1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * src/backend/optimizer/plan/planner.c
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/plancat.h"
30 #include "optimizer/planmain.h"
31 #include "optimizer/planner.h"
32 #include "optimizer/prep.h"
33 #include "optimizer/subselect.h"
34 #include "optimizer/tlist.h"
35 #include "optimizer/var.h"
36 #ifdef OPTIMIZER_DEBUG
37 #include "nodes/print.h"
39 #include "parser/analyze.h"
40 #include "parser/parse_expr.h"
41 #include "parser/parse_oper.h"
42 #include "parser/parsetree.h"
43 #include "utils/lsyscache.h"
44 #include "utils/rel.h"
45 #include "utils/syscache.h"
49 double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
51 /* Hook for plugins to get control in planner() */
52 planner_hook_type planner_hook = NULL;
55 /* Expression kind codes for preprocess_expression */
56 #define EXPRKIND_QUAL 0
57 #define EXPRKIND_TARGET 1
58 #define EXPRKIND_RTFUNC 2
59 #define EXPRKIND_VALUES 3
60 #define EXPRKIND_LIMIT 4
61 #define EXPRKIND_APPINFO 5
64 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
65 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
66 static Plan *inheritance_planner(PlannerInfo *root);
67 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
68 static bool is_dummy_plan(Plan *plan);
69 static void preprocess_rowmarks(PlannerInfo *root);
70 static double preprocess_limit(PlannerInfo *root,
71 double tuple_fraction,
72 int64 *offset_est, int64 *count_est);
73 static void preprocess_groupclause(PlannerInfo *root);
74 static bool choose_hashed_grouping(PlannerInfo *root,
75 double tuple_fraction, double limit_tuples,
76 double path_rows, int path_width,
77 Path *cheapest_path, Path *sorted_path,
78 double dNumGroups, AggClauseCosts *agg_costs);
79 static bool choose_hashed_distinct(PlannerInfo *root,
80 double tuple_fraction, double limit_tuples,
81 double path_rows, int path_width,
82 Cost cheapest_startup_cost, Cost cheapest_total_cost,
83 Cost sorted_startup_cost, Cost sorted_total_cost,
84 List *sorted_pathkeys,
85 double dNumDistinctRows);
86 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
87 AttrNumber **groupColIdx, bool *need_tlist_eval);
88 static void locate_grouping_columns(PlannerInfo *root,
91 AttrNumber *groupColIdx);
92 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
93 static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
94 static List *add_volatile_sort_exprs(List *window_tlist, List *tlist,
96 static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
97 List *tlist, bool canonicalize);
98 static void get_column_info_for_window(PlannerInfo *root, WindowClause *wc,
100 int numSortCols, AttrNumber *sortColIdx,
102 AttrNumber **partColIdx,
105 AttrNumber **ordColIdx,
109 /*****************************************************************************
111 * Query optimizer entry point
113 * To support loadable plugins that monitor or modify planner behavior,
114 * we provide a hook variable that lets a plugin get control before and
115 * after the standard planning process. The plugin would normally call
116 * standard_planner().
118 * Note to plugin authors: standard_planner() scribbles on its Query input,
119 * so you'd better copy that data structure if you want to plan more than once.
121 *****************************************************************************/
123 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
128 result = (*planner_hook) (parse, cursorOptions, boundParams);
130 result = standard_planner(parse, cursorOptions, boundParams);
135 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
139 double tuple_fraction;
146 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
147 if (parse->utilityStmt &&
148 IsA(parse->utilityStmt, DeclareCursorStmt))
149 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
152 * Set up global state for this planner invocation. This data is needed
153 * across all levels of sub-Query that might exist in the given command,
154 * so we keep it in a separate struct that's linked to by each per-Query
157 glob = makeNode(PlannerGlobal);
159 glob->boundParams = boundParams;
160 glob->paramlist = NIL;
161 glob->subplans = NIL;
162 glob->subrtables = NIL;
163 glob->subrowmarks = NIL;
164 glob->rewindPlanIDs = NULL;
165 glob->finalrtable = NIL;
166 glob->finalrowmarks = NIL;
167 glob->resultRelations = NIL;
168 glob->relationOids = NIL;
169 glob->invalItems = NIL;
171 glob->lastRowMarkId = 0;
172 glob->transientPlan = false;
174 /* Determine what fraction of the plan is likely to be scanned */
175 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
178 * We have no real idea how many tuples the user will ultimately FETCH
179 * from a cursor, but it is often the case that he doesn't want 'em
180 * all, or would prefer a fast-start plan anyway so that he can
181 * process some of the tuples sooner. Use a GUC parameter to decide
182 * what fraction to optimize for.
184 tuple_fraction = cursor_tuple_fraction;
187 * We document cursor_tuple_fraction as simply being a fraction, which
188 * means the edge cases 0 and 1 have to be treated specially here. We
189 * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
191 if (tuple_fraction >= 1.0)
192 tuple_fraction = 0.0;
193 else if (tuple_fraction <= 0.0)
194 tuple_fraction = 1e-10;
198 /* Default assumption is we need all the tuples */
199 tuple_fraction = 0.0;
202 /* primary planning entry point (may recurse for subqueries) */
203 top_plan = subquery_planner(glob, parse, NULL,
204 false, tuple_fraction, &root);
207 * If creating a plan for a scrollable cursor, make sure it can run
208 * backwards on demand. Add a Material node at the top at need.
210 if (cursorOptions & CURSOR_OPT_SCROLL)
212 if (!ExecSupportsBackwardScan(top_plan))
213 top_plan = materialize_finished_plan(top_plan);
216 /* final cleanup of the plan */
217 Assert(glob->finalrtable == NIL);
218 Assert(glob->finalrowmarks == NIL);
219 Assert(glob->resultRelations == NIL);
220 top_plan = set_plan_references(glob, top_plan,
223 /* ... and the subplans (both regular subplans and initplans) */
224 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
225 Assert(list_length(glob->subplans) == list_length(glob->subrowmarks));
226 lrt = list_head(glob->subrtables);
227 lrm = list_head(glob->subrowmarks);
228 foreach(lp, glob->subplans)
230 Plan *subplan = (Plan *) lfirst(lp);
231 List *subrtable = (List *) lfirst(lrt);
232 List *subrowmark = (List *) lfirst(lrm);
234 lfirst(lp) = set_plan_references(glob, subplan,
235 subrtable, subrowmark);
240 /* build the PlannedStmt result */
241 result = makeNode(PlannedStmt);
243 result->commandType = parse->commandType;
244 result->hasReturning = (parse->returningList != NIL);
245 result->hasModifyingCTE = parse->hasModifyingCTE;
246 result->canSetTag = parse->canSetTag;
247 result->transientPlan = glob->transientPlan;
248 result->planTree = top_plan;
249 result->rtable = glob->finalrtable;
250 result->resultRelations = glob->resultRelations;
251 result->utilityStmt = parse->utilityStmt;
252 result->intoClause = parse->intoClause;
253 result->subplans = glob->subplans;
254 result->rewindPlanIDs = glob->rewindPlanIDs;
255 result->rowMarks = glob->finalrowmarks;
256 result->relationOids = glob->relationOids;
257 result->invalItems = glob->invalItems;
258 result->nParamExec = list_length(glob->paramlist);
264 /*--------------------
266 * Invokes the planner on a subquery. We recurse to here for each
267 * sub-SELECT found in the query tree.
269 * glob is the global state for the current planner run.
270 * parse is the querytree produced by the parser & rewriter.
271 * parent_root is the immediate parent Query's info (NULL at the top level).
272 * hasRecursion is true if this is a recursive WITH query.
273 * tuple_fraction is the fraction of tuples we expect will be retrieved.
274 * tuple_fraction is interpreted as explained for grouping_planner, below.
276 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
277 * among other things this tells the output sort ordering of the plan.
279 * Basically, this routine does the stuff that should only be done once
280 * per Query object. It then calls grouping_planner. At one time,
281 * grouping_planner could be invoked recursively on the same Query object;
282 * that's not currently true, but we keep the separation between the two
283 * routines anyway, in case we need it again someday.
285 * subquery_planner will be called recursively to handle sub-Query nodes
286 * found within the query's expressions and rangetable.
288 * Returns a query plan.
289 *--------------------
292 subquery_planner(PlannerGlobal *glob, Query *parse,
293 PlannerInfo *parent_root,
294 bool hasRecursion, double tuple_fraction,
295 PlannerInfo **subroot)
297 int num_old_subplans = list_length(glob->subplans);
304 /* Create a PlannerInfo data structure for this subquery */
305 root = makeNode(PlannerInfo);
308 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
309 root->parent_root = parent_root;
310 root->planner_cxt = CurrentMemoryContext;
311 root->init_plans = NIL;
312 root->cte_plan_ids = NIL;
313 root->eq_classes = NIL;
314 root->append_rel_list = NIL;
315 root->rowMarks = NIL;
316 root->hasInheritedTarget = false;
318 root->hasRecursion = hasRecursion;
320 root->wt_param_id = SS_assign_special_param(root);
322 root->wt_param_id = -1;
323 root->non_recursive_plan = NULL;
326 * If there is a WITH list, process each WITH query and build an initplan
327 * SubPlan structure for it.
330 SS_process_ctes(root);
333 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
334 * to transform them into joins. Note that this step does not descend
335 * into subqueries; if we pull up any subqueries below, their SubLinks are
336 * processed just before pulling them up.
338 if (parse->hasSubLinks)
339 pull_up_sublinks(root);
342 * Scan the rangetable for set-returning functions, and inline them if
343 * possible (producing subqueries that might get pulled up next).
344 * Recursion issues here are handled in the same way as for SubLinks.
346 inline_set_returning_functions(root);
349 * Check to see if any subqueries in the jointree can be merged into this
352 parse->jointree = (FromExpr *)
353 pull_up_subqueries(root, (Node *) parse->jointree, NULL, NULL);
356 * If this is a simple UNION ALL query, flatten it into an appendrel. We
357 * do this now because it requires applying pull_up_subqueries to the leaf
358 * queries of the UNION ALL, which weren't touched above because they
359 * weren't referenced by the jointree (they will be after we do this).
361 if (parse->setOperations)
362 flatten_simple_union_all(root);
365 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
366 * avoid the expense of doing flatten_join_alias_vars(). Also check for
367 * outer joins --- if none, we can skip reduce_outer_joins(). This must be
368 * done after we have done pull_up_subqueries, of course.
370 root->hasJoinRTEs = false;
371 hasOuterJoins = false;
372 foreach(l, parse->rtable)
374 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
376 if (rte->rtekind == RTE_JOIN)
378 root->hasJoinRTEs = true;
379 if (IS_OUTER_JOIN(rte->jointype))
381 hasOuterJoins = true;
382 /* Can quit scanning once we find an outer join */
389 * Preprocess RowMark information. We need to do this after subquery
390 * pullup (so that all non-inherited RTEs are present) and before
391 * inheritance expansion (so that the info is available for
392 * expand_inherited_tables to examine and modify).
394 preprocess_rowmarks(root);
397 * Expand any rangetable entries that are inheritance sets into "append
398 * relations". This can add entries to the rangetable, but they must be
399 * plain base relations not joins, so it's OK (and marginally more
400 * efficient) to do it after checking for join RTEs. We must do it after
401 * pulling up subqueries, else we'd fail to handle inherited tables in
404 expand_inherited_tables(root);
407 * Set hasHavingQual to remember if HAVING clause is present. Needed
408 * because preprocess_expression will reduce a constant-true condition to
409 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
411 root->hasHavingQual = (parse->havingQual != NULL);
413 /* Clear this flag; might get set in distribute_qual_to_rels */
414 root->hasPseudoConstantQuals = false;
417 * Do expression preprocessing on targetlist and quals, as well as other
418 * random expressions in the querytree. Note that we do not need to
419 * handle sort/group expressions explicitly, because they are actually
420 * part of the targetlist.
422 parse->targetList = (List *)
423 preprocess_expression(root, (Node *) parse->targetList,
426 parse->returningList = (List *)
427 preprocess_expression(root, (Node *) parse->returningList,
430 preprocess_qual_conditions(root, (Node *) parse->jointree);
432 parse->havingQual = preprocess_expression(root, parse->havingQual,
435 foreach(l, parse->windowClause)
437 WindowClause *wc = (WindowClause *) lfirst(l);
439 /* partitionClause/orderClause are sort/group expressions */
440 wc->startOffset = preprocess_expression(root, wc->startOffset,
442 wc->endOffset = preprocess_expression(root, wc->endOffset,
446 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
448 parse->limitCount = preprocess_expression(root, parse->limitCount,
451 root->append_rel_list = (List *)
452 preprocess_expression(root, (Node *) root->append_rel_list,
455 /* Also need to preprocess expressions for function and values RTEs */
456 foreach(l, parse->rtable)
458 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
460 if (rte->rtekind == RTE_FUNCTION)
461 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
463 else if (rte->rtekind == RTE_VALUES)
464 rte->values_lists = (List *)
465 preprocess_expression(root, (Node *) rte->values_lists,
470 * In some cases we may want to transfer a HAVING clause into WHERE. We
471 * cannot do so if the HAVING clause contains aggregates (obviously) or
472 * volatile functions (since a HAVING clause is supposed to be executed
473 * only once per group). Also, it may be that the clause is so expensive
474 * to execute that we're better off doing it only once per group, despite
475 * the loss of selectivity. This is hard to estimate short of doing the
476 * entire planning process twice, so we use a heuristic: clauses
477 * containing subplans are left in HAVING. Otherwise, we move or copy the
478 * HAVING clause into WHERE, in hopes of eliminating tuples before
479 * aggregation instead of after.
481 * If the query has explicit grouping then we can simply move such a
482 * clause into WHERE; any group that fails the clause will not be in the
483 * output because none of its tuples will reach the grouping or
484 * aggregation stage. Otherwise we must have a degenerate (variable-free)
485 * HAVING clause, which we put in WHERE so that query_planner() can use it
486 * in a gating Result node, but also keep in HAVING to ensure that we
487 * don't emit a bogus aggregated row. (This could be done better, but it
488 * seems not worth optimizing.)
490 * Note that both havingQual and parse->jointree->quals are in
491 * implicitly-ANDed-list form at this point, even though they are declared
495 foreach(l, (List *) parse->havingQual)
497 Node *havingclause = (Node *) lfirst(l);
499 if (contain_agg_clause(havingclause) ||
500 contain_volatile_functions(havingclause) ||
501 contain_subplans(havingclause))
503 /* keep it in HAVING */
504 newHaving = lappend(newHaving, havingclause);
506 else if (parse->groupClause)
508 /* move it to WHERE */
509 parse->jointree->quals = (Node *)
510 lappend((List *) parse->jointree->quals, havingclause);
514 /* put a copy in WHERE, keep it in HAVING */
515 parse->jointree->quals = (Node *)
516 lappend((List *) parse->jointree->quals,
517 copyObject(havingclause));
518 newHaving = lappend(newHaving, havingclause);
521 parse->havingQual = (Node *) newHaving;
524 * If we have any outer joins, try to reduce them to plain inner joins.
525 * This step is most easily done after we've done expression
529 reduce_outer_joins(root);
532 * Do the main planning. If we have an inherited target relation, that
533 * needs special processing, else go straight to grouping_planner.
535 if (parse->resultRelation &&
536 rt_fetch(parse->resultRelation, parse->rtable)->inh)
537 plan = inheritance_planner(root);
540 plan = grouping_planner(root, tuple_fraction);
541 /* If it's not SELECT, we need a ModifyTable node */
542 if (parse->commandType != CMD_SELECT)
544 List *returningLists;
548 * Deal with the RETURNING clause if any. It's convenient to pass
549 * the returningList through setrefs.c now rather than at top
550 * level (if we waited, handling inherited UPDATE/DELETE would be
553 if (parse->returningList)
557 Assert(parse->resultRelation);
558 rlist = set_returning_clause_references(root->glob,
559 parse->returningList,
561 parse->resultRelation);
562 returningLists = list_make1(rlist);
565 returningLists = NIL;
568 * If there was a FOR UPDATE/SHARE clause, the LockRows node will
569 * have dealt with fetching non-locked marked rows, else we need
570 * to have ModifyTable do that.
575 rowMarks = root->rowMarks;
577 plan = (Plan *) make_modifytable(parse->commandType,
579 list_make1_int(parse->resultRelation),
583 SS_assign_special_param(root));
588 * If any subplans were generated, or if there are any parameters to worry
589 * about, build initPlan list and extParam/allParam sets for plan nodes,
590 * and attach the initPlans to the top plan node.
592 if (list_length(glob->subplans) != num_old_subplans ||
593 root->glob->paramlist != NIL)
594 SS_finalize_plan(root, plan, true);
596 /* Return internal info if caller wants it */
604 * preprocess_expression
605 * Do subquery_planner's preprocessing work for an expression,
606 * which can be a targetlist, a WHERE clause (including JOIN/ON
607 * conditions), or a HAVING clause.
610 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
613 * Fall out quickly if expression is empty. This occurs often enough to
614 * be worth checking. Note that null->null is the correct conversion for
615 * implicit-AND result format, too.
621 * If the query has any join RTEs, replace join alias variables with
622 * base-relation variables. We must do this before sublink processing,
623 * else sublinks expanded out from join aliases wouldn't get processed. We
624 * can skip it in VALUES lists, however, since they can't contain any Vars
627 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
628 expr = flatten_join_alias_vars(root, expr);
631 * Simplify constant expressions.
633 * Note: an essential effect of this is to convert named-argument function
634 * calls to positional notation and insert the current actual values of
635 * any default arguments for functions. To ensure that happens, we *must*
636 * process all expressions here. Previous PG versions sometimes skipped
637 * const-simplification if it didn't seem worth the trouble, but we can't
640 * Note: this also flattens nested AND and OR expressions into N-argument
641 * form. All processing of a qual expression after this point must be
642 * careful to maintain AND/OR flatness --- that is, do not generate a tree
643 * with AND directly under AND, nor OR directly under OR.
645 expr = eval_const_expressions(root, expr);
648 * If it's a qual or havingQual, canonicalize it.
650 if (kind == EXPRKIND_QUAL)
652 expr = (Node *) canonicalize_qual((Expr *) expr);
654 #ifdef OPTIMIZER_DEBUG
655 printf("After canonicalize_qual()\n");
660 /* Expand SubLinks to SubPlans */
661 if (root->parse->hasSubLinks)
662 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
665 * XXX do not insert anything here unless you have grokked the comments in
666 * SS_replace_correlation_vars ...
669 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
670 if (root->query_level > 1)
671 expr = SS_replace_correlation_vars(root, expr);
674 * If it's a qual or havingQual, convert it to implicit-AND format. (We
675 * don't want to do this before eval_const_expressions, since the latter
676 * would be unable to simplify a top-level AND correctly. Also,
677 * SS_process_sublinks expects explicit-AND format.)
679 if (kind == EXPRKIND_QUAL)
680 expr = (Node *) make_ands_implicit((Expr *) expr);
686 * preprocess_qual_conditions
687 * Recursively scan the query's jointree and do subquery_planner's
688 * preprocessing work on each qual condition found therein.
691 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
695 if (IsA(jtnode, RangeTblRef))
697 /* nothing to do here */
699 else if (IsA(jtnode, FromExpr))
701 FromExpr *f = (FromExpr *) jtnode;
704 foreach(l, f->fromlist)
705 preprocess_qual_conditions(root, lfirst(l));
707 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
709 else if (IsA(jtnode, JoinExpr))
711 JoinExpr *j = (JoinExpr *) jtnode;
713 preprocess_qual_conditions(root, j->larg);
714 preprocess_qual_conditions(root, j->rarg);
716 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
719 elog(ERROR, "unrecognized node type: %d",
720 (int) nodeTag(jtnode));
724 * inheritance_planner
725 * Generate a plan in the case where the result relation is an
728 * We have to handle this case differently from cases where a source relation
729 * is an inheritance set. Source inheritance is expanded at the bottom of the
730 * plan tree (see allpaths.c), but target inheritance has to be expanded at
731 * the top. The reason is that for UPDATE, each target relation needs a
732 * different targetlist matching its own column set. Fortunately,
733 * the UPDATE/DELETE target can never be the nullable side of an outer join,
734 * so it's OK to generate the plan this way.
736 * Returns a query plan.
739 inheritance_planner(PlannerInfo *root)
741 Query *parse = root->parse;
742 int parentRTindex = parse->resultRelation;
743 List *subplans = NIL;
744 List *resultRelations = NIL;
745 List *returningLists = NIL;
752 foreach(l, root->append_rel_list)
754 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
757 /* append_rel_list contains all append rels; ignore others */
758 if (appinfo->parent_relid != parentRTindex)
762 * Generate modified query with this rel as target.
764 memcpy(&subroot, root, sizeof(PlannerInfo));
765 subroot.parse = (Query *)
766 adjust_appendrel_attrs((Node *) parse,
768 subroot.init_plans = NIL;
769 subroot.hasInheritedTarget = true;
770 /* We needn't modify the child's append_rel_list */
771 /* There shouldn't be any OJ info to translate, as yet */
772 Assert(subroot.join_info_list == NIL);
773 /* and we haven't created PlaceHolderInfos, either */
774 Assert(subroot.placeholder_list == NIL);
777 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
780 * If this child rel was excluded by constraint exclusion, exclude it
783 if (is_dummy_plan(subplan))
786 /* Save rtable from first rel for use below */
788 rtable = subroot.parse->rtable;
790 subplans = lappend(subplans, subplan);
792 /* Make sure any initplans from this rel get into the outer list */
793 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
795 /* Build list of target-relation RT indexes */
796 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
798 /* Build list of per-relation RETURNING targetlists */
799 if (parse->returningList)
803 rlist = set_returning_clause_references(root->glob,
804 subroot.parse->returningList,
806 appinfo->child_relid);
807 returningLists = lappend(returningLists, rlist);
811 /* Mark result as unordered (probably unnecessary) */
812 root->query_pathkeys = NIL;
815 * If we managed to exclude every child rel, return a dummy plan; it
816 * doesn't even need a ModifyTable node.
820 /* although dummy, it must have a valid tlist for executor */
821 tlist = preprocess_targetlist(root, parse->targetList);
822 return (Plan *) make_result(root,
824 (Node *) list_make1(makeBoolConst(false,
830 * Planning might have modified the rangetable, due to changes of the
831 * Query structures inside subquery RTEs. We have to ensure that this
832 * gets propagated back to the master copy. But can't do this until we
833 * are done planning, because all the calls to grouping_planner need
834 * virgin sub-Queries to work from. (We are effectively assuming that
835 * sub-Queries will get planned identically each time, or at least that
836 * the impacts on their rangetables will be the same each time.)
838 * XXX should clean this up someday
840 parse->rtable = rtable;
843 * If there was a FOR UPDATE/SHARE clause, the LockRows node will have
844 * dealt with fetching non-locked marked rows, else we need to have
845 * ModifyTable do that.
850 rowMarks = root->rowMarks;
852 /* And last, tack on a ModifyTable node to do the UPDATE/DELETE work */
853 return (Plan *) make_modifytable(parse->commandType,
859 SS_assign_special_param(root));
862 /*--------------------
864 * Perform planning steps related to grouping, aggregation, etc.
865 * This primarily means adding top-level processing to the basic
866 * query plan produced by query_planner.
868 * tuple_fraction is the fraction of tuples we expect will be retrieved
870 * tuple_fraction is interpreted as follows:
871 * 0: expect all tuples to be retrieved (normal case)
872 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
873 * from the plan to be retrieved
874 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
875 * expected to be retrieved (ie, a LIMIT specification)
877 * Returns a query plan. Also, root->query_pathkeys is returned as the
878 * actual output ordering of the plan (in pathkey format).
879 *--------------------
882 grouping_planner(PlannerInfo *root, double tuple_fraction)
884 Query *parse = root->parse;
885 List *tlist = parse->targetList;
886 int64 offset_est = 0;
888 double limit_tuples = -1.0;
890 List *current_pathkeys;
891 double dNumGroups = 0;
892 bool use_hashed_distinct = false;
893 bool tested_hashed_distinct = false;
895 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
896 if (parse->limitCount || parse->limitOffset)
898 tuple_fraction = preprocess_limit(root, tuple_fraction,
899 &offset_est, &count_est);
902 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
903 * estimate the effects of using a bounded sort.
905 if (count_est > 0 && offset_est >= 0)
906 limit_tuples = (double) count_est + (double) offset_est;
909 if (parse->setOperations)
911 List *set_sortclauses;
914 * If there's a top-level ORDER BY, assume we have to fetch all the
915 * tuples. This might be too simplistic given all the hackery below
916 * to possibly avoid the sort; but the odds of accurate estimates here
917 * are pretty low anyway.
919 if (parse->sortClause)
920 tuple_fraction = 0.0;
923 * Construct the plan for set operations. The result will not need
924 * any work except perhaps a top-level sort and/or LIMIT. Note that
925 * any special work for recursive unions is the responsibility of
926 * plan_set_operations.
928 result_plan = plan_set_operations(root, tuple_fraction,
932 * Calculate pathkeys representing the sort order (if any) of the set
933 * operation's result. We have to do this before overwriting the sort
936 current_pathkeys = make_pathkeys_for_sortclauses(root,
938 result_plan->targetlist,
942 * We should not need to call preprocess_targetlist, since we must be
943 * in a SELECT query node. Instead, use the targetlist returned by
944 * plan_set_operations (since this tells whether it returned any
945 * resjunk columns!), and transfer any sort key information from the
948 Assert(parse->commandType == CMD_SELECT);
950 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
954 * Can't handle FOR UPDATE/SHARE here (parser should have checked
955 * already, but let's make sure).
959 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
960 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
963 * Calculate pathkeys that represent result ordering requirements
965 Assert(parse->distinctClause == NIL);
966 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
973 /* No set operations, do regular planning */
975 double sub_limit_tuples;
976 AttrNumber *groupColIdx = NULL;
977 bool need_tlist_eval = true;
983 AggClauseCosts agg_costs;
987 bool use_hashed_grouping = false;
988 WindowFuncLists *wflists = NULL;
989 List *activeWindows = NIL;
991 MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
993 /* A recursive query should always have setOperations */
994 Assert(!root->hasRecursion);
996 /* Preprocess GROUP BY clause, if any */
997 if (parse->groupClause)
998 preprocess_groupclause(root);
999 numGroupCols = list_length(parse->groupClause);
1001 /* Preprocess targetlist */
1002 tlist = preprocess_targetlist(root, tlist);
1005 * Locate any window functions in the tlist. (We don't need to look
1006 * anywhere else, since expressions used in ORDER BY will be in there
1007 * too.) Note that they could all have been eliminated by constant
1008 * folding, in which case we don't need to do any more work.
1010 if (parse->hasWindowFuncs)
1012 wflists = find_window_functions((Node *) tlist,
1013 list_length(parse->windowClause));
1014 if (wflists->numWindowFuncs > 0)
1015 activeWindows = select_active_windows(root, wflists);
1017 parse->hasWindowFuncs = false;
1021 * Generate appropriate target list for subplan; may be different from
1022 * tlist if grouping or aggregation is needed.
1024 sub_tlist = make_subplanTargetList(root, tlist,
1025 &groupColIdx, &need_tlist_eval);
1028 * Do aggregate preprocessing, if the query has any aggs.
1030 * Note: think not that we can turn off hasAggs if we find no aggs. It
1031 * is possible for constant-expression simplification to remove all
1032 * explicit references to aggs, but we still have to follow the
1033 * aggregate semantics (eg, producing only one output row).
1038 * Collect statistics about aggregates for estimating costs. Note:
1039 * we do not attempt to detect duplicate aggregates here; a
1040 * somewhat-overestimated cost is okay for our present purposes.
1042 count_agg_clauses(root, (Node *) tlist, &agg_costs);
1043 count_agg_clauses(root, parse->havingQual, &agg_costs);
1046 * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1047 * adding logic between here and the optimize_minmax_aggregates
1048 * call. Anything that is needed in MIN/MAX-optimizable cases
1049 * will have to be duplicated in planagg.c.
1051 preprocess_minmax_aggregates(root, tlist);
1055 * Calculate pathkeys that represent grouping/ordering requirements.
1056 * Stash them in PlannerInfo so that query_planner can canonicalize
1057 * them after EquivalenceClasses have been formed. The sortClause is
1058 * certainly sort-able, but GROUP BY and DISTINCT might not be, in
1059 * which case we just leave their pathkeys empty.
1061 if (parse->groupClause &&
1062 grouping_is_sortable(parse->groupClause))
1063 root->group_pathkeys =
1064 make_pathkeys_for_sortclauses(root,
1069 root->group_pathkeys = NIL;
1071 /* We consider only the first (bottom) window in pathkeys logic */
1072 if (activeWindows != NIL)
1074 WindowClause *wc = (WindowClause *) linitial(activeWindows);
1076 root->window_pathkeys = make_pathkeys_for_window(root,
1082 root->window_pathkeys = NIL;
1084 if (parse->distinctClause &&
1085 grouping_is_sortable(parse->distinctClause))
1086 root->distinct_pathkeys =
1087 make_pathkeys_for_sortclauses(root,
1088 parse->distinctClause,
1092 root->distinct_pathkeys = NIL;
1094 root->sort_pathkeys =
1095 make_pathkeys_for_sortclauses(root,
1101 * Figure out whether we want a sorted result from query_planner.
1103 * If we have a sortable GROUP BY clause, then we want a result sorted
1104 * properly for grouping. Otherwise, if we have window functions to
1105 * evaluate, we try to sort for the first window. Otherwise, if
1106 * there's a sortable DISTINCT clause that's more rigorous than the
1107 * ORDER BY clause, we try to produce output that's sufficiently well
1108 * sorted for the DISTINCT. Otherwise, if there is an ORDER BY
1109 * clause, we want to sort by the ORDER BY clause.
1111 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
1112 * superset of GROUP BY, it would be tempting to request sort by ORDER
1113 * BY --- but that might just leave us failing to exploit an available
1114 * sort order at all. Needs more thought. The choice for DISTINCT
1115 * versus ORDER BY is much easier, since we know that the parser
1116 * ensured that one is a superset of the other.
1118 if (root->group_pathkeys)
1119 root->query_pathkeys = root->group_pathkeys;
1120 else if (root->window_pathkeys)
1121 root->query_pathkeys = root->window_pathkeys;
1122 else if (list_length(root->distinct_pathkeys) >
1123 list_length(root->sort_pathkeys))
1124 root->query_pathkeys = root->distinct_pathkeys;
1125 else if (root->sort_pathkeys)
1126 root->query_pathkeys = root->sort_pathkeys;
1128 root->query_pathkeys = NIL;
1131 * Figure out whether there's a hard limit on the number of rows that
1132 * query_planner's result subplan needs to return. Even if we know a
1133 * hard limit overall, it doesn't apply if the query has any
1134 * grouping/aggregation operations.
1136 if (parse->groupClause ||
1137 parse->distinctClause ||
1139 parse->hasWindowFuncs ||
1140 root->hasHavingQual)
1141 sub_limit_tuples = -1.0;
1143 sub_limit_tuples = limit_tuples;
1146 * Generate the best unsorted and presorted paths for this Query (but
1147 * note there may not be any presorted path). query_planner will also
1148 * estimate the number of groups in the query, and canonicalize all
1151 query_planner(root, sub_tlist, tuple_fraction, sub_limit_tuples,
1152 &cheapest_path, &sorted_path, &dNumGroups);
1155 * Extract rowcount and width estimates for possible use in grouping
1156 * decisions. Beware here of the possibility that
1157 * cheapest_path->parent is NULL (ie, there is no FROM clause).
1159 if (cheapest_path->parent)
1161 path_rows = cheapest_path->parent->rows;
1162 path_width = cheapest_path->parent->width;
1166 path_rows = 1; /* assume non-set result */
1167 path_width = 100; /* arbitrary */
1170 if (parse->groupClause)
1173 * If grouping, decide whether to use sorted or hashed grouping.
1175 use_hashed_grouping =
1176 choose_hashed_grouping(root,
1177 tuple_fraction, limit_tuples,
1178 path_rows, path_width,
1179 cheapest_path, sorted_path,
1180 dNumGroups, &agg_costs);
1181 /* Also convert # groups to long int --- but 'ware overflow! */
1182 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
1184 else if (parse->distinctClause && sorted_path &&
1185 !root->hasHavingQual && !parse->hasAggs && !activeWindows)
1188 * We'll reach the DISTINCT stage without any intermediate
1189 * processing, so figure out whether we will want to hash or not
1190 * so we can choose whether to use cheapest or sorted path.
1192 use_hashed_distinct =
1193 choose_hashed_distinct(root,
1194 tuple_fraction, limit_tuples,
1195 path_rows, path_width,
1196 cheapest_path->startup_cost,
1197 cheapest_path->total_cost,
1198 sorted_path->startup_cost,
1199 sorted_path->total_cost,
1200 sorted_path->pathkeys,
1202 tested_hashed_distinct = true;
1206 * Select the best path. If we are doing hashed grouping, we will
1207 * always read all the input tuples, so use the cheapest-total path.
1208 * Otherwise, trust query_planner's decision about which to use.
1210 if (use_hashed_grouping || use_hashed_distinct || !sorted_path)
1211 best_path = cheapest_path;
1213 best_path = sorted_path;
1216 * Check to see if it's possible to optimize MIN/MAX aggregates. If
1217 * so, we will forget all the work we did so far to choose a "regular"
1218 * path ... but we had to do it anyway to be able to tell which way is
1221 result_plan = optimize_minmax_aggregates(root,
1225 if (result_plan != NULL)
1228 * optimize_minmax_aggregates generated the full plan, with the
1229 * right tlist, and it has no sort order.
1231 current_pathkeys = NIL;
1236 * Normal case --- create a plan according to query_planner's
1239 bool need_sort_for_grouping = false;
1241 result_plan = create_plan(root, best_path);
1242 current_pathkeys = best_path->pathkeys;
1244 /* Detect if we'll need an explicit sort for grouping */
1245 if (parse->groupClause && !use_hashed_grouping &&
1246 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1248 need_sort_for_grouping = true;
1251 * Always override query_planner's tlist, so that we don't
1252 * sort useless data from a "physical" tlist.
1254 need_tlist_eval = true;
1258 * create_plan() returns a plan with just a "flat" tlist of
1259 * required Vars. Usually we need to insert the sub_tlist as the
1260 * tlist of the top plan node. However, we can skip that if we
1261 * determined that whatever query_planner chose to return will be
1264 if (need_tlist_eval)
1267 * If the top-level plan node is one that cannot do expression
1268 * evaluation, we must insert a Result node to project the
1271 if (!is_projection_capable_plan(result_plan))
1273 result_plan = (Plan *) make_result(root,
1281 * Otherwise, just replace the subplan's flat tlist with
1282 * the desired tlist.
1284 result_plan->targetlist = sub_tlist;
1288 * Also, account for the cost of evaluation of the sub_tlist.
1290 * Up to now, we have only been dealing with "flat" tlists,
1291 * containing just Vars. So their evaluation cost is zero
1292 * according to the model used by cost_qual_eval() (or if you
1293 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1294 * can avoid accounting for tlist cost throughout
1295 * query_planner() and subroutines. But now we've inserted a
1296 * tlist that might contain actual operators, sub-selects, etc
1297 * --- so we'd better account for its cost.
1299 * Below this point, any tlist eval cost for added-on nodes
1300 * should be accounted for as we create those nodes.
1301 * Presently, of the node types we can add on, only Agg,
1302 * WindowAgg, and Group project new tlists (the rest just copy
1303 * their input tuples) --- so make_agg(), make_windowagg() and
1304 * make_group() are responsible for computing the added cost.
1306 cost_qual_eval(&tlist_cost, sub_tlist, root);
1307 result_plan->startup_cost += tlist_cost.startup;
1308 result_plan->total_cost += tlist_cost.startup +
1309 tlist_cost.per_tuple * result_plan->plan_rows;
1314 * Since we're using query_planner's tlist and not the one
1315 * make_subplanTargetList calculated, we have to refigure any
1316 * grouping-column indexes make_subplanTargetList computed.
1318 locate_grouping_columns(root, tlist, result_plan->targetlist,
1323 * Insert AGG or GROUP node if needed, plus an explicit sort step
1326 * HAVING clause, if any, becomes qual of the Agg or Group node.
1328 if (use_hashed_grouping)
1330 /* Hashed aggregate plan --- no sort needed */
1331 result_plan = (Plan *) make_agg(root,
1333 (List *) parse->havingQual,
1338 extract_grouping_ops(parse->groupClause),
1341 /* Hashed aggregation produces randomly-ordered results */
1342 current_pathkeys = NIL;
1344 else if (parse->hasAggs)
1346 /* Plain aggregate plan --- sort if needed */
1347 AggStrategy aggstrategy;
1349 if (parse->groupClause)
1351 if (need_sort_for_grouping)
1353 result_plan = (Plan *)
1354 make_sort_from_groupcols(root,
1358 current_pathkeys = root->group_pathkeys;
1360 aggstrategy = AGG_SORTED;
1363 * The AGG node will not change the sort ordering of its
1364 * groups, so current_pathkeys describes the result too.
1369 aggstrategy = AGG_PLAIN;
1370 /* Result will be only one row anyway; no sort order */
1371 current_pathkeys = NIL;
1374 result_plan = (Plan *) make_agg(root,
1376 (List *) parse->havingQual,
1381 extract_grouping_ops(parse->groupClause),
1385 else if (parse->groupClause)
1388 * GROUP BY without aggregation, so insert a group node (plus
1389 * the appropriate sort node, if necessary).
1391 * Add an explicit sort if we couldn't make the path come out
1392 * the way the GROUP node needs it.
1394 if (need_sort_for_grouping)
1396 result_plan = (Plan *)
1397 make_sort_from_groupcols(root,
1401 current_pathkeys = root->group_pathkeys;
1404 result_plan = (Plan *) make_group(root,
1406 (List *) parse->havingQual,
1409 extract_grouping_ops(parse->groupClause),
1412 /* The Group node won't change sort ordering */
1414 else if (root->hasHavingQual)
1417 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1418 * This is a degenerate case in which we are supposed to emit
1419 * either 0 or 1 row depending on whether HAVING succeeds.
1420 * Furthermore, there cannot be any variables in either HAVING
1421 * or the targetlist, so we actually do not need the FROM
1422 * table at all! We can just throw away the plan-so-far and
1423 * generate a Result node. This is a sufficiently unusual
1424 * corner case that it's not worth contorting the structure of
1425 * this routine to avoid having to generate the plan in the
1428 result_plan = (Plan *) make_result(root,
1433 } /* end of non-minmax-aggregate case */
1436 * Since each window function could require a different sort order, we
1437 * stack up a WindowAgg node for each window, with sort steps between
1446 * If the top-level plan node is one that cannot do expression
1447 * evaluation, we must insert a Result node to project the desired
1448 * tlist. (In some cases this might not really be required, but
1449 * it's not worth trying to avoid it.) Note that on second and
1450 * subsequent passes through the following loop, the top-level
1451 * node will be a WindowAgg which we know can project; so we only
1452 * need to check once.
1454 if (!is_projection_capable_plan(result_plan))
1456 result_plan = (Plan *) make_result(root,
1463 * The "base" targetlist for all steps of the windowing process is
1464 * a flat tlist of all Vars and Aggs needed in the result. (In
1465 * some cases we wouldn't need to propagate all of these all the
1466 * way to the top, since they might only be needed as inputs to
1467 * WindowFuncs. It's probably not worth trying to optimize that
1468 * though.) We also need any volatile sort expressions, because
1469 * make_sort_from_pathkeys won't add those on its own, and anyway
1470 * we want them evaluated only once at the bottom of the stack. As
1471 * we climb up the stack, we add outputs for the WindowFuncs
1472 * computed at each level. Also, each input tlist has to present
1473 * all the columns needed to sort the data for the next WindowAgg
1474 * step. That's handled internally by make_sort_from_pathkeys,
1475 * but we need the copyObject steps here to ensure that each plan
1476 * node has a separately modifiable tlist.
1478 * Note: it's essential here to use PVC_INCLUDE_AGGREGATES so that
1479 * Vars mentioned only in aggregate expressions aren't pulled out
1480 * as separate targetlist entries. Otherwise we could be putting
1481 * ungrouped Vars directly into an Agg node's tlist, resulting in
1482 * undefined behavior.
1484 window_tlist = flatten_tlist(tlist,
1485 PVC_INCLUDE_AGGREGATES,
1486 PVC_INCLUDE_PLACEHOLDERS);
1487 window_tlist = add_volatile_sort_exprs(window_tlist, tlist,
1489 result_plan->targetlist = (List *) copyObject(window_tlist);
1491 foreach(l, activeWindows)
1493 WindowClause *wc = (WindowClause *) lfirst(l);
1494 List *window_pathkeys;
1496 AttrNumber *partColIdx;
1499 AttrNumber *ordColIdx;
1502 window_pathkeys = make_pathkeys_for_window(root,
1508 * This is a bit tricky: we build a sort node even if we don't
1509 * really have to sort. Even when no explicit sort is needed,
1510 * we need to have suitable resjunk items added to the input
1511 * plan's tlist for any partitioning or ordering columns that
1512 * aren't plain Vars. Furthermore, this way we can use
1513 * existing infrastructure to identify which input columns are
1514 * the interesting ones.
1516 if (window_pathkeys)
1520 sort_plan = make_sort_from_pathkeys(root,
1524 if (!pathkeys_contained_in(window_pathkeys,
1527 /* we do indeed need to sort */
1528 result_plan = (Plan *) sort_plan;
1529 current_pathkeys = window_pathkeys;
1531 /* In either case, extract the per-column information */
1532 get_column_info_for_window(root, wc, tlist,
1534 sort_plan->sortColIdx,
1544 /* empty window specification, nothing to sort */
1547 partOperators = NULL;
1550 ordOperators = NULL;
1555 /* Add the current WindowFuncs to the running tlist */
1556 window_tlist = add_to_flat_tlist(window_tlist,
1557 wflists->windowFuncs[wc->winref]);
1561 /* Install the original tlist in the topmost WindowAgg */
1562 window_tlist = tlist;
1565 /* ... and make the WindowAgg plan node */
1566 result_plan = (Plan *)
1567 make_windowagg(root,
1568 (List *) copyObject(window_tlist),
1569 wflists->windowFuncs[wc->winref],
1583 } /* end of if (setOperations) */
1586 * If there is a DISTINCT clause, add the necessary node(s).
1588 if (parse->distinctClause)
1590 double dNumDistinctRows;
1591 long numDistinctRows;
1594 * If there was grouping or aggregation, use the current number of
1595 * rows as the estimated number of DISTINCT rows (ie, assume the
1596 * result was already mostly unique). If not, use the number of
1597 * distinct-groups calculated by query_planner.
1599 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1600 dNumDistinctRows = result_plan->plan_rows;
1602 dNumDistinctRows = dNumGroups;
1604 /* Also convert to long int --- but 'ware overflow! */
1605 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1607 /* Choose implementation method if we didn't already */
1608 if (!tested_hashed_distinct)
1611 * At this point, either hashed or sorted grouping will have to
1612 * work from result_plan, so we pass that as both "cheapest" and
1615 use_hashed_distinct =
1616 choose_hashed_distinct(root,
1617 tuple_fraction, limit_tuples,
1618 result_plan->plan_rows,
1619 result_plan->plan_width,
1620 result_plan->startup_cost,
1621 result_plan->total_cost,
1622 result_plan->startup_cost,
1623 result_plan->total_cost,
1628 if (use_hashed_distinct)
1630 /* Hashed aggregate plan --- no sort needed */
1631 result_plan = (Plan *) make_agg(root,
1632 result_plan->targetlist,
1636 list_length(parse->distinctClause),
1637 extract_grouping_cols(parse->distinctClause,
1638 result_plan->targetlist),
1639 extract_grouping_ops(parse->distinctClause),
1642 /* Hashed aggregation produces randomly-ordered results */
1643 current_pathkeys = NIL;
1648 * Use a Unique node to implement DISTINCT. Add an explicit sort
1649 * if we couldn't make the path come out the way the Unique node
1650 * needs it. If we do have to sort, always sort by the more
1651 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1652 * below. However, for regular DISTINCT, don't sort now if we
1653 * don't have to --- sorting afterwards will likely be cheaper,
1654 * and also has the possibility of optimizing via LIMIT. But for
1655 * DISTINCT ON, we *must* force the final sort now, else it won't
1656 * have the desired behavior.
1658 List *needed_pathkeys;
1660 if (parse->hasDistinctOn &&
1661 list_length(root->distinct_pathkeys) <
1662 list_length(root->sort_pathkeys))
1663 needed_pathkeys = root->sort_pathkeys;
1665 needed_pathkeys = root->distinct_pathkeys;
1667 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1669 if (list_length(root->distinct_pathkeys) >=
1670 list_length(root->sort_pathkeys))
1671 current_pathkeys = root->distinct_pathkeys;
1674 current_pathkeys = root->sort_pathkeys;
1675 /* Assert checks that parser didn't mess up... */
1676 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1680 result_plan = (Plan *) make_sort_from_pathkeys(root,
1686 result_plan = (Plan *) make_unique(result_plan,
1687 parse->distinctClause);
1688 result_plan->plan_rows = dNumDistinctRows;
1689 /* The Unique node won't change sort ordering */
1694 * If ORDER BY was given and we were not able to make the plan come out in
1695 * the right order, add an explicit sort step.
1697 if (parse->sortClause)
1699 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1701 result_plan = (Plan *) make_sort_from_pathkeys(root,
1703 root->sort_pathkeys,
1705 current_pathkeys = root->sort_pathkeys;
1710 * If there is a FOR UPDATE/SHARE clause, add the LockRows node. (Note: we
1711 * intentionally test parse->rowMarks not root->rowMarks here. If there
1712 * are only non-locking rowmarks, they should be handled by the
1713 * ModifyTable node instead.)
1715 if (parse->rowMarks)
1717 result_plan = (Plan *) make_lockrows(result_plan,
1719 SS_assign_special_param(root));
1722 * The result can no longer be assumed sorted, since locking might
1723 * cause the sort key columns to be replaced with new values.
1725 current_pathkeys = NIL;
1729 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1731 if (parse->limitCount || parse->limitOffset)
1733 result_plan = (Plan *) make_limit(result_plan,
1741 * Return the actual output ordering in query_pathkeys for possible use by
1742 * an outer query level.
1744 root->query_pathkeys = current_pathkeys;
1750 * Detect whether a plan node is a "dummy" plan created when a relation
1751 * is deemed not to need scanning due to constraint exclusion.
1753 * Currently, such dummy plans are Result nodes with constant FALSE
1757 is_dummy_plan(Plan *plan)
1759 if (IsA(plan, Result))
1761 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1763 if (list_length(rcqual) == 1)
1765 Const *constqual = (Const *) linitial(rcqual);
1767 if (constqual && IsA(constqual, Const))
1769 if (!constqual->constisnull &&
1770 !DatumGetBool(constqual->constvalue))
1779 * Create a bitmapset of the RT indexes of live base relations
1781 * Helper for preprocess_rowmarks ... at this point in the proceedings,
1782 * the only good way to distinguish baserels from appendrel children
1783 * is to see what is in the join tree.
1786 get_base_rel_indexes(Node *jtnode)
1792 if (IsA(jtnode, RangeTblRef))
1794 int varno = ((RangeTblRef *) jtnode)->rtindex;
1796 result = bms_make_singleton(varno);
1798 else if (IsA(jtnode, FromExpr))
1800 FromExpr *f = (FromExpr *) jtnode;
1804 foreach(l, f->fromlist)
1805 result = bms_join(result,
1806 get_base_rel_indexes(lfirst(l)));
1808 else if (IsA(jtnode, JoinExpr))
1810 JoinExpr *j = (JoinExpr *) jtnode;
1812 result = bms_join(get_base_rel_indexes(j->larg),
1813 get_base_rel_indexes(j->rarg));
1817 elog(ERROR, "unrecognized node type: %d",
1818 (int) nodeTag(jtnode));
1819 result = NULL; /* keep compiler quiet */
1825 * preprocess_rowmarks - set up PlanRowMarks if needed
1828 preprocess_rowmarks(PlannerInfo *root)
1830 Query *parse = root->parse;
1836 if (parse->rowMarks)
1839 * We've got trouble if FOR UPDATE/SHARE appears inside grouping,
1840 * since grouping renders a reference to individual tuple CTIDs
1841 * invalid. This is also checked at parse time, but that's
1842 * insufficient because of rule substitution, query pullup, etc.
1844 CheckSelectLocking(parse);
1849 * We only need rowmarks for UPDATE, DELETE, or FOR UPDATE/SHARE.
1851 if (parse->commandType != CMD_UPDATE &&
1852 parse->commandType != CMD_DELETE)
1857 * We need to have rowmarks for all base relations except the target. We
1858 * make a bitmapset of all base rels and then remove the items we don't
1859 * need or have FOR UPDATE/SHARE marks for.
1861 rels = get_base_rel_indexes((Node *) parse->jointree);
1862 if (parse->resultRelation)
1863 rels = bms_del_member(rels, parse->resultRelation);
1866 * Convert RowMarkClauses to PlanRowMark representation.
1869 foreach(l, parse->rowMarks)
1871 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
1872 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
1876 * Currently, it is syntactically impossible to have FOR UPDATE
1877 * applied to an update/delete target rel. If that ever becomes
1878 * possible, we should drop the target from the PlanRowMark list.
1880 Assert(rc->rti != parse->resultRelation);
1883 * Ignore RowMarkClauses for subqueries; they aren't real tables and
1884 * can't support true locking. Subqueries that got flattened into the
1885 * main query should be ignored completely. Any that didn't will get
1886 * ROW_MARK_COPY items in the next loop.
1888 if (rte->rtekind != RTE_RELATION)
1891 rels = bms_del_member(rels, rc->rti);
1893 newrc = makeNode(PlanRowMark);
1894 newrc->rti = newrc->prti = rc->rti;
1895 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
1897 newrc->markType = ROW_MARK_EXCLUSIVE;
1899 newrc->markType = ROW_MARK_SHARE;
1900 newrc->noWait = rc->noWait;
1901 newrc->isParent = false;
1903 prowmarks = lappend(prowmarks, newrc);
1907 * Now, add rowmarks for any non-target, non-locked base relations.
1910 foreach(l, parse->rtable)
1912 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
1916 if (!bms_is_member(i, rels))
1919 newrc = makeNode(PlanRowMark);
1920 newrc->rti = newrc->prti = i;
1921 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
1922 /* real tables support REFERENCE, anything else needs COPY */
1923 if (rte->rtekind == RTE_RELATION &&
1924 rte->relkind != RELKIND_FOREIGN_TABLE)
1925 newrc->markType = ROW_MARK_REFERENCE;
1927 newrc->markType = ROW_MARK_COPY;
1928 newrc->noWait = false; /* doesn't matter */
1929 newrc->isParent = false;
1931 prowmarks = lappend(prowmarks, newrc);
1934 root->rowMarks = prowmarks;
1938 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1940 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1941 * results back in *count_est and *offset_est. These variables are set to
1942 * 0 if the corresponding clause is not present, and -1 if it's present
1943 * but we couldn't estimate the value for it. (The "0" convention is OK
1944 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1945 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1946 * usual practice of never estimating less than one row.) These values will
1947 * be passed to make_limit, which see if you change this code.
1949 * The return value is the suitably adjusted tuple_fraction to use for
1950 * planning the query. This adjustment is not overridable, since it reflects
1951 * plan actions that grouping_planner() will certainly take, not assumptions
1955 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1956 int64 *offset_est, int64 *count_est)
1958 Query *parse = root->parse;
1960 double limit_fraction;
1962 /* Should not be called unless LIMIT or OFFSET */
1963 Assert(parse->limitCount || parse->limitOffset);
1966 * Try to obtain the clause values. We use estimate_expression_value
1967 * primarily because it can sometimes do something useful with Params.
1969 if (parse->limitCount)
1971 est = estimate_expression_value(root, parse->limitCount);
1972 if (est && IsA(est, Const))
1974 if (((Const *) est)->constisnull)
1976 /* NULL indicates LIMIT ALL, ie, no limit */
1977 *count_est = 0; /* treat as not present */
1981 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1982 if (*count_est <= 0)
1983 *count_est = 1; /* force to at least 1 */
1987 *count_est = -1; /* can't estimate */
1990 *count_est = 0; /* not present */
1992 if (parse->limitOffset)
1994 est = estimate_expression_value(root, parse->limitOffset);
1995 if (est && IsA(est, Const))
1997 if (((Const *) est)->constisnull)
1999 /* Treat NULL as no offset; the executor will too */
2000 *offset_est = 0; /* treat as not present */
2004 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2005 if (*offset_est < 0)
2006 *offset_est = 0; /* less than 0 is same as 0 */
2010 *offset_est = -1; /* can't estimate */
2013 *offset_est = 0; /* not present */
2015 if (*count_est != 0)
2018 * A LIMIT clause limits the absolute number of tuples returned.
2019 * However, if it's not a constant LIMIT then we have to guess; for
2020 * lack of a better idea, assume 10% of the plan's result is wanted.
2022 if (*count_est < 0 || *offset_est < 0)
2024 /* LIMIT or OFFSET is an expression ... punt ... */
2025 limit_fraction = 0.10;
2029 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2030 limit_fraction = (double) *count_est + (double) *offset_est;
2034 * If we have absolute limits from both caller and LIMIT, use the
2035 * smaller value; likewise if they are both fractional. If one is
2036 * fractional and the other absolute, we can't easily determine which
2037 * is smaller, but we use the heuristic that the absolute will usually
2040 if (tuple_fraction >= 1.0)
2042 if (limit_fraction >= 1.0)
2045 tuple_fraction = Min(tuple_fraction, limit_fraction);
2049 /* caller absolute, limit fractional; use caller's value */
2052 else if (tuple_fraction > 0.0)
2054 if (limit_fraction >= 1.0)
2056 /* caller fractional, limit absolute; use limit */
2057 tuple_fraction = limit_fraction;
2061 /* both fractional */
2062 tuple_fraction = Min(tuple_fraction, limit_fraction);
2067 /* no info from caller, just use limit */
2068 tuple_fraction = limit_fraction;
2071 else if (*offset_est != 0 && tuple_fraction > 0.0)
2074 * We have an OFFSET but no LIMIT. This acts entirely differently
2075 * from the LIMIT case: here, we need to increase rather than decrease
2076 * the caller's tuple_fraction, because the OFFSET acts to cause more
2077 * tuples to be fetched instead of fewer. This only matters if we got
2078 * a tuple_fraction > 0, however.
2080 * As above, use 10% if OFFSET is present but unestimatable.
2082 if (*offset_est < 0)
2083 limit_fraction = 0.10;
2085 limit_fraction = (double) *offset_est;
2088 * If we have absolute counts from both caller and OFFSET, add them
2089 * together; likewise if they are both fractional. If one is
2090 * fractional and the other absolute, we want to take the larger, and
2091 * we heuristically assume that's the fractional one.
2093 if (tuple_fraction >= 1.0)
2095 if (limit_fraction >= 1.0)
2097 /* both absolute, so add them together */
2098 tuple_fraction += limit_fraction;
2102 /* caller absolute, limit fractional; use limit */
2103 tuple_fraction = limit_fraction;
2108 if (limit_fraction >= 1.0)
2110 /* caller fractional, limit absolute; use caller's value */
2114 /* both fractional, so add them together */
2115 tuple_fraction += limit_fraction;
2116 if (tuple_fraction >= 1.0)
2117 tuple_fraction = 0.0; /* assume fetch all */
2122 return tuple_fraction;
2127 * preprocess_groupclause - do preparatory work on GROUP BY clause
2129 * The idea here is to adjust the ordering of the GROUP BY elements
2130 * (which in itself is semantically insignificant) to match ORDER BY,
2131 * thereby allowing a single sort operation to both implement the ORDER BY
2132 * requirement and set up for a Unique step that implements GROUP BY.
2134 * In principle it might be interesting to consider other orderings of the
2135 * GROUP BY elements, which could match the sort ordering of other
2136 * possible plans (eg an indexscan) and thereby reduce cost. We don't
2137 * bother with that, though. Hashed grouping will frequently win anyway.
2139 * Note: we need no comparable processing of the distinctClause because
2140 * the parser already enforced that that matches ORDER BY.
2143 preprocess_groupclause(PlannerInfo *root)
2145 Query *parse = root->parse;
2146 List *new_groupclause;
2151 /* If no ORDER BY, nothing useful to do here */
2152 if (parse->sortClause == NIL)
2156 * Scan the ORDER BY clause and construct a list of matching GROUP BY
2157 * items, but only as far as we can make a matching prefix.
2159 * This code assumes that the sortClause contains no duplicate items.
2161 new_groupclause = NIL;
2162 foreach(sl, parse->sortClause)
2164 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
2166 foreach(gl, parse->groupClause)
2168 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2172 new_groupclause = lappend(new_groupclause, gc);
2177 break; /* no match, so stop scanning */
2180 /* Did we match all of the ORDER BY list, or just some of it? */
2181 partial_match = (sl != NULL);
2183 /* If no match at all, no point in reordering GROUP BY */
2184 if (new_groupclause == NIL)
2188 * Add any remaining GROUP BY items to the new list, but only if we were
2189 * able to make a complete match. In other words, we only rearrange the
2190 * GROUP BY list if the result is that one list is a prefix of the other
2191 * --- otherwise there's no possibility of a common sort. Also, give up
2192 * if there are any non-sortable GROUP BY items, since then there's no
2195 foreach(gl, parse->groupClause)
2197 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2199 if (list_member_ptr(new_groupclause, gc))
2200 continue; /* it matched an ORDER BY item */
2202 return; /* give up, no common sort possible */
2203 if (!OidIsValid(gc->sortop))
2204 return; /* give up, GROUP BY can't be sorted */
2205 new_groupclause = lappend(new_groupclause, gc);
2208 /* Success --- install the rearranged GROUP BY list */
2209 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2210 parse->groupClause = new_groupclause;
2214 * choose_hashed_grouping - should we use hashed grouping?
2216 * Returns TRUE to select hashing, FALSE to select sorting.
2219 choose_hashed_grouping(PlannerInfo *root,
2220 double tuple_fraction, double limit_tuples,
2221 double path_rows, int path_width,
2222 Path *cheapest_path, Path *sorted_path,
2223 double dNumGroups, AggClauseCosts *agg_costs)
2225 Query *parse = root->parse;
2226 int numGroupCols = list_length(parse->groupClause);
2230 List *target_pathkeys;
2231 List *current_pathkeys;
2236 * Executor doesn't support hashed aggregation with DISTINCT or ORDER BY
2237 * aggregates. (Doing so would imply storing *all* the input values in
2238 * the hash table, and/or running many sorts in parallel, either of which
2239 * seems like a certain loser.)
2241 can_hash = (agg_costs->numOrderedAggs == 0 &&
2242 grouping_is_hashable(parse->groupClause));
2243 can_sort = grouping_is_sortable(parse->groupClause);
2245 /* Quick out if only one choice is workable */
2246 if (!(can_hash && can_sort))
2254 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2255 errmsg("could not implement GROUP BY"),
2256 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2259 /* Prefer sorting when enable_hashagg is off */
2260 if (!enable_hashagg)
2264 * Don't do it if it doesn't look like the hashtable will fit into
2268 /* Estimate per-hash-entry space at tuple width... */
2269 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2270 /* plus space for pass-by-ref transition values... */
2271 hashentrysize += agg_costs->transitionSpace;
2272 /* plus the per-hash-entry overhead */
2273 hashentrysize += hash_agg_entry_size(agg_costs->numAggs);
2275 if (hashentrysize * dNumGroups > work_mem * 1024L)
2279 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
2280 * DISTINCT and ORDER BY as the assumed required output sort order. This
2281 * is an oversimplification because the DISTINCT might get implemented via
2282 * hashing, but it's not clear that the case is common enough (or that our
2283 * estimates are good enough) to justify trying to solve it exactly.
2285 if (list_length(root->distinct_pathkeys) >
2286 list_length(root->sort_pathkeys))
2287 target_pathkeys = root->distinct_pathkeys;
2289 target_pathkeys = root->sort_pathkeys;
2292 * See if the estimated cost is no more than doing it the other way. While
2293 * avoiding the need for sorted input is usually a win, the fact that the
2294 * output won't be sorted may be a loss; so we need to do an actual cost
2297 * We need to consider cheapest_path + hashagg [+ final sort] versus
2298 * either cheapest_path [+ sort] + group or agg [+ final sort] or
2299 * presorted_path + group or agg [+ final sort] where brackets indicate a
2300 * step that may not be needed. We assume query_planner() will have
2301 * returned a presorted path only if it's a winner compared to
2302 * cheapest_path for this purpose.
2304 * These path variables are dummies that just hold cost fields; we don't
2305 * make actual Paths for these steps.
2307 cost_agg(&hashed_p, root, AGG_HASHED, agg_costs,
2308 numGroupCols, dNumGroups,
2309 cheapest_path->startup_cost, cheapest_path->total_cost,
2311 /* Result of hashed agg is always unsorted */
2312 if (target_pathkeys)
2313 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
2314 dNumGroups, path_width,
2315 0.0, work_mem, limit_tuples);
2319 sorted_p.startup_cost = sorted_path->startup_cost;
2320 sorted_p.total_cost = sorted_path->total_cost;
2321 current_pathkeys = sorted_path->pathkeys;
2325 sorted_p.startup_cost = cheapest_path->startup_cost;
2326 sorted_p.total_cost = cheapest_path->total_cost;
2327 current_pathkeys = cheapest_path->pathkeys;
2329 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
2331 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
2332 path_rows, path_width,
2333 0.0, work_mem, -1.0);
2334 current_pathkeys = root->group_pathkeys;
2338 cost_agg(&sorted_p, root, AGG_SORTED, agg_costs,
2339 numGroupCols, dNumGroups,
2340 sorted_p.startup_cost, sorted_p.total_cost,
2343 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
2344 sorted_p.startup_cost, sorted_p.total_cost,
2346 /* The Agg or Group node will preserve ordering */
2347 if (target_pathkeys &&
2348 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
2349 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
2350 dNumGroups, path_width,
2351 0.0, work_mem, limit_tuples);
2354 * Now make the decision using the top-level tuple fraction. First we
2355 * have to convert an absolute count (LIMIT) into fractional form.
2357 if (tuple_fraction >= 1.0)
2358 tuple_fraction /= dNumGroups;
2360 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2361 tuple_fraction) < 0)
2363 /* Hashed is cheaper, so use it */
2370 * choose_hashed_distinct - should we use hashing for DISTINCT?
2372 * This is fairly similar to choose_hashed_grouping, but there are enough
2373 * differences that it doesn't seem worth trying to unify the two functions.
2374 * (One difference is that we sometimes apply this after forming a Plan,
2375 * so the input alternatives can't be represented as Paths --- instead we
2376 * pass in the costs as individual variables.)
2378 * But note that making the two choices independently is a bit bogus in
2379 * itself. If the two could be combined into a single choice operation
2380 * it'd probably be better, but that seems far too unwieldy to be practical,
2381 * especially considering that the combination of GROUP BY and DISTINCT
2382 * isn't very common in real queries. By separating them, we are giving
2383 * extra preference to using a sorting implementation when a common sort key
2384 * is available ... and that's not necessarily wrong anyway.
2386 * Returns TRUE to select hashing, FALSE to select sorting.
2389 choose_hashed_distinct(PlannerInfo *root,
2390 double tuple_fraction, double limit_tuples,
2391 double path_rows, int path_width,
2392 Cost cheapest_startup_cost, Cost cheapest_total_cost,
2393 Cost sorted_startup_cost, Cost sorted_total_cost,
2394 List *sorted_pathkeys,
2395 double dNumDistinctRows)
2397 Query *parse = root->parse;
2398 int numDistinctCols = list_length(parse->distinctClause);
2402 List *current_pathkeys;
2403 List *needed_pathkeys;
2408 * If we have a sortable DISTINCT ON clause, we always use sorting. This
2409 * enforces the expected behavior of DISTINCT ON.
2411 can_sort = grouping_is_sortable(parse->distinctClause);
2412 if (can_sort && parse->hasDistinctOn)
2415 can_hash = grouping_is_hashable(parse->distinctClause);
2417 /* Quick out if only one choice is workable */
2418 if (!(can_hash && can_sort))
2426 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2427 errmsg("could not implement DISTINCT"),
2428 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2431 /* Prefer sorting when enable_hashagg is off */
2432 if (!enable_hashagg)
2436 * Don't do it if it doesn't look like the hashtable will fit into
2439 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2441 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
2445 * See if the estimated cost is no more than doing it the other way. While
2446 * avoiding the need for sorted input is usually a win, the fact that the
2447 * output won't be sorted may be a loss; so we need to do an actual cost
2450 * We need to consider cheapest_path + hashagg [+ final sort] versus
2451 * sorted_path [+ sort] + group [+ final sort] where brackets indicate a
2452 * step that may not be needed.
2454 * These path variables are dummies that just hold cost fields; we don't
2455 * make actual Paths for these steps.
2457 cost_agg(&hashed_p, root, AGG_HASHED, NULL,
2458 numDistinctCols, dNumDistinctRows,
2459 cheapest_startup_cost, cheapest_total_cost,
2463 * Result of hashed agg is always unsorted, so if ORDER BY is present we
2464 * need to charge for the final sort.
2466 if (parse->sortClause)
2467 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
2468 dNumDistinctRows, path_width,
2469 0.0, work_mem, limit_tuples);
2472 * Now for the GROUP case. See comments in grouping_planner about the
2473 * sorting choices here --- this code should match that code.
2475 sorted_p.startup_cost = sorted_startup_cost;
2476 sorted_p.total_cost = sorted_total_cost;
2477 current_pathkeys = sorted_pathkeys;
2478 if (parse->hasDistinctOn &&
2479 list_length(root->distinct_pathkeys) <
2480 list_length(root->sort_pathkeys))
2481 needed_pathkeys = root->sort_pathkeys;
2483 needed_pathkeys = root->distinct_pathkeys;
2484 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
2486 if (list_length(root->distinct_pathkeys) >=
2487 list_length(root->sort_pathkeys))
2488 current_pathkeys = root->distinct_pathkeys;
2490 current_pathkeys = root->sort_pathkeys;
2491 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
2492 path_rows, path_width,
2493 0.0, work_mem, -1.0);
2495 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
2496 sorted_p.startup_cost, sorted_p.total_cost,
2498 if (parse->sortClause &&
2499 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
2500 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
2501 dNumDistinctRows, path_width,
2502 0.0, work_mem, limit_tuples);
2505 * Now make the decision using the top-level tuple fraction. First we
2506 * have to convert an absolute count (LIMIT) into fractional form.
2508 if (tuple_fraction >= 1.0)
2509 tuple_fraction /= dNumDistinctRows;
2511 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2512 tuple_fraction) < 0)
2514 /* Hashed is cheaper, so use it */
2521 * make_subplanTargetList
2522 * Generate appropriate target list when grouping is required.
2524 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
2525 * above the result of query_planner, we typically want to pass a different
2526 * target list to query_planner than the outer plan nodes should have.
2527 * This routine generates the correct target list for the subplan.
2529 * The initial target list passed from the parser already contains entries
2530 * for all ORDER BY and GROUP BY expressions, but it will not have entries
2531 * for variables used only in HAVING clauses; so we need to add those
2532 * variables to the subplan target list. Also, we flatten all expressions
2533 * except GROUP BY items into their component variables; the other expressions
2534 * will be computed by the inserted nodes rather than by the subplan.
2535 * For example, given a query like
2536 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
2537 * we want to pass this targetlist to the subplan:
2539 * where the a+b target will be used by the Sort/Group steps, and the
2540 * other targets will be used for computing the final results. (In the
2541 * above example we could theoretically suppress the a and b targets and
2542 * pass down only c,d,a+b, but it's not really worth the trouble to
2543 * eliminate simple var references from the subplan. We will avoid doing
2544 * the extra computation to recompute a+b at the outer level; see
2545 * fix_upper_expr() in setrefs.c.)
2547 * If we are grouping or aggregating, *and* there are no non-Var grouping
2548 * expressions, then the returned tlist is effectively dummy; we do not
2549 * need to force it to be evaluated, because all the Vars it contains
2550 * should be present in the output of query_planner anyway.
2552 * 'tlist' is the query's target list.
2553 * 'groupColIdx' receives an array of column numbers for the GROUP BY
2554 * expressions (if there are any) in the subplan's target list.
2555 * 'need_tlist_eval' is set true if we really need to evaluate the
2558 * The result is the targetlist to be passed to the subplan.
2561 make_subplanTargetList(PlannerInfo *root,
2563 AttrNumber **groupColIdx,
2564 bool *need_tlist_eval)
2566 Query *parse = root->parse;
2571 *groupColIdx = NULL;
2574 * If we're not grouping or aggregating, there's nothing to do here;
2575 * query_planner should receive the unmodified target list.
2577 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual &&
2578 !parse->hasWindowFuncs)
2580 *need_tlist_eval = true;
2585 * Otherwise, start with a "flattened" tlist (having just the Vars
2586 * mentioned in the targetlist and HAVING qual). Note this includes Vars
2587 * used in resjunk items, so we are covering the needs of ORDER BY and
2588 * window specifications. Vars used within Aggrefs will be pulled out
2591 sub_tlist = flatten_tlist(tlist,
2592 PVC_RECURSE_AGGREGATES,
2593 PVC_INCLUDE_PLACEHOLDERS);
2594 extravars = pull_var_clause(parse->havingQual,
2595 PVC_RECURSE_AGGREGATES,
2596 PVC_INCLUDE_PLACEHOLDERS);
2597 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2598 list_free(extravars);
2599 *need_tlist_eval = false; /* only eval if not flat tlist */
2602 * If grouping, create sub_tlist entries for all GROUP BY expressions
2603 * (GROUP BY items that are simple Vars should be in the list already),
2604 * and make an array showing where the group columns are in the sub_tlist.
2606 numCols = list_length(parse->groupClause);
2610 AttrNumber *grpColIdx;
2613 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2614 *groupColIdx = grpColIdx;
2616 foreach(gl, parse->groupClause)
2618 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2619 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2623 * Find or make a matching sub_tlist entry. If the groupexpr
2624 * isn't a Var, no point in searching. (Note that the parser
2625 * won't make multiple groupClause entries for the same TLE.)
2627 if (groupexpr && IsA(groupexpr, Var))
2628 te = tlist_member(groupexpr, sub_tlist);
2634 te = makeTargetEntry((Expr *) groupexpr,
2635 list_length(sub_tlist) + 1,
2638 sub_tlist = lappend(sub_tlist, te);
2639 *need_tlist_eval = true; /* it's not flat anymore */
2642 /* and save its resno */
2643 grpColIdx[keyno++] = te->resno;
2651 * locate_grouping_columns
2652 * Locate grouping columns in the tlist chosen by query_planner.
2654 * This is only needed if we don't use the sub_tlist chosen by
2655 * make_subplanTargetList. We have to forget the column indexes found
2656 * by that routine and re-locate the grouping exprs in the real sub_tlist.
2659 locate_grouping_columns(PlannerInfo *root,
2662 AttrNumber *groupColIdx)
2668 * No work unless grouping.
2670 if (!root->parse->groupClause)
2672 Assert(groupColIdx == NULL);
2675 Assert(groupColIdx != NULL);
2677 foreach(gl, root->parse->groupClause)
2679 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2680 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2681 TargetEntry *te = tlist_member(groupexpr, sub_tlist);
2684 elog(ERROR, "failed to locate grouping columns");
2685 groupColIdx[keyno++] = te->resno;
2690 * postprocess_setop_tlist
2691 * Fix up targetlist returned by plan_set_operations().
2693 * We need to transpose sort key info from the orig_tlist into new_tlist.
2694 * NOTE: this would not be good enough if we supported resjunk sort keys
2695 * for results of set operations --- then, we'd need to project a whole
2696 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2697 * find any resjunk columns in orig_tlist.
2700 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2703 ListCell *orig_tlist_item = list_head(orig_tlist);
2705 foreach(l, new_tlist)
2707 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2708 TargetEntry *orig_tle;
2710 /* ignore resjunk columns in setop result */
2711 if (new_tle->resjunk)
2714 Assert(orig_tlist_item != NULL);
2715 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2716 orig_tlist_item = lnext(orig_tlist_item);
2717 if (orig_tle->resjunk) /* should not happen */
2718 elog(ERROR, "resjunk output columns are not implemented");
2719 Assert(new_tle->resno == orig_tle->resno);
2720 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2722 if (orig_tlist_item != NULL)
2723 elog(ERROR, "resjunk output columns are not implemented");
2728 * select_active_windows
2729 * Create a list of the "active" window clauses (ie, those referenced
2730 * by non-deleted WindowFuncs) in the order they are to be executed.
2733 select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
2739 /* First, make a list of the active windows */
2741 foreach(lc, root->parse->windowClause)
2743 WindowClause *wc = (WindowClause *) lfirst(lc);
2745 /* It's only active if wflists shows some related WindowFuncs */
2746 Assert(wc->winref <= wflists->maxWinRef);
2747 if (wflists->windowFuncs[wc->winref] != NIL)
2748 actives = lappend(actives, wc);
2752 * Now, ensure that windows with identical partitioning/ordering clauses
2753 * are adjacent in the list. This is required by the SQL standard, which
2754 * says that only one sort is to be used for such windows, even if they
2755 * are otherwise distinct (eg, different names or framing clauses).
2757 * There is room to be much smarter here, for example detecting whether
2758 * one window's sort keys are a prefix of another's (so that sorting for
2759 * the latter would do for the former), or putting windows first that
2760 * match a sort order available for the underlying query. For the moment
2761 * we are content with meeting the spec.
2764 while (actives != NIL)
2766 WindowClause *wc = (WindowClause *) linitial(actives);
2770 /* Move wc from actives to result */
2771 actives = list_delete_first(actives);
2772 result = lappend(result, wc);
2774 /* Now move any matching windows from actives to result */
2776 for (lc = list_head(actives); lc; lc = next)
2778 WindowClause *wc2 = (WindowClause *) lfirst(lc);
2781 /* framing options are NOT to be compared here! */
2782 if (equal(wc->partitionClause, wc2->partitionClause) &&
2783 equal(wc->orderClause, wc2->orderClause))
2785 actives = list_delete_cell(actives, lc, prev);
2786 result = lappend(result, wc2);
2797 * add_volatile_sort_exprs
2798 * Identify any volatile sort/group expressions used by the active
2799 * windows, and add them to window_tlist if not already present.
2800 * Return the modified window_tlist.
2803 add_volatile_sort_exprs(List *window_tlist, List *tlist, List *activeWindows)
2805 Bitmapset *sgrefs = NULL;
2808 /* First, collect the sortgrouprefs of the windows into a bitmapset */
2809 foreach(lc, activeWindows)
2811 WindowClause *wc = (WindowClause *) lfirst(lc);
2814 foreach(lc2, wc->partitionClause)
2816 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2818 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2820 foreach(lc2, wc->orderClause)
2822 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2824 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2829 * Now scan the original tlist to find the referenced expressions. Any
2830 * that are volatile must be added to window_tlist.
2832 * Note: we know that the input window_tlist contains no items marked with
2833 * ressortgrouprefs, so we don't have to worry about collisions of the
2834 * reference numbers.
2838 TargetEntry *tle = (TargetEntry *) lfirst(lc);
2840 if (tle->ressortgroupref != 0 &&
2841 bms_is_member(tle->ressortgroupref, sgrefs) &&
2842 contain_volatile_functions((Node *) tle->expr))
2844 TargetEntry *newtle;
2846 newtle = makeTargetEntry(tle->expr,
2847 list_length(window_tlist) + 1,
2850 newtle->ressortgroupref = tle->ressortgroupref;
2851 window_tlist = lappend(window_tlist, newtle);
2855 return window_tlist;
2859 * make_pathkeys_for_window
2860 * Create a pathkeys list describing the required input ordering
2861 * for the given WindowClause.
2863 * The required ordering is first the PARTITION keys, then the ORDER keys.
2864 * In the future we might try to implement windowing using hashing, in which
2865 * case the ordering could be relaxed, but for now we always sort.
2868 make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
2869 List *tlist, bool canonicalize)
2871 List *window_pathkeys;
2872 List *window_sortclauses;
2874 /* Throw error if can't sort */
2875 if (!grouping_is_sortable(wc->partitionClause))
2877 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2878 errmsg("could not implement window PARTITION BY"),
2879 errdetail("Window partitioning columns must be of sortable datatypes.")));
2880 if (!grouping_is_sortable(wc->orderClause))
2882 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2883 errmsg("could not implement window ORDER BY"),
2884 errdetail("Window ordering columns must be of sortable datatypes.")));
2886 /* Okay, make the combined pathkeys */
2887 window_sortclauses = list_concat(list_copy(wc->partitionClause),
2888 list_copy(wc->orderClause));
2889 window_pathkeys = make_pathkeys_for_sortclauses(root,
2893 list_free(window_sortclauses);
2894 return window_pathkeys;
2898 * get_column_info_for_window
2899 * Get the partitioning/ordering column numbers and equality operators
2900 * for a WindowAgg node.
2902 * This depends on the behavior of make_pathkeys_for_window()!
2904 * We are given the target WindowClause and an array of the input column
2905 * numbers associated with the resulting pathkeys. In the easy case, there
2906 * are the same number of pathkey columns as partitioning + ordering columns
2907 * and we just have to copy some data around. However, it's possible that
2908 * some of the original partitioning + ordering columns were eliminated as
2909 * redundant during the transformation to pathkeys. (This can happen even
2910 * though the parser gets rid of obvious duplicates. A typical scenario is a
2911 * window specification "PARTITION BY x ORDER BY y" coupled with a clause
2912 * "WHERE x = y" that causes the two sort columns to be recognized as
2913 * redundant.) In that unusual case, we have to work a lot harder to
2914 * determine which keys are significant.
2916 * The method used here is a bit brute-force: add the sort columns to a list
2917 * one at a time and note when the resulting pathkey list gets longer. But
2918 * it's a sufficiently uncommon case that a faster way doesn't seem worth
2919 * the amount of code refactoring that'd be needed.
2923 get_column_info_for_window(PlannerInfo *root, WindowClause *wc, List *tlist,
2924 int numSortCols, AttrNumber *sortColIdx,
2926 AttrNumber **partColIdx,
2927 Oid **partOperators,
2929 AttrNumber **ordColIdx,
2932 int numPart = list_length(wc->partitionClause);
2933 int numOrder = list_length(wc->orderClause);
2935 if (numSortCols == numPart + numOrder)
2938 *partNumCols = numPart;
2939 *partColIdx = sortColIdx;
2940 *partOperators = extract_grouping_ops(wc->partitionClause);
2941 *ordNumCols = numOrder;
2942 *ordColIdx = sortColIdx + numPart;
2943 *ordOperators = extract_grouping_ops(wc->orderClause);
2952 /* first, allocate what's certainly enough space for the arrays */
2954 *partColIdx = (AttrNumber *) palloc(numPart * sizeof(AttrNumber));
2955 *partOperators = (Oid *) palloc(numPart * sizeof(Oid));
2957 *ordColIdx = (AttrNumber *) palloc(numOrder * sizeof(AttrNumber));
2958 *ordOperators = (Oid *) palloc(numOrder * sizeof(Oid));
2962 foreach(lc, wc->partitionClause)
2964 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2967 sortclauses = lappend(sortclauses, sgc);
2968 new_pathkeys = make_pathkeys_for_sortclauses(root,
2972 if (list_length(new_pathkeys) > list_length(pathkeys))
2974 /* this sort clause is actually significant */
2975 (*partColIdx)[*partNumCols] = sortColIdx[scidx++];
2976 (*partOperators)[*partNumCols] = sgc->eqop;
2978 pathkeys = new_pathkeys;
2981 foreach(lc, wc->orderClause)
2983 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2986 sortclauses = lappend(sortclauses, sgc);
2987 new_pathkeys = make_pathkeys_for_sortclauses(root,
2991 if (list_length(new_pathkeys) > list_length(pathkeys))
2993 /* this sort clause is actually significant */
2994 (*ordColIdx)[*ordNumCols] = sortColIdx[scidx++];
2995 (*ordOperators)[*ordNumCols] = sgc->eqop;
2997 pathkeys = new_pathkeys;
3000 /* complain if we didn't eat exactly the right number of sort cols */
3001 if (scidx != numSortCols)
3002 elog(ERROR, "failed to deconstruct sort operators into partitioning/ordering operators");
3008 * expression_planner
3009 * Perform planner's transformations on a standalone expression.
3011 * Various utility commands need to evaluate expressions that are not part
3012 * of a plannable query. They can do so using the executor's regular
3013 * expression-execution machinery, but first the expression has to be fed
3014 * through here to transform it from parser output to something executable.
3016 * Currently, we disallow sublinks in standalone expressions, so there's no
3017 * real "planning" involved here. (That might not always be true though.)
3018 * What we must do is run eval_const_expressions to ensure that any function
3019 * calls are converted to positional notation and function default arguments
3020 * get inserted. The fact that constant subexpressions get simplified is a
3021 * side-effect that is useful when the expression will get evaluated more than
3022 * once. Also, we must fix operator function IDs.
3024 * Note: this must not make any damaging changes to the passed-in expression
3025 * tree. (It would actually be okay to apply fix_opfuncids to it, but since
3026 * we first do an expression_tree_mutator-based walk, what is returned will
3027 * be a new node tree.)
3030 expression_planner(Expr *expr)
3035 * Convert named-argument function calls, insert default arguments and
3036 * simplify constant subexprs
3038 result = eval_const_expressions(NULL, (Node *) expr);
3040 /* Fill in opfuncid values if missing */
3041 fix_opfuncids(result);
3043 return (Expr *) result;
3048 * plan_cluster_use_sort
3049 * Use the planner to decide how CLUSTER should implement sorting
3051 * tableOid is the OID of a table to be clustered on its index indexOid
3052 * (which is already known to be a btree index). Decide whether it's
3053 * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
3054 * Return TRUE to use sorting, FALSE to use an indexscan.
3056 * Note: caller had better already hold some type of lock on the table.
3059 plan_cluster_use_sort(Oid tableOid, Oid indexOid)
3063 PlannerGlobal *glob;
3066 IndexOptInfo *indexInfo;
3067 QualCost indexExprCost;
3068 Cost comparisonCost;
3070 Path seqScanAndSortPath;
3071 IndexPath *indexScanPath;
3074 /* Set up mostly-dummy planner state */
3075 query = makeNode(Query);
3076 query->commandType = CMD_SELECT;
3078 glob = makeNode(PlannerGlobal);
3080 root = makeNode(PlannerInfo);
3081 root->parse = query;
3083 root->query_level = 1;
3084 root->planner_cxt = CurrentMemoryContext;
3085 root->wt_param_id = -1;
3087 /* Build a minimal RTE for the rel */
3088 rte = makeNode(RangeTblEntry);
3089 rte->rtekind = RTE_RELATION;
3090 rte->relid = tableOid;
3091 rte->relkind = RELKIND_RELATION;
3093 rte->inFromCl = true;
3094 query->rtable = list_make1(rte);
3096 /* ... and insert it into PlannerInfo */
3097 root->simple_rel_array_size = 2;
3098 root->simple_rel_array = (RelOptInfo **)
3099 palloc0(root->simple_rel_array_size * sizeof(RelOptInfo *));
3100 root->simple_rte_array = (RangeTblEntry **)
3101 palloc0(root->simple_rel_array_size * sizeof(RangeTblEntry *));
3102 root->simple_rte_array[1] = rte;
3104 /* Build RelOptInfo */
3105 rel = build_simple_rel(root, 1, RELOPT_BASEREL);
3107 /* Locate IndexOptInfo for the target index */
3109 foreach(lc, rel->indexlist)
3111 indexInfo = (IndexOptInfo *) lfirst(lc);
3112 if (indexInfo->indexoid == indexOid)
3117 * It's possible that get_relation_info did not generate an IndexOptInfo
3118 * for the desired index; this could happen if it's not yet reached its
3119 * indcheckxmin usability horizon, or if it's a system index and we're
3120 * ignoring system indexes. In such cases we should tell CLUSTER to not
3121 * trust the index contents but use seqscan-and-sort.
3123 if (lc == NULL) /* not in the list? */
3124 return true; /* use sort */
3127 * Rather than doing all the pushups that would be needed to use
3128 * set_baserel_size_estimates, just do a quick hack for rows and width.
3130 rel->rows = rel->tuples;
3131 rel->width = get_relation_data_width(tableOid, NULL);
3133 root->total_table_pages = rel->pages;
3136 * Determine eval cost of the index expressions, if any. We need to
3137 * charge twice that amount for each tuple comparison that happens during
3138 * the sort, since tuplesort.c will have to re-evaluate the index
3139 * expressions each time. (XXX that's pretty inefficient...)
3141 cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
3142 comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
3144 /* Estimate the cost of seq scan + sort */
3145 seqScanPath = create_seqscan_path(root, rel);
3146 cost_sort(&seqScanAndSortPath, root, NIL,
3147 seqScanPath->total_cost, rel->tuples, rel->width,
3148 comparisonCost, maintenance_work_mem, -1.0);
3150 /* Estimate the cost of index scan */
3151 indexScanPath = create_index_path(root, indexInfo,
3153 ForwardScanDirection, NULL);
3155 return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);