1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.246 2008/10/22 20:17:51 tgl Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "executor/executor.h"
22 #include "executor/nodeAgg.h"
23 #include "miscadmin.h"
24 #include "nodes/makefuncs.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/cost.h"
27 #include "optimizer/pathnode.h"
28 #include "optimizer/paths.h"
29 #include "optimizer/planmain.h"
30 #include "optimizer/planner.h"
31 #include "optimizer/prep.h"
32 #include "optimizer/subselect.h"
33 #include "optimizer/tlist.h"
34 #include "optimizer/var.h"
35 #ifdef OPTIMIZER_DEBUG
36 #include "nodes/print.h"
38 #include "parser/parse_expr.h"
39 #include "parser/parse_oper.h"
40 #include "parser/parsetree.h"
41 #include "utils/lsyscache.h"
42 #include "utils/syscache.h"
46 double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
48 /* Hook for plugins to get control in planner() */
49 planner_hook_type planner_hook = NULL;
52 /* Expression kind codes for preprocess_expression */
53 #define EXPRKIND_QUAL 0
54 #define EXPRKIND_TARGET 1
55 #define EXPRKIND_RTFUNC 2
56 #define EXPRKIND_VALUES 3
57 #define EXPRKIND_LIMIT 4
58 #define EXPRKIND_APPINFO 5
61 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
62 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
63 static Plan *inheritance_planner(PlannerInfo *root);
64 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
65 static bool is_dummy_plan(Plan *plan);
66 static double preprocess_limit(PlannerInfo *root,
67 double tuple_fraction,
68 int64 *offset_est, int64 *count_est);
69 static void preprocess_groupclause(PlannerInfo *root);
70 static bool choose_hashed_grouping(PlannerInfo *root,
71 double tuple_fraction, double limit_tuples,
72 Path *cheapest_path, Path *sorted_path,
73 double dNumGroups, AggClauseCounts *agg_counts);
74 static bool choose_hashed_distinct(PlannerInfo *root,
75 Plan *input_plan, List *input_pathkeys,
76 double tuple_fraction, double limit_tuples,
77 double dNumDistinctRows);
78 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
79 AttrNumber **groupColIdx, bool *need_tlist_eval);
80 static void locate_grouping_columns(PlannerInfo *root,
83 AttrNumber *groupColIdx);
84 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
87 /*****************************************************************************
89 * Query optimizer entry point
91 * To support loadable plugins that monitor or modify planner behavior,
92 * we provide a hook variable that lets a plugin get control before and
93 * after the standard planning process. The plugin would normally call
96 * Note to plugin authors: standard_planner() scribbles on its Query input,
97 * so you'd better copy that data structure if you want to plan more than once.
99 *****************************************************************************/
101 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
106 result = (*planner_hook) (parse, cursorOptions, boundParams);
108 result = standard_planner(parse, cursorOptions, boundParams);
113 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
117 double tuple_fraction;
123 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
124 if (parse->utilityStmt &&
125 IsA(parse->utilityStmt, DeclareCursorStmt))
126 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
129 * Set up global state for this planner invocation. This data is needed
130 * across all levels of sub-Query that might exist in the given command,
131 * so we keep it in a separate struct that's linked to by each per-Query
134 glob = makeNode(PlannerGlobal);
136 glob->boundParams = boundParams;
137 glob->paramlist = NIL;
138 glob->subplans = NIL;
139 glob->subrtables = NIL;
140 glob->rewindPlanIDs = NULL;
141 glob->finalrtable = NIL;
142 glob->relationOids = NIL;
143 glob->invalItems = NIL;
145 glob->transientPlan = false;
147 /* Determine what fraction of the plan is likely to be scanned */
148 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
151 * We have no real idea how many tuples the user will ultimately FETCH
152 * from a cursor, but it is often the case that he doesn't want 'em
153 * all, or would prefer a fast-start plan anyway so that he can
154 * process some of the tuples sooner. Use a GUC parameter to decide
155 * what fraction to optimize for.
157 tuple_fraction = cursor_tuple_fraction;
160 * We document cursor_tuple_fraction as simply being a fraction,
161 * which means the edge cases 0 and 1 have to be treated specially
162 * here. We convert 1 to 0 ("all the tuples") and 0 to a very small
165 if (tuple_fraction >= 1.0)
166 tuple_fraction = 0.0;
167 else if (tuple_fraction <= 0.0)
168 tuple_fraction = 1e-10;
172 /* Default assumption is we need all the tuples */
173 tuple_fraction = 0.0;
176 /* primary planning entry point (may recurse for subqueries) */
177 top_plan = subquery_planner(glob, parse, NULL,
178 false, tuple_fraction, &root);
181 * If creating a plan for a scrollable cursor, make sure it can run
182 * backwards on demand. Add a Material node at the top at need.
184 if (cursorOptions & CURSOR_OPT_SCROLL)
186 if (!ExecSupportsBackwardScan(top_plan))
187 top_plan = materialize_finished_plan(top_plan);
190 /* final cleanup of the plan */
191 Assert(glob->finalrtable == NIL);
192 top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
193 /* ... and the subplans (both regular subplans and initplans) */
194 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
195 forboth(lp, glob->subplans, lr, glob->subrtables)
197 Plan *subplan = (Plan *) lfirst(lp);
198 List *subrtable = (List *) lfirst(lr);
200 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
203 /* build the PlannedStmt result */
204 result = makeNode(PlannedStmt);
206 result->commandType = parse->commandType;
207 result->canSetTag = parse->canSetTag;
208 result->transientPlan = glob->transientPlan;
209 result->planTree = top_plan;
210 result->rtable = glob->finalrtable;
211 result->resultRelations = root->resultRelations;
212 result->utilityStmt = parse->utilityStmt;
213 result->intoClause = parse->intoClause;
214 result->subplans = glob->subplans;
215 result->rewindPlanIDs = glob->rewindPlanIDs;
216 result->returningLists = root->returningLists;
217 result->rowMarks = parse->rowMarks;
218 result->relationOids = glob->relationOids;
219 result->invalItems = glob->invalItems;
220 result->nParamExec = list_length(glob->paramlist);
226 /*--------------------
228 * Invokes the planner on a subquery. We recurse to here for each
229 * sub-SELECT found in the query tree.
231 * glob is the global state for the current planner run.
232 * parse is the querytree produced by the parser & rewriter.
233 * parent_root is the immediate parent Query's info (NULL at the top level).
234 * hasRecursion is true if this is a recursive WITH query.
235 * tuple_fraction is the fraction of tuples we expect will be retrieved.
236 * tuple_fraction is interpreted as explained for grouping_planner, below.
238 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
239 * among other things this tells the output sort ordering of the plan.
241 * Basically, this routine does the stuff that should only be done once
242 * per Query object. It then calls grouping_planner. At one time,
243 * grouping_planner could be invoked recursively on the same Query object;
244 * that's not currently true, but we keep the separation between the two
245 * routines anyway, in case we need it again someday.
247 * subquery_planner will be called recursively to handle sub-Query nodes
248 * found within the query's expressions and rangetable.
250 * Returns a query plan.
251 *--------------------
254 subquery_planner(PlannerGlobal *glob, Query *parse,
255 PlannerInfo *parent_root,
256 bool hasRecursion, double tuple_fraction,
257 PlannerInfo **subroot)
259 int num_old_subplans = list_length(glob->subplans);
266 /* Create a PlannerInfo data structure for this subquery */
267 root = makeNode(PlannerInfo);
270 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
271 root->parent_root = parent_root;
272 root->planner_cxt = CurrentMemoryContext;
273 root->init_plans = NIL;
274 root->cte_plan_ids = NIL;
275 root->eq_classes = NIL;
276 root->append_rel_list = NIL;
278 root->hasRecursion = hasRecursion;
280 root->wt_param_id = SS_assign_worktable_param(root);
282 root->wt_param_id = -1;
283 root->non_recursive_plan = NULL;
286 * If there is a WITH list, process each WITH query and build an
287 * initplan SubPlan structure for it.
290 SS_process_ctes(root);
293 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
294 * to transform them into joins. Note that this step does not descend
295 * into subqueries; if we pull up any subqueries below, their SubLinks are
296 * processed just before pulling them up.
298 if (parse->hasSubLinks)
299 pull_up_sublinks(root);
302 * Scan the rangetable for set-returning functions, and inline them
303 * if possible (producing subqueries that might get pulled up next).
304 * Recursion issues here are handled in the same way as for SubLinks.
306 inline_set_returning_functions(root);
309 * Check to see if any subqueries in the rangetable can be merged into
312 parse->jointree = (FromExpr *)
313 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
316 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
317 * avoid the expense of doing flatten_join_alias_vars(). Also check for
318 * outer joins --- if none, we can skip reduce_outer_joins().
319 * This must be done after we have done pull_up_subqueries, of course.
321 root->hasJoinRTEs = false;
322 hasOuterJoins = false;
323 foreach(l, parse->rtable)
325 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
327 if (rte->rtekind == RTE_JOIN)
329 root->hasJoinRTEs = true;
330 if (IS_OUTER_JOIN(rte->jointype))
332 hasOuterJoins = true;
333 /* Can quit scanning once we find an outer join */
340 * Expand any rangetable entries that are inheritance sets into "append
341 * relations". This can add entries to the rangetable, but they must be
342 * plain base relations not joins, so it's OK (and marginally more
343 * efficient) to do it after checking for join RTEs. We must do it after
344 * pulling up subqueries, else we'd fail to handle inherited tables in
347 expand_inherited_tables(root);
350 * Set hasHavingQual to remember if HAVING clause is present. Needed
351 * because preprocess_expression will reduce a constant-true condition to
352 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
354 root->hasHavingQual = (parse->havingQual != NULL);
356 /* Clear this flag; might get set in distribute_qual_to_rels */
357 root->hasPseudoConstantQuals = false;
360 * Do expression preprocessing on targetlist and quals.
362 parse->targetList = (List *)
363 preprocess_expression(root, (Node *) parse->targetList,
366 parse->returningList = (List *)
367 preprocess_expression(root, (Node *) parse->returningList,
370 preprocess_qual_conditions(root, (Node *) parse->jointree);
372 parse->havingQual = preprocess_expression(root, parse->havingQual,
375 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
377 parse->limitCount = preprocess_expression(root, parse->limitCount,
380 root->append_rel_list = (List *)
381 preprocess_expression(root, (Node *) root->append_rel_list,
384 /* Also need to preprocess expressions for function and values RTEs */
385 foreach(l, parse->rtable)
387 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
389 if (rte->rtekind == RTE_FUNCTION)
390 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
392 else if (rte->rtekind == RTE_VALUES)
393 rte->values_lists = (List *)
394 preprocess_expression(root, (Node *) rte->values_lists,
399 * In some cases we may want to transfer a HAVING clause into WHERE. We
400 * cannot do so if the HAVING clause contains aggregates (obviously) or
401 * volatile functions (since a HAVING clause is supposed to be executed
402 * only once per group). Also, it may be that the clause is so expensive
403 * to execute that we're better off doing it only once per group, despite
404 * the loss of selectivity. This is hard to estimate short of doing the
405 * entire planning process twice, so we use a heuristic: clauses
406 * containing subplans are left in HAVING. Otherwise, we move or copy the
407 * HAVING clause into WHERE, in hopes of eliminating tuples before
408 * aggregation instead of after.
410 * If the query has explicit grouping then we can simply move such a
411 * clause into WHERE; any group that fails the clause will not be in the
412 * output because none of its tuples will reach the grouping or
413 * aggregation stage. Otherwise we must have a degenerate (variable-free)
414 * HAVING clause, which we put in WHERE so that query_planner() can use it
415 * in a gating Result node, but also keep in HAVING to ensure that we
416 * don't emit a bogus aggregated row. (This could be done better, but it
417 * seems not worth optimizing.)
419 * Note that both havingQual and parse->jointree->quals are in
420 * implicitly-ANDed-list form at this point, even though they are declared
424 foreach(l, (List *) parse->havingQual)
426 Node *havingclause = (Node *) lfirst(l);
428 if (contain_agg_clause(havingclause) ||
429 contain_volatile_functions(havingclause) ||
430 contain_subplans(havingclause))
432 /* keep it in HAVING */
433 newHaving = lappend(newHaving, havingclause);
435 else if (parse->groupClause)
437 /* move it to WHERE */
438 parse->jointree->quals = (Node *)
439 lappend((List *) parse->jointree->quals, havingclause);
443 /* put a copy in WHERE, keep it in HAVING */
444 parse->jointree->quals = (Node *)
445 lappend((List *) parse->jointree->quals,
446 copyObject(havingclause));
447 newHaving = lappend(newHaving, havingclause);
450 parse->havingQual = (Node *) newHaving;
453 * If we have any outer joins, try to reduce them to plain inner joins.
454 * This step is most easily done after we've done expression
458 reduce_outer_joins(root);
461 * Do the main planning. If we have an inherited target relation, that
462 * needs special processing, else go straight to grouping_planner.
464 if (parse->resultRelation &&
465 rt_fetch(parse->resultRelation, parse->rtable)->inh)
466 plan = inheritance_planner(root);
468 plan = grouping_planner(root, tuple_fraction);
471 * If any subplans were generated, or if we're inside a subplan, build
472 * initPlan list and extParam/allParam sets for plan nodes, and attach the
473 * initPlans to the top plan node.
475 if (list_length(glob->subplans) != num_old_subplans ||
476 root->query_level > 1)
477 SS_finalize_plan(root, plan, true);
479 /* Return internal info if caller wants it */
487 * preprocess_expression
488 * Do subquery_planner's preprocessing work for an expression,
489 * which can be a targetlist, a WHERE clause (including JOIN/ON
490 * conditions), or a HAVING clause.
493 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
496 * Fall out quickly if expression is empty. This occurs often enough to
497 * be worth checking. Note that null->null is the correct conversion for
498 * implicit-AND result format, too.
504 * If the query has any join RTEs, replace join alias variables with
505 * base-relation variables. We must do this before sublink processing,
506 * else sublinks expanded out from join aliases wouldn't get processed. We
507 * can skip it in VALUES lists, however, since they can't contain any Vars
510 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
511 expr = flatten_join_alias_vars(root, expr);
514 * Simplify constant expressions.
516 * Note: this also flattens nested AND and OR expressions into N-argument
517 * form. All processing of a qual expression after this point must be
518 * careful to maintain AND/OR flatness --- that is, do not generate a tree
519 * with AND directly under AND, nor OR directly under OR.
521 * Because this is a relatively expensive process, we skip it when the
522 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
523 * expression will only be evaluated once anyway, so no point in
524 * pre-simplifying; we can't execute it any faster than the executor can,
525 * and we will waste cycles copying the tree. Notice however that we
526 * still must do it for quals (to get AND/OR flatness); and if we are in a
527 * subquery we should not assume it will be done only once.
529 * For VALUES lists we never do this at all, again on the grounds that we
530 * should optimize for one-time evaluation.
532 if (kind != EXPRKIND_VALUES &&
533 (root->parse->jointree->fromlist != NIL ||
534 kind == EXPRKIND_QUAL ||
535 root->query_level > 1))
536 expr = eval_const_expressions(root, expr);
539 * If it's a qual or havingQual, canonicalize it.
541 if (kind == EXPRKIND_QUAL)
543 expr = (Node *) canonicalize_qual((Expr *) expr);
545 #ifdef OPTIMIZER_DEBUG
546 printf("After canonicalize_qual()\n");
551 /* Expand SubLinks to SubPlans */
552 if (root->parse->hasSubLinks)
553 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
556 * XXX do not insert anything here unless you have grokked the comments in
557 * SS_replace_correlation_vars ...
560 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
561 if (root->query_level > 1)
562 expr = SS_replace_correlation_vars(root, expr);
565 * If it's a qual or havingQual, convert it to implicit-AND format. (We
566 * don't want to do this before eval_const_expressions, since the latter
567 * would be unable to simplify a top-level AND correctly. Also,
568 * SS_process_sublinks expects explicit-AND format.)
570 if (kind == EXPRKIND_QUAL)
571 expr = (Node *) make_ands_implicit((Expr *) expr);
577 * preprocess_qual_conditions
578 * Recursively scan the query's jointree and do subquery_planner's
579 * preprocessing work on each qual condition found therein.
582 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
586 if (IsA(jtnode, RangeTblRef))
588 /* nothing to do here */
590 else if (IsA(jtnode, FromExpr))
592 FromExpr *f = (FromExpr *) jtnode;
595 foreach(l, f->fromlist)
596 preprocess_qual_conditions(root, lfirst(l));
598 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
600 else if (IsA(jtnode, JoinExpr))
602 JoinExpr *j = (JoinExpr *) jtnode;
604 preprocess_qual_conditions(root, j->larg);
605 preprocess_qual_conditions(root, j->rarg);
607 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
610 elog(ERROR, "unrecognized node type: %d",
611 (int) nodeTag(jtnode));
615 * inheritance_planner
616 * Generate a plan in the case where the result relation is an
619 * We have to handle this case differently from cases where a source relation
620 * is an inheritance set. Source inheritance is expanded at the bottom of the
621 * plan tree (see allpaths.c), but target inheritance has to be expanded at
622 * the top. The reason is that for UPDATE, each target relation needs a
623 * different targetlist matching its own column set. Also, for both UPDATE
624 * and DELETE, the executor needs the Append plan node at the top, else it
625 * can't keep track of which table is the current target table. Fortunately,
626 * the UPDATE/DELETE target can never be the nullable side of an outer join,
627 * so it's OK to generate the plan this way.
629 * Returns a query plan.
632 inheritance_planner(PlannerInfo *root)
634 Query *parse = root->parse;
635 int parentRTindex = parse->resultRelation;
636 List *subplans = NIL;
637 List *resultRelations = NIL;
638 List *returningLists = NIL;
644 foreach(l, root->append_rel_list)
646 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
649 /* append_rel_list contains all append rels; ignore others */
650 if (appinfo->parent_relid != parentRTindex)
654 * Generate modified query with this rel as target.
656 memcpy(&subroot, root, sizeof(PlannerInfo));
657 subroot.parse = (Query *)
658 adjust_appendrel_attrs((Node *) parse,
660 subroot.returningLists = NIL;
661 subroot.init_plans = NIL;
662 /* We needn't modify the child's append_rel_list */
663 /* There shouldn't be any OJ info to translate, as yet */
664 Assert(subroot.join_info_list == NIL);
665 /* and we haven't created PlaceHolderInfos, either */
666 Assert(subroot.placeholder_list == NIL);
669 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
672 * If this child rel was excluded by constraint exclusion, exclude it
675 if (is_dummy_plan(subplan))
678 /* Save rtable and tlist from first rel for use below */
681 rtable = subroot.parse->rtable;
682 tlist = subplan->targetlist;
685 subplans = lappend(subplans, subplan);
687 /* Make sure any initplans from this rel get into the outer list */
688 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
690 /* Build target-relations list for the executor */
691 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
693 /* Build list of per-relation RETURNING targetlists */
694 if (parse->returningList)
696 Assert(list_length(subroot.returningLists) == 1);
697 returningLists = list_concat(returningLists,
698 subroot.returningLists);
702 root->resultRelations = resultRelations;
703 root->returningLists = returningLists;
705 /* Mark result as unordered (probably unnecessary) */
706 root->query_pathkeys = NIL;
709 * If we managed to exclude every child rel, return a dummy plan
713 root->resultRelations = list_make1_int(parentRTindex);
714 /* although dummy, it must have a valid tlist for executor */
715 tlist = preprocess_targetlist(root, parse->targetList);
716 return (Plan *) make_result(root,
718 (Node *) list_make1(makeBoolConst(false,
724 * Planning might have modified the rangetable, due to changes of the
725 * Query structures inside subquery RTEs. We have to ensure that this
726 * gets propagated back to the master copy. But can't do this until we
727 * are done planning, because all the calls to grouping_planner need
728 * virgin sub-Queries to work from. (We are effectively assuming that
729 * sub-Queries will get planned identically each time, or at least that
730 * the impacts on their rangetables will be the same each time.)
732 * XXX should clean this up someday
734 parse->rtable = rtable;
736 /* Suppress Append if there's only one surviving child rel */
737 if (list_length(subplans) == 1)
738 return (Plan *) linitial(subplans);
740 return (Plan *) make_append(subplans, true, tlist);
743 /*--------------------
745 * Perform planning steps related to grouping, aggregation, etc.
746 * This primarily means adding top-level processing to the basic
747 * query plan produced by query_planner.
749 * tuple_fraction is the fraction of tuples we expect will be retrieved
751 * tuple_fraction is interpreted as follows:
752 * 0: expect all tuples to be retrieved (normal case)
753 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
754 * from the plan to be retrieved
755 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
756 * expected to be retrieved (ie, a LIMIT specification)
758 * Returns a query plan. Also, root->query_pathkeys is returned as the
759 * actual output ordering of the plan (in pathkey format).
760 *--------------------
763 grouping_planner(PlannerInfo *root, double tuple_fraction)
765 Query *parse = root->parse;
766 List *tlist = parse->targetList;
767 int64 offset_est = 0;
769 double limit_tuples = -1.0;
771 List *current_pathkeys;
772 double dNumGroups = 0;
774 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
775 if (parse->limitCount || parse->limitOffset)
777 tuple_fraction = preprocess_limit(root, tuple_fraction,
778 &offset_est, &count_est);
781 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
782 * estimate the effects of using a bounded sort.
784 if (count_est > 0 && offset_est >= 0)
785 limit_tuples = (double) count_est + (double) offset_est;
788 if (parse->setOperations)
790 List *set_sortclauses;
793 * If there's a top-level ORDER BY, assume we have to fetch all the
794 * tuples. This might be too simplistic given all the hackery below
795 * to possibly avoid the sort; but the odds of accurate estimates
796 * here are pretty low anyway.
798 if (parse->sortClause)
799 tuple_fraction = 0.0;
802 * Construct the plan for set operations. The result will not need
803 * any work except perhaps a top-level sort and/or LIMIT. Note that
804 * any special work for recursive unions is the responsibility of
805 * plan_set_operations.
807 result_plan = plan_set_operations(root, tuple_fraction,
811 * Calculate pathkeys representing the sort order (if any) of the set
812 * operation's result. We have to do this before overwriting the sort
815 current_pathkeys = make_pathkeys_for_sortclauses(root,
817 result_plan->targetlist,
821 * We should not need to call preprocess_targetlist, since we must be
822 * in a SELECT query node. Instead, use the targetlist returned by
823 * plan_set_operations (since this tells whether it returned any
824 * resjunk columns!), and transfer any sort key information from the
827 Assert(parse->commandType == CMD_SELECT);
829 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
833 * Can't handle FOR UPDATE/SHARE here (parser should have checked
834 * already, but let's make sure).
838 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
839 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
842 * Calculate pathkeys that represent result ordering requirements
844 Assert(parse->distinctClause == NIL);
845 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
852 /* No set operations, do regular planning */
854 AttrNumber *groupColIdx = NULL;
855 bool need_tlist_eval = true;
861 AggClauseCounts agg_counts;
863 bool use_hashed_grouping = false;
865 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
867 /* A recursive query should always have setOperations */
868 Assert(!root->hasRecursion);
870 /* Preprocess GROUP BY clause, if any */
871 if (parse->groupClause)
872 preprocess_groupclause(root);
873 numGroupCols = list_length(parse->groupClause);
875 /* Preprocess targetlist */
876 tlist = preprocess_targetlist(root, tlist);
879 * Generate appropriate target list for subplan; may be different from
880 * tlist if grouping or aggregation is needed.
882 sub_tlist = make_subplanTargetList(root, tlist,
883 &groupColIdx, &need_tlist_eval);
886 * Calculate pathkeys that represent grouping/ordering requirements.
887 * Stash them in PlannerInfo so that query_planner can canonicalize
888 * them after EquivalenceClasses have been formed. The sortClause
889 * is certainly sort-able, but GROUP BY and DISTINCT might not be,
890 * in which case we just leave their pathkeys empty.
892 if (parse->groupClause &&
893 grouping_is_sortable(parse->groupClause))
894 root->group_pathkeys =
895 make_pathkeys_for_sortclauses(root,
900 root->group_pathkeys = NIL;
902 if (parse->distinctClause &&
903 grouping_is_sortable(parse->distinctClause))
904 root->distinct_pathkeys =
905 make_pathkeys_for_sortclauses(root,
906 parse->distinctClause,
910 root->distinct_pathkeys = NIL;
912 root->sort_pathkeys =
913 make_pathkeys_for_sortclauses(root,
919 * Will need actual number of aggregates for estimating costs.
921 * Note: we do not attempt to detect duplicate aggregates here; a
922 * somewhat-overestimated count is okay for our present purposes.
924 * Note: think not that we can turn off hasAggs if we find no aggs. It
925 * is possible for constant-expression simplification to remove all
926 * explicit references to aggs, but we still have to follow the
927 * aggregate semantics (eg, producing only one output row).
931 count_agg_clauses((Node *) tlist, &agg_counts);
932 count_agg_clauses(parse->havingQual, &agg_counts);
936 * Figure out whether we want a sorted result from query_planner.
938 * If we have a sortable GROUP BY clause, then we want a result sorted
939 * properly for grouping. Otherwise, if there's a sortable DISTINCT
940 * clause that's more rigorous than the ORDER BY clause, we try to
941 * produce output that's sufficiently well sorted for the DISTINCT.
942 * Otherwise, if there is an ORDER BY clause, we want to sort by the
945 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
946 * superset of GROUP BY, it would be tempting to request sort by ORDER
947 * BY --- but that might just leave us failing to exploit an available
948 * sort order at all. Needs more thought. The choice for DISTINCT
949 * versus ORDER BY is much easier, since we know that the parser
950 * ensured that one is a superset of the other.
952 if (root->group_pathkeys)
953 root->query_pathkeys = root->group_pathkeys;
954 else if (list_length(root->distinct_pathkeys) >
955 list_length(root->sort_pathkeys))
956 root->query_pathkeys = root->distinct_pathkeys;
957 else if (root->sort_pathkeys)
958 root->query_pathkeys = root->sort_pathkeys;
960 root->query_pathkeys = NIL;
963 * Generate the best unsorted and presorted paths for this Query (but
964 * note there may not be any presorted path). query_planner will also
965 * estimate the number of groups in the query, and canonicalize all
968 query_planner(root, sub_tlist, tuple_fraction, limit_tuples,
969 &cheapest_path, &sorted_path, &dNumGroups);
972 * If grouping, decide whether to use sorted or hashed grouping.
974 if (parse->groupClause)
980 * Executor doesn't support hashed aggregation with DISTINCT
981 * aggregates. (Doing so would imply storing *all* the input
982 * values in the hash table, which seems like a certain loser.)
984 can_hash = (agg_counts.numDistinctAggs == 0 &&
985 grouping_is_hashable(parse->groupClause));
986 can_sort = grouping_is_sortable(parse->groupClause);
987 if (can_hash && can_sort)
989 /* we have a meaningful choice to make ... */
990 use_hashed_grouping =
991 choose_hashed_grouping(root,
992 tuple_fraction, limit_tuples,
993 cheapest_path, sorted_path,
994 dNumGroups, &agg_counts);
997 use_hashed_grouping = true;
999 use_hashed_grouping = false;
1002 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1003 errmsg("could not implement GROUP BY"),
1004 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
1006 /* Also convert # groups to long int --- but 'ware overflow! */
1007 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
1011 * Select the best path. If we are doing hashed grouping, we will
1012 * always read all the input tuples, so use the cheapest-total path.
1013 * Otherwise, trust query_planner's decision about which to use.
1015 if (use_hashed_grouping || !sorted_path)
1016 best_path = cheapest_path;
1018 best_path = sorted_path;
1021 * Check to see if it's possible to optimize MIN/MAX aggregates. If
1022 * so, we will forget all the work we did so far to choose a "regular"
1023 * path ... but we had to do it anyway to be able to tell which way is
1026 result_plan = optimize_minmax_aggregates(root,
1029 if (result_plan != NULL)
1032 * optimize_minmax_aggregates generated the full plan, with the
1033 * right tlist, and it has no sort order.
1035 current_pathkeys = NIL;
1040 * Normal case --- create a plan according to query_planner's
1043 bool need_sort_for_grouping = false;
1045 result_plan = create_plan(root, best_path);
1046 current_pathkeys = best_path->pathkeys;
1048 /* Detect if we'll need an explicit sort for grouping */
1049 if (parse->groupClause && !use_hashed_grouping &&
1050 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1052 need_sort_for_grouping = true;
1054 * Always override query_planner's tlist, so that we don't
1055 * sort useless data from a "physical" tlist.
1057 need_tlist_eval = true;
1061 * create_plan() returns a plan with just a "flat" tlist of
1062 * required Vars. Usually we need to insert the sub_tlist as the
1063 * tlist of the top plan node. However, we can skip that if we
1064 * determined that whatever query_planner chose to return will be
1067 if (need_tlist_eval)
1070 * If the top-level plan node is one that cannot do expression
1071 * evaluation, we must insert a Result node to project the
1074 if (!is_projection_capable_plan(result_plan))
1076 result_plan = (Plan *) make_result(root,
1084 * Otherwise, just replace the subplan's flat tlist with
1085 * the desired tlist.
1087 result_plan->targetlist = sub_tlist;
1091 * Also, account for the cost of evaluation of the sub_tlist.
1093 * Up to now, we have only been dealing with "flat" tlists,
1094 * containing just Vars. So their evaluation cost is zero
1095 * according to the model used by cost_qual_eval() (or if you
1096 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1097 * can avoid accounting for tlist cost throughout
1098 * query_planner() and subroutines. But now we've inserted a
1099 * tlist that might contain actual operators, sub-selects, etc
1100 * --- so we'd better account for its cost.
1102 * Below this point, any tlist eval cost for added-on nodes
1103 * should be accounted for as we create those nodes.
1104 * Presently, of the node types we can add on, only Agg and
1105 * Group project new tlists (the rest just copy their input
1106 * tuples) --- so make_agg() and make_group() are responsible
1107 * for computing the added cost.
1109 cost_qual_eval(&tlist_cost, sub_tlist, root);
1110 result_plan->startup_cost += tlist_cost.startup;
1111 result_plan->total_cost += tlist_cost.startup +
1112 tlist_cost.per_tuple * result_plan->plan_rows;
1117 * Since we're using query_planner's tlist and not the one
1118 * make_subplanTargetList calculated, we have to refigure any
1119 * grouping-column indexes make_subplanTargetList computed.
1121 locate_grouping_columns(root, tlist, result_plan->targetlist,
1126 * Insert AGG or GROUP node if needed, plus an explicit sort step
1129 * HAVING clause, if any, becomes qual of the Agg or Group node.
1131 if (use_hashed_grouping)
1133 /* Hashed aggregate plan --- no sort needed */
1134 result_plan = (Plan *) make_agg(root,
1136 (List *) parse->havingQual,
1140 extract_grouping_ops(parse->groupClause),
1144 /* Hashed aggregation produces randomly-ordered results */
1145 current_pathkeys = NIL;
1147 else if (parse->hasAggs)
1149 /* Plain aggregate plan --- sort if needed */
1150 AggStrategy aggstrategy;
1152 if (parse->groupClause)
1154 if (need_sort_for_grouping)
1156 result_plan = (Plan *)
1157 make_sort_from_groupcols(root,
1161 current_pathkeys = root->group_pathkeys;
1163 aggstrategy = AGG_SORTED;
1166 * The AGG node will not change the sort ordering of its
1167 * groups, so current_pathkeys describes the result too.
1172 aggstrategy = AGG_PLAIN;
1173 /* Result will be only one row anyway; no sort order */
1174 current_pathkeys = NIL;
1177 result_plan = (Plan *) make_agg(root,
1179 (List *) parse->havingQual,
1183 extract_grouping_ops(parse->groupClause),
1188 else if (parse->groupClause)
1191 * GROUP BY without aggregation, so insert a group node (plus
1192 * the appropriate sort node, if necessary).
1194 * Add an explicit sort if we couldn't make the path come out
1195 * the way the GROUP node needs it.
1197 if (need_sort_for_grouping)
1199 result_plan = (Plan *)
1200 make_sort_from_groupcols(root,
1204 current_pathkeys = root->group_pathkeys;
1207 result_plan = (Plan *) make_group(root,
1209 (List *) parse->havingQual,
1212 extract_grouping_ops(parse->groupClause),
1215 /* The Group node won't change sort ordering */
1217 else if (root->hasHavingQual)
1220 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1221 * This is a degenerate case in which we are supposed to emit
1222 * either 0 or 1 row depending on whether HAVING succeeds.
1223 * Furthermore, there cannot be any variables in either HAVING
1224 * or the targetlist, so we actually do not need the FROM
1225 * table at all! We can just throw away the plan-so-far and
1226 * generate a Result node. This is a sufficiently unusual
1227 * corner case that it's not worth contorting the structure of
1228 * this routine to avoid having to generate the plan in the
1231 result_plan = (Plan *) make_result(root,
1236 } /* end of non-minmax-aggregate case */
1237 } /* end of if (setOperations) */
1240 * If there is a DISTINCT clause, add the necessary node(s).
1242 if (parse->distinctClause)
1244 double dNumDistinctRows;
1245 long numDistinctRows;
1246 bool use_hashed_distinct;
1251 * If there was grouping or aggregation, use the current number of
1252 * rows as the estimated number of DISTINCT rows (ie, assume the
1253 * result was already mostly unique). If not, use the number of
1254 * distinct-groups calculated by query_planner.
1256 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1257 dNumDistinctRows = result_plan->plan_rows;
1259 dNumDistinctRows = dNumGroups;
1261 /* Also convert to long int --- but 'ware overflow! */
1262 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1265 * If we have a sortable DISTINCT ON clause, we always use sorting.
1266 * This enforces the expected behavior of DISTINCT ON.
1268 can_sort = grouping_is_sortable(parse->distinctClause);
1269 if (can_sort && parse->hasDistinctOn)
1270 use_hashed_distinct = false;
1273 can_hash = grouping_is_hashable(parse->distinctClause);
1274 if (can_hash && can_sort)
1276 /* we have a meaningful choice to make ... */
1277 use_hashed_distinct =
1278 choose_hashed_distinct(root,
1279 result_plan, current_pathkeys,
1280 tuple_fraction, limit_tuples,
1284 use_hashed_distinct = true;
1286 use_hashed_distinct = false;
1290 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1291 errmsg("could not implement DISTINCT"),
1292 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
1293 use_hashed_distinct = false; /* keep compiler quiet */
1297 if (use_hashed_distinct)
1299 /* Hashed aggregate plan --- no sort needed */
1300 result_plan = (Plan *) make_agg(root,
1301 result_plan->targetlist,
1304 list_length(parse->distinctClause),
1305 extract_grouping_cols(parse->distinctClause,
1306 result_plan->targetlist),
1307 extract_grouping_ops(parse->distinctClause),
1311 /* Hashed aggregation produces randomly-ordered results */
1312 current_pathkeys = NIL;
1317 * Use a Unique node to implement DISTINCT. Add an explicit sort
1318 * if we couldn't make the path come out the way the Unique node
1319 * needs it. If we do have to sort, always sort by the more
1320 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1321 * below. However, for regular DISTINCT, don't sort now if we
1322 * don't have to --- sorting afterwards will likely be cheaper,
1323 * and also has the possibility of optimizing via LIMIT. But
1324 * for DISTINCT ON, we *must* force the final sort now, else
1325 * it won't have the desired behavior.
1327 List *needed_pathkeys;
1329 if (parse->hasDistinctOn &&
1330 list_length(root->distinct_pathkeys) <
1331 list_length(root->sort_pathkeys))
1332 needed_pathkeys = root->sort_pathkeys;
1334 needed_pathkeys = root->distinct_pathkeys;
1336 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1338 if (list_length(root->distinct_pathkeys) >=
1339 list_length(root->sort_pathkeys))
1340 current_pathkeys = root->distinct_pathkeys;
1343 current_pathkeys = root->sort_pathkeys;
1344 /* Assert checks that parser didn't mess up... */
1345 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1349 result_plan = (Plan *) make_sort_from_pathkeys(root,
1355 result_plan = (Plan *) make_unique(result_plan,
1356 parse->distinctClause);
1357 result_plan->plan_rows = dNumDistinctRows;
1358 /* The Unique node won't change sort ordering */
1363 * If ORDER BY was given and we were not able to make the plan come out in
1364 * the right order, add an explicit sort step.
1366 if (parse->sortClause)
1368 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1370 result_plan = (Plan *) make_sort_from_pathkeys(root,
1372 root->sort_pathkeys,
1374 current_pathkeys = root->sort_pathkeys;
1379 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1381 if (parse->limitCount || parse->limitOffset)
1383 result_plan = (Plan *) make_limit(result_plan,
1391 * Deal with the RETURNING clause if any. It's convenient to pass the
1392 * returningList through setrefs.c now rather than at top level (if we
1393 * waited, handling inherited UPDATE/DELETE would be much harder).
1395 if (parse->returningList)
1399 Assert(parse->resultRelation);
1400 rlist = set_returning_clause_references(root->glob,
1401 parse->returningList,
1403 parse->resultRelation);
1404 root->returningLists = list_make1(rlist);
1407 root->returningLists = NIL;
1409 /* Compute result-relations list if needed */
1410 if (parse->resultRelation)
1411 root->resultRelations = list_make1_int(parse->resultRelation);
1413 root->resultRelations = NIL;
1416 * Return the actual output ordering in query_pathkeys for possible use by
1417 * an outer query level.
1419 root->query_pathkeys = current_pathkeys;
1425 * Detect whether a plan node is a "dummy" plan created when a relation
1426 * is deemed not to need scanning due to constraint exclusion.
1428 * Currently, such dummy plans are Result nodes with constant FALSE
1432 is_dummy_plan(Plan *plan)
1434 if (IsA(plan, Result))
1436 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1438 if (list_length(rcqual) == 1)
1440 Const *constqual = (Const *) linitial(rcqual);
1442 if (constqual && IsA(constqual, Const))
1444 if (!constqual->constisnull &&
1445 !DatumGetBool(constqual->constvalue))
1454 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1456 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1457 * results back in *count_est and *offset_est. These variables are set to
1458 * 0 if the corresponding clause is not present, and -1 if it's present
1459 * but we couldn't estimate the value for it. (The "0" convention is OK
1460 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1461 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1462 * usual practice of never estimating less than one row.) These values will
1463 * be passed to make_limit, which see if you change this code.
1465 * The return value is the suitably adjusted tuple_fraction to use for
1466 * planning the query. This adjustment is not overridable, since it reflects
1467 * plan actions that grouping_planner() will certainly take, not assumptions
1471 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1472 int64 *offset_est, int64 *count_est)
1474 Query *parse = root->parse;
1476 double limit_fraction;
1478 /* Should not be called unless LIMIT or OFFSET */
1479 Assert(parse->limitCount || parse->limitOffset);
1482 * Try to obtain the clause values. We use estimate_expression_value
1483 * primarily because it can sometimes do something useful with Params.
1485 if (parse->limitCount)
1487 est = estimate_expression_value(root, parse->limitCount);
1488 if (est && IsA(est, Const))
1490 if (((Const *) est)->constisnull)
1492 /* NULL indicates LIMIT ALL, ie, no limit */
1493 *count_est = 0; /* treat as not present */
1497 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1498 if (*count_est <= 0)
1499 *count_est = 1; /* force to at least 1 */
1503 *count_est = -1; /* can't estimate */
1506 *count_est = 0; /* not present */
1508 if (parse->limitOffset)
1510 est = estimate_expression_value(root, parse->limitOffset);
1511 if (est && IsA(est, Const))
1513 if (((Const *) est)->constisnull)
1515 /* Treat NULL as no offset; the executor will too */
1516 *offset_est = 0; /* treat as not present */
1520 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1521 if (*offset_est < 0)
1522 *offset_est = 0; /* less than 0 is same as 0 */
1526 *offset_est = -1; /* can't estimate */
1529 *offset_est = 0; /* not present */
1531 if (*count_est != 0)
1534 * A LIMIT clause limits the absolute number of tuples returned.
1535 * However, if it's not a constant LIMIT then we have to guess; for
1536 * lack of a better idea, assume 10% of the plan's result is wanted.
1538 if (*count_est < 0 || *offset_est < 0)
1540 /* LIMIT or OFFSET is an expression ... punt ... */
1541 limit_fraction = 0.10;
1545 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1546 limit_fraction = (double) *count_est + (double) *offset_est;
1550 * If we have absolute limits from both caller and LIMIT, use the
1551 * smaller value; likewise if they are both fractional. If one is
1552 * fractional and the other absolute, we can't easily determine which
1553 * is smaller, but we use the heuristic that the absolute will usually
1556 if (tuple_fraction >= 1.0)
1558 if (limit_fraction >= 1.0)
1561 tuple_fraction = Min(tuple_fraction, limit_fraction);
1565 /* caller absolute, limit fractional; use caller's value */
1568 else if (tuple_fraction > 0.0)
1570 if (limit_fraction >= 1.0)
1572 /* caller fractional, limit absolute; use limit */
1573 tuple_fraction = limit_fraction;
1577 /* both fractional */
1578 tuple_fraction = Min(tuple_fraction, limit_fraction);
1583 /* no info from caller, just use limit */
1584 tuple_fraction = limit_fraction;
1587 else if (*offset_est != 0 && tuple_fraction > 0.0)
1590 * We have an OFFSET but no LIMIT. This acts entirely differently
1591 * from the LIMIT case: here, we need to increase rather than decrease
1592 * the caller's tuple_fraction, because the OFFSET acts to cause more
1593 * tuples to be fetched instead of fewer. This only matters if we got
1594 * a tuple_fraction > 0, however.
1596 * As above, use 10% if OFFSET is present but unestimatable.
1598 if (*offset_est < 0)
1599 limit_fraction = 0.10;
1601 limit_fraction = (double) *offset_est;
1604 * If we have absolute counts from both caller and OFFSET, add them
1605 * together; likewise if they are both fractional. If one is
1606 * fractional and the other absolute, we want to take the larger, and
1607 * we heuristically assume that's the fractional one.
1609 if (tuple_fraction >= 1.0)
1611 if (limit_fraction >= 1.0)
1613 /* both absolute, so add them together */
1614 tuple_fraction += limit_fraction;
1618 /* caller absolute, limit fractional; use limit */
1619 tuple_fraction = limit_fraction;
1624 if (limit_fraction >= 1.0)
1626 /* caller fractional, limit absolute; use caller's value */
1630 /* both fractional, so add them together */
1631 tuple_fraction += limit_fraction;
1632 if (tuple_fraction >= 1.0)
1633 tuple_fraction = 0.0; /* assume fetch all */
1638 return tuple_fraction;
1643 * preprocess_groupclause - do preparatory work on GROUP BY clause
1645 * The idea here is to adjust the ordering of the GROUP BY elements
1646 * (which in itself is semantically insignificant) to match ORDER BY,
1647 * thereby allowing a single sort operation to both implement the ORDER BY
1648 * requirement and set up for a Unique step that implements GROUP BY.
1650 * In principle it might be interesting to consider other orderings of the
1651 * GROUP BY elements, which could match the sort ordering of other
1652 * possible plans (eg an indexscan) and thereby reduce cost. We don't
1653 * bother with that, though. Hashed grouping will frequently win anyway.
1655 * Note: we need no comparable processing of the distinctClause because
1656 * the parser already enforced that that matches ORDER BY.
1659 preprocess_groupclause(PlannerInfo *root)
1661 Query *parse = root->parse;
1662 List *new_groupclause;
1667 /* If no ORDER BY, nothing useful to do here */
1668 if (parse->sortClause == NIL)
1672 * Scan the ORDER BY clause and construct a list of matching GROUP BY
1673 * items, but only as far as we can make a matching prefix.
1675 * This code assumes that the sortClause contains no duplicate items.
1677 new_groupclause = NIL;
1678 foreach(sl, parse->sortClause)
1680 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
1682 foreach(gl, parse->groupClause)
1684 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1688 new_groupclause = lappend(new_groupclause, gc);
1693 break; /* no match, so stop scanning */
1696 /* Did we match all of the ORDER BY list, or just some of it? */
1697 partial_match = (sl != NULL);
1699 /* If no match at all, no point in reordering GROUP BY */
1700 if (new_groupclause == NIL)
1704 * Add any remaining GROUP BY items to the new list, but only if we
1705 * were able to make a complete match. In other words, we only
1706 * rearrange the GROUP BY list if the result is that one list is a
1707 * prefix of the other --- otherwise there's no possibility of a
1708 * common sort. Also, give up if there are any non-sortable GROUP BY
1709 * items, since then there's no hope anyway.
1711 foreach(gl, parse->groupClause)
1713 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
1715 if (list_member_ptr(new_groupclause, gc))
1716 continue; /* it matched an ORDER BY item */
1718 return; /* give up, no common sort possible */
1719 if (!OidIsValid(gc->sortop))
1720 return; /* give up, GROUP BY can't be sorted */
1721 new_groupclause = lappend(new_groupclause, gc);
1724 /* Success --- install the rearranged GROUP BY list */
1725 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
1726 parse->groupClause = new_groupclause;
1730 * choose_hashed_grouping - should we use hashed grouping?
1732 * Note: this is only applied when both alternatives are actually feasible.
1735 choose_hashed_grouping(PlannerInfo *root,
1736 double tuple_fraction, double limit_tuples,
1737 Path *cheapest_path, Path *sorted_path,
1738 double dNumGroups, AggClauseCounts *agg_counts)
1740 int numGroupCols = list_length(root->parse->groupClause);
1741 double cheapest_path_rows;
1742 int cheapest_path_width;
1744 List *target_pathkeys;
1745 List *current_pathkeys;
1749 /* Prefer sorting when enable_hashagg is off */
1750 if (!enable_hashagg)
1754 * Don't do it if it doesn't look like the hashtable will fit into
1757 * Beware here of the possibility that cheapest_path->parent is NULL. This
1758 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1760 if (cheapest_path->parent)
1762 cheapest_path_rows = cheapest_path->parent->rows;
1763 cheapest_path_width = cheapest_path->parent->width;
1767 cheapest_path_rows = 1; /* assume non-set result */
1768 cheapest_path_width = 100; /* arbitrary */
1771 /* Estimate per-hash-entry space at tuple width... */
1772 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1773 /* plus space for pass-by-ref transition values... */
1774 hashentrysize += agg_counts->transitionSpace;
1775 /* plus the per-hash-entry overhead */
1776 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1778 if (hashentrysize * dNumGroups > work_mem * 1024L)
1782 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
1783 * DISTINCT and ORDER BY as the assumed required output sort order.
1784 * This is an oversimplification because the DISTINCT might get
1785 * implemented via hashing, but it's not clear that the case is common
1786 * enough (or that our estimates are good enough) to justify trying to
1789 if (list_length(root->distinct_pathkeys) >
1790 list_length(root->sort_pathkeys))
1791 target_pathkeys = root->distinct_pathkeys;
1793 target_pathkeys = root->sort_pathkeys;
1796 * See if the estimated cost is no more than doing it the other way. While
1797 * avoiding the need for sorted input is usually a win, the fact that the
1798 * output won't be sorted may be a loss; so we need to do an actual cost
1801 * We need to consider cheapest_path + hashagg [+ final sort] versus
1802 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1803 * presorted_path + group or agg [+ final sort] where brackets indicate a
1804 * step that may not be needed. We assume query_planner() will have
1805 * returned a presorted path only if it's a winner compared to
1806 * cheapest_path for this purpose.
1808 * These path variables are dummies that just hold cost fields; we don't
1809 * make actual Paths for these steps.
1811 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1812 numGroupCols, dNumGroups,
1813 cheapest_path->startup_cost, cheapest_path->total_cost,
1814 cheapest_path_rows);
1815 /* Result of hashed agg is always unsorted */
1816 if (target_pathkeys)
1817 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
1818 dNumGroups, cheapest_path_width, limit_tuples);
1822 sorted_p.startup_cost = sorted_path->startup_cost;
1823 sorted_p.total_cost = sorted_path->total_cost;
1824 current_pathkeys = sorted_path->pathkeys;
1828 sorted_p.startup_cost = cheapest_path->startup_cost;
1829 sorted_p.total_cost = cheapest_path->total_cost;
1830 current_pathkeys = cheapest_path->pathkeys;
1832 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1834 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1835 cheapest_path_rows, cheapest_path_width, -1.0);
1836 current_pathkeys = root->group_pathkeys;
1839 if (root->parse->hasAggs)
1840 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1841 numGroupCols, dNumGroups,
1842 sorted_p.startup_cost, sorted_p.total_cost,
1843 cheapest_path_rows);
1845 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1846 sorted_p.startup_cost, sorted_p.total_cost,
1847 cheapest_path_rows);
1848 /* The Agg or Group node will preserve ordering */
1849 if (target_pathkeys &&
1850 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
1851 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
1852 dNumGroups, cheapest_path_width, limit_tuples);
1855 * Now make the decision using the top-level tuple fraction. First we
1856 * have to convert an absolute count (LIMIT) into fractional form.
1858 if (tuple_fraction >= 1.0)
1859 tuple_fraction /= dNumGroups;
1861 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1862 tuple_fraction) < 0)
1864 /* Hashed is cheaper, so use it */
1871 * choose_hashed_distinct - should we use hashing for DISTINCT?
1873 * This is fairly similar to choose_hashed_grouping, but there are enough
1874 * differences that it doesn't seem worth trying to unify the two functions.
1876 * But note that making the two choices independently is a bit bogus in
1877 * itself. If the two could be combined into a single choice operation
1878 * it'd probably be better, but that seems far too unwieldy to be practical,
1879 * especially considering that the combination of GROUP BY and DISTINCT
1880 * isn't very common in real queries. By separating them, we are giving
1881 * extra preference to using a sorting implementation when a common sort key
1882 * is available ... and that's not necessarily wrong anyway.
1884 * Note: this is only applied when both alternatives are actually feasible.
1887 choose_hashed_distinct(PlannerInfo *root,
1888 Plan *input_plan, List *input_pathkeys,
1889 double tuple_fraction, double limit_tuples,
1890 double dNumDistinctRows)
1892 int numDistinctCols = list_length(root->parse->distinctClause);
1894 List *current_pathkeys;
1895 List *needed_pathkeys;
1899 /* Prefer sorting when enable_hashagg is off */
1900 if (!enable_hashagg)
1904 * Don't do it if it doesn't look like the hashtable will fit into
1907 hashentrysize = MAXALIGN(input_plan->plan_width) + MAXALIGN(sizeof(MinimalTupleData));
1909 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
1913 * See if the estimated cost is no more than doing it the other way. While
1914 * avoiding the need for sorted input is usually a win, the fact that the
1915 * output won't be sorted may be a loss; so we need to do an actual cost
1918 * We need to consider input_plan + hashagg [+ final sort] versus
1919 * input_plan [+ sort] + group [+ final sort] where brackets indicate
1920 * a step that may not be needed.
1922 * These path variables are dummies that just hold cost fields; we don't
1923 * make actual Paths for these steps.
1925 cost_agg(&hashed_p, root, AGG_HASHED, 0,
1926 numDistinctCols, dNumDistinctRows,
1927 input_plan->startup_cost, input_plan->total_cost,
1928 input_plan->plan_rows);
1930 * Result of hashed agg is always unsorted, so if ORDER BY is present
1931 * we need to charge for the final sort.
1933 if (root->parse->sortClause)
1934 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1935 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1938 * Now for the GROUP case. See comments in grouping_planner about the
1939 * sorting choices here --- this code should match that code.
1941 sorted_p.startup_cost = input_plan->startup_cost;
1942 sorted_p.total_cost = input_plan->total_cost;
1943 current_pathkeys = input_pathkeys;
1944 if (root->parse->hasDistinctOn &&
1945 list_length(root->distinct_pathkeys) <
1946 list_length(root->sort_pathkeys))
1947 needed_pathkeys = root->sort_pathkeys;
1949 needed_pathkeys = root->distinct_pathkeys;
1950 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1952 if (list_length(root->distinct_pathkeys) >=
1953 list_length(root->sort_pathkeys))
1954 current_pathkeys = root->distinct_pathkeys;
1956 current_pathkeys = root->sort_pathkeys;
1957 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
1958 input_plan->plan_rows, input_plan->plan_width, -1.0);
1960 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
1961 sorted_p.startup_cost, sorted_p.total_cost,
1962 input_plan->plan_rows);
1963 if (root->parse->sortClause &&
1964 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1965 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1966 dNumDistinctRows, input_plan->plan_width, limit_tuples);
1969 * Now make the decision using the top-level tuple fraction. First we
1970 * have to convert an absolute count (LIMIT) into fractional form.
1972 if (tuple_fraction >= 1.0)
1973 tuple_fraction /= dNumDistinctRows;
1975 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1976 tuple_fraction) < 0)
1978 /* Hashed is cheaper, so use it */
1985 * make_subplanTargetList
1986 * Generate appropriate target list when grouping is required.
1988 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1989 * above the result of query_planner, we typically want to pass a different
1990 * target list to query_planner than the outer plan nodes should have.
1991 * This routine generates the correct target list for the subplan.
1993 * The initial target list passed from the parser already contains entries
1994 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1995 * for variables used only in HAVING clauses; so we need to add those
1996 * variables to the subplan target list. Also, we flatten all expressions
1997 * except GROUP BY items into their component variables; the other expressions
1998 * will be computed by the inserted nodes rather than by the subplan.
1999 * For example, given a query like
2000 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
2001 * we want to pass this targetlist to the subplan:
2003 * where the a+b target will be used by the Sort/Group steps, and the
2004 * other targets will be used for computing the final results. (In the
2005 * above example we could theoretically suppress the a and b targets and
2006 * pass down only c,d,a+b, but it's not really worth the trouble to
2007 * eliminate simple var references from the subplan. We will avoid doing
2008 * the extra computation to recompute a+b at the outer level; see
2009 * fix_upper_expr() in setrefs.c.)
2011 * If we are grouping or aggregating, *and* there are no non-Var grouping
2012 * expressions, then the returned tlist is effectively dummy; we do not
2013 * need to force it to be evaluated, because all the Vars it contains
2014 * should be present in the output of query_planner anyway.
2016 * 'tlist' is the query's target list.
2017 * 'groupColIdx' receives an array of column numbers for the GROUP BY
2018 * expressions (if there are any) in the subplan's target list.
2019 * 'need_tlist_eval' is set true if we really need to evaluate the
2022 * The result is the targetlist to be passed to the subplan.
2026 make_subplanTargetList(PlannerInfo *root,
2028 AttrNumber **groupColIdx,
2029 bool *need_tlist_eval)
2031 Query *parse = root->parse;
2036 *groupColIdx = NULL;
2039 * If we're not grouping or aggregating, there's nothing to do here;
2040 * query_planner should receive the unmodified target list.
2042 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
2044 *need_tlist_eval = true;
2049 * Otherwise, start with a "flattened" tlist (having just the vars
2050 * mentioned in the targetlist and HAVING qual --- but not upper-level
2051 * Vars; they will be replaced by Params later on).
2053 sub_tlist = flatten_tlist(tlist);
2054 extravars = pull_var_clause(parse->havingQual, true);
2055 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2056 list_free(extravars);
2057 *need_tlist_eval = false; /* only eval if not flat tlist */
2060 * If grouping, create sub_tlist entries for all GROUP BY expressions
2061 * (GROUP BY items that are simple Vars should be in the list already),
2062 * and make an array showing where the group columns are in the sub_tlist.
2064 numCols = list_length(parse->groupClause);
2068 AttrNumber *grpColIdx;
2071 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2072 *groupColIdx = grpColIdx;
2074 foreach(gl, parse->groupClause)
2076 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2077 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2078 TargetEntry *te = NULL;
2081 * Find or make a matching sub_tlist entry. If the groupexpr
2082 * isn't a Var, no point in searching. (Note that the parser
2083 * won't make multiple groupClause entries for the same TLE.)
2085 if (groupexpr && IsA(groupexpr, Var))
2089 foreach(sl, sub_tlist)
2091 TargetEntry *lte = (TargetEntry *) lfirst(sl);
2093 if (equal(groupexpr, lte->expr))
2102 te = makeTargetEntry((Expr *) groupexpr,
2103 list_length(sub_tlist) + 1,
2106 sub_tlist = lappend(sub_tlist, te);
2107 *need_tlist_eval = true; /* it's not flat anymore */
2110 /* and save its resno */
2111 grpColIdx[keyno++] = te->resno;
2119 * locate_grouping_columns
2120 * Locate grouping columns in the tlist chosen by query_planner.
2122 * This is only needed if we don't use the sub_tlist chosen by
2123 * make_subplanTargetList. We have to forget the column indexes found
2124 * by that routine and re-locate the grouping vars in the real sub_tlist.
2127 locate_grouping_columns(PlannerInfo *root,
2130 AttrNumber *groupColIdx)
2136 * No work unless grouping.
2138 if (!root->parse->groupClause)
2140 Assert(groupColIdx == NULL);
2143 Assert(groupColIdx != NULL);
2145 foreach(gl, root->parse->groupClause)
2147 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2148 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2149 TargetEntry *te = NULL;
2152 foreach(sl, sub_tlist)
2154 te = (TargetEntry *) lfirst(sl);
2155 if (equal(groupexpr, te->expr))
2159 elog(ERROR, "failed to locate grouping columns");
2161 groupColIdx[keyno++] = te->resno;
2166 * postprocess_setop_tlist
2167 * Fix up targetlist returned by plan_set_operations().
2169 * We need to transpose sort key info from the orig_tlist into new_tlist.
2170 * NOTE: this would not be good enough if we supported resjunk sort keys
2171 * for results of set operations --- then, we'd need to project a whole
2172 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2173 * find any resjunk columns in orig_tlist.
2176 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2179 ListCell *orig_tlist_item = list_head(orig_tlist);
2181 foreach(l, new_tlist)
2183 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2184 TargetEntry *orig_tle;
2186 /* ignore resjunk columns in setop result */
2187 if (new_tle->resjunk)
2190 Assert(orig_tlist_item != NULL);
2191 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2192 orig_tlist_item = lnext(orig_tlist_item);
2193 if (orig_tle->resjunk) /* should not happen */
2194 elog(ERROR, "resjunk output columns are not implemented");
2195 Assert(new_tle->resno == orig_tle->resno);
2196 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2198 if (orig_tlist_item != NULL)
2199 elog(ERROR, "resjunk output columns are not implemented");