1 /*-------------------------------------------------------------------------
4 * The query optimizer external interface.
6 * Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * src/backend/optimizer/plan/planner.c
13 *-------------------------------------------------------------------------
21 #include "access/htup_details.h"
22 #include "access/parallel.h"
23 #include "access/sysattr.h"
24 #include "access/xact.h"
25 #include "catalog/pg_constraint_fn.h"
26 #include "catalog/pg_proc.h"
27 #include "catalog/pg_type.h"
28 #include "executor/executor.h"
29 #include "executor/nodeAgg.h"
30 #include "foreign/fdwapi.h"
31 #include "miscadmin.h"
32 #include "lib/bipartite_match.h"
33 #include "nodes/makefuncs.h"
34 #include "nodes/nodeFuncs.h"
35 #ifdef OPTIMIZER_DEBUG
36 #include "nodes/print.h"
38 #include "optimizer/clauses.h"
39 #include "optimizer/cost.h"
40 #include "optimizer/pathnode.h"
41 #include "optimizer/paths.h"
42 #include "optimizer/plancat.h"
43 #include "optimizer/planmain.h"
44 #include "optimizer/planner.h"
45 #include "optimizer/prep.h"
46 #include "optimizer/subselect.h"
47 #include "optimizer/tlist.h"
48 #include "optimizer/var.h"
49 #include "parser/analyze.h"
50 #include "parser/parsetree.h"
51 #include "parser/parse_agg.h"
52 #include "rewrite/rewriteManip.h"
53 #include "storage/dsm_impl.h"
54 #include "utils/rel.h"
55 #include "utils/selfuncs.h"
56 #include "utils/lsyscache.h"
57 #include "utils/syscache.h"
61 double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
62 int force_parallel_mode = FORCE_PARALLEL_OFF;
64 /* Hook for plugins to get control in planner() */
65 planner_hook_type planner_hook = NULL;
67 /* Hook for plugins to get control when grouping_planner() plans upper rels */
68 create_upper_paths_hook_type create_upper_paths_hook = NULL;
71 /* Expression kind codes for preprocess_expression */
72 #define EXPRKIND_QUAL 0
73 #define EXPRKIND_TARGET 1
74 #define EXPRKIND_RTFUNC 2
75 #define EXPRKIND_RTFUNC_LATERAL 3
76 #define EXPRKIND_VALUES 4
77 #define EXPRKIND_VALUES_LATERAL 5
78 #define EXPRKIND_LIMIT 6
79 #define EXPRKIND_APPINFO 7
80 #define EXPRKIND_PHV 8
81 #define EXPRKIND_TABLESAMPLE 9
82 #define EXPRKIND_ARBITER_ELEM 10
84 /* Passthrough data for standard_qp_callback */
87 List *tlist; /* preprocessed query targetlist */
88 List *activeWindows; /* active windows, if any */
89 List *groupClause; /* overrides parse->groupClause */
93 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
94 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
95 static void inheritance_planner(PlannerInfo *root);
96 static void grouping_planner(PlannerInfo *root, bool inheritance_update,
97 double tuple_fraction);
98 static void preprocess_rowmarks(PlannerInfo *root);
99 static double preprocess_limit(PlannerInfo *root,
100 double tuple_fraction,
101 int64 *offset_est, int64 *count_est);
102 static bool limit_needed(Query *parse);
103 static void remove_useless_groupby_columns(PlannerInfo *root);
104 static List *preprocess_groupclause(PlannerInfo *root, List *force);
105 static List *extract_rollup_sets(List *groupingSets);
106 static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
107 static void standard_qp_callback(PlannerInfo *root, void *extra);
108 static double get_number_of_groups(PlannerInfo *root,
111 List *rollup_groupclauses);
112 static Size estimate_hashagg_tablesize(Path *path,
113 const AggClauseCosts *agg_costs,
115 static RelOptInfo *create_grouping_paths(PlannerInfo *root,
116 RelOptInfo *input_rel,
118 const AggClauseCosts *agg_costs,
120 List *rollup_groupclauses);
121 static RelOptInfo *create_window_paths(PlannerInfo *root,
122 RelOptInfo *input_rel,
123 PathTarget *input_target,
124 PathTarget *output_target,
126 WindowFuncLists *wflists,
127 List *activeWindows);
128 static void create_one_window_path(PlannerInfo *root,
129 RelOptInfo *window_rel,
131 PathTarget *input_target,
132 PathTarget *output_target,
134 WindowFuncLists *wflists,
135 List *activeWindows);
136 static RelOptInfo *create_distinct_paths(PlannerInfo *root,
137 RelOptInfo *input_rel);
138 static RelOptInfo *create_ordered_paths(PlannerInfo *root,
139 RelOptInfo *input_rel,
141 double limit_tuples);
142 static PathTarget *make_group_input_target(PlannerInfo *root,
143 PathTarget *final_target);
144 static PathTarget *make_partial_grouping_target(PlannerInfo *root,
145 PathTarget *grouping_target);
146 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
147 static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
148 static PathTarget *make_window_input_target(PlannerInfo *root,
149 PathTarget *final_target,
150 List *activeWindows);
151 static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
153 static PathTarget *make_sort_input_target(PlannerInfo *root,
154 PathTarget *final_target,
155 bool *have_postponed_srfs);
158 /*****************************************************************************
160 * Query optimizer entry point
162 * To support loadable plugins that monitor or modify planner behavior,
163 * we provide a hook variable that lets a plugin get control before and
164 * after the standard planning process. The plugin would normally call
165 * standard_planner().
167 * Note to plugin authors: standard_planner() scribbles on its Query input,
168 * so you'd better copy that data structure if you want to plan more than once.
170 *****************************************************************************/
172 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
177 result = (*planner_hook) (parse, cursorOptions, boundParams);
179 result = standard_planner(parse, cursorOptions, boundParams);
184 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
188 double tuple_fraction;
190 RelOptInfo *final_rel;
196 /* Cursor options may come from caller or from DECLARE CURSOR stmt */
197 if (parse->utilityStmt &&
198 IsA(parse->utilityStmt, DeclareCursorStmt))
199 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
202 * Set up global state for this planner invocation. This data is needed
203 * across all levels of sub-Query that might exist in the given command,
204 * so we keep it in a separate struct that's linked to by each per-Query
207 glob = makeNode(PlannerGlobal);
209 glob->boundParams = boundParams;
210 glob->subplans = NIL;
211 glob->subroots = NIL;
212 glob->rewindPlanIDs = NULL;
213 glob->finalrtable = NIL;
214 glob->finalrowmarks = NIL;
215 glob->resultRelations = NIL;
216 glob->relationOids = NIL;
217 glob->invalItems = NIL;
218 glob->nParamExec = 0;
220 glob->lastRowMarkId = 0;
221 glob->lastPlanNodeId = 0;
222 glob->transientPlan = false;
223 glob->dependsOnRole = false;
226 * Assess whether it's feasible to use parallel mode for this query. We
227 * can't do this in a standalone backend, or if the command will try to
228 * modify any data, or if this is a cursor operation, or if GUCs are set
229 * to values that don't permit parallelism, or if parallel-unsafe
230 * functions are present in the query tree.
232 * For now, we don't try to use parallel mode if we're running inside a
233 * parallel worker. We might eventually be able to relax this
234 * restriction, but for now it seems best not to have parallel workers
235 * trying to create their own parallel workers.
237 * We can't use parallelism in serializable mode because the predicate
238 * locking code is not parallel-aware. It's not catastrophic if someone
239 * tries to run a parallel plan in serializable mode; it just won't get
240 * any workers and will run serially. But it seems like a good heuristic
241 * to assume that the same serialization level will be in effect at plan
242 * time and execution time, so don't generate a parallel plan if we're in
245 if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
247 dynamic_shared_memory_type != DSM_IMPL_NONE &&
248 parse->commandType == CMD_SELECT &&
249 parse->utilityStmt == NULL &&
250 !parse->hasModifyingCTE &&
251 max_parallel_workers_per_gather > 0 &&
252 !IsParallelWorker() &&
253 !IsolationIsSerializable())
255 /* all the cheap tests pass, so scan the query tree */
256 glob->maxParallelHazard = max_parallel_hazard(parse);
257 glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
261 /* skip the query tree scan, just assume it's unsafe */
262 glob->maxParallelHazard = PROPARALLEL_UNSAFE;
263 glob->parallelModeOK = false;
267 * glob->parallelModeNeeded should tell us whether it's necessary to
268 * impose the parallel mode restrictions, but we don't actually want to
269 * impose them unless we choose a parallel plan, so it is normally set
270 * only if a parallel plan is chosen (see create_gather_plan). That way,
271 * people who mislabel their functions but don't use parallelism anyway
272 * aren't harmed. But when force_parallel_mode is set, we enable the
273 * restrictions whenever possible for testing purposes.
275 glob->parallelModeNeeded = glob->parallelModeOK &&
276 (force_parallel_mode != FORCE_PARALLEL_OFF);
278 /* Determine what fraction of the plan is likely to be scanned */
279 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
282 * We have no real idea how many tuples the user will ultimately FETCH
283 * from a cursor, but it is often the case that he doesn't want 'em
284 * all, or would prefer a fast-start plan anyway so that he can
285 * process some of the tuples sooner. Use a GUC parameter to decide
286 * what fraction to optimize for.
288 tuple_fraction = cursor_tuple_fraction;
291 * We document cursor_tuple_fraction as simply being a fraction, which
292 * means the edge cases 0 and 1 have to be treated specially here. We
293 * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
295 if (tuple_fraction >= 1.0)
296 tuple_fraction = 0.0;
297 else if (tuple_fraction <= 0.0)
298 tuple_fraction = 1e-10;
302 /* Default assumption is we need all the tuples */
303 tuple_fraction = 0.0;
306 /* primary planning entry point (may recurse for subqueries) */
307 root = subquery_planner(glob, parse, NULL,
308 false, tuple_fraction);
310 /* Select best Path and turn it into a Plan */
311 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
312 best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
314 top_plan = create_plan(root, best_path);
317 * If creating a plan for a scrollable cursor, make sure it can run
318 * backwards on demand. Add a Material node at the top at need.
320 if (cursorOptions & CURSOR_OPT_SCROLL)
322 if (!ExecSupportsBackwardScan(top_plan))
324 Plan *sub_plan = top_plan;
326 top_plan = materialize_finished_plan(sub_plan);
329 * XXX horrid kluge: if there are any initPlans attached to the
330 * formerly-top plan node, move them up to the Material node. This
331 * prevents failure in SS_finalize_plan, which see for comments.
332 * We don't bother adjusting the sub_plan's cost estimate for
335 top_plan->initPlan = sub_plan->initPlan;
336 sub_plan->initPlan = NIL;
341 * Optionally add a Gather node for testing purposes, provided this is
342 * actually a safe thing to do. (Note: we assume adding a Material node
343 * above did not change the parallel safety of the plan, so we can still
344 * rely on best_path->parallel_safe. However, that flag doesn't account
345 * for subplans, which we are unable to transmit to workers presently.)
347 if (force_parallel_mode != FORCE_PARALLEL_OFF &&
348 best_path->parallel_safe &&
349 glob->subplans == NIL)
351 Gather *gather = makeNode(Gather);
353 gather->plan.targetlist = top_plan->targetlist;
354 gather->plan.qual = NIL;
355 gather->plan.lefttree = top_plan;
356 gather->plan.righttree = NULL;
357 gather->num_workers = 1;
358 gather->single_copy = true;
359 gather->invisible = (force_parallel_mode == FORCE_PARALLEL_REGRESS);
362 * Ideally we'd use cost_gather here, but setting up dummy path data
363 * to satisfy it doesn't seem much cleaner than knowing what it does.
365 gather->plan.startup_cost = top_plan->startup_cost +
367 gather->plan.total_cost = top_plan->total_cost +
368 parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
369 gather->plan.plan_rows = top_plan->plan_rows;
370 gather->plan.plan_width = top_plan->plan_width;
371 gather->plan.parallel_aware = false;
373 /* use parallel mode for parallel plans. */
374 root->glob->parallelModeNeeded = true;
376 top_plan = &gather->plan;
380 * If any Params were generated, run through the plan tree and compute
381 * each plan node's extParam/allParam sets. Ideally we'd merge this into
382 * set_plan_references' tree traversal, but for now it has to be separate
383 * because we need to visit subplans before not after main plan.
385 if (glob->nParamExec > 0)
387 Assert(list_length(glob->subplans) == list_length(glob->subroots));
388 forboth(lp, glob->subplans, lr, glob->subroots)
390 Plan *subplan = (Plan *) lfirst(lp);
391 PlannerInfo *subroot = (PlannerInfo *) lfirst(lr);
393 SS_finalize_plan(subroot, subplan);
395 SS_finalize_plan(root, top_plan);
398 /* final cleanup of the plan */
399 Assert(glob->finalrtable == NIL);
400 Assert(glob->finalrowmarks == NIL);
401 Assert(glob->resultRelations == NIL);
402 top_plan = set_plan_references(root, top_plan);
403 /* ... and the subplans (both regular subplans and initplans) */
404 Assert(list_length(glob->subplans) == list_length(glob->subroots));
405 forboth(lp, glob->subplans, lr, glob->subroots)
407 Plan *subplan = (Plan *) lfirst(lp);
408 PlannerInfo *subroot = (PlannerInfo *) lfirst(lr);
410 lfirst(lp) = set_plan_references(subroot, subplan);
413 /* build the PlannedStmt result */
414 result = makeNode(PlannedStmt);
416 result->commandType = parse->commandType;
417 result->queryId = parse->queryId;
418 result->hasReturning = (parse->returningList != NIL);
419 result->hasModifyingCTE = parse->hasModifyingCTE;
420 result->canSetTag = parse->canSetTag;
421 result->transientPlan = glob->transientPlan;
422 result->dependsOnRole = glob->dependsOnRole;
423 result->parallelModeNeeded = glob->parallelModeNeeded;
424 result->planTree = top_plan;
425 result->rtable = glob->finalrtable;
426 result->resultRelations = glob->resultRelations;
427 result->utilityStmt = parse->utilityStmt;
428 result->subplans = glob->subplans;
429 result->rewindPlanIDs = glob->rewindPlanIDs;
430 result->rowMarks = glob->finalrowmarks;
431 result->relationOids = glob->relationOids;
432 result->invalItems = glob->invalItems;
433 result->nParamExec = glob->nParamExec;
439 /*--------------------
441 * Invokes the planner on a subquery. We recurse to here for each
442 * sub-SELECT found in the query tree.
444 * glob is the global state for the current planner run.
445 * parse is the querytree produced by the parser & rewriter.
446 * parent_root is the immediate parent Query's info (NULL at the top level).
447 * hasRecursion is true if this is a recursive WITH query.
448 * tuple_fraction is the fraction of tuples we expect will be retrieved.
449 * tuple_fraction is interpreted as explained for grouping_planner, below.
451 * Basically, this routine does the stuff that should only be done once
452 * per Query object. It then calls grouping_planner. At one time,
453 * grouping_planner could be invoked recursively on the same Query object;
454 * that's not currently true, but we keep the separation between the two
455 * routines anyway, in case we need it again someday.
457 * subquery_planner will be called recursively to handle sub-Query nodes
458 * found within the query's expressions and rangetable.
460 * Returns the PlannerInfo struct ("root") that contains all data generated
461 * while planning the subquery. In particular, the Path(s) attached to
462 * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
463 * cheapest way(s) to implement the query. The top level will select the
464 * best Path and pass it through createplan.c to produce a finished Plan.
465 *--------------------
468 subquery_planner(PlannerGlobal *glob, Query *parse,
469 PlannerInfo *parent_root,
470 bool hasRecursion, double tuple_fraction)
473 List *newWithCheckOptions;
476 RelOptInfo *final_rel;
479 /* Create a PlannerInfo data structure for this subquery */
480 root = makeNode(PlannerInfo);
483 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
484 root->parent_root = parent_root;
485 root->plan_params = NIL;
486 root->outer_params = NULL;
487 root->planner_cxt = CurrentMemoryContext;
488 root->init_plans = NIL;
489 root->cte_plan_ids = NIL;
490 root->multiexpr_params = NIL;
491 root->eq_classes = NIL;
492 root->append_rel_list = NIL;
493 root->rowMarks = NIL;
494 memset(root->upper_rels, 0, sizeof(root->upper_rels));
495 memset(root->upper_targets, 0, sizeof(root->upper_targets));
496 root->processed_tlist = NIL;
497 root->grouping_map = NULL;
498 root->minmax_aggs = NIL;
499 root->hasInheritedTarget = false;
500 root->hasRecursion = hasRecursion;
502 root->wt_param_id = SS_assign_special_param(root);
504 root->wt_param_id = -1;
505 root->non_recursive_path = NULL;
508 * If there is a WITH list, process each WITH query and build an initplan
509 * SubPlan structure for it.
512 SS_process_ctes(root);
515 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
516 * to transform them into joins. Note that this step does not descend
517 * into subqueries; if we pull up any subqueries below, their SubLinks are
518 * processed just before pulling them up.
520 if (parse->hasSubLinks)
521 pull_up_sublinks(root);
524 * Scan the rangetable for set-returning functions, and inline them if
525 * possible (producing subqueries that might get pulled up next).
526 * Recursion issues here are handled in the same way as for SubLinks.
528 inline_set_returning_functions(root);
531 * Check to see if any subqueries in the jointree can be merged into this
534 pull_up_subqueries(root);
537 * If this is a simple UNION ALL query, flatten it into an appendrel. We
538 * do this now because it requires applying pull_up_subqueries to the leaf
539 * queries of the UNION ALL, which weren't touched above because they
540 * weren't referenced by the jointree (they will be after we do this).
542 if (parse->setOperations)
543 flatten_simple_union_all(root);
546 * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
547 * avoid the expense of doing flatten_join_alias_vars(). Also check for
548 * outer joins --- if none, we can skip reduce_outer_joins(). And check
549 * for LATERAL RTEs, too. This must be done after we have done
550 * pull_up_subqueries(), of course.
552 root->hasJoinRTEs = false;
553 root->hasLateralRTEs = false;
554 hasOuterJoins = false;
555 foreach(l, parse->rtable)
557 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
559 if (rte->rtekind == RTE_JOIN)
561 root->hasJoinRTEs = true;
562 if (IS_OUTER_JOIN(rte->jointype))
563 hasOuterJoins = true;
566 root->hasLateralRTEs = true;
570 * Preprocess RowMark information. We need to do this after subquery
571 * pullup (so that all non-inherited RTEs are present) and before
572 * inheritance expansion (so that the info is available for
573 * expand_inherited_tables to examine and modify).
575 preprocess_rowmarks(root);
578 * Expand any rangetable entries that are inheritance sets into "append
579 * relations". This can add entries to the rangetable, but they must be
580 * plain base relations not joins, so it's OK (and marginally more
581 * efficient) to do it after checking for join RTEs. We must do it after
582 * pulling up subqueries, else we'd fail to handle inherited tables in
585 expand_inherited_tables(root);
588 * Set hasHavingQual to remember if HAVING clause is present. Needed
589 * because preprocess_expression will reduce a constant-true condition to
590 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
592 root->hasHavingQual = (parse->havingQual != NULL);
594 /* Clear this flag; might get set in distribute_qual_to_rels */
595 root->hasPseudoConstantQuals = false;
598 * Do expression preprocessing on targetlist and quals, as well as other
599 * random expressions in the querytree. Note that we do not need to
600 * handle sort/group expressions explicitly, because they are actually
601 * part of the targetlist.
603 parse->targetList = (List *)
604 preprocess_expression(root, (Node *) parse->targetList,
607 /* Constant-folding might have removed all set-returning functions */
608 if (parse->hasTargetSRFs)
609 parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
611 newWithCheckOptions = NIL;
612 foreach(l, parse->withCheckOptions)
614 WithCheckOption *wco = (WithCheckOption *) lfirst(l);
616 wco->qual = preprocess_expression(root, wco->qual,
618 if (wco->qual != NULL)
619 newWithCheckOptions = lappend(newWithCheckOptions, wco);
621 parse->withCheckOptions = newWithCheckOptions;
623 parse->returningList = (List *)
624 preprocess_expression(root, (Node *) parse->returningList,
627 preprocess_qual_conditions(root, (Node *) parse->jointree);
629 parse->havingQual = preprocess_expression(root, parse->havingQual,
632 foreach(l, parse->windowClause)
634 WindowClause *wc = (WindowClause *) lfirst(l);
636 /* partitionClause/orderClause are sort/group expressions */
637 wc->startOffset = preprocess_expression(root, wc->startOffset,
639 wc->endOffset = preprocess_expression(root, wc->endOffset,
643 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
645 parse->limitCount = preprocess_expression(root, parse->limitCount,
648 if (parse->onConflict)
650 parse->onConflict->arbiterElems = (List *)
651 preprocess_expression(root,
652 (Node *) parse->onConflict->arbiterElems,
653 EXPRKIND_ARBITER_ELEM);
654 parse->onConflict->arbiterWhere =
655 preprocess_expression(root,
656 parse->onConflict->arbiterWhere,
658 parse->onConflict->onConflictSet = (List *)
659 preprocess_expression(root,
660 (Node *) parse->onConflict->onConflictSet,
662 parse->onConflict->onConflictWhere =
663 preprocess_expression(root,
664 parse->onConflict->onConflictWhere,
666 /* exclRelTlist contains only Vars, so no preprocessing needed */
669 root->append_rel_list = (List *)
670 preprocess_expression(root, (Node *) root->append_rel_list,
673 /* Also need to preprocess expressions within RTEs */
674 foreach(l, parse->rtable)
676 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
679 if (rte->rtekind == RTE_RELATION)
681 if (rte->tablesample)
682 rte->tablesample = (TableSampleClause *)
683 preprocess_expression(root,
684 (Node *) rte->tablesample,
685 EXPRKIND_TABLESAMPLE);
687 else if (rte->rtekind == RTE_SUBQUERY)
690 * We don't want to do all preprocessing yet on the subquery's
691 * expressions, since that will happen when we plan it. But if it
692 * contains any join aliases of our level, those have to get
693 * expanded now, because planning of the subquery won't do it.
694 * That's only possible if the subquery is LATERAL.
696 if (rte->lateral && root->hasJoinRTEs)
697 rte->subquery = (Query *)
698 flatten_join_alias_vars(root, (Node *) rte->subquery);
700 else if (rte->rtekind == RTE_FUNCTION)
702 /* Preprocess the function expression(s) fully */
703 kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
704 rte->functions = (List *) preprocess_expression(root, (Node *) rte->functions, kind);
706 else if (rte->rtekind == RTE_VALUES)
708 /* Preprocess the values lists fully */
709 kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
710 rte->values_lists = (List *)
711 preprocess_expression(root, (Node *) rte->values_lists, kind);
716 * In some cases we may want to transfer a HAVING clause into WHERE. We
717 * cannot do so if the HAVING clause contains aggregates (obviously) or
718 * volatile functions (since a HAVING clause is supposed to be executed
719 * only once per group). We also can't do this if there are any nonempty
720 * grouping sets; moving such a clause into WHERE would potentially change
721 * the results, if any referenced column isn't present in all the grouping
722 * sets. (If there are only empty grouping sets, then the HAVING clause
723 * must be degenerate as discussed below.)
725 * Also, it may be that the clause is so expensive to execute that we're
726 * better off doing it only once per group, despite the loss of
727 * selectivity. This is hard to estimate short of doing the entire
728 * planning process twice, so we use a heuristic: clauses containing
729 * subplans are left in HAVING. Otherwise, we move or copy the HAVING
730 * clause into WHERE, in hopes of eliminating tuples before aggregation
733 * If the query has explicit grouping then we can simply move such a
734 * clause into WHERE; any group that fails the clause will not be in the
735 * output because none of its tuples will reach the grouping or
736 * aggregation stage. Otherwise we must have a degenerate (variable-free)
737 * HAVING clause, which we put in WHERE so that query_planner() can use it
738 * in a gating Result node, but also keep in HAVING to ensure that we
739 * don't emit a bogus aggregated row. (This could be done better, but it
740 * seems not worth optimizing.)
742 * Note that both havingQual and parse->jointree->quals are in
743 * implicitly-ANDed-list form at this point, even though they are declared
747 foreach(l, (List *) parse->havingQual)
749 Node *havingclause = (Node *) lfirst(l);
751 if ((parse->groupClause && parse->groupingSets) ||
752 contain_agg_clause(havingclause) ||
753 contain_volatile_functions(havingclause) ||
754 contain_subplans(havingclause))
756 /* keep it in HAVING */
757 newHaving = lappend(newHaving, havingclause);
759 else if (parse->groupClause && !parse->groupingSets)
761 /* move it to WHERE */
762 parse->jointree->quals = (Node *)
763 lappend((List *) parse->jointree->quals, havingclause);
767 /* put a copy in WHERE, keep it in HAVING */
768 parse->jointree->quals = (Node *)
769 lappend((List *) parse->jointree->quals,
770 copyObject(havingclause));
771 newHaving = lappend(newHaving, havingclause);
774 parse->havingQual = (Node *) newHaving;
776 /* Remove any redundant GROUP BY columns */
777 remove_useless_groupby_columns(root);
780 * If we have any outer joins, try to reduce them to plain inner joins.
781 * This step is most easily done after we've done expression
785 reduce_outer_joins(root);
788 * Do the main planning. If we have an inherited target relation, that
789 * needs special processing, else go straight to grouping_planner.
791 if (parse->resultRelation &&
792 rt_fetch(parse->resultRelation, parse->rtable)->inh)
793 inheritance_planner(root);
795 grouping_planner(root, false, tuple_fraction);
798 * Capture the set of outer-level param IDs we have access to, for use in
799 * extParam/allParam calculations later.
801 SS_identify_outer_params(root);
804 * If any initPlans were created in this query level, increment the
805 * surviving Paths' costs to account for them. They won't actually get
806 * attached to the plan tree till create_plan() runs, but we want to be
807 * sure their costs are included now.
809 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
810 SS_charge_for_initplans(root, final_rel);
813 * Make sure we've identified the cheapest Path for the final rel. (By
814 * doing this here not in grouping_planner, we include initPlan costs in
815 * the decision, though it's unlikely that will change anything.)
817 set_cheapest(final_rel);
823 * preprocess_expression
824 * Do subquery_planner's preprocessing work for an expression,
825 * which can be a targetlist, a WHERE clause (including JOIN/ON
826 * conditions), a HAVING clause, or a few other things.
829 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
832 * Fall out quickly if expression is empty. This occurs often enough to
833 * be worth checking. Note that null->null is the correct conversion for
834 * implicit-AND result format, too.
840 * If the query has any join RTEs, replace join alias variables with
841 * base-relation variables. We must do this before sublink processing,
842 * else sublinks expanded out from join aliases would not get processed.
843 * We can skip it in non-lateral RTE functions, VALUES lists, and
844 * TABLESAMPLE clauses, however, since they can't contain any Vars of the
845 * current query level.
847 if (root->hasJoinRTEs &&
848 !(kind == EXPRKIND_RTFUNC ||
849 kind == EXPRKIND_VALUES ||
850 kind == EXPRKIND_TABLESAMPLE))
851 expr = flatten_join_alias_vars(root, expr);
854 * Simplify constant expressions.
856 * Note: an essential effect of this is to convert named-argument function
857 * calls to positional notation and insert the current actual values of
858 * any default arguments for functions. To ensure that happens, we *must*
859 * process all expressions here. Previous PG versions sometimes skipped
860 * const-simplification if it didn't seem worth the trouble, but we can't
863 * Note: this also flattens nested AND and OR expressions into N-argument
864 * form. All processing of a qual expression after this point must be
865 * careful to maintain AND/OR flatness --- that is, do not generate a tree
866 * with AND directly under AND, nor OR directly under OR.
868 expr = eval_const_expressions(root, expr);
871 * If it's a qual or havingQual, canonicalize it.
873 if (kind == EXPRKIND_QUAL)
875 expr = (Node *) canonicalize_qual((Expr *) expr);
877 #ifdef OPTIMIZER_DEBUG
878 printf("After canonicalize_qual()\n");
883 /* Expand SubLinks to SubPlans */
884 if (root->parse->hasSubLinks)
885 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
888 * XXX do not insert anything here unless you have grokked the comments in
889 * SS_replace_correlation_vars ...
892 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
893 if (root->query_level > 1)
894 expr = SS_replace_correlation_vars(root, expr);
897 * If it's a qual or havingQual, convert it to implicit-AND format. (We
898 * don't want to do this before eval_const_expressions, since the latter
899 * would be unable to simplify a top-level AND correctly. Also,
900 * SS_process_sublinks expects explicit-AND format.)
902 if (kind == EXPRKIND_QUAL)
903 expr = (Node *) make_ands_implicit((Expr *) expr);
909 * preprocess_qual_conditions
910 * Recursively scan the query's jointree and do subquery_planner's
911 * preprocessing work on each qual condition found therein.
914 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
918 if (IsA(jtnode, RangeTblRef))
920 /* nothing to do here */
922 else if (IsA(jtnode, FromExpr))
924 FromExpr *f = (FromExpr *) jtnode;
927 foreach(l, f->fromlist)
928 preprocess_qual_conditions(root, lfirst(l));
930 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
932 else if (IsA(jtnode, JoinExpr))
934 JoinExpr *j = (JoinExpr *) jtnode;
936 preprocess_qual_conditions(root, j->larg);
937 preprocess_qual_conditions(root, j->rarg);
939 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
942 elog(ERROR, "unrecognized node type: %d",
943 (int) nodeTag(jtnode));
947 * preprocess_phv_expression
948 * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
950 * If a LATERAL subquery references an output of another subquery, and that
951 * output must be wrapped in a PlaceHolderVar because of an intermediate outer
952 * join, then we'll push the PlaceHolderVar expression down into the subquery
953 * and later pull it back up during find_lateral_references, which runs after
954 * subquery_planner has preprocessed all the expressions that were in the
955 * current query level to start with. So we need to preprocess it then.
958 preprocess_phv_expression(PlannerInfo *root, Expr *expr)
960 return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
964 * inheritance_planner
965 * Generate Paths in the case where the result relation is an
968 * We have to handle this case differently from cases where a source relation
969 * is an inheritance set. Source inheritance is expanded at the bottom of the
970 * plan tree (see allpaths.c), but target inheritance has to be expanded at
971 * the top. The reason is that for UPDATE, each target relation needs a
972 * different targetlist matching its own column set. Fortunately,
973 * the UPDATE/DELETE target can never be the nullable side of an outer join,
974 * so it's OK to generate the plan this way.
976 * Returns nothing; the useful output is in the Paths we attach to
977 * the (UPPERREL_FINAL, NULL) upperrel stored in *root.
979 * Note that we have not done set_cheapest() on the final rel; it's convenient
980 * to leave this to the caller.
983 inheritance_planner(PlannerInfo *root)
985 Query *parse = root->parse;
986 int parentRTindex = parse->resultRelation;
987 Bitmapset *resultRTindexes;
988 Bitmapset *subqueryRTindexes;
989 Bitmapset *modifiableARIindexes;
990 int nominalRelation = -1;
991 List *final_rtable = NIL;
992 int save_rel_array_size = 0;
993 RelOptInfo **save_rel_array = NULL;
994 List *subpaths = NIL;
995 List *subroots = NIL;
996 List *resultRelations = NIL;
997 List *withCheckOptionLists = NIL;
998 List *returningLists = NIL;
1000 RelOptInfo *final_rel;
1004 Assert(parse->commandType != CMD_INSERT);
1007 * We generate a modified instance of the original Query for each target
1008 * relation, plan that, and put all the plans into a list that will be
1009 * controlled by a single ModifyTable node. All the instances share the
1010 * same rangetable, but each instance must have its own set of subquery
1011 * RTEs within the finished rangetable because (1) they are likely to get
1012 * scribbled on during planning, and (2) it's not inconceivable that
1013 * subqueries could get planned differently in different cases. We need
1014 * not create duplicate copies of other RTE kinds, in particular not the
1015 * target relations, because they don't have either of those issues. Not
1016 * having to duplicate the target relations is important because doing so
1017 * (1) would result in a rangetable of length O(N^2) for N targets, with
1018 * at least O(N^3) work expended here; and (2) would greatly complicate
1019 * management of the rowMarks list.
1021 * Note that any RTEs with security barrier quals will be turned into
1022 * subqueries during planning, and so we must create copies of them too,
1023 * except where they are target relations, which will each only be used in
1026 * To begin with, we'll need a bitmapset of the target relation relids.
1028 resultRTindexes = bms_make_singleton(parentRTindex);
1029 foreach(lc, root->append_rel_list)
1031 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(lc);
1033 if (appinfo->parent_relid == parentRTindex)
1034 resultRTindexes = bms_add_member(resultRTindexes,
1035 appinfo->child_relid);
1039 * Now, generate a bitmapset of the relids of the subquery RTEs, including
1040 * security-barrier RTEs that will become subqueries, as just explained.
1042 subqueryRTindexes = NULL;
1044 foreach(lc, parse->rtable)
1046 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lc);
1048 if (rte->rtekind == RTE_SUBQUERY ||
1049 (rte->securityQuals != NIL &&
1050 !bms_is_member(rti, resultRTindexes)))
1051 subqueryRTindexes = bms_add_member(subqueryRTindexes, rti);
1056 * Next, we want to identify which AppendRelInfo items contain references
1057 * to any of the aforesaid subquery RTEs. These items will need to be
1058 * copied and modified to adjust their subquery references; whereas the
1059 * other ones need not be touched. It's worth being tense over this
1060 * because we can usually avoid processing most of the AppendRelInfo
1061 * items, thereby saving O(N^2) space and time when the target is a large
1062 * inheritance tree. We can identify AppendRelInfo items by their
1063 * child_relid, since that should be unique within the list.
1065 modifiableARIindexes = NULL;
1066 if (subqueryRTindexes != NULL)
1068 foreach(lc, root->append_rel_list)
1070 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(lc);
1072 if (bms_is_member(appinfo->parent_relid, subqueryRTindexes) ||
1073 bms_is_member(appinfo->child_relid, subqueryRTindexes) ||
1074 bms_overlap(pull_varnos((Node *) appinfo->translated_vars),
1076 modifiableARIindexes = bms_add_member(modifiableARIindexes,
1077 appinfo->child_relid);
1082 * And now we can get on with generating a plan for each child table.
1084 foreach(lc, root->append_rel_list)
1086 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(lc);
1087 PlannerInfo *subroot;
1088 RelOptInfo *sub_final_rel;
1091 /* append_rel_list contains all append rels; ignore others */
1092 if (appinfo->parent_relid != parentRTindex)
1096 * We need a working copy of the PlannerInfo so that we can control
1097 * propagation of information back to the main copy.
1099 subroot = makeNode(PlannerInfo);
1100 memcpy(subroot, root, sizeof(PlannerInfo));
1103 * Generate modified query with this rel as target. We first apply
1104 * adjust_appendrel_attrs, which copies the Query and changes
1105 * references to the parent RTE to refer to the current child RTE,
1106 * then fool around with subquery RTEs.
1108 subroot->parse = (Query *)
1109 adjust_appendrel_attrs(root,
1114 * The rowMarks list might contain references to subquery RTEs, so
1115 * make a copy that we can apply ChangeVarNodes to. (Fortunately, the
1116 * executor doesn't need to see the modified copies --- we can just
1117 * pass it the original rowMarks list.)
1119 subroot->rowMarks = (List *) copyObject(root->rowMarks);
1122 * The append_rel_list likewise might contain references to subquery
1123 * RTEs (if any subqueries were flattenable UNION ALLs). So prepare
1124 * to apply ChangeVarNodes to that, too. As explained above, we only
1125 * want to copy items that actually contain such references; the rest
1126 * can just get linked into the subroot's append_rel_list.
1128 * If we know there are no such references, we can just use the outer
1129 * append_rel_list unmodified.
1131 if (modifiableARIindexes != NULL)
1135 subroot->append_rel_list = NIL;
1136 foreach(lc2, root->append_rel_list)
1138 AppendRelInfo *appinfo2 = (AppendRelInfo *) lfirst(lc2);
1140 if (bms_is_member(appinfo2->child_relid, modifiableARIindexes))
1141 appinfo2 = (AppendRelInfo *) copyObject(appinfo2);
1143 subroot->append_rel_list = lappend(subroot->append_rel_list,
1149 * Add placeholders to the child Query's rangetable list to fill the
1150 * RT indexes already reserved for subqueries in previous children.
1151 * These won't be referenced, so there's no need to make them very
1154 while (list_length(subroot->parse->rtable) < list_length(final_rtable))
1155 subroot->parse->rtable = lappend(subroot->parse->rtable,
1156 makeNode(RangeTblEntry));
1159 * If this isn't the first child Query, generate duplicates of all
1160 * subquery (or subquery-to-be) RTEs, and adjust Var numbering to
1161 * reference the duplicates. To simplify the loop logic, we scan the
1162 * original rtable not the copy just made by adjust_appendrel_attrs;
1163 * that should be OK since subquery RTEs couldn't contain any
1164 * references to the target rel.
1166 if (final_rtable != NIL && subqueryRTindexes != NULL)
1171 foreach(lr, parse->rtable)
1173 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lr);
1175 if (bms_is_member(rti, subqueryRTindexes))
1180 * The RTE can't contain any references to its own RT
1181 * index, except in the security barrier quals, so we can
1182 * save a few cycles by applying ChangeVarNodes before we
1183 * append the RTE to the rangetable.
1185 newrti = list_length(subroot->parse->rtable) + 1;
1186 ChangeVarNodes((Node *) subroot->parse, rti, newrti, 0);
1187 ChangeVarNodes((Node *) subroot->rowMarks, rti, newrti, 0);
1188 /* Skip processing unchanging parts of append_rel_list */
1189 if (modifiableARIindexes != NULL)
1193 foreach(lc2, subroot->append_rel_list)
1195 AppendRelInfo *appinfo2 = (AppendRelInfo *) lfirst(lc2);
1197 if (bms_is_member(appinfo2->child_relid,
1198 modifiableARIindexes))
1199 ChangeVarNodes((Node *) appinfo2, rti, newrti, 0);
1202 rte = copyObject(rte);
1203 ChangeVarNodes((Node *) rte->securityQuals, rti, newrti, 0);
1204 subroot->parse->rtable = lappend(subroot->parse->rtable,
1211 /* There shouldn't be any OJ info to translate, as yet */
1212 Assert(subroot->join_info_list == NIL);
1213 /* and we haven't created PlaceHolderInfos, either */
1214 Assert(subroot->placeholder_list == NIL);
1215 /* hack to mark target relation as an inheritance partition */
1216 subroot->hasInheritedTarget = true;
1218 /* Generate Path(s) for accessing this result relation */
1219 grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1222 * Planning may have modified the query result relation (if there were
1223 * security barrier quals on the result RTE).
1225 appinfo->child_relid = subroot->parse->resultRelation;
1228 * We'll use the first child relation (even if it's excluded) as the
1229 * nominal target relation of the ModifyTable node. Because of the
1230 * way expand_inherited_rtentry works, this should always be the RTE
1231 * representing the parent table in its role as a simple member of the
1232 * inheritance set. (It would be logically cleaner to use the
1233 * inheritance parent RTE as the nominal target; but since that RTE
1234 * will not be otherwise referenced in the plan, doing so would give
1235 * rise to confusing use of multiple aliases in EXPLAIN output for
1236 * what the user will think is the "same" table.)
1238 if (nominalRelation < 0)
1239 nominalRelation = appinfo->child_relid;
1242 * Select cheapest path in case there's more than one. We always run
1243 * modification queries to conclusion, so we care only for the
1244 * cheapest-total path.
1246 sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
1247 set_cheapest(sub_final_rel);
1248 subpath = sub_final_rel->cheapest_total_path;
1251 * If this child rel was excluded by constraint exclusion, exclude it
1252 * from the result plan.
1254 if (IS_DUMMY_PATH(subpath))
1258 * If this is the first non-excluded child, its post-planning rtable
1259 * becomes the initial contents of final_rtable; otherwise, append
1260 * just its modified subquery RTEs to final_rtable.
1262 if (final_rtable == NIL)
1263 final_rtable = subroot->parse->rtable;
1266 List *tmp_rtable = NIL;
1271 * Check to see if any of the original RTEs were turned into
1272 * subqueries during planning. Currently, this should only ever
1273 * happen due to securityQuals being involved which push a
1274 * relation down under a subquery, to ensure that the security
1275 * barrier quals are evaluated first.
1277 * When this happens, we want to use the new subqueries in the
1280 forboth(cell1, final_rtable, cell2, subroot->parse->rtable)
1282 RangeTblEntry *rte1 = (RangeTblEntry *) lfirst(cell1);
1283 RangeTblEntry *rte2 = (RangeTblEntry *) lfirst(cell2);
1285 if (rte1->rtekind == RTE_RELATION &&
1286 rte2->rtekind == RTE_SUBQUERY)
1288 /* Should only be when there are securityQuals today */
1289 Assert(rte1->securityQuals != NIL);
1290 tmp_rtable = lappend(tmp_rtable, rte2);
1293 tmp_rtable = lappend(tmp_rtable, rte1);
1296 final_rtable = list_concat(tmp_rtable,
1297 list_copy_tail(subroot->parse->rtable,
1298 list_length(final_rtable)));
1302 * We need to collect all the RelOptInfos from all child plans into
1303 * the main PlannerInfo, since setrefs.c will need them. We use the
1304 * last child's simple_rel_array (previous ones are too short), so we
1305 * have to propagate forward the RelOptInfos that were already built
1306 * in previous children.
1308 Assert(subroot->simple_rel_array_size >= save_rel_array_size);
1309 for (rti = 1; rti < save_rel_array_size; rti++)
1311 RelOptInfo *brel = save_rel_array[rti];
1314 subroot->simple_rel_array[rti] = brel;
1316 save_rel_array_size = subroot->simple_rel_array_size;
1317 save_rel_array = subroot->simple_rel_array;
1319 /* Make sure any initplans from this rel get into the outer list */
1320 root->init_plans = subroot->init_plans;
1322 /* Build list of sub-paths */
1323 subpaths = lappend(subpaths, subpath);
1325 /* Build list of modified subroots, too */
1326 subroots = lappend(subroots, subroot);
1328 /* Build list of target-relation RT indexes */
1329 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
1331 /* Build lists of per-relation WCO and RETURNING targetlists */
1332 if (parse->withCheckOptions)
1333 withCheckOptionLists = lappend(withCheckOptionLists,
1334 subroot->parse->withCheckOptions);
1335 if (parse->returningList)
1336 returningLists = lappend(returningLists,
1337 subroot->parse->returningList);
1339 Assert(!parse->onConflict);
1342 /* Result path must go into outer query's FINAL upperrel */
1343 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1346 * We don't currently worry about setting final_rel's consider_parallel
1347 * flag in this case, nor about allowing FDWs or create_upper_paths_hook
1348 * to get control here.
1352 * If we managed to exclude every child rel, return a dummy plan; it
1353 * doesn't even need a ModifyTable node.
1355 if (subpaths == NIL)
1357 set_dummy_rel_pathlist(final_rel);
1362 * Put back the final adjusted rtable into the master copy of the Query.
1363 * (We mustn't do this if we found no non-excluded children.)
1365 parse->rtable = final_rtable;
1366 root->simple_rel_array_size = save_rel_array_size;
1367 root->simple_rel_array = save_rel_array;
1368 /* Must reconstruct master's simple_rte_array, too */
1369 root->simple_rte_array = (RangeTblEntry **)
1370 palloc0((list_length(final_rtable) + 1) * sizeof(RangeTblEntry *));
1372 foreach(lc, final_rtable)
1374 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lc);
1376 root->simple_rte_array[rti++] = rte;
1380 * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node will
1381 * have dealt with fetching non-locked marked rows, else we need to have
1382 * ModifyTable do that.
1384 if (parse->rowMarks)
1387 rowMarks = root->rowMarks;
1389 /* Create Path representing a ModifyTable to do the UPDATE/DELETE work */
1390 add_path(final_rel, (Path *)
1391 create_modifytable_path(root, final_rel,
1398 withCheckOptionLists,
1402 SS_assign_special_param(root)));
1405 /*--------------------
1407 * Perform planning steps related to grouping, aggregation, etc.
1409 * This function adds all required top-level processing to the scan/join
1410 * Path(s) produced by query_planner.
1412 * If inheritance_update is true, we're being called from inheritance_planner
1413 * and should not include a ModifyTable step in the resulting Path(s).
1414 * (inheritance_planner will create a single ModifyTable node covering all the
1417 * tuple_fraction is the fraction of tuples we expect will be retrieved.
1418 * tuple_fraction is interpreted as follows:
1419 * 0: expect all tuples to be retrieved (normal case)
1420 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1421 * from the plan to be retrieved
1422 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1423 * expected to be retrieved (ie, a LIMIT specification)
1425 * Returns nothing; the useful output is in the Paths we attach to the
1426 * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1427 * root->processed_tlist contains the final processed targetlist.
1429 * Note that we have not done set_cheapest() on the final rel; it's convenient
1430 * to leave this to the caller.
1431 *--------------------
1434 grouping_planner(PlannerInfo *root, bool inheritance_update,
1435 double tuple_fraction)
1437 Query *parse = root->parse;
1438 List *tlist = parse->targetList;
1439 int64 offset_est = 0;
1440 int64 count_est = 0;
1441 double limit_tuples = -1.0;
1442 bool have_postponed_srfs = false;
1444 PathTarget *final_target;
1445 RelOptInfo *current_rel;
1446 RelOptInfo *final_rel;
1449 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1450 if (parse->limitCount || parse->limitOffset)
1452 tuple_fraction = preprocess_limit(root, tuple_fraction,
1453 &offset_est, &count_est);
1456 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1457 * estimate the effects of using a bounded sort.
1459 if (count_est > 0 && offset_est >= 0)
1460 limit_tuples = (double) count_est + (double) offset_est;
1463 /* Make tuple_fraction accessible to lower-level routines */
1464 root->tuple_fraction = tuple_fraction;
1466 if (parse->setOperations)
1469 * If there's a top-level ORDER BY, assume we have to fetch all the
1470 * tuples. This might be too simplistic given all the hackery below
1471 * to possibly avoid the sort; but the odds of accurate estimates here
1472 * are pretty low anyway. XXX try to get rid of this in favor of
1473 * letting plan_set_operations generate both fast-start and
1474 * cheapest-total paths.
1476 if (parse->sortClause)
1477 root->tuple_fraction = 0.0;
1480 * Construct Paths for set operations. The results will not need any
1481 * work except perhaps a top-level sort and/or LIMIT. Note that any
1482 * special work for recursive unions is the responsibility of
1483 * plan_set_operations.
1485 current_rel = plan_set_operations(root);
1488 * We should not need to call preprocess_targetlist, since we must be
1489 * in a SELECT query node. Instead, use the targetlist returned by
1490 * plan_set_operations (since this tells whether it returned any
1491 * resjunk columns!), and transfer any sort key information from the
1494 Assert(parse->commandType == CMD_SELECT);
1496 tlist = root->processed_tlist; /* from plan_set_operations */
1498 /* for safety, copy processed_tlist instead of modifying in-place */
1499 tlist = postprocess_setop_tlist(copyObject(tlist), parse->targetList);
1501 /* Save aside the final decorated tlist */
1502 root->processed_tlist = tlist;
1504 /* Also extract the PathTarget form of the setop result tlist */
1505 final_target = current_rel->cheapest_total_path->pathtarget;
1508 * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1509 * checked already, but let's make sure).
1511 if (parse->rowMarks)
1513 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1515 translator: %s is a SQL row locking clause such as FOR UPDATE */
1516 errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1517 LCS_asString(((RowMarkClause *)
1518 linitial(parse->rowMarks))->strength))));
1521 * Calculate pathkeys that represent result ordering requirements
1523 Assert(parse->distinctClause == NIL);
1524 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1530 /* No set operations, do regular planning */
1531 PathTarget *sort_input_target;
1532 PathTarget *grouping_target;
1533 PathTarget *scanjoin_target;
1535 AggClauseCosts agg_costs;
1536 WindowFuncLists *wflists = NULL;
1537 List *activeWindows = NIL;
1538 List *rollup_lists = NIL;
1539 List *rollup_groupclauses = NIL;
1540 standard_qp_extra qp_extra;
1542 /* A recursive query should always have setOperations */
1543 Assert(!root->hasRecursion);
1545 /* Preprocess grouping sets and GROUP BY clause, if any */
1546 if (parse->groupingSets)
1548 int *tleref_to_colnum_map;
1555 parse->groupingSets = expand_grouping_sets(parse->groupingSets, -1);
1557 /* Identify max SortGroupRef in groupClause, for array sizing */
1559 foreach(lc, parse->groupClause)
1561 SortGroupClause *gc = lfirst(lc);
1563 if (gc->tleSortGroupRef > maxref)
1564 maxref = gc->tleSortGroupRef;
1567 /* Allocate workspace array for remapping */
1568 tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
1570 /* Examine the rollup sets */
1571 sets = extract_rollup_sets(parse->groupingSets);
1573 foreach(lc_set, sets)
1575 List *current_sets = (List *) lfirst(lc_set);
1580 * Reorder the current list of grouping sets into correct
1581 * prefix order. If only one aggregation pass is needed, try
1582 * to make the list match the ORDER BY clause; if more than
1583 * one pass is needed, we don't bother with that.
1585 current_sets = reorder_grouping_sets(current_sets,
1586 (list_length(sets) == 1
1591 * Order the groupClause appropriately. If the first grouping
1592 * set is empty, this can match regular GROUP BY
1593 * preprocessing, otherwise we have to force the groupClause
1594 * to match that grouping set's order.
1596 groupclause = preprocess_groupclause(root,
1597 linitial(current_sets));
1600 * Now that we've pinned down an order for the groupClause for
1601 * this list of grouping sets, we need to remap the entries in
1602 * the grouping sets from sortgrouprefs to plain indices
1603 * (0-based) into the groupClause for this collection of
1607 foreach(lc, groupclause)
1609 SortGroupClause *gc = lfirst(lc);
1611 tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
1614 foreach(lc, current_sets)
1616 foreach(lc2, (List *) lfirst(lc))
1618 lfirst_int(lc2) = tleref_to_colnum_map[lfirst_int(lc2)];
1622 /* Save the reordered sets and corresponding groupclauses */
1623 rollup_lists = lcons(current_sets, rollup_lists);
1624 rollup_groupclauses = lcons(groupclause, rollup_groupclauses);
1629 /* Preprocess regular GROUP BY clause, if any */
1630 if (parse->groupClause)
1631 parse->groupClause = preprocess_groupclause(root, NIL);
1634 /* Preprocess targetlist */
1635 tlist = preprocess_targetlist(root, tlist);
1637 if (parse->onConflict)
1638 parse->onConflict->onConflictSet =
1639 preprocess_onconflict_targetlist(parse->onConflict->onConflictSet,
1640 parse->resultRelation,
1644 * Expand any rangetable entries that have security barrier quals.
1645 * This may add new security barrier subquery RTEs to the rangetable.
1647 expand_security_quals(root, tlist);
1650 * We are now done hacking up the query's targetlist. Most of the
1651 * remaining planning work will be done with the PathTarget
1652 * representation of tlists, but save aside the full representation so
1653 * that we can transfer its decoration (resnames etc) to the topmost
1654 * tlist of the finished Plan.
1656 root->processed_tlist = tlist;
1659 * Collect statistics about aggregates for estimating costs, and mark
1660 * all the aggregates with resolved aggtranstypes. We must do this
1661 * before slicing and dicing the tlist into various pathtargets, else
1662 * some copies of the Aggref nodes might escape being marked with the
1663 * correct transtypes.
1665 * Note: currently, we do not detect duplicate aggregates here. This
1666 * may result in somewhat-overestimated cost, which is fine for our
1667 * purposes since all Paths will get charged the same. But at some
1668 * point we might wish to do that detection in the planner, rather
1669 * than during executor startup.
1671 MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1674 get_agg_clause_costs(root, (Node *) tlist, AGGSPLIT_SIMPLE,
1676 get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
1681 * Locate any window functions in the tlist. (We don't need to look
1682 * anywhere else, since expressions used in ORDER BY will be in there
1683 * too.) Note that they could all have been eliminated by constant
1684 * folding, in which case we don't need to do any more work.
1686 if (parse->hasWindowFuncs)
1688 wflists = find_window_functions((Node *) tlist,
1689 list_length(parse->windowClause));
1690 if (wflists->numWindowFuncs > 0)
1691 activeWindows = select_active_windows(root, wflists);
1693 parse->hasWindowFuncs = false;
1697 * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1698 * adding logic between here and the query_planner() call. Anything
1699 * that is needed in MIN/MAX-optimizable cases will have to be
1700 * duplicated in planagg.c.
1703 preprocess_minmax_aggregates(root, tlist);
1706 * Figure out whether there's a hard limit on the number of rows that
1707 * query_planner's result subplan needs to return. Even if we know a
1708 * hard limit overall, it doesn't apply if the query has any
1709 * grouping/aggregation operations, or SRFs in the tlist.
1711 if (parse->groupClause ||
1712 parse->groupingSets ||
1713 parse->distinctClause ||
1715 parse->hasWindowFuncs ||
1716 parse->hasTargetSRFs ||
1717 root->hasHavingQual)
1718 root->limit_tuples = -1.0;
1720 root->limit_tuples = limit_tuples;
1722 /* Set up data needed by standard_qp_callback */
1723 qp_extra.tlist = tlist;
1724 qp_extra.activeWindows = activeWindows;
1725 qp_extra.groupClause =
1726 parse->groupingSets ? llast(rollup_groupclauses) : parse->groupClause;
1729 * Generate the best unsorted and presorted paths for the scan/join
1730 * portion of this Query, ie the processing represented by the
1731 * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1732 * We also generate (in standard_qp_callback) pathkey representations
1733 * of the query's sort clause, distinct clause, etc.
1735 current_rel = query_planner(root, tlist,
1736 standard_qp_callback, &qp_extra);
1739 * Convert the query's result tlist into PathTarget format.
1741 * Note: it's desirable to not do this till after query_planner(),
1742 * because the target width estimates can use per-Var width numbers
1743 * that were obtained within query_planner().
1745 final_target = create_pathtarget(root, tlist);
1748 * If ORDER BY was given, consider whether we should use a post-sort
1749 * projection, and compute the adjusted target for preceding steps if
1752 if (parse->sortClause)
1753 sort_input_target = make_sort_input_target(root,
1755 &have_postponed_srfs);
1757 sort_input_target = final_target;
1760 * If we have window functions to deal with, the output from any
1761 * grouping step needs to be what the window functions want;
1762 * otherwise, it should be sort_input_target.
1765 grouping_target = make_window_input_target(root,
1769 grouping_target = sort_input_target;
1772 * If we have grouping or aggregation to do, the topmost scan/join
1773 * plan node must emit what the grouping step wants; otherwise, it
1774 * should emit grouping_target.
1776 have_grouping = (parse->groupClause || parse->groupingSets ||
1777 parse->hasAggs || root->hasHavingQual);
1779 scanjoin_target = make_group_input_target(root, final_target);
1781 scanjoin_target = grouping_target;
1784 * Forcibly apply scan/join target to all the Paths for the scan/join
1787 * In principle we should re-run set_cheapest() here to identify the
1788 * cheapest path, but it seems unlikely that adding the same tlist
1789 * eval costs to all the paths would change that, so we don't bother.
1790 * Instead, just assume that the cheapest-startup and cheapest-total
1791 * paths remain so. (There should be no parameterized paths anymore,
1792 * so we needn't worry about updating cheapest_parameterized_paths.)
1794 foreach(lc, current_rel->pathlist)
1796 Path *subpath = (Path *) lfirst(lc);
1799 Assert(subpath->param_info == NULL);
1800 path = apply_projection_to_path(root, current_rel,
1801 subpath, scanjoin_target);
1802 /* If we had to add a Result, path is different from subpath */
1803 if (path != subpath)
1806 if (subpath == current_rel->cheapest_startup_path)
1807 current_rel->cheapest_startup_path = path;
1808 if (subpath == current_rel->cheapest_total_path)
1809 current_rel->cheapest_total_path = path;
1814 * Upper planning steps which make use of the top scan/join rel's
1815 * partial pathlist will expect partial paths for that rel to produce
1816 * the same output as complete paths ... and we just changed the
1817 * output for the complete paths, so we'll need to do the same thing
1818 * for partial paths. But only parallel-safe expressions can be
1819 * computed by partial paths.
1821 if (current_rel->partial_pathlist &&
1822 is_parallel_safe(root, (Node *) scanjoin_target->exprs))
1824 /* Apply the scan/join target to each partial path */
1825 foreach(lc, current_rel->partial_pathlist)
1827 Path *subpath = (Path *) lfirst(lc);
1830 /* Shouldn't have any parameterized paths anymore */
1831 Assert(subpath->param_info == NULL);
1834 * Don't use apply_projection_to_path() here, because there
1835 * could be other pointers to these paths, and therefore we
1836 * mustn't modify them in place.
1838 newpath = (Path *) create_projection_path(root,
1842 lfirst(lc) = newpath;
1848 * In the unfortunate event that scanjoin_target is not
1849 * parallel-safe, we can't apply it to the partial paths; in that
1850 * case, we'll need to forget about the partial paths, which
1851 * aren't valid input for upper planning steps.
1853 current_rel->partial_pathlist = NIL;
1857 * Save the various upper-rel PathTargets we just computed into
1858 * root->upper_targets[]. The core code doesn't use this, but it
1859 * provides a convenient place for extensions to get at the info. For
1860 * consistency, we save all the intermediate targets, even though some
1861 * of the corresponding upperrels might not be needed for this query.
1863 root->upper_targets[UPPERREL_FINAL] = final_target;
1864 root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1865 root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1868 * If we have grouping and/or aggregation, consider ways to implement
1869 * that. We build a new upperrel representing the output of this
1874 current_rel = create_grouping_paths(root,
1879 rollup_groupclauses);
1883 * If we have window functions, consider ways to implement those. We
1884 * build a new upperrel representing the output of this phase.
1888 current_rel = create_window_paths(root,
1898 * If there is a DISTINCT clause, consider ways to implement that. We
1899 * build a new upperrel representing the output of this phase.
1901 if (parse->distinctClause)
1903 current_rel = create_distinct_paths(root,
1907 } /* end of if (setOperations) */
1910 * If ORDER BY was given, consider ways to implement that, and generate a
1911 * new upperrel containing only paths that emit the correct ordering and
1912 * project the correct final_target. We can apply the original
1913 * limit_tuples limit in sort costing here, but only if there are no
1916 if (parse->sortClause)
1918 current_rel = create_ordered_paths(root,
1921 have_postponed_srfs ? -1.0 :
1926 * If there are set-returning functions in the tlist, scale up the output
1927 * rowcounts of all surviving Paths to account for that. Note that if any
1928 * SRFs appear in sorting or grouping columns, we'll have underestimated
1929 * the numbers of rows passing through earlier steps; but that's such a
1930 * weird usage that it doesn't seem worth greatly complicating matters to
1933 if (parse->hasTargetSRFs)
1934 tlist_rows = tlist_returns_set_rows(tlist);
1940 foreach(lc, current_rel->pathlist)
1942 Path *path = (Path *) lfirst(lc);
1945 * We assume that execution costs of the tlist as such were
1946 * already accounted for. However, it still seems appropriate to
1947 * charge something more for the executor's general costs of
1948 * processing the added tuples. The cost is probably less than
1949 * cpu_tuple_cost, though, so we arbitrarily use half of that.
1951 path->total_cost += path->rows * (tlist_rows - 1) *
1954 path->rows *= tlist_rows;
1956 /* No need to run set_cheapest; we're keeping all paths anyway. */
1960 * Now we are prepared to build the final-output upperrel.
1962 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1965 * If the input rel is marked consider_parallel and there's nothing that's
1966 * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1967 * consider_parallel as well. Note that if the query has rowMarks or is
1968 * not a SELECT, consider_parallel will be false for every relation in the
1971 if (current_rel->consider_parallel &&
1972 is_parallel_safe(root, parse->limitOffset) &&
1973 is_parallel_safe(root, parse->limitCount))
1974 final_rel->consider_parallel = true;
1977 * If the current_rel belongs to a single FDW, so does the final_rel.
1979 final_rel->serverid = current_rel->serverid;
1980 final_rel->userid = current_rel->userid;
1981 final_rel->useridiscurrent = current_rel->useridiscurrent;
1982 final_rel->fdwroutine = current_rel->fdwroutine;
1985 * Generate paths for the final_rel. Insert all surviving paths, with
1986 * LockRows, Limit, and/or ModifyTable steps added if needed.
1988 foreach(lc, current_rel->pathlist)
1990 Path *path = (Path *) lfirst(lc);
1993 * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1994 * (Note: we intentionally test parse->rowMarks not root->rowMarks
1995 * here. If there are only non-locking rowmarks, they should be
1996 * handled by the ModifyTable node instead. However, root->rowMarks
1997 * is what goes into the LockRows node.)
1999 if (parse->rowMarks)
2001 path = (Path *) create_lockrows_path(root, final_rel, path,
2003 SS_assign_special_param(root));
2007 * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2009 if (limit_needed(parse))
2011 path = (Path *) create_limit_path(root, final_rel, path,
2014 offset_est, count_est);
2018 * If this is an INSERT/UPDATE/DELETE, and we're not being called from
2019 * inheritance_planner, add the ModifyTable node.
2021 if (parse->commandType != CMD_SELECT && !inheritance_update)
2023 List *withCheckOptionLists;
2024 List *returningLists;
2028 * Set up the WITH CHECK OPTION and RETURNING lists-of-lists, if
2031 if (parse->withCheckOptions)
2032 withCheckOptionLists = list_make1(parse->withCheckOptions);
2034 withCheckOptionLists = NIL;
2036 if (parse->returningList)
2037 returningLists = list_make1(parse->returningList);
2039 returningLists = NIL;
2042 * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2043 * will have dealt with fetching non-locked marked rows, else we
2044 * need to have ModifyTable do that.
2046 if (parse->rowMarks)
2049 rowMarks = root->rowMarks;
2052 create_modifytable_path(root, final_rel,
2055 parse->resultRelation,
2056 list_make1_int(parse->resultRelation),
2059 withCheckOptionLists,
2063 SS_assign_special_param(root));
2066 /* And shove it into final_rel */
2067 add_path(final_rel, path);
2071 * If there is an FDW that's responsible for all baserels of the query,
2072 * let it consider adding ForeignPaths.
2074 if (final_rel->fdwroutine &&
2075 final_rel->fdwroutine->GetForeignUpperPaths)
2076 final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2077 current_rel, final_rel);
2079 /* Let extensions possibly add some more paths */
2080 if (create_upper_paths_hook)
2081 (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2082 current_rel, final_rel);
2084 /* Note: currently, we leave it to callers to do set_cheapest() */
2089 * Detect whether a plan node is a "dummy" plan created when a relation
2090 * is deemed not to need scanning due to constraint exclusion.
2092 * Currently, such dummy plans are Result nodes with constant FALSE
2093 * filter quals (see set_dummy_rel_pathlist and create_append_plan).
2095 * XXX this probably ought to be somewhere else, but not clear where.
2098 is_dummy_plan(Plan *plan)
2100 if (IsA(plan, Result))
2102 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
2104 if (list_length(rcqual) == 1)
2106 Const *constqual = (Const *) linitial(rcqual);
2108 if (constqual && IsA(constqual, Const))
2110 if (!constqual->constisnull &&
2111 !DatumGetBool(constqual->constvalue))
2120 * Create a bitmapset of the RT indexes of live base relations
2122 * Helper for preprocess_rowmarks ... at this point in the proceedings,
2123 * the only good way to distinguish baserels from appendrel children
2124 * is to see what is in the join tree.
2127 get_base_rel_indexes(Node *jtnode)
2133 if (IsA(jtnode, RangeTblRef))
2135 int varno = ((RangeTblRef *) jtnode)->rtindex;
2137 result = bms_make_singleton(varno);
2139 else if (IsA(jtnode, FromExpr))
2141 FromExpr *f = (FromExpr *) jtnode;
2145 foreach(l, f->fromlist)
2146 result = bms_join(result,
2147 get_base_rel_indexes(lfirst(l)));
2149 else if (IsA(jtnode, JoinExpr))
2151 JoinExpr *j = (JoinExpr *) jtnode;
2153 result = bms_join(get_base_rel_indexes(j->larg),
2154 get_base_rel_indexes(j->rarg));
2158 elog(ERROR, "unrecognized node type: %d",
2159 (int) nodeTag(jtnode));
2160 result = NULL; /* keep compiler quiet */
2166 * preprocess_rowmarks - set up PlanRowMarks if needed
2169 preprocess_rowmarks(PlannerInfo *root)
2171 Query *parse = root->parse;
2177 if (parse->rowMarks)
2180 * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2181 * grouping, since grouping renders a reference to individual tuple
2182 * CTIDs invalid. This is also checked at parse time, but that's
2183 * insufficient because of rule substitution, query pullup, etc.
2185 CheckSelectLocking(parse, ((RowMarkClause *)
2186 linitial(parse->rowMarks))->strength);
2191 * We only need rowmarks for UPDATE, DELETE, or FOR [KEY]
2194 if (parse->commandType != CMD_UPDATE &&
2195 parse->commandType != CMD_DELETE)
2200 * We need to have rowmarks for all base relations except the target. We
2201 * make a bitmapset of all base rels and then remove the items we don't
2202 * need or have FOR [KEY] UPDATE/SHARE marks for.
2204 rels = get_base_rel_indexes((Node *) parse->jointree);
2205 if (parse->resultRelation)
2206 rels = bms_del_member(rels, parse->resultRelation);
2209 * Convert RowMarkClauses to PlanRowMark representation.
2212 foreach(l, parse->rowMarks)
2214 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
2215 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2219 * Currently, it is syntactically impossible to have FOR UPDATE et al
2220 * applied to an update/delete target rel. If that ever becomes
2221 * possible, we should drop the target from the PlanRowMark list.
2223 Assert(rc->rti != parse->resultRelation);
2226 * Ignore RowMarkClauses for subqueries; they aren't real tables and
2227 * can't support true locking. Subqueries that got flattened into the
2228 * main query should be ignored completely. Any that didn't will get
2229 * ROW_MARK_COPY items in the next loop.
2231 if (rte->rtekind != RTE_RELATION)
2234 rels = bms_del_member(rels, rc->rti);
2236 newrc = makeNode(PlanRowMark);
2237 newrc->rti = newrc->prti = rc->rti;
2238 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2239 newrc->markType = select_rowmark_type(rte, rc->strength);
2240 newrc->allMarkTypes = (1 << newrc->markType);
2241 newrc->strength = rc->strength;
2242 newrc->waitPolicy = rc->waitPolicy;
2243 newrc->isParent = false;
2245 prowmarks = lappend(prowmarks, newrc);
2249 * Now, add rowmarks for any non-target, non-locked base relations.
2252 foreach(l, parse->rtable)
2254 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
2258 if (!bms_is_member(i, rels))
2261 newrc = makeNode(PlanRowMark);
2262 newrc->rti = newrc->prti = i;
2263 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2264 newrc->markType = select_rowmark_type(rte, LCS_NONE);
2265 newrc->allMarkTypes = (1 << newrc->markType);
2266 newrc->strength = LCS_NONE;
2267 newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2268 newrc->isParent = false;
2270 prowmarks = lappend(prowmarks, newrc);
2273 root->rowMarks = prowmarks;
2277 * Select RowMarkType to use for a given table
2280 select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2282 if (rte->rtekind != RTE_RELATION)
2284 /* If it's not a table at all, use ROW_MARK_COPY */
2285 return ROW_MARK_COPY;
2287 else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2289 /* Let the FDW select the rowmark type, if it wants to */
2290 FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2292 if (fdwroutine->GetForeignRowMarkType != NULL)
2293 return fdwroutine->GetForeignRowMarkType(rte, strength);
2294 /* Otherwise, use ROW_MARK_COPY by default */
2295 return ROW_MARK_COPY;
2299 /* Regular table, apply the appropriate lock type */
2305 * We don't need a tuple lock, only the ability to re-fetch
2306 * the row. Regular tables support ROW_MARK_REFERENCE, but if
2307 * this RTE has security barrier quals, it will be turned into
2308 * a subquery during planning, so use ROW_MARK_COPY.
2310 * This is only necessary for LCS_NONE, since real tuple locks
2311 * on an RTE with security barrier quals are supported by
2312 * pushing the lock down into the subquery --- see
2313 * expand_security_qual.
2315 if (rte->securityQuals != NIL)
2316 return ROW_MARK_COPY;
2317 return ROW_MARK_REFERENCE;
2319 case LCS_FORKEYSHARE:
2320 return ROW_MARK_KEYSHARE;
2323 return ROW_MARK_SHARE;
2325 case LCS_FORNOKEYUPDATE:
2326 return ROW_MARK_NOKEYEXCLUSIVE;
2329 return ROW_MARK_EXCLUSIVE;
2332 elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2333 return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2338 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2340 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2341 * results back in *count_est and *offset_est. These variables are set to
2342 * 0 if the corresponding clause is not present, and -1 if it's present
2343 * but we couldn't estimate the value for it. (The "0" convention is OK
2344 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2345 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2346 * usual practice of never estimating less than one row.) These values will
2347 * be passed to create_limit_path, which see if you change this code.
2349 * The return value is the suitably adjusted tuple_fraction to use for
2350 * planning the query. This adjustment is not overridable, since it reflects
2351 * plan actions that grouping_planner() will certainly take, not assumptions
2355 preprocess_limit(PlannerInfo *root, double tuple_fraction,
2356 int64 *offset_est, int64 *count_est)
2358 Query *parse = root->parse;
2360 double limit_fraction;
2362 /* Should not be called unless LIMIT or OFFSET */
2363 Assert(parse->limitCount || parse->limitOffset);
2366 * Try to obtain the clause values. We use estimate_expression_value
2367 * primarily because it can sometimes do something useful with Params.
2369 if (parse->limitCount)
2371 est = estimate_expression_value(root, parse->limitCount);
2372 if (est && IsA(est, Const))
2374 if (((Const *) est)->constisnull)
2376 /* NULL indicates LIMIT ALL, ie, no limit */
2377 *count_est = 0; /* treat as not present */
2381 *count_est = DatumGetInt64(((Const *) est)->constvalue);
2382 if (*count_est <= 0)
2383 *count_est = 1; /* force to at least 1 */
2387 *count_est = -1; /* can't estimate */
2390 *count_est = 0; /* not present */
2392 if (parse->limitOffset)
2394 est = estimate_expression_value(root, parse->limitOffset);
2395 if (est && IsA(est, Const))
2397 if (((Const *) est)->constisnull)
2399 /* Treat NULL as no offset; the executor will too */
2400 *offset_est = 0; /* treat as not present */
2404 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2405 if (*offset_est < 0)
2406 *offset_est = 0; /* treat as not present */
2410 *offset_est = -1; /* can't estimate */
2413 *offset_est = 0; /* not present */
2415 if (*count_est != 0)
2418 * A LIMIT clause limits the absolute number of tuples returned.
2419 * However, if it's not a constant LIMIT then we have to guess; for
2420 * lack of a better idea, assume 10% of the plan's result is wanted.
2422 if (*count_est < 0 || *offset_est < 0)
2424 /* LIMIT or OFFSET is an expression ... punt ... */
2425 limit_fraction = 0.10;
2429 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2430 limit_fraction = (double) *count_est + (double) *offset_est;
2434 * If we have absolute limits from both caller and LIMIT, use the
2435 * smaller value; likewise if they are both fractional. If one is
2436 * fractional and the other absolute, we can't easily determine which
2437 * is smaller, but we use the heuristic that the absolute will usually
2440 if (tuple_fraction >= 1.0)
2442 if (limit_fraction >= 1.0)
2445 tuple_fraction = Min(tuple_fraction, limit_fraction);
2449 /* caller absolute, limit fractional; use caller's value */
2452 else if (tuple_fraction > 0.0)
2454 if (limit_fraction >= 1.0)
2456 /* caller fractional, limit absolute; use limit */
2457 tuple_fraction = limit_fraction;
2461 /* both fractional */
2462 tuple_fraction = Min(tuple_fraction, limit_fraction);
2467 /* no info from caller, just use limit */
2468 tuple_fraction = limit_fraction;
2471 else if (*offset_est != 0 && tuple_fraction > 0.0)
2474 * We have an OFFSET but no LIMIT. This acts entirely differently
2475 * from the LIMIT case: here, we need to increase rather than decrease
2476 * the caller's tuple_fraction, because the OFFSET acts to cause more
2477 * tuples to be fetched instead of fewer. This only matters if we got
2478 * a tuple_fraction > 0, however.
2480 * As above, use 10% if OFFSET is present but unestimatable.
2482 if (*offset_est < 0)
2483 limit_fraction = 0.10;
2485 limit_fraction = (double) *offset_est;
2488 * If we have absolute counts from both caller and OFFSET, add them
2489 * together; likewise if they are both fractional. If one is
2490 * fractional and the other absolute, we want to take the larger, and
2491 * we heuristically assume that's the fractional one.
2493 if (tuple_fraction >= 1.0)
2495 if (limit_fraction >= 1.0)
2497 /* both absolute, so add them together */
2498 tuple_fraction += limit_fraction;
2502 /* caller absolute, limit fractional; use limit */
2503 tuple_fraction = limit_fraction;
2508 if (limit_fraction >= 1.0)
2510 /* caller fractional, limit absolute; use caller's value */
2514 /* both fractional, so add them together */
2515 tuple_fraction += limit_fraction;
2516 if (tuple_fraction >= 1.0)
2517 tuple_fraction = 0.0; /* assume fetch all */
2522 return tuple_fraction;
2526 * limit_needed - do we actually need a Limit plan node?
2528 * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2529 * a Limit node. This is worth checking for because "OFFSET 0" is a common
2530 * locution for an optimization fence. (Because other places in the planner
2531 * merely check whether parse->limitOffset isn't NULL, it will still work as
2532 * an optimization fence --- we're just suppressing unnecessary run-time
2535 * This might look like it could be merged into preprocess_limit, but there's
2536 * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2537 * in preprocess_limit it's good enough to consider estimated values.
2540 limit_needed(Query *parse)
2544 node = parse->limitCount;
2547 if (IsA(node, Const))
2549 /* NULL indicates LIMIT ALL, ie, no limit */
2550 if (!((Const *) node)->constisnull)
2551 return true; /* LIMIT with a constant value */
2554 return true; /* non-constant LIMIT */
2557 node = parse->limitOffset;
2560 if (IsA(node, Const))
2562 /* Treat NULL as no offset; the executor would too */
2563 if (!((Const *) node)->constisnull)
2565 int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2568 return true; /* OFFSET with a nonzero value */
2572 return true; /* non-constant OFFSET */
2575 return false; /* don't need a Limit plan node */
2580 * remove_useless_groupby_columns
2581 * Remove any columns in the GROUP BY clause that are redundant due to
2582 * being functionally dependent on other GROUP BY columns.
2584 * Since some other DBMSes do not allow references to ungrouped columns, it's
2585 * not unusual to find all columns listed in GROUP BY even though listing the
2586 * primary-key columns would be sufficient. Deleting such excess columns
2587 * avoids redundant sorting work, so it's worth doing. When we do this, we
2588 * must mark the plan as dependent on the pkey constraint (compare the
2589 * parser's check_ungrouped_columns() and check_functional_grouping()).
2591 * In principle, we could treat any NOT-NULL columns appearing in a UNIQUE
2592 * index as the determining columns. But as with check_functional_grouping(),
2593 * there's currently no way to represent dependency on a NOT NULL constraint,
2594 * so we consider only the pkey for now.
2597 remove_useless_groupby_columns(PlannerInfo *root)
2599 Query *parse = root->parse;
2600 Bitmapset **groupbyattnos;
2601 Bitmapset **surplusvars;
2605 /* No chance to do anything if there are less than two GROUP BY items */
2606 if (list_length(parse->groupClause) < 2)
2609 /* Don't fiddle with the GROUP BY clause if the query has grouping sets */
2610 if (parse->groupingSets)
2614 * Scan the GROUP BY clause to find GROUP BY items that are simple Vars.
2615 * Fill groupbyattnos[k] with a bitmapset of the column attnos of RTE k
2616 * that are GROUP BY items.
2618 groupbyattnos = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
2619 (list_length(parse->rtable) + 1));
2620 foreach(lc, parse->groupClause)
2622 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2623 TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
2624 Var *var = (Var *) tle->expr;
2627 * Ignore non-Vars and Vars from other query levels.
2629 * XXX in principle, stable expressions containing Vars could also be
2630 * removed, if all the Vars are functionally dependent on other GROUP
2631 * BY items. But it's not clear that such cases occur often enough to
2632 * be worth troubling over.
2634 if (!IsA(var, Var) ||
2635 var->varlevelsup > 0)
2638 /* OK, remember we have this Var */
2640 Assert(relid <= list_length(parse->rtable));
2641 groupbyattnos[relid] = bms_add_member(groupbyattnos[relid],
2642 var->varattno - FirstLowInvalidHeapAttributeNumber);
2646 * Consider each relation and see if it is possible to remove some of its
2647 * Vars from GROUP BY. For simplicity and speed, we do the actual removal
2648 * in a separate pass. Here, we just fill surplusvars[k] with a bitmapset
2649 * of the column attnos of RTE k that are removable GROUP BY items.
2651 surplusvars = NULL; /* don't allocate array unless required */
2653 foreach(lc, parse->rtable)
2655 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lc);
2656 Bitmapset *relattnos;
2657 Bitmapset *pkattnos;
2662 /* Only plain relations could have primary-key constraints */
2663 if (rte->rtekind != RTE_RELATION)
2666 /* Nothing to do unless this rel has multiple Vars in GROUP BY */
2667 relattnos = groupbyattnos[relid];
2668 if (bms_membership(relattnos) != BMS_MULTIPLE)
2672 * Can't remove any columns for this rel if there is no suitable
2673 * (i.e., nondeferrable) primary key constraint.
2675 pkattnos = get_primary_key_attnos(rte->relid, false, &constraintOid);
2676 if (pkattnos == NULL)
2680 * If the primary key is a proper subset of relattnos then we have
2681 * some items in the GROUP BY that can be removed.
2683 if (bms_subset_compare(pkattnos, relattnos) == BMS_SUBSET1)
2686 * To easily remember whether we've found anything to do, we don't
2687 * allocate the surplusvars[] array until we find something.
2689 if (surplusvars == NULL)
2690 surplusvars = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
2691 (list_length(parse->rtable) + 1));
2693 /* Remember the attnos of the removable columns */
2694 surplusvars[relid] = bms_difference(relattnos, pkattnos);
2696 /* Also, mark the resulting plan as dependent on this constraint */
2697 parse->constraintDeps = lappend_oid(parse->constraintDeps,
2703 * If we found any surplus Vars, build a new GROUP BY clause without them.
2704 * (Note: this may leave some TLEs with unreferenced ressortgroupref
2705 * markings, but that's harmless.)
2707 if (surplusvars != NULL)
2709 List *new_groupby = NIL;
2711 foreach(lc, parse->groupClause)
2713 SortGroupClause *sgc = (SortGroupClause *) lfirst(lc);
2714 TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
2715 Var *var = (Var *) tle->expr;
2718 * New list must include non-Vars, outer Vars, and anything not
2719 * marked as surplus.
2721 if (!IsA(var, Var) ||
2722 var->varlevelsup > 0 ||
2723 !bms_is_member(var->varattno - FirstLowInvalidHeapAttributeNumber,
2724 surplusvars[var->varno]))
2725 new_groupby = lappend(new_groupby, sgc);
2728 parse->groupClause = new_groupby;
2733 * preprocess_groupclause - do preparatory work on GROUP BY clause
2735 * The idea here is to adjust the ordering of the GROUP BY elements
2736 * (which in itself is semantically insignificant) to match ORDER BY,
2737 * thereby allowing a single sort operation to both implement the ORDER BY
2738 * requirement and set up for a Unique step that implements GROUP BY.
2740 * In principle it might be interesting to consider other orderings of the
2741 * GROUP BY elements, which could match the sort ordering of other
2742 * possible plans (eg an indexscan) and thereby reduce cost. We don't
2743 * bother with that, though. Hashed grouping will frequently win anyway.
2745 * Note: we need no comparable processing of the distinctClause because
2746 * the parser already enforced that that matches ORDER BY.
2748 * For grouping sets, the order of items is instead forced to agree with that
2749 * of the grouping set (and items not in the grouping set are skipped). The
2750 * work of sorting the order of grouping set elements to match the ORDER BY if
2751 * possible is done elsewhere.
2754 preprocess_groupclause(PlannerInfo *root, List *force)
2756 Query *parse = root->parse;
2757 List *new_groupclause = NIL;
2762 /* For grouping sets, we need to force the ordering */
2767 Index ref = lfirst_int(sl);
2768 SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2770 new_groupclause = lappend(new_groupclause, cl);
2773 return new_groupclause;
2776 /* If no ORDER BY, nothing useful to do here */
2777 if (parse->sortClause == NIL)
2778 return parse->groupClause;
2781 * Scan the ORDER BY clause and construct a list of matching GROUP BY
2782 * items, but only as far as we can make a matching prefix.
2784 * This code assumes that the sortClause contains no duplicate items.
2786 foreach(sl, parse->sortClause)
2788 SortGroupClause *sc = (SortGroupClause *) lfirst(sl);
2790 foreach(gl, parse->groupClause)
2792 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2796 new_groupclause = lappend(new_groupclause, gc);
2801 break; /* no match, so stop scanning */
2804 /* Did we match all of the ORDER BY list, or just some of it? */
2805 partial_match = (sl != NULL);
2807 /* If no match at all, no point in reordering GROUP BY */
2808 if (new_groupclause == NIL)
2809 return parse->groupClause;
2812 * Add any remaining GROUP BY items to the new list, but only if we were
2813 * able to make a complete match. In other words, we only rearrange the
2814 * GROUP BY list if the result is that one list is a prefix of the other
2815 * --- otherwise there's no possibility of a common sort. Also, give up
2816 * if there are any non-sortable GROUP BY items, since then there's no
2819 foreach(gl, parse->groupClause)
2821 SortGroupClause *gc = (SortGroupClause *) lfirst(gl);
2823 if (list_member_ptr(new_groupclause, gc))
2824 continue; /* it matched an ORDER BY item */
2826 return parse->groupClause; /* give up, no common sort possible */
2827 if (!OidIsValid(gc->sortop))
2828 return parse->groupClause; /* give up, GROUP BY can't be sorted */
2829 new_groupclause = lappend(new_groupclause, gc);
2832 /* Success --- install the rearranged GROUP BY list */
2833 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2834 return new_groupclause;
2838 * Extract lists of grouping sets that can be implemented using a single
2839 * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2841 * Input must be sorted with smallest sets first. Result has each sublist
2842 * sorted with smallest sets first.
2844 * We want to produce the absolute minimum possible number of lists here to
2845 * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2846 * of finding the minimal partition of a partially-ordered set into chains
2847 * (which is what we need, taking the list of grouping sets as a poset ordered
2848 * by set inclusion) can be mapped to the problem of finding the maximum
2849 * cardinality matching on a bipartite graph, which is solvable in polynomial
2850 * time with a worst case of no worse than O(n^2.5) and usually much
2851 * better. Since our N is at most 4096, we don't need to consider fallbacks to
2852 * heuristic or approximate methods. (Planning time for a 12-d cube is under
2853 * half a second on my modest system even with optimization off and assertions
2857 extract_rollup_sets(List *groupingSets)
2859 int num_sets_raw = list_length(groupingSets);
2861 int num_sets = 0; /* distinct sets */
2866 Bitmapset **set_masks;
2869 short *adjacency_buf;
2870 BipartiteMatchState *state;
2874 ListCell *lc1 = list_head(groupingSets);
2878 * Start by stripping out empty sets. The algorithm doesn't require this,
2879 * but the planner currently needs all empty sets to be returned in the
2880 * first list, so we strip them here and add them back after.
2882 while (lc1 && lfirst(lc1) == NIL)
2888 /* bail out now if it turns out that all we had were empty sets. */
2890 return list_make1(groupingSets);
2893 * We don't strictly need to remove duplicate sets here, but if we don't,
2894 * they tend to become scattered through the result, which is a bit
2895 * confusing (and irritating if we ever decide to optimize them out).
2896 * So we remove them here and add them back after.
2898 * For each non-duplicate set, we fill in the following:
2900 * orig_sets[i] = list of the original set lists
2901 * set_masks[i] = bitmapset for testing inclusion
2902 * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2904 * chains[i] will be the result group this set is assigned to.
2906 * We index all of these from 1 rather than 0 because it is convenient
2907 * to leave 0 free for the NIL node in the graph algorithm.
2910 orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
2911 set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
2912 adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
2913 adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
2919 for_each_cell(lc, lc1)
2921 List *candidate = lfirst(lc);
2922 Bitmapset *candidate_set = NULL;
2926 foreach(lc2, candidate)
2928 candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
2931 /* we can only be a dup if we're the same length as a previous set */
2932 if (j_size == list_length(candidate))
2936 for (k = j; k < i; ++k)
2938 if (bms_equal(set_masks[k], candidate_set))
2945 else if (j_size < list_length(candidate))
2947 j_size = list_length(candidate);
2953 orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
2954 bms_free(candidate_set);
2961 orig_sets[i] = list_make1(candidate);
2962 set_masks[i] = candidate_set;
2964 /* fill in adjacency list; no need to compare equal-size sets */
2966 for (k = j - 1; k > 0; --k)
2968 if (bms_is_subset(set_masks[k], candidate_set))
2969 adjacency_buf[++n_adj] = k;
2974 adjacency_buf[0] = n_adj;
2975 adjacency[i] = palloc((n_adj + 1) * sizeof(short));
2976 memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
2979 adjacency[i] = NULL;
2988 * Apply the graph matching algorithm to do the work.
2990 state = BipartiteMatch(num_sets, num_sets, adjacency);
2993 * Now, the state->pair* fields have the info we need to assign sets to
2994 * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
2995 * pair_vu[v] = u (both will be true, but we check both so that we can do
2998 chains = palloc0((num_sets + 1) * sizeof(int));
3000 for (i = 1; i <= num_sets; ++i)
3002 int u = state->pair_vu[i];
3003 int v = state->pair_uv[i];
3006 chains[i] = chains[u];
3007 else if (v > 0 && v < i)
3008 chains[i] = chains[v];
3010 chains[i] = ++num_chains;
3013 /* build result lists. */
3014 results = palloc0((num_chains + 1) * sizeof(List *));
3016 for (i = 1; i <= num_sets; ++i)
3022 results[c] = list_concat(results[c], orig_sets[i]);
3025 /* push any empty sets back on the first list. */
3026 while (num_empty-- > 0)
3027 results[1] = lcons(NIL, results[1]);
3029 /* make result list */
3030 for (i = 1; i <= num_chains; ++i)
3031 result = lappend(result, results[i]);
3034 * Free all the things.
3036 * (This is over-fussy for small sets but for large sets we could have
3037 * tied up a nontrivial amount of memory.)
3039 BipartiteMatchFree(state);
3042 for (i = 1; i <= num_sets; ++i)
3044 pfree(adjacency[i]);
3046 pfree(adjacency_buf);
3048 for (i = 1; i <= num_sets; ++i)
3049 bms_free(set_masks[i]);
3056 * Reorder the elements of a list of grouping sets such that they have correct
3057 * prefix relationships.
3059 * The input must be ordered with smallest sets first; the result is returned
3060 * with largest sets first. Note that the result shares no list substructure
3061 * with the input, so it's safe for the caller to modify it later.
3063 * If we're passed in a sortclause, we follow its order of columns to the
3064 * extent possible, to minimize the chance that we add unnecessary sorts.
3065 * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3066 * gets implemented in one pass.)
3069 reorder_grouping_sets(List *groupingsets, List *sortclause)
3073 List *previous = NIL;
3076 foreach(lc, groupingsets)
3078 List *candidate = lfirst(lc);
3079 List *new_elems = list_difference_int(candidate, previous);
3081 if (list_length(new_elems) > 0)
3083 while (list_length(sortclause) > list_length(previous))
3085 SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3086 int ref = sc->tleSortGroupRef;
3088 if (list_member_int(new_elems, ref))
3090 previous = lappend_int(previous, ref);
3091 new_elems = list_delete_int(new_elems, ref);
3095 /* diverged from the sortclause; give up on it */
3101 foreach(lc2, new_elems)
3103 previous = lappend_int(previous, lfirst_int(lc2));
3107 result = lcons(list_copy(previous), result);
3108 list_free(new_elems);
3111 list_free(previous);
3117 * Compute query_pathkeys and other pathkeys during plan generation
3120 standard_qp_callback(PlannerInfo *root, void *extra)
3122 Query *parse = root->parse;
3123 standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3124 List *tlist = qp_extra->tlist;
3125 List *activeWindows = qp_extra->activeWindows;
3128 * Calculate pathkeys that represent grouping/ordering requirements. The
3129 * sortClause is certainly sort-able, but GROUP BY and DISTINCT might not
3130 * be, in which case we just leave their pathkeys empty.
3132 if (qp_extra->groupClause &&
3133 grouping_is_sortable(qp_extra->groupClause))
3134 root->group_pathkeys =
3135 make_pathkeys_for_sortclauses(root,
3136 qp_extra->groupClause,
3139 root->group_pathkeys = NIL;
3141 /* We consider only the first (bottom) window in pathkeys logic */
3142 if (activeWindows != NIL)
3144 WindowClause *wc = (WindowClause *) linitial(activeWindows);
3146 root->window_pathkeys = make_pathkeys_for_window(root,
3151 root->window_pathkeys = NIL;
3153 if (parse->distinctClause &&
3154 grouping_is_sortable(parse->distinctClause))
3155 root->distinct_pathkeys =
3156 make_pathkeys_for_sortclauses(root,
3157 parse->distinctClause,
3160 root->distinct_pathkeys = NIL;
3162 root->sort_pathkeys =
3163 make_pathkeys_for_sortclauses(root,
3168 * Figure out whether we want a sorted result from query_planner.
3170 * If we have a sortable GROUP BY clause, then we want a result sorted
3171 * properly for grouping. Otherwise, if we have window functions to
3172 * evaluate, we try to sort for the first window. Otherwise, if there's a
3173 * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3174 * we try to produce output that's sufficiently well sorted for the
3175 * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3176 * by the ORDER BY clause.
3178 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3179 * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3180 * that might just leave us failing to exploit an available sort order at
3181 * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3182 * much easier, since we know that the parser ensured that one is a
3183 * superset of the other.
3185 if (root->group_pathkeys)
3186 root->query_pathkeys = root->group_pathkeys;
3187 else if (root->window_pathkeys)
3188 root->query_pathkeys = root->window_pathkeys;
3189 else if (list_length(root->distinct_pathkeys) >
3190 list_length(root->sort_pathkeys))
3191 root->query_pathkeys = root->distinct_pathkeys;
3192 else if (root->sort_pathkeys)
3193 root->query_pathkeys = root->sort_pathkeys;
3195 root->query_pathkeys = NIL;
3199 * Estimate number of groups produced by grouping clauses (1 if not grouping)
3201 * path_rows: number of output rows from scan/join step
3202 * rollup_lists: list of grouping sets, or NIL if not doing grouping sets
3203 * rollup_groupclauses: list of grouping clauses for grouping sets,
3204 * or NIL if not doing grouping sets
3207 get_number_of_groups(PlannerInfo *root,
3210 List *rollup_groupclauses)
3212 Query *parse = root->parse;
3215 if (parse->groupClause)
3219 if (parse->groupingSets)
3221 /* Add up the estimates for each grouping set */
3226 forboth(lc, rollup_groupclauses, lc2, rollup_lists)
3228 List *groupClause = (List *) lfirst(lc);
3229 List *gsets = (List *) lfirst(lc2);
3232 groupExprs = get_sortgrouplist_exprs(groupClause,
3237 List *gset = (List *) lfirst(lc3);
3239 dNumGroups += estimate_num_groups(root,
3248 /* Plain GROUP BY */
3249 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3252 dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3256 else if (parse->groupingSets)
3258 /* Empty grouping sets ... one result row for each one */
3259 dNumGroups = list_length(parse->groupingSets);
3261 else if (parse->hasAggs || root->hasHavingQual)
3263 /* Plain aggregation, one result row */
3276 * estimate_hashagg_tablesize
3277 * estimate the number of bytes that a hash aggregate hashtable will
3278 * require based on the agg_costs, path width and dNumGroups.
3281 estimate_hashagg_tablesize(Path *path, const AggClauseCosts *agg_costs,
3286 /* Estimate per-hash-entry space at tuple width... */
3287 hashentrysize = MAXALIGN(path->pathtarget->width) +
3288 MAXALIGN(SizeofMinimalTupleHeader);
3290 /* plus space for pass-by-ref transition values... */
3291 hashentrysize += agg_costs->transitionSpace;
3292 /* plus the per-hash-entry overhead */
3293 hashentrysize += hash_agg_entry_size(agg_costs->numAggs);
3296 * Note that this disregards the effect of fill-factor and growth policy
3297 * of the hash-table. That's probably ok, given default the default
3298 * fill-factor is relatively high. It'd be hard to meaningfully factor in
3299 * "double-in-size" growth policies here.
3301 return hashentrysize * dNumGroups;
3305 * create_grouping_paths
3307 * Build a new upperrel containing Paths for grouping and/or aggregation.
3309 * input_rel: contains the source-data Paths
3310 * target: the pathtarget for the result Paths to compute
3311 * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
3312 * rollup_lists: list of grouping sets, or NIL if not doing grouping sets
3313 * rollup_groupclauses: list of grouping clauses for grouping sets,
3314 * or NIL if not doing grouping sets
3316 * Note: all Paths in input_rel are expected to return the target computed
3317 * by make_group_input_target.
3319 * We need to consider sorted and hashed aggregation in the same function,
3320 * because otherwise (1) it would be harder to throw an appropriate error
3321 * message if neither way works, and (2) we should not allow hashtable size
3322 * considerations to dissuade us from using hashing if sorting is not possible.
3325 create_grouping_paths(PlannerInfo *root,
3326 RelOptInfo *input_rel,
3328 const AggClauseCosts *agg_costs,
3330 List *rollup_groupclauses)
3332 Query *parse = root->parse;
3333 Path *cheapest_path = input_rel->cheapest_total_path;
3334 RelOptInfo *grouped_rel;
3335 PathTarget *partial_grouping_target = NULL;
3336 AggClauseCosts agg_partial_costs; /* parallel only */
3337 AggClauseCosts agg_final_costs; /* parallel only */
3338 Size hashaggtablesize;
3340 double dNumPartialGroups = 0;
3343 bool try_parallel_aggregation;
3347 /* For now, do all work in the (GROUP_AGG, NULL) upperrel */
3348 grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3351 * If the input relation is not parallel-safe, then the grouped relation
3352 * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3353 * target list and HAVING quals are parallel-safe.
3355 if (input_rel->consider_parallel &&
3356 is_parallel_safe(root, (Node *) target->exprs) &&
3357 is_parallel_safe(root, (Node *) parse->havingQual))
3358 grouped_rel->consider_parallel = true;
3361 * If the input rel belongs to a single FDW, so does the grouped rel.
3363 grouped_rel->serverid = input_rel->serverid;
3364 grouped_rel->userid = input_rel->userid;
3365 grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3366 grouped_rel->fdwroutine = input_rel->fdwroutine;
3369 * Check for degenerate grouping.
3371 if ((root->hasHavingQual || parse->groupingSets) &&
3372 !parse->hasAggs && parse->groupClause == NIL)
3375 * We have a HAVING qual and/or grouping sets, but no aggregates and
3376 * no GROUP BY (which implies that the grouping sets are all empty).
3378 * This is a degenerate case in which we are supposed to emit either
3379 * zero or one row for each grouping set depending on whether HAVING
3380 * succeeds. Furthermore, there cannot be any variables in either
3381 * HAVING or the targetlist, so we actually do not need the FROM table
3382 * at all! We can just throw away the plan-so-far and generate a
3383 * Result node. This is a sufficiently unusual corner case that it's
3384 * not worth contorting the structure of this module to avoid having
3385 * to generate the earlier paths in the first place.
3387 int nrows = list_length(parse->groupingSets);
3393 * Doesn't seem worthwhile writing code to cons up a
3394 * generate_series or a values scan to emit multiple rows. Instead
3395 * just make N clones and append them. (With a volatile HAVING
3396 * clause, this means you might get between 0 and N output rows.
3397 * Offhand I think that's desired.)
3401 while (--nrows >= 0)
3404 create_result_path(root, grouped_rel,
3406 (List *) parse->havingQual);
3407 paths = lappend(paths, path);
3410 create_append_path(grouped_rel,
3414 path->pathtarget = target;
3418 /* No grouping sets, or just one, so one output row */
3420 create_result_path(root, grouped_rel,
3422 (List *) parse->havingQual);
3425 add_path(grouped_rel, path);
3427 /* No need to consider any other alternatives. */
3428 set_cheapest(grouped_rel);
3434 * Estimate number of groups.
3436 dNumGroups = get_number_of_groups(root,
3437 cheapest_path->rows,
3439 rollup_groupclauses);
3442 * Determine whether it's possible to perform sort-based implementations
3443 * of grouping. (Note that if groupClause is empty,
3444 * grouping_is_sortable() is trivially true, and all the
3445 * pathkeys_contained_in() tests will succeed too, so that we'll consider
3446 * every surviving input path.)
3448 can_sort = grouping_is_sortable(parse->groupClause);
3451 * Determine whether we should consider hash-based implementations of
3454 * Hashed aggregation only applies if we're grouping. We currently can't
3455 * hash if there are grouping sets, though.
3457 * Executor doesn't support hashed aggregation with DISTINCT or ORDER BY
3458 * aggregates. (Doing so would imply storing *all* the input values in
3459 * the hash table, and/or running many sorts in parallel, either of which
3460 * seems like a certain loser.) We similarly don't support ordered-set
3461 * aggregates in hashed aggregation, but that case is also included in the
3462 * numOrderedAggs count.
3464 * Note: grouping_is_hashable() is much more expensive to check than the
3465 * other gating conditions, so we want to do it last.
3467 can_hash = (parse->groupClause != NIL &&
3468 parse->groupingSets == NIL &&
3469 agg_costs->numOrderedAggs == 0 &&
3470 grouping_is_hashable(parse->groupClause));
3473 * If grouped_rel->consider_parallel is true, then paths that we generate
3474 * for this grouping relation could be run inside of a worker, but that
3475 * doesn't mean we can actually use the PartialAggregate/FinalizeAggregate
3476 * execution strategy. Figure that out.
3478 if (!grouped_rel->consider_parallel)
3480 /* Not even parallel-safe. */
3481 try_parallel_aggregation = false;
3483 else if (input_rel->partial_pathlist == NIL)
3485 /* Nothing to use as input for partial aggregate. */
3486 try_parallel_aggregation = false;
3488 else if (!parse->hasAggs && parse->groupClause == NIL)
3491 * We don't know how to do parallel aggregation unless we have either
3492 * some aggregates or a grouping clause.
3494 try_parallel_aggregation = false;
3496 else if (parse->groupingSets)
3498 /* We don't know how to do grouping sets in parallel. */
3499 try_parallel_aggregation = false;
3501 else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
3503 /* Insufficient support for partial mode. */
3504 try_parallel_aggregation = false;
3508 /* Everything looks good. */
3509 try_parallel_aggregation = true;
3513 * Before generating paths for grouped_rel, we first generate any possible
3514 * partial paths; that way, later code can easily consider both parallel
3515 * and non-parallel approaches to grouping. Note that the partial paths
3516 * we generate here are also partially aggregated, so simply pushing a
3517 * Gather node on top is insufficient to create a final path, as would be
3518 * the case for a scan/join rel.
3520 if (try_parallel_aggregation)
3522 Path *cheapest_partial_path = linitial(input_rel->partial_pathlist);
3525 * Build target list for partial aggregate paths. These paths cannot
3526 * just emit the same tlist as regular aggregate paths, because (1) we
3527 * must include Vars and Aggrefs needed in HAVING, which might not
3528 * appear in the result tlist, and (2) the Aggrefs must be set in
3531 partial_grouping_target = make_partial_grouping_target(root, target);
3533 /* Estimate number of partial groups. */
3534 dNumPartialGroups = get_number_of_groups(root,
3535 cheapest_partial_path->rows,
3540 * Collect statistics about aggregates for estimating costs of
3541 * performing aggregation in parallel.
3543 MemSet(&agg_partial_costs, 0, sizeof(AggClauseCosts));
3544 MemSet(&agg_final_costs, 0, sizeof(AggClauseCosts));
3548 get_agg_clause_costs(root, (Node *) partial_grouping_target->exprs,
3549 AGGSPLIT_INITIAL_SERIAL,
3550 &agg_partial_costs);
3553 get_agg_clause_costs(root, (Node *) target->exprs,
3554 AGGSPLIT_FINAL_DESERIAL,
3556 get_agg_clause_costs(root, parse->havingQual,
3557 AGGSPLIT_FINAL_DESERIAL,
3563 /* This was checked before setting try_parallel_aggregation */
3564 Assert(parse->hasAggs || parse->groupClause);
3567 * Use any available suitably-sorted path as input, and also
3568 * consider sorting the cheapest partial path.
3570 foreach(lc, input_rel->partial_pathlist)
3572 Path *path = (Path *) lfirst(lc);
3575 is_sorted = pathkeys_contained_in(root->group_pathkeys,
3577 if (path == cheapest_partial_path || is_sorted)
3579 /* Sort the cheapest partial path, if it isn't already */
3581 path = (Path *) create_sort_path(root,
3584 root->group_pathkeys,
3588 add_partial_path(grouped_rel, (Path *)
3589 create_agg_path(root,
3592 partial_grouping_target,
3593 parse->groupClause ? AGG_SORTED : AGG_PLAIN,
3594 AGGSPLIT_INITIAL_SERIAL,
3598 dNumPartialGroups));
3600 add_partial_path(grouped_rel, (Path *)
3601 create_group_path(root,
3604 partial_grouping_target,
3607 dNumPartialGroups));
3615 Assert(parse->hasAggs || parse->groupClause);
3618 estimate_hashagg_tablesize(cheapest_partial_path,
3623 * Tentatively produce a partial HashAgg Path, depending on if it
3624 * looks as if the hash table will fit in work_mem.
3626 if (hashaggtablesize < work_mem * 1024L)
3628 add_partial_path(grouped_rel, (Path *)
3629 create_agg_path(root,
3631 cheapest_partial_path,
3632 partial_grouping_target,
3634 AGGSPLIT_INITIAL_SERIAL,
3638 dNumPartialGroups));
3643 /* Build final grouping paths */
3647 * Use any available suitably-sorted path as input, and also consider
3648 * sorting the cheapest-total path.
3650 foreach(lc, input_rel->pathlist)
3652 Path *path = (Path *) lfirst(lc);
3655 is_sorted = pathkeys_contained_in(root->group_pathkeys,
3657 if (path == cheapest_path || is_sorted)
3659 /* Sort the cheapest-total path if it isn't already sorted */
3661 path = (Path *) create_sort_path(root,
3664 root->group_pathkeys,
3667 /* Now decide what to stick atop it */
3668 if (parse->groupingSets)
3671 * We have grouping sets, possibly with aggregation. Make
3672 * a GroupingSetsPath.
3674 add_path(grouped_rel, (Path *)
3675 create_groupingsets_path(root,
3679 (List *) parse->havingQual,
3681 rollup_groupclauses,
3685 else if (parse->hasAggs)
3688 * We have aggregation, possibly with plain GROUP BY. Make
3691 add_path(grouped_rel, (Path *)
3692 create_agg_path(root,
3696 parse->groupClause ? AGG_SORTED : AGG_PLAIN,
3699 (List *) parse->havingQual,
3703 else if (parse->groupClause)
3706 * We have GROUP BY without aggregation or grouping sets.
3709 add_path(grouped_rel, (Path *)
3710 create_group_path(root,
3715 (List *) parse->havingQual,
3720 /* Other cases should have been handled above */
3727 * Now generate a complete GroupAgg Path atop of the cheapest partial
3728 * path. We need only bother with the cheapest path here, as the
3729 * output of Gather is never sorted.
3731 if (grouped_rel->partial_pathlist)
3733 Path *path = (Path *) linitial(grouped_rel->partial_pathlist);
3734 double total_groups = path->rows * path->parallel_workers;
3736 path = (Path *) create_gather_path(root,
3739 partial_grouping_target,
3744 * Since Gather's output is always unsorted, we'll need to sort,
3745 * unless there's no GROUP BY clause or a degenerate (constant)
3746 * one, in which case there will only be a single group.
3748 if (root->group_pathkeys)
3749 path = (Path *) create_sort_path(root,
3752 root->group_pathkeys,
3756 add_path(grouped_rel, (Path *)
3757 create_agg_path(root,
3761 parse->groupClause ? AGG_SORTED : AGG_PLAIN,
3762 AGGSPLIT_FINAL_DESERIAL,
3764 (List *) parse->havingQual,
3768 add_path(grouped_rel, (Path *)
3769 create_group_path(root,
3774 (List *) parse->havingQual,
3781 hashaggtablesize = estimate_hashagg_tablesize(cheapest_path,
3786 * Provided that the estimated size of the hashtable does not exceed
3787 * work_mem, we'll generate a HashAgg Path, although if we were unable
3788 * to sort above, then we'd better generate a Path, so that we at
3791 if (hashaggtablesize < work_mem * 1024L ||
3792 grouped_rel->pathlist == NIL)
3795 * We just need an Agg over the cheapest-total input path, since
3796 * input order won't matter.
3798 add_path(grouped_rel, (Path *)
3799 create_agg_path(root, grouped_rel,
3805 (List *) parse->havingQual,
3811 * Generate a HashAgg Path atop of the cheapest partial path. Once
3812 * again, we'll only do this if it looks as though the hash table
3813 * won't exceed work_mem.
3815 if (grouped_rel->partial_pathlist)
3817 Path *path = (Path *) linitial(grouped_rel->partial_pathlist);
3819 hashaggtablesize = estimate_hashagg_tablesize(path,
3823 if (hashaggtablesize < work_mem * 1024L)
3825 double total_groups = path->rows * path->parallel_workers;
3827 path = (Path *) create_gather_path(root,
3830 partial_grouping_target,
3834 add_path(grouped_rel, (Path *)
3835 create_agg_path(root,
3840 AGGSPLIT_FINAL_DESERIAL,
3842 (List *) parse->havingQual,
3849 /* Give a helpful error if we failed to find any implementation */
3850 if (grouped_rel->pathlist == NIL)
3852 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
3853 errmsg("could not implement GROUP BY"),
3854 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
3857 * If there is an FDW that's responsible for all baserels of the query,
3858 * let it consider adding ForeignPaths.
3860 if (grouped_rel->fdwroutine &&
3861 grouped_rel->fdwroutine->GetForeignUpperPaths)
3862 grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
3863 input_rel, grouped_rel);
3865 /* Let extensions possibly add some more paths */
3866 if (create_upper_paths_hook)
3867 (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
3868 input_rel, grouped_rel);
3870 /* Now choose the best path(s) */
3871 set_cheapest(grouped_rel);
3877 * create_window_paths
3879 * Build a new upperrel containing Paths for window-function evaluation.
3881 * input_rel: contains the source-data Paths
3882 * input_target: result of make_window_input_target
3883 * output_target: what the topmost WindowAggPath should return
3884 * tlist: query's target list (needed to look up pathkeys)
3885 * wflists: result of find_window_functions
3886 * activeWindows: result of select_active_windows
3888 * Note: all Paths in input_rel are expected to return input_target.
3891 create_window_paths(PlannerInfo *root,
3892 RelOptInfo *input_rel,
3893 PathTarget *input_target,
3894 PathTarget *output_target,
3896 WindowFuncLists *wflists,
3897 List *activeWindows)
3899 RelOptInfo *window_rel;
3902 /* For now, do all work in the (WINDOW, NULL) upperrel */
3903 window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
3906 * If the input relation is not parallel-safe, then the window relation
3907 * can't be parallel-safe, either. Otherwise, we need to examine the
3908 * target list and active windows for non-parallel-safe constructs.
3910 if (input_rel->consider_parallel &&
3911 is_parallel_safe(root, (Node *) output_target->exprs) &&
3912 is_parallel_safe(root, (Node *) activeWindows))
3913 window_rel->consider_parallel = true;
3916 * If the input rel belongs to a single FDW, so does the window rel.
3918 window_rel->serverid = input_rel->serverid;
3919 window_rel->userid = input_rel->userid;
3920 window_rel->useridiscurrent = input_rel->useridiscurrent;
3921 window_rel->fdwroutine = input_rel->fdwroutine;
3924 * Consider computing window functions starting from the existing
3925 * cheapest-total path (which will likely require a sort) as well as any
3926 * existing paths that satisfy root->window_pathkeys (which won't).
3928 foreach(lc, input_rel->pathlist)
3930 Path *path = (Path *) lfirst(lc);
3932 if (path == input_rel->cheapest_total_path ||
3933 pathkeys_contained_in(root->window_pathkeys, path->pathkeys))
3934 create_one_window_path(root,
3945 * If there is an FDW that's responsible for all baserels of the query,
3946 * let it consider adding ForeignPaths.
3948 if (window_rel->fdwroutine &&
3949 window_rel->fdwroutine->GetForeignUpperPaths)
3950 window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
3951 input_rel, window_rel);
3953 /* Let extensions possibly add some more paths */
3954 if (create_upper_paths_hook)
3955 (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
3956 input_rel, window_rel);
3958 /* Now choose the best path(s) */
3959 set_cheapest(window_rel);
3965 * Stack window-function implementation steps atop the given Path, and
3966 * add the result to window_rel.
3968 * window_rel: upperrel to contain result
3969 * path: input Path to use (must return input_target)
3970 * input_target: result of make_window_input_target
3971 * output_target: what the topmost WindowAggPath should return
3972 * tlist: query's target list (needed to look up pathkeys)
3973 * wflists: result of find_window_functions
3974 * activeWindows: result of select_active_windows
3977 create_one_window_path(PlannerInfo *root,
3978 RelOptInfo *window_rel,
3980 PathTarget *input_target,
3981 PathTarget *output_target,
3983 WindowFuncLists *wflists,
3984 List *activeWindows)
3986 PathTarget *window_target;
3990 * Since each window clause could require a different sort order, we stack
3991 * up a WindowAgg node for each clause, with sort steps between them as
3992 * needed. (We assume that select_active_windows chose a good order for
3993 * executing the clauses in.)
3995 * input_target should contain all Vars and Aggs needed for the result.
3996 * (In some cases we wouldn't need to propagate all of these all the way
3997 * to the top, since they might only be needed as inputs to WindowFuncs.
3998 * It's probably not worth trying to optimize that though.) It must also
3999 * contain all window partitioning and sorting expressions, to ensure
4000 * they're computed only once at the bottom of the stack (that's critical
4001 * for volatile functions). As we climb up the stack, we'll add outputs
4002 * for the WindowFuncs computed at each level.
4004 window_target = input_target;
4006 foreach(l, activeWindows)
4008 WindowClause *wc = (WindowClause *) lfirst(l);
4009 List *window_pathkeys;
4011 window_pathkeys = make_pathkeys_for_window(root,
4015 /* Sort if necessary */
4016 if (!pathkeys_contained_in(window_pathkeys, path->pathkeys))
4018 path = (Path *) create_sort_path(root, window_rel,
4027 * Add the current WindowFuncs to the output target for this
4028 * intermediate WindowAggPath. We must copy window_target to
4029 * avoid changing the previous path's target.
4031 * Note: a WindowFunc adds nothing to the target's eval costs; but
4032 * we do need to account for the increase in tlist width.
4036 window_target = copy_pathtarget(window_target);
4037 foreach(lc2, wflists->windowFuncs[wc->winref])
4039 WindowFunc *wfunc = (WindowFunc *) lfirst(lc2);
4041 Assert(IsA(wfunc, WindowFunc));
4042 add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4043 window_target->width += get_typavgwidth(wfunc->wintype, -1);
4048 /* Install the goal target in the topmost WindowAgg */
4049 window_target = output_target;
4053 create_windowagg_path(root, window_rel, path, window_target,
4054 wflists->windowFuncs[wc->winref],
4059 add_path(window_rel, path);
4063 * create_distinct_paths
4065 * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4067 * input_rel: contains the source-data Paths
4069 * Note: input paths should already compute the desired pathtarget, since
4070 * Sort/Unique won't project anything.
4073 create_distinct_paths(PlannerInfo *root,
4074 RelOptInfo *input_rel)
4076 Query *parse = root->parse;
4077 Path *cheapest_input_path = input_rel->cheapest_total_path;
4078 RelOptInfo *distinct_rel;
4079 double numDistinctRows;
4084 /* For now, do all work in the (DISTINCT, NULL) upperrel */
4085 distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4088 * We don't compute anything at this level, so distinct_rel will be
4089 * parallel-safe if the input rel is parallel-safe. In particular, if
4090 * there is a DISTINCT ON (...) clause, any path for the input_rel will
4091 * output those expressions, and will not be parallel-safe unless those
4092 * expressions are parallel-safe.
4094 distinct_rel->consider_parallel = input_rel->consider_parallel;
4097 * If the input rel belongs to a single FDW, so does the distinct_rel.
4099 distinct_rel->serverid = input_rel->serverid;
4100 distinct_rel->userid = input_rel->userid;
4101 distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4102 distinct_rel->fdwroutine = input_rel->fdwroutine;
4104 /* Estimate number of distinct rows there will be */
4105 if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4106 root->hasHavingQual)
4109 * If there was grouping or aggregation, use the number of input rows
4110 * as the estimated number of DISTINCT rows (ie, assume the input is
4111 * already mostly unique).
4113 numDistinctRows = cheapest_input_path->rows;
4118 * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4120 List *distinctExprs;
4122 distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
4124 numDistinctRows = estimate_num_groups(root, distinctExprs,
4125 cheapest_input_path->rows,
4130 * Consider sort-based implementations of DISTINCT, if possible.
4132 if (grouping_is_sortable(parse->distinctClause))
4135 * First, if we have any adequately-presorted paths, just stick a
4136 * Unique node on those. Then consider doing an explicit sort of the
4137 * cheapest input path and Unique'ing that.
4139 * When we have DISTINCT ON, we must sort by the more rigorous of
4140 * DISTINCT and ORDER BY, else it won't have the desired behavior.
4141 * Also, if we do have to do an explicit sort, we might as well use
4142 * the more rigorous ordering to avoid a second sort later. (Note
4143 * that the parser will have ensured that one clause is a prefix of
4146 List *needed_pathkeys;
4148 if (parse->hasDistinctOn &&
4149 list_length(root->distinct_pathkeys) <
4150 list_length(root->sort_pathkeys))
4151 needed_pathkeys = root->sort_pathkeys;
4153 needed_pathkeys = root->distinct_pathkeys;
4155 foreach(lc, input_rel->pathlist)
4157 Path *path = (Path *) lfirst(lc);
4159 if (pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4161 add_path(distinct_rel, (Path *)
4162 create_upper_unique_path(root, distinct_rel,
4164 list_length(root->distinct_pathkeys),
4169 /* For explicit-sort case, always use the more rigorous clause */
4170 if (list_length(root->distinct_pathkeys) <
4171 list_length(root->sort_pathkeys))
4173 needed_pathkeys = root->sort_pathkeys;
4174 /* Assert checks that parser didn't mess up... */
4175 Assert(pathkeys_contained_in(root->distinct_pathkeys,
4179 needed_pathkeys = root->distinct_pathkeys;
4181 path = cheapest_input_path;
4182 if (!pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4183 path = (Path *) create_sort_path(root, distinct_rel,
4188 add_path(distinct_rel, (Path *)
4189 create_upper_unique_path(root, distinct_rel,
4191 list_length(root->distinct_pathkeys),
4196 * Consider hash-based implementations of DISTINCT, if possible.
4198 * If we were not able to make any other types of path, we *must* hash or
4199 * die trying. If we do have other choices, there are several things that
4200 * should prevent selection of hashing: if the query uses DISTINCT ON
4201 * (because it won't really have the expected behavior if we hash), or if
4202 * enable_hashagg is off, or if it looks like the hashtable will exceed
4205 * Note: grouping_is_hashable() is much more expensive to check than the
4206 * other gating conditions, so we want to do it last.
4208 if (distinct_rel->pathlist == NIL)
4209 allow_hash = true; /* we have no alternatives */
4210 else if (parse->hasDistinctOn || !enable_hashagg)
4211 allow_hash = false; /* policy-based decision not to hash */
4216 /* Estimate per-hash-entry space at tuple width... */
4217 hashentrysize = MAXALIGN(cheapest_input_path->pathtarget->width) +
4218 MAXALIGN(SizeofMinimalTupleHeader);
4219 /* plus the per-hash-entry overhead */
4220 hashentrysize += hash_agg_entry_size(0);
4222 /* Allow hashing only if hashtable is predicted to fit in work_mem */
4223 allow_hash = (hashentrysize * numDistinctRows <= work_mem * 1024L);
4226 if (allow_hash && grouping_is_hashable(parse->distinctClause))
4228 /* Generate hashed aggregate path --- no sort needed */
4229 add_path(distinct_rel, (Path *)
4230 create_agg_path(root,
4232 cheapest_input_path,
4233 cheapest_input_path->pathtarget,
4236 parse->distinctClause,
4242 /* Give a helpful error if we failed to find any implementation */
4243 if (distinct_rel->pathlist == NIL)
4245 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4246 errmsg("could not implement DISTINCT"),
4247 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4250 * If there is an FDW that's responsible for all baserels of the query,
4251 * let it consider adding ForeignPaths.
4253 if (distinct_rel->fdwroutine &&
4254 distinct_rel->fdwroutine->GetForeignUpperPaths)
4255 distinct_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_DISTINCT,
4256 input_rel, distinct_rel);
4258 /* Let extensions possibly add some more paths */
4259 if (create_upper_paths_hook)
4260 (*create_upper_paths_hook) (root, UPPERREL_DISTINCT,
4261 input_rel, distinct_rel);
4263 /* Now choose the best path(s) */
4264 set_cheapest(distinct_rel);
4266 return distinct_rel;
4270 * create_ordered_paths
4272 * Build a new upperrel containing Paths for ORDER BY evaluation.
4274 * All paths in the result must satisfy the ORDER BY ordering.
4275 * The only new path we need consider is an explicit sort on the
4276 * cheapest-total existing path.
4278 * input_rel: contains the source-data Paths
4279 * target: the output tlist the result Paths must emit
4280 * limit_tuples: estimated bound on the number of output tuples,
4281 * or -1 if no LIMIT or couldn't estimate
4284 create_ordered_paths(PlannerInfo *root,
4285 RelOptInfo *input_rel,
4287 double limit_tuples)
4289 Path *cheapest_input_path = input_rel->cheapest_total_path;
4290 RelOptInfo *ordered_rel;
4293 /* For now, do all work in the (ORDERED, NULL) upperrel */
4294 ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
4297 * If the input relation is not parallel-safe, then the ordered relation
4298 * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4299 * target list is parallel-safe.
4301 if (input_rel->consider_parallel &&
4302 is_parallel_safe(root, (Node *) target->exprs))
4303 ordered_rel->consider_parallel = true;
4306 * If the input rel belongs to a single FDW, so does the ordered_rel.
4308 ordered_rel->serverid = input_rel->serverid;
4309 ordered_rel->userid = input_rel->userid;
4310 ordered_rel->useridiscurrent = input_rel->useridiscurrent;
4311 ordered_rel->fdwroutine = input_rel->fdwroutine;
4313 foreach(lc, input_rel->pathlist)
4315 Path *path = (Path *) lfirst(lc);
4318 is_sorted = pathkeys_contained_in(root->sort_pathkeys,
4320 if (path == cheapest_input_path || is_sorted)
4324 /* An explicit sort here can take advantage of LIMIT */
4325 path = (Path *) create_sort_path(root,
4328 root->sort_pathkeys,
4332 /* Add projection step if needed */
4333 if (path->pathtarget != target)
4334 path = apply_projection_to_path(root, ordered_rel,
4337 add_path(ordered_rel, path);
4342 * If there is an FDW that's responsible for all baserels of the query,
4343 * let it consider adding ForeignPaths.
4345 if (ordered_rel->fdwroutine &&
4346 ordered_rel->fdwroutine->GetForeignUpperPaths)
4347 ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
4348 input_rel, ordered_rel);
4350 /* Let extensions possibly add some more paths */
4351 if (create_upper_paths_hook)
4352 (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
4353 input_rel, ordered_rel);
4356 * No need to bother with set_cheapest here; grouping_planner does not
4359 Assert(ordered_rel->pathlist != NIL);
4366 * make_group_input_target
4367 * Generate appropriate PathTarget for initial input to grouping nodes.
4369 * If there is grouping or aggregation, the scan/join subplan cannot emit
4370 * the query's final targetlist; for example, it certainly can't emit any
4371 * aggregate function calls. This routine generates the correct target
4372 * for the scan/join subplan.
4374 * The query target list passed from the parser already contains entries
4375 * for all ORDER BY and GROUP BY expressions, but it will not have entries
4376 * for variables used only in HAVING clauses; so we need to add those
4377 * variables to the subplan target list. Also, we flatten all expressions
4378 * except GROUP BY items into their component variables; other expressions
4379 * will be computed by the upper plan nodes rather than by the subplan.
4380 * For example, given a query like
4381 * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
4382 * we want to pass this targetlist to the subplan:
4384 * where the a+b target will be used by the Sort/Group steps, and the
4385 * other targets will be used for computing the final results.
4387 * 'final_target' is the query's final target list (in PathTarget form)
4389 * The result is the PathTarget to be computed by the Paths returned from
4393 make_group_input_target(PlannerInfo *root, PathTarget *final_target)
4395 Query *parse = root->parse;
4396 PathTarget *input_target;
4397 List *non_group_cols;
4398 List *non_group_vars;
4403 * We must build a target containing all grouping columns, plus any other
4404 * Vars mentioned in the query's targetlist and HAVING qual.
4406 input_target = create_empty_pathtarget();
4407 non_group_cols = NIL;
4410 foreach(lc, final_target->exprs)
4412 Expr *expr = (Expr *) lfirst(lc);
4413 Index sgref = get_pathtarget_sortgroupref(final_target, i);
4415 if (sgref && parse->groupClause &&
4416 get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
4419 * It's a grouping column, so add it to the input target as-is.
4421 add_column_to_pathtarget(input_target, expr, sgref);
4426 * Non-grouping column, so just remember the expression for later
4427 * call to pull_var_clause.
4429 non_group_cols = lappend(non_group_cols, expr);
4436 * If there's a HAVING clause, we'll need the Vars it uses, too.
4438 if (parse->havingQual)
4439 non_group_cols = lappend(non_group_cols, parse->havingQual);
4442 * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
4443 * add them to the input target if not already present. (A Var used
4444 * directly as a GROUP BY item will be present already.) Note this
4445 * includes Vars used in resjunk items, so we are covering the needs of
4446 * ORDER BY and window specifications. Vars used within Aggrefs and
4447 * WindowFuncs will be pulled out here, too.
4449 non_group_vars = pull_var_clause((Node *) non_group_cols,
4450 PVC_RECURSE_AGGREGATES |
4451 PVC_RECURSE_WINDOWFUNCS |
4452 PVC_INCLUDE_PLACEHOLDERS);
4453 add_new_columns_to_pathtarget(input_target, non_group_vars);
4455 /* clean up cruft */
4456 list_free(non_group_vars);
4457 list_free(non_group_cols);
4459 /* XXX this causes some redundant cost calculation ... */
4460 return set_pathtarget_cost_width(root, input_target);
4464 * make_partial_grouping_target
4465 * Generate appropriate PathTarget for output of partial aggregate
4466 * (or partial grouping, if there are no aggregates) nodes.
4468 * A partial aggregation node needs to emit all the same aggregates that
4469 * a regular aggregation node would, plus any aggregates used in HAVING;
4470 * except that the Aggref nodes should be marked as partial aggregates.
4472 * In addition, we'd better emit any Vars and PlaceholderVars that are
4473 * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
4474 * these would be Vars that are grouped by or used in grouping expressions.)
4476 * grouping_target is the tlist to be emitted by the topmost aggregation step.
4477 * We get the HAVING clause out of *root.
4480 make_partial_grouping_target(PlannerInfo *root, PathTarget *grouping_target)
4482 Query *parse = root->parse;
4483 PathTarget *partial_target;
4484 List *non_group_cols;
4485 List *non_group_exprs;
4489 partial_target = create_empty_pathtarget();
4490 non_group_cols = NIL;
4493 foreach(lc, grouping_target->exprs)
4495 Expr *expr = (Expr *) lfirst(lc);
4496 Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
4498 if (sgref && parse->groupClause &&
4499 get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
4502 * It's a grouping column, so add it to the partial_target as-is.
4503 * (This allows the upper agg step to repeat the grouping calcs.)
4505 add_column_to_pathtarget(partial_target, expr, sgref);
4510 * Non-grouping column, so just remember the expression for later
4511 * call to pull_var_clause.
4513 non_group_cols = lappend(non_group_cols, expr);
4520 * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
4522 if (parse->havingQual)
4523 non_group_cols = lappend(non_group_cols, parse->havingQual);
4526 * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
4527 * non-group cols (plus HAVING), and add them to the partial_target if not
4528 * already present. (An expression used directly as a GROUP BY item will
4529 * be present already.) Note this includes Vars used in resjunk items, so
4530 * we are covering the needs of ORDER BY and window specifications.
4532 non_group_exprs = pull_var_clause((Node *) non_group_cols,
4533 PVC_INCLUDE_AGGREGATES |
4534 PVC_RECURSE_WINDOWFUNCS |
4535 PVC_INCLUDE_PLACEHOLDERS);
4537 add_new_columns_to_pathtarget(partial_target, non_group_exprs);
4540 * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
4541 * are at the top level of the target list, so we can just scan the list
4542 * rather than recursing through the expression trees.
4544 foreach(lc, partial_target->exprs)
4546 Aggref *aggref = (Aggref *) lfirst(lc);
4548 if (IsA(aggref, Aggref))
4553 * We shouldn't need to copy the substructure of the Aggref node,
4554 * but flat-copy the node itself to avoid damaging other trees.
4556 newaggref = makeNode(Aggref);
4557 memcpy(newaggref, aggref, sizeof(Aggref));
4559 /* For now, assume serialization is required */
4560 mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
4562 lfirst(lc) = newaggref;
4566 /* clean up cruft */
4567 list_free(non_group_exprs);
4568 list_free(non_group_cols);
4570 /* XXX this causes some redundant cost calculation ... */
4571 return set_pathtarget_cost_width(root, partial_target);
4575 * mark_partial_aggref
4576 * Adjust an Aggref to make it represent a partial-aggregation step.
4578 * The Aggref node is modified in-place; caller must do any copying required.
4581 mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
4583 /* aggtranstype should be computed by this point */
4584 Assert(OidIsValid(agg->aggtranstype));
4585 /* ... but aggsplit should still be as the parser left it */
4586 Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
4588 /* Mark the Aggref with the intended partial-aggregation mode */
4589 agg->aggsplit = aggsplit;
4592 * Adjust result type if needed. Normally, a partial aggregate returns
4593 * the aggregate's transition type; but if that's INTERNAL and we're
4594 * serializing, it returns BYTEA instead.
4596 if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
4598 if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
4599 agg->aggtype = BYTEAOID;
4601 agg->aggtype = agg->aggtranstype;
4606 * postprocess_setop_tlist
4607 * Fix up targetlist returned by plan_set_operations().
4609 * We need to transpose sort key info from the orig_tlist into new_tlist.
4610 * NOTE: this would not be good enough if we supported resjunk sort keys
4611 * for results of set operations --- then, we'd need to project a whole
4612 * new tlist to evaluate the resjunk columns. For now, just ereport if we
4613 * find any resjunk columns in orig_tlist.
4616 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
4619 ListCell *orig_tlist_item = list_head(orig_tlist);
4621 foreach(l, new_tlist)
4623 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
4624 TargetEntry *orig_tle;
4626 /* ignore resjunk columns in setop result */
4627 if (new_tle->resjunk)
4630 Assert(orig_tlist_item != NULL);
4631 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
4632 orig_tlist_item = lnext(orig_tlist_item);
4633 if (orig_tle->resjunk) /* should not happen */
4634 elog(ERROR, "resjunk output columns are not implemented");
4635 Assert(new_tle->resno == orig_tle->resno);
4636 new_tle->ressortgroupref = orig_tle->ressortgroupref;
4638 if (orig_tlist_item != NULL)
4639 elog(ERROR, "resjunk output columns are not implemented");
4644 * select_active_windows
4645 * Create a list of the "active" window clauses (ie, those referenced
4646 * by non-deleted WindowFuncs) in the order they are to be executed.
4649 select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
4655 /* First, make a list of the active windows */
4657 foreach(lc, root->parse->windowClause)
4659 WindowClause *wc = (WindowClause *) lfirst(lc);
4661 /* It's only active if wflists shows some related WindowFuncs */
4662 Assert(wc->winref <= wflists->maxWinRef);
4663 if (wflists->windowFuncs[wc->winref] != NIL)
4664 actives = lappend(actives, wc);
4668 * Now, ensure that windows with identical partitioning/ordering clauses
4669 * are adjacent in the list. This is required by the SQL standard, which
4670 * says that only one sort is to be used for such windows, even if they
4671 * are otherwise distinct (eg, different names or framing clauses).
4673 * There is room to be much smarter here, for example detecting whether
4674 * one window's sort keys are a prefix of another's (so that sorting for
4675 * the latter would do for the former), or putting windows first that
4676 * match a sort order available for the underlying query. For the moment
4677 * we are content with meeting the spec.
4680 while (actives != NIL)
4682 WindowClause *wc = (WindowClause *) linitial(actives);
4686 /* Move wc from actives to result */
4687 actives = list_delete_first(actives);
4688 result = lappend(result, wc);
4690 /* Now move any matching windows from actives to result */
4692 for (lc = list_head(actives); lc; lc = next)
4694 WindowClause *wc2 = (WindowClause *) lfirst(lc);
4697 /* framing options are NOT to be compared here! */
4698 if (equal(wc->partitionClause, wc2->partitionClause) &&
4699 equal(wc->orderClause, wc2->orderClause))
4701 actives = list_delete_cell(actives, lc, prev);
4702 result = lappend(result, wc2);
4713 * make_window_input_target
4714 * Generate appropriate PathTarget for initial input to WindowAgg nodes.
4716 * When the query has window functions, this function computes the desired
4717 * target to be computed by the node just below the first WindowAgg.
4718 * This tlist must contain all values needed to evaluate the window functions,
4719 * compute the final target list, and perform any required final sort step.
4720 * If multiple WindowAggs are needed, each intermediate one adds its window
4721 * function results onto this base tlist; only the topmost WindowAgg computes
4722 * the actual desired target list.
4724 * This function is much like make_group_input_target, though not quite enough
4725 * like it to share code. As in that function, we flatten most expressions
4726 * into their component variables. But we do not want to flatten window
4727 * PARTITION BY/ORDER BY clauses, since that might result in multiple
4728 * evaluations of them, which would be bad (possibly even resulting in
4729 * inconsistent answers, if they contain volatile functions).
4730 * Also, we must not flatten GROUP BY clauses that were left unflattened by
4731 * make_group_input_target, because we may no longer have access to the
4732 * individual Vars in them.
4734 * Another key difference from make_group_input_target is that we don't
4735 * flatten Aggref expressions, since those are to be computed below the
4736 * window functions and just referenced like Vars above that.
4738 * 'final_target' is the query's final target list (in PathTarget form)
4739 * 'activeWindows' is the list of active windows previously identified by
4740 * select_active_windows.
4742 * The result is the PathTarget to be computed by the plan node immediately
4743 * below the first WindowAgg node.
4746 make_window_input_target(PlannerInfo *root,
4747 PathTarget *final_target,
4748 List *activeWindows)
4750 Query *parse = root->parse;
4751 PathTarget *input_target;
4753 List *flattenable_cols;
4754 List *flattenable_vars;
4758 Assert(parse->hasWindowFuncs);
4761 * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
4762 * into a bitmapset for convenient reference below.
4765 foreach(lc, activeWindows)
4767 WindowClause *wc = (WindowClause *) lfirst(lc);
4770 foreach(lc2, wc->partitionClause)
4772 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
4774 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
4776 foreach(lc2, wc->orderClause)
4778 SortGroupClause *sortcl = (SortGroupClause *) lfirst(lc2);
4780 sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
4784 /* Add in sortgroupref numbers of GROUP BY clauses, too */
4785 foreach(lc, parse->groupClause)
4787 SortGroupClause *grpcl = (SortGroupClause *) lfirst(lc);
4789 sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
4793 * Construct a target containing all the non-flattenable targetlist items,
4794 * and save aside the others for a moment.
4796 input_target = create_empty_pathtarget();
4797 flattenable_cols = NIL;
4800 foreach(lc, final_target->exprs)
4802 Expr *expr = (Expr *) lfirst(lc);
4803 Index sgref = get_pathtarget_sortgroupref(final_target, i);
4806 * Don't want to deconstruct window clauses or GROUP BY items. (Note
4807 * that such items can't contain window functions, so it's okay to
4808 * compute them below the WindowAgg nodes.)
4810 if (sgref != 0 && bms_is_member(sgref, sgrefs))
4813 * Don't want to deconstruct this value, so add it to the input
4816 add_column_to_pathtarget(input_target, expr, sgref);
4821 * Column is to be flattened, so just remember the expression for
4822 * later call to pull_var_clause.
4824 flattenable_cols = lappend(flattenable_cols, expr);
4831 * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
4832 * add them to the input target if not already present. (Some might be
4833 * there already because they're used directly as window/group clauses.)
4835 * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
4836 * Aggrefs are placed in the Agg node's tlist and not left to be computed
4837 * at higher levels. On the other hand, we should recurse into
4838 * WindowFuncs to make sure their input expressions are available.
4840 flattenable_vars = pull_var_clause((Node *) flattenable_cols,
4841 PVC_INCLUDE_AGGREGATES |
4842 PVC_RECURSE_WINDOWFUNCS |
4843 PVC_INCLUDE_PLACEHOLDERS);
4844 add_new_columns_to_pathtarget(input_target, flattenable_vars);
4846 /* clean up cruft */
4847 list_free(flattenable_vars);
4848 list_free(flattenable_cols);
4850 /* XXX this causes some redundant cost calculation ... */
4851 return set_pathtarget_cost_width(root, input_target);
4855 * make_pathkeys_for_window
4856 * Create a pathkeys list describing the required input ordering
4857 * for the given WindowClause.
4859 * The required ordering is first the PARTITION keys, then the ORDER keys.
4860 * In the future we might try to implement windowing using hashing, in which
4861 * case the ordering could be relaxed, but for now we always sort.
4863 * Caution: if you change this, see createplan.c's get_column_info_for_window!
4866 make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
4869 List *window_pathkeys;
4870 List *window_sortclauses;
4872 /* Throw error if can't sort */
4873 if (!grouping_is_sortable(wc->partitionClause))
4875 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4876 errmsg("could not implement window PARTITION BY"),
4877 errdetail("Window partitioning columns must be of sortable datatypes.")));
4878 if (!grouping_is_sortable(wc->orderClause))
4880 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4881 errmsg("could not implement window ORDER BY"),
4882 errdetail("Window ordering columns must be of sortable datatypes.")));
4884 /* Okay, make the combined pathkeys */
4885 window_sortclauses = list_concat(list_copy(wc->partitionClause),
4886 list_copy(wc->orderClause));
4887 window_pathkeys = make_pathkeys_for_sortclauses(root,
4890 list_free(window_sortclauses);
4891 return window_pathkeys;
4895 * make_sort_input_target
4896 * Generate appropriate PathTarget for initial input to Sort step.
4898 * If the query has ORDER BY, this function chooses the target to be computed
4899 * by the node just below the Sort (and DISTINCT, if any, since Unique can't
4900 * project) steps. This might or might not be identical to the query's final
4903 * The main argument for keeping the sort-input tlist the same as the final
4904 * is that we avoid a separate projection node (which will be needed if
4905 * they're different, because Sort can't project). However, there are also
4906 * advantages to postponing tlist evaluation till after the Sort: it ensures
4907 * a consistent order of evaluation for any volatile functions in the tlist,
4908 * and if there's also a LIMIT, we can stop the query without ever computing
4909 * tlist functions for later rows, which is beneficial for both volatile and
4910 * expensive functions.
4912 * Our current policy is to postpone volatile expressions till after the sort
4913 * unconditionally (assuming that that's possible, ie they are in plain tlist
4914 * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
4915 * postpone set-returning expressions, because running them beforehand would
4916 * bloat the sort dataset, and because it might cause unexpected output order
4917 * if the sort isn't stable. However there's a constraint on that: all SRFs
4918 * in the tlist should be evaluated at the same plan step, so that they can
4919 * run in sync in ExecTargetList. So if any SRFs are in sort columns, we
4920 * mustn't postpone any SRFs. (Note that in principle that policy should
4921 * probably get applied to the group/window input targetlists too, but we
4922 * have not done that historically.) Lastly, expensive expressions are
4923 * postponed if there is a LIMIT, or if root->tuple_fraction shows that
4924 * partial evaluation of the query is possible (if neither is true, we expect
4925 * to have to evaluate the expressions for every row anyway), or if there are
4926 * any volatile or set-returning expressions (since once we've put in a
4927 * projection at all, it won't cost any more to postpone more stuff).
4929 * Another issue that could potentially be considered here is that
4930 * evaluating tlist expressions could result in data that's either wider
4931 * or narrower than the input Vars, thus changing the volume of data that
4932 * has to go through the Sort. However, we usually have only a very bad
4933 * idea of the output width of any expression more complex than a Var,
4934 * so for now it seems too risky to try to optimize on that basis.
4936 * Note that if we do produce a modified sort-input target, and then the
4937 * query ends up not using an explicit Sort, no particular harm is done:
4938 * we'll initially use the modified target for the preceding path nodes,
4939 * but then change them to the final target with apply_projection_to_path.
4940 * Moreover, in such a case the guarantees about evaluation order of
4941 * volatile functions still hold, since the rows are sorted already.
4943 * This function has some things in common with make_group_input_target and
4944 * make_window_input_target, though the detailed rules for what to do are
4945 * different. We never flatten/postpone any grouping or ordering columns;
4946 * those are needed before the sort. If we do flatten a particular
4947 * expression, we leave Aggref and WindowFunc nodes alone, since those were
4950 * 'final_target' is the query's final target list (in PathTarget form)
4951 * 'have_postponed_srfs' is an output argument, see below
4953 * The result is the PathTarget to be computed by the plan node immediately
4954 * below the Sort step (and the Distinct step, if any). This will be
4955 * exactly final_target if we decide a projection step wouldn't be helpful.
4957 * In addition, *have_postponed_srfs is set to TRUE if we choose to postpone
4958 * any set-returning functions to after the Sort.
4961 make_sort_input_target(PlannerInfo *root,
4962 PathTarget *final_target,
4963 bool *have_postponed_srfs)
4965 Query *parse = root->parse;
4966 PathTarget *input_target;
4972 bool have_expensive;
4973 bool have_srf_sortcols;
4975 List *postponable_cols;
4976 List *postponable_vars;
4980 /* Shouldn't get here unless query has ORDER BY */
4981 Assert(parse->sortClause);
4983 *have_postponed_srfs = false; /* default result */
4985 /* Inspect tlist and collect per-column information */
4986 ncols = list_length(final_target->exprs);
4987 col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
4988 postpone_col = (bool *) palloc0(ncols * sizeof(bool));
4989 have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
4992 foreach(lc, final_target->exprs)
4994 Expr *expr = (Expr *) lfirst(lc);
4997 * If the column has a sortgroupref, assume it has to be evaluated
4998 * before sorting. Generally such columns would be ORDER BY, GROUP
4999 * BY, etc targets. One exception is columns that were removed from
5000 * GROUP BY by remove_useless_groupby_columns() ... but those would
5001 * only be Vars anyway. There don't seem to be any cases where it
5002 * would be worth the trouble to double-check.
5004 if (get_pathtarget_sortgroupref(final_target, i) == 0)
5007 * Check for SRF or volatile functions. Check the SRF case first
5008 * because we must know whether we have any postponed SRFs.
5010 if (parse->hasTargetSRFs &&
5011 expression_returns_set((Node *) expr))
5013 /* We'll decide below whether these are postponable */
5014 col_is_srf[i] = true;
5017 else if (contain_volatile_functions((Node *) expr))
5019 /* Unconditionally postpone */
5020 postpone_col[i] = true;
5021 have_volatile = true;
5026 * Else check the cost. XXX it's annoying to have to do this
5027 * when set_pathtarget_cost_width() just did it. Refactor to
5028 * allow sharing the work?
5032 cost_qual_eval_node(&cost, (Node *) expr, root);
5035 * We arbitrarily define "expensive" as "more than 10X
5036 * cpu_operator_cost". Note this will take in any PL function
5037 * with default cost.
5039 if (cost.per_tuple > 10 * cpu_operator_cost)
5041 postpone_col[i] = true;
5042 have_expensive = true;
5048 /* For sortgroupref cols, just check if any contain SRFs */
5049 if (!have_srf_sortcols &&
5050 parse->hasTargetSRFs &&
5051 expression_returns_set((Node *) expr))
5052 have_srf_sortcols = true;
5059 * We can postpone SRFs if we have some but none are in sortgroupref cols.
5061 postpone_srfs = (have_srf && !have_srf_sortcols);
5064 * If we don't need a post-sort projection, just return final_target.
5066 if (!(postpone_srfs || have_volatile ||
5068 (parse->limitCount || root->tuple_fraction > 0))))
5069 return final_target;
5072 * Report whether the post-sort projection will contain set-returning
5073 * functions. This is important because it affects whether the Sort can
5074 * rely on the query's LIMIT (if any) to bound the number of rows it needs
5077 *have_postponed_srfs = postpone_srfs;
5080 * Construct the sort-input target, taking all non-postponable columns and
5081 * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
5082 * the postponable ones.
5084 input_target = create_empty_pathtarget();
5085 postponable_cols = NIL;
5088 foreach(lc, final_target->exprs)
5090 Expr *expr = (Expr *) lfirst(lc);
5092 if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
5093 postponable_cols = lappend(postponable_cols, expr);
5095 add_column_to_pathtarget(input_target, expr,
5096 get_pathtarget_sortgroupref(final_target, i));
5102 * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
5103 * postponable columns, and add them to the sort-input target if not
5104 * already present. (Some might be there already.) We mustn't
5105 * deconstruct Aggrefs or WindowFuncs here, since the projection node
5106 * would be unable to recompute them.
5108 postponable_vars = pull_var_clause((Node *) postponable_cols,
5109 PVC_INCLUDE_AGGREGATES |
5110 PVC_INCLUDE_WINDOWFUNCS |
5111 PVC_INCLUDE_PLACEHOLDERS);
5112 add_new_columns_to_pathtarget(input_target, postponable_vars);
5114 /* clean up cruft */
5115 list_free(postponable_vars);
5116 list_free(postponable_cols);
5118 /* XXX this represents even more redundant cost calculation ... */
5119 return set_pathtarget_cost_width(root, input_target);
5123 * get_cheapest_fractional_path
5124 * Find the cheapest path for retrieving a specified fraction of all
5125 * the tuples expected to be returned by the given relation.
5127 * We interpret tuple_fraction the same way as grouping_planner.
5129 * We assume set_cheapest() has been run on the given rel.
5132 get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
5134 Path *best_path = rel->cheapest_total_path;
5137 /* If all tuples will be retrieved, just return the cheapest-total path */
5138 if (tuple_fraction <= 0.0)
5141 /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
5142 if (tuple_fraction >= 1.0 && best_path->rows > 0)
5143 tuple_fraction /= best_path->rows;
5145 foreach(l, rel->pathlist)
5147 Path *path = (Path *) lfirst(l);
5149 if (path == rel->cheapest_total_path ||
5150 compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
5160 * expression_planner
5161 * Perform planner's transformations on a standalone expression.
5163 * Various utility commands need to evaluate expressions that are not part
5164 * of a plannable query. They can do so using the executor's regular
5165 * expression-execution machinery, but first the expression has to be fed
5166 * through here to transform it from parser output to something executable.
5168 * Currently, we disallow sublinks in standalone expressions, so there's no
5169 * real "planning" involved here. (That might not always be true though.)
5170 * What we must do is run eval_const_expressions to ensure that any function
5171 * calls are converted to positional notation and function default arguments
5172 * get inserted. The fact that constant subexpressions get simplified is a
5173 * side-effect that is useful when the expression will get evaluated more than
5174 * once. Also, we must fix operator function IDs.
5176 * Note: this must not make any damaging changes to the passed-in expression
5177 * tree. (It would actually be okay to apply fix_opfuncids to it, but since
5178 * we first do an expression_tree_mutator-based walk, what is returned will
5179 * be a new node tree.)
5182 expression_planner(Expr *expr)
5187 * Convert named-argument function calls, insert default arguments and
5188 * simplify constant subexprs
5190 result = eval_const_expressions(NULL, (Node *) expr);
5192 /* Fill in opfuncid values if missing */
5193 fix_opfuncids(result);
5195 return (Expr *) result;
5200 * plan_cluster_use_sort
5201 * Use the planner to decide how CLUSTER should implement sorting
5203 * tableOid is the OID of a table to be clustered on its index indexOid
5204 * (which is already known to be a btree index). Decide whether it's
5205 * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
5206 * Return TRUE to use sorting, FALSE to use an indexscan.
5208 * Note: caller had better already hold some type of lock on the table.
5211 plan_cluster_use_sort(Oid tableOid, Oid indexOid)
5215 PlannerGlobal *glob;
5218 IndexOptInfo *indexInfo;
5219 QualCost indexExprCost;
5220 Cost comparisonCost;
5222 Path seqScanAndSortPath;
5223 IndexPath *indexScanPath;
5226 /* We can short-circuit the cost comparison if indexscans are disabled */
5227 if (!enable_indexscan)
5228 return true; /* use sort */
5230 /* Set up mostly-dummy planner state */
5231 query = makeNode(Query);
5232 query->commandType = CMD_SELECT;
5234 glob = makeNode(PlannerGlobal);
5236 root = makeNode(PlannerInfo);
5237 root->parse = query;
5239 root->query_level = 1;
5240 root->planner_cxt = CurrentMemoryContext;
5241 root->wt_param_id = -1;
5243 /* Build a minimal RTE for the rel */
5244 rte = makeNode(RangeTblEntry);
5245 rte->rtekind = RTE_RELATION;
5246 rte->relid = tableOid;
5247 rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
5248 rte->lateral = false;
5250 rte->inFromCl = true;
5251 query->rtable = list_make1(rte);
5253 /* Set up RTE/RelOptInfo arrays */
5254 setup_simple_rel_arrays(root);
5256 /* Build RelOptInfo */
5257 rel = build_simple_rel(root, 1, RELOPT_BASEREL);
5259 /* Locate IndexOptInfo for the target index */
5261 foreach(lc, rel->indexlist)
5263 indexInfo = (IndexOptInfo *) lfirst(lc);
5264 if (indexInfo->indexoid == indexOid)
5269 * It's possible that get_relation_info did not generate an IndexOptInfo
5270 * for the desired index; this could happen if it's not yet reached its
5271 * indcheckxmin usability horizon, or if it's a system index and we're
5272 * ignoring system indexes. In such cases we should tell CLUSTER to not
5273 * trust the index contents but use seqscan-and-sort.
5275 if (lc == NULL) /* not in the list? */
5276 return true; /* use sort */
5279 * Rather than doing all the pushups that would be needed to use
5280 * set_baserel_size_estimates, just do a quick hack for rows and width.
5282 rel->rows = rel->tuples;
5283 rel->reltarget->width = get_relation_data_width(tableOid, NULL);
5285 root->total_table_pages = rel->pages;
5288 * Determine eval cost of the index expressions, if any. We need to
5289 * charge twice that amount for each tuple comparison that happens during
5290 * the sort, since tuplesort.c will have to re-evaluate the index
5291 * expressions each time. (XXX that's pretty inefficient...)
5293 cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
5294 comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
5296 /* Estimate the cost of seq scan + sort */
5297 seqScanPath = create_seqscan_path(root, rel, NULL, 0);
5298 cost_sort(&seqScanAndSortPath, root, NIL,
5299 seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
5300 comparisonCost, maintenance_work_mem, -1.0);
5302 /* Estimate the cost of index scan */
5303 indexScanPath = create_index_path(root, indexInfo,
5304 NIL, NIL, NIL, NIL, NIL,
5305 ForwardScanDirection, false,
5308 return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
projects / postgresql / blob