1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2001, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.112 2001/10/30 19:58:58 tgl Exp $
13 *-------------------------------------------------------------------------
18 #include "catalog/pg_type.h"
19 #include "nodes/makefuncs.h"
20 #ifdef OPTIMIZER_DEBUG
21 #include "nodes/print.h"
23 #include "optimizer/clauses.h"
24 #include "optimizer/paths.h"
25 #include "optimizer/planmain.h"
26 #include "optimizer/planner.h"
27 #include "optimizer/prep.h"
28 #include "optimizer/subselect.h"
29 #include "optimizer/tlist.h"
30 #include "optimizer/var.h"
31 #include "parser/analyze.h"
32 #include "parser/parsetree.h"
33 #include "parser/parse_expr.h"
34 #include "rewrite/rewriteManip.h"
35 #include "utils/lsyscache.h"
38 /* Expression kind codes for preprocess_expression */
39 #define EXPRKIND_TARGET 0
40 #define EXPRKIND_WHERE 1
41 #define EXPRKIND_HAVING 2
44 static Node *pull_up_subqueries(Query *parse, Node *jtnode);
45 static bool is_simple_subquery(Query *subquery);
46 static void resolvenew_in_jointree(Node *jtnode, int varno, List *subtlist);
47 static Node *preprocess_jointree(Query *parse, Node *jtnode);
48 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
49 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
50 static Plan *inheritance_planner(Query *parse, List *inheritlist);
51 static Plan *grouping_planner(Query *parse, double tuple_fraction);
52 static List *make_subplanTargetList(Query *parse, List *tlist,
53 AttrNumber **groupColIdx);
54 static Plan *make_groupplan(Query *parse,
55 List *group_tlist, bool tuplePerGroup,
56 List *groupClause, AttrNumber *grpColIdx,
57 bool is_presorted, Plan *subplan);
58 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
61 /*****************************************************************************
63 * Query optimizer entry point
65 *****************************************************************************/
70 Index save_PlannerQueryLevel;
71 List *save_PlannerParamVar;
74 * The planner can be called recursively (an example is when
75 * eval_const_expressions tries to pre-evaluate an SQL function). So,
76 * these global state variables must be saved and restored.
78 * These vars cannot be moved into the Query structure since their whole
79 * purpose is communication across multiple sub-Queries.
81 * Note we do NOT save and restore PlannerPlanId: it exists to assign
82 * unique IDs to SubPlan nodes, and we want those IDs to be unique for
83 * the life of a backend. Also, PlannerInitPlan is saved/restored in
84 * subquery_planner, not here.
86 save_PlannerQueryLevel = PlannerQueryLevel;
87 save_PlannerParamVar = PlannerParamVar;
89 /* Initialize state for handling outer-level references and params */
90 PlannerQueryLevel = 0; /* will be 1 in top-level subquery_planner */
91 PlannerParamVar = NIL;
93 /* primary planning entry point (may recurse for subqueries) */
94 result_plan = subquery_planner(parse, -1.0 /* default case */ );
96 Assert(PlannerQueryLevel == 0);
98 /* executor wants to know total number of Params used overall */
99 result_plan->nParamExec = length(PlannerParamVar);
101 /* final cleanup of the plan */
102 set_plan_references(result_plan);
104 /* restore state for outer planner, if any */
105 PlannerQueryLevel = save_PlannerQueryLevel;
106 PlannerParamVar = save_PlannerParamVar;
112 /*--------------------
114 * Invokes the planner on a subquery. We recurse to here for each
115 * sub-SELECT found in the query tree.
117 * parse is the querytree produced by the parser & rewriter.
118 * tuple_fraction is the fraction of tuples we expect will be retrieved.
119 * tuple_fraction is interpreted as explained for grouping_planner, below.
121 * Basically, this routine does the stuff that should only be done once
122 * per Query object. It then calls grouping_planner. At one time,
123 * grouping_planner could be invoked recursively on the same Query object;
124 * that's not currently true, but we keep the separation between the two
125 * routines anyway, in case we need it again someday.
127 * subquery_planner will be called recursively to handle sub-Query nodes
128 * found within the query's expressions and rangetable.
130 * Returns a query plan.
131 *--------------------
134 subquery_planner(Query *parse, double tuple_fraction)
136 List *saved_initplan = PlannerInitPlan;
137 int saved_planid = PlannerPlanId;
142 /* Set up for a new level of subquery */
144 PlannerInitPlan = NIL;
146 #ifdef ENABLE_KEY_SET_QUERY
147 /* this should go away sometime soon */
148 transformKeySetQuery(parse);
152 * Check to see if any subqueries in the rangetable can be merged into
155 parse->jointree = (FromExpr *)
156 pull_up_subqueries(parse, (Node *) parse->jointree);
159 * If so, we may have created opportunities to simplify the jointree.
161 parse->jointree = (FromExpr *)
162 preprocess_jointree(parse, (Node *) parse->jointree);
165 * Do expression preprocessing on targetlist and quals.
167 parse->targetList = (List *)
168 preprocess_expression(parse, (Node *) parse->targetList,
171 preprocess_qual_conditions(parse, (Node *) parse->jointree);
173 parse->havingQual = preprocess_expression(parse, parse->havingQual,
177 * Check for ungrouped variables passed to subplans in targetlist and
178 * HAVING clause (but not in WHERE or JOIN/ON clauses, since those are
179 * evaluated before grouping). We can't do this any earlier because
180 * we must use the preprocessed targetlist for comparisons of grouped
183 if (parse->hasSubLinks &&
184 (parse->groupClause != NIL || parse->hasAggs))
185 check_subplans_for_ungrouped_vars(parse);
188 * A HAVING clause without aggregates is equivalent to a WHERE clause
189 * (except it can only refer to grouped fields). Transfer any
190 * agg-free clauses of the HAVING qual into WHERE. This may seem like
191 * wasting cycles to cater to stupidly-written queries, but there are
192 * other reasons for doing it. Firstly, if the query contains no aggs
193 * at all, then we aren't going to generate an Agg plan node, and so
194 * there'll be no place to execute HAVING conditions; without this
195 * transfer, we'd lose the HAVING condition entirely, which is wrong.
196 * Secondly, when we push down a qual condition into a sub-query, it's
197 * easiest to push the qual into HAVING always, in case it contains
198 * aggs, and then let this code sort it out.
200 * Note that both havingQual and parse->jointree->quals are in
201 * implicitly-ANDed-list form at this point, even though they are
202 * declared as Node *. Also note that contain_agg_clause does not
203 * recurse into sub-selects, which is exactly what we need here.
206 foreach(lst, (List *) parse->havingQual)
208 Node *havingclause = (Node *) lfirst(lst);
210 if (contain_agg_clause(havingclause))
211 newHaving = lappend(newHaving, havingclause);
213 parse->jointree->quals = (Node *)
214 lappend((List *) parse->jointree->quals, havingclause);
216 parse->havingQual = (Node *) newHaving;
219 * Do the main planning. If we have an inherited target relation,
220 * that needs special processing, else go straight to
223 if (parse->resultRelation &&
224 (lst = expand_inherted_rtentry(parse, parse->resultRelation, false))
226 plan = inheritance_planner(parse, lst);
228 plan = grouping_planner(parse, tuple_fraction);
231 * If any subplans were generated, or if we're inside a subplan, build
232 * subPlan, extParam and locParam lists for plan nodes.
234 if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
236 (void) SS_finalize_plan(plan);
239 * At the moment, SS_finalize_plan doesn't handle initPlans and so
240 * we assign them to the topmost plan node.
242 plan->initPlan = PlannerInitPlan;
243 /* Must add the initPlans' extParams to the topmost node's, too */
244 foreach(lst, plan->initPlan)
246 SubPlan *subplan = (SubPlan *) lfirst(lst);
248 plan->extParam = set_unioni(plan->extParam,
249 subplan->plan->extParam);
253 /* Return to outer subquery context */
255 PlannerInitPlan = saved_initplan;
256 /* we do NOT restore PlannerPlanId; that's not an oversight! */
263 * Look for subqueries in the rangetable that can be pulled up into
264 * the parent query. If the subquery has no special features like
265 * grouping/aggregation then we can merge it into the parent's jointree.
267 * A tricky aspect of this code is that if we pull up a subquery we have
268 * to replace Vars that reference the subquery's outputs throughout the
269 * parent query, including quals attached to jointree nodes above the one
270 * we are currently processing! We handle this by being careful not to
271 * change the jointree structure while recursing: no nodes other than
272 * subquery RangeTblRef entries will be replaced. Also, we can't turn
273 * ResolveNew loose on the whole jointree, because it'll return a mutated
274 * copy of the tree; we have to invoke it just on the quals, instead.
277 pull_up_subqueries(Query *parse, Node *jtnode)
281 if (IsA(jtnode, RangeTblRef))
283 int varno = ((RangeTblRef *) jtnode)->rtindex;
284 RangeTblEntry *rte = rt_fetch(varno, parse->rtable);
285 Query *subquery = rte->subquery;
288 * Is this a subquery RTE, and if so, is the subquery simple
289 * enough to pull up? (If not, do nothing at this node.)
291 * Note: even if the subquery itself is simple enough, we can't pull
292 * it up if there is a reference to its whole tuple result.
294 if (subquery && is_simple_subquery(subquery) &&
295 !contain_whole_tuple_var((Node *) parse, varno, 0))
303 * First, recursively pull up the subquery's subqueries, so
304 * that this routine's processing is complete for its jointree
305 * and rangetable. NB: if the same subquery is referenced
306 * from multiple jointree items (which can't happen normally,
307 * but might after rule rewriting), then we will invoke this
308 * processing multiple times on that subquery. OK because
309 * nothing will happen after the first time. We do have to be
310 * careful to copy everything we pull up, however, or risk
311 * having chunks of structure multiply linked.
313 subquery->jointree = (FromExpr *)
314 pull_up_subqueries(subquery, (Node *) subquery->jointree);
317 * Append the subquery's rangetable to mine (currently, no
318 * adjustments will be needed in the subquery's rtable).
320 rtoffset = length(parse->rtable);
321 parse->rtable = nconc(parse->rtable,
322 copyObject(subquery->rtable));
325 * Make copies of the subquery's jointree and targetlist with
326 * varnos adjusted to match the merged rangetable.
328 subjointree = copyObject(subquery->jointree);
329 OffsetVarNodes(subjointree, rtoffset, 0);
330 subtlist = copyObject(subquery->targetList);
331 OffsetVarNodes((Node *) subtlist, rtoffset, 0);
334 * Replace all of the top query's references to the subquery's
335 * outputs with copies of the adjusted subtlist items, being
336 * careful not to replace any of the jointree structure.
338 parse->targetList = (List *)
339 ResolveNew((Node *) parse->targetList,
340 varno, 0, subtlist, CMD_SELECT, 0);
341 resolvenew_in_jointree((Node *) parse->jointree, varno, subtlist);
343 ResolveNew(parse->havingQual,
344 varno, 0, subtlist, CMD_SELECT, 0);
347 * Pull up any FOR UPDATE markers, too.
349 foreach(l, subquery->rowMarks)
351 int submark = lfirsti(l);
353 parse->rowMarks = lappendi(parse->rowMarks,
358 * Miscellaneous housekeeping.
360 parse->hasSubLinks |= subquery->hasSubLinks;
361 /* subquery won't be pulled up if it hasAggs, so no work there */
364 * Return the adjusted subquery jointree to replace the
365 * RangeTblRef entry in my jointree.
370 else if (IsA(jtnode, FromExpr))
372 FromExpr *f = (FromExpr *) jtnode;
375 foreach(l, f->fromlist)
376 lfirst(l) = pull_up_subqueries(parse, lfirst(l));
378 else if (IsA(jtnode, JoinExpr))
380 JoinExpr *j = (JoinExpr *) jtnode;
383 * At the moment, we can't pull up subqueries that are inside the
384 * nullable side of an outer join, because substituting their
385 * target list entries for upper Var references wouldn't do the
386 * right thing (the entries wouldn't go to NULL when they're
387 * supposed to). Suppressing the pullup is an ugly,
388 * performance-losing hack, but I see no alternative for now. Find
389 * a better way to handle this when we redesign query trees ---
395 j->larg = pull_up_subqueries(parse, j->larg);
396 j->rarg = pull_up_subqueries(parse, j->rarg);
399 j->larg = pull_up_subqueries(parse, j->larg);
404 j->rarg = pull_up_subqueries(parse, j->rarg);
409 * This is where we fail if upper levels of planner
410 * haven't rewritten UNION JOIN as an Append ...
412 elog(ERROR, "UNION JOIN is not implemented yet");
415 elog(ERROR, "pull_up_subqueries: unexpected join type %d",
421 elog(ERROR, "pull_up_subqueries: unexpected node type %d",
428 * Check a subquery in the range table to see if it's simple enough
429 * to pull up into the parent query.
432 is_simple_subquery(Query *subquery)
435 * Let's just make sure it's a valid subselect ...
437 if (!IsA(subquery, Query) ||
438 subquery->commandType != CMD_SELECT ||
439 subquery->resultRelation != 0 ||
440 subquery->into != NULL ||
442 elog(ERROR, "is_simple_subquery: subquery is bogus");
445 * Can't currently pull up a query with setops. Maybe after querytree
448 if (subquery->setOperations)
452 * Can't pull up a subquery involving grouping, aggregation, sorting,
455 if (subquery->hasAggs ||
456 subquery->groupClause ||
457 subquery->havingQual ||
458 subquery->sortClause ||
459 subquery->distinctClause ||
460 subquery->limitOffset ||
461 subquery->limitCount)
465 * Hack: don't try to pull up a subquery with an empty jointree.
466 * query_planner() will correctly generate a Result plan for a
467 * jointree that's totally empty, but I don't think the right things
468 * happen if an empty FromExpr appears lower down in a jointree. Not
469 * worth working hard on this, just to collapse SubqueryScan/Result
472 if (subquery->jointree->fromlist == NIL)
479 * Helper routine for pull_up_subqueries: do ResolveNew on every expression
480 * in the jointree, without changing the jointree structure itself. Ugly,
481 * but there's no other way...
484 resolvenew_in_jointree(Node *jtnode, int varno, List *subtlist)
488 if (IsA(jtnode, RangeTblRef))
490 /* nothing to do here */
492 else if (IsA(jtnode, FromExpr))
494 FromExpr *f = (FromExpr *) jtnode;
497 foreach(l, f->fromlist)
498 resolvenew_in_jointree(lfirst(l), varno, subtlist);
499 f->quals = ResolveNew(f->quals,
500 varno, 0, subtlist, CMD_SELECT, 0);
502 else if (IsA(jtnode, JoinExpr))
504 JoinExpr *j = (JoinExpr *) jtnode;
506 resolvenew_in_jointree(j->larg, varno, subtlist);
507 resolvenew_in_jointree(j->rarg, varno, subtlist);
508 j->quals = ResolveNew(j->quals,
509 varno, 0, subtlist, CMD_SELECT, 0);
512 * We don't bother to update the colvars list, since it won't be
517 elog(ERROR, "resolvenew_in_jointree: unexpected node type %d",
522 * preprocess_jointree
523 * Attempt to simplify a query's jointree.
525 * If we succeed in pulling up a subquery then we might form a jointree
526 * in which a FromExpr is a direct child of another FromExpr. In that
527 * case we can consider collapsing the two FromExprs into one. This is
528 * an optional conversion, since the planner will work correctly either
529 * way. But we may find a better plan (at the cost of more planning time)
530 * if we merge the two nodes.
532 * NOTE: don't try to do this in the same jointree scan that does subquery
533 * pullup! Since we're changing the jointree structure here, that wouldn't
534 * work reliably --- see comments for pull_up_subqueries().
537 preprocess_jointree(Query *parse, Node *jtnode)
541 if (IsA(jtnode, RangeTblRef))
543 /* nothing to do here... */
545 else if (IsA(jtnode, FromExpr))
547 FromExpr *f = (FromExpr *) jtnode;
551 foreach(l, f->fromlist)
553 Node *child = (Node *) lfirst(l);
555 /* Recursively simplify the child... */
556 child = preprocess_jointree(parse, child);
557 /* Now, is it a FromExpr? */
558 if (child && IsA(child, FromExpr))
561 * Yes, so do we want to merge it into parent? Always do
562 * so if child has just one element (since that doesn't
563 * make the parent's list any longer). Otherwise we have
564 * to be careful about the increase in planning time
565 * caused by combining the two join search spaces into
566 * one. Our heuristic is to merge if the merge will
567 * produce a join list no longer than GEQO_RELS/2.
568 * (Perhaps need an additional user parameter?)
570 FromExpr *subf = (FromExpr *) child;
571 int childlen = length(subf->fromlist);
572 int myothers = length(newlist) + length(lnext(l));
574 if (childlen <= 1 || (childlen + myothers) <= geqo_rels / 2)
576 newlist = nconc(newlist, subf->fromlist);
577 f->quals = make_and_qual(f->quals, subf->quals);
580 newlist = lappend(newlist, child);
583 newlist = lappend(newlist, child);
585 f->fromlist = newlist;
587 else if (IsA(jtnode, JoinExpr))
589 JoinExpr *j = (JoinExpr *) jtnode;
591 /* Can't usefully change the JoinExpr, but recurse on children */
592 j->larg = preprocess_jointree(parse, j->larg);
593 j->rarg = preprocess_jointree(parse, j->rarg);
596 elog(ERROR, "preprocess_jointree: unexpected node type %d",
602 * preprocess_expression
603 * Do subquery_planner's preprocessing work for an expression,
604 * which can be a targetlist, a WHERE clause (including JOIN/ON
605 * conditions), or a HAVING clause.
608 preprocess_expression(Query *parse, Node *expr, int kind)
611 * Simplify constant expressions.
613 * Note that at this point quals have not yet been converted to
614 * implicit-AND form, so we can apply eval_const_expressions directly.
615 * Also note that we need to do this before SS_process_sublinks,
616 * because that routine inserts bogus "Const" nodes.
618 expr = eval_const_expressions(expr);
621 * If it's a qual or havingQual, canonicalize it, and convert it to
622 * implicit-AND format.
624 * XXX Is there any value in re-applying eval_const_expressions after
627 if (kind != EXPRKIND_TARGET)
629 expr = (Node *) canonicalize_qual((Expr *) expr, true);
631 #ifdef OPTIMIZER_DEBUG
632 printf("After canonicalize_qual()\n");
637 /* Expand SubLinks to SubPlans */
638 if (parse->hasSubLinks)
639 expr = SS_process_sublinks(expr);
641 /* Replace uplevel vars with Param nodes */
642 if (PlannerQueryLevel > 1)
643 expr = SS_replace_correlation_vars(expr);
649 * preprocess_qual_conditions
650 * Recursively scan the query's jointree and do subquery_planner's
651 * preprocessing work on each qual condition found therein.
654 preprocess_qual_conditions(Query *parse, Node *jtnode)
658 if (IsA(jtnode, RangeTblRef))
660 /* nothing to do here */
662 else if (IsA(jtnode, FromExpr))
664 FromExpr *f = (FromExpr *) jtnode;
667 foreach(l, f->fromlist)
668 preprocess_qual_conditions(parse, lfirst(l));
670 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_WHERE);
672 else if (IsA(jtnode, JoinExpr))
674 JoinExpr *j = (JoinExpr *) jtnode;
676 preprocess_qual_conditions(parse, j->larg);
677 preprocess_qual_conditions(parse, j->rarg);
679 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_WHERE);
682 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
686 /*--------------------
687 * inheritance_planner
688 * Generate a plan in the case where the result relation is an
691 * We have to handle this case differently from cases where a source
692 * relation is an inheritance set. Source inheritance is expanded at
693 * the bottom of the plan tree (see allpaths.c), but target inheritance
694 * has to be expanded at the top. The reason is that for UPDATE, each
695 * target relation needs a different targetlist matching its own column
696 * set. (This is not so critical for DELETE, but for simplicity we treat
697 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
698 * can never be the nullable side of an outer join, so it's OK to generate
701 * parse is the querytree produced by the parser & rewriter.
702 * inheritlist is an integer list of RT indexes for the result relation set.
704 * Returns a query plan.
705 *--------------------
708 inheritance_planner(Query *parse, List *inheritlist)
710 int parentRTindex = parse->resultRelation;
711 Oid parentOID = getrelid(parentRTindex, parse->rtable);
712 List *subplans = NIL;
716 foreach(l, inheritlist)
718 int childRTindex = lfirsti(l);
719 Oid childOID = getrelid(childRTindex, parse->rtable);
723 /* Generate modified query with this rel as target */
724 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
725 parentRTindex, parentOID,
726 childRTindex, childOID);
728 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
729 subplans = lappend(subplans, subplan);
730 /* Save preprocessed tlist from first rel for use in Append */
732 tlist = subplan->targetlist;
735 /* Save the target-relations list for the executor, too */
736 parse->resultRelations = inheritlist;
738 return (Plan *) make_append(subplans, true, tlist);
741 /*--------------------
743 * Perform planning steps related to grouping, aggregation, etc.
744 * This primarily means adding top-level processing to the basic
745 * query plan produced by query_planner.
747 * parse is the querytree produced by the parser & rewriter.
748 * tuple_fraction is the fraction of tuples we expect will be retrieved
750 * tuple_fraction is interpreted as follows:
751 * < 0: determine fraction by inspection of query (normal case)
752 * 0: expect all tuples to be retrieved
753 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
754 * from the plan to be retrieved
755 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
756 * expected to be retrieved (ie, a LIMIT specification)
757 * The normal case is to pass -1, but some callers pass values >= 0 to
758 * override this routine's determination of the appropriate fraction.
760 * Returns a query plan.
761 *--------------------
764 grouping_planner(Query *parse, double tuple_fraction)
766 List *tlist = parse->targetList;
768 List *current_pathkeys;
769 List *group_pathkeys;
771 AttrNumber *groupColIdx = NULL;
773 if (parse->setOperations)
776 * Construct the plan for set operations. The result will not
777 * need any work except perhaps a top-level sort and/or LIMIT.
779 result_plan = plan_set_operations(parse);
782 * We should not need to call preprocess_targetlist, since we must
783 * be in a SELECT query node. Instead, use the targetlist
784 * returned by plan_set_operations (since this tells whether it
785 * returned any resjunk columns!), and transfer any sort key
786 * information from the original tlist.
788 Assert(parse->commandType == CMD_SELECT);
790 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
793 * Can't handle FOR UPDATE here (parser should have checked
794 * already, but let's make sure).
797 elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
800 * We set current_pathkeys NIL indicating we do not know sort
801 * order. This is correct when the top set operation is UNION
802 * ALL, since the appended-together results are unsorted even if
803 * the subplans were sorted. For other set operations we could be
804 * smarter --- room for future improvement!
806 current_pathkeys = NIL;
809 * Calculate pathkeys that represent grouping/ordering
810 * requirements (grouping should always be null, but...)
812 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
814 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
821 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
822 tlist = preprocess_targetlist(tlist,
824 parse->resultRelation,
828 * Add TID targets for rels selected FOR UPDATE (should this be
829 * done in preprocess_targetlist?). The executor uses the TID to
830 * know which rows to lock, much as for UPDATE or DELETE.
837 * We've got trouble if the FOR UPDATE appears inside
838 * grouping, since grouping renders a reference to individual
839 * tuple CTIDs invalid. This is also checked at parse time,
840 * but that's insufficient because of rule substitution, query
843 CheckSelectForUpdate(parse);
846 * Currently the executor only supports FOR UPDATE at top
849 if (PlannerQueryLevel > 1)
850 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
852 foreach(l, parse->rowMarks)
854 Index rti = lfirsti(l);
860 resname = (char *) palloc(32);
861 sprintf(resname, "ctid%u", rti);
862 resdom = makeResdom(length(tlist) + 1,
869 SelfItemPointerAttributeNumber,
874 ctid = makeTargetEntry(resdom, (Node *) var);
875 tlist = lappend(tlist, ctid);
880 * Generate appropriate target list for subplan; may be different
881 * from tlist if grouping or aggregation is needed.
883 sub_tlist = make_subplanTargetList(parse, tlist, &groupColIdx);
886 * Calculate pathkeys that represent grouping/ordering
889 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
891 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
895 * Figure out whether we need a sorted result from query_planner.
897 * If we have a GROUP BY clause, then we want a result sorted
898 * properly for grouping. Otherwise, if there is an ORDER BY
899 * clause, we want to sort by the ORDER BY clause. (Note: if we
900 * have both, and ORDER BY is a superset of GROUP BY, it would be
901 * tempting to request sort by ORDER BY --- but that might just
902 * leave us failing to exploit an available sort order at all.
903 * Needs more thought...)
905 if (parse->groupClause)
906 parse->query_pathkeys = group_pathkeys;
907 else if (parse->sortClause)
908 parse->query_pathkeys = sort_pathkeys;
910 parse->query_pathkeys = NIL;
913 * Figure out whether we expect to retrieve all the tuples that
914 * the plan can generate, or to stop early due to outside factors
915 * such as a cursor. If the caller passed a value >= 0, believe
916 * that value, else do our own examination of the query context.
918 if (tuple_fraction < 0.0)
920 /* Initial assumption is we need all the tuples */
921 tuple_fraction = 0.0;
924 * Check for retrieve-into-portal, ie DECLARE CURSOR.
926 * We have no real idea how many tuples the user will ultimately
927 * FETCH from a cursor, but it seems a good bet that he
928 * doesn't want 'em all. Optimize for 10% retrieval (you
929 * gotta better number? Should this be a SETtable parameter?)
932 tuple_fraction = 0.10;
936 * Adjust tuple_fraction if we see that we are going to apply
937 * limiting/grouping/aggregation/etc. This is not overridable by
938 * the caller, since it reflects plan actions that this routine
939 * will certainly take, not assumptions about context.
941 if (parse->limitCount != NULL)
944 * A LIMIT clause limits the absolute number of tuples
945 * returned. However, if it's not a constant LIMIT then we
946 * have to punt; for lack of a better idea, assume 10% of the
947 * plan's result is wanted.
949 double limit_fraction = 0.0;
951 if (IsA(parse->limitCount, Const))
953 Const *limitc = (Const *) parse->limitCount;
954 int32 count = DatumGetInt32(limitc->constvalue);
957 * A NULL-constant LIMIT represents "LIMIT ALL", which we
958 * treat the same as no limit (ie, expect to retrieve all
961 if (!limitc->constisnull && count > 0)
963 limit_fraction = (double) count;
964 /* We must also consider the OFFSET, if present */
965 if (parse->limitOffset != NULL)
967 if (IsA(parse->limitOffset, Const))
971 limitc = (Const *) parse->limitOffset;
972 offset = DatumGetInt32(limitc->constvalue);
973 if (!limitc->constisnull && offset > 0)
974 limit_fraction += (double) offset;
978 /* OFFSET is an expression ... punt ... */
979 limit_fraction = 0.10;
986 /* LIMIT is an expression ... punt ... */
987 limit_fraction = 0.10;
990 if (limit_fraction > 0.0)
993 * If we have absolute limits from both caller and LIMIT,
994 * use the smaller value; if one is fractional and the
995 * other absolute, treat the fraction as a fraction of the
996 * absolute value; else we can multiply the two fractions
999 if (tuple_fraction >= 1.0)
1001 if (limit_fraction >= 1.0)
1004 tuple_fraction = Min(tuple_fraction, limit_fraction);
1008 /* caller absolute, limit fractional */
1009 tuple_fraction *= limit_fraction;
1010 if (tuple_fraction < 1.0)
1011 tuple_fraction = 1.0;
1014 else if (tuple_fraction > 0.0)
1016 if (limit_fraction >= 1.0)
1018 /* caller fractional, limit absolute */
1019 tuple_fraction *= limit_fraction;
1020 if (tuple_fraction < 1.0)
1021 tuple_fraction = 1.0;
1025 /* both fractional */
1026 tuple_fraction *= limit_fraction;
1031 /* no info from caller, just use limit */
1032 tuple_fraction = limit_fraction;
1037 if (parse->groupClause)
1040 * In GROUP BY mode, we have the little problem that we don't
1041 * really know how many input tuples will be needed to make a
1042 * group, so we can't translate an output LIMIT count into an
1043 * input count. For lack of a better idea, assume 25% of the
1044 * input data will be processed if there is any output limit.
1045 * However, if the caller gave us a fraction rather than an
1046 * absolute count, we can keep using that fraction (which
1047 * amounts to assuming that all the groups are about the same
1050 if (tuple_fraction >= 1.0)
1051 tuple_fraction = 0.25;
1054 * If both GROUP BY and ORDER BY are specified, we will need
1055 * two levels of sort --- and, therefore, certainly need to
1056 * read all the input tuples --- unless ORDER BY is a subset
1057 * of GROUP BY. (We have not yet canonicalized the pathkeys,
1058 * so must use the slower noncanonical comparison method.)
1060 if (parse->groupClause && parse->sortClause &&
1061 !noncanonical_pathkeys_contained_in(sort_pathkeys,
1063 tuple_fraction = 0.0;
1065 else if (parse->hasAggs)
1068 * Ungrouped aggregate will certainly want all the input
1071 tuple_fraction = 0.0;
1073 else if (parse->distinctClause)
1076 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
1077 * number of input tuples per output tuple. Handle the same
1080 if (tuple_fraction >= 1.0)
1081 tuple_fraction = 0.25;
1084 /* Generate the basic plan for this Query */
1085 result_plan = query_planner(parse,
1090 * query_planner returns actual sort order (which is not
1091 * necessarily what we requested) in query_pathkeys.
1093 current_pathkeys = parse->query_pathkeys;
1097 * We couldn't canonicalize group_pathkeys and sort_pathkeys before
1098 * running query_planner(), so do it now.
1100 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
1101 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
1104 * If we have a GROUP BY clause, insert a group node (plus the
1105 * appropriate sort node, if necessary).
1107 if (parse->groupClause)
1114 * Decide whether how many tuples per group the Group node needs
1115 * to return. (Needs only one tuple per group if no aggregate is
1116 * present. Otherwise, need every tuple from the group to do the
1117 * aggregation.) Note tuplePerGroup is named backwards :-(
1119 tuplePerGroup = parse->hasAggs;
1122 * If there are aggregates then the Group node should just return
1123 * the same set of vars as the subplan did. If there are no aggs
1124 * then the Group node had better compute the final tlist.
1127 group_tlist = new_unsorted_tlist(result_plan->targetlist);
1129 group_tlist = tlist;
1132 * Figure out whether the path result is already ordered the way
1133 * we need it --- if so, no need for an explicit sort step.
1135 if (pathkeys_contained_in(group_pathkeys, current_pathkeys))
1137 is_sorted = true; /* no sort needed now */
1138 /* current_pathkeys remains unchanged */
1143 * We will need to do an explicit sort by the GROUP BY clause.
1144 * make_groupplan will do the work, but set current_pathkeys
1145 * to indicate the resulting order.
1148 current_pathkeys = group_pathkeys;
1151 result_plan = make_groupplan(parse,
1161 * If aggregate is present, insert the Agg node
1163 * HAVING clause, if any, becomes qual of the Agg node
1167 result_plan = (Plan *) make_agg(tlist,
1168 (List *) parse->havingQual,
1170 /* Note: Agg does not affect any existing sort order of the tuples */
1174 /* If there are no Aggs, we shouldn't have any HAVING qual anymore */
1175 Assert(parse->havingQual == NULL);
1179 * If we were not able to make the plan come out in the right order,
1180 * add an explicit sort step.
1182 if (parse->sortClause)
1184 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1185 result_plan = make_sortplan(parse, tlist, result_plan,
1190 * If there is a DISTINCT clause, add the UNIQUE node.
1192 if (parse->distinctClause)
1194 result_plan = (Plan *) make_unique(tlist, result_plan,
1195 parse->distinctClause);
1199 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1201 if (parse->limitOffset || parse->limitCount)
1203 result_plan = (Plan *) make_limit(tlist, result_plan,
1212 * make_subplanTargetList
1213 * Generate appropriate target list when grouping is required.
1215 * When grouping_planner inserts Aggregate and/or Group plan nodes above
1216 * the result of query_planner, we typically want to pass a different
1217 * target list to query_planner than the outer plan nodes should have.
1218 * This routine generates the correct target list for the subplan.
1220 * The initial target list passed from the parser already contains entries
1221 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1222 * for variables used only in HAVING clauses; so we need to add those
1223 * variables to the subplan target list. Also, if we are doing either
1224 * grouping or aggregation, we flatten all expressions except GROUP BY items
1225 * into their component variables; the other expressions will be computed by
1226 * the inserted nodes rather than by the subplan. For example,
1227 * given a query like
1228 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1229 * we want to pass this targetlist to the subplan:
1231 * where the a+b target will be used by the Sort/Group steps, and the
1232 * other targets will be used for computing the final results. (In the
1233 * above example we could theoretically suppress the a and b targets and
1234 * pass down only c,d,a+b, but it's not really worth the trouble to
1235 * eliminate simple var references from the subplan. We will avoid doing
1236 * the extra computation to recompute a+b at the outer level; see
1237 * replace_vars_with_subplan_refs() in setrefs.c.)
1239 * 'parse' is the query being processed.
1240 * 'tlist' is the query's target list.
1241 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1242 * expressions (if there are any) in the subplan's target list.
1244 * The result is the targetlist to be passed to the subplan.
1248 make_subplanTargetList(Query *parse,
1250 AttrNumber **groupColIdx)
1256 *groupColIdx = NULL;
1259 * If we're not grouping or aggregating, nothing to do here;
1260 * query_planner should receive the unmodified target list.
1262 if (!parse->hasAggs && !parse->groupClause && !parse->havingQual)
1266 * Otherwise, start with a "flattened" tlist (having just the vars
1267 * mentioned in the targetlist and HAVING qual --- but not upper-
1268 * level Vars; they will be replaced by Params later on).
1270 sub_tlist = flatten_tlist(tlist);
1271 extravars = pull_var_clause(parse->havingQual, false);
1272 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1273 freeList(extravars);
1276 * If grouping, create sub_tlist entries for all GROUP BY expressions
1277 * (GROUP BY items that are simple Vars should be in the list
1278 * already), and make an array showing where the group columns are in
1281 numCols = length(parse->groupClause);
1285 AttrNumber *grpColIdx;
1288 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1289 *groupColIdx = grpColIdx;
1291 foreach(gl, parse->groupClause)
1293 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1294 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1295 TargetEntry *te = NULL;
1298 /* Find or make a matching sub_tlist entry */
1299 foreach(sl, sub_tlist)
1301 te = (TargetEntry *) lfirst(sl);
1302 if (equal(groupexpr, te->expr))
1307 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1308 exprType(groupexpr),
1309 exprTypmod(groupexpr),
1313 sub_tlist = lappend(sub_tlist, te);
1316 /* and save its resno */
1317 grpColIdx[keyno++] = te->resdom->resno;
1326 * Add a Group node for GROUP BY processing.
1327 * If we couldn't make the subplan produce presorted output for grouping,
1328 * first add an explicit Sort node.
1331 make_groupplan(Query *parse,
1335 AttrNumber *grpColIdx,
1339 int numCols = length(groupClause);
1344 * The Sort node always just takes a copy of the subplan's tlist
1345 * plus ordering information. (This might seem inefficient if the
1346 * subplan contains complex GROUP BY expressions, but in fact Sort
1347 * does not evaluate its targetlist --- it only outputs the same
1348 * tuples in a new order. So the expressions we might be copying
1349 * are just dummies with no extra execution cost.)
1351 List *sort_tlist = new_unsorted_tlist(subplan->targetlist);
1355 foreach(gl, groupClause)
1357 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1358 TargetEntry *te = nth(grpColIdx[keyno] - 1, sort_tlist);
1359 Resdom *resdom = te->resdom;
1362 * Check for the possibility of duplicate group-by clauses ---
1363 * the parser should have removed 'em, but the Sort executor
1364 * will get terribly confused if any get through!
1366 if (resdom->reskey == 0)
1368 /* OK, insert the ordering info needed by the executor. */
1369 resdom->reskey = ++keyno;
1370 resdom->reskeyop = grpcl->sortop;
1376 subplan = (Plan *) make_sort(parse, sort_tlist, subplan, keyno);
1379 return (Plan *) make_group(group_tlist, tuplePerGroup, numCols,
1380 grpColIdx, subplan);
1385 * Add a Sort node to implement an explicit ORDER BY clause.
1388 make_sortplan(Query *parse, List *tlist, Plan *plannode, List *sortcls)
1395 * First make a copy of the tlist so that we don't corrupt the
1398 sort_tlist = new_unsorted_tlist(tlist);
1402 SortClause *sortcl = (SortClause *) lfirst(i);
1403 TargetEntry *tle = get_sortgroupclause_tle(sortcl, sort_tlist);
1404 Resdom *resdom = tle->resdom;
1407 * Check for the possibility of duplicate order-by clauses --- the
1408 * parser should have removed 'em, but the executor will get
1409 * terribly confused if any get through!
1411 if (resdom->reskey == 0)
1413 /* OK, insert the ordering info needed by the executor. */
1414 resdom->reskey = ++keyno;
1415 resdom->reskeyop = sortcl->sortop;
1421 return (Plan *) make_sort(parse, sort_tlist, plannode, keyno);
1425 * postprocess_setop_tlist
1426 * Fix up targetlist returned by plan_set_operations().
1428 * We need to transpose sort key info from the orig_tlist into new_tlist.
1429 * NOTE: this would not be good enough if we supported resjunk sort keys
1430 * for results of set operations --- then, we'd need to project a whole
1431 * new tlist to evaluate the resjunk columns. For now, just elog if we
1432 * find any resjunk columns in orig_tlist.
1435 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1439 foreach(l, new_tlist)
1441 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1442 TargetEntry *orig_tle;
1444 /* ignore resjunk columns in setop result */
1445 if (new_tle->resdom->resjunk)
1448 Assert(orig_tlist != NIL);
1449 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1450 orig_tlist = lnext(orig_tlist);
1451 if (orig_tle->resdom->resjunk)
1452 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1453 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1454 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1455 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1457 if (orig_tlist != NIL)
1458 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");