1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2005, 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 1.196 2005/12/20 02:30:36 tgl Exp $
13 *-------------------------------------------------------------------------
20 #include "catalog/pg_operator.h"
21 #include "catalog/pg_type.h"
22 #include "executor/executor.h"
23 #include "executor/nodeAgg.h"
24 #include "miscadmin.h"
25 #include "nodes/makefuncs.h"
26 #ifdef OPTIMIZER_DEBUG
27 #include "nodes/print.h"
29 #include "optimizer/clauses.h"
30 #include "optimizer/cost.h"
31 #include "optimizer/pathnode.h"
32 #include "optimizer/paths.h"
33 #include "optimizer/planmain.h"
34 #include "optimizer/planner.h"
35 #include "optimizer/prep.h"
36 #include "optimizer/subselect.h"
37 #include "optimizer/tlist.h"
38 #include "optimizer/var.h"
39 #include "parser/parsetree.h"
40 #include "parser/parse_expr.h"
41 #include "parser/parse_oper.h"
42 #include "utils/selfuncs.h"
43 #include "utils/syscache.h"
46 ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
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_LIMIT 3
54 #define EXPRKIND_ININFO 4
57 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
58 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
59 static Plan *inheritance_planner(PlannerInfo *root, List *inheritlist);
60 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
61 static double preprocess_limit(PlannerInfo *root,
62 double tuple_fraction,
63 int *offset_est, int *count_est);
64 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
65 Path *cheapest_path, Path *sorted_path,
66 double dNumGroups, AggClauseCounts *agg_counts);
67 static bool hash_safe_grouping(PlannerInfo *root);
68 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
69 AttrNumber **groupColIdx, bool *need_tlist_eval);
70 static void locate_grouping_columns(PlannerInfo *root,
73 AttrNumber *groupColIdx);
74 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
77 /*****************************************************************************
79 * Query optimizer entry point
81 *****************************************************************************/
83 planner(Query *parse, bool isCursor, int cursorOptions,
84 ParamListInfo boundParams)
86 double tuple_fraction;
88 Index save_PlannerQueryLevel;
89 List *save_PlannerParamList;
90 ParamListInfo save_PlannerBoundParamList;
93 * The planner can be called recursively (an example is when
94 * eval_const_expressions tries to pre-evaluate an SQL function). So,
95 * these global state variables must be saved and restored.
97 * Query level and the param list cannot be moved into the per-query
98 * PlannerInfo structure since their whole purpose is communication across
99 * multiple sub-queries. Also, boundParams is explicitly info from outside
100 * the query, and so is likewise better handled as a global variable.
102 * Note we do NOT save and restore PlannerPlanId: it exists to assign
103 * unique IDs to SubPlan nodes, and we want those IDs to be unique for the
104 * life of a backend. Also, PlannerInitPlan is saved/restored in
105 * subquery_planner, not here.
107 save_PlannerQueryLevel = PlannerQueryLevel;
108 save_PlannerParamList = PlannerParamList;
109 save_PlannerBoundParamList = PlannerBoundParamList;
111 /* Initialize state for handling outer-level references and params */
112 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
113 PlannerParamList = NIL;
114 PlannerBoundParamList = boundParams;
116 /* Determine what fraction of the plan is likely to be scanned */
120 * We have no real idea how many tuples the user will ultimately FETCH
121 * from a cursor, but it seems a good bet that he doesn't want 'em
122 * all. Optimize for 10% retrieval (you gotta better number? Should
123 * this be a SETtable parameter?)
125 tuple_fraction = 0.10;
129 /* Default assumption is we need all the tuples */
130 tuple_fraction = 0.0;
133 /* primary planning entry point (may recurse for subqueries) */
134 result_plan = subquery_planner(parse, tuple_fraction, NULL);
136 /* check we popped out the right number of levels */
137 Assert(PlannerQueryLevel == 0);
140 * If creating a plan for a scrollable cursor, make sure it can run
141 * backwards on demand. Add a Material node at the top at need.
143 if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
145 if (!ExecSupportsBackwardScan(result_plan))
146 result_plan = materialize_finished_plan(result_plan);
149 /* final cleanup of the plan */
150 result_plan = set_plan_references(result_plan, parse->rtable);
152 /* executor wants to know total number of Params used overall */
153 result_plan->nParamExec = list_length(PlannerParamList);
155 /* restore state for outer planner, if any */
156 PlannerQueryLevel = save_PlannerQueryLevel;
157 PlannerParamList = save_PlannerParamList;
158 PlannerBoundParamList = save_PlannerBoundParamList;
164 /*--------------------
166 * Invokes the planner on a subquery. We recurse to here for each
167 * sub-SELECT found in the query tree.
169 * parse is the querytree produced by the parser & rewriter.
170 * tuple_fraction is the fraction of tuples we expect will be retrieved.
171 * tuple_fraction is interpreted as explained for grouping_planner, below.
173 * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
174 * the output sort ordering of the completed plan.
176 * Basically, this routine does the stuff that should only be done once
177 * per Query object. It then calls grouping_planner. At one time,
178 * grouping_planner could be invoked recursively on the same Query object;
179 * that's not currently true, but we keep the separation between the two
180 * routines anyway, in case we need it again someday.
182 * subquery_planner will be called recursively to handle sub-Query nodes
183 * found within the query's expressions and rangetable.
185 * Returns a query plan.
186 *--------------------
189 subquery_planner(Query *parse, double tuple_fraction,
190 List **subquery_pathkeys)
192 List *saved_initplan = PlannerInitPlan;
193 int saved_planid = PlannerPlanId;
200 /* Set up for a new level of subquery */
202 PlannerInitPlan = NIL;
204 /* Create a PlannerInfo data structure for this subquery */
205 root = makeNode(PlannerInfo);
209 * Look for IN clauses at the top level of WHERE, and transform them into
210 * joins. Note that this step only handles IN clauses originally at top
211 * level of WHERE; if we pull up any subqueries in the next step, their
212 * INs are processed just before pulling them up.
214 root->in_info_list = NIL;
215 if (parse->hasSubLinks)
216 parse->jointree->quals = pull_up_IN_clauses(root,
217 parse->jointree->quals);
220 * Check to see if any subqueries in the rangetable can be merged into
223 parse->jointree = (FromExpr *)
224 pull_up_subqueries(root, (Node *) parse->jointree, false);
227 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
228 * avoid the expense of doing flatten_join_alias_vars(). Also check for
229 * outer joins --- if none, we can skip reduce_outer_joins() and some
230 * other processing. This must be done after we have done
231 * pull_up_subqueries, of course.
233 * Note: if reduce_outer_joins manages to eliminate all outer joins,
234 * root->hasOuterJoins is not reset currently. This is OK since its
235 * purpose is merely to suppress unnecessary processing in simple cases.
237 root->hasJoinRTEs = false;
238 root->hasOuterJoins = false;
239 foreach(l, parse->rtable)
241 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
243 if (rte->rtekind == RTE_JOIN)
245 root->hasJoinRTEs = true;
246 if (IS_OUTER_JOIN(rte->jointype))
248 root->hasOuterJoins = true;
249 /* Can quit scanning once we find an outer join */
256 * Set hasHavingQual to remember if HAVING clause is present. Needed
257 * because preprocess_expression will reduce a constant-true condition to
258 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
260 root->hasHavingQual = (parse->havingQual != NULL);
263 * Do expression preprocessing on targetlist and quals.
265 parse->targetList = (List *)
266 preprocess_expression(root, (Node *) parse->targetList,
269 preprocess_qual_conditions(root, (Node *) parse->jointree);
271 parse->havingQual = preprocess_expression(root, parse->havingQual,
274 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
276 parse->limitCount = preprocess_expression(root, parse->limitCount,
279 root->in_info_list = (List *)
280 preprocess_expression(root, (Node *) root->in_info_list,
283 /* Also need to preprocess expressions for function RTEs */
284 foreach(l, parse->rtable)
286 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
288 if (rte->rtekind == RTE_FUNCTION)
289 rte->funcexpr = preprocess_expression(root, rte->funcexpr,
294 * In some cases we may want to transfer a HAVING clause into WHERE. We
295 * cannot do so if the HAVING clause contains aggregates (obviously) or
296 * volatile functions (since a HAVING clause is supposed to be executed
297 * only once per group). Also, it may be that the clause is so expensive
298 * to execute that we're better off doing it only once per group, despite
299 * the loss of selectivity. This is hard to estimate short of doing the
300 * entire planning process twice, so we use a heuristic: clauses
301 * containing subplans are left in HAVING. Otherwise, we move or copy the
302 * HAVING clause into WHERE, in hopes of eliminating tuples before
303 * aggregation instead of after.
305 * If the query has explicit grouping then we can simply move such a
306 * clause into WHERE; any group that fails the clause will not be in the
307 * output because none of its tuples will reach the grouping or
308 * aggregation stage. Otherwise we must have a degenerate (variable-free)
309 * HAVING clause, which we put in WHERE so that query_planner() can use it
310 * in a gating Result node, but also keep in HAVING to ensure that we
311 * don't emit a bogus aggregated row. (This could be done better, but it
312 * seems not worth optimizing.)
314 * Note that both havingQual and parse->jointree->quals are in
315 * implicitly-ANDed-list form at this point, even though they are declared
319 foreach(l, (List *) parse->havingQual)
321 Node *havingclause = (Node *) lfirst(l);
323 if (contain_agg_clause(havingclause) ||
324 contain_volatile_functions(havingclause) ||
325 contain_subplans(havingclause))
327 /* keep it in HAVING */
328 newHaving = lappend(newHaving, havingclause);
330 else if (parse->groupClause)
332 /* move it to WHERE */
333 parse->jointree->quals = (Node *)
334 lappend((List *) parse->jointree->quals, havingclause);
338 /* put a copy in WHERE, keep it in HAVING */
339 parse->jointree->quals = (Node *)
340 lappend((List *) parse->jointree->quals,
341 copyObject(havingclause));
342 newHaving = lappend(newHaving, havingclause);
345 parse->havingQual = (Node *) newHaving;
348 * If we have any outer joins, try to reduce them to plain inner joins.
349 * This step is most easily done after we've done expression
352 if (root->hasOuterJoins)
353 reduce_outer_joins(root);
356 * Do the main planning. If we have an inherited target relation, that
357 * needs special processing, else go straight to grouping_planner.
359 if (parse->resultRelation &&
360 (lst = expand_inherited_rtentry(root, parse->resultRelation)) != NIL)
361 plan = inheritance_planner(root, lst);
363 plan = grouping_planner(root, tuple_fraction);
366 * If any subplans were generated, or if we're inside a subplan, build
367 * initPlan list and extParam/allParam sets for plan nodes, and attach the
368 * initPlans to the top plan node.
370 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
371 SS_finalize_plan(plan, parse->rtable);
373 /* Return sort ordering info if caller wants it */
374 if (subquery_pathkeys)
375 *subquery_pathkeys = root->query_pathkeys;
377 /* Return to outer subquery context */
379 PlannerInitPlan = saved_initplan;
380 /* we do NOT restore PlannerPlanId; that's not an oversight! */
386 * preprocess_expression
387 * Do subquery_planner's preprocessing work for an expression,
388 * which can be a targetlist, a WHERE clause (including JOIN/ON
389 * conditions), or a HAVING clause.
392 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
395 * Fall out quickly if expression is empty. This occurs often enough to
396 * be worth checking. Note that null->null is the correct conversion for
397 * implicit-AND result format, too.
403 * If the query has any join RTEs, replace join alias variables with
404 * base-relation variables. We must do this before sublink processing,
405 * else sublinks expanded out from join aliases wouldn't get processed.
407 if (root->hasJoinRTEs)
408 expr = flatten_join_alias_vars(root, expr);
411 * Simplify constant expressions.
413 * Note: this also flattens nested AND and OR expressions into N-argument
414 * form. All processing of a qual expression after this point must be
415 * careful to maintain AND/OR flatness --- that is, do not generate a tree
416 * with AND directly under AND, nor OR directly under OR.
418 * Because this is a relatively expensive process, we skip it when the
419 * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
420 * expression will only be evaluated once anyway, so no point in
421 * pre-simplifying; we can't execute it any faster than the executor can,
422 * and we will waste cycles copying the tree. Notice however that we
423 * still must do it for quals (to get AND/OR flatness); and if we are in a
424 * subquery we should not assume it will be done only once.
426 if (root->parse->jointree->fromlist != NIL ||
427 kind == EXPRKIND_QUAL ||
428 PlannerQueryLevel > 1)
429 expr = eval_const_expressions(expr);
432 * If it's a qual or havingQual, canonicalize it.
434 if (kind == EXPRKIND_QUAL)
436 expr = (Node *) canonicalize_qual((Expr *) expr);
438 #ifdef OPTIMIZER_DEBUG
439 printf("After canonicalize_qual()\n");
444 /* Expand SubLinks to SubPlans */
445 if (root->parse->hasSubLinks)
446 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
449 * XXX do not insert anything here unless you have grokked the comments in
450 * SS_replace_correlation_vars ...
453 /* Replace uplevel vars with Param nodes */
454 if (PlannerQueryLevel > 1)
455 expr = SS_replace_correlation_vars(expr);
458 * If it's a qual or havingQual, convert it to implicit-AND format. (We
459 * don't want to do this before eval_const_expressions, since the latter
460 * would be unable to simplify a top-level AND correctly. Also,
461 * SS_process_sublinks expects explicit-AND format.)
463 if (kind == EXPRKIND_QUAL)
464 expr = (Node *) make_ands_implicit((Expr *) expr);
470 * preprocess_qual_conditions
471 * Recursively scan the query's jointree and do subquery_planner's
472 * preprocessing work on each qual condition found therein.
475 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
479 if (IsA(jtnode, RangeTblRef))
481 /* nothing to do here */
483 else if (IsA(jtnode, FromExpr))
485 FromExpr *f = (FromExpr *) jtnode;
488 foreach(l, f->fromlist)
489 preprocess_qual_conditions(root, lfirst(l));
491 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
493 else if (IsA(jtnode, JoinExpr))
495 JoinExpr *j = (JoinExpr *) jtnode;
497 preprocess_qual_conditions(root, j->larg);
498 preprocess_qual_conditions(root, j->rarg);
500 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
503 elog(ERROR, "unrecognized node type: %d",
504 (int) nodeTag(jtnode));
507 /*--------------------
508 * inheritance_planner
509 * Generate a plan in the case where the result relation is an
512 * We have to handle this case differently from cases where a source
513 * relation is an inheritance set. Source inheritance is expanded at
514 * the bottom of the plan tree (see allpaths.c), but target inheritance
515 * has to be expanded at the top. The reason is that for UPDATE, each
516 * target relation needs a different targetlist matching its own column
517 * set. (This is not so critical for DELETE, but for simplicity we treat
518 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
519 * can never be the nullable side of an outer join, so it's OK to generate
522 * inheritlist is an integer list of RT indexes for the result relation set.
524 * Returns a query plan.
525 *--------------------
528 inheritance_planner(PlannerInfo *root, List *inheritlist)
530 Query *parse = root->parse;
531 int parentRTindex = parse->resultRelation;
532 Oid parentOID = getrelid(parentRTindex, parse->rtable);
533 int mainrtlength = list_length(parse->rtable);
534 List *subplans = NIL;
538 foreach(l, inheritlist)
540 int childRTindex = lfirst_int(l);
541 Oid childOID = getrelid(childRTindex, parse->rtable);
546 * Generate modified query with this rel as target. We have to be
547 * prepared to translate varnos in in_info_list as well as in the
550 memcpy(&subroot, root, sizeof(PlannerInfo));
551 subroot.parse = (Query *)
552 adjust_inherited_attrs((Node *) parse,
553 parentRTindex, parentOID,
554 childRTindex, childOID);
555 subroot.in_info_list = (List *)
556 adjust_inherited_attrs((Node *) root->in_info_list,
557 parentRTindex, parentOID,
558 childRTindex, childOID);
559 /* There shouldn't be any OJ info to translate, though */
560 Assert(subroot.oj_info_list == NIL);
563 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
565 subplans = lappend(subplans, subplan);
568 * XXX my goodness this next bit is ugly. Really need to think about
569 * ways to rein in planner's habit of scribbling on its input.
571 * Planning of the subquery might have modified the rangetable, either
572 * by addition of RTEs due to expansion of inherited source tables, or
573 * by changes of the Query structures inside subquery RTEs. We have
574 * to ensure that this gets propagated back to the master copy.
575 * However, if we aren't done planning yet, we also need to ensure
576 * that subsequent calls to grouping_planner have virgin sub-Queries
577 * to work from. So, if we are at the last list entry, just copy the
578 * subquery rangetable back to the master copy; if we are not, then
579 * extend the master copy by adding whatever the subquery added. (We
580 * assume these added entries will go untouched by the future
581 * grouping_planner calls. We are also effectively assuming that
582 * sub-Queries will get planned identically each time, or at least
583 * that the impacts on their rangetables will be the same each time.
584 * Did I say this is ugly?)
586 if (lnext(l) == NULL)
587 parse->rtable = subroot.parse->rtable;
590 int subrtlength = list_length(subroot.parse->rtable);
592 if (subrtlength > mainrtlength)
596 subrt = list_copy_tail(subroot.parse->rtable, mainrtlength);
597 parse->rtable = list_concat(parse->rtable, subrt);
598 mainrtlength = subrtlength;
602 /* Save preprocessed tlist from first rel for use in Append */
604 tlist = subplan->targetlist;
607 /* Save the target-relations list for the executor, too */
608 parse->resultRelations = inheritlist;
610 /* Mark result as unordered (probably unnecessary) */
611 root->query_pathkeys = NIL;
613 return (Plan *) make_append(subplans, true, tlist);
616 /*--------------------
618 * Perform planning steps related to grouping, aggregation, etc.
619 * This primarily means adding top-level processing to the basic
620 * query plan produced by query_planner.
622 * tuple_fraction is the fraction of tuples we expect will be retrieved
624 * tuple_fraction is interpreted as follows:
625 * 0: expect all tuples to be retrieved (normal case)
626 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
627 * from the plan to be retrieved
628 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
629 * expected to be retrieved (ie, a LIMIT specification)
631 * Returns a query plan. Also, root->query_pathkeys is returned as the
632 * actual output ordering of the plan (in pathkey format).
633 *--------------------
636 grouping_planner(PlannerInfo *root, double tuple_fraction)
638 Query *parse = root->parse;
639 List *tlist = parse->targetList;
643 List *current_pathkeys;
645 double dNumGroups = 0;
647 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
648 if (parse->limitCount || parse->limitOffset)
649 tuple_fraction = preprocess_limit(root, tuple_fraction,
650 &offset_est, &count_est);
652 if (parse->setOperations)
654 List *set_sortclauses;
657 * If there's a top-level ORDER BY, assume we have to fetch all the
658 * tuples. This might seem too simplistic given all the hackery below
659 * to possibly avoid the sort ... but a nonzero tuple_fraction is only
660 * of use to plan_set_operations() when the setop is UNION ALL, and
661 * the result of UNION ALL is always unsorted.
663 if (parse->sortClause)
664 tuple_fraction = 0.0;
667 * Construct the plan for set operations. The result will not need
668 * any work except perhaps a top-level sort and/or LIMIT.
670 result_plan = plan_set_operations(root, tuple_fraction,
674 * Calculate pathkeys representing the sort order (if any) of the set
675 * operation's result. We have to do this before overwriting the sort
678 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
679 result_plan->targetlist);
680 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
683 * We should not need to call preprocess_targetlist, since we must be
684 * in a SELECT query node. Instead, use the targetlist returned by
685 * plan_set_operations (since this tells whether it returned any
686 * resjunk columns!), and transfer any sort key information from the
689 Assert(parse->commandType == CMD_SELECT);
691 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
694 * Can't handle FOR UPDATE/SHARE here (parser should have checked
695 * already, but let's make sure).
699 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
700 errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
703 * Calculate pathkeys that represent result ordering requirements
705 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
707 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
711 /* No set operations, do regular planning */
713 List *group_pathkeys;
714 AttrNumber *groupColIdx = NULL;
715 bool need_tlist_eval = true;
721 AggClauseCounts agg_counts;
722 int numGroupCols = list_length(parse->groupClause);
723 bool use_hashed_grouping = false;
725 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
727 /* Preprocess targetlist */
728 tlist = preprocess_targetlist(root, tlist);
731 * Generate appropriate target list for subplan; may be different from
732 * tlist if grouping or aggregation is needed.
734 sub_tlist = make_subplanTargetList(root, tlist,
735 &groupColIdx, &need_tlist_eval);
738 * Calculate pathkeys that represent grouping/ordering requirements.
739 * Stash them in PlannerInfo so that query_planner can canonicalize
742 root->group_pathkeys =
743 make_pathkeys_for_sortclauses(parse->groupClause, tlist);
744 root->sort_pathkeys =
745 make_pathkeys_for_sortclauses(parse->sortClause, tlist);
748 * Will need actual number of aggregates for estimating costs.
750 * Note: we do not attempt to detect duplicate aggregates here; a
751 * somewhat-overestimated count is okay for our present purposes.
753 * Note: think not that we can turn off hasAggs if we find no aggs. It
754 * is possible for constant-expression simplification to remove all
755 * explicit references to aggs, but we still have to follow the
756 * aggregate semantics (eg, producing only one output row).
760 count_agg_clauses((Node *) tlist, &agg_counts);
761 count_agg_clauses(parse->havingQual, &agg_counts);
765 * Figure out whether we need a sorted result from query_planner.
767 * If we have a GROUP BY clause, then we want a result sorted properly
768 * for grouping. Otherwise, if there is an ORDER BY clause, we want
769 * to sort by the ORDER BY clause. (Note: if we have both, and ORDER
770 * BY is a superset of GROUP BY, it would be tempting to request sort
771 * by ORDER BY --- but that might just leave us failing to exploit an
772 * available sort order at all. Needs more thought...)
774 if (parse->groupClause)
775 root->query_pathkeys = root->group_pathkeys;
776 else if (parse->sortClause)
777 root->query_pathkeys = root->sort_pathkeys;
779 root->query_pathkeys = NIL;
782 * Generate the best unsorted and presorted paths for this Query (but
783 * note there may not be any presorted path). query_planner will also
784 * estimate the number of groups in the query, and canonicalize all
787 query_planner(root, sub_tlist, tuple_fraction,
788 &cheapest_path, &sorted_path, &dNumGroups);
790 group_pathkeys = root->group_pathkeys;
791 sort_pathkeys = root->sort_pathkeys;
794 * If grouping, decide whether we want to use hashed grouping.
796 if (parse->groupClause)
798 use_hashed_grouping =
799 choose_hashed_grouping(root, tuple_fraction,
800 cheapest_path, sorted_path,
801 dNumGroups, &agg_counts);
803 /* Also convert # groups to long int --- but 'ware overflow! */
804 numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
808 * Select the best path. If we are doing hashed grouping, we will
809 * always read all the input tuples, so use the cheapest-total path.
810 * Otherwise, trust query_planner's decision about which to use.
812 if (use_hashed_grouping || !sorted_path)
813 best_path = cheapest_path;
815 best_path = sorted_path;
818 * Check to see if it's possible to optimize MIN/MAX aggregates. If
819 * so, we will forget all the work we did so far to choose a "regular"
820 * path ... but we had to do it anyway to be able to tell which way is
823 result_plan = optimize_minmax_aggregates(root,
826 if (result_plan != NULL)
829 * optimize_minmax_aggregates generated the full plan, with the
830 * right tlist, and it has no sort order.
832 current_pathkeys = NIL;
837 * Normal case --- create a plan according to query_planner's
840 result_plan = create_plan(root, best_path);
841 current_pathkeys = best_path->pathkeys;
844 * create_plan() returns a plan with just a "flat" tlist of
845 * required Vars. Usually we need to insert the sub_tlist as the
846 * tlist of the top plan node. However, we can skip that if we
847 * determined that whatever query_planner chose to return will be
853 * If the top-level plan node is one that cannot do expression
854 * evaluation, we must insert a Result node to project the
857 if (!is_projection_capable_plan(result_plan))
859 result_plan = (Plan *) make_result(sub_tlist, NULL,
865 * Otherwise, just replace the subplan's flat tlist with
868 result_plan->targetlist = sub_tlist;
872 * Also, account for the cost of evaluation of the sub_tlist.
874 * Up to now, we have only been dealing with "flat" tlists,
875 * containing just Vars. So their evaluation cost is zero
876 * according to the model used by cost_qual_eval() (or if you
877 * prefer, the cost is factored into cpu_tuple_cost). Thus we
878 * can avoid accounting for tlist cost throughout
879 * query_planner() and subroutines. But now we've inserted a
880 * tlist that might contain actual operators, sub-selects, etc
881 * --- so we'd better account for its cost.
883 * Below this point, any tlist eval cost for added-on nodes
884 * should be accounted for as we create those nodes.
885 * Presently, of the node types we can add on, only Agg and
886 * Group project new tlists (the rest just copy their input
887 * tuples) --- so make_agg() and make_group() are responsible
888 * for computing the added cost.
890 cost_qual_eval(&tlist_cost, sub_tlist);
891 result_plan->startup_cost += tlist_cost.startup;
892 result_plan->total_cost += tlist_cost.startup +
893 tlist_cost.per_tuple * result_plan->plan_rows;
898 * Since we're using query_planner's tlist and not the one
899 * make_subplanTargetList calculated, we have to refigure any
900 * grouping-column indexes make_subplanTargetList computed.
902 locate_grouping_columns(root, tlist, result_plan->targetlist,
907 * Insert AGG or GROUP node if needed, plus an explicit sort step
910 * HAVING clause, if any, becomes qual of the Agg or Group node.
912 if (use_hashed_grouping)
914 /* Hashed aggregate plan --- no sort needed */
915 result_plan = (Plan *) make_agg(root,
917 (List *) parse->havingQual,
924 /* Hashed aggregation produces randomly-ordered results */
925 current_pathkeys = NIL;
927 else if (parse->hasAggs)
929 /* Plain aggregate plan --- sort if needed */
930 AggStrategy aggstrategy;
932 if (parse->groupClause)
934 if (!pathkeys_contained_in(group_pathkeys,
937 result_plan = (Plan *)
938 make_sort_from_groupcols(root,
942 current_pathkeys = group_pathkeys;
944 aggstrategy = AGG_SORTED;
947 * The AGG node will not change the sort ordering of its
948 * groups, so current_pathkeys describes the result too.
953 aggstrategy = AGG_PLAIN;
954 /* Result will be only one row anyway; no sort order */
955 current_pathkeys = NIL;
958 result_plan = (Plan *) make_agg(root,
960 (List *) parse->havingQual,
968 else if (parse->groupClause)
971 * GROUP BY without aggregation, so insert a group node (plus
972 * the appropriate sort node, if necessary).
974 * Add an explicit sort if we couldn't make the path come out
975 * the way the GROUP node needs it.
977 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
979 result_plan = (Plan *)
980 make_sort_from_groupcols(root,
984 current_pathkeys = group_pathkeys;
987 result_plan = (Plan *) make_group(root,
989 (List *) parse->havingQual,
994 /* The Group node won't change sort ordering */
996 else if (root->hasHavingQual)
999 * No aggregates, and no GROUP BY, but we have a HAVING qual.
1000 * This is a degenerate case in which we are supposed to emit
1001 * either 0 or 1 row depending on whether HAVING succeeds.
1002 * Furthermore, there cannot be any variables in either HAVING
1003 * or the targetlist, so we actually do not need the FROM
1004 * table at all! We can just throw away the plan-so-far and
1005 * generate a Result node. This is a sufficiently unusual
1006 * corner case that it's not worth contorting the structure of
1007 * this routine to avoid having to generate the plan in the
1010 result_plan = (Plan *) make_result(tlist,
1014 } /* end of non-minmax-aggregate case */
1015 } /* end of if (setOperations) */
1018 * If we were not able to make the plan come out in the right order, add
1019 * an explicit sort step.
1021 if (parse->sortClause)
1023 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1025 result_plan = (Plan *)
1026 make_sort_from_sortclauses(root,
1029 current_pathkeys = sort_pathkeys;
1034 * If there is a DISTINCT clause, add the UNIQUE node.
1036 if (parse->distinctClause)
1038 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1041 * If there was grouping or aggregation, leave plan_rows as-is (ie,
1042 * assume the result was already mostly unique). If not, use the
1043 * number of distinct-groups calculated by query_planner.
1045 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1046 result_plan->plan_rows = dNumGroups;
1050 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1052 if (parse->limitCount || parse->limitOffset)
1054 result_plan = (Plan *) make_limit(result_plan,
1062 * Return the actual output ordering in query_pathkeys for possible use by
1063 * an outer query level.
1065 root->query_pathkeys = current_pathkeys;
1071 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1073 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1074 * results back in *count_est and *offset_est. These variables are set to
1075 * 0 if the corresponding clause is not present, and -1 if it's present
1076 * but we couldn't estimate the value for it. (The "0" convention is OK
1077 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1078 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
1079 * usual practice of never estimating less than one row.) These values will
1080 * be passed to make_limit, which see if you change this code.
1082 * The return value is the suitably adjusted tuple_fraction to use for
1083 * planning the query. This adjustment is not overridable, since it reflects
1084 * plan actions that grouping_planner() will certainly take, not assumptions
1088 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1089 int *offset_est, int *count_est)
1091 Query *parse = root->parse;
1093 double limit_fraction;
1095 /* Should not be called unless LIMIT or OFFSET */
1096 Assert(parse->limitCount || parse->limitOffset);
1099 * Try to obtain the clause values. We use estimate_expression_value
1100 * primarily because it can sometimes do something useful with Params.
1102 if (parse->limitCount)
1104 est = estimate_expression_value(parse->limitCount);
1105 if (est && IsA(est, Const))
1107 if (((Const *) est)->constisnull)
1109 /* NULL indicates LIMIT ALL, ie, no limit */
1110 *count_est = 0; /* treat as not present */
1114 *count_est = DatumGetInt32(((Const *) est)->constvalue);
1115 if (*count_est <= 0)
1116 *count_est = 1; /* force to at least 1 */
1120 *count_est = -1; /* can't estimate */
1123 *count_est = 0; /* not present */
1125 if (parse->limitOffset)
1127 est = estimate_expression_value(parse->limitOffset);
1128 if (est && IsA(est, Const))
1130 if (((Const *) est)->constisnull)
1132 /* Treat NULL as no offset; the executor will too */
1133 *offset_est = 0; /* treat as not present */
1137 *offset_est = DatumGetInt32(((Const *) est)->constvalue);
1138 if (*offset_est < 0)
1139 *offset_est = 0; /* less than 0 is same as 0 */
1143 *offset_est = -1; /* can't estimate */
1146 *offset_est = 0; /* not present */
1148 if (*count_est != 0)
1151 * A LIMIT clause limits the absolute number of tuples returned.
1152 * However, if it's not a constant LIMIT then we have to guess; for
1153 * lack of a better idea, assume 10% of the plan's result is wanted.
1155 if (*count_est < 0 || *offset_est < 0)
1157 /* LIMIT or OFFSET is an expression ... punt ... */
1158 limit_fraction = 0.10;
1162 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1163 limit_fraction = (double) *count_est + (double) *offset_est;
1167 * If we have absolute limits from both caller and LIMIT, use the
1168 * smaller value; likewise if they are both fractional. If one is
1169 * fractional and the other absolute, we can't easily determine which
1170 * is smaller, but we use the heuristic that the absolute will usually
1173 if (tuple_fraction >= 1.0)
1175 if (limit_fraction >= 1.0)
1178 tuple_fraction = Min(tuple_fraction, limit_fraction);
1182 /* caller absolute, limit fractional; use caller's value */
1185 else if (tuple_fraction > 0.0)
1187 if (limit_fraction >= 1.0)
1189 /* caller fractional, limit absolute; use limit */
1190 tuple_fraction = limit_fraction;
1194 /* both fractional */
1195 tuple_fraction = Min(tuple_fraction, limit_fraction);
1200 /* no info from caller, just use limit */
1201 tuple_fraction = limit_fraction;
1204 else if (*offset_est != 0 && tuple_fraction > 0.0)
1207 * We have an OFFSET but no LIMIT. This acts entirely differently
1208 * from the LIMIT case: here, we need to increase rather than decrease
1209 * the caller's tuple_fraction, because the OFFSET acts to cause more
1210 * tuples to be fetched instead of fewer. This only matters if we got
1211 * a tuple_fraction > 0, however.
1213 * As above, use 10% if OFFSET is present but unestimatable.
1215 if (*offset_est < 0)
1216 limit_fraction = 0.10;
1218 limit_fraction = (double) *offset_est;
1221 * If we have absolute counts from both caller and OFFSET, add them
1222 * together; likewise if they are both fractional. If one is
1223 * fractional and the other absolute, we want to take the larger, and
1224 * we heuristically assume that's the fractional one.
1226 if (tuple_fraction >= 1.0)
1228 if (limit_fraction >= 1.0)
1230 /* both absolute, so add them together */
1231 tuple_fraction += limit_fraction;
1235 /* caller absolute, limit fractional; use limit */
1236 tuple_fraction = limit_fraction;
1241 if (limit_fraction >= 1.0)
1243 /* caller fractional, limit absolute; use caller's value */
1247 /* both fractional, so add them together */
1248 tuple_fraction += limit_fraction;
1249 if (tuple_fraction >= 1.0)
1250 tuple_fraction = 0.0; /* assume fetch all */
1255 return tuple_fraction;
1259 * choose_hashed_grouping - should we use hashed grouping?
1262 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1263 Path *cheapest_path, Path *sorted_path,
1264 double dNumGroups, AggClauseCounts *agg_counts)
1266 int numGroupCols = list_length(root->parse->groupClause);
1267 double cheapest_path_rows;
1268 int cheapest_path_width;
1270 List *current_pathkeys;
1275 * Check can't-do-it conditions, including whether the grouping operators
1278 * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1279 * (Doing so would imply storing *all* the input values in the hash table,
1280 * which seems like a certain loser.)
1282 if (!enable_hashagg)
1284 if (agg_counts->numDistinctAggs != 0)
1286 if (!hash_safe_grouping(root))
1290 * Don't do it if it doesn't look like the hashtable will fit into
1293 * Beware here of the possibility that cheapest_path->parent is NULL. This
1294 * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1296 if (cheapest_path->parent)
1298 cheapest_path_rows = cheapest_path->parent->rows;
1299 cheapest_path_width = cheapest_path->parent->width;
1303 cheapest_path_rows = 1; /* assume non-set result */
1304 cheapest_path_width = 100; /* arbitrary */
1307 /* Estimate per-hash-entry space at tuple width... */
1308 hashentrysize = cheapest_path_width;
1309 /* plus space for pass-by-ref transition values... */
1310 hashentrysize += agg_counts->transitionSpace;
1311 /* plus the per-hash-entry overhead */
1312 hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1314 if (hashentrysize * dNumGroups > work_mem * 1024L)
1318 * See if the estimated cost is no more than doing it the other way. While
1319 * avoiding the need for sorted input is usually a win, the fact that the
1320 * output won't be sorted may be a loss; so we need to do an actual cost
1323 * We need to consider cheapest_path + hashagg [+ final sort] versus
1324 * either cheapest_path [+ sort] + group or agg [+ final sort] or
1325 * presorted_path + group or agg [+ final sort] where brackets indicate a
1326 * step that may not be needed. We assume query_planner() will have
1327 * returned a presorted path only if it's a winner compared to
1328 * cheapest_path for this purpose.
1330 * These path variables are dummies that just hold cost fields; we don't
1331 * make actual Paths for these steps.
1333 cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1334 numGroupCols, dNumGroups,
1335 cheapest_path->startup_cost, cheapest_path->total_cost,
1336 cheapest_path_rows);
1337 /* Result of hashed agg is always unsorted */
1338 if (root->sort_pathkeys)
1339 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1340 dNumGroups, cheapest_path_width);
1344 sorted_p.startup_cost = sorted_path->startup_cost;
1345 sorted_p.total_cost = sorted_path->total_cost;
1346 current_pathkeys = sorted_path->pathkeys;
1350 sorted_p.startup_cost = cheapest_path->startup_cost;
1351 sorted_p.total_cost = cheapest_path->total_cost;
1352 current_pathkeys = cheapest_path->pathkeys;
1354 if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1356 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1357 cheapest_path_rows, cheapest_path_width);
1358 current_pathkeys = root->group_pathkeys;
1361 if (root->parse->hasAggs)
1362 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1363 numGroupCols, dNumGroups,
1364 sorted_p.startup_cost, sorted_p.total_cost,
1365 cheapest_path_rows);
1367 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1368 sorted_p.startup_cost, sorted_p.total_cost,
1369 cheapest_path_rows);
1370 /* The Agg or Group node will preserve ordering */
1371 if (root->sort_pathkeys &&
1372 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1373 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1374 dNumGroups, cheapest_path_width);
1377 * Now make the decision using the top-level tuple fraction. First we
1378 * have to convert an absolute count (LIMIT) into fractional form.
1380 if (tuple_fraction >= 1.0)
1381 tuple_fraction /= dNumGroups;
1383 if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1384 tuple_fraction) < 0)
1386 /* Hashed is cheaper, so use it */
1393 * hash_safe_grouping - are grouping operators hashable?
1395 * We assume hashed aggregation will work if the datatype's equality operator
1396 * is marked hashjoinable.
1399 hash_safe_grouping(PlannerInfo *root)
1403 foreach(gl, root->parse->groupClause)
1405 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1406 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1407 root->parse->targetList);
1411 optup = equality_oper(exprType((Node *) tle->expr), true);
1414 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1415 ReleaseSysCache(optup);
1423 * make_subplanTargetList
1424 * Generate appropriate target list when grouping is required.
1426 * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1427 * above the result of query_planner, we typically want to pass a different
1428 * target list to query_planner than the outer plan nodes should have.
1429 * This routine generates the correct target list for the subplan.
1431 * The initial target list passed from the parser already contains entries
1432 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1433 * for variables used only in HAVING clauses; so we need to add those
1434 * variables to the subplan target list. Also, we flatten all expressions
1435 * except GROUP BY items into their component variables; the other expressions
1436 * will be computed by the inserted nodes rather than by the subplan.
1437 * For example, given a query like
1438 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1439 * we want to pass this targetlist to the subplan:
1441 * where the a+b target will be used by the Sort/Group steps, and the
1442 * other targets will be used for computing the final results. (In the
1443 * above example we could theoretically suppress the a and b targets and
1444 * pass down only c,d,a+b, but it's not really worth the trouble to
1445 * eliminate simple var references from the subplan. We will avoid doing
1446 * the extra computation to recompute a+b at the outer level; see
1447 * replace_vars_with_subplan_refs() in setrefs.c.)
1449 * If we are grouping or aggregating, *and* there are no non-Var grouping
1450 * expressions, then the returned tlist is effectively dummy; we do not
1451 * need to force it to be evaluated, because all the Vars it contains
1452 * should be present in the output of query_planner anyway.
1454 * 'tlist' is the query's target list.
1455 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1456 * expressions (if there are any) in the subplan's target list.
1457 * 'need_tlist_eval' is set true if we really need to evaluate the
1460 * The result is the targetlist to be passed to the subplan.
1464 make_subplanTargetList(PlannerInfo *root,
1466 AttrNumber **groupColIdx,
1467 bool *need_tlist_eval)
1469 Query *parse = root->parse;
1474 *groupColIdx = NULL;
1477 * If we're not grouping or aggregating, there's nothing to do here;
1478 * query_planner should receive the unmodified target list.
1480 if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1482 *need_tlist_eval = true;
1487 * Otherwise, start with a "flattened" tlist (having just the vars
1488 * mentioned in the targetlist and HAVING qual --- but not upper- level
1489 * Vars; they will be replaced by Params later on).
1491 sub_tlist = flatten_tlist(tlist);
1492 extravars = pull_var_clause(parse->havingQual, false);
1493 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1494 list_free(extravars);
1495 *need_tlist_eval = false; /* only eval if not flat tlist */
1498 * If grouping, create sub_tlist entries for all GROUP BY expressions
1499 * (GROUP BY items that are simple Vars should be in the list already),
1500 * and make an array showing where the group columns are in the sub_tlist.
1502 numCols = list_length(parse->groupClause);
1506 AttrNumber *grpColIdx;
1509 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1510 *groupColIdx = grpColIdx;
1512 foreach(gl, parse->groupClause)
1514 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1515 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1516 TargetEntry *te = NULL;
1519 /* Find or make a matching sub_tlist entry */
1520 foreach(sl, sub_tlist)
1522 te = (TargetEntry *) lfirst(sl);
1523 if (equal(groupexpr, te->expr))
1528 te = makeTargetEntry((Expr *) groupexpr,
1529 list_length(sub_tlist) + 1,
1532 sub_tlist = lappend(sub_tlist, te);
1533 *need_tlist_eval = true; /* it's not flat anymore */
1536 /* and save its resno */
1537 grpColIdx[keyno++] = te->resno;
1545 * locate_grouping_columns
1546 * Locate grouping columns in the tlist chosen by query_planner.
1548 * This is only needed if we don't use the sub_tlist chosen by
1549 * make_subplanTargetList. We have to forget the column indexes found
1550 * by that routine and re-locate the grouping vars in the real sub_tlist.
1553 locate_grouping_columns(PlannerInfo *root,
1556 AttrNumber *groupColIdx)
1562 * No work unless grouping.
1564 if (!root->parse->groupClause)
1566 Assert(groupColIdx == NULL);
1569 Assert(groupColIdx != NULL);
1571 foreach(gl, root->parse->groupClause)
1573 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1574 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1575 TargetEntry *te = NULL;
1578 foreach(sl, sub_tlist)
1580 te = (TargetEntry *) lfirst(sl);
1581 if (equal(groupexpr, te->expr))
1585 elog(ERROR, "failed to locate grouping columns");
1587 groupColIdx[keyno++] = te->resno;
1592 * postprocess_setop_tlist
1593 * Fix up targetlist returned by plan_set_operations().
1595 * We need to transpose sort key info from the orig_tlist into new_tlist.
1596 * NOTE: this would not be good enough if we supported resjunk sort keys
1597 * for results of set operations --- then, we'd need to project a whole
1598 * new tlist to evaluate the resjunk columns. For now, just ereport if we
1599 * find any resjunk columns in orig_tlist.
1602 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1605 ListCell *orig_tlist_item = list_head(orig_tlist);
1607 foreach(l, new_tlist)
1609 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1610 TargetEntry *orig_tle;
1612 /* ignore resjunk columns in setop result */
1613 if (new_tle->resjunk)
1616 Assert(orig_tlist_item != NULL);
1617 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1618 orig_tlist_item = lnext(orig_tlist_item);
1619 if (orig_tle->resjunk) /* should not happen */
1620 elog(ERROR, "resjunk output columns are not implemented");
1621 Assert(new_tle->resno == orig_tle->resno);
1622 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1624 if (orig_tlist_item != NULL)
1625 elog(ERROR, "resjunk output columns are not implemented");