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 1.200 2006/06/28 20:04:38 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
55 #define EXPRKIND_APPINFO 5
58 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
59 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
60 static Plan *inheritance_planner(PlannerInfo *root);
61 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
62 static double preprocess_limit(PlannerInfo *root,
63 double tuple_fraction,
64 int *offset_est, int *count_est);
65 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
66 Path *cheapest_path, Path *sorted_path,
67 double dNumGroups, AggClauseCounts *agg_counts);
68 static bool hash_safe_grouping(PlannerInfo *root);
69 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
70 AttrNumber **groupColIdx, bool *need_tlist_eval);
71 static void locate_grouping_columns(PlannerInfo *root,
74 AttrNumber *groupColIdx);
75 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
78 /*****************************************************************************
80 * Query optimizer entry point
82 *****************************************************************************/
84 planner(Query *parse, bool isCursor, int cursorOptions,
85 ParamListInfo boundParams)
87 double tuple_fraction;
89 Index save_PlannerQueryLevel;
90 List *save_PlannerParamList;
91 ParamListInfo save_PlannerBoundParamList;
94 * The planner can be called recursively (an example is when
95 * eval_const_expressions tries to pre-evaluate an SQL function). So,
96 * these global state variables must be saved and restored.
98 * Query level and the param list cannot be moved into the per-query
99 * PlannerInfo structure since their whole purpose is communication across
100 * multiple sub-queries. Also, boundParams is explicitly info from outside
101 * the query, and so is likewise better handled as a global variable.
103 * Note we do NOT save and restore PlannerPlanId: it exists to assign
104 * unique IDs to SubPlan nodes, and we want those IDs to be unique for the
105 * life of a backend. Also, PlannerInitPlan is saved/restored in
106 * subquery_planner, not here.
108 save_PlannerQueryLevel = PlannerQueryLevel;
109 save_PlannerParamList = PlannerParamList;
110 save_PlannerBoundParamList = PlannerBoundParamList;
112 /* Initialize state for handling outer-level references and params */
113 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
114 PlannerParamList = NIL;
115 PlannerBoundParamList = boundParams;
117 /* Determine what fraction of the plan is likely to be scanned */
121 * We have no real idea how many tuples the user will ultimately FETCH
122 * from a cursor, but it seems a good bet that he doesn't want 'em
123 * all. Optimize for 10% retrieval (you gotta better number? Should
124 * this be a SETtable parameter?)
126 tuple_fraction = 0.10;
130 /* Default assumption is we need all the tuples */
131 tuple_fraction = 0.0;
134 /* primary planning entry point (may recurse for subqueries) */
135 result_plan = subquery_planner(parse, tuple_fraction, NULL);
137 /* check we popped out the right number of levels */
138 Assert(PlannerQueryLevel == 0);
141 * If creating a plan for a scrollable cursor, make sure it can run
142 * backwards on demand. Add a Material node at the top at need.
144 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
146 if (!ExecSupportsBackwardScan(result_plan))
147 result_plan = materialize_finished_plan(result_plan);
150 /* final cleanup of the plan */
151 result_plan = set_plan_references(result_plan, parse->rtable);
153 /* executor wants to know total number of Params used overall */
154 result_plan->nParamExec = list_length(PlannerParamList);
156 /* restore state for outer planner, if any */
157 PlannerQueryLevel = save_PlannerQueryLevel;
158 PlannerParamList = save_PlannerParamList;
159 PlannerBoundParamList = save_PlannerBoundParamList;
165 /*--------------------
167 * Invokes the planner on a subquery. We recurse to here for each
168 * sub-SELECT found in the query tree.
170 * parse is the querytree produced by the parser & rewriter.
171 * tuple_fraction is the fraction of tuples we expect will be retrieved.
172 * tuple_fraction is interpreted as explained for grouping_planner, below.
174 * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
175 * the output sort ordering of the completed plan.
177 * Basically, this routine does the stuff that should only be done once
178 * per Query object. It then calls grouping_planner. At one time,
179 * grouping_planner could be invoked recursively on the same Query object;
180 * that's not currently true, but we keep the separation between the two
181 * routines anyway, in case we need it again someday.
183 * subquery_planner will be called recursively to handle sub-Query nodes
184 * found within the query's expressions and rangetable.
186 * Returns a query plan.
187 *--------------------
190 subquery_planner(Query *parse, double tuple_fraction,
191 List **subquery_pathkeys)
193 List *saved_initplan = PlannerInitPlan;
194 int saved_planid = PlannerPlanId;
200 /* Set up for a new level of subquery */
202 PlannerInitPlan = NIL;
204 /* Create a PlannerInfo data structure for this subquery */
205 root = makeNode(PlannerInfo);
207 root->in_info_list = NIL;
208 root->append_rel_list = NIL;
211 * Look for IN clauses at the top level of WHERE, and transform them into
212 * joins. Note that this step only handles IN clauses originally at top
213 * level of WHERE; if we pull up any subqueries in the next step, their
214 * INs are processed just before pulling them up.
216 if (parse->hasSubLinks)
217 parse->jointree->quals = pull_up_IN_clauses(root,
218 parse->jointree->quals);
221 * Check to see if any subqueries in the rangetable can be merged into
224 parse->jointree = (FromExpr *)
225 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
228 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
229 * avoid the expense of doing flatten_join_alias_vars(). Also check for
230 * outer joins --- if none, we can skip reduce_outer_joins() and some
231 * other processing. This must be done after we have done
232 * pull_up_subqueries, of course.
234 * Note: if reduce_outer_joins manages to eliminate all outer joins,
235 * root->hasOuterJoins is not reset currently. This is OK since its
236 * purpose is merely to suppress unnecessary processing in simple cases.
238 root->hasJoinRTEs = false;
239 root->hasOuterJoins = false;
240 foreach(l, parse->rtable)
242 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
244 if (rte->rtekind == RTE_JOIN)
246 root->hasJoinRTEs = true;
247 if (IS_OUTER_JOIN(rte->jointype))
249 root->hasOuterJoins = true;
250 /* Can quit scanning once we find an outer join */
257 * Expand any rangetable entries that are inheritance sets into "append
258 * relations". This can add entries to the rangetable, but they must be
259 * plain base relations not joins, so it's OK (and marginally more
260 * efficient) to do it after checking for join RTEs. We must do it after
261 * pulling up subqueries, else we'd fail to handle inherited tables in
264 expand_inherited_tables(root);
267 * Set hasHavingQual to remember if HAVING clause is present. Needed
268 * because preprocess_expression will reduce a constant-true condition to
269 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
271 root->hasHavingQual = (parse->havingQual != NULL);
274 * Do expression preprocessing on targetlist and quals.
276 parse->targetList = (List *)
277 preprocess_expression(root, (Node *) parse->targetList,
280 preprocess_qual_conditions(root, (Node *) parse->jointree);
282 parse->havingQual = preprocess_expression(root, parse->havingQual,
285 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
287 parse->limitCount = preprocess_expression(root, parse->limitCount,
290 root->in_info_list = (List *)
291 preprocess_expression(root, (Node *) root->in_info_list,
293 root->append_rel_list = (List *)
294 preprocess_expression(root, (Node *) root->append_rel_list,
297 /* Also need to preprocess expressions for function RTEs */
298 foreach(l, parse->rtable)
300 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
302 if (rte->rtekind == RTE_FUNCTION)
303 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
308 * In some cases we may want to transfer a HAVING clause into WHERE. We
309 * cannot do so if the HAVING clause contains aggregates (obviously) or
310 * volatile functions (since a HAVING clause is supposed to be executed
311 * only once per group). Also, it may be that the clause is so expensive
312 * to execute that we're better off doing it only once per group, despite
313 * the loss of selectivity. This is hard to estimate short of doing the
314 * entire planning process twice, so we use a heuristic: clauses
315 * containing subplans are left in HAVING. Otherwise, we move or copy the
316 * HAVING clause into WHERE, in hopes of eliminating tuples before
317 * aggregation instead of after.
319 * If the query has explicit grouping then we can simply move such a
320 * clause into WHERE; any group that fails the clause will not be in the
321 * output because none of its tuples will reach the grouping or
322 * aggregation stage. Otherwise we must have a degenerate (variable-free)
323 * HAVING clause, which we put in WHERE so that query_planner() can use it
324 * in a gating Result node, but also keep in HAVING to ensure that we
325 * don't emit a bogus aggregated row. (This could be done better, but it
326 * seems not worth optimizing.)
328 * Note that both havingQual and parse->jointree->quals are in
329 * implicitly-ANDed-list form at this point, even though they are declared
333 foreach(l, (List *) parse->havingQual)
335 Node *havingclause = (Node *) lfirst(l);
337 if (contain_agg_clause(havingclause) ||
338 contain_volatile_functions(havingclause) ||
339 contain_subplans(havingclause))
341 /* keep it in HAVING */
342 newHaving = lappend(newHaving, havingclause);
344 else if (parse->groupClause)
346 /* move it to WHERE */
347 parse->jointree->quals = (Node *)
348 lappend((List *) parse->jointree->quals, havingclause);
352 /* put a copy in WHERE, keep it in HAVING */
353 parse->jointree->quals = (Node *)
354 lappend((List *) parse->jointree->quals,
355 copyObject(havingclause));
356 newHaving = lappend(newHaving, havingclause);
359 parse->havingQual = (Node *) newHaving;
362 * If we have any outer joins, try to reduce them to plain inner joins.
363 * This step is most easily done after we've done expression
366 if (root->hasOuterJoins)
367 reduce_outer_joins(root);
370 * Do the main planning. If we have an inherited target relation, that
371 * needs special processing, else go straight to grouping_planner.
373 if (parse->resultRelation &&
374 rt_fetch(parse->resultRelation, parse->rtable)->inh)
375 plan = inheritance_planner(root);
377 plan = grouping_planner(root, tuple_fraction);
380 * If any subplans were generated, or if we're inside a subplan, build
381 * initPlan list and extParam/allParam sets for plan nodes, and attach the
382 * initPlans to the top plan node.
384 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
385 SS_finalize_plan(plan, parse->rtable);
387 /* Return sort ordering info if caller wants it */
388 if (subquery_pathkeys)
389 *subquery_pathkeys = root->query_pathkeys;
391 /* Return to outer subquery context */
393 PlannerInitPlan = saved_initplan;
394 /* we do NOT restore PlannerPlanId; that's not an oversight! */
400 * preprocess_expression
401 * Do subquery_planner's preprocessing work for an expression,
402 * which can be a targetlist, a WHERE clause (including JOIN/ON
403 * conditions), or a HAVING clause.
406 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
409 * Fall out quickly if expression is empty. This occurs often enough to
410 * be worth checking. Note that null->null is the correct conversion for
411 * implicit-AND result format, too.
417 * If the query has any join RTEs, replace join alias variables with
418 * base-relation variables. We must do this before sublink processing,
419 * else sublinks expanded out from join aliases wouldn't get processed.
421 if (root->hasJoinRTEs)
422 expr = flatten_join_alias_vars(root, 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 * Because this is a relatively expensive process, we skip it when the
433 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
434 * expression will only be evaluated once anyway, so no point in
435 * pre-simplifying; we can't execute it any faster than the executor can,
436 * and we will waste cycles copying the tree. Notice however that we
437 * still must do it for quals (to get AND/OR flatness); and if we are in a
438 * subquery we should not assume it will be done only once.
440 if (root->parse->jointree->fromlist != NIL ||
441 kind == EXPRKIND_QUAL ||
442 PlannerQueryLevel > 1)
443 expr = eval_const_expressions(expr);
446 * If it's a qual or havingQual, canonicalize it.
448 if (kind == EXPRKIND_QUAL)
450 expr = (Node *) canonicalize_qual((Expr *) expr);
452 #ifdef OPTIMIZER_DEBUG
453 printf("After canonicalize_qual()\n");
458 /* Expand SubLinks to SubPlans */
459 if (root->parse->hasSubLinks)
460 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
463 * XXX do not insert anything here unless you have grokked the comments in
464 * SS_replace_correlation_vars ...
467 /* Replace uplevel vars with Param nodes */
468 if (PlannerQueryLevel > 1)
469 expr = SS_replace_correlation_vars(expr);
472 * If it's a qual or havingQual, convert it to implicit-AND format. (We
473 * don't want to do this before eval_const_expressions, since the latter
474 * would be unable to simplify a top-level AND correctly. Also,
475 * SS_process_sublinks expects explicit-AND format.)
477 if (kind == EXPRKIND_QUAL)
478 expr = (Node *) make_ands_implicit((Expr *) expr);
484 * preprocess_qual_conditions
485 * Recursively scan the query's jointree and do subquery_planner's
486 * preprocessing work on each qual condition found therein.
489 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
493 if (IsA(jtnode, RangeTblRef))
495 /* nothing to do here */
497 else if (IsA(jtnode, FromExpr))
499 FromExpr *f = (FromExpr *) jtnode;
502 foreach(l, f->fromlist)
503 preprocess_qual_conditions(root, lfirst(l));
505 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
507 else if (IsA(jtnode, JoinExpr))
509 JoinExpr *j = (JoinExpr *) jtnode;
511 preprocess_qual_conditions(root, j->larg);
512 preprocess_qual_conditions(root, j->rarg);
514 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
517 elog(ERROR, "unrecognized node type: %d",
518 (int) nodeTag(jtnode));
522 * inheritance_planner
523 * Generate a plan in the case where the result relation is an
526 * We have to handle this case differently from cases where a source relation
527 * is an inheritance set. Source inheritance is expanded at the bottom of the
528 * plan tree (see allpaths.c), but target inheritance has to be expanded at
529 * the top. The reason is that for UPDATE, each target relation needs a
530 * different targetlist matching its own column set. Also, for both UPDATE
531 * and DELETE, the executor needs the Append plan node at the top, else it
532 * can't keep track of which table is the current target table. Fortunately,
533 * the UPDATE/DELETE target can never be the nullable side of an outer join,
534 * so it's OK to generate the plan this way.
536 * Returns a query plan.
539 inheritance_planner(PlannerInfo *root)
541 Query *parse = root->parse;
542 int parentRTindex = parse->resultRelation;
543 List *subplans = NIL;
548 subroot.parse = NULL; /* catch it if no matches in loop */
550 parse->resultRelations = NIL;
552 foreach(l, root->append_rel_list)
554 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
557 /* append_rel_list contains all append rels; ignore others */
558 if (appinfo->parent_relid != parentRTindex)
561 /* Build target-relations list for the executor */
562 parse->resultRelations = lappend_int(parse->resultRelations,
563 appinfo->child_relid);
566 * Generate modified query with this rel as target. We have to be
567 * prepared to translate varnos in in_info_list as well as in the
570 memcpy(&subroot, root, sizeof(PlannerInfo));
571 subroot.parse = (Query *)
572 adjust_appendrel_attrs((Node *) parse,
574 subroot.in_info_list = (List *)
575 adjust_appendrel_attrs((Node *) root->in_info_list,
577 /* There shouldn't be any OJ info to translate, as yet */
578 Assert(subroot.oj_info_list == NIL);
581 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
583 subplans = lappend(subplans, subplan);
585 /* Save preprocessed tlist from first rel for use in Append */
587 tlist = subplan->targetlist;
591 * Planning might have modified the rangetable, due to changes of the
592 * Query structures inside subquery RTEs. We have to ensure that this
593 * gets propagated back to the master copy. But can't do this until we
594 * are done planning, because all the calls to grouping_planner need
595 * virgin sub-Queries to work from. (We are effectively assuming that
596 * sub-Queries will get planned identically each time, or at least that
597 * the impacts on their rangetables will be the same each time.)
599 * XXX should clean this up someday
601 parse->rtable = subroot.parse->rtable;
603 /* Mark result as unordered (probably unnecessary) */
604 root->query_pathkeys = NIL;
606 return (Plan *) make_append(subplans, true, tlist);
609 /*--------------------
611 * Perform planning steps related to grouping, aggregation, etc.
612 * This primarily means adding top-level processing to the basic
613 * query plan produced by query_planner.
615 * tuple_fraction is the fraction of tuples we expect will be retrieved
617 * tuple_fraction is interpreted as follows:
618 * 0: expect all tuples to be retrieved (normal case)
619 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
620 * from the plan to be retrieved
621 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
622 * expected to be retrieved (ie, a LIMIT specification)
624 * Returns a query plan. Also, root->query_pathkeys is returned as the
625 * actual output ordering of the plan (in pathkey format).
626 *--------------------
629 grouping_planner(PlannerInfo *root, double tuple_fraction)
631 Query *parse = root->parse;
632 List *tlist = parse->targetList;
636 List *current_pathkeys;
638 double dNumGroups = 0;
640 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
641 if (parse->limitCount || parse->limitOffset)
642 tuple_fraction = preprocess_limit(root, tuple_fraction,
643 &offset_est, &count_est);
645 if (parse->setOperations)
647 List *set_sortclauses;
650 * If there's a top-level ORDER BY, assume we have to fetch all the
651 * tuples. This might seem too simplistic given all the hackery below
652 * to possibly avoid the sort ... but a nonzero tuple_fraction is only
653 * of use to plan_set_operations() when the setop is UNION ALL, and
654 * the result of UNION ALL is always unsorted.
656 if (parse->sortClause)
657 tuple_fraction = 0.0;
660 * Construct the plan for set operations. The result will not need
661 * any work except perhaps a top-level sort and/or LIMIT.
663 result_plan = plan_set_operations(root, tuple_fraction,
667 * Calculate pathkeys representing the sort order (if any) of the set
668 * operation's result. We have to do this before overwriting the sort
671 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
672 result_plan->targetlist);
673 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
676 * We should not need to call preprocess_targetlist, since we must be
677 * in a SELECT query node. Instead, use the targetlist returned by
678 * plan_set_operations (since this tells whether it returned any
679 * resjunk columns!), and transfer any sort key information from the
682 Assert(parse->commandType == CMD_SELECT);
684 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
687 * Can't handle FOR UPDATE/SHARE here (parser should have checked
688 * already, but let's make sure).
692 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
693 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
696 * Calculate pathkeys that represent result ordering requirements
698 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
700 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
704 /* No set operations, do regular planning */
706 List *group_pathkeys;
707 AttrNumber *groupColIdx = NULL;
708 bool need_tlist_eval = true;
714 AggClauseCounts agg_counts;
715 int numGroupCols = list_length(parse->groupClause);
716 bool use_hashed_grouping = false;
718 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
720 /* Preprocess targetlist */
721 tlist = preprocess_targetlist(root, tlist);
724 * Generate appropriate target list for subplan; may be different from
725 * tlist if grouping or aggregation is needed.
727 sub_tlist = make_subplanTargetList(root, tlist,
728 &groupColIdx, &need_tlist_eval);
731 * Calculate pathkeys that represent grouping/ordering requirements.
732 * Stash them in PlannerInfo so that query_planner can canonicalize
735 root->group_pathkeys =
736 make_pathkeys_for_sortclauses(parse->groupClause, tlist);
737 root->sort_pathkeys =
738 make_pathkeys_for_sortclauses(parse->sortClause, tlist);
741 * Will need actual number of aggregates for estimating costs.
743 * Note: we do not attempt to detect duplicate aggregates here; a
744 * somewhat-overestimated count is okay for our present purposes.
746 * Note: think not that we can turn off hasAggs if we find no aggs. It
747 * is possible for constant-expression simplification to remove all
748 * explicit references to aggs, but we still have to follow the
749 * aggregate semantics (eg, producing only one output row).
753 count_agg_clauses((Node *) tlist, &agg_counts);
754 count_agg_clauses(parse->havingQual, &agg_counts);
758 * Figure out whether we need a sorted result from query_planner.
760 * If we have a GROUP BY clause, then we want a result sorted properly
761 * for grouping. Otherwise, if there is an ORDER BY clause, we want
762 * to sort by the ORDER BY clause. (Note: if we have both, and ORDER
763 * BY is a superset of GROUP BY, it would be tempting to request sort
764 * by ORDER BY --- but that might just leave us failing to exploit an
765 * available sort order at all. Needs more thought...)
767 if (parse->groupClause)
768 root->query_pathkeys = root->group_pathkeys;
769 else if (parse->sortClause)
770 root->query_pathkeys = root->sort_pathkeys;
772 root->query_pathkeys = NIL;
775 * Generate the best unsorted and presorted paths for this Query (but
776 * note there may not be any presorted path). query_planner will also
777 * estimate the number of groups in the query, and canonicalize all
780 query_planner(root, sub_tlist, tuple_fraction,
781 &cheapest_path, &sorted_path, &dNumGroups);
783 group_pathkeys = root->group_pathkeys;
784 sort_pathkeys = root->sort_pathkeys;
787 * If grouping, decide whether we want to use hashed grouping.
789 if (parse->groupClause)
791 use_hashed_grouping =
792 choose_hashed_grouping(root, tuple_fraction,
793 cheapest_path, sorted_path,
794 dNumGroups, &agg_counts);
796 /* Also convert # groups to long int --- but 'ware overflow! */
797 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
801 * Select the best path. If we are doing hashed grouping, we will
802 * always read all the input tuples, so use the cheapest-total path.
803 * Otherwise, trust query_planner's decision about which to use.
805 if (use_hashed_grouping || !sorted_path)
806 best_path = cheapest_path;
808 best_path = sorted_path;
811 * Check to see if it's possible to optimize MIN/MAX aggregates. If
812 * so, we will forget all the work we did so far to choose a "regular"
813 * path ... but we had to do it anyway to be able to tell which way is
816 result_plan = optimize_minmax_aggregates(root,
819 if (result_plan != NULL)
822 * optimize_minmax_aggregates generated the full plan, with the
823 * right tlist, and it has no sort order.
825 current_pathkeys = NIL;
830 * Normal case --- create a plan according to query_planner's
833 result_plan = create_plan(root, best_path);
834 current_pathkeys = best_path->pathkeys;
837 * create_plan() returns a plan with just a "flat" tlist of
838 * required Vars. Usually we need to insert the sub_tlist as the
839 * tlist of the top plan node. However, we can skip that if we
840 * determined that whatever query_planner chose to return will be
846 * If the top-level plan node is one that cannot do expression
847 * evaluation, we must insert a Result node to project the
850 if (!is_projection_capable_plan(result_plan))
852 result_plan = (Plan *) make_result(sub_tlist, NULL,
858 * Otherwise, just replace the subplan's flat tlist with
861 result_plan->targetlist = sub_tlist;
865 * Also, account for the cost of evaluation of the sub_tlist.
867 * Up to now, we have only been dealing with "flat" tlists,
868 * containing just Vars. So their evaluation cost is zero
869 * according to the model used by cost_qual_eval() (or if you
870 * prefer, the cost is factored into cpu_tuple_cost). Thus we
871 * can avoid accounting for tlist cost throughout
872 * query_planner() and subroutines. But now we've inserted a
873 * tlist that might contain actual operators, sub-selects, etc
874 * --- so we'd better account for its cost.
876 * Below this point, any tlist eval cost for added-on nodes
877 * should be accounted for as we create those nodes.
878 * Presently, of the node types we can add on, only Agg and
879 * Group project new tlists (the rest just copy their input
880 * tuples) --- so make_agg() and make_group() are responsible
881 * for computing the added cost.
883 cost_qual_eval(&tlist_cost, sub_tlist);
884 result_plan->startup_cost += tlist_cost.startup;
885 result_plan->total_cost += tlist_cost.startup +
886 tlist_cost.per_tuple * result_plan->plan_rows;
891 * Since we're using query_planner's tlist and not the one
892 * make_subplanTargetList calculated, we have to refigure any
893 * grouping-column indexes make_subplanTargetList computed.
895 locate_grouping_columns(root, tlist, result_plan->targetlist,
900 * Insert AGG or GROUP node if needed, plus an explicit sort step
903 * HAVING clause, if any, becomes qual of the Agg or Group node.
905 if (use_hashed_grouping)
907 /* Hashed aggregate plan --- no sort needed */
908 result_plan = (Plan *) make_agg(root,
910 (List *) parse->havingQual,
917 /* Hashed aggregation produces randomly-ordered results */
918 current_pathkeys = NIL;
920 else if (parse->hasAggs)
922 /* Plain aggregate plan --- sort if needed */
923 AggStrategy aggstrategy;
925 if (parse->groupClause)
927 if (!pathkeys_contained_in(group_pathkeys,
930 result_plan = (Plan *)
931 make_sort_from_groupcols(root,
935 current_pathkeys = group_pathkeys;
937 aggstrategy = AGG_SORTED;
940 * The AGG node will not change the sort ordering of its
941 * groups, so current_pathkeys describes the result too.
946 aggstrategy = AGG_PLAIN;
947 /* Result will be only one row anyway; no sort order */
948 current_pathkeys = NIL;
951 result_plan = (Plan *) make_agg(root,
953 (List *) parse->havingQual,
961 else if (parse->groupClause)
964 * GROUP BY without aggregation, so insert a group node (plus
965 * the appropriate sort node, if necessary).
967 * Add an explicit sort if we couldn't make the path come out
968 * the way the GROUP node needs it.
970 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
972 result_plan = (Plan *)
973 make_sort_from_groupcols(root,
977 current_pathkeys = group_pathkeys;
980 result_plan = (Plan *) make_group(root,
982 (List *) parse->havingQual,
987 /* The Group node won't change sort ordering */
989 else if (root->hasHavingQual)
992 * No aggregates, and no GROUP BY, but we have a HAVING qual.
993 * This is a degenerate case in which we are supposed to emit
994 * either 0 or 1 row depending on whether HAVING succeeds.
995 * Furthermore, there cannot be any variables in either HAVING
996 * or the targetlist, so we actually do not need the FROM
997 * table at all! We can just throw away the plan-so-far and
998 * generate a Result node. This is a sufficiently unusual
999 * corner case that it's not worth contorting the structure of
1000 * this routine to avoid having to generate the plan in the
1003 result_plan = (Plan *) make_result(tlist,
1007 } /* end of non-minmax-aggregate case */
1008 } /* end of if (setOperations) */
1011 * If we were not able to make the plan come out in the right order, add
1012 * an explicit sort step.
1014 if (parse->sortClause)
1016 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1018 result_plan = (Plan *)
1019 make_sort_from_sortclauses(root,
1022 current_pathkeys = sort_pathkeys;
1027 * If there is a DISTINCT clause, add the UNIQUE node.
1029 if (parse->distinctClause)
1031 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1034 * If there was grouping or aggregation, leave plan_rows as-is (ie,
1035 * assume the result was already mostly unique). If not, use the
1036 * number of distinct-groups calculated by query_planner.
1038 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1039 result_plan->plan_rows = dNumGroups;
1043 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1045 if (parse->limitCount || parse->limitOffset)
1047 result_plan = (Plan *) make_limit(result_plan,
1055 * Return the actual output ordering in query_pathkeys for possible use by
1056 * an outer query level.
1058 root->query_pathkeys = current_pathkeys;
1064 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1066 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1067 * results back in *count_est and *offset_est. These variables are set to
1068 * 0 if the corresponding clause is not present, and -1 if it's present
1069 * but we couldn't estimate the value for it. (The "0" convention is OK
1070 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1071 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1072 * usual practice of never estimating less than one row.) These values will
1073 * be passed to make_limit, which see if you change this code.
1075 * The return value is the suitably adjusted tuple_fraction to use for
1076 * planning the query. This adjustment is not overridable, since it reflects
1077 * plan actions that grouping_planner() will certainly take, not assumptions
1081 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1082 int *offset_est, int *count_est)
1084 Query *parse = root->parse;
1086 double limit_fraction;
1088 /* Should not be called unless LIMIT or OFFSET */
1089 Assert(parse->limitCount || parse->limitOffset);
1092 * Try to obtain the clause values. We use estimate_expression_value
1093 * primarily because it can sometimes do something useful with Params.
1095 if (parse->limitCount)
1097 est = estimate_expression_value(parse->limitCount);
1098 if (est && IsA(est, Const))
1100 if (((Const *) est)->constisnull)
1102 /* NULL indicates LIMIT ALL, ie, no limit */
1103 *count_est = 0; /* treat as not present */
1107 *count_est = DatumGetInt32(((Const *) est)->constvalue);
1108 if (*count_est <= 0)
1109 *count_est = 1; /* force to at least 1 */
1113 *count_est = -1; /* can't estimate */
1116 *count_est = 0; /* not present */
1118 if (parse->limitOffset)
1120 est = estimate_expression_value(parse->limitOffset);
1121 if (est && IsA(est, Const))
1123 if (((Const *) est)->constisnull)
1125 /* Treat NULL as no offset; the executor will too */
1126 *offset_est = 0; /* treat as not present */
1130 *offset_est = DatumGetInt32(((Const *) est)->constvalue);
1131 if (*offset_est < 0)
1132 *offset_est = 0; /* less than 0 is same as 0 */
1136 *offset_est = -1; /* can't estimate */
1139 *offset_est = 0; /* not present */
1141 if (*count_est != 0)
1144 * A LIMIT clause limits the absolute number of tuples returned.
1145 * However, if it's not a constant LIMIT then we have to guess; for
1146 * lack of a better idea, assume 10% of the plan's result is wanted.
1148 if (*count_est < 0 || *offset_est < 0)
1150 /* LIMIT or OFFSET is an expression ... punt ... */
1151 limit_fraction = 0.10;
1155 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1156 limit_fraction = (double) *count_est + (double) *offset_est;
1160 * If we have absolute limits from both caller and LIMIT, use the
1161 * smaller value; likewise if they are both fractional. If one is
1162 * fractional and the other absolute, we can't easily determine which
1163 * is smaller, but we use the heuristic that the absolute will usually
1166 if (tuple_fraction >= 1.0)
1168 if (limit_fraction >= 1.0)
1171 tuple_fraction = Min(tuple_fraction, limit_fraction);
1175 /* caller absolute, limit fractional; use caller's value */
1178 else if (tuple_fraction > 0.0)
1180 if (limit_fraction >= 1.0)
1182 /* caller fractional, limit absolute; use limit */
1183 tuple_fraction = limit_fraction;
1187 /* both fractional */
1188 tuple_fraction = Min(tuple_fraction, limit_fraction);
1193 /* no info from caller, just use limit */
1194 tuple_fraction = limit_fraction;
1197 else if (*offset_est != 0 && tuple_fraction > 0.0)
1200 * We have an OFFSET but no LIMIT. This acts entirely differently
1201 * from the LIMIT case: here, we need to increase rather than decrease
1202 * the caller's tuple_fraction, because the OFFSET acts to cause more
1203 * tuples to be fetched instead of fewer. This only matters if we got
1204 * a tuple_fraction > 0, however.
1206 * As above, use 10% if OFFSET is present but unestimatable.
1208 if (*offset_est < 0)
1209 limit_fraction = 0.10;
1211 limit_fraction = (double) *offset_est;
1214 * If we have absolute counts from both caller and OFFSET, add them
1215 * together; likewise if they are both fractional. If one is
1216 * fractional and the other absolute, we want to take the larger, and
1217 * we heuristically assume that's the fractional one.
1219 if (tuple_fraction >= 1.0)
1221 if (limit_fraction >= 1.0)
1223 /* both absolute, so add them together */
1224 tuple_fraction += limit_fraction;
1228 /* caller absolute, limit fractional; use limit */
1229 tuple_fraction = limit_fraction;
1234 if (limit_fraction >= 1.0)
1236 /* caller fractional, limit absolute; use caller's value */
1240 /* both fractional, so add them together */
1241 tuple_fraction += limit_fraction;
1242 if (tuple_fraction >= 1.0)
1243 tuple_fraction = 0.0; /* assume fetch all */
1248 return tuple_fraction;
1252 * choose_hashed_grouping - should we use hashed grouping?
1255 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1256 Path *cheapest_path, Path *sorted_path,
1257 double dNumGroups, AggClauseCounts *agg_counts)
1259 int numGroupCols = list_length(root->parse->groupClause);
1260 double cheapest_path_rows;
1261 int cheapest_path_width;
1263 List *current_pathkeys;
1268 * Check can't-do-it conditions, including whether the grouping operators
1271 * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1272 * (Doing so would imply storing *all* the input values in the hash table,
1273 * which seems like a certain loser.)
1275 if (!enable_hashagg)
1277 if (agg_counts->numDistinctAggs != 0)
1279 if (!hash_safe_grouping(root))
1283 * Don't do it if it doesn't look like the hashtable will fit into
1286 * Beware here of the possibility that cheapest_path->parent is NULL. This
1287 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1289 if (cheapest_path->parent)
1291 cheapest_path_rows = cheapest_path->parent->rows;
1292 cheapest_path_width = cheapest_path->parent->width;
1296 cheapest_path_rows = 1; /* assume non-set result */
1297 cheapest_path_width = 100; /* arbitrary */
1300 /* Estimate per-hash-entry space at tuple width... */
1301 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1302 /* plus space for pass-by-ref transition values... */
1303 hashentrysize += agg_counts->transitionSpace;
1304 /* plus the per-hash-entry overhead */
1305 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1307 if (hashentrysize * dNumGroups > work_mem * 1024L)
1311 * See if the estimated cost is no more than doing it the other way. While
1312 * avoiding the need for sorted input is usually a win, the fact that the
1313 * output won't be sorted may be a loss; so we need to do an actual cost
1316 * We need to consider cheapest_path + hashagg [+ final sort] versus
1317 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1318 * presorted_path + group or agg [+ final sort] where brackets indicate a
1319 * step that may not be needed. We assume query_planner() will have
1320 * returned a presorted path only if it's a winner compared to
1321 * cheapest_path for this purpose.
1323 * These path variables are dummies that just hold cost fields; we don't
1324 * make actual Paths for these steps.
1326 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1327 numGroupCols, dNumGroups,
1328 cheapest_path->startup_cost, cheapest_path->total_cost,
1329 cheapest_path_rows);
1330 /* Result of hashed agg is always unsorted */
1331 if (root->sort_pathkeys)
1332 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1333 dNumGroups, cheapest_path_width);
1337 sorted_p.startup_cost = sorted_path->startup_cost;
1338 sorted_p.total_cost = sorted_path->total_cost;
1339 current_pathkeys = sorted_path->pathkeys;
1343 sorted_p.startup_cost = cheapest_path->startup_cost;
1344 sorted_p.total_cost = cheapest_path->total_cost;
1345 current_pathkeys = cheapest_path->pathkeys;
1347 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1349 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1350 cheapest_path_rows, cheapest_path_width);
1351 current_pathkeys = root->group_pathkeys;
1354 if (root->parse->hasAggs)
1355 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1356 numGroupCols, dNumGroups,
1357 sorted_p.startup_cost, sorted_p.total_cost,
1358 cheapest_path_rows);
1360 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1361 sorted_p.startup_cost, sorted_p.total_cost,
1362 cheapest_path_rows);
1363 /* The Agg or Group node will preserve ordering */
1364 if (root->sort_pathkeys &&
1365 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1366 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1367 dNumGroups, cheapest_path_width);
1370 * Now make the decision using the top-level tuple fraction. First we
1371 * have to convert an absolute count (LIMIT) into fractional form.
1373 if (tuple_fraction >= 1.0)
1374 tuple_fraction /= dNumGroups;
1376 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1377 tuple_fraction) < 0)
1379 /* Hashed is cheaper, so use it */
1386 * hash_safe_grouping - are grouping operators hashable?
1388 * We assume hashed aggregation will work if the datatype's equality operator
1389 * is marked hashjoinable.
1392 hash_safe_grouping(PlannerInfo *root)
1396 foreach(gl, root->parse->groupClause)
1398 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1399 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1400 root->parse->targetList);
1404 optup = equality_oper(exprType((Node *) tle->expr), true);
1407 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1408 ReleaseSysCache(optup);
1416 * make_subplanTargetList
1417 * Generate appropriate target list when grouping is required.
1419 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1420 * above the result of query_planner, we typically want to pass a different
1421 * target list to query_planner than the outer plan nodes should have.
1422 * This routine generates the correct target list for the subplan.
1424 * The initial target list passed from the parser already contains entries
1425 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1426 * for variables used only in HAVING clauses; so we need to add those
1427 * variables to the subplan target list. Also, we flatten all expressions
1428 * except GROUP BY items into their component variables; the other expressions
1429 * will be computed by the inserted nodes rather than by the subplan.
1430 * For example, given a query like
1431 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1432 * we want to pass this targetlist to the subplan:
1434 * where the a+b target will be used by the Sort/Group steps, and the
1435 * other targets will be used for computing the final results. (In the
1436 * above example we could theoretically suppress the a and b targets and
1437 * pass down only c,d,a+b, but it's not really worth the trouble to
1438 * eliminate simple var references from the subplan. We will avoid doing
1439 * the extra computation to recompute a+b at the outer level; see
1440 * replace_vars_with_subplan_refs() in setrefs.c.)
1442 * If we are grouping or aggregating, *and* there are no non-Var grouping
1443 * expressions, then the returned tlist is effectively dummy; we do not
1444 * need to force it to be evaluated, because all the Vars it contains
1445 * should be present in the output of query_planner anyway.
1447 * 'tlist' is the query's target list.
1448 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1449 * expressions (if there are any) in the subplan's target list.
1450 * 'need_tlist_eval' is set true if we really need to evaluate the
1453 * The result is the targetlist to be passed to the subplan.
1457 make_subplanTargetList(PlannerInfo *root,
1459 AttrNumber **groupColIdx,
1460 bool *need_tlist_eval)
1462 Query *parse = root->parse;
1467 *groupColIdx = NULL;
1470 * If we're not grouping or aggregating, there's nothing to do here;
1471 * query_planner should receive the unmodified target list.
1473 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1475 *need_tlist_eval = true;
1480 * Otherwise, start with a "flattened" tlist (having just the vars
1481 * mentioned in the targetlist and HAVING qual --- but not upper- level
1482 * Vars; they will be replaced by Params later on).
1484 sub_tlist = flatten_tlist(tlist);
1485 extravars = pull_var_clause(parse->havingQual, false);
1486 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1487 list_free(extravars);
1488 *need_tlist_eval = false; /* only eval if not flat tlist */
1491 * If grouping, create sub_tlist entries for all GROUP BY expressions
1492 * (GROUP BY items that are simple Vars should be in the list already),
1493 * and make an array showing where the group columns are in the sub_tlist.
1495 numCols = list_length(parse->groupClause);
1499 AttrNumber *grpColIdx;
1502 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1503 *groupColIdx = grpColIdx;
1505 foreach(gl, parse->groupClause)
1507 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1508 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1509 TargetEntry *te = NULL;
1512 /* Find or make a matching sub_tlist entry */
1513 foreach(sl, sub_tlist)
1515 te = (TargetEntry *) lfirst(sl);
1516 if (equal(groupexpr, te->expr))
1521 te = makeTargetEntry((Expr *) groupexpr,
1522 list_length(sub_tlist) + 1,
1525 sub_tlist = lappend(sub_tlist, te);
1526 *need_tlist_eval = true; /* it's not flat anymore */
1529 /* and save its resno */
1530 grpColIdx[keyno++] = te->resno;
1538 * locate_grouping_columns
1539 * Locate grouping columns in the tlist chosen by query_planner.
1541 * This is only needed if we don't use the sub_tlist chosen by
1542 * make_subplanTargetList. We have to forget the column indexes found
1543 * by that routine and re-locate the grouping vars in the real sub_tlist.
1546 locate_grouping_columns(PlannerInfo *root,
1549 AttrNumber *groupColIdx)
1555 * No work unless grouping.
1557 if (!root->parse->groupClause)
1559 Assert(groupColIdx == NULL);
1562 Assert(groupColIdx != NULL);
1564 foreach(gl, root->parse->groupClause)
1566 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1567 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1568 TargetEntry *te = NULL;
1571 foreach(sl, sub_tlist)
1573 te = (TargetEntry *) lfirst(sl);
1574 if (equal(groupexpr, te->expr))
1578 elog(ERROR, "failed to locate grouping columns");
1580 groupColIdx[keyno++] = te->resno;
1585 * postprocess_setop_tlist
1586 * Fix up targetlist returned by plan_set_operations().
1588 * We need to transpose sort key info from the orig_tlist into new_tlist.
1589 * NOTE: this would not be good enough if we supported resjunk sort keys
1590 * for results of set operations --- then, we'd need to project a whole
1591 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1592 * find any resjunk columns in orig_tlist.
1595 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1598 ListCell *orig_tlist_item = list_head(orig_tlist);
1600 foreach(l, new_tlist)
1602 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1603 TargetEntry *orig_tle;
1605 /* ignore resjunk columns in setop result */
1606 if (new_tle->resjunk)
1609 Assert(orig_tlist_item != NULL);
1610 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1611 orig_tlist_item = lnext(orig_tlist_item);
1612 if (orig_tle->resjunk) /* should not happen */
1613 elog(ERROR, "resjunk output columns are not implemented");
1614 Assert(new_tle->resno == orig_tle->resno);
1615 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1617 if (orig_tlist_item != NULL)
1618 elog(ERROR, "resjunk output columns are not implemented");