1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2006, 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 1.204 2006/07/26 00:34:48 momjian Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "executor/executor.h"
22 #include "executor/nodeAgg.h"
23 #include "miscadmin.h"
24 #include "nodes/makefuncs.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/cost.h"
27 #include "optimizer/pathnode.h"
28 #include "optimizer/paths.h"
29 #include "optimizer/planmain.h"
30 #include "optimizer/planner.h"
31 #include "optimizer/prep.h"
32 #include "optimizer/subselect.h"
33 #include "optimizer/tlist.h"
34 #include "optimizer/var.h"
35 #ifdef OPTIMIZER_DEBUG
36 #include "nodes/print.h"
38 #include "parser/parse_expr.h"
39 #include "parser/parse_oper.h"
40 #include "parser/parsetree.h"
41 #include "utils/syscache.h"
44 ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
47 /* Expression kind codes for preprocess_expression */
48 #define EXPRKIND_QUAL 0
49 #define EXPRKIND_TARGET 1
50 #define EXPRKIND_RTFUNC 2
51 #define EXPRKIND_LIMIT 3
52 #define EXPRKIND_ININFO 4
53 #define EXPRKIND_APPINFO 5
56 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
57 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
58 static Plan *inheritance_planner(PlannerInfo *root);
59 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
60 static double preprocess_limit(PlannerInfo *root,
61 double tuple_fraction,
62 int64 *offset_est, int64 *count_est);
63 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
64 Path *cheapest_path, Path *sorted_path,
65 double dNumGroups, AggClauseCounts *agg_counts);
66 static bool hash_safe_grouping(PlannerInfo *root);
67 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
68 AttrNumber **groupColIdx, bool *need_tlist_eval);
69 static void locate_grouping_columns(PlannerInfo *root,
72 AttrNumber *groupColIdx);
73 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
76 /*****************************************************************************
78 * Query optimizer entry point
80 *****************************************************************************/
82 planner(Query *parse, bool isCursor, int cursorOptions,
83 ParamListInfo boundParams)
85 double tuple_fraction;
87 Index save_PlannerQueryLevel;
88 List *save_PlannerParamList;
89 ParamListInfo save_PlannerBoundParamList;
92 * The planner can be called recursively (an example is when
93 * eval_const_expressions tries to pre-evaluate an SQL function). So,
94 * these global state variables must be saved and restored.
96 * Query level and the param list cannot be moved into the per-query
97 * PlannerInfo structure since their whole purpose is communication across
98 * multiple sub-queries. Also, boundParams is explicitly info from outside
99 * the query, and so is likewise better handled as a global variable.
101 * Note we do NOT save and restore PlannerPlanId: it exists to assign
102 * unique IDs to SubPlan nodes, and we want those IDs to be unique for the
103 * life of a backend. Also, PlannerInitPlan is saved/restored in
104 * subquery_planner, not here.
106 save_PlannerQueryLevel = PlannerQueryLevel;
107 save_PlannerParamList = PlannerParamList;
108 save_PlannerBoundParamList = PlannerBoundParamList;
110 /* Initialize state for handling outer-level references and params */
111 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
112 PlannerParamList = NIL;
113 PlannerBoundParamList = boundParams;
115 /* Determine what fraction of the plan is likely to be scanned */
119 * We have no real idea how many tuples the user will ultimately FETCH
120 * from a cursor, but it seems a good bet that he doesn't want 'em
121 * all. Optimize for 10% retrieval (you gotta better number? Should
122 * this be a SETtable parameter?)
124 tuple_fraction = 0.10;
128 /* Default assumption is we need all the tuples */
129 tuple_fraction = 0.0;
132 /* primary planning entry point (may recurse for subqueries) */
133 result_plan = subquery_planner(parse, tuple_fraction, NULL);
135 /* check we popped out the right number of levels */
136 Assert(PlannerQueryLevel == 0);
139 * If creating a plan for a scrollable cursor, make sure it can run
140 * backwards on demand. Add a Material node at the top at need.
142 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
144 if (!ExecSupportsBackwardScan(result_plan))
145 result_plan = materialize_finished_plan(result_plan);
148 /* final cleanup of the plan */
149 result_plan = set_plan_references(result_plan, parse->rtable);
151 /* executor wants to know total number of Params used overall */
152 result_plan->nParamExec = list_length(PlannerParamList);
154 /* restore state for outer planner, if any */
155 PlannerQueryLevel = save_PlannerQueryLevel;
156 PlannerParamList = save_PlannerParamList;
157 PlannerBoundParamList = save_PlannerBoundParamList;
163 /*--------------------
165 * Invokes the planner on a subquery. We recurse to here for each
166 * sub-SELECT found in the query tree.
168 * parse is the querytree produced by the parser & rewriter.
169 * tuple_fraction is the fraction of tuples we expect will be retrieved.
170 * tuple_fraction is interpreted as explained for grouping_planner, below.
172 * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
173 * the output sort ordering of the completed plan.
175 * Basically, this routine does the stuff that should only be done once
176 * per Query object. It then calls grouping_planner. At one time,
177 * grouping_planner could be invoked recursively on the same Query object;
178 * that's not currently true, but we keep the separation between the two
179 * routines anyway, in case we need it again someday.
181 * subquery_planner will be called recursively to handle sub-Query nodes
182 * found within the query's expressions and rangetable.
184 * Returns a query plan.
185 *--------------------
188 subquery_planner(Query *parse, double tuple_fraction,
189 List **subquery_pathkeys)
191 List *saved_initplan = PlannerInitPlan;
192 int saved_planid = PlannerPlanId;
198 /* Set up for a new level of subquery */
200 PlannerInitPlan = NIL;
202 /* Create a PlannerInfo data structure for this subquery */
203 root = makeNode(PlannerInfo);
205 root->in_info_list = NIL;
206 root->append_rel_list = NIL;
209 * Look for IN clauses at the top level of WHERE, and transform them into
210 * joins. Note that this step only handles IN clauses originally at top
211 * level of WHERE; if we pull up any subqueries in the next step, their
212 * INs are processed just before pulling them up.
214 if (parse->hasSubLinks)
215 parse->jointree->quals = pull_up_IN_clauses(root,
216 parse->jointree->quals);
219 * Check to see if any subqueries in the rangetable can be merged into
222 parse->jointree = (FromExpr *)
223 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
226 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
227 * avoid the expense of doing flatten_join_alias_vars(). Also check for
228 * outer joins --- if none, we can skip reduce_outer_joins() and some
229 * other processing. This must be done after we have done
230 * pull_up_subqueries, of course.
232 * Note: if reduce_outer_joins manages to eliminate all outer joins,
233 * root->hasOuterJoins is not reset currently. This is OK since its
234 * purpose is merely to suppress unnecessary processing in simple cases.
236 root->hasJoinRTEs = false;
237 root->hasOuterJoins = false;
238 foreach(l, parse->rtable)
240 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
242 if (rte->rtekind == RTE_JOIN)
244 root->hasJoinRTEs = true;
245 if (IS_OUTER_JOIN(rte->jointype))
247 root->hasOuterJoins = true;
248 /* Can quit scanning once we find an outer join */
255 * Expand any rangetable entries that are inheritance sets into "append
256 * relations". This can add entries to the rangetable, but they must be
257 * plain base relations not joins, so it's OK (and marginally more
258 * efficient) to do it after checking for join RTEs. We must do it after
259 * pulling up subqueries, else we'd fail to handle inherited tables in
262 expand_inherited_tables(root);
265 * Set hasHavingQual to remember if HAVING clause is present. Needed
266 * because preprocess_expression will reduce a constant-true condition to
267 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
269 root->hasHavingQual = (parse->havingQual != NULL);
271 /* Clear this flag; might get set in distribute_qual_to_rels */
272 root->hasPseudoConstantQuals = false;
275 * Do expression preprocessing on targetlist and quals.
277 parse->targetList = (List *)
278 preprocess_expression(root, (Node *) parse->targetList,
281 preprocess_qual_conditions(root, (Node *) parse->jointree);
283 parse->havingQual = preprocess_expression(root, parse->havingQual,
286 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
288 parse->limitCount = preprocess_expression(root, parse->limitCount,
291 root->in_info_list = (List *)
292 preprocess_expression(root, (Node *) root->in_info_list,
294 root->append_rel_list = (List *)
295 preprocess_expression(root, (Node *) root->append_rel_list,
298 /* Also need to preprocess expressions for function RTEs */
299 foreach(l, parse->rtable)
301 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
303 if (rte->rtekind == RTE_FUNCTION)
304 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
309 * In some cases we may want to transfer a HAVING clause into WHERE. We
310 * cannot do so if the HAVING clause contains aggregates (obviously) or
311 * volatile functions (since a HAVING clause is supposed to be executed
312 * only once per group). Also, it may be that the clause is so expensive
313 * to execute that we're better off doing it only once per group, despite
314 * the loss of selectivity. This is hard to estimate short of doing the
315 * entire planning process twice, so we use a heuristic: clauses
316 * containing subplans are left in HAVING. Otherwise, we move or copy the
317 * HAVING clause into WHERE, in hopes of eliminating tuples before
318 * aggregation instead of after.
320 * If the query has explicit grouping then we can simply move such a
321 * clause into WHERE; any group that fails the clause will not be in the
322 * output because none of its tuples will reach the grouping or
323 * aggregation stage. Otherwise we must have a degenerate (variable-free)
324 * HAVING clause, which we put in WHERE so that query_planner() can use it
325 * in a gating Result node, but also keep in HAVING to ensure that we
326 * don't emit a bogus aggregated row. (This could be done better, but it
327 * seems not worth optimizing.)
329 * Note that both havingQual and parse->jointree->quals are in
330 * implicitly-ANDed-list form at this point, even though they are declared
334 foreach(l, (List *) parse->havingQual)
336 Node *havingclause = (Node *) lfirst(l);
338 if (contain_agg_clause(havingclause) ||
339 contain_volatile_functions(havingclause) ||
340 contain_subplans(havingclause))
342 /* keep it in HAVING */
343 newHaving = lappend(newHaving, havingclause);
345 else if (parse->groupClause)
347 /* move it to WHERE */
348 parse->jointree->quals = (Node *)
349 lappend((List *) parse->jointree->quals, havingclause);
353 /* put a copy in WHERE, keep it in HAVING */
354 parse->jointree->quals = (Node *)
355 lappend((List *) parse->jointree->quals,
356 copyObject(havingclause));
357 newHaving = lappend(newHaving, havingclause);
360 parse->havingQual = (Node *) newHaving;
363 * If we have any outer joins, try to reduce them to plain inner joins.
364 * This step is most easily done after we've done expression
367 if (root->hasOuterJoins)
368 reduce_outer_joins(root);
371 * Do the main planning. If we have an inherited target relation, that
372 * needs special processing, else go straight to grouping_planner.
374 if (parse->resultRelation &&
375 rt_fetch(parse->resultRelation, parse->rtable)->inh)
376 plan = inheritance_planner(root);
378 plan = grouping_planner(root, tuple_fraction);
381 * If any subplans were generated, or if we're inside a subplan, build
382 * initPlan list and extParam/allParam sets for plan nodes, and attach the
383 * initPlans to the top plan node.
385 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
386 SS_finalize_plan(plan, parse->rtable);
388 /* Return sort ordering info if caller wants it */
389 if (subquery_pathkeys)
390 *subquery_pathkeys = root->query_pathkeys;
392 /* Return to outer subquery context */
394 PlannerInitPlan = saved_initplan;
395 /* we do NOT restore PlannerPlanId; that's not an oversight! */
401 * preprocess_expression
402 * Do subquery_planner's preprocessing work for an expression,
403 * which can be a targetlist, a WHERE clause (including JOIN/ON
404 * conditions), or a HAVING clause.
407 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
410 * Fall out quickly if expression is empty. This occurs often enough to
411 * be worth checking. Note that null->null is the correct conversion for
412 * implicit-AND result format, too.
418 * If the query has any join RTEs, replace join alias variables with
419 * base-relation variables. We must do this before sublink processing,
420 * else sublinks expanded out from join aliases wouldn't get processed.
422 if (root->hasJoinRTEs)
423 expr = flatten_join_alias_vars(root, expr);
426 * Simplify constant expressions.
428 * Note: this also flattens nested AND and OR expressions into N-argument
429 * form. All processing of a qual expression after this point must be
430 * careful to maintain AND/OR flatness --- that is, do not generate a tree
431 * with AND directly under AND, nor OR directly under OR.
433 * Because this is a relatively expensive process, we skip it when the
434 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
435 * expression will only be evaluated once anyway, so no point in
436 * pre-simplifying; we can't execute it any faster than the executor can,
437 * and we will waste cycles copying the tree. Notice however that we
438 * still must do it for quals (to get AND/OR flatness); and if we are in a
439 * subquery we should not assume it will be done only once.
441 if (root->parse->jointree->fromlist != NIL ||
442 kind == EXPRKIND_QUAL ||
443 PlannerQueryLevel > 1)
444 expr = eval_const_expressions(expr);
447 * If it's a qual or havingQual, canonicalize it.
449 if (kind == EXPRKIND_QUAL)
451 expr = (Node *) canonicalize_qual((Expr *) expr);
453 #ifdef OPTIMIZER_DEBUG
454 printf("After canonicalize_qual()\n");
459 /* Expand SubLinks to SubPlans */
460 if (root->parse->hasSubLinks)
461 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
464 * XXX do not insert anything here unless you have grokked the comments in
465 * SS_replace_correlation_vars ...
468 /* Replace uplevel vars with Param nodes */
469 if (PlannerQueryLevel > 1)
470 expr = SS_replace_correlation_vars(expr);
473 * If it's a qual or havingQual, convert it to implicit-AND format. (We
474 * don't want to do this before eval_const_expressions, since the latter
475 * would be unable to simplify a top-level AND correctly. Also,
476 * SS_process_sublinks expects explicit-AND format.)
478 if (kind == EXPRKIND_QUAL)
479 expr = (Node *) make_ands_implicit((Expr *) expr);
485 * preprocess_qual_conditions
486 * Recursively scan the query's jointree and do subquery_planner's
487 * preprocessing work on each qual condition found therein.
490 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
494 if (IsA(jtnode, RangeTblRef))
496 /* nothing to do here */
498 else if (IsA(jtnode, FromExpr))
500 FromExpr *f = (FromExpr *) jtnode;
503 foreach(l, f->fromlist)
504 preprocess_qual_conditions(root, lfirst(l));
506 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
508 else if (IsA(jtnode, JoinExpr))
510 JoinExpr *j = (JoinExpr *) jtnode;
512 preprocess_qual_conditions(root, j->larg);
513 preprocess_qual_conditions(root, j->rarg);
515 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
518 elog(ERROR, "unrecognized node type: %d",
519 (int) nodeTag(jtnode));
523 * inheritance_planner
524 * Generate a plan in the case where the result relation is an
527 * We have to handle this case differently from cases where a source relation
528 * is an inheritance set. Source inheritance is expanded at the bottom of the
529 * plan tree (see allpaths.c), but target inheritance has to be expanded at
530 * the top. The reason is that for UPDATE, each target relation needs a
531 * different targetlist matching its own column set. Also, for both UPDATE
532 * and DELETE, the executor needs the Append plan node at the top, else it
533 * can't keep track of which table is the current target table. Fortunately,
534 * the UPDATE/DELETE target can never be the nullable side of an outer join,
535 * so it's OK to generate the plan this way.
537 * Returns a query plan.
540 inheritance_planner(PlannerInfo *root)
542 Query *parse = root->parse;
543 int parentRTindex = parse->resultRelation;
544 List *subplans = NIL;
549 subroot.parse = NULL; /* catch it if no matches in loop */
551 parse->resultRelations = NIL;
553 foreach(l, root->append_rel_list)
555 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
558 /* append_rel_list contains all append rels; ignore others */
559 if (appinfo->parent_relid != parentRTindex)
562 /* Build target-relations list for the executor */
563 parse->resultRelations = lappend_int(parse->resultRelations,
564 appinfo->child_relid);
567 * Generate modified query with this rel as target. We have to be
568 * prepared to translate varnos in in_info_list as well as in the
571 memcpy(&subroot, root, sizeof(PlannerInfo));
572 subroot.parse = (Query *)
573 adjust_appendrel_attrs((Node *) parse,
575 subroot.in_info_list = (List *)
576 adjust_appendrel_attrs((Node *) root->in_info_list,
578 /* There shouldn't be any OJ info to translate, as yet */
579 Assert(subroot.oj_info_list == NIL);
582 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
584 subplans = lappend(subplans, subplan);
586 /* Save preprocessed tlist from first rel for use in Append */
588 tlist = subplan->targetlist;
592 * Planning might have modified the rangetable, due to changes of the
593 * Query structures inside subquery RTEs. We have to ensure that this
594 * gets propagated back to the master copy. But can't do this until we
595 * are done planning, because all the calls to grouping_planner need
596 * virgin sub-Queries to work from. (We are effectively assuming that
597 * sub-Queries will get planned identically each time, or at least that
598 * the impacts on their rangetables will be the same each time.)
600 * XXX should clean this up someday
602 parse->rtable = subroot.parse->rtable;
604 /* Mark result as unordered (probably unnecessary) */
605 root->query_pathkeys = NIL;
607 return (Plan *) make_append(subplans, true, tlist);
610 /*--------------------
612 * Perform planning steps related to grouping, aggregation, etc.
613 * This primarily means adding top-level processing to the basic
614 * query plan produced by query_planner.
616 * tuple_fraction is the fraction of tuples we expect will be retrieved
618 * tuple_fraction is interpreted as follows:
619 * 0: expect all tuples to be retrieved (normal case)
620 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
621 * from the plan to be retrieved
622 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
623 * expected to be retrieved (ie, a LIMIT specification)
625 * Returns a query plan. Also, root->query_pathkeys is returned as the
626 * actual output ordering of the plan (in pathkey format).
627 *--------------------
630 grouping_planner(PlannerInfo *root, double tuple_fraction)
632 Query *parse = root->parse;
633 List *tlist = parse->targetList;
634 int64 offset_est = 0;
637 List *current_pathkeys;
639 double dNumGroups = 0;
641 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
642 if (parse->limitCount || parse->limitOffset)
643 tuple_fraction = preprocess_limit(root, tuple_fraction,
644 &offset_est, &count_est);
646 if (parse->setOperations)
648 List *set_sortclauses;
651 * If there's a top-level ORDER BY, assume we have to fetch all the
652 * tuples. This might seem too simplistic given all the hackery below
653 * to possibly avoid the sort ... but a nonzero tuple_fraction is only
654 * of use to plan_set_operations() when the setop is UNION ALL, and
655 * the result of UNION ALL is always unsorted.
657 if (parse->sortClause)
658 tuple_fraction = 0.0;
661 * Construct the plan for set operations. The result will not need
662 * any work except perhaps a top-level sort and/or LIMIT.
664 result_plan = plan_set_operations(root, tuple_fraction,
668 * Calculate pathkeys representing the sort order (if any) of the set
669 * operation's result. We have to do this before overwriting the sort
672 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
673 result_plan->targetlist);
674 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
677 * We should not need to call preprocess_targetlist, since we must be
678 * in a SELECT query node. Instead, use the targetlist returned by
679 * plan_set_operations (since this tells whether it returned any
680 * resjunk columns!), and transfer any sort key information from the
683 Assert(parse->commandType == CMD_SELECT);
685 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
688 * Can't handle FOR UPDATE/SHARE here (parser should have checked
689 * already, but let's make sure).
693 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
694 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
697 * Calculate pathkeys that represent result ordering requirements
699 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
701 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
705 /* No set operations, do regular planning */
707 List *group_pathkeys;
708 AttrNumber *groupColIdx = NULL;
709 bool need_tlist_eval = true;
715 AggClauseCounts agg_counts;
716 int numGroupCols = list_length(parse->groupClause);
717 bool use_hashed_grouping = false;
719 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
721 /* Preprocess targetlist */
722 tlist = preprocess_targetlist(root, tlist);
725 * Generate appropriate target list for subplan; may be different from
726 * tlist if grouping or aggregation is needed.
728 sub_tlist = make_subplanTargetList(root, tlist,
729 &groupColIdx, &need_tlist_eval);
732 * Calculate pathkeys that represent grouping/ordering requirements.
733 * Stash them in PlannerInfo so that query_planner can canonicalize
736 root->group_pathkeys =
737 make_pathkeys_for_sortclauses(parse->groupClause, tlist);
738 root->sort_pathkeys =
739 make_pathkeys_for_sortclauses(parse->sortClause, tlist);
742 * Will need actual number of aggregates for estimating costs.
744 * Note: we do not attempt to detect duplicate aggregates here; a
745 * somewhat-overestimated count is okay for our present purposes.
747 * Note: think not that we can turn off hasAggs if we find no aggs. It
748 * is possible for constant-expression simplification to remove all
749 * explicit references to aggs, but we still have to follow the
750 * aggregate semantics (eg, producing only one output row).
754 count_agg_clauses((Node *) tlist, &agg_counts);
755 count_agg_clauses(parse->havingQual, &agg_counts);
759 * Figure out whether we need a sorted result from query_planner.
761 * If we have a GROUP BY clause, then we want a result sorted properly
762 * for grouping. Otherwise, if there is an ORDER BY clause, we want
763 * to sort by the ORDER BY clause. (Note: if we have both, and ORDER
764 * BY is a superset of GROUP BY, it would be tempting to request sort
765 * by ORDER BY --- but that might just leave us failing to exploit an
766 * available sort order at all. Needs more thought...)
768 if (parse->groupClause)
769 root->query_pathkeys = root->group_pathkeys;
770 else if (parse->sortClause)
771 root->query_pathkeys = root->sort_pathkeys;
773 root->query_pathkeys = NIL;
776 * Generate the best unsorted and presorted paths for this Query (but
777 * note there may not be any presorted path). query_planner will also
778 * estimate the number of groups in the query, and canonicalize all
781 query_planner(root, sub_tlist, tuple_fraction,
782 &cheapest_path, &sorted_path, &dNumGroups);
784 group_pathkeys = root->group_pathkeys;
785 sort_pathkeys = root->sort_pathkeys;
788 * If grouping, decide whether we want to use hashed grouping.
790 if (parse->groupClause)
792 use_hashed_grouping =
793 choose_hashed_grouping(root, tuple_fraction,
794 cheapest_path, sorted_path,
795 dNumGroups, &agg_counts);
797 /* Also convert # groups to long int --- but 'ware overflow! */
798 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
802 * Select the best path. If we are doing hashed grouping, we will
803 * always read all the input tuples, so use the cheapest-total path.
804 * Otherwise, trust query_planner's decision about which to use.
806 if (use_hashed_grouping || !sorted_path)
807 best_path = cheapest_path;
809 best_path = sorted_path;
812 * Check to see if it's possible to optimize MIN/MAX aggregates. If
813 * so, we will forget all the work we did so far to choose a "regular"
814 * path ... but we had to do it anyway to be able to tell which way is
817 result_plan = optimize_minmax_aggregates(root,
820 if (result_plan != NULL)
823 * optimize_minmax_aggregates generated the full plan, with the
824 * right tlist, and it has no sort order.
826 current_pathkeys = NIL;
831 * Normal case --- create a plan according to query_planner's
834 result_plan = create_plan(root, best_path);
835 current_pathkeys = best_path->pathkeys;
838 * create_plan() returns a plan with just a "flat" tlist of
839 * required Vars. Usually we need to insert the sub_tlist as the
840 * tlist of the top plan node. However, we can skip that if we
841 * determined that whatever query_planner chose to return will be
847 * If the top-level plan node is one that cannot do expression
848 * evaluation, we must insert a Result node to project the
851 if (!is_projection_capable_plan(result_plan))
853 result_plan = (Plan *) make_result(sub_tlist, NULL,
859 * Otherwise, just replace the subplan's flat tlist with
862 result_plan->targetlist = sub_tlist;
866 * Also, account for the cost of evaluation of the sub_tlist.
868 * Up to now, we have only been dealing with "flat" tlists,
869 * containing just Vars. So their evaluation cost is zero
870 * according to the model used by cost_qual_eval() (or if you
871 * prefer, the cost is factored into cpu_tuple_cost). Thus we
872 * can avoid accounting for tlist cost throughout
873 * query_planner() and subroutines. But now we've inserted a
874 * tlist that might contain actual operators, sub-selects, etc
875 * --- so we'd better account for its cost.
877 * Below this point, any tlist eval cost for added-on nodes
878 * should be accounted for as we create those nodes.
879 * Presently, of the node types we can add on, only Agg and
880 * Group project new tlists (the rest just copy their input
881 * tuples) --- so make_agg() and make_group() are responsible
882 * for computing the added cost.
884 cost_qual_eval(&tlist_cost, sub_tlist);
885 result_plan->startup_cost += tlist_cost.startup;
886 result_plan->total_cost += tlist_cost.startup +
887 tlist_cost.per_tuple * result_plan->plan_rows;
892 * Since we're using query_planner's tlist and not the one
893 * make_subplanTargetList calculated, we have to refigure any
894 * grouping-column indexes make_subplanTargetList computed.
896 locate_grouping_columns(root, tlist, result_plan->targetlist,
901 * Insert AGG or GROUP node if needed, plus an explicit sort step
904 * HAVING clause, if any, becomes qual of the Agg or Group node.
906 if (use_hashed_grouping)
908 /* Hashed aggregate plan --- no sort needed */
909 result_plan = (Plan *) make_agg(root,
911 (List *) parse->havingQual,
918 /* Hashed aggregation produces randomly-ordered results */
919 current_pathkeys = NIL;
921 else if (parse->hasAggs)
923 /* Plain aggregate plan --- sort if needed */
924 AggStrategy aggstrategy;
926 if (parse->groupClause)
928 if (!pathkeys_contained_in(group_pathkeys,
931 result_plan = (Plan *)
932 make_sort_from_groupcols(root,
936 current_pathkeys = group_pathkeys;
938 aggstrategy = AGG_SORTED;
941 * The AGG node will not change the sort ordering of its
942 * groups, so current_pathkeys describes the result too.
947 aggstrategy = AGG_PLAIN;
948 /* Result will be only one row anyway; no sort order */
949 current_pathkeys = NIL;
952 result_plan = (Plan *) make_agg(root,
954 (List *) parse->havingQual,
962 else if (parse->groupClause)
965 * GROUP BY without aggregation, so insert a group node (plus
966 * the appropriate sort node, if necessary).
968 * Add an explicit sort if we couldn't make the path come out
969 * the way the GROUP node needs it.
971 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
973 result_plan = (Plan *)
974 make_sort_from_groupcols(root,
978 current_pathkeys = group_pathkeys;
981 result_plan = (Plan *) make_group(root,
983 (List *) parse->havingQual,
988 /* The Group node won't change sort ordering */
990 else if (root->hasHavingQual)
993 * No aggregates, and no GROUP BY, but we have a HAVING qual.
994 * This is a degenerate case in which we are supposed to emit
995 * either 0 or 1 row depending on whether HAVING succeeds.
996 * Furthermore, there cannot be any variables in either HAVING
997 * or the targetlist, so we actually do not need the FROM
998 * table at all! We can just throw away the plan-so-far and
999 * generate a Result node. This is a sufficiently unusual
1000 * corner case that it's not worth contorting the structure of
1001 * this routine to avoid having to generate the plan in the
1004 result_plan = (Plan *) make_result(tlist,
1008 } /* end of non-minmax-aggregate case */
1009 } /* end of if (setOperations) */
1012 * If we were not able to make the plan come out in the right order, add
1013 * an explicit sort step.
1015 if (parse->sortClause)
1017 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1019 result_plan = (Plan *)
1020 make_sort_from_sortclauses(root,
1023 current_pathkeys = sort_pathkeys;
1028 * If there is a DISTINCT clause, add the UNIQUE node.
1030 if (parse->distinctClause)
1032 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1035 * If there was grouping or aggregation, leave plan_rows as-is (ie,
1036 * assume the result was already mostly unique). If not, use the
1037 * number of distinct-groups calculated by query_planner.
1039 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1040 result_plan->plan_rows = dNumGroups;
1044 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1046 if (parse->limitCount || parse->limitOffset)
1048 result_plan = (Plan *) make_limit(result_plan,
1056 * Return the actual output ordering in query_pathkeys for possible use by
1057 * an outer query level.
1059 root->query_pathkeys = current_pathkeys;
1065 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1067 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1068 * results back in *count_est and *offset_est. These variables are set to
1069 * 0 if the corresponding clause is not present, and -1 if it's present
1070 * but we couldn't estimate the value for it. (The "0" convention is OK
1071 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1072 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1073 * usual practice of never estimating less than one row.) These values will
1074 * be passed to make_limit, which see if you change this code.
1076 * The return value is the suitably adjusted tuple_fraction to use for
1077 * planning the query. This adjustment is not overridable, since it reflects
1078 * plan actions that grouping_planner() will certainly take, not assumptions
1082 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1083 int64 *offset_est, int64 *count_est)
1085 Query *parse = root->parse;
1087 double limit_fraction;
1089 /* Should not be called unless LIMIT or OFFSET */
1090 Assert(parse->limitCount || parse->limitOffset);
1093 * Try to obtain the clause values. We use estimate_expression_value
1094 * primarily because it can sometimes do something useful with Params.
1096 if (parse->limitCount)
1098 est = estimate_expression_value(parse->limitCount);
1099 if (est && IsA(est, Const))
1101 if (((Const *) est)->constisnull)
1103 /* NULL indicates LIMIT ALL, ie, no limit */
1104 *count_est = 0; /* treat as not present */
1108 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1109 if (*count_est <= 0)
1110 *count_est = 1; /* force to at least 1 */
1114 *count_est = -1; /* can't estimate */
1117 *count_est = 0; /* not present */
1119 if (parse->limitOffset)
1121 est = estimate_expression_value(parse->limitOffset);
1122 if (est && IsA(est, Const))
1124 if (((Const *) est)->constisnull)
1126 /* Treat NULL as no offset; the executor will too */
1127 *offset_est = 0; /* treat as not present */
1131 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1133 if (*offset_est < 0)
1134 *offset_est = 0; /* less than 0 is same as 0 */
1138 *offset_est = -1; /* can't estimate */
1141 *offset_est = 0; /* not present */
1143 if (*count_est != 0)
1146 * A LIMIT clause limits the absolute number of tuples returned.
1147 * However, if it's not a constant LIMIT then we have to guess; for
1148 * lack of a better idea, assume 10% of the plan's result is wanted.
1150 if (*count_est < 0 || *offset_est < 0)
1152 /* LIMIT or OFFSET is an expression ... punt ... */
1153 limit_fraction = 0.10;
1157 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1158 limit_fraction = (double) *count_est + (double) *offset_est;
1162 * If we have absolute limits from both caller and LIMIT, use the
1163 * smaller value; likewise if they are both fractional. If one is
1164 * fractional and the other absolute, we can't easily determine which
1165 * is smaller, but we use the heuristic that the absolute will usually
1168 if (tuple_fraction >= 1.0)
1170 if (limit_fraction >= 1.0)
1173 tuple_fraction = Min(tuple_fraction, limit_fraction);
1177 /* caller absolute, limit fractional; use caller's value */
1180 else if (tuple_fraction > 0.0)
1182 if (limit_fraction >= 1.0)
1184 /* caller fractional, limit absolute; use limit */
1185 tuple_fraction = limit_fraction;
1189 /* both fractional */
1190 tuple_fraction = Min(tuple_fraction, limit_fraction);
1195 /* no info from caller, just use limit */
1196 tuple_fraction = limit_fraction;
1199 else if (*offset_est != 0 && tuple_fraction > 0.0)
1202 * We have an OFFSET but no LIMIT. This acts entirely differently
1203 * from the LIMIT case: here, we need to increase rather than decrease
1204 * the caller's tuple_fraction, because the OFFSET acts to cause more
1205 * tuples to be fetched instead of fewer. This only matters if we got
1206 * a tuple_fraction > 0, however.
1208 * As above, use 10% if OFFSET is present but unestimatable.
1210 if (*offset_est < 0)
1211 limit_fraction = 0.10;
1213 limit_fraction = (double) *offset_est;
1216 * If we have absolute counts from both caller and OFFSET, add them
1217 * together; likewise if they are both fractional. If one is
1218 * fractional and the other absolute, we want to take the larger, and
1219 * we heuristically assume that's the fractional one.
1221 if (tuple_fraction >= 1.0)
1223 if (limit_fraction >= 1.0)
1225 /* both absolute, so add them together */
1226 tuple_fraction += limit_fraction;
1230 /* caller absolute, limit fractional; use limit */
1231 tuple_fraction = limit_fraction;
1236 if (limit_fraction >= 1.0)
1238 /* caller fractional, limit absolute; use caller's value */
1242 /* both fractional, so add them together */
1243 tuple_fraction += limit_fraction;
1244 if (tuple_fraction >= 1.0)
1245 tuple_fraction = 0.0; /* assume fetch all */
1250 return tuple_fraction;
1254 * choose_hashed_grouping - should we use hashed grouping?
1257 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1258 Path *cheapest_path, Path *sorted_path,
1259 double dNumGroups, AggClauseCounts *agg_counts)
1261 int numGroupCols = list_length(root->parse->groupClause);
1262 double cheapest_path_rows;
1263 int cheapest_path_width;
1265 List *current_pathkeys;
1270 * Check can't-do-it conditions, including whether the grouping operators
1273 * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1274 * (Doing so would imply storing *all* the input values in the hash table,
1275 * which seems like a certain loser.)
1277 if (!enable_hashagg)
1279 if (agg_counts->numDistinctAggs != 0)
1281 if (!hash_safe_grouping(root))
1285 * Don't do it if it doesn't look like the hashtable will fit into
1288 * Beware here of the possibility that cheapest_path->parent is NULL. This
1289 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1291 if (cheapest_path->parent)
1293 cheapest_path_rows = cheapest_path->parent->rows;
1294 cheapest_path_width = cheapest_path->parent->width;
1298 cheapest_path_rows = 1; /* assume non-set result */
1299 cheapest_path_width = 100; /* arbitrary */
1302 /* Estimate per-hash-entry space at tuple width... */
1303 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1304 /* plus space for pass-by-ref transition values... */
1305 hashentrysize += agg_counts->transitionSpace;
1306 /* plus the per-hash-entry overhead */
1307 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1309 if (hashentrysize * dNumGroups > work_mem * 1024L)
1313 * See if the estimated cost is no more than doing it the other way. While
1314 * avoiding the need for sorted input is usually a win, the fact that the
1315 * output won't be sorted may be a loss; so we need to do an actual cost
1318 * We need to consider cheapest_path + hashagg [+ final sort] versus
1319 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1320 * presorted_path + group or agg [+ final sort] where brackets indicate a
1321 * step that may not be needed. We assume query_planner() will have
1322 * returned a presorted path only if it's a winner compared to
1323 * cheapest_path for this purpose.
1325 * These path variables are dummies that just hold cost fields; we don't
1326 * make actual Paths for these steps.
1328 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1329 numGroupCols, dNumGroups,
1330 cheapest_path->startup_cost, cheapest_path->total_cost,
1331 cheapest_path_rows);
1332 /* Result of hashed agg is always unsorted */
1333 if (root->sort_pathkeys)
1334 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1335 dNumGroups, cheapest_path_width);
1339 sorted_p.startup_cost = sorted_path->startup_cost;
1340 sorted_p.total_cost = sorted_path->total_cost;
1341 current_pathkeys = sorted_path->pathkeys;
1345 sorted_p.startup_cost = cheapest_path->startup_cost;
1346 sorted_p.total_cost = cheapest_path->total_cost;
1347 current_pathkeys = cheapest_path->pathkeys;
1349 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1351 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1352 cheapest_path_rows, cheapest_path_width);
1353 current_pathkeys = root->group_pathkeys;
1356 if (root->parse->hasAggs)
1357 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1358 numGroupCols, dNumGroups,
1359 sorted_p.startup_cost, sorted_p.total_cost,
1360 cheapest_path_rows);
1362 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1363 sorted_p.startup_cost, sorted_p.total_cost,
1364 cheapest_path_rows);
1365 /* The Agg or Group node will preserve ordering */
1366 if (root->sort_pathkeys &&
1367 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1368 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1369 dNumGroups, cheapest_path_width);
1372 * Now make the decision using the top-level tuple fraction. First we
1373 * have to convert an absolute count (LIMIT) into fractional form.
1375 if (tuple_fraction >= 1.0)
1376 tuple_fraction /= dNumGroups;
1378 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1379 tuple_fraction) < 0)
1381 /* Hashed is cheaper, so use it */
1388 * hash_safe_grouping - are grouping operators hashable?
1390 * We assume hashed aggregation will work if the datatype's equality operator
1391 * is marked hashjoinable.
1394 hash_safe_grouping(PlannerInfo *root)
1398 foreach(gl, root->parse->groupClause)
1400 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1401 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1402 root->parse->targetList);
1406 optup = equality_oper(exprType((Node *) tle->expr), true);
1409 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1410 ReleaseSysCache(optup);
1418 * make_subplanTargetList
1419 * Generate appropriate target list when grouping is required.
1421 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1422 * above the result of query_planner, we typically want to pass a different
1423 * target list to query_planner than the outer plan nodes should have.
1424 * This routine generates the correct target list for the subplan.
1426 * The initial target list passed from the parser already contains entries
1427 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1428 * for variables used only in HAVING clauses; so we need to add those
1429 * variables to the subplan target list. Also, we flatten all expressions
1430 * except GROUP BY items into their component variables; the other expressions
1431 * will be computed by the inserted nodes rather than by the subplan.
1432 * For example, given a query like
1433 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1434 * we want to pass this targetlist to the subplan:
1436 * where the a+b target will be used by the Sort/Group steps, and the
1437 * other targets will be used for computing the final results. (In the
1438 * above example we could theoretically suppress the a and b targets and
1439 * pass down only c,d,a+b, but it's not really worth the trouble to
1440 * eliminate simple var references from the subplan. We will avoid doing
1441 * the extra computation to recompute a+b at the outer level; see
1442 * replace_vars_with_subplan_refs() in setrefs.c.)
1444 * If we are grouping or aggregating, *and* there are no non-Var grouping
1445 * expressions, then the returned tlist is effectively dummy; we do not
1446 * need to force it to be evaluated, because all the Vars it contains
1447 * should be present in the output of query_planner anyway.
1449 * 'tlist' is the query's target list.
1450 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1451 * expressions (if there are any) in the subplan's target list.
1452 * 'need_tlist_eval' is set true if we really need to evaluate the
1455 * The result is the targetlist to be passed to the subplan.
1459 make_subplanTargetList(PlannerInfo *root,
1461 AttrNumber **groupColIdx,
1462 bool *need_tlist_eval)
1464 Query *parse = root->parse;
1469 *groupColIdx = NULL;
1472 * If we're not grouping or aggregating, there's nothing to do here;
1473 * query_planner should receive the unmodified target list.
1475 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1477 *need_tlist_eval = true;
1482 * Otherwise, start with a "flattened" tlist (having just the vars
1483 * mentioned in the targetlist and HAVING qual --- but not upper- level
1484 * Vars; they will be replaced by Params later on).
1486 sub_tlist = flatten_tlist(tlist);
1487 extravars = pull_var_clause(parse->havingQual, false);
1488 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1489 list_free(extravars);
1490 *need_tlist_eval = false; /* only eval if not flat tlist */
1493 * If grouping, create sub_tlist entries for all GROUP BY expressions
1494 * (GROUP BY items that are simple Vars should be in the list already),
1495 * and make an array showing where the group columns are in the sub_tlist.
1497 numCols = list_length(parse->groupClause);
1501 AttrNumber *grpColIdx;
1504 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1505 *groupColIdx = grpColIdx;
1507 foreach(gl, parse->groupClause)
1509 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1510 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1511 TargetEntry *te = NULL;
1514 /* Find or make a matching sub_tlist entry */
1515 foreach(sl, sub_tlist)
1517 te = (TargetEntry *) lfirst(sl);
1518 if (equal(groupexpr, te->expr))
1523 te = makeTargetEntry((Expr *) groupexpr,
1524 list_length(sub_tlist) + 1,
1527 sub_tlist = lappend(sub_tlist, te);
1528 *need_tlist_eval = true; /* it's not flat anymore */
1531 /* and save its resno */
1532 grpColIdx[keyno++] = te->resno;
1540 * locate_grouping_columns
1541 * Locate grouping columns in the tlist chosen by query_planner.
1543 * This is only needed if we don't use the sub_tlist chosen by
1544 * make_subplanTargetList. We have to forget the column indexes found
1545 * by that routine and re-locate the grouping vars in the real sub_tlist.
1548 locate_grouping_columns(PlannerInfo *root,
1551 AttrNumber *groupColIdx)
1557 * No work unless grouping.
1559 if (!root->parse->groupClause)
1561 Assert(groupColIdx == NULL);
1564 Assert(groupColIdx != NULL);
1566 foreach(gl, root->parse->groupClause)
1568 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1569 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1570 TargetEntry *te = NULL;
1573 foreach(sl, sub_tlist)
1575 te = (TargetEntry *) lfirst(sl);
1576 if (equal(groupexpr, te->expr))
1580 elog(ERROR, "failed to locate grouping columns");
1582 groupColIdx[keyno++] = te->resno;
1587 * postprocess_setop_tlist
1588 * Fix up targetlist returned by plan_set_operations().
1590 * We need to transpose sort key info from the orig_tlist into new_tlist.
1591 * NOTE: this would not be good enough if we supported resjunk sort keys
1592 * for results of set operations --- then, we'd need to project a whole
1593 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1594 * find any resjunk columns in orig_tlist.
1597 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1600 ListCell *orig_tlist_item = list_head(orig_tlist);
1602 foreach(l, new_tlist)
1604 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1605 TargetEntry *orig_tle;
1607 /* ignore resjunk columns in setop result */
1608 if (new_tle->resjunk)
1611 Assert(orig_tlist_item != NULL);
1612 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1613 orig_tlist_item = lnext(orig_tlist_item);
1614 if (orig_tle->resjunk) /* should not happen */
1615 elog(ERROR, "resjunk output columns are not implemented");
1616 Assert(new_tle->resno == orig_tle->resno);
1617 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1619 if (orig_tlist_item != NULL)
1620 elog(ERROR, "resjunk output columns are not implemented");