1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * src/backend/optimizer/plan/planner.c
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "executor/executor.h"
22 #include "executor/nodeAgg.h"
23 #include "miscadmin.h"
24 #include "nodes/makefuncs.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/cost.h"
27 #include "optimizer/pathnode.h"
28 #include "optimizer/paths.h"
29 #include "optimizer/plancat.h"
30 #include "optimizer/planmain.h"
31 #include "optimizer/planner.h"
32 #include "optimizer/prep.h"
33 #include "optimizer/subselect.h"
34 #include "optimizer/tlist.h"
35 #include "optimizer/var.h"
36 #ifdef OPTIMIZER_DEBUG
37 #include "nodes/print.h"
39 #include "parser/analyze.h"
40 #include "parser/parse_expr.h"
41 #include "parser/parse_oper.h"
42 #include "parser/parsetree.h"
43 #include "utils/lsyscache.h"
44 #include "utils/syscache.h"
48 double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
50 /* Hook for plugins to get control in planner() */
51 planner_hook_type planner_hook = NULL;
54 /* Expression kind codes for preprocess_expression */
55 #define EXPRKIND_QUAL 0
56 #define EXPRKIND_TARGET 1
57 #define EXPRKIND_RTFUNC 2
58 #define EXPRKIND_VALUES 3
59 #define EXPRKIND_LIMIT 4
60 #define EXPRKIND_APPINFO 5
63 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
64 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
65 static Plan *inheritance_planner(PlannerInfo *root);
66 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
67 static bool is_dummy_plan(Plan *plan);
68 static void preprocess_rowmarks(PlannerInfo *root);
69 static double preprocess_limit(PlannerInfo *root,
70 double tuple_fraction,
71 int64 *offset_est, int64 *count_est);
72 static void preprocess_groupclause(PlannerInfo *root);
73 static bool choose_hashed_grouping(PlannerInfo *root,
74 double tuple_fraction, double limit_tuples,
75 double path_rows, int path_width,
76 Path *cheapest_path, Path *sorted_path,
77 double dNumGroups, AggClauseCounts *agg_counts);
78 static bool choose_hashed_distinct(PlannerInfo *root,
79 double tuple_fraction, double limit_tuples,
80 double path_rows, int path_width,
81 Cost cheapest_startup_cost, Cost cheapest_total_cost,
82 Cost sorted_startup_cost, Cost sorted_total_cost,
83 List *sorted_pathkeys,
84 double dNumDistinctRows);
85 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
86 AttrNumber **groupColIdx, bool *need_tlist_eval);
87 static void locate_grouping_columns(PlannerInfo *root,
90 AttrNumber *groupColIdx);
91 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
92 static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
93 static List *add_volatile_sort_exprs(List *window_tlist, List *tlist,
95 static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
96 List *tlist, bool canonicalize);
97 static void get_column_info_for_window(PlannerInfo *root, WindowClause *wc,
99 int numSortCols, AttrNumber *sortColIdx,
101 AttrNumber **partColIdx,
104 AttrNumber **ordColIdx,
108 /*****************************************************************************
110 * Query optimizer entry point
112 * To support loadable plugins that monitor or modify planner behavior,
113 * we provide a hook variable that lets a plugin get control before and
114 * after the standard planning process. The plugin would normally call
115 * standard_planner().
117 * Note to plugin authors: standard_planner() scribbles on its Query input,
118 * so you'd better copy that data structure if you want to plan more than once.
120 *****************************************************************************/
122 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
127 result = (*planner_hook) (parse, cursorOptions, boundParams);
129 result = standard_planner(parse, cursorOptions, boundParams);
134 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
138 double tuple_fraction;
145 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
146 if (parse->utilityStmt &&
147 IsA(parse->utilityStmt, DeclareCursorStmt))
148 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
151 * Set up global state for this planner invocation. This data is needed
152 * across all levels of sub-Query that might exist in the given command,
153 * so we keep it in a separate struct that's linked to by each per-Query
156 glob = makeNode(PlannerGlobal);
158 glob->boundParams = boundParams;
159 glob->paramlist = NIL;
160 glob->subplans = NIL;
161 glob->subrtables = NIL;
162 glob->subrowmarks = NIL;
163 glob->rewindPlanIDs = NULL;
164 glob->finalrtable = NIL;
165 glob->finalrowmarks = NIL;
166 glob->relationOids = NIL;
167 glob->invalItems = NIL;
169 glob->transientPlan = false;
171 /* Determine what fraction of the plan is likely to be scanned */
172 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
175 * We have no real idea how many tuples the user will ultimately FETCH
176 * from a cursor, but it is often the case that he doesn't want 'em
177 * all, or would prefer a fast-start plan anyway so that he can
178 * process some of the tuples sooner. Use a GUC parameter to decide
179 * what fraction to optimize for.
181 tuple_fraction = cursor_tuple_fraction;
184 * We document cursor_tuple_fraction as simply being a fraction, which
185 * means the edge cases 0 and 1 have to be treated specially here. We
186 * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
188 if (tuple_fraction >= 1.0)
189 tuple_fraction = 0.0;
190 else if (tuple_fraction <= 0.0)
191 tuple_fraction = 1e-10;
195 /* Default assumption is we need all the tuples */
196 tuple_fraction = 0.0;
199 /* primary planning entry point (may recurse for subqueries) */
200 top_plan = subquery_planner(glob, parse, NULL,
201 false, tuple_fraction, &root);
204 * If creating a plan for a scrollable cursor, make sure it can run
205 * backwards on demand. Add a Material node at the top at need.
207 if (cursorOptions & CURSOR_OPT_SCROLL)
209 if (!ExecSupportsBackwardScan(top_plan))
210 top_plan = materialize_finished_plan(top_plan);
213 /* final cleanup of the plan */
214 Assert(glob->finalrtable == NIL);
215 Assert(glob->finalrowmarks == NIL);
216 top_plan = set_plan_references(glob, top_plan,
219 /* ... and the subplans (both regular subplans and initplans) */
220 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
221 Assert(list_length(glob->subplans) == list_length(glob->subrowmarks));
222 lrt = list_head(glob->subrtables);
223 lrm = list_head(glob->subrowmarks);
224 foreach(lp, glob->subplans)
226 Plan *subplan = (Plan *) lfirst(lp);
227 List *subrtable = (List *) lfirst(lrt);
228 List *subrowmark = (List *) lfirst(lrm);
230 lfirst(lp) = set_plan_references(glob, subplan,
231 subrtable, subrowmark);
236 /* build the PlannedStmt result */
237 result = makeNode(PlannedStmt);
239 result->commandType = parse->commandType;
240 result->hasReturning = (parse->returningList != NIL);
241 result->canSetTag = parse->canSetTag;
242 result->transientPlan = glob->transientPlan;
243 result->planTree = top_plan;
244 result->rtable = glob->finalrtable;
245 result->resultRelations = root->resultRelations;
246 result->utilityStmt = parse->utilityStmt;
247 result->intoClause = parse->intoClause;
248 result->subplans = glob->subplans;
249 result->rewindPlanIDs = glob->rewindPlanIDs;
250 result->rowMarks = glob->finalrowmarks;
251 result->relationOids = glob->relationOids;
252 result->invalItems = glob->invalItems;
253 result->nParamExec = list_length(glob->paramlist);
259 /*--------------------
261 * Invokes the planner on a subquery. We recurse to here for each
262 * sub-SELECT found in the query tree.
264 * glob is the global state for the current planner run.
265 * parse is the querytree produced by the parser & rewriter.
266 * parent_root is the immediate parent Query's info (NULL at the top level).
267 * hasRecursion is true if this is a recursive WITH query.
268 * tuple_fraction is the fraction of tuples we expect will be retrieved.
269 * tuple_fraction is interpreted as explained for grouping_planner, below.
271 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
272 * among other things this tells the output sort ordering of the plan.
274 * Basically, this routine does the stuff that should only be done once
275 * per Query object. It then calls grouping_planner. At one time,
276 * grouping_planner could be invoked recursively on the same Query object;
277 * that's not currently true, but we keep the separation between the two
278 * routines anyway, in case we need it again someday.
280 * subquery_planner will be called recursively to handle sub-Query nodes
281 * found within the query's expressions and rangetable.
283 * Returns a query plan.
284 *--------------------
287 subquery_planner(PlannerGlobal *glob, Query *parse,
288 PlannerInfo *parent_root,
289 bool hasRecursion, double tuple_fraction,
290 PlannerInfo **subroot)
292 int num_old_subplans = list_length(glob->subplans);
299 /* Create a PlannerInfo data structure for this subquery */
300 root = makeNode(PlannerInfo);
303 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
304 root->parent_root = parent_root;
305 root->planner_cxt = CurrentMemoryContext;
306 root->init_plans = NIL;
307 root->cte_plan_ids = NIL;
308 root->eq_classes = NIL;
309 root->append_rel_list = NIL;
310 root->rowMarks = NIL;
311 root->hasInheritedTarget = false;
313 root->hasRecursion = hasRecursion;
315 root->wt_param_id = SS_assign_special_param(root);
317 root->wt_param_id = -1;
318 root->non_recursive_plan = NULL;
321 * If there is a WITH list, process each WITH query and build an initplan
322 * SubPlan structure for it.
325 SS_process_ctes(root);
328 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
329 * to transform them into joins. Note that this step does not descend
330 * into subqueries; if we pull up any subqueries below, their SubLinks are
331 * processed just before pulling them up.
333 if (parse->hasSubLinks)
334 pull_up_sublinks(root);
337 * Scan the rangetable for set-returning functions, and inline them if
338 * possible (producing subqueries that might get pulled up next).
339 * Recursion issues here are handled in the same way as for SubLinks.
341 inline_set_returning_functions(root);
344 * Check to see if any subqueries in the jointree can be merged into
347 parse->jointree = (FromExpr *)
348 pull_up_subqueries(root, (Node *) parse->jointree, NULL, NULL);
351 * If this is a simple UNION ALL query, flatten it into an appendrel.
352 * We do this now because it requires applying pull_up_subqueries to the
353 * leaf queries of the UNION ALL, which weren't touched above because they
354 * weren't referenced by the jointree (they will be after we do this).
356 if (parse->setOperations)
357 flatten_simple_union_all(root);
360 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
361 * avoid the expense of doing flatten_join_alias_vars(). Also check for
362 * outer joins --- if none, we can skip reduce_outer_joins(). This must be
363 * done after we have done pull_up_subqueries, of course.
365 root->hasJoinRTEs = false;
366 hasOuterJoins = false;
367 foreach(l, parse->rtable)
369 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
371 if (rte->rtekind == RTE_JOIN)
373 root->hasJoinRTEs = true;
374 if (IS_OUTER_JOIN(rte->jointype))
376 hasOuterJoins = true;
377 /* Can quit scanning once we find an outer join */
384 * Preprocess RowMark information. We need to do this after subquery
385 * pullup (so that all non-inherited RTEs are present) and before
386 * inheritance expansion (so that the info is available for
387 * expand_inherited_tables to examine and modify).
389 preprocess_rowmarks(root);
392 * Expand any rangetable entries that are inheritance sets into "append
393 * relations". This can add entries to the rangetable, but they must be
394 * plain base relations not joins, so it's OK (and marginally more
395 * efficient) to do it after checking for join RTEs. We must do it after
396 * pulling up subqueries, else we'd fail to handle inherited tables in
399 expand_inherited_tables(root);
402 * Set hasHavingQual to remember if HAVING clause is present. Needed
403 * because preprocess_expression will reduce a constant-true condition to
404 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
406 root->hasHavingQual = (parse->havingQual != NULL);
408 /* Clear this flag; might get set in distribute_qual_to_rels */
409 root->hasPseudoConstantQuals = false;
412 * Do expression preprocessing on targetlist and quals, as well as other
413 * random expressions in the querytree. Note that we do not need to
414 * handle sort/group expressions explicitly, because they are actually
415 * part of the targetlist.
417 parse->targetList = (List *)
418 preprocess_expression(root, (Node *) parse->targetList,
421 parse->returningList = (List *)
422 preprocess_expression(root, (Node *) parse->returningList,
425 preprocess_qual_conditions(root, (Node *) parse->jointree);
427 parse->havingQual = preprocess_expression(root, parse->havingQual,
430 foreach(l, parse->windowClause)
432 WindowClause *wc = (WindowClause *) lfirst(l);
434 /* partitionClause/orderClause are sort/group expressions */
435 wc->startOffset = preprocess_expression(root, wc->startOffset,
437 wc->endOffset = preprocess_expression(root, wc->endOffset,
441 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
443 parse->limitCount = preprocess_expression(root, parse->limitCount,
446 root->append_rel_list = (List *)
447 preprocess_expression(root, (Node *) root->append_rel_list,
450 /* Also need to preprocess expressions for function and values RTEs */
451 foreach(l, parse->rtable)
453 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
455 if (rte->rtekind == RTE_FUNCTION)
456 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
458 else if (rte->rtekind == RTE_VALUES)
459 rte->values_lists = (List *)
460 preprocess_expression(root, (Node *) rte->values_lists,
465 * In some cases we may want to transfer a HAVING clause into WHERE. We
466 * cannot do so if the HAVING clause contains aggregates (obviously) or
467 * volatile functions (since a HAVING clause is supposed to be executed
468 * only once per group). Also, it may be that the clause is so expensive
469 * to execute that we're better off doing it only once per group, despite
470 * the loss of selectivity. This is hard to estimate short of doing the
471 * entire planning process twice, so we use a heuristic: clauses
472 * containing subplans are left in HAVING. Otherwise, we move or copy the
473 * HAVING clause into WHERE, in hopes of eliminating tuples before
474 * aggregation instead of after.
476 * If the query has explicit grouping then we can simply move such a
477 * clause into WHERE; any group that fails the clause will not be in the
478 * output because none of its tuples will reach the grouping or
479 * aggregation stage. Otherwise we must have a degenerate (variable-free)
480 * HAVING clause, which we put in WHERE so that query_planner() can use it
481 * in a gating Result node, but also keep in HAVING to ensure that we
482 * don't emit a bogus aggregated row. (This could be done better, but it
483 * seems not worth optimizing.)
485 * Note that both havingQual and parse->jointree->quals are in
486 * implicitly-ANDed-list form at this point, even though they are declared
490 foreach(l, (List *) parse->havingQual)
492 Node *havingclause = (Node *) lfirst(l);
494 if (contain_agg_clause(havingclause) ||
495 contain_volatile_functions(havingclause) ||
496 contain_subplans(havingclause))
498 /* keep it in HAVING */
499 newHaving = lappend(newHaving, havingclause);
501 else if (parse->groupClause)
503 /* move it to WHERE */
504 parse->jointree->quals = (Node *)
505 lappend((List *) parse->jointree->quals, havingclause);
509 /* put a copy in WHERE, keep it in HAVING */
510 parse->jointree->quals = (Node *)
511 lappend((List *) parse->jointree->quals,
512 copyObject(havingclause));
513 newHaving = lappend(newHaving, havingclause);
516 parse->havingQual = (Node *) newHaving;
519 * If we have any outer joins, try to reduce them to plain inner joins.
520 * This step is most easily done after we've done expression
524 reduce_outer_joins(root);
527 * Do the main planning. If we have an inherited target relation, that
528 * needs special processing, else go straight to grouping_planner.
530 if (parse->resultRelation &&
531 rt_fetch(parse->resultRelation, parse->rtable)->inh)
532 plan = inheritance_planner(root);
535 plan = grouping_planner(root, tuple_fraction);
536 /* If it's not SELECT, we need a ModifyTable node */
537 if (parse->commandType != CMD_SELECT)
539 List *returningLists;
543 * Deal with the RETURNING clause if any. It's convenient to pass
544 * the returningList through setrefs.c now rather than at top
545 * level (if we waited, handling inherited UPDATE/DELETE would be
548 if (parse->returningList)
552 Assert(parse->resultRelation);
553 rlist = set_returning_clause_references(root->glob,
554 parse->returningList,
556 parse->resultRelation);
557 returningLists = list_make1(rlist);
560 returningLists = NIL;
563 * If there was a FOR UPDATE/SHARE clause, the LockRows node will
564 * have dealt with fetching non-locked marked rows, else we need
565 * to have ModifyTable do that.
570 rowMarks = root->rowMarks;
572 plan = (Plan *) make_modifytable(parse->commandType,
573 copyObject(root->resultRelations),
577 SS_assign_special_param(root));
582 * If any subplans were generated, or if there are any parameters to worry
583 * about, build initPlan list and extParam/allParam sets for plan nodes,
584 * and attach the initPlans to the top plan node.
586 if (list_length(glob->subplans) != num_old_subplans ||
587 root->glob->paramlist != NIL)
588 SS_finalize_plan(root, plan, true);
590 /* Return internal info if caller wants it */
598 * preprocess_expression
599 * Do subquery_planner's preprocessing work for an expression,
600 * which can be a targetlist, a WHERE clause (including JOIN/ON
601 * conditions), or a HAVING clause.
604 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
607 * Fall out quickly if expression is empty. This occurs often enough to
608 * be worth checking. Note that null->null is the correct conversion for
609 * implicit-AND result format, too.
615 * If the query has any join RTEs, replace join alias variables with
616 * base-relation variables. We must do this before sublink processing,
617 * else sublinks expanded out from join aliases wouldn't get processed. We
618 * can skip it in VALUES lists, however, since they can't contain any Vars
621 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
622 expr = flatten_join_alias_vars(root, expr);
625 * Simplify constant expressions.
627 * Note: an essential effect of this is to convert named-argument function
628 * calls to positional notation and insert the current actual values of
629 * any default arguments for functions. To ensure that happens, we *must*
630 * process all expressions here. Previous PG versions sometimes skipped
631 * const-simplification if it didn't seem worth the trouble, but we can't
634 * Note: this also flattens nested AND and OR expressions into N-argument
635 * form. All processing of a qual expression after this point must be
636 * careful to maintain AND/OR flatness --- that is, do not generate a tree
637 * with AND directly under AND, nor OR directly under OR.
639 expr = eval_const_expressions(root, expr);
642 * If it's a qual or havingQual, canonicalize it.
644 if (kind == EXPRKIND_QUAL)
646 expr = (Node *) canonicalize_qual((Expr *) expr);
648 #ifdef OPTIMIZER_DEBUG
649 printf("After canonicalize_qual()\n");
654 /* Expand SubLinks to SubPlans */
655 if (root->parse->hasSubLinks)
656 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
659 * XXX do not insert anything here unless you have grokked the comments in
660 * SS_replace_correlation_vars ...
663 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
664 if (root->query_level > 1)
665 expr = SS_replace_correlation_vars(root, expr);
668 * If it's a qual or havingQual, convert it to implicit-AND format. (We
669 * don't want to do this before eval_const_expressions, since the latter
670 * would be unable to simplify a top-level AND correctly. Also,
671 * SS_process_sublinks expects explicit-AND format.)
673 if (kind == EXPRKIND_QUAL)
674 expr = (Node *) make_ands_implicit((Expr *) expr);
680 * preprocess_qual_conditions
681 * Recursively scan the query's jointree and do subquery_planner's
682 * preprocessing work on each qual condition found therein.
685 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
689 if (IsA(jtnode, RangeTblRef))
691 /* nothing to do here */
693 else if (IsA(jtnode, FromExpr))
695 FromExpr *f = (FromExpr *) jtnode;
698 foreach(l, f->fromlist)
699 preprocess_qual_conditions(root, lfirst(l));
701 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
703 else if (IsA(jtnode, JoinExpr))
705 JoinExpr *j = (JoinExpr *) jtnode;
707 preprocess_qual_conditions(root, j->larg);
708 preprocess_qual_conditions(root, j->rarg);
710 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
713 elog(ERROR, "unrecognized node type: %d",
714 (int) nodeTag(jtnode));
718 * inheritance_planner
719 * Generate a plan in the case where the result relation is an
722 * We have to handle this case differently from cases where a source relation
723 * is an inheritance set. Source inheritance is expanded at the bottom of the
724 * plan tree (see allpaths.c), but target inheritance has to be expanded at
725 * the top. The reason is that for UPDATE, each target relation needs a
726 * different targetlist matching its own column set. Fortunately,
727 * the UPDATE/DELETE target can never be the nullable side of an outer join,
728 * so it's OK to generate the plan this way.
730 * Returns a query plan.
733 inheritance_planner(PlannerInfo *root)
735 Query *parse = root->parse;
736 int parentRTindex = parse->resultRelation;
737 List *subplans = NIL;
738 List *resultRelations = NIL;
739 List *returningLists = NIL;
746 foreach(l, root->append_rel_list)
748 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
751 /* append_rel_list contains all append rels; ignore others */
752 if (appinfo->parent_relid != parentRTindex)
756 * Generate modified query with this rel as target.
758 memcpy(&subroot, root, sizeof(PlannerInfo));
759 subroot.parse = (Query *)
760 adjust_appendrel_attrs((Node *) parse,
762 subroot.init_plans = NIL;
763 subroot.hasInheritedTarget = true;
764 /* We needn't modify the child's append_rel_list */
765 /* There shouldn't be any OJ info to translate, as yet */
766 Assert(subroot.join_info_list == NIL);
767 /* and we haven't created PlaceHolderInfos, either */
768 Assert(subroot.placeholder_list == NIL);
771 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
774 * If this child rel was excluded by constraint exclusion, exclude it
777 if (is_dummy_plan(subplan))
780 /* Save rtable from first rel for use below */
782 rtable = subroot.parse->rtable;
784 subplans = lappend(subplans, subplan);
786 /* Make sure any initplans from this rel get into the outer list */
787 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
789 /* Build target-relations list for the executor */
790 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
792 /* Build list of per-relation RETURNING targetlists */
793 if (parse->returningList)
797 rlist = set_returning_clause_references(root->glob,
798 subroot.parse->returningList,
800 appinfo->child_relid);
801 returningLists = lappend(returningLists, rlist);
805 root->resultRelations = resultRelations;
807 /* Mark result as unordered (probably unnecessary) */
808 root->query_pathkeys = NIL;
811 * If we managed to exclude every child rel, return a dummy plan; it
812 * doesn't even need a ModifyTable node.
816 root->resultRelations = list_make1_int(parentRTindex);
817 /* although dummy, it must have a valid tlist for executor */
818 tlist = preprocess_targetlist(root, parse->targetList);
819 return (Plan *) make_result(root,
821 (Node *) list_make1(makeBoolConst(false,
827 * Planning might have modified the rangetable, due to changes of the
828 * Query structures inside subquery RTEs. We have to ensure that this
829 * gets propagated back to the master copy. But can't do this until we
830 * are done planning, because all the calls to grouping_planner need
831 * virgin sub-Queries to work from. (We are effectively assuming that
832 * sub-Queries will get planned identically each time, or at least that
833 * the impacts on their rangetables will be the same each time.)
835 * XXX should clean this up someday
837 parse->rtable = rtable;
840 * If there was a FOR UPDATE/SHARE clause, the LockRows node will have
841 * dealt with fetching non-locked marked rows, else we need to have
842 * ModifyTable do that.
847 rowMarks = root->rowMarks;
849 /* And last, tack on a ModifyTable node to do the UPDATE/DELETE work */
850 return (Plan *) make_modifytable(parse->commandType,
851 copyObject(root->resultRelations),
855 SS_assign_special_param(root));
858 /*--------------------
860 * Perform planning steps related to grouping, aggregation, etc.
861 * This primarily means adding top-level processing to the basic
862 * query plan produced by query_planner.
864 * tuple_fraction is the fraction of tuples we expect will be retrieved
866 * tuple_fraction is interpreted as follows:
867 * 0: expect all tuples to be retrieved (normal case)
868 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
869 * from the plan to be retrieved
870 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
871 * expected to be retrieved (ie, a LIMIT specification)
873 * Returns a query plan. Also, root->query_pathkeys is returned as the
874 * actual output ordering of the plan (in pathkey format).
875 *--------------------
878 grouping_planner(PlannerInfo *root, double tuple_fraction)
880 Query *parse = root->parse;
881 List *tlist = parse->targetList;
882 int64 offset_est = 0;
884 double limit_tuples = -1.0;
886 List *current_pathkeys;
887 double dNumGroups = 0;
888 bool use_hashed_distinct = false;
889 bool tested_hashed_distinct = false;
891 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
892 if (parse->limitCount || parse->limitOffset)
894 tuple_fraction = preprocess_limit(root, tuple_fraction,
895 &offset_est, &count_est);
898 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
899 * estimate the effects of using a bounded sort.
901 if (count_est > 0 && offset_est >= 0)
902 limit_tuples = (double) count_est + (double) offset_est;
905 if (parse->setOperations)
907 List *set_sortclauses;
910 * If there's a top-level ORDER BY, assume we have to fetch all the
911 * tuples. This might be too simplistic given all the hackery below
912 * to possibly avoid the sort; but the odds of accurate estimates here
913 * are pretty low anyway.
915 if (parse->sortClause)
916 tuple_fraction = 0.0;
919 * Construct the plan for set operations. The result will not need
920 * any work except perhaps a top-level sort and/or LIMIT. Note that
921 * any special work for recursive unions is the responsibility of
922 * plan_set_operations.
924 result_plan = plan_set_operations(root, tuple_fraction,
928 * Calculate pathkeys representing the sort order (if any) of the set
929 * operation's result. We have to do this before overwriting the sort
932 current_pathkeys = make_pathkeys_for_sortclauses(root,
934 result_plan->targetlist,
938 * We should not need to call preprocess_targetlist, since we must be
939 * in a SELECT query node. Instead, use the targetlist returned by
940 * plan_set_operations (since this tells whether it returned any
941 * resjunk columns!), and transfer any sort key information from the
944 Assert(parse->commandType == CMD_SELECT);
946 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
950 * Can't handle FOR UPDATE/SHARE here (parser should have checked
951 * already, but let's make sure).
955 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
956 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
959 * Calculate pathkeys that represent result ordering requirements
961 Assert(parse->distinctClause == NIL);
962 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
969 /* No set operations, do regular planning */
971 double sub_limit_tuples;
972 AttrNumber *groupColIdx = NULL;
973 bool need_tlist_eval = true;
979 AggClauseCounts agg_counts;
983 bool use_hashed_grouping = false;
984 WindowFuncLists *wflists = NULL;
985 List *activeWindows = NIL;
987 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
989 /* A recursive query should always have setOperations */
990 Assert(!root->hasRecursion);
992 /* Preprocess GROUP BY clause, if any */
993 if (parse->groupClause)
994 preprocess_groupclause(root);
995 numGroupCols = list_length(parse->groupClause);
997 /* Preprocess targetlist */
998 tlist = preprocess_targetlist(root, tlist);
1001 * Locate any window functions in the tlist. (We don't need to look
1002 * anywhere else, since expressions used in ORDER BY will be in there
1003 * too.) Note that they could all have been eliminated by constant
1004 * folding, in which case we don't need to do any more work.
1006 if (parse->hasWindowFuncs)
1008 wflists = find_window_functions((Node *) tlist,
1009 list_length(parse->windowClause));
1010 if (wflists->numWindowFuncs > 0)
1011 activeWindows = select_active_windows(root, wflists);
1013 parse->hasWindowFuncs = false;
1017 * Generate appropriate target list for subplan; may be different from
1018 * tlist if grouping or aggregation is needed.
1020 sub_tlist = make_subplanTargetList(root, tlist,
1021 &groupColIdx, &need_tlist_eval);
1024 * Do aggregate preprocessing, if the query has any aggs.
1026 * Note: think not that we can turn off hasAggs if we find no aggs. It
1027 * is possible for constant-expression simplification to remove all
1028 * explicit references to aggs, but we still have to follow the
1029 * aggregate semantics (eg, producing only one output row).
1034 * Will need actual number of aggregates for estimating costs.
1035 * Note: we do not attempt to detect duplicate aggregates here; a
1036 * somewhat-overestimated count is okay for our present purposes.
1038 count_agg_clauses((Node *) tlist, &agg_counts);
1039 count_agg_clauses(parse->havingQual, &agg_counts);
1042 * Preprocess MIN/MAX aggregates, if any.
1044 preprocess_minmax_aggregates(root, tlist);
1048 * Calculate pathkeys that represent grouping/ordering requirements.
1049 * Stash them in PlannerInfo so that query_planner can canonicalize
1050 * them after EquivalenceClasses have been formed. The sortClause is
1051 * certainly sort-able, but GROUP BY and DISTINCT might not be, in
1052 * which case we just leave their pathkeys empty.
1054 if (parse->groupClause &&
1055 grouping_is_sortable(parse->groupClause))
1056 root->group_pathkeys =
1057 make_pathkeys_for_sortclauses(root,
1062 root->group_pathkeys = NIL;
1064 /* We consider only the first (bottom) window in pathkeys logic */
1065 if (activeWindows != NIL)
1067 WindowClause *wc = (WindowClause *) linitial(activeWindows);
1069 root->window_pathkeys = make_pathkeys_for_window(root,
1075 root->window_pathkeys = NIL;
1077 if (parse->distinctClause &&
1078 grouping_is_sortable(parse->distinctClause))
1079 root->distinct_pathkeys =
1080 make_pathkeys_for_sortclauses(root,
1081 parse->distinctClause,
1085 root->distinct_pathkeys = NIL;
1087 root->sort_pathkeys =
1088 make_pathkeys_for_sortclauses(root,
1094 * Figure out whether we want a sorted result from query_planner.
1096 * If we have a sortable GROUP BY clause, then we want a result sorted
1097 * properly for grouping. Otherwise, if we have window functions to
1098 * evaluate, we try to sort for the first window. Otherwise, if
1099 * there's a sortable DISTINCT clause that's more rigorous than the
1100 * ORDER BY clause, we try to produce output that's sufficiently well
1101 * sorted for the DISTINCT. Otherwise, if there is an ORDER BY
1102 * clause, we want to sort by the ORDER BY clause.
1104 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
1105 * superset of GROUP BY, it would be tempting to request sort by ORDER
1106 * BY --- but that might just leave us failing to exploit an available
1107 * sort order at all. Needs more thought. The choice for DISTINCT
1108 * versus ORDER BY is much easier, since we know that the parser
1109 * ensured that one is a superset of the other.
1111 if (root->group_pathkeys)
1112 root->query_pathkeys = root->group_pathkeys;
1113 else if (root->window_pathkeys)
1114 root->query_pathkeys = root->window_pathkeys;
1115 else if (list_length(root->distinct_pathkeys) >
1116 list_length(root->sort_pathkeys))
1117 root->query_pathkeys = root->distinct_pathkeys;
1118 else if (root->sort_pathkeys)
1119 root->query_pathkeys = root->sort_pathkeys;
1121 root->query_pathkeys = NIL;
1124 * Figure out whether there's a hard limit on the number of rows that
1125 * query_planner's result subplan needs to return. Even if we know a
1126 * hard limit overall, it doesn't apply if the query has any
1127 * grouping/aggregation operations.
1129 if (parse->groupClause ||
1130 parse->distinctClause ||
1132 parse->hasWindowFuncs ||
1133 root->hasHavingQual)
1134 sub_limit_tuples = -1.0;
1136 sub_limit_tuples = limit_tuples;
1139 * Generate the best unsorted and presorted paths for this Query (but
1140 * note there may not be any presorted path). query_planner will also
1141 * estimate the number of groups in the query, and canonicalize all
1144 query_planner(root, sub_tlist, tuple_fraction, sub_limit_tuples,
1145 &cheapest_path, &sorted_path, &dNumGroups);
1148 * Extract rowcount and width estimates for possible use in grouping
1149 * decisions. Beware here of the possibility that
1150 * cheapest_path->parent is NULL (ie, there is no FROM clause).
1152 if (cheapest_path->parent)
1154 path_rows = cheapest_path->parent->rows;
1155 path_width = cheapest_path->parent->width;
1159 path_rows = 1; /* assume non-set result */
1160 path_width = 100; /* arbitrary */
1163 if (parse->groupClause)
1166 * If grouping, decide whether to use sorted or hashed grouping.
1168 use_hashed_grouping =
1169 choose_hashed_grouping(root,
1170 tuple_fraction, limit_tuples,
1171 path_rows, path_width,
1172 cheapest_path, sorted_path,
1173 dNumGroups, &agg_counts);
1174 /* Also convert # groups to long int --- but 'ware overflow! */
1175 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
1177 else if (parse->distinctClause && sorted_path &&
1178 !root->hasHavingQual && !parse->hasAggs && !activeWindows)
1181 * We'll reach the DISTINCT stage without any intermediate
1182 * processing, so figure out whether we will want to hash or not
1183 * so we can choose whether to use cheapest or sorted path.
1185 use_hashed_distinct =
1186 choose_hashed_distinct(root,
1187 tuple_fraction, limit_tuples,
1188 path_rows, path_width,
1189 cheapest_path->startup_cost,
1190 cheapest_path->total_cost,
1191 sorted_path->startup_cost,
1192 sorted_path->total_cost,
1193 sorted_path->pathkeys,
1195 tested_hashed_distinct = true;
1199 * Select the best path. If we are doing hashed grouping, we will
1200 * always read all the input tuples, so use the cheapest-total path.
1201 * Otherwise, trust query_planner's decision about which to use.
1203 if (use_hashed_grouping || use_hashed_distinct || !sorted_path)
1204 best_path = cheapest_path;
1206 best_path = sorted_path;
1209 * Check to see if it's possible to optimize MIN/MAX aggregates. If
1210 * so, we will forget all the work we did so far to choose a "regular"
1211 * path ... but we had to do it anyway to be able to tell which way is
1214 result_plan = optimize_minmax_aggregates(root,
1217 if (result_plan != NULL)
1220 * optimize_minmax_aggregates generated the full plan, with the
1221 * right tlist, and it has no sort order.
1223 current_pathkeys = NIL;
1228 * Normal case --- create a plan according to query_planner's
1231 bool need_sort_for_grouping = false;
1233 result_plan = create_plan(root, best_path);
1234 current_pathkeys = best_path->pathkeys;
1236 /* Detect if we'll need an explicit sort for grouping */
1237 if (parse->groupClause && !use_hashed_grouping &&
1238 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1240 need_sort_for_grouping = true;
1243 * Always override query_planner's tlist, so that we don't
1244 * sort useless data from a "physical" tlist.
1246 need_tlist_eval = true;
1250 * create_plan() returns a plan with just a "flat" tlist of
1251 * required Vars. Usually we need to insert the sub_tlist as the
1252 * tlist of the top plan node. However, we can skip that if we
1253 * determined that whatever query_planner chose to return will be
1256 if (need_tlist_eval)
1259 * If the top-level plan node is one that cannot do expression
1260 * evaluation, we must insert a Result node to project the
1263 if (!is_projection_capable_plan(result_plan))
1265 result_plan = (Plan *) make_result(root,
1273 * Otherwise, just replace the subplan's flat tlist with
1274 * the desired tlist.
1276 result_plan->targetlist = sub_tlist;
1280 * Also, account for the cost of evaluation of the sub_tlist.
1282 * Up to now, we have only been dealing with "flat" tlists,
1283 * containing just Vars. So their evaluation cost is zero
1284 * according to the model used by cost_qual_eval() (or if you
1285 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1286 * can avoid accounting for tlist cost throughout
1287 * query_planner() and subroutines. But now we've inserted a
1288 * tlist that might contain actual operators, sub-selects, etc
1289 * --- so we'd better account for its cost.
1291 * Below this point, any tlist eval cost for added-on nodes
1292 * should be accounted for as we create those nodes.
1293 * Presently, of the node types we can add on, only Agg,
1294 * WindowAgg, and Group project new tlists (the rest just copy
1295 * their input tuples) --- so make_agg(), make_windowagg() and
1296 * make_group() are responsible for computing the added cost.
1298 cost_qual_eval(&tlist_cost, sub_tlist, root);
1299 result_plan->startup_cost += tlist_cost.startup;
1300 result_plan->total_cost += tlist_cost.startup +
1301 tlist_cost.per_tuple * result_plan->plan_rows;
1306 * Since we're using query_planner's tlist and not the one
1307 * make_subplanTargetList calculated, we have to refigure any
1308 * grouping-column indexes make_subplanTargetList computed.
1310 locate_grouping_columns(root, tlist, result_plan->targetlist,
1315 * Insert AGG or GROUP node if needed, plus an explicit sort step
1318 * HAVING clause, if any, becomes qual of the Agg or Group node.
1320 if (use_hashed_grouping)
1322 /* Hashed aggregate plan --- no sort needed */
1323 result_plan = (Plan *) make_agg(root,
1325 (List *) parse->havingQual,
1329 extract_grouping_ops(parse->groupClause),
1333 /* Hashed aggregation produces randomly-ordered results */
1334 current_pathkeys = NIL;
1336 else if (parse->hasAggs)
1338 /* Plain aggregate plan --- sort if needed */
1339 AggStrategy aggstrategy;
1341 if (parse->groupClause)
1343 if (need_sort_for_grouping)
1345 result_plan = (Plan *)
1346 make_sort_from_groupcols(root,
1350 current_pathkeys = root->group_pathkeys;
1352 aggstrategy = AGG_SORTED;
1355 * The AGG node will not change the sort ordering of its
1356 * groups, so current_pathkeys describes the result too.
1361 aggstrategy = AGG_PLAIN;
1362 /* Result will be only one row anyway; no sort order */
1363 current_pathkeys = NIL;
1366 result_plan = (Plan *) make_agg(root,
1368 (List *) parse->havingQual,
1372 extract_grouping_ops(parse->groupClause),
1377 else if (parse->groupClause)
1380 * GROUP BY without aggregation, so insert a group node (plus
1381 * the appropriate sort node, if necessary).
1383 * Add an explicit sort if we couldn't make the path come out
1384 * the way the GROUP node needs it.
1386 if (need_sort_for_grouping)
1388 result_plan = (Plan *)
1389 make_sort_from_groupcols(root,
1393 current_pathkeys = root->group_pathkeys;
1396 result_plan = (Plan *) make_group(root,
1398 (List *) parse->havingQual,
1401 extract_grouping_ops(parse->groupClause),
1404 /* The Group node won't change sort ordering */
1406 else if (root->hasHavingQual)
1409 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1410 * This is a degenerate case in which we are supposed to emit
1411 * either 0 or 1 row depending on whether HAVING succeeds.
1412 * Furthermore, there cannot be any variables in either HAVING
1413 * or the targetlist, so we actually do not need the FROM
1414 * table at all! We can just throw away the plan-so-far and
1415 * generate a Result node. This is a sufficiently unusual
1416 * corner case that it's not worth contorting the structure of
1417 * this routine to avoid having to generate the plan in the
1420 result_plan = (Plan *) make_result(root,
1425 } /* end of non-minmax-aggregate case */
1428 * Since each window function could require a different sort order, we
1429 * stack up a WindowAgg node for each window, with sort steps between
1438 * If the top-level plan node is one that cannot do expression
1439 * evaluation, we must insert a Result node to project the desired
1440 * tlist. (In some cases this might not really be required, but
1441 * it's not worth trying to avoid it.) Note that on second and
1442 * subsequent passes through the following loop, the top-level
1443 * node will be a WindowAgg which we know can project; so we only
1444 * need to check once.
1446 if (!is_projection_capable_plan(result_plan))
1448 result_plan = (Plan *) make_result(root,
1455 * The "base" targetlist for all steps of the windowing process is
1456 * a flat tlist of all Vars and Aggs needed in the result. (In
1457 * some cases we wouldn't need to propagate all of these all the
1458 * way to the top, since they might only be needed as inputs to
1459 * WindowFuncs. It's probably not worth trying to optimize that
1460 * though.) We also need any volatile sort expressions, because
1461 * make_sort_from_pathkeys won't add those on its own, and anyway
1462 * we want them evaluated only once at the bottom of the stack. As
1463 * we climb up the stack, we add outputs for the WindowFuncs
1464 * computed at each level. Also, each input tlist has to present
1465 * all the columns needed to sort the data for the next WindowAgg
1466 * step. That's handled internally by make_sort_from_pathkeys,
1467 * but we need the copyObject steps here to ensure that each plan
1468 * node has a separately modifiable tlist.
1470 window_tlist = flatten_tlist(tlist);
1472 window_tlist = add_to_flat_tlist(window_tlist,
1473 pull_agg_clause((Node *) tlist));
1474 window_tlist = add_volatile_sort_exprs(window_tlist, tlist,
1476 result_plan->targetlist = (List *) copyObject(window_tlist);
1478 foreach(l, activeWindows)
1480 WindowClause *wc = (WindowClause *) lfirst(l);
1481 List *window_pathkeys;
1483 AttrNumber *partColIdx;
1486 AttrNumber *ordColIdx;
1489 window_pathkeys = make_pathkeys_for_window(root,
1495 * This is a bit tricky: we build a sort node even if we don't
1496 * really have to sort. Even when no explicit sort is needed,
1497 * we need to have suitable resjunk items added to the input
1498 * plan's tlist for any partitioning or ordering columns that
1499 * aren't plain Vars. Furthermore, this way we can use
1500 * existing infrastructure to identify which input columns are
1501 * the interesting ones.
1503 if (window_pathkeys)
1507 sort_plan = make_sort_from_pathkeys(root,
1511 if (!pathkeys_contained_in(window_pathkeys,
1514 /* we do indeed need to sort */
1515 result_plan = (Plan *) sort_plan;
1516 current_pathkeys = window_pathkeys;
1518 /* In either case, extract the per-column information */
1519 get_column_info_for_window(root, wc, tlist,
1521 sort_plan->sortColIdx,
1531 /* empty window specification, nothing to sort */
1534 partOperators = NULL;
1537 ordOperators = NULL;
1542 /* Add the current WindowFuncs to the running tlist */
1543 window_tlist = add_to_flat_tlist(window_tlist,
1544 wflists->windowFuncs[wc->winref]);
1548 /* Install the original tlist in the topmost WindowAgg */
1549 window_tlist = tlist;
1552 /* ... and make the WindowAgg plan node */
1553 result_plan = (Plan *)
1554 make_windowagg(root,
1555 (List *) copyObject(window_tlist),
1556 list_length(wflists->windowFuncs[wc->winref]),
1570 } /* end of if (setOperations) */
1573 * If there is a DISTINCT clause, add the necessary node(s).
1575 if (parse->distinctClause)
1577 double dNumDistinctRows;
1578 long numDistinctRows;
1581 * If there was grouping or aggregation, use the current number of
1582 * rows as the estimated number of DISTINCT rows (ie, assume the
1583 * result was already mostly unique). If not, use the number of
1584 * distinct-groups calculated by query_planner.
1586 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1587 dNumDistinctRows = result_plan->plan_rows;
1589 dNumDistinctRows = dNumGroups;
1591 /* Also convert to long int --- but 'ware overflow! */
1592 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1594 /* Choose implementation method if we didn't already */
1595 if (!tested_hashed_distinct)
1598 * At this point, either hashed or sorted grouping will have to
1599 * work from result_plan, so we pass that as both "cheapest" and
1602 use_hashed_distinct =
1603 choose_hashed_distinct(root,
1604 tuple_fraction, limit_tuples,
1605 result_plan->plan_rows,
1606 result_plan->plan_width,
1607 result_plan->startup_cost,
1608 result_plan->total_cost,
1609 result_plan->startup_cost,
1610 result_plan->total_cost,
1615 if (use_hashed_distinct)
1617 /* Hashed aggregate plan --- no sort needed */
1618 result_plan = (Plan *) make_agg(root,
1619 result_plan->targetlist,
1622 list_length(parse->distinctClause),
1623 extract_grouping_cols(parse->distinctClause,
1624 result_plan->targetlist),
1625 extract_grouping_ops(parse->distinctClause),
1629 /* Hashed aggregation produces randomly-ordered results */
1630 current_pathkeys = NIL;
1635 * Use a Unique node to implement DISTINCT. Add an explicit sort
1636 * if we couldn't make the path come out the way the Unique node
1637 * needs it. If we do have to sort, always sort by the more
1638 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1639 * below. However, for regular DISTINCT, don't sort now if we
1640 * don't have to --- sorting afterwards will likely be cheaper,
1641 * and also has the possibility of optimizing via LIMIT. But for
1642 * DISTINCT ON, we *must* force the final sort now, else it won't
1643 * have the desired behavior.
1645 List *needed_pathkeys;
1647 if (parse->hasDistinctOn &&
1648 list_length(root->distinct_pathkeys) <
1649 list_length(root->sort_pathkeys))
1650 needed_pathkeys = root->sort_pathkeys;
1652 needed_pathkeys = root->distinct_pathkeys;
1654 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1656 if (list_length(root->distinct_pathkeys) >=
1657 list_length(root->sort_pathkeys))
1658 current_pathkeys = root->distinct_pathkeys;
1661 current_pathkeys = root->sort_pathkeys;
1662 /* Assert checks that parser didn't mess up... */
1663 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1667 result_plan = (Plan *) make_sort_from_pathkeys(root,
1673 result_plan = (Plan *) make_unique(result_plan,
1674 parse->distinctClause);
1675 result_plan->plan_rows = dNumDistinctRows;
1676 /* The Unique node won't change sort ordering */
1681 * If ORDER BY was given and we were not able to make the plan come out in
1682 * the right order, add an explicit sort step.
1684 if (parse->sortClause)
1686 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1688 result_plan = (Plan *) make_sort_from_pathkeys(root,
1690 root->sort_pathkeys,
1692 current_pathkeys = root->sort_pathkeys;
1697 * If there is a FOR UPDATE/SHARE clause, add the LockRows node. (Note: we
1698 * intentionally test parse->rowMarks not root->rowMarks here. If there
1699 * are only non-locking rowmarks, they should be handled by the
1700 * ModifyTable node instead.)
1702 if (parse->rowMarks)
1704 result_plan = (Plan *) make_lockrows(result_plan,
1706 SS_assign_special_param(root));
1709 * The result can no longer be assumed sorted, since locking might
1710 * cause the sort key columns to be replaced with new values.
1712 current_pathkeys = NIL;
1716 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1718 if (parse->limitCount || parse->limitOffset)
1720 result_plan = (Plan *) make_limit(result_plan,
1727 /* Compute result-relations list if needed */
1728 if (parse->resultRelation)
1729 root->resultRelations = list_make1_int(parse->resultRelation);
1731 root->resultRelations = NIL;
1734 * Return the actual output ordering in query_pathkeys for possible use by
1735 * an outer query level.
1737 root->query_pathkeys = current_pathkeys;
1743 * Detect whether a plan node is a "dummy" plan created when a relation
1744 * is deemed not to need scanning due to constraint exclusion.
1746 * Currently, such dummy plans are Result nodes with constant FALSE
1750 is_dummy_plan(Plan *plan)
1752 if (IsA(plan, Result))
1754 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1756 if (list_length(rcqual) == 1)
1758 Const *constqual = (Const *) linitial(rcqual);
1760 if (constqual && IsA(constqual, Const))
1762 if (!constqual->constisnull &&
1763 !DatumGetBool(constqual->constvalue))
1772 * Create a bitmapset of the RT indexes of live base relations
1774 * Helper for preprocess_rowmarks ... at this point in the proceedings,
1775 * the only good way to distinguish baserels from appendrel children
1776 * is to see what is in the join tree.
1779 get_base_rel_indexes(Node *jtnode)
1785 if (IsA(jtnode, RangeTblRef))
1787 int varno = ((RangeTblRef *) jtnode)->rtindex;
1789 result = bms_make_singleton(varno);
1791 else if (IsA(jtnode, FromExpr))
1793 FromExpr *f = (FromExpr *) jtnode;
1797 foreach(l, f->fromlist)
1798 result = bms_join(result,
1799 get_base_rel_indexes(lfirst(l)));
1801 else if (IsA(jtnode, JoinExpr))
1803 JoinExpr *j = (JoinExpr *) jtnode;
1805 result = bms_join(get_base_rel_indexes(j->larg),
1806 get_base_rel_indexes(j->rarg));
1810 elog(ERROR, "unrecognized node type: %d",
1811 (int) nodeTag(jtnode));
1812 result = NULL; /* keep compiler quiet */
1818 * preprocess_rowmarks - set up PlanRowMarks if needed
1821 preprocess_rowmarks(PlannerInfo *root)
1823 Query *parse = root->parse;
1829 if (parse->rowMarks)
1832 * We've got trouble if FOR UPDATE/SHARE appears inside grouping,
1833 * since grouping renders a reference to individual tuple CTIDs
1834 * invalid. This is also checked at parse time, but that's
1835 * insufficient because of rule substitution, query pullup, etc.
1837 CheckSelectLocking(parse);
1842 * We only need rowmarks for UPDATE, DELETE, or FOR UPDATE/SHARE.
1844 if (parse->commandType != CMD_UPDATE &&
1845 parse->commandType != CMD_DELETE)
1850 * We need to have rowmarks for all base relations except the target. We
1851 * make a bitmapset of all base rels and then remove the items we don't
1852 * need or have FOR UPDATE/SHARE marks for.
1854 rels = get_base_rel_indexes((Node *) parse->jointree);
1855 if (parse->resultRelation)
1856 rels = bms_del_member(rels, parse->resultRelation);
1859 * Convert RowMarkClauses to PlanRowMark representation.
1862 foreach(l, parse->rowMarks)
1864 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
1865 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
1869 * Currently, it is syntactically impossible to have FOR UPDATE
1870 * applied to an update/delete target rel. If that ever becomes
1871 * possible, we should drop the target from the PlanRowMark list.
1873 Assert(rc->rti != parse->resultRelation);
1876 * Ignore RowMarkClauses for subqueries; they aren't real tables and
1877 * can't support true locking. Subqueries that got flattened into the
1878 * main query should be ignored completely. Any that didn't will get
1879 * ROW_MARK_COPY items in the next loop.
1881 if (rte->rtekind != RTE_RELATION)
1884 rels = bms_del_member(rels, rc->rti);
1886 newrc = makeNode(PlanRowMark);
1887 newrc->rti = newrc->prti = rc->rti;
1889 newrc->markType = ROW_MARK_EXCLUSIVE;
1891 newrc->markType = ROW_MARK_SHARE;
1892 newrc->noWait = rc->noWait;
1893 newrc->isParent = false;
1895 prowmarks = lappend(prowmarks, newrc);
1899 * Now, add rowmarks for any non-target, non-locked base relations.
1902 foreach(l, parse->rtable)
1904 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
1908 if (!bms_is_member(i, rels))
1911 newrc = makeNode(PlanRowMark);
1912 newrc->rti = newrc->prti = i;
1913 /* real tables support REFERENCE, anything else needs COPY */
1914 if (rte->rtekind == RTE_RELATION)
1915 newrc->markType = ROW_MARK_REFERENCE;
1917 newrc->markType = ROW_MARK_COPY;
1918 newrc->noWait = false; /* doesn't matter */
1919 newrc->isParent = false;
1921 prowmarks = lappend(prowmarks, newrc);
1924 root->rowMarks = prowmarks;
1928 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1930 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1931 * results back in *count_est and *offset_est. These variables are set to
1932 * 0 if the corresponding clause is not present, and -1 if it's present
1933 * but we couldn't estimate the value for it. (The "0" convention is OK
1934 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1935 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1936 * usual practice of never estimating less than one row.) These values will
1937 * be passed to make_limit, which see if you change this code.
1939 * The return value is the suitably adjusted tuple_fraction to use for
1940 * planning the query. This adjustment is not overridable, since it reflects
1941 * plan actions that grouping_planner() will certainly take, not assumptions
1945 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1946 int64 *offset_est, int64 *count_est)
1948 Query *parse = root->parse;
1950 double limit_fraction;
1952 /* Should not be called unless LIMIT or OFFSET */
1953 Assert(parse->limitCount || parse->limitOffset);
1956 * Try to obtain the clause values. We use estimate_expression_value
1957 * primarily because it can sometimes do something useful with Params.
1959 if (parse->limitCount)
1961 est = estimate_expression_value(root, parse->limitCount);
1962 if (est && IsA(est, Const))
1964 if (((Const *) est)->constisnull)
1966 /* NULL indicates LIMIT ALL, ie, no limit */
1967 *count_est = 0; /* treat as not present */
1971 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1972 if (*count_est <= 0)
1973 *count_est = 1; /* force to at least 1 */
1977 *count_est = -1; /* can't estimate */
1980 *count_est = 0; /* not present */
1982 if (parse->limitOffset)
1984 est = estimate_expression_value(root, parse->limitOffset);
1985 if (est && IsA(est, Const))
1987 if (((Const *) est)->constisnull)
1989 /* Treat NULL as no offset; the executor will too */
1990 *offset_est = 0; /* treat as not present */
1994 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1995 if (*offset_est < 0)
1996 *offset_est = 0; /* less than 0 is same as 0 */
2000 *offset_est = -1; /* can't estimate */
2003 *offset_est = 0; /* not present */
2005 if (*count_est != 0)
2008 * A LIMIT clause limits the absolute number of tuples returned.
2009 * However, if it's not a constant LIMIT then we have to guess; for
2010 * lack of a better idea, assume 10% of the plan's result is wanted.
2012 if (*count_est < 0 || *offset_est < 0)
2014 /* LIMIT or OFFSET is an expression ... punt ... */
2015 limit_fraction = 0.10;
2019 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2020 limit_fraction = (double) *count_est + (double) *offset_est;
2024 * If we have absolute limits from both caller and LIMIT, use the
2025 * smaller value; likewise if they are both fractional. If one is
2026 * fractional and the other absolute, we can't easily determine which
2027 * is smaller, but we use the heuristic that the absolute will usually
2030 if (tuple_fraction >= 1.0)
2032 if (limit_fraction >= 1.0)
2035 tuple_fraction = Min(tuple_fraction, limit_fraction);
2039 /* caller absolute, limit fractional; use caller's value */
2042 else if (tuple_fraction > 0.0)
2044 if (limit_fraction >= 1.0)
2046 /* caller fractional, limit absolute; use limit */
2047 tuple_fraction = limit_fraction;
2051 /* both fractional */
2052 tuple_fraction = Min(tuple_fraction, limit_fraction);
2057 /* no info from caller, just use limit */
2058 tuple_fraction = limit_fraction;
2061 else if (*offset_est != 0 && tuple_fraction > 0.0)
2064 * We have an OFFSET but no LIMIT. This acts entirely differently
2065 * from the LIMIT case: here, we need to increase rather than decrease
2066 * the caller's tuple_fraction, because the OFFSET acts to cause more
2067 * tuples to be fetched instead of fewer. This only matters if we got
2068 * a tuple_fraction > 0, however.
2070 * As above, use 10% if OFFSET is present but unestimatable.
2072 if (*offset_est < 0)
2073 limit_fraction = 0.10;
2075 limit_fraction = (double) *offset_est;
2078 * If we have absolute counts from both caller and OFFSET, add them
2079 * together; likewise if they are both fractional. If one is
2080 * fractional and the other absolute, we want to take the larger, and
2081 * we heuristically assume that's the fractional one.
2083 if (tuple_fraction >= 1.0)
2085 if (limit_fraction >= 1.0)
2087 /* both absolute, so add them together */
2088 tuple_fraction += limit_fraction;
2092 /* caller absolute, limit fractional; use limit */
2093 tuple_fraction = limit_fraction;
2098 if (limit_fraction >= 1.0)
2100 /* caller fractional, limit absolute; use caller's value */
2104 /* both fractional, so add them together */
2105 tuple_fraction += limit_fraction;
2106 if (tuple_fraction >= 1.0)
2107 tuple_fraction = 0.0; /* assume fetch all */
2112 return tuple_fraction;
2117 * preprocess_groupclause - do preparatory work on GROUP BY clause
2119 * The idea here is to adjust the ordering of the GROUP BY elements
2120 * (which in itself is semantically insignificant) to match ORDER BY,
2121 * thereby allowing a single sort operation to both implement the ORDER BY
2122 * requirement and set up for a Unique step that implements GROUP BY.
2124 * In principle it might be interesting to consider other orderings of the
2125 * GROUP BY elements, which could match the sort ordering of other
2126 * possible plans (eg an indexscan) and thereby reduce cost. We don't
2127 * bother with that, though. Hashed grouping will frequently win anyway.
2129 * Note: we need no comparable processing of the distinctClause because
2130 * the parser already enforced that that matches ORDER BY.
2133 preprocess_groupclause(PlannerInfo *root)
2135 Query *parse = root->parse;
2136 List *new_groupclause;
2141 /* If no ORDER BY, nothing useful to do here */
2142 if (parse->sortClause == NIL)
2146 * Scan the ORDER BY clause and construct a list of matching GROUP BY
2147 * items, but only as far as we can make a matching prefix.
2149 * This code assumes that the sortClause contains no duplicate items.
2151 new_groupclause = NIL;
2152 foreach(sl, parse->sortClause)
2154 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
2156 foreach(gl, parse->groupClause)
2158 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2162 new_groupclause = lappend(new_groupclause, gc);
2167 break; /* no match, so stop scanning */
2170 /* Did we match all of the ORDER BY list, or just some of it? */
2171 partial_match = (sl != NULL);
2173 /* If no match at all, no point in reordering GROUP BY */
2174 if (new_groupclause == NIL)
2178 * Add any remaining GROUP BY items to the new list, but only if we were
2179 * able to make a complete match. In other words, we only rearrange the
2180 * GROUP BY list if the result is that one list is a prefix of the other
2181 * --- otherwise there's no possibility of a common sort. Also, give up
2182 * if there are any non-sortable GROUP BY items, since then there's no
2185 foreach(gl, parse->groupClause)
2187 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2189 if (list_member_ptr(new_groupclause, gc))
2190 continue; /* it matched an ORDER BY item */
2192 return; /* give up, no common sort possible */
2193 if (!OidIsValid(gc->sortop))
2194 return; /* give up, GROUP BY can't be sorted */
2195 new_groupclause = lappend(new_groupclause, gc);
2198 /* Success --- install the rearranged GROUP BY list */
2199 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2200 parse->groupClause = new_groupclause;
2204 * choose_hashed_grouping - should we use hashed grouping?
2206 * Returns TRUE to select hashing, FALSE to select sorting.
2209 choose_hashed_grouping(PlannerInfo *root,
2210 double tuple_fraction, double limit_tuples,
2211 double path_rows, int path_width,
2212 Path *cheapest_path, Path *sorted_path,
2213 double dNumGroups, AggClauseCounts *agg_counts)
2215 Query *parse = root->parse;
2216 int numGroupCols = list_length(parse->groupClause);
2220 List *target_pathkeys;
2221 List *current_pathkeys;
2226 * Executor doesn't support hashed aggregation with DISTINCT or ORDER BY
2227 * aggregates. (Doing so would imply storing *all* the input values in
2228 * the hash table, and/or running many sorts in parallel, either of which
2229 * seems like a certain loser.)
2231 can_hash = (agg_counts->numOrderedAggs == 0 &&
2232 grouping_is_hashable(parse->groupClause));
2233 can_sort = grouping_is_sortable(parse->groupClause);
2235 /* Quick out if only one choice is workable */
2236 if (!(can_hash && can_sort))
2244 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2245 errmsg("could not implement GROUP BY"),
2246 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2249 /* Prefer sorting when enable_hashagg is off */
2250 if (!enable_hashagg)
2254 * Don't do it if it doesn't look like the hashtable will fit into
2258 /* Estimate per-hash-entry space at tuple width... */
2259 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2260 /* plus space for pass-by-ref transition values... */
2261 hashentrysize += agg_counts->transitionSpace;
2262 /* plus the per-hash-entry overhead */
2263 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
2265 if (hashentrysize * dNumGroups > work_mem * 1024L)
2269 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
2270 * DISTINCT and ORDER BY as the assumed required output sort order. This
2271 * is an oversimplification because the DISTINCT might get implemented via
2272 * hashing, but it's not clear that the case is common enough (or that our
2273 * estimates are good enough) to justify trying to solve it exactly.
2275 if (list_length(root->distinct_pathkeys) >
2276 list_length(root->sort_pathkeys))
2277 target_pathkeys = root->distinct_pathkeys;
2279 target_pathkeys = root->sort_pathkeys;
2282 * See if the estimated cost is no more than doing it the other way. While
2283 * avoiding the need for sorted input is usually a win, the fact that the
2284 * output won't be sorted may be a loss; so we need to do an actual cost
2287 * We need to consider cheapest_path + hashagg [+ final sort] versus
2288 * either cheapest_path [+ sort] + group or agg [+ final sort] or
2289 * presorted_path + group or agg [+ final sort] where brackets indicate a
2290 * step that may not be needed. We assume query_planner() will have
2291 * returned a presorted path only if it's a winner compared to
2292 * cheapest_path for this purpose.
2294 * These path variables are dummies that just hold cost fields; we don't
2295 * make actual Paths for these steps.
2297 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
2298 numGroupCols, dNumGroups,
2299 cheapest_path->startup_cost, cheapest_path->total_cost,
2301 /* Result of hashed agg is always unsorted */
2302 if (target_pathkeys)
2303 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
2304 dNumGroups, path_width,
2305 0.0, work_mem, limit_tuples);
2309 sorted_p.startup_cost = sorted_path->startup_cost;
2310 sorted_p.total_cost = sorted_path->total_cost;
2311 current_pathkeys = sorted_path->pathkeys;
2315 sorted_p.startup_cost = cheapest_path->startup_cost;
2316 sorted_p.total_cost = cheapest_path->total_cost;
2317 current_pathkeys = cheapest_path->pathkeys;
2319 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
2321 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
2322 path_rows, path_width,
2323 0.0, work_mem, -1.0);
2324 current_pathkeys = root->group_pathkeys;
2328 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
2329 numGroupCols, dNumGroups,
2330 sorted_p.startup_cost, sorted_p.total_cost,
2333 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
2334 sorted_p.startup_cost, sorted_p.total_cost,
2336 /* The Agg or Group node will preserve ordering */
2337 if (target_pathkeys &&
2338 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
2339 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
2340 dNumGroups, path_width,
2341 0.0, work_mem, limit_tuples);
2344 * Now make the decision using the top-level tuple fraction. First we
2345 * have to convert an absolute count (LIMIT) into fractional form.
2347 if (tuple_fraction >= 1.0)
2348 tuple_fraction /= dNumGroups;
2350 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2351 tuple_fraction) < 0)
2353 /* Hashed is cheaper, so use it */
2360 * choose_hashed_distinct - should we use hashing for DISTINCT?
2362 * This is fairly similar to choose_hashed_grouping, but there are enough
2363 * differences that it doesn't seem worth trying to unify the two functions.
2364 * (One difference is that we sometimes apply this after forming a Plan,
2365 * so the input alternatives can't be represented as Paths --- instead we
2366 * pass in the costs as individual variables.)
2368 * But note that making the two choices independently is a bit bogus in
2369 * itself. If the two could be combined into a single choice operation
2370 * it'd probably be better, but that seems far too unwieldy to be practical,
2371 * especially considering that the combination of GROUP BY and DISTINCT
2372 * isn't very common in real queries. By separating them, we are giving
2373 * extra preference to using a sorting implementation when a common sort key
2374 * is available ... and that's not necessarily wrong anyway.
2376 * Returns TRUE to select hashing, FALSE to select sorting.
2379 choose_hashed_distinct(PlannerInfo *root,
2380 double tuple_fraction, double limit_tuples,
2381 double path_rows, int path_width,
2382 Cost cheapest_startup_cost, Cost cheapest_total_cost,
2383 Cost sorted_startup_cost, Cost sorted_total_cost,
2384 List *sorted_pathkeys,
2385 double dNumDistinctRows)
2387 Query *parse = root->parse;
2388 int numDistinctCols = list_length(parse->distinctClause);
2392 List *current_pathkeys;
2393 List *needed_pathkeys;
2398 * If we have a sortable DISTINCT ON clause, we always use sorting. This
2399 * enforces the expected behavior of DISTINCT ON.
2401 can_sort = grouping_is_sortable(parse->distinctClause);
2402 if (can_sort && parse->hasDistinctOn)
2405 can_hash = grouping_is_hashable(parse->distinctClause);
2407 /* Quick out if only one choice is workable */
2408 if (!(can_hash && can_sort))
2416 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2417 errmsg("could not implement DISTINCT"),
2418 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2421 /* Prefer sorting when enable_hashagg is off */
2422 if (!enable_hashagg)
2426 * Don't do it if it doesn't look like the hashtable will fit into
2429 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2431 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
2435 * See if the estimated cost is no more than doing it the other way. While
2436 * avoiding the need for sorted input is usually a win, the fact that the
2437 * output won't be sorted may be a loss; so we need to do an actual cost
2440 * We need to consider cheapest_path + hashagg [+ final sort] versus
2441 * sorted_path [+ sort] + group [+ final sort] where brackets indicate a
2442 * step that may not be needed.
2444 * These path variables are dummies that just hold cost fields; we don't
2445 * make actual Paths for these steps.
2447 cost_agg(&hashed_p, root, AGG_HASHED, 0,
2448 numDistinctCols, dNumDistinctRows,
2449 cheapest_startup_cost, cheapest_total_cost,
2453 * Result of hashed agg is always unsorted, so if ORDER BY is present we
2454 * need to charge for the final sort.
2456 if (parse->sortClause)
2457 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
2458 dNumDistinctRows, path_width,
2459 0.0, work_mem, limit_tuples);
2462 * Now for the GROUP case. See comments in grouping_planner about the
2463 * sorting choices here --- this code should match that code.
2465 sorted_p.startup_cost = sorted_startup_cost;
2466 sorted_p.total_cost = sorted_total_cost;
2467 current_pathkeys = sorted_pathkeys;
2468 if (parse->hasDistinctOn &&
2469 list_length(root->distinct_pathkeys) <
2470 list_length(root->sort_pathkeys))
2471 needed_pathkeys = root->sort_pathkeys;
2473 needed_pathkeys = root->distinct_pathkeys;
2474 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
2476 if (list_length(root->distinct_pathkeys) >=
2477 list_length(root->sort_pathkeys))
2478 current_pathkeys = root->distinct_pathkeys;
2480 current_pathkeys = root->sort_pathkeys;
2481 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
2482 path_rows, path_width,
2483 0.0, work_mem, -1.0);
2485 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
2486 sorted_p.startup_cost, sorted_p.total_cost,
2488 if (parse->sortClause &&
2489 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
2490 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
2491 dNumDistinctRows, path_width,
2492 0.0, work_mem, limit_tuples);
2495 * Now make the decision using the top-level tuple fraction. First we
2496 * have to convert an absolute count (LIMIT) into fractional form.
2498 if (tuple_fraction >= 1.0)
2499 tuple_fraction /= dNumDistinctRows;
2501 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2502 tuple_fraction) < 0)
2504 /* Hashed is cheaper, so use it */
2511 * make_subplanTargetList
2512 * Generate appropriate target list when grouping is required.
2514 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
2515 * above the result of query_planner, we typically want to pass a different
2516 * target list to query_planner than the outer plan nodes should have.
2517 * This routine generates the correct target list for the subplan.
2519 * The initial target list passed from the parser already contains entries
2520 * for all ORDER BY and GROUP BY expressions, but it will not have entries
2521 * for variables used only in HAVING clauses; so we need to add those
2522 * variables to the subplan target list. Also, we flatten all expressions
2523 * except GROUP BY items into their component variables; the other expressions
2524 * will be computed by the inserted nodes rather than by the subplan.
2525 * For example, given a query like
2526 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
2527 * we want to pass this targetlist to the subplan:
2529 * where the a+b target will be used by the Sort/Group steps, and the
2530 * other targets will be used for computing the final results. (In the
2531 * above example we could theoretically suppress the a and b targets and
2532 * pass down only c,d,a+b, but it's not really worth the trouble to
2533 * eliminate simple var references from the subplan. We will avoid doing
2534 * the extra computation to recompute a+b at the outer level; see
2535 * fix_upper_expr() in setrefs.c.)
2537 * If we are grouping or aggregating, *and* there are no non-Var grouping
2538 * expressions, then the returned tlist is effectively dummy; we do not
2539 * need to force it to be evaluated, because all the Vars it contains
2540 * should be present in the output of query_planner anyway.
2542 * 'tlist' is the query's target list.
2543 * 'groupColIdx' receives an array of column numbers for the GROUP BY
2544 * expressions (if there are any) in the subplan's target list.
2545 * 'need_tlist_eval' is set true if we really need to evaluate the
2548 * The result is the targetlist to be passed to the subplan.
2551 make_subplanTargetList(PlannerInfo *root,
2553 AttrNumber **groupColIdx,
2554 bool *need_tlist_eval)
2556 Query *parse = root->parse;
2561 *groupColIdx = NULL;
2564 * If we're not grouping or aggregating, there's nothing to do here;
2565 * query_planner should receive the unmodified target list.
2567 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual &&
2568 !parse->hasWindowFuncs)
2570 *need_tlist_eval = true;
2575 * Otherwise, start with a "flattened" tlist (having just the vars
2576 * mentioned in the targetlist and HAVING qual --- but not upper-level
2577 * Vars; they will be replaced by Params later on). Note this includes
2578 * vars used in resjunk items, so we are covering the needs of ORDER BY
2579 * and window specifications.
2581 sub_tlist = flatten_tlist(tlist);
2582 extravars = pull_var_clause(parse->havingQual, PVC_INCLUDE_PLACEHOLDERS);
2583 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2584 list_free(extravars);
2585 *need_tlist_eval = false; /* only eval if not flat tlist */
2588 * If grouping, create sub_tlist entries for all GROUP BY expressions
2589 * (GROUP BY items that are simple Vars should be in the list already),
2590 * and make an array showing where the group columns are in the sub_tlist.
2592 numCols = list_length(parse->groupClause);
2596 AttrNumber *grpColIdx;
2599 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2600 *groupColIdx = grpColIdx;
2602 foreach(gl, parse->groupClause)
2604 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2605 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2609 * Find or make a matching sub_tlist entry. If the groupexpr
2610 * isn't a Var, no point in searching. (Note that the parser
2611 * won't make multiple groupClause entries for the same TLE.)
2613 if (groupexpr && IsA(groupexpr, Var))
2614 te = tlist_member(groupexpr, sub_tlist);
2620 te = makeTargetEntry((Expr *) groupexpr,
2621 list_length(sub_tlist) + 1,
2624 sub_tlist = lappend(sub_tlist, te);
2625 *need_tlist_eval = true; /* it's not flat anymore */
2628 /* and save its resno */
2629 grpColIdx[keyno++] = te->resno;
2637 * locate_grouping_columns
2638 * Locate grouping columns in the tlist chosen by query_planner.
2640 * This is only needed if we don't use the sub_tlist chosen by
2641 * make_subplanTargetList. We have to forget the column indexes found
2642 * by that routine and re-locate the grouping exprs in the real sub_tlist.
2645 locate_grouping_columns(PlannerInfo *root,
2648 AttrNumber *groupColIdx)
2654 * No work unless grouping.
2656 if (!root->parse->groupClause)
2658 Assert(groupColIdx == NULL);
2661 Assert(groupColIdx != NULL);
2663 foreach(gl, root->parse->groupClause)
2665 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2666 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2667 TargetEntry *te = tlist_member(groupexpr, sub_tlist);
2670 elog(ERROR, "failed to locate grouping columns");
2671 groupColIdx[keyno++] = te->resno;
2676 * postprocess_setop_tlist
2677 * Fix up targetlist returned by plan_set_operations().
2679 * We need to transpose sort key info from the orig_tlist into new_tlist.
2680 * NOTE: this would not be good enough if we supported resjunk sort keys
2681 * for results of set operations --- then, we'd need to project a whole
2682 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2683 * find any resjunk columns in orig_tlist.
2686 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2689 ListCell *orig_tlist_item = list_head(orig_tlist);
2691 foreach(l, new_tlist)
2693 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2694 TargetEntry *orig_tle;
2696 /* ignore resjunk columns in setop result */
2697 if (new_tle->resjunk)
2700 Assert(orig_tlist_item != NULL);
2701 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2702 orig_tlist_item = lnext(orig_tlist_item);
2703 if (orig_tle->resjunk) /* should not happen */
2704 elog(ERROR, "resjunk output columns are not implemented");
2705 Assert(new_tle->resno == orig_tle->resno);
2706 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2708 if (orig_tlist_item != NULL)
2709 elog(ERROR, "resjunk output columns are not implemented");
2714 * select_active_windows
2715 * Create a list of the "active" window clauses (ie, those referenced
2716 * by non-deleted WindowFuncs) in the order they are to be executed.
2719 select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
2725 /* First, make a list of the active windows */
2727 foreach(lc, root->parse->windowClause)
2729 WindowClause *wc = (WindowClause *) lfirst(lc);
2731 /* It's only active if wflists shows some related WindowFuncs */
2732 Assert(wc->winref <= wflists->maxWinRef);
2733 if (wflists->windowFuncs[wc->winref] != NIL)
2734 actives = lappend(actives, wc);
2738 * Now, ensure that windows with identical partitioning/ordering clauses
2739 * are adjacent in the list. This is required by the SQL standard, which
2740 * says that only one sort is to be used for such windows, even if they
2741 * are otherwise distinct (eg, different names or framing clauses).
2743 * There is room to be much smarter here, for example detecting whether
2744 * one window's sort keys are a prefix of another's (so that sorting for
2745 * the latter would do for the former), or putting windows first that
2746 * match a sort order available for the underlying query. For the moment
2747 * we are content with meeting the spec.
2750 while (actives != NIL)
2752 WindowClause *wc = (WindowClause *) linitial(actives);
2756 /* Move wc from actives to result */
2757 actives = list_delete_first(actives);
2758 result = lappend(result, wc);
2760 /* Now move any matching windows from actives to result */
2762 for (lc = list_head(actives); lc; lc = next)
2764 WindowClause *wc2 = (WindowClause *) lfirst(lc);
2767 /* framing options are NOT to be compared here! */
2768 if (equal(wc->partitionClause, wc2->partitionClause) &&
2769 equal(wc->orderClause, wc2->orderClause))
2771 actives = list_delete_cell(actives, lc, prev);
2772 result = lappend(result, wc2);
2783 * add_volatile_sort_exprs
2784 * Identify any volatile sort/group expressions used by the active
2785 * windows, and add them to window_tlist if not already present.
2786 * Return the modified window_tlist.
2789 add_volatile_sort_exprs(List *window_tlist, List *tlist, List *activeWindows)
2791 Bitmapset *sgrefs = NULL;
2794 /* First, collect the sortgrouprefs of the windows into a bitmapset */
2795 foreach(lc, activeWindows)
2797 WindowClause *wc = (WindowClause *) lfirst(lc);
2800 foreach(lc2, wc->partitionClause)
2802 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2804 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2806 foreach(lc2, wc->orderClause)
2808 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2810 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2815 * Now scan the original tlist to find the referenced expressions. Any
2816 * that are volatile must be added to window_tlist.
2818 * Note: we know that the input window_tlist contains no items marked with
2819 * ressortgrouprefs, so we don't have to worry about collisions of the
2820 * reference numbers.
2824 TargetEntry *tle = (TargetEntry *) lfirst(lc);
2826 if (tle->ressortgroupref != 0 &&
2827 bms_is_member(tle->ressortgroupref, sgrefs) &&
2828 contain_volatile_functions((Node *) tle->expr))
2830 TargetEntry *newtle;
2832 newtle = makeTargetEntry(tle->expr,
2833 list_length(window_tlist) + 1,
2836 newtle->ressortgroupref = tle->ressortgroupref;
2837 window_tlist = lappend(window_tlist, newtle);
2841 return window_tlist;
2845 * make_pathkeys_for_window
2846 * Create a pathkeys list describing the required input ordering
2847 * for the given WindowClause.
2849 * The required ordering is first the PARTITION keys, then the ORDER keys.
2850 * In the future we might try to implement windowing using hashing, in which
2851 * case the ordering could be relaxed, but for now we always sort.
2854 make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
2855 List *tlist, bool canonicalize)
2857 List *window_pathkeys;
2858 List *window_sortclauses;
2860 /* Throw error if can't sort */
2861 if (!grouping_is_sortable(wc->partitionClause))
2863 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2864 errmsg("could not implement window PARTITION BY"),
2865 errdetail("Window partitioning columns must be of sortable datatypes.")));
2866 if (!grouping_is_sortable(wc->orderClause))
2868 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2869 errmsg("could not implement window ORDER BY"),
2870 errdetail("Window ordering columns must be of sortable datatypes.")));
2872 /* Okay, make the combined pathkeys */
2873 window_sortclauses = list_concat(list_copy(wc->partitionClause),
2874 list_copy(wc->orderClause));
2875 window_pathkeys = make_pathkeys_for_sortclauses(root,
2879 list_free(window_sortclauses);
2880 return window_pathkeys;
2884 * get_column_info_for_window
2885 * Get the partitioning/ordering column numbers and equality operators
2886 * for a WindowAgg node.
2888 * This depends on the behavior of make_pathkeys_for_window()!
2890 * We are given the target WindowClause and an array of the input column
2891 * numbers associated with the resulting pathkeys. In the easy case, there
2892 * are the same number of pathkey columns as partitioning + ordering columns
2893 * and we just have to copy some data around. However, it's possible that
2894 * some of the original partitioning + ordering columns were eliminated as
2895 * redundant during the transformation to pathkeys. (This can happen even
2896 * though the parser gets rid of obvious duplicates. A typical scenario is a
2897 * window specification "PARTITION BY x ORDER BY y" coupled with a clause
2898 * "WHERE x = y" that causes the two sort columns to be recognized as
2899 * redundant.) In that unusual case, we have to work a lot harder to
2900 * determine which keys are significant.
2902 * The method used here is a bit brute-force: add the sort columns to a list
2903 * one at a time and note when the resulting pathkey list gets longer. But
2904 * it's a sufficiently uncommon case that a faster way doesn't seem worth
2905 * the amount of code refactoring that'd be needed.
2909 get_column_info_for_window(PlannerInfo *root, WindowClause *wc, List *tlist,
2910 int numSortCols, AttrNumber *sortColIdx,
2912 AttrNumber **partColIdx,
2913 Oid **partOperators,
2915 AttrNumber **ordColIdx,
2918 int numPart = list_length(wc->partitionClause);
2919 int numOrder = list_length(wc->orderClause);
2921 if (numSortCols == numPart + numOrder)
2924 *partNumCols = numPart;
2925 *partColIdx = sortColIdx;
2926 *partOperators = extract_grouping_ops(wc->partitionClause);
2927 *ordNumCols = numOrder;
2928 *ordColIdx = sortColIdx + numPart;
2929 *ordOperators = extract_grouping_ops(wc->orderClause);
2938 /* first, allocate what's certainly enough space for the arrays */
2940 *partColIdx = (AttrNumber *) palloc(numPart * sizeof(AttrNumber));
2941 *partOperators = (Oid *) palloc(numPart * sizeof(Oid));
2943 *ordColIdx = (AttrNumber *) palloc(numOrder * sizeof(AttrNumber));
2944 *ordOperators = (Oid *) palloc(numOrder * sizeof(Oid));
2948 foreach(lc, wc->partitionClause)
2950 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2953 sortclauses = lappend(sortclauses, sgc);
2954 new_pathkeys = make_pathkeys_for_sortclauses(root,
2958 if (list_length(new_pathkeys) > list_length(pathkeys))
2960 /* this sort clause is actually significant */
2961 (*partColIdx)[*partNumCols] = sortColIdx[scidx++];
2962 (*partOperators)[*partNumCols] = sgc->eqop;
2964 pathkeys = new_pathkeys;
2967 foreach(lc, wc->orderClause)
2969 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2972 sortclauses = lappend(sortclauses, sgc);
2973 new_pathkeys = make_pathkeys_for_sortclauses(root,
2977 if (list_length(new_pathkeys) > list_length(pathkeys))
2979 /* this sort clause is actually significant */
2980 (*ordColIdx)[*ordNumCols] = sortColIdx[scidx++];
2981 (*ordOperators)[*ordNumCols] = sgc->eqop;
2983 pathkeys = new_pathkeys;
2986 /* complain if we didn't eat exactly the right number of sort cols */
2987 if (scidx != numSortCols)
2988 elog(ERROR, "failed to deconstruct sort operators into partitioning/ordering operators");
2994 * expression_planner
2995 * Perform planner's transformations on a standalone expression.
2997 * Various utility commands need to evaluate expressions that are not part
2998 * of a plannable query. They can do so using the executor's regular
2999 * expression-execution machinery, but first the expression has to be fed
3000 * through here to transform it from parser output to something executable.
3002 * Currently, we disallow sublinks in standalone expressions, so there's no
3003 * real "planning" involved here. (That might not always be true though.)
3004 * What we must do is run eval_const_expressions to ensure that any function
3005 * calls are converted to positional notation and function default arguments
3006 * get inserted. The fact that constant subexpressions get simplified is a
3007 * side-effect that is useful when the expression will get evaluated more than
3008 * once. Also, we must fix operator function IDs.
3010 * Note: this must not make any damaging changes to the passed-in expression
3011 * tree. (It would actually be okay to apply fix_opfuncids to it, but since
3012 * we first do an expression_tree_mutator-based walk, what is returned will
3013 * be a new node tree.)
3016 expression_planner(Expr *expr)
3021 * Convert named-argument function calls, insert default arguments and
3022 * simplify constant subexprs
3024 result = eval_const_expressions(NULL, (Node *) expr);
3026 /* Fill in opfuncid values if missing */
3027 fix_opfuncids(result);
3029 return (Expr *) result;
3034 * plan_cluster_use_sort
3035 * Use the planner to decide how CLUSTER should implement sorting
3037 * tableOid is the OID of a table to be clustered on its index indexOid
3038 * (which is already known to be a btree index). Decide whether it's
3039 * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
3040 * Return TRUE to use sorting, FALSE to use an indexscan.
3042 * Note: caller had better already hold some type of lock on the table.
3045 plan_cluster_use_sort(Oid tableOid, Oid indexOid)
3049 PlannerGlobal *glob;
3052 IndexOptInfo *indexInfo;
3053 QualCost indexExprCost;
3054 Cost comparisonCost;
3056 Path seqScanAndSortPath;
3057 IndexPath *indexScanPath;
3060 /* Set up mostly-dummy planner state */
3061 query = makeNode(Query);
3062 query->commandType = CMD_SELECT;
3064 glob = makeNode(PlannerGlobal);
3066 root = makeNode(PlannerInfo);
3067 root->parse = query;
3069 root->query_level = 1;
3070 root->planner_cxt = CurrentMemoryContext;
3071 root->wt_param_id = -1;
3073 /* Build a minimal RTE for the rel */
3074 rte = makeNode(RangeTblEntry);
3075 rte->rtekind = RTE_RELATION;
3076 rte->relid = tableOid;
3078 rte->inFromCl = true;
3079 query->rtable = list_make1(rte);
3081 /* ... and insert it into PlannerInfo */
3082 root->simple_rel_array_size = 2;
3083 root->simple_rel_array = (RelOptInfo **)
3084 palloc0(root->simple_rel_array_size * sizeof(RelOptInfo *));
3085 root->simple_rte_array = (RangeTblEntry **)
3086 palloc0(root->simple_rel_array_size * sizeof(RangeTblEntry *));
3087 root->simple_rte_array[1] = rte;
3089 /* Build RelOptInfo */
3090 rel = build_simple_rel(root, 1, RELOPT_BASEREL);
3093 * Rather than doing all the pushups that would be needed to use
3094 * set_baserel_size_estimates, just do a quick hack for rows and width.
3096 rel->rows = rel->tuples;
3097 rel->width = get_relation_data_width(tableOid, NULL);
3099 root->total_table_pages = rel->pages;
3101 /* Locate IndexOptInfo for the target index */
3103 foreach(lc, rel->indexlist)
3105 indexInfo = (IndexOptInfo *) lfirst(lc);
3106 if (indexInfo->indexoid == indexOid)
3109 if (lc == NULL) /* not in the list? */
3110 elog(ERROR, "index %u does not belong to table %u",
3111 indexOid, tableOid);
3114 * Determine eval cost of the index expressions, if any. We need to
3115 * charge twice that amount for each tuple comparison that happens
3116 * during the sort, since tuplesort.c will have to re-evaluate the
3117 * index expressions each time. (XXX that's pretty inefficient...)
3119 cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
3120 comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
3122 /* Estimate the cost of seq scan + sort */
3123 seqScanPath = create_seqscan_path(root, rel);
3124 cost_sort(&seqScanAndSortPath, root, NIL,
3125 seqScanPath->total_cost, rel->tuples, rel->width,
3126 comparisonCost, maintenance_work_mem, -1.0);
3128 /* Estimate the cost of index scan */
3129 indexScanPath = create_index_path(root, indexInfo,
3131 ForwardScanDirection, NULL);
3133 return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);