1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2008, 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 1.243 2008/09/09 18:58:08 tgl 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/lsyscache.h"
42 #include "utils/syscache.h"
46 double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
48 /* Hook for plugins to get control in planner() */
49 planner_hook_type planner_hook = NULL;
52 /* Expression kind codes for preprocess_expression */
53 #define EXPRKIND_QUAL 0
54 #define EXPRKIND_TARGET 1
55 #define EXPRKIND_RTFUNC 2
56 #define EXPRKIND_VALUES 3
57 #define EXPRKIND_LIMIT 4
58 #define EXPRKIND_APPINFO 5
61 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
62 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
63 static Plan *inheritance_planner(PlannerInfo *root);
64 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
65 static bool is_dummy_plan(Plan *plan);
66 static double preprocess_limit(PlannerInfo *root,
67 double tuple_fraction,
68 int64 *offset_est, int64 *count_est);
69 static void preprocess_groupclause(PlannerInfo *root);
70 static bool choose_hashed_grouping(PlannerInfo *root,
71 double tuple_fraction, double limit_tuples,
72 Path *cheapest_path, Path *sorted_path,
73 double dNumGroups, AggClauseCounts *agg_counts);
74 static bool choose_hashed_distinct(PlannerInfo *root,
75 Plan *input_plan, List *input_pathkeys,
76 double tuple_fraction, double limit_tuples,
77 double dNumDistinctRows);
78 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
79 AttrNumber **groupColIdx, bool *need_tlist_eval);
80 static void locate_grouping_columns(PlannerInfo *root,
83 AttrNumber *groupColIdx);
84 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
87 /*****************************************************************************
89 * Query optimizer entry point
91 * To support loadable plugins that monitor or modify planner behavior,
92 * we provide a hook variable that lets a plugin get control before and
93 * after the standard planning process. The plugin would normally call
96 * Note to plugin authors: standard_planner() scribbles on its Query input,
97 * so you'd better copy that data structure if you want to plan more than once.
99 *****************************************************************************/
101 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
106 result = (*planner_hook) (parse, cursorOptions, boundParams);
108 result = standard_planner(parse, cursorOptions, boundParams);
113 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
117 double tuple_fraction;
123 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
124 if (parse->utilityStmt &&
125 IsA(parse->utilityStmt, DeclareCursorStmt))
126 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
129 * Set up global state for this planner invocation. This data is needed
130 * across all levels of sub-Query that might exist in the given command,
131 * so we keep it in a separate struct that's linked to by each per-Query
134 glob = makeNode(PlannerGlobal);
136 glob->boundParams = boundParams;
137 glob->paramlist = NIL;
138 glob->subplans = NIL;
139 glob->subrtables = NIL;
140 glob->rewindPlanIDs = NULL;
141 glob->finalrtable = NIL;
142 glob->relationOids = NIL;
143 glob->invalItems = NIL;
144 glob->transientPlan = false;
146 /* Determine what fraction of the plan is likely to be scanned */
147 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
150 * We have no real idea how many tuples the user will ultimately FETCH
151 * from a cursor, but it is often the case that he doesn't want 'em
152 * all, or would prefer a fast-start plan anyway so that he can
153 * process some of the tuples sooner. Use a GUC parameter to decide
154 * what fraction to optimize for.
156 tuple_fraction = cursor_tuple_fraction;
159 * We document cursor_tuple_fraction as simply being a fraction,
160 * which means the edge cases 0 and 1 have to be treated specially
161 * here. We convert 1 to 0 ("all the tuples") and 0 to a very small
164 if (tuple_fraction >= 1.0)
165 tuple_fraction = 0.0;
166 else if (tuple_fraction <= 0.0)
167 tuple_fraction = 1e-10;
171 /* Default assumption is we need all the tuples */
172 tuple_fraction = 0.0;
175 /* primary planning entry point (may recurse for subqueries) */
176 top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
179 * If creating a plan for a scrollable cursor, make sure it can run
180 * backwards on demand. Add a Material node at the top at need.
182 if (cursorOptions & CURSOR_OPT_SCROLL)
184 if (!ExecSupportsBackwardScan(top_plan))
185 top_plan = materialize_finished_plan(top_plan);
188 /* final cleanup of the plan */
189 Assert(glob->finalrtable == NIL);
190 top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
191 /* ... and the subplans (both regular subplans and initplans) */
192 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
193 forboth(lp, glob->subplans, lr, glob->subrtables)
195 Plan *subplan = (Plan *) lfirst(lp);
196 List *subrtable = (List *) lfirst(lr);
198 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
201 /* build the PlannedStmt result */
202 result = makeNode(PlannedStmt);
204 result->commandType = parse->commandType;
205 result->canSetTag = parse->canSetTag;
206 result->transientPlan = glob->transientPlan;
207 result->planTree = top_plan;
208 result->rtable = glob->finalrtable;
209 result->resultRelations = root->resultRelations;
210 result->utilityStmt = parse->utilityStmt;
211 result->intoClause = parse->intoClause;
212 result->subplans = glob->subplans;
213 result->rewindPlanIDs = glob->rewindPlanIDs;
214 result->returningLists = root->returningLists;
215 result->rowMarks = parse->rowMarks;
216 result->relationOids = glob->relationOids;
217 result->invalItems = glob->invalItems;
218 result->nParamExec = list_length(glob->paramlist);
224 /*--------------------
226 * Invokes the planner on a subquery. We recurse to here for each
227 * sub-SELECT found in the query tree.
229 * glob is the global state for the current planner run.
230 * parse is the querytree produced by the parser & rewriter.
231 * level is the current recursion depth (1 at the top-level Query).
232 * tuple_fraction is the fraction of tuples we expect will be retrieved.
233 * tuple_fraction is interpreted as explained for grouping_planner, below.
235 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
236 * among other things this tells the output sort ordering of the plan.
238 * Basically, this routine does the stuff that should only be done once
239 * per Query object. It then calls grouping_planner. At one time,
240 * grouping_planner could be invoked recursively on the same Query object;
241 * that's not currently true, but we keep the separation between the two
242 * routines anyway, in case we need it again someday.
244 * subquery_planner will be called recursively to handle sub-Query nodes
245 * found within the query's expressions and rangetable.
247 * Returns a query plan.
248 *--------------------
251 subquery_planner(PlannerGlobal *glob, Query *parse,
252 Index level, double tuple_fraction,
253 PlannerInfo **subroot)
255 int num_old_subplans = list_length(glob->subplans);
262 /* Create a PlannerInfo data structure for this subquery */
263 root = makeNode(PlannerInfo);
266 root->query_level = level;
267 root->planner_cxt = CurrentMemoryContext;
268 root->init_plans = NIL;
269 root->eq_classes = NIL;
270 root->append_rel_list = NIL;
273 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
274 * to transform them into joins. Note that this step does not descend
275 * into subqueries; if we pull up any subqueries below, their SubLinks are
276 * processed just before pulling them up.
278 if (parse->hasSubLinks)
279 pull_up_sublinks(root);
282 * Scan the rangetable for set-returning functions, and inline them
283 * if possible (producing subqueries that might get pulled up next).
284 * Recursion issues here are handled in the same way as for SubLinks.
286 inline_set_returning_functions(root);
289 * Check to see if any subqueries in the rangetable can be merged into
292 parse->jointree = (FromExpr *)
293 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
296 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
297 * avoid the expense of doing flatten_join_alias_vars(). Also check for
298 * outer joins --- if none, we can skip reduce_outer_joins().
299 * This must be done after we have done pull_up_subqueries, of course.
301 root->hasJoinRTEs = false;
302 hasOuterJoins = false;
303 foreach(l, parse->rtable)
305 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
307 if (rte->rtekind == RTE_JOIN)
309 root->hasJoinRTEs = true;
310 if (IS_OUTER_JOIN(rte->jointype))
312 hasOuterJoins = true;
313 /* Can quit scanning once we find an outer join */
320 * Expand any rangetable entries that are inheritance sets into "append
321 * relations". This can add entries to the rangetable, but they must be
322 * plain base relations not joins, so it's OK (and marginally more
323 * efficient) to do it after checking for join RTEs. We must do it after
324 * pulling up subqueries, else we'd fail to handle inherited tables in
327 expand_inherited_tables(root);
330 * Set hasHavingQual to remember if HAVING clause is present. Needed
331 * because preprocess_expression will reduce a constant-true condition to
332 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
334 root->hasHavingQual = (parse->havingQual != NULL);
336 /* Clear this flag; might get set in distribute_qual_to_rels */
337 root->hasPseudoConstantQuals = false;
340 * Do expression preprocessing on targetlist and quals.
342 parse->targetList = (List *)
343 preprocess_expression(root, (Node *) parse->targetList,
346 parse->returningList = (List *)
347 preprocess_expression(root, (Node *) parse->returningList,
350 preprocess_qual_conditions(root, (Node *) parse->jointree);
352 parse->havingQual = preprocess_expression(root, parse->havingQual,
355 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
357 parse->limitCount = preprocess_expression(root, parse->limitCount,
360 root->append_rel_list = (List *)
361 preprocess_expression(root, (Node *) root->append_rel_list,
364 /* Also need to preprocess expressions for function and values RTEs */
365 foreach(l, parse->rtable)
367 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
369 if (rte->rtekind == RTE_FUNCTION)
370 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
372 else if (rte->rtekind == RTE_VALUES)
373 rte->values_lists = (List *)
374 preprocess_expression(root, (Node *) rte->values_lists,
379 * In some cases we may want to transfer a HAVING clause into WHERE. We
380 * cannot do so if the HAVING clause contains aggregates (obviously) or
381 * volatile functions (since a HAVING clause is supposed to be executed
382 * only once per group). Also, it may be that the clause is so expensive
383 * to execute that we're better off doing it only once per group, despite
384 * the loss of selectivity. This is hard to estimate short of doing the
385 * entire planning process twice, so we use a heuristic: clauses
386 * containing subplans are left in HAVING. Otherwise, we move or copy the
387 * HAVING clause into WHERE, in hopes of eliminating tuples before
388 * aggregation instead of after.
390 * If the query has explicit grouping then we can simply move such a
391 * clause into WHERE; any group that fails the clause will not be in the
392 * output because none of its tuples will reach the grouping or
393 * aggregation stage. Otherwise we must have a degenerate (variable-free)
394 * HAVING clause, which we put in WHERE so that query_planner() can use it
395 * in a gating Result node, but also keep in HAVING to ensure that we
396 * don't emit a bogus aggregated row. (This could be done better, but it
397 * seems not worth optimizing.)
399 * Note that both havingQual and parse->jointree->quals are in
400 * implicitly-ANDed-list form at this point, even though they are declared
404 foreach(l, (List *) parse->havingQual)
406 Node *havingclause = (Node *) lfirst(l);
408 if (contain_agg_clause(havingclause) ||
409 contain_volatile_functions(havingclause) ||
410 contain_subplans(havingclause))
412 /* keep it in HAVING */
413 newHaving = lappend(newHaving, havingclause);
415 else if (parse->groupClause)
417 /* move it to WHERE */
418 parse->jointree->quals = (Node *)
419 lappend((List *) parse->jointree->quals, havingclause);
423 /* put a copy in WHERE, keep it in HAVING */
424 parse->jointree->quals = (Node *)
425 lappend((List *) parse->jointree->quals,
426 copyObject(havingclause));
427 newHaving = lappend(newHaving, havingclause);
430 parse->havingQual = (Node *) newHaving;
433 * If we have any outer joins, try to reduce them to plain inner joins.
434 * This step is most easily done after we've done expression
438 reduce_outer_joins(root);
441 * Do the main planning. If we have an inherited target relation, that
442 * needs special processing, else go straight to grouping_planner.
444 if (parse->resultRelation &&
445 rt_fetch(parse->resultRelation, parse->rtable)->inh)
446 plan = inheritance_planner(root);
448 plan = grouping_planner(root, tuple_fraction);
451 * If any subplans were generated, or if we're inside a subplan, build
452 * initPlan list and extParam/allParam sets for plan nodes, and attach the
453 * initPlans to the top plan node.
455 if (list_length(glob->subplans) != num_old_subplans ||
456 root->query_level > 1)
457 SS_finalize_plan(root, plan, true);
459 /* Return internal info if caller wants it */
467 * preprocess_expression
468 * Do subquery_planner's preprocessing work for an expression,
469 * which can be a targetlist, a WHERE clause (including JOIN/ON
470 * conditions), or a HAVING clause.
473 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
476 * Fall out quickly if expression is empty. This occurs often enough to
477 * be worth checking. Note that null->null is the correct conversion for
478 * implicit-AND result format, too.
484 * If the query has any join RTEs, replace join alias variables with
485 * base-relation variables. We must do this before sublink processing,
486 * else sublinks expanded out from join aliases wouldn't get processed. We
487 * can skip it in VALUES lists, however, since they can't contain any Vars
490 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
491 expr = flatten_join_alias_vars(root, expr);
494 * Simplify constant expressions.
496 * Note: this also flattens nested AND and OR expressions into N-argument
497 * form. All processing of a qual expression after this point must be
498 * careful to maintain AND/OR flatness --- that is, do not generate a tree
499 * with AND directly under AND, nor OR directly under OR.
501 * Because this is a relatively expensive process, we skip it when the
502 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
503 * expression will only be evaluated once anyway, so no point in
504 * pre-simplifying; we can't execute it any faster than the executor can,
505 * and we will waste cycles copying the tree. Notice however that we
506 * still must do it for quals (to get AND/OR flatness); and if we are in a
507 * subquery we should not assume it will be done only once.
509 * For VALUES lists we never do this at all, again on the grounds that we
510 * should optimize for one-time evaluation.
512 if (kind != EXPRKIND_VALUES &&
513 (root->parse->jointree->fromlist != NIL ||
514 kind == EXPRKIND_QUAL ||
515 root->query_level > 1))
516 expr = eval_const_expressions(root, expr);
519 * If it's a qual or havingQual, canonicalize it.
521 if (kind == EXPRKIND_QUAL)
523 expr = (Node *) canonicalize_qual((Expr *) expr);
525 #ifdef OPTIMIZER_DEBUG
526 printf("After canonicalize_qual()\n");
531 /* Expand SubLinks to SubPlans */
532 if (root->parse->hasSubLinks)
533 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
536 * XXX do not insert anything here unless you have grokked the comments in
537 * SS_replace_correlation_vars ...
540 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
541 if (root->query_level > 1)
542 expr = SS_replace_correlation_vars(root, expr);
545 * If it's a qual or havingQual, convert it to implicit-AND format. (We
546 * don't want to do this before eval_const_expressions, since the latter
547 * would be unable to simplify a top-level AND correctly. Also,
548 * SS_process_sublinks expects explicit-AND format.)
550 if (kind == EXPRKIND_QUAL)
551 expr = (Node *) make_ands_implicit((Expr *) expr);
557 * preprocess_qual_conditions
558 * Recursively scan the query's jointree and do subquery_planner's
559 * preprocessing work on each qual condition found therein.
562 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
566 if (IsA(jtnode, RangeTblRef))
568 /* nothing to do here */
570 else if (IsA(jtnode, FromExpr))
572 FromExpr *f = (FromExpr *) jtnode;
575 foreach(l, f->fromlist)
576 preprocess_qual_conditions(root, lfirst(l));
578 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
580 else if (IsA(jtnode, JoinExpr))
582 JoinExpr *j = (JoinExpr *) jtnode;
584 preprocess_qual_conditions(root, j->larg);
585 preprocess_qual_conditions(root, j->rarg);
587 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
590 elog(ERROR, "unrecognized node type: %d",
591 (int) nodeTag(jtnode));
595 * inheritance_planner
596 * Generate a plan in the case where the result relation is an
599 * We have to handle this case differently from cases where a source relation
600 * is an inheritance set. Source inheritance is expanded at the bottom of the
601 * plan tree (see allpaths.c), but target inheritance has to be expanded at
602 * the top. The reason is that for UPDATE, each target relation needs a
603 * different targetlist matching its own column set. Also, for both UPDATE
604 * and DELETE, the executor needs the Append plan node at the top, else it
605 * can't keep track of which table is the current target table. Fortunately,
606 * the UPDATE/DELETE target can never be the nullable side of an outer join,
607 * so it's OK to generate the plan this way.
609 * Returns a query plan.
612 inheritance_planner(PlannerInfo *root)
614 Query *parse = root->parse;
615 int parentRTindex = parse->resultRelation;
616 List *subplans = NIL;
617 List *resultRelations = NIL;
618 List *returningLists = NIL;
624 foreach(l, root->append_rel_list)
626 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
629 /* append_rel_list contains all append rels; ignore others */
630 if (appinfo->parent_relid != parentRTindex)
634 * Generate modified query with this rel as target.
636 memcpy(&subroot, root, sizeof(PlannerInfo));
637 subroot.parse = (Query *)
638 adjust_appendrel_attrs((Node *) parse,
640 subroot.init_plans = NIL;
641 /* There shouldn't be any OJ info to translate, as yet */
642 Assert(subroot.join_info_list == NIL);
645 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
648 * If this child rel was excluded by constraint exclusion, exclude it
651 if (is_dummy_plan(subplan))
654 /* Save rtable and tlist from first rel for use below */
657 rtable = subroot.parse->rtable;
658 tlist = subplan->targetlist;
661 subplans = lappend(subplans, subplan);
663 /* Make sure any initplans from this rel get into the outer list */
664 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
666 /* Build target-relations list for the executor */
667 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
669 /* Build list of per-relation RETURNING targetlists */
670 if (parse->returningList)
672 Assert(list_length(subroot.returningLists) == 1);
673 returningLists = list_concat(returningLists,
674 subroot.returningLists);
678 root->resultRelations = resultRelations;
679 root->returningLists = returningLists;
681 /* Mark result as unordered (probably unnecessary) */
682 root->query_pathkeys = NIL;
685 * If we managed to exclude every child rel, return a dummy plan
689 root->resultRelations = list_make1_int(parentRTindex);
690 /* although dummy, it must have a valid tlist for executor */
691 tlist = preprocess_targetlist(root, parse->targetList);
692 return (Plan *) make_result(root,
694 (Node *) list_make1(makeBoolConst(false,
700 * Planning might have modified the rangetable, due to changes of the
701 * Query structures inside subquery RTEs. We have to ensure that this
702 * gets propagated back to the master copy. But can't do this until we
703 * are done planning, because all the calls to grouping_planner need
704 * virgin sub-Queries to work from. (We are effectively assuming that
705 * sub-Queries will get planned identically each time, or at least that
706 * the impacts on their rangetables will be the same each time.)
708 * XXX should clean this up someday
710 parse->rtable = rtable;
712 /* Suppress Append if there's only one surviving child rel */
713 if (list_length(subplans) == 1)
714 return (Plan *) linitial(subplans);
716 return (Plan *) make_append(subplans, true, tlist);
719 /*--------------------
721 * Perform planning steps related to grouping, aggregation, etc.
722 * This primarily means adding top-level processing to the basic
723 * query plan produced by query_planner.
725 * tuple_fraction is the fraction of tuples we expect will be retrieved
727 * tuple_fraction is interpreted as follows:
728 * 0: expect all tuples to be retrieved (normal case)
729 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
730 * from the plan to be retrieved
731 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
732 * expected to be retrieved (ie, a LIMIT specification)
734 * Returns a query plan. Also, root->query_pathkeys is returned as the
735 * actual output ordering of the plan (in pathkey format).
736 *--------------------
739 grouping_planner(PlannerInfo *root, double tuple_fraction)
741 Query *parse = root->parse;
742 List *tlist = parse->targetList;
743 int64 offset_est = 0;
745 double limit_tuples = -1.0;
747 List *current_pathkeys;
748 double dNumGroups = 0;
750 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
751 if (parse->limitCount || parse->limitOffset)
753 tuple_fraction = preprocess_limit(root, tuple_fraction,
754 &offset_est, &count_est);
757 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
758 * estimate the effects of using a bounded sort.
760 if (count_est > 0 && offset_est >= 0)
761 limit_tuples = (double) count_est + (double) offset_est;
764 if (parse->setOperations)
766 List *set_sortclauses;
769 * If there's a top-level ORDER BY, assume we have to fetch all the
770 * tuples. This might be too simplistic given all the hackery below
771 * to possibly avoid the sort; but the odds of accurate estimates
772 * here are pretty low anyway.
774 if (parse->sortClause)
775 tuple_fraction = 0.0;
778 * Construct the plan for set operations. The result will not need
779 * any work except perhaps a top-level sort and/or LIMIT.
781 result_plan = plan_set_operations(root, tuple_fraction,
785 * Calculate pathkeys representing the sort order (if any) of the set
786 * operation's result. We have to do this before overwriting the sort
789 current_pathkeys = make_pathkeys_for_sortclauses(root,
791 result_plan->targetlist,
795 * We should not need to call preprocess_targetlist, since we must be
796 * in a SELECT query node. Instead, use the targetlist returned by
797 * plan_set_operations (since this tells whether it returned any
798 * resjunk columns!), and transfer any sort key information from the
801 Assert(parse->commandType == CMD_SELECT);
803 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
807 * Can't handle FOR UPDATE/SHARE here (parser should have checked
808 * already, but let's make sure).
812 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
813 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
816 * Calculate pathkeys that represent result ordering requirements
818 Assert(parse->distinctClause == NIL);
819 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
826 /* No set operations, do regular planning */
828 AttrNumber *groupColIdx = NULL;
829 bool need_tlist_eval = true;
835 AggClauseCounts agg_counts;
837 bool use_hashed_grouping = false;
839 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
841 /* Preprocess GROUP BY clause, if any */
842 if (parse->groupClause)
843 preprocess_groupclause(root);
844 numGroupCols = list_length(parse->groupClause);
846 /* Preprocess targetlist */
847 tlist = preprocess_targetlist(root, tlist);
850 * Generate appropriate target list for subplan; may be different from
851 * tlist if grouping or aggregation is needed.
853 sub_tlist = make_subplanTargetList(root, tlist,
854 &groupColIdx, &need_tlist_eval);
857 * Calculate pathkeys that represent grouping/ordering requirements.
858 * Stash them in PlannerInfo so that query_planner can canonicalize
859 * them after EquivalenceClasses have been formed. The sortClause
860 * is certainly sort-able, but GROUP BY and DISTINCT might not be,
861 * in which case we just leave their pathkeys empty.
863 if (parse->groupClause &&
864 grouping_is_sortable(parse->groupClause))
865 root->group_pathkeys =
866 make_pathkeys_for_sortclauses(root,
871 root->group_pathkeys = NIL;
873 if (parse->distinctClause &&
874 grouping_is_sortable(parse->distinctClause))
875 root->distinct_pathkeys =
876 make_pathkeys_for_sortclauses(root,
877 parse->distinctClause,
881 root->distinct_pathkeys = NIL;
883 root->sort_pathkeys =
884 make_pathkeys_for_sortclauses(root,
890 * Will need actual number of aggregates for estimating costs.
892 * Note: we do not attempt to detect duplicate aggregates here; a
893 * somewhat-overestimated count is okay for our present purposes.
895 * Note: think not that we can turn off hasAggs if we find no aggs. It
896 * is possible for constant-expression simplification to remove all
897 * explicit references to aggs, but we still have to follow the
898 * aggregate semantics (eg, producing only one output row).
902 count_agg_clauses((Node *) tlist, &agg_counts);
903 count_agg_clauses(parse->havingQual, &agg_counts);
907 * Figure out whether we want a sorted result from query_planner.
909 * If we have a sortable GROUP BY clause, then we want a result sorted
910 * properly for grouping. Otherwise, if there's a sortable DISTINCT
911 * clause that's more rigorous than the ORDER BY clause, we try to
912 * produce output that's sufficiently well sorted for the DISTINCT.
913 * Otherwise, if there is an ORDER BY clause, we want to sort by the
916 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
917 * superset of GROUP BY, it would be tempting to request sort by ORDER
918 * BY --- but that might just leave us failing to exploit an available
919 * sort order at all. Needs more thought. The choice for DISTINCT
920 * versus ORDER BY is much easier, since we know that the parser
921 * ensured that one is a superset of the other.
923 if (root->group_pathkeys)
924 root->query_pathkeys = root->group_pathkeys;
925 else if (list_length(root->distinct_pathkeys) >
926 list_length(root->sort_pathkeys))
927 root->query_pathkeys = root->distinct_pathkeys;
928 else if (root->sort_pathkeys)
929 root->query_pathkeys = root->sort_pathkeys;
931 root->query_pathkeys = NIL;
934 * Generate the best unsorted and presorted paths for this Query (but
935 * note there may not be any presorted path). query_planner will also
936 * estimate the number of groups in the query, and canonicalize all
939 query_planner(root, sub_tlist, tuple_fraction, limit_tuples,
940 &cheapest_path, &sorted_path, &dNumGroups);
943 * If grouping, decide whether to use sorted or hashed grouping.
945 if (parse->groupClause)
951 * Executor doesn't support hashed aggregation with DISTINCT
952 * aggregates. (Doing so would imply storing *all* the input
953 * values in the hash table, which seems like a certain loser.)
955 can_hash = (agg_counts.numDistinctAggs == 0 &&
956 grouping_is_hashable(parse->groupClause));
957 can_sort = grouping_is_sortable(parse->groupClause);
958 if (can_hash && can_sort)
960 /* we have a meaningful choice to make ... */
961 use_hashed_grouping =
962 choose_hashed_grouping(root,
963 tuple_fraction, limit_tuples,
964 cheapest_path, sorted_path,
965 dNumGroups, &agg_counts);
968 use_hashed_grouping = true;
970 use_hashed_grouping = false;
973 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
974 errmsg("could not implement GROUP BY"),
975 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
977 /* Also convert # groups to long int --- but 'ware overflow! */
978 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
982 * Select the best path. If we are doing hashed grouping, we will
983 * always read all the input tuples, so use the cheapest-total path.
984 * Otherwise, trust query_planner's decision about which to use.
986 if (use_hashed_grouping || !sorted_path)
987 best_path = cheapest_path;
989 best_path = sorted_path;
992 * Check to see if it's possible to optimize MIN/MAX aggregates. If
993 * so, we will forget all the work we did so far to choose a "regular"
994 * path ... but we had to do it anyway to be able to tell which way is
997 result_plan = optimize_minmax_aggregates(root,
1000 if (result_plan != NULL)
1003 * optimize_minmax_aggregates generated the full plan, with the
1004 * right tlist, and it has no sort order.
1006 current_pathkeys = NIL;
1011 * Normal case --- create a plan according to query_planner's
1014 bool need_sort_for_grouping = false;
1016 result_plan = create_plan(root, best_path);
1017 current_pathkeys = best_path->pathkeys;
1019 /* Detect if we'll need an explicit sort for grouping */
1020 if (parse->groupClause && !use_hashed_grouping &&
1021 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1023 need_sort_for_grouping = true;
1025 * Always override query_planner's tlist, so that we don't
1026 * sort useless data from a "physical" tlist.
1028 need_tlist_eval = true;
1032 * create_plan() returns a plan with just a "flat" tlist of
1033 * required Vars. Usually we need to insert the sub_tlist as the
1034 * tlist of the top plan node. However, we can skip that if we
1035 * determined that whatever query_planner chose to return will be
1038 if (need_tlist_eval)
1041 * If the top-level plan node is one that cannot do expression
1042 * evaluation, we must insert a Result node to project the
1045 if (!is_projection_capable_plan(result_plan))
1047 result_plan = (Plan *) make_result(root,
1055 * Otherwise, just replace the subplan's flat tlist with
1056 * the desired tlist.
1058 result_plan->targetlist = sub_tlist;
1062 * Also, account for the cost of evaluation of the sub_tlist.
1064 * Up to now, we have only been dealing with "flat" tlists,
1065 * containing just Vars. So their evaluation cost is zero
1066 * according to the model used by cost_qual_eval() (or if you
1067 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1068 * can avoid accounting for tlist cost throughout
1069 * query_planner() and subroutines. But now we've inserted a
1070 * tlist that might contain actual operators, sub-selects, etc
1071 * --- so we'd better account for its cost.
1073 * Below this point, any tlist eval cost for added-on nodes
1074 * should be accounted for as we create those nodes.
1075 * Presently, of the node types we can add on, only Agg and
1076 * Group project new tlists (the rest just copy their input
1077 * tuples) --- so make_agg() and make_group() are responsible
1078 * for computing the added cost.
1080 cost_qual_eval(&tlist_cost, sub_tlist, root);
1081 result_plan->startup_cost += tlist_cost.startup;
1082 result_plan->total_cost += tlist_cost.startup +
1083 tlist_cost.per_tuple * result_plan->plan_rows;
1088 * Since we're using query_planner's tlist and not the one
1089 * make_subplanTargetList calculated, we have to refigure any
1090 * grouping-column indexes make_subplanTargetList computed.
1092 locate_grouping_columns(root, tlist, result_plan->targetlist,
1097 * Insert AGG or GROUP node if needed, plus an explicit sort step
1100 * HAVING clause, if any, becomes qual of the Agg or Group node.
1102 if (use_hashed_grouping)
1104 /* Hashed aggregate plan --- no sort needed */
1105 result_plan = (Plan *) make_agg(root,
1107 (List *) parse->havingQual,
1111 extract_grouping_ops(parse->groupClause),
1115 /* Hashed aggregation produces randomly-ordered results */
1116 current_pathkeys = NIL;
1118 else if (parse->hasAggs)
1120 /* Plain aggregate plan --- sort if needed */
1121 AggStrategy aggstrategy;
1123 if (parse->groupClause)
1125 if (need_sort_for_grouping)
1127 result_plan = (Plan *)
1128 make_sort_from_groupcols(root,
1132 current_pathkeys = root->group_pathkeys;
1134 aggstrategy = AGG_SORTED;
1137 * The AGG node will not change the sort ordering of its
1138 * groups, so current_pathkeys describes the result too.
1143 aggstrategy = AGG_PLAIN;
1144 /* Result will be only one row anyway; no sort order */
1145 current_pathkeys = NIL;
1148 result_plan = (Plan *) make_agg(root,
1150 (List *) parse->havingQual,
1154 extract_grouping_ops(parse->groupClause),
1159 else if (parse->groupClause)
1162 * GROUP BY without aggregation, so insert a group node (plus
1163 * the appropriate sort node, if necessary).
1165 * Add an explicit sort if we couldn't make the path come out
1166 * the way the GROUP node needs it.
1168 if (need_sort_for_grouping)
1170 result_plan = (Plan *)
1171 make_sort_from_groupcols(root,
1175 current_pathkeys = root->group_pathkeys;
1178 result_plan = (Plan *) make_group(root,
1180 (List *) parse->havingQual,
1183 extract_grouping_ops(parse->groupClause),
1186 /* The Group node won't change sort ordering */
1188 else if (root->hasHavingQual)
1191 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1192 * This is a degenerate case in which we are supposed to emit
1193 * either 0 or 1 row depending on whether HAVING succeeds.
1194 * Furthermore, there cannot be any variables in either HAVING
1195 * or the targetlist, so we actually do not need the FROM
1196 * table at all! We can just throw away the plan-so-far and
1197 * generate a Result node. This is a sufficiently unusual
1198 * corner case that it's not worth contorting the structure of
1199 * this routine to avoid having to generate the plan in the
1202 result_plan = (Plan *) make_result(root,
1207 } /* end of non-minmax-aggregate case */
1208 } /* end of if (setOperations) */
1211 * If there is a DISTINCT clause, add the necessary node(s).
1213 if (parse->distinctClause)
1215 double dNumDistinctRows;
1216 long numDistinctRows;
1217 bool use_hashed_distinct;
1222 * If there was grouping or aggregation, use the current number of
1223 * rows as the estimated number of DISTINCT rows (ie, assume the
1224 * result was already mostly unique). If not, use the number of
1225 * distinct-groups calculated by query_planner.
1227 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1228 dNumDistinctRows = result_plan->plan_rows;
1230 dNumDistinctRows = dNumGroups;
1232 /* Also convert to long int --- but 'ware overflow! */
1233 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1236 * If we have a sortable DISTINCT ON clause, we always use sorting.
1237 * This enforces the expected behavior of DISTINCT ON.
1239 can_sort = grouping_is_sortable(parse->distinctClause);
1240 if (can_sort && parse->hasDistinctOn)
1241 use_hashed_distinct = false;
1244 can_hash = grouping_is_hashable(parse->distinctClause);
1245 if (can_hash && can_sort)
1247 /* we have a meaningful choice to make ... */
1248 use_hashed_distinct =
1249 choose_hashed_distinct(root,
1250 result_plan, current_pathkeys,
1251 tuple_fraction, limit_tuples,
1255 use_hashed_distinct = true;
1257 use_hashed_distinct = false;
1261 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1262 errmsg("could not implement DISTINCT"),
1263 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
1264 use_hashed_distinct = false; /* keep compiler quiet */
1268 if (use_hashed_distinct)
1270 /* Hashed aggregate plan --- no sort needed */
1271 result_plan = (Plan *) make_agg(root,
1272 result_plan->targetlist,
1275 list_length(parse->distinctClause),
1276 extract_grouping_cols(parse->distinctClause,
1277 result_plan->targetlist),
1278 extract_grouping_ops(parse->distinctClause),
1282 /* Hashed aggregation produces randomly-ordered results */
1283 current_pathkeys = NIL;
1288 * Use a Unique node to implement DISTINCT. Add an explicit sort
1289 * if we couldn't make the path come out the way the Unique node
1290 * needs it. If we do have to sort, always sort by the more
1291 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1292 * below. However, for regular DISTINCT, don't sort now if we
1293 * don't have to --- sorting afterwards will likely be cheaper,
1294 * and also has the possibility of optimizing via LIMIT. But
1295 * for DISTINCT ON, we *must* force the final sort now, else
1296 * it won't have the desired behavior.
1298 List *needed_pathkeys;
1300 if (parse->hasDistinctOn &&
1301 list_length(root->distinct_pathkeys) <
1302 list_length(root->sort_pathkeys))
1303 needed_pathkeys = root->sort_pathkeys;
1305 needed_pathkeys = root->distinct_pathkeys;
1307 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1309 if (list_length(root->distinct_pathkeys) >=
1310 list_length(root->sort_pathkeys))
1311 current_pathkeys = root->distinct_pathkeys;
1314 current_pathkeys = root->sort_pathkeys;
1315 /* Assert checks that parser didn't mess up... */
1316 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1320 result_plan = (Plan *) make_sort_from_pathkeys(root,
1326 result_plan = (Plan *) make_unique(result_plan,
1327 parse->distinctClause);
1328 result_plan->plan_rows = dNumDistinctRows;
1329 /* The Unique node won't change sort ordering */
1334 * If ORDER BY was given and we were not able to make the plan come out in
1335 * the right order, add an explicit sort step.
1337 if (parse->sortClause)
1339 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1341 result_plan = (Plan *) make_sort_from_pathkeys(root,
1343 root->sort_pathkeys,
1345 current_pathkeys = root->sort_pathkeys;
1350 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1352 if (parse->limitCount || parse->limitOffset)
1354 result_plan = (Plan *) make_limit(result_plan,
1362 * Deal with the RETURNING clause if any. It's convenient to pass the
1363 * returningList through setrefs.c now rather than at top level (if we
1364 * waited, handling inherited UPDATE/DELETE would be much harder).
1366 if (parse->returningList)
1370 Assert(parse->resultRelation);
1371 rlist = set_returning_clause_references(root->glob,
1372 parse->returningList,
1374 parse->resultRelation);
1375 root->returningLists = list_make1(rlist);
1378 root->returningLists = NIL;
1380 /* Compute result-relations list if needed */
1381 if (parse->resultRelation)
1382 root->resultRelations = list_make1_int(parse->resultRelation);
1384 root->resultRelations = NIL;
1387 * Return the actual output ordering in query_pathkeys for possible use by
1388 * an outer query level.
1390 root->query_pathkeys = current_pathkeys;
1396 * Detect whether a plan node is a "dummy" plan created when a relation
1397 * is deemed not to need scanning due to constraint exclusion.
1399 * Currently, such dummy plans are Result nodes with constant FALSE
1403 is_dummy_plan(Plan *plan)
1405 if (IsA(plan, Result))
1407 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1409 if (list_length(rcqual) == 1)
1411 Const *constqual = (Const *) linitial(rcqual);
1413 if (constqual && IsA(constqual, Const))
1415 if (!constqual->constisnull &&
1416 !DatumGetBool(constqual->constvalue))
1425 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1427 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1428 * results back in *count_est and *offset_est. These variables are set to
1429 * 0 if the corresponding clause is not present, and -1 if it's present
1430 * but we couldn't estimate the value for it. (The "0" convention is OK
1431 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1432 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1433 * usual practice of never estimating less than one row.) These values will
1434 * be passed to make_limit, which see if you change this code.
1436 * The return value is the suitably adjusted tuple_fraction to use for
1437 * planning the query. This adjustment is not overridable, since it reflects
1438 * plan actions that grouping_planner() will certainly take, not assumptions
1442 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1443 int64 *offset_est, int64 *count_est)
1445 Query *parse = root->parse;
1447 double limit_fraction;
1449 /* Should not be called unless LIMIT or OFFSET */
1450 Assert(parse->limitCount || parse->limitOffset);
1453 * Try to obtain the clause values. We use estimate_expression_value
1454 * primarily because it can sometimes do something useful with Params.
1456 if (parse->limitCount)
1458 est = estimate_expression_value(root, parse->limitCount);
1459 if (est && IsA(est, Const))
1461 if (((Const *) est)->constisnull)
1463 /* NULL indicates LIMIT ALL, ie, no limit */
1464 *count_est = 0; /* treat as not present */
1468 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1469 if (*count_est <= 0)
1470 *count_est = 1; /* force to at least 1 */
1474 *count_est = -1; /* can't estimate */
1477 *count_est = 0; /* not present */
1479 if (parse->limitOffset)
1481 est = estimate_expression_value(root, parse->limitOffset);
1482 if (est && IsA(est, Const))
1484 if (((Const *) est)->constisnull)
1486 /* Treat NULL as no offset; the executor will too */
1487 *offset_est = 0; /* treat as not present */
1491 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1492 if (*offset_est < 0)
1493 *offset_est = 0; /* less than 0 is same as 0 */
1497 *offset_est = -1; /* can't estimate */
1500 *offset_est = 0; /* not present */
1502 if (*count_est != 0)
1505 * A LIMIT clause limits the absolute number of tuples returned.
1506 * However, if it's not a constant LIMIT then we have to guess; for
1507 * lack of a better idea, assume 10% of the plan's result is wanted.
1509 if (*count_est < 0 || *offset_est < 0)
1511 /* LIMIT or OFFSET is an expression ... punt ... */
1512 limit_fraction = 0.10;
1516 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1517 limit_fraction = (double) *count_est + (double) *offset_est;
1521 * If we have absolute limits from both caller and LIMIT, use the
1522 * smaller value; likewise if they are both fractional. If one is
1523 * fractional and the other absolute, we can't easily determine which
1524 * is smaller, but we use the heuristic that the absolute will usually
1527 if (tuple_fraction >= 1.0)
1529 if (limit_fraction >= 1.0)
1532 tuple_fraction = Min(tuple_fraction, limit_fraction);
1536 /* caller absolute, limit fractional; use caller's value */
1539 else if (tuple_fraction > 0.0)
1541 if (limit_fraction >= 1.0)
1543 /* caller fractional, limit absolute; use limit */
1544 tuple_fraction = limit_fraction;
1548 /* both fractional */
1549 tuple_fraction = Min(tuple_fraction, limit_fraction);
1554 /* no info from caller, just use limit */
1555 tuple_fraction = limit_fraction;
1558 else if (*offset_est != 0 && tuple_fraction > 0.0)
1561 * We have an OFFSET but no LIMIT. This acts entirely differently
1562 * from the LIMIT case: here, we need to increase rather than decrease
1563 * the caller's tuple_fraction, because the OFFSET acts to cause more
1564 * tuples to be fetched instead of fewer. This only matters if we got
1565 * a tuple_fraction > 0, however.
1567 * As above, use 10% if OFFSET is present but unestimatable.
1569 if (*offset_est < 0)
1570 limit_fraction = 0.10;
1572 limit_fraction = (double) *offset_est;
1575 * If we have absolute counts from both caller and OFFSET, add them
1576 * together; likewise if they are both fractional. If one is
1577 * fractional and the other absolute, we want to take the larger, and
1578 * we heuristically assume that's the fractional one.
1580 if (tuple_fraction >= 1.0)
1582 if (limit_fraction >= 1.0)
1584 /* both absolute, so add them together */
1585 tuple_fraction += limit_fraction;
1589 /* caller absolute, limit fractional; use limit */
1590 tuple_fraction = limit_fraction;
1595 if (limit_fraction >= 1.0)
1597 /* caller fractional, limit absolute; use caller's value */
1601 /* both fractional, so add them together */
1602 tuple_fraction += limit_fraction;
1603 if (tuple_fraction >= 1.0)
1604 tuple_fraction = 0.0; /* assume fetch all */
1609 return tuple_fraction;
1614 * preprocess_groupclause - do preparatory work on GROUP BY clause
1616 * The idea here is to adjust the ordering of the GROUP BY elements
1617 * (which in itself is semantically insignificant) to match ORDER BY,
1618 * thereby allowing a single sort operation to both implement the ORDER BY
1619 * requirement and set up for a Unique step that implements GROUP BY.
1621 * In principle it might be interesting to consider other orderings of the
1622 * GROUP BY elements, which could match the sort ordering of other
1623 * possible plans (eg an indexscan) and thereby reduce cost. We don't
1624 * bother with that, though. Hashed grouping will frequently win anyway.
1626 * Note: we need no comparable processing of the distinctClause because
1627 * the parser already enforced that that matches ORDER BY.
1630 preprocess_groupclause(PlannerInfo *root)
1632 Query *parse = root->parse;
1633 List *new_groupclause;
1638 /* If no ORDER BY, nothing useful to do here */
1639 if (parse->sortClause == NIL)
1643 * Scan the ORDER BY clause and construct a list of matching GROUP BY
1644 * items, but only as far as we can make a matching prefix.
1646 * This code assumes that the sortClause contains no duplicate items.
1648 new_groupclause = NIL;
1649 foreach(sl, parse->sortClause)
1651 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
1653 foreach(gl, parse->groupClause)
1655 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1659 new_groupclause = lappend(new_groupclause, gc);
1664 break; /* no match, so stop scanning */
1667 /* Did we match all of the ORDER BY list, or just some of it? */
1668 partial_match = (sl != NULL);
1670 /* If no match at all, no point in reordering GROUP BY */
1671 if (new_groupclause == NIL)
1675 * Add any remaining GROUP BY items to the new list, but only if we
1676 * were able to make a complete match. In other words, we only
1677 * rearrange the GROUP BY list if the result is that one list is a
1678 * prefix of the other --- otherwise there's no possibility of a
1679 * common sort. Also, give up if there are any non-sortable GROUP BY
1680 * items, since then there's no hope anyway.
1682 foreach(gl, parse->groupClause)
1684 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1686 if (list_member_ptr(new_groupclause, gc))
1687 continue; /* it matched an ORDER BY item */
1689 return; /* give up, no common sort possible */
1690 if (!OidIsValid(gc->sortop))
1691 return; /* give up, GROUP BY can't be sorted */
1692 new_groupclause = lappend(new_groupclause, gc);
1695 /* Success --- install the rearranged GROUP BY list */
1696 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
1697 parse->groupClause = new_groupclause;
1701 * choose_hashed_grouping - should we use hashed grouping?
1703 * Note: this is only applied when both alternatives are actually feasible.
1706 choose_hashed_grouping(PlannerInfo *root,
1707 double tuple_fraction, double limit_tuples,
1708 Path *cheapest_path, Path *sorted_path,
1709 double dNumGroups, AggClauseCounts *agg_counts)
1711 int numGroupCols = list_length(root->parse->groupClause);
1712 double cheapest_path_rows;
1713 int cheapest_path_width;
1715 List *target_pathkeys;
1716 List *current_pathkeys;
1720 /* Prefer sorting when enable_hashagg is off */
1721 if (!enable_hashagg)
1725 * Don't do it if it doesn't look like the hashtable will fit into
1728 * Beware here of the possibility that cheapest_path->parent is NULL. This
1729 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1731 if (cheapest_path->parent)
1733 cheapest_path_rows = cheapest_path->parent->rows;
1734 cheapest_path_width = cheapest_path->parent->width;
1738 cheapest_path_rows = 1; /* assume non-set result */
1739 cheapest_path_width = 100; /* arbitrary */
1742 /* Estimate per-hash-entry space at tuple width... */
1743 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1744 /* plus space for pass-by-ref transition values... */
1745 hashentrysize += agg_counts->transitionSpace;
1746 /* plus the per-hash-entry overhead */
1747 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1749 if (hashentrysize * dNumGroups > work_mem * 1024L)
1753 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
1754 * DISTINCT and ORDER BY as the assumed required output sort order.
1755 * This is an oversimplification because the DISTINCT might get
1756 * implemented via hashing, but it's not clear that the case is common
1757 * enough (or that our estimates are good enough) to justify trying to
1760 if (list_length(root->distinct_pathkeys) >
1761 list_length(root->sort_pathkeys))
1762 target_pathkeys = root->distinct_pathkeys;
1764 target_pathkeys = root->sort_pathkeys;
1767 * See if the estimated cost is no more than doing it the other way. While
1768 * avoiding the need for sorted input is usually a win, the fact that the
1769 * output won't be sorted may be a loss; so we need to do an actual cost
1772 * We need to consider cheapest_path + hashagg [+ final sort] versus
1773 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1774 * presorted_path + group or agg [+ final sort] where brackets indicate a
1775 * step that may not be needed. We assume query_planner() will have
1776 * returned a presorted path only if it's a winner compared to
1777 * cheapest_path for this purpose.
1779 * These path variables are dummies that just hold cost fields; we don't
1780 * make actual Paths for these steps.
1782 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1783 numGroupCols, dNumGroups,
1784 cheapest_path->startup_cost, cheapest_path->total_cost,
1785 cheapest_path_rows);
1786 /* Result of hashed agg is always unsorted */
1787 if (target_pathkeys)
1788 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
1789 dNumGroups, cheapest_path_width, limit_tuples);
1793 sorted_p.startup_cost = sorted_path->startup_cost;
1794 sorted_p.total_cost = sorted_path->total_cost;
1795 current_pathkeys = sorted_path->pathkeys;
1799 sorted_p.startup_cost = cheapest_path->startup_cost;
1800 sorted_p.total_cost = cheapest_path->total_cost;
1801 current_pathkeys = cheapest_path->pathkeys;
1803 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1805 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1806 cheapest_path_rows, cheapest_path_width, -1.0);
1807 current_pathkeys = root->group_pathkeys;
1810 if (root->parse->hasAggs)
1811 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1812 numGroupCols, dNumGroups,
1813 sorted_p.startup_cost, sorted_p.total_cost,
1814 cheapest_path_rows);
1816 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1817 sorted_p.startup_cost, sorted_p.total_cost,
1818 cheapest_path_rows);
1819 /* The Agg or Group node will preserve ordering */
1820 if (target_pathkeys &&
1821 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
1822 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
1823 dNumGroups, cheapest_path_width, limit_tuples);
1826 * Now make the decision using the top-level tuple fraction. First we
1827 * have to convert an absolute count (LIMIT) into fractional form.
1829 if (tuple_fraction >= 1.0)
1830 tuple_fraction /= dNumGroups;
1832 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1833 tuple_fraction) < 0)
1835 /* Hashed is cheaper, so use it */
1842 * choose_hashed_distinct - should we use hashing for DISTINCT?
1844 * This is fairly similar to choose_hashed_grouping, but there are enough
1845 * differences that it doesn't seem worth trying to unify the two functions.
1847 * But note that making the two choices independently is a bit bogus in
1848 * itself. If the two could be combined into a single choice operation
1849 * it'd probably be better, but that seems far too unwieldy to be practical,
1850 * especially considering that the combination of GROUP BY and DISTINCT
1851 * isn't very common in real queries. By separating them, we are giving
1852 * extra preference to using a sorting implementation when a common sort key
1853 * is available ... and that's not necessarily wrong anyway.
1855 * Note: this is only applied when both alternatives are actually feasible.
1858 choose_hashed_distinct(PlannerInfo *root,
1859 Plan *input_plan, List *input_pathkeys,
1860 double tuple_fraction, double limit_tuples,
1861 double dNumDistinctRows)
1863 int numDistinctCols = list_length(root->parse->distinctClause);
1865 List *current_pathkeys;
1866 List *needed_pathkeys;
1870 /* Prefer sorting when enable_hashagg is off */
1871 if (!enable_hashagg)
1875 * Don't do it if it doesn't look like the hashtable will fit into
1878 hashentrysize = MAXALIGN(input_plan->plan_width) + MAXALIGN(sizeof(MinimalTupleData));
1880 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
1884 * See if the estimated cost is no more than doing it the other way. While
1885 * avoiding the need for sorted input is usually a win, the fact that the
1886 * output won't be sorted may be a loss; so we need to do an actual cost
1889 * We need to consider input_plan + hashagg [+ final sort] versus
1890 * input_plan [+ sort] + group [+ final sort] where brackets indicate
1891 * a step that may not be needed.
1893 * These path variables are dummies that just hold cost fields; we don't
1894 * make actual Paths for these steps.
1896 cost_agg(&hashed_p, root, AGG_HASHED, 0,
1897 numDistinctCols, dNumDistinctRows,
1898 input_plan->startup_cost, input_plan->total_cost,
1899 input_plan->plan_rows);
1901 * Result of hashed agg is always unsorted, so if ORDER BY is present
1902 * we need to charge for the final sort.
1904 if (root->parse->sortClause)
1905 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1906 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1909 * Now for the GROUP case. See comments in grouping_planner about the
1910 * sorting choices here --- this code should match that code.
1912 sorted_p.startup_cost = input_plan->startup_cost;
1913 sorted_p.total_cost = input_plan->total_cost;
1914 current_pathkeys = input_pathkeys;
1915 if (root->parse->hasDistinctOn &&
1916 list_length(root->distinct_pathkeys) <
1917 list_length(root->sort_pathkeys))
1918 needed_pathkeys = root->sort_pathkeys;
1920 needed_pathkeys = root->distinct_pathkeys;
1921 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1923 if (list_length(root->distinct_pathkeys) >=
1924 list_length(root->sort_pathkeys))
1925 current_pathkeys = root->distinct_pathkeys;
1927 current_pathkeys = root->sort_pathkeys;
1928 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
1929 input_plan->plan_rows, input_plan->plan_width, -1.0);
1931 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
1932 sorted_p.startup_cost, sorted_p.total_cost,
1933 input_plan->plan_rows);
1934 if (root->parse->sortClause &&
1935 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1936 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1937 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1940 * Now make the decision using the top-level tuple fraction. First we
1941 * have to convert an absolute count (LIMIT) into fractional form.
1943 if (tuple_fraction >= 1.0)
1944 tuple_fraction /= dNumDistinctRows;
1946 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1947 tuple_fraction) < 0)
1949 /* Hashed is cheaper, so use it */
1956 * make_subplanTargetList
1957 * Generate appropriate target list when grouping is required.
1959 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1960 * above the result of query_planner, we typically want to pass a different
1961 * target list to query_planner than the outer plan nodes should have.
1962 * This routine generates the correct target list for the subplan.
1964 * The initial target list passed from the parser already contains entries
1965 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1966 * for variables used only in HAVING clauses; so we need to add those
1967 * variables to the subplan target list. Also, we flatten all expressions
1968 * except GROUP BY items into their component variables; the other expressions
1969 * will be computed by the inserted nodes rather than by the subplan.
1970 * For example, given a query like
1971 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1972 * we want to pass this targetlist to the subplan:
1974 * where the a+b target will be used by the Sort/Group steps, and the
1975 * other targets will be used for computing the final results. (In the
1976 * above example we could theoretically suppress the a and b targets and
1977 * pass down only c,d,a+b, but it's not really worth the trouble to
1978 * eliminate simple var references from the subplan. We will avoid doing
1979 * the extra computation to recompute a+b at the outer level; see
1980 * fix_upper_expr() in setrefs.c.)
1982 * If we are grouping or aggregating, *and* there are no non-Var grouping
1983 * expressions, then the returned tlist is effectively dummy; we do not
1984 * need to force it to be evaluated, because all the Vars it contains
1985 * should be present in the output of query_planner anyway.
1987 * 'tlist' is the query's target list.
1988 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1989 * expressions (if there are any) in the subplan's target list.
1990 * 'need_tlist_eval' is set true if we really need to evaluate the
1993 * The result is the targetlist to be passed to the subplan.
1997 make_subplanTargetList(PlannerInfo *root,
1999 AttrNumber **groupColIdx,
2000 bool *need_tlist_eval)
2002 Query *parse = root->parse;
2007 *groupColIdx = NULL;
2010 * If we're not grouping or aggregating, there's nothing to do here;
2011 * query_planner should receive the unmodified target list.
2013 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
2015 *need_tlist_eval = true;
2020 * Otherwise, start with a "flattened" tlist (having just the vars
2021 * mentioned in the targetlist and HAVING qual --- but not upper-level
2022 * Vars; they will be replaced by Params later on).
2024 sub_tlist = flatten_tlist(tlist);
2025 extravars = pull_var_clause(parse->havingQual, false);
2026 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2027 list_free(extravars);
2028 *need_tlist_eval = false; /* only eval if not flat tlist */
2031 * If grouping, create sub_tlist entries for all GROUP BY expressions
2032 * (GROUP BY items that are simple Vars should be in the list already),
2033 * and make an array showing where the group columns are in the sub_tlist.
2035 numCols = list_length(parse->groupClause);
2039 AttrNumber *grpColIdx;
2042 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2043 *groupColIdx = grpColIdx;
2045 foreach(gl, parse->groupClause)
2047 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2048 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2049 TargetEntry *te = NULL;
2052 * Find or make a matching sub_tlist entry. If the groupexpr
2053 * isn't a Var, no point in searching. (Note that the parser
2054 * won't make multiple groupClause entries for the same TLE.)
2056 if (groupexpr && IsA(groupexpr, Var))
2060 foreach(sl, sub_tlist)
2062 TargetEntry *lte = (TargetEntry *) lfirst(sl);
2064 if (equal(groupexpr, lte->expr))
2073 te = makeTargetEntry((Expr *) groupexpr,
2074 list_length(sub_tlist) + 1,
2077 sub_tlist = lappend(sub_tlist, te);
2078 *need_tlist_eval = true; /* it's not flat anymore */
2081 /* and save its resno */
2082 grpColIdx[keyno++] = te->resno;
2090 * locate_grouping_columns
2091 * Locate grouping columns in the tlist chosen by query_planner.
2093 * This is only needed if we don't use the sub_tlist chosen by
2094 * make_subplanTargetList. We have to forget the column indexes found
2095 * by that routine and re-locate the grouping vars in the real sub_tlist.
2098 locate_grouping_columns(PlannerInfo *root,
2101 AttrNumber *groupColIdx)
2107 * No work unless grouping.
2109 if (!root->parse->groupClause)
2111 Assert(groupColIdx == NULL);
2114 Assert(groupColIdx != NULL);
2116 foreach(gl, root->parse->groupClause)
2118 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2119 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2120 TargetEntry *te = NULL;
2123 foreach(sl, sub_tlist)
2125 te = (TargetEntry *) lfirst(sl);
2126 if (equal(groupexpr, te->expr))
2130 elog(ERROR, "failed to locate grouping columns");
2132 groupColIdx[keyno++] = te->resno;
2137 * postprocess_setop_tlist
2138 * Fix up targetlist returned by plan_set_operations().
2140 * We need to transpose sort key info from the orig_tlist into new_tlist.
2141 * NOTE: this would not be good enough if we supported resjunk sort keys
2142 * for results of set operations --- then, we'd need to project a whole
2143 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2144 * find any resjunk columns in orig_tlist.
2147 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2150 ListCell *orig_tlist_item = list_head(orig_tlist);
2152 foreach(l, new_tlist)
2154 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2155 TargetEntry *orig_tle;
2157 /* ignore resjunk columns in setop result */
2158 if (new_tle->resjunk)
2161 Assert(orig_tlist_item != NULL);
2162 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2163 orig_tlist_item = lnext(orig_tlist_item);
2164 if (orig_tle->resjunk) /* should not happen */
2165 elog(ERROR, "resjunk output columns are not implemented");
2166 Assert(new_tle->resno == orig_tle->resno);
2167 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2169 if (orig_tlist_item != NULL)
2170 elog(ERROR, "resjunk output columns are not implemented");