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 1.182 2005/04/06 16:34:05 tgl 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 * Simplify constant expressions.
427 * Note: this also flattens nested AND and OR expressions into N-argument
428 * form. All processing of a qual expression after this point must be
429 * careful to maintain AND/OR flatness --- that is, do not generate a tree
430 * with AND directly under AND, nor OR directly under OR.
432 expr = eval_const_expressions(expr);
435 * If it's a qual or havingQual, canonicalize it.
437 if (kind == EXPRKIND_QUAL)
439 expr = (Node *) canonicalize_qual((Expr *) expr);
441 #ifdef OPTIMIZER_DEBUG
442 printf("After canonicalize_qual()\n");
447 /* Expand SubLinks to SubPlans */
448 if (parse->hasSubLinks)
449 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
452 * XXX do not insert anything here unless you have grokked the
453 * comments in SS_replace_correlation_vars ...
456 /* Replace uplevel vars with Param nodes */
457 if (PlannerQueryLevel > 1)
458 expr = SS_replace_correlation_vars(expr);
461 * If it's a qual or havingQual, convert it to implicit-AND format.
462 * (We don't want to do this before eval_const_expressions, since the
463 * latter would be unable to simplify a top-level AND correctly. Also,
464 * SS_process_sublinks expects explicit-AND format.)
466 if (kind == EXPRKIND_QUAL)
467 expr = (Node *) make_ands_implicit((Expr *) expr);
473 * preprocess_qual_conditions
474 * Recursively scan the query's jointree and do subquery_planner's
475 * preprocessing work on each qual condition found therein.
478 preprocess_qual_conditions(Query *parse, Node *jtnode)
482 if (IsA(jtnode, RangeTblRef))
484 /* nothing to do here */
486 else if (IsA(jtnode, FromExpr))
488 FromExpr *f = (FromExpr *) jtnode;
491 foreach(l, f->fromlist)
492 preprocess_qual_conditions(parse, lfirst(l));
494 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
496 else if (IsA(jtnode, JoinExpr))
498 JoinExpr *j = (JoinExpr *) jtnode;
500 preprocess_qual_conditions(parse, j->larg);
501 preprocess_qual_conditions(parse, j->rarg);
503 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
506 elog(ERROR, "unrecognized node type: %d",
507 (int) nodeTag(jtnode));
510 /*--------------------
511 * inheritance_planner
512 * Generate a plan in the case where the result relation is an
515 * We have to handle this case differently from cases where a source
516 * relation is an inheritance set. Source inheritance is expanded at
517 * the bottom of the plan tree (see allpaths.c), but target inheritance
518 * has to be expanded at the top. The reason is that for UPDATE, each
519 * target relation needs a different targetlist matching its own column
520 * set. (This is not so critical for DELETE, but for simplicity we treat
521 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
522 * can never be the nullable side of an outer join, so it's OK to generate
525 * parse is the querytree produced by the parser & rewriter.
526 * inheritlist is an integer list of RT indexes for the result relation set.
528 * Returns a query plan.
529 *--------------------
532 inheritance_planner(Query *parse, List *inheritlist)
534 int parentRTindex = parse->resultRelation;
535 Oid parentOID = getrelid(parentRTindex, parse->rtable);
536 int mainrtlength = list_length(parse->rtable);
537 List *subplans = NIL;
541 foreach(l, inheritlist)
543 int childRTindex = lfirst_int(l);
544 Oid childOID = getrelid(childRTindex, parse->rtable);
548 /* Generate modified query with this rel as target */
549 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
550 parentRTindex, parentOID,
551 childRTindex, childOID);
553 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
554 subplans = lappend(subplans, subplan);
557 * XXX my goodness this next bit is ugly. Really need to think about
558 * ways to rein in planner's habit of scribbling on its input.
560 * Planning of the subquery might have modified the rangetable,
561 * either by addition of RTEs due to expansion of inherited source
562 * tables, or by changes of the Query structures inside subquery
563 * RTEs. We have to ensure that this gets propagated back to the
564 * master copy. However, if we aren't done planning yet, we also
565 * need to ensure that subsequent calls to grouping_planner have
566 * virgin sub-Queries to work from. So, if we are at the last
567 * list entry, just copy the subquery rangetable back to the master
568 * copy; if we are not, then extend the master copy by adding
569 * whatever the subquery added. (We assume these added entries
570 * will go untouched by the future grouping_planner calls. We are
571 * also effectively assuming that sub-Queries will get planned
572 * identically each time, or at least that the impacts on their
573 * rangetables will be the same each time. Did I say this is ugly?)
575 if (lnext(l) == NULL)
576 parse->rtable = subquery->rtable;
579 int subrtlength = list_length(subquery->rtable);
581 if (subrtlength > mainrtlength)
585 subrt = list_copy_tail(subquery->rtable, mainrtlength);
586 parse->rtable = list_concat(parse->rtable, subrt);
587 mainrtlength = subrtlength;
591 /* Save preprocessed tlist from first rel for use in Append */
593 tlist = subplan->targetlist;
596 /* Save the target-relations list for the executor, too */
597 parse->resultRelations = inheritlist;
599 /* Mark result as unordered (probably unnecessary) */
600 parse->query_pathkeys = NIL;
602 return (Plan *) make_append(subplans, true, tlist);
605 /*--------------------
607 * Perform planning steps related to grouping, aggregation, etc.
608 * This primarily means adding top-level processing to the basic
609 * query plan produced by query_planner.
611 * parse is the querytree produced by the parser & rewriter.
612 * tuple_fraction is the fraction of tuples we expect will be retrieved
614 * tuple_fraction is interpreted as follows:
615 * 0: expect all tuples to be retrieved (normal case)
616 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
617 * from the plan to be retrieved
618 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
619 * expected to be retrieved (ie, a LIMIT specification)
621 * Returns a query plan. Also, parse->query_pathkeys is returned as the
622 * actual output ordering of the plan (in pathkey format).
623 *--------------------
626 grouping_planner(Query *parse, double tuple_fraction)
628 List *tlist = parse->targetList;
630 List *current_pathkeys;
633 if (parse->setOperations)
635 List *set_sortclauses;
638 * Construct the plan for set operations. The result will not
639 * need any work except perhaps a top-level sort and/or LIMIT.
641 result_plan = plan_set_operations(parse,
645 * Calculate pathkeys representing the sort order (if any) of the
646 * set operation's result. We have to do this before overwriting
647 * the sort key information...
649 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
650 result_plan->targetlist);
651 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
654 * We should not need to call preprocess_targetlist, since we must
655 * be in a SELECT query node. Instead, use the targetlist
656 * returned by plan_set_operations (since this tells whether it
657 * returned any resjunk columns!), and transfer any sort key
658 * information from the original tlist.
660 Assert(parse->commandType == CMD_SELECT);
662 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
665 * Can't handle FOR UPDATE here (parser should have checked
666 * already, but let's make sure).
670 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
671 errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
674 * Calculate pathkeys that represent result ordering requirements
676 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
678 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
682 /* No set operations, do regular planning */
684 List *group_pathkeys;
685 AttrNumber *groupColIdx = NULL;
686 bool need_tlist_eval = true;
688 double sub_tuple_fraction;
691 double dNumGroups = 0;
693 AggClauseCounts agg_counts;
694 int numGroupCols = list_length(parse->groupClause);
695 bool use_hashed_grouping = false;
697 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
699 /* Preprocess targetlist */
700 tlist = preprocess_targetlist(parse, tlist);
703 * Generate appropriate target list for subplan; may be different
704 * from tlist if grouping or aggregation is needed.
706 sub_tlist = make_subplanTargetList(parse, tlist,
707 &groupColIdx, &need_tlist_eval);
710 * Calculate pathkeys that represent grouping/ordering
713 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
715 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
719 * Will need actual number of aggregates for estimating costs.
721 * Note: we do not attempt to detect duplicate aggregates here; a
722 * somewhat-overestimated count is okay for our present purposes.
724 * Note: think not that we can turn off hasAggs if we find no aggs.
725 * It is possible for constant-expression simplification to remove
726 * all explicit references to aggs, but we still have to follow
727 * the aggregate semantics (eg, producing only one output row).
731 count_agg_clauses((Node *) tlist, &agg_counts);
732 count_agg_clauses(parse->havingQual, &agg_counts);
736 * Figure out whether we need a sorted result from query_planner.
738 * If we have a GROUP BY clause, then we want a result sorted
739 * properly for grouping. Otherwise, if there is an ORDER BY
740 * clause, we want to sort by the ORDER BY clause. (Note: if we
741 * have both, and ORDER BY is a superset of GROUP BY, it would be
742 * tempting to request sort by ORDER BY --- but that might just
743 * leave us failing to exploit an available sort order at all.
744 * Needs more thought...)
746 if (parse->groupClause)
747 parse->query_pathkeys = group_pathkeys;
748 else if (parse->sortClause)
749 parse->query_pathkeys = sort_pathkeys;
751 parse->query_pathkeys = NIL;
754 * Adjust tuple_fraction if we see that we are going to apply
755 * limiting/grouping/aggregation/etc. This is not overridable by
756 * the caller, since it reflects plan actions that this routine
757 * will certainly take, not assumptions about context.
759 if (parse->limitCount != NULL)
762 * A LIMIT clause limits the absolute number of tuples
763 * returned. However, if it's not a constant LIMIT then we
764 * have to punt; for lack of a better idea, assume 10% of the
765 * plan's result is wanted.
767 double limit_fraction = 0.0;
769 if (IsA(parse->limitCount, Const))
771 Const *limitc = (Const *) parse->limitCount;
772 int32 count = DatumGetInt32(limitc->constvalue);
775 * A NULL-constant LIMIT represents "LIMIT ALL", which we
776 * treat the same as no limit (ie, expect to retrieve all
779 if (!limitc->constisnull && count > 0)
781 limit_fraction = (double) count;
782 /* We must also consider the OFFSET, if present */
783 if (parse->limitOffset != NULL)
785 if (IsA(parse->limitOffset, Const))
789 limitc = (Const *) parse->limitOffset;
790 offset = DatumGetInt32(limitc->constvalue);
791 if (!limitc->constisnull && offset > 0)
792 limit_fraction += (double) offset;
796 /* OFFSET is an expression ... punt ... */
797 limit_fraction = 0.10;
804 /* LIMIT is an expression ... punt ... */
805 limit_fraction = 0.10;
808 if (limit_fraction > 0.0)
811 * If we have absolute limits from both caller and LIMIT,
812 * use the smaller value; if one is fractional and the
813 * other absolute, treat the fraction as a fraction of the
814 * absolute value; else we can multiply the two fractions
817 if (tuple_fraction >= 1.0)
819 if (limit_fraction >= 1.0)
822 tuple_fraction = Min(tuple_fraction, limit_fraction);
826 /* caller absolute, limit fractional */
827 tuple_fraction *= limit_fraction;
828 if (tuple_fraction < 1.0)
829 tuple_fraction = 1.0;
832 else if (tuple_fraction > 0.0)
834 if (limit_fraction >= 1.0)
836 /* caller fractional, limit absolute */
837 tuple_fraction *= limit_fraction;
838 if (tuple_fraction < 1.0)
839 tuple_fraction = 1.0;
843 /* both fractional */
844 tuple_fraction *= limit_fraction;
849 /* no info from caller, just use limit */
850 tuple_fraction = limit_fraction;
856 * With grouping or aggregation, the tuple fraction to pass to
857 * query_planner() may be different from what it is at top level.
859 sub_tuple_fraction = tuple_fraction;
861 if (parse->groupClause)
864 * In GROUP BY mode, we have the little problem that we don't
865 * really know how many input tuples will be needed to make a
866 * group, so we can't translate an output LIMIT count into an
867 * input count. For lack of a better idea, assume 25% of the
868 * input data will be processed if there is any output limit.
869 * However, if the caller gave us a fraction rather than an
870 * absolute count, we can keep using that fraction (which
871 * amounts to assuming that all the groups are about the same
874 if (sub_tuple_fraction >= 1.0)
875 sub_tuple_fraction = 0.25;
878 * If both GROUP BY and ORDER BY are specified, we will need
879 * two levels of sort --- and, therefore, certainly need to
880 * read all the input tuples --- unless ORDER BY is a subset
881 * of GROUP BY. (We have not yet canonicalized the pathkeys,
882 * so must use the slower noncanonical comparison method.)
884 if (parse->groupClause && parse->sortClause &&
885 !noncanonical_pathkeys_contained_in(sort_pathkeys,
887 sub_tuple_fraction = 0.0;
889 else if (parse->hasAggs)
892 * Ungrouped aggregate will certainly want all the input
895 sub_tuple_fraction = 0.0;
897 else if (parse->distinctClause)
900 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
901 * number of input tuples per output tuple. Handle the same
904 if (sub_tuple_fraction >= 1.0)
905 sub_tuple_fraction = 0.25;
909 * Generate the best unsorted and presorted paths for this Query
910 * (but note there may not be any presorted path).
912 query_planner(parse, sub_tlist, sub_tuple_fraction,
913 &cheapest_path, &sorted_path);
916 * We couldn't canonicalize group_pathkeys and sort_pathkeys
917 * before running query_planner(), so do it now.
919 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
920 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
923 * Consider whether we might want to use hashed grouping.
925 if (parse->groupClause)
928 double cheapest_path_rows;
929 int cheapest_path_width;
932 * Beware in this section of the possibility that
933 * cheapest_path->parent is NULL. This could happen if user
934 * does something silly like SELECT 'foo' GROUP BY 1;
936 if (cheapest_path->parent)
938 cheapest_path_rows = cheapest_path->parent->rows;
939 cheapest_path_width = cheapest_path->parent->width;
943 cheapest_path_rows = 1; /* assume non-set result */
944 cheapest_path_width = 100; /* arbitrary */
948 * Always estimate the number of groups. We can't do this
949 * until after running query_planner(), either.
951 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
953 dNumGroups = estimate_num_groups(parse,
956 /* Also want it as a long int --- but 'ware overflow! */
957 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
960 * Check can't-do-it conditions, including whether the
961 * grouping operators are hashjoinable.
963 * Executor doesn't support hashed aggregation with DISTINCT
964 * aggregates. (Doing so would imply storing *all* the input
965 * values in the hash table, which seems like a certain
968 if (!enable_hashagg || !hash_safe_grouping(parse))
969 use_hashed_grouping = false;
970 else if (agg_counts.numDistinctAggs != 0)
971 use_hashed_grouping = false;
975 * Use hashed grouping if (a) we think we can fit the
976 * hashtable into work_mem, *and* (b) the estimated cost
977 * is no more than doing it the other way. While avoiding
978 * the need for sorted input is usually a win, the fact
979 * that the output won't be sorted may be a loss; so we
980 * need to do an actual cost comparison.
984 /* Estimate per-hash-entry space at tuple width... */
985 hashentrysize = cheapest_path_width;
986 /* plus space for pass-by-ref transition values... */
987 hashentrysize += agg_counts.transitionSpace;
988 /* plus the per-hash-entry overhead */
989 hashentrysize += hash_agg_entry_size(agg_counts.numAggs);
991 if (hashentrysize * dNumGroups <= work_mem * 1024L)
994 * Okay, do the cost comparison. We need to consider
995 * cheapest_path + hashagg [+ final sort] versus
996 * either cheapest_path [+ sort] + group or agg [+
997 * final sort] or presorted_path + group or agg [+
998 * final sort] where brackets indicate a step that may
999 * not be needed. We assume query_planner() will have
1000 * returned a presorted path only if it's a winner
1001 * compared to cheapest_path for this purpose.
1003 * These path variables are dummies that just hold cost
1004 * fields; we don't make actual Paths for these steps.
1009 cost_agg(&hashed_p, parse,
1010 AGG_HASHED, agg_counts.numAggs,
1011 numGroupCols, dNumGroups,
1012 cheapest_path->startup_cost,
1013 cheapest_path->total_cost,
1014 cheapest_path_rows);
1015 /* Result of hashed agg is always unsorted */
1017 cost_sort(&hashed_p, parse, sort_pathkeys,
1018 hashed_p.total_cost,
1020 cheapest_path_width);
1024 sorted_p.startup_cost = sorted_path->startup_cost;
1025 sorted_p.total_cost = sorted_path->total_cost;
1026 current_pathkeys = sorted_path->pathkeys;
1030 sorted_p.startup_cost = cheapest_path->startup_cost;
1031 sorted_p.total_cost = cheapest_path->total_cost;
1032 current_pathkeys = cheapest_path->pathkeys;
1034 if (!pathkeys_contained_in(group_pathkeys,
1037 cost_sort(&sorted_p, parse, group_pathkeys,
1038 sorted_p.total_cost,
1040 cheapest_path_width);
1041 current_pathkeys = group_pathkeys;
1044 cost_agg(&sorted_p, parse,
1045 AGG_SORTED, agg_counts.numAggs,
1046 numGroupCols, dNumGroups,
1047 sorted_p.startup_cost,
1048 sorted_p.total_cost,
1049 cheapest_path_rows);
1051 cost_group(&sorted_p, parse,
1052 numGroupCols, dNumGroups,
1053 sorted_p.startup_cost,
1054 sorted_p.total_cost,
1055 cheapest_path_rows);
1056 /* The Agg or Group node will preserve ordering */
1057 if (sort_pathkeys &&
1058 !pathkeys_contained_in(sort_pathkeys,
1061 cost_sort(&sorted_p, parse, sort_pathkeys,
1062 sorted_p.total_cost,
1064 cheapest_path_width);
1068 * Now make the decision using the top-level tuple
1069 * fraction. First we have to convert an absolute
1070 * count (LIMIT) into fractional form.
1072 if (tuple_fraction >= 1.0)
1073 tuple_fraction /= dNumGroups;
1075 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1076 tuple_fraction) < 0)
1078 /* Hashed is cheaper, so use it */
1079 use_hashed_grouping = true;
1086 * Select the best path and create a plan to execute it.
1088 * If we are doing hashed grouping, we will always read all the input
1089 * tuples, so use the cheapest-total path. Otherwise, trust
1090 * query_planner's decision about which to use.
1092 if (sorted_path && !use_hashed_grouping)
1094 result_plan = create_plan(parse, sorted_path);
1095 current_pathkeys = sorted_path->pathkeys;
1099 result_plan = create_plan(parse, cheapest_path);
1100 current_pathkeys = cheapest_path->pathkeys;
1104 * create_plan() returns a plan with just a "flat" tlist of
1105 * required Vars. Usually we need to insert the sub_tlist as the
1106 * tlist of the top plan node. However, we can skip that if we
1107 * determined that whatever query_planner chose to return will be
1110 if (need_tlist_eval)
1113 * If the top-level plan node is one that cannot do expression
1114 * evaluation, we must insert a Result node to project the
1117 if (!is_projection_capable_plan(result_plan))
1119 result_plan = (Plan *) make_result(sub_tlist, NULL,
1125 * Otherwise, just replace the subplan's flat tlist with
1126 * the desired tlist.
1128 result_plan->targetlist = sub_tlist;
1132 * Also, account for the cost of evaluation of the sub_tlist.
1134 * Up to now, we have only been dealing with "flat" tlists,
1135 * containing just Vars. So their evaluation cost is zero
1136 * according to the model used by cost_qual_eval() (or if you
1137 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1138 * can avoid accounting for tlist cost throughout
1139 * query_planner() and subroutines. But now we've inserted a
1140 * tlist that might contain actual operators, sub-selects, etc
1141 * --- so we'd better account for its cost.
1143 * Below this point, any tlist eval cost for added-on nodes
1144 * should be accounted for as we create those nodes.
1145 * Presently, of the node types we can add on, only Agg and
1146 * Group project new tlists (the rest just copy their input
1147 * tuples) --- so make_agg() and make_group() are responsible
1148 * for computing the added cost.
1150 cost_qual_eval(&tlist_cost, sub_tlist);
1151 result_plan->startup_cost += tlist_cost.startup;
1152 result_plan->total_cost += tlist_cost.startup +
1153 tlist_cost.per_tuple * result_plan->plan_rows;
1158 * Since we're using query_planner's tlist and not the one
1159 * make_subplanTargetList calculated, we have to refigure any
1160 * grouping-column indexes make_subplanTargetList computed.
1162 locate_grouping_columns(parse, tlist, result_plan->targetlist,
1167 * Insert AGG or GROUP node if needed, plus an explicit sort step
1170 * HAVING clause, if any, becomes qual of the Agg or Group node.
1172 if (use_hashed_grouping)
1174 /* Hashed aggregate plan --- no sort needed */
1175 result_plan = (Plan *) make_agg(parse,
1177 (List *) parse->havingQual,
1184 /* Hashed aggregation produces randomly-ordered results */
1185 current_pathkeys = NIL;
1187 else if (parse->hasAggs)
1189 /* Plain aggregate plan --- sort if needed */
1190 AggStrategy aggstrategy;
1192 if (parse->groupClause)
1194 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1196 result_plan = (Plan *)
1197 make_sort_from_groupcols(parse,
1201 current_pathkeys = group_pathkeys;
1203 aggstrategy = AGG_SORTED;
1206 * The AGG node will not change the sort ordering of its
1207 * groups, so current_pathkeys describes the result too.
1212 aggstrategy = AGG_PLAIN;
1213 /* Result will be only one row anyway; no sort order */
1214 current_pathkeys = NIL;
1217 result_plan = (Plan *) make_agg(parse,
1219 (List *) parse->havingQual,
1227 else if (parse->groupClause)
1230 * GROUP BY without aggregation, so insert a group node (plus the
1231 * appropriate sort node, if necessary).
1233 * Add an explicit sort if we couldn't make the path come
1234 * out the way the GROUP node needs it.
1236 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1238 result_plan = (Plan *)
1239 make_sort_from_groupcols(parse,
1243 current_pathkeys = group_pathkeys;
1246 result_plan = (Plan *) make_group(parse,
1248 (List *) parse->havingQual,
1253 /* The Group node won't change sort ordering */
1255 else if (parse->hasHavingQual)
1258 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1259 * This is a degenerate case in which we are supposed to emit
1260 * either 0 or 1 row depending on whether HAVING succeeds.
1261 * Furthermore, there cannot be any variables in either HAVING
1262 * or the targetlist, so we actually do not need the FROM table
1263 * at all! We can just throw away the plan-so-far and generate
1264 * a Result node. This is a sufficiently unusual corner case
1265 * that it's not worth contorting the structure of this routine
1266 * to avoid having to generate the plan in the first place.
1268 result_plan = (Plan *) make_result(tlist,
1272 } /* end of if (setOperations) */
1275 * If we were not able to make the plan come out in the right order,
1276 * add an explicit sort step.
1278 if (parse->sortClause)
1280 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1282 result_plan = (Plan *)
1283 make_sort_from_sortclauses(parse,
1286 current_pathkeys = sort_pathkeys;
1291 * If there is a DISTINCT clause, add the UNIQUE node.
1293 if (parse->distinctClause)
1295 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1298 * If there was grouping or aggregation, leave plan_rows as-is
1299 * (ie, assume the result was already mostly unique). If not,
1300 * it's reasonable to assume the UNIQUE filter has effects
1301 * comparable to GROUP BY.
1303 if (!parse->groupClause && !parse->hasHavingQual && !parse->hasAggs)
1305 List *distinctExprs;
1307 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1309 result_plan->plan_rows = estimate_num_groups(parse,
1311 result_plan->plan_rows);
1316 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1318 if (parse->limitOffset || parse->limitCount)
1320 result_plan = (Plan *) make_limit(result_plan,
1326 * Return the actual output ordering in query_pathkeys for possible
1327 * use by an outer query level.
1329 parse->query_pathkeys = current_pathkeys;
1335 * hash_safe_grouping - are grouping operators hashable?
1337 * We assume hashed aggregation will work if the datatype's equality operator
1338 * is marked hashjoinable.
1341 hash_safe_grouping(Query *parse)
1345 foreach(gl, parse->groupClause)
1347 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1348 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1352 optup = equality_oper(exprType((Node *) tle->expr), true);
1355 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1356 ReleaseSysCache(optup);
1364 * make_subplanTargetList
1365 * Generate appropriate target list when grouping is required.
1367 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1368 * above the result of query_planner, we typically want to pass a different
1369 * target list to query_planner than the outer plan nodes should have.
1370 * This routine generates the correct target list for the subplan.
1372 * The initial target list passed from the parser already contains entries
1373 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1374 * for variables used only in HAVING clauses; so we need to add those
1375 * variables to the subplan target list. Also, we flatten all expressions
1376 * except GROUP BY items into their component variables; the other expressions
1377 * will be computed by the inserted nodes rather than by the subplan.
1378 * For example, 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, there's nothing to do here;
1419 * query_planner should receive the unmodified target list.
1421 if (!parse->hasAggs && !parse->groupClause && !parse->hasHavingQual)
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((Expr *) groupexpr,
1471 list_length(sub_tlist) + 1,
1474 sub_tlist = lappend(sub_tlist, te);
1475 *need_tlist_eval = true; /* it's not flat anymore */
1478 /* and save its resno */
1479 grpColIdx[keyno++] = te->resno;
1487 * locate_grouping_columns
1488 * Locate grouping columns in the tlist chosen by query_planner.
1490 * This is only needed if we don't use the sub_tlist chosen by
1491 * make_subplanTargetList. We have to forget the column indexes found
1492 * by that routine and re-locate the grouping vars in the real sub_tlist.
1495 locate_grouping_columns(Query *parse,
1498 AttrNumber *groupColIdx)
1504 * No work unless grouping.
1506 if (!parse->groupClause)
1508 Assert(groupColIdx == NULL);
1511 Assert(groupColIdx != NULL);
1513 foreach(gl, parse->groupClause)
1515 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1516 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1517 TargetEntry *te = NULL;
1520 foreach(sl, sub_tlist)
1522 te = (TargetEntry *) lfirst(sl);
1523 if (equal(groupexpr, te->expr))
1527 elog(ERROR, "failed to locate grouping columns");
1529 groupColIdx[keyno++] = te->resno;
1534 * postprocess_setop_tlist
1535 * Fix up targetlist returned by plan_set_operations().
1537 * We need to transpose sort key info from the orig_tlist into new_tlist.
1538 * NOTE: this would not be good enough if we supported resjunk sort keys
1539 * for results of set operations --- then, we'd need to project a whole
1540 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1541 * find any resjunk columns in orig_tlist.
1544 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1547 ListCell *orig_tlist_item = list_head(orig_tlist);
1549 foreach(l, new_tlist)
1551 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1552 TargetEntry *orig_tle;
1554 /* ignore resjunk columns in setop result */
1555 if (new_tle->resjunk)
1558 Assert(orig_tlist_item != NULL);
1559 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1560 orig_tlist_item = lnext(orig_tlist_item);
1561 if (orig_tle->resjunk) /* should not happen */
1562 elog(ERROR, "resjunk output columns are not implemented");
1563 Assert(new_tle->resno == orig_tle->resno);
1564 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1566 if (orig_tlist_item != NULL)
1567 elog(ERROR, "resjunk output columns are not implemented");