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 1.153 2003/05/06 00:20:32 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_PlannerParamVar;
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_PlannerParamVar = PlannerParamVar;
96 /* Initialize state for handling outer-level references and params */
97 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
98 PlannerParamVar = 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(PlannerParamVar);
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 PlannerParamVar = save_PlannerParamVar;
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 *. Also note that contain_agg_clause does not
265 * recurse into sub-selects, which is exactly what we need here.
268 foreach(lst, (List *) parse->havingQual)
270 Node *havingclause = (Node *) lfirst(lst);
272 if (contain_agg_clause(havingclause))
273 newHaving = lappend(newHaving, havingclause);
275 parse->jointree->quals = (Node *)
276 lappend((List *) parse->jointree->quals, havingclause);
278 parse->havingQual = (Node *) newHaving;
281 * If we have any outer joins, try to reduce them to plain inner joins.
282 * This step is most easily done after we've done expression preprocessing.
285 reduce_outer_joins(parse);
288 * See if we can simplify the jointree; opportunities for this may come
289 * from having pulled up subqueries, or from flattening explicit JOIN
290 * syntax. We must do this after flattening JOIN alias variables, since
291 * eliminating explicit JOIN nodes from the jointree will cause
292 * get_relids_for_join() to fail. But it should happen after
293 * reduce_outer_joins, anyway.
295 parse->jointree = (FromExpr *)
296 simplify_jointree(parse, (Node *) parse->jointree);
299 * Do the main planning. If we have an inherited target relation,
300 * that needs special processing, else go straight to
303 if (parse->resultRelation &&
304 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
306 plan = inheritance_planner(parse, lst);
308 plan = grouping_planner(parse, tuple_fraction);
311 * If any subplans were generated, or if we're inside a subplan, build
312 * initPlan list and extParam/allParam sets for plan nodes.
314 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
316 Cost initplan_cost = 0;
318 /* Prepare extParam/allParam sets for all nodes in tree */
319 SS_finalize_plan(plan, parse->rtable);
322 * SS_finalize_plan doesn't handle initPlans, so we have to manually
323 * attach them to the topmost plan node, and add their extParams to
324 * the topmost node's, too.
326 * We also add the total_cost of each initPlan to the startup cost
327 * of the top node. This is a conservative overestimate, since in
328 * fact each initPlan might be executed later than plan startup, or
331 plan->initPlan = PlannerInitPlan;
333 foreach(lst, plan->initPlan)
335 SubPlan *initplan = (SubPlan *) lfirst(lst);
337 plan->extParam = bms_add_members(plan->extParam,
338 initplan->plan->extParam);
339 initplan_cost += initplan->plan->total_cost;
342 plan->startup_cost += initplan_cost;
343 plan->total_cost += initplan_cost;
346 /* Return to outer subquery context */
348 PlannerInitPlan = saved_initplan;
349 /* we do NOT restore PlannerPlanId; that's not an oversight! */
355 * preprocess_expression
356 * Do subquery_planner's preprocessing work for an expression,
357 * which can be a targetlist, a WHERE clause (including JOIN/ON
358 * conditions), or a HAVING clause.
361 preprocess_expression(Query *parse, Node *expr, int kind)
364 * If the query has any join RTEs, replace join alias variables with
365 * base-relation variables. We must do this before sublink processing,
366 * else sublinks expanded out from join aliases wouldn't get processed.
368 if (parse->hasJoinRTEs)
369 expr = flatten_join_alias_vars(parse, expr);
372 * Simplify constant expressions.
374 * Note that at this point quals have not yet been converted to
375 * implicit-AND form, so we can apply eval_const_expressions directly.
377 expr = eval_const_expressions(expr);
380 * If it's a qual or havingQual, canonicalize it, and convert it to
381 * implicit-AND format.
383 * XXX Is there any value in re-applying eval_const_expressions after
386 if (kind == EXPRKIND_QUAL)
388 expr = (Node *) canonicalize_qual((Expr *) expr, true);
390 #ifdef OPTIMIZER_DEBUG
391 printf("After canonicalize_qual()\n");
396 /* Expand SubLinks to SubPlans */
397 if (parse->hasSubLinks)
398 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
400 /* Replace uplevel vars with Param nodes */
401 if (PlannerQueryLevel > 1)
402 expr = SS_replace_correlation_vars(expr);
408 * preprocess_qual_conditions
409 * Recursively scan the query's jointree and do subquery_planner's
410 * preprocessing work on each qual condition found therein.
413 preprocess_qual_conditions(Query *parse, Node *jtnode)
417 if (IsA(jtnode, RangeTblRef))
419 /* nothing to do here */
421 else if (IsA(jtnode, FromExpr))
423 FromExpr *f = (FromExpr *) jtnode;
426 foreach(l, f->fromlist)
427 preprocess_qual_conditions(parse, lfirst(l));
429 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
431 else if (IsA(jtnode, JoinExpr))
433 JoinExpr *j = (JoinExpr *) jtnode;
435 preprocess_qual_conditions(parse, j->larg);
436 preprocess_qual_conditions(parse, j->rarg);
438 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
441 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
445 /*--------------------
446 * inheritance_planner
447 * Generate a plan in the case where the result relation is an
450 * We have to handle this case differently from cases where a source
451 * relation is an inheritance set. Source inheritance is expanded at
452 * the bottom of the plan tree (see allpaths.c), but target inheritance
453 * has to be expanded at the top. The reason is that for UPDATE, each
454 * target relation needs a different targetlist matching its own column
455 * set. (This is not so critical for DELETE, but for simplicity we treat
456 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
457 * can never be the nullable side of an outer join, so it's OK to generate
460 * parse is the querytree produced by the parser & rewriter.
461 * inheritlist is an integer list of RT indexes for the result relation set.
463 * Returns a query plan.
464 *--------------------
467 inheritance_planner(Query *parse, List *inheritlist)
469 int parentRTindex = parse->resultRelation;
470 Oid parentOID = getrelid(parentRTindex, parse->rtable);
471 int mainrtlength = length(parse->rtable);
472 List *subplans = NIL;
476 foreach(l, inheritlist)
478 int childRTindex = lfirsti(l);
479 Oid childOID = getrelid(childRTindex, parse->rtable);
484 /* Generate modified query with this rel as target */
485 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
486 parentRTindex, parentOID,
487 childRTindex, childOID);
489 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
490 subplans = lappend(subplans, subplan);
492 * It's possible that additional RTEs got added to the rangetable
493 * due to expansion of inherited source tables (see allpaths.c).
494 * If so, we must copy 'em back to the main parse tree's rtable.
496 * XXX my goodness this is ugly. Really need to think about ways
497 * to rein in planner's habit of scribbling on its input.
499 subrtlength = length(subquery->rtable);
500 if (subrtlength > mainrtlength)
502 List *subrt = subquery->rtable;
504 while (mainrtlength-- > 0) /* wish we had nthcdr() */
505 subrt = lnext(subrt);
506 parse->rtable = nconc(parse->rtable, subrt);
507 mainrtlength = subrtlength;
509 /* Save preprocessed tlist from first rel for use in Append */
511 tlist = subplan->targetlist;
514 /* Save the target-relations list for the executor, too */
515 parse->resultRelations = inheritlist;
517 /* Mark result as unordered (probably unnecessary) */
518 parse->query_pathkeys = NIL;
520 return (Plan *) make_append(subplans, true, tlist);
523 /*--------------------
525 * Perform planning steps related to grouping, aggregation, etc.
526 * This primarily means adding top-level processing to the basic
527 * query plan produced by query_planner.
529 * parse is the querytree produced by the parser & rewriter.
530 * tuple_fraction is the fraction of tuples we expect will be retrieved
532 * tuple_fraction is interpreted as follows:
533 * 0: expect all tuples to be retrieved (normal case)
534 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
535 * from the plan to be retrieved
536 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
537 * expected to be retrieved (ie, a LIMIT specification)
539 * Returns a query plan. Also, parse->query_pathkeys is returned as the
540 * actual output ordering of the plan (in pathkey format).
541 *--------------------
544 grouping_planner(Query *parse, double tuple_fraction)
546 List *tlist = parse->targetList;
548 List *current_pathkeys;
551 if (parse->setOperations)
554 * Construct the plan for set operations. The result will not
555 * need any work except perhaps a top-level sort and/or LIMIT.
557 result_plan = plan_set_operations(parse);
560 * We should not need to call preprocess_targetlist, since we must
561 * be in a SELECT query node. Instead, use the targetlist
562 * returned by plan_set_operations (since this tells whether it
563 * returned any resjunk columns!), and transfer any sort key
564 * information from the original tlist.
566 Assert(parse->commandType == CMD_SELECT);
568 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
571 * Can't handle FOR UPDATE here (parser should have checked
572 * already, but let's make sure).
575 elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
578 * We set current_pathkeys NIL indicating we do not know sort
579 * order. This is correct when the top set operation is UNION
580 * ALL, since the appended-together results are unsorted even if
581 * the subplans were sorted. For other set operations we could be
582 * smarter --- room for future improvement!
584 current_pathkeys = NIL;
587 * Calculate pathkeys that represent ordering requirements
589 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
591 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
595 /* No set operations, do regular planning */
597 List *group_pathkeys;
598 AttrNumber *groupColIdx = NULL;
599 bool need_tlist_eval = true;
601 double sub_tuple_fraction;
604 double dNumGroups = 0;
607 int numGroupCols = length(parse->groupClause);
608 bool use_hashed_grouping = false;
610 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
611 tlist = preprocess_targetlist(tlist,
613 parse->resultRelation,
617 * Add TID targets for rels selected FOR UPDATE (should this be
618 * done in preprocess_targetlist?). The executor uses the TID to
619 * know which rows to lock, much as for UPDATE or DELETE.
626 * We've got trouble if the FOR UPDATE appears inside
627 * grouping, since grouping renders a reference to individual
628 * tuple CTIDs invalid. This is also checked at parse time,
629 * but that's insufficient because of rule substitution, query
632 CheckSelectForUpdate(parse);
635 * Currently the executor only supports FOR UPDATE at top
638 if (PlannerQueryLevel > 1)
639 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
641 foreach(l, parse->rowMarks)
643 Index rti = lfirsti(l);
649 resname = (char *) palloc(32);
650 snprintf(resname, 32, "ctid%u", rti);
651 resdom = makeResdom(length(tlist) + 1,
658 SelfItemPointerAttributeNumber,
663 ctid = makeTargetEntry(resdom, (Expr *) var);
664 tlist = lappend(tlist, ctid);
669 * Generate appropriate target list for subplan; may be different
670 * from tlist if grouping or aggregation is needed.
672 sub_tlist = make_subplanTargetList(parse, tlist,
673 &groupColIdx, &need_tlist_eval);
676 * Calculate pathkeys that represent grouping/ordering
679 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
681 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
685 * Will need actual number of aggregates for estimating costs.
686 * Also, it's possible that optimization has eliminated all
687 * aggregates, and we may as well check for that here.
689 * Note: we do not attempt to detect duplicate aggregates here;
690 * a somewhat-overestimated count is okay for our present purposes.
694 numAggs = count_agg_clause((Node *) tlist) +
695 count_agg_clause(parse->havingQual);
697 parse->hasAggs = false;
701 * Figure out whether we need a sorted result from query_planner.
703 * If we have a GROUP BY clause, then we want a result sorted
704 * properly for grouping. Otherwise, if there is an ORDER BY
705 * clause, we want to sort by the ORDER BY clause. (Note: if we
706 * have both, and ORDER BY is a superset of GROUP BY, it would be
707 * tempting to request sort by ORDER BY --- but that might just
708 * leave us failing to exploit an available sort order at all.
709 * Needs more thought...)
711 if (parse->groupClause)
712 parse->query_pathkeys = group_pathkeys;
713 else if (parse->sortClause)
714 parse->query_pathkeys = sort_pathkeys;
716 parse->query_pathkeys = NIL;
719 * Adjust tuple_fraction if we see that we are going to apply
720 * limiting/grouping/aggregation/etc. This is not overridable by
721 * the caller, since it reflects plan actions that this routine
722 * will certainly take, not assumptions about context.
724 if (parse->limitCount != NULL)
727 * A LIMIT clause limits the absolute number of tuples
728 * returned. However, if it's not a constant LIMIT then we
729 * have to punt; for lack of a better idea, assume 10% of the
730 * plan's result is wanted.
732 double limit_fraction = 0.0;
734 if (IsA(parse->limitCount, Const))
736 Const *limitc = (Const *) parse->limitCount;
737 int32 count = DatumGetInt32(limitc->constvalue);
740 * A NULL-constant LIMIT represents "LIMIT ALL", which we
741 * treat the same as no limit (ie, expect to retrieve all
744 if (!limitc->constisnull && count > 0)
746 limit_fraction = (double) count;
747 /* We must also consider the OFFSET, if present */
748 if (parse->limitOffset != NULL)
750 if (IsA(parse->limitOffset, Const))
754 limitc = (Const *) parse->limitOffset;
755 offset = DatumGetInt32(limitc->constvalue);
756 if (!limitc->constisnull && offset > 0)
757 limit_fraction += (double) offset;
761 /* OFFSET is an expression ... punt ... */
762 limit_fraction = 0.10;
769 /* LIMIT is an expression ... punt ... */
770 limit_fraction = 0.10;
773 if (limit_fraction > 0.0)
776 * If we have absolute limits from both caller and LIMIT,
777 * use the smaller value; if one is fractional and the
778 * other absolute, treat the fraction as a fraction of the
779 * absolute value; else we can multiply the two fractions
782 if (tuple_fraction >= 1.0)
784 if (limit_fraction >= 1.0)
787 tuple_fraction = Min(tuple_fraction, limit_fraction);
791 /* caller absolute, limit fractional */
792 tuple_fraction *= limit_fraction;
793 if (tuple_fraction < 1.0)
794 tuple_fraction = 1.0;
797 else if (tuple_fraction > 0.0)
799 if (limit_fraction >= 1.0)
801 /* caller fractional, limit absolute */
802 tuple_fraction *= limit_fraction;
803 if (tuple_fraction < 1.0)
804 tuple_fraction = 1.0;
808 /* both fractional */
809 tuple_fraction *= limit_fraction;
814 /* no info from caller, just use limit */
815 tuple_fraction = limit_fraction;
821 * With grouping or aggregation, the tuple fraction to pass to
822 * query_planner() may be different from what it is at top level.
824 sub_tuple_fraction = tuple_fraction;
826 if (parse->groupClause)
829 * In GROUP BY mode, we have the little problem that we don't
830 * really know how many input tuples will be needed to make a
831 * group, so we can't translate an output LIMIT count into an
832 * input count. For lack of a better idea, assume 25% of the
833 * input data will be processed if there is any output limit.
834 * However, if the caller gave us a fraction rather than an
835 * absolute count, we can keep using that fraction (which
836 * amounts to assuming that all the groups are about the same
839 if (sub_tuple_fraction >= 1.0)
840 sub_tuple_fraction = 0.25;
843 * If both GROUP BY and ORDER BY are specified, we will need
844 * two levels of sort --- and, therefore, certainly need to
845 * read all the input tuples --- unless ORDER BY is a subset
846 * of GROUP BY. (We have not yet canonicalized the pathkeys,
847 * so must use the slower noncanonical comparison method.)
849 if (parse->groupClause && parse->sortClause &&
850 !noncanonical_pathkeys_contained_in(sort_pathkeys,
852 sub_tuple_fraction = 0.0;
854 else if (parse->hasAggs)
857 * Ungrouped aggregate will certainly want all the input
860 sub_tuple_fraction = 0.0;
862 else if (parse->distinctClause)
865 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
866 * number of input tuples per output tuple. Handle the same
869 if (sub_tuple_fraction >= 1.0)
870 sub_tuple_fraction = 0.25;
874 * Generate the best unsorted and presorted paths for this Query
875 * (but note there may not be any presorted path).
877 query_planner(parse, sub_tlist, sub_tuple_fraction,
878 &cheapest_path, &sorted_path);
881 * We couldn't canonicalize group_pathkeys and sort_pathkeys before
882 * running query_planner(), so do it now.
884 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
885 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
888 * Consider whether we might want to use hashed grouping.
890 if (parse->groupClause)
895 * Always estimate the number of groups. We can't do this until
896 * after running query_planner(), either.
898 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
900 dNumGroups = estimate_num_groups(parse,
902 cheapest_path->parent->rows);
903 /* Also want it as a long int --- but 'ware overflow! */
904 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
907 * Check can't-do-it conditions, including whether the grouping
908 * operators are hashjoinable.
910 * Executor doesn't support hashed aggregation with DISTINCT
911 * aggregates. (Doing so would imply storing *all* the input
912 * values in the hash table, which seems like a certain loser.)
914 if (!enable_hashagg || !hash_safe_grouping(parse))
915 use_hashed_grouping = false;
916 else if (parse->hasAggs &&
917 (contain_distinct_agg_clause((Node *) tlist) ||
918 contain_distinct_agg_clause(parse->havingQual)))
919 use_hashed_grouping = false;
923 * Use hashed grouping if (a) we think we can fit the
924 * hashtable into SortMem, *and* (b) the estimated cost
925 * is no more than doing it the other way. While avoiding
926 * the need for sorted input is usually a win, the fact
927 * that the output won't be sorted may be a loss; so we
928 * need to do an actual cost comparison.
930 * In most cases we have no good way to estimate the size of
931 * the transition value needed by an aggregate; arbitrarily
932 * assume it is 100 bytes. Also set the overhead per hashtable
935 int hashentrysize = cheapest_path->parent->width + 64 +
938 if (hashentrysize * dNumGroups <= SortMem * 1024L)
941 * Okay, do the cost comparison. We need to consider
942 * cheapest_path + hashagg [+ final sort]
944 * cheapest_path [+ sort] + group or agg [+ final sort]
946 * presorted_path + group or agg [+ final sort]
947 * where brackets indicate a step that may not be needed.
948 * We assume query_planner() will have returned a
949 * presorted path only if it's a winner compared to
950 * cheapest_path for this purpose.
952 * These path variables are dummies that just hold cost
953 * fields; we don't make actual Paths for these steps.
958 cost_agg(&hashed_p, parse,
960 numGroupCols, dNumGroups,
961 cheapest_path->startup_cost,
962 cheapest_path->total_cost,
963 cheapest_path->parent->rows);
964 /* Result of hashed agg is always unsorted */
966 cost_sort(&hashed_p, parse, sort_pathkeys,
969 cheapest_path->parent->width);
973 sorted_p.startup_cost = sorted_path->startup_cost;
974 sorted_p.total_cost = sorted_path->total_cost;
975 current_pathkeys = sorted_path->pathkeys;
979 sorted_p.startup_cost = cheapest_path->startup_cost;
980 sorted_p.total_cost = cheapest_path->total_cost;
981 current_pathkeys = cheapest_path->pathkeys;
983 if (!pathkeys_contained_in(group_pathkeys,
986 cost_sort(&sorted_p, parse, group_pathkeys,
988 cheapest_path->parent->rows,
989 cheapest_path->parent->width);
990 current_pathkeys = group_pathkeys;
993 cost_agg(&sorted_p, parse,
995 numGroupCols, dNumGroups,
996 sorted_p.startup_cost,
998 cheapest_path->parent->rows);
1000 cost_group(&sorted_p, parse,
1001 numGroupCols, dNumGroups,
1002 sorted_p.startup_cost,
1003 sorted_p.total_cost,
1004 cheapest_path->parent->rows);
1005 /* The Agg or Group node will preserve ordering */
1006 if (sort_pathkeys &&
1007 !pathkeys_contained_in(sort_pathkeys,
1010 cost_sort(&sorted_p, parse, sort_pathkeys,
1011 sorted_p.total_cost,
1013 cheapest_path->parent->width);
1017 * Now make the decision using the top-level tuple
1018 * fraction. First we have to convert an absolute
1019 * count (LIMIT) into fractional form.
1021 if (tuple_fraction >= 1.0)
1022 tuple_fraction /= dNumGroups;
1024 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1025 tuple_fraction) < 0)
1027 /* Hashed is cheaper, so use it */
1028 use_hashed_grouping = true;
1035 * Select the best path and create a plan to execute it.
1037 * If we are doing hashed grouping, we will always read all the
1038 * input tuples, so use the cheapest-total path. Otherwise,
1039 * trust query_planner's decision about which to use.
1041 if (sorted_path && !use_hashed_grouping)
1043 result_plan = create_plan(parse, sorted_path);
1044 current_pathkeys = sorted_path->pathkeys;
1048 result_plan = create_plan(parse, cheapest_path);
1049 current_pathkeys = cheapest_path->pathkeys;
1053 * create_plan() returns a plan with just a "flat" tlist of required
1054 * Vars. Usually we need to insert the sub_tlist as the tlist of the
1055 * top plan node. However, we can skip that if we determined that
1056 * whatever query_planner chose to return will be good enough.
1058 if (need_tlist_eval)
1061 * If the top-level plan node is one that cannot do expression
1062 * evaluation, we must insert a Result node to project the desired
1064 * Currently, the only plan node we might see here that falls into
1065 * that category is Append.
1067 if (IsA(result_plan, Append))
1069 result_plan = (Plan *) make_result(sub_tlist, NULL,
1075 * Otherwise, just replace the subplan's flat tlist with
1076 * the desired tlist.
1078 result_plan->targetlist = sub_tlist;
1081 * Also, account for the cost of evaluation of the sub_tlist.
1083 * Up to now, we have only been dealing with "flat" tlists,
1084 * containing just Vars. So their evaluation cost is zero
1085 * according to the model used by cost_qual_eval() (or if you
1086 * prefer, the cost is factored into cpu_tuple_cost). Thus we can
1087 * avoid accounting for tlist cost throughout query_planner() and
1088 * subroutines. But now we've inserted a tlist that might contain
1089 * actual operators, sub-selects, etc --- so we'd better account
1092 * Below this point, any tlist eval cost for added-on nodes should
1093 * be accounted for as we create those nodes. Presently, of the
1094 * node types we can add on, only Agg and Group project new tlists
1095 * (the rest just copy their input tuples) --- so make_agg() and
1096 * make_group() are responsible for computing the added cost.
1098 cost_qual_eval(&tlist_cost, sub_tlist);
1099 result_plan->startup_cost += tlist_cost.startup;
1100 result_plan->total_cost += tlist_cost.startup +
1101 tlist_cost.per_tuple * result_plan->plan_rows;
1106 * Since we're using query_planner's tlist and not the one
1107 * make_subplanTargetList calculated, we have to refigure
1108 * any grouping-column indexes make_subplanTargetList computed.
1110 locate_grouping_columns(parse, tlist, result_plan->targetlist,
1115 * Insert AGG or GROUP node if needed, plus an explicit sort step
1118 * HAVING clause, if any, becomes qual of the Agg node
1120 if (use_hashed_grouping)
1122 /* Hashed aggregate plan --- no sort needed */
1123 result_plan = (Plan *) make_agg(parse,
1125 (List *) parse->havingQual,
1132 /* Hashed aggregation produces randomly-ordered results */
1133 current_pathkeys = NIL;
1135 else if (parse->hasAggs)
1137 /* Plain aggregate plan --- sort if needed */
1138 AggStrategy aggstrategy;
1140 if (parse->groupClause)
1142 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1144 result_plan = (Plan *)
1145 make_sort_from_groupcols(parse,
1149 current_pathkeys = group_pathkeys;
1151 aggstrategy = AGG_SORTED;
1153 * The AGG node will not change the sort ordering of its
1154 * groups, so current_pathkeys describes the result too.
1159 aggstrategy = AGG_PLAIN;
1160 /* Result will be only one row anyway; no sort order */
1161 current_pathkeys = NIL;
1164 result_plan = (Plan *) make_agg(parse,
1166 (List *) parse->havingQual,
1177 * If there are no Aggs, we shouldn't have any HAVING qual anymore
1179 Assert(parse->havingQual == NULL);
1182 * If we have a GROUP BY clause, insert a group node (plus the
1183 * appropriate sort node, if necessary).
1185 if (parse->groupClause)
1188 * Add an explicit sort if we couldn't make the path come out
1189 * the way the GROUP node needs it.
1191 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1193 result_plan = (Plan *)
1194 make_sort_from_groupcols(parse,
1198 current_pathkeys = group_pathkeys;
1201 result_plan = (Plan *) make_group(parse,
1207 /* The Group node won't change sort ordering */
1210 } /* end of if (setOperations) */
1213 * If we were not able to make the plan come out in the right order,
1214 * add an explicit sort step.
1216 if (parse->sortClause)
1218 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1220 result_plan = (Plan *)
1221 make_sort_from_sortclauses(parse,
1225 current_pathkeys = sort_pathkeys;
1230 * If there is a DISTINCT clause, add the UNIQUE node.
1232 if (parse->distinctClause)
1234 result_plan = (Plan *) make_unique(tlist, result_plan,
1235 parse->distinctClause);
1237 * If there was grouping or aggregation, leave plan_rows as-is
1238 * (ie, assume the result was already mostly unique). If not,
1239 * it's reasonable to assume the UNIQUE filter has effects
1240 * comparable to GROUP BY.
1242 if (!parse->groupClause && !parse->hasAggs)
1244 List *distinctExprs;
1246 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1248 result_plan->plan_rows = estimate_num_groups(parse,
1250 result_plan->plan_rows);
1255 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1257 if (parse->limitOffset || parse->limitCount)
1259 result_plan = (Plan *) make_limit(tlist, result_plan,
1265 * Return the actual output ordering in query_pathkeys for possible
1266 * use by an outer query level.
1268 parse->query_pathkeys = current_pathkeys;
1274 * hash_safe_grouping - are grouping operators hashable?
1276 * We assume hashed aggregation will work if the datatype's equality operator
1277 * is marked hashjoinable.
1280 hash_safe_grouping(Query *parse)
1284 foreach(gl, parse->groupClause)
1286 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1287 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1291 optup = equality_oper(tle->resdom->restype, false);
1292 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1293 ReleaseSysCache(optup);
1301 * make_subplanTargetList
1302 * Generate appropriate target list when grouping is required.
1304 * When grouping_planner inserts Aggregate or Group plan nodes above
1305 * the result of query_planner, we typically want to pass a different
1306 * target list to query_planner than the outer plan nodes should have.
1307 * This routine generates the correct target list for the subplan.
1309 * The initial target list passed from the parser already contains entries
1310 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1311 * for variables used only in HAVING clauses; so we need to add those
1312 * variables to the subplan target list. Also, if we are doing either
1313 * grouping or aggregation, we flatten all expressions except GROUP BY items
1314 * into their component variables; the other expressions will be computed by
1315 * the inserted nodes rather than by the subplan. For example,
1316 * given a query like
1317 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1318 * we want to pass this targetlist to the subplan:
1320 * where the a+b target will be used by the Sort/Group steps, and the
1321 * other targets will be used for computing the final results. (In the
1322 * above example we could theoretically suppress the a and b targets and
1323 * pass down only c,d,a+b, but it's not really worth the trouble to
1324 * eliminate simple var references from the subplan. We will avoid doing
1325 * the extra computation to recompute a+b at the outer level; see
1326 * replace_vars_with_subplan_refs() in setrefs.c.)
1328 * If we are grouping or aggregating, *and* there are no non-Var grouping
1329 * expressions, then the returned tlist is effectively dummy; we do not
1330 * need to force it to be evaluated, because all the Vars it contains
1331 * should be present in the output of query_planner anyway.
1333 * 'parse' is the query being processed.
1334 * 'tlist' is the query's target list.
1335 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1336 * expressions (if there are any) in the subplan's target list.
1337 * 'need_tlist_eval' is set true if we really need to evaluate the
1340 * The result is the targetlist to be passed to the subplan.
1344 make_subplanTargetList(Query *parse,
1346 AttrNumber **groupColIdx,
1347 bool *need_tlist_eval)
1353 *groupColIdx = NULL;
1356 * If we're not grouping or aggregating, nothing to do here;
1357 * query_planner should receive the unmodified target list.
1359 if (!parse->hasAggs && !parse->groupClause && !parse->havingQual)
1361 *need_tlist_eval = true;
1366 * Otherwise, start with a "flattened" tlist (having just the vars
1367 * mentioned in the targetlist and HAVING qual --- but not upper-
1368 * level Vars; they will be replaced by Params later on).
1370 sub_tlist = flatten_tlist(tlist);
1371 extravars = pull_var_clause(parse->havingQual, false);
1372 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1373 freeList(extravars);
1374 *need_tlist_eval = false; /* only eval if not flat tlist */
1377 * If grouping, create sub_tlist entries for all GROUP BY expressions
1378 * (GROUP BY items that are simple Vars should be in the list
1379 * already), and make an array showing where the group columns are in
1382 numCols = length(parse->groupClause);
1386 AttrNumber *grpColIdx;
1389 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1390 *groupColIdx = grpColIdx;
1392 foreach(gl, parse->groupClause)
1394 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1395 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1396 TargetEntry *te = NULL;
1399 /* Find or make a matching sub_tlist entry */
1400 foreach(sl, sub_tlist)
1402 te = (TargetEntry *) lfirst(sl);
1403 if (equal(groupexpr, te->expr))
1408 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1409 exprType(groupexpr),
1410 exprTypmod(groupexpr),
1413 (Expr *) groupexpr);
1414 sub_tlist = lappend(sub_tlist, te);
1415 *need_tlist_eval = true; /* it's not flat anymore */
1418 /* and save its resno */
1419 grpColIdx[keyno++] = te->resdom->resno;
1427 * locate_grouping_columns
1428 * Locate grouping columns in the tlist chosen by query_planner.
1430 * This is only needed if we don't use the sub_tlist chosen by
1431 * make_subplanTargetList. We have to forget the column indexes found
1432 * by that routine and re-locate the grouping vars in the real sub_tlist.
1435 locate_grouping_columns(Query *parse,
1438 AttrNumber *groupColIdx)
1444 * No work unless grouping.
1446 if (!parse->groupClause)
1448 Assert(groupColIdx == NULL);
1451 Assert(groupColIdx != NULL);
1453 foreach(gl, parse->groupClause)
1455 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1456 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1457 TargetEntry *te = NULL;
1460 foreach(sl, sub_tlist)
1462 te = (TargetEntry *) lfirst(sl);
1463 if (equal(groupexpr, te->expr))
1467 elog(ERROR, "locate_grouping_columns: failed");
1469 groupColIdx[keyno++] = te->resdom->resno;
1474 * postprocess_setop_tlist
1475 * Fix up targetlist returned by plan_set_operations().
1477 * We need to transpose sort key info from the orig_tlist into new_tlist.
1478 * NOTE: this would not be good enough if we supported resjunk sort keys
1479 * for results of set operations --- then, we'd need to project a whole
1480 * new tlist to evaluate the resjunk columns. For now, just elog if we
1481 * find any resjunk columns in orig_tlist.
1484 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1488 foreach(l, new_tlist)
1490 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1491 TargetEntry *orig_tle;
1493 /* ignore resjunk columns in setop result */
1494 if (new_tle->resdom->resjunk)
1497 Assert(orig_tlist != NIL);
1498 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1499 orig_tlist = lnext(orig_tlist);
1500 if (orig_tle->resdom->resjunk)
1501 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1502 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1503 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1504 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1506 if (orig_tlist != NIL)
1507 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");