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 1.242 2008/08/17 01:19:59 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->transientPlan = false;
145 /* Determine what fraction of the plan is likely to be scanned */
146 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
149 * We have no real idea how many tuples the user will ultimately FETCH
150 * from a cursor, but it is often the case that he doesn't want 'em
151 * all, or would prefer a fast-start plan anyway so that he can
152 * process some of the tuples sooner. Use a GUC parameter to decide
153 * what fraction to optimize for.
155 tuple_fraction = cursor_tuple_fraction;
158 * We document cursor_tuple_fraction as simply being a fraction,
159 * which means the edge cases 0 and 1 have to be treated specially
160 * here. We convert 1 to 0 ("all the tuples") and 0 to a very small
163 if (tuple_fraction >= 1.0)
164 tuple_fraction = 0.0;
165 else if (tuple_fraction <= 0.0)
166 tuple_fraction = 1e-10;
170 /* Default assumption is we need all the tuples */
171 tuple_fraction = 0.0;
174 /* primary planning entry point (may recurse for subqueries) */
175 top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
178 * If creating a plan for a scrollable cursor, make sure it can run
179 * backwards on demand. Add a Material node at the top at need.
181 if (cursorOptions & CURSOR_OPT_SCROLL)
183 if (!ExecSupportsBackwardScan(top_plan))
184 top_plan = materialize_finished_plan(top_plan);
187 /* final cleanup of the plan */
188 Assert(glob->finalrtable == NIL);
189 top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
190 /* ... and the subplans (both regular subplans and initplans) */
191 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
192 forboth(lp, glob->subplans, lr, glob->subrtables)
194 Plan *subplan = (Plan *) lfirst(lp);
195 List *subrtable = (List *) lfirst(lr);
197 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
200 /* build the PlannedStmt result */
201 result = makeNode(PlannedStmt);
203 result->commandType = parse->commandType;
204 result->canSetTag = parse->canSetTag;
205 result->transientPlan = glob->transientPlan;
206 result->planTree = top_plan;
207 result->rtable = glob->finalrtable;
208 result->resultRelations = root->resultRelations;
209 result->utilityStmt = parse->utilityStmt;
210 result->intoClause = parse->intoClause;
211 result->subplans = glob->subplans;
212 result->rewindPlanIDs = glob->rewindPlanIDs;
213 result->returningLists = root->returningLists;
214 result->rowMarks = parse->rowMarks;
215 result->relationOids = glob->relationOids;
216 result->nParamExec = list_length(glob->paramlist);
222 /*--------------------
224 * Invokes the planner on a subquery. We recurse to here for each
225 * sub-SELECT found in the query tree.
227 * glob is the global state for the current planner run.
228 * parse is the querytree produced by the parser & rewriter.
229 * level is the current recursion depth (1 at the top-level Query).
230 * tuple_fraction is the fraction of tuples we expect will be retrieved.
231 * tuple_fraction is interpreted as explained for grouping_planner, below.
233 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
234 * among other things this tells the output sort ordering of the plan.
236 * Basically, this routine does the stuff that should only be done once
237 * per Query object. It then calls grouping_planner. At one time,
238 * grouping_planner could be invoked recursively on the same Query object;
239 * that's not currently true, but we keep the separation between the two
240 * routines anyway, in case we need it again someday.
242 * subquery_planner will be called recursively to handle sub-Query nodes
243 * found within the query's expressions and rangetable.
245 * Returns a query plan.
246 *--------------------
249 subquery_planner(PlannerGlobal *glob, Query *parse,
250 Index level, double tuple_fraction,
251 PlannerInfo **subroot)
253 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->append_rel_list = NIL;
271 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
272 * to transform them into joins. Note that this step does not descend
273 * into subqueries; if we pull up any subqueries below, their SubLinks are
274 * processed just before pulling them up.
276 if (parse->hasSubLinks)
277 pull_up_sublinks(root);
280 * Scan the rangetable for set-returning functions, and inline them
281 * if possible (producing subqueries that might get pulled up next).
282 * Recursion issues here are handled in the same way as for SubLinks.
284 inline_set_returning_functions(root);
287 * Check to see if any subqueries in the rangetable can be merged into
290 parse->jointree = (FromExpr *)
291 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
294 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
295 * avoid the expense of doing flatten_join_alias_vars(). Also check for
296 * outer joins --- if none, we can skip reduce_outer_joins().
297 * This must be done after we have done pull_up_subqueries, of course.
299 root->hasJoinRTEs = false;
300 hasOuterJoins = false;
301 foreach(l, parse->rtable)
303 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
305 if (rte->rtekind == RTE_JOIN)
307 root->hasJoinRTEs = true;
308 if (IS_OUTER_JOIN(rte->jointype))
310 hasOuterJoins = true;
311 /* Can quit scanning once we find an outer join */
318 * Expand any rangetable entries that are inheritance sets into "append
319 * relations". This can add entries to the rangetable, but they must be
320 * plain base relations not joins, so it's OK (and marginally more
321 * efficient) to do it after checking for join RTEs. We must do it after
322 * pulling up subqueries, else we'd fail to handle inherited tables in
325 expand_inherited_tables(root);
328 * Set hasHavingQual to remember if HAVING clause is present. Needed
329 * because preprocess_expression will reduce a constant-true condition to
330 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
332 root->hasHavingQual = (parse->havingQual != NULL);
334 /* Clear this flag; might get set in distribute_qual_to_rels */
335 root->hasPseudoConstantQuals = false;
338 * Do expression preprocessing on targetlist and quals.
340 parse->targetList = (List *)
341 preprocess_expression(root, (Node *) parse->targetList,
344 parse->returningList = (List *)
345 preprocess_expression(root, (Node *) parse->returningList,
348 preprocess_qual_conditions(root, (Node *) parse->jointree);
350 parse->havingQual = preprocess_expression(root, parse->havingQual,
353 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
355 parse->limitCount = preprocess_expression(root, parse->limitCount,
358 root->append_rel_list = (List *)
359 preprocess_expression(root, (Node *) root->append_rel_list,
362 /* Also need to preprocess expressions for function and values RTEs */
363 foreach(l, parse->rtable)
365 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
367 if (rte->rtekind == RTE_FUNCTION)
368 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
370 else if (rte->rtekind == RTE_VALUES)
371 rte->values_lists = (List *)
372 preprocess_expression(root, (Node *) rte->values_lists,
377 * In some cases we may want to transfer a HAVING clause into WHERE. We
378 * cannot do so if the HAVING clause contains aggregates (obviously) or
379 * volatile functions (since a HAVING clause is supposed to be executed
380 * only once per group). Also, it may be that the clause is so expensive
381 * to execute that we're better off doing it only once per group, despite
382 * the loss of selectivity. This is hard to estimate short of doing the
383 * entire planning process twice, so we use a heuristic: clauses
384 * containing subplans are left in HAVING. Otherwise, we move or copy the
385 * HAVING clause into WHERE, in hopes of eliminating tuples before
386 * aggregation instead of after.
388 * If the query has explicit grouping then we can simply move such a
389 * clause into WHERE; any group that fails the clause will not be in the
390 * output because none of its tuples will reach the grouping or
391 * aggregation stage. Otherwise we must have a degenerate (variable-free)
392 * HAVING clause, which we put in WHERE so that query_planner() can use it
393 * in a gating Result node, but also keep in HAVING to ensure that we
394 * don't emit a bogus aggregated row. (This could be done better, but it
395 * seems not worth optimizing.)
397 * Note that both havingQual and parse->jointree->quals are in
398 * implicitly-ANDed-list form at this point, even though they are declared
402 foreach(l, (List *) parse->havingQual)
404 Node *havingclause = (Node *) lfirst(l);
406 if (contain_agg_clause(havingclause) ||
407 contain_volatile_functions(havingclause) ||
408 contain_subplans(havingclause))
410 /* keep it in HAVING */
411 newHaving = lappend(newHaving, havingclause);
413 else if (parse->groupClause)
415 /* move it to WHERE */
416 parse->jointree->quals = (Node *)
417 lappend((List *) parse->jointree->quals, havingclause);
421 /* put a copy in WHERE, keep it in HAVING */
422 parse->jointree->quals = (Node *)
423 lappend((List *) parse->jointree->quals,
424 copyObject(havingclause));
425 newHaving = lappend(newHaving, havingclause);
428 parse->havingQual = (Node *) newHaving;
431 * If we have any outer joins, try to reduce them to plain inner joins.
432 * This step is most easily done after we've done expression
436 reduce_outer_joins(root);
439 * Do the main planning. If we have an inherited target relation, that
440 * needs special processing, else go straight to grouping_planner.
442 if (parse->resultRelation &&
443 rt_fetch(parse->resultRelation, parse->rtable)->inh)
444 plan = inheritance_planner(root);
446 plan = grouping_planner(root, tuple_fraction);
449 * If any subplans were generated, or if we're inside a subplan, build
450 * initPlan list and extParam/allParam sets for plan nodes, and attach the
451 * initPlans to the top plan node.
453 if (list_length(glob->subplans) != num_old_subplans ||
454 root->query_level > 1)
455 SS_finalize_plan(root, plan, true);
457 /* Return internal info if caller wants it */
465 * preprocess_expression
466 * Do subquery_planner's preprocessing work for an expression,
467 * which can be a targetlist, a WHERE clause (including JOIN/ON
468 * conditions), or a HAVING clause.
471 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
474 * Fall out quickly if expression is empty. This occurs often enough to
475 * be worth checking. Note that null->null is the correct conversion for
476 * implicit-AND result format, too.
482 * If the query has any join RTEs, replace join alias variables with
483 * base-relation variables. We must do this before sublink processing,
484 * else sublinks expanded out from join aliases wouldn't get processed. We
485 * can skip it in VALUES lists, however, since they can't contain any Vars
488 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
489 expr = flatten_join_alias_vars(root, expr);
492 * Simplify constant expressions.
494 * Note: this also flattens nested AND and OR expressions into N-argument
495 * form. All processing of a qual expression after this point must be
496 * careful to maintain AND/OR flatness --- that is, do not generate a tree
497 * with AND directly under AND, nor OR directly under OR.
499 * Because this is a relatively expensive process, we skip it when the
500 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
501 * expression will only be evaluated once anyway, so no point in
502 * pre-simplifying; we can't execute it any faster than the executor can,
503 * and we will waste cycles copying the tree. Notice however that we
504 * still must do it for quals (to get AND/OR flatness); and if we are in a
505 * subquery we should not assume it will be done only once.
507 * For VALUES lists we never do this at all, again on the grounds that we
508 * should optimize for one-time evaluation.
510 if (kind != EXPRKIND_VALUES &&
511 (root->parse->jointree->fromlist != NIL ||
512 kind == EXPRKIND_QUAL ||
513 root->query_level > 1))
514 expr = eval_const_expressions(root, expr);
517 * If it's a qual or havingQual, canonicalize it.
519 if (kind == EXPRKIND_QUAL)
521 expr = (Node *) canonicalize_qual((Expr *) expr);
523 #ifdef OPTIMIZER_DEBUG
524 printf("After canonicalize_qual()\n");
529 /* Expand SubLinks to SubPlans */
530 if (root->parse->hasSubLinks)
531 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
534 * XXX do not insert anything here unless you have grokked the comments in
535 * SS_replace_correlation_vars ...
538 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
539 if (root->query_level > 1)
540 expr = SS_replace_correlation_vars(root, expr);
543 * If it's a qual or havingQual, convert it to implicit-AND format. (We
544 * don't want to do this before eval_const_expressions, since the latter
545 * would be unable to simplify a top-level AND correctly. Also,
546 * SS_process_sublinks expects explicit-AND format.)
548 if (kind == EXPRKIND_QUAL)
549 expr = (Node *) make_ands_implicit((Expr *) expr);
555 * preprocess_qual_conditions
556 * Recursively scan the query's jointree and do subquery_planner's
557 * preprocessing work on each qual condition found therein.
560 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
564 if (IsA(jtnode, RangeTblRef))
566 /* nothing to do here */
568 else if (IsA(jtnode, FromExpr))
570 FromExpr *f = (FromExpr *) jtnode;
573 foreach(l, f->fromlist)
574 preprocess_qual_conditions(root, lfirst(l));
576 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
578 else if (IsA(jtnode, JoinExpr))
580 JoinExpr *j = (JoinExpr *) jtnode;
582 preprocess_qual_conditions(root, j->larg);
583 preprocess_qual_conditions(root, j->rarg);
585 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
588 elog(ERROR, "unrecognized node type: %d",
589 (int) nodeTag(jtnode));
593 * inheritance_planner
594 * Generate a plan in the case where the result relation is an
597 * We have to handle this case differently from cases where a source relation
598 * is an inheritance set. Source inheritance is expanded at the bottom of the
599 * plan tree (see allpaths.c), but target inheritance has to be expanded at
600 * the top. The reason is that for UPDATE, each target relation needs a
601 * different targetlist matching its own column set. Also, for both UPDATE
602 * and DELETE, the executor needs the Append plan node at the top, else it
603 * can't keep track of which table is the current target table. Fortunately,
604 * the UPDATE/DELETE target can never be the nullable side of an outer join,
605 * so it's OK to generate the plan this way.
607 * Returns a query plan.
610 inheritance_planner(PlannerInfo *root)
612 Query *parse = root->parse;
613 int parentRTindex = parse->resultRelation;
614 List *subplans = NIL;
615 List *resultRelations = NIL;
616 List *returningLists = NIL;
622 foreach(l, root->append_rel_list)
624 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
627 /* append_rel_list contains all append rels; ignore others */
628 if (appinfo->parent_relid != parentRTindex)
632 * Generate modified query with this rel as target.
634 memcpy(&subroot, root, sizeof(PlannerInfo));
635 subroot.parse = (Query *)
636 adjust_appendrel_attrs((Node *) parse,
638 subroot.init_plans = NIL;
639 /* There shouldn't be any OJ info to translate, as yet */
640 Assert(subroot.join_info_list == NIL);
643 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
646 * If this child rel was excluded by constraint exclusion, exclude it
649 if (is_dummy_plan(subplan))
652 /* Save rtable and tlist from first rel for use below */
655 rtable = subroot.parse->rtable;
656 tlist = subplan->targetlist;
659 subplans = lappend(subplans, subplan);
661 /* Make sure any initplans from this rel get into the outer list */
662 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
664 /* Build target-relations list for the executor */
665 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
667 /* Build list of per-relation RETURNING targetlists */
668 if (parse->returningList)
670 Assert(list_length(subroot.returningLists) == 1);
671 returningLists = list_concat(returningLists,
672 subroot.returningLists);
676 root->resultRelations = resultRelations;
677 root->returningLists = returningLists;
679 /* Mark result as unordered (probably unnecessary) */
680 root->query_pathkeys = NIL;
683 * If we managed to exclude every child rel, return a dummy plan
687 root->resultRelations = list_make1_int(parentRTindex);
688 /* although dummy, it must have a valid tlist for executor */
689 tlist = preprocess_targetlist(root, parse->targetList);
690 return (Plan *) make_result(root,
692 (Node *) list_make1(makeBoolConst(false,
698 * Planning might have modified the rangetable, due to changes of the
699 * Query structures inside subquery RTEs. We have to ensure that this
700 * gets propagated back to the master copy. But can't do this until we
701 * are done planning, because all the calls to grouping_planner need
702 * virgin sub-Queries to work from. (We are effectively assuming that
703 * sub-Queries will get planned identically each time, or at least that
704 * the impacts on their rangetables will be the same each time.)
706 * XXX should clean this up someday
708 parse->rtable = rtable;
710 /* Suppress Append if there's only one surviving child rel */
711 if (list_length(subplans) == 1)
712 return (Plan *) linitial(subplans);
714 return (Plan *) make_append(subplans, true, tlist);
717 /*--------------------
719 * Perform planning steps related to grouping, aggregation, etc.
720 * This primarily means adding top-level processing to the basic
721 * query plan produced by query_planner.
723 * tuple_fraction is the fraction of tuples we expect will be retrieved
725 * tuple_fraction is interpreted as follows:
726 * 0: expect all tuples to be retrieved (normal case)
727 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
728 * from the plan to be retrieved
729 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
730 * expected to be retrieved (ie, a LIMIT specification)
732 * Returns a query plan. Also, root->query_pathkeys is returned as the
733 * actual output ordering of the plan (in pathkey format).
734 *--------------------
737 grouping_planner(PlannerInfo *root, double tuple_fraction)
739 Query *parse = root->parse;
740 List *tlist = parse->targetList;
741 int64 offset_est = 0;
743 double limit_tuples = -1.0;
745 List *current_pathkeys;
746 double dNumGroups = 0;
748 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
749 if (parse->limitCount || parse->limitOffset)
751 tuple_fraction = preprocess_limit(root, tuple_fraction,
752 &offset_est, &count_est);
755 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
756 * estimate the effects of using a bounded sort.
758 if (count_est > 0 && offset_est >= 0)
759 limit_tuples = (double) count_est + (double) offset_est;
762 if (parse->setOperations)
764 List *set_sortclauses;
767 * If there's a top-level ORDER BY, assume we have to fetch all the
768 * tuples. This might be too simplistic given all the hackery below
769 * to possibly avoid the sort; but the odds of accurate estimates
770 * here are pretty low anyway.
772 if (parse->sortClause)
773 tuple_fraction = 0.0;
776 * Construct the plan for set operations. The result will not need
777 * any work except perhaps a top-level sort and/or LIMIT.
779 result_plan = plan_set_operations(root, tuple_fraction,
783 * Calculate pathkeys representing the sort order (if any) of the set
784 * operation's result. We have to do this before overwriting the sort
787 current_pathkeys = make_pathkeys_for_sortclauses(root,
789 result_plan->targetlist,
793 * We should not need to call preprocess_targetlist, since we must be
794 * in a SELECT query node. Instead, use the targetlist returned by
795 * plan_set_operations (since this tells whether it returned any
796 * resjunk columns!), and transfer any sort key information from the
799 Assert(parse->commandType == CMD_SELECT);
801 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
805 * Can't handle FOR UPDATE/SHARE here (parser should have checked
806 * already, but let's make sure).
810 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
811 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
814 * Calculate pathkeys that represent result ordering requirements
816 Assert(parse->distinctClause == NIL);
817 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
824 /* No set operations, do regular planning */
826 AttrNumber *groupColIdx = NULL;
827 bool need_tlist_eval = true;
833 AggClauseCounts agg_counts;
835 bool use_hashed_grouping = false;
837 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
839 /* Preprocess GROUP BY clause, if any */
840 if (parse->groupClause)
841 preprocess_groupclause(root);
842 numGroupCols = list_length(parse->groupClause);
844 /* Preprocess targetlist */
845 tlist = preprocess_targetlist(root, tlist);
848 * Generate appropriate target list for subplan; may be different from
849 * tlist if grouping or aggregation is needed.
851 sub_tlist = make_subplanTargetList(root, tlist,
852 &groupColIdx, &need_tlist_eval);
855 * Calculate pathkeys that represent grouping/ordering requirements.
856 * Stash them in PlannerInfo so that query_planner can canonicalize
857 * them after EquivalenceClasses have been formed. The sortClause
858 * is certainly sort-able, but GROUP BY and DISTINCT might not be,
859 * in which case we just leave their pathkeys empty.
861 if (parse->groupClause &&
862 grouping_is_sortable(parse->groupClause))
863 root->group_pathkeys =
864 make_pathkeys_for_sortclauses(root,
869 root->group_pathkeys = NIL;
871 if (parse->distinctClause &&
872 grouping_is_sortable(parse->distinctClause))
873 root->distinct_pathkeys =
874 make_pathkeys_for_sortclauses(root,
875 parse->distinctClause,
879 root->distinct_pathkeys = NIL;
881 root->sort_pathkeys =
882 make_pathkeys_for_sortclauses(root,
888 * Will need actual number of aggregates for estimating costs.
890 * Note: we do not attempt to detect duplicate aggregates here; a
891 * somewhat-overestimated count is okay for our present purposes.
893 * Note: think not that we can turn off hasAggs if we find no aggs. It
894 * is possible for constant-expression simplification to remove all
895 * explicit references to aggs, but we still have to follow the
896 * aggregate semantics (eg, producing only one output row).
900 count_agg_clauses((Node *) tlist, &agg_counts);
901 count_agg_clauses(parse->havingQual, &agg_counts);
905 * Figure out whether we want a sorted result from query_planner.
907 * If we have a sortable GROUP BY clause, then we want a result sorted
908 * properly for grouping. Otherwise, if there's a sortable DISTINCT
909 * clause that's more rigorous than the ORDER BY clause, we try to
910 * produce output that's sufficiently well sorted for the DISTINCT.
911 * Otherwise, if there is an ORDER BY clause, we want to sort by the
914 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
915 * superset of GROUP BY, it would be tempting to request sort by ORDER
916 * BY --- but that might just leave us failing to exploit an available
917 * sort order at all. Needs more thought. The choice for DISTINCT
918 * versus ORDER BY is much easier, since we know that the parser
919 * ensured that one is a superset of the other.
921 if (root->group_pathkeys)
922 root->query_pathkeys = root->group_pathkeys;
923 else if (list_length(root->distinct_pathkeys) >
924 list_length(root->sort_pathkeys))
925 root->query_pathkeys = root->distinct_pathkeys;
926 else if (root->sort_pathkeys)
927 root->query_pathkeys = root->sort_pathkeys;
929 root->query_pathkeys = NIL;
932 * Generate the best unsorted and presorted paths for this Query (but
933 * note there may not be any presorted path). query_planner will also
934 * estimate the number of groups in the query, and canonicalize all
937 query_planner(root, sub_tlist, tuple_fraction, limit_tuples,
938 &cheapest_path, &sorted_path, &dNumGroups);
941 * If grouping, decide whether to use sorted or hashed grouping.
943 if (parse->groupClause)
949 * Executor doesn't support hashed aggregation with DISTINCT
950 * aggregates. (Doing so would imply storing *all* the input
951 * values in the hash table, which seems like a certain loser.)
953 can_hash = (agg_counts.numDistinctAggs == 0 &&
954 grouping_is_hashable(parse->groupClause));
955 can_sort = grouping_is_sortable(parse->groupClause);
956 if (can_hash && can_sort)
958 /* we have a meaningful choice to make ... */
959 use_hashed_grouping =
960 choose_hashed_grouping(root,
961 tuple_fraction, limit_tuples,
962 cheapest_path, sorted_path,
963 dNumGroups, &agg_counts);
966 use_hashed_grouping = true;
968 use_hashed_grouping = false;
971 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
972 errmsg("could not implement GROUP BY"),
973 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
975 /* Also convert # groups to long int --- but 'ware overflow! */
976 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
980 * Select the best path. If we are doing hashed grouping, we will
981 * always read all the input tuples, so use the cheapest-total path.
982 * Otherwise, trust query_planner's decision about which to use.
984 if (use_hashed_grouping || !sorted_path)
985 best_path = cheapest_path;
987 best_path = sorted_path;
990 * Check to see if it's possible to optimize MIN/MAX aggregates. If
991 * so, we will forget all the work we did so far to choose a "regular"
992 * path ... but we had to do it anyway to be able to tell which way is
995 result_plan = optimize_minmax_aggregates(root,
998 if (result_plan != NULL)
1001 * optimize_minmax_aggregates generated the full plan, with the
1002 * right tlist, and it has no sort order.
1004 current_pathkeys = NIL;
1009 * Normal case --- create a plan according to query_planner's
1012 bool need_sort_for_grouping = false;
1014 result_plan = create_plan(root, best_path);
1015 current_pathkeys = best_path->pathkeys;
1017 /* Detect if we'll need an explicit sort for grouping */
1018 if (parse->groupClause && !use_hashed_grouping &&
1019 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1021 need_sort_for_grouping = true;
1023 * Always override query_planner's tlist, so that we don't
1024 * sort useless data from a "physical" tlist.
1026 need_tlist_eval = true;
1030 * create_plan() returns a plan with just a "flat" tlist of
1031 * required Vars. Usually we need to insert the sub_tlist as the
1032 * tlist of the top plan node. However, we can skip that if we
1033 * determined that whatever query_planner chose to return will be
1036 if (need_tlist_eval)
1039 * If the top-level plan node is one that cannot do expression
1040 * evaluation, we must insert a Result node to project the
1043 if (!is_projection_capable_plan(result_plan))
1045 result_plan = (Plan *) make_result(root,
1053 * Otherwise, just replace the subplan's flat tlist with
1054 * the desired tlist.
1056 result_plan->targetlist = sub_tlist;
1060 * Also, account for the cost of evaluation of the sub_tlist.
1062 * Up to now, we have only been dealing with "flat" tlists,
1063 * containing just Vars. So their evaluation cost is zero
1064 * according to the model used by cost_qual_eval() (or if you
1065 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1066 * can avoid accounting for tlist cost throughout
1067 * query_planner() and subroutines. But now we've inserted a
1068 * tlist that might contain actual operators, sub-selects, etc
1069 * --- so we'd better account for its cost.
1071 * Below this point, any tlist eval cost for added-on nodes
1072 * should be accounted for as we create those nodes.
1073 * Presently, of the node types we can add on, only Agg and
1074 * Group project new tlists (the rest just copy their input
1075 * tuples) --- so make_agg() and make_group() are responsible
1076 * for computing the added cost.
1078 cost_qual_eval(&tlist_cost, sub_tlist, root);
1079 result_plan->startup_cost += tlist_cost.startup;
1080 result_plan->total_cost += tlist_cost.startup +
1081 tlist_cost.per_tuple * result_plan->plan_rows;
1086 * Since we're using query_planner's tlist and not the one
1087 * make_subplanTargetList calculated, we have to refigure any
1088 * grouping-column indexes make_subplanTargetList computed.
1090 locate_grouping_columns(root, tlist, result_plan->targetlist,
1095 * Insert AGG or GROUP node if needed, plus an explicit sort step
1098 * HAVING clause, if any, becomes qual of the Agg or Group node.
1100 if (use_hashed_grouping)
1102 /* Hashed aggregate plan --- no sort needed */
1103 result_plan = (Plan *) make_agg(root,
1105 (List *) parse->havingQual,
1109 extract_grouping_ops(parse->groupClause),
1113 /* Hashed aggregation produces randomly-ordered results */
1114 current_pathkeys = NIL;
1116 else if (parse->hasAggs)
1118 /* Plain aggregate plan --- sort if needed */
1119 AggStrategy aggstrategy;
1121 if (parse->groupClause)
1123 if (need_sort_for_grouping)
1125 result_plan = (Plan *)
1126 make_sort_from_groupcols(root,
1130 current_pathkeys = root->group_pathkeys;
1132 aggstrategy = AGG_SORTED;
1135 * The AGG node will not change the sort ordering of its
1136 * groups, so current_pathkeys describes the result too.
1141 aggstrategy = AGG_PLAIN;
1142 /* Result will be only one row anyway; no sort order */
1143 current_pathkeys = NIL;
1146 result_plan = (Plan *) make_agg(root,
1148 (List *) parse->havingQual,
1152 extract_grouping_ops(parse->groupClause),
1157 else if (parse->groupClause)
1160 * GROUP BY without aggregation, so insert a group node (plus
1161 * the appropriate sort node, if necessary).
1163 * Add an explicit sort if we couldn't make the path come out
1164 * the way the GROUP node needs it.
1166 if (need_sort_for_grouping)
1168 result_plan = (Plan *)
1169 make_sort_from_groupcols(root,
1173 current_pathkeys = root->group_pathkeys;
1176 result_plan = (Plan *) make_group(root,
1178 (List *) parse->havingQual,
1181 extract_grouping_ops(parse->groupClause),
1184 /* The Group node won't change sort ordering */
1186 else if (root->hasHavingQual)
1189 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1190 * This is a degenerate case in which we are supposed to emit
1191 * either 0 or 1 row depending on whether HAVING succeeds.
1192 * Furthermore, there cannot be any variables in either HAVING
1193 * or the targetlist, so we actually do not need the FROM
1194 * table at all! We can just throw away the plan-so-far and
1195 * generate a Result node. This is a sufficiently unusual
1196 * corner case that it's not worth contorting the structure of
1197 * this routine to avoid having to generate the plan in the
1200 result_plan = (Plan *) make_result(root,
1205 } /* end of non-minmax-aggregate case */
1206 } /* end of if (setOperations) */
1209 * If there is a DISTINCT clause, add the necessary node(s).
1211 if (parse->distinctClause)
1213 double dNumDistinctRows;
1214 long numDistinctRows;
1215 bool use_hashed_distinct;
1220 * If there was grouping or aggregation, use the current number of
1221 * rows as the estimated number of DISTINCT rows (ie, assume the
1222 * result was already mostly unique). If not, use the number of
1223 * distinct-groups calculated by query_planner.
1225 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1226 dNumDistinctRows = result_plan->plan_rows;
1228 dNumDistinctRows = dNumGroups;
1230 /* Also convert to long int --- but 'ware overflow! */
1231 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1234 * If we have a sortable DISTINCT ON clause, we always use sorting.
1235 * This enforces the expected behavior of DISTINCT ON.
1237 can_sort = grouping_is_sortable(parse->distinctClause);
1238 if (can_sort && parse->hasDistinctOn)
1239 use_hashed_distinct = false;
1242 can_hash = grouping_is_hashable(parse->distinctClause);
1243 if (can_hash && can_sort)
1245 /* we have a meaningful choice to make ... */
1246 use_hashed_distinct =
1247 choose_hashed_distinct(root,
1248 result_plan, current_pathkeys,
1249 tuple_fraction, limit_tuples,
1253 use_hashed_distinct = true;
1255 use_hashed_distinct = false;
1259 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1260 errmsg("could not implement DISTINCT"),
1261 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
1262 use_hashed_distinct = false; /* keep compiler quiet */
1266 if (use_hashed_distinct)
1268 /* Hashed aggregate plan --- no sort needed */
1269 result_plan = (Plan *) make_agg(root,
1270 result_plan->targetlist,
1273 list_length(parse->distinctClause),
1274 extract_grouping_cols(parse->distinctClause,
1275 result_plan->targetlist),
1276 extract_grouping_ops(parse->distinctClause),
1280 /* Hashed aggregation produces randomly-ordered results */
1281 current_pathkeys = NIL;
1286 * Use a Unique node to implement DISTINCT. Add an explicit sort
1287 * if we couldn't make the path come out the way the Unique node
1288 * needs it. If we do have to sort, always sort by the more
1289 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1290 * below. However, for regular DISTINCT, don't sort now if we
1291 * don't have to --- sorting afterwards will likely be cheaper,
1292 * and also has the possibility of optimizing via LIMIT. But
1293 * for DISTINCT ON, we *must* force the final sort now, else
1294 * it won't have the desired behavior.
1296 List *needed_pathkeys;
1298 if (parse->hasDistinctOn &&
1299 list_length(root->distinct_pathkeys) <
1300 list_length(root->sort_pathkeys))
1301 needed_pathkeys = root->sort_pathkeys;
1303 needed_pathkeys = root->distinct_pathkeys;
1305 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1307 if (list_length(root->distinct_pathkeys) >=
1308 list_length(root->sort_pathkeys))
1309 current_pathkeys = root->distinct_pathkeys;
1312 current_pathkeys = root->sort_pathkeys;
1313 /* Assert checks that parser didn't mess up... */
1314 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1318 result_plan = (Plan *) make_sort_from_pathkeys(root,
1324 result_plan = (Plan *) make_unique(result_plan,
1325 parse->distinctClause);
1326 result_plan->plan_rows = dNumDistinctRows;
1327 /* The Unique node won't change sort ordering */
1332 * If ORDER BY was given and we were not able to make the plan come out in
1333 * the right order, add an explicit sort step.
1335 if (parse->sortClause)
1337 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1339 result_plan = (Plan *) make_sort_from_pathkeys(root,
1341 root->sort_pathkeys,
1343 current_pathkeys = root->sort_pathkeys;
1348 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1350 if (parse->limitCount || parse->limitOffset)
1352 result_plan = (Plan *) make_limit(result_plan,
1360 * Deal with the RETURNING clause if any. It's convenient to pass the
1361 * returningList through setrefs.c now rather than at top level (if we
1362 * waited, handling inherited UPDATE/DELETE would be much harder).
1364 if (parse->returningList)
1368 Assert(parse->resultRelation);
1369 rlist = set_returning_clause_references(root->glob,
1370 parse->returningList,
1372 parse->resultRelation);
1373 root->returningLists = list_make1(rlist);
1376 root->returningLists = NIL;
1378 /* Compute result-relations list if needed */
1379 if (parse->resultRelation)
1380 root->resultRelations = list_make1_int(parse->resultRelation);
1382 root->resultRelations = NIL;
1385 * Return the actual output ordering in query_pathkeys for possible use by
1386 * an outer query level.
1388 root->query_pathkeys = current_pathkeys;
1394 * Detect whether a plan node is a "dummy" plan created when a relation
1395 * is deemed not to need scanning due to constraint exclusion.
1397 * Currently, such dummy plans are Result nodes with constant FALSE
1401 is_dummy_plan(Plan *plan)
1403 if (IsA(plan, Result))
1405 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1407 if (list_length(rcqual) == 1)
1409 Const *constqual = (Const *) linitial(rcqual);
1411 if (constqual && IsA(constqual, Const))
1413 if (!constqual->constisnull &&
1414 !DatumGetBool(constqual->constvalue))
1423 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1425 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1426 * results back in *count_est and *offset_est. These variables are set to
1427 * 0 if the corresponding clause is not present, and -1 if it's present
1428 * but we couldn't estimate the value for it. (The "0" convention is OK
1429 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1430 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1431 * usual practice of never estimating less than one row.) These values will
1432 * be passed to make_limit, which see if you change this code.
1434 * The return value is the suitably adjusted tuple_fraction to use for
1435 * planning the query. This adjustment is not overridable, since it reflects
1436 * plan actions that grouping_planner() will certainly take, not assumptions
1440 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1441 int64 *offset_est, int64 *count_est)
1443 Query *parse = root->parse;
1445 double limit_fraction;
1447 /* Should not be called unless LIMIT or OFFSET */
1448 Assert(parse->limitCount || parse->limitOffset);
1451 * Try to obtain the clause values. We use estimate_expression_value
1452 * primarily because it can sometimes do something useful with Params.
1454 if (parse->limitCount)
1456 est = estimate_expression_value(root, parse->limitCount);
1457 if (est && IsA(est, Const))
1459 if (((Const *) est)->constisnull)
1461 /* NULL indicates LIMIT ALL, ie, no limit */
1462 *count_est = 0; /* treat as not present */
1466 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1467 if (*count_est <= 0)
1468 *count_est = 1; /* force to at least 1 */
1472 *count_est = -1; /* can't estimate */
1475 *count_est = 0; /* not present */
1477 if (parse->limitOffset)
1479 est = estimate_expression_value(root, parse->limitOffset);
1480 if (est && IsA(est, Const))
1482 if (((Const *) est)->constisnull)
1484 /* Treat NULL as no offset; the executor will too */
1485 *offset_est = 0; /* treat as not present */
1489 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1490 if (*offset_est < 0)
1491 *offset_est = 0; /* less than 0 is same as 0 */
1495 *offset_est = -1; /* can't estimate */
1498 *offset_est = 0; /* not present */
1500 if (*count_est != 0)
1503 * A LIMIT clause limits the absolute number of tuples returned.
1504 * However, if it's not a constant LIMIT then we have to guess; for
1505 * lack of a better idea, assume 10% of the plan's result is wanted.
1507 if (*count_est < 0 || *offset_est < 0)
1509 /* LIMIT or OFFSET is an expression ... punt ... */
1510 limit_fraction = 0.10;
1514 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1515 limit_fraction = (double) *count_est + (double) *offset_est;
1519 * If we have absolute limits from both caller and LIMIT, use the
1520 * smaller value; likewise if they are both fractional. If one is
1521 * fractional and the other absolute, we can't easily determine which
1522 * is smaller, but we use the heuristic that the absolute will usually
1525 if (tuple_fraction >= 1.0)
1527 if (limit_fraction >= 1.0)
1530 tuple_fraction = Min(tuple_fraction, limit_fraction);
1534 /* caller absolute, limit fractional; use caller's value */
1537 else if (tuple_fraction > 0.0)
1539 if (limit_fraction >= 1.0)
1541 /* caller fractional, limit absolute; use limit */
1542 tuple_fraction = limit_fraction;
1546 /* both fractional */
1547 tuple_fraction = Min(tuple_fraction, limit_fraction);
1552 /* no info from caller, just use limit */
1553 tuple_fraction = limit_fraction;
1556 else if (*offset_est != 0 && tuple_fraction > 0.0)
1559 * We have an OFFSET but no LIMIT. This acts entirely differently
1560 * from the LIMIT case: here, we need to increase rather than decrease
1561 * the caller's tuple_fraction, because the OFFSET acts to cause more
1562 * tuples to be fetched instead of fewer. This only matters if we got
1563 * a tuple_fraction > 0, however.
1565 * As above, use 10% if OFFSET is present but unestimatable.
1567 if (*offset_est < 0)
1568 limit_fraction = 0.10;
1570 limit_fraction = (double) *offset_est;
1573 * If we have absolute counts from both caller and OFFSET, add them
1574 * together; likewise if they are both fractional. If one is
1575 * fractional and the other absolute, we want to take the larger, and
1576 * we heuristically assume that's the fractional one.
1578 if (tuple_fraction >= 1.0)
1580 if (limit_fraction >= 1.0)
1582 /* both absolute, so add them together */
1583 tuple_fraction += limit_fraction;
1587 /* caller absolute, limit fractional; use limit */
1588 tuple_fraction = limit_fraction;
1593 if (limit_fraction >= 1.0)
1595 /* caller fractional, limit absolute; use caller's value */
1599 /* both fractional, so add them together */
1600 tuple_fraction += limit_fraction;
1601 if (tuple_fraction >= 1.0)
1602 tuple_fraction = 0.0; /* assume fetch all */
1607 return tuple_fraction;
1612 * preprocess_groupclause - do preparatory work on GROUP BY clause
1614 * The idea here is to adjust the ordering of the GROUP BY elements
1615 * (which in itself is semantically insignificant) to match ORDER BY,
1616 * thereby allowing a single sort operation to both implement the ORDER BY
1617 * requirement and set up for a Unique step that implements GROUP BY.
1619 * In principle it might be interesting to consider other orderings of the
1620 * GROUP BY elements, which could match the sort ordering of other
1621 * possible plans (eg an indexscan) and thereby reduce cost. We don't
1622 * bother with that, though. Hashed grouping will frequently win anyway.
1624 * Note: we need no comparable processing of the distinctClause because
1625 * the parser already enforced that that matches ORDER BY.
1628 preprocess_groupclause(PlannerInfo *root)
1630 Query *parse = root->parse;
1631 List *new_groupclause;
1636 /* If no ORDER BY, nothing useful to do here */
1637 if (parse->sortClause == NIL)
1641 * Scan the ORDER BY clause and construct a list of matching GROUP BY
1642 * items, but only as far as we can make a matching prefix.
1644 * This code assumes that the sortClause contains no duplicate items.
1646 new_groupclause = NIL;
1647 foreach(sl, parse->sortClause)
1649 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
1651 foreach(gl, parse->groupClause)
1653 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1657 new_groupclause = lappend(new_groupclause, gc);
1662 break; /* no match, so stop scanning */
1665 /* Did we match all of the ORDER BY list, or just some of it? */
1666 partial_match = (sl != NULL);
1668 /* If no match at all, no point in reordering GROUP BY */
1669 if (new_groupclause == NIL)
1673 * Add any remaining GROUP BY items to the new list, but only if we
1674 * were able to make a complete match. In other words, we only
1675 * rearrange the GROUP BY list if the result is that one list is a
1676 * prefix of the other --- otherwise there's no possibility of a
1677 * common sort. Also, give up if there are any non-sortable GROUP BY
1678 * items, since then there's no hope anyway.
1680 foreach(gl, parse->groupClause)
1682 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1684 if (list_member_ptr(new_groupclause, gc))
1685 continue; /* it matched an ORDER BY item */
1687 return; /* give up, no common sort possible */
1688 if (!OidIsValid(gc->sortop))
1689 return; /* give up, GROUP BY can't be sorted */
1690 new_groupclause = lappend(new_groupclause, gc);
1693 /* Success --- install the rearranged GROUP BY list */
1694 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
1695 parse->groupClause = new_groupclause;
1699 * choose_hashed_grouping - should we use hashed grouping?
1701 * Note: this is only applied when both alternatives are actually feasible.
1704 choose_hashed_grouping(PlannerInfo *root,
1705 double tuple_fraction, double limit_tuples,
1706 Path *cheapest_path, Path *sorted_path,
1707 double dNumGroups, AggClauseCounts *agg_counts)
1709 int numGroupCols = list_length(root->parse->groupClause);
1710 double cheapest_path_rows;
1711 int cheapest_path_width;
1713 List *target_pathkeys;
1714 List *current_pathkeys;
1718 /* Prefer sorting when enable_hashagg is off */
1719 if (!enable_hashagg)
1723 * Don't do it if it doesn't look like the hashtable will fit into
1726 * Beware here of the possibility that cheapest_path->parent is NULL. This
1727 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1729 if (cheapest_path->parent)
1731 cheapest_path_rows = cheapest_path->parent->rows;
1732 cheapest_path_width = cheapest_path->parent->width;
1736 cheapest_path_rows = 1; /* assume non-set result */
1737 cheapest_path_width = 100; /* arbitrary */
1740 /* Estimate per-hash-entry space at tuple width... */
1741 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1742 /* plus space for pass-by-ref transition values... */
1743 hashentrysize += agg_counts->transitionSpace;
1744 /* plus the per-hash-entry overhead */
1745 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1747 if (hashentrysize * dNumGroups > work_mem * 1024L)
1751 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
1752 * DISTINCT and ORDER BY as the assumed required output sort order.
1753 * This is an oversimplification because the DISTINCT might get
1754 * implemented via hashing, but it's not clear that the case is common
1755 * enough (or that our estimates are good enough) to justify trying to
1758 if (list_length(root->distinct_pathkeys) >
1759 list_length(root->sort_pathkeys))
1760 target_pathkeys = root->distinct_pathkeys;
1762 target_pathkeys = root->sort_pathkeys;
1765 * See if the estimated cost is no more than doing it the other way. While
1766 * avoiding the need for sorted input is usually a win, the fact that the
1767 * output won't be sorted may be a loss; so we need to do an actual cost
1770 * We need to consider cheapest_path + hashagg [+ final sort] versus
1771 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1772 * presorted_path + group or agg [+ final sort] where brackets indicate a
1773 * step that may not be needed. We assume query_planner() will have
1774 * returned a presorted path only if it's a winner compared to
1775 * cheapest_path for this purpose.
1777 * These path variables are dummies that just hold cost fields; we don't
1778 * make actual Paths for these steps.
1780 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1781 numGroupCols, dNumGroups,
1782 cheapest_path->startup_cost, cheapest_path->total_cost,
1783 cheapest_path_rows);
1784 /* Result of hashed agg is always unsorted */
1785 if (target_pathkeys)
1786 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
1787 dNumGroups, cheapest_path_width, limit_tuples);
1791 sorted_p.startup_cost = sorted_path->startup_cost;
1792 sorted_p.total_cost = sorted_path->total_cost;
1793 current_pathkeys = sorted_path->pathkeys;
1797 sorted_p.startup_cost = cheapest_path->startup_cost;
1798 sorted_p.total_cost = cheapest_path->total_cost;
1799 current_pathkeys = cheapest_path->pathkeys;
1801 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1803 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1804 cheapest_path_rows, cheapest_path_width, -1.0);
1805 current_pathkeys = root->group_pathkeys;
1808 if (root->parse->hasAggs)
1809 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1810 numGroupCols, dNumGroups,
1811 sorted_p.startup_cost, sorted_p.total_cost,
1812 cheapest_path_rows);
1814 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1815 sorted_p.startup_cost, sorted_p.total_cost,
1816 cheapest_path_rows);
1817 /* The Agg or Group node will preserve ordering */
1818 if (target_pathkeys &&
1819 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
1820 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
1821 dNumGroups, cheapest_path_width, limit_tuples);
1824 * Now make the decision using the top-level tuple fraction. First we
1825 * have to convert an absolute count (LIMIT) into fractional form.
1827 if (tuple_fraction >= 1.0)
1828 tuple_fraction /= dNumGroups;
1830 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1831 tuple_fraction) < 0)
1833 /* Hashed is cheaper, so use it */
1840 * choose_hashed_distinct - should we use hashing for DISTINCT?
1842 * This is fairly similar to choose_hashed_grouping, but there are enough
1843 * differences that it doesn't seem worth trying to unify the two functions.
1845 * But note that making the two choices independently is a bit bogus in
1846 * itself. If the two could be combined into a single choice operation
1847 * it'd probably be better, but that seems far too unwieldy to be practical,
1848 * especially considering that the combination of GROUP BY and DISTINCT
1849 * isn't very common in real queries. By separating them, we are giving
1850 * extra preference to using a sorting implementation when a common sort key
1851 * is available ... and that's not necessarily wrong anyway.
1853 * Note: this is only applied when both alternatives are actually feasible.
1856 choose_hashed_distinct(PlannerInfo *root,
1857 Plan *input_plan, List *input_pathkeys,
1858 double tuple_fraction, double limit_tuples,
1859 double dNumDistinctRows)
1861 int numDistinctCols = list_length(root->parse->distinctClause);
1863 List *current_pathkeys;
1864 List *needed_pathkeys;
1868 /* Prefer sorting when enable_hashagg is off */
1869 if (!enable_hashagg)
1873 * Don't do it if it doesn't look like the hashtable will fit into
1876 hashentrysize = MAXALIGN(input_plan->plan_width) + MAXALIGN(sizeof(MinimalTupleData));
1878 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
1882 * See if the estimated cost is no more than doing it the other way. While
1883 * avoiding the need for sorted input is usually a win, the fact that the
1884 * output won't be sorted may be a loss; so we need to do an actual cost
1887 * We need to consider input_plan + hashagg [+ final sort] versus
1888 * input_plan [+ sort] + group [+ final sort] where brackets indicate
1889 * a step that may not be needed.
1891 * These path variables are dummies that just hold cost fields; we don't
1892 * make actual Paths for these steps.
1894 cost_agg(&hashed_p, root, AGG_HASHED, 0,
1895 numDistinctCols, dNumDistinctRows,
1896 input_plan->startup_cost, input_plan->total_cost,
1897 input_plan->plan_rows);
1899 * Result of hashed agg is always unsorted, so if ORDER BY is present
1900 * we need to charge for the final sort.
1902 if (root->parse->sortClause)
1903 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1904 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1907 * Now for the GROUP case. See comments in grouping_planner about the
1908 * sorting choices here --- this code should match that code.
1910 sorted_p.startup_cost = input_plan->startup_cost;
1911 sorted_p.total_cost = input_plan->total_cost;
1912 current_pathkeys = input_pathkeys;
1913 if (root->parse->hasDistinctOn &&
1914 list_length(root->distinct_pathkeys) <
1915 list_length(root->sort_pathkeys))
1916 needed_pathkeys = root->sort_pathkeys;
1918 needed_pathkeys = root->distinct_pathkeys;
1919 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1921 if (list_length(root->distinct_pathkeys) >=
1922 list_length(root->sort_pathkeys))
1923 current_pathkeys = root->distinct_pathkeys;
1925 current_pathkeys = root->sort_pathkeys;
1926 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
1927 input_plan->plan_rows, input_plan->plan_width, -1.0);
1929 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
1930 sorted_p.startup_cost, sorted_p.total_cost,
1931 input_plan->plan_rows);
1932 if (root->parse->sortClause &&
1933 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1934 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1935 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1938 * Now make the decision using the top-level tuple fraction. First we
1939 * have to convert an absolute count (LIMIT) into fractional form.
1941 if (tuple_fraction >= 1.0)
1942 tuple_fraction /= dNumDistinctRows;
1944 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1945 tuple_fraction) < 0)
1947 /* Hashed is cheaper, so use it */
1954 * make_subplanTargetList
1955 * Generate appropriate target list when grouping is required.
1957 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1958 * above the result of query_planner, we typically want to pass a different
1959 * target list to query_planner than the outer plan nodes should have.
1960 * This routine generates the correct target list for the subplan.
1962 * The initial target list passed from the parser already contains entries
1963 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1964 * for variables used only in HAVING clauses; so we need to add those
1965 * variables to the subplan target list. Also, we flatten all expressions
1966 * except GROUP BY items into their component variables; the other expressions
1967 * will be computed by the inserted nodes rather than by the subplan.
1968 * For example, given a query like
1969 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1970 * we want to pass this targetlist to the subplan:
1972 * where the a+b target will be used by the Sort/Group steps, and the
1973 * other targets will be used for computing the final results. (In the
1974 * above example we could theoretically suppress the a and b targets and
1975 * pass down only c,d,a+b, but it's not really worth the trouble to
1976 * eliminate simple var references from the subplan. We will avoid doing
1977 * the extra computation to recompute a+b at the outer level; see
1978 * fix_upper_expr() in setrefs.c.)
1980 * If we are grouping or aggregating, *and* there are no non-Var grouping
1981 * expressions, then the returned tlist is effectively dummy; we do not
1982 * need to force it to be evaluated, because all the Vars it contains
1983 * should be present in the output of query_planner anyway.
1985 * 'tlist' is the query's target list.
1986 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1987 * expressions (if there are any) in the subplan's target list.
1988 * 'need_tlist_eval' is set true if we really need to evaluate the
1991 * The result is the targetlist to be passed to the subplan.
1995 make_subplanTargetList(PlannerInfo *root,
1997 AttrNumber **groupColIdx,
1998 bool *need_tlist_eval)
2000 Query *parse = root->parse;
2005 *groupColIdx = NULL;
2008 * If we're not grouping or aggregating, there's nothing to do here;
2009 * query_planner should receive the unmodified target list.
2011 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
2013 *need_tlist_eval = true;
2018 * Otherwise, start with a "flattened" tlist (having just the vars
2019 * mentioned in the targetlist and HAVING qual --- but not upper-level
2020 * Vars; they will be replaced by Params later on).
2022 sub_tlist = flatten_tlist(tlist);
2023 extravars = pull_var_clause(parse->havingQual, false);
2024 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2025 list_free(extravars);
2026 *need_tlist_eval = false; /* only eval if not flat tlist */
2029 * If grouping, create sub_tlist entries for all GROUP BY expressions
2030 * (GROUP BY items that are simple Vars should be in the list already),
2031 * and make an array showing where the group columns are in the sub_tlist.
2033 numCols = list_length(parse->groupClause);
2037 AttrNumber *grpColIdx;
2040 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2041 *groupColIdx = grpColIdx;
2043 foreach(gl, parse->groupClause)
2045 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2046 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2047 TargetEntry *te = NULL;
2050 * Find or make a matching sub_tlist entry. If the groupexpr
2051 * isn't a Var, no point in searching. (Note that the parser
2052 * won't make multiple groupClause entries for the same TLE.)
2054 if (groupexpr && IsA(groupexpr, Var))
2058 foreach(sl, sub_tlist)
2060 TargetEntry *lte = (TargetEntry *) lfirst(sl);
2062 if (equal(groupexpr, lte->expr))
2071 te = makeTargetEntry((Expr *) groupexpr,
2072 list_length(sub_tlist) + 1,
2075 sub_tlist = lappend(sub_tlist, te);
2076 *need_tlist_eval = true; /* it's not flat anymore */
2079 /* and save its resno */
2080 grpColIdx[keyno++] = te->resno;
2088 * locate_grouping_columns
2089 * Locate grouping columns in the tlist chosen by query_planner.
2091 * This is only needed if we don't use the sub_tlist chosen by
2092 * make_subplanTargetList. We have to forget the column indexes found
2093 * by that routine and re-locate the grouping vars in the real sub_tlist.
2096 locate_grouping_columns(PlannerInfo *root,
2099 AttrNumber *groupColIdx)
2105 * No work unless grouping.
2107 if (!root->parse->groupClause)
2109 Assert(groupColIdx == NULL);
2112 Assert(groupColIdx != NULL);
2114 foreach(gl, root->parse->groupClause)
2116 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2117 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2118 TargetEntry *te = NULL;
2121 foreach(sl, sub_tlist)
2123 te = (TargetEntry *) lfirst(sl);
2124 if (equal(groupexpr, te->expr))
2128 elog(ERROR, "failed to locate grouping columns");
2130 groupColIdx[keyno++] = te->resno;
2135 * postprocess_setop_tlist
2136 * Fix up targetlist returned by plan_set_operations().
2138 * We need to transpose sort key info from the orig_tlist into new_tlist.
2139 * NOTE: this would not be good enough if we supported resjunk sort keys
2140 * for results of set operations --- then, we'd need to project a whole
2141 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2142 * find any resjunk columns in orig_tlist.
2145 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2148 ListCell *orig_tlist_item = list_head(orig_tlist);
2150 foreach(l, new_tlist)
2152 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2153 TargetEntry *orig_tle;
2155 /* ignore resjunk columns in setop result */
2156 if (new_tle->resjunk)
2159 Assert(orig_tlist_item != NULL);
2160 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2161 orig_tlist_item = lnext(orig_tlist_item);
2162 if (orig_tle->resjunk) /* should not happen */
2163 elog(ERROR, "resjunk output columns are not implemented");
2164 Assert(new_tle->resno == orig_tle->resno);
2165 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2167 if (orig_tlist_item != NULL)
2168 elog(ERROR, "resjunk output columns are not implemented");