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 1.240 2008/08/07 01:11:50 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_ININFO 5
59 #define EXPRKIND_APPINFO 6
62 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
63 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
64 static Plan *inheritance_planner(PlannerInfo *root);
65 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
66 static bool is_dummy_plan(Plan *plan);
67 static double preprocess_limit(PlannerInfo *root,
68 double tuple_fraction,
69 int64 *offset_est, int64 *count_est);
70 static void preprocess_groupclause(PlannerInfo *root);
71 static bool choose_hashed_grouping(PlannerInfo *root,
72 double tuple_fraction, double limit_tuples,
73 Path *cheapest_path, Path *sorted_path,
74 double dNumGroups, AggClauseCounts *agg_counts);
75 static bool choose_hashed_distinct(PlannerInfo *root,
76 Plan *input_plan, List *input_pathkeys,
77 double tuple_fraction, double limit_tuples,
78 double dNumDistinctRows);
79 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
80 AttrNumber **groupColIdx, bool *need_tlist_eval);
81 static void locate_grouping_columns(PlannerInfo *root,
84 AttrNumber *groupColIdx);
85 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
88 /*****************************************************************************
90 * Query optimizer entry point
92 * To support loadable plugins that monitor or modify planner behavior,
93 * we provide a hook variable that lets a plugin get control before and
94 * after the standard planning process. The plugin would normally call
97 * Note to plugin authors: standard_planner() scribbles on its Query input,
98 * so you'd better copy that data structure if you want to plan more than once.
100 *****************************************************************************/
102 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
107 result = (*planner_hook) (parse, cursorOptions, boundParams);
109 result = standard_planner(parse, cursorOptions, boundParams);
114 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
118 double tuple_fraction;
124 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
125 if (parse->utilityStmt &&
126 IsA(parse->utilityStmt, DeclareCursorStmt))
127 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
130 * Set up global state for this planner invocation. This data is needed
131 * across all levels of sub-Query that might exist in the given command,
132 * so we keep it in a separate struct that's linked to by each per-Query
135 glob = makeNode(PlannerGlobal);
137 glob->boundParams = boundParams;
138 glob->paramlist = NIL;
139 glob->subplans = NIL;
140 glob->subrtables = NIL;
141 glob->rewindPlanIDs = NULL;
142 glob->finalrtable = NIL;
143 glob->relationOids = 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->nParamExec = list_length(glob->paramlist);
223 /*--------------------
225 * Invokes the planner on a subquery. We recurse to here for each
226 * sub-SELECT found in the query tree.
228 * glob is the global state for the current planner run.
229 * parse is the querytree produced by the parser & rewriter.
230 * level is the current recursion depth (1 at the top-level Query).
231 * tuple_fraction is the fraction of tuples we expect will be retrieved.
232 * tuple_fraction is interpreted as explained for grouping_planner, below.
234 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
235 * among other things this tells the output sort ordering of the plan.
237 * Basically, this routine does the stuff that should only be done once
238 * per Query object. It then calls grouping_planner. At one time,
239 * grouping_planner could be invoked recursively on the same Query object;
240 * that's not currently true, but we keep the separation between the two
241 * routines anyway, in case we need it again someday.
243 * subquery_planner will be called recursively to handle sub-Query nodes
244 * found within the query's expressions and rangetable.
246 * Returns a query plan.
247 *--------------------
250 subquery_planner(PlannerGlobal *glob, Query *parse,
251 Index level, double tuple_fraction,
252 PlannerInfo **subroot)
254 int num_old_subplans = list_length(glob->subplans);
260 /* Create a PlannerInfo data structure for this subquery */
261 root = makeNode(PlannerInfo);
264 root->query_level = level;
265 root->planner_cxt = CurrentMemoryContext;
266 root->init_plans = NIL;
267 root->eq_classes = NIL;
268 root->in_info_list = NIL;
269 root->append_rel_list = NIL;
272 * Look for IN clauses at the top level of WHERE, and transform them into
273 * joins. Note that this step only handles IN clauses originally at top
274 * level of WHERE; if we pull up any subqueries below, their INs are
275 * processed just before pulling them up.
277 if (parse->hasSubLinks)
278 parse->jointree->quals = pull_up_IN_clauses(root,
279 parse->jointree->quals);
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 IN clauses.
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() and some
299 * other processing. This must be done after we have done
300 * pull_up_subqueries, of course.
302 * Note: if reduce_outer_joins manages to eliminate all outer joins,
303 * root->hasOuterJoins is not reset currently. This is OK since its
304 * purpose is merely to suppress unnecessary processing in simple cases.
306 root->hasJoinRTEs = false;
307 root->hasOuterJoins = false;
308 foreach(l, parse->rtable)
310 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
312 if (rte->rtekind == RTE_JOIN)
314 root->hasJoinRTEs = true;
315 if (IS_OUTER_JOIN(rte->jointype))
317 root->hasOuterJoins = true;
318 /* Can quit scanning once we find an outer join */
325 * Expand any rangetable entries that are inheritance sets into "append
326 * relations". This can add entries to the rangetable, but they must be
327 * plain base relations not joins, so it's OK (and marginally more
328 * efficient) to do it after checking for join RTEs. We must do it after
329 * pulling up subqueries, else we'd fail to handle inherited tables in
332 expand_inherited_tables(root);
335 * Set hasHavingQual to remember if HAVING clause is present. Needed
336 * because preprocess_expression will reduce a constant-true condition to
337 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
339 root->hasHavingQual = (parse->havingQual != NULL);
341 /* Clear this flag; might get set in distribute_qual_to_rels */
342 root->hasPseudoConstantQuals = false;
345 * Do expression preprocessing on targetlist and quals.
347 parse->targetList = (List *)
348 preprocess_expression(root, (Node *) parse->targetList,
351 parse->returningList = (List *)
352 preprocess_expression(root, (Node *) parse->returningList,
355 preprocess_qual_conditions(root, (Node *) parse->jointree);
357 parse->havingQual = preprocess_expression(root, parse->havingQual,
360 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
362 parse->limitCount = preprocess_expression(root, parse->limitCount,
365 root->in_info_list = (List *)
366 preprocess_expression(root, (Node *) root->in_info_list,
368 root->append_rel_list = (List *)
369 preprocess_expression(root, (Node *) root->append_rel_list,
372 /* Also need to preprocess expressions for function and values RTEs */
373 foreach(l, parse->rtable)
375 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
377 if (rte->rtekind == RTE_FUNCTION)
378 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
380 else if (rte->rtekind == RTE_VALUES)
381 rte->values_lists = (List *)
382 preprocess_expression(root, (Node *) rte->values_lists,
387 * In some cases we may want to transfer a HAVING clause into WHERE. We
388 * cannot do so if the HAVING clause contains aggregates (obviously) or
389 * volatile functions (since a HAVING clause is supposed to be executed
390 * only once per group). Also, it may be that the clause is so expensive
391 * to execute that we're better off doing it only once per group, despite
392 * the loss of selectivity. This is hard to estimate short of doing the
393 * entire planning process twice, so we use a heuristic: clauses
394 * containing subplans are left in HAVING. Otherwise, we move or copy the
395 * HAVING clause into WHERE, in hopes of eliminating tuples before
396 * aggregation instead of after.
398 * If the query has explicit grouping then we can simply move such a
399 * clause into WHERE; any group that fails the clause will not be in the
400 * output because none of its tuples will reach the grouping or
401 * aggregation stage. Otherwise we must have a degenerate (variable-free)
402 * HAVING clause, which we put in WHERE so that query_planner() can use it
403 * in a gating Result node, but also keep in HAVING to ensure that we
404 * don't emit a bogus aggregated row. (This could be done better, but it
405 * seems not worth optimizing.)
407 * Note that both havingQual and parse->jointree->quals are in
408 * implicitly-ANDed-list form at this point, even though they are declared
412 foreach(l, (List *) parse->havingQual)
414 Node *havingclause = (Node *) lfirst(l);
416 if (contain_agg_clause(havingclause) ||
417 contain_volatile_functions(havingclause) ||
418 contain_subplans(havingclause))
420 /* keep it in HAVING */
421 newHaving = lappend(newHaving, havingclause);
423 else if (parse->groupClause)
425 /* move it to WHERE */
426 parse->jointree->quals = (Node *)
427 lappend((List *) parse->jointree->quals, havingclause);
431 /* put a copy in WHERE, keep it in HAVING */
432 parse->jointree->quals = (Node *)
433 lappend((List *) parse->jointree->quals,
434 copyObject(havingclause));
435 newHaving = lappend(newHaving, havingclause);
438 parse->havingQual = (Node *) newHaving;
441 * If we have any outer joins, try to reduce them to plain inner joins.
442 * This step is most easily done after we've done expression
445 if (root->hasOuterJoins)
446 reduce_outer_joins(root);
449 * Do the main planning. If we have an inherited target relation, that
450 * needs special processing, else go straight to grouping_planner.
452 if (parse->resultRelation &&
453 rt_fetch(parse->resultRelation, parse->rtable)->inh)
454 plan = inheritance_planner(root);
456 plan = grouping_planner(root, tuple_fraction);
459 * If any subplans were generated, or if we're inside a subplan, build
460 * initPlan list and extParam/allParam sets for plan nodes, and attach the
461 * initPlans to the top plan node.
463 if (list_length(glob->subplans) != num_old_subplans ||
464 root->query_level > 1)
465 SS_finalize_plan(root, plan, true);
467 /* Return internal info if caller wants it */
475 * preprocess_expression
476 * Do subquery_planner's preprocessing work for an expression,
477 * which can be a targetlist, a WHERE clause (including JOIN/ON
478 * conditions), or a HAVING clause.
481 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
484 * Fall out quickly if expression is empty. This occurs often enough to
485 * be worth checking. Note that null->null is the correct conversion for
486 * implicit-AND result format, too.
492 * If the query has any join RTEs, replace join alias variables with
493 * base-relation variables. We must do this before sublink processing,
494 * else sublinks expanded out from join aliases wouldn't get processed. We
495 * can skip it in VALUES lists, however, since they can't contain any Vars
498 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
499 expr = flatten_join_alias_vars(root, expr);
502 * Simplify constant expressions.
504 * Note: this also flattens nested AND and OR expressions into N-argument
505 * form. All processing of a qual expression after this point must be
506 * careful to maintain AND/OR flatness --- that is, do not generate a tree
507 * with AND directly under AND, nor OR directly under OR.
509 * Because this is a relatively expensive process, we skip it when the
510 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
511 * expression will only be evaluated once anyway, so no point in
512 * pre-simplifying; we can't execute it any faster than the executor can,
513 * and we will waste cycles copying the tree. Notice however that we
514 * still must do it for quals (to get AND/OR flatness); and if we are in a
515 * subquery we should not assume it will be done only once.
517 * For VALUES lists we never do this at all, again on the grounds that we
518 * should optimize for one-time evaluation.
520 if (kind != EXPRKIND_VALUES &&
521 (root->parse->jointree->fromlist != NIL ||
522 kind == EXPRKIND_QUAL ||
523 root->query_level > 1))
524 expr = eval_const_expressions(root, expr);
527 * If it's a qual or havingQual, canonicalize it.
529 if (kind == EXPRKIND_QUAL)
531 expr = (Node *) canonicalize_qual((Expr *) expr);
533 #ifdef OPTIMIZER_DEBUG
534 printf("After canonicalize_qual()\n");
539 /* Expand SubLinks to SubPlans */
540 if (root->parse->hasSubLinks)
541 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
544 * XXX do not insert anything here unless you have grokked the comments in
545 * SS_replace_correlation_vars ...
548 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
549 if (root->query_level > 1)
550 expr = SS_replace_correlation_vars(root, expr);
553 * If it's a qual or havingQual, convert it to implicit-AND format. (We
554 * don't want to do this before eval_const_expressions, since the latter
555 * would be unable to simplify a top-level AND correctly. Also,
556 * SS_process_sublinks expects explicit-AND format.)
558 if (kind == EXPRKIND_QUAL)
559 expr = (Node *) make_ands_implicit((Expr *) expr);
565 * preprocess_qual_conditions
566 * Recursively scan the query's jointree and do subquery_planner's
567 * preprocessing work on each qual condition found therein.
570 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
574 if (IsA(jtnode, RangeTblRef))
576 /* nothing to do here */
578 else if (IsA(jtnode, FromExpr))
580 FromExpr *f = (FromExpr *) jtnode;
583 foreach(l, f->fromlist)
584 preprocess_qual_conditions(root, lfirst(l));
586 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
588 else if (IsA(jtnode, JoinExpr))
590 JoinExpr *j = (JoinExpr *) jtnode;
592 preprocess_qual_conditions(root, j->larg);
593 preprocess_qual_conditions(root, j->rarg);
595 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
598 elog(ERROR, "unrecognized node type: %d",
599 (int) nodeTag(jtnode));
603 * inheritance_planner
604 * Generate a plan in the case where the result relation is an
607 * We have to handle this case differently from cases where a source relation
608 * is an inheritance set. Source inheritance is expanded at the bottom of the
609 * plan tree (see allpaths.c), but target inheritance has to be expanded at
610 * the top. The reason is that for UPDATE, each target relation needs a
611 * different targetlist matching its own column set. Also, for both UPDATE
612 * and DELETE, the executor needs the Append plan node at the top, else it
613 * can't keep track of which table is the current target table. Fortunately,
614 * the UPDATE/DELETE target can never be the nullable side of an outer join,
615 * so it's OK to generate the plan this way.
617 * Returns a query plan.
620 inheritance_planner(PlannerInfo *root)
622 Query *parse = root->parse;
623 int parentRTindex = parse->resultRelation;
624 List *subplans = NIL;
625 List *resultRelations = NIL;
626 List *returningLists = NIL;
632 foreach(l, root->append_rel_list)
634 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
637 /* append_rel_list contains all append rels; ignore others */
638 if (appinfo->parent_relid != parentRTindex)
642 * Generate modified query with this rel as target. We have to be
643 * prepared to translate varnos in in_info_list as well as in the
646 memcpy(&subroot, root, sizeof(PlannerInfo));
647 subroot.parse = (Query *)
648 adjust_appendrel_attrs((Node *) parse,
650 subroot.in_info_list = (List *)
651 adjust_appendrel_attrs((Node *) root->in_info_list,
653 subroot.init_plans = NIL;
654 /* There shouldn't be any OJ info to translate, as yet */
655 Assert(subroot.oj_info_list == NIL);
658 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
661 * If this child rel was excluded by constraint exclusion, exclude it
664 if (is_dummy_plan(subplan))
667 /* Save rtable and tlist from first rel for use below */
670 rtable = subroot.parse->rtable;
671 tlist = subplan->targetlist;
674 subplans = lappend(subplans, subplan);
676 /* Make sure any initplans from this rel get into the outer list */
677 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
679 /* Build target-relations list for the executor */
680 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
682 /* Build list of per-relation RETURNING targetlists */
683 if (parse->returningList)
685 Assert(list_length(subroot.returningLists) == 1);
686 returningLists = list_concat(returningLists,
687 subroot.returningLists);
691 root->resultRelations = resultRelations;
692 root->returningLists = returningLists;
694 /* Mark result as unordered (probably unnecessary) */
695 root->query_pathkeys = NIL;
698 * If we managed to exclude every child rel, return a dummy plan
702 root->resultRelations = list_make1_int(parentRTindex);
703 /* although dummy, it must have a valid tlist for executor */
704 tlist = preprocess_targetlist(root, parse->targetList);
705 return (Plan *) make_result(root,
707 (Node *) list_make1(makeBoolConst(false,
713 * Planning might have modified the rangetable, due to changes of the
714 * Query structures inside subquery RTEs. We have to ensure that this
715 * gets propagated back to the master copy. But can't do this until we
716 * are done planning, because all the calls to grouping_planner need
717 * virgin sub-Queries to work from. (We are effectively assuming that
718 * sub-Queries will get planned identically each time, or at least that
719 * the impacts on their rangetables will be the same each time.)
721 * XXX should clean this up someday
723 parse->rtable = rtable;
725 /* Suppress Append if there's only one surviving child rel */
726 if (list_length(subplans) == 1)
727 return (Plan *) linitial(subplans);
729 return (Plan *) make_append(subplans, true, tlist);
732 /*--------------------
734 * Perform planning steps related to grouping, aggregation, etc.
735 * This primarily means adding top-level processing to the basic
736 * query plan produced by query_planner.
738 * tuple_fraction is the fraction of tuples we expect will be retrieved
740 * tuple_fraction is interpreted as follows:
741 * 0: expect all tuples to be retrieved (normal case)
742 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
743 * from the plan to be retrieved
744 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
745 * expected to be retrieved (ie, a LIMIT specification)
747 * Returns a query plan. Also, root->query_pathkeys is returned as the
748 * actual output ordering of the plan (in pathkey format).
749 *--------------------
752 grouping_planner(PlannerInfo *root, double tuple_fraction)
754 Query *parse = root->parse;
755 List *tlist = parse->targetList;
756 int64 offset_est = 0;
758 double limit_tuples = -1.0;
760 List *current_pathkeys;
761 double dNumGroups = 0;
763 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
764 if (parse->limitCount || parse->limitOffset)
766 tuple_fraction = preprocess_limit(root, tuple_fraction,
767 &offset_est, &count_est);
770 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
771 * estimate the effects of using a bounded sort.
773 if (count_est > 0 && offset_est >= 0)
774 limit_tuples = (double) count_est + (double) offset_est;
777 if (parse->setOperations)
779 List *set_sortclauses;
782 * If there's a top-level ORDER BY, assume we have to fetch all the
783 * tuples. This might be too simplistic given all the hackery below
784 * to possibly avoid the sort; but the odds of accurate estimates
785 * here are pretty low anyway.
787 if (parse->sortClause)
788 tuple_fraction = 0.0;
791 * Construct the plan for set operations. The result will not need
792 * any work except perhaps a top-level sort and/or LIMIT.
794 result_plan = plan_set_operations(root, tuple_fraction,
798 * Calculate pathkeys representing the sort order (if any) of the set
799 * operation's result. We have to do this before overwriting the sort
802 current_pathkeys = make_pathkeys_for_sortclauses(root,
804 result_plan->targetlist,
808 * We should not need to call preprocess_targetlist, since we must be
809 * in a SELECT query node. Instead, use the targetlist returned by
810 * plan_set_operations (since this tells whether it returned any
811 * resjunk columns!), and transfer any sort key information from the
814 Assert(parse->commandType == CMD_SELECT);
816 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
820 * Can't handle FOR UPDATE/SHARE here (parser should have checked
821 * already, but let's make sure).
825 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
826 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
829 * Calculate pathkeys that represent result ordering requirements
831 Assert(parse->distinctClause == NIL);
832 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
839 /* No set operations, do regular planning */
841 AttrNumber *groupColIdx = NULL;
842 bool need_tlist_eval = true;
848 AggClauseCounts agg_counts;
850 bool use_hashed_grouping = false;
852 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
854 /* Preprocess GROUP BY clause, if any */
855 if (parse->groupClause)
856 preprocess_groupclause(root);
857 numGroupCols = list_length(parse->groupClause);
859 /* Preprocess targetlist */
860 tlist = preprocess_targetlist(root, tlist);
863 * Generate appropriate target list for subplan; may be different from
864 * tlist if grouping or aggregation is needed.
866 sub_tlist = make_subplanTargetList(root, tlist,
867 &groupColIdx, &need_tlist_eval);
870 * Calculate pathkeys that represent grouping/ordering requirements.
871 * Stash them in PlannerInfo so that query_planner can canonicalize
872 * them after EquivalenceClasses have been formed. The sortClause
873 * is certainly sort-able, but GROUP BY and DISTINCT might not be,
874 * in which case we just leave their pathkeys empty.
876 if (parse->groupClause &&
877 grouping_is_sortable(parse->groupClause))
878 root->group_pathkeys =
879 make_pathkeys_for_sortclauses(root,
884 root->group_pathkeys = NIL;
886 if (parse->distinctClause &&
887 grouping_is_sortable(parse->distinctClause))
888 root->distinct_pathkeys =
889 make_pathkeys_for_sortclauses(root,
890 parse->distinctClause,
894 root->distinct_pathkeys = NIL;
896 root->sort_pathkeys =
897 make_pathkeys_for_sortclauses(root,
903 * Will need actual number of aggregates for estimating costs.
905 * Note: we do not attempt to detect duplicate aggregates here; a
906 * somewhat-overestimated count is okay for our present purposes.
908 * Note: think not that we can turn off hasAggs if we find no aggs. It
909 * is possible for constant-expression simplification to remove all
910 * explicit references to aggs, but we still have to follow the
911 * aggregate semantics (eg, producing only one output row).
915 count_agg_clauses((Node *) tlist, &agg_counts);
916 count_agg_clauses(parse->havingQual, &agg_counts);
920 * Figure out whether we want a sorted result from query_planner.
922 * If we have a sortable GROUP BY clause, then we want a result sorted
923 * properly for grouping. Otherwise, if there's a sortable DISTINCT
924 * clause that's more rigorous than the ORDER BY clause, we try to
925 * produce output that's sufficiently well sorted for the DISTINCT.
926 * Otherwise, if there is an ORDER BY clause, we want to sort by the
929 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
930 * superset of GROUP BY, it would be tempting to request sort by ORDER
931 * BY --- but that might just leave us failing to exploit an available
932 * sort order at all. Needs more thought. The choice for DISTINCT
933 * versus ORDER BY is much easier, since we know that the parser
934 * ensured that one is a superset of the other.
936 if (root->group_pathkeys)
937 root->query_pathkeys = root->group_pathkeys;
938 else if (list_length(root->distinct_pathkeys) >
939 list_length(root->sort_pathkeys))
940 root->query_pathkeys = root->distinct_pathkeys;
941 else if (root->sort_pathkeys)
942 root->query_pathkeys = root->sort_pathkeys;
944 root->query_pathkeys = NIL;
947 * Generate the best unsorted and presorted paths for this Query (but
948 * note there may not be any presorted path). query_planner will also
949 * estimate the number of groups in the query, and canonicalize all
952 query_planner(root, sub_tlist, tuple_fraction, limit_tuples,
953 &cheapest_path, &sorted_path, &dNumGroups);
956 * If grouping, decide whether to use sorted or hashed grouping.
958 if (parse->groupClause)
964 * Executor doesn't support hashed aggregation with DISTINCT
965 * aggregates. (Doing so would imply storing *all* the input
966 * values in the hash table, which seems like a certain loser.)
968 can_hash = (agg_counts.numDistinctAggs == 0 &&
969 grouping_is_hashable(parse->groupClause));
970 can_sort = grouping_is_sortable(parse->groupClause);
971 if (can_hash && can_sort)
973 /* we have a meaningful choice to make ... */
974 use_hashed_grouping =
975 choose_hashed_grouping(root,
976 tuple_fraction, limit_tuples,
977 cheapest_path, sorted_path,
978 dNumGroups, &agg_counts);
981 use_hashed_grouping = true;
983 use_hashed_grouping = false;
986 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
987 errmsg("could not implement GROUP BY"),
988 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
990 /* Also convert # groups to long int --- but 'ware overflow! */
991 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
995 * Select the best path. If we are doing hashed grouping, we will
996 * always read all the input tuples, so use the cheapest-total path.
997 * Otherwise, trust query_planner's decision about which to use.
999 if (use_hashed_grouping || !sorted_path)
1000 best_path = cheapest_path;
1002 best_path = sorted_path;
1005 * Check to see if it's possible to optimize MIN/MAX aggregates. If
1006 * so, we will forget all the work we did so far to choose a "regular"
1007 * path ... but we had to do it anyway to be able to tell which way is
1010 result_plan = optimize_minmax_aggregates(root,
1013 if (result_plan != NULL)
1016 * optimize_minmax_aggregates generated the full plan, with the
1017 * right tlist, and it has no sort order.
1019 current_pathkeys = NIL;
1024 * Normal case --- create a plan according to query_planner's
1027 bool need_sort_for_grouping = false;
1029 result_plan = create_plan(root, best_path);
1030 current_pathkeys = best_path->pathkeys;
1032 /* Detect if we'll need an explicit sort for grouping */
1033 if (parse->groupClause && !use_hashed_grouping &&
1034 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1036 need_sort_for_grouping = true;
1038 * Always override query_planner's tlist, so that we don't
1039 * sort useless data from a "physical" tlist.
1041 need_tlist_eval = true;
1045 * create_plan() returns a plan with just a "flat" tlist of
1046 * required Vars. Usually we need to insert the sub_tlist as the
1047 * tlist of the top plan node. However, we can skip that if we
1048 * determined that whatever query_planner chose to return will be
1051 if (need_tlist_eval)
1054 * If the top-level plan node is one that cannot do expression
1055 * evaluation, we must insert a Result node to project the
1058 if (!is_projection_capable_plan(result_plan))
1060 result_plan = (Plan *) make_result(root,
1068 * Otherwise, just replace the subplan's flat tlist with
1069 * the desired tlist.
1071 result_plan->targetlist = sub_tlist;
1075 * Also, account for the cost of evaluation of the sub_tlist.
1077 * Up to now, we have only been dealing with "flat" tlists,
1078 * containing just Vars. So their evaluation cost is zero
1079 * according to the model used by cost_qual_eval() (or if you
1080 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1081 * can avoid accounting for tlist cost throughout
1082 * query_planner() and subroutines. But now we've inserted a
1083 * tlist that might contain actual operators, sub-selects, etc
1084 * --- so we'd better account for its cost.
1086 * Below this point, any tlist eval cost for added-on nodes
1087 * should be accounted for as we create those nodes.
1088 * Presently, of the node types we can add on, only Agg and
1089 * Group project new tlists (the rest just copy their input
1090 * tuples) --- so make_agg() and make_group() are responsible
1091 * for computing the added cost.
1093 cost_qual_eval(&tlist_cost, sub_tlist, root);
1094 result_plan->startup_cost += tlist_cost.startup;
1095 result_plan->total_cost += tlist_cost.startup +
1096 tlist_cost.per_tuple * result_plan->plan_rows;
1101 * Since we're using query_planner's tlist and not the one
1102 * make_subplanTargetList calculated, we have to refigure any
1103 * grouping-column indexes make_subplanTargetList computed.
1105 locate_grouping_columns(root, tlist, result_plan->targetlist,
1110 * Insert AGG or GROUP node if needed, plus an explicit sort step
1113 * HAVING clause, if any, becomes qual of the Agg or Group node.
1115 if (use_hashed_grouping)
1117 /* Hashed aggregate plan --- no sort needed */
1118 result_plan = (Plan *) make_agg(root,
1120 (List *) parse->havingQual,
1124 extract_grouping_ops(parse->groupClause),
1128 /* Hashed aggregation produces randomly-ordered results */
1129 current_pathkeys = NIL;
1131 else if (parse->hasAggs)
1133 /* Plain aggregate plan --- sort if needed */
1134 AggStrategy aggstrategy;
1136 if (parse->groupClause)
1138 if (need_sort_for_grouping)
1140 result_plan = (Plan *)
1141 make_sort_from_groupcols(root,
1145 current_pathkeys = root->group_pathkeys;
1147 aggstrategy = AGG_SORTED;
1150 * The AGG node will not change the sort ordering of its
1151 * groups, so current_pathkeys describes the result too.
1156 aggstrategy = AGG_PLAIN;
1157 /* Result will be only one row anyway; no sort order */
1158 current_pathkeys = NIL;
1161 result_plan = (Plan *) make_agg(root,
1163 (List *) parse->havingQual,
1167 extract_grouping_ops(parse->groupClause),
1172 else if (parse->groupClause)
1175 * GROUP BY without aggregation, so insert a group node (plus
1176 * the appropriate sort node, if necessary).
1178 * Add an explicit sort if we couldn't make the path come out
1179 * the way the GROUP node needs it.
1181 if (need_sort_for_grouping)
1183 result_plan = (Plan *)
1184 make_sort_from_groupcols(root,
1188 current_pathkeys = root->group_pathkeys;
1191 result_plan = (Plan *) make_group(root,
1193 (List *) parse->havingQual,
1196 extract_grouping_ops(parse->groupClause),
1199 /* The Group node won't change sort ordering */
1201 else if (root->hasHavingQual)
1204 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1205 * This is a degenerate case in which we are supposed to emit
1206 * either 0 or 1 row depending on whether HAVING succeeds.
1207 * Furthermore, there cannot be any variables in either HAVING
1208 * or the targetlist, so we actually do not need the FROM
1209 * table at all! We can just throw away the plan-so-far and
1210 * generate a Result node. This is a sufficiently unusual
1211 * corner case that it's not worth contorting the structure of
1212 * this routine to avoid having to generate the plan in the
1215 result_plan = (Plan *) make_result(root,
1220 } /* end of non-minmax-aggregate case */
1221 } /* end of if (setOperations) */
1224 * If there is a DISTINCT clause, add the necessary node(s).
1226 if (parse->distinctClause)
1228 double dNumDistinctRows;
1229 long numDistinctRows;
1230 bool use_hashed_distinct;
1235 * If there was grouping or aggregation, use the current number of
1236 * rows as the estimated number of DISTINCT rows (ie, assume the
1237 * result was already mostly unique). If not, use the number of
1238 * distinct-groups calculated by query_planner.
1240 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1241 dNumDistinctRows = result_plan->plan_rows;
1243 dNumDistinctRows = dNumGroups;
1245 /* Also convert to long int --- but 'ware overflow! */
1246 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1249 * If we have a sortable DISTINCT ON clause, we always use sorting.
1250 * This enforces the expected behavior of DISTINCT ON.
1252 can_sort = grouping_is_sortable(parse->distinctClause);
1253 if (can_sort && parse->hasDistinctOn)
1254 use_hashed_distinct = false;
1257 can_hash = grouping_is_hashable(parse->distinctClause);
1258 if (can_hash && can_sort)
1260 /* we have a meaningful choice to make ... */
1261 use_hashed_distinct =
1262 choose_hashed_distinct(root,
1263 result_plan, current_pathkeys,
1264 tuple_fraction, limit_tuples,
1268 use_hashed_distinct = true;
1270 use_hashed_distinct = false;
1274 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1275 errmsg("could not implement DISTINCT"),
1276 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
1277 use_hashed_distinct = false; /* keep compiler quiet */
1281 if (use_hashed_distinct)
1283 /* Hashed aggregate plan --- no sort needed */
1284 result_plan = (Plan *) make_agg(root,
1285 result_plan->targetlist,
1288 list_length(parse->distinctClause),
1289 extract_grouping_cols(parse->distinctClause,
1290 result_plan->targetlist),
1291 extract_grouping_ops(parse->distinctClause),
1295 /* Hashed aggregation produces randomly-ordered results */
1296 current_pathkeys = NIL;
1301 * Use a Unique node to implement DISTINCT. Add an explicit sort
1302 * if we couldn't make the path come out the way the Unique node
1303 * needs it. If we do have to sort, always sort by the more
1304 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1305 * below. However, for regular DISTINCT, don't sort now if we
1306 * don't have to --- sorting afterwards will likely be cheaper,
1307 * and also has the possibility of optimizing via LIMIT. But
1308 * for DISTINCT ON, we *must* force the final sort now, else
1309 * it won't have the desired behavior.
1311 List *needed_pathkeys;
1313 if (parse->hasDistinctOn &&
1314 list_length(root->distinct_pathkeys) <
1315 list_length(root->sort_pathkeys))
1316 needed_pathkeys = root->sort_pathkeys;
1318 needed_pathkeys = root->distinct_pathkeys;
1320 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1322 if (list_length(root->distinct_pathkeys) >=
1323 list_length(root->sort_pathkeys))
1324 current_pathkeys = root->distinct_pathkeys;
1327 current_pathkeys = root->sort_pathkeys;
1328 /* Assert checks that parser didn't mess up... */
1329 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1333 result_plan = (Plan *) make_sort_from_pathkeys(root,
1339 result_plan = (Plan *) make_unique(result_plan,
1340 parse->distinctClause);
1341 result_plan->plan_rows = dNumDistinctRows;
1342 /* The Unique node won't change sort ordering */
1347 * If ORDER BY was given and we were not able to make the plan come out in
1348 * the right order, add an explicit sort step.
1350 if (parse->sortClause)
1352 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1354 result_plan = (Plan *) make_sort_from_pathkeys(root,
1356 root->sort_pathkeys,
1358 current_pathkeys = root->sort_pathkeys;
1363 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1365 if (parse->limitCount || parse->limitOffset)
1367 result_plan = (Plan *) make_limit(result_plan,
1375 * Deal with the RETURNING clause if any. It's convenient to pass the
1376 * returningList through setrefs.c now rather than at top level (if we
1377 * waited, handling inherited UPDATE/DELETE would be much harder).
1379 if (parse->returningList)
1383 Assert(parse->resultRelation);
1384 rlist = set_returning_clause_references(root->glob,
1385 parse->returningList,
1387 parse->resultRelation);
1388 root->returningLists = list_make1(rlist);
1391 root->returningLists = NIL;
1393 /* Compute result-relations list if needed */
1394 if (parse->resultRelation)
1395 root->resultRelations = list_make1_int(parse->resultRelation);
1397 root->resultRelations = NIL;
1400 * Return the actual output ordering in query_pathkeys for possible use by
1401 * an outer query level.
1403 root->query_pathkeys = current_pathkeys;
1409 * Detect whether a plan node is a "dummy" plan created when a relation
1410 * is deemed not to need scanning due to constraint exclusion.
1412 * Currently, such dummy plans are Result nodes with constant FALSE
1416 is_dummy_plan(Plan *plan)
1418 if (IsA(plan, Result))
1420 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1422 if (list_length(rcqual) == 1)
1424 Const *constqual = (Const *) linitial(rcqual);
1426 if (constqual && IsA(constqual, Const))
1428 if (!constqual->constisnull &&
1429 !DatumGetBool(constqual->constvalue))
1438 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1440 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1441 * results back in *count_est and *offset_est. These variables are set to
1442 * 0 if the corresponding clause is not present, and -1 if it's present
1443 * but we couldn't estimate the value for it. (The "0" convention is OK
1444 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1445 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1446 * usual practice of never estimating less than one row.) These values will
1447 * be passed to make_limit, which see if you change this code.
1449 * The return value is the suitably adjusted tuple_fraction to use for
1450 * planning the query. This adjustment is not overridable, since it reflects
1451 * plan actions that grouping_planner() will certainly take, not assumptions
1455 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1456 int64 *offset_est, int64 *count_est)
1458 Query *parse = root->parse;
1460 double limit_fraction;
1462 /* Should not be called unless LIMIT or OFFSET */
1463 Assert(parse->limitCount || parse->limitOffset);
1466 * Try to obtain the clause values. We use estimate_expression_value
1467 * primarily because it can sometimes do something useful with Params.
1469 if (parse->limitCount)
1471 est = estimate_expression_value(root, parse->limitCount);
1472 if (est && IsA(est, Const))
1474 if (((Const *) est)->constisnull)
1476 /* NULL indicates LIMIT ALL, ie, no limit */
1477 *count_est = 0; /* treat as not present */
1481 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1482 if (*count_est <= 0)
1483 *count_est = 1; /* force to at least 1 */
1487 *count_est = -1; /* can't estimate */
1490 *count_est = 0; /* not present */
1492 if (parse->limitOffset)
1494 est = estimate_expression_value(root, parse->limitOffset);
1495 if (est && IsA(est, Const))
1497 if (((Const *) est)->constisnull)
1499 /* Treat NULL as no offset; the executor will too */
1500 *offset_est = 0; /* treat as not present */
1504 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1505 if (*offset_est < 0)
1506 *offset_est = 0; /* less than 0 is same as 0 */
1510 *offset_est = -1; /* can't estimate */
1513 *offset_est = 0; /* not present */
1515 if (*count_est != 0)
1518 * A LIMIT clause limits the absolute number of tuples returned.
1519 * However, if it's not a constant LIMIT then we have to guess; for
1520 * lack of a better idea, assume 10% of the plan's result is wanted.
1522 if (*count_est < 0 || *offset_est < 0)
1524 /* LIMIT or OFFSET is an expression ... punt ... */
1525 limit_fraction = 0.10;
1529 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1530 limit_fraction = (double) *count_est + (double) *offset_est;
1534 * If we have absolute limits from both caller and LIMIT, use the
1535 * smaller value; likewise if they are both fractional. If one is
1536 * fractional and the other absolute, we can't easily determine which
1537 * is smaller, but we use the heuristic that the absolute will usually
1540 if (tuple_fraction >= 1.0)
1542 if (limit_fraction >= 1.0)
1545 tuple_fraction = Min(tuple_fraction, limit_fraction);
1549 /* caller absolute, limit fractional; use caller's value */
1552 else if (tuple_fraction > 0.0)
1554 if (limit_fraction >= 1.0)
1556 /* caller fractional, limit absolute; use limit */
1557 tuple_fraction = limit_fraction;
1561 /* both fractional */
1562 tuple_fraction = Min(tuple_fraction, limit_fraction);
1567 /* no info from caller, just use limit */
1568 tuple_fraction = limit_fraction;
1571 else if (*offset_est != 0 && tuple_fraction > 0.0)
1574 * We have an OFFSET but no LIMIT. This acts entirely differently
1575 * from the LIMIT case: here, we need to increase rather than decrease
1576 * the caller's tuple_fraction, because the OFFSET acts to cause more
1577 * tuples to be fetched instead of fewer. This only matters if we got
1578 * a tuple_fraction > 0, however.
1580 * As above, use 10% if OFFSET is present but unestimatable.
1582 if (*offset_est < 0)
1583 limit_fraction = 0.10;
1585 limit_fraction = (double) *offset_est;
1588 * If we have absolute counts from both caller and OFFSET, add them
1589 * together; likewise if they are both fractional. If one is
1590 * fractional and the other absolute, we want to take the larger, and
1591 * we heuristically assume that's the fractional one.
1593 if (tuple_fraction >= 1.0)
1595 if (limit_fraction >= 1.0)
1597 /* both absolute, so add them together */
1598 tuple_fraction += limit_fraction;
1602 /* caller absolute, limit fractional; use limit */
1603 tuple_fraction = limit_fraction;
1608 if (limit_fraction >= 1.0)
1610 /* caller fractional, limit absolute; use caller's value */
1614 /* both fractional, so add them together */
1615 tuple_fraction += limit_fraction;
1616 if (tuple_fraction >= 1.0)
1617 tuple_fraction = 0.0; /* assume fetch all */
1622 return tuple_fraction;
1627 * preprocess_groupclause - do preparatory work on GROUP BY clause
1629 * The idea here is to adjust the ordering of the GROUP BY elements
1630 * (which in itself is semantically insignificant) to match ORDER BY,
1631 * thereby allowing a single sort operation to both implement the ORDER BY
1632 * requirement and set up for a Unique step that implements GROUP BY.
1634 * In principle it might be interesting to consider other orderings of the
1635 * GROUP BY elements, which could match the sort ordering of other
1636 * possible plans (eg an indexscan) and thereby reduce cost. We don't
1637 * bother with that, though. Hashed grouping will frequently win anyway.
1639 * Note: we need no comparable processing of the distinctClause because
1640 * the parser already enforced that that matches ORDER BY.
1643 preprocess_groupclause(PlannerInfo *root)
1645 Query *parse = root->parse;
1646 List *new_groupclause;
1651 /* If no ORDER BY, nothing useful to do here */
1652 if (parse->sortClause == NIL)
1656 * Scan the ORDER BY clause and construct a list of matching GROUP BY
1657 * items, but only as far as we can make a matching prefix.
1659 * This code assumes that the sortClause contains no duplicate items.
1661 new_groupclause = NIL;
1662 foreach(sl, parse->sortClause)
1664 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
1666 foreach(gl, parse->groupClause)
1668 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1672 new_groupclause = lappend(new_groupclause, gc);
1677 break; /* no match, so stop scanning */
1680 /* Did we match all of the ORDER BY list, or just some of it? */
1681 partial_match = (sl != NULL);
1683 /* If no match at all, no point in reordering GROUP BY */
1684 if (new_groupclause == NIL)
1688 * Add any remaining GROUP BY items to the new list, but only if we
1689 * were able to make a complete match. In other words, we only
1690 * rearrange the GROUP BY list if the result is that one list is a
1691 * prefix of the other --- otherwise there's no possibility of a
1692 * common sort. Also, give up if there are any non-sortable GROUP BY
1693 * items, since then there's no hope anyway.
1695 foreach(gl, parse->groupClause)
1697 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1699 if (list_member_ptr(new_groupclause, gc))
1700 continue; /* it matched an ORDER BY item */
1702 return; /* give up, no common sort possible */
1703 if (!OidIsValid(gc->sortop))
1704 return; /* give up, GROUP BY can't be sorted */
1705 new_groupclause = lappend(new_groupclause, gc);
1708 /* Success --- install the rearranged GROUP BY list */
1709 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
1710 parse->groupClause = new_groupclause;
1714 * choose_hashed_grouping - should we use hashed grouping?
1716 * Note: this is only applied when both alternatives are actually feasible.
1719 choose_hashed_grouping(PlannerInfo *root,
1720 double tuple_fraction, double limit_tuples,
1721 Path *cheapest_path, Path *sorted_path,
1722 double dNumGroups, AggClauseCounts *agg_counts)
1724 int numGroupCols = list_length(root->parse->groupClause);
1725 double cheapest_path_rows;
1726 int cheapest_path_width;
1728 List *target_pathkeys;
1729 List *current_pathkeys;
1733 /* Prefer sorting when enable_hashagg is off */
1734 if (!enable_hashagg)
1738 * Don't do it if it doesn't look like the hashtable will fit into
1741 * Beware here of the possibility that cheapest_path->parent is NULL. This
1742 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1744 if (cheapest_path->parent)
1746 cheapest_path_rows = cheapest_path->parent->rows;
1747 cheapest_path_width = cheapest_path->parent->width;
1751 cheapest_path_rows = 1; /* assume non-set result */
1752 cheapest_path_width = 100; /* arbitrary */
1755 /* Estimate per-hash-entry space at tuple width... */
1756 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1757 /* plus space for pass-by-ref transition values... */
1758 hashentrysize += agg_counts->transitionSpace;
1759 /* plus the per-hash-entry overhead */
1760 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1762 if (hashentrysize * dNumGroups > work_mem * 1024L)
1766 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
1767 * DISTINCT and ORDER BY as the assumed required output sort order.
1768 * This is an oversimplification because the DISTINCT might get
1769 * implemented via hashing, but it's not clear that the case is common
1770 * enough (or that our estimates are good enough) to justify trying to
1773 if (list_length(root->distinct_pathkeys) >
1774 list_length(root->sort_pathkeys))
1775 target_pathkeys = root->distinct_pathkeys;
1777 target_pathkeys = root->sort_pathkeys;
1780 * See if the estimated cost is no more than doing it the other way. While
1781 * avoiding the need for sorted input is usually a win, the fact that the
1782 * output won't be sorted may be a loss; so we need to do an actual cost
1785 * We need to consider cheapest_path + hashagg [+ final sort] versus
1786 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1787 * presorted_path + group or agg [+ final sort] where brackets indicate a
1788 * step that may not be needed. We assume query_planner() will have
1789 * returned a presorted path only if it's a winner compared to
1790 * cheapest_path for this purpose.
1792 * These path variables are dummies that just hold cost fields; we don't
1793 * make actual Paths for these steps.
1795 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1796 numGroupCols, dNumGroups,
1797 cheapest_path->startup_cost, cheapest_path->total_cost,
1798 cheapest_path_rows);
1799 /* Result of hashed agg is always unsorted */
1800 if (target_pathkeys)
1801 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
1802 dNumGroups, cheapest_path_width, limit_tuples);
1806 sorted_p.startup_cost = sorted_path->startup_cost;
1807 sorted_p.total_cost = sorted_path->total_cost;
1808 current_pathkeys = sorted_path->pathkeys;
1812 sorted_p.startup_cost = cheapest_path->startup_cost;
1813 sorted_p.total_cost = cheapest_path->total_cost;
1814 current_pathkeys = cheapest_path->pathkeys;
1816 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1818 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1819 cheapest_path_rows, cheapest_path_width, -1.0);
1820 current_pathkeys = root->group_pathkeys;
1823 if (root->parse->hasAggs)
1824 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1825 numGroupCols, dNumGroups,
1826 sorted_p.startup_cost, sorted_p.total_cost,
1827 cheapest_path_rows);
1829 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1830 sorted_p.startup_cost, sorted_p.total_cost,
1831 cheapest_path_rows);
1832 /* The Agg or Group node will preserve ordering */
1833 if (target_pathkeys &&
1834 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
1835 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
1836 dNumGroups, cheapest_path_width, limit_tuples);
1839 * Now make the decision using the top-level tuple fraction. First we
1840 * have to convert an absolute count (LIMIT) into fractional form.
1842 if (tuple_fraction >= 1.0)
1843 tuple_fraction /= dNumGroups;
1845 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1846 tuple_fraction) < 0)
1848 /* Hashed is cheaper, so use it */
1855 * choose_hashed_distinct - should we use hashing for DISTINCT?
1857 * This is fairly similar to choose_hashed_grouping, but there are enough
1858 * differences that it doesn't seem worth trying to unify the two functions.
1860 * But note that making the two choices independently is a bit bogus in
1861 * itself. If the two could be combined into a single choice operation
1862 * it'd probably be better, but that seems far too unwieldy to be practical,
1863 * especially considering that the combination of GROUP BY and DISTINCT
1864 * isn't very common in real queries. By separating them, we are giving
1865 * extra preference to using a sorting implementation when a common sort key
1866 * is available ... and that's not necessarily wrong anyway.
1868 * Note: this is only applied when both alternatives are actually feasible.
1871 choose_hashed_distinct(PlannerInfo *root,
1872 Plan *input_plan, List *input_pathkeys,
1873 double tuple_fraction, double limit_tuples,
1874 double dNumDistinctRows)
1876 int numDistinctCols = list_length(root->parse->distinctClause);
1878 List *current_pathkeys;
1879 List *needed_pathkeys;
1883 /* Prefer sorting when enable_hashagg is off */
1884 if (!enable_hashagg)
1888 * Don't do it if it doesn't look like the hashtable will fit into
1891 hashentrysize = MAXALIGN(input_plan->plan_width) + MAXALIGN(sizeof(MinimalTupleData));
1893 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
1897 * See if the estimated cost is no more than doing it the other way. While
1898 * avoiding the need for sorted input is usually a win, the fact that the
1899 * output won't be sorted may be a loss; so we need to do an actual cost
1902 * We need to consider input_plan + hashagg [+ final sort] versus
1903 * input_plan [+ sort] + group [+ final sort] where brackets indicate
1904 * a step that may not be needed.
1906 * These path variables are dummies that just hold cost fields; we don't
1907 * make actual Paths for these steps.
1909 cost_agg(&hashed_p, root, AGG_HASHED, 0,
1910 numDistinctCols, dNumDistinctRows,
1911 input_plan->startup_cost, input_plan->total_cost,
1912 input_plan->plan_rows);
1914 * Result of hashed agg is always unsorted, so if ORDER BY is present
1915 * we need to charge for the final sort.
1917 if (root->parse->sortClause)
1918 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1919 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1922 * Now for the GROUP case. See comments in grouping_planner about the
1923 * sorting choices here --- this code should match that code.
1925 sorted_p.startup_cost = input_plan->startup_cost;
1926 sorted_p.total_cost = input_plan->total_cost;
1927 current_pathkeys = input_pathkeys;
1928 if (root->parse->hasDistinctOn &&
1929 list_length(root->distinct_pathkeys) <
1930 list_length(root->sort_pathkeys))
1931 needed_pathkeys = root->sort_pathkeys;
1933 needed_pathkeys = root->distinct_pathkeys;
1934 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1936 if (list_length(root->distinct_pathkeys) >=
1937 list_length(root->sort_pathkeys))
1938 current_pathkeys = root->distinct_pathkeys;
1940 current_pathkeys = root->sort_pathkeys;
1941 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
1942 input_plan->plan_rows, input_plan->plan_width, -1.0);
1944 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
1945 sorted_p.startup_cost, sorted_p.total_cost,
1946 input_plan->plan_rows);
1947 if (root->parse->sortClause &&
1948 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1949 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1950 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1953 * Now make the decision using the top-level tuple fraction. First we
1954 * have to convert an absolute count (LIMIT) into fractional form.
1956 if (tuple_fraction >= 1.0)
1957 tuple_fraction /= dNumDistinctRows;
1959 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1960 tuple_fraction) < 0)
1962 /* Hashed is cheaper, so use it */
1969 * make_subplanTargetList
1970 * Generate appropriate target list when grouping is required.
1972 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1973 * above the result of query_planner, we typically want to pass a different
1974 * target list to query_planner than the outer plan nodes should have.
1975 * This routine generates the correct target list for the subplan.
1977 * The initial target list passed from the parser already contains entries
1978 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1979 * for variables used only in HAVING clauses; so we need to add those
1980 * variables to the subplan target list. Also, we flatten all expressions
1981 * except GROUP BY items into their component variables; the other expressions
1982 * will be computed by the inserted nodes rather than by the subplan.
1983 * For example, given a query like
1984 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1985 * we want to pass this targetlist to the subplan:
1987 * where the a+b target will be used by the Sort/Group steps, and the
1988 * other targets will be used for computing the final results. (In the
1989 * above example we could theoretically suppress the a and b targets and
1990 * pass down only c,d,a+b, but it's not really worth the trouble to
1991 * eliminate simple var references from the subplan. We will avoid doing
1992 * the extra computation to recompute a+b at the outer level; see
1993 * fix_upper_expr() in setrefs.c.)
1995 * If we are grouping or aggregating, *and* there are no non-Var grouping
1996 * expressions, then the returned tlist is effectively dummy; we do not
1997 * need to force it to be evaluated, because all the Vars it contains
1998 * should be present in the output of query_planner anyway.
2000 * 'tlist' is the query's target list.
2001 * 'groupColIdx' receives an array of column numbers for the GROUP BY
2002 * expressions (if there are any) in the subplan's target list.
2003 * 'need_tlist_eval' is set true if we really need to evaluate the
2006 * The result is the targetlist to be passed to the subplan.
2010 make_subplanTargetList(PlannerInfo *root,
2012 AttrNumber **groupColIdx,
2013 bool *need_tlist_eval)
2015 Query *parse = root->parse;
2020 *groupColIdx = NULL;
2023 * If we're not grouping or aggregating, there's nothing to do here;
2024 * query_planner should receive the unmodified target list.
2026 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
2028 *need_tlist_eval = true;
2033 * Otherwise, start with a "flattened" tlist (having just the vars
2034 * mentioned in the targetlist and HAVING qual --- but not upper-level
2035 * Vars; they will be replaced by Params later on).
2037 sub_tlist = flatten_tlist(tlist);
2038 extravars = pull_var_clause(parse->havingQual, false);
2039 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2040 list_free(extravars);
2041 *need_tlist_eval = false; /* only eval if not flat tlist */
2044 * If grouping, create sub_tlist entries for all GROUP BY expressions
2045 * (GROUP BY items that are simple Vars should be in the list already),
2046 * and make an array showing where the group columns are in the sub_tlist.
2048 numCols = list_length(parse->groupClause);
2052 AttrNumber *grpColIdx;
2055 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2056 *groupColIdx = grpColIdx;
2058 foreach(gl, parse->groupClause)
2060 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2061 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2062 TargetEntry *te = NULL;
2065 * Find or make a matching sub_tlist entry. If the groupexpr
2066 * isn't a Var, no point in searching. (Note that the parser
2067 * won't make multiple groupClause entries for the same TLE.)
2069 if (groupexpr && IsA(groupexpr, Var))
2073 foreach(sl, sub_tlist)
2075 TargetEntry *lte = (TargetEntry *) lfirst(sl);
2077 if (equal(groupexpr, lte->expr))
2086 te = makeTargetEntry((Expr *) groupexpr,
2087 list_length(sub_tlist) + 1,
2090 sub_tlist = lappend(sub_tlist, te);
2091 *need_tlist_eval = true; /* it's not flat anymore */
2094 /* and save its resno */
2095 grpColIdx[keyno++] = te->resno;
2103 * locate_grouping_columns
2104 * Locate grouping columns in the tlist chosen by query_planner.
2106 * This is only needed if we don't use the sub_tlist chosen by
2107 * make_subplanTargetList. We have to forget the column indexes found
2108 * by that routine and re-locate the grouping vars in the real sub_tlist.
2111 locate_grouping_columns(PlannerInfo *root,
2114 AttrNumber *groupColIdx)
2120 * No work unless grouping.
2122 if (!root->parse->groupClause)
2124 Assert(groupColIdx == NULL);
2127 Assert(groupColIdx != NULL);
2129 foreach(gl, root->parse->groupClause)
2131 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2132 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2133 TargetEntry *te = NULL;
2136 foreach(sl, sub_tlist)
2138 te = (TargetEntry *) lfirst(sl);
2139 if (equal(groupexpr, te->expr))
2143 elog(ERROR, "failed to locate grouping columns");
2145 groupColIdx[keyno++] = te->resno;
2150 * postprocess_setop_tlist
2151 * Fix up targetlist returned by plan_set_operations().
2153 * We need to transpose sort key info from the orig_tlist into new_tlist.
2154 * NOTE: this would not be good enough if we supported resjunk sort keys
2155 * for results of set operations --- then, we'd need to project a whole
2156 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2157 * find any resjunk columns in orig_tlist.
2160 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2163 ListCell *orig_tlist_item = list_head(orig_tlist);
2165 foreach(l, new_tlist)
2167 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2168 TargetEntry *orig_tle;
2170 /* ignore resjunk columns in setop result */
2171 if (new_tle->resjunk)
2174 Assert(orig_tlist_item != NULL);
2175 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2176 orig_tlist_item = lnext(orig_tlist_item);
2177 if (orig_tle->resjunk) /* should not happen */
2178 elog(ERROR, "resjunk output columns are not implemented");
2179 Assert(new_tle->resno == orig_tle->resno);
2180 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2182 if (orig_tlist_item != NULL)
2183 elog(ERROR, "resjunk output columns are not implemented");