1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2005, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.180 2005/03/17 23:44:26 neilc Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "catalog/pg_type.h"
22 #include "executor/executor.h"
23 #include "executor/nodeAgg.h"
24 #include "miscadmin.h"
25 #include "nodes/makefuncs.h"
26 #ifdef OPTIMIZER_DEBUG
27 #include "nodes/print.h"
29 #include "optimizer/clauses.h"
30 #include "optimizer/cost.h"
31 #include "optimizer/pathnode.h"
32 #include "optimizer/paths.h"
33 #include "optimizer/planmain.h"
34 #include "optimizer/planner.h"
35 #include "optimizer/prep.h"
36 #include "optimizer/subselect.h"
37 #include "optimizer/tlist.h"
38 #include "optimizer/var.h"
39 #include "parser/parsetree.h"
40 #include "parser/parse_expr.h"
41 #include "parser/parse_oper.h"
42 #include "utils/selfuncs.h"
43 #include "utils/syscache.h"
46 ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
49 /* Expression kind codes for preprocess_expression */
50 #define EXPRKIND_QUAL 0
51 #define EXPRKIND_TARGET 1
52 #define EXPRKIND_RTFUNC 2
53 #define EXPRKIND_LIMIT 3
54 #define EXPRKIND_ININFO 4
57 static Node *preprocess_expression(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 * Set hasHavingQual to remember if HAVING clause is present. Needed
239 * because preprocess_expression will reduce a constant-true condition
240 * to an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
242 parse->hasHavingQual = (parse->havingQual != NULL);
245 * Do expression preprocessing on targetlist and quals.
247 parse->targetList = (List *)
248 preprocess_expression(parse, (Node *) parse->targetList,
251 preprocess_qual_conditions(parse, (Node *) parse->jointree);
253 parse->havingQual = preprocess_expression(parse, parse->havingQual,
256 parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
258 parse->limitCount = preprocess_expression(parse, parse->limitCount,
261 parse->in_info_list = (List *)
262 preprocess_expression(parse, (Node *) parse->in_info_list,
265 /* Also need to preprocess expressions for function RTEs */
266 foreach(l, parse->rtable)
268 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
270 if (rte->rtekind == RTE_FUNCTION)
271 rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
276 * In some cases we may want to transfer a HAVING clause into WHERE.
277 * We cannot do so if the HAVING clause contains aggregates (obviously)
278 * or volatile functions (since a HAVING clause is supposed to be executed
279 * only once per group). Also, it may be that the clause is so expensive
280 * to execute that we're better off doing it only once per group, despite
281 * the loss of selectivity. This is hard to estimate short of doing the
282 * entire planning process twice, so we use a heuristic: clauses
283 * containing subplans are left in HAVING. Otherwise, we move or copy
284 * the HAVING clause into WHERE, in hopes of eliminating tuples before
285 * aggregation instead of after.
287 * If the query has explicit grouping then we can simply move such a
288 * clause into WHERE; any group that fails the clause will not be
289 * in the output because none of its tuples will reach the grouping
290 * or aggregation stage. Otherwise we must have a degenerate
291 * (variable-free) HAVING clause, which we put in WHERE so that
292 * query_planner() can use it in a gating Result node, but also keep
293 * in HAVING to ensure that we don't emit a bogus aggregated row.
294 * (This could be done better, but it seems not worth optimizing.)
296 * Note that both havingQual and parse->jointree->quals are in
297 * implicitly-ANDed-list form at this point, even though they are
298 * declared as Node *.
301 foreach(l, (List *) parse->havingQual)
303 Node *havingclause = (Node *) lfirst(l);
305 if (contain_agg_clause(havingclause) ||
306 contain_volatile_functions(havingclause) ||
307 contain_subplans(havingclause))
309 /* keep it in HAVING */
310 newHaving = lappend(newHaving, havingclause);
312 else if (parse->groupClause)
314 /* move it to WHERE */
315 parse->jointree->quals = (Node *)
316 lappend((List *) parse->jointree->quals, havingclause);
320 /* put a copy in WHERE, keep it in HAVING */
321 parse->jointree->quals = (Node *)
322 lappend((List *) parse->jointree->quals,
323 copyObject(havingclause));
324 newHaving = lappend(newHaving, havingclause);
327 parse->havingQual = (Node *) newHaving;
330 * If we have any outer joins, try to reduce them to plain inner
331 * joins. This step is most easily done after we've done expression
335 reduce_outer_joins(parse);
338 * See if we can simplify the jointree; opportunities for this may
339 * come from having pulled up subqueries, or from flattening explicit
340 * JOIN syntax. We must do this after flattening JOIN alias
341 * variables, since eliminating explicit JOIN nodes from the jointree
342 * will cause get_relids_for_join() to fail. But it should happen
343 * after reduce_outer_joins, anyway.
345 parse->jointree = (FromExpr *)
346 simplify_jointree(parse, (Node *) parse->jointree);
349 * Do the main planning. If we have an inherited target relation,
350 * that needs special processing, else go straight to
353 if (parse->resultRelation &&
354 (lst = expand_inherited_rtentry(parse, parse->resultRelation)) != NIL)
355 plan = inheritance_planner(parse, lst);
357 plan = grouping_planner(parse, tuple_fraction);
360 * If any subplans were generated, or if we're inside a subplan, build
361 * initPlan list and extParam/allParam sets for plan nodes.
363 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
365 Cost initplan_cost = 0;
367 /* Prepare extParam/allParam sets for all nodes in tree */
368 SS_finalize_plan(plan, parse->rtable);
371 * SS_finalize_plan doesn't handle initPlans, so we have to
372 * manually attach them to the topmost plan node, and add their
373 * extParams to the topmost node's, too.
375 * We also add the total_cost of each initPlan to the startup cost of
376 * the top node. This is a conservative overestimate, since in
377 * fact each initPlan might be executed later than plan startup,
378 * or even not at all.
380 plan->initPlan = PlannerInitPlan;
382 foreach(l, plan->initPlan)
384 SubPlan *initplan = (SubPlan *) lfirst(l);
386 plan->extParam = bms_add_members(plan->extParam,
387 initplan->plan->extParam);
388 /* allParam must include all members of extParam */
389 plan->allParam = bms_add_members(plan->allParam,
391 initplan_cost += initplan->plan->total_cost;
394 plan->startup_cost += initplan_cost;
395 plan->total_cost += initplan_cost;
398 /* Return to outer subquery context */
400 PlannerInitPlan = saved_initplan;
401 /* we do NOT restore PlannerPlanId; that's not an oversight! */
407 * preprocess_expression
408 * Do subquery_planner's preprocessing work for an expression,
409 * which can be a targetlist, a WHERE clause (including JOIN/ON
410 * conditions), or a HAVING clause.
413 preprocess_expression(Query *parse, Node *expr, int kind)
416 * If the query has any join RTEs, replace join alias variables with
417 * base-relation variables. We must do this before sublink processing,
418 * else sublinks expanded out from join aliases wouldn't get
421 if (parse->hasJoinRTEs)
422 expr = flatten_join_alias_vars(parse, expr);
425 * If it's a qual or havingQual, canonicalize it. It seems most
426 * useful to do this before applying eval_const_expressions, since the
427 * latter can optimize flattened AND/ORs better than unflattened ones.
429 * Note: all processing of a qual expression after this point must be
430 * careful to maintain AND/OR flatness --- that is, do not generate a
431 * tree with AND directly under AND, nor OR directly under OR.
433 if (kind == EXPRKIND_QUAL)
435 expr = (Node *) canonicalize_qual((Expr *) expr);
437 #ifdef OPTIMIZER_DEBUG
438 printf("After canonicalize_qual()\n");
444 * Simplify constant expressions.
446 expr = eval_const_expressions(expr);
448 /* Expand SubLinks to SubPlans */
449 if (parse->hasSubLinks)
450 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
453 * XXX do not insert anything here unless you have grokked the
454 * comments in SS_replace_correlation_vars ...
457 /* Replace uplevel vars with Param nodes */
458 if (PlannerQueryLevel > 1)
459 expr = SS_replace_correlation_vars(expr);
462 * If it's a qual or havingQual, convert it to implicit-AND format.
463 * (We don't want to do this before eval_const_expressions, since the
464 * latter would be unable to simplify a top-level AND correctly. Also,
465 * SS_process_sublinks expects explicit-AND format.)
467 if (kind == EXPRKIND_QUAL)
468 expr = (Node *) make_ands_implicit((Expr *) expr);
474 * preprocess_qual_conditions
475 * Recursively scan the query's jointree and do subquery_planner's
476 * preprocessing work on each qual condition found therein.
479 preprocess_qual_conditions(Query *parse, Node *jtnode)
483 if (IsA(jtnode, RangeTblRef))
485 /* nothing to do here */
487 else if (IsA(jtnode, FromExpr))
489 FromExpr *f = (FromExpr *) jtnode;
492 foreach(l, f->fromlist)
493 preprocess_qual_conditions(parse, lfirst(l));
495 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
497 else if (IsA(jtnode, JoinExpr))
499 JoinExpr *j = (JoinExpr *) jtnode;
501 preprocess_qual_conditions(parse, j->larg);
502 preprocess_qual_conditions(parse, j->rarg);
504 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
507 elog(ERROR, "unrecognized node type: %d",
508 (int) nodeTag(jtnode));
511 /*--------------------
512 * inheritance_planner
513 * Generate a plan in the case where the result relation is an
516 * We have to handle this case differently from cases where a source
517 * relation is an inheritance set. Source inheritance is expanded at
518 * the bottom of the plan tree (see allpaths.c), but target inheritance
519 * has to be expanded at the top. The reason is that for UPDATE, each
520 * target relation needs a different targetlist matching its own column
521 * set. (This is not so critical for DELETE, but for simplicity we treat
522 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
523 * can never be the nullable side of an outer join, so it's OK to generate
526 * parse is the querytree produced by the parser & rewriter.
527 * inheritlist is an integer list of RT indexes for the result relation set.
529 * Returns a query plan.
530 *--------------------
533 inheritance_planner(Query *parse, List *inheritlist)
535 int parentRTindex = parse->resultRelation;
536 Oid parentOID = getrelid(parentRTindex, parse->rtable);
537 int mainrtlength = list_length(parse->rtable);
538 List *subplans = NIL;
542 foreach(l, inheritlist)
544 int childRTindex = lfirst_int(l);
545 Oid childOID = getrelid(childRTindex, parse->rtable);
549 /* Generate modified query with this rel as target */
550 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
551 parentRTindex, parentOID,
552 childRTindex, childOID);
554 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
555 subplans = lappend(subplans, subplan);
558 * XXX my goodness this next bit is ugly. Really need to think about
559 * ways to rein in planner's habit of scribbling on its input.
561 * Planning of the subquery might have modified the rangetable,
562 * either by addition of RTEs due to expansion of inherited source
563 * tables, or by changes of the Query structures inside subquery
564 * RTEs. We have to ensure that this gets propagated back to the
565 * master copy. However, if we aren't done planning yet, we also
566 * need to ensure that subsequent calls to grouping_planner have
567 * virgin sub-Queries to work from. So, if we are at the last
568 * list entry, just copy the subquery rangetable back to the master
569 * copy; if we are not, then extend the master copy by adding
570 * whatever the subquery added. (We assume these added entries
571 * will go untouched by the future grouping_planner calls. We are
572 * also effectively assuming that sub-Queries will get planned
573 * identically each time, or at least that the impacts on their
574 * rangetables will be the same each time. Did I say this is ugly?)
576 if (lnext(l) == NULL)
577 parse->rtable = subquery->rtable;
580 int subrtlength = list_length(subquery->rtable);
582 if (subrtlength > mainrtlength)
586 subrt = list_copy_tail(subquery->rtable, mainrtlength);
587 parse->rtable = list_concat(parse->rtable, subrt);
588 mainrtlength = subrtlength;
592 /* Save preprocessed tlist from first rel for use in Append */
594 tlist = subplan->targetlist;
597 /* Save the target-relations list for the executor, too */
598 parse->resultRelations = inheritlist;
600 /* Mark result as unordered (probably unnecessary) */
601 parse->query_pathkeys = NIL;
603 return (Plan *) make_append(subplans, true, tlist);
606 /*--------------------
608 * Perform planning steps related to grouping, aggregation, etc.
609 * This primarily means adding top-level processing to the basic
610 * query plan produced by query_planner.
612 * parse is the querytree produced by the parser & rewriter.
613 * tuple_fraction is the fraction of tuples we expect will be retrieved
615 * tuple_fraction is interpreted as follows:
616 * 0: expect all tuples to be retrieved (normal case)
617 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
618 * from the plan to be retrieved
619 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
620 * expected to be retrieved (ie, a LIMIT specification)
622 * Returns a query plan. Also, parse->query_pathkeys is returned as the
623 * actual output ordering of the plan (in pathkey format).
624 *--------------------
627 grouping_planner(Query *parse, double tuple_fraction)
629 List *tlist = parse->targetList;
631 List *current_pathkeys;
634 if (parse->setOperations)
636 List *set_sortclauses;
639 * Construct the plan for set operations. The result will not
640 * need any work except perhaps a top-level sort and/or LIMIT.
642 result_plan = plan_set_operations(parse,
646 * Calculate pathkeys representing the sort order (if any) of the
647 * set operation's result. We have to do this before overwriting
648 * the sort key information...
650 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
651 result_plan->targetlist);
652 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
655 * We should not need to call preprocess_targetlist, since we must
656 * be in a SELECT query node. Instead, use the targetlist
657 * returned by plan_set_operations (since this tells whether it
658 * returned any resjunk columns!), and transfer any sort key
659 * information from the original tlist.
661 Assert(parse->commandType == CMD_SELECT);
663 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
666 * Can't handle FOR UPDATE here (parser should have checked
667 * already, but let's make sure).
671 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
672 errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
675 * Calculate pathkeys that represent result ordering requirements
677 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
679 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
683 /* No set operations, do regular planning */
685 List *group_pathkeys;
686 AttrNumber *groupColIdx = NULL;
687 bool need_tlist_eval = true;
689 double sub_tuple_fraction;
692 double dNumGroups = 0;
694 AggClauseCounts agg_counts;
695 int numGroupCols = list_length(parse->groupClause);
696 bool use_hashed_grouping = false;
698 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
700 /* Preprocess targetlist */
701 tlist = preprocess_targetlist(parse, tlist);
704 * Generate appropriate target list for subplan; may be different
705 * from tlist if grouping or aggregation is needed.
707 sub_tlist = make_subplanTargetList(parse, tlist,
708 &groupColIdx, &need_tlist_eval);
711 * Calculate pathkeys that represent grouping/ordering
714 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
716 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
720 * Will need actual number of aggregates for estimating costs.
722 * Note: we do not attempt to detect duplicate aggregates here; a
723 * somewhat-overestimated count is okay for our present purposes.
725 * Note: think not that we can turn off hasAggs if we find no aggs.
726 * It is possible for constant-expression simplification to remove
727 * all explicit references to aggs, but we still have to follow
728 * the aggregate semantics (eg, producing only one output row).
732 count_agg_clauses((Node *) tlist, &agg_counts);
733 count_agg_clauses(parse->havingQual, &agg_counts);
737 * Figure out whether we need a sorted result from query_planner.
739 * If we have a GROUP BY clause, then we want a result sorted
740 * properly for grouping. Otherwise, if there is an ORDER BY
741 * clause, we want to sort by the ORDER BY clause. (Note: if we
742 * have both, and ORDER BY is a superset of GROUP BY, it would be
743 * tempting to request sort by ORDER BY --- but that might just
744 * leave us failing to exploit an available sort order at all.
745 * Needs more thought...)
747 if (parse->groupClause)
748 parse->query_pathkeys = group_pathkeys;
749 else if (parse->sortClause)
750 parse->query_pathkeys = sort_pathkeys;
752 parse->query_pathkeys = NIL;
755 * Adjust tuple_fraction if we see that we are going to apply
756 * limiting/grouping/aggregation/etc. This is not overridable by
757 * the caller, since it reflects plan actions that this routine
758 * will certainly take, not assumptions about context.
760 if (parse->limitCount != NULL)
763 * A LIMIT clause limits the absolute number of tuples
764 * returned. However, if it's not a constant LIMIT then we
765 * have to punt; for lack of a better idea, assume 10% of the
766 * plan's result is wanted.
768 double limit_fraction = 0.0;
770 if (IsA(parse->limitCount, Const))
772 Const *limitc = (Const *) parse->limitCount;
773 int32 count = DatumGetInt32(limitc->constvalue);
776 * A NULL-constant LIMIT represents "LIMIT ALL", which we
777 * treat the same as no limit (ie, expect to retrieve all
780 if (!limitc->constisnull && count > 0)
782 limit_fraction = (double) count;
783 /* We must also consider the OFFSET, if present */
784 if (parse->limitOffset != NULL)
786 if (IsA(parse->limitOffset, Const))
790 limitc = (Const *) parse->limitOffset;
791 offset = DatumGetInt32(limitc->constvalue);
792 if (!limitc->constisnull && offset > 0)
793 limit_fraction += (double) offset;
797 /* OFFSET is an expression ... punt ... */
798 limit_fraction = 0.10;
805 /* LIMIT is an expression ... punt ... */
806 limit_fraction = 0.10;
809 if (limit_fraction > 0.0)
812 * If we have absolute limits from both caller and LIMIT,
813 * use the smaller value; if one is fractional and the
814 * other absolute, treat the fraction as a fraction of the
815 * absolute value; else we can multiply the two fractions
818 if (tuple_fraction >= 1.0)
820 if (limit_fraction >= 1.0)
823 tuple_fraction = Min(tuple_fraction, limit_fraction);
827 /* caller absolute, limit fractional */
828 tuple_fraction *= limit_fraction;
829 if (tuple_fraction < 1.0)
830 tuple_fraction = 1.0;
833 else if (tuple_fraction > 0.0)
835 if (limit_fraction >= 1.0)
837 /* caller fractional, limit absolute */
838 tuple_fraction *= limit_fraction;
839 if (tuple_fraction < 1.0)
840 tuple_fraction = 1.0;
844 /* both fractional */
845 tuple_fraction *= limit_fraction;
850 /* no info from caller, just use limit */
851 tuple_fraction = limit_fraction;
857 * With grouping or aggregation, the tuple fraction to pass to
858 * query_planner() may be different from what it is at top level.
860 sub_tuple_fraction = tuple_fraction;
862 if (parse->groupClause)
865 * In GROUP BY mode, we have the little problem that we don't
866 * really know how many input tuples will be needed to make a
867 * group, so we can't translate an output LIMIT count into an
868 * input count. For lack of a better idea, assume 25% of the
869 * input data will be processed if there is any output limit.
870 * However, if the caller gave us a fraction rather than an
871 * absolute count, we can keep using that fraction (which
872 * amounts to assuming that all the groups are about the same
875 if (sub_tuple_fraction >= 1.0)
876 sub_tuple_fraction = 0.25;
879 * If both GROUP BY and ORDER BY are specified, we will need
880 * two levels of sort --- and, therefore, certainly need to
881 * read all the input tuples --- unless ORDER BY is a subset
882 * of GROUP BY. (We have not yet canonicalized the pathkeys,
883 * so must use the slower noncanonical comparison method.)
885 if (parse->groupClause && parse->sortClause &&
886 !noncanonical_pathkeys_contained_in(sort_pathkeys,
888 sub_tuple_fraction = 0.0;
890 else if (parse->hasAggs)
893 * Ungrouped aggregate will certainly want all the input
896 sub_tuple_fraction = 0.0;
898 else if (parse->distinctClause)
901 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
902 * number of input tuples per output tuple. Handle the same
905 if (sub_tuple_fraction >= 1.0)
906 sub_tuple_fraction = 0.25;
910 * Generate the best unsorted and presorted paths for this Query
911 * (but note there may not be any presorted path).
913 query_planner(parse, sub_tlist, sub_tuple_fraction,
914 &cheapest_path, &sorted_path);
917 * We couldn't canonicalize group_pathkeys and sort_pathkeys
918 * before running query_planner(), so do it now.
920 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
921 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
924 * Consider whether we might want to use hashed grouping.
926 if (parse->groupClause)
929 double cheapest_path_rows;
930 int cheapest_path_width;
933 * Beware in this section of the possibility that
934 * cheapest_path->parent is NULL. This could happen if user
935 * does something silly like SELECT 'foo' GROUP BY 1;
937 if (cheapest_path->parent)
939 cheapest_path_rows = cheapest_path->parent->rows;
940 cheapest_path_width = cheapest_path->parent->width;
944 cheapest_path_rows = 1; /* assume non-set result */
945 cheapest_path_width = 100; /* arbitrary */
949 * Always estimate the number of groups. We can't do this
950 * until after running query_planner(), either.
952 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
954 dNumGroups = estimate_num_groups(parse,
957 /* Also want it as a long int --- but 'ware overflow! */
958 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
961 * Check can't-do-it conditions, including whether the
962 * grouping operators are hashjoinable.
964 * Executor doesn't support hashed aggregation with DISTINCT
965 * aggregates. (Doing so would imply storing *all* the input
966 * values in the hash table, which seems like a certain
969 if (!enable_hashagg || !hash_safe_grouping(parse))
970 use_hashed_grouping = false;
971 else if (agg_counts.numDistinctAggs != 0)
972 use_hashed_grouping = false;
976 * Use hashed grouping if (a) we think we can fit the
977 * hashtable into work_mem, *and* (b) the estimated cost
978 * is no more than doing it the other way. While avoiding
979 * the need for sorted input is usually a win, the fact
980 * that the output won't be sorted may be a loss; so we
981 * need to do an actual cost comparison.
985 /* Estimate per-hash-entry space at tuple width... */
986 hashentrysize = cheapest_path_width;
987 /* plus space for pass-by-ref transition values... */
988 hashentrysize += agg_counts.transitionSpace;
989 /* plus the per-hash-entry overhead */
990 hashentrysize += hash_agg_entry_size(agg_counts.numAggs);
992 if (hashentrysize * dNumGroups <= work_mem * 1024L)
995 * Okay, do the cost comparison. We need to consider
996 * cheapest_path + hashagg [+ final sort] versus
997 * either cheapest_path [+ sort] + group or agg [+
998 * final sort] or presorted_path + group or agg [+
999 * final sort] where brackets indicate a step that may
1000 * not be needed. We assume query_planner() will have
1001 * returned a presorted path only if it's a winner
1002 * compared to cheapest_path for this purpose.
1004 * These path variables are dummies that just hold cost
1005 * fields; we don't make actual Paths for these steps.
1010 cost_agg(&hashed_p, parse,
1011 AGG_HASHED, agg_counts.numAggs,
1012 numGroupCols, dNumGroups,
1013 cheapest_path->startup_cost,
1014 cheapest_path->total_cost,
1015 cheapest_path_rows);
1016 /* Result of hashed agg is always unsorted */
1018 cost_sort(&hashed_p, parse, sort_pathkeys,
1019 hashed_p.total_cost,
1021 cheapest_path_width);
1025 sorted_p.startup_cost = sorted_path->startup_cost;
1026 sorted_p.total_cost = sorted_path->total_cost;
1027 current_pathkeys = sorted_path->pathkeys;
1031 sorted_p.startup_cost = cheapest_path->startup_cost;
1032 sorted_p.total_cost = cheapest_path->total_cost;
1033 current_pathkeys = cheapest_path->pathkeys;
1035 if (!pathkeys_contained_in(group_pathkeys,
1038 cost_sort(&sorted_p, parse, group_pathkeys,
1039 sorted_p.total_cost,
1041 cheapest_path_width);
1042 current_pathkeys = group_pathkeys;
1045 cost_agg(&sorted_p, parse,
1046 AGG_SORTED, agg_counts.numAggs,
1047 numGroupCols, dNumGroups,
1048 sorted_p.startup_cost,
1049 sorted_p.total_cost,
1050 cheapest_path_rows);
1052 cost_group(&sorted_p, parse,
1053 numGroupCols, dNumGroups,
1054 sorted_p.startup_cost,
1055 sorted_p.total_cost,
1056 cheapest_path_rows);
1057 /* The Agg or Group node will preserve ordering */
1058 if (sort_pathkeys &&
1059 !pathkeys_contained_in(sort_pathkeys,
1062 cost_sort(&sorted_p, parse, sort_pathkeys,
1063 sorted_p.total_cost,
1065 cheapest_path_width);
1069 * Now make the decision using the top-level tuple
1070 * fraction. First we have to convert an absolute
1071 * count (LIMIT) into fractional form.
1073 if (tuple_fraction >= 1.0)
1074 tuple_fraction /= dNumGroups;
1076 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1077 tuple_fraction) < 0)
1079 /* Hashed is cheaper, so use it */
1080 use_hashed_grouping = true;
1087 * Select the best path and create a plan to execute it.
1089 * If we are doing hashed grouping, we will always read all the input
1090 * tuples, so use the cheapest-total path. Otherwise, trust
1091 * query_planner's decision about which to use.
1093 if (sorted_path && !use_hashed_grouping)
1095 result_plan = create_plan(parse, sorted_path);
1096 current_pathkeys = sorted_path->pathkeys;
1100 result_plan = create_plan(parse, cheapest_path);
1101 current_pathkeys = cheapest_path->pathkeys;
1105 * create_plan() returns a plan with just a "flat" tlist of
1106 * required Vars. Usually we need to insert the sub_tlist as the
1107 * tlist of the top plan node. However, we can skip that if we
1108 * determined that whatever query_planner chose to return will be
1111 if (need_tlist_eval)
1114 * If the top-level plan node is one that cannot do expression
1115 * evaluation, we must insert a Result node to project the
1118 if (!is_projection_capable_plan(result_plan))
1120 result_plan = (Plan *) make_result(sub_tlist, NULL,
1126 * Otherwise, just replace the subplan's flat tlist with
1127 * the desired tlist.
1129 result_plan->targetlist = sub_tlist;
1133 * Also, account for the cost of evaluation of the sub_tlist.
1135 * Up to now, we have only been dealing with "flat" tlists,
1136 * containing just Vars. So their evaluation cost is zero
1137 * according to the model used by cost_qual_eval() (or if you
1138 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1139 * can avoid accounting for tlist cost throughout
1140 * query_planner() and subroutines. But now we've inserted a
1141 * tlist that might contain actual operators, sub-selects, etc
1142 * --- so we'd better account for its cost.
1144 * Below this point, any tlist eval cost for added-on nodes
1145 * should be accounted for as we create those nodes.
1146 * Presently, of the node types we can add on, only Agg and
1147 * Group project new tlists (the rest just copy their input
1148 * tuples) --- so make_agg() and make_group() are responsible
1149 * for computing the added cost.
1151 cost_qual_eval(&tlist_cost, sub_tlist);
1152 result_plan->startup_cost += tlist_cost.startup;
1153 result_plan->total_cost += tlist_cost.startup +
1154 tlist_cost.per_tuple * result_plan->plan_rows;
1159 * Since we're using query_planner's tlist and not the one
1160 * make_subplanTargetList calculated, we have to refigure any
1161 * grouping-column indexes make_subplanTargetList computed.
1163 locate_grouping_columns(parse, tlist, result_plan->targetlist,
1168 * Insert AGG or GROUP node if needed, plus an explicit sort step
1171 * HAVING clause, if any, becomes qual of the Agg or Group node.
1173 if (use_hashed_grouping)
1175 /* Hashed aggregate plan --- no sort needed */
1176 result_plan = (Plan *) make_agg(parse,
1178 (List *) parse->havingQual,
1185 /* Hashed aggregation produces randomly-ordered results */
1186 current_pathkeys = NIL;
1188 else if (parse->hasAggs)
1190 /* Plain aggregate plan --- sort if needed */
1191 AggStrategy aggstrategy;
1193 if (parse->groupClause)
1195 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1197 result_plan = (Plan *)
1198 make_sort_from_groupcols(parse,
1202 current_pathkeys = group_pathkeys;
1204 aggstrategy = AGG_SORTED;
1207 * The AGG node will not change the sort ordering of its
1208 * groups, so current_pathkeys describes the result too.
1213 aggstrategy = AGG_PLAIN;
1214 /* Result will be only one row anyway; no sort order */
1215 current_pathkeys = NIL;
1218 result_plan = (Plan *) make_agg(parse,
1220 (List *) parse->havingQual,
1228 else if (parse->groupClause)
1231 * GROUP BY without aggregation, so insert a group node (plus the
1232 * appropriate sort node, if necessary).
1234 * Add an explicit sort if we couldn't make the path come
1235 * out the way the GROUP node needs it.
1237 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1239 result_plan = (Plan *)
1240 make_sort_from_groupcols(parse,
1244 current_pathkeys = group_pathkeys;
1247 result_plan = (Plan *) make_group(parse,
1249 (List *) parse->havingQual,
1254 /* The Group node won't change sort ordering */
1256 else if (parse->hasHavingQual)
1259 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1260 * This is a degenerate case in which we are supposed to emit
1261 * either 0 or 1 row depending on whether HAVING succeeds.
1262 * Furthermore, there cannot be any variables in either HAVING
1263 * or the targetlist, so we actually do not need the FROM table
1264 * at all! We can just throw away the plan-so-far and generate
1265 * a Result node. This is a sufficiently unusual corner case
1266 * that it's not worth contorting the structure of this routine
1267 * to avoid having to generate the plan in the first place.
1269 result_plan = (Plan *) make_result(tlist,
1273 } /* end of if (setOperations) */
1276 * If we were not able to make the plan come out in the right order,
1277 * add an explicit sort step.
1279 if (parse->sortClause)
1281 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1283 result_plan = (Plan *)
1284 make_sort_from_sortclauses(parse,
1287 current_pathkeys = sort_pathkeys;
1292 * If there is a DISTINCT clause, add the UNIQUE node.
1294 if (parse->distinctClause)
1296 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1299 * If there was grouping or aggregation, leave plan_rows as-is
1300 * (ie, assume the result was already mostly unique). If not,
1301 * it's reasonable to assume the UNIQUE filter has effects
1302 * comparable to GROUP BY.
1304 if (!parse->groupClause && !parse->hasHavingQual && !parse->hasAggs)
1306 List *distinctExprs;
1308 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1310 result_plan->plan_rows = estimate_num_groups(parse,
1312 result_plan->plan_rows);
1317 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1319 if (parse->limitOffset || parse->limitCount)
1321 result_plan = (Plan *) make_limit(result_plan,
1327 * Return the actual output ordering in query_pathkeys for possible
1328 * use by an outer query level.
1330 parse->query_pathkeys = current_pathkeys;
1336 * hash_safe_grouping - are grouping operators hashable?
1338 * We assume hashed aggregation will work if the datatype's equality operator
1339 * is marked hashjoinable.
1342 hash_safe_grouping(Query *parse)
1346 foreach(gl, parse->groupClause)
1348 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1349 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1353 optup = equality_oper(tle->resdom->restype, true);
1356 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1357 ReleaseSysCache(optup);
1365 * make_subplanTargetList
1366 * Generate appropriate target list when grouping is required.
1368 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1369 * above the result of query_planner, we typically want to pass a different
1370 * target list to query_planner than the outer plan nodes should have.
1371 * This routine generates the correct target list for the subplan.
1373 * The initial target list passed from the parser already contains entries
1374 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1375 * for variables used only in HAVING clauses; so we need to add those
1376 * variables to the subplan target list. Also, we flatten all expressions
1377 * except GROUP BY items into their component variables; the other expressions
1378 * will be computed by the inserted nodes rather than by the subplan.
1379 * For example, given a query like
1380 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1381 * we want to pass this targetlist to the subplan:
1383 * where the a+b target will be used by the Sort/Group steps, and the
1384 * other targets will be used for computing the final results. (In the
1385 * above example we could theoretically suppress the a and b targets and
1386 * pass down only c,d,a+b, but it's not really worth the trouble to
1387 * eliminate simple var references from the subplan. We will avoid doing
1388 * the extra computation to recompute a+b at the outer level; see
1389 * replace_vars_with_subplan_refs() in setrefs.c.)
1391 * If we are grouping or aggregating, *and* there are no non-Var grouping
1392 * expressions, then the returned tlist is effectively dummy; we do not
1393 * need to force it to be evaluated, because all the Vars it contains
1394 * should be present in the output of query_planner anyway.
1396 * 'parse' is the query being processed.
1397 * 'tlist' is the query's target list.
1398 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1399 * expressions (if there are any) in the subplan's target list.
1400 * 'need_tlist_eval' is set true if we really need to evaluate the
1403 * The result is the targetlist to be passed to the subplan.
1407 make_subplanTargetList(Query *parse,
1409 AttrNumber **groupColIdx,
1410 bool *need_tlist_eval)
1416 *groupColIdx = NULL;
1419 * If we're not grouping or aggregating, there's nothing to do here;
1420 * query_planner should receive the unmodified target list.
1422 if (!parse->hasAggs && !parse->groupClause && !parse->hasHavingQual)
1424 *need_tlist_eval = true;
1429 * Otherwise, start with a "flattened" tlist (having just the vars
1430 * mentioned in the targetlist and HAVING qual --- but not upper-
1431 * level Vars; they will be replaced by Params later on).
1433 sub_tlist = flatten_tlist(tlist);
1434 extravars = pull_var_clause(parse->havingQual, false);
1435 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1436 list_free(extravars);
1437 *need_tlist_eval = false; /* only eval if not flat tlist */
1440 * If grouping, create sub_tlist entries for all GROUP BY expressions
1441 * (GROUP BY items that are simple Vars should be in the list
1442 * already), and make an array showing where the group columns are in
1445 numCols = list_length(parse->groupClause);
1449 AttrNumber *grpColIdx;
1452 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1453 *groupColIdx = grpColIdx;
1455 foreach(gl, parse->groupClause)
1457 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1458 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1459 TargetEntry *te = NULL;
1462 /* Find or make a matching sub_tlist entry */
1463 foreach(sl, sub_tlist)
1465 te = (TargetEntry *) lfirst(sl);
1466 if (equal(groupexpr, te->expr))
1471 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1472 exprType(groupexpr),
1473 exprTypmod(groupexpr),
1476 (Expr *) groupexpr);
1477 sub_tlist = lappend(sub_tlist, te);
1478 *need_tlist_eval = true; /* it's not flat anymore */
1481 /* and save its resno */
1482 grpColIdx[keyno++] = te->resdom->resno;
1490 * locate_grouping_columns
1491 * Locate grouping columns in the tlist chosen by query_planner.
1493 * This is only needed if we don't use the sub_tlist chosen by
1494 * make_subplanTargetList. We have to forget the column indexes found
1495 * by that routine and re-locate the grouping vars in the real sub_tlist.
1498 locate_grouping_columns(Query *parse,
1501 AttrNumber *groupColIdx)
1507 * No work unless grouping.
1509 if (!parse->groupClause)
1511 Assert(groupColIdx == NULL);
1514 Assert(groupColIdx != NULL);
1516 foreach(gl, parse->groupClause)
1518 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1519 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1520 TargetEntry *te = NULL;
1523 foreach(sl, sub_tlist)
1525 te = (TargetEntry *) lfirst(sl);
1526 if (equal(groupexpr, te->expr))
1530 elog(ERROR, "failed to locate grouping columns");
1532 groupColIdx[keyno++] = te->resdom->resno;
1537 * postprocess_setop_tlist
1538 * Fix up targetlist returned by plan_set_operations().
1540 * We need to transpose sort key info from the orig_tlist into new_tlist.
1541 * NOTE: this would not be good enough if we supported resjunk sort keys
1542 * for results of set operations --- then, we'd need to project a whole
1543 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1544 * find any resjunk columns in orig_tlist.
1547 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1550 ListCell *orig_tlist_item = list_head(orig_tlist);
1552 foreach(l, new_tlist)
1554 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1555 TargetEntry *orig_tle;
1557 /* ignore resjunk columns in setop result */
1558 if (new_tle->resdom->resjunk)
1561 Assert(orig_tlist_item != NULL);
1562 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1563 orig_tlist_item = lnext(orig_tlist_item);
1564 if (orig_tle->resdom->resjunk) /* should not happen */
1565 elog(ERROR, "resjunk output columns are not implemented");
1566 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1567 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1568 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1570 if (orig_tlist_item != NULL)
1571 elog(ERROR, "resjunk output columns are not implemented");