 limit_fraction = 0.10;
1220 if (limit_fraction > 0.0)
1223 * If we have absolute limits from both caller and LIMIT, use the
1224 * smaller value; if one is fractional and the other absolute,
1225 * treat the fraction as a fraction of the absolute value;
1226 * else we can multiply the two fractions together.
1228 if (tuple_fraction >= 1.0)
1230 if (limit_fraction >= 1.0)
1233 tuple_fraction = Min(tuple_fraction, limit_fraction);
1237 /* caller absolute, limit fractional */
1238 tuple_fraction *= limit_fraction;
1239 if (tuple_fraction < 1.0)
1240 tuple_fraction = 1.0;
1243 else if (tuple_fraction > 0.0)
1245 if (limit_fraction >= 1.0)
1247 /* caller fractional, limit absolute */
1248 tuple_fraction *= limit_fraction;
1249 if (tuple_fraction < 1.0)
1250 tuple_fraction = 1.0;
1254 /* both fractional */
1255 tuple_fraction *= limit_fraction;
1260 /* no info from caller, just use limit */
1261 tuple_fraction = limit_fraction;
1265 return tuple_fraction;
1269 * choose_hashed_grouping - should we use hashed grouping?
1272 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1273 Path *cheapest_path, Path *sorted_path,
1274 List *sort_pathkeys, List *group_pathkeys,
1275 double dNumGroups, AggClauseCounts *agg_counts)
1277 int numGroupCols = list_length(root->parse->groupClause);
1278 double cheapest_path_rows;
1279 int cheapest_path_width;
1281 List *current_pathkeys;
1286 * Check can't-do-it conditions, including whether the grouping operators
1289 * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1290 * (Doing so would imply storing *all* the input values in the hash table,
1291 * which seems like a certain loser.)
1293 if (!enable_hashagg)
1295 if (agg_counts->numDistinctAggs != 0)
1297 if (!hash_safe_grouping(root))
1301 * Don't do it if it doesn't look like the hashtable will fit into
1304 * Beware here of the possibility that cheapest_path->parent is NULL.
1305 * This could happen if user does something silly like
1306 * SELECT 'foo' GROUP BY 1;
1308 if (cheapest_path->parent)
1310 cheapest_path_rows = cheapest_path->parent->rows;
1311 cheapest_path_width = cheapest_path->parent->width;
1315 cheapest_path_rows = 1; /* assume non-set result */
1316 cheapest_path_width = 100; /* arbitrary */
1319 /* Estimate per-hash-entry space at tuple width... */
1320 hashentrysize = cheapest_path_width;
1321 /* plus space for pass-by-ref transition values... */
1322 hashentrysize += agg_counts->transitionSpace;
1323 /* plus the per-hash-entry overhead */
1324 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1326 if (hashentrysize * dNumGroups > work_mem * 1024L)
1330 * See if the estimated cost is no more than doing it the other way.
1331 * While avoiding the need for sorted input is usually a win, the fact
1332 * that the output won't be sorted may be a loss; so we need to do an
1333 * actual cost comparison.
1335 * We need to consider
1336 * cheapest_path + hashagg [+ final sort]
1338 * cheapest_path [+ sort] + group or agg [+ final sort]
1340 * presorted_path + group or agg [+ final sort]
1341 * where brackets indicate a step that may not be needed. We assume
1342 * query_planner() will have returned a presorted path only if it's a
1343 * winner compared to cheapest_path for this purpose.
1345 * These path variables are dummies that just hold cost fields; we don't
1346 * make actual Paths for these steps.
1348 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1349 numGroupCols, dNumGroups,
1350 cheapest_path->startup_cost, cheapest_path->total_cost,
1351 cheapest_path_rows);
1352 /* Result of hashed agg is always unsorted */
1354 cost_sort(&hashed_p, root, sort_pathkeys, hashed_p.total_cost,
1355 dNumGroups, cheapest_path_width);
1359 sorted_p.startup_cost = sorted_path->startup_cost;
1360 sorted_p.total_cost = sorted_path->total_cost;
1361 current_pathkeys = sorted_path->pathkeys;
1365 sorted_p.startup_cost = cheapest_path->startup_cost;
1366 sorted_p.total_cost = cheapest_path->total_cost;
1367 current_pathkeys = cheapest_path->pathkeys;
1369 if (!pathkeys_contained_in(group_pathkeys,
1372 cost_sort(&sorted_p, root, group_pathkeys, sorted_p.total_cost,
1373 cheapest_path_rows, cheapest_path_width);
1374 current_pathkeys = group_pathkeys;
1377 if (root->parse->hasAggs)
1378 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1379 numGroupCols, dNumGroups,
1380 sorted_p.startup_cost, sorted_p.total_cost,
1381 cheapest_path_rows);
1383 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1384 sorted_p.startup_cost, sorted_p.total_cost,
1385 cheapest_path_rows);
1386 /* The Agg or Group node will preserve ordering */
1387 if (sort_pathkeys &&
1388 !pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1389 cost_sort(&sorted_p, root, sort_pathkeys, sorted_p.total_cost,
1390 dNumGroups, cheapest_path_width);
1393 * Now make the decision using the top-level tuple fraction. First we
1394 * have to convert an absolute count (LIMIT) into fractional form.
1396 if (tuple_fraction >= 1.0)
1397 tuple_fraction /= dNumGroups;
1399 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1400 tuple_fraction) < 0)
1402 /* Hashed is cheaper, so use it */
1409 * hash_safe_grouping - are grouping operators hashable?
1411 * We assume hashed aggregation will work if the datatype's equality operator
1412 * is marked hashjoinable.
1415 hash_safe_grouping(PlannerInfo *root)
1419 foreach(gl, root->parse->groupClause)
1421 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1422 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1423 root->parse->targetList);
1427 optup = equality_oper(exprType((Node *) tle->expr), true);
1430 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1431 ReleaseSysCache(optup);
1439 * make_subplanTargetList
1440 * Generate appropriate target list when grouping is required.
1442 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1443 * above the result of query_planner, we typically want to pass a different
1444 * target list to query_planner than the outer plan nodes should have.
1445 * This routine generates the correct target list for the subplan.
1447 * The initial target list passed from the parser already contains entries
1448 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1449 * for variables used only in HAVING clauses; so we need to add those
1450 * variables to the subplan target list. Also, we flatten all expressions
1451 * except GROUP BY items into their component variables; the other expressions
1452 * will be computed by the inserted nodes rather than by the subplan.
1453 * For example, given a query like
1454 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1455 * we want to pass this targetlist to the subplan:
1457 * where the a+b target will be used by the Sort/Group steps, and the
1458 * other targets will be used for computing the final results. (In the
1459 * above example we could theoretically suppress the a and b targets and
1460 * pass down only c,d,a+b, but it's not really worth the trouble to
1461 * eliminate simple var references from the subplan. We will avoid doing
1462 * the extra computation to recompute a+b at the outer level; see
1463 * replace_vars_with_subplan_refs() in setrefs.c.)
1465 * If we are grouping or aggregating, *and* there are no non-Var grouping
1466 * expressions, then the returned tlist is effectively dummy; we do not
1467 * need to force it to be evaluated, because all the Vars it contains
1468 * should be present in the output of query_planner anyway.
1470 * 'tlist' is the query's target list.
1471 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1472 * expressions (if there are any) in the subplan's target list.
1473 * 'need_tlist_eval' is set true if we really need to evaluate the
1476 * The result is the targetlist to be passed to the subplan.
1480 make_subplanTargetList(PlannerInfo *root,
1482 AttrNumber **groupColIdx,
1483 bool *need_tlist_eval)
1485 Query *parse = root->parse;
1490 *groupColIdx = NULL;
1493 * If we're not grouping or aggregating, there's nothing to do here;
1494 * query_planner should receive the unmodified target list.
1496 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1498 *need_tlist_eval = true;
1503 * Otherwise, start with a "flattened" tlist (having just the vars
1504 * mentioned in the targetlist and HAVING qual --- but not upper-
1505 * level Vars; they will be replaced by Params later on).
1507 sub_tlist = flatten_tlist(tlist);
1508 extravars = pull_var_clause(parse->havingQual, false);
1509 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1510 list_free(extravars);
1511 *need_tlist_eval = false; /* only eval if not flat tlist */
1514 * If grouping, create sub_tlist entries for all GROUP BY expressions
1515 * (GROUP BY items that are simple Vars should be in the list
1516 * already), and make an array showing where the group columns are in
1519 numCols = list_length(parse->groupClause);
1523 AttrNumber *grpColIdx;
1526 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1527 *groupColIdx = grpColIdx;
1529 foreach(gl, parse->groupClause)
1531 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1532 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1533 TargetEntry *te = NULL;
1536 /* Find or make a matching sub_tlist entry */
1537 foreach(sl, sub_tlist)
1539 te = (TargetEntry *) lfirst(sl);
1540 if (equal(groupexpr, te->expr))
1545 te = makeTargetEntry((Expr *) groupexpr,
1546 list_length(sub_tlist) + 1,
1549 sub_tlist = lappend(sub_tlist, te);
1550 *need_tlist_eval = true; /* it's not flat anymore */
1553 /* and save its resno */
1554 grpColIdx[keyno++] = te->resno;
1562 * locate_grouping_columns
1563 * Locate grouping columns in the tlist chosen by query_planner.
1565 * This is only needed if we don't use the sub_tlist chosen by
1566 * make_subplanTargetList. We have to forget the column indexes found
1567 * by that routine and re-locate the grouping vars in the real sub_tlist.
1570 locate_grouping_columns(PlannerInfo *root,
1573 AttrNumber *groupColIdx)
1579 * No work unless grouping.
1581 if (!root->parse->groupClause)
1583 Assert(groupColIdx == NULL);
1586 Assert(groupColIdx != NULL);
1588 foreach(gl, root->parse->groupClause)
1590 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1591 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1592 TargetEntry *te = NULL;
1595 foreach(sl, sub_tlist)
1597 te = (TargetEntry *) lfirst(sl);
1598 if (equal(groupexpr, te->expr))
1602 elog(ERROR, "failed to locate grouping columns");
1604 groupColIdx[keyno++] = te->resno;
1609 * postprocess_setop_tlist
1610 * Fix up targetlist returned by plan_set_operations().
1612 * We need to transpose sort key info from the orig_tlist into new_tlist.
1613 * NOTE: this would not be good enough if we supported resjunk sort keys
1614 * for results of set operations --- then, we'd need to project a whole
1615 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1616 * find any resjunk columns in orig_tlist.
1619 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1622 ListCell *orig_tlist_item = list_head(orig_tlist);
1624 foreach(l, new_tlist)
1626 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1627 TargetEntry *orig_tle;
1629 /* ignore resjunk columns in setop result */
1630 if (new_tle->resjunk)
1633 Assert(orig_tlist_item != NULL);
1634 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1635 orig_tlist_item = lnext(orig_tlist_item);
1636 if (orig_tle->resjunk) /* should not happen */
1637 elog(ERROR, "resjunk output columns are not implemented");
1638 Assert(new_tle->resno == orig_tle->resno);
1639 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1641 if (orig_tlist_item != NULL)
1642 elog(ERROR, "resjunk output columns are not implemented");