1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2007, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.220 2007/05/25 17:54:25 tgl Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "executor/executor.h"
22 #include "executor/nodeAgg.h"
23 #include "miscadmin.h"
24 #include "nodes/makefuncs.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/cost.h"
27 #include "optimizer/pathnode.h"
28 #include "optimizer/paths.h"
29 #include "optimizer/planmain.h"
30 #include "optimizer/planner.h"
31 #include "optimizer/prep.h"
32 #include "optimizer/subselect.h"
33 #include "optimizer/tlist.h"
34 #include "optimizer/var.h"
35 #ifdef OPTIMIZER_DEBUG
36 #include "nodes/print.h"
38 #include "parser/parse_expr.h"
39 #include "parser/parse_oper.h"
40 #include "parser/parsetree.h"
41 #include "utils/lsyscache.h"
42 #include "utils/syscache.h"
45 /* Hook for plugins to get control in planner() */
46 planner_hook_type planner_hook = NULL;
49 /* Expression kind codes for preprocess_expression */
50 #define EXPRKIND_QUAL 0
51 #define EXPRKIND_TARGET 1
52 #define EXPRKIND_RTFUNC 2
53 #define EXPRKIND_VALUES 3
54 #define EXPRKIND_LIMIT 4
55 #define EXPRKIND_ININFO 5
56 #define EXPRKIND_APPINFO 6
59 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
60 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
61 static Plan *inheritance_planner(PlannerInfo *root);
62 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
63 static bool is_dummy_plan(Plan *plan);
64 static double preprocess_limit(PlannerInfo *root,
65 double tuple_fraction,
66 int64 *offset_est, int64 *count_est);
67 static Oid *extract_grouping_ops(List *groupClause);
68 static bool choose_hashed_grouping(PlannerInfo *root,
69 double tuple_fraction, double limit_tuples,
70 Path *cheapest_path, Path *sorted_path,
71 Oid *groupOperators, double dNumGroups,
72 AggClauseCounts *agg_counts);
73 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
74 AttrNumber **groupColIdx, bool *need_tlist_eval);
75 static void locate_grouping_columns(PlannerInfo *root,
78 AttrNumber *groupColIdx);
79 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
82 /*****************************************************************************
84 * Query optimizer entry point
86 * To support loadable plugins that monitor or modify planner behavior,
87 * we provide a hook variable that lets a plugin get control before and
88 * after the standard planning process. The plugin would normally call
91 * Note to plugin authors: standard_planner() scribbles on its Query input,
92 * so you'd better copy that data structure if you want to plan more than once.
94 *****************************************************************************/
96 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
101 result = (*planner_hook) (parse, cursorOptions, boundParams);
103 result = standard_planner(parse, cursorOptions, boundParams);
108 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
112 double tuple_fraction;
118 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
119 if (parse->utilityStmt &&
120 IsA(parse->utilityStmt, DeclareCursorStmt))
121 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
124 * Set up global state for this planner invocation. This data is needed
125 * across all levels of sub-Query that might exist in the given command,
126 * so we keep it in a separate struct that's linked to by each per-Query
129 glob = makeNode(PlannerGlobal);
131 glob->boundParams = boundParams;
132 glob->paramlist = NIL;
133 glob->subplans = NIL;
134 glob->subrtables = NIL;
135 glob->rewindPlanIDs = NULL;
136 glob->finalrtable = NIL;
138 /* Determine what fraction of the plan is likely to be scanned */
139 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
142 * We have no real idea how many tuples the user will ultimately FETCH
143 * from a cursor, but it seems a good bet that he doesn't want 'em
144 * all. Optimize for 10% retrieval (you gotta better number? Should
145 * this be a SETtable parameter?)
147 tuple_fraction = 0.10;
151 /* Default assumption is we need all the tuples */
152 tuple_fraction = 0.0;
155 /* primary planning entry point (may recurse for subqueries) */
156 top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
159 * If creating a plan for a scrollable cursor, make sure it can run
160 * backwards on demand. Add a Material node at the top at need.
162 if (cursorOptions & CURSOR_OPT_SCROLL)
164 if (!ExecSupportsBackwardScan(top_plan))
165 top_plan = materialize_finished_plan(top_plan);
168 /* final cleanup of the plan */
169 Assert(glob->finalrtable == NIL);
170 top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
171 /* ... and the subplans (both regular subplans and initplans) */
172 Assert(list_length(glob->subplans) == list_length(glob->subrtables));
173 forboth(lp, glob->subplans, lr, glob->subrtables)
175 Plan *subplan = (Plan *) lfirst(lp);
176 List *subrtable = (List *) lfirst(lr);
178 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
181 /* build the PlannedStmt result */
182 result = makeNode(PlannedStmt);
184 result->commandType = parse->commandType;
185 result->canSetTag = parse->canSetTag;
186 result->planTree = top_plan;
187 result->rtable = glob->finalrtable;
188 result->resultRelations = root->resultRelations;
189 result->utilityStmt = parse->utilityStmt;
190 result->intoClause = parse->intoClause;
191 result->subplans = glob->subplans;
192 result->rewindPlanIDs = glob->rewindPlanIDs;
193 result->returningLists = root->returningLists;
194 result->rowMarks = parse->rowMarks;
195 result->nParamExec = list_length(glob->paramlist);
201 /*--------------------
203 * Invokes the planner on a subquery. We recurse to here for each
204 * sub-SELECT found in the query tree.
206 * glob is the global state for the current planner run.
207 * parse is the querytree produced by the parser & rewriter.
208 * level is the current recursion depth (1 at the top-level Query).
209 * tuple_fraction is the fraction of tuples we expect will be retrieved.
210 * tuple_fraction is interpreted as explained for grouping_planner, below.
212 * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
213 * among other things this tells the output sort ordering of the plan.
215 * Basically, this routine does the stuff that should only be done once
216 * per Query object. It then calls grouping_planner. At one time,
217 * grouping_planner could be invoked recursively on the same Query object;
218 * that's not currently true, but we keep the separation between the two
219 * routines anyway, in case we need it again someday.
221 * subquery_planner will be called recursively to handle sub-Query nodes
222 * found within the query's expressions and rangetable.
224 * Returns a query plan.
225 *--------------------
228 subquery_planner(PlannerGlobal *glob, Query *parse,
229 Index level, double tuple_fraction,
230 PlannerInfo **subroot)
232 int num_old_subplans = list_length(glob->subplans);
238 /* Create a PlannerInfo data structure for this subquery */
239 root = makeNode(PlannerInfo);
242 root->query_level = level;
243 root->planner_cxt = CurrentMemoryContext;
244 root->init_plans = NIL;
245 root->eq_classes = NIL;
246 root->in_info_list = NIL;
247 root->append_rel_list = NIL;
250 * Look for IN clauses at the top level of WHERE, and transform them into
251 * joins. Note that this step only handles IN clauses originally at top
252 * level of WHERE; if we pull up any subqueries in the next step, their
253 * INs are processed just before pulling them up.
255 if (parse->hasSubLinks)
256 parse->jointree->quals = pull_up_IN_clauses(root,
257 parse->jointree->quals);
260 * Check to see if any subqueries in the rangetable can be merged into
263 parse->jointree = (FromExpr *)
264 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
267 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
268 * avoid the expense of doing flatten_join_alias_vars(). Also check for
269 * outer joins --- if none, we can skip reduce_outer_joins() and some
270 * other processing. This must be done after we have done
271 * pull_up_subqueries, of course.
273 * Note: if reduce_outer_joins manages to eliminate all outer joins,
274 * root->hasOuterJoins is not reset currently. This is OK since its
275 * purpose is merely to suppress unnecessary processing in simple cases.
277 root->hasJoinRTEs = false;
278 root->hasOuterJoins = false;
279 foreach(l, parse->rtable)
281 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
283 if (rte->rtekind == RTE_JOIN)
285 root->hasJoinRTEs = true;
286 if (IS_OUTER_JOIN(rte->jointype))
288 root->hasOuterJoins = true;
289 /* Can quit scanning once we find an outer join */
296 * Expand any rangetable entries that are inheritance sets into "append
297 * relations". This can add entries to the rangetable, but they must be
298 * plain base relations not joins, so it's OK (and marginally more
299 * efficient) to do it after checking for join RTEs. We must do it after
300 * pulling up subqueries, else we'd fail to handle inherited tables in
303 expand_inherited_tables(root);
306 * Set hasHavingQual to remember if HAVING clause is present. Needed
307 * because preprocess_expression will reduce a constant-true condition to
308 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
310 root->hasHavingQual = (parse->havingQual != NULL);
312 /* Clear this flag; might get set in distribute_qual_to_rels */
313 root->hasPseudoConstantQuals = false;
316 * Do expression preprocessing on targetlist and quals.
318 parse->targetList = (List *)
319 preprocess_expression(root, (Node *) parse->targetList,
322 parse->returningList = (List *)
323 preprocess_expression(root, (Node *) parse->returningList,
326 preprocess_qual_conditions(root, (Node *) parse->jointree);
328 parse->havingQual = preprocess_expression(root, parse->havingQual,
331 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
333 parse->limitCount = preprocess_expression(root, parse->limitCount,
336 root->in_info_list = (List *)
337 preprocess_expression(root, (Node *) root->in_info_list,
339 root->append_rel_list = (List *)
340 preprocess_expression(root, (Node *) root->append_rel_list,
343 /* Also need to preprocess expressions for function and values RTEs */
344 foreach(l, parse->rtable)
346 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
348 if (rte->rtekind == RTE_FUNCTION)
349 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
351 else if (rte->rtekind == RTE_VALUES)
352 rte->values_lists = (List *)
353 preprocess_expression(root, (Node *) rte->values_lists,
358 * In some cases we may want to transfer a HAVING clause into WHERE. We
359 * cannot do so if the HAVING clause contains aggregates (obviously) or
360 * volatile functions (since a HAVING clause is supposed to be executed
361 * only once per group). Also, it may be that the clause is so expensive
362 * to execute that we're better off doing it only once per group, despite
363 * the loss of selectivity. This is hard to estimate short of doing the
364 * entire planning process twice, so we use a heuristic: clauses
365 * containing subplans are left in HAVING. Otherwise, we move or copy the
366 * HAVING clause into WHERE, in hopes of eliminating tuples before
367 * aggregation instead of after.
369 * If the query has explicit grouping then we can simply move such a
370 * clause into WHERE; any group that fails the clause will not be in the
371 * output because none of its tuples will reach the grouping or
372 * aggregation stage. Otherwise we must have a degenerate (variable-free)
373 * HAVING clause, which we put in WHERE so that query_planner() can use it
374 * in a gating Result node, but also keep in HAVING to ensure that we
375 * don't emit a bogus aggregated row. (This could be done better, but it
376 * seems not worth optimizing.)
378 * Note that both havingQual and parse->jointree->quals are in
379 * implicitly-ANDed-list form at this point, even though they are declared
383 foreach(l, (List *) parse->havingQual)
385 Node *havingclause = (Node *) lfirst(l);
387 if (contain_agg_clause(havingclause) ||
388 contain_volatile_functions(havingclause) ||
389 contain_subplans(havingclause))
391 /* keep it in HAVING */
392 newHaving = lappend(newHaving, havingclause);
394 else if (parse->groupClause)
396 /* move it to WHERE */
397 parse->jointree->quals = (Node *)
398 lappend((List *) parse->jointree->quals, havingclause);
402 /* put a copy in WHERE, keep it in HAVING */
403 parse->jointree->quals = (Node *)
404 lappend((List *) parse->jointree->quals,
405 copyObject(havingclause));
406 newHaving = lappend(newHaving, havingclause);
409 parse->havingQual = (Node *) newHaving;
412 * If we have any outer joins, try to reduce them to plain inner joins.
413 * This step is most easily done after we've done expression
416 if (root->hasOuterJoins)
417 reduce_outer_joins(root);
420 * Do the main planning. If we have an inherited target relation, that
421 * needs special processing, else go straight to grouping_planner.
423 if (parse->resultRelation &&
424 rt_fetch(parse->resultRelation, parse->rtable)->inh)
425 plan = inheritance_planner(root);
427 plan = grouping_planner(root, tuple_fraction);
430 * If any subplans were generated, or if we're inside a subplan, build
431 * initPlan list and extParam/allParam sets for plan nodes, and attach the
432 * initPlans to the top plan node.
434 if (list_length(glob->subplans) != num_old_subplans ||
435 root->query_level > 1)
436 SS_finalize_plan(root, plan);
438 /* Return internal info if caller wants it */
446 * preprocess_expression
447 * Do subquery_planner's preprocessing work for an expression,
448 * which can be a targetlist, a WHERE clause (including JOIN/ON
449 * conditions), or a HAVING clause.
452 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
455 * Fall out quickly if expression is empty. This occurs often enough to
456 * be worth checking. Note that null->null is the correct conversion for
457 * implicit-AND result format, too.
463 * If the query has any join RTEs, replace join alias variables with
464 * base-relation variables. We must do this before sublink processing,
465 * else sublinks expanded out from join aliases wouldn't get processed. We
466 * can skip it in VALUES lists, however, since they can't contain any Vars
469 if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
470 expr = flatten_join_alias_vars(root, expr);
473 * Simplify constant expressions.
475 * Note: this also flattens nested AND and OR expressions into N-argument
476 * form. All processing of a qual expression after this point must be
477 * careful to maintain AND/OR flatness --- that is, do not generate a tree
478 * with AND directly under AND, nor OR directly under OR.
480 * Because this is a relatively expensive process, we skip it when the
481 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
482 * expression will only be evaluated once anyway, so no point in
483 * pre-simplifying; we can't execute it any faster than the executor can,
484 * and we will waste cycles copying the tree. Notice however that we
485 * still must do it for quals (to get AND/OR flatness); and if we are in a
486 * subquery we should not assume it will be done only once.
488 * For VALUES lists we never do this at all, again on the grounds that we
489 * should optimize for one-time evaluation.
491 if (kind != EXPRKIND_VALUES &&
492 (root->parse->jointree->fromlist != NIL ||
493 kind == EXPRKIND_QUAL ||
494 root->query_level > 1))
495 expr = eval_const_expressions(expr);
498 * If it's a qual or havingQual, canonicalize it.
500 if (kind == EXPRKIND_QUAL)
502 expr = (Node *) canonicalize_qual((Expr *) expr);
504 #ifdef OPTIMIZER_DEBUG
505 printf("After canonicalize_qual()\n");
510 /* Expand SubLinks to SubPlans */
511 if (root->parse->hasSubLinks)
512 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
515 * XXX do not insert anything here unless you have grokked the comments in
516 * SS_replace_correlation_vars ...
519 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
520 if (root->query_level > 1)
521 expr = SS_replace_correlation_vars(root, expr);
524 * If it's a qual or havingQual, convert it to implicit-AND format. (We
525 * don't want to do this before eval_const_expressions, since the latter
526 * would be unable to simplify a top-level AND correctly. Also,
527 * SS_process_sublinks expects explicit-AND format.)
529 if (kind == EXPRKIND_QUAL)
530 expr = (Node *) make_ands_implicit((Expr *) expr);
536 * preprocess_qual_conditions
537 * Recursively scan the query's jointree and do subquery_planner's
538 * preprocessing work on each qual condition found therein.
541 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
545 if (IsA(jtnode, RangeTblRef))
547 /* nothing to do here */
549 else if (IsA(jtnode, FromExpr))
551 FromExpr *f = (FromExpr *) jtnode;
554 foreach(l, f->fromlist)
555 preprocess_qual_conditions(root, lfirst(l));
557 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
559 else if (IsA(jtnode, JoinExpr))
561 JoinExpr *j = (JoinExpr *) jtnode;
563 preprocess_qual_conditions(root, j->larg);
564 preprocess_qual_conditions(root, j->rarg);
566 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
569 elog(ERROR, "unrecognized node type: %d",
570 (int) nodeTag(jtnode));
574 * inheritance_planner
575 * Generate a plan in the case where the result relation is an
578 * We have to handle this case differently from cases where a source relation
579 * is an inheritance set. Source inheritance is expanded at the bottom of the
580 * plan tree (see allpaths.c), but target inheritance has to be expanded at
581 * the top. The reason is that for UPDATE, each target relation needs a
582 * different targetlist matching its own column set. Also, for both UPDATE
583 * and DELETE, the executor needs the Append plan node at the top, else it
584 * can't keep track of which table is the current target table. Fortunately,
585 * the UPDATE/DELETE target can never be the nullable side of an outer join,
586 * so it's OK to generate the plan this way.
588 * Returns a query plan.
591 inheritance_planner(PlannerInfo *root)
593 Query *parse = root->parse;
594 int parentRTindex = parse->resultRelation;
595 List *subplans = NIL;
596 List *resultRelations = NIL;
597 List *returningLists = NIL;
603 foreach(l, root->append_rel_list)
605 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
608 /* append_rel_list contains all append rels; ignore others */
609 if (appinfo->parent_relid != parentRTindex)
613 * Generate modified query with this rel as target. We have to be
614 * prepared to translate varnos in in_info_list as well as in the
617 memcpy(&subroot, root, sizeof(PlannerInfo));
618 subroot.parse = (Query *)
619 adjust_appendrel_attrs((Node *) parse,
621 subroot.in_info_list = (List *)
622 adjust_appendrel_attrs((Node *) root->in_info_list,
624 subroot.init_plans = NIL;
625 /* There shouldn't be any OJ info to translate, as yet */
626 Assert(subroot.oj_info_list == NIL);
629 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
632 * If this child rel was excluded by constraint exclusion, exclude it
635 if (is_dummy_plan(subplan))
638 /* Save rtable and tlist from first rel for use below */
641 rtable = subroot.parse->rtable;
642 tlist = subplan->targetlist;
645 subplans = lappend(subplans, subplan);
647 /* Make sure any initplans from this rel get into the outer list */
648 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
650 /* Build target-relations list for the executor */
651 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
653 /* Build list of per-relation RETURNING targetlists */
654 if (parse->returningList)
656 Assert(list_length(subroot.returningLists) == 1);
657 returningLists = list_concat(returningLists,
658 subroot.returningLists);
662 root->resultRelations = resultRelations;
663 root->returningLists = returningLists;
665 /* Mark result as unordered (probably unnecessary) */
666 root->query_pathkeys = NIL;
669 * If we managed to exclude every child rel, return a dummy plan
672 return (Plan *) make_result(root,
674 (Node *) list_make1(makeBoolConst(false,
679 * Planning might have modified the rangetable, due to changes of the
680 * Query structures inside subquery RTEs. We have to ensure that this
681 * gets propagated back to the master copy. But can't do this until we
682 * are done planning, because all the calls to grouping_planner need
683 * virgin sub-Queries to work from. (We are effectively assuming that
684 * sub-Queries will get planned identically each time, or at least that
685 * the impacts on their rangetables will be the same each time.)
687 * XXX should clean this up someday
689 parse->rtable = rtable;
691 /* Suppress Append if there's only one surviving child rel */
692 if (list_length(subplans) == 1)
693 return (Plan *) linitial(subplans);
695 return (Plan *) make_append(subplans, true, tlist);
698 /*--------------------
700 * Perform planning steps related to grouping, aggregation, etc.
701 * This primarily means adding top-level processing to the basic
702 * query plan produced by query_planner.
704 * tuple_fraction is the fraction of tuples we expect will be retrieved
706 * tuple_fraction is interpreted as follows:
707 * 0: expect all tuples to be retrieved (normal case)
708 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
709 * from the plan to be retrieved
710 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
711 * expected to be retrieved (ie, a LIMIT specification)
713 * Returns a query plan. Also, root->query_pathkeys is returned as the
714 * actual output ordering of the plan (in pathkey format).
715 *--------------------
718 grouping_planner(PlannerInfo *root, double tuple_fraction)
720 Query *parse = root->parse;
721 List *tlist = parse->targetList;
722 int64 offset_est = 0;
724 double limit_tuples = -1.0;
726 List *current_pathkeys;
728 double dNumGroups = 0;
730 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
731 if (parse->limitCount || parse->limitOffset)
733 tuple_fraction = preprocess_limit(root, tuple_fraction,
734 &offset_est, &count_est);
736 * If we have a known LIMIT, and don't have an unknown OFFSET,
737 * we can estimate the effects of using a bounded sort.
739 if (count_est > 0 && offset_est >= 0)
740 limit_tuples = (double) count_est + (double) offset_est;
743 if (parse->setOperations)
745 List *set_sortclauses;
748 * If there's a top-level ORDER BY, assume we have to fetch all the
749 * tuples. This might seem too simplistic given all the hackery below
750 * to possibly avoid the sort ... but a nonzero tuple_fraction is only
751 * of use to plan_set_operations() when the setop is UNION ALL, and
752 * the result of UNION ALL is always unsorted.
754 if (parse->sortClause)
755 tuple_fraction = 0.0;
758 * Construct the plan for set operations. The result will not need
759 * any work except perhaps a top-level sort and/or LIMIT.
761 result_plan = plan_set_operations(root, tuple_fraction,
765 * Calculate pathkeys representing the sort order (if any) of the set
766 * operation's result. We have to do this before overwriting the sort
769 current_pathkeys = make_pathkeys_for_sortclauses(root,
771 result_plan->targetlist,
775 * We should not need to call preprocess_targetlist, since we must be
776 * in a SELECT query node. Instead, use the targetlist returned by
777 * plan_set_operations (since this tells whether it returned any
778 * resjunk columns!), and transfer any sort key information from the
781 Assert(parse->commandType == CMD_SELECT);
783 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
786 * Can't handle FOR UPDATE/SHARE here (parser should have checked
787 * already, but let's make sure).
791 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
792 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
795 * Calculate pathkeys that represent result ordering requirements
797 sort_pathkeys = make_pathkeys_for_sortclauses(root,
804 /* No set operations, do regular planning */
806 List *group_pathkeys;
807 AttrNumber *groupColIdx = NULL;
808 Oid *groupOperators = NULL;
809 bool need_tlist_eval = true;
815 AggClauseCounts agg_counts;
816 int numGroupCols = list_length(parse->groupClause);
817 bool use_hashed_grouping = false;
819 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
821 /* Preprocess targetlist */
822 tlist = preprocess_targetlist(root, tlist);
825 * Generate appropriate target list for subplan; may be different from
826 * tlist if grouping or aggregation is needed.
828 sub_tlist = make_subplanTargetList(root, tlist,
829 &groupColIdx, &need_tlist_eval);
832 * Calculate pathkeys that represent grouping/ordering requirements.
833 * Stash them in PlannerInfo so that query_planner can canonicalize
834 * them after EquivalenceClasses have been formed.
836 root->group_pathkeys =
837 make_pathkeys_for_sortclauses(root,
841 root->sort_pathkeys =
842 make_pathkeys_for_sortclauses(root,
848 * Will need actual number of aggregates for estimating costs.
850 * Note: we do not attempt to detect duplicate aggregates here; a
851 * somewhat-overestimated count is okay for our present purposes.
853 * Note: think not that we can turn off hasAggs if we find no aggs. It
854 * is possible for constant-expression simplification to remove all
855 * explicit references to aggs, but we still have to follow the
856 * aggregate semantics (eg, producing only one output row).
860 count_agg_clauses((Node *) tlist, &agg_counts);
861 count_agg_clauses(parse->havingQual, &agg_counts);
865 * Figure out whether we need a sorted result from query_planner.
867 * If we have a GROUP BY clause, then we want a result sorted properly
868 * for grouping. Otherwise, if there is an ORDER BY clause, we want
869 * to sort by the ORDER BY clause. (Note: if we have both, and ORDER
870 * BY is a superset of GROUP BY, it would be tempting to request sort
871 * by ORDER BY --- but that might just leave us failing to exploit an
872 * available sort order at all. Needs more thought...)
874 if (parse->groupClause)
875 root->query_pathkeys = root->group_pathkeys;
876 else if (parse->sortClause)
877 root->query_pathkeys = root->sort_pathkeys;
879 root->query_pathkeys = NIL;
882 * Generate the best unsorted and presorted paths for this Query (but
883 * note there may not be any presorted path). query_planner will also
884 * estimate the number of groups in the query, and canonicalize all
887 query_planner(root, sub_tlist, tuple_fraction, limit_tuples,
888 &cheapest_path, &sorted_path, &dNumGroups);
890 group_pathkeys = root->group_pathkeys;
891 sort_pathkeys = root->sort_pathkeys;
894 * If grouping, extract the grouping operators and decide whether we
895 * want to use hashed grouping.
897 if (parse->groupClause)
899 groupOperators = extract_grouping_ops(parse->groupClause);
900 use_hashed_grouping =
901 choose_hashed_grouping(root, tuple_fraction, limit_tuples,
902 cheapest_path, sorted_path,
903 groupOperators, dNumGroups,
906 /* Also convert # groups to long int --- but 'ware overflow! */
907 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
911 * Select the best path. If we are doing hashed grouping, we will
912 * always read all the input tuples, so use the cheapest-total path.
913 * Otherwise, trust query_planner's decision about which to use.
915 if (use_hashed_grouping || !sorted_path)
916 best_path = cheapest_path;
918 best_path = sorted_path;
921 * Check to see if it's possible to optimize MIN/MAX aggregates. If
922 * so, we will forget all the work we did so far to choose a "regular"
923 * path ... but we had to do it anyway to be able to tell which way is
926 result_plan = optimize_minmax_aggregates(root,
929 if (result_plan != NULL)
932 * optimize_minmax_aggregates generated the full plan, with the
933 * right tlist, and it has no sort order.
935 current_pathkeys = NIL;
940 * Normal case --- create a plan according to query_planner's
943 result_plan = create_plan(root, best_path);
944 current_pathkeys = best_path->pathkeys;
947 * create_plan() returns a plan with just a "flat" tlist of
948 * required Vars. Usually we need to insert the sub_tlist as the
949 * tlist of the top plan node. However, we can skip that if we
950 * determined that whatever query_planner chose to return will be
956 * If the top-level plan node is one that cannot do expression
957 * evaluation, we must insert a Result node to project the
960 if (!is_projection_capable_plan(result_plan))
962 result_plan = (Plan *) make_result(root,
970 * Otherwise, just replace the subplan's flat tlist with
973 result_plan->targetlist = sub_tlist;
977 * Also, account for the cost of evaluation of the sub_tlist.
979 * Up to now, we have only been dealing with "flat" tlists,
980 * containing just Vars. So their evaluation cost is zero
981 * according to the model used by cost_qual_eval() (or if you
982 * prefer, the cost is factored into cpu_tuple_cost). Thus we
983 * can avoid accounting for tlist cost throughout
984 * query_planner() and subroutines. But now we've inserted a
985 * tlist that might contain actual operators, sub-selects, etc
986 * --- so we'd better account for its cost.
988 * Below this point, any tlist eval cost for added-on nodes
989 * should be accounted for as we create those nodes.
990 * Presently, of the node types we can add on, only Agg and
991 * Group project new tlists (the rest just copy their input
992 * tuples) --- so make_agg() and make_group() are responsible
993 * for computing the added cost.
995 cost_qual_eval(&tlist_cost, sub_tlist, root);
996 result_plan->startup_cost += tlist_cost.startup;
997 result_plan->total_cost += tlist_cost.startup +
998 tlist_cost.per_tuple * result_plan->plan_rows;
1003 * Since we're using query_planner's tlist and not the one
1004 * make_subplanTargetList calculated, we have to refigure any
1005 * grouping-column indexes make_subplanTargetList computed.
1007 locate_grouping_columns(root, tlist, result_plan->targetlist,
1012 * Insert AGG or GROUP node if needed, plus an explicit sort step
1015 * HAVING clause, if any, becomes qual of the Agg or Group node.
1017 if (use_hashed_grouping)
1019 /* Hashed aggregate plan --- no sort needed */
1020 result_plan = (Plan *) make_agg(root,
1022 (List *) parse->havingQual,
1030 /* Hashed aggregation produces randomly-ordered results */
1031 current_pathkeys = NIL;
1033 else if (parse->hasAggs)
1035 /* Plain aggregate plan --- sort if needed */
1036 AggStrategy aggstrategy;
1038 if (parse->groupClause)
1040 if (!pathkeys_contained_in(group_pathkeys,
1043 result_plan = (Plan *)
1044 make_sort_from_groupcols(root,
1048 current_pathkeys = group_pathkeys;
1050 aggstrategy = AGG_SORTED;
1053 * The AGG node will not change the sort ordering of its
1054 * groups, so current_pathkeys describes the result too.
1059 aggstrategy = AGG_PLAIN;
1060 /* Result will be only one row anyway; no sort order */
1061 current_pathkeys = NIL;
1064 result_plan = (Plan *) make_agg(root,
1066 (List *) parse->havingQual,
1075 else if (parse->groupClause)
1078 * GROUP BY without aggregation, so insert a group node (plus
1079 * the appropriate sort node, if necessary).
1081 * Add an explicit sort if we couldn't make the path come out
1082 * the way the GROUP node needs it.
1084 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1086 result_plan = (Plan *)
1087 make_sort_from_groupcols(root,
1091 current_pathkeys = group_pathkeys;
1094 result_plan = (Plan *) make_group(root,
1096 (List *) parse->havingQual,
1102 /* The Group node won't change sort ordering */
1104 else if (root->hasHavingQual)
1107 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1108 * This is a degenerate case in which we are supposed to emit
1109 * either 0 or 1 row depending on whether HAVING succeeds.
1110 * Furthermore, there cannot be any variables in either HAVING
1111 * or the targetlist, so we actually do not need the FROM
1112 * table at all! We can just throw away the plan-so-far and
1113 * generate a Result node. This is a sufficiently unusual
1114 * corner case that it's not worth contorting the structure of
1115 * this routine to avoid having to generate the plan in the
1118 result_plan = (Plan *) make_result(root,
1123 } /* end of non-minmax-aggregate case */
1124 } /* end of if (setOperations) */
1127 * If we were not able to make the plan come out in the right order, add
1128 * an explicit sort step.
1130 if (parse->sortClause)
1132 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1134 result_plan = (Plan *) make_sort_from_pathkeys(root,
1138 current_pathkeys = sort_pathkeys;
1143 * If there is a DISTINCT clause, add the UNIQUE node.
1145 if (parse->distinctClause)
1147 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1150 * If there was grouping or aggregation, leave plan_rows as-is (ie,
1151 * assume the result was already mostly unique). If not, use the
1152 * number of distinct-groups calculated by query_planner.
1154 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1155 result_plan->plan_rows = dNumGroups;
1159 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1161 if (parse->limitCount || parse->limitOffset)
1163 result_plan = (Plan *) make_limit(result_plan,
1171 * Deal with the RETURNING clause if any. It's convenient to pass the
1172 * returningList through setrefs.c now rather than at top level (if we
1173 * waited, handling inherited UPDATE/DELETE would be much harder).
1175 if (parse->returningList)
1179 Assert(parse->resultRelation);
1180 rlist = set_returning_clause_references(parse->returningList,
1182 parse->resultRelation);
1183 root->returningLists = list_make1(rlist);
1186 root->returningLists = NIL;
1188 /* Compute result-relations list if needed */
1189 if (parse->resultRelation)
1190 root->resultRelations = list_make1_int(parse->resultRelation);
1192 root->resultRelations = NIL;
1195 * Return the actual output ordering in query_pathkeys for possible use by
1196 * an outer query level.
1198 root->query_pathkeys = current_pathkeys;
1204 * Detect whether a plan node is a "dummy" plan created when a relation
1205 * is deemed not to need scanning due to constraint exclusion.
1207 * Currently, such dummy plans are Result nodes with constant FALSE
1211 is_dummy_plan(Plan *plan)
1213 if (IsA(plan, Result))
1215 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1217 if (list_length(rcqual) == 1)
1219 Const *constqual = (Const *) linitial(rcqual);
1221 if (constqual && IsA(constqual, Const))
1223 if (!constqual->constisnull &&
1224 !DatumGetBool(constqual->constvalue))
1233 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1235 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1236 * results back in *count_est and *offset_est. These variables are set to
1237 * 0 if the corresponding clause is not present, and -1 if it's present
1238 * but we couldn't estimate the value for it. (The "0" convention is OK
1239 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1240 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1241 * usual practice of never estimating less than one row.) These values will
1242 * be passed to make_limit, which see if you change this code.
1244 * The return value is the suitably adjusted tuple_fraction to use for
1245 * planning the query. This adjustment is not overridable, since it reflects
1246 * plan actions that grouping_planner() will certainly take, not assumptions
1250 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1251 int64 *offset_est, int64 *count_est)
1253 Query *parse = root->parse;
1255 double limit_fraction;
1257 /* Should not be called unless LIMIT or OFFSET */
1258 Assert(parse->limitCount || parse->limitOffset);
1261 * Try to obtain the clause values. We use estimate_expression_value
1262 * primarily because it can sometimes do something useful with Params.
1264 if (parse->limitCount)
1266 est = estimate_expression_value(root, parse->limitCount);
1267 if (est && IsA(est, Const))
1269 if (((Const *) est)->constisnull)
1271 /* NULL indicates LIMIT ALL, ie, no limit */
1272 *count_est = 0; /* treat as not present */
1276 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1277 if (*count_est <= 0)
1278 *count_est = 1; /* force to at least 1 */
1282 *count_est = -1; /* can't estimate */
1285 *count_est = 0; /* not present */
1287 if (parse->limitOffset)
1289 est = estimate_expression_value(root, parse->limitOffset);
1290 if (est && IsA(est, Const))
1292 if (((Const *) est)->constisnull)
1294 /* Treat NULL as no offset; the executor will too */
1295 *offset_est = 0; /* treat as not present */
1299 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1300 if (*offset_est < 0)
1301 *offset_est = 0; /* less than 0 is same as 0 */
1305 *offset_est = -1; /* can't estimate */
1308 *offset_est = 0; /* not present */
1310 if (*count_est != 0)
1313 * A LIMIT clause limits the absolute number of tuples returned.
1314 * However, if it's not a constant LIMIT then we have to guess; for
1315 * lack of a better idea, assume 10% of the plan's result is wanted.
1317 if (*count_est < 0 || *offset_est < 0)
1319 /* LIMIT or OFFSET is an expression ... punt ... */
1320 limit_fraction = 0.10;
1324 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1325 limit_fraction = (double) *count_est + (double) *offset_est;
1329 * If we have absolute limits from both caller and LIMIT, use the
1330 * smaller value; likewise if they are both fractional. If one is
1331 * fractional and the other absolute, we can't easily determine which
1332 * is smaller, but we use the heuristic that the absolute will usually
1335 if (tuple_fraction >= 1.0)
1337 if (limit_fraction >= 1.0)
1340 tuple_fraction = Min(tuple_fraction, limit_fraction);
1344 /* caller absolute, limit fractional; use caller's value */
1347 else if (tuple_fraction > 0.0)
1349 if (limit_fraction >= 1.0)
1351 /* caller fractional, limit absolute; use limit */
1352 tuple_fraction = limit_fraction;
1356 /* both fractional */
1357 tuple_fraction = Min(tuple_fraction, limit_fraction);
1362 /* no info from caller, just use limit */
1363 tuple_fraction = limit_fraction;
1366 else if (*offset_est != 0 && tuple_fraction > 0.0)
1369 * We have an OFFSET but no LIMIT. This acts entirely differently
1370 * from the LIMIT case: here, we need to increase rather than decrease
1371 * the caller's tuple_fraction, because the OFFSET acts to cause more
1372 * tuples to be fetched instead of fewer. This only matters if we got
1373 * a tuple_fraction > 0, however.
1375 * As above, use 10% if OFFSET is present but unestimatable.
1377 if (*offset_est < 0)
1378 limit_fraction = 0.10;
1380 limit_fraction = (double) *offset_est;
1383 * If we have absolute counts from both caller and OFFSET, add them
1384 * together; likewise if they are both fractional. If one is
1385 * fractional and the other absolute, we want to take the larger, and
1386 * we heuristically assume that's the fractional one.
1388 if (tuple_fraction >= 1.0)
1390 if (limit_fraction >= 1.0)
1392 /* both absolute, so add them together */
1393 tuple_fraction += limit_fraction;
1397 /* caller absolute, limit fractional; use limit */
1398 tuple_fraction = limit_fraction;
1403 if (limit_fraction >= 1.0)
1405 /* caller fractional, limit absolute; use caller's value */
1409 /* both fractional, so add them together */
1410 tuple_fraction += limit_fraction;
1411 if (tuple_fraction >= 1.0)
1412 tuple_fraction = 0.0; /* assume fetch all */
1417 return tuple_fraction;
1421 * extract_grouping_ops - make an array of the equality operator OIDs
1422 * for the GROUP BY clause
1425 extract_grouping_ops(List *groupClause)
1427 int numCols = list_length(groupClause);
1429 Oid *groupOperators;
1432 groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1434 foreach(glitem, groupClause)
1436 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1438 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1439 if (!OidIsValid(groupOperators[colno])) /* shouldn't happen */
1440 elog(ERROR, "could not find equality operator for ordering operator %u",
1445 return groupOperators;
1449 * choose_hashed_grouping - should we use hashed grouping?
1452 choose_hashed_grouping(PlannerInfo *root,
1453 double tuple_fraction, double limit_tuples,
1454 Path *cheapest_path, Path *sorted_path,
1455 Oid *groupOperators, double dNumGroups,
1456 AggClauseCounts *agg_counts)
1458 int numGroupCols = list_length(root->parse->groupClause);
1459 double cheapest_path_rows;
1460 int cheapest_path_width;
1462 List *current_pathkeys;
1468 * Check can't-do-it conditions, including whether the grouping operators
1469 * are hashjoinable. (We assume hashing is OK if they are marked
1470 * oprcanhash. If there isn't actually a supporting hash function,
1471 * the executor will complain at runtime.)
1473 * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1474 * (Doing so would imply storing *all* the input values in the hash table,
1475 * which seems like a certain loser.)
1477 if (!enable_hashagg)
1479 if (agg_counts->numDistinctAggs != 0)
1481 for (i = 0; i < numGroupCols; i++)
1483 if (!op_hashjoinable(groupOperators[i]))
1488 * Don't do it if it doesn't look like the hashtable will fit into
1491 * Beware here of the possibility that cheapest_path->parent is NULL. This
1492 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1494 if (cheapest_path->parent)
1496 cheapest_path_rows = cheapest_path->parent->rows;
1497 cheapest_path_width = cheapest_path->parent->width;
1501 cheapest_path_rows = 1; /* assume non-set result */
1502 cheapest_path_width = 100; /* arbitrary */
1505 /* Estimate per-hash-entry space at tuple width... */
1506 hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1507 /* plus space for pass-by-ref transition values... */
1508 hashentrysize += agg_counts->transitionSpace;
1509 /* plus the per-hash-entry overhead */
1510 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1512 if (hashentrysize * dNumGroups > work_mem * 1024L)
1516 * See if the estimated cost is no more than doing it the other way. While
1517 * avoiding the need for sorted input is usually a win, the fact that the
1518 * output won't be sorted may be a loss; so we need to do an actual cost
1521 * We need to consider cheapest_path + hashagg [+ final sort] versus
1522 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1523 * presorted_path + group or agg [+ final sort] where brackets indicate a
1524 * step that may not be needed. We assume query_planner() will have
1525 * returned a presorted path only if it's a winner compared to
1526 * cheapest_path for this purpose.
1528 * These path variables are dummies that just hold cost fields; we don't
1529 * make actual Paths for these steps.
1531 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1532 numGroupCols, dNumGroups,
1533 cheapest_path->startup_cost, cheapest_path->total_cost,
1534 cheapest_path_rows);
1535 /* Result of hashed agg is always unsorted */
1536 if (root->sort_pathkeys)
1537 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1538 dNumGroups, cheapest_path_width, limit_tuples);
1542 sorted_p.startup_cost = sorted_path->startup_cost;
1543 sorted_p.total_cost = sorted_path->total_cost;
1544 current_pathkeys = sorted_path->pathkeys;
1548 sorted_p.startup_cost = cheapest_path->startup_cost;
1549 sorted_p.total_cost = cheapest_path->total_cost;
1550 current_pathkeys = cheapest_path->pathkeys;
1552 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1554 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1555 cheapest_path_rows, cheapest_path_width, -1.0);
1556 current_pathkeys = root->group_pathkeys;
1559 if (root->parse->hasAggs)
1560 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1561 numGroupCols, dNumGroups,
1562 sorted_p.startup_cost, sorted_p.total_cost,
1563 cheapest_path_rows);
1565 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1566 sorted_p.startup_cost, sorted_p.total_cost,
1567 cheapest_path_rows);
1568 /* The Agg or Group node will preserve ordering */
1569 if (root->sort_pathkeys &&
1570 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1571 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1572 dNumGroups, cheapest_path_width, limit_tuples);
1575 * Now make the decision using the top-level tuple fraction. First we
1576 * have to convert an absolute count (LIMIT) into fractional form.
1578 if (tuple_fraction >= 1.0)
1579 tuple_fraction /= dNumGroups;
1581 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1582 tuple_fraction) < 0)
1584 /* Hashed is cheaper, so use it */
1591 * make_subplanTargetList
1592 * Generate appropriate target list when grouping is required.
1594 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1595 * above the result of query_planner, we typically want to pass a different
1596 * target list to query_planner than the outer plan nodes should have.
1597 * This routine generates the correct target list for the subplan.
1599 * The initial target list passed from the parser already contains entries
1600 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1601 * for variables used only in HAVING clauses; so we need to add those
1602 * variables to the subplan target list. Also, we flatten all expressions
1603 * except GROUP BY items into their component variables; the other expressions
1604 * will be computed by the inserted nodes rather than by the subplan.
1605 * For example, given a query like
1606 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1607 * we want to pass this targetlist to the subplan:
1609 * where the a+b target will be used by the Sort/Group steps, and the
1610 * other targets will be used for computing the final results. (In the
1611 * above example we could theoretically suppress the a and b targets and
1612 * pass down only c,d,a+b, but it's not really worth the trouble to
1613 * eliminate simple var references from the subplan. We will avoid doing
1614 * the extra computation to recompute a+b at the outer level; see
1615 * fix_upper_expr() in setrefs.c.)
1617 * If we are grouping or aggregating, *and* there are no non-Var grouping
1618 * expressions, then the returned tlist is effectively dummy; we do not
1619 * need to force it to be evaluated, because all the Vars it contains
1620 * should be present in the output of query_planner anyway.
1622 * 'tlist' is the query's target list.
1623 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1624 * expressions (if there are any) in the subplan's target list.
1625 * 'need_tlist_eval' is set true if we really need to evaluate the
1628 * The result is the targetlist to be passed to the subplan.
1632 make_subplanTargetList(PlannerInfo *root,
1634 AttrNumber **groupColIdx,
1635 bool *need_tlist_eval)
1637 Query *parse = root->parse;
1642 *groupColIdx = NULL;
1645 * If we're not grouping or aggregating, there's nothing to do here;
1646 * query_planner should receive the unmodified target list.
1648 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1650 *need_tlist_eval = true;
1655 * Otherwise, start with a "flattened" tlist (having just the vars
1656 * mentioned in the targetlist and HAVING qual --- but not upper- level
1657 * Vars; they will be replaced by Params later on).
1659 sub_tlist = flatten_tlist(tlist);
1660 extravars = pull_var_clause(parse->havingQual, false);
1661 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1662 list_free(extravars);
1663 *need_tlist_eval = false; /* only eval if not flat tlist */
1666 * If grouping, create sub_tlist entries for all GROUP BY expressions
1667 * (GROUP BY items that are simple Vars should be in the list already),
1668 * and make an array showing where the group columns are in the sub_tlist.
1670 numCols = list_length(parse->groupClause);
1674 AttrNumber *grpColIdx;
1677 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1678 *groupColIdx = grpColIdx;
1680 foreach(gl, parse->groupClause)
1682 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1683 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1684 TargetEntry *te = NULL;
1687 /* Find or make a matching sub_tlist entry */
1688 foreach(sl, sub_tlist)
1690 te = (TargetEntry *) lfirst(sl);
1691 if (equal(groupexpr, te->expr))
1696 te = makeTargetEntry((Expr *) groupexpr,
1697 list_length(sub_tlist) + 1,
1700 sub_tlist = lappend(sub_tlist, te);
1701 *need_tlist_eval = true; /* it's not flat anymore */
1704 /* and save its resno */
1705 grpColIdx[keyno++] = te->resno;
1713 * locate_grouping_columns
1714 * Locate grouping columns in the tlist chosen by query_planner.
1716 * This is only needed if we don't use the sub_tlist chosen by
1717 * make_subplanTargetList. We have to forget the column indexes found
1718 * by that routine and re-locate the grouping vars in the real sub_tlist.
1721 locate_grouping_columns(PlannerInfo *root,
1724 AttrNumber *groupColIdx)
1730 * No work unless grouping.
1732 if (!root->parse->groupClause)
1734 Assert(groupColIdx == NULL);
1737 Assert(groupColIdx != NULL);
1739 foreach(gl, root->parse->groupClause)
1741 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1742 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1743 TargetEntry *te = NULL;
1746 foreach(sl, sub_tlist)
1748 te = (TargetEntry *) lfirst(sl);
1749 if (equal(groupexpr, te->expr))
1753 elog(ERROR, "failed to locate grouping columns");
1755 groupColIdx[keyno++] = te->resno;
1760 * postprocess_setop_tlist
1761 * Fix up targetlist returned by plan_set_operations().
1763 * We need to transpose sort key info from the orig_tlist into new_tlist.
1764 * NOTE: this would not be good enough if we supported resjunk sort keys
1765 * for results of set operations --- then, we'd need to project a whole
1766 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1767 * find any resjunk columns in orig_tlist.
1770 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1773 ListCell *orig_tlist_item = list_head(orig_tlist);
1775 foreach(l, new_tlist)
1777 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1778 TargetEntry *orig_tle;
1780 /* ignore resjunk columns in setop result */
1781 if (new_tle->resjunk)
1784 Assert(orig_tlist_item != NULL);
1785 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1786 orig_tlist_item = lnext(orig_tlist_item);
1787 if (orig_tle->resjunk) /* should not happen */
1788 elog(ERROR, "resjunk output columns are not implemented");
1789 Assert(new_tle->resno == orig_tle->resno);
1790 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1792 if (orig_tlist_item != NULL)
1793 elog(ERROR, "resjunk output columns are not implemented");