1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2004, 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 1.173 2004/08/29 04:12:33 momjian 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 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(Query *parse, Node *expr, int kind);
58 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
59 static Plan *inheritance_planner(Query *parse, List *inheritlist);
60 static Plan *grouping_planner(Query *parse, double tuple_fraction);
61 static bool hash_safe_grouping(Query *parse);
62 static List *make_subplanTargetList(Query *parse, List *tlist,
63 AttrNumber **groupColIdx, bool *need_tlist_eval);
64 static void locate_grouping_columns(Query *parse,
67 AttrNumber *groupColIdx);
68 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
71 /*****************************************************************************
73 * Query optimizer entry point
75 *****************************************************************************/
77 planner(Query *parse, bool isCursor, int cursorOptions,
78 ParamListInfo boundParams)
80 double tuple_fraction;
82 Index save_PlannerQueryLevel;
83 List *save_PlannerParamList;
84 ParamListInfo save_PlannerBoundParamList;
87 * The planner can be called recursively (an example is when
88 * eval_const_expressions tries to pre-evaluate an SQL function). So,
89 * these global state variables must be saved and restored.
91 * Query level and the param list cannot be moved into the Query structure
92 * since their whole purpose is communication across multiple sub-Queries.
93 * Also, boundParams is explicitly info from outside the Query, and so
94 * is likewise better handled as a global variable.
96 * Note we do NOT save and restore PlannerPlanId: it exists to assign
97 * unique IDs to SubPlan nodes, and we want those IDs to be unique for
98 * the life of a backend. Also, PlannerInitPlan is saved/restored in
99 * subquery_planner, not here.
101 save_PlannerQueryLevel = PlannerQueryLevel;
102 save_PlannerParamList = PlannerParamList;
103 save_PlannerBoundParamList = PlannerBoundParamList;
105 /* Initialize state for handling outer-level references and params */
106 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
107 PlannerParamList = NIL;
108 PlannerBoundParamList = boundParams;
110 /* Determine what fraction of the plan is likely to be scanned */
114 * We have no real idea how many tuples the user will ultimately
115 * FETCH from a cursor, but it seems a good bet that he doesn't
116 * want 'em all. Optimize for 10% retrieval (you gotta better
117 * number? Should this be a SETtable parameter?)
119 tuple_fraction = 0.10;
123 /* Default assumption is we need all the tuples */
124 tuple_fraction = 0.0;
127 /* primary planning entry point (may recurse for subqueries) */
128 result_plan = subquery_planner(parse, tuple_fraction);
130 Assert(PlannerQueryLevel == 0);
133 * If creating a plan for a scrollable cursor, make sure it can run
134 * backwards on demand. Add a Material node at the top at need.
136 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
138 if (!ExecSupportsBackwardScan(result_plan))
139 result_plan = materialize_finished_plan(result_plan);
142 /* executor wants to know total number of Params used overall */
143 result_plan->nParamExec = list_length(PlannerParamList);
145 /* final cleanup of the plan */
146 set_plan_references(result_plan, parse->rtable);
148 /* restore state for outer planner, if any */
149 PlannerQueryLevel = save_PlannerQueryLevel;
150 PlannerParamList = save_PlannerParamList;
151 PlannerBoundParamList = save_PlannerBoundParamList;
157 /*--------------------
159 * Invokes the planner on a subquery. We recurse to here for each
160 * sub-SELECT found in the query tree.
162 * parse is the querytree produced by the parser & rewriter.
163 * tuple_fraction is the fraction of tuples we expect will be retrieved.
164 * tuple_fraction is interpreted as explained for grouping_planner, below.
166 * Basically, this routine does the stuff that should only be done once
167 * per Query object. It then calls grouping_planner. At one time,
168 * grouping_planner could be invoked recursively on the same Query object;
169 * that's not currently true, but we keep the separation between the two
170 * routines anyway, in case we need it again someday.
172 * subquery_planner will be called recursively to handle sub-Query nodes
173 * found within the query's expressions and rangetable.
175 * Returns a query plan.
176 *--------------------
179 subquery_planner(Query *parse, double tuple_fraction)
181 List *saved_initplan = PlannerInitPlan;
182 int saved_planid = PlannerPlanId;
189 /* Set up for a new level of subquery */
191 PlannerInitPlan = NIL;
194 * Look for IN clauses at the top level of WHERE, and transform them
195 * into joins. Note that this step only handles IN clauses originally
196 * at top level of WHERE; if we pull up any subqueries in the next
197 * step, their INs are processed just before pulling them up.
199 parse->in_info_list = NIL;
200 if (parse->hasSubLinks)
201 parse->jointree->quals = pull_up_IN_clauses(parse,
202 parse->jointree->quals);
205 * Check to see if any subqueries in the rangetable can be merged into
208 parse->jointree = (FromExpr *)
209 pull_up_subqueries(parse, (Node *) parse->jointree, false);
212 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
213 * can avoid the expense of doing flatten_join_alias_vars(). Also
214 * check for outer joins --- if none, we can skip
215 * reduce_outer_joins(). This must be done after we have done
216 * pull_up_subqueries, of course.
218 parse->hasJoinRTEs = false;
219 hasOuterJoins = false;
220 foreach(l, parse->rtable)
222 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
224 if (rte->rtekind == RTE_JOIN)
226 parse->hasJoinRTEs = true;
227 if (IS_OUTER_JOIN(rte->jointype))
229 hasOuterJoins = true;
230 /* Can quit scanning once we find an outer join */
237 * Do expression preprocessing on targetlist and quals.
239 parse->targetList = (List *)
240 preprocess_expression(parse, (Node *) parse->targetList,
243 preprocess_qual_conditions(parse, (Node *) parse->jointree);
245 parse->havingQual = preprocess_expression(parse, parse->havingQual,
248 parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
250 parse->limitCount = preprocess_expression(parse, parse->limitCount,
253 parse->in_info_list = (List *)
254 preprocess_expression(parse, (Node *) parse->in_info_list,
257 /* Also need to preprocess expressions for function RTEs */
258 foreach(l, parse->rtable)
260 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
262 if (rte->rtekind == RTE_FUNCTION)
263 rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
268 * A HAVING clause without aggregates is equivalent to a WHERE clause
269 * (except it can only refer to grouped fields). Transfer any
270 * agg-free clauses of the HAVING qual into WHERE. This may seem like
271 * wasting cycles to cater to stupidly-written queries, but there are
272 * other reasons for doing it. Firstly, if the query contains no aggs
273 * at all, then we aren't going to generate an Agg plan node, and so
274 * there'll be no place to execute HAVING conditions; without this
275 * transfer, we'd lose the HAVING condition entirely, which is wrong.
276 * Secondly, when we push down a qual condition into a sub-query, it's
277 * easiest to push the qual into HAVING always, in case it contains
278 * aggs, and then let this code sort it out.
280 * Note that both havingQual and parse->jointree->quals are in
281 * implicitly-ANDed-list form at this point, even though they are
282 * declared as Node *.
285 foreach(l, (List *) parse->havingQual)
287 Node *havingclause = (Node *) lfirst(l);
289 if (contain_agg_clause(havingclause))
290 newHaving = lappend(newHaving, havingclause);
292 parse->jointree->quals = (Node *)
293 lappend((List *) parse->jointree->quals, havingclause);
295 parse->havingQual = (Node *) newHaving;
298 * If we have any outer joins, try to reduce them to plain inner
299 * joins. This step is most easily done after we've done expression
303 reduce_outer_joins(parse);
306 * See if we can simplify the jointree; opportunities for this may
307 * come from having pulled up subqueries, or from flattening explicit
308 * JOIN syntax. We must do this after flattening JOIN alias
309 * variables, since eliminating explicit JOIN nodes from the jointree
310 * will cause get_relids_for_join() to fail. But it should happen
311 * after reduce_outer_joins, anyway.
313 parse->jointree = (FromExpr *)
314 simplify_jointree(parse, (Node *) parse->jointree);
317 * Do the main planning. If we have an inherited target relation,
318 * that needs special processing, else go straight to
321 if (parse->resultRelation &&
322 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
324 plan = inheritance_planner(parse, lst);
326 plan = grouping_planner(parse, tuple_fraction);
329 * If any subplans were generated, or if we're inside a subplan, build
330 * initPlan list and extParam/allParam sets for plan nodes.
332 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
334 Cost initplan_cost = 0;
336 /* Prepare extParam/allParam sets for all nodes in tree */
337 SS_finalize_plan(plan, parse->rtable);
340 * SS_finalize_plan doesn't handle initPlans, so we have to
341 * manually attach them to the topmost plan node, and add their
342 * extParams to the topmost node's, too.
344 * We also add the total_cost of each initPlan to the startup cost of
345 * the top node. This is a conservative overestimate, since in
346 * fact each initPlan might be executed later than plan startup,
347 * or even not at all.
349 plan->initPlan = PlannerInitPlan;
351 foreach(l, plan->initPlan)
353 SubPlan *initplan = (SubPlan *) lfirst(l);
355 plan->extParam = bms_add_members(plan->extParam,
356 initplan->plan->extParam);
357 /* allParam must include all members of extParam */
358 plan->allParam = bms_add_members(plan->allParam,
360 initplan_cost += initplan->plan->total_cost;
363 plan->startup_cost += initplan_cost;
364 plan->total_cost += initplan_cost;
367 /* Return to outer subquery context */
369 PlannerInitPlan = saved_initplan;
370 /* we do NOT restore PlannerPlanId; that's not an oversight! */
376 * preprocess_expression
377 * Do subquery_planner's preprocessing work for an expression,
378 * which can be a targetlist, a WHERE clause (including JOIN/ON
379 * conditions), or a HAVING clause.
382 preprocess_expression(Query *parse, Node *expr, int kind)
385 * If the query has any join RTEs, replace join alias variables with
386 * base-relation variables. We must do this before sublink processing,
387 * else sublinks expanded out from join aliases wouldn't get
390 if (parse->hasJoinRTEs)
391 expr = flatten_join_alias_vars(parse, expr);
394 * If it's a qual or havingQual, canonicalize it. It seems most useful
395 * to do this before applying eval_const_expressions, since the latter
396 * can optimize flattened AND/ORs better than unflattened ones.
398 * Note: all processing of a qual expression after this point must be
399 * careful to maintain AND/OR flatness --- that is, do not generate a
400 * tree with AND directly under AND, nor OR directly under OR.
402 if (kind == EXPRKIND_QUAL)
404 expr = (Node *) canonicalize_qual((Expr *) expr);
406 #ifdef OPTIMIZER_DEBUG
407 printf("After canonicalize_qual()\n");
413 * Simplify constant expressions.
415 expr = eval_const_expressions(expr);
417 /* Expand SubLinks to SubPlans */
418 if (parse->hasSubLinks)
419 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
422 * XXX do not insert anything here unless you have grokked the
423 * comments in SS_replace_correlation_vars ...
426 /* Replace uplevel vars with Param nodes */
427 if (PlannerQueryLevel > 1)
428 expr = SS_replace_correlation_vars(expr);
431 * If it's a qual or havingQual, convert it to implicit-AND format.
432 * (We don't want to do this before eval_const_expressions, since the
433 * latter would be unable to simplify a top-level AND correctly. Also,
434 * SS_process_sublinks expects explicit-AND format.)
436 if (kind == EXPRKIND_QUAL)
437 expr = (Node *) make_ands_implicit((Expr *) expr);
443 * preprocess_qual_conditions
444 * Recursively scan the query's jointree and do subquery_planner's
445 * preprocessing work on each qual condition found therein.
448 preprocess_qual_conditions(Query *parse, Node *jtnode)
452 if (IsA(jtnode, RangeTblRef))
454 /* nothing to do here */
456 else if (IsA(jtnode, FromExpr))
458 FromExpr *f = (FromExpr *) jtnode;
461 foreach(l, f->fromlist)
462 preprocess_qual_conditions(parse, lfirst(l));
464 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
466 else if (IsA(jtnode, JoinExpr))
468 JoinExpr *j = (JoinExpr *) jtnode;
470 preprocess_qual_conditions(parse, j->larg);
471 preprocess_qual_conditions(parse, j->rarg);
473 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
476 elog(ERROR, "unrecognized node type: %d",
477 (int) nodeTag(jtnode));
480 /*--------------------
481 * inheritance_planner
482 * Generate a plan in the case where the result relation is an
485 * We have to handle this case differently from cases where a source
486 * relation is an inheritance set. Source inheritance is expanded at
487 * the bottom of the plan tree (see allpaths.c), but target inheritance
488 * has to be expanded at the top. The reason is that for UPDATE, each
489 * target relation needs a different targetlist matching its own column
490 * set. (This is not so critical for DELETE, but for simplicity we treat
491 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
492 * can never be the nullable side of an outer join, so it's OK to generate
495 * parse is the querytree produced by the parser & rewriter.
496 * inheritlist is an integer list of RT indexes for the result relation set.
498 * Returns a query plan.
499 *--------------------
502 inheritance_planner(Query *parse, List *inheritlist)
504 int parentRTindex = parse->resultRelation;
505 Oid parentOID = getrelid(parentRTindex, parse->rtable);
506 int mainrtlength = list_length(parse->rtable);
507 List *subplans = NIL;
511 foreach(l, inheritlist)
513 int childRTindex = lfirst_int(l);
514 Oid childOID = getrelid(childRTindex, parse->rtable);
519 /* Generate modified query with this rel as target */
520 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
521 parentRTindex, parentOID,
522 childRTindex, childOID);
524 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
525 subplans = lappend(subplans, subplan);
528 * It's possible that additional RTEs got added to the rangetable
529 * due to expansion of inherited source tables (see allpaths.c).
530 * If so, we must copy 'em back to the main parse tree's rtable.
532 * XXX my goodness this is ugly. Really need to think about ways to
533 * rein in planner's habit of scribbling on its input.
535 subrtlength = list_length(subquery->rtable);
536 if (subrtlength > mainrtlength)
540 subrt = list_copy_tail(subquery->rtable, mainrtlength);
541 parse->rtable = list_concat(parse->rtable, subrt);
542 mainrtlength = subrtlength;
544 /* Save preprocessed tlist from first rel for use in Append */
546 tlist = subplan->targetlist;
549 /* Save the target-relations list for the executor, too */
550 parse->resultRelations = inheritlist;
552 /* Mark result as unordered (probably unnecessary) */
553 parse->query_pathkeys = NIL;
555 return (Plan *) make_append(subplans, true, tlist);
558 /*--------------------
560 * Perform planning steps related to grouping, aggregation, etc.
561 * This primarily means adding top-level processing to the basic
562 * query plan produced by query_planner.
564 * parse is the querytree produced by the parser & rewriter.
565 * tuple_fraction is the fraction of tuples we expect will be retrieved
567 * tuple_fraction is interpreted as follows:
568 * 0: expect all tuples to be retrieved (normal case)
569 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
570 * from the plan to be retrieved
571 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
572 * expected to be retrieved (ie, a LIMIT specification)
574 * Returns a query plan. Also, parse->query_pathkeys is returned as the
575 * actual output ordering of the plan (in pathkey format).
576 *--------------------
579 grouping_planner(Query *parse, double tuple_fraction)
581 List *tlist = parse->targetList;
583 List *current_pathkeys;
586 if (parse->setOperations)
588 List *set_sortclauses;
591 * Construct the plan for set operations. The result will not
592 * need any work except perhaps a top-level sort and/or LIMIT.
594 result_plan = plan_set_operations(parse,
598 * Calculate pathkeys representing the sort order (if any) of the
599 * set operation's result. We have to do this before overwriting
600 * the sort key information...
602 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
603 result_plan->targetlist);
604 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
607 * We should not need to call preprocess_targetlist, since we must
608 * be in a SELECT query node. Instead, use the targetlist
609 * returned by plan_set_operations (since this tells whether it
610 * returned any resjunk columns!), and transfer any sort key
611 * information from the original tlist.
613 Assert(parse->commandType == CMD_SELECT);
615 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
618 * Can't handle FOR UPDATE here (parser should have checked
619 * already, but let's make sure).
623 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
624 errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
627 * Calculate pathkeys that represent result ordering requirements
629 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
631 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
635 /* No set operations, do regular planning */
637 List *group_pathkeys;
638 AttrNumber *groupColIdx = NULL;
639 bool need_tlist_eval = true;
641 double sub_tuple_fraction;
644 double dNumGroups = 0;
647 int numGroupCols = list_length(parse->groupClause);
648 bool use_hashed_grouping = false;
650 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
651 tlist = preprocess_targetlist(tlist,
653 parse->resultRelation,
657 * Add TID targets for rels selected FOR UPDATE (should this be
658 * done in preprocess_targetlist?). The executor uses the TID to
659 * know which rows to lock, much as for UPDATE or DELETE.
666 * We've got trouble if the FOR UPDATE appears inside
667 * grouping, since grouping renders a reference to individual
668 * tuple CTIDs invalid. This is also checked at parse time,
669 * but that's insufficient because of rule substitution, query
672 CheckSelectForUpdate(parse);
675 * Currently the executor only supports FOR UPDATE at top
678 if (PlannerQueryLevel > 1)
680 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
681 errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
683 foreach(l, parse->rowMarks)
685 Index rti = lfirst_int(l);
691 resname = (char *) palloc(32);
692 snprintf(resname, 32, "ctid%u", rti);
693 resdom = makeResdom(list_length(tlist) + 1,
700 SelfItemPointerAttributeNumber,
705 ctid = makeTargetEntry(resdom, (Expr *) var);
706 tlist = lappend(tlist, ctid);
711 * Generate appropriate target list for subplan; may be different
712 * from tlist if grouping or aggregation is needed.
714 sub_tlist = make_subplanTargetList(parse, tlist,
715 &groupColIdx, &need_tlist_eval);
718 * Calculate pathkeys that represent grouping/ordering
721 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
723 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
727 * Will need actual number of aggregates for estimating costs.
729 * Note: we do not attempt to detect duplicate aggregates here; a
730 * somewhat-overestimated count is okay for our present purposes.
732 * Note: think not that we can turn off hasAggs if we find no aggs.
733 * It is possible for constant-expression simplification to remove
734 * all explicit references to aggs, but we still have to follow the
735 * aggregate semantics (eg, producing only one output row).
738 numAggs = count_agg_clause((Node *) tlist) +
739 count_agg_clause(parse->havingQual);
742 * Figure out whether we need a sorted result from query_planner.
744 * If we have a GROUP BY clause, then we want a result sorted
745 * properly for grouping. Otherwise, if there is an ORDER BY
746 * clause, we want to sort by the ORDER BY clause. (Note: if we
747 * have both, and ORDER BY is a superset of GROUP BY, it would be
748 * tempting to request sort by ORDER BY --- but that might just
749 * leave us failing to exploit an available sort order at all.
750 * Needs more thought...)
752 if (parse->groupClause)
753 parse->query_pathkeys = group_pathkeys;
754 else if (parse->sortClause)
755 parse->query_pathkeys = sort_pathkeys;
757 parse->query_pathkeys = NIL;
760 * Adjust tuple_fraction if we see that we are going to apply
761 * limiting/grouping/aggregation/etc. This is not overridable by
762 * the caller, since it reflects plan actions that this routine
763 * will certainly take, not assumptions about context.
765 if (parse->limitCount != NULL)
768 * A LIMIT clause limits the absolute number of tuples
769 * returned. However, if it's not a constant LIMIT then we
770 * have to punt; for lack of a better idea, assume 10% of the
771 * plan's result is wanted.
773 double limit_fraction = 0.0;
775 if (IsA(parse->limitCount, Const))
777 Const *limitc = (Const *) parse->limitCount;
778 int32 count = DatumGetInt32(limitc->constvalue);
781 * A NULL-constant LIMIT represents "LIMIT ALL", which we
782 * treat the same as no limit (ie, expect to retrieve all
785 if (!limitc->constisnull && count > 0)
787 limit_fraction = (double) count;
788 /* We must also consider the OFFSET, if present */
789 if (parse->limitOffset != NULL)
791 if (IsA(parse->limitOffset, Const))
795 limitc = (Const *) parse->limitOffset;
796 offset = DatumGetInt32(limitc->constvalue);
797 if (!limitc->constisnull && offset > 0)
798 limit_fraction += (double) offset;
802 /* OFFSET is an expression ... punt ... */
803 limit_fraction = 0.10;
810 /* LIMIT is an expression ... punt ... */
811 limit_fraction = 0.10;
814 if (limit_fraction > 0.0)
817 * If we have absolute limits from both caller and LIMIT,
818 * use the smaller value; if one is fractional and the
819 * other absolute, treat the fraction as a fraction of the
820 * absolute value; else we can multiply the two fractions
823 if (tuple_fraction >= 1.0)
825 if (limit_fraction >= 1.0)
828 tuple_fraction = Min(tuple_fraction, limit_fraction);
832 /* caller absolute, limit fractional */
833 tuple_fraction *= limit_fraction;
834 if (tuple_fraction < 1.0)
835 tuple_fraction = 1.0;
838 else if (tuple_fraction > 0.0)
840 if (limit_fraction >= 1.0)
842 /* caller fractional, limit absolute */
843 tuple_fraction *= limit_fraction;
844 if (tuple_fraction < 1.0)
845 tuple_fraction = 1.0;
849 /* both fractional */
850 tuple_fraction *= limit_fraction;
855 /* no info from caller, just use limit */
856 tuple_fraction = limit_fraction;
862 * With grouping or aggregation, the tuple fraction to pass to
863 * query_planner() may be different from what it is at top level.
865 sub_tuple_fraction = tuple_fraction;
867 if (parse->groupClause)
870 * In GROUP BY mode, we have the little problem that we don't
871 * really know how many input tuples will be needed to make a
872 * group, so we can't translate an output LIMIT count into an
873 * input count. For lack of a better idea, assume 25% of the
874 * input data will be processed if there is any output limit.
875 * However, if the caller gave us a fraction rather than an
876 * absolute count, we can keep using that fraction (which
877 * amounts to assuming that all the groups are about the same
880 if (sub_tuple_fraction >= 1.0)
881 sub_tuple_fraction = 0.25;
884 * If both GROUP BY and ORDER BY are specified, we will need
885 * two levels of sort --- and, therefore, certainly need to
886 * read all the input tuples --- unless ORDER BY is a subset
887 * of GROUP BY. (We have not yet canonicalized the pathkeys,
888 * so must use the slower noncanonical comparison method.)
890 if (parse->groupClause && parse->sortClause &&
891 !noncanonical_pathkeys_contained_in(sort_pathkeys,
893 sub_tuple_fraction = 0.0;
895 else if (parse->hasAggs)
898 * Ungrouped aggregate will certainly want all the input
901 sub_tuple_fraction = 0.0;
903 else if (parse->distinctClause)
906 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
907 * number of input tuples per output tuple. Handle the same
910 if (sub_tuple_fraction >= 1.0)
911 sub_tuple_fraction = 0.25;
915 * Generate the best unsorted and presorted paths for this Query
916 * (but note there may not be any presorted path).
918 query_planner(parse, sub_tlist, sub_tuple_fraction,
919 &cheapest_path, &sorted_path);
922 * We couldn't canonicalize group_pathkeys and sort_pathkeys
923 * before running query_planner(), so do it now.
925 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
926 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
929 * Consider whether we might want to use hashed grouping.
931 if (parse->groupClause)
934 double cheapest_path_rows;
935 int cheapest_path_width;
938 * Beware in this section of the possibility that
939 * cheapest_path->parent is NULL. This could happen if user
940 * does something silly like SELECT 'foo' GROUP BY 1;
942 if (cheapest_path->parent)
944 cheapest_path_rows = cheapest_path->parent->rows;
945 cheapest_path_width = cheapest_path->parent->width;
949 cheapest_path_rows = 1; /* assume non-set result */
950 cheapest_path_width = 100; /* arbitrary */
954 * Always estimate the number of groups. We can't do this
955 * until after running query_planner(), either.
957 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
959 dNumGroups = estimate_num_groups(parse,
962 /* Also want it as a long int --- but 'ware overflow! */
963 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
966 * Check can't-do-it conditions, including whether the
967 * grouping operators are hashjoinable.
969 * Executor doesn't support hashed aggregation with DISTINCT
970 * aggregates. (Doing so would imply storing *all* the input
971 * values in the hash table, which seems like a certain
974 if (!enable_hashagg || !hash_safe_grouping(parse))
975 use_hashed_grouping = false;
976 else if (parse->hasAggs &&
977 (contain_distinct_agg_clause((Node *) tlist) ||
978 contain_distinct_agg_clause(parse->havingQual)))
979 use_hashed_grouping = false;
983 * Use hashed grouping if (a) we think we can fit the
984 * hashtable into work_mem, *and* (b) the estimated cost is
985 * no more than doing it the other way. While avoiding
986 * the need for sorted input is usually a win, the fact
987 * that the output won't be sorted may be a loss; so we
988 * need to do an actual cost comparison.
990 * In most cases we have no good way to estimate the size of
991 * the transition value needed by an aggregate;
992 * arbitrarily assume it is 100 bytes. Also set the
993 * overhead per hashtable entry at 64 bytes.
995 int hashentrysize = cheapest_path_width + 64 + numAggs * 100;
997 if (hashentrysize * dNumGroups <= work_mem * 1024L)
1000 * Okay, do the cost comparison. We need to consider
1001 * cheapest_path + hashagg [+ final sort] versus
1002 * either cheapest_path [+ sort] + group or agg [+
1003 * final sort] or presorted_path + group or agg [+
1004 * final sort] where brackets indicate a step that may
1005 * not be needed. We assume query_planner() will have
1006 * returned a presorted path only if it's a winner
1007 * compared to cheapest_path for this purpose.
1009 * These path variables are dummies that just hold cost
1010 * fields; we don't make actual Paths for these steps.
1015 cost_agg(&hashed_p, parse,
1016 AGG_HASHED, numAggs,
1017 numGroupCols, dNumGroups,
1018 cheapest_path->startup_cost,
1019 cheapest_path->total_cost,
1020 cheapest_path_rows);
1021 /* Result of hashed agg is always unsorted */
1023 cost_sort(&hashed_p, parse, sort_pathkeys,
1024 hashed_p.total_cost,
1026 cheapest_path_width);
1030 sorted_p.startup_cost = sorted_path->startup_cost;
1031 sorted_p.total_cost = sorted_path->total_cost;
1032 current_pathkeys = sorted_path->pathkeys;
1036 sorted_p.startup_cost = cheapest_path->startup_cost;
1037 sorted_p.total_cost = cheapest_path->total_cost;
1038 current_pathkeys = cheapest_path->pathkeys;
1040 if (!pathkeys_contained_in(group_pathkeys,
1043 cost_sort(&sorted_p, parse, group_pathkeys,
1044 sorted_p.total_cost,
1046 cheapest_path_width);
1047 current_pathkeys = group_pathkeys;
1050 cost_agg(&sorted_p, parse,
1051 AGG_SORTED, numAggs,
1052 numGroupCols, dNumGroups,
1053 sorted_p.startup_cost,
1054 sorted_p.total_cost,
1055 cheapest_path_rows);
1057 cost_group(&sorted_p, parse,
1058 numGroupCols, dNumGroups,
1059 sorted_p.startup_cost,
1060 sorted_p.total_cost,
1061 cheapest_path_rows);
1062 /* The Agg or Group node will preserve ordering */
1063 if (sort_pathkeys &&
1064 !pathkeys_contained_in(sort_pathkeys,
1067 cost_sort(&sorted_p, parse, sort_pathkeys,
1068 sorted_p.total_cost,
1070 cheapest_path_width);
1074 * Now make the decision using the top-level tuple
1075 * fraction. First we have to convert an absolute
1076 * count (LIMIT) into fractional form.
1078 if (tuple_fraction >= 1.0)
1079 tuple_fraction /= dNumGroups;
1081 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1082 tuple_fraction) < 0)
1084 /* Hashed is cheaper, so use it */
1085 use_hashed_grouping = true;
1092 * Select the best path and create a plan to execute it.
1094 * If we are doing hashed grouping, we will always read all the input
1095 * tuples, so use the cheapest-total path. Otherwise, trust
1096 * query_planner's decision about which to use.
1098 if (sorted_path && !use_hashed_grouping)
1100 result_plan = create_plan(parse, sorted_path);
1101 current_pathkeys = sorted_path->pathkeys;
1105 result_plan = create_plan(parse, cheapest_path);
1106 current_pathkeys = cheapest_path->pathkeys;
1110 * create_plan() returns a plan with just a "flat" tlist of
1111 * required Vars. Usually we need to insert the sub_tlist as the
1112 * tlist of the top plan node. However, we can skip that if we
1113 * determined that whatever query_planner chose to return will be
1116 if (need_tlist_eval)
1119 * If the top-level plan node is one that cannot do expression
1120 * evaluation, we must insert a Result node to project the
1123 if (!is_projection_capable_plan(result_plan))
1125 result_plan = (Plan *) make_result(sub_tlist, NULL,
1131 * Otherwise, just replace the subplan's flat tlist with
1132 * the desired tlist.
1134 result_plan->targetlist = sub_tlist;
1138 * Also, account for the cost of evaluation of the sub_tlist.
1140 * Up to now, we have only been dealing with "flat" tlists,
1141 * containing just Vars. So their evaluation cost is zero
1142 * according to the model used by cost_qual_eval() (or if you
1143 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1144 * can avoid accounting for tlist cost throughout
1145 * query_planner() and subroutines. But now we've inserted a
1146 * tlist that might contain actual operators, sub-selects, etc
1147 * --- so we'd better account for its cost.
1149 * Below this point, any tlist eval cost for added-on nodes
1150 * should be accounted for as we create those nodes.
1151 * Presently, of the node types we can add on, only Agg and
1152 * Group project new tlists (the rest just copy their input
1153 * tuples) --- so make_agg() and make_group() are responsible
1154 * for computing the added cost.
1156 cost_qual_eval(&tlist_cost, sub_tlist);
1157 result_plan->startup_cost += tlist_cost.startup;
1158 result_plan->total_cost += tlist_cost.startup +
1159 tlist_cost.per_tuple * result_plan->plan_rows;
1164 * Since we're using query_planner's tlist and not the one
1165 * make_subplanTargetList calculated, we have to refigure any
1166 * grouping-column indexes make_subplanTargetList computed.
1168 locate_grouping_columns(parse, tlist, result_plan->targetlist,
1173 * Insert AGG or GROUP node if needed, plus an explicit sort step
1176 * HAVING clause, if any, becomes qual of the Agg node
1178 if (use_hashed_grouping)
1180 /* Hashed aggregate plan --- no sort needed */
1181 result_plan = (Plan *) make_agg(parse,
1183 (List *) parse->havingQual,
1190 /* Hashed aggregation produces randomly-ordered results */
1191 current_pathkeys = NIL;
1193 else if (parse->hasAggs)
1195 /* Plain aggregate plan --- sort if needed */
1196 AggStrategy aggstrategy;
1198 if (parse->groupClause)
1200 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1202 result_plan = (Plan *)
1203 make_sort_from_groupcols(parse,
1207 current_pathkeys = group_pathkeys;
1209 aggstrategy = AGG_SORTED;
1212 * The AGG node will not change the sort ordering of its
1213 * groups, so current_pathkeys describes the result too.
1218 aggstrategy = AGG_PLAIN;
1219 /* Result will be only one row anyway; no sort order */
1220 current_pathkeys = NIL;
1223 result_plan = (Plan *) make_agg(parse,
1225 (List *) parse->havingQual,
1236 * If there are no Aggs, we shouldn't have any HAVING qual
1239 Assert(parse->havingQual == NULL);
1242 * If we have a GROUP BY clause, insert a group node (plus the
1243 * appropriate sort node, if necessary).
1245 if (parse->groupClause)
1248 * Add an explicit sort if we couldn't make the path come
1249 * out the way the GROUP node needs it.
1251 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1253 result_plan = (Plan *)
1254 make_sort_from_groupcols(parse,
1258 current_pathkeys = group_pathkeys;
1261 result_plan = (Plan *) make_group(parse,
1267 /* The Group node won't change sort ordering */
1270 } /* end of if (setOperations) */
1273 * If we were not able to make the plan come out in the right order,
1274 * add an explicit sort step.
1276 if (parse->sortClause)
1278 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1280 result_plan = (Plan *)
1281 make_sort_from_sortclauses(parse,
1284 current_pathkeys = sort_pathkeys;
1289 * If there is a DISTINCT clause, add the UNIQUE node.
1291 if (parse->distinctClause)
1293 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1296 * If there was grouping or aggregation, leave plan_rows as-is
1297 * (ie, assume the result was already mostly unique). If not,
1298 * it's reasonable to assume the UNIQUE filter has effects
1299 * comparable to GROUP BY.
1301 if (!parse->groupClause && !parse->hasAggs)
1303 List *distinctExprs;
1305 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1307 result_plan->plan_rows = estimate_num_groups(parse,
1309 result_plan->plan_rows);
1314 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1316 if (parse->limitOffset || parse->limitCount)
1318 result_plan = (Plan *) make_limit(result_plan,
1324 * Return the actual output ordering in query_pathkeys for possible
1325 * use by an outer query level.
1327 parse->query_pathkeys = current_pathkeys;
1333 * hash_safe_grouping - are grouping operators hashable?
1335 * We assume hashed aggregation will work if the datatype's equality operator
1336 * is marked hashjoinable.
1339 hash_safe_grouping(Query *parse)
1343 foreach(gl, parse->groupClause)
1345 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1346 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1350 optup = equality_oper(tle->resdom->restype, true);
1353 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1354 ReleaseSysCache(optup);
1362 * make_subplanTargetList
1363 * Generate appropriate target list when grouping is required.
1365 * When grouping_planner inserts Aggregate or Group plan nodes above
1366 * the result of query_planner, we typically want to pass a different
1367 * target list to query_planner than the outer plan nodes should have.
1368 * This routine generates the correct target list for the subplan.
1370 * The initial target list passed from the parser already contains entries
1371 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1372 * for variables used only in HAVING clauses; so we need to add those
1373 * variables to the subplan target list. Also, if we are doing either
1374 * grouping or aggregation, we flatten all expressions except GROUP BY items
1375 * into their component variables; the other expressions will be computed by
1376 * the inserted nodes rather than by the subplan. For example,
1377 * given a query like
1378 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1379 * we want to pass this targetlist to the subplan:
1381 * where the a+b target will be used by the Sort/Group steps, and the
1382 * other targets will be used for computing the final results. (In the
1383 * above example we could theoretically suppress the a and b targets and
1384 * pass down only c,d,a+b, but it's not really worth the trouble to
1385 * eliminate simple var references from the subplan. We will avoid doing
1386 * the extra computation to recompute a+b at the outer level; see
1387 * replace_vars_with_subplan_refs() in setrefs.c.)
1389 * If we are grouping or aggregating, *and* there are no non-Var grouping
1390 * expressions, then the returned tlist is effectively dummy; we do not
1391 * need to force it to be evaluated, because all the Vars it contains
1392 * should be present in the output of query_planner anyway.
1394 * 'parse' is the query being processed.
1395 * 'tlist' is the query's target list.
1396 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1397 * expressions (if there are any) in the subplan's target list.
1398 * 'need_tlist_eval' is set true if we really need to evaluate the
1401 * The result is the targetlist to be passed to the subplan.
1405 make_subplanTargetList(Query *parse,
1407 AttrNumber **groupColIdx,
1408 bool *need_tlist_eval)
1414 *groupColIdx = NULL;
1417 * If we're not grouping or aggregating, nothing to do here;
1418 * query_planner should receive the unmodified target list.
1420 if (!parse->hasAggs && !parse->groupClause)
1422 *need_tlist_eval = true;
1427 * Otherwise, start with a "flattened" tlist (having just the vars
1428 * mentioned in the targetlist and HAVING qual --- but not upper-
1429 * level Vars; they will be replaced by Params later on).
1431 sub_tlist = flatten_tlist(tlist);
1432 extravars = pull_var_clause(parse->havingQual, false);
1433 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1434 list_free(extravars);
1435 *need_tlist_eval = false; /* only eval if not flat tlist */
1438 * If grouping, create sub_tlist entries for all GROUP BY expressions
1439 * (GROUP BY items that are simple Vars should be in the list
1440 * already), and make an array showing where the group columns are in
1443 numCols = list_length(parse->groupClause);
1447 AttrNumber *grpColIdx;
1450 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1451 *groupColIdx = grpColIdx;
1453 foreach(gl, parse->groupClause)
1455 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1456 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1457 TargetEntry *te = NULL;
1460 /* Find or make a matching sub_tlist entry */
1461 foreach(sl, sub_tlist)
1463 te = (TargetEntry *) lfirst(sl);
1464 if (equal(groupexpr, te->expr))
1469 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1470 exprType(groupexpr),
1471 exprTypmod(groupexpr),
1474 (Expr *) groupexpr);
1475 sub_tlist = lappend(sub_tlist, te);
1476 *need_tlist_eval = true; /* it's not flat anymore */
1479 /* and save its resno */
1480 grpColIdx[keyno++] = te->resdom->resno;
1488 * locate_grouping_columns
1489 * Locate grouping columns in the tlist chosen by query_planner.
1491 * This is only needed if we don't use the sub_tlist chosen by
1492 * make_subplanTargetList. We have to forget the column indexes found
1493 * by that routine and re-locate the grouping vars in the real sub_tlist.
1496 locate_grouping_columns(Query *parse,
1499 AttrNumber *groupColIdx)
1505 * No work unless grouping.
1507 if (!parse->groupClause)
1509 Assert(groupColIdx == NULL);
1512 Assert(groupColIdx != NULL);
1514 foreach(gl, parse->groupClause)
1516 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1517 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1518 TargetEntry *te = NULL;
1521 foreach(sl, sub_tlist)
1523 te = (TargetEntry *) lfirst(sl);
1524 if (equal(groupexpr, te->expr))
1528 elog(ERROR, "failed to locate grouping columns");
1530 groupColIdx[keyno++] = te->resdom->resno;
1535 * postprocess_setop_tlist
1536 * Fix up targetlist returned by plan_set_operations().
1538 * We need to transpose sort key info from the orig_tlist into new_tlist.
1539 * NOTE: this would not be good enough if we supported resjunk sort keys
1540 * for results of set operations --- then, we'd need to project a whole
1541 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1542 * find any resjunk columns in orig_tlist.
1545 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1548 ListCell *orig_tlist_item = list_head(orig_tlist);
1550 foreach(l, new_tlist)
1552 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1553 TargetEntry *orig_tle;
1555 /* ignore resjunk columns in setop result */
1556 if (new_tle->resdom->resjunk)
1559 Assert(orig_tlist_item != NULL);
1560 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1561 orig_tlist_item = lnext(orig_tlist_item);
1562 if (orig_tle->resdom->resjunk) /* should not happen */
1563 elog(ERROR, "resjunk output columns are not implemented");
1564 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1565 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1566 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1568 if (orig_tlist_item != NULL)
1569 elog(ERROR, "resjunk output columns are not implemented");