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, AggClauseCosts *agg_costs);
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->resultRelations = NIL;
167 glob->relationOids = NIL;
168 glob->invalItems = NIL;
170 glob->lastRowMarkId = 0;
171 glob->transientPlan = false;
173 /* Determine what fraction of the plan is likely to be scanned */
174 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
177 * We have no real idea how many tuples the user will ultimately FETCH
178 * from a cursor, but it is often the case that he doesn't want 'em
179 * all, or would prefer a fast-start plan anyway so that he can
180 * process some of the tuples sooner. Use a GUC parameter to decide
181 * what fraction to optimize for.
183 tuple_fraction = cursor_tuple_fraction;
186 * We document cursor_tuple_fraction as simply being a fraction, which
187 * means the edge cases 0 and 1 have to be treated specially here. We
188 * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
190 if (tuple_fraction >= 1.0)
191 tuple_fraction = 0.0;
192 else if (tuple_fraction <= 0.0)
193 tuple_fraction = 1e-10;
197 /* Default assumption is we need all the tuples */
198 tuple_fraction = 0.0;
201 /* primary planning entry point (may recurse for subqueries) */
202 top_plan = subquery_planner(glob, parse, NULL,
203 false, tuple_fraction, &root);
206 * If creating a plan for a scrollable cursor, make sure it can run
207 * backwards on demand. Add a Material node at the top at need.
209 if (cursorOptions & CURSOR_OPT_SCROLL)
211 if (!ExecSupportsBackwardScan(top_plan))
212 top_plan = materialize_finished_plan(top_plan);
215 /* final cleanup of the plan */
216 Assert(glob->finalrtable == NIL);
217 Assert(glob->finalrowmarks == NIL);
218 Assert(glob->resultRelations == NIL);
219 top_plan = set_plan_references(glob, top_plan,
222 /* ... and the subplans (both regular subplans and initplans) */
223 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
224 Assert(list_length(glob->subplans) == list_length(glob->subrowmarks));
225 lrt = list_head(glob->subrtables);
226 lrm = list_head(glob->subrowmarks);
227 foreach(lp, glob->subplans)
229 Plan *subplan = (Plan *) lfirst(lp);
230 List *subrtable = (List *) lfirst(lrt);
231 List *subrowmark = (List *) lfirst(lrm);
233 lfirst(lp) = set_plan_references(glob, subplan,
234 subrtable, subrowmark);
239 /* build the PlannedStmt result */
240 result = makeNode(PlannedStmt);
242 result->commandType = parse->commandType;
243 result->hasReturning = (parse->returningList != NIL);
244 result->hasModifyingCTE = parse->hasModifyingCTE;
245 result->canSetTag = parse->canSetTag;
246 result->transientPlan = glob->transientPlan;
247 result->planTree = top_plan;
248 result->rtable = glob->finalrtable;
249 result->resultRelations = glob->resultRelations;
250 result->utilityStmt = parse->utilityStmt;
251 result->intoClause = parse->intoClause;
252 result->subplans = glob->subplans;
253 result->rewindPlanIDs = glob->rewindPlanIDs;
254 result->rowMarks = glob->finalrowmarks;
255 result->relationOids = glob->relationOids;
256 result->invalItems = glob->invalItems;
257 result->nParamExec = list_length(glob->paramlist);
263 /*--------------------
265 * Invokes the planner on a subquery. We recurse to here for each
266 * sub-SELECT found in the query tree.
268 * glob is the global state for the current planner run.
269 * parse is the querytree produced by the parser & rewriter.
270 * parent_root is the immediate parent Query's info (NULL at the top level).
271 * hasRecursion is true if this is a recursive WITH query.
272 * tuple_fraction is the fraction of tuples we expect will be retrieved.
273 * tuple_fraction is interpreted as explained for grouping_planner, below.
275 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
276 * among other things this tells the output sort ordering of the plan.
278 * Basically, this routine does the stuff that should only be done once
279 * per Query object. It then calls grouping_planner. At one time,
280 * grouping_planner could be invoked recursively on the same Query object;
281 * that's not currently true, but we keep the separation between the two
282 * routines anyway, in case we need it again someday.
284 * subquery_planner will be called recursively to handle sub-Query nodes
285 * found within the query's expressions and rangetable.
287 * Returns a query plan.
288 *--------------------
291 subquery_planner(PlannerGlobal *glob, Query *parse,
292 PlannerInfo *parent_root,
293 bool hasRecursion, double tuple_fraction,
294 PlannerInfo **subroot)
296 int num_old_subplans = list_length(glob->subplans);
303 /* Create a PlannerInfo data structure for this subquery */
304 root = makeNode(PlannerInfo);
307 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
308 root->parent_root = parent_root;
309 root->planner_cxt = CurrentMemoryContext;
310 root->init_plans = NIL;
311 root->cte_plan_ids = NIL;
312 root->eq_classes = NIL;
313 root->append_rel_list = NIL;
314 root->rowMarks = NIL;
315 root->hasInheritedTarget = false;
317 root->hasRecursion = hasRecursion;
319 root->wt_param_id = SS_assign_special_param(root);
321 root->wt_param_id = -1;
322 root->non_recursive_plan = NULL;
325 * If there is a WITH list, process each WITH query and build an initplan
326 * SubPlan structure for it.
329 SS_process_ctes(root);
332 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
333 * to transform them into joins. Note that this step does not descend
334 * into subqueries; if we pull up any subqueries below, their SubLinks are
335 * processed just before pulling them up.
337 if (parse->hasSubLinks)
338 pull_up_sublinks(root);
341 * Scan the rangetable for set-returning functions, and inline them if
342 * possible (producing subqueries that might get pulled up next).
343 * Recursion issues here are handled in the same way as for SubLinks.
345 inline_set_returning_functions(root);
348 * Check to see if any subqueries in the jointree can be merged into this
351 parse->jointree = (FromExpr *)
352 pull_up_subqueries(root, (Node *) parse->jointree, NULL, NULL);
355 * If this is a simple UNION ALL query, flatten it into an appendrel. We
356 * do this now because it requires applying pull_up_subqueries to the leaf
357 * queries of the UNION ALL, which weren't touched above because they
358 * weren't referenced by the jointree (they will be after we do this).
360 if (parse->setOperations)
361 flatten_simple_union_all(root);
364 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
365 * avoid the expense of doing flatten_join_alias_vars(). Also check for
366 * outer joins --- if none, we can skip reduce_outer_joins(). This must be
367 * done after we have done pull_up_subqueries, of course.
369 root->hasJoinRTEs = false;
370 hasOuterJoins = false;
371 foreach(l, parse->rtable)
373 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
375 if (rte->rtekind == RTE_JOIN)
377 root->hasJoinRTEs = true;
378 if (IS_OUTER_JOIN(rte->jointype))
380 hasOuterJoins = true;
381 /* Can quit scanning once we find an outer join */
388 * Preprocess RowMark information. We need to do this after subquery
389 * pullup (so that all non-inherited RTEs are present) and before
390 * inheritance expansion (so that the info is available for
391 * expand_inherited_tables to examine and modify).
393 preprocess_rowmarks(root);
396 * Expand any rangetable entries that are inheritance sets into "append
397 * relations". This can add entries to the rangetable, but they must be
398 * plain base relations not joins, so it's OK (and marginally more
399 * efficient) to do it after checking for join RTEs. We must do it after
400 * pulling up subqueries, else we'd fail to handle inherited tables in
403 expand_inherited_tables(root);
406 * Set hasHavingQual to remember if HAVING clause is present. Needed
407 * because preprocess_expression will reduce a constant-true condition to
408 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
410 root->hasHavingQual = (parse->havingQual != NULL);
412 /* Clear this flag; might get set in distribute_qual_to_rels */
413 root->hasPseudoConstantQuals = false;
416 * Do expression preprocessing on targetlist and quals, as well as other
417 * random expressions in the querytree. Note that we do not need to
418 * handle sort/group expressions explicitly, because they are actually
419 * part of the targetlist.
421 parse->targetList = (List *)
422 preprocess_expression(root, (Node *) parse->targetList,
425 parse->returningList = (List *)
426 preprocess_expression(root, (Node *) parse->returningList,
429 preprocess_qual_conditions(root, (Node *) parse->jointree);
431 parse->havingQual = preprocess_expression(root, parse->havingQual,
434 foreach(l, parse->windowClause)
436 WindowClause *wc = (WindowClause *) lfirst(l);
438 /* partitionClause/orderClause are sort/group expressions */
439 wc->startOffset = preprocess_expression(root, wc->startOffset,
441 wc->endOffset = preprocess_expression(root, wc->endOffset,
445 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
447 parse->limitCount = preprocess_expression(root, parse->limitCount,
450 root->append_rel_list = (List *)
451 preprocess_expression(root, (Node *) root->append_rel_list,
454 /* Also need to preprocess expressions for function and values RTEs */
455 foreach(l, parse->rtable)
457 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
459 if (rte->rtekind == RTE_FUNCTION)
460 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
462 else if (rte->rtekind == RTE_VALUES)
463 rte->values_lists = (List *)
464 preprocess_expression(root, (Node *) rte->values_lists,
469 * In some cases we may want to transfer a HAVING clause into WHERE. We
470 * cannot do so if the HAVING clause contains aggregates (obviously) or
471 * volatile functions (since a HAVING clause is supposed to be executed
472 * only once per group). Also, it may be that the clause is so expensive
473 * to execute that we're better off doing it only once per group, despite
474 * the loss of selectivity. This is hard to estimate short of doing the
475 * entire planning process twice, so we use a heuristic: clauses
476 * containing subplans are left in HAVING. Otherwise, we move or copy the
477 * HAVING clause into WHERE, in hopes of eliminating tuples before
478 * aggregation instead of after.
480 * If the query has explicit grouping then we can simply move such a
481 * clause into WHERE; any group that fails the clause will not be in the
482 * output because none of its tuples will reach the grouping or
483 * aggregation stage. Otherwise we must have a degenerate (variable-free)
484 * HAVING clause, which we put in WHERE so that query_planner() can use it
485 * in a gating Result node, but also keep in HAVING to ensure that we
486 * don't emit a bogus aggregated row. (This could be done better, but it
487 * seems not worth optimizing.)
489 * Note that both havingQual and parse->jointree->quals are in
490 * implicitly-ANDed-list form at this point, even though they are declared
494 foreach(l, (List *) parse->havingQual)
496 Node *havingclause = (Node *) lfirst(l);
498 if (contain_agg_clause(havingclause) ||
499 contain_volatile_functions(havingclause) ||
500 contain_subplans(havingclause))
502 /* keep it in HAVING */
503 newHaving = lappend(newHaving, havingclause);
505 else if (parse->groupClause)
507 /* move it to WHERE */
508 parse->jointree->quals = (Node *)
509 lappend((List *) parse->jointree->quals, havingclause);
513 /* put a copy in WHERE, keep it in HAVING */
514 parse->jointree->quals = (Node *)
515 lappend((List *) parse->jointree->quals,
516 copyObject(havingclause));
517 newHaving = lappend(newHaving, havingclause);
520 parse->havingQual = (Node *) newHaving;
523 * If we have any outer joins, try to reduce them to plain inner joins.
524 * This step is most easily done after we've done expression
528 reduce_outer_joins(root);
531 * Do the main planning. If we have an inherited target relation, that
532 * needs special processing, else go straight to grouping_planner.
534 if (parse->resultRelation &&
535 rt_fetch(parse->resultRelation, parse->rtable)->inh)
536 plan = inheritance_planner(root);
539 plan = grouping_planner(root, tuple_fraction);
540 /* If it's not SELECT, we need a ModifyTable node */
541 if (parse->commandType != CMD_SELECT)
543 List *returningLists;
547 * Deal with the RETURNING clause if any. It's convenient to pass
548 * the returningList through setrefs.c now rather than at top
549 * level (if we waited, handling inherited UPDATE/DELETE would be
552 if (parse->returningList)
556 Assert(parse->resultRelation);
557 rlist = set_returning_clause_references(root->glob,
558 parse->returningList,
560 parse->resultRelation);
561 returningLists = list_make1(rlist);
564 returningLists = NIL;
567 * If there was a FOR UPDATE/SHARE clause, the LockRows node will
568 * have dealt with fetching non-locked marked rows, else we need
569 * to have ModifyTable do that.
574 rowMarks = root->rowMarks;
576 plan = (Plan *) make_modifytable(parse->commandType,
578 list_make1_int(parse->resultRelation),
582 SS_assign_special_param(root));
587 * If any subplans were generated, or if there are any parameters to worry
588 * about, build initPlan list and extParam/allParam sets for plan nodes,
589 * and attach the initPlans to the top plan node.
591 if (list_length(glob->subplans) != num_old_subplans ||
592 root->glob->paramlist != NIL)
593 SS_finalize_plan(root, plan, true);
595 /* Return internal info if caller wants it */
603 * preprocess_expression
604 * Do subquery_planner's preprocessing work for an expression,
605 * which can be a targetlist, a WHERE clause (including JOIN/ON
606 * conditions), or a HAVING clause.
609 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
612 * Fall out quickly if expression is empty. This occurs often enough to
613 * be worth checking. Note that null->null is the correct conversion for
614 * implicit-AND result format, too.
620 * If the query has any join RTEs, replace join alias variables with
621 * base-relation variables. We must do this before sublink processing,
622 * else sublinks expanded out from join aliases wouldn't get processed. We
623 * can skip it in VALUES lists, however, since they can't contain any Vars
626 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
627 expr = flatten_join_alias_vars(root, expr);
630 * Simplify constant expressions.
632 * Note: an essential effect of this is to convert named-argument function
633 * calls to positional notation and insert the current actual values of
634 * any default arguments for functions. To ensure that happens, we *must*
635 * process all expressions here. Previous PG versions sometimes skipped
636 * const-simplification if it didn't seem worth the trouble, but we can't
639 * Note: this also flattens nested AND and OR expressions into N-argument
640 * form. All processing of a qual expression after this point must be
641 * careful to maintain AND/OR flatness --- that is, do not generate a tree
642 * with AND directly under AND, nor OR directly under OR.
644 expr = eval_const_expressions(root, expr);
647 * If it's a qual or havingQual, canonicalize it.
649 if (kind == EXPRKIND_QUAL)
651 expr = (Node *) canonicalize_qual((Expr *) expr);
653 #ifdef OPTIMIZER_DEBUG
654 printf("After canonicalize_qual()\n");
659 /* Expand SubLinks to SubPlans */
660 if (root->parse->hasSubLinks)
661 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
664 * XXX do not insert anything here unless you have grokked the comments in
665 * SS_replace_correlation_vars ...
668 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
669 if (root->query_level > 1)
670 expr = SS_replace_correlation_vars(root, expr);
673 * If it's a qual or havingQual, convert it to implicit-AND format. (We
674 * don't want to do this before eval_const_expressions, since the latter
675 * would be unable to simplify a top-level AND correctly. Also,
676 * SS_process_sublinks expects explicit-AND format.)
678 if (kind == EXPRKIND_QUAL)
679 expr = (Node *) make_ands_implicit((Expr *) expr);
685 * preprocess_qual_conditions
686 * Recursively scan the query's jointree and do subquery_planner's
687 * preprocessing work on each qual condition found therein.
690 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
694 if (IsA(jtnode, RangeTblRef))
696 /* nothing to do here */
698 else if (IsA(jtnode, FromExpr))
700 FromExpr *f = (FromExpr *) jtnode;
703 foreach(l, f->fromlist)
704 preprocess_qual_conditions(root, lfirst(l));
706 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
708 else if (IsA(jtnode, JoinExpr))
710 JoinExpr *j = (JoinExpr *) jtnode;
712 preprocess_qual_conditions(root, j->larg);
713 preprocess_qual_conditions(root, j->rarg);
715 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
718 elog(ERROR, "unrecognized node type: %d",
719 (int) nodeTag(jtnode));
723 * inheritance_planner
724 * Generate a plan in the case where the result relation is an
727 * We have to handle this case differently from cases where a source relation
728 * is an inheritance set. Source inheritance is expanded at the bottom of the
729 * plan tree (see allpaths.c), but target inheritance has to be expanded at
730 * the top. The reason is that for UPDATE, each target relation needs a
731 * different targetlist matching its own column set. Fortunately,
732 * the UPDATE/DELETE target can never be the nullable side of an outer join,
733 * so it's OK to generate the plan this way.
735 * Returns a query plan.
738 inheritance_planner(PlannerInfo *root)
740 Query *parse = root->parse;
741 int parentRTindex = parse->resultRelation;
742 List *subplans = NIL;
743 List *resultRelations = NIL;
744 List *returningLists = NIL;
751 foreach(l, root->append_rel_list)
753 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
756 /* append_rel_list contains all append rels; ignore others */
757 if (appinfo->parent_relid != parentRTindex)
761 * Generate modified query with this rel as target.
763 memcpy(&subroot, root, sizeof(PlannerInfo));
764 subroot.parse = (Query *)
765 adjust_appendrel_attrs((Node *) parse,
767 subroot.init_plans = NIL;
768 subroot.hasInheritedTarget = true;
769 /* We needn't modify the child's append_rel_list */
770 /* There shouldn't be any OJ info to translate, as yet */
771 Assert(subroot.join_info_list == NIL);
772 /* and we haven't created PlaceHolderInfos, either */
773 Assert(subroot.placeholder_list == NIL);
776 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
779 * If this child rel was excluded by constraint exclusion, exclude it
782 if (is_dummy_plan(subplan))
785 /* Save rtable from first rel for use below */
787 rtable = subroot.parse->rtable;
789 subplans = lappend(subplans, subplan);
791 /* Make sure any initplans from this rel get into the outer list */
792 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
794 /* Build list of target-relation RT indexes */
795 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
797 /* Build list of per-relation RETURNING targetlists */
798 if (parse->returningList)
802 rlist = set_returning_clause_references(root->glob,
803 subroot.parse->returningList,
805 appinfo->child_relid);
806 returningLists = lappend(returningLists, rlist);
810 /* Mark result as unordered (probably unnecessary) */
811 root->query_pathkeys = NIL;
814 * If we managed to exclude every child rel, return a dummy plan; it
815 * doesn't even need a ModifyTable node.
819 /* although dummy, it must have a valid tlist for executor */
820 tlist = preprocess_targetlist(root, parse->targetList);
821 return (Plan *) make_result(root,
823 (Node *) list_make1(makeBoolConst(false,
829 * Planning might have modified the rangetable, due to changes of the
830 * Query structures inside subquery RTEs. We have to ensure that this
831 * gets propagated back to the master copy. But can't do this until we
832 * are done planning, because all the calls to grouping_planner need
833 * virgin sub-Queries to work from. (We are effectively assuming that
834 * sub-Queries will get planned identically each time, or at least that
835 * the impacts on their rangetables will be the same each time.)
837 * XXX should clean this up someday
839 parse->rtable = rtable;
842 * If there was a FOR UPDATE/SHARE clause, the LockRows node will have
843 * dealt with fetching non-locked marked rows, else we need to have
844 * ModifyTable do that.
849 rowMarks = root->rowMarks;
851 /* And last, tack on a ModifyTable node to do the UPDATE/DELETE work */
852 return (Plan *) make_modifytable(parse->commandType,
858 SS_assign_special_param(root));
861 /*--------------------
863 * Perform planning steps related to grouping, aggregation, etc.
864 * This primarily means adding top-level processing to the basic
865 * query plan produced by query_planner.
867 * tuple_fraction is the fraction of tuples we expect will be retrieved
869 * tuple_fraction is interpreted as follows:
870 * 0: expect all tuples to be retrieved (normal case)
871 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
872 * from the plan to be retrieved
873 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
874 * expected to be retrieved (ie, a LIMIT specification)
876 * Returns a query plan. Also, root->query_pathkeys is returned as the
877 * actual output ordering of the plan (in pathkey format).
878 *--------------------
881 grouping_planner(PlannerInfo *root, double tuple_fraction)
883 Query *parse = root->parse;
884 List *tlist = parse->targetList;
885 int64 offset_est = 0;
887 double limit_tuples = -1.0;
889 List *current_pathkeys;
890 double dNumGroups = 0;
891 bool use_hashed_distinct = false;
892 bool tested_hashed_distinct = false;
894 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
895 if (parse->limitCount || parse->limitOffset)
897 tuple_fraction = preprocess_limit(root, tuple_fraction,
898 &offset_est, &count_est);
901 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
902 * estimate the effects of using a bounded sort.
904 if (count_est > 0 && offset_est >= 0)
905 limit_tuples = (double) count_est + (double) offset_est;
908 if (parse->setOperations)
910 List *set_sortclauses;
913 * If there's a top-level ORDER BY, assume we have to fetch all the
914 * tuples. This might be too simplistic given all the hackery below
915 * to possibly avoid the sort; but the odds of accurate estimates here
916 * are pretty low anyway.
918 if (parse->sortClause)
919 tuple_fraction = 0.0;
922 * Construct the plan for set operations. The result will not need
923 * any work except perhaps a top-level sort and/or LIMIT. Note that
924 * any special work for recursive unions is the responsibility of
925 * plan_set_operations.
927 result_plan = plan_set_operations(root, tuple_fraction,
931 * Calculate pathkeys representing the sort order (if any) of the set
932 * operation's result. We have to do this before overwriting the sort
935 current_pathkeys = make_pathkeys_for_sortclauses(root,
937 result_plan->targetlist,
941 * We should not need to call preprocess_targetlist, since we must be
942 * in a SELECT query node. Instead, use the targetlist returned by
943 * plan_set_operations (since this tells whether it returned any
944 * resjunk columns!), and transfer any sort key information from the
947 Assert(parse->commandType == CMD_SELECT);
949 tlist = postprocess_setop_tlist(copyObject(result_plan->targetlist),
953 * Can't handle FOR UPDATE/SHARE here (parser should have checked
954 * already, but let's make sure).
958 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
959 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
962 * Calculate pathkeys that represent result ordering requirements
964 Assert(parse->distinctClause == NIL);
965 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
972 /* No set operations, do regular planning */
974 double sub_limit_tuples;
975 AttrNumber *groupColIdx = NULL;
976 bool need_tlist_eval = true;
982 AggClauseCosts agg_costs;
986 bool use_hashed_grouping = false;
987 WindowFuncLists *wflists = NULL;
988 List *activeWindows = NIL;
990 MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
992 /* A recursive query should always have setOperations */
993 Assert(!root->hasRecursion);
995 /* Preprocess GROUP BY clause, if any */
996 if (parse->groupClause)
997 preprocess_groupclause(root);
998 numGroupCols = list_length(parse->groupClause);
1000 /* Preprocess targetlist */
1001 tlist = preprocess_targetlist(root, tlist);
1004 * Locate any window functions in the tlist. (We don't need to look
1005 * anywhere else, since expressions used in ORDER BY will be in there
1006 * too.) Note that they could all have been eliminated by constant
1007 * folding, in which case we don't need to do any more work.
1009 if (parse->hasWindowFuncs)
1011 wflists = find_window_functions((Node *) tlist,
1012 list_length(parse->windowClause));
1013 if (wflists->numWindowFuncs > 0)
1014 activeWindows = select_active_windows(root, wflists);
1016 parse->hasWindowFuncs = false;
1020 * Generate appropriate target list for subplan; may be different from
1021 * tlist if grouping or aggregation is needed.
1023 sub_tlist = make_subplanTargetList(root, tlist,
1024 &groupColIdx, &need_tlist_eval);
1027 * Do aggregate preprocessing, if the query has any aggs.
1029 * Note: think not that we can turn off hasAggs if we find no aggs. It
1030 * is possible for constant-expression simplification to remove all
1031 * explicit references to aggs, but we still have to follow the
1032 * aggregate semantics (eg, producing only one output row).
1037 * Collect statistics about aggregates for estimating costs. Note:
1038 * we do not attempt to detect duplicate aggregates here; a
1039 * somewhat-overestimated cost is okay for our present purposes.
1041 count_agg_clauses(root, (Node *) tlist, &agg_costs);
1042 count_agg_clauses(root, parse->havingQual, &agg_costs);
1045 * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1046 * adding logic between here and the optimize_minmax_aggregates
1047 * call. Anything that is needed in MIN/MAX-optimizable cases
1048 * will have to be duplicated in planagg.c.
1050 preprocess_minmax_aggregates(root, tlist);
1054 * Calculate pathkeys that represent grouping/ordering requirements.
1055 * Stash them in PlannerInfo so that query_planner can canonicalize
1056 * them after EquivalenceClasses have been formed. The sortClause is
1057 * certainly sort-able, but GROUP BY and DISTINCT might not be, in
1058 * which case we just leave their pathkeys empty.
1060 if (parse->groupClause &&
1061 grouping_is_sortable(parse->groupClause))
1062 root->group_pathkeys =
1063 make_pathkeys_for_sortclauses(root,
1068 root->group_pathkeys = NIL;
1070 /* We consider only the first (bottom) window in pathkeys logic */
1071 if (activeWindows != NIL)
1073 WindowClause *wc = (WindowClause *) linitial(activeWindows);
1075 root->window_pathkeys = make_pathkeys_for_window(root,
1081 root->window_pathkeys = NIL;
1083 if (parse->distinctClause &&
1084 grouping_is_sortable(parse->distinctClause))
1085 root->distinct_pathkeys =
1086 make_pathkeys_for_sortclauses(root,
1087 parse->distinctClause,
1091 root->distinct_pathkeys = NIL;
1093 root->sort_pathkeys =
1094 make_pathkeys_for_sortclauses(root,
1100 * Figure out whether we want a sorted result from query_planner.
1102 * If we have a sortable GROUP BY clause, then we want a result sorted
1103 * properly for grouping. Otherwise, if we have window functions to
1104 * evaluate, we try to sort for the first window. Otherwise, if
1105 * there's a sortable DISTINCT clause that's more rigorous than the
1106 * ORDER BY clause, we try to produce output that's sufficiently well
1107 * sorted for the DISTINCT. Otherwise, if there is an ORDER BY
1108 * clause, we want to sort by the ORDER BY clause.
1110 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a
1111 * superset of GROUP BY, it would be tempting to request sort by ORDER
1112 * BY --- but that might just leave us failing to exploit an available
1113 * sort order at all. Needs more thought. The choice for DISTINCT
1114 * versus ORDER BY is much easier, since we know that the parser
1115 * ensured that one is a superset of the other.
1117 if (root->group_pathkeys)
1118 root->query_pathkeys = root->group_pathkeys;
1119 else if (root->window_pathkeys)
1120 root->query_pathkeys = root->window_pathkeys;
1121 else if (list_length(root->distinct_pathkeys) >
1122 list_length(root->sort_pathkeys))
1123 root->query_pathkeys = root->distinct_pathkeys;
1124 else if (root->sort_pathkeys)
1125 root->query_pathkeys = root->sort_pathkeys;
1127 root->query_pathkeys = NIL;
1130 * Figure out whether there's a hard limit on the number of rows that
1131 * query_planner's result subplan needs to return. Even if we know a
1132 * hard limit overall, it doesn't apply if the query has any
1133 * grouping/aggregation operations.
1135 if (parse->groupClause ||
1136 parse->distinctClause ||
1138 parse->hasWindowFuncs ||
1139 root->hasHavingQual)
1140 sub_limit_tuples = -1.0;
1142 sub_limit_tuples = limit_tuples;
1145 * Generate the best unsorted and presorted paths for this Query (but
1146 * note there may not be any presorted path). query_planner will also
1147 * estimate the number of groups in the query, and canonicalize all
1150 query_planner(root, sub_tlist, tuple_fraction, sub_limit_tuples,
1151 &cheapest_path, &sorted_path, &dNumGroups);
1154 * Extract rowcount and width estimates for possible use in grouping
1155 * decisions. Beware here of the possibility that
1156 * cheapest_path->parent is NULL (ie, there is no FROM clause).
1158 if (cheapest_path->parent)
1160 path_rows = cheapest_path->parent->rows;
1161 path_width = cheapest_path->parent->width;
1165 path_rows = 1; /* assume non-set result */
1166 path_width = 100; /* arbitrary */
1169 if (parse->groupClause)
1172 * If grouping, decide whether to use sorted or hashed grouping.
1174 use_hashed_grouping =
1175 choose_hashed_grouping(root,
1176 tuple_fraction, limit_tuples,
1177 path_rows, path_width,
1178 cheapest_path, sorted_path,
1179 dNumGroups, &agg_costs);
1180 /* Also convert # groups to long int --- but 'ware overflow! */
1181 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
1183 else if (parse->distinctClause && sorted_path &&
1184 !root->hasHavingQual && !parse->hasAggs && !activeWindows)
1187 * We'll reach the DISTINCT stage without any intermediate
1188 * processing, so figure out whether we will want to hash or not
1189 * so we can choose whether to use cheapest or sorted path.
1191 use_hashed_distinct =
1192 choose_hashed_distinct(root,
1193 tuple_fraction, limit_tuples,
1194 path_rows, path_width,
1195 cheapest_path->startup_cost,
1196 cheapest_path->total_cost,
1197 sorted_path->startup_cost,
1198 sorted_path->total_cost,
1199 sorted_path->pathkeys,
1201 tested_hashed_distinct = true;
1205 * Select the best path. If we are doing hashed grouping, we will
1206 * always read all the input tuples, so use the cheapest-total path.
1207 * Otherwise, trust query_planner's decision about which to use.
1209 if (use_hashed_grouping || use_hashed_distinct || !sorted_path)
1210 best_path = cheapest_path;
1212 best_path = sorted_path;
1215 * Check to see if it's possible to optimize MIN/MAX aggregates. If
1216 * so, we will forget all the work we did so far to choose a "regular"
1217 * path ... but we had to do it anyway to be able to tell which way is
1220 result_plan = optimize_minmax_aggregates(root,
1224 if (result_plan != NULL)
1227 * optimize_minmax_aggregates generated the full plan, with the
1228 * right tlist, and it has no sort order.
1230 current_pathkeys = NIL;
1235 * Normal case --- create a plan according to query_planner's
1238 bool need_sort_for_grouping = false;
1240 result_plan = create_plan(root, best_path);
1241 current_pathkeys = best_path->pathkeys;
1243 /* Detect if we'll need an explicit sort for grouping */
1244 if (parse->groupClause && !use_hashed_grouping &&
1245 !pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1247 need_sort_for_grouping = true;
1250 * Always override query_planner's tlist, so that we don't
1251 * sort useless data from a "physical" tlist.
1253 need_tlist_eval = true;
1257 * create_plan() returns a plan with just a "flat" tlist of
1258 * required Vars. Usually we need to insert the sub_tlist as the
1259 * tlist of the top plan node. However, we can skip that if we
1260 * determined that whatever query_planner chose to return will be
1263 if (need_tlist_eval)
1266 * If the top-level plan node is one that cannot do expression
1267 * evaluation, we must insert a Result node to project the
1270 if (!is_projection_capable_plan(result_plan))
1272 result_plan = (Plan *) make_result(root,
1280 * Otherwise, just replace the subplan's flat tlist with
1281 * the desired tlist.
1283 result_plan->targetlist = sub_tlist;
1287 * Also, account for the cost of evaluation of the sub_tlist.
1289 * Up to now, we have only been dealing with "flat" tlists,
1290 * containing just Vars. So their evaluation cost is zero
1291 * according to the model used by cost_qual_eval() (or if you
1292 * prefer, the cost is factored into cpu_tuple_cost). Thus we
1293 * can avoid accounting for tlist cost throughout
1294 * query_planner() and subroutines. But now we've inserted a
1295 * tlist that might contain actual operators, sub-selects, etc
1296 * --- so we'd better account for its cost.
1298 * Below this point, any tlist eval cost for added-on nodes
1299 * should be accounted for as we create those nodes.
1300 * Presently, of the node types we can add on, only Agg,
1301 * WindowAgg, and Group project new tlists (the rest just copy
1302 * their input tuples) --- so make_agg(), make_windowagg() and
1303 * make_group() are responsible for computing the added cost.
1305 cost_qual_eval(&tlist_cost, sub_tlist, root);
1306 result_plan->startup_cost += tlist_cost.startup;
1307 result_plan->total_cost += tlist_cost.startup +
1308 tlist_cost.per_tuple * result_plan->plan_rows;
1313 * Since we're using query_planner's tlist and not the one
1314 * make_subplanTargetList calculated, we have to refigure any
1315 * grouping-column indexes make_subplanTargetList computed.
1317 locate_grouping_columns(root, tlist, result_plan->targetlist,
1322 * Insert AGG or GROUP node if needed, plus an explicit sort step
1325 * HAVING clause, if any, becomes qual of the Agg or Group node.
1327 if (use_hashed_grouping)
1329 /* Hashed aggregate plan --- no sort needed */
1330 result_plan = (Plan *) make_agg(root,
1332 (List *) parse->havingQual,
1337 extract_grouping_ops(parse->groupClause),
1340 /* Hashed aggregation produces randomly-ordered results */
1341 current_pathkeys = NIL;
1343 else if (parse->hasAggs)
1345 /* Plain aggregate plan --- sort if needed */
1346 AggStrategy aggstrategy;
1348 if (parse->groupClause)
1350 if (need_sort_for_grouping)
1352 result_plan = (Plan *)
1353 make_sort_from_groupcols(root,
1357 current_pathkeys = root->group_pathkeys;
1359 aggstrategy = AGG_SORTED;
1362 * The AGG node will not change the sort ordering of its
1363 * groups, so current_pathkeys describes the result too.
1368 aggstrategy = AGG_PLAIN;
1369 /* Result will be only one row anyway; no sort order */
1370 current_pathkeys = NIL;
1373 result_plan = (Plan *) make_agg(root,
1375 (List *) parse->havingQual,
1380 extract_grouping_ops(parse->groupClause),
1384 else if (parse->groupClause)
1387 * GROUP BY without aggregation, so insert a group node (plus
1388 * the appropriate sort node, if necessary).
1390 * Add an explicit sort if we couldn't make the path come out
1391 * the way the GROUP node needs it.
1393 if (need_sort_for_grouping)
1395 result_plan = (Plan *)
1396 make_sort_from_groupcols(root,
1400 current_pathkeys = root->group_pathkeys;
1403 result_plan = (Plan *) make_group(root,
1405 (List *) parse->havingQual,
1408 extract_grouping_ops(parse->groupClause),
1411 /* The Group node won't change sort ordering */
1413 else if (root->hasHavingQual)
1416 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1417 * This is a degenerate case in which we are supposed to emit
1418 * either 0 or 1 row depending on whether HAVING succeeds.
1419 * Furthermore, there cannot be any variables in either HAVING
1420 * or the targetlist, so we actually do not need the FROM
1421 * table at all! We can just throw away the plan-so-far and
1422 * generate a Result node. This is a sufficiently unusual
1423 * corner case that it's not worth contorting the structure of
1424 * this routine to avoid having to generate the plan in the
1427 result_plan = (Plan *) make_result(root,
1432 } /* end of non-minmax-aggregate case */
1435 * Since each window function could require a different sort order, we
1436 * stack up a WindowAgg node for each window, with sort steps between
1445 * If the top-level plan node is one that cannot do expression
1446 * evaluation, we must insert a Result node to project the desired
1447 * tlist. (In some cases this might not really be required, but
1448 * it's not worth trying to avoid it.) Note that on second and
1449 * subsequent passes through the following loop, the top-level
1450 * node will be a WindowAgg which we know can project; so we only
1451 * need to check once.
1453 if (!is_projection_capable_plan(result_plan))
1455 result_plan = (Plan *) make_result(root,
1462 * The "base" targetlist for all steps of the windowing process is
1463 * a flat tlist of all Vars and Aggs needed in the result. (In
1464 * some cases we wouldn't need to propagate all of these all the
1465 * way to the top, since they might only be needed as inputs to
1466 * WindowFuncs. It's probably not worth trying to optimize that
1467 * though.) We also need any volatile sort expressions, because
1468 * make_sort_from_pathkeys won't add those on its own, and anyway
1469 * we want them evaluated only once at the bottom of the stack. As
1470 * we climb up the stack, we add outputs for the WindowFuncs
1471 * computed at each level. Also, each input tlist has to present
1472 * all the columns needed to sort the data for the next WindowAgg
1473 * step. That's handled internally by make_sort_from_pathkeys,
1474 * but we need the copyObject steps here to ensure that each plan
1475 * node has a separately modifiable tlist.
1477 window_tlist = flatten_tlist(tlist);
1479 window_tlist = add_to_flat_tlist(window_tlist,
1480 pull_agg_clause((Node *) tlist));
1481 window_tlist = add_volatile_sort_exprs(window_tlist, tlist,
1483 result_plan->targetlist = (List *) copyObject(window_tlist);
1485 foreach(l, activeWindows)
1487 WindowClause *wc = (WindowClause *) lfirst(l);
1488 List *window_pathkeys;
1490 AttrNumber *partColIdx;
1493 AttrNumber *ordColIdx;
1496 window_pathkeys = make_pathkeys_for_window(root,
1502 * This is a bit tricky: we build a sort node even if we don't
1503 * really have to sort. Even when no explicit sort is needed,
1504 * we need to have suitable resjunk items added to the input
1505 * plan's tlist for any partitioning or ordering columns that
1506 * aren't plain Vars. Furthermore, this way we can use
1507 * existing infrastructure to identify which input columns are
1508 * the interesting ones.
1510 if (window_pathkeys)
1514 sort_plan = make_sort_from_pathkeys(root,
1518 if (!pathkeys_contained_in(window_pathkeys,
1521 /* we do indeed need to sort */
1522 result_plan = (Plan *) sort_plan;
1523 current_pathkeys = window_pathkeys;
1525 /* In either case, extract the per-column information */
1526 get_column_info_for_window(root, wc, tlist,
1528 sort_plan->sortColIdx,
1538 /* empty window specification, nothing to sort */
1541 partOperators = NULL;
1544 ordOperators = NULL;
1549 /* Add the current WindowFuncs to the running tlist */
1550 window_tlist = add_to_flat_tlist(window_tlist,
1551 wflists->windowFuncs[wc->winref]);
1555 /* Install the original tlist in the topmost WindowAgg */
1556 window_tlist = tlist;
1559 /* ... and make the WindowAgg plan node */
1560 result_plan = (Plan *)
1561 make_windowagg(root,
1562 (List *) copyObject(window_tlist),
1563 wflists->windowFuncs[wc->winref],
1577 } /* end of if (setOperations) */
1580 * If there is a DISTINCT clause, add the necessary node(s).
1582 if (parse->distinctClause)
1584 double dNumDistinctRows;
1585 long numDistinctRows;
1588 * If there was grouping or aggregation, use the current number of
1589 * rows as the estimated number of DISTINCT rows (ie, assume the
1590 * result was already mostly unique). If not, use the number of
1591 * distinct-groups calculated by query_planner.
1593 if (parse->groupClause || root->hasHavingQual || parse->hasAggs)
1594 dNumDistinctRows = result_plan->plan_rows;
1596 dNumDistinctRows = dNumGroups;
1598 /* Also convert to long int --- but 'ware overflow! */
1599 numDistinctRows = (long) Min(dNumDistinctRows, (double) LONG_MAX);
1601 /* Choose implementation method if we didn't already */
1602 if (!tested_hashed_distinct)
1605 * At this point, either hashed or sorted grouping will have to
1606 * work from result_plan, so we pass that as both "cheapest" and
1609 use_hashed_distinct =
1610 choose_hashed_distinct(root,
1611 tuple_fraction, limit_tuples,
1612 result_plan->plan_rows,
1613 result_plan->plan_width,
1614 result_plan->startup_cost,
1615 result_plan->total_cost,
1616 result_plan->startup_cost,
1617 result_plan->total_cost,
1622 if (use_hashed_distinct)
1624 /* Hashed aggregate plan --- no sort needed */
1625 result_plan = (Plan *) make_agg(root,
1626 result_plan->targetlist,
1630 list_length(parse->distinctClause),
1631 extract_grouping_cols(parse->distinctClause,
1632 result_plan->targetlist),
1633 extract_grouping_ops(parse->distinctClause),
1636 /* Hashed aggregation produces randomly-ordered results */
1637 current_pathkeys = NIL;
1642 * Use a Unique node to implement DISTINCT. Add an explicit sort
1643 * if we couldn't make the path come out the way the Unique node
1644 * needs it. If we do have to sort, always sort by the more
1645 * rigorous of DISTINCT and ORDER BY, to avoid a second sort
1646 * below. However, for regular DISTINCT, don't sort now if we
1647 * don't have to --- sorting afterwards will likely be cheaper,
1648 * and also has the possibility of optimizing via LIMIT. But for
1649 * DISTINCT ON, we *must* force the final sort now, else it won't
1650 * have the desired behavior.
1652 List *needed_pathkeys;
1654 if (parse->hasDistinctOn &&
1655 list_length(root->distinct_pathkeys) <
1656 list_length(root->sort_pathkeys))
1657 needed_pathkeys = root->sort_pathkeys;
1659 needed_pathkeys = root->distinct_pathkeys;
1661 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
1663 if (list_length(root->distinct_pathkeys) >=
1664 list_length(root->sort_pathkeys))
1665 current_pathkeys = root->distinct_pathkeys;
1668 current_pathkeys = root->sort_pathkeys;
1669 /* Assert checks that parser didn't mess up... */
1670 Assert(pathkeys_contained_in(root->distinct_pathkeys,
1674 result_plan = (Plan *) make_sort_from_pathkeys(root,
1680 result_plan = (Plan *) make_unique(result_plan,
1681 parse->distinctClause);
1682 result_plan->plan_rows = dNumDistinctRows;
1683 /* The Unique node won't change sort ordering */
1688 * If ORDER BY was given and we were not able to make the plan come out in
1689 * the right order, add an explicit sort step.
1691 if (parse->sortClause)
1693 if (!pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1695 result_plan = (Plan *) make_sort_from_pathkeys(root,
1697 root->sort_pathkeys,
1699 current_pathkeys = root->sort_pathkeys;
1704 * If there is a FOR UPDATE/SHARE clause, add the LockRows node. (Note: we
1705 * intentionally test parse->rowMarks not root->rowMarks here. If there
1706 * are only non-locking rowmarks, they should be handled by the
1707 * ModifyTable node instead.)
1709 if (parse->rowMarks)
1711 result_plan = (Plan *) make_lockrows(result_plan,
1713 SS_assign_special_param(root));
1716 * The result can no longer be assumed sorted, since locking might
1717 * cause the sort key columns to be replaced with new values.
1719 current_pathkeys = NIL;
1723 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1725 if (parse->limitCount || parse->limitOffset)
1727 result_plan = (Plan *) make_limit(result_plan,
1735 * Return the actual output ordering in query_pathkeys for possible use by
1736 * an outer query level.
1738 root->query_pathkeys = current_pathkeys;
1744 * Detect whether a plan node is a "dummy" plan created when a relation
1745 * is deemed not to need scanning due to constraint exclusion.
1747 * Currently, such dummy plans are Result nodes with constant FALSE
1751 is_dummy_plan(Plan *plan)
1753 if (IsA(plan, Result))
1755 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1757 if (list_length(rcqual) == 1)
1759 Const *constqual = (Const *) linitial(rcqual);
1761 if (constqual && IsA(constqual, Const))
1763 if (!constqual->constisnull &&
1764 !DatumGetBool(constqual->constvalue))
1773 * Create a bitmapset of the RT indexes of live base relations
1775 * Helper for preprocess_rowmarks ... at this point in the proceedings,
1776 * the only good way to distinguish baserels from appendrel children
1777 * is to see what is in the join tree.
1780 get_base_rel_indexes(Node *jtnode)
1786 if (IsA(jtnode, RangeTblRef))
1788 int varno = ((RangeTblRef *) jtnode)->rtindex;
1790 result = bms_make_singleton(varno);
1792 else if (IsA(jtnode, FromExpr))
1794 FromExpr *f = (FromExpr *) jtnode;
1798 foreach(l, f->fromlist)
1799 result = bms_join(result,
1800 get_base_rel_indexes(lfirst(l)));
1802 else if (IsA(jtnode, JoinExpr))
1804 JoinExpr *j = (JoinExpr *) jtnode;
1806 result = bms_join(get_base_rel_indexes(j->larg),
1807 get_base_rel_indexes(j->rarg));
1811 elog(ERROR, "unrecognized node type: %d",
1812 (int) nodeTag(jtnode));
1813 result = NULL; /* keep compiler quiet */
1819 * preprocess_rowmarks - set up PlanRowMarks if needed
1822 preprocess_rowmarks(PlannerInfo *root)
1824 Query *parse = root->parse;
1830 if (parse->rowMarks)
1833 * We've got trouble if FOR UPDATE/SHARE appears inside grouping,
1834 * since grouping renders a reference to individual tuple CTIDs
1835 * invalid. This is also checked at parse time, but that's
1836 * insufficient because of rule substitution, query pullup, etc.
1838 CheckSelectLocking(parse);
1843 * We only need rowmarks for UPDATE, DELETE, or FOR UPDATE/SHARE.
1845 if (parse->commandType != CMD_UPDATE &&
1846 parse->commandType != CMD_DELETE)
1851 * We need to have rowmarks for all base relations except the target. We
1852 * make a bitmapset of all base rels and then remove the items we don't
1853 * need or have FOR UPDATE/SHARE marks for.
1855 rels = get_base_rel_indexes((Node *) parse->jointree);
1856 if (parse->resultRelation)
1857 rels = bms_del_member(rels, parse->resultRelation);
1860 * Convert RowMarkClauses to PlanRowMark representation.
1863 foreach(l, parse->rowMarks)
1865 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
1866 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
1870 * Currently, it is syntactically impossible to have FOR UPDATE
1871 * applied to an update/delete target rel. If that ever becomes
1872 * possible, we should drop the target from the PlanRowMark list.
1874 Assert(rc->rti != parse->resultRelation);
1877 * Ignore RowMarkClauses for subqueries; they aren't real tables and
1878 * can't support true locking. Subqueries that got flattened into the
1879 * main query should be ignored completely. Any that didn't will get
1880 * ROW_MARK_COPY items in the next loop.
1882 if (rte->rtekind != RTE_RELATION)
1885 rels = bms_del_member(rels, rc->rti);
1887 newrc = makeNode(PlanRowMark);
1888 newrc->rti = newrc->prti = rc->rti;
1889 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
1891 newrc->markType = ROW_MARK_EXCLUSIVE;
1893 newrc->markType = ROW_MARK_SHARE;
1894 newrc->noWait = rc->noWait;
1895 newrc->isParent = false;
1897 prowmarks = lappend(prowmarks, newrc);
1901 * Now, add rowmarks for any non-target, non-locked base relations.
1904 foreach(l, parse->rtable)
1906 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
1910 if (!bms_is_member(i, rels))
1913 newrc = makeNode(PlanRowMark);
1914 newrc->rti = newrc->prti = i;
1915 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
1916 /* real tables support REFERENCE, anything else needs COPY */
1917 if (rte->rtekind == RTE_RELATION &&
1918 rte->relkind != RELKIND_FOREIGN_TABLE)
1919 newrc->markType = ROW_MARK_REFERENCE;
1921 newrc->markType = ROW_MARK_COPY;
1922 newrc->noWait = false; /* doesn't matter */
1923 newrc->isParent = false;
1925 prowmarks = lappend(prowmarks, newrc);
1928 root->rowMarks = prowmarks;
1932 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1934 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1935 * results back in *count_est and *offset_est. These variables are set to
1936 * 0 if the corresponding clause is not present, and -1 if it's present
1937 * but we couldn't estimate the value for it. (The "0" convention is OK
1938 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1939 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1940 * usual practice of never estimating less than one row.) These values will
1941 * be passed to make_limit, which see if you change this code.
1943 * The return value is the suitably adjusted tuple_fraction to use for
1944 * planning the query. This adjustment is not overridable, since it reflects
1945 * plan actions that grouping_planner() will certainly take, not assumptions
1949 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1950 int64 *offset_est, int64 *count_est)
1952 Query *parse = root->parse;
1954 double limit_fraction;
1956 /* Should not be called unless LIMIT or OFFSET */
1957 Assert(parse->limitCount || parse->limitOffset);
1960 * Try to obtain the clause values. We use estimate_expression_value
1961 * primarily because it can sometimes do something useful with Params.
1963 if (parse->limitCount)
1965 est = estimate_expression_value(root, parse->limitCount);
1966 if (est && IsA(est, Const))
1968 if (((Const *) est)->constisnull)
1970 /* NULL indicates LIMIT ALL, ie, no limit */
1971 *count_est = 0; /* treat as not present */
1975 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1976 if (*count_est <= 0)
1977 *count_est = 1; /* force to at least 1 */
1981 *count_est = -1; /* can't estimate */
1984 *count_est = 0; /* not present */
1986 if (parse->limitOffset)
1988 est = estimate_expression_value(root, parse->limitOffset);
1989 if (est && IsA(est, Const))
1991 if (((Const *) est)->constisnull)
1993 /* Treat NULL as no offset; the executor will too */
1994 *offset_est = 0; /* treat as not present */
1998 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1999 if (*offset_est < 0)
2000 *offset_est = 0; /* less than 0 is same as 0 */
2004 *offset_est = -1; /* can't estimate */
2007 *offset_est = 0; /* not present */
2009 if (*count_est != 0)
2012 * A LIMIT clause limits the absolute number of tuples returned.
2013 * However, if it's not a constant LIMIT then we have to guess; for
2014 * lack of a better idea, assume 10% of the plan's result is wanted.
2016 if (*count_est < 0 || *offset_est < 0)
2018 /* LIMIT or OFFSET is an expression ... punt ... */
2019 limit_fraction = 0.10;
2023 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2024 limit_fraction = (double) *count_est + (double) *offset_est;
2028 * If we have absolute limits from both caller and LIMIT, use the
2029 * smaller value; likewise if they are both fractional. If one is
2030 * fractional and the other absolute, we can't easily determine which
2031 * is smaller, but we use the heuristic that the absolute will usually
2034 if (tuple_fraction >= 1.0)
2036 if (limit_fraction >= 1.0)
2039 tuple_fraction = Min(tuple_fraction, limit_fraction);
2043 /* caller absolute, limit fractional; use caller's value */
2046 else if (tuple_fraction > 0.0)
2048 if (limit_fraction >= 1.0)
2050 /* caller fractional, limit absolute; use limit */
2051 tuple_fraction = limit_fraction;
2055 /* both fractional */
2056 tuple_fraction = Min(tuple_fraction, limit_fraction);
2061 /* no info from caller, just use limit */
2062 tuple_fraction = limit_fraction;
2065 else if (*offset_est != 0 && tuple_fraction > 0.0)
2068 * We have an OFFSET but no LIMIT. This acts entirely differently
2069 * from the LIMIT case: here, we need to increase rather than decrease
2070 * the caller's tuple_fraction, because the OFFSET acts to cause more
2071 * tuples to be fetched instead of fewer. This only matters if we got
2072 * a tuple_fraction > 0, however.
2074 * As above, use 10% if OFFSET is present but unestimatable.
2076 if (*offset_est < 0)
2077 limit_fraction = 0.10;
2079 limit_fraction = (double) *offset_est;
2082 * If we have absolute counts from both caller and OFFSET, add them
2083 * together; likewise if they are both fractional. If one is
2084 * fractional and the other absolute, we want to take the larger, and
2085 * we heuristically assume that's the fractional one.
2087 if (tuple_fraction >= 1.0)
2089 if (limit_fraction >= 1.0)
2091 /* both absolute, so add them together */
2092 tuple_fraction += limit_fraction;
2096 /* caller absolute, limit fractional; use limit */
2097 tuple_fraction = limit_fraction;
2102 if (limit_fraction >= 1.0)
2104 /* caller fractional, limit absolute; use caller's value */
2108 /* both fractional, so add them together */
2109 tuple_fraction += limit_fraction;
2110 if (tuple_fraction >= 1.0)
2111 tuple_fraction = 0.0; /* assume fetch all */
2116 return tuple_fraction;
2121 * preprocess_groupclause - do preparatory work on GROUP BY clause
2123 * The idea here is to adjust the ordering of the GROUP BY elements
2124 * (which in itself is semantically insignificant) to match ORDER BY,
2125 * thereby allowing a single sort operation to both implement the ORDER BY
2126 * requirement and set up for a Unique step that implements GROUP BY.
2128 * In principle it might be interesting to consider other orderings of the
2129 * GROUP BY elements, which could match the sort ordering of other
2130 * possible plans (eg an indexscan) and thereby reduce cost. We don't
2131 * bother with that, though. Hashed grouping will frequently win anyway.
2133 * Note: we need no comparable processing of the distinctClause because
2134 * the parser already enforced that that matches ORDER BY.
2137 preprocess_groupclause(PlannerInfo *root)
2139 Query *parse = root->parse;
2140 List *new_groupclause;
2145 /* If no ORDER BY, nothing useful to do here */
2146 if (parse->sortClause == NIL)
2150 * Scan the ORDER BY clause and construct a list of matching GROUP BY
2151 * items, but only as far as we can make a matching prefix.
2153 * This code assumes that the sortClause contains no duplicate items.
2155 new_groupclause = NIL;
2156 foreach(sl, parse->sortClause)
2158 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
2160 foreach(gl, parse->groupClause)
2162 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2166 new_groupclause = lappend(new_groupclause, gc);
2171 break; /* no match, so stop scanning */
2174 /* Did we match all of the ORDER BY list, or just some of it? */
2175 partial_match = (sl != NULL);
2177 /* If no match at all, no point in reordering GROUP BY */
2178 if (new_groupclause == NIL)
2182 * Add any remaining GROUP BY items to the new list, but only if we were
2183 * able to make a complete match. In other words, we only rearrange the
2184 * GROUP BY list if the result is that one list is a prefix of the other
2185 * --- otherwise there's no possibility of a common sort. Also, give up
2186 * if there are any non-sortable GROUP BY items, since then there's no
2189 foreach(gl, parse->groupClause)
2191 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2193 if (list_member_ptr(new_groupclause, gc))
2194 continue; /* it matched an ORDER BY item */
2196 return; /* give up, no common sort possible */
2197 if (!OidIsValid(gc->sortop))
2198 return; /* give up, GROUP BY can't be sorted */
2199 new_groupclause = lappend(new_groupclause, gc);
2202 /* Success --- install the rearranged GROUP BY list */
2203 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2204 parse->groupClause = new_groupclause;
2208 * choose_hashed_grouping - should we use hashed grouping?
2210 * Returns TRUE to select hashing, FALSE to select sorting.
2213 choose_hashed_grouping(PlannerInfo *root,
2214 double tuple_fraction, double limit_tuples,
2215 double path_rows, int path_width,
2216 Path *cheapest_path, Path *sorted_path,
2217 double dNumGroups, AggClauseCosts *agg_costs)
2219 Query *parse = root->parse;
2220 int numGroupCols = list_length(parse->groupClause);
2224 List *target_pathkeys;
2225 List *current_pathkeys;
2230 * Executor doesn't support hashed aggregation with DISTINCT or ORDER BY
2231 * aggregates. (Doing so would imply storing *all* the input values in
2232 * the hash table, and/or running many sorts in parallel, either of which
2233 * seems like a certain loser.)
2235 can_hash = (agg_costs->numOrderedAggs == 0 &&
2236 grouping_is_hashable(parse->groupClause));
2237 can_sort = grouping_is_sortable(parse->groupClause);
2239 /* Quick out if only one choice is workable */
2240 if (!(can_hash && can_sort))
2248 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2249 errmsg("could not implement GROUP BY"),
2250 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2253 /* Prefer sorting when enable_hashagg is off */
2254 if (!enable_hashagg)
2258 * Don't do it if it doesn't look like the hashtable will fit into
2262 /* Estimate per-hash-entry space at tuple width... */
2263 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2264 /* plus space for pass-by-ref transition values... */
2265 hashentrysize += agg_costs->transitionSpace;
2266 /* plus the per-hash-entry overhead */
2267 hashentrysize += hash_agg_entry_size(agg_costs->numAggs);
2269 if (hashentrysize * dNumGroups > work_mem * 1024L)
2273 * When we have both GROUP BY and DISTINCT, use the more-rigorous of
2274 * DISTINCT and ORDER BY as the assumed required output sort order. This
2275 * is an oversimplification because the DISTINCT might get implemented via
2276 * hashing, but it's not clear that the case is common enough (or that our
2277 * estimates are good enough) to justify trying to solve it exactly.
2279 if (list_length(root->distinct_pathkeys) >
2280 list_length(root->sort_pathkeys))
2281 target_pathkeys = root->distinct_pathkeys;
2283 target_pathkeys = root->sort_pathkeys;
2286 * See if the estimated cost is no more than doing it the other way. While
2287 * avoiding the need for sorted input is usually a win, the fact that the
2288 * output won't be sorted may be a loss; so we need to do an actual cost
2291 * We need to consider cheapest_path + hashagg [+ final sort] versus
2292 * either cheapest_path [+ sort] + group or agg [+ final sort] or
2293 * presorted_path + group or agg [+ final sort] where brackets indicate a
2294 * step that may not be needed. We assume query_planner() will have
2295 * returned a presorted path only if it's a winner compared to
2296 * cheapest_path for this purpose.
2298 * These path variables are dummies that just hold cost fields; we don't
2299 * make actual Paths for these steps.
2301 cost_agg(&hashed_p, root, AGG_HASHED, agg_costs,
2302 numGroupCols, dNumGroups,
2303 cheapest_path->startup_cost, cheapest_path->total_cost,
2305 /* Result of hashed agg is always unsorted */
2306 if (target_pathkeys)
2307 cost_sort(&hashed_p, root, target_pathkeys, hashed_p.total_cost,
2308 dNumGroups, path_width,
2309 0.0, work_mem, limit_tuples);
2313 sorted_p.startup_cost = sorted_path->startup_cost;
2314 sorted_p.total_cost = sorted_path->total_cost;
2315 current_pathkeys = sorted_path->pathkeys;
2319 sorted_p.startup_cost = cheapest_path->startup_cost;
2320 sorted_p.total_cost = cheapest_path->total_cost;
2321 current_pathkeys = cheapest_path->pathkeys;
2323 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
2325 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
2326 path_rows, path_width,
2327 0.0, work_mem, -1.0);
2328 current_pathkeys = root->group_pathkeys;
2332 cost_agg(&sorted_p, root, AGG_SORTED, agg_costs,
2333 numGroupCols, dNumGroups,
2334 sorted_p.startup_cost, sorted_p.total_cost,
2337 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
2338 sorted_p.startup_cost, sorted_p.total_cost,
2340 /* The Agg or Group node will preserve ordering */
2341 if (target_pathkeys &&
2342 !pathkeys_contained_in(target_pathkeys, current_pathkeys))
2343 cost_sort(&sorted_p, root, target_pathkeys, sorted_p.total_cost,
2344 dNumGroups, path_width,
2345 0.0, work_mem, limit_tuples);
2348 * Now make the decision using the top-level tuple fraction. First we
2349 * have to convert an absolute count (LIMIT) into fractional form.
2351 if (tuple_fraction >= 1.0)
2352 tuple_fraction /= dNumGroups;
2354 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2355 tuple_fraction) < 0)
2357 /* Hashed is cheaper, so use it */
2364 * choose_hashed_distinct - should we use hashing for DISTINCT?
2366 * This is fairly similar to choose_hashed_grouping, but there are enough
2367 * differences that it doesn't seem worth trying to unify the two functions.
2368 * (One difference is that we sometimes apply this after forming a Plan,
2369 * so the input alternatives can't be represented as Paths --- instead we
2370 * pass in the costs as individual variables.)
2372 * But note that making the two choices independently is a bit bogus in
2373 * itself. If the two could be combined into a single choice operation
2374 * it'd probably be better, but that seems far too unwieldy to be practical,
2375 * especially considering that the combination of GROUP BY and DISTINCT
2376 * isn't very common in real queries. By separating them, we are giving
2377 * extra preference to using a sorting implementation when a common sort key
2378 * is available ... and that's not necessarily wrong anyway.
2380 * Returns TRUE to select hashing, FALSE to select sorting.
2383 choose_hashed_distinct(PlannerInfo *root,
2384 double tuple_fraction, double limit_tuples,
2385 double path_rows, int path_width,
2386 Cost cheapest_startup_cost, Cost cheapest_total_cost,
2387 Cost sorted_startup_cost, Cost sorted_total_cost,
2388 List *sorted_pathkeys,
2389 double dNumDistinctRows)
2391 Query *parse = root->parse;
2392 int numDistinctCols = list_length(parse->distinctClause);
2396 List *current_pathkeys;
2397 List *needed_pathkeys;
2402 * If we have a sortable DISTINCT ON clause, we always use sorting. This
2403 * enforces the expected behavior of DISTINCT ON.
2405 can_sort = grouping_is_sortable(parse->distinctClause);
2406 if (can_sort && parse->hasDistinctOn)
2409 can_hash = grouping_is_hashable(parse->distinctClause);
2411 /* Quick out if only one choice is workable */
2412 if (!(can_hash && can_sort))
2420 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2421 errmsg("could not implement DISTINCT"),
2422 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2425 /* Prefer sorting when enable_hashagg is off */
2426 if (!enable_hashagg)
2430 * Don't do it if it doesn't look like the hashtable will fit into
2433 hashentrysize = MAXALIGN(path_width) + MAXALIGN(sizeof(MinimalTupleData));
2435 if (hashentrysize * dNumDistinctRows > work_mem * 1024L)
2439 * See if the estimated cost is no more than doing it the other way. While
2440 * avoiding the need for sorted input is usually a win, the fact that the
2441 * output won't be sorted may be a loss; so we need to do an actual cost
2444 * We need to consider cheapest_path + hashagg [+ final sort] versus
2445 * sorted_path [+ sort] + group [+ final sort] where brackets indicate a
2446 * step that may not be needed.
2448 * These path variables are dummies that just hold cost fields; we don't
2449 * make actual Paths for these steps.
2451 cost_agg(&hashed_p, root, AGG_HASHED, NULL,
2452 numDistinctCols, dNumDistinctRows,
2453 cheapest_startup_cost, cheapest_total_cost,
2457 * Result of hashed agg is always unsorted, so if ORDER BY is present we
2458 * need to charge for the final sort.
2460 if (parse->sortClause)
2461 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
2462 dNumDistinctRows, path_width,
2463 0.0, work_mem, limit_tuples);
2466 * Now for the GROUP case. See comments in grouping_planner about the
2467 * sorting choices here --- this code should match that code.
2469 sorted_p.startup_cost = sorted_startup_cost;
2470 sorted_p.total_cost = sorted_total_cost;
2471 current_pathkeys = sorted_pathkeys;
2472 if (parse->hasDistinctOn &&
2473 list_length(root->distinct_pathkeys) <
2474 list_length(root->sort_pathkeys))
2475 needed_pathkeys = root->sort_pathkeys;
2477 needed_pathkeys = root->distinct_pathkeys;
2478 if (!pathkeys_contained_in(needed_pathkeys, current_pathkeys))
2480 if (list_length(root->distinct_pathkeys) >=
2481 list_length(root->sort_pathkeys))
2482 current_pathkeys = root->distinct_pathkeys;
2484 current_pathkeys = root->sort_pathkeys;
2485 cost_sort(&sorted_p, root, current_pathkeys, sorted_p.total_cost,
2486 path_rows, path_width,
2487 0.0, work_mem, -1.0);
2489 cost_group(&sorted_p, root, numDistinctCols, dNumDistinctRows,
2490 sorted_p.startup_cost, sorted_p.total_cost,
2492 if (parse->sortClause &&
2493 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
2494 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
2495 dNumDistinctRows, path_width,
2496 0.0, work_mem, limit_tuples);
2499 * Now make the decision using the top-level tuple fraction. First we
2500 * have to convert an absolute count (LIMIT) into fractional form.
2502 if (tuple_fraction >= 1.0)
2503 tuple_fraction /= dNumDistinctRows;
2505 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
2506 tuple_fraction) < 0)
2508 /* Hashed is cheaper, so use it */
2515 * make_subplanTargetList
2516 * Generate appropriate target list when grouping is required.
2518 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
2519 * above the result of query_planner, we typically want to pass a different
2520 * target list to query_planner than the outer plan nodes should have.
2521 * This routine generates the correct target list for the subplan.
2523 * The initial target list passed from the parser already contains entries
2524 * for all ORDER BY and GROUP BY expressions, but it will not have entries
2525 * for variables used only in HAVING clauses; so we need to add those
2526 * variables to the subplan target list. Also, we flatten all expressions
2527 * except GROUP BY items into their component variables; the other expressions
2528 * will be computed by the inserted nodes rather than by the subplan.
2529 * For example, given a query like
2530 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
2531 * we want to pass this targetlist to the subplan:
2533 * where the a+b target will be used by the Sort/Group steps, and the
2534 * other targets will be used for computing the final results. (In the
2535 * above example we could theoretically suppress the a and b targets and
2536 * pass down only c,d,a+b, but it's not really worth the trouble to
2537 * eliminate simple var references from the subplan. We will avoid doing
2538 * the extra computation to recompute a+b at the outer level; see
2539 * fix_upper_expr() in setrefs.c.)
2541 * If we are grouping or aggregating, *and* there are no non-Var grouping
2542 * expressions, then the returned tlist is effectively dummy; we do not
2543 * need to force it to be evaluated, because all the Vars it contains
2544 * should be present in the output of query_planner anyway.
2546 * 'tlist' is the query's target list.
2547 * 'groupColIdx' receives an array of column numbers for the GROUP BY
2548 * expressions (if there are any) in the subplan's target list.
2549 * 'need_tlist_eval' is set true if we really need to evaluate the
2552 * The result is the targetlist to be passed to the subplan.
2555 make_subplanTargetList(PlannerInfo *root,
2557 AttrNumber **groupColIdx,
2558 bool *need_tlist_eval)
2560 Query *parse = root->parse;
2565 *groupColIdx = NULL;
2568 * If we're not grouping or aggregating, there's nothing to do here;
2569 * query_planner should receive the unmodified target list.
2571 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual &&
2572 !parse->hasWindowFuncs)
2574 *need_tlist_eval = true;
2579 * Otherwise, start with a "flattened" tlist (having just the vars
2580 * mentioned in the targetlist and HAVING qual --- but not upper-level
2581 * Vars; they will be replaced by Params later on). Note this includes
2582 * vars used in resjunk items, so we are covering the needs of ORDER BY
2583 * and window specifications.
2585 sub_tlist = flatten_tlist(tlist);
2586 extravars = pull_var_clause(parse->havingQual, PVC_INCLUDE_PLACEHOLDERS);
2587 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
2588 list_free(extravars);
2589 *need_tlist_eval = false; /* only eval if not flat tlist */
2592 * If grouping, create sub_tlist entries for all GROUP BY expressions
2593 * (GROUP BY items that are simple Vars should be in the list already),
2594 * and make an array showing where the group columns are in the sub_tlist.
2596 numCols = list_length(parse->groupClause);
2600 AttrNumber *grpColIdx;
2603 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
2604 *groupColIdx = grpColIdx;
2606 foreach(gl, parse->groupClause)
2608 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2609 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2613 * Find or make a matching sub_tlist entry. If the groupexpr
2614 * isn't a Var, no point in searching. (Note that the parser
2615 * won't make multiple groupClause entries for the same TLE.)
2617 if (groupexpr && IsA(groupexpr, Var))
2618 te = tlist_member(groupexpr, sub_tlist);
2624 te = makeTargetEntry((Expr *) groupexpr,
2625 list_length(sub_tlist) + 1,
2628 sub_tlist = lappend(sub_tlist, te);
2629 *need_tlist_eval = true; /* it's not flat anymore */
2632 /* and save its resno */
2633 grpColIdx[keyno++] = te->resno;
2641 * locate_grouping_columns
2642 * Locate grouping columns in the tlist chosen by query_planner.
2644 * This is only needed if we don't use the sub_tlist chosen by
2645 * make_subplanTargetList. We have to forget the column indexes found
2646 * by that routine and re-locate the grouping exprs in the real sub_tlist.
2649 locate_grouping_columns(PlannerInfo *root,
2652 AttrNumber *groupColIdx)
2658 * No work unless grouping.
2660 if (!root->parse->groupClause)
2662 Assert(groupColIdx == NULL);
2665 Assert(groupColIdx != NULL);
2667 foreach(gl, root->parse->groupClause)
2669 SortGroupClause *grpcl = (SortGroupClause *) lfirst(gl);
2670 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
2671 TargetEntry *te = tlist_member(groupexpr, sub_tlist);
2674 elog(ERROR, "failed to locate grouping columns");
2675 groupColIdx[keyno++] = te->resno;
2680 * postprocess_setop_tlist
2681 * Fix up targetlist returned by plan_set_operations().
2683 * We need to transpose sort key info from the orig_tlist into new_tlist.
2684 * NOTE: this would not be good enough if we supported resjunk sort keys
2685 * for results of set operations --- then, we'd need to project a whole
2686 * new tlist to evaluate the resjunk columns. For now, just ereport if we
2687 * find any resjunk columns in orig_tlist.
2690 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
2693 ListCell *orig_tlist_item = list_head(orig_tlist);
2695 foreach(l, new_tlist)
2697 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
2698 TargetEntry *orig_tle;
2700 /* ignore resjunk columns in setop result */
2701 if (new_tle->resjunk)
2704 Assert(orig_tlist_item != NULL);
2705 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
2706 orig_tlist_item = lnext(orig_tlist_item);
2707 if (orig_tle->resjunk) /* should not happen */
2708 elog(ERROR, "resjunk output columns are not implemented");
2709 Assert(new_tle->resno == orig_tle->resno);
2710 new_tle->ressortgroupref = orig_tle->ressortgroupref;
2712 if (orig_tlist_item != NULL)
2713 elog(ERROR, "resjunk output columns are not implemented");
2718 * select_active_windows
2719 * Create a list of the "active" window clauses (ie, those referenced
2720 * by non-deleted WindowFuncs) in the order they are to be executed.
2723 select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
2729 /* First, make a list of the active windows */
2731 foreach(lc, root->parse->windowClause)
2733 WindowClause *wc = (WindowClause *) lfirst(lc);
2735 /* It's only active if wflists shows some related WindowFuncs */
2736 Assert(wc->winref <= wflists->maxWinRef);
2737 if (wflists->windowFuncs[wc->winref] != NIL)
2738 actives = lappend(actives, wc);
2742 * Now, ensure that windows with identical partitioning/ordering clauses
2743 * are adjacent in the list. This is required by the SQL standard, which
2744 * says that only one sort is to be used for such windows, even if they
2745 * are otherwise distinct (eg, different names or framing clauses).
2747 * There is room to be much smarter here, for example detecting whether
2748 * one window's sort keys are a prefix of another's (so that sorting for
2749 * the latter would do for the former), or putting windows first that
2750 * match a sort order available for the underlying query. For the moment
2751 * we are content with meeting the spec.
2754 while (actives != NIL)
2756 WindowClause *wc = (WindowClause *) linitial(actives);
2760 /* Move wc from actives to result */
2761 actives = list_delete_first(actives);
2762 result = lappend(result, wc);
2764 /* Now move any matching windows from actives to result */
2766 for (lc = list_head(actives); lc; lc = next)
2768 WindowClause *wc2 = (WindowClause *) lfirst(lc);
2771 /* framing options are NOT to be compared here! */
2772 if (equal(wc->partitionClause, wc2->partitionClause) &&
2773 equal(wc->orderClause, wc2->orderClause))
2775 actives = list_delete_cell(actives, lc, prev);
2776 result = lappend(result, wc2);
2787 * add_volatile_sort_exprs
2788 * Identify any volatile sort/group expressions used by the active
2789 * windows, and add them to window_tlist if not already present.
2790 * Return the modified window_tlist.
2793 add_volatile_sort_exprs(List *window_tlist, List *tlist, List *activeWindows)
2795 Bitmapset *sgrefs = NULL;
2798 /* First, collect the sortgrouprefs of the windows into a bitmapset */
2799 foreach(lc, activeWindows)
2801 WindowClause *wc = (WindowClause *) lfirst(lc);
2804 foreach(lc2, wc->partitionClause)
2806 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2808 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2810 foreach(lc2, wc->orderClause)
2812 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
2814 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
2819 * Now scan the original tlist to find the referenced expressions. Any
2820 * that are volatile must be added to window_tlist.
2822 * Note: we know that the input window_tlist contains no items marked with
2823 * ressortgrouprefs, so we don't have to worry about collisions of the
2824 * reference numbers.
2828 TargetEntry *tle = (TargetEntry *) lfirst(lc);
2830 if (tle->ressortgroupref != 0 &&
2831 bms_is_member(tle->ressortgroupref, sgrefs) &&
2832 contain_volatile_functions((Node *) tle->expr))
2834 TargetEntry *newtle;
2836 newtle = makeTargetEntry(tle->expr,
2837 list_length(window_tlist) + 1,
2840 newtle->ressortgroupref = tle->ressortgroupref;
2841 window_tlist = lappend(window_tlist, newtle);
2845 return window_tlist;
2849 * make_pathkeys_for_window
2850 * Create a pathkeys list describing the required input ordering
2851 * for the given WindowClause.
2853 * The required ordering is first the PARTITION keys, then the ORDER keys.
2854 * In the future we might try to implement windowing using hashing, in which
2855 * case the ordering could be relaxed, but for now we always sort.
2858 make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
2859 List *tlist, bool canonicalize)
2861 List *window_pathkeys;
2862 List *window_sortclauses;
2864 /* Throw error if can't sort */
2865 if (!grouping_is_sortable(wc->partitionClause))
2867 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2868 errmsg("could not implement window PARTITION BY"),
2869 errdetail("Window partitioning columns must be of sortable datatypes.")));
2870 if (!grouping_is_sortable(wc->orderClause))
2872 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2873 errmsg("could not implement window ORDER BY"),
2874 errdetail("Window ordering columns must be of sortable datatypes.")));
2876 /* Okay, make the combined pathkeys */
2877 window_sortclauses = list_concat(list_copy(wc->partitionClause),
2878 list_copy(wc->orderClause));
2879 window_pathkeys = make_pathkeys_for_sortclauses(root,
2883 list_free(window_sortclauses);
2884 return window_pathkeys;
2888 * get_column_info_for_window
2889 * Get the partitioning/ordering column numbers and equality operators
2890 * for a WindowAgg node.
2892 * This depends on the behavior of make_pathkeys_for_window()!
2894 * We are given the target WindowClause and an array of the input column
2895 * numbers associated with the resulting pathkeys. In the easy case, there
2896 * are the same number of pathkey columns as partitioning + ordering columns
2897 * and we just have to copy some data around. However, it's possible that
2898 * some of the original partitioning + ordering columns were eliminated as
2899 * redundant during the transformation to pathkeys. (This can happen even
2900 * though the parser gets rid of obvious duplicates. A typical scenario is a
2901 * window specification "PARTITION BY x ORDER BY y" coupled with a clause
2902 * "WHERE x = y" that causes the two sort columns to be recognized as
2903 * redundant.) In that unusual case, we have to work a lot harder to
2904 * determine which keys are significant.
2906 * The method used here is a bit brute-force: add the sort columns to a list
2907 * one at a time and note when the resulting pathkey list gets longer. But
2908 * it's a sufficiently uncommon case that a faster way doesn't seem worth
2909 * the amount of code refactoring that'd be needed.
2913 get_column_info_for_window(PlannerInfo *root, WindowClause *wc, List *tlist,
2914 int numSortCols, AttrNumber *sortColIdx,
2916 AttrNumber **partColIdx,
2917 Oid **partOperators,
2919 AttrNumber **ordColIdx,
2922 int numPart = list_length(wc->partitionClause);
2923 int numOrder = list_length(wc->orderClause);
2925 if (numSortCols == numPart + numOrder)
2928 *partNumCols = numPart;
2929 *partColIdx = sortColIdx;
2930 *partOperators = extract_grouping_ops(wc->partitionClause);
2931 *ordNumCols = numOrder;
2932 *ordColIdx = sortColIdx + numPart;
2933 *ordOperators = extract_grouping_ops(wc->orderClause);
2942 /* first, allocate what's certainly enough space for the arrays */
2944 *partColIdx = (AttrNumber *) palloc(numPart * sizeof(AttrNumber));
2945 *partOperators = (Oid *) palloc(numPart * sizeof(Oid));
2947 *ordColIdx = (AttrNumber *) palloc(numOrder * sizeof(AttrNumber));
2948 *ordOperators = (Oid *) palloc(numOrder * sizeof(Oid));
2952 foreach(lc, wc->partitionClause)
2954 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2957 sortclauses = lappend(sortclauses, sgc);
2958 new_pathkeys = make_pathkeys_for_sortclauses(root,
2962 if (list_length(new_pathkeys) > list_length(pathkeys))
2964 /* this sort clause is actually significant */
2965 (*partColIdx)[*partNumCols] = sortColIdx[scidx++];
2966 (*partOperators)[*partNumCols] = sgc->eqop;
2968 pathkeys = new_pathkeys;
2971 foreach(lc, wc->orderClause)
2973 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2976 sortclauses = lappend(sortclauses, sgc);
2977 new_pathkeys = make_pathkeys_for_sortclauses(root,
2981 if (list_length(new_pathkeys) > list_length(pathkeys))
2983 /* this sort clause is actually significant */
2984 (*ordColIdx)[*ordNumCols] = sortColIdx[scidx++];
2985 (*ordOperators)[*ordNumCols] = sgc->eqop;
2987 pathkeys = new_pathkeys;
2990 /* complain if we didn't eat exactly the right number of sort cols */
2991 if (scidx != numSortCols)
2992 elog(ERROR, "failed to deconstruct sort operators into partitioning/ordering operators");
2998 * expression_planner
2999 * Perform planner's transformations on a standalone expression.
3001 * Various utility commands need to evaluate expressions that are not part
3002 * of a plannable query. They can do so using the executor's regular
3003 * expression-execution machinery, but first the expression has to be fed
3004 * through here to transform it from parser output to something executable.
3006 * Currently, we disallow sublinks in standalone expressions, so there's no
3007 * real "planning" involved here. (That might not always be true though.)
3008 * What we must do is run eval_const_expressions to ensure that any function
3009 * calls are converted to positional notation and function default arguments
3010 * get inserted. The fact that constant subexpressions get simplified is a
3011 * side-effect that is useful when the expression will get evaluated more than
3012 * once. Also, we must fix operator function IDs.
3014 * Note: this must not make any damaging changes to the passed-in expression
3015 * tree. (It would actually be okay to apply fix_opfuncids to it, but since
3016 * we first do an expression_tree_mutator-based walk, what is returned will
3017 * be a new node tree.)
3020 expression_planner(Expr *expr)
3025 * Convert named-argument function calls, insert default arguments and
3026 * simplify constant subexprs
3028 result = eval_const_expressions(NULL, (Node *) expr);
3030 /* Fill in opfuncid values if missing */
3031 fix_opfuncids(result);
3033 return (Expr *) result;
3038 * plan_cluster_use_sort
3039 * Use the planner to decide how CLUSTER should implement sorting
3041 * tableOid is the OID of a table to be clustered on its index indexOid
3042 * (which is already known to be a btree index). Decide whether it's
3043 * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
3044 * Return TRUE to use sorting, FALSE to use an indexscan.
3046 * Note: caller had better already hold some type of lock on the table.
3049 plan_cluster_use_sort(Oid tableOid, Oid indexOid)
3053 PlannerGlobal *glob;
3056 IndexOptInfo *indexInfo;
3057 QualCost indexExprCost;
3058 Cost comparisonCost;
3060 Path seqScanAndSortPath;
3061 IndexPath *indexScanPath;
3064 /* Set up mostly-dummy planner state */
3065 query = makeNode(Query);
3066 query->commandType = CMD_SELECT;
3068 glob = makeNode(PlannerGlobal);
3070 root = makeNode(PlannerInfo);
3071 root->parse = query;
3073 root->query_level = 1;
3074 root->planner_cxt = CurrentMemoryContext;
3075 root->wt_param_id = -1;
3077 /* Build a minimal RTE for the rel */
3078 rte = makeNode(RangeTblEntry);
3079 rte->rtekind = RTE_RELATION;
3080 rte->relid = tableOid;
3081 rte->relkind = RELKIND_RELATION;
3083 rte->inFromCl = true;
3084 query->rtable = list_make1(rte);
3086 /* ... and insert it into PlannerInfo */
3087 root->simple_rel_array_size = 2;
3088 root->simple_rel_array = (RelOptInfo **)
3089 palloc0(root->simple_rel_array_size * sizeof(RelOptInfo *));
3090 root->simple_rte_array = (RangeTblEntry **)
3091 palloc0(root->simple_rel_array_size * sizeof(RangeTblEntry *));
3092 root->simple_rte_array[1] = rte;
3094 /* Build RelOptInfo */
3095 rel = build_simple_rel(root, 1, RELOPT_BASEREL);
3097 /* Locate IndexOptInfo for the target index */
3099 foreach(lc, rel->indexlist)
3101 indexInfo = (IndexOptInfo *) lfirst(lc);
3102 if (indexInfo->indexoid == indexOid)
3107 * It's possible that get_relation_info did not generate an IndexOptInfo
3108 * for the desired index; this could happen if it's not yet reached its
3109 * indcheckxmin usability horizon, or if it's a system index and we're
3110 * ignoring system indexes. In such cases we should tell CLUSTER to not
3111 * trust the index contents but use seqscan-and-sort.
3113 if (lc == NULL) /* not in the list? */
3114 return true; /* use sort */
3117 * Rather than doing all the pushups that would be needed to use
3118 * set_baserel_size_estimates, just do a quick hack for rows and width.
3120 rel->rows = rel->tuples;
3121 rel->width = get_relation_data_width(tableOid, NULL);
3123 root->total_table_pages = rel->pages;
3126 * Determine eval cost of the index expressions, if any. We need to
3127 * charge twice that amount for each tuple comparison that happens during
3128 * the sort, since tuplesort.c will have to re-evaluate the index
3129 * expressions each time. (XXX that's pretty inefficient...)
3131 cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
3132 comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
3134 /* Estimate the cost of seq scan + sort */
3135 seqScanPath = create_seqscan_path(root, rel);
3136 cost_sort(&seqScanAndSortPath, root, NIL,
3137 seqScanPath->total_cost, rel->tuples, rel->width,
3138 comparisonCost, maintenance_work_mem, -1.0);
3140 /* Estimate the cost of index scan */
3141 indexScanPath = create_index_path(root, indexInfo,
3143 ForwardScanDirection, NULL);
3145 return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);