1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2012, 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 "executor/executor.h"
21 #include "executor/nodeAgg.h"
22 #include "miscadmin.h"
23 #include "nodes/makefuncs.h"
24 #ifdef OPTIMIZER_DEBUG
25 #include "nodes/print.h"
27 #include "optimizer/clauses.h"
28 #include "optimizer/cost.h"
29 #include "optimizer/pathnode.h"
30 #include "optimizer/paths.h"
31 #include "optimizer/plancat.h"
32 #include "optimizer/planmain.h"
33 #include "optimizer/planner.h"
34 #include "optimizer/prep.h"
35 #include "optimizer/subselect.h"
36 #include "optimizer/tlist.h"
37 #include "parser/analyze.h"
38 #include "parser/parsetree.h"
39 #include "rewrite/rewriteManip.h"
40 #include "utils/rel.h"
44 double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
46 /* Hook for plugins to get control in planner() */
47 planner_hook_type planner_hook = NULL;
50 /* Expression kind codes for preprocess_expression */
51 #define EXPRKIND_QUAL 0
52 #define EXPRKIND_TARGET 1
53 #define EXPRKIND_RTFUNC 2
54 #define EXPRKIND_VALUES 3
55 #define EXPRKIND_LIMIT 4
56 #define EXPRKIND_APPINFO 5
59 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
60 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
61 static Plan *inheritance_planner(PlannerInfo *root);
62 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
63 static void preprocess_rowmarks(PlannerInfo *root);
64 static double preprocess_limit(PlannerInfo *root,
65 double tuple_fraction,
66 int64 *offset_est, int64 *count_est);
67 static void preprocess_groupclause(PlannerInfo *root);
68 static bool choose_hashed_grouping(PlannerInfo *root,
69 double tuple_fraction, double limit_tuples,
70 double path_rows, int path_width,
71 Path *cheapest_path, Path *sorted_path,
72 double dNumGroups, AggClauseCosts *agg_costs);
73 static bool choose_hashed_distinct(PlannerInfo *root,
74 double tuple_fraction, double limit_tuples,
75 double path_rows, int path_width,
76 Cost cheapest_startup_cost, Cost cheapest_total_cost,
77 Cost sorted_startup_cost, Cost sorted_total_cost,
78 List *sorted_pathkeys,
79 double dNumDistinctRows);
80 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
81 AttrNumber **groupColIdx, bool *need_tlist_eval);
82 static int get_grouping_column_index(Query *parse, TargetEntry *tle);
83 static void locate_grouping_columns(PlannerInfo *root,
86 AttrNumber *groupColIdx);
87 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
88 static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
89 static List *add_volatile_sort_exprs(List *window_tlist, List *tlist,
91 static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
92 List *tlist, bool canonicalize);
93 static void get_column_info_for_window(PlannerInfo *root, WindowClause *wc,
95 int numSortCols, AttrNumber *sortColIdx,
97 AttrNumber **partColIdx,
100 AttrNumber **ordColIdx,
104 /*****************************************************************************
106 * Query optimizer entry point
108 * To support loadable plugins that monitor or modify planner behavior,
109 * we provide a hook variable that lets a plugin get control before and
110 * after the standard planning process. The plugin would normally call
111 * standard_planner().
113 * Note to plugin authors: standard_planner() scribbles on its Query input,
114 * so you'd better copy that data structure if you want to plan more than once.
116 *****************************************************************************/
118 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
123 result = (*planner_hook) (parse, cursorOptions, boundParams);
125 result = standard_planner(parse, cursorOptions, boundParams);
130 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
134 double tuple_fraction;
140 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
141 if (parse->utilityStmt &&
142 IsA(parse->utilityStmt, DeclareCursorStmt))
143 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
146 * Set up global state for this planner invocation. This data is needed
147 * across all levels of sub-Query that might exist in the given command,
148 * so we keep it in a separate struct that's linked to by each per-Query
151 glob = makeNode(PlannerGlobal);
153 glob->boundParams = boundParams;
154 glob->paramlist = NIL;
155 glob->subplans = NIL;
156 glob->subroots = NIL;
157 glob->rewindPlanIDs = NULL;
158 glob->finalrtable = NIL;
159 glob->finalrowmarks = NIL;
160 glob->resultRelations = NIL;
161 glob->relationOids = NIL;
162 glob->invalItems = NIL;
164 glob->lastRowMarkId = 0;
165 glob->transientPlan = false;
167 /* Determine what fraction of the plan is likely to be scanned */
168 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
171 * We have no real idea how many tuples the user will ultimately FETCH
172 * from a cursor, but it is often the case that he doesn't want 'em
173 * all, or would prefer a fast-start plan anyway so that he can
174 * process some of the tuples sooner. Use a GUC parameter to decide
175 * what fraction to optimize for.
177 tuple_fraction = cursor_tuple_fraction;
180 * We document cursor_tuple_fraction as simply being a fraction, which
181 * means the edge cases 0 and 1 have to be treated specially here. We
182 * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
184 if (tuple_fraction >= 1.0)
185 tuple_fraction = 0.0;
186 else if (tuple_fraction <= 0.0)
187 tuple_fraction = 1e-10;
191 /* Default assumption is we need all the tuples */
192 tuple_fraction = 0.0;
195 /* primary planning entry point (may recurse for subqueries) */
196 top_plan = subquery_planner(glob, parse, NULL,
197 false, tuple_fraction, &root);
200 * If creating a plan for a scrollable cursor, make sure it can run
201 * backwards on demand. Add a Material node at the top at need.
203 if (cursorOptions & CURSOR_OPT_SCROLL)
205 if (!ExecSupportsBackwardScan(top_plan))
206 top_plan = materialize_finished_plan(top_plan);
209 /* final cleanup of the plan */
210 Assert(glob->finalrtable == NIL);
211 Assert(glob->finalrowmarks == NIL);
212 Assert(glob->resultRelations == NIL);
213 top_plan = set_plan_references(root, top_plan);
214 /* ... and the subplans (both regular subplans and initplans) */
215 Assert(list_length(glob->subplans) == list_length(glob->subroots));
216 forboth(lp, glob->subplans, lr, glob->subroots)
218 Plan *subplan = (Plan *) lfirst(lp);
219 PlannerInfo *subroot = (PlannerInfo *) lfirst(lr);
221 lfirst(lp) = set_plan_references(subroot, subplan);
224 /* build the PlannedStmt result */
225 result = makeNode(PlannedStmt);
227 result->commandType = parse->commandType;
228 result->hasReturning = (parse->returningList != NIL);
229 result->hasModifyingCTE = parse->hasModifyingCTE;
230 result->canSetTag = parse->canSetTag;
231 result->transientPlan = glob->transientPlan;
232 result->planTree = top_plan;
233 result->rtable = glob->finalrtable;
234 result->resultRelations = glob->resultRelations;
235 result->utilityStmt = parse->utilityStmt;
236 result->intoClause = parse->intoClause;
237 result->subplans = glob->subplans;
238 result->rewindPlanIDs = glob->rewindPlanIDs;
239 result->rowMarks = glob->finalrowmarks;
240 result->relationOids = glob->relationOids;
241 result->invalItems = glob->invalItems;
242 result->nParamExec = list_length(glob->paramlist);
248 /*--------------------
250 * Invokes the planner on a subquery. We recurse to here for each
251 * sub-SELECT found in the query tree.
253 * glob is the global state for the current planner run.
254 * parse is the querytree produced by the parser & rewriter.
255 * parent_root is the immediate parent Query's info (NULL at the top level).
256 * hasRecursion is true if this is a recursive WITH query.
257 * tuple_fraction is the fraction of tuples we expect will be retrieved.
258 * tuple_fraction is interpreted as explained for grouping_planner, below.
260 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
261 * among other things this tells the output sort ordering of the plan.
263 * Basically, this routine does the stuff that should only be done once
264 * per Query object. It then calls grouping_planner. At one time,
265 * grouping_planner could be invoked recursively on the same Query object;
266 * that's not currently true, but we keep the separation between the two
267 * routines anyway, in case we need it again someday.
269 * subquery_planner will be called recursively to handle sub-Query nodes
270 * found within the query's expressions and rangetable.
272 * Returns a query plan.
273 *--------------------
276 subquery_planner(PlannerGlobal *glob, Query *parse,
277 PlannerInfo *parent_root,
278 bool hasRecursion, double tuple_fraction,
279 PlannerInfo **subroot)
281 int num_old_subplans = list_length(glob->subplans);
288 /* Create a PlannerInfo data structure for this subquery */
289 root = makeNode(PlannerInfo);
292 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
293 root->parent_root = parent_root;
294 root->planner_cxt = CurrentMemoryContext;
295 root->init_plans = NIL;
296 root->cte_plan_ids = NIL;
297 root->eq_classes = NIL;
298 root->append_rel_list = NIL;
299 root->rowMarks = NIL;
300 root->hasInheritedTarget = false;
302 root->hasRecursion = hasRecursion;
304 root->wt_param_id = SS_assign_special_param(root);
306 root->wt_param_id = -1;
307 root->non_recursive_plan = NULL;
310 * If there is a WITH list, process each WITH query and build an initplan
311 * SubPlan structure for it.
314 SS_process_ctes(root);
317 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
318 * to transform them into joins. Note that this step does not descend
319 * into subqueries; if we pull up any subqueries below, their SubLinks are
320 * processed just before pulling them up.
322 if (parse->hasSubLinks)
323 pull_up_sublinks(root);
326 * Scan the rangetable for set-returning functions, and inline them if
327 * possible (producing subqueries that might get pulled up next).
328 * Recursion issues here are handled in the same way as for SubLinks.
330 inline_set_returning_functions(root);
333 * Check to see if any subqueries in the jointree can be merged into this
336 parse->jointree = (FromExpr *)
337 pull_up_subqueries(root, (Node *) parse->jointree, NULL, NULL);
340 * If this is a simple UNION ALL query, flatten it into an appendrel. We
341 * do this now because it requires applying pull_up_subqueries to the leaf
342 * queries of the UNION ALL, which weren't touched above because they
343 * weren't referenced by the jointree (they will be after we do this).
345 if (parse->setOperations)
346 flatten_simple_union_all(root);
349 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
350 * avoid the expense of doing flatten_join_alias_vars(). Also check for
351 * outer joins --- if none, we can skip reduce_outer_joins(). This must be
352 * done after we have done pull_up_subqueries, of course.
354 root->hasJoinRTEs = false;
355 hasOuterJoins = false;
356 foreach(l, parse->rtable)
358 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
360 if (rte->rtekind == RTE_JOIN)
362 root->hasJoinRTEs = true;
363 if (IS_OUTER_JOIN(rte->jointype))
365 hasOuterJoins = true;
366 /* Can quit scanning once we find an outer join */
373 * Preprocess RowMark information. We need to do this after subquery
374 * pullup (so that all non-inherited RTEs are present) and before
375 * inheritance expansion (so that the info is available for
376 * expand_inherited_tables to examine and modify).
378 preprocess_rowmarks(root);
381 * Expand any rangetable entries that are inheritance sets into "append
382 * relations". This can add entries to the rangetable, but they must be
383 * plain base relations not joins, so it's OK (and marginally more
384 * efficient) to do it after checking for join RTEs. We must do it after
385 * pulling up subqueries, else we'd fail to handle inherited tables in
388 expand_inherited_tables(root);
391 * Set hasHavingQual to remember if HAVING clause is present. Needed
392 * because preprocess_expression will reduce a constant-true condition to
393 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
395 root->hasHavingQual = (parse->havingQual != NULL);
397 /* Clear this flag; might get set in distribute_qual_to_rels */
398 root->hasPseudoConstantQuals = false;
401 * Do expression preprocessing on targetlist and quals, as well as other
402 * random expressions in the querytree. Note that we do not need to
403 * handle sort/group expressions explicitly, because they are actually
404 * part of the targetlist.
406 parse->targetList = (List *)
407 preprocess_expression(root, (Node *) parse->targetList,
410 parse->returningList = (List *)
411 preprocess_expression(root, (Node *) parse->returningList,
414 preprocess_qual_conditions(root, (Node *) parse->jointree);
416 parse->havingQual = preprocess_expression(root, parse->havingQual,
419 foreach(l, parse->windowClause)
421 WindowClause *wc = (WindowClause *) lfirst(l);
423 /* partitionClause/orderClause are sort/group expressions */
424 wc->startOffset = preprocess_expression(root, wc->startOffset,
426 wc->endOffset = preprocess_expression(root, wc->endOffset,
430 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
432 parse->limitCount = preprocess_expression(root, parse->limitCount,
435 root->append_rel_list = (List *)
436 preprocess_expression(root, (Node *) root->append_rel_list,
439 /* Also need to preprocess expressions for function and values RTEs */
440 foreach(l, parse->rtable)
442 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
444 if (rte->rtekind == RTE_FUNCTION)
445 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
447 else if (rte->rtekind == RTE_VALUES)
448 rte->values_lists = (List *)
449 preprocess_expression(root, (Node *) rte->values_lists,
454 * In some cases we may want to transfer a HAVING clause into WHERE. We
455 * cannot do so if the HAVING clause contains aggregates (obviously) or
456 * volatile functions (since a HAVING clause is supposed to be executed
457 * only once per group). Also, it may be that the clause is so expensive
458 * to execute that we're better off doing it only once per group, despite
459 * the loss of selectivity. This is hard to estimate short of doing the
460 * entire planning process twice, so we use a heuristic: clauses
461 * containing subplans are left in HAVING. Otherwise, we move or copy the
462 * HAVING clause into WHERE, in hopes of eliminating tuples before
463 * aggregation instead of after.
465 * If the query has explicit grouping then we can simply move such a
466 * clause into WHERE; any group that fails the clause will not be in the
467 * output because none of its tuples will reach the grouping or
468 * aggregation stage. Otherwise we must have a degenerate (variable-free)
469 * HAVING clause, which we put in WHERE so that query_planner() can use it
470 * in a gating Result node, but also keep in HAVING to ensure that we
471 * don't emit a bogus aggregated row. (This could be done better, but it
472 * seems not worth optimizing.)
474 * Note that both havingQual and parse->jointree->quals are in
475 * implicitly-ANDed-list form at this point, even though they are declared
479 foreach(l, (List *) parse->havingQual)
481 Node *havingclause = (Node *) lfirst(l);
483 if (contain_agg_clause(havingclause) ||
484 contain_volatile_functions(havingclause) ||
485 contain_subplans(havingclause))
487 /* keep it in HAVING */
488 newHaving = lappend(newHaving, havingclause);
490 else if (parse->groupClause)
492 /* move it to WHERE */
493 parse->jointree->quals = (Node *)
494 lappend((List *) parse->jointree->quals, havingclause);
498 /* put a copy in WHERE, keep it in HAVING */
499 parse->jointree->quals = (Node *)
500 lappend((List *) parse->jointree->quals,
501 copyObject(havingclause));
502 newHaving = lappend(newHaving, havingclause);
505 parse->havingQual = (Node *) newHaving;
508 * If we have any outer joins, try to reduce them to plain inner joins.
509 * This step is most easily done after we've done expression
513 reduce_outer_joins(root);
516 * Do the main planning. If we have an inherited target relation, that
517 * needs special processing, else go straight to grouping_planner.
519 if (parse->resultRelation &&
520 rt_fetch(parse->resultRelation, parse->rtable)->inh)
521 plan = inheritance_planner(root);
524 plan = grouping_planner(root, tuple_fraction);
525 /* If it's not SELECT, we need a ModifyTable node */
526 if (parse->commandType != CMD_SELECT)
528 List *returningLists;
532 * Deal with the RETURNING clause if any. It's convenient to pass
533 * the returningList through setrefs.c now rather than at top
534 * level (if we waited, handling inherited UPDATE/DELETE would be
537 if (parse->returningList)
541 Assert(parse->resultRelation);
542 rlist = set_returning_clause_references(root,
543 parse->returningList,
545 parse->resultRelation);
546 returningLists = list_make1(rlist);
549 returningLists = NIL;
552 * If there was a FOR UPDATE/SHARE clause, the LockRows node will
553 * have dealt with fetching non-locked marked rows, else we need
554 * to have ModifyTable do that.
559 rowMarks = root->rowMarks;
561 plan = (Plan *) make_modifytable(parse->commandType,
563 list_make1_int(parse->resultRelation),
567 SS_assign_special_param(root));
572 * If any subplans were generated, or if there are any parameters to worry
573 * about, build initPlan list and extParam/allParam sets for plan nodes,
574 * and attach the initPlans to the top plan node.
576 if (list_length(glob->subplans) != num_old_subplans ||
577 root->glob->paramlist != NIL)
578 SS_finalize_plan(root, plan, true);
580 /* Return internal info if caller wants it */
588 * preprocess_expression
589 * Do subquery_planner's preprocessing work for an expression,
590 * which can be a targetlist, a WHERE clause (including JOIN/ON
591 * conditions), or a HAVING clause.
594 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
597 * Fall out quickly if expression is empty. This occurs often enough to
598 * be worth checking. Note that null->null is the correct conversion for
599 * implicit-AND result format, too.
605 * If the query has any join RTEs, replace join alias variables with
606 * base-relation variables. We must do this before sublink processing,
607 * else sublinks expanded out from join aliases wouldn't get processed. We
608 * can skip it in VALUES lists, however, since they can't contain any Vars
611 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
612 expr = flatten_join_alias_vars(root, expr);
615 * Simplify constant expressions.
617 * Note: an essential effect of this is to convert named-argument function
618 * calls to positional notation and insert the current actual values of
619 * any default arguments for functions. To ensure that happens, we *must*
620 * process all expressions here. Previous PG versions sometimes skipped
621 * const-simplification if it didn't seem worth the trouble, but we can't
624 * Note: this also flattens nested AND and OR expressions into N-argument
625 * form. All processing of a qual expression after this point must be
626 * careful to maintain AND/OR flatness --- that is, do not generate a tree
627 * with AND directly under AND, nor OR directly under OR.
629 expr = eval_const_expressions(root, expr);
632 * If it's a qual or havingQual, canonicalize it.
634 if (kind == EXPRKIND_QUAL)
636 expr = (Node *) canonicalize_qual((Expr *) expr);
638 #ifdef OPTIMIZER_DEBUG
639 printf("After canonicalize_qual()\n");
644 /* Expand SubLinks to SubPlans */
645 if (root->parse->hasSubLinks)
646 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
649 * XXX do not insert anything here unless you have grokked the comments in
650 * SS_replace_correlation_vars ...
653 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
654 if (root->query_level > 1)
655 expr = SS_replace_correlation_vars(root, expr);
658 * If it's a qual or havingQual, convert it to implicit-AND format. (We
659 * don't want to do this before eval_const_expressions, since the latter
660 * would be unable to simplify a top-level AND correctly. Also,
661 * SS_process_sublinks expects explicit-AND format.)
663 if (kind == EXPRKIND_QUAL)
664 expr = (Node *) make_ands_implicit((Expr *) expr);
670 * preprocess_qual_conditions
671 * Recursively scan the query's jointree and do subquery_planner's
672 * preprocessing work on each qual condition found therein.
675 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
679 if (IsA(jtnode, RangeTblRef))
681 /* nothing to do here */
683 else if (IsA(jtnode, FromExpr))
685 FromExpr *f = (FromExpr *) jtnode;
688 foreach(l, f->fromlist)
689 preprocess_qual_conditions(root, lfirst(l));
691 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
693 else if (IsA(jtnode, JoinExpr))
695 JoinExpr *j = (JoinExpr *) jtnode;
697 preprocess_qual_conditions(root, j->larg);
698 preprocess_qual_conditions(root, j->rarg);
700 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
703 elog(ERROR, "unrecognized node type: %d",
704 (int) nodeTag(jtnode));
708 * inheritance_planner
709 * Generate a plan in the case where the result relation is an
712 * We have to handle this case differently from cases where a source relation
713 * is an inheritance set. Source inheritance is expanded at the bottom of the
714 * plan tree (see allpaths.c), but target inheritance has to be expanded at
715 * the top. The reason is that for UPDATE, each target relation needs a
716 * different targetlist matching its own column set. Fortunately,
717 * the UPDATE/DELETE target can never be the nullable side of an outer join,
718 * so it's OK to generate the plan this way.
720 * Returns a query plan.
723 inheritance_planner(PlannerInfo *root)
725 Query *parse = root->parse;
726 int parentRTindex = parse->resultRelation;
727 List *final_rtable = NIL;
728 int save_rel_array_size = 0;
729 RelOptInfo **save_rel_array = NULL;
730 List *subplans = NIL;
731 List *resultRelations = NIL;
732 List *returningLists = NIL;
737 * We generate a modified instance of the original Query for each target
738 * relation, plan that, and put all the plans into a list that will be
739 * controlled by a single ModifyTable node. All the instances share the
740 * same rangetable, but each instance must have its own set of subquery
741 * RTEs within the finished rangetable because (1) they are likely to get
742 * scribbled on during planning, and (2) it's not inconceivable that
743 * subqueries could get planned differently in different cases. We need
744 * not create duplicate copies of other RTE kinds, in particular not the
745 * target relations, because they don't have either of those issues. Not
746 * having to duplicate the target relations is important because doing so
747 * (1) would result in a rangetable of length O(N^2) for N targets, with
748 * at least O(N^3) work expended here; and (2) would greatly complicate
749 * management of the rowMarks list.
751 foreach(lc, root->append_rel_list)
753 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(lc);
758 /* append_rel_list contains all append rels; ignore others */
759 if (appinfo->parent_relid != parentRTindex)
763 * We need a working copy of the PlannerInfo so that we can control
764 * propagation of information back to the main copy.
766 memcpy(&subroot, root, sizeof(PlannerInfo));
769 * Generate modified query with this rel as target. We first apply
770 * adjust_appendrel_attrs, which copies the Query and changes
771 * references to the parent RTE to refer to the current child RTE,
772 * then fool around with subquery RTEs.
774 subroot.parse = (Query *)
775 adjust_appendrel_attrs((Node *) parse,
779 * The rowMarks list might contain references to subquery RTEs, so
780 * make a copy that we can apply ChangeVarNodes to. (Fortunately,
781 * the executor doesn't need to see the modified copies --- we can
782 * just pass it the original rowMarks list.)
784 subroot.rowMarks = (List *) copyObject(root->rowMarks);
787 * Add placeholders to the child Query's rangetable list to fill the
788 * RT indexes already reserved for subqueries in previous children.
789 * These won't be referenced, so there's no need to make them very
792 while (list_length(subroot.parse->rtable) < list_length(final_rtable))
793 subroot.parse->rtable = lappend(subroot.parse->rtable,
794 makeNode(RangeTblEntry));
797 * If this isn't the first child Query, generate duplicates of all
798 * subquery RTEs, and adjust Var numbering to reference the duplicates.
799 * To simplify the loop logic, we scan the original rtable not the
800 * copy just made by adjust_appendrel_attrs; that should be OK since
801 * subquery RTEs couldn't contain any references to the target rel.
803 if (final_rtable != NIL)
808 foreach(lr, parse->rtable)
810 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lr);
812 if (rte->rtekind == RTE_SUBQUERY)
817 * The RTE can't contain any references to its own RT
818 * index, so we can save a few cycles by applying
819 * ChangeVarNodes before we append the RTE to the
822 newrti = list_length(subroot.parse->rtable) + 1;
823 ChangeVarNodes((Node *) subroot.parse, rti, newrti, 0);
824 ChangeVarNodes((Node *) subroot.rowMarks, rti, newrti, 0);
825 rte = copyObject(rte);
826 subroot.parse->rtable = lappend(subroot.parse->rtable,
833 /* We needn't modify the child's append_rel_list */
834 /* There shouldn't be any OJ info to translate, as yet */
835 Assert(subroot.join_info_list == NIL);
836 /* and we haven't created PlaceHolderInfos, either */
837 Assert(subroot.placeholder_list == NIL);
838 /* build a separate list of initplans for each child */
839 subroot.init_plans = NIL;
840 /* hack to mark target relation as an inheritance partition */
841 subroot.hasInheritedTarget = true;
844 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
847 * If this child rel was excluded by constraint exclusion, exclude it
848 * from the result plan.
850 if (is_dummy_plan(subplan))
853 subplans = lappend(subplans, subplan);
856 * If this is the first non-excluded child, its post-planning rtable
857 * becomes the initial contents of final_rtable; otherwise, append
858 * just its modified subquery RTEs to final_rtable.
860 if (final_rtable == NIL)
861 final_rtable = subroot.parse->rtable;
863 final_rtable = list_concat(final_rtable,
864 list_copy_tail(subroot.parse->rtable,
865 list_length(final_rtable)));
868 * We need to collect all the RelOptInfos from all child plans into
869 * the main PlannerInfo, since setrefs.c will need them. We use the
870 * last child's simple_rel_array (previous ones are too short), so we
871 * have to propagate forward the RelOptInfos that were already built
872 * in previous children.
874 Assert(subroot.simple_rel_array_size >= save_rel_array_size);
875 for (rti = 1; rti < save_rel_array_size; rti++)
877 RelOptInfo *brel = save_rel_array[rti];
880 subroot.simple_rel_array[rti] = brel;
882 save_rel_array_size = subroot.simple_rel_array_size;
883 save_rel_array = subroot.simple_rel_array;
885 /* Make sure any initplans from this rel get into the outer list */
886 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
888 /* Build list of target-relation RT indexes */
889 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
891 /* Build list of per-relation RETURNING targetlists */
892 if (parse->returningList)
896 rlist = set_returning_clause_references(&subroot,
897 subroot.parse->returningList,
899 appinfo->child_relid);
900 returningLists = lappend(returningLists, rlist);
904 /* Mark result as unordered (probably unnecessary) */
905 root->query_pathkeys = NIL;
908 * If we managed to exclude every child rel, return a dummy plan; it
909 * doesn't even need a ModifyTable node.
913 /* although dummy, it must have a valid tlist for executor */
916 tlist = preprocess_targetlist(root, parse->targetList);
917 return (Plan *) make_result(root,
919 (Node *) list_make1(makeBoolConst(false,
925 * Put back the final adjusted rtable into the master copy of the Query.
927 parse->rtable = final_rtable;
928 root->simple_rel_array_size = save_rel_array_size;
929 root->simple_rel_array = save_rel_array;
932 * If there was a FOR UPDATE/SHARE clause, the LockRows node will have
933 * dealt with fetching non-locked marked rows, else we need to have
934 * ModifyTable do that.
939 rowMarks = root->rowMarks;
941 /* And last, tack on a ModifyTable node to do the UPDATE/DELETE work */
942 return (Plan *) make_modifytable(parse->commandType,
948 SS_assign_special_param(root));
951 /*--------------------
953 * Perform planning steps related to grouping, aggregation, etc.
954 * This primarily means adding top-level processing to the basic
955 * query plan produced by query_planner.
957 * tuple_fraction is the fraction of tuples we expect will be retrieved
959 * tuple_fraction is interpreted as follows:
960 * 0: expect all tuples to be retrieved (normal case)
961 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
962 * from the plan to be retrieved
963 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
964 * expected to be retrieved (ie, a LIMIT specification)
966 * Returns a query plan. Also, root->query_pathkeys is returned as the
967 * actual output ordering of the plan (in pathkey format).
968 *--------------------
971 grouping_planner(PlannerInfo *root, double tuple_fraction)
973 Query *parse = root->parse;
974 List *tlist = parse->targetList;
975 int64 offset_est = 0;
977 double limit_tuples = -1.0;
979 List *current_pathkeys;
980 double dNumGroups = 0;
981 bool use_hashed_distinct = false;
982 bool tested_hashed_distinct = false;
984 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
985 if (parse->limitCount || parse->limitOffset)
987 tuple_fraction = preprocess_limit(root, tuple_fraction,
988 &offset_est, &count_est);
991 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
992 * estimate the effects of using a bounded sort.
994 if (count_est > 0 && offset_est >= 0)
995 limit_tuples = (double) count_est + (double) offset_est;
998 if (parse->setOperations)
1000 List *set_sortclauses;
1003 * If there's a top-level ORDER BY, assume we have to fetch all the
1004 * tuples. This might be too simplistic given all the hackery below
1005 * to possibly avoid the sort; but the odds of accurate estimates here
1006 * are pretty low anyway.
1008 if (parse->sortClause)
1009 tuple_fraction = 0.0;
1012 * Construct the plan for set operations. The result will not need
1013 * any work except perhaps a top-level sort and/or LIMIT. Note that
1014 * any special work for recursive unions is the responsibility of
1015 * plan_set_operations.
1017 result_plan = plan_set_operations(root, tuple_fraction,
1021 * Calculate pathkeys representing the sort order (if any) of the set
1022 * operation's result. We have to do this before overwriting the sort
1023 * key information...
1025 current_pathkeys = make_pathkeys_for_sortclauses(root,
1027 result_plan->targetlist,
1031 * We should not need to call preprocess_targetlist, since we must be
1032 * in a SELECT query node. Instead, use the targetlist returned by
1033 * plan_set_operations (since this tells whether it returned any
1034 * resjunk columns!), and transfer any sort key information from the
1037 Assert(parse->commandType == CMD_SELECT);
1039 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
1043 * Can't handle FOR UPDATE/SHARE here (parser should have checked
1044 * already, but let's make sure).
1046 if (parse->rowMarks)
1048 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1049 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
1052 * Calculate pathkeys that represent result ordering requirements
1054 Assert(parse->distinctClause == NIL);
1055 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1062 /* No set operations, do regular planning */
1064 double sub_limit_tuples;
1065 AttrNumber *groupColIdx = NULL;
1066 bool need_tlist_eval = true;
1067 QualCost tlist_cost;
1068 Path *cheapest_path;
1072 AggClauseCosts agg_costs;
1076 bool use_hashed_grouping = false;
1077 WindowFuncLists *wflists = NULL;
1078 List *activeWindows = NIL;
1080 MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1082 /* A recursive query should always have setOperations */
1083 Assert(!root->hasRecursion);
1085 /* Preprocess GROUP BY clause, if any */
1086 if (parse->groupClause)
1087 preprocess_groupclause(root);
1088 numGroupCols = list_length(parse->groupClause);
1090 /* Preprocess targetlist */
1091 tlist = preprocess_targetlist(root, tlist);
1094 * Locate any window functions in the tlist. (We don't need to look
1095 * anywhere else, since expressions used in ORDER BY will be in there
1096 * too.) Note that they could all have been eliminated by constant
1097 * folding, in which case we don't need to do any more work.
1099 if (parse->hasWindowFuncs)
1101 wflists = find_window_functions((Node *) tlist,
1102 list_length(parse->windowClause));
1103 if (wflists->numWindowFuncs > 0)
1104 activeWindows = select_active_windows(root, wflists);
1106 parse->hasWindowFuncs = false;
1110 * Generate appropriate target list for subplan; may be different from
1111 * tlist if grouping or aggregation is needed.
1113 sub_tlist = make_subplanTargetList(root, tlist,
1114 &groupColIdx, &need_tlist_eval);
1117 * Do aggregate preprocessing, if the query has any aggs.
1119 * Note: think not that we can turn off hasAggs if we find no aggs. It
1120 * is possible for constant-expression simplification to remove all
1121 * explicit references to aggs, but we still have to follow the
1122 * aggregate semantics (eg, producing only one output row).
1127 * Collect statistics about aggregates for estimating costs. Note:
1128 * we do not attempt to detect duplicate aggregates here; a
1129 * somewhat-overestimated cost is okay for our present purposes.
1131 count_agg_clauses(root, (Node *) tlist, &agg_costs);
1132 count_agg_clauses(root, parse->havingQual, &agg_costs);
1135 * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1136 * adding logic between here and the optimize_minmax_aggregates
1137 * call. Anything that is needed in MIN/MAX-optimizable cases
1138 * will have to be duplicated in planagg.c.
1140 preprocess_minmax_aggregates(root, tlist);
1144 * Calculate pathkeys that represent grouping/ordering requirements.
1145 * Stash them in PlannerInfo so that query_planner can canonicalize
1146 * them after EquivalenceClasses have been formed. The sortClause is
1147 * certainly sort-able, but GROUP BY and DISTINCT might not be, in
1148 * which case we just leave their pathkeys empty.
1150 if (parse->groupClause &&
1151 grouping_is_sortable(parse->groupClause))
1152 root->group_pathkeys =
1153 make_pathkeys_for_sortclauses(root,
1158 root->group_pathkeys = NIL;
1160 /* We consider only the first (bottom) window in pathkeys logic */
1161 if (activeWindows != NIL)
1163 WindowClause *wc = (WindowClause *) linitial(activeWindows);
1165 root->window_pathkeys = make_pathkeys_for_window(root,
1171 root->window_pathkeys = NIL;
1173 if (parse->distinctClause &&
1174 grouping_is_sortable(parse->distinctClause))
1175 root->distinct_pathkeys =
1176 make_pathkeys_for_sortclauses(root,
1177 parse->distinctClause,
1181 root->distinct_pathkeys = NIL;
1183 root->sort_pathkeys =
1184 make_pathkeys_for_sortclauses(root,
1190 * Figure out whether we want a sorted result from query_planner.
1192 * If we have a sortable GROUP BY clause, then we want a result sorted
1193 * properly for grouping. Otherwise, if we have window functions to
1194 * evaluate, we try to sort for the first window. Otherwise, if
1195 * there's a sortable DISTINCT clause that's more rigorous than the
1196 * ORDER BY clause, we try to produce output that's sufficiently well
1197 * sorted for the DISTINCT. Otherwise, if there is an ORDER BY
1198 * clause, we want to sort by the ORDER BY clause.
1200 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
1201 * superset of GROUP BY, it would be tempting to request sort by ORDER
1202 * BY --- but that might just leave us failing to exploit an available
1203 * sort order at all. Needs more thought. The choice for DISTINCT
1204 * versus ORDER BY is much easier, since we know that the parser
1205 * ensured that one is a superset of the other.
1207 if (root->group_pathkeys)
1208 root->query_pathkeys = root->group_pathkeys;
1209 else if (root->window_pathkeys)
1210 root->query_pathkeys = root->window_pathkeys;
1211 else if (list_length(root->distinct_pathkeys) >
1212 list_length(root->sort_pathkeys))
1213 root->query_pathkeys = root->distinct_pathkeys;
1214 else if (root->sort_pathkeys)
1215 root->query_pathkeys = root->sort_pathkeys;
1217 root->query_pathkeys = NIL;
1220 * Figure out whether there's a hard limit on the number of rows that
1221 * query_planner's result subplan needs to return. Even if we know a
1222 * hard limit overall, it doesn't apply if the query has any
1223 * grouping/aggregation operations.
1225 if (parse->groupClause ||
1226 parse->distinctClause ||
1228 parse->hasWindowFuncs ||
1229 root->hasHavingQual)
1230 sub_limit_tuples = -1.0;
1232 sub_limit_tuples = limit_tuples;
1235 * Generate the best unsorted and presorted paths for this Query (but
1236 * note there may not be any presorted path). query_planner will also
1237 * estimate the number of groups in the query, and canonicalize all
1240 query_planner(root, sub_tlist, tuple_fraction, sub_limit_tuples,
1241 &cheapest_path, &sorted_path, &dNumGroups);
1244 * Extract rowcount and width estimates for possible use in grouping
1245 * decisions. Beware here of the possibility that
1246 * cheapest_path->parent is NULL (ie, there is no FROM clause).
1248 if (cheapest_path->parent)
1250 path_rows = cheapest_path->parent->rows;
1251 path_width = cheapest_path->parent->width;
1255 path_rows = 1; /* assume non-set result */
1256 path_width = 100; /* arbitrary */
1259 if (parse->groupClause)
1262 * If grouping, decide whether to use sorted or hashed grouping.
1264 use_hashed_grouping =
1265 choose_hashed_grouping(root,
1266 tuple_fraction, limit_tuples,
1267 path_rows, path_width,
1268 cheapest_path, sorted_path,
1269 dNumGroups, &agg_costs);
1270 /* Also convert # groups to long int --- but 'ware overflow! */
1271 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
1273 else if (parse->distinctClause && sorted_path &&
1274 !root->hasHavingQual && !parse->hasAggs && !activeWindows)
1277 * We'll reach the DISTINCT stage without any intermediate
1278 * processing, so figure out whether we will want to hash or not
1279 * so we can choose whether to use cheapest or sorted path.
1281 use_hashed_distinct =
1282 choose_hashed_distinct(root,
1283 tuple_fraction, limit_tuples,
1284 path_rows, path_width,
1285 cheapest_path->startup_cost,
1286 cheapest_path->total_cost,
1287 sorted_path->startup_cost,
1288 sorted_path->total_cost,
1289 sorted_path->pathkeys,
1291 tested_hashed_distinct = true;
1295 * Select the best path. If we are doing hashed grouping, we will
1296 * always read all the input tuples, so use the cheapest-total path.
1297 * Otherwise, trust query_planner's decision about which to use.
1299 if (use_hashed_grouping || use_hashed_distinct || !sorted_path)
1300 best_path = cheapest_path;
1302 best_path = sorted_path;
1305 * Check to see if it's possible to optimize MIN/MAX aggregates. If
1306 * so, we will forget all the work we did so far to choose a "regular"
1307 * path ... but we had to do it anyway to be able to tell which way is
1310 result_plan = optimize_minmax_aggregates(root,
1314 if (result_plan != NULL)
1317 * optimize_minmax_aggregates generated the full plan, with the
1318 * right tlist, and it has no sort order.
1320 current_pathkeys = NIL;
1325 * Normal case --- create a plan according to query_planner's
1328 bool need_sort_for_grouping = false;
1330 result_plan = create_plan(root, best_path);
1331 current_pathkeys = best_path->pathkeys;
1333 /* Detect if we'll need an explicit sort for grouping */
1334 if (parse->groupClause && !use_hashed_grouping &&
1335 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1337 need_sort_for_grouping = true;
1340 * Always override create_plan's tlist, so that we don't
1341 * sort useless data from a "physical" tlist.
1343 need_tlist_eval = true;
1347 * create_plan returns a plan with just a "flat" tlist of
1348 * required Vars. Usually we need to insert the sub_tlist as the
1349 * tlist of the top plan node. However, we can skip that if we
1350 * determined that whatever create_plan chose to return will be
1353 if (need_tlist_eval)
1356 * If the top-level plan node is one that cannot do expression
1357 * evaluation, we must insert a Result node to project the
1360 if (!is_projection_capable_plan(result_plan))
1362 result_plan = (Plan *) make_result(root,
1370 * Otherwise, just replace the subplan's flat tlist with
1371 * the desired tlist.
1373 result_plan->targetlist = sub_tlist;
1377 * Also, account for the cost of evaluation of the sub_tlist.
1379 * Up to now, we have only been dealing with "flat" tlists,
1380 * containing just Vars. So their evaluation cost is zero
1381 * according to the model used by cost_qual_eval() (or if you
1382 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1383 * can avoid accounting for tlist cost throughout
1384 * query_planner() and subroutines. But now we've inserted a
1385 * tlist that might contain actual operators, sub-selects, etc
1386 * --- so we'd better account for its cost.
1388 * Below this point, any tlist eval cost for added-on nodes
1389 * should be accounted for as we create those nodes.
1390 * Presently, of the node types we can add on, only Agg,
1391 * WindowAgg, and Group project new tlists (the rest just copy
1392 * their input tuples) --- so make_agg(), make_windowagg() and
1393 * make_group() are responsible for computing the added cost.
1395 cost_qual_eval(&tlist_cost, sub_tlist, root);
1396 result_plan->startup_cost += tlist_cost.startup;
1397 result_plan->total_cost += tlist_cost.startup +
1398 tlist_cost.per_tuple * result_plan->plan_rows;
1403 * Since we're using create_plan's tlist and not the one
1404 * make_subplanTargetList calculated, we have to refigure any
1405 * grouping-column indexes make_subplanTargetList computed.
1407 locate_grouping_columns(root, tlist, result_plan->targetlist,
1412 * Insert AGG or GROUP node if needed, plus an explicit sort step
1415 * HAVING clause, if any, becomes qual of the Agg or Group node.
1417 if (use_hashed_grouping)
1419 /* Hashed aggregate plan --- no sort needed */
1420 result_plan = (Plan *) make_agg(root,
1422 (List *) parse->havingQual,
1427 extract_grouping_ops(parse->groupClause),
1430 /* Hashed aggregation produces randomly-ordered results */
1431 current_pathkeys = NIL;
1433 else if (parse->hasAggs)
1435 /* Plain aggregate plan --- sort if needed */
1436 AggStrategy aggstrategy;
1438 if (parse->groupClause)
1440 if (need_sort_for_grouping)
1442 result_plan = (Plan *)
1443 make_sort_from_groupcols(root,
1447 current_pathkeys = root->group_pathkeys;
1449 aggstrategy = AGG_SORTED;
1452 * The AGG node will not change the sort ordering of its
1453 * groups, so current_pathkeys describes the result too.
1458 aggstrategy = AGG_PLAIN;
1459 /* Result will be only one row anyway; no sort order */
1460 current_pathkeys = NIL;
1463 result_plan = (Plan *) make_agg(root,
1465 (List *) parse->havingQual,
1470 extract_grouping_ops(parse->groupClause),
1474 else if (parse->groupClause)
1477 * GROUP BY without aggregation, so insert a group node (plus
1478 * the appropriate sort node, if necessary).
1480 * Add an explicit sort if we couldn't make the path come out
1481 * the way the GROUP node needs it.
1483 if (need_sort_for_grouping)
1485 result_plan = (Plan *)
1486 make_sort_from_groupcols(root,
1490 current_pathkeys = root->group_pathkeys;
1493 result_plan = (Plan *) make_group(root,
1495 (List *) parse->havingQual,
1498 extract_grouping_ops(parse->groupClause),
1501 /* The Group node won't change sort ordering */
1503 else if (root->hasHavingQual)
1506 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1507 * This is a degenerate case in which we are supposed to emit
1508 * either 0 or 1 row depending on whether HAVING succeeds.
1509 * Furthermore, there cannot be any variables in either HAVING
1510 * or the targetlist, so we actually do not need the FROM
1511 * table at all! We can just throw away the plan-so-far and
1512 * generate a Result node. This is a sufficiently unusual
1513 * corner case that it's not worth contorting the structure of
1514 * this routine to avoid having to generate the plan in the
1517 result_plan = (Plan *) make_result(root,
1522 } /* end of non-minmax-aggregate case */
1525 * Since each window function could require a different sort order, we
1526 * stack up a WindowAgg node for each window, with sort steps between
1535 * If the top-level plan node is one that cannot do expression
1536 * evaluation, we must insert a Result node to project the desired
1537 * tlist. (In some cases this might not really be required, but
1538 * it's not worth trying to avoid it.) Note that on second and
1539 * subsequent passes through the following loop, the top-level
1540 * node will be a WindowAgg which we know can project; so we only
1541 * need to check once.
1543 if (!is_projection_capable_plan(result_plan))
1545 result_plan = (Plan *) make_result(root,
1552 * The "base" targetlist for all steps of the windowing process is
1553 * a flat tlist of all Vars and Aggs needed in the result. (In
1554 * some cases we wouldn't need to propagate all of these all the
1555 * way to the top, since they might only be needed as inputs to
1556 * WindowFuncs. It's probably not worth trying to optimize that
1557 * though.) We also need any volatile sort expressions, because
1558 * make_sort_from_pathkeys won't add those on its own, and anyway
1559 * we want them evaluated only once at the bottom of the stack. As
1560 * we climb up the stack, we add outputs for the WindowFuncs
1561 * computed at each level. Also, each input tlist has to present
1562 * all the columns needed to sort the data for the next WindowAgg
1563 * step. That's handled internally by make_sort_from_pathkeys,
1564 * but we need the copyObject steps here to ensure that each plan
1565 * node has a separately modifiable tlist.
1567 * Note: it's essential here to use PVC_INCLUDE_AGGREGATES so that
1568 * Vars mentioned only in aggregate expressions aren't pulled out
1569 * as separate targetlist entries. Otherwise we could be putting
1570 * ungrouped Vars directly into an Agg node's tlist, resulting in
1571 * undefined behavior.
1573 window_tlist = flatten_tlist(tlist,
1574 PVC_INCLUDE_AGGREGATES,
1575 PVC_INCLUDE_PLACEHOLDERS);
1576 window_tlist = add_volatile_sort_exprs(window_tlist, tlist,
1578 result_plan->targetlist = (List *) copyObject(window_tlist);
1580 foreach(l, activeWindows)
1582 WindowClause *wc = (WindowClause *) lfirst(l);
1583 List *window_pathkeys;
1585 AttrNumber *partColIdx;
1588 AttrNumber *ordColIdx;
1591 window_pathkeys = make_pathkeys_for_window(root,
1597 * This is a bit tricky: we build a sort node even if we don't
1598 * really have to sort. Even when no explicit sort is needed,
1599 * we need to have suitable resjunk items added to the input
1600 * plan's tlist for any partitioning or ordering columns that
1601 * aren't plain Vars. Furthermore, this way we can use
1602 * existing infrastructure to identify which input columns are
1603 * the interesting ones.
1605 if (window_pathkeys)
1609 sort_plan = make_sort_from_pathkeys(root,
1613 if (!pathkeys_contained_in(window_pathkeys,
1616 /* we do indeed need to sort */
1617 result_plan = (Plan *) sort_plan;
1618 current_pathkeys = window_pathkeys;
1620 /* In either case, extract the per-column information */
1621 get_column_info_for_window(root, wc, tlist,
1623 sort_plan->sortColIdx,
1633 /* empty window specification, nothing to sort */
1636 partOperators = NULL;
1639 ordOperators = NULL;
1644 /* Add the current WindowFuncs to the running tlist */
1645 window_tlist = add_to_flat_tlist(window_tlist,
1646 wflists->windowFuncs[wc->winref]);
1650 /* Install the original tlist in the topmost WindowAgg */
1651 window_tlist = tlist;
1654 /* ... and make the WindowAgg plan node */
1655 result_plan = (Plan *)
1656 make_windowagg(root,
1657 (List *) copyObject(window_tlist),
1658 wflists->windowFuncs[wc->winref],
1672 } /* end of if (setOperations) */
1675 * If there is a DISTINCT clause, add the necessary node(s).
1677 if (parse->distinctClause)
1679 double dNumDistinctRows;
1680 long numDistinctRows;
1683 * If there was grouping or aggregation, use the current number of
1684 * rows as the estimated number of DISTINCT rows (ie, assume the
1685 * result was already mostly unique). If not, use the number of
1686 * distinct-groups calculated by query_planner.
1688 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1689 dNumDistinctRows = result_plan->plan_rows;
1691 dNumDistinctRows = dNumGroups;
1693 /* Also convert to long int --- but 'ware overflow! */
1694 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1696 /* Choose implementation method if we didn't already */
1697 if (!tested_hashed_distinct)
1700 * At this point, either hashed or sorted grouping will have to
1701 * work from result_plan, so we pass that as both "cheapest" and
1704 use_hashed_distinct =
1705 choose_hashed_distinct(root,
1706 tuple_fraction, limit_tuples,
1707 result_plan->plan_rows,
1708 result_plan->plan_width,
1709 result_plan->startup_cost,
1710 result_plan->total_cost,
1711 result_plan->startup_cost,
1712 result_plan->total_cost,
1717 if (use_hashed_distinct)
1719 /* Hashed aggregate plan --- no sort needed */
1720 result_plan = (Plan *) make_agg(root,
1721 result_plan->targetlist,
1725 list_length(parse->distinctClause),
1726 extract_grouping_cols(parse->distinctClause,
1727 result_plan->targetlist),
1728 extract_grouping_ops(parse->distinctClause),
1731 /* Hashed aggregation produces randomly-ordered results */
1732 current_pathkeys = NIL;
1737 * Use a Unique node to implement DISTINCT. Add an explicit sort
1738 * if we couldn't make the path come out the way the Unique node
1739 * needs it. If we do have to sort, always sort by the more
1740 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1741 * below. However, for regular DISTINCT, don't sort now if we
1742 * don't have to --- sorting afterwards will likely be cheaper,
1743 * and also has the possibility of optimizing via LIMIT. But for
1744 * DISTINCT ON, we *must* force the final sort now, else it won't
1745 * have the desired behavior.
1747 List *needed_pathkeys;
1749 if (parse->hasDistinctOn &&
1750 list_length(root->distinct_pathkeys) <
1751 list_length(root->sort_pathkeys))
1752 needed_pathkeys = root->sort_pathkeys;
1754 needed_pathkeys = root->distinct_pathkeys;
1756 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1758 if (list_length(root->distinct_pathkeys) >=
1759 list_length(root->sort_pathkeys))
1760 current_pathkeys = root->distinct_pathkeys;
1763 current_pathkeys = root->sort_pathkeys;
1764 /* Assert checks that parser didn't mess up... */
1765 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1769 result_plan = (Plan *) make_sort_from_pathkeys(root,
1775 result_plan = (Plan *) make_unique(result_plan,
1776 parse->distinctClause);
1777 result_plan->plan_rows = dNumDistinctRows;
1778 /* The Unique node won't change sort ordering */
1783 * If ORDER BY was given and we were not able to make the plan come out in
1784 * the right order, add an explicit sort step.
1786 if (parse->sortClause)
1788 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1790 result_plan = (Plan *) make_sort_from_pathkeys(root,
1792 root->sort_pathkeys,
1794 current_pathkeys = root->sort_pathkeys;
1799 * If there is a FOR UPDATE/SHARE clause, add the LockRows node. (Note: we
1800 * intentionally test parse->rowMarks not root->rowMarks here. If there
1801 * are only non-locking rowmarks, they should be handled by the
1802 * ModifyTable node instead.)
1804 if (parse->rowMarks)
1806 result_plan = (Plan *) make_lockrows(result_plan,
1808 SS_assign_special_param(root));
1811 * The result can no longer be assumed sorted, since locking might
1812 * cause the sort key columns to be replaced with new values.
1814 current_pathkeys = NIL;
1818 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1820 if (parse->limitCount || parse->limitOffset)
1822 result_plan = (Plan *) make_limit(result_plan,
1830 * Return the actual output ordering in query_pathkeys for possible use by
1831 * an outer query level.
1833 root->query_pathkeys = current_pathkeys;
1839 * Detect whether a plan node is a "dummy" plan created when a relation
1840 * is deemed not to need scanning due to constraint exclusion.
1842 * Currently, such dummy plans are Result nodes with constant FALSE
1843 * filter quals (see set_dummy_rel_pathlist and create_append_plan).
1845 * XXX this probably ought to be somewhere else, but not clear where.
1848 is_dummy_plan(Plan *plan)
1850 if (IsA(plan, Result))
1852 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1854 if (list_length(rcqual) == 1)
1856 Const *constqual = (Const *) linitial(rcqual);
1858 if (constqual && IsA(constqual, Const))
1860 if (!constqual->constisnull &&
1861 !DatumGetBool(constqual->constvalue))
1870 * Create a bitmapset of the RT indexes of live base relations
1872 * Helper for preprocess_rowmarks ... at this point in the proceedings,
1873 * the only good way to distinguish baserels from appendrel children
1874 * is to see what is in the join tree.
1877 get_base_rel_indexes(Node *jtnode)
1883 if (IsA(jtnode, RangeTblRef))
1885 int varno = ((RangeTblRef *) jtnode)->rtindex;
1887 result = bms_make_singleton(varno);
1889 else if (IsA(jtnode, FromExpr))
1891 FromExpr *f = (FromExpr *) jtnode;
1895 foreach(l, f->fromlist)
1896 result = bms_join(result,
1897 get_base_rel_indexes(lfirst(l)));
1899 else if (IsA(jtnode, JoinExpr))
1901 JoinExpr *j = (JoinExpr *) jtnode;
1903 result = bms_join(get_base_rel_indexes(j->larg),
1904 get_base_rel_indexes(j->rarg));
1908 elog(ERROR, "unrecognized node type: %d",
1909 (int) nodeTag(jtnode));
1910 result = NULL; /* keep compiler quiet */
1916 * preprocess_rowmarks - set up PlanRowMarks if needed
1919 preprocess_rowmarks(PlannerInfo *root)
1921 Query *parse = root->parse;
1927 if (parse->rowMarks)
1930 * We've got trouble if FOR UPDATE/SHARE appears inside grouping,
1931 * since grouping renders a reference to individual tuple CTIDs
1932 * invalid. This is also checked at parse time, but that's
1933 * insufficient because of rule substitution, query pullup, etc.
1935 CheckSelectLocking(parse);
1940 * We only need rowmarks for UPDATE, DELETE, or FOR UPDATE/SHARE.
1942 if (parse->commandType != CMD_UPDATE &&
1943 parse->commandType != CMD_DELETE)
1948 * We need to have rowmarks for all base relations except the target. We
1949 * make a bitmapset of all base rels and then remove the items we don't
1950 * need or have FOR UPDATE/SHARE marks for.
1952 rels = get_base_rel_indexes((Node *) parse->jointree);
1953 if (parse->resultRelation)
1954 rels = bms_del_member(rels, parse->resultRelation);
1957 * Convert RowMarkClauses to PlanRowMark representation.
1960 foreach(l, parse->rowMarks)
1962 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
1963 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
1967 * Currently, it is syntactically impossible to have FOR UPDATE
1968 * applied to an update/delete target rel. If that ever becomes
1969 * possible, we should drop the target from the PlanRowMark list.
1971 Assert(rc->rti != parse->resultRelation);
1974 * Ignore RowMarkClauses for subqueries; they aren't real tables and
1975 * can't support true locking. Subqueries that got flattened into the
1976 * main query should be ignored completely. Any that didn't will get
1977 * ROW_MARK_COPY items in the next loop.
1979 if (rte->rtekind != RTE_RELATION)
1982 rels = bms_del_member(rels, rc->rti);
1984 newrc = makeNode(PlanRowMark);
1985 newrc->rti = newrc->prti = rc->rti;
1986 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
1988 newrc->markType = ROW_MARK_EXCLUSIVE;
1990 newrc->markType = ROW_MARK_SHARE;
1991 newrc->noWait = rc->noWait;
1992 newrc->isParent = false;
1994 prowmarks = lappend(prowmarks, newrc);
1998 * Now, add rowmarks for any non-target, non-locked base relations.
2001 foreach(l, parse->rtable)
2003 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
2007 if (!bms_is_member(i, rels))
2010 newrc = makeNode(PlanRowMark);
2011 newrc->rti = newrc->prti = i;
2012 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2013 /* real tables support REFERENCE, anything else needs COPY */
2014 if (rte->rtekind == RTE_RELATION &&
2015 rte->relkind != RELKIND_FOREIGN_TABLE)
2016 newrc->markType = ROW_MARK_REFERENCE;
2018 newrc->markType = ROW_MARK_COPY;
2019 newrc->noWait = false; /* doesn't matter */
2020 newrc->isParent = false;
2022 prowmarks = lappend(prowmarks, newrc);
2025 root->rowMarks = prowmarks;
2029 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2031 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2032 * results back in *count_est and *offset_est. These variables are set to
2033 * 0 if the corresponding clause is not present, and -1 if it's present
2034 * but we couldn't estimate the value for it. (The "0" convention is OK
2035 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2036 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2037 * usual practice of never estimating less than one row.) These values will
2038 * be passed to make_limit, which see if you change this code.
2040 * The return value is the suitably adjusted tuple_fraction to use for
2041 * planning the query. This adjustment is not overridable, since it reflects
2042 * plan actions that grouping_planner() will certainly take, not assumptions
2046 preprocess_limit(PlannerInfo *root, double tuple_fraction,
2047 int64 *offset_est, int64 *count_est)
2049 Query *parse = root->parse;
2051 double limit_fraction;
2053 /* Should not be called unless LIMIT or OFFSET */
2054 Assert(parse->limitCount || parse->limitOffset);
2057 * Try to obtain the clause values. We use estimate_expression_value
2058 * primarily because it can sometimes do something useful with Params.
2060 if (parse->limitCount)
2062 est = estimate_expression_value(root, parse->limitCount);
2063 if (est && IsA(est, Const))
2065 if (((Const *) est)->constisnull)
2067 /* NULL indicates LIMIT ALL, ie, no limit */
2068 *count_est = 0; /* treat as not present */
2072 *count_est = DatumGetInt64(((Const *) est)->constvalue);
2073 if (*count_est <= 0)
2074 *count_est = 1; /* force to at least 1 */
2078 *count_est = -1; /* can't estimate */
2081 *count_est = 0; /* not present */
2083 if (parse->limitOffset)
2085 est = estimate_expression_value(root, parse->limitOffset);
2086 if (est && IsA(est, Const))
2088 if (((Const *) est)->constisnull)
2090 /* Treat NULL as no offset; the executor will too */
2091 *offset_est = 0; /* treat as not present */
2095 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2096 if (*offset_est < 0)
2097 *offset_est = 0; /* less than 0 is same as 0 */
2101 *offset_est = -1; /* can't estimate */
2104 *offset_est = 0; /* not present */
2106 if (*count_est != 0)
2109 * A LIMIT clause limits the absolute number of tuples returned.
2110 * However, if it's not a constant LIMIT then we have to guess; for
2111 * lack of a better idea, assume 10% of the plan's result is wanted.
2113 if (*count_est < 0 || *offset_est < 0)
2115 /* LIMIT or OFFSET is an expression ... punt ... */
2116 limit_fraction = 0.10;
2120 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2121 limit_fraction = (double) *count_est + (double) *offset_est;
2125 * If we have absolute limits from both caller and LIMIT, use the
2126 * smaller value; likewise if they are both fractional. If one is
2127 * fractional and the other absolute, we can't easily determine which
2128 * is smaller, but we use the heuristic that the absolute will usually
2131 if (tuple_fraction >= 1.0)
2133 if (limit_fraction >= 1.0)
2136 tuple_fraction = Min(tuple_fraction, limit_fraction);
2140 /* caller absolute, limit fractional; use caller's value */
2143 else if (tuple_fraction > 0.0)
2145 if (limit_fraction >= 1.0)
2147 /* caller fractional, limit absolute; use limit */
2148 tuple_fraction = limit_fraction;
2152 /* both fractional */
2153 tuple_fraction = Min(tuple_fraction, limit_fraction);
2158 /* no info from caller, just use limit */
2159 tuple_fraction = limit_fraction;
2162 else if (*offset_est != 0 && tuple_fraction > 0.0)
2165 * We have an OFFSET but no LIMIT. This acts entirely differently
2166 * from the LIMIT case: here, we need to increase rather than decrease
2167 * the caller's tuple_fraction, because the OFFSET acts to cause more
2168 * tuples to be fetched instead of fewer. This only matters if we got
2169 * a tuple_fraction > 0, however.
2171 * As above, use 10% if OFFSET is present but unestimatable.
2173 if (*offset_est < 0)
2174 limit_fraction = 0.10;
2176 limit_fraction = (double) *offset_est;
2179 * If we have absolute counts from both caller and OFFSET, add them
2180 * together; likewise if they are both fractional. If one is
2181 * fractional and the other absolute, we want to take the larger, and
2182 * we heuristically assume that's the fractional one.
2184 if (tuple_fraction >= 1.0)
2186 if (limit_fraction >= 1.0)
2188 /* both absolute, so add them together */
2189 tuple_fraction += limit_fraction;
2193 /* caller absolute, limit fractional; use limit */
2194 tuple_fraction = limit_fraction;
2199 if (limit_fraction >= 1.0)
2201 /* caller fractional, limit absolute; use caller's value */
2205 /* both fractional, so add them together */
2206 tuple_fraction += limit_fraction;
2207 if (tuple_fraction >= 1.0)
2208 tuple_fraction = 0.0; /* assume fetch all */
2213 return tuple_fraction;
2218 * preprocess_groupclause - do preparatory work on GROUP BY clause
2220 * The idea here is to adjust the ordering of the GROUP BY elements
2221 * (which in itself is semantically insignificant) to match ORDER BY,
2222 * thereby allowing a single sort operation to both implement the ORDER BY
2223 * requirement and set up for a Unique step that implements GROUP BY.
2225 * In principle it might be interesting to consider other orderings of the
2226 * GROUP BY elements, which could match the sort ordering of other
2227 * possible plans (eg an indexscan) and thereby reduce cost. We don't
2228 * bother with that, though. Hashed grouping will frequently win anyway.
2230 * Note: we need no comparable processing of the distinctClause because
2231 * the parser already enforced that that matches ORDER BY.
2234 preprocess_groupclause(PlannerInfo *root)
2236 Query *parse = root->parse;
2237 List *new_groupclause;
2242 /* If no ORDER BY, nothing useful to do here */
2243 if (parse->sortClause == NIL)
2247 * Scan the ORDER BY clause and construct a list of matching GROUP BY
2248 * items, but only as far as we can make a matching prefix.
2250 * This code assumes that the sortClause contains no duplicate items.
2252 new_groupclause = NIL;
2253 foreach(sl, parse->sortClause)
2255 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
2257 foreach(gl, parse->groupClause)
2259 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2263 new_groupclause = lappend(new_groupclause, gc);
2268 break; /* no match, so stop scanning */
2271 /* Did we match all of the ORDER BY list, or just some of it? */
2272 partial_match = (sl != NULL);
2274 /* If no match at all, no point in reordering GROUP BY */
2275 if (new_groupclause == NIL)
2279 * Add any remaining GROUP BY items to the new list, but only if we were
2280 * able to make a complete match. In other words, we only rearrange the
2281 * GROUP BY list if the result is that one list is a prefix of the other
2282 * --- otherwise there's no possibility of a common sort. Also, give up
2283 * if there are any non-sortable GROUP BY items, since then there's no
2286 foreach(gl, parse->groupClause)
2288 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2290 if (list_member_ptr(new_groupclause, gc))
2291 continue; /* it matched an ORDER BY item */
2293 return; /* give up, no common sort possible */
2294 if (!OidIsValid(gc->sortop))
2295 return; /* give up, GROUP BY can't be sorted */
2296 new_groupclause = lappend(new_groupclause, gc);
2299 /* Success --- install the rearranged GROUP BY list */
2300 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2301 parse->groupClause = new_groupclause;
2305 * choose_hashed_grouping - should we use hashed grouping?
2307 * Returns TRUE to select hashing, FALSE to select sorting.
2310 choose_hashed_grouping(PlannerInfo *root,
2311 double tuple_fraction, double limit_tuples,
2312 double path_rows, int path_width,
2313 Path *cheapest_path, Path *sorted_path,
2314 double dNumGroups, AggClauseCosts *agg_costs)
2316 Query *parse = root->parse;
2317 int numGroupCols = list_length(parse->groupClause);
2321 List *target_pathkeys;
2322 List *current_pathkeys;
2327 * Executor doesn't support hashed aggregation with DISTINCT or ORDER BY
2328 * aggregates. (Doing so would imply storing *all* the input values in
2329 * the hash table, and/or running many sorts in parallel, either of which
2330 * seems like a certain loser.)
2332 can_hash = (agg_costs->numOrderedAggs == 0 &&
2333 grouping_is_hashable(parse->groupClause));
2334 can_sort = grouping_is_sortable(parse->groupClause);
2336 /* Quick out if only one choice is workable */
2337 if (!(can_hash && can_sort))
2345 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2346 errmsg("could not implement GROUP BY"),
2347 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2350 /* Prefer sorting when enable_hashagg is off */
2351 if (!enable_hashagg)
2355 * Don't do it if it doesn't look like the hashtable will fit into
2359 /* Estimate per-hash-entry space at tuple width... */
2360 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2361 /* plus space for pass-by-ref transition values... */
2362 hashentrysize += agg_costs->transitionSpace;
2363 /* plus the per-hash-entry overhead */
2364 hashentrysize += hash_agg_entry_size(agg_costs->numAggs);
2366 if (hashentrysize * dNumGroups > work_mem * 1024L)
2370 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
2371 * DISTINCT and ORDER BY as the assumed required output sort order. This
2372 * is an oversimplification because the DISTINCT might get implemented via
2373 * hashing, but it's not clear that the case is common enough (or that our
2374 * estimates are good enough) to justify trying to solve it exactly.
2376 if (list_length(root->distinct_pathkeys) >
2377 list_length(root->sort_pathkeys))
2378 target_pathkeys = root->distinct_pathkeys;
2380 target_pathkeys = root->sort_pathkeys;
2383 * See if the estimated cost is no more than doing it the other way. While
2384 * avoiding the need for sorted input is usually a win, the fact that the
2385 * output won't be sorted may be a loss; so we need to do an actual cost
2388 * We need to consider cheapest_path + hashagg [+ final sort] versus
2389 * either cheapest_path [+ sort] + group or agg [+ final sort] or
2390 * presorted_path + group or agg [+ final sort] where brackets indicate a
2391 * step that may not be needed. We assume query_planner() will have
2392 * returned a presorted path only if it's a winner compared to
2393 * cheapest_path for this purpose.
2395 * These path variables are dummies that just hold cost fields; we don't
2396 * make actual Paths for these steps.
2398 cost_agg(&hashed_p, root, AGG_HASHED, agg_costs,
2399 numGroupCols, dNumGroups,
2400 cheapest_path->startup_cost, cheapest_path->total_cost,
2402 /* Result of hashed agg is always unsorted */
2403 if (target_pathkeys)
2404 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
2405 dNumGroups, path_width,
2406 0.0, work_mem, limit_tuples);
2410 sorted_p.startup_cost = sorted_path->startup_cost;
2411 sorted_p.total_cost = sorted_path->total_cost;
2412 current_pathkeys = sorted_path->pathkeys;
2416 sorted_p.startup_cost = cheapest_path->startup_cost;
2417 sorted_p.total_cost = cheapest_path->total_cost;
2418 current_pathkeys = cheapest_path->pathkeys;
2420 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
2422 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
2423 path_rows, path_width,
2424 0.0, work_mem, -1.0);
2425 current_pathkeys = root->group_pathkeys;
2429 cost_agg(&sorted_p, root, AGG_SORTED, agg_costs,
2430 numGroupCols, dNumGroups,
2431 sorted_p.startup_cost, sorted_p.total_cost,
2434 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
2435 sorted_p.startup_cost, sorted_p.total_cost,
2437 /* The Agg or Group node will preserve ordering */
2438 if (target_pathkeys &&
2439 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
2440 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
2441 dNumGroups, path_width,
2442 0.0, work_mem, limit_tuples);
2445 * Now make the decision using the top-level tuple fraction. First we
2446 * have to convert an absolute count (LIMIT) into fractional form.
2448 if (tuple_fraction >= 1.0)
2449 tuple_fraction /= dNumGroups;
2451 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2452 tuple_fraction) < 0)
2454 /* Hashed is cheaper, so use it */
2461 * choose_hashed_distinct - should we use hashing for DISTINCT?
2463 * This is fairly similar to choose_hashed_grouping, but there are enough
2464 * differences that it doesn't seem worth trying to unify the two functions.
2465 * (One difference is that we sometimes apply this after forming a Plan,
2466 * so the input alternatives can't be represented as Paths --- instead we
2467 * pass in the costs as individual variables.)
2469 * But note that making the two choices independently is a bit bogus in
2470 * itself. If the two could be combined into a single choice operation
2471 * it'd probably be better, but that seems far too unwieldy to be practical,
2472 * especially considering that the combination of GROUP BY and DISTINCT
2473 * isn't very common in real queries. By separating them, we are giving
2474 * extra preference to using a sorting implementation when a common sort key
2475 * is available ... and that's not necessarily wrong anyway.
2477 * Returns TRUE to select hashing, FALSE to select sorting.
2480 choose_hashed_distinct(PlannerInfo *root,
2481 double tuple_fraction, double limit_tuples,
2482 double path_rows, int path_width,
2483 Cost cheapest_startup_cost, Cost cheapest_total_cost,
2484 Cost sorted_startup_cost, Cost sorted_total_cost,
2485 List *sorted_pathkeys,
2486 double dNumDistinctRows)
2488 Query *parse = root->parse;
2489 int numDistinctCols = list_length(parse->distinctClause);
2493 List *current_pathkeys;
2494 List *needed_pathkeys;
2499 * If we have a sortable DISTINCT ON clause, we always use sorting. This
2500 * enforces the expected behavior of DISTINCT ON.
2502 can_sort = grouping_is_sortable(parse->distinctClause);
2503 if (can_sort && parse->hasDistinctOn)
2506 can_hash = grouping_is_hashable(parse->distinctClause);
2508 /* Quick out if only one choice is workable */
2509 if (!(can_hash && can_sort))
2517 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2518 errmsg("could not implement DISTINCT"),
2519 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2522 /* Prefer sorting when enable_hashagg is off */
2523 if (!enable_hashagg)
2527 * Don't do it if it doesn't look like the hashtable will fit into
2530 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2532 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
2536 * See if the estimated cost is no more than doing it the other way. While
2537 * avoiding the need for sorted input is usually a win, the fact that the
2538 * output won't be sorted may be a loss; so we need to do an actual cost
2541 * We need to consider cheapest_path + hashagg [+ final sort] versus
2542 * sorted_path [+ sort] + group [+ final sort] where brackets indicate a
2543 * step that may not be needed.
2545 * These path variables are dummies that just hold cost fields; we don't
2546 * make actual Paths for these steps.
2548 cost_agg(&hashed_p, root, AGG_HASHED, NULL,
2549 numDistinctCols, dNumDistinctRows,
2550 cheapest_startup_cost, cheapest_total_cost,
2554 * Result of hashed agg is always unsorted, so if ORDER BY is present we
2555 * need to charge for the final sort.
2557 if (parse->sortClause)
2558 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
2559 dNumDistinctRows, path_width,
2560 0.0, work_mem, limit_tuples);
2563 * Now for the GROUP case. See comments in grouping_planner about the
2564 * sorting choices here --- this code should match that code.
2566 sorted_p.startup_cost = sorted_startup_cost;
2567 sorted_p.total_cost = sorted_total_cost;
2568 current_pathkeys = sorted_pathkeys;
2569 if (parse->hasDistinctOn &&
2570 list_length(root->distinct_pathkeys) <
2571 list_length(root->sort_pathkeys))
2572 needed_pathkeys = root->sort_pathkeys;
2574 needed_pathkeys = root->distinct_pathkeys;
2575 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
2577 if (list_length(root->distinct_pathkeys) >=
2578 list_length(root->sort_pathkeys))
2579 current_pathkeys = root->distinct_pathkeys;
2581 current_pathkeys = root->sort_pathkeys;
2582 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
2583 path_rows, path_width,
2584 0.0, work_mem, -1.0);
2586 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
2587 sorted_p.startup_cost, sorted_p.total_cost,
2589 if (parse->sortClause &&
2590 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
2591 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
2592 dNumDistinctRows, path_width,
2593 0.0, work_mem, limit_tuples);
2596 * Now make the decision using the top-level tuple fraction. First we
2597 * have to convert an absolute count (LIMIT) into fractional form.
2599 if (tuple_fraction >= 1.0)
2600 tuple_fraction /= dNumDistinctRows;
2602 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2603 tuple_fraction) < 0)
2605 /* Hashed is cheaper, so use it */
2612 * make_subplanTargetList
2613 * Generate appropriate target list when grouping is required.
2615 * When grouping_planner inserts grouping or aggregation plan nodes
2616 * above the scan/join plan constructed by query_planner+create_plan,
2617 * we typically want the scan/join plan to emit a different target list
2618 * than the outer plan nodes should have. This routine generates the
2619 * correct target list for the scan/join subplan.
2621 * The initial target list passed from the parser already contains entries
2622 * for all ORDER BY and GROUP BY expressions, but it will not have entries
2623 * for variables used only in HAVING clauses; so we need to add those
2624 * variables to the subplan target list. Also, we flatten all expressions
2625 * except GROUP BY items into their component variables; the other expressions
2626 * will be computed by the inserted nodes rather than by the subplan.
2627 * For example, given a query like
2628 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
2629 * we want to pass this targetlist to the subplan:
2631 * where the a+b target will be used by the Sort/Group steps, and the
2632 * other targets will be used for computing the final results.
2634 * If we are grouping or aggregating, *and* there are no non-Var grouping
2635 * expressions, then the returned tlist is effectively dummy; we do not
2636 * need to force it to be evaluated, because all the Vars it contains
2637 * should be present in the "flat" tlist generated by create_plan, though
2638 * possibly in a different order. In that case we'll use create_plan's tlist,
2639 * and the tlist made here is only needed as input to query_planner to tell
2640 * it which Vars are needed in the output of the scan/join plan.
2642 * 'tlist' is the query's target list.
2643 * 'groupColIdx' receives an array of column numbers for the GROUP BY
2644 * expressions (if there are any) in the returned target list.
2645 * 'need_tlist_eval' is set true if we really need to evaluate the
2646 * returned tlist as-is.
2648 * The result is the targetlist to be passed to query_planner.
2651 make_subplanTargetList(PlannerInfo *root,
2653 AttrNumber **groupColIdx,
2654 bool *need_tlist_eval)
2656 Query *parse = root->parse;
2658 List *non_group_cols;
2659 List *non_group_vars;
2662 *groupColIdx = NULL;
2665 * If we're not grouping or aggregating, there's nothing to do here;
2666 * query_planner should receive the unmodified target list.
2668 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual &&
2669 !parse->hasWindowFuncs)
2671 *need_tlist_eval = true;
2676 * Otherwise, we must build a tlist containing all grouping columns,
2677 * plus any other Vars mentioned in the targetlist and HAVING qual.
2680 non_group_cols = NIL;
2681 *need_tlist_eval = false; /* only eval if not flat tlist */
2683 numCols = list_length(parse->groupClause);
2687 * If grouping, create sub_tlist entries for all GROUP BY columns, and
2688 * make an array showing where the group columns are in the sub_tlist.
2690 * Note: with this implementation, the array entries will always be
2691 * 1..N, but we don't want callers to assume that.
2693 AttrNumber *grpColIdx;
2696 grpColIdx = (AttrNumber *) palloc0(sizeof(AttrNumber) * numCols);
2697 *groupColIdx = grpColIdx;
2701 TargetEntry *tle = (TargetEntry *) lfirst(tl);
2704 colno = get_grouping_column_index(parse, tle);
2708 * It's a grouping column, so add it to the result tlist and
2709 * remember its resno in grpColIdx[].
2711 TargetEntry *newtle;
2713 newtle = makeTargetEntry(tle->expr,
2714 list_length(sub_tlist) + 1,
2717 sub_tlist = lappend(sub_tlist, newtle);
2719 Assert(grpColIdx[colno] == 0); /* no dups expected */
2720 grpColIdx[colno] = newtle->resno;
2722 if (!(newtle->expr && IsA(newtle->expr, Var)))
2723 *need_tlist_eval = true; /* tlist contains non Vars */
2728 * Non-grouping column, so just remember the expression
2729 * for later call to pull_var_clause. There's no need for
2730 * pull_var_clause to examine the TargetEntry node itself.
2732 non_group_cols = lappend(non_group_cols, tle->expr);
2739 * With no grouping columns, just pass whole tlist to pull_var_clause.
2740 * Need (shallow) copy to avoid damaging input tlist below.
2742 non_group_cols = list_copy(tlist);
2746 * If there's a HAVING clause, we'll need the Vars it uses, too.
2748 if (parse->havingQual)
2749 non_group_cols = lappend(non_group_cols, parse->havingQual);
2752 * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
2753 * add them to the result tlist if not already present. (A Var used
2754 * directly as a GROUP BY item will be present already.) Note this
2755 * includes Vars used in resjunk items, so we are covering the needs of
2756 * ORDER BY and window specifications. Vars used within Aggrefs will be
2757 * pulled out here, too.
2759 non_group_vars = pull_var_clause((Node *) non_group_cols,
2760 PVC_RECURSE_AGGREGATES,
2761 PVC_INCLUDE_PLACEHOLDERS);
2762 sub_tlist = add_to_flat_tlist(sub_tlist, non_group_vars);
2764 /* clean up cruft */
2765 list_free(non_group_vars);
2766 list_free(non_group_cols);
2772 * get_grouping_column_index
2773 * Get the GROUP BY column position, if any, of a targetlist entry.
2775 * Returns the index (counting from 0) of the TLE in the GROUP BY list, or -1
2776 * if it's not a grouping column. Note: the result is unique because the
2777 * parser won't make multiple groupClause entries for the same TLE.
2780 get_grouping_column_index(Query *parse, TargetEntry *tle)
2783 Index ressortgroupref = tle->ressortgroupref;
2786 /* No need to search groupClause if TLE hasn't got a sortgroupref */
2787 if (ressortgroupref == 0)
2790 foreach(gl, parse->groupClause)
2792 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2794 if (grpcl->tleSortGroupRef == ressortgroupref)
2803 * locate_grouping_columns
2804 * Locate grouping columns in the tlist chosen by create_plan.
2806 * This is only needed if we don't use the sub_tlist chosen by
2807 * make_subplanTargetList. We have to forget the column indexes found
2808 * by that routine and re-locate the grouping exprs in the real sub_tlist.
2811 locate_grouping_columns(PlannerInfo *root,
2814 AttrNumber *groupColIdx)
2820 * No work unless grouping.
2822 if (!root->parse->groupClause)
2824 Assert(groupColIdx == NULL);
2827 Assert(groupColIdx != NULL);
2829 foreach(gl, root->parse->groupClause)
2831 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2832 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2833 TargetEntry *te = tlist_member(groupexpr, sub_tlist);
2836 elog(ERROR, "failed to locate grouping columns");
2837 groupColIdx[keyno++] = te->resno;
2842 * postprocess_setop_tlist
2843 * Fix up targetlist returned by plan_set_operations().
2845 * We need to transpose sort key info from the orig_tlist into new_tlist.
2846 * NOTE: this would not be good enough if we supported resjunk sort keys
2847 * for results of set operations --- then, we'd need to project a whole
2848 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2849 * find any resjunk columns in orig_tlist.
2852 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2855 ListCell *orig_tlist_item = list_head(orig_tlist);
2857 foreach(l, new_tlist)
2859 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2860 TargetEntry *orig_tle;
2862 /* ignore resjunk columns in setop result */
2863 if (new_tle->resjunk)
2866 Assert(orig_tlist_item != NULL);
2867 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2868 orig_tlist_item = lnext(orig_tlist_item);
2869 if (orig_tle->resjunk) /* should not happen */
2870 elog(ERROR, "resjunk output columns are not implemented");
2871 Assert(new_tle->resno == orig_tle->resno);
2872 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2874 if (orig_tlist_item != NULL)
2875 elog(ERROR, "resjunk output columns are not implemented");
2880 * select_active_windows
2881 * Create a list of the "active" window clauses (ie, those referenced
2882 * by non-deleted WindowFuncs) in the order they are to be executed.
2885 select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
2891 /* First, make a list of the active windows */
2893 foreach(lc, root->parse->windowClause)
2895 WindowClause *wc = (WindowClause *) lfirst(lc);
2897 /* It's only active if wflists shows some related WindowFuncs */
2898 Assert(wc->winref <= wflists->maxWinRef);
2899 if (wflists->windowFuncs[wc->winref] != NIL)
2900 actives = lappend(actives, wc);
2904 * Now, ensure that windows with identical partitioning/ordering clauses
2905 * are adjacent in the list. This is required by the SQL standard, which
2906 * says that only one sort is to be used for such windows, even if they
2907 * are otherwise distinct (eg, different names or framing clauses).
2909 * There is room to be much smarter here, for example detecting whether
2910 * one window's sort keys are a prefix of another's (so that sorting for
2911 * the latter would do for the former), or putting windows first that
2912 * match a sort order available for the underlying query. For the moment
2913 * we are content with meeting the spec.
2916 while (actives != NIL)
2918 WindowClause *wc = (WindowClause *) linitial(actives);
2922 /* Move wc from actives to result */
2923 actives = list_delete_first(actives);
2924 result = lappend(result, wc);
2926 /* Now move any matching windows from actives to result */
2928 for (lc = list_head(actives); lc; lc = next)
2930 WindowClause *wc2 = (WindowClause *) lfirst(lc);
2933 /* framing options are NOT to be compared here! */
2934 if (equal(wc->partitionClause, wc2->partitionClause) &&
2935 equal(wc->orderClause, wc2->orderClause))
2937 actives = list_delete_cell(actives, lc, prev);
2938 result = lappend(result, wc2);
2949 * add_volatile_sort_exprs
2950 * Identify any volatile sort/group expressions used by the active
2951 * windows, and add them to window_tlist if not already present.
2952 * Return the modified window_tlist.
2955 add_volatile_sort_exprs(List *window_tlist, List *tlist, List *activeWindows)
2957 Bitmapset *sgrefs = NULL;
2960 /* First, collect the sortgrouprefs of the windows into a bitmapset */
2961 foreach(lc, activeWindows)
2963 WindowClause *wc = (WindowClause *) lfirst(lc);
2966 foreach(lc2, wc->partitionClause)
2968 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2970 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2972 foreach(lc2, wc->orderClause)
2974 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2976 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2981 * Now scan the original tlist to find the referenced expressions. Any
2982 * that are volatile must be added to window_tlist.
2984 * Note: we know that the input window_tlist contains no items marked with
2985 * ressortgrouprefs, so we don't have to worry about collisions of the
2986 * reference numbers.
2990 TargetEntry *tle = (TargetEntry *) lfirst(lc);
2992 if (tle->ressortgroupref != 0 &&
2993 bms_is_member(tle->ressortgroupref, sgrefs) &&
2994 contain_volatile_functions((Node *) tle->expr))
2996 TargetEntry *newtle;
2998 newtle = makeTargetEntry(tle->expr,
2999 list_length(window_tlist) + 1,
3002 newtle->ressortgroupref = tle->ressortgroupref;
3003 window_tlist = lappend(window_tlist, newtle);
3007 return window_tlist;
3011 * make_pathkeys_for_window
3012 * Create a pathkeys list describing the required input ordering
3013 * for the given WindowClause.
3015 * The required ordering is first the PARTITION keys, then the ORDER keys.
3016 * In the future we might try to implement windowing using hashing, in which
3017 * case the ordering could be relaxed, but for now we always sort.
3020 make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
3021 List *tlist, bool canonicalize)
3023 List *window_pathkeys;
3024 List *window_sortclauses;
3026 /* Throw error if can't sort */
3027 if (!grouping_is_sortable(wc->partitionClause))
3029 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
3030 errmsg("could not implement window PARTITION BY"),
3031 errdetail("Window partitioning columns must be of sortable datatypes.")));
3032 if (!grouping_is_sortable(wc->orderClause))
3034 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
3035 errmsg("could not implement window ORDER BY"),
3036 errdetail("Window ordering columns must be of sortable datatypes.")));
3038 /* Okay, make the combined pathkeys */
3039 window_sortclauses = list_concat(list_copy(wc->partitionClause),
3040 list_copy(wc->orderClause));
3041 window_pathkeys = make_pathkeys_for_sortclauses(root,
3045 list_free(window_sortclauses);
3046 return window_pathkeys;
3050 * get_column_info_for_window
3051 * Get the partitioning/ordering column numbers and equality operators
3052 * for a WindowAgg node.
3054 * This depends on the behavior of make_pathkeys_for_window()!
3056 * We are given the target WindowClause and an array of the input column
3057 * numbers associated with the resulting pathkeys. In the easy case, there
3058 * are the same number of pathkey columns as partitioning + ordering columns
3059 * and we just have to copy some data around. However, it's possible that
3060 * some of the original partitioning + ordering columns were eliminated as
3061 * redundant during the transformation to pathkeys. (This can happen even
3062 * though the parser gets rid of obvious duplicates. A typical scenario is a
3063 * window specification "PARTITION BY x ORDER BY y" coupled with a clause
3064 * "WHERE x = y" that causes the two sort columns to be recognized as
3065 * redundant.) In that unusual case, we have to work a lot harder to
3066 * determine which keys are significant.
3068 * The method used here is a bit brute-force: add the sort columns to a list
3069 * one at a time and note when the resulting pathkey list gets longer. But
3070 * it's a sufficiently uncommon case that a faster way doesn't seem worth
3071 * the amount of code refactoring that'd be needed.
3075 get_column_info_for_window(PlannerInfo *root, WindowClause *wc, List *tlist,
3076 int numSortCols, AttrNumber *sortColIdx,
3078 AttrNumber **partColIdx,
3079 Oid **partOperators,
3081 AttrNumber **ordColIdx,
3084 int numPart = list_length(wc->partitionClause);
3085 int numOrder = list_length(wc->orderClause);
3087 if (numSortCols == numPart + numOrder)
3090 *partNumCols = numPart;
3091 *partColIdx = sortColIdx;
3092 *partOperators = extract_grouping_ops(wc->partitionClause);
3093 *ordNumCols = numOrder;
3094 *ordColIdx = sortColIdx + numPart;
3095 *ordOperators = extract_grouping_ops(wc->orderClause);
3104 /* first, allocate what's certainly enough space for the arrays */
3106 *partColIdx = (AttrNumber *) palloc(numPart * sizeof(AttrNumber));
3107 *partOperators = (Oid *) palloc(numPart * sizeof(Oid));
3109 *ordColIdx = (AttrNumber *) palloc(numOrder * sizeof(AttrNumber));
3110 *ordOperators = (Oid *) palloc(numOrder * sizeof(Oid));
3114 foreach(lc, wc->partitionClause)
3116 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
3119 sortclauses = lappend(sortclauses, sgc);
3120 new_pathkeys = make_pathkeys_for_sortclauses(root,
3124 if (list_length(new_pathkeys) > list_length(pathkeys))
3126 /* this sort clause is actually significant */
3127 (*partColIdx)[*partNumCols] = sortColIdx[scidx++];
3128 (*partOperators)[*partNumCols] = sgc->eqop;
3130 pathkeys = new_pathkeys;
3133 foreach(lc, wc->orderClause)
3135 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
3138 sortclauses = lappend(sortclauses, sgc);
3139 new_pathkeys = make_pathkeys_for_sortclauses(root,
3143 if (list_length(new_pathkeys) > list_length(pathkeys))
3145 /* this sort clause is actually significant */
3146 (*ordColIdx)[*ordNumCols] = sortColIdx[scidx++];
3147 (*ordOperators)[*ordNumCols] = sgc->eqop;
3149 pathkeys = new_pathkeys;
3152 /* complain if we didn't eat exactly the right number of sort cols */
3153 if (scidx != numSortCols)
3154 elog(ERROR, "failed to deconstruct sort operators into partitioning/ordering operators");
3160 * expression_planner
3161 * Perform planner's transformations on a standalone expression.
3163 * Various utility commands need to evaluate expressions that are not part
3164 * of a plannable query. They can do so using the executor's regular
3165 * expression-execution machinery, but first the expression has to be fed
3166 * through here to transform it from parser output to something executable.
3168 * Currently, we disallow sublinks in standalone expressions, so there's no
3169 * real "planning" involved here. (That might not always be true though.)
3170 * What we must do is run eval_const_expressions to ensure that any function
3171 * calls are converted to positional notation and function default arguments
3172 * get inserted. The fact that constant subexpressions get simplified is a
3173 * side-effect that is useful when the expression will get evaluated more than
3174 * once. Also, we must fix operator function IDs.
3176 * Note: this must not make any damaging changes to the passed-in expression
3177 * tree. (It would actually be okay to apply fix_opfuncids to it, but since
3178 * we first do an expression_tree_mutator-based walk, what is returned will
3179 * be a new node tree.)
3182 expression_planner(Expr *expr)
3187 * Convert named-argument function calls, insert default arguments and
3188 * simplify constant subexprs
3190 result = eval_const_expressions(NULL, (Node *) expr);
3192 /* Fill in opfuncid values if missing */
3193 fix_opfuncids(result);
3195 return (Expr *) result;
3200 * plan_cluster_use_sort
3201 * Use the planner to decide how CLUSTER should implement sorting
3203 * tableOid is the OID of a table to be clustered on its index indexOid
3204 * (which is already known to be a btree index). Decide whether it's
3205 * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
3206 * Return TRUE to use sorting, FALSE to use an indexscan.
3208 * Note: caller had better already hold some type of lock on the table.
3211 plan_cluster_use_sort(Oid tableOid, Oid indexOid)
3215 PlannerGlobal *glob;
3218 IndexOptInfo *indexInfo;
3219 QualCost indexExprCost;
3220 Cost comparisonCost;
3222 Path seqScanAndSortPath;
3223 IndexPath *indexScanPath;
3226 /* Set up mostly-dummy planner state */
3227 query = makeNode(Query);
3228 query->commandType = CMD_SELECT;
3230 glob = makeNode(PlannerGlobal);
3232 root = makeNode(PlannerInfo);
3233 root->parse = query;
3235 root->query_level = 1;
3236 root->planner_cxt = CurrentMemoryContext;
3237 root->wt_param_id = -1;
3239 /* Build a minimal RTE for the rel */
3240 rte = makeNode(RangeTblEntry);
3241 rte->rtekind = RTE_RELATION;
3242 rte->relid = tableOid;
3243 rte->relkind = RELKIND_RELATION;
3245 rte->inFromCl = true;
3246 query->rtable = list_make1(rte);
3248 /* Set up RTE/RelOptInfo arrays */
3249 setup_simple_rel_arrays(root);
3251 /* Build RelOptInfo */
3252 rel = build_simple_rel(root, 1, RELOPT_BASEREL);
3254 /* Locate IndexOptInfo for the target index */
3256 foreach(lc, rel->indexlist)
3258 indexInfo = (IndexOptInfo *) lfirst(lc);
3259 if (indexInfo->indexoid == indexOid)
3264 * It's possible that get_relation_info did not generate an IndexOptInfo
3265 * for the desired index; this could happen if it's not yet reached its
3266 * indcheckxmin usability horizon, or if it's a system index and we're
3267 * ignoring system indexes. In such cases we should tell CLUSTER to not
3268 * trust the index contents but use seqscan-and-sort.
3270 if (lc == NULL) /* not in the list? */
3271 return true; /* use sort */
3274 * Rather than doing all the pushups that would be needed to use
3275 * set_baserel_size_estimates, just do a quick hack for rows and width.
3277 rel->rows = rel->tuples;
3278 rel->width = get_relation_data_width(tableOid, NULL);
3280 root->total_table_pages = rel->pages;
3283 * Determine eval cost of the index expressions, if any. We need to
3284 * charge twice that amount for each tuple comparison that happens during
3285 * the sort, since tuplesort.c will have to re-evaluate the index
3286 * expressions each time. (XXX that's pretty inefficient...)
3288 cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
3289 comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
3291 /* Estimate the cost of seq scan + sort */
3292 seqScanPath = create_seqscan_path(root, rel);
3293 cost_sort(&seqScanAndSortPath, root, NIL,
3294 seqScanPath->total_cost, rel->tuples, rel->width,
3295 comparisonCost, maintenance_work_mem, -1.0);
3297 /* Estimate the cost of index scan */
3298 indexScanPath = create_index_path(root, indexInfo,
3299 NIL, NIL, NIL, NIL, NIL,
3300 ForwardScanDirection, false, NULL);
3302 return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
projects / postgresql / blob