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 1.114 2001/12/10 22:54:12 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 * Don't pull up a subquery that has any set-returning functions in
466 * its targetlist. Otherwise we might well wind up inserting
467 * set-returning functions into places where they mustn't go,
468 * such as quals of higher queries.
470 if (contain_iter_clause((Node *) subquery->targetList))
474 * Hack: don't try to pull up a subquery with an empty jointree.
475 * query_planner() will correctly generate a Result plan for a
476 * jointree that's totally empty, but I don't think the right things
477 * happen if an empty FromExpr appears lower down in a jointree. Not
478 * worth working hard on this, just to collapse SubqueryScan/Result
481 if (subquery->jointree->fromlist == NIL)
488 * Helper routine for pull_up_subqueries: do ResolveNew on every expression
489 * in the jointree, without changing the jointree structure itself. Ugly,
490 * but there's no other way...
493 resolvenew_in_jointree(Node *jtnode, int varno, List *subtlist)
497 if (IsA(jtnode, RangeTblRef))
499 /* nothing to do here */
501 else if (IsA(jtnode, FromExpr))
503 FromExpr *f = (FromExpr *) jtnode;
506 foreach(l, f->fromlist)
507 resolvenew_in_jointree(lfirst(l), varno, subtlist);
508 f->quals = ResolveNew(f->quals,
509 varno, 0, subtlist, CMD_SELECT, 0);
511 else if (IsA(jtnode, JoinExpr))
513 JoinExpr *j = (JoinExpr *) jtnode;
515 resolvenew_in_jointree(j->larg, varno, subtlist);
516 resolvenew_in_jointree(j->rarg, varno, subtlist);
517 j->quals = ResolveNew(j->quals,
518 varno, 0, subtlist, CMD_SELECT, 0);
521 * We don't bother to update the colvars list, since it won't be
526 elog(ERROR, "resolvenew_in_jointree: unexpected node type %d",
531 * preprocess_jointree
532 * Attempt to simplify a query's jointree.
534 * If we succeed in pulling up a subquery then we might form a jointree
535 * in which a FromExpr is a direct child of another FromExpr. In that
536 * case we can consider collapsing the two FromExprs into one. This is
537 * an optional conversion, since the planner will work correctly either
538 * way. But we may find a better plan (at the cost of more planning time)
539 * if we merge the two nodes.
541 * NOTE: don't try to do this in the same jointree scan that does subquery
542 * pullup! Since we're changing the jointree structure here, that wouldn't
543 * work reliably --- see comments for pull_up_subqueries().
546 preprocess_jointree(Query *parse, Node *jtnode)
550 if (IsA(jtnode, RangeTblRef))
552 /* nothing to do here... */
554 else if (IsA(jtnode, FromExpr))
556 FromExpr *f = (FromExpr *) jtnode;
560 foreach(l, f->fromlist)
562 Node *child = (Node *) lfirst(l);
564 /* Recursively simplify the child... */
565 child = preprocess_jointree(parse, child);
566 /* Now, is it a FromExpr? */
567 if (child && IsA(child, FromExpr))
570 * Yes, so do we want to merge it into parent? Always do
571 * so if child has just one element (since that doesn't
572 * make the parent's list any longer). Otherwise we have
573 * to be careful about the increase in planning time
574 * caused by combining the two join search spaces into
575 * one. Our heuristic is to merge if the merge will
576 * produce a join list no longer than GEQO_RELS/2.
577 * (Perhaps need an additional user parameter?)
579 FromExpr *subf = (FromExpr *) child;
580 int childlen = length(subf->fromlist);
581 int myothers = length(newlist) + length(lnext(l));
583 if (childlen <= 1 || (childlen + myothers) <= geqo_rels / 2)
585 newlist = nconc(newlist, subf->fromlist);
586 f->quals = make_and_qual(f->quals, subf->quals);
589 newlist = lappend(newlist, child);
592 newlist = lappend(newlist, child);
594 f->fromlist = newlist;
596 else if (IsA(jtnode, JoinExpr))
598 JoinExpr *j = (JoinExpr *) jtnode;
600 /* Can't usefully change the JoinExpr, but recurse on children */
601 j->larg = preprocess_jointree(parse, j->larg);
602 j->rarg = preprocess_jointree(parse, j->rarg);
605 elog(ERROR, "preprocess_jointree: unexpected node type %d",
611 * preprocess_expression
612 * Do subquery_planner's preprocessing work for an expression,
613 * which can be a targetlist, a WHERE clause (including JOIN/ON
614 * conditions), or a HAVING clause.
617 preprocess_expression(Query *parse, Node *expr, int kind)
620 * Simplify constant expressions.
622 * Note that at this point quals have not yet been converted to
623 * implicit-AND form, so we can apply eval_const_expressions directly.
624 * Also note that we need to do this before SS_process_sublinks,
625 * because that routine inserts bogus "Const" nodes.
627 expr = eval_const_expressions(expr);
630 * If it's a qual or havingQual, canonicalize it, and convert it to
631 * implicit-AND format.
633 * XXX Is there any value in re-applying eval_const_expressions after
636 if (kind != EXPRKIND_TARGET)
638 expr = (Node *) canonicalize_qual((Expr *) expr, true);
640 #ifdef OPTIMIZER_DEBUG
641 printf("After canonicalize_qual()\n");
646 /* Expand SubLinks to SubPlans */
647 if (parse->hasSubLinks)
648 expr = SS_process_sublinks(expr);
650 /* Replace uplevel vars with Param nodes */
651 if (PlannerQueryLevel > 1)
652 expr = SS_replace_correlation_vars(expr);
658 * preprocess_qual_conditions
659 * Recursively scan the query's jointree and do subquery_planner's
660 * preprocessing work on each qual condition found therein.
663 preprocess_qual_conditions(Query *parse, Node *jtnode)
667 if (IsA(jtnode, RangeTblRef))
669 /* nothing to do here */
671 else if (IsA(jtnode, FromExpr))
673 FromExpr *f = (FromExpr *) jtnode;
676 foreach(l, f->fromlist)
677 preprocess_qual_conditions(parse, lfirst(l));
679 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_WHERE);
681 else if (IsA(jtnode, JoinExpr))
683 JoinExpr *j = (JoinExpr *) jtnode;
685 preprocess_qual_conditions(parse, j->larg);
686 preprocess_qual_conditions(parse, j->rarg);
688 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_WHERE);
691 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
695 /*--------------------
696 * inheritance_planner
697 * Generate a plan in the case where the result relation is an
700 * We have to handle this case differently from cases where a source
701 * relation is an inheritance set. Source inheritance is expanded at
702 * the bottom of the plan tree (see allpaths.c), but target inheritance
703 * has to be expanded at the top. The reason is that for UPDATE, each
704 * target relation needs a different targetlist matching its own column
705 * set. (This is not so critical for DELETE, but for simplicity we treat
706 * inherited DELETE the same way.) Fortunately, the UPDATE/DELETE target
707 * can never be the nullable side of an outer join, so it's OK to generate
710 * parse is the querytree produced by the parser & rewriter.
711 * inheritlist is an integer list of RT indexes for the result relation set.
713 * Returns a query plan.
714 *--------------------
717 inheritance_planner(Query *parse, List *inheritlist)
719 int parentRTindex = parse->resultRelation;
720 Oid parentOID = getrelid(parentRTindex, parse->rtable);
721 List *subplans = NIL;
725 foreach(l, inheritlist)
727 int childRTindex = lfirsti(l);
728 Oid childOID = getrelid(childRTindex, parse->rtable);
732 /* Generate modified query with this rel as target */
733 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
734 parentRTindex, parentOID,
735 childRTindex, childOID);
737 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
738 subplans = lappend(subplans, subplan);
739 /* Save preprocessed tlist from first rel for use in Append */
741 tlist = subplan->targetlist;
744 /* Save the target-relations list for the executor, too */
745 parse->resultRelations = inheritlist;
747 return (Plan *) make_append(subplans, true, tlist);
750 /*--------------------
752 * Perform planning steps related to grouping, aggregation, etc.
753 * This primarily means adding top-level processing to the basic
754 * query plan produced by query_planner.
756 * parse is the querytree produced by the parser & rewriter.
757 * tuple_fraction is the fraction of tuples we expect will be retrieved
759 * tuple_fraction is interpreted as follows:
760 * < 0: determine fraction by inspection of query (normal case)
761 * 0: expect all tuples to be retrieved
762 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
763 * from the plan to be retrieved
764 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
765 * expected to be retrieved (ie, a LIMIT specification)
766 * The normal case is to pass -1, but some callers pass values >= 0 to
767 * override this routine's determination of the appropriate fraction.
769 * Returns a query plan.
770 *--------------------
773 grouping_planner(Query *parse, double tuple_fraction)
775 List *tlist = parse->targetList;
777 List *current_pathkeys;
778 List *group_pathkeys;
780 AttrNumber *groupColIdx = NULL;
782 if (parse->setOperations)
785 * Construct the plan for set operations. The result will not
786 * need any work except perhaps a top-level sort and/or LIMIT.
788 result_plan = plan_set_operations(parse);
791 * We should not need to call preprocess_targetlist, since we must
792 * be in a SELECT query node. Instead, use the targetlist
793 * returned by plan_set_operations (since this tells whether it
794 * returned any resjunk columns!), and transfer any sort key
795 * information from the original tlist.
797 Assert(parse->commandType == CMD_SELECT);
799 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
802 * Can't handle FOR UPDATE here (parser should have checked
803 * already, but let's make sure).
806 elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
809 * We set current_pathkeys NIL indicating we do not know sort
810 * order. This is correct when the top set operation is UNION
811 * ALL, since the appended-together results are unsorted even if
812 * the subplans were sorted. For other set operations we could be
813 * smarter --- room for future improvement!
815 current_pathkeys = NIL;
818 * Calculate pathkeys that represent grouping/ordering
819 * requirements (grouping should always be null, but...)
821 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
823 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
830 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
831 tlist = preprocess_targetlist(tlist,
833 parse->resultRelation,
837 * Add TID targets for rels selected FOR UPDATE (should this be
838 * done in preprocess_targetlist?). The executor uses the TID to
839 * know which rows to lock, much as for UPDATE or DELETE.
846 * We've got trouble if the FOR UPDATE appears inside
847 * grouping, since grouping renders a reference to individual
848 * tuple CTIDs invalid. This is also checked at parse time,
849 * but that's insufficient because of rule substitution, query
852 CheckSelectForUpdate(parse);
855 * Currently the executor only supports FOR UPDATE at top
858 if (PlannerQueryLevel > 1)
859 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
861 foreach(l, parse->rowMarks)
863 Index rti = lfirsti(l);
869 resname = (char *) palloc(32);
870 sprintf(resname, "ctid%u", rti);
871 resdom = makeResdom(length(tlist) + 1,
878 SelfItemPointerAttributeNumber,
883 ctid = makeTargetEntry(resdom, (Node *) var);
884 tlist = lappend(tlist, ctid);
889 * Generate appropriate target list for subplan; may be different
890 * from tlist if grouping or aggregation is needed.
892 sub_tlist = make_subplanTargetList(parse, tlist, &groupColIdx);
895 * Calculate pathkeys that represent grouping/ordering
898 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
900 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
904 * Figure out whether we need a sorted result from query_planner.
906 * If we have a GROUP BY clause, then we want a result sorted
907 * properly for grouping. Otherwise, if there is an ORDER BY
908 * clause, we want to sort by the ORDER BY clause. (Note: if we
909 * have both, and ORDER BY is a superset of GROUP BY, it would be
910 * tempting to request sort by ORDER BY --- but that might just
911 * leave us failing to exploit an available sort order at all.
912 * Needs more thought...)
914 if (parse->groupClause)
915 parse->query_pathkeys = group_pathkeys;
916 else if (parse->sortClause)
917 parse->query_pathkeys = sort_pathkeys;
919 parse->query_pathkeys = NIL;
922 * Figure out whether we expect to retrieve all the tuples that
923 * the plan can generate, or to stop early due to outside factors
924 * such as a cursor. If the caller passed a value >= 0, believe
925 * that value, else do our own examination of the query context.
927 if (tuple_fraction < 0.0)
929 /* Initial assumption is we need all the tuples */
930 tuple_fraction = 0.0;
933 * Check for retrieve-into-portal, ie DECLARE CURSOR.
935 * We have no real idea how many tuples the user will ultimately
936 * FETCH from a cursor, but it seems a good bet that he
937 * doesn't want 'em all. Optimize for 10% retrieval (you
938 * gotta better number? Should this be a SETtable parameter?)
941 tuple_fraction = 0.10;
945 * Adjust tuple_fraction if we see that we are going to apply
946 * limiting/grouping/aggregation/etc. This is not overridable by
947 * the caller, since it reflects plan actions that this routine
948 * will certainly take, not assumptions about context.
950 if (parse->limitCount != NULL)
953 * A LIMIT clause limits the absolute number of tuples
954 * returned. However, if it's not a constant LIMIT then we
955 * have to punt; for lack of a better idea, assume 10% of the
956 * plan's result is wanted.
958 double limit_fraction = 0.0;
960 if (IsA(parse->limitCount, Const))
962 Const *limitc = (Const *) parse->limitCount;
963 int32 count = DatumGetInt32(limitc->constvalue);
966 * A NULL-constant LIMIT represents "LIMIT ALL", which we
967 * treat the same as no limit (ie, expect to retrieve all
970 if (!limitc->constisnull && count > 0)
972 limit_fraction = (double) count;
973 /* We must also consider the OFFSET, if present */
974 if (parse->limitOffset != NULL)
976 if (IsA(parse->limitOffset, Const))
980 limitc = (Const *) parse->limitOffset;
981 offset = DatumGetInt32(limitc->constvalue);
982 if (!limitc->constisnull && offset > 0)
983 limit_fraction += (double) offset;
987 /* OFFSET is an expression ... punt ... */
988 limit_fraction = 0.10;
995 /* LIMIT is an expression ... punt ... */
996 limit_fraction = 0.10;
999 if (limit_fraction > 0.0)
1002 * If we have absolute limits from both caller and LIMIT,
1003 * use the smaller value; if one is fractional and the
1004 * other absolute, treat the fraction as a fraction of the
1005 * absolute value; else we can multiply the two fractions
1008 if (tuple_fraction >= 1.0)
1010 if (limit_fraction >= 1.0)
1013 tuple_fraction = Min(tuple_fraction, limit_fraction);
1017 /* caller absolute, limit fractional */
1018 tuple_fraction *= limit_fraction;
1019 if (tuple_fraction < 1.0)
1020 tuple_fraction = 1.0;
1023 else if (tuple_fraction > 0.0)
1025 if (limit_fraction >= 1.0)
1027 /* caller fractional, limit absolute */
1028 tuple_fraction *= limit_fraction;
1029 if (tuple_fraction < 1.0)
1030 tuple_fraction = 1.0;
1034 /* both fractional */
1035 tuple_fraction *= limit_fraction;
1040 /* no info from caller, just use limit */
1041 tuple_fraction = limit_fraction;
1046 if (parse->groupClause)
1049 * In GROUP BY mode, we have the little problem that we don't
1050 * really know how many input tuples will be needed to make a
1051 * group, so we can't translate an output LIMIT count into an
1052 * input count. For lack of a better idea, assume 25% of the
1053 * input data will be processed if there is any output limit.
1054 * However, if the caller gave us a fraction rather than an
1055 * absolute count, we can keep using that fraction (which
1056 * amounts to assuming that all the groups are about the same
1059 if (tuple_fraction >= 1.0)
1060 tuple_fraction = 0.25;
1063 * If both GROUP BY and ORDER BY are specified, we will need
1064 * two levels of sort --- and, therefore, certainly need to
1065 * read all the input tuples --- unless ORDER BY is a subset
1066 * of GROUP BY. (We have not yet canonicalized the pathkeys,
1067 * so must use the slower noncanonical comparison method.)
1069 if (parse->groupClause && parse->sortClause &&
1070 !noncanonical_pathkeys_contained_in(sort_pathkeys,
1072 tuple_fraction = 0.0;
1074 else if (parse->hasAggs)
1077 * Ungrouped aggregate will certainly want all the input
1080 tuple_fraction = 0.0;
1082 else if (parse->distinctClause)
1085 * SELECT DISTINCT, like GROUP, will absorb an unpredictable
1086 * number of input tuples per output tuple. Handle the same
1089 if (tuple_fraction >= 1.0)
1090 tuple_fraction = 0.25;
1093 /* Generate the basic plan for this Query */
1094 result_plan = query_planner(parse,
1099 * query_planner returns actual sort order (which is not
1100 * necessarily what we requested) in query_pathkeys.
1102 current_pathkeys = parse->query_pathkeys;
1106 * We couldn't canonicalize group_pathkeys and sort_pathkeys before
1107 * running query_planner(), so do it now.
1109 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
1110 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
1113 * If we have a GROUP BY clause, insert a group node (plus the
1114 * appropriate sort node, if necessary).
1116 if (parse->groupClause)
1123 * Decide whether how many tuples per group the Group node needs
1124 * to return. (Needs only one tuple per group if no aggregate is
1125 * present. Otherwise, need every tuple from the group to do the
1126 * aggregation.) Note tuplePerGroup is named backwards :-(
1128 tuplePerGroup = parse->hasAggs;
1131 * If there are aggregates then the Group node should just return
1132 * the same set of vars as the subplan did. If there are no aggs
1133 * then the Group node had better compute the final tlist.
1136 group_tlist = new_unsorted_tlist(result_plan->targetlist);
1138 group_tlist = tlist;
1141 * Figure out whether the path result is already ordered the way
1142 * we need it --- if so, no need for an explicit sort step.
1144 if (pathkeys_contained_in(group_pathkeys, current_pathkeys))
1146 is_sorted = true; /* no sort needed now */
1147 /* current_pathkeys remains unchanged */
1152 * We will need to do an explicit sort by the GROUP BY clause.
1153 * make_groupplan will do the work, but set current_pathkeys
1154 * to indicate the resulting order.
1157 current_pathkeys = group_pathkeys;
1160 result_plan = make_groupplan(parse,
1170 * If aggregate is present, insert the Agg node
1172 * HAVING clause, if any, becomes qual of the Agg node
1176 result_plan = (Plan *) make_agg(tlist,
1177 (List *) parse->havingQual,
1179 /* Note: Agg does not affect any existing sort order of the tuples */
1183 /* If there are no Aggs, we shouldn't have any HAVING qual anymore */
1184 Assert(parse->havingQual == NULL);
1188 * If we were not able to make the plan come out in the right order,
1189 * add an explicit sort step.
1191 if (parse->sortClause)
1193 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1194 result_plan = make_sortplan(parse, tlist, result_plan,
1199 * If there is a DISTINCT clause, add the UNIQUE node.
1201 if (parse->distinctClause)
1203 result_plan = (Plan *) make_unique(tlist, result_plan,
1204 parse->distinctClause);
1208 * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1210 if (parse->limitOffset || parse->limitCount)
1212 result_plan = (Plan *) make_limit(tlist, result_plan,
1221 * make_subplanTargetList
1222 * Generate appropriate target list when grouping is required.
1224 * When grouping_planner inserts Aggregate and/or Group plan nodes above
1225 * the result of query_planner, we typically want to pass a different
1226 * target list to query_planner than the outer plan nodes should have.
1227 * This routine generates the correct target list for the subplan.
1229 * The initial target list passed from the parser already contains entries
1230 * for all ORDER BY and GROUP BY expressions, but it will not have entries
1231 * for variables used only in HAVING clauses; so we need to add those
1232 * variables to the subplan target list. Also, if we are doing either
1233 * grouping or aggregation, we flatten all expressions except GROUP BY items
1234 * into their component variables; the other expressions will be computed by
1235 * the inserted nodes rather than by the subplan. For example,
1236 * given a query like
1237 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1238 * we want to pass this targetlist to the subplan:
1240 * where the a+b target will be used by the Sort/Group steps, and the
1241 * other targets will be used for computing the final results. (In the
1242 * above example we could theoretically suppress the a and b targets and
1243 * pass down only c,d,a+b, but it's not really worth the trouble to
1244 * eliminate simple var references from the subplan. We will avoid doing
1245 * the extra computation to recompute a+b at the outer level; see
1246 * replace_vars_with_subplan_refs() in setrefs.c.)
1248 * 'parse' is the query being processed.
1249 * 'tlist' is the query's target list.
1250 * 'groupColIdx' receives an array of column numbers for the GROUP BY
1251 * expressions (if there are any) in the subplan's target list.
1253 * The result is the targetlist to be passed to the subplan.
1257 make_subplanTargetList(Query *parse,
1259 AttrNumber **groupColIdx)
1265 *groupColIdx = NULL;
1268 * If we're not grouping or aggregating, nothing to do here;
1269 * query_planner should receive the unmodified target list.
1271 if (!parse->hasAggs && !parse->groupClause && !parse->havingQual)
1275 * Otherwise, start with a "flattened" tlist (having just the vars
1276 * mentioned in the targetlist and HAVING qual --- but not upper-
1277 * level Vars; they will be replaced by Params later on).
1279 sub_tlist = flatten_tlist(tlist);
1280 extravars = pull_var_clause(parse->havingQual, false);
1281 sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1282 freeList(extravars);
1285 * If grouping, create sub_tlist entries for all GROUP BY expressions
1286 * (GROUP BY items that are simple Vars should be in the list
1287 * already), and make an array showing where the group columns are in
1290 numCols = length(parse->groupClause);
1294 AttrNumber *grpColIdx;
1297 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1298 *groupColIdx = grpColIdx;
1300 foreach(gl, parse->groupClause)
1302 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1303 Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1304 TargetEntry *te = NULL;
1307 /* Find or make a matching sub_tlist entry */
1308 foreach(sl, sub_tlist)
1310 te = (TargetEntry *) lfirst(sl);
1311 if (equal(groupexpr, te->expr))
1316 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1317 exprType(groupexpr),
1318 exprTypmod(groupexpr),
1322 sub_tlist = lappend(sub_tlist, te);
1325 /* and save its resno */
1326 grpColIdx[keyno++] = te->resdom->resno;
1335 * Add a Group node for GROUP BY processing.
1336 * If we couldn't make the subplan produce presorted output for grouping,
1337 * first add an explicit Sort node.
1340 make_groupplan(Query *parse,
1344 AttrNumber *grpColIdx,
1348 int numCols = length(groupClause);
1353 * The Sort node always just takes a copy of the subplan's tlist
1354 * plus ordering information. (This might seem inefficient if the
1355 * subplan contains complex GROUP BY expressions, but in fact Sort
1356 * does not evaluate its targetlist --- it only outputs the same
1357 * tuples in a new order. So the expressions we might be copying
1358 * are just dummies with no extra execution cost.)
1360 List *sort_tlist = new_unsorted_tlist(subplan->targetlist);
1364 foreach(gl, groupClause)
1366 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1367 TargetEntry *te = nth(grpColIdx[keyno] - 1, sort_tlist);
1368 Resdom *resdom = te->resdom;
1371 * Check for the possibility of duplicate group-by clauses ---
1372 * the parser should have removed 'em, but the Sort executor
1373 * will get terribly confused if any get through!
1375 if (resdom->reskey == 0)
1377 /* OK, insert the ordering info needed by the executor. */
1378 resdom->reskey = ++keyno;
1379 resdom->reskeyop = grpcl->sortop;
1385 subplan = (Plan *) make_sort(parse, sort_tlist, subplan, keyno);
1388 return (Plan *) make_group(group_tlist, tuplePerGroup, numCols,
1389 grpColIdx, subplan);
1394 * Add a Sort node to implement an explicit ORDER BY clause.
1397 make_sortplan(Query *parse, List *tlist, Plan *plannode, List *sortcls)
1404 * First make a copy of the tlist so that we don't corrupt the
1407 sort_tlist = new_unsorted_tlist(tlist);
1411 SortClause *sortcl = (SortClause *) lfirst(i);
1412 TargetEntry *tle = get_sortgroupclause_tle(sortcl, sort_tlist);
1413 Resdom *resdom = tle->resdom;
1416 * Check for the possibility of duplicate order-by clauses --- the
1417 * parser should have removed 'em, but the executor will get
1418 * terribly confused if any get through!
1420 if (resdom->reskey == 0)
1422 /* OK, insert the ordering info needed by the executor. */
1423 resdom->reskey = ++keyno;
1424 resdom->reskeyop = sortcl->sortop;
1430 return (Plan *) make_sort(parse, sort_tlist, plannode, keyno);
1434 * postprocess_setop_tlist
1435 * Fix up targetlist returned by plan_set_operations().
1437 * We need to transpose sort key info from the orig_tlist into new_tlist.
1438 * NOTE: this would not be good enough if we supported resjunk sort keys
1439 * for results of set operations --- then, we'd need to project a whole
1440 * new tlist to evaluate the resjunk columns. For now, just elog if we
1441 * find any resjunk columns in orig_tlist.
1444 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1448 foreach(l, new_tlist)
1450 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1451 TargetEntry *orig_tle;
1453 /* ignore resjunk columns in setop result */
1454 if (new_tle->resdom->resjunk)
1457 Assert(orig_tlist != NIL);
1458 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1459 orig_tlist = lnext(orig_tlist);
1460 if (orig_tle->resdom->resjunk)
1461 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1462 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1463 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1464 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1466 if (orig_tlist != NIL)
1467 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");