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 1.175 2004/08/30 02:54:38 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
92 * structure since their whole purpose is communication across
93 * multiple sub-Queries. Also, boundParams is explicitly info from
94 * outside the Query, and so is likewise better handled as a global
97 * Note we do NOT save and restore PlannerPlanId: it exists to assign
98 * unique IDs to SubPlan nodes, and we want those IDs to be unique for
99 * the life of a backend. Also, PlannerInitPlan is saved/restored in
100 * subquery_planner, not here.
102 save_PlannerQueryLevel = PlannerQueryLevel;
103 save_PlannerParamList = PlannerParamList;
104 save_PlannerBoundParamList = PlannerBoundParamList;
106 /* Initialize state for handling outer-level references and params */
107 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
108 PlannerParamList = NIL;
109 PlannerBoundParamList = boundParams;
111 /* Determine what fraction of the plan is likely to be scanned */
115 * We have no real idea how many tuples the user will ultimately
116 * FETCH from a cursor, but it seems a good bet that he doesn't
117 * want 'em all. Optimize for 10% retrieval (you gotta better
118 * number? Should this be a SETtable parameter?)
120 tuple_fraction = 0.10;
124 /* Default assumption is we need all the tuples */
125 tuple_fraction = 0.0;
128 /* primary planning entry point (may recurse for subqueries) */
129 result_plan = subquery_planner(parse, tuple_fraction);
131 Assert(PlannerQueryLevel == 0);
134 * If creating a plan for a scrollable cursor, make sure it can run
135 * backwards on demand. Add a Material node at the top at need.
137 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
139 if (!ExecSupportsBackwardScan(result_plan))
140 result_plan = materialize_finished_plan(result_plan);
143 /* executor wants to know total number of Params used overall */
144 result_plan->nParamExec = list_length(PlannerParamList);
146 /* final cleanup of the plan */
147 set_plan_references(result_plan, parse->rtable);
149 /* restore state for outer planner, if any */
150 PlannerQueryLevel = save_PlannerQueryLevel;
151 PlannerParamList = save_PlannerParamList;
152 PlannerBoundParamList = save_PlannerBoundParamList;
158 /*--------------------
160 * Invokes the planner on a subquery. We recurse to here for each
161 * sub-SELECT found in the query tree.
163 * parse is the querytree produced by the parser & rewriter.
164 * tuple_fraction is the fraction of tuples we expect will be retrieved.
165 * tuple_fraction is interpreted as explained for grouping_planner, below.
167 * Basically, this routine does the stuff that should only be done once
168 * per Query object. It then calls grouping_planner. At one time,
169 * grouping_planner could be invoked recursively on the same Query object;
170 * that's not currently true, but we keep the separation between the two
171 * routines anyway, in case we need it again someday.
173 * subquery_planner will be called recursively to handle sub-Query nodes
174 * found within the query's expressions and rangetable.
176 * Returns a query plan.
177 *--------------------
180 subquery_planner(Query *parse, double tuple_fraction)
182 List *saved_initplan = PlannerInitPlan;
183 int saved_planid = PlannerPlanId;
190 /* Set up for a new level of subquery */
192 PlannerInitPlan = NIL;
195 * Look for IN clauses at the top level of WHERE, and transform them
196 * into joins. Note that this step only handles IN clauses originally
197 * at top level of WHERE; if we pull up any subqueries in the next
198 * step, their INs are processed just before pulling them up.
200 parse->in_info_list = NIL;
201 if (parse->hasSubLinks)
202 parse->jointree->quals = pull_up_IN_clauses(parse,
203 parse->jointree->quals);
206 * Check to see if any subqueries in the rangetable can be merged into
209 parse->jointree = (FromExpr *)
210 pull_up_subqueries(parse, (Node *) parse->jointree, false);
213 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
214 * can avoid the expense of doing flatten_join_alias_vars(). Also
215 * check for outer joins --- if none, we can skip
216 * reduce_outer_joins(). This must be done after we have done
217 * pull_up_subqueries, of course.
219 parse->hasJoinRTEs = false;
220 hasOuterJoins = false;
221 foreach(l, parse->rtable)
223 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
225 if (rte->rtekind == RTE_JOIN)
227 parse->hasJoinRTEs = true;
228 if (IS_OUTER_JOIN(rte->jointype))
230 hasOuterJoins = true;
231 /* Can quit scanning once we find an outer join */
238 * Do expression preprocessing on targetlist and quals.
240 parse->targetList = (List *)
241 preprocess_expression(parse, (Node *) parse->targetList,
244 preprocess_qual_conditions(parse, (Node *) parse->jointree);
246 parse->havingQual = preprocess_expression(parse, parse->havingQual,
249 parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
251 parse->limitCount = preprocess_expression(parse, parse->limitCount,
254 parse->in_info_list = (List *)
255 preprocess_expression(parse, (Node *) parse->in_info_list,
258 /* Also need to preprocess expressions for function RTEs */
259 foreach(l, parse->rtable)
261 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
263 if (rte->rtekind == RTE_FUNCTION)
264 rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
269 * A HAVING clause without aggregates is equivalent to a WHERE clause
270 * (except it can only refer to grouped fields). Transfer any
271 * agg-free clauses of the HAVING qual into WHERE. This may seem like
272 * wasting cycles to cater to stupidly-written queries, but there are
273 * other reasons for doing it. Firstly, if the query contains no aggs
274 * at all, then we aren't going to generate an Agg plan node, and so
275 * there'll be no place to execute HAVING conditions; without this
276 * transfer, we'd lose the HAVING condition entirely, which is wrong.
277 * Secondly, when we push down a qual condition into a sub-query, it's
278 * easiest to push the qual into HAVING always, in case it contains
279 * aggs, and then let this code sort it out.
281 * Note that both havingQual and parse->jointree->quals are in
282 * implicitly-ANDed-list form at this point, even though they are
283 * declared as Node *.
286 foreach(l, (List *) parse->havingQual)
288 Node *havingclause = (Node *) lfirst(l);
290 if (contain_agg_clause(havingclause))
291 newHaving = lappend(newHaving, havingclause);
293 parse->jointree->quals = (Node *)
294 lappend((List *) parse->jointree->quals, havingclause);
296 parse->havingQual = (Node *) newHaving;
299 * If we have any outer joins, try to reduce them to plain inner
300 * joins. This step is most easily done after we've done expression
304 reduce_outer_joins(parse);
307 * See if we can simplify the jointree; opportunities for this may
308 * come from having pulled up subqueries, or from flattening explicit
309 * JOIN syntax. We must do this after flattening JOIN alias
310 * variables, since eliminating explicit JOIN nodes from the jointree
311 * will cause get_relids_for_join() to fail. But it should happen
312 * after reduce_outer_joins, anyway.
314 parse->jointree = (FromExpr *)
315 simplify_jointree(parse, (Node *) parse->jointree);
318 * Do the main planning. If we have an inherited target relation,
319 * that needs special processing, else go straight to
322 if (parse->resultRelation &&
323 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
325 plan = inheritance_planner(parse, lst);
327 plan = grouping_planner(parse, tuple_fraction);
330 * If any subplans were generated, or if we're inside a subplan, build
331 * initPlan list and extParam/allParam sets for plan nodes.
333 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
335 Cost initplan_cost = 0;
337 /* Prepare extParam/allParam sets for all nodes in tree */
338 SS_finalize_plan(plan, parse->rtable);
341 * SS_finalize_plan doesn't handle initPlans, so we have to
342 * manually attach them to the topmost plan node, and add their
343 * extParams to the topmost node's, too.
345 * We also add the total_cost of each initPlan to the startup cost of
346 * the top node. This is a conservative overestimate, since in
347 * fact each initPlan might be executed later than plan startup,
348 * or even not at all.
350 plan->initPlan = PlannerInitPlan;
352 foreach(l, plan->initPlan)
354 SubPlan *initplan = (SubPlan *) lfirst(l);
356 plan->extParam = bms_add_members(plan->extParam,
357 initplan->plan->extParam);
358 /* allParam must include all members of extParam */
359 plan->allParam = bms_add_members(plan->allParam,
361 initplan_cost += initplan->plan->total_cost;
364 plan->startup_cost += initplan_cost;
365 plan->total_cost += initplan_cost;
368 /* Return to outer subquery context */
370 PlannerInitPlan = saved_initplan;
371 /* we do NOT restore PlannerPlanId; that's not an oversight! */
377 * preprocess_expression
378 * Do subquery_planner's preprocessing work for an expression,
379 * which can be a targetlist, a WHERE clause (including JOIN/ON
380 * conditions), or a HAVING clause.
383 preprocess_expression(Query *parse, Node *expr, int kind)
386 * If the query has any join RTEs, replace join alias variables with
387 * base-relation variables. We must do this before sublink processing,
388 * else sublinks expanded out from join aliases wouldn't get
391 if (parse->hasJoinRTEs)
392 expr = flatten_join_alias_vars(parse, expr);
395 * If it's a qual or havingQual, canonicalize it. It seems most
396 * useful to do this before applying eval_const_expressions, since the
397 * latter can optimize flattened AND/ORs better than unflattened ones.
399 * Note: all processing of a qual expression after this point must be
400 * careful to maintain AND/OR flatness --- that is, do not generate a
401 * tree with AND directly under AND, nor OR directly under OR.
403 if (kind == EXPRKIND_QUAL)
405 expr = (Node *) canonicalize_qual((Expr *) expr);
407 #ifdef OPTIMIZER_DEBUG
408 printf("After canonicalize_qual()\n");
414 * Simplify constant expressions.
416 expr = eval_const_expressions(expr);
418 /* Expand SubLinks to SubPlans */
419 if (parse->hasSubLinks)
420 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
423 * XXX do not insert anything here unless you have grokked the
424 * comments in SS_replace_correlation_vars ...
427 /* Replace uplevel vars with Param nodes */
428 if (PlannerQueryLevel > 1)
429 expr = SS_replace_correlation_vars(expr);
432 * If it's a qual or havingQual, convert it to implicit-AND format.
433 * (We don't want to do this before eval_const_expressions, since the
434 * latter would be unable to simplify a top-level AND correctly. Also,
435 * SS_process_sublinks expects explicit-AND format.)
437 if (kind == EXPRKIND_QUAL)
438 expr = (Node *) make_ands_implicit((Expr *) expr);
444 * preprocess_qual_conditions
445 * Recursively scan the query's jointree and do subquery_planner's
446 * preprocessing work on each qual condition found therein.
449 preprocess_qual_conditions(Query *parse, Node *jtnode)
453 if (IsA(jtnode, RangeTblRef))
455 /* nothing to do here */
457 else if (IsA(jtnode, FromExpr))
459 FromExpr *f = (FromExpr *) jtnode;
462 foreach(l, f->fromlist)
463 preprocess_qual_conditions(parse, lfirst(l));
465 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
467 else if (IsA(jtnode, JoinExpr))
469 JoinExpr *j = (JoinExpr *) jtnode;
471 preprocess_qual_conditions(parse, j->larg);
472 preprocess_qual_conditions(parse, j->rarg);
474 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
477 elog(ERROR, "unrecognized node type: %d",
478 (int) nodeTag(jtnode));
481 /*--------------------
482 * inheritance_planner
483 * Generate a plan in the case where the result relation is an
486 * We have to handle this case differently from cases where a source
487 * relation is an inheritance set. Source inheritance is expanded at
488 * the bottom of the plan tree (see allpaths.c), but target inheritance
489 * has to be expanded at the top. The reason is that for UPDATE, each
490 * target relation needs a different targetlist matching its own column
491 * set. (This is not so critical for DELETE, but for simplicity we treat
492 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
493 * can never be the nullable side of an outer join, so it's OK to generate
496 * parse is the querytree produced by the parser & rewriter.
497 * inheritlist is an integer list of RT indexes for the result relation set.
499 * Returns a query plan.
500 *--------------------
503 inheritance_planner(Query *parse, List *inheritlist)
505 int parentRTindex = parse->resultRelation;
506 Oid parentOID = getrelid(parentRTindex, parse->rtable);
507 int mainrtlength = list_length(parse->rtable);
508 List *subplans = NIL;
512 foreach(l, inheritlist)
514 int childRTindex = lfirst_int(l);
515 Oid childOID = getrelid(childRTindex, parse->rtable);
520 /* Generate modified query with this rel as target */
521 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
522 parentRTindex, parentOID,
523 childRTindex, childOID);
525 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
526 subplans = lappend(subplans, subplan);
529 * It's possible that additional RTEs got added to the rangetable
530 * due to expansion of inherited source tables (see allpaths.c).
531 * If so, we must copy 'em back to the main parse tree's rtable.
533 * XXX my goodness this is ugly. Really need to think about ways to
534 * rein in planner's habit of scribbling on its input.
536 subrtlength = list_length(subquery->rtable);
537 if (subrtlength > mainrtlength)
541 subrt = list_copy_tail(subquery->rtable, mainrtlength);
542 parse->rtable = list_concat(parse->rtable, subrt);
543 mainrtlength = subrtlength;
545 /* Save preprocessed tlist from first rel for use in Append */
547 tlist = subplan->targetlist;
550 /* Save the target-relations list for the executor, too */
551 parse->resultRelations = inheritlist;
553 /* Mark result as unordered (probably unnecessary) */
554 parse->query_pathkeys = NIL;
556 return (Plan *) make_append(subplans, true, tlist);
559 /*--------------------
561 * Perform planning steps related to grouping, aggregation, etc.
562 * This primarily means adding top-level processing to the basic
563 * query plan produced by query_planner.
565 * parse is the querytree produced by the parser & rewriter.
566 * tuple_fraction is the fraction of tuples we expect will be retrieved
568 * tuple_fraction is interpreted as follows:
569 * 0: expect all tuples to be retrieved (normal case)
570 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
571 * from the plan to be retrieved
572 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
573 * expected to be retrieved (ie, a LIMIT specification)
575 * Returns a query plan. Also, parse->query_pathkeys is returned as the
576 * actual output ordering of the plan (in pathkey format).
577 *--------------------
580 grouping_planner(Query *parse, double tuple_fraction)
582 List *tlist = parse->targetList;
584 List *current_pathkeys;
587 if (parse->setOperations)
589 List *set_sortclauses;
592 * Construct the plan for set operations. The result will not
593 * need any work except perhaps a top-level sort and/or LIMIT.
595 result_plan = plan_set_operations(parse,
599 * Calculate pathkeys representing the sort order (if any) of the
600 * set operation's result. We have to do this before overwriting
601 * the sort key information...
603 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
604 result_plan->targetlist);
605 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
608 * We should not need to call preprocess_targetlist, since we must
609 * be in a SELECT query node. Instead, use the targetlist
610 * returned by plan_set_operations (since this tells whether it
611 * returned any resjunk columns!), and transfer any sort key
612 * information from the original tlist.
614 Assert(parse->commandType == CMD_SELECT);
616 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
619 * Can't handle FOR UPDATE here (parser should have checked
620 * already, but let's make sure).
624 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
625 errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
628 * Calculate pathkeys that represent result ordering requirements
630 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
632 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
636 /* No set operations, do regular planning */
638 List *group_pathkeys;
639 AttrNumber *groupColIdx = NULL;
640 bool need_tlist_eval = true;
642 double sub_tuple_fraction;
645 double dNumGroups = 0;
648 int numGroupCols = list_length(parse->groupClause);
649 bool use_hashed_grouping = false;
651 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
652 tlist = preprocess_targetlist(tlist,
654 parse->resultRelation,
658 * Add TID targets for rels selected FOR UPDATE (should this be
659 * done in preprocess_targetlist?). The executor uses the TID to
660 * know which rows to lock, much as for UPDATE or DELETE.
667 * We've got trouble if the FOR UPDATE appears inside
668 * grouping, since grouping renders a reference to individual
669 * tuple CTIDs invalid. This is also checked at parse time,
670 * but that's insufficient because of rule substitution, query
673 CheckSelectForUpdate(parse);
676 * Currently the executor only supports FOR UPDATE at top
679 if (PlannerQueryLevel > 1)
681 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
682 errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
684 foreach(l, parse->rowMarks)
686 Index rti = lfirst_int(l);
692 resname = (char *) palloc(32);
693 snprintf(resname, 32, "ctid%u", rti);
694 resdom = makeResdom(list_length(tlist) + 1,
701 SelfItemPointerAttributeNumber,
706 ctid = makeTargetEntry(resdom, (Expr *) var);
707 tlist = lappend(tlist, ctid);
712 * Generate appropriate target list for subplan; may be different
713 * from tlist if grouping or aggregation is needed.
715 sub_tlist = make_subplanTargetList(parse, tlist,
716 &groupColIdx, &need_tlist_eval);
719 * Calculate pathkeys that represent grouping/ordering
722 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
724 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
728 * Will need actual number of aggregates for estimating costs.
730 * Note: we do not attempt to detect duplicate aggregates here; a
731 * somewhat-overestimated count is okay for our present purposes.
733 * Note: think not that we can turn off hasAggs if we find no aggs.
734 * It is possible for constant-expression simplification to remove
735 * all explicit references to aggs, but we still have to follow
736 * the aggregate semantics (eg, producing only one output row).
739 numAggs = count_agg_clause((Node *) tlist) +
740 count_agg_clause(parse->havingQual);
743 * Figure out whether we need a sorted result from query_planner.
745 * If we have a GROUP BY clause, then we want a result sorted
746 * properly for grouping. Otherwise, if there is an ORDER BY
747 * clause, we want to sort by the ORDER BY clause. (Note: if we
748 * have both, and ORDER BY is a superset of GROUP BY, it would be
749 * tempting to request sort by ORDER BY --- but that might just
750 * leave us failing to exploit an available sort order at all.
751 * Needs more thought...)
753 if (parse->groupClause)
754 parse->query_pathkeys = group_pathkeys;
755 else if (parse->sortClause)
756 parse->query_pathkeys = sort_pathkeys;
758 parse->query_pathkeys = NIL;
761 * Adjust tuple_fraction if we see that we are going to apply
762 * limiting/grouping/aggregation/etc. This is not overridable by
763 * the caller, since it reflects plan actions that this routine
764 * will certainly take, not assumptions about context.
766 if (parse->limitCount != NULL)
769 * A LIMIT clause limits the absolute number of tuples
770 * returned. However, if it's not a constant LIMIT then we
771 * have to punt; for lack of a better idea, assume 10% of the
772 * plan's result is wanted.
774 double limit_fraction = 0.0;
776 if (IsA(parse->limitCount, Const))
778 Const *limitc = (Const *) parse->limitCount;
779 int32 count = DatumGetInt32(limitc->constvalue);
782 * A NULL-constant LIMIT represents "LIMIT ALL", which we
783 * treat the same as no limit (ie, expect to retrieve all
786 if (!limitc->constisnull && count > 0)
788 limit_fraction = (double) count;
789 /* We must also consider the OFFSET, if present */
790 if (parse->limitOffset != NULL)
792 if (IsA(parse->limitOffset, Const))
796 limitc = (Const *) parse->limitOffset;
797 offset = DatumGetInt32(limitc->constvalue);
798 if (!limitc->constisnull && offset > 0)
799 limit_fraction += (double) offset;
803 /* OFFSET is an expression ... punt ... */
804 limit_fraction = 0.10;
811 /* LIMIT is an expression ... punt ... */
812 limit_fraction = 0.10;
815 if (limit_fraction > 0.0)
818 * If we have absolute limits from both caller and LIMIT,
819 * use the smaller value; if one is fractional and the
820 * other absolute, treat the fraction as a fraction of the
821 * absolute value; else we can multiply the two fractions
824 if (tuple_fraction >= 1.0)
826 if (limit_fraction >= 1.0)
829 tuple_fraction = Min(tuple_fraction, limit_fraction);
833 /* caller absolute, limit fractional */
834 tuple_fraction *= limit_fraction;
835 if (tuple_fraction < 1.0)
836 tuple_fraction = 1.0;
839 else if (tuple_fraction > 0.0)
841 if (limit_fraction >= 1.0)
843 /* caller fractional, limit absolute */
844 tuple_fraction *= limit_fraction;
845 if (tuple_fraction < 1.0)
846 tuple_fraction = 1.0;
850 /* both fractional */
851 tuple_fraction *= limit_fraction;
856 /* no info from caller, just use limit */
857 tuple_fraction = limit_fraction;
863 * With grouping or aggregation, the tuple fraction to pass to
864 * query_planner() may be different from what it is at top level.
866 sub_tuple_fraction = tuple_fraction;
868 if (parse->groupClause)
871 * In GROUP BY mode, we have the little problem that we don't
872 * really know how many input tuples will be needed to make a
873 * group, so we can't translate an output LIMIT count into an
874 * input count. For lack of a better idea, assume 25% of the
875 * input data will be processed if there is any output limit.
876 * However, if the caller gave us a fraction rather than an
877 * absolute count, we can keep using that fraction (which
878 * amounts to assuming that all the groups are about the same
881 if (sub_tuple_fraction >= 1.0)
882 sub_tuple_fraction = 0.25;
885 * If both GROUP BY and ORDER BY are specified, we will need
886 * two levels of sort --- and, therefore, certainly need to
887 * read all the input tuples --- unless ORDER BY is a subset
888 * of GROUP BY. (We have not yet canonicalized the pathkeys,
889 * so must use the slower noncanonical comparison method.)
891 if (parse->groupClause && parse->sortClause &&
892 !noncanonical_pathkeys_contained_in(sort_pathkeys,
894 sub_tuple_fraction = 0.0;
896 else if (parse->hasAggs)
899 * Ungrouped aggregate will certainly want all the input
902 sub_tuple_fraction = 0.0;
904 else if (parse->distinctClause)
907 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
908 * number of input tuples per output tuple. Handle the same
911 if (sub_tuple_fraction >= 1.0)
912 sub_tuple_fraction = 0.25;
916 * Generate the best unsorted and presorted paths for this Query
917 * (but note there may not be any presorted path).
919 query_planner(parse, sub_tlist, sub_tuple_fraction,
920 &cheapest_path, &sorted_path);
923 * We couldn't canonicalize group_pathkeys and sort_pathkeys
924 * before running query_planner(), so do it now.
926 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
927 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
930 * Consider whether we might want to use hashed grouping.
932 if (parse->groupClause)
935 double cheapest_path_rows;
936 int cheapest_path_width;
939 * Beware in this section of the possibility that
940 * cheapest_path->parent is NULL. This could happen if user
941 * does something silly like SELECT 'foo' GROUP BY 1;
943 if (cheapest_path->parent)
945 cheapest_path_rows = cheapest_path->parent->rows;
946 cheapest_path_width = cheapest_path->parent->width;
950 cheapest_path_rows = 1; /* assume non-set result */
951 cheapest_path_width = 100; /* arbitrary */
955 * Always estimate the number of groups. We can't do this
956 * until after running query_planner(), either.
958 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
960 dNumGroups = estimate_num_groups(parse,
963 /* Also want it as a long int --- but 'ware overflow! */
964 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
967 * Check can't-do-it conditions, including whether the
968 * grouping operators are hashjoinable.
970 * Executor doesn't support hashed aggregation with DISTINCT
971 * aggregates. (Doing so would imply storing *all* the input
972 * values in the hash table, which seems like a certain
975 if (!enable_hashagg || !hash_safe_grouping(parse))
976 use_hashed_grouping = false;
977 else if (parse->hasAggs &&
978 (contain_distinct_agg_clause((Node *) tlist) ||
979 contain_distinct_agg_clause(parse->havingQual)))
980 use_hashed_grouping = false;
984 * Use hashed grouping if (a) we think we can fit the
985 * hashtable into work_mem, *and* (b) the estimated cost
986 * is no more than doing it the other way. While avoiding
987 * the need for sorted input is usually a win, the fact
988 * that the output won't be sorted may be a loss; so we
989 * need to do an actual cost comparison.
991 * In most cases we have no good way to estimate the size of
992 * the transition value needed by an aggregate;
993 * arbitrarily assume it is 100 bytes. Also set the
994 * overhead per hashtable entry at 64 bytes.
996 int hashentrysize = cheapest_path_width + 64 + numAggs * 100;
998 if (hashentrysize * dNumGroups <= work_mem * 1024L)
1001 * Okay, do the cost comparison. We need to consider
1002 * cheapest_path + hashagg [+ final sort] versus
1003 * either cheapest_path [+ sort] + group or agg [+
1004 * final sort] or presorted_path + group or agg [+
1005 * final sort] where brackets indicate a step that may
1006 * not be needed. We assume query_planner() will have
1007 * returned a presorted path only if it's a winner
1008 * compared to cheapest_path for this purpose.
1010 * These path variables are dummies that just hold cost
1011 * fields; we don't make actual Paths for these steps.
1016 cost_agg(&hashed_p, parse,
1017 AGG_HASHED, numAggs,
1018 numGroupCols, dNumGroups,
1019 cheapest_path->startup_cost,
1020 cheapest_path->total_cost,
1021 cheapest_path_rows);
1022 /* Result of hashed agg is always unsorted */
1024 cost_sort(&hashed_p, parse, sort_pathkeys,
1025 hashed_p.total_cost,
1027 cheapest_path_width);
1031 sorted_p.startup_cost = sorted_path->startup_cost;
1032 sorted_p.total_cost = sorted_path->total_cost;
1033 current_pathkeys = sorted_path->pathkeys;
1037 sorted_p.startup_cost = cheapest_path->startup_cost;
1038 sorted_p.total_cost = cheapest_path->total_cost;
1039 current_pathkeys = cheapest_path->pathkeys;
1041 if (!pathkeys_contained_in(group_pathkeys,
1044 cost_sort(&sorted_p, parse, group_pathkeys,
1045 sorted_p.total_cost,
1047 cheapest_path_width);
1048 current_pathkeys = group_pathkeys;
1051 cost_agg(&sorted_p, parse,
1052 AGG_SORTED, numAggs,
1053 numGroupCols, dNumGroups,
1054 sorted_p.startup_cost,
1055 sorted_p.total_cost,
1056 cheapest_path_rows);
1058 cost_group(&sorted_p, parse,
1059 numGroupCols, dNumGroups,
1060 sorted_p.startup_cost,
1061 sorted_p.total_cost,
1062 cheapest_path_rows);
1063 /* The Agg or Group node will preserve ordering */
1064 if (sort_pathkeys &&
1065 !pathkeys_contained_in(sort_pathkeys,
1068 cost_sort(&sorted_p, parse, sort_pathkeys,
1069 sorted_p.total_cost,
1071 cheapest_path_width);
1075 * Now make the decision using the top-level tuple
1076 * fraction. First we have to convert an absolute
1077 * count (LIMIT) into fractional form.
1079 if (tuple_fraction >= 1.0)
1080 tuple_fraction /= dNumGroups;
1082 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1083 tuple_fraction) < 0)
1085 /* Hashed is cheaper, so use it */
1086 use_hashed_grouping = true;
1093 * Select the best path and create a plan to execute it.
1095 * If we are doing hashed grouping, we will always read all the input
1096 * tuples, so use the cheapest-total path. Otherwise, trust
1097 * query_planner's decision about which to use.
1099 if (sorted_path && !use_hashed_grouping)
1101 result_plan = create_plan(parse, sorted_path);
1102 current_pathkeys = sorted_path->pathkeys;
1106 result_plan = create_plan(parse, cheapest_path);
1107 current_pathkeys = cheapest_path->pathkeys;
1111 * create_plan() returns a plan with just a "flat" tlist of
1112 * required Vars. Usually we need to insert the sub_tlist as the
1113 * tlist of the top plan node. However, we can skip that if we
1114 * determined that whatever query_planner chose to return will be
1117 if (need_tlist_eval)
1120 * If the top-level plan node is one that cannot do expression
1121 * evaluation, we must insert a Result node to project the
1124 if (!is_projection_capable_plan(result_plan))
1126 result_plan = (Plan *) make_result(sub_tlist, NULL,
1132 * Otherwise, just replace the subplan's flat tlist with
1133 * the desired tlist.
1135 result_plan->targetlist = sub_tlist;
1139 * Also, account for the cost of evaluation of the sub_tlist.
1141 * Up to now, we have only been dealing with "flat" tlists,
1142 * containing just Vars. So their evaluation cost is zero
1143 * according to the model used by cost_qual_eval() (or if you
1144 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1145 * can avoid accounting for tlist cost throughout
1146 * query_planner() and subroutines. But now we've inserted a
1147 * tlist that might contain actual operators, sub-selects, etc
1148 * --- so we'd better account for its cost.
1150 * Below this point, any tlist eval cost for added-on nodes
1151 * should be accounted for as we create those nodes.
1152 * Presently, of the node types we can add on, only Agg and
1153 * Group project new tlists (the rest just copy their input
1154 * tuples) --- so make_agg() and make_group() are responsible
1155 * for computing the added cost.
1157 cost_qual_eval(&tlist_cost, sub_tlist);
1158 result_plan->startup_cost += tlist_cost.startup;
1159 result_plan->total_cost += tlist_cost.startup +
1160 tlist_cost.per_tuple * result_plan->plan_rows;
1165 * Since we're using query_planner's tlist and not the one
1166 * make_subplanTargetList calculated, we have to refigure any
1167 * grouping-column indexes make_subplanTargetList computed.
1169 locate_grouping_columns(parse, tlist, result_plan->targetlist,
1174 * Insert AGG or GROUP node if needed, plus an explicit sort step
1177 * HAVING clause, if any, becomes qual of the Agg node
1179 if (use_hashed_grouping)
1181 /* Hashed aggregate plan --- no sort needed */
1182 result_plan = (Plan *) make_agg(parse,
1184 (List *) parse->havingQual,
1191 /* Hashed aggregation produces randomly-ordered results */
1192 current_pathkeys = NIL;
1194 else if (parse->hasAggs)
1196 /* Plain aggregate plan --- sort if needed */
1197 AggStrategy aggstrategy;
1199 if (parse->groupClause)
1201 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1203 result_plan = (Plan *)
1204 make_sort_from_groupcols(parse,
1208 current_pathkeys = group_pathkeys;
1210 aggstrategy = AGG_SORTED;
1213 * The AGG node will not change the sort ordering of its
1214 * groups, so current_pathkeys describes the result too.
1219 aggstrategy = AGG_PLAIN;
1220 /* Result will be only one row anyway; no sort order */
1221 current_pathkeys = NIL;
1224 result_plan = (Plan *) make_agg(parse,
1226 (List *) parse->havingQual,
1237 * If there are no Aggs, we shouldn't have any HAVING qual
1240 Assert(parse->havingQual == NULL);
1243 * If we have a GROUP BY clause, insert a group node (plus the
1244 * appropriate sort node, if necessary).
1246 if (parse->groupClause)
1249 * Add an explicit sort if we couldn't make the path come
1250 * out the way the GROUP node needs it.
1252 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1254 result_plan = (Plan *)
1255 make_sort_from_groupcols(parse,
1259 current_pathkeys = group_pathkeys;
1262 result_plan = (Plan *) make_group(parse,
1268 /* The Group node won't change sort ordering */
1271 } /* end of if (setOperations) */
1274 * If we were not able to make the plan come out in the right order,
1275 * add an explicit sort step.
1277 if (parse->sortClause)
1279 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1281 result_plan = (Plan *)
1282 make_sort_from_sortclauses(parse,
1285 current_pathkeys = sort_pathkeys;
1290 * If there is a DISTINCT clause, add the UNIQUE node.
1292 if (parse->distinctClause)
1294 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1297 * If there was grouping or aggregation, leave plan_rows as-is
1298 * (ie, assume the result was already mostly unique). If not,
1299 * it's reasonable to assume the UNIQUE filter has effects
1300 * comparable to GROUP BY.
1302 if (!parse->groupClause && !parse->hasAggs)
1304 List *distinctExprs;
1306 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1308 result_plan->plan_rows = estimate_num_groups(parse,
1310 result_plan->plan_rows);
1315 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1317 if (parse->limitOffset || parse->limitCount)
1319 result_plan = (Plan *) make_limit(result_plan,
1325 * Return the actual output ordering in query_pathkeys for possible
1326 * use by an outer query level.
1328 parse->query_pathkeys = current_pathkeys;
1334 * hash_safe_grouping - are grouping operators hashable?
1336 * We assume hashed aggregation will work if the datatype's equality operator
1337 * is marked hashjoinable.
1340 hash_safe_grouping(Query *parse)
1344 foreach(gl, parse->groupClause)
1346 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1347 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1351 optup = equality_oper(tle->resdom->restype, true);
1354 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1355 ReleaseSysCache(optup);
1363 * make_subplanTargetList
1364 * Generate appropriate target list when grouping is required.
1366 * When grouping_planner inserts Aggregate or Group plan nodes above
1367 * the result of query_planner, we typically want to pass a different
1368 * target list to query_planner than the outer plan nodes should have.
1369 * This routine generates the correct target list for the subplan.
1371 * The initial target list passed from the parser already contains entries
1372 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1373 * for variables used only in HAVING clauses; so we need to add those
1374 * variables to the subplan target list. Also, if we are doing either
1375 * grouping or aggregation, we flatten all expressions except GROUP BY items
1376 * into their component variables; the other expressions will be computed by
1377 * the inserted nodes rather than by the subplan. For example,
1378 * given a query like
1379 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1380 * we want to pass this targetlist to the subplan:
1382 * where the a+b target will be used by the Sort/Group steps, and the
1383 * other targets will be used for computing the final results. (In the
1384 * above example we could theoretically suppress the a and b targets and
1385 * pass down only c,d,a+b, but it's not really worth the trouble to
1386 * eliminate simple var references from the subplan. We will avoid doing
1387 * the extra computation to recompute a+b at the outer level; see
1388 * replace_vars_with_subplan_refs() in setrefs.c.)
1390 * If we are grouping or aggregating, *and* there are no non-Var grouping
1391 * expressions, then the returned tlist is effectively dummy; we do not
1392 * need to force it to be evaluated, because all the Vars it contains
1393 * should be present in the output of query_planner anyway.
1395 * 'parse' is the query being processed.
1396 * 'tlist' is the query's target list.
1397 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1398 * expressions (if there are any) in the subplan's target list.
1399 * 'need_tlist_eval' is set true if we really need to evaluate the
1402 * The result is the targetlist to be passed to the subplan.
1406 make_subplanTargetList(Query *parse,
1408 AttrNumber **groupColIdx,
1409 bool *need_tlist_eval)
1415 *groupColIdx = NULL;
1418 * If we're not grouping or aggregating, nothing to do here;
1419 * query_planner should receive the unmodified target list.
1421 if (!parse->hasAggs && !parse->groupClause)
1423 *need_tlist_eval = true;
1428 * Otherwise, start with a "flattened" tlist (having just the vars
1429 * mentioned in the targetlist and HAVING qual --- but not upper-
1430 * level Vars; they will be replaced by Params later on).
1432 sub_tlist = flatten_tlist(tlist);
1433 extravars = pull_var_clause(parse->havingQual, false);
1434 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1435 list_free(extravars);
1436 *need_tlist_eval = false; /* only eval if not flat tlist */
1439 * If grouping, create sub_tlist entries for all GROUP BY expressions
1440 * (GROUP BY items that are simple Vars should be in the list
1441 * already), and make an array showing where the group columns are in
1444 numCols = list_length(parse->groupClause);
1448 AttrNumber *grpColIdx;
1451 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1452 *groupColIdx = grpColIdx;
1454 foreach(gl, parse->groupClause)
1456 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1457 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1458 TargetEntry *te = NULL;
1461 /* Find or make a matching sub_tlist entry */
1462 foreach(sl, sub_tlist)
1464 te = (TargetEntry *) lfirst(sl);
1465 if (equal(groupexpr, te->expr))
1470 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1471 exprType(groupexpr),
1472 exprTypmod(groupexpr),
1475 (Expr *) groupexpr);
1476 sub_tlist = lappend(sub_tlist, te);
1477 *need_tlist_eval = true; /* it's not flat anymore */
1480 /* and save its resno */
1481 grpColIdx[keyno++] = te->resdom->resno;
1489 * locate_grouping_columns
1490 * Locate grouping columns in the tlist chosen by query_planner.
1492 * This is only needed if we don't use the sub_tlist chosen by
1493 * make_subplanTargetList. We have to forget the column indexes found
1494 * by that routine and re-locate the grouping vars in the real sub_tlist.
1497 locate_grouping_columns(Query *parse,
1500 AttrNumber *groupColIdx)
1506 * No work unless grouping.
1508 if (!parse->groupClause)
1510 Assert(groupColIdx == NULL);
1513 Assert(groupColIdx != NULL);
1515 foreach(gl, parse->groupClause)
1517 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1518 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1519 TargetEntry *te = NULL;
1522 foreach(sl, sub_tlist)
1524 te = (TargetEntry *) lfirst(sl);
1525 if (equal(groupexpr, te->expr))
1529 elog(ERROR, "failed to locate grouping columns");
1531 groupColIdx[keyno++] = te->resdom->resno;
1536 * postprocess_setop_tlist
1537 * Fix up targetlist returned by plan_set_operations().
1539 * We need to transpose sort key info from the orig_tlist into new_tlist.
1540 * NOTE: this would not be good enough if we supported resjunk sort keys
1541 * for results of set operations --- then, we'd need to project a whole
1542 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1543 * find any resjunk columns in orig_tlist.
1546 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1549 ListCell *orig_tlist_item = list_head(orig_tlist);
1551 foreach(l, new_tlist)
1553 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1554 TargetEntry *orig_tle;
1556 /* ignore resjunk columns in setop result */
1557 if (new_tle->resdom->resjunk)
1560 Assert(orig_tlist_item != NULL);
1561 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1562 orig_tlist_item = lnext(orig_tlist_item);
1563 if (orig_tle->resdom->resjunk) /* should not happen */
1564 elog(ERROR, "resjunk output columns are not implemented");
1565 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1566 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1567 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1569 if (orig_tlist_item != NULL)
1570 elog(ERROR, "resjunk output columns are not implemented");