1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.154 2003/06/06 15:04:02 tgl Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "catalog/pg_type.h"
22 #include "executor/executor.h"
23 #include "miscadmin.h"
24 #include "nodes/makefuncs.h"
25 #ifdef OPTIMIZER_DEBUG
26 #include "nodes/print.h"
28 #include "optimizer/clauses.h"
29 #include "optimizer/cost.h"
30 #include "optimizer/pathnode.h"
31 #include "optimizer/paths.h"
32 #include "optimizer/planmain.h"
33 #include "optimizer/planner.h"
34 #include "optimizer/prep.h"
35 #include "optimizer/subselect.h"
36 #include "optimizer/tlist.h"
37 #include "optimizer/var.h"
38 #include "parser/analyze.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 /* Expression kind codes for preprocess_expression */
47 #define EXPRKIND_QUAL 0
48 #define EXPRKIND_TARGET 1
49 #define EXPRKIND_RTFUNC 2
50 #define EXPRKIND_ININFO 3
53 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
54 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
55 static Plan *inheritance_planner(Query *parse, List *inheritlist);
56 static Plan *grouping_planner(Query *parse, double tuple_fraction);
57 static bool hash_safe_grouping(Query *parse);
58 static List *make_subplanTargetList(Query *parse, List *tlist,
59 AttrNumber **groupColIdx, bool *need_tlist_eval);
60 static void locate_grouping_columns(Query *parse,
63 AttrNumber *groupColIdx);
64 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
67 /*****************************************************************************
69 * Query optimizer entry point
71 *****************************************************************************/
73 planner(Query *parse, bool isCursor, int cursorOptions)
75 double tuple_fraction;
77 Index save_PlannerQueryLevel;
78 List *save_PlannerParamList;
81 * The planner can be called recursively (an example is when
82 * eval_const_expressions tries to pre-evaluate an SQL function). So,
83 * these global state variables must be saved and restored.
85 * These vars cannot be moved into the Query structure since their whole
86 * purpose is communication across multiple sub-Queries.
88 * Note we do NOT save and restore PlannerPlanId: it exists to assign
89 * unique IDs to SubPlan nodes, and we want those IDs to be unique for
90 * the life of a backend. Also, PlannerInitPlan is saved/restored in
91 * subquery_planner, not here.
93 save_PlannerQueryLevel = PlannerQueryLevel;
94 save_PlannerParamList = PlannerParamList;
96 /* Initialize state for handling outer-level references and params */
97 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
98 PlannerParamList = NIL;
100 /* Determine what fraction of the plan is likely to be scanned */
104 * We have no real idea how many tuples the user will ultimately
105 * FETCH from a cursor, but it seems a good bet that he
106 * doesn't want 'em all. Optimize for 10% retrieval (you
107 * gotta better number? Should this be a SETtable parameter?)
109 tuple_fraction = 0.10;
113 /* Default assumption is we need all the tuples */
114 tuple_fraction = 0.0;
117 /* primary planning entry point (may recurse for subqueries) */
118 result_plan = subquery_planner(parse, tuple_fraction);
120 Assert(PlannerQueryLevel == 0);
123 * If creating a plan for a scrollable cursor, make sure it can
124 * run backwards on demand. Add a Material node at the top at need.
126 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
128 if (!ExecSupportsBackwardScan(result_plan))
129 result_plan = materialize_finished_plan(result_plan);
132 /* executor wants to know total number of Params used overall */
133 result_plan->nParamExec = length(PlannerParamList);
135 /* final cleanup of the plan */
136 set_plan_references(result_plan, parse->rtable);
138 /* restore state for outer planner, if any */
139 PlannerQueryLevel = save_PlannerQueryLevel;
140 PlannerParamList = save_PlannerParamList;
146 /*--------------------
148 * Invokes the planner on a subquery. We recurse to here for each
149 * sub-SELECT found in the query tree.
151 * parse is the querytree produced by the parser & rewriter.
152 * tuple_fraction is the fraction of tuples we expect will be retrieved.
153 * tuple_fraction is interpreted as explained for grouping_planner, below.
155 * Basically, this routine does the stuff that should only be done once
156 * per Query object. It then calls grouping_planner. At one time,
157 * grouping_planner could be invoked recursively on the same Query object;
158 * that's not currently true, but we keep the separation between the two
159 * routines anyway, in case we need it again someday.
161 * subquery_planner will be called recursively to handle sub-Query nodes
162 * found within the query's expressions and rangetable.
164 * Returns a query plan.
165 *--------------------
168 subquery_planner(Query *parse, double tuple_fraction)
170 List *saved_initplan = PlannerInitPlan;
171 int saved_planid = PlannerPlanId;
177 /* Set up for a new level of subquery */
179 PlannerInitPlan = NIL;
182 * Look for IN clauses at the top level of WHERE, and transform them
183 * into joins. Note that this step only handles IN clauses originally
184 * at top level of WHERE; if we pull up any subqueries in the next step,
185 * their INs are processed just before pulling them up.
187 parse->in_info_list = NIL;
188 if (parse->hasSubLinks)
189 parse->jointree->quals = pull_up_IN_clauses(parse,
190 parse->jointree->quals);
193 * Check to see if any subqueries in the rangetable can be merged into
196 parse->jointree = (FromExpr *)
197 pull_up_subqueries(parse, (Node *) parse->jointree, false);
200 * Detect whether any rangetable entries are RTE_JOIN kind; if not,
201 * we can avoid the expense of doing flatten_join_alias_vars(). Also
202 * check for outer joins --- if none, we can skip reduce_outer_joins().
203 * This must be done after we have done pull_up_subqueries, of course.
205 parse->hasJoinRTEs = false;
206 hasOuterJoins = false;
207 foreach(lst, parse->rtable)
209 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
211 if (rte->rtekind == RTE_JOIN)
213 parse->hasJoinRTEs = true;
214 if (IS_OUTER_JOIN(rte->jointype))
216 hasOuterJoins = true;
217 /* Can quit scanning once we find an outer join */
224 * Do expression preprocessing on targetlist and quals.
226 parse->targetList = (List *)
227 preprocess_expression(parse, (Node *) parse->targetList,
230 preprocess_qual_conditions(parse, (Node *) parse->jointree);
232 parse->havingQual = preprocess_expression(parse, parse->havingQual,
235 parse->in_info_list = (List *)
236 preprocess_expression(parse, (Node *) parse->in_info_list,
239 /* Also need to preprocess expressions for function RTEs */
240 foreach(lst, parse->rtable)
242 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
244 if (rte->rtekind == RTE_FUNCTION)
245 rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
250 * A HAVING clause without aggregates is equivalent to a WHERE clause
251 * (except it can only refer to grouped fields). Transfer any
252 * agg-free clauses of the HAVING qual into WHERE. This may seem like
253 * wasting cycles to cater to stupidly-written queries, but there are
254 * other reasons for doing it. Firstly, if the query contains no aggs
255 * at all, then we aren't going to generate an Agg plan node, and so
256 * there'll be no place to execute HAVING conditions; without this
257 * transfer, we'd lose the HAVING condition entirely, which is wrong.
258 * Secondly, when we push down a qual condition into a sub-query, it's
259 * easiest to push the qual into HAVING always, in case it contains
260 * aggs, and then let this code sort it out.
262 * Note that both havingQual and parse->jointree->quals are in
263 * implicitly-ANDed-list form at this point, even though they are
264 * declared as Node *.
267 foreach(lst, (List *) parse->havingQual)
269 Node *havingclause = (Node *) lfirst(lst);
271 if (contain_agg_clause(havingclause))
272 newHaving = lappend(newHaving, havingclause);
274 parse->jointree->quals = (Node *)
275 lappend((List *) parse->jointree->quals, havingclause);
277 parse->havingQual = (Node *) newHaving;
280 * If we have any outer joins, try to reduce them to plain inner joins.
281 * This step is most easily done after we've done expression preprocessing.
284 reduce_outer_joins(parse);
287 * See if we can simplify the jointree; opportunities for this may come
288 * from having pulled up subqueries, or from flattening explicit JOIN
289 * syntax. We must do this after flattening JOIN alias variables, since
290 * eliminating explicit JOIN nodes from the jointree will cause
291 * get_relids_for_join() to fail. But it should happen after
292 * reduce_outer_joins, anyway.
294 parse->jointree = (FromExpr *)
295 simplify_jointree(parse, (Node *) parse->jointree);
298 * Do the main planning. If we have an inherited target relation,
299 * that needs special processing, else go straight to
302 if (parse->resultRelation &&
303 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
305 plan = inheritance_planner(parse, lst);
307 plan = grouping_planner(parse, tuple_fraction);
310 * If any subplans were generated, or if we're inside a subplan, build
311 * initPlan list and extParam/allParam sets for plan nodes.
313 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
315 Cost initplan_cost = 0;
317 /* Prepare extParam/allParam sets for all nodes in tree */
318 SS_finalize_plan(plan, parse->rtable);
321 * SS_finalize_plan doesn't handle initPlans, so we have to manually
322 * attach them to the topmost plan node, and add their extParams to
323 * the topmost node's, too.
325 * We also add the total_cost of each initPlan to the startup cost
326 * of the top node. This is a conservative overestimate, since in
327 * fact each initPlan might be executed later than plan startup, or
330 plan->initPlan = PlannerInitPlan;
332 foreach(lst, plan->initPlan)
334 SubPlan *initplan = (SubPlan *) lfirst(lst);
336 plan->extParam = bms_add_members(plan->extParam,
337 initplan->plan->extParam);
338 initplan_cost += initplan->plan->total_cost;
341 plan->startup_cost += initplan_cost;
342 plan->total_cost += initplan_cost;
345 /* Return to outer subquery context */
347 PlannerInitPlan = saved_initplan;
348 /* we do NOT restore PlannerPlanId; that's not an oversight! */
354 * preprocess_expression
355 * Do subquery_planner's preprocessing work for an expression,
356 * which can be a targetlist, a WHERE clause (including JOIN/ON
357 * conditions), or a HAVING clause.
360 preprocess_expression(Query *parse, Node *expr, int kind)
363 * If the query has any join RTEs, replace join alias variables with
364 * base-relation variables. We must do this before sublink processing,
365 * else sublinks expanded out from join aliases wouldn't get processed.
367 if (parse->hasJoinRTEs)
368 expr = flatten_join_alias_vars(parse, expr);
371 * Simplify constant expressions.
373 * Note that at this point quals have not yet been converted to
374 * implicit-AND form, so we can apply eval_const_expressions directly.
376 expr = eval_const_expressions(expr);
379 * If it's a qual or havingQual, canonicalize it, and convert it to
380 * implicit-AND format.
382 * XXX Is there any value in re-applying eval_const_expressions after
385 if (kind == EXPRKIND_QUAL)
387 expr = (Node *) canonicalize_qual((Expr *) expr, true);
389 #ifdef OPTIMIZER_DEBUG
390 printf("After canonicalize_qual()\n");
395 /* Expand SubLinks to SubPlans */
396 if (parse->hasSubLinks)
397 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
400 * XXX do not insert anything here unless you have grokked the comments
401 * in SS_replace_correlation_vars ...
404 /* Replace uplevel vars with Param nodes */
405 if (PlannerQueryLevel > 1)
406 expr = SS_replace_correlation_vars(expr);
412 * preprocess_qual_conditions
413 * Recursively scan the query's jointree and do subquery_planner's
414 * preprocessing work on each qual condition found therein.
417 preprocess_qual_conditions(Query *parse, Node *jtnode)
421 if (IsA(jtnode, RangeTblRef))
423 /* nothing to do here */
425 else if (IsA(jtnode, FromExpr))
427 FromExpr *f = (FromExpr *) jtnode;
430 foreach(l, f->fromlist)
431 preprocess_qual_conditions(parse, lfirst(l));
433 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
435 else if (IsA(jtnode, JoinExpr))
437 JoinExpr *j = (JoinExpr *) jtnode;
439 preprocess_qual_conditions(parse, j->larg);
440 preprocess_qual_conditions(parse, j->rarg);
442 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
445 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
449 /*--------------------
450 * inheritance_planner
451 * Generate a plan in the case where the result relation is an
454 * We have to handle this case differently from cases where a source
455 * relation is an inheritance set. Source inheritance is expanded at
456 * the bottom of the plan tree (see allpaths.c), but target inheritance
457 * has to be expanded at the top. The reason is that for UPDATE, each
458 * target relation needs a different targetlist matching its own column
459 * set. (This is not so critical for DELETE, but for simplicity we treat
460 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
461 * can never be the nullable side of an outer join, so it's OK to generate
464 * parse is the querytree produced by the parser & rewriter.
465 * inheritlist is an integer list of RT indexes for the result relation set.
467 * Returns a query plan.
468 *--------------------
471 inheritance_planner(Query *parse, List *inheritlist)
473 int parentRTindex = parse->resultRelation;
474 Oid parentOID = getrelid(parentRTindex, parse->rtable);
475 int mainrtlength = length(parse->rtable);
476 List *subplans = NIL;
480 foreach(l, inheritlist)
482 int childRTindex = lfirsti(l);
483 Oid childOID = getrelid(childRTindex, parse->rtable);
488 /* Generate modified query with this rel as target */
489 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
490 parentRTindex, parentOID,
491 childRTindex, childOID);
493 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
494 subplans = lappend(subplans, subplan);
496 * It's possible that additional RTEs got added to the rangetable
497 * due to expansion of inherited source tables (see allpaths.c).
498 * If so, we must copy 'em back to the main parse tree's rtable.
500 * XXX my goodness this is ugly. Really need to think about ways
501 * to rein in planner's habit of scribbling on its input.
503 subrtlength = length(subquery->rtable);
504 if (subrtlength > mainrtlength)
506 List *subrt = subquery->rtable;
508 while (mainrtlength-- > 0) /* wish we had nthcdr() */
509 subrt = lnext(subrt);
510 parse->rtable = nconc(parse->rtable, subrt);
511 mainrtlength = subrtlength;
513 /* Save preprocessed tlist from first rel for use in Append */
515 tlist = subplan->targetlist;
518 /* Save the target-relations list for the executor, too */
519 parse->resultRelations = inheritlist;
521 /* Mark result as unordered (probably unnecessary) */
522 parse->query_pathkeys = NIL;
524 return (Plan *) make_append(subplans, true, tlist);
527 /*--------------------
529 * Perform planning steps related to grouping, aggregation, etc.
530 * This primarily means adding top-level processing to the basic
531 * query plan produced by query_planner.
533 * parse is the querytree produced by the parser & rewriter.
534 * tuple_fraction is the fraction of tuples we expect will be retrieved
536 * tuple_fraction is interpreted as follows:
537 * 0: expect all tuples to be retrieved (normal case)
538 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
539 * from the plan to be retrieved
540 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
541 * expected to be retrieved (ie, a LIMIT specification)
543 * Returns a query plan. Also, parse->query_pathkeys is returned as the
544 * actual output ordering of the plan (in pathkey format).
545 *--------------------
548 grouping_planner(Query *parse, double tuple_fraction)
550 List *tlist = parse->targetList;
552 List *current_pathkeys;
555 if (parse->setOperations)
558 * Construct the plan for set operations. The result will not
559 * need any work except perhaps a top-level sort and/or LIMIT.
561 result_plan = plan_set_operations(parse);
564 * We should not need to call preprocess_targetlist, since we must
565 * be in a SELECT query node. Instead, use the targetlist
566 * returned by plan_set_operations (since this tells whether it
567 * returned any resjunk columns!), and transfer any sort key
568 * information from the original tlist.
570 Assert(parse->commandType == CMD_SELECT);
572 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
575 * Can't handle FOR UPDATE here (parser should have checked
576 * already, but let's make sure).
579 elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
582 * We set current_pathkeys NIL indicating we do not know sort
583 * order. This is correct when the top set operation is UNION
584 * ALL, since the appended-together results are unsorted even if
585 * the subplans were sorted. For other set operations we could be
586 * smarter --- room for future improvement!
588 current_pathkeys = NIL;
591 * Calculate pathkeys that represent ordering requirements
593 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
595 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
599 /* No set operations, do regular planning */
601 List *group_pathkeys;
602 AttrNumber *groupColIdx = NULL;
603 bool need_tlist_eval = true;
605 double sub_tuple_fraction;
608 double dNumGroups = 0;
611 int numGroupCols = length(parse->groupClause);
612 bool use_hashed_grouping = false;
614 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
615 tlist = preprocess_targetlist(tlist,
617 parse->resultRelation,
621 * Add TID targets for rels selected FOR UPDATE (should this be
622 * done in preprocess_targetlist?). The executor uses the TID to
623 * know which rows to lock, much as for UPDATE or DELETE.
630 * We've got trouble if the FOR UPDATE appears inside
631 * grouping, since grouping renders a reference to individual
632 * tuple CTIDs invalid. This is also checked at parse time,
633 * but that's insufficient because of rule substitution, query
636 CheckSelectForUpdate(parse);
639 * Currently the executor only supports FOR UPDATE at top
642 if (PlannerQueryLevel > 1)
643 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
645 foreach(l, parse->rowMarks)
647 Index rti = lfirsti(l);
653 resname = (char *) palloc(32);
654 snprintf(resname, 32, "ctid%u", rti);
655 resdom = makeResdom(length(tlist) + 1,
662 SelfItemPointerAttributeNumber,
667 ctid = makeTargetEntry(resdom, (Expr *) var);
668 tlist = lappend(tlist, ctid);
673 * Generate appropriate target list for subplan; may be different
674 * from tlist if grouping or aggregation is needed.
676 sub_tlist = make_subplanTargetList(parse, tlist,
677 &groupColIdx, &need_tlist_eval);
680 * Calculate pathkeys that represent grouping/ordering
683 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
685 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
689 * Will need actual number of aggregates for estimating costs.
690 * Also, it's possible that optimization has eliminated all
691 * aggregates, and we may as well check for that here.
693 * Note: we do not attempt to detect duplicate aggregates here;
694 * a somewhat-overestimated count is okay for our present purposes.
698 numAggs = count_agg_clause((Node *) tlist) +
699 count_agg_clause(parse->havingQual);
701 parse->hasAggs = false;
705 * Figure out whether we need a sorted result from query_planner.
707 * If we have a GROUP BY clause, then we want a result sorted
708 * properly for grouping. Otherwise, if there is an ORDER BY
709 * clause, we want to sort by the ORDER BY clause. (Note: if we
710 * have both, and ORDER BY is a superset of GROUP BY, it would be
711 * tempting to request sort by ORDER BY --- but that might just
712 * leave us failing to exploit an available sort order at all.
713 * Needs more thought...)
715 if (parse->groupClause)
716 parse->query_pathkeys = group_pathkeys;
717 else if (parse->sortClause)
718 parse->query_pathkeys = sort_pathkeys;
720 parse->query_pathkeys = NIL;
723 * Adjust tuple_fraction if we see that we are going to apply
724 * limiting/grouping/aggregation/etc. This is not overridable by
725 * the caller, since it reflects plan actions that this routine
726 * will certainly take, not assumptions about context.
728 if (parse->limitCount != NULL)
731 * A LIMIT clause limits the absolute number of tuples
732 * returned. However, if it's not a constant LIMIT then we
733 * have to punt; for lack of a better idea, assume 10% of the
734 * plan's result is wanted.
736 double limit_fraction = 0.0;
738 if (IsA(parse->limitCount, Const))
740 Const *limitc = (Const *) parse->limitCount;
741 int32 count = DatumGetInt32(limitc->constvalue);
744 * A NULL-constant LIMIT represents "LIMIT ALL", which we
745 * treat the same as no limit (ie, expect to retrieve all
748 if (!limitc->constisnull && count > 0)
750 limit_fraction = (double) count;
751 /* We must also consider the OFFSET, if present */
752 if (parse->limitOffset != NULL)
754 if (IsA(parse->limitOffset, Const))
758 limitc = (Const *) parse->limitOffset;
759 offset = DatumGetInt32(limitc->constvalue);
760 if (!limitc->constisnull && offset > 0)
761 limit_fraction += (double) offset;
765 /* OFFSET is an expression ... punt ... */
766 limit_fraction = 0.10;
773 /* LIMIT is an expression ... punt ... */
774 limit_fraction = 0.10;
777 if (limit_fraction > 0.0)
780 * If we have absolute limits from both caller and LIMIT,
781 * use the smaller value; if one is fractional and the
782 * other absolute, treat the fraction as a fraction of the
783 * absolute value; else we can multiply the two fractions
786 if (tuple_fraction >= 1.0)
788 if (limit_fraction >= 1.0)
791 tuple_fraction = Min(tuple_fraction, limit_fraction);
795 /* caller absolute, limit fractional */
796 tuple_fraction *= limit_fraction;
797 if (tuple_fraction < 1.0)
798 tuple_fraction = 1.0;
801 else if (tuple_fraction > 0.0)
803 if (limit_fraction >= 1.0)
805 /* caller fractional, limit absolute */
806 tuple_fraction *= limit_fraction;
807 if (tuple_fraction < 1.0)
808 tuple_fraction = 1.0;
812 /* both fractional */
813 tuple_fraction *= limit_fraction;
818 /* no info from caller, just use limit */
819 tuple_fraction = limit_fraction;
825 * With grouping or aggregation, the tuple fraction to pass to
826 * query_planner() may be different from what it is at top level.
828 sub_tuple_fraction = tuple_fraction;
830 if (parse->groupClause)
833 * In GROUP BY mode, we have the little problem that we don't
834 * really know how many input tuples will be needed to make a
835 * group, so we can't translate an output LIMIT count into an
836 * input count. For lack of a better idea, assume 25% of the
837 * input data will be processed if there is any output limit.
838 * However, if the caller gave us a fraction rather than an
839 * absolute count, we can keep using that fraction (which
840 * amounts to assuming that all the groups are about the same
843 if (sub_tuple_fraction >= 1.0)
844 sub_tuple_fraction = 0.25;
847 * If both GROUP BY and ORDER BY are specified, we will need
848 * two levels of sort --- and, therefore, certainly need to
849 * read all the input tuples --- unless ORDER BY is a subset
850 * of GROUP BY. (We have not yet canonicalized the pathkeys,
851 * so must use the slower noncanonical comparison method.)
853 if (parse->groupClause && parse->sortClause &&
854 !noncanonical_pathkeys_contained_in(sort_pathkeys,
856 sub_tuple_fraction = 0.0;
858 else if (parse->hasAggs)
861 * Ungrouped aggregate will certainly want all the input
864 sub_tuple_fraction = 0.0;
866 else if (parse->distinctClause)
869 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
870 * number of input tuples per output tuple. Handle the same
873 if (sub_tuple_fraction >= 1.0)
874 sub_tuple_fraction = 0.25;
878 * Generate the best unsorted and presorted paths for this Query
879 * (but note there may not be any presorted path).
881 query_planner(parse, sub_tlist, sub_tuple_fraction,
882 &cheapest_path, &sorted_path);
885 * We couldn't canonicalize group_pathkeys and sort_pathkeys before
886 * running query_planner(), so do it now.
888 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
889 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
892 * Consider whether we might want to use hashed grouping.
894 if (parse->groupClause)
899 * Always estimate the number of groups. We can't do this until
900 * after running query_planner(), either.
902 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
904 dNumGroups = estimate_num_groups(parse,
906 cheapest_path->parent->rows);
907 /* Also want it as a long int --- but 'ware overflow! */
908 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
911 * Check can't-do-it conditions, including whether the grouping
912 * operators are hashjoinable.
914 * Executor doesn't support hashed aggregation with DISTINCT
915 * aggregates. (Doing so would imply storing *all* the input
916 * values in the hash table, which seems like a certain loser.)
918 if (!enable_hashagg || !hash_safe_grouping(parse))
919 use_hashed_grouping = false;
920 else if (parse->hasAggs &&
921 (contain_distinct_agg_clause((Node *) tlist) ||
922 contain_distinct_agg_clause(parse->havingQual)))
923 use_hashed_grouping = false;
927 * Use hashed grouping if (a) we think we can fit the
928 * hashtable into SortMem, *and* (b) the estimated cost
929 * is no more than doing it the other way. While avoiding
930 * the need for sorted input is usually a win, the fact
931 * that the output won't be sorted may be a loss; so we
932 * need to do an actual cost comparison.
934 * In most cases we have no good way to estimate the size of
935 * the transition value needed by an aggregate; arbitrarily
936 * assume it is 100 bytes. Also set the overhead per hashtable
939 int hashentrysize = cheapest_path->parent->width + 64 +
942 if (hashentrysize * dNumGroups <= SortMem * 1024L)
945 * Okay, do the cost comparison. We need to consider
946 * cheapest_path + hashagg [+ final sort]
948 * cheapest_path [+ sort] + group or agg [+ final sort]
950 * presorted_path + group or agg [+ final sort]
951 * where brackets indicate a step that may not be needed.
952 * We assume query_planner() will have returned a
953 * presorted path only if it's a winner compared to
954 * cheapest_path for this purpose.
956 * These path variables are dummies that just hold cost
957 * fields; we don't make actual Paths for these steps.
962 cost_agg(&hashed_p, parse,
964 numGroupCols, dNumGroups,
965 cheapest_path->startup_cost,
966 cheapest_path->total_cost,
967 cheapest_path->parent->rows);
968 /* Result of hashed agg is always unsorted */
970 cost_sort(&hashed_p, parse, sort_pathkeys,
973 cheapest_path->parent->width);
977 sorted_p.startup_cost = sorted_path->startup_cost;
978 sorted_p.total_cost = sorted_path->total_cost;
979 current_pathkeys = sorted_path->pathkeys;
983 sorted_p.startup_cost = cheapest_path->startup_cost;
984 sorted_p.total_cost = cheapest_path->total_cost;
985 current_pathkeys = cheapest_path->pathkeys;
987 if (!pathkeys_contained_in(group_pathkeys,
990 cost_sort(&sorted_p, parse, group_pathkeys,
992 cheapest_path->parent->rows,
993 cheapest_path->parent->width);
994 current_pathkeys = group_pathkeys;
997 cost_agg(&sorted_p, parse,
999 numGroupCols, dNumGroups,
1000 sorted_p.startup_cost,
1001 sorted_p.total_cost,
1002 cheapest_path->parent->rows);
1004 cost_group(&sorted_p, parse,
1005 numGroupCols, dNumGroups,
1006 sorted_p.startup_cost,
1007 sorted_p.total_cost,
1008 cheapest_path->parent->rows);
1009 /* The Agg or Group node will preserve ordering */
1010 if (sort_pathkeys &&
1011 !pathkeys_contained_in(sort_pathkeys,
1014 cost_sort(&sorted_p, parse, sort_pathkeys,
1015 sorted_p.total_cost,
1017 cheapest_path->parent->width);
1021 * Now make the decision using the top-level tuple
1022 * fraction. First we have to convert an absolute
1023 * count (LIMIT) into fractional form.
1025 if (tuple_fraction >= 1.0)
1026 tuple_fraction /= dNumGroups;
1028 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1029 tuple_fraction) < 0)
1031 /* Hashed is cheaper, so use it */
1032 use_hashed_grouping = true;
1039 * Select the best path and create a plan to execute it.
1041 * If we are doing hashed grouping, we will always read all the
1042 * input tuples, so use the cheapest-total path. Otherwise,
1043 * trust query_planner's decision about which to use.
1045 if (sorted_path && !use_hashed_grouping)
1047 result_plan = create_plan(parse, sorted_path);
1048 current_pathkeys = sorted_path->pathkeys;
1052 result_plan = create_plan(parse, cheapest_path);
1053 current_pathkeys = cheapest_path->pathkeys;
1057 * create_plan() returns a plan with just a "flat" tlist of required
1058 * Vars. Usually we need to insert the sub_tlist as the tlist of the
1059 * top plan node. However, we can skip that if we determined that
1060 * whatever query_planner chose to return will be good enough.
1062 if (need_tlist_eval)
1065 * If the top-level plan node is one that cannot do expression
1066 * evaluation, we must insert a Result node to project the desired
1068 * Currently, the only plan node we might see here that falls into
1069 * that category is Append.
1071 if (IsA(result_plan, Append))
1073 result_plan = (Plan *) make_result(sub_tlist, NULL,
1079 * Otherwise, just replace the subplan's flat tlist with
1080 * the desired tlist.
1082 result_plan->targetlist = sub_tlist;
1085 * Also, account for the cost of evaluation of the sub_tlist.
1087 * Up to now, we have only been dealing with "flat" tlists,
1088 * containing just Vars. So their evaluation cost is zero
1089 * according to the model used by cost_qual_eval() (or if you
1090 * prefer, the cost is factored into cpu_tuple_cost). Thus we can
1091 * avoid accounting for tlist cost throughout query_planner() and
1092 * subroutines. But now we've inserted a tlist that might contain
1093 * actual operators, sub-selects, etc --- so we'd better account
1096 * Below this point, any tlist eval cost for added-on nodes should
1097 * be accounted for as we create those nodes. Presently, of the
1098 * node types we can add on, only Agg and Group project new tlists
1099 * (the rest just copy their input tuples) --- so make_agg() and
1100 * make_group() are responsible for computing the added cost.
1102 cost_qual_eval(&tlist_cost, sub_tlist);
1103 result_plan->startup_cost += tlist_cost.startup;
1104 result_plan->total_cost += tlist_cost.startup +
1105 tlist_cost.per_tuple * result_plan->plan_rows;
1110 * Since we're using query_planner's tlist and not the one
1111 * make_subplanTargetList calculated, we have to refigure
1112 * any grouping-column indexes make_subplanTargetList computed.
1114 locate_grouping_columns(parse, tlist, result_plan->targetlist,
1119 * Insert AGG or GROUP node if needed, plus an explicit sort step
1122 * HAVING clause, if any, becomes qual of the Agg node
1124 if (use_hashed_grouping)
1126 /* Hashed aggregate plan --- no sort needed */
1127 result_plan = (Plan *) make_agg(parse,
1129 (List *) parse->havingQual,
1136 /* Hashed aggregation produces randomly-ordered results */
1137 current_pathkeys = NIL;
1139 else if (parse->hasAggs)
1141 /* Plain aggregate plan --- sort if needed */
1142 AggStrategy aggstrategy;
1144 if (parse->groupClause)
1146 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1148 result_plan = (Plan *)
1149 make_sort_from_groupcols(parse,
1153 current_pathkeys = group_pathkeys;
1155 aggstrategy = AGG_SORTED;
1157 * The AGG node will not change the sort ordering of its
1158 * groups, so current_pathkeys describes the result too.
1163 aggstrategy = AGG_PLAIN;
1164 /* Result will be only one row anyway; no sort order */
1165 current_pathkeys = NIL;
1168 result_plan = (Plan *) make_agg(parse,
1170 (List *) parse->havingQual,
1181 * If there are no Aggs, we shouldn't have any HAVING qual anymore
1183 Assert(parse->havingQual == NULL);
1186 * If we have a GROUP BY clause, insert a group node (plus the
1187 * appropriate sort node, if necessary).
1189 if (parse->groupClause)
1192 * Add an explicit sort if we couldn't make the path come out
1193 * the way the GROUP node needs it.
1195 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1197 result_plan = (Plan *)
1198 make_sort_from_groupcols(parse,
1202 current_pathkeys = group_pathkeys;
1205 result_plan = (Plan *) make_group(parse,
1211 /* The Group node won't change sort ordering */
1214 } /* end of if (setOperations) */
1217 * If we were not able to make the plan come out in the right order,
1218 * add an explicit sort step.
1220 if (parse->sortClause)
1222 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1224 result_plan = (Plan *)
1225 make_sort_from_sortclauses(parse,
1229 current_pathkeys = sort_pathkeys;
1234 * If there is a DISTINCT clause, add the UNIQUE node.
1236 if (parse->distinctClause)
1238 result_plan = (Plan *) make_unique(tlist, result_plan,
1239 parse->distinctClause);
1241 * If there was grouping or aggregation, leave plan_rows as-is
1242 * (ie, assume the result was already mostly unique). If not,
1243 * it's reasonable to assume the UNIQUE filter has effects
1244 * comparable to GROUP BY.
1246 if (!parse->groupClause && !parse->hasAggs)
1248 List *distinctExprs;
1250 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1252 result_plan->plan_rows = estimate_num_groups(parse,
1254 result_plan->plan_rows);
1259 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1261 if (parse->limitOffset || parse->limitCount)
1263 result_plan = (Plan *) make_limit(tlist, result_plan,
1269 * Return the actual output ordering in query_pathkeys for possible
1270 * use by an outer query level.
1272 parse->query_pathkeys = current_pathkeys;
1278 * hash_safe_grouping - are grouping operators hashable?
1280 * We assume hashed aggregation will work if the datatype's equality operator
1281 * is marked hashjoinable.
1284 hash_safe_grouping(Query *parse)
1288 foreach(gl, parse->groupClause)
1290 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1291 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1295 optup = equality_oper(tle->resdom->restype, false);
1296 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1297 ReleaseSysCache(optup);
1305 * make_subplanTargetList
1306 * Generate appropriate target list when grouping is required.
1308 * When grouping_planner inserts Aggregate or Group plan nodes above
1309 * the result of query_planner, we typically want to pass a different
1310 * target list to query_planner than the outer plan nodes should have.
1311 * This routine generates the correct target list for the subplan.
1313 * The initial target list passed from the parser already contains entries
1314 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1315 * for variables used only in HAVING clauses; so we need to add those
1316 * variables to the subplan target list. Also, if we are doing either
1317 * grouping or aggregation, we flatten all expressions except GROUP BY items
1318 * into their component variables; the other expressions will be computed by
1319 * the inserted nodes rather than by the subplan. For example,
1320 * given a query like
1321 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1322 * we want to pass this targetlist to the subplan:
1324 * where the a+b target will be used by the Sort/Group steps, and the
1325 * other targets will be used for computing the final results. (In the
1326 * above example we could theoretically suppress the a and b targets and
1327 * pass down only c,d,a+b, but it's not really worth the trouble to
1328 * eliminate simple var references from the subplan. We will avoid doing
1329 * the extra computation to recompute a+b at the outer level; see
1330 * replace_vars_with_subplan_refs() in setrefs.c.)
1332 * If we are grouping or aggregating, *and* there are no non-Var grouping
1333 * expressions, then the returned tlist is effectively dummy; we do not
1334 * need to force it to be evaluated, because all the Vars it contains
1335 * should be present in the output of query_planner anyway.
1337 * 'parse' is the query being processed.
1338 * 'tlist' is the query's target list.
1339 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1340 * expressions (if there are any) in the subplan's target list.
1341 * 'need_tlist_eval' is set true if we really need to evaluate the
1344 * The result is the targetlist to be passed to the subplan.
1348 make_subplanTargetList(Query *parse,
1350 AttrNumber **groupColIdx,
1351 bool *need_tlist_eval)
1357 *groupColIdx = NULL;
1360 * If we're not grouping or aggregating, nothing to do here;
1361 * query_planner should receive the unmodified target list.
1363 if (!parse->hasAggs && !parse->groupClause)
1365 *need_tlist_eval = true;
1370 * Otherwise, start with a "flattened" tlist (having just the vars
1371 * mentioned in the targetlist and HAVING qual --- but not upper-
1372 * level Vars; they will be replaced by Params later on).
1374 sub_tlist = flatten_tlist(tlist);
1375 extravars = pull_var_clause(parse->havingQual, false);
1376 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1377 freeList(extravars);
1378 *need_tlist_eval = false; /* only eval if not flat tlist */
1381 * If grouping, create sub_tlist entries for all GROUP BY expressions
1382 * (GROUP BY items that are simple Vars should be in the list
1383 * already), and make an array showing where the group columns are in
1386 numCols = length(parse->groupClause);
1390 AttrNumber *grpColIdx;
1393 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1394 *groupColIdx = grpColIdx;
1396 foreach(gl, parse->groupClause)
1398 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1399 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1400 TargetEntry *te = NULL;
1403 /* Find or make a matching sub_tlist entry */
1404 foreach(sl, sub_tlist)
1406 te = (TargetEntry *) lfirst(sl);
1407 if (equal(groupexpr, te->expr))
1412 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1413 exprType(groupexpr),
1414 exprTypmod(groupexpr),
1417 (Expr *) groupexpr);
1418 sub_tlist = lappend(sub_tlist, te);
1419 *need_tlist_eval = true; /* it's not flat anymore */
1422 /* and save its resno */
1423 grpColIdx[keyno++] = te->resdom->resno;
1431 * locate_grouping_columns
1432 * Locate grouping columns in the tlist chosen by query_planner.
1434 * This is only needed if we don't use the sub_tlist chosen by
1435 * make_subplanTargetList. We have to forget the column indexes found
1436 * by that routine and re-locate the grouping vars in the real sub_tlist.
1439 locate_grouping_columns(Query *parse,
1442 AttrNumber *groupColIdx)
1448 * No work unless grouping.
1450 if (!parse->groupClause)
1452 Assert(groupColIdx == NULL);
1455 Assert(groupColIdx != NULL);
1457 foreach(gl, parse->groupClause)
1459 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1460 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1461 TargetEntry *te = NULL;
1464 foreach(sl, sub_tlist)
1466 te = (TargetEntry *) lfirst(sl);
1467 if (equal(groupexpr, te->expr))
1471 elog(ERROR, "locate_grouping_columns: failed");
1473 groupColIdx[keyno++] = te->resdom->resno;
1478 * postprocess_setop_tlist
1479 * Fix up targetlist returned by plan_set_operations().
1481 * We need to transpose sort key info from the orig_tlist into new_tlist.
1482 * NOTE: this would not be good enough if we supported resjunk sort keys
1483 * for results of set operations --- then, we'd need to project a whole
1484 * new tlist to evaluate the resjunk columns. For now, just elog if we
1485 * find any resjunk columns in orig_tlist.
1488 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1492 foreach(l, new_tlist)
1494 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1495 TargetEntry *orig_tle;
1497 /* ignore resjunk columns in setop result */
1498 if (new_tle->resdom->resjunk)
1501 Assert(orig_tlist != NIL);
1502 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1503 orig_tlist = lnext(orig_tlist);
1504 if (orig_tle->resdom->resjunk)
1505 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1506 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1507 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1508 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1510 if (orig_tlist != NIL)
1511 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");