PostgreSQL Source Code git master
planner.c
Go to the documentation of this file.
1/*-------------------------------------------------------------------------
2 *
3 * planner.c
4 * The query optimizer external interface.
5 *
6 * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 *
10 * IDENTIFICATION
11 * src/backend/optimizer/plan/planner.c
12 *
13 *-------------------------------------------------------------------------
14 */
15
16#include "postgres.h"
17
18#include <limits.h>
19#include <math.h>
20
21#include "access/genam.h"
22#include "access/parallel.h"
23#include "access/sysattr.h"
24#include "access/table.h"
25#include "catalog/pg_aggregate.h"
26#include "catalog/pg_inherits.h"
27#include "catalog/pg_proc.h"
28#include "catalog/pg_type.h"
29#include "executor/executor.h"
30#include "foreign/fdwapi.h"
31#include "jit/jit.h"
32#include "lib/bipartite_match.h"
33#include "lib/knapsack.h"
34#include "miscadmin.h"
35#include "nodes/makefuncs.h"
36#include "nodes/nodeFuncs.h"
37#ifdef OPTIMIZER_DEBUG
38#include "nodes/print.h"
39#endif
40#include "nodes/supportnodes.h"
41#include "optimizer/appendinfo.h"
42#include "optimizer/clauses.h"
43#include "optimizer/cost.h"
44#include "optimizer/optimizer.h"
45#include "optimizer/paramassign.h"
46#include "optimizer/pathnode.h"
47#include "optimizer/paths.h"
48#include "optimizer/plancat.h"
49#include "optimizer/planmain.h"
50#include "optimizer/planner.h"
51#include "optimizer/prep.h"
52#include "optimizer/subselect.h"
53#include "optimizer/tlist.h"
54#include "parser/analyze.h"
55#include "parser/parse_agg.h"
56#include "parser/parse_clause.h"
57#include "parser/parse_relation.h"
58#include "parser/parsetree.h"
59#include "partitioning/partdesc.h"
60#include "rewrite/rewriteManip.h"
61#include "utils/lsyscache.h"
62#include "utils/rel.h"
63#include "utils/selfuncs.h"
64
65/* GUC parameters */
66double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
67int debug_parallel_query = DEBUG_PARALLEL_OFF;
68bool parallel_leader_participation = true;
69bool enable_distinct_reordering = true;
70
71/* Hook for plugins to get control in planner() */
72planner_hook_type planner_hook = NULL;
73
74/* Hook for plugins to get control when grouping_planner() plans upper rels */
75create_upper_paths_hook_type create_upper_paths_hook = NULL;
76
77
78/* Expression kind codes for preprocess_expression */
79#define EXPRKIND_QUAL 0
80#define EXPRKIND_TARGET 1
81#define EXPRKIND_RTFUNC 2
82#define EXPRKIND_RTFUNC_LATERAL 3
83#define EXPRKIND_VALUES 4
84#define EXPRKIND_VALUES_LATERAL 5
85#define EXPRKIND_LIMIT 6
86#define EXPRKIND_APPINFO 7
87#define EXPRKIND_PHV 8
88#define EXPRKIND_TABLESAMPLE 9
89#define EXPRKIND_ARBITER_ELEM 10
90#define EXPRKIND_TABLEFUNC 11
91#define EXPRKIND_TABLEFUNC_LATERAL 12
92#define EXPRKIND_GROUPEXPR 13
93
94/*
95 * Data specific to grouping sets
96 */
97typedef struct
98{
99 List *rollups;
100 List *hash_sets_idx;
101 double dNumHashGroups;
102 bool any_hashable;
103 Bitmapset *unsortable_refs;
104 Bitmapset *unhashable_refs;
105 List *unsortable_sets;
106 int *tleref_to_colnum_map;
107} grouping_sets_data;
108
109/*
110 * Temporary structure for use during WindowClause reordering in order to be
111 * able to sort WindowClauses on partitioning/ordering prefix.
112 */
113typedef struct
114{
115 WindowClause *wc;
116 List *uniqueOrder; /* A List of unique ordering/partitioning
117 * clauses per Window */
118} WindowClauseSortData;
119
120/* Passthrough data for standard_qp_callback */
121typedef struct
122{
123 List *activeWindows; /* active windows, if any */
124 grouping_sets_data *gset_data; /* grouping sets data, if any */
125 SetOperationStmt *setop; /* parent set operation or NULL if not a
126 * subquery belonging to a set operation */
127} standard_qp_extra;
128
129/* Local functions */
130static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
131static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
132static void grouping_planner(PlannerInfo *root, double tuple_fraction,
133 SetOperationStmt *setops);
134static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
135static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
136 int *tleref_to_colnum_map);
137static void preprocess_rowmarks(PlannerInfo *root);
138static double preprocess_limit(PlannerInfo *root,
139 double tuple_fraction,
140 int64 *offset_est, int64 *count_est);
141static List *preprocess_groupclause(PlannerInfo *root, List *force);
142static List *extract_rollup_sets(List *groupingSets);
143static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
144static void standard_qp_callback(PlannerInfo *root, void *extra);
145static double get_number_of_groups(PlannerInfo *root,
146 double path_rows,
147 grouping_sets_data *gd,
148 List *target_list);
149static RelOptInfo *create_grouping_paths(PlannerInfo *root,
150 RelOptInfo *input_rel,
151 PathTarget *target,
152 bool target_parallel_safe,
153 grouping_sets_data *gd);
154static bool is_degenerate_grouping(PlannerInfo *root);
155static void create_degenerate_grouping_paths(PlannerInfo *root,
156 RelOptInfo *input_rel,
157 RelOptInfo *grouped_rel);
158static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
159 PathTarget *target, bool target_parallel_safe,
160 Node *havingQual);
161static void create_ordinary_grouping_paths(PlannerInfo *root,
162 RelOptInfo *input_rel,
163 RelOptInfo *grouped_rel,
164 const AggClauseCosts *agg_costs,
165 grouping_sets_data *gd,
166 GroupPathExtraData *extra,
167 RelOptInfo **partially_grouped_rel_p);
168static void consider_groupingsets_paths(PlannerInfo *root,
169 RelOptInfo *grouped_rel,
170 Path *path,
171 bool is_sorted,
172 bool can_hash,
173 grouping_sets_data *gd,
174 const AggClauseCosts *agg_costs,
175 double dNumGroups);
176static RelOptInfo *create_window_paths(PlannerInfo *root,
177 RelOptInfo *input_rel,
178 PathTarget *input_target,
179 PathTarget *output_target,
180 bool output_target_parallel_safe,
181 WindowFuncLists *wflists,
182 List *activeWindows);
183static void create_one_window_path(PlannerInfo *root,
184 RelOptInfo *window_rel,
185 Path *path,
186 PathTarget *input_target,
187 PathTarget *output_target,
188 WindowFuncLists *wflists,
189 List *activeWindows);
190static RelOptInfo *create_distinct_paths(PlannerInfo *root,
191 RelOptInfo *input_rel,
192 PathTarget *target);
193static void create_partial_distinct_paths(PlannerInfo *root,
194 RelOptInfo *input_rel,
195 RelOptInfo *final_distinct_rel,
196 PathTarget *target);
197static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
198 RelOptInfo *input_rel,
199 RelOptInfo *distinct_rel);
200static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
201 List *needed_pathkeys,
202 List *path_pathkeys);
203static RelOptInfo *create_ordered_paths(PlannerInfo *root,
204 RelOptInfo *input_rel,
205 PathTarget *target,
206 bool target_parallel_safe,
207 double limit_tuples);
208static PathTarget *make_group_input_target(PlannerInfo *root,
209 PathTarget *final_target);
210static PathTarget *make_partial_grouping_target(PlannerInfo *root,
211 PathTarget *grouping_target,
212 Node *havingQual);
213static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
214static void optimize_window_clauses(PlannerInfo *root,
215 WindowFuncLists *wflists);
216static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
217static PathTarget *make_window_input_target(PlannerInfo *root,
218 PathTarget *final_target,
219 List *activeWindows);
220static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
221 List *tlist);
222static PathTarget *make_sort_input_target(PlannerInfo *root,
223 PathTarget *final_target,
224 bool *have_postponed_srfs);
225static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
226 List *targets, List *targets_contain_srfs);
227static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
228 RelOptInfo *grouped_rel,
229 RelOptInfo *partially_grouped_rel,
230 const AggClauseCosts *agg_costs,
231 grouping_sets_data *gd,
232 double dNumGroups,
233 GroupPathExtraData *extra);
234static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
235 RelOptInfo *grouped_rel,
236 RelOptInfo *input_rel,
237 grouping_sets_data *gd,
238 GroupPathExtraData *extra,
239 bool force_rel_creation);
240static Path *make_ordered_path(PlannerInfo *root,
241 RelOptInfo *rel,
242 Path *path,
243 Path *cheapest_path,
244 List *pathkeys,
245 double limit_tuples);
246static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
247static bool can_partial_agg(PlannerInfo *root);
248static void apply_scanjoin_target_to_paths(PlannerInfo *root,
249 RelOptInfo *rel,
250 List *scanjoin_targets,
251 List *scanjoin_targets_contain_srfs,
252 bool scanjoin_target_parallel_safe,
253 bool tlist_same_exprs);
254static void create_partitionwise_grouping_paths(PlannerInfo *root,
255 RelOptInfo *input_rel,
256 RelOptInfo *grouped_rel,
257 RelOptInfo *partially_grouped_rel,
258 const AggClauseCosts *agg_costs,
259 grouping_sets_data *gd,
260 PartitionwiseAggregateType patype,
261 GroupPathExtraData *extra);
262static bool group_by_has_partkey(RelOptInfo *input_rel,
263 List *targetList,
264 List *groupClause);
265static int common_prefix_cmp(const void *a, const void *b);
266static List *generate_setop_child_grouplist(SetOperationStmt *op,
267 List *targetlist);
268
269
270/*****************************************************************************
271 *
272 * Query optimizer entry point
273 *
274 * To support loadable plugins that monitor or modify planner behavior,
275 * we provide a hook variable that lets a plugin get control before and
276 * after the standard planning process. The plugin would normally call
277 * standard_planner().
278 *
279 * Note to plugin authors: standard_planner() scribbles on its Query input,
280 * so you'd better copy that data structure if you want to plan more than once.
281 *
282 *****************************************************************************/
283PlannedStmt *
284planner(Query *parse, const char *query_string, int cursorOptions,
285 ParamListInfo boundParams)
286{
287 PlannedStmt *result;
288
289 if (planner_hook)
290 result = (*planner_hook) (parse, query_string, cursorOptions, boundParams);
291 else
292 result = standard_planner(parse, query_string, cursorOptions, boundParams);
293 return result;
294}
295
296PlannedStmt *
297standard_planner(Query *parse, const char *query_string, int cursorOptions,
298 ParamListInfo boundParams)
299{
300 PlannedStmt *result;
301 PlannerGlobal *glob;
302 double tuple_fraction;
303 PlannerInfo *root;
304 RelOptInfo *final_rel;
305 Path *best_path;
306 Plan *top_plan;
307 ListCell *lp,
308 *lr;
309
310 /*
311 * Set up global state for this planner invocation. This data is needed
312 * across all levels of sub-Query that might exist in the given command,
313 * so we keep it in a separate struct that's linked to by each per-Query
314 * PlannerInfo.
315 */
316 glob = makeNode(PlannerGlobal);
317
318 glob->boundParams = boundParams;
319 glob->subplans = NIL;
320 glob->subpaths = NIL;
321 glob->subroots = NIL;
322 glob->rewindPlanIDs = NULL;
323 glob->finalrtable = NIL;
324 glob->finalrteperminfos = NIL;
325 glob->finalrowmarks = NIL;
326 glob->resultRelations = NIL;
327 glob->appendRelations = NIL;
328 glob->relationOids = NIL;
329 glob->invalItems = NIL;
330 glob->paramExecTypes = NIL;
331 glob->lastPHId = 0;
332 glob->lastRowMarkId = 0;
333 glob->lastPlanNodeId = 0;
334 glob->transientPlan = false;
335 glob->dependsOnRole = false;
336
337 /*
338 * Assess whether it's feasible to use parallel mode for this query. We
339 * can't do this in a standalone backend, or if the command will try to
340 * modify any data, or if this is a cursor operation, or if GUCs are set
341 * to values that don't permit parallelism, or if parallel-unsafe
342 * functions are present in the query tree.
343 *
344 * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
345 * MATERIALIZED VIEW to use parallel plans, but this is safe only because
346 * the command is writing into a completely new table which workers won't
347 * be able to see. If the workers could see the table, the fact that
348 * group locking would cause them to ignore the leader's heavyweight GIN
349 * page locks would make this unsafe. We'll have to fix that somehow if
350 * we want to allow parallel inserts in general; updates and deletes have
351 * additional problems especially around combo CIDs.)
352 *
353 * For now, we don't try to use parallel mode if we're running inside a
354 * parallel worker. We might eventually be able to relax this
355 * restriction, but for now it seems best not to have parallel workers
356 * trying to create their own parallel workers.
357 */
358 if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
359 IsUnderPostmaster &&
360 parse->commandType == CMD_SELECT &&
361 !parse->hasModifyingCTE &&
362 max_parallel_workers_per_gather > 0 &&
363 !IsParallelWorker())
364 {
365 /* all the cheap tests pass, so scan the query tree */
366 glob->maxParallelHazard = max_parallel_hazard(parse);
367 glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
368 }
369 else
370 {
371 /* skip the query tree scan, just assume it's unsafe */
372 glob->maxParallelHazard = PROPARALLEL_UNSAFE;
373 glob->parallelModeOK = false;
374 }
375
376 /*
377 * glob->parallelModeNeeded is normally set to false here and changed to
378 * true during plan creation if a Gather or Gather Merge plan is actually
379 * created (cf. create_gather_plan, create_gather_merge_plan).
380 *
381 * However, if debug_parallel_query = on or debug_parallel_query =
382 * regress, then we impose parallel mode whenever it's safe to do so, even
383 * if the final plan doesn't use parallelism. It's not safe to do so if
384 * the query contains anything parallel-unsafe; parallelModeOK will be
385 * false in that case. Note that parallelModeOK can't change after this
386 * point. Otherwise, everything in the query is either parallel-safe or
387 * parallel-restricted, and in either case it should be OK to impose
388 * parallel-mode restrictions. If that ends up breaking something, then
389 * either some function the user included in the query is incorrectly
390 * labeled as parallel-safe or parallel-restricted when in reality it's
391 * parallel-unsafe, or else the query planner itself has a bug.
392 */
393 glob->parallelModeNeeded = glob->parallelModeOK &&
394 (debug_parallel_query != DEBUG_PARALLEL_OFF);
395
396 /* Determine what fraction of the plan is likely to be scanned */
397 if (cursorOptions & CURSOR_OPT_FAST_PLAN)
398 {
399 /*
400 * We have no real idea how many tuples the user will ultimately FETCH
401 * from a cursor, but it is often the case that he doesn't want 'em
402 * all, or would prefer a fast-start plan anyway so that he can
403 * process some of the tuples sooner. Use a GUC parameter to decide
404 * what fraction to optimize for.
405 */
406 tuple_fraction = cursor_tuple_fraction;
407
408 /*
409 * We document cursor_tuple_fraction as simply being a fraction, which
410 * means the edge cases 0 and 1 have to be treated specially here. We
411 * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
412 */
413 if (tuple_fraction >= 1.0)
414 tuple_fraction = 0.0;
415 else if (tuple_fraction <= 0.0)
416 tuple_fraction = 1e-10;
417 }
418 else
419 {
420 /* Default assumption is we need all the tuples */
421 tuple_fraction = 0.0;
422 }
423
424 /* primary planning entry point (may recurse for subqueries) */
425 root = subquery_planner(glob, parse, NULL, false, tuple_fraction, NULL);
426
427 /* Select best Path and turn it into a Plan */
428 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
429 best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
430
431 top_plan = create_plan(root, best_path);
432
433 /*
434 * If creating a plan for a scrollable cursor, make sure it can run
435 * backwards on demand. Add a Material node at the top at need.
436 */
437 if (cursorOptions & CURSOR_OPT_SCROLL)
438 {
439 if (!ExecSupportsBackwardScan(top_plan))
440 top_plan = materialize_finished_plan(top_plan);
441 }
442
443 /*
444 * Optionally add a Gather node for testing purposes, provided this is
445 * actually a safe thing to do.
446 *
447 * We can add Gather even when top_plan has parallel-safe initPlans, but
448 * then we have to move the initPlans to the Gather node because of
449 * SS_finalize_plan's limitations. That would cause cosmetic breakage of
450 * regression tests when debug_parallel_query = regress, because initPlans
451 * that would normally appear on the top_plan move to the Gather, causing
452 * them to disappear from EXPLAIN output. That doesn't seem worth kluging
453 * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
454 */
455 if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
456 top_plan->parallel_safe &&
457 (top_plan->initPlan == NIL ||
458 debug_parallel_query != DEBUG_PARALLEL_REGRESS))
459 {
460 Gather *gather = makeNode(Gather);
461 Cost initplan_cost;
462 bool unsafe_initplans;
463
464 gather->plan.targetlist = top_plan->targetlist;
465 gather->plan.qual = NIL;
466 gather->plan.lefttree = top_plan;
467 gather->plan.righttree = NULL;
468 gather->num_workers = 1;
469 gather->single_copy = true;
470 gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
471
472 /* Transfer any initPlans to the new top node */
473 gather->plan.initPlan = top_plan->initPlan;
474 top_plan->initPlan = NIL;
475
476 /*
477 * Since this Gather has no parallel-aware descendants to signal to,
478 * we don't need a rescan Param.
479 */
480 gather->rescan_param = -1;
481
482 /*
483 * Ideally we'd use cost_gather here, but setting up dummy path data
484 * to satisfy it doesn't seem much cleaner than knowing what it does.
485 */
486 gather->plan.startup_cost = top_plan->startup_cost +
487 parallel_setup_cost;
488 gather->plan.total_cost = top_plan->total_cost +
489 parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
490 gather->plan.plan_rows = top_plan->plan_rows;
491 gather->plan.plan_width = top_plan->plan_width;
492 gather->plan.parallel_aware = false;
493 gather->plan.parallel_safe = false;
494
495 /*
496 * Delete the initplans' cost from top_plan. We needn't add it to the
497 * Gather node, since the above coding already included it there.
498 */
499 SS_compute_initplan_cost(gather->plan.initPlan,
500 &initplan_cost, &unsafe_initplans);
501 top_plan->startup_cost -= initplan_cost;
502 top_plan->total_cost -= initplan_cost;
503
504 /* use parallel mode for parallel plans. */
505 root->glob->parallelModeNeeded = true;
506
507 top_plan = &gather->plan;
508 }
509
510 /*
511 * If any Params were generated, run through the plan tree and compute
512 * each plan node's extParam/allParam sets. Ideally we'd merge this into
513 * set_plan_references' tree traversal, but for now it has to be separate
514 * because we need to visit subplans before not after main plan.
515 */
516 if (glob->paramExecTypes != NIL)
517 {
518 Assert(list_length(glob->subplans) == list_length(glob->subroots));
519 forboth(lp, glob->subplans, lr, glob->subroots)
520 {
521 Plan *subplan = (Plan *) lfirst(lp);
522 PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
523
524 SS_finalize_plan(subroot, subplan);
525 }
526 SS_finalize_plan(root, top_plan);
527 }
528
529 /* final cleanup of the plan */
530 Assert(glob->finalrtable == NIL);
531 Assert(glob->finalrteperminfos == NIL);
532 Assert(glob->finalrowmarks == NIL);
533 Assert(glob->resultRelations == NIL);
534 Assert(glob->appendRelations == NIL);
535 top_plan = set_plan_references(root, top_plan);
536 /* ... and the subplans (both regular subplans and initplans) */
537 Assert(list_length(glob->subplans) == list_length(glob->subroots));
538 forboth(lp, glob->subplans, lr, glob->subroots)
539 {
540 Plan *subplan = (Plan *) lfirst(lp);
541 PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
542
543 lfirst(lp) = set_plan_references(subroot, subplan);
544 }
545
546 /* build the PlannedStmt result */
547 result = makeNode(PlannedStmt);
548
549 result->commandType = parse->commandType;
550 result->queryId = parse->queryId;
551 result->hasReturning = (parse->returningList != NIL);
552 result->hasModifyingCTE = parse->hasModifyingCTE;
553 result->canSetTag = parse->canSetTag;
554 result->transientPlan = glob->transientPlan;
555 result->dependsOnRole = glob->dependsOnRole;
556 result->parallelModeNeeded = glob->parallelModeNeeded;
557 result->planTree = top_plan;
558 result->rtable = glob->finalrtable;
559 result->permInfos = glob->finalrteperminfos;
560 result->resultRelations = glob->resultRelations;
561 result->appendRelations = glob->appendRelations;
562 result->subplans = glob->subplans;
563 result->rewindPlanIDs = glob->rewindPlanIDs;
564 result->rowMarks = glob->finalrowmarks;
565 result->relationOids = glob->relationOids;
566 result->invalItems = glob->invalItems;
567 result->paramExecTypes = glob->paramExecTypes;
568 /* utilityStmt should be null, but we might as well copy it */
569 result->utilityStmt = parse->utilityStmt;
570 result->stmt_location = parse->stmt_location;
571 result->stmt_len = parse->stmt_len;
572
573 result->jitFlags = PGJIT_NONE;
574 if (jit_enabled && jit_above_cost >= 0 &&
575 top_plan->total_cost > jit_above_cost)
576 {
577 result->jitFlags |= PGJIT_PERFORM;
578
579 /*
580 * Decide how much effort should be put into generating better code.
581 */
582 if (jit_optimize_above_cost >= 0 &&
583 top_plan->total_cost > jit_optimize_above_cost)
584 result->jitFlags |= PGJIT_OPT3;
585 if (jit_inline_above_cost >= 0 &&
586 top_plan->total_cost > jit_inline_above_cost)
587 result->jitFlags |= PGJIT_INLINE;
588
589 /*
590 * Decide which operations should be JITed.
591 */
592 if (jit_expressions)
593 result->jitFlags |= PGJIT_EXPR;
594 if (jit_tuple_deforming)
595 result->jitFlags |= PGJIT_DEFORM;
596 }
597
598 if (glob->partition_directory != NULL)
599 DestroyPartitionDirectory(glob->partition_directory);
600
601 return result;
602}
603
604
605/*--------------------
606 * subquery_planner
607 * Invokes the planner on a subquery. We recurse to here for each
608 * sub-SELECT found in the query tree.
609 *
610 * glob is the global state for the current planner run.
611 * parse is the querytree produced by the parser & rewriter.
612 * parent_root is the immediate parent Query's info (NULL at the top level).
613 * hasRecursion is true if this is a recursive WITH query.
614 * tuple_fraction is the fraction of tuples we expect will be retrieved.
615 * tuple_fraction is interpreted as explained for grouping_planner, below.
616 * setops is used for set operation subqueries to provide the subquery with
617 * the context in which it's being used so that Paths correctly sorted for the
618 * set operation can be generated. NULL when not planning a set operation
619 * child, or when a child of a set op that isn't interested in sorted input.
620 *
621 * Basically, this routine does the stuff that should only be done once
622 * per Query object. It then calls grouping_planner. At one time,
623 * grouping_planner could be invoked recursively on the same Query object;
624 * that's not currently true, but we keep the separation between the two
625 * routines anyway, in case we need it again someday.
626 *
627 * subquery_planner will be called recursively to handle sub-Query nodes
628 * found within the query's expressions and rangetable.
629 *
630 * Returns the PlannerInfo struct ("root") that contains all data generated
631 * while planning the subquery. In particular, the Path(s) attached to
632 * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
633 * cheapest way(s) to implement the query. The top level will select the
634 * best Path and pass it through createplan.c to produce a finished Plan.
635 *--------------------
636 */
637PlannerInfo *
638subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root,
639 bool hasRecursion, double tuple_fraction,
640 SetOperationStmt *setops)
641{
642 PlannerInfo *root;
643 List *newWithCheckOptions;
644 List *newHaving;
645 bool hasOuterJoins;
646 bool hasResultRTEs;
647 RelOptInfo *final_rel;
648 ListCell *l;
649
650 /* Create a PlannerInfo data structure for this subquery */
651 root = makeNode(PlannerInfo);
652 root->parse = parse;
653 root->glob = glob;
654 root->query_level = parent_root ? parent_root->query_level + 1 : 1;
655 root->parent_root = parent_root;
656 root->plan_params = NIL;
657 root->outer_params = NULL;
658 root->planner_cxt = CurrentMemoryContext;
659 root->init_plans = NIL;
660 root->cte_plan_ids = NIL;
661 root->multiexpr_params = NIL;
662 root->join_domains = NIL;
663 root->eq_classes = NIL;
664 root->ec_merging_done = false;
665 root->last_rinfo_serial = 0;
666 root->all_result_relids =
667 parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
668 root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
669 root->append_rel_list = NIL;
670 root->row_identity_vars = NIL;
671 root->rowMarks = NIL;
672 memset(root->upper_rels, 0, sizeof(root->upper_rels));
673 memset(root->upper_targets, 0, sizeof(root->upper_targets));
674 root->processed_groupClause = NIL;
675 root->processed_distinctClause = NIL;
676 root->processed_tlist = NIL;
677 root->update_colnos = NIL;
678 root->grouping_map = NULL;
679 root->minmax_aggs = NIL;
680 root->qual_security_level = 0;
681 root->hasPseudoConstantQuals = false;
682 root->hasAlternativeSubPlans = false;
683 root->placeholdersFrozen = false;
684 root->hasRecursion = hasRecursion;
685 if (hasRecursion)
686 root->wt_param_id = assign_special_exec_param(root);
687 else
688 root->wt_param_id = -1;
689 root->non_recursive_path = NULL;
690 root->partColsUpdated = false;
691
692 /*
693 * Create the top-level join domain. This won't have valid contents until
694 * deconstruct_jointree fills it in, but the node needs to exist before
695 * that so we can build EquivalenceClasses referencing it.
696 */
697 root->join_domains = list_make1(makeNode(JoinDomain));
698
699 /*
700 * If there is a WITH list, process each WITH query and either convert it
701 * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
702 */
703 if (parse->cteList)
704 SS_process_ctes(root);
705
706 /*
707 * If it's a MERGE command, transform the joinlist as appropriate.
708 */
709 transform_MERGE_to_join(parse);
710
711 /*
712 * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
713 * that we don't need so many special cases to deal with that situation.
714 */
715 replace_empty_jointree(parse);
716
717 /*
718 * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
719 * to transform them into joins. Note that this step does not descend
720 * into subqueries; if we pull up any subqueries below, their SubLinks are
721 * processed just before pulling them up.
722 */
723 if (parse->hasSubLinks)
724 pull_up_sublinks(root);
725
726 /*
727 * Scan the rangetable for function RTEs, do const-simplification on them,
728 * and then inline them if possible (producing subqueries that might get
729 * pulled up next). Recursion issues here are handled in the same way as
730 * for SubLinks.
731 */
732 preprocess_function_rtes(root);
733
734 /*
735 * Check to see if any subqueries in the jointree can be merged into this
736 * query.
737 */
738 pull_up_subqueries(root);
739
740 /*
741 * If this is a simple UNION ALL query, flatten it into an appendrel. We
742 * do this now because it requires applying pull_up_subqueries to the leaf
743 * queries of the UNION ALL, which weren't touched above because they
744 * weren't referenced by the jointree (they will be after we do this).
745 */
746 if (parse->setOperations)
747 flatten_simple_union_all(root);
748
749 /*
750 * Survey the rangetable to see what kinds of entries are present. We can
751 * skip some later processing if relevant SQL features are not used; for
752 * example if there are no JOIN RTEs we can avoid the expense of doing
753 * flatten_join_alias_vars(). This must be done after we have finished
754 * adding rangetable entries, of course. (Note: actually, processing of
755 * inherited or partitioned rels can cause RTEs for their child tables to
756 * get added later; but those must all be RTE_RELATION entries, so they
757 * don't invalidate the conclusions drawn here.)
758 */
759 root->hasJoinRTEs = false;
760 root->hasLateralRTEs = false;
761 root->group_rtindex = 0;
762 hasOuterJoins = false;
763 hasResultRTEs = false;
764 foreach(l, parse->rtable)
765 {
766 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
767
768 switch (rte->rtekind)
769 {
770 case RTE_RELATION:
771 if (rte->inh)
772 {
773 /*
774 * Check to see if the relation actually has any children;
775 * if not, clear the inh flag so we can treat it as a
776 * plain base relation.
777 *
778 * Note: this could give a false-positive result, if the
779 * rel once had children but no longer does. We used to
780 * be able to clear rte->inh later on when we discovered
781 * that, but no more; we have to handle such cases as
782 * full-fledged inheritance.
783 */
784 rte->inh = has_subclass(rte->relid);
785 }
786 break;
787 case RTE_JOIN:
788 root->hasJoinRTEs = true;
789 if (IS_OUTER_JOIN(rte->jointype))
790 hasOuterJoins = true;
791 break;
792 case RTE_RESULT:
793 hasResultRTEs = true;
794 break;
795 case RTE_GROUP:
796 Assert(parse->hasGroupRTE);
797 root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
798 break;
799 default:
800 /* No work here for other RTE types */
801 break;
802 }
803
804 if (rte->lateral)
805 root->hasLateralRTEs = true;
806
807 /*
808 * We can also determine the maximum security level required for any
809 * securityQuals now. Addition of inheritance-child RTEs won't affect
810 * this, because child tables don't have their own securityQuals; see
811 * expand_single_inheritance_child().
812 */
813 if (rte->securityQuals)
814 root->qual_security_level = Max(root->qual_security_level,
815 list_length(rte->securityQuals));
816 }
817
818 /*
819 * If we have now verified that the query target relation is
820 * non-inheriting, mark it as a leaf target.
821 */
822 if (parse->resultRelation)
823 {
824 RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
825
826 if (!rte->inh)
827 root->leaf_result_relids =
828 bms_make_singleton(parse->resultRelation);
829 }
830
831 /*
832 * Preprocess RowMark information. We need to do this after subquery
833 * pullup, so that all base relations are present.
834 */
835 preprocess_rowmarks(root);
836
837 /*
838 * Set hasHavingQual to remember if HAVING clause is present. Needed
839 * because preprocess_expression will reduce a constant-true condition to
840 * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
841 */
842 root->hasHavingQual = (parse->havingQual != NULL);
843
844 /*
845 * Do expression preprocessing on targetlist and quals, as well as other
846 * random expressions in the querytree. Note that we do not need to
847 * handle sort/group expressions explicitly, because they are actually
848 * part of the targetlist.
849 */
850 parse->targetList = (List *)
851 preprocess_expression(root, (Node *) parse->targetList,
852 EXPRKIND_TARGET);
853
854 newWithCheckOptions = NIL;
855 foreach(l, parse->withCheckOptions)
856 {
857 WithCheckOption *wco = lfirst_node(WithCheckOption, l);
858
859 wco->qual = preprocess_expression(root, wco->qual,
860 EXPRKIND_QUAL);
861 if (wco->qual != NULL)
862 newWithCheckOptions = lappend(newWithCheckOptions, wco);
863 }
864 parse->withCheckOptions = newWithCheckOptions;
865
866 parse->returningList = (List *)
867 preprocess_expression(root, (Node *) parse->returningList,
868 EXPRKIND_TARGET);
869
870 preprocess_qual_conditions(root, (Node *) parse->jointree);
871
872 parse->havingQual = preprocess_expression(root, parse->havingQual,
873 EXPRKIND_QUAL);
874
875 foreach(l, parse->windowClause)
876 {
877 WindowClause *wc = lfirst_node(WindowClause, l);
878
879 /* partitionClause/orderClause are sort/group expressions */
880 wc->startOffset = preprocess_expression(root, wc->startOffset,
881 EXPRKIND_LIMIT);
882 wc->endOffset = preprocess_expression(root, wc->endOffset,
883 EXPRKIND_LIMIT);
884 }
885
886 parse->limitOffset = preprocess_expression(root, parse->limitOffset,
887 EXPRKIND_LIMIT);
888 parse->limitCount = preprocess_expression(root, parse->limitCount,
889 EXPRKIND_LIMIT);
890
891 if (parse->onConflict)
892 {
893 parse->onConflict->arbiterElems = (List *)
894 preprocess_expression(root,
895 (Node *) parse->onConflict->arbiterElems,
896 EXPRKIND_ARBITER_ELEM);
897 parse->onConflict->arbiterWhere =
898 preprocess_expression(root,
899 parse->onConflict->arbiterWhere,
900 EXPRKIND_QUAL);
901 parse->onConflict->onConflictSet = (List *)
902 preprocess_expression(root,
903 (Node *) parse->onConflict->onConflictSet,
904 EXPRKIND_TARGET);
905 parse->onConflict->onConflictWhere =
906 preprocess_expression(root,
907 parse->onConflict->onConflictWhere,
908 EXPRKIND_QUAL);
909 /* exclRelTlist contains only Vars, so no preprocessing needed */
910 }
911
912 foreach(l, parse->mergeActionList)
913 {
914 MergeAction *action = (MergeAction *) lfirst(l);
915
916 action->targetList = (List *)
917 preprocess_expression(root,
918 (Node *) action->targetList,
919 EXPRKIND_TARGET);
920 action->qual =
921 preprocess_expression(root,
922 (Node *) action->qual,
923 EXPRKIND_QUAL);
924 }
925
926 parse->mergeJoinCondition =
927 preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
928
929 root->append_rel_list = (List *)
930 preprocess_expression(root, (Node *) root->append_rel_list,
931 EXPRKIND_APPINFO);
932
933 /* Also need to preprocess expressions within RTEs */
934 foreach(l, parse->rtable)
935 {
936 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
937 int kind;
938 ListCell *lcsq;
939
940 if (rte->rtekind == RTE_RELATION)
941 {
942 if (rte->tablesample)
943 rte->tablesample = (TableSampleClause *)
944 preprocess_expression(root,
945 (Node *) rte->tablesample,
946 EXPRKIND_TABLESAMPLE);
947 }
948 else if (rte->rtekind == RTE_SUBQUERY)
949 {
950 /*
951 * We don't want to do all preprocessing yet on the subquery's
952 * expressions, since that will happen when we plan it. But if it
953 * contains any join aliases of our level, those have to get
954 * expanded now, because planning of the subquery won't do it.
955 * That's only possible if the subquery is LATERAL.
956 */
957 if (rte->lateral && root->hasJoinRTEs)
958 rte->subquery = (Query *)
959 flatten_join_alias_vars(root, root->parse,
960 (Node *) rte->subquery);
961 }
962 else if (rte->rtekind == RTE_FUNCTION)
963 {
964 /* Preprocess the function expression(s) fully */
965 kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
966 rte->functions = (List *)
967 preprocess_expression(root, (Node *) rte->functions, kind);
968 }
969 else if (rte->rtekind == RTE_TABLEFUNC)
970 {
971 /* Preprocess the function expression(s) fully */
972 kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
973 rte->tablefunc = (TableFunc *)
974 preprocess_expression(root, (Node *) rte->tablefunc, kind);
975 }
976 else if (rte->rtekind == RTE_VALUES)
977 {
978 /* Preprocess the values lists fully */
979 kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
980 rte->values_lists = (List *)
981 preprocess_expression(root, (Node *) rte->values_lists, kind);
982 }
983 else if (rte->rtekind == RTE_GROUP)
984 {
985 /* Preprocess the groupexprs list fully */
986 rte->groupexprs = (List *)
987 preprocess_expression(root, (Node *) rte->groupexprs,
988 EXPRKIND_GROUPEXPR);
989 }
990
991 /*
992 * Process each element of the securityQuals list as if it were a
993 * separate qual expression (as indeed it is). We need to do it this
994 * way to get proper canonicalization of AND/OR structure. Note that
995 * this converts each element into an implicit-AND sublist.
996 */
997 foreach(lcsq, rte->securityQuals)
998 {
999 lfirst(lcsq) = preprocess_expression(root,
1000 (Node *) lfirst(lcsq),
1001 EXPRKIND_QUAL);
1002 }
1003 }
1004
1005 /*
1006 * Now that we are done preprocessing expressions, and in particular done
1007 * flattening join alias variables, get rid of the joinaliasvars lists.
1008 * They no longer match what expressions in the rest of the tree look
1009 * like, because we have not preprocessed expressions in those lists (and
1010 * do not want to; for example, expanding a SubLink there would result in
1011 * a useless unreferenced subplan). Leaving them in place simply creates
1012 * a hazard for later scans of the tree. We could try to prevent that by
1013 * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1014 * but that doesn't sound very reliable.
1015 */
1016 if (root->hasJoinRTEs)
1017 {
1018 foreach(l, parse->rtable)
1019 {
1020 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1021
1022 rte->joinaliasvars = NIL;
1023 }
1024 }
1025
1026 /*
1027 * Replace any Vars in the subquery's targetlist and havingQual that
1028 * reference GROUP outputs with the underlying grouping expressions.
1029 *
1030 * Note that we need to perform this replacement after we've preprocessed
1031 * the grouping expressions. This is to ensure that there is only one
1032 * instance of SubPlan for each SubLink contained within the grouping
1033 * expressions.
1034 */
1035 if (parse->hasGroupRTE)
1036 {
1037 parse->targetList = (List *)
1038 flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1039 parse->havingQual =
1040 flatten_group_exprs(root, root->parse, parse->havingQual);
1041 }
1042
1043 /* Constant-folding might have removed all set-returning functions */
1044 if (parse->hasTargetSRFs)
1045 parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1046
1047 /*
1048 * In some cases we may want to transfer a HAVING clause into WHERE. We
1049 * cannot do so if the HAVING clause contains aggregates (obviously) or
1050 * volatile functions (since a HAVING clause is supposed to be executed
1051 * only once per group). We also can't do this if there are any nonempty
1052 * grouping sets and the clause references any columns that are nullable
1053 * by the grouping sets; moving such a clause into WHERE would potentially
1054 * change the results. (If there are only empty grouping sets, then the
1055 * HAVING clause must be degenerate as discussed below.)
1056 *
1057 * Also, it may be that the clause is so expensive to execute that we're
1058 * better off doing it only once per group, despite the loss of
1059 * selectivity. This is hard to estimate short of doing the entire
1060 * planning process twice, so we use a heuristic: clauses containing
1061 * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1062 * clause into WHERE, in hopes of eliminating tuples before aggregation
1063 * instead of after.
1064 *
1065 * If the query has explicit grouping then we can simply move such a
1066 * clause into WHERE; any group that fails the clause will not be in the
1067 * output because none of its tuples will reach the grouping or
1068 * aggregation stage. Otherwise we must have a degenerate (variable-free)
1069 * HAVING clause, which we put in WHERE so that query_planner() can use it
1070 * in a gating Result node, but also keep in HAVING to ensure that we
1071 * don't emit a bogus aggregated row. (This could be done better, but it
1072 * seems not worth optimizing.)
1073 *
1074 * Note that a HAVING clause may contain expressions that are not fully
1075 * preprocessed. This can happen if these expressions are part of
1076 * grouping items. In such cases, they are replaced with GROUP Vars in
1077 * the parser and then replaced back after we've done with expression
1078 * preprocessing on havingQual. This is not an issue if the clause
1079 * remains in HAVING, because these expressions will be matched to lower
1080 * target items in setrefs.c. However, if the clause is moved or copied
1081 * into WHERE, we need to ensure that these expressions are fully
1082 * preprocessed.
1083 *
1084 * Note that both havingQual and parse->jointree->quals are in
1085 * implicitly-ANDed-list form at this point, even though they are declared
1086 * as Node *.
1087 */
1088 newHaving = NIL;
1089 foreach(l, (List *) parse->havingQual)
1090 {
1091 Node *havingclause = (Node *) lfirst(l);
1092
1093 if (contain_agg_clause(havingclause) ||
1094 contain_volatile_functions(havingclause) ||
1095 contain_subplans(havingclause) ||
1096 (parse->groupClause && parse->groupingSets &&
1097 bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1098 {
1099 /* keep it in HAVING */
1100 newHaving = lappend(newHaving, havingclause);
1101 }
1102 else if (parse->groupClause)
1103 {
1104 Node *whereclause;
1105
1106 /* Preprocess the HAVING clause fully */
1107 whereclause = preprocess_expression(root, havingclause,
1108 EXPRKIND_QUAL);
1109 /* ... and move it to WHERE */
1110 parse->jointree->quals = (Node *)
1111 list_concat((List *) parse->jointree->quals,
1112 (List *) whereclause);
1113 }
1114 else
1115 {
1116 Node *whereclause;
1117
1118 /* Preprocess the HAVING clause fully */
1119 whereclause = preprocess_expression(root, copyObject(havingclause),
1120 EXPRKIND_QUAL);
1121 /* ... and put a copy in WHERE */
1122 parse->jointree->quals = (Node *)
1123 list_concat((List *) parse->jointree->quals,
1124 (List *) whereclause);
1125 /* ... and also keep it in HAVING */
1126 newHaving = lappend(newHaving, havingclause);
1127 }
1128 }
1129 parse->havingQual = (Node *) newHaving;
1130
1131 /*
1132 * If we have any outer joins, try to reduce them to plain inner joins.
1133 * This step is most easily done after we've done expression
1134 * preprocessing.
1135 */
1136 if (hasOuterJoins)
1137 reduce_outer_joins(root);
1138
1139 /*
1140 * If we have any RTE_RESULT relations, see if they can be deleted from
1141 * the jointree. We also rely on this processing to flatten single-child
1142 * FromExprs underneath outer joins. This step is most effectively done
1143 * after we've done expression preprocessing and outer join reduction.
1144 */
1145 if (hasResultRTEs || hasOuterJoins)
1146 remove_useless_result_rtes(root);
1147
1148 /*
1149 * Do the main planning.
1150 */
1151 grouping_planner(root, tuple_fraction, setops);
1152
1153 /*
1154 * Capture the set of outer-level param IDs we have access to, for use in
1155 * extParam/allParam calculations later.
1156 */
1157 SS_identify_outer_params(root);
1158
1159 /*
1160 * If any initPlans were created in this query level, adjust the surviving
1161 * Paths' costs and parallel-safety flags to account for them. The
1162 * initPlans won't actually get attached to the plan tree till
1163 * create_plan() runs, but we must include their effects now.
1164 */
1165 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1166 SS_charge_for_initplans(root, final_rel);
1167
1168 /*
1169 * Make sure we've identified the cheapest Path for the final rel. (By
1170 * doing this here not in grouping_planner, we include initPlan costs in
1171 * the decision, though it's unlikely that will change anything.)
1172 */
1173 set_cheapest(final_rel);
1174
1175 return root;
1176}
1177
1178/*
1179 * preprocess_expression
1180 * Do subquery_planner's preprocessing work for an expression,
1181 * which can be a targetlist, a WHERE clause (including JOIN/ON
1182 * conditions), a HAVING clause, or a few other things.
1183 */
1184static Node *
1185preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1186{
1187 /*
1188 * Fall out quickly if expression is empty. This occurs often enough to
1189 * be worth checking. Note that null->null is the correct conversion for
1190 * implicit-AND result format, too.
1191 */
1192 if (expr == NULL)
1193 return NULL;
1194
1195 /*
1196 * If the query has any join RTEs, replace join alias variables with
1197 * base-relation variables. We must do this first, since any expressions
1198 * we may extract from the joinaliasvars lists have not been preprocessed.
1199 * For example, if we did this after sublink processing, sublinks expanded
1200 * out from join aliases would not get processed. But we can skip this in
1201 * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1202 * they can't contain any Vars of the current query level.
1203 */
1204 if (root->hasJoinRTEs &&
1205 !(kind == EXPRKIND_RTFUNC ||
1206 kind == EXPRKIND_VALUES ||
1207 kind == EXPRKIND_TABLESAMPLE ||
1208 kind == EXPRKIND_TABLEFUNC))
1209 expr = flatten_join_alias_vars(root, root->parse, expr);
1210
1211 /*
1212 * Simplify constant expressions. For function RTEs, this was already
1213 * done by preprocess_function_rtes. (But note we must do it again for
1214 * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1215 * un-simplified subexpressions inserted by flattening of subqueries or
1216 * join alias variables.)
1217 *
1218 * Note: an essential effect of this is to convert named-argument function
1219 * calls to positional notation and insert the current actual values of
1220 * any default arguments for functions. To ensure that happens, we *must*
1221 * process all expressions here. Previous PG versions sometimes skipped
1222 * const-simplification if it didn't seem worth the trouble, but we can't
1223 * do that anymore.
1224 *
1225 * Note: this also flattens nested AND and OR expressions into N-argument
1226 * form. All processing of a qual expression after this point must be
1227 * careful to maintain AND/OR flatness --- that is, do not generate a tree
1228 * with AND directly under AND, nor OR directly under OR.
1229 */
1230 if (kind != EXPRKIND_RTFUNC)
1231 expr = eval_const_expressions(root, expr);
1232
1233 /*
1234 * If it's a qual or havingQual, canonicalize it.
1235 */
1236 if (kind == EXPRKIND_QUAL)
1237 {
1238 expr = (Node *) canonicalize_qual((Expr *) expr, false);
1239
1240#ifdef OPTIMIZER_DEBUG
1241 printf("After canonicalize_qual()\n");
1242 pprint(expr);
1243#endif
1244 }
1245
1246 /*
1247 * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1248 * hashfuncid of any that might execute more quickly by using hash lookups
1249 * instead of a linear search.
1250 */
1251 if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1252 {
1253 convert_saop_to_hashed_saop(expr);
1254 }
1255
1256 /* Expand SubLinks to SubPlans */
1257 if (root->parse->hasSubLinks)
1258 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1259
1260 /*
1261 * XXX do not insert anything here unless you have grokked the comments in
1262 * SS_replace_correlation_vars ...
1263 */
1264
1265 /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1266 if (root->query_level > 1)
1267 expr = SS_replace_correlation_vars(root, expr);
1268
1269 /*
1270 * If it's a qual or havingQual, convert it to implicit-AND format. (We
1271 * don't want to do this before eval_const_expressions, since the latter
1272 * would be unable to simplify a top-level AND correctly. Also,
1273 * SS_process_sublinks expects explicit-AND format.)
1274 */
1275 if (kind == EXPRKIND_QUAL)
1276 expr = (Node *) make_ands_implicit((Expr *) expr);
1277
1278 return expr;
1279}
1280
1281/*
1282 * preprocess_qual_conditions
1283 * Recursively scan the query's jointree and do subquery_planner's
1284 * preprocessing work on each qual condition found therein.
1285 */
1286static void
1287preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1288{
1289 if (jtnode == NULL)
1290 return;
1291 if (IsA(jtnode, RangeTblRef))
1292 {
1293 /* nothing to do here */
1294 }
1295 else if (IsA(jtnode, FromExpr))
1296 {
1297 FromExpr *f = (FromExpr *) jtnode;
1298 ListCell *l;
1299
1300 foreach(l, f->fromlist)
1301 preprocess_qual_conditions(root, lfirst(l));
1302
1303 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1304 }
1305 else if (IsA(jtnode, JoinExpr))
1306 {
1307 JoinExpr *j = (JoinExpr *) jtnode;
1308
1309 preprocess_qual_conditions(root, j->larg);
1310 preprocess_qual_conditions(root, j->rarg);
1311
1312 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1313 }
1314 else
1315 elog(ERROR, "unrecognized node type: %d",
1316 (int) nodeTag(jtnode));
1317}
1318
1319/*
1320 * preprocess_phv_expression
1321 * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1322 *
1323 * If a LATERAL subquery references an output of another subquery, and that
1324 * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1325 * join, then we'll push the PlaceHolderVar expression down into the subquery
1326 * and later pull it back up during find_lateral_references, which runs after
1327 * subquery_planner has preprocessed all the expressions that were in the
1328 * current query level to start with. So we need to preprocess it then.
1329 */
1330Expr *
1331preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1332{
1333 return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1334}
1335
1336/*--------------------
1337 * grouping_planner
1338 * Perform planning steps related to grouping, aggregation, etc.
1339 *
1340 * This function adds all required top-level processing to the scan/join
1341 * Path(s) produced by query_planner.
1342 *
1343 * tuple_fraction is the fraction of tuples we expect will be retrieved.
1344 * tuple_fraction is interpreted as follows:
1345 * 0: expect all tuples to be retrieved (normal case)
1346 * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1347 * from the plan to be retrieved
1348 * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1349 * expected to be retrieved (ie, a LIMIT specification).
1350 * setops is used for set operation subqueries to provide the subquery with
1351 * the context in which it's being used so that Paths correctly sorted for the
1352 * set operation can be generated. NULL when not planning a set operation
1353 * child, or when a child of a set op that isn't interested in sorted input.
1354 *
1355 * Returns nothing; the useful output is in the Paths we attach to the
1356 * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1357 * root->processed_tlist contains the final processed targetlist.
1358 *
1359 * Note that we have not done set_cheapest() on the final rel; it's convenient
1360 * to leave this to the caller.
1361 *--------------------
1362 */
1363static void
1364grouping_planner(PlannerInfo *root, double tuple_fraction,
1365 SetOperationStmt *setops)
1366{
1367 Query *parse = root->parse;
1368 int64 offset_est = 0;
1369 int64 count_est = 0;
1370 double limit_tuples = -1.0;
1371 bool have_postponed_srfs = false;
1372 PathTarget *final_target;
1373 List *final_targets;
1374 List *final_targets_contain_srfs;
1375 bool final_target_parallel_safe;
1376 RelOptInfo *current_rel;
1377 RelOptInfo *final_rel;
1378 FinalPathExtraData extra;
1379 ListCell *lc;
1380
1381 /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1382 if (parse->limitCount || parse->limitOffset)
1383 {
1384 tuple_fraction = preprocess_limit(root, tuple_fraction,
1385 &offset_est, &count_est);
1386
1387 /*
1388 * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1389 * estimate the effects of using a bounded sort.
1390 */
1391 if (count_est > 0 && offset_est >= 0)
1392 limit_tuples = (double) count_est + (double) offset_est;
1393 }
1394
1395 /* Make tuple_fraction accessible to lower-level routines */
1396 root->tuple_fraction = tuple_fraction;
1397
1398 if (parse->setOperations)
1399 {
1400 /*
1401 * Construct Paths for set operations. The results will not need any
1402 * work except perhaps a top-level sort and/or LIMIT. Note that any
1403 * special work for recursive unions is the responsibility of
1404 * plan_set_operations.
1405 */
1406 current_rel = plan_set_operations(root);
1407
1408 /*
1409 * We should not need to call preprocess_targetlist, since we must be
1410 * in a SELECT query node. Instead, use the processed_tlist returned
1411 * by plan_set_operations (since this tells whether it returned any
1412 * resjunk columns!), and transfer any sort key information from the
1413 * original tlist.
1414 */
1415 Assert(parse->commandType == CMD_SELECT);
1416
1417 /* for safety, copy processed_tlist instead of modifying in-place */
1418 root->processed_tlist =
1419 postprocess_setop_tlist(copyObject(root->processed_tlist),
1420 parse->targetList);
1421
1422 /* Also extract the PathTarget form of the setop result tlist */
1423 final_target = current_rel->cheapest_total_path->pathtarget;
1424
1425 /* And check whether it's parallel safe */
1426 final_target_parallel_safe =
1427 is_parallel_safe(root, (Node *) final_target->exprs);
1428
1429 /* The setop result tlist couldn't contain any SRFs */
1430 Assert(!parse->hasTargetSRFs);
1431 final_targets = final_targets_contain_srfs = NIL;
1432
1433 /*
1434 * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1435 * checked already, but let's make sure).
1436 */
1437 if (parse->rowMarks)
1438 ereport(ERROR,
1439 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1440 /*------
1441 translator: %s is a SQL row locking clause such as FOR UPDATE */
1442 errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1443 LCS_asString(linitial_node(RowMarkClause,
1444 parse->rowMarks)->strength))));
1445
1446 /*
1447 * Calculate pathkeys that represent result ordering requirements
1448 */
1449 Assert(parse->distinctClause == NIL);
1450 root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1451 parse->sortClause,
1452 root->processed_tlist);
1453 }
1454 else
1455 {
1456 /* No set operations, do regular planning */
1457 PathTarget *sort_input_target;
1458 List *sort_input_targets;
1459 List *sort_input_targets_contain_srfs;
1460 bool sort_input_target_parallel_safe;
1461 PathTarget *grouping_target;
1462 List *grouping_targets;
1463 List *grouping_targets_contain_srfs;
1464 bool grouping_target_parallel_safe;
1465 PathTarget *scanjoin_target;
1466 List *scanjoin_targets;
1467 List *scanjoin_targets_contain_srfs;
1468 bool scanjoin_target_parallel_safe;
1469 bool scanjoin_target_same_exprs;
1470 bool have_grouping;
1471 WindowFuncLists *wflists = NULL;
1472 List *activeWindows = NIL;
1473 grouping_sets_data *gset_data = NULL;
1474 standard_qp_extra qp_extra;
1475
1476 /* A recursive query should always have setOperations */
1477 Assert(!root->hasRecursion);
1478
1479 /* Preprocess grouping sets and GROUP BY clause, if any */
1480 if (parse->groupingSets)
1481 {
1482 gset_data = preprocess_grouping_sets(root);
1483 }
1484 else if (parse->groupClause)
1485 {
1486 /* Preprocess regular GROUP BY clause, if any */
1487 root->processed_groupClause = preprocess_groupclause(root, NIL);
1488 }
1489
1490 /*
1491 * Preprocess targetlist. Note that much of the remaining planning
1492 * work will be done with the PathTarget representation of tlists, but
1493 * we must also maintain the full representation of the final tlist so
1494 * that we can transfer its decoration (resnames etc) to the topmost
1495 * tlist of the finished Plan. This is kept in processed_tlist.
1496 */
1497 preprocess_targetlist(root);
1498
1499 /*
1500 * Mark all the aggregates with resolved aggtranstypes, and detect
1501 * aggregates that are duplicates or can share transition state. We
1502 * must do this before slicing and dicing the tlist into various
1503 * pathtargets, else some copies of the Aggref nodes might escape
1504 * being marked.
1505 */
1506 if (parse->hasAggs)
1507 {
1508 preprocess_aggrefs(root, (Node *) root->processed_tlist);
1509 preprocess_aggrefs(root, (Node *) parse->havingQual);
1510 }
1511
1512 /*
1513 * Locate any window functions in the tlist. (We don't need to look
1514 * anywhere else, since expressions used in ORDER BY will be in there
1515 * too.) Note that they could all have been eliminated by constant
1516 * folding, in which case we don't need to do any more work.
1517 */
1518 if (parse->hasWindowFuncs)
1519 {
1520 wflists = find_window_functions((Node *) root->processed_tlist,
1521 list_length(parse->windowClause));
1522 if (wflists->numWindowFuncs > 0)
1523 {
1524 /*
1525 * See if any modifications can be made to each WindowClause
1526 * to allow the executor to execute the WindowFuncs more
1527 * quickly.
1528 */
1529 optimize_window_clauses(root, wflists);
1530
1531 activeWindows = select_active_windows(root, wflists);
1532 }
1533 else
1534 parse->hasWindowFuncs = false;
1535 }
1536
1537 /*
1538 * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1539 * adding logic between here and the query_planner() call. Anything
1540 * that is needed in MIN/MAX-optimizable cases will have to be
1541 * duplicated in planagg.c.
1542 */
1543 if (parse->hasAggs)
1544 preprocess_minmax_aggregates(root);
1545
1546 /*
1547 * Figure out whether there's a hard limit on the number of rows that
1548 * query_planner's result subplan needs to return. Even if we know a
1549 * hard limit overall, it doesn't apply if the query has any
1550 * grouping/aggregation operations, or SRFs in the tlist.
1551 */
1552 if (parse->groupClause ||
1553 parse->groupingSets ||
1554 parse->distinctClause ||
1555 parse->hasAggs ||
1556 parse->hasWindowFuncs ||
1557 parse->hasTargetSRFs ||
1558 root->hasHavingQual)
1559 root->limit_tuples = -1.0;
1560 else
1561 root->limit_tuples = limit_tuples;
1562
1563 /* Set up data needed by standard_qp_callback */
1564 qp_extra.activeWindows = activeWindows;
1565 qp_extra.gset_data = gset_data;
1566
1567 /*
1568 * If we're a subquery for a set operation, store the SetOperationStmt
1569 * in qp_extra.
1570 */
1571 qp_extra.setop = setops;
1572
1573 /*
1574 * Generate the best unsorted and presorted paths for the scan/join
1575 * portion of this Query, ie the processing represented by the
1576 * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1577 * We also generate (in standard_qp_callback) pathkey representations
1578 * of the query's sort clause, distinct clause, etc.
1579 */
1580 current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1581
1582 /*
1583 * Convert the query's result tlist into PathTarget format.
1584 *
1585 * Note: this cannot be done before query_planner() has performed
1586 * appendrel expansion, because that might add resjunk entries to
1587 * root->processed_tlist. Waiting till afterwards is also helpful
1588 * because the target width estimates can use per-Var width numbers
1589 * that were obtained within query_planner().
1590 */
1591 final_target = create_pathtarget(root, root->processed_tlist);
1592 final_target_parallel_safe =
1593 is_parallel_safe(root, (Node *) final_target->exprs);
1594
1595 /*
1596 * If ORDER BY was given, consider whether we should use a post-sort
1597 * projection, and compute the adjusted target for preceding steps if
1598 * so.
1599 */
1600 if (parse->sortClause)
1601 {
1602 sort_input_target = make_sort_input_target(root,
1603 final_target,
1604 &have_postponed_srfs);
1605 sort_input_target_parallel_safe =
1606 is_parallel_safe(root, (Node *) sort_input_target->exprs);
1607 }
1608 else
1609 {
1610 sort_input_target = final_target;
1611 sort_input_target_parallel_safe = final_target_parallel_safe;
1612 }
1613
1614 /*
1615 * If we have window functions to deal with, the output from any
1616 * grouping step needs to be what the window functions want;
1617 * otherwise, it should be sort_input_target.
1618 */
1619 if (activeWindows)
1620 {
1621 grouping_target = make_window_input_target(root,
1622 final_target,
1623 activeWindows);
1624 grouping_target_parallel_safe =
1625 is_parallel_safe(root, (Node *) grouping_target->exprs);
1626 }
1627 else
1628 {
1629 grouping_target = sort_input_target;
1630 grouping_target_parallel_safe = sort_input_target_parallel_safe;
1631 }
1632
1633 /*
1634 * If we have grouping or aggregation to do, the topmost scan/join
1635 * plan node must emit what the grouping step wants; otherwise, it
1636 * should emit grouping_target.
1637 */
1638 have_grouping = (parse->groupClause || parse->groupingSets ||
1639 parse->hasAggs || root->hasHavingQual);
1640 if (have_grouping)
1641 {
1642 scanjoin_target = make_group_input_target(root, final_target);
1643 scanjoin_target_parallel_safe =
1644 is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1645 }
1646 else
1647 {
1648 scanjoin_target = grouping_target;
1649 scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1650 }
1651
1652 /*
1653 * If there are any SRFs in the targetlist, we must separate each of
1654 * these PathTargets into SRF-computing and SRF-free targets. Replace
1655 * each of the named targets with a SRF-free version, and remember the
1656 * list of additional projection steps we need to add afterwards.
1657 */
1658 if (parse->hasTargetSRFs)
1659 {
1660 /* final_target doesn't recompute any SRFs in sort_input_target */
1661 split_pathtarget_at_srfs(root, final_target, sort_input_target,
1662 &final_targets,
1663 &final_targets_contain_srfs);
1664 final_target = linitial_node(PathTarget, final_targets);
1665 Assert(!linitial_int(final_targets_contain_srfs));
1666 /* likewise for sort_input_target vs. grouping_target */
1667 split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1668 &sort_input_targets,
1669 &sort_input_targets_contain_srfs);
1670 sort_input_target = linitial_node(PathTarget, sort_input_targets);
1671 Assert(!linitial_int(sort_input_targets_contain_srfs));
1672 /* likewise for grouping_target vs. scanjoin_target */
1673 split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
1674 &grouping_targets,
1675 &grouping_targets_contain_srfs);
1676 grouping_target = linitial_node(PathTarget, grouping_targets);
1677 Assert(!linitial_int(grouping_targets_contain_srfs));
1678 /* scanjoin_target will not have any SRFs precomputed for it */
1679 split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1680 &scanjoin_targets,
1681 &scanjoin_targets_contain_srfs);
1682 scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1683 Assert(!linitial_int(scanjoin_targets_contain_srfs));
1684 }
1685 else
1686 {
1687 /* initialize lists; for most of these, dummy values are OK */
1688 final_targets = final_targets_contain_srfs = NIL;
1689 sort_input_targets = sort_input_targets_contain_srfs = NIL;
1690 grouping_targets = grouping_targets_contain_srfs = NIL;
1691 scanjoin_targets = list_make1(scanjoin_target);
1692 scanjoin_targets_contain_srfs = NIL;
1693 }
1694
1695 /* Apply scan/join target. */
1696 scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1697 && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1698 apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1699 scanjoin_targets_contain_srfs,
1700 scanjoin_target_parallel_safe,
1701 scanjoin_target_same_exprs);
1702
1703 /*
1704 * Save the various upper-rel PathTargets we just computed into
1705 * root->upper_targets[]. The core code doesn't use this, but it
1706 * provides a convenient place for extensions to get at the info. For
1707 * consistency, we save all the intermediate targets, even though some
1708 * of the corresponding upperrels might not be needed for this query.
1709 */
1710 root->upper_targets[UPPERREL_FINAL] = final_target;
1711 root->upper_targets[UPPERREL_ORDERED] = final_target;
1712 root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1713 root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1714 root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1715 root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1716
1717 /*
1718 * If we have grouping and/or aggregation, consider ways to implement
1719 * that. We build a new upperrel representing the output of this
1720 * phase.
1721 */
1722 if (have_grouping)
1723 {
1724 current_rel = create_grouping_paths(root,
1725 current_rel,
1726 grouping_target,
1727 grouping_target_parallel_safe,
1728 gset_data);
1729 /* Fix things up if grouping_target contains SRFs */
1730 if (parse->hasTargetSRFs)
1731 adjust_paths_for_srfs(root, current_rel,
1732 grouping_targets,
1733 grouping_targets_contain_srfs);
1734 }
1735
1736 /*
1737 * If we have window functions, consider ways to implement those. We
1738 * build a new upperrel representing the output of this phase.
1739 */
1740 if (activeWindows)
1741 {
1742 current_rel = create_window_paths(root,
1743 current_rel,
1744 grouping_target,
1745 sort_input_target,
1746 sort_input_target_parallel_safe,
1747 wflists,
1748 activeWindows);
1749 /* Fix things up if sort_input_target contains SRFs */
1750 if (parse->hasTargetSRFs)
1751 adjust_paths_for_srfs(root, current_rel,
1752 sort_input_targets,
1753 sort_input_targets_contain_srfs);
1754 }
1755
1756 /*
1757 * If there is a DISTINCT clause, consider ways to implement that. We
1758 * build a new upperrel representing the output of this phase.
1759 */
1760 if (parse->distinctClause)
1761 {
1762 current_rel = create_distinct_paths(root,
1763 current_rel,
1764 sort_input_target);
1765 }
1766 } /* end of if (setOperations) */
1767
1768 /*
1769 * If ORDER BY was given, consider ways to implement that, and generate a
1770 * new upperrel containing only paths that emit the correct ordering and
1771 * project the correct final_target. We can apply the original
1772 * limit_tuples limit in sort costing here, but only if there are no
1773 * postponed SRFs.
1774 */
1775 if (parse->sortClause)
1776 {
1777 current_rel = create_ordered_paths(root,
1778 current_rel,
1779 final_target,
1780 final_target_parallel_safe,
1781 have_postponed_srfs ? -1.0 :
1782 limit_tuples);
1783 /* Fix things up if final_target contains SRFs */
1784 if (parse->hasTargetSRFs)
1785 adjust_paths_for_srfs(root, current_rel,
1786 final_targets,
1787 final_targets_contain_srfs);
1788 }
1789
1790 /*
1791 * Now we are prepared to build the final-output upperrel.
1792 */
1793 final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1794
1795 /*
1796 * If the input rel is marked consider_parallel and there's nothing that's
1797 * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1798 * consider_parallel as well. Note that if the query has rowMarks or is
1799 * not a SELECT, consider_parallel will be false for every relation in the
1800 * query.
1801 */
1802 if (current_rel->consider_parallel &&
1803 is_parallel_safe(root, parse->limitOffset) &&
1804 is_parallel_safe(root, parse->limitCount))
1805 final_rel->consider_parallel = true;
1806
1807 /*
1808 * If the current_rel belongs to a single FDW, so does the final_rel.
1809 */
1810 final_rel->serverid = current_rel->serverid;
1811 final_rel->userid = current_rel->userid;
1812 final_rel->useridiscurrent = current_rel->useridiscurrent;
1813 final_rel->fdwroutine = current_rel->fdwroutine;
1814
1815 /*
1816 * Generate paths for the final_rel. Insert all surviving paths, with
1817 * LockRows, Limit, and/or ModifyTable steps added if needed.
1818 */
1819 foreach(lc, current_rel->pathlist)
1820 {
1821 Path *path = (Path *) lfirst(lc);
1822
1823 /*
1824 * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1825 * (Note: we intentionally test parse->rowMarks not root->rowMarks
1826 * here. If there are only non-locking rowmarks, they should be
1827 * handled by the ModifyTable node instead. However, root->rowMarks
1828 * is what goes into the LockRows node.)
1829 */
1830 if (parse->rowMarks)
1831 {
1832 path = (Path *) create_lockrows_path(root, final_rel, path,
1833 root->rowMarks,
1834 assign_special_exec_param(root));
1835 }
1836
1837 /*
1838 * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1839 */
1840 if (limit_needed(parse))
1841 {
1842 path = (Path *) create_limit_path(root, final_rel, path,
1843 parse->limitOffset,
1844 parse->limitCount,
1845 parse->limitOption,
1846 offset_est, count_est);
1847 }
1848
1849 /*
1850 * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
1851 */
1852 if (parse->commandType != CMD_SELECT)
1853 {
1854 Index rootRelation;
1855 List *resultRelations = NIL;
1856 List *updateColnosLists = NIL;
1857 List *withCheckOptionLists = NIL;
1858 List *returningLists = NIL;
1859 List *mergeActionLists = NIL;
1860 List *mergeJoinConditions = NIL;
1861 List *rowMarks;
1862
1863 if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
1864 {
1865 /* Inherited UPDATE/DELETE/MERGE */
1866 RelOptInfo *top_result_rel = find_base_rel(root,
1867 parse->resultRelation);
1868 int resultRelation = -1;
1869
1870 /* Pass the root result rel forward to the executor. */
1871 rootRelation = parse->resultRelation;
1872
1873 /* Add only leaf children to ModifyTable. */
1874 while ((resultRelation = bms_next_member(root->leaf_result_relids,
1875 resultRelation)) >= 0)
1876 {
1877 RelOptInfo *this_result_rel = find_base_rel(root,
1878 resultRelation);
1879
1880 /*
1881 * Also exclude any leaf rels that have turned dummy since
1882 * being added to the list, for example, by being excluded
1883 * by constraint exclusion.
1884 */
1885 if (IS_DUMMY_REL(this_result_rel))
1886 continue;
1887
1888 /* Build per-target-rel lists needed by ModifyTable */
1889 resultRelations = lappend_int(resultRelations,
1890 resultRelation);
1891 if (parse->commandType == CMD_UPDATE)
1892 {
1893 List *update_colnos = root->update_colnos;
1894
1895 if (this_result_rel != top_result_rel)
1896 update_colnos =
1897 adjust_inherited_attnums_multilevel(root,
1898 update_colnos,
1899 this_result_rel->relid,
1900 top_result_rel->relid);
1901 updateColnosLists = lappend(updateColnosLists,
1902 update_colnos);
1903 }
1904 if (parse->withCheckOptions)
1905 {
1906 List *withCheckOptions = parse->withCheckOptions;
1907
1908 if (this_result_rel != top_result_rel)
1909 withCheckOptions = (List *)
1910 adjust_appendrel_attrs_multilevel(root,
1911 (Node *) withCheckOptions,
1912 this_result_rel,
1913 top_result_rel);
1914 withCheckOptionLists = lappend(withCheckOptionLists,
1915 withCheckOptions);
1916 }
1917 if (parse->returningList)
1918 {
1919 List *returningList = parse->returningList;
1920
1921 if (this_result_rel != top_result_rel)
1922 returningList = (List *)
1923 adjust_appendrel_attrs_multilevel(root,
1924 (Node *) returningList,
1925 this_result_rel,
1926 top_result_rel);
1927 returningLists = lappend(returningLists,
1928 returningList);
1929 }
1930 if (parse->mergeActionList)
1931 {
1932 ListCell *l;
1933 List *mergeActionList = NIL;
1934
1935 /*
1936 * Copy MergeActions and translate stuff that
1937 * references attribute numbers.
1938 */
1939 foreach(l, parse->mergeActionList)
1940 {
1941 MergeAction *action = lfirst(l),
1942 *leaf_action = copyObject(action);
1943
1944 leaf_action->qual =
1945 adjust_appendrel_attrs_multilevel(root,
1946 (Node *) action->qual,
1947 this_result_rel,
1948 top_result_rel);
1949 leaf_action->targetList = (List *)
1950 adjust_appendrel_attrs_multilevel(root,
1951 (Node *) action->targetList,
1952 this_result_rel,
1953 top_result_rel);
1954 if (leaf_action->commandType == CMD_UPDATE)
1955 leaf_action->updateColnos =
1956 adjust_inherited_attnums_multilevel(root,
1957 action->updateColnos,
1958 this_result_rel->relid,
1959 top_result_rel->relid);
1960 mergeActionList = lappend(mergeActionList,
1961 leaf_action);
1962 }
1963
1964 mergeActionLists = lappend(mergeActionLists,
1965 mergeActionList);
1966 }
1967 if (parse->commandType == CMD_MERGE)
1968 {
1969 Node *mergeJoinCondition = parse->mergeJoinCondition;
1970
1971 if (this_result_rel != top_result_rel)
1972 mergeJoinCondition =
1973 adjust_appendrel_attrs_multilevel(root,
1974 mergeJoinCondition,
1975 this_result_rel,
1976 top_result_rel);
1977 mergeJoinConditions = lappend(mergeJoinConditions,
1978 mergeJoinCondition);
1979 }
1980 }
1981
1982 if (resultRelations == NIL)
1983 {
1984 /*
1985 * We managed to exclude every child rel, so generate a
1986 * dummy one-relation plan using info for the top target
1987 * rel (even though that may not be a leaf target).
1988 * Although it's clear that no data will be updated or
1989 * deleted, we still need to have a ModifyTable node so
1990 * that any statement triggers will be executed. (This
1991 * could be cleaner if we fixed nodeModifyTable.c to allow
1992 * zero target relations, but that probably wouldn't be a
1993 * net win.)
1994 */
1995 resultRelations = list_make1_int(parse->resultRelation);
1996 if (parse->commandType == CMD_UPDATE)
1997 updateColnosLists = list_make1(root->update_colnos);
1998 if (parse->withCheckOptions)
1999 withCheckOptionLists = list_make1(parse->withCheckOptions);
2000 if (parse->returningList)
2001 returningLists = list_make1(parse->returningList);
2002 if (parse->mergeActionList)
2003 mergeActionLists = list_make1(parse->mergeActionList);
2004 if (parse->commandType == CMD_MERGE)
2005 mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2006 }
2007 }
2008 else
2009 {
2010 /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2011 rootRelation = 0; /* there's no separate root rel */
2012 resultRelations = list_make1_int(parse->resultRelation);
2013 if (parse->commandType == CMD_UPDATE)
2014 updateColnosLists = list_make1(root->update_colnos);
2015 if (parse->withCheckOptions)
2016 withCheckOptionLists = list_make1(parse->withCheckOptions);
2017 if (parse->returningList)
2018 returningLists = list_make1(parse->returningList);
2019 if (parse->mergeActionList)
2020 mergeActionLists = list_make1(parse->mergeActionList);
2021 if (parse->commandType == CMD_MERGE)
2022 mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2023 }
2024
2025 /*
2026 * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2027 * will have dealt with fetching non-locked marked rows, else we
2028 * need to have ModifyTable do that.
2029 */
2030 if (parse->rowMarks)
2031 rowMarks = NIL;
2032 else
2033 rowMarks = root->rowMarks;
2034
2035 path = (Path *)
2036 create_modifytable_path(root, final_rel,
2037 path,
2038 parse->commandType,
2039 parse->canSetTag,
2040 parse->resultRelation,
2041 rootRelation,
2042 root->partColsUpdated,
2043 resultRelations,
2044 updateColnosLists,
2045 withCheckOptionLists,
2046 returningLists,
2047 rowMarks,
2048 parse->onConflict,
2049 mergeActionLists,
2050 mergeJoinConditions,
2051 assign_special_exec_param(root));
2052 }
2053
2054 /* And shove it into final_rel */
2055 add_path(final_rel, path);
2056 }
2057
2058 /*
2059 * Generate partial paths for final_rel, too, if outer query levels might
2060 * be able to make use of them.
2061 */
2062 if (final_rel->consider_parallel && root->query_level > 1 &&
2063 !limit_needed(parse))
2064 {
2065 Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2066 foreach(lc, current_rel->partial_pathlist)
2067 {
2068 Path *partial_path = (Path *) lfirst(lc);
2069
2070 add_partial_path(final_rel, partial_path);
2071 }
2072 }
2073
2074 extra.limit_needed = limit_needed(parse);
2075 extra.limit_tuples = limit_tuples;
2076 extra.count_est = count_est;
2077 extra.offset_est = offset_est;
2078
2079 /*
2080 * If there is an FDW that's responsible for all baserels of the query,
2081 * let it consider adding ForeignPaths.
2082 */
2083 if (final_rel->fdwroutine &&
2084 final_rel->fdwroutine->GetForeignUpperPaths)
2085 final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2086 current_rel, final_rel,
2087 &extra);
2088
2089 /* Let extensions possibly add some more paths */
2090 if (create_upper_paths_hook)
2091 (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2092 current_rel, final_rel, &extra);
2093
2094 /* Note: currently, we leave it to callers to do set_cheapest() */
2095}
2096
2097/*
2098 * Do preprocessing for groupingSets clause and related data. This handles the
2099 * preliminary steps of expanding the grouping sets, organizing them into lists
2100 * of rollups, and preparing annotations which will later be filled in with
2101 * size estimates.
2102 */
2103static grouping_sets_data *
2104preprocess_grouping_sets(PlannerInfo *root)
2105{
2106 Query *parse = root->parse;
2107 List *sets;
2108 int maxref = 0;
2109 ListCell *lc_set;
2110 grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2111
2112 parse->groupingSets = expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
2113
2114 gd->any_hashable = false;
2115 gd->unhashable_refs = NULL;
2116 gd->unsortable_refs = NULL;
2117 gd->unsortable_sets = NIL;
2118
2119 /*
2120 * We don't currently make any attempt to optimize the groupClause when
2121 * there are grouping sets, so just duplicate it in processed_groupClause.
2122 */
2123 root->processed_groupClause = parse->groupClause;
2124
2125 if (parse->groupClause)
2126 {
2127 ListCell *lc;
2128
2129 foreach(lc, parse->groupClause)
2130 {
2131 SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2132 Index ref = gc->tleSortGroupRef;
2133
2134 if (ref > maxref)
2135 maxref = ref;
2136
2137 if (!gc->hashable)
2138 gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2139
2140 if (!OidIsValid(gc->sortop))
2141 gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2142 }
2143 }
2144
2145 /* Allocate workspace array for remapping */
2146 gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2147
2148 /*
2149 * If we have any unsortable sets, we must extract them before trying to
2150 * prepare rollups. Unsortable sets don't go through
2151 * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2152 * here.
2153 */
2154 if (!bms_is_empty(gd->unsortable_refs))
2155 {
2156 List *sortable_sets = NIL;
2157 ListCell *lc;
2158
2159 foreach(lc, parse->groupingSets)
2160 {
2161 List *gset = (List *) lfirst(lc);
2162
2163 if (bms_overlap_list(gd->unsortable_refs, gset))
2164 {
2165 GroupingSetData *gs = makeNode(GroupingSetData);
2166
2167 gs->set = gset;
2168 gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2169
2170 /*
2171 * We must enforce here that an unsortable set is hashable;
2172 * later code assumes this. Parse analysis only checks that
2173 * every individual column is either hashable or sortable.
2174 *
2175 * Note that passing this test doesn't guarantee we can
2176 * generate a plan; there might be other showstoppers.
2177 */
2178 if (bms_overlap_list(gd->unhashable_refs, gset))
2179 ereport(ERROR,
2180 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2181 errmsg("could not implement GROUP BY"),
2182 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2183 }
2184 else
2185 sortable_sets = lappend(sortable_sets, gset);
2186 }
2187
2188 if (sortable_sets)
2189 sets = extract_rollup_sets(sortable_sets);
2190 else
2191 sets = NIL;
2192 }
2193 else
2194 sets = extract_rollup_sets(parse->groupingSets);
2195
2196 foreach(lc_set, sets)
2197 {
2198 List *current_sets = (List *) lfirst(lc_set);
2199 RollupData *rollup = makeNode(RollupData);
2200 GroupingSetData *gs;
2201
2202 /*
2203 * Reorder the current list of grouping sets into correct prefix
2204 * order. If only one aggregation pass is needed, try to make the
2205 * list match the ORDER BY clause; if more than one pass is needed, we
2206 * don't bother with that.
2207 *
2208 * Note that this reorders the sets from smallest-member-first to
2209 * largest-member-first, and applies the GroupingSetData annotations,
2210 * though the data will be filled in later.
2211 */
2212 current_sets = reorder_grouping_sets(current_sets,
2213 (list_length(sets) == 1
2214 ? parse->sortClause
2215 : NIL));
2216
2217 /*
2218 * Get the initial (and therefore largest) grouping set.
2219 */
2220 gs = linitial_node(GroupingSetData, current_sets);
2221
2222 /*
2223 * Order the groupClause appropriately. If the first grouping set is
2224 * empty, then the groupClause must also be empty; otherwise we have
2225 * to force the groupClause to match that grouping set's order.
2226 *
2227 * (The first grouping set can be empty even though parse->groupClause
2228 * is not empty only if all non-empty grouping sets are unsortable.
2229 * The groupClauses for hashed grouping sets are built later on.)
2230 */
2231 if (gs->set)
2232 rollup->groupClause = preprocess_groupclause(root, gs->set);
2233 else
2234 rollup->groupClause = NIL;
2235
2236 /*
2237 * Is it hashable? We pretend empty sets are hashable even though we
2238 * actually force them not to be hashed later. But don't bother if
2239 * there's nothing but empty sets (since in that case we can't hash
2240 * anything).
2241 */
2242 if (gs->set &&
2243 !bms_overlap_list(gd->unhashable_refs, gs->set))
2244 {
2245 rollup->hashable = true;
2246 gd->any_hashable = true;
2247 }
2248
2249 /*
2250 * Now that we've pinned down an order for the groupClause for this
2251 * list of grouping sets, we need to remap the entries in the grouping
2252 * sets from sortgrouprefs to plain indices (0-based) into the
2253 * groupClause for this collection of grouping sets. We keep the
2254 * original form for later use, though.
2255 */
2256 rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2257 current_sets,
2258 gd->tleref_to_colnum_map);
2259 rollup->gsets_data = current_sets;
2260
2261 gd->rollups = lappend(gd->rollups, rollup);
2262 }
2263
2264 if (gd->unsortable_sets)
2265 {
2266 /*
2267 * We have not yet pinned down a groupclause for this, but we will
2268 * need index-based lists for estimation purposes. Construct
2269 * hash_sets_idx based on the entire original groupclause for now.
2270 */
2271 gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2272 gd->unsortable_sets,
2273 gd->tleref_to_colnum_map);
2274 gd->any_hashable = true;
2275 }
2276
2277 return gd;
2278}
2279
2280/*
2281 * Given a groupclause and a list of GroupingSetData, return equivalent sets
2282 * (without annotation) mapped to indexes into the given groupclause.
2283 */
2284static List *
2285remap_to_groupclause_idx(List *groupClause,
2286 List *gsets,
2287 int *tleref_to_colnum_map)
2288{
2289 int ref = 0;
2290 List *result = NIL;
2291 ListCell *lc;
2292
2293 foreach(lc, groupClause)
2294 {
2295 SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2296
2297 tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2298 }
2299
2300 foreach(lc, gsets)
2301 {
2302 List *set = NIL;
2303 ListCell *lc2;
2304 GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2305
2306 foreach(lc2, gs->set)
2307 {
2308 set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2309 }
2310
2311 result = lappend(result, set);
2312 }
2313
2314 return result;
2315}
2316
2317
2318/*
2319 * preprocess_rowmarks - set up PlanRowMarks if needed
2320 */
2321static void
2322preprocess_rowmarks(PlannerInfo *root)
2323{
2324 Query *parse = root->parse;
2325 Bitmapset *rels;
2326 List *prowmarks;
2327 ListCell *l;
2328 int i;
2329
2330 if (parse->rowMarks)
2331 {
2332 /*
2333 * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2334 * grouping, since grouping renders a reference to individual tuple
2335 * CTIDs invalid. This is also checked at parse time, but that's
2336 * insufficient because of rule substitution, query pullup, etc.
2337 */
2338 CheckSelectLocking(parse, linitial_node(RowMarkClause,
2339 parse->rowMarks)->strength);
2340 }
2341 else
2342 {
2343 /*
2344 * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2345 * UPDATE/SHARE.
2346 */
2347 if (parse->commandType != CMD_UPDATE &&
2348 parse->commandType != CMD_DELETE &&
2349 parse->commandType != CMD_MERGE)
2350 return;
2351 }
2352
2353 /*
2354 * We need to have rowmarks for all base relations except the target. We
2355 * make a bitmapset of all base rels and then remove the items we don't
2356 * need or have FOR [KEY] UPDATE/SHARE marks for.
2357 */
2358 rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2359 if (parse->resultRelation)
2360 rels = bms_del_member(rels, parse->resultRelation);
2361
2362 /*
2363 * Convert RowMarkClauses to PlanRowMark representation.
2364 */
2365 prowmarks = NIL;
2366 foreach(l, parse->rowMarks)
2367 {
2368 RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2369 RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2370 PlanRowMark *newrc;
2371
2372 /*
2373 * Currently, it is syntactically impossible to have FOR UPDATE et al
2374 * applied to an update/delete target rel. If that ever becomes
2375 * possible, we should drop the target from the PlanRowMark list.
2376 */
2377 Assert(rc->rti != parse->resultRelation);
2378
2379 /*
2380 * Ignore RowMarkClauses for subqueries; they aren't real tables and
2381 * can't support true locking. Subqueries that got flattened into the
2382 * main query should be ignored completely. Any that didn't will get
2383 * ROW_MARK_COPY items in the next loop.
2384 */
2385 if (rte->rtekind != RTE_RELATION)
2386 continue;
2387
2388 rels = bms_del_member(rels, rc->rti);
2389
2390 newrc = makeNode(PlanRowMark);
2391 newrc->rti = newrc->prti = rc->rti;
2392 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2393 newrc->markType = select_rowmark_type(rte, rc->strength);
2394 newrc->allMarkTypes = (1 << newrc->markType);
2395 newrc->strength = rc->strength;
2396 newrc->waitPolicy = rc->waitPolicy;
2397 newrc->isParent = false;
2398
2399 prowmarks = lappend(prowmarks, newrc);
2400 }
2401
2402 /*
2403 * Now, add rowmarks for any non-target, non-locked base relations.
2404 */
2405 i = 0;
2406 foreach(l, parse->rtable)
2407 {
2408 RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2409 PlanRowMark *newrc;
2410
2411 i++;
2412 if (!bms_is_member(i, rels))
2413 continue;
2414
2415 newrc = makeNode(PlanRowMark);
2416 newrc->rti = newrc->prti = i;
2417 newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2418 newrc->markType = select_rowmark_type(rte, LCS_NONE);
2419 newrc->allMarkTypes = (1 << newrc->markType);
2420 newrc->strength = LCS_NONE;
2421 newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2422 newrc->isParent = false;
2423
2424 prowmarks = lappend(prowmarks, newrc);
2425 }
2426
2427 root->rowMarks = prowmarks;
2428}
2429
2430/*
2431 * Select RowMarkType to use for a given table
2432 */
2433RowMarkType
2434select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2435{
2436 if (rte->rtekind != RTE_RELATION)
2437 {
2438 /* If it's not a table at all, use ROW_MARK_COPY */
2439 return ROW_MARK_COPY;
2440 }
2441 else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2442 {
2443 /* Let the FDW select the rowmark type, if it wants to */
2444 FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2445
2446 if (fdwroutine->GetForeignRowMarkType != NULL)
2447 return fdwroutine->GetForeignRowMarkType(rte, strength);
2448 /* Otherwise, use ROW_MARK_COPY by default */
2449 return ROW_MARK_COPY;
2450 }
2451 else
2452 {
2453 /* Regular table, apply the appropriate lock type */
2454 switch (strength)
2455 {
2456 case LCS_NONE:
2457
2458 /*
2459 * We don't need a tuple lock, only the ability to re-fetch
2460 * the row.
2461 */
2462 return ROW_MARK_REFERENCE;
2463 break;
2464 case LCS_FORKEYSHARE:
2465 return ROW_MARK_KEYSHARE;
2466 break;
2467 case LCS_FORSHARE:
2468 return ROW_MARK_SHARE;
2469 break;
2470 case LCS_FORNOKEYUPDATE:
2471 return ROW_MARK_NOKEYEXCLUSIVE;
2472 break;
2473 case LCS_FORUPDATE:
2474 return ROW_MARK_EXCLUSIVE;
2475 break;
2476 }
2477 elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2478 return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2479 }
2480}
2481
2482/*
2483 * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2484 *
2485 * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2486 * results back in *count_est and *offset_est. These variables are set to
2487 * 0 if the corresponding clause is not present, and -1 if it's present
2488 * but we couldn't estimate the value for it. (The "0" convention is OK
2489 * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2490 * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2491 * usual practice of never estimating less than one row.) These values will
2492 * be passed to create_limit_path, which see if you change this code.
2493 *
2494 * The return value is the suitably adjusted tuple_fraction to use for
2495 * planning the query. This adjustment is not overridable, since it reflects
2496 * plan actions that grouping_planner() will certainly take, not assumptions
2497 * about context.
2498 */
2499static double
2500preprocess_limit(PlannerInfo *root, double tuple_fraction,
2501 int64 *offset_est, int64 *count_est)
2502{
2503 Query *parse = root->parse;
2504 Node *est;
2505 double limit_fraction;
2506
2507 /* Should not be called unless LIMIT or OFFSET */
2508 Assert(parse->limitCount || parse->limitOffset);
2509
2510 /*
2511 * Try to obtain the clause values. We use estimate_expression_value
2512 * primarily because it can sometimes do something useful with Params.
2513 */
2514 if (parse->limitCount)
2515 {
2516 est = estimate_expression_value(root, parse->limitCount);
2517 if (est && IsA(est, Const))
2518 {
2519 if (((Const *) est)->constisnull)
2520 {
2521 /* NULL indicates LIMIT ALL, ie, no limit */
2522 *count_est = 0; /* treat as not present */
2523 }
2524 else
2525 {
2526 *count_est = DatumGetInt64(((Const *) est)->constvalue);
2527 if (*count_est <= 0)
2528 *count_est = 1; /* force to at least 1 */
2529 }
2530 }
2531 else
2532 *count_est = -1; /* can't estimate */
2533 }
2534 else
2535 *count_est = 0; /* not present */
2536
2537 if (parse->limitOffset)
2538 {
2539 est = estimate_expression_value(root, parse->limitOffset);
2540 if (est && IsA(est, Const))
2541 {
2542 if (((Const *) est)->constisnull)
2543 {
2544 /* Treat NULL as no offset; the executor will too */
2545 *offset_est = 0; /* treat as not present */
2546 }
2547 else
2548 {
2549 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2550 if (*offset_est < 0)
2551 *offset_est = 0; /* treat as not present */
2552 }
2553 }
2554 else
2555 *offset_est = -1; /* can't estimate */
2556 }
2557 else
2558 *offset_est = 0; /* not present */
2559
2560 if (*count_est != 0)
2561 {
2562 /*
2563 * A LIMIT clause limits the absolute number of tuples returned.
2564 * However, if it's not a constant LIMIT then we have to guess; for
2565 * lack of a better idea, assume 10% of the plan's result is wanted.
2566 */
2567 if (*count_est < 0 || *offset_est < 0)
2568 {
2569 /* LIMIT or OFFSET is an expression ... punt ... */
2570 limit_fraction = 0.10;
2571 }
2572 else
2573 {
2574 /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2575 limit_fraction = (double) *count_est + (double) *offset_est;
2576 }
2577
2578 /*
2579 * If we have absolute limits from both caller and LIMIT, use the
2580 * smaller value; likewise if they are both fractional. If one is
2581 * fractional and the other absolute, we can't easily determine which
2582 * is smaller, but we use the heuristic that the absolute will usually
2583 * be smaller.
2584 */
2585 if (tuple_fraction >= 1.0)
2586 {
2587 if (limit_fraction >= 1.0)
2588 {
2589 /* both absolute */
2590 tuple_fraction = Min(tuple_fraction, limit_fraction);
2591 }
2592 else
2593 {
2594 /* caller absolute, limit fractional; use caller's value */
2595 }
2596 }
2597 else if (tuple_fraction > 0.0)
2598 {
2599 if (limit_fraction >= 1.0)
2600 {
2601 /* caller fractional, limit absolute; use limit */
2602 tuple_fraction = limit_fraction;
2603 }
2604 else
2605 {
2606 /* both fractional */
2607 tuple_fraction = Min(tuple_fraction, limit_fraction);
2608 }
2609 }
2610 else
2611 {
2612 /* no info from caller, just use limit */
2613 tuple_fraction = limit_fraction;
2614 }
2615 }
2616 else if (*offset_est != 0 && tuple_fraction > 0.0)
2617 {
2618 /*
2619 * We have an OFFSET but no LIMIT. This acts entirely differently
2620 * from the LIMIT case: here, we need to increase rather than decrease
2621 * the caller's tuple_fraction, because the OFFSET acts to cause more
2622 * tuples to be fetched instead of fewer. This only matters if we got
2623 * a tuple_fraction > 0, however.
2624 *
2625 * As above, use 10% if OFFSET is present but unestimatable.
2626 */
2627 if (*offset_est < 0)
2628 limit_fraction = 0.10;
2629 else
2630 limit_fraction = (double) *offset_est;
2631
2632 /*
2633 * If we have absolute counts from both caller and OFFSET, add them
2634 * together; likewise if they are both fractional. If one is
2635 * fractional and the other absolute, we want to take the larger, and
2636 * we heuristically assume that's the fractional one.
2637 */
2638 if (tuple_fraction >= 1.0)
2639 {
2640 if (limit_fraction >= 1.0)
2641 {
2642 /* both absolute, so add them together */
2643 tuple_fraction += limit_fraction;
2644 }
2645 else
2646 {
2647 /* caller absolute, limit fractional; use limit */
2648 tuple_fraction = limit_fraction;
2649 }
2650 }
2651 else
2652 {
2653 if (limit_fraction >= 1.0)
2654 {
2655 /* caller fractional, limit absolute; use caller's value */
2656 }
2657 else
2658 {
2659 /* both fractional, so add them together */
2660 tuple_fraction += limit_fraction;
2661 if (tuple_fraction >= 1.0)
2662 tuple_fraction = 0.0; /* assume fetch all */
2663 }
2664 }
2665 }
2666
2667 return tuple_fraction;
2668}
2669
2670/*
2671 * limit_needed - do we actually need a Limit plan node?
2672 *
2673 * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2674 * a Limit node. This is worth checking for because "OFFSET 0" is a common
2675 * locution for an optimization fence. (Because other places in the planner
2676 * merely check whether parse->limitOffset isn't NULL, it will still work as
2677 * an optimization fence --- we're just suppressing unnecessary run-time
2678 * overhead.)
2679 *
2680 * This might look like it could be merged into preprocess_limit, but there's
2681 * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2682 * in preprocess_limit it's good enough to consider estimated values.
2683 */
2684bool
2685limit_needed(Query *parse)
2686{
2687 Node *node;
2688
2689 node = parse->limitCount;
2690 if (node)
2691 {
2692 if (IsA(node, Const))
2693 {
2694 /* NULL indicates LIMIT ALL, ie, no limit */
2695 if (!((Const *) node)->constisnull)
2696 return true; /* LIMIT with a constant value */
2697 }
2698 else
2699 return true; /* non-constant LIMIT */
2700 }
2701
2702 node = parse->limitOffset;
2703 if (node)
2704 {
2705 if (IsA(node, Const))
2706 {
2707 /* Treat NULL as no offset; the executor would too */
2708 if (!((Const *) node)->constisnull)
2709 {
2710 int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2711
2712 if (offset != 0)
2713 return true; /* OFFSET with a nonzero value */
2714 }
2715 }
2716 else
2717 return true; /* non-constant OFFSET */
2718 }
2719
2720 return false; /* don't need a Limit plan node */
2721}
2722
2723/*
2724 * preprocess_groupclause - do preparatory work on GROUP BY clause
2725 *
2726 * The idea here is to adjust the ordering of the GROUP BY elements
2727 * (which in itself is semantically insignificant) to match ORDER BY,
2728 * thereby allowing a single sort operation to both implement the ORDER BY
2729 * requirement and set up for a Unique step that implements GROUP BY.
2730 * We also consider partial match between GROUP BY and ORDER BY elements,
2731 * which could allow to implement ORDER BY using the incremental sort.
2732 *
2733 * We also consider other orderings of the GROUP BY elements, which could
2734 * match the sort ordering of other possible plans (eg an indexscan) and
2735 * thereby reduce cost. This is implemented during the generation of grouping
2736 * paths. See get_useful_group_keys_orderings() for details.
2737 *
2738 * Note: we need no comparable processing of the distinctClause because
2739 * the parser already enforced that that matches ORDER BY.
2740 *
2741 * Note: we return a fresh List, but its elements are the same
2742 * SortGroupClauses appearing in parse->groupClause. This is important
2743 * because later processing may modify the processed_groupClause list.
2744 *
2745 * For grouping sets, the order of items is instead forced to agree with that
2746 * of the grouping set (and items not in the grouping set are skipped). The
2747 * work of sorting the order of grouping set elements to match the ORDER BY if
2748 * possible is done elsewhere.
2749 */
2750static List *
2751preprocess_groupclause(PlannerInfo *root, List *force)
2752{
2753 Query *parse = root->parse;
2754 List *new_groupclause = NIL;
2755 ListCell *sl;
2756 ListCell *gl;
2757
2758 /* For grouping sets, we need to force the ordering */
2759 if (force)
2760 {
2761 foreach(sl, force)
2762 {
2763 Index ref = lfirst_int(sl);
2764 SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2765
2766 new_groupclause = lappend(new_groupclause, cl);
2767 }
2768
2769 return new_groupclause;
2770 }
2771
2772 /* If no ORDER BY, nothing useful to do here */
2773 if (parse->sortClause == NIL)
2774 return list_copy(parse->groupClause);
2775
2776 /*
2777 * Scan the ORDER BY clause and construct a list of matching GROUP BY
2778 * items, but only as far as we can make a matching prefix.
2779 *
2780 * This code assumes that the sortClause contains no duplicate items.
2781 */
2782 foreach(sl, parse->sortClause)
2783 {
2784 SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2785
2786 foreach(gl, parse->groupClause)
2787 {
2788 SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2789
2790 if (equal(gc, sc))
2791 {
2792 new_groupclause = lappend(new_groupclause, gc);
2793 break;
2794 }
2795 }
2796 if (gl == NULL)
2797 break; /* no match, so stop scanning */
2798 }
2799
2800
2801 /* If no match at all, no point in reordering GROUP BY */
2802 if (new_groupclause == NIL)
2803 return list_copy(parse->groupClause);
2804
2805 /*
2806 * Add any remaining GROUP BY items to the new list. We don't require a
2807 * complete match, because even partial match allows ORDER BY to be
2808 * implemented using incremental sort. Also, give up if there are any
2809 * non-sortable GROUP BY items, since then there's no hope anyway.
2810 */
2811 foreach(gl, parse->groupClause)
2812 {
2813 SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2814
2815 if (list_member_ptr(new_groupclause, gc))
2816 continue; /* it matched an ORDER BY item */
2817 if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2818 return list_copy(parse->groupClause);
2819 new_groupclause = lappend(new_groupclause, gc);
2820 }
2821
2822 /* Success --- install the rearranged GROUP BY list */
2823 Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2824 return new_groupclause;
2825}
2826
2827/*
2828 * Extract lists of grouping sets that can be implemented using a single
2829 * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2830 *
2831 * Input must be sorted with smallest sets first. Result has each sublist
2832 * sorted with smallest sets first.
2833 *
2834 * We want to produce the absolute minimum possible number of lists here to
2835 * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2836 * of finding the minimal partition of a partially-ordered set into chains
2837 * (which is what we need, taking the list of grouping sets as a poset ordered
2838 * by set inclusion) can be mapped to the problem of finding the maximum
2839 * cardinality matching on a bipartite graph, which is solvable in polynomial
2840 * time with a worst case of no worse than O(n^2.5) and usually much
2841 * better. Since our N is at most 4096, we don't need to consider fallbacks to
2842 * heuristic or approximate methods. (Planning time for a 12-d cube is under
2843 * half a second on my modest system even with optimization off and assertions
2844 * on.)
2845 */
2846static List *
2847extract_rollup_sets(List *groupingSets)
2848{
2849 int num_sets_raw = list_length(groupingSets);
2850 int num_empty = 0;
2851 int num_sets = 0; /* distinct sets */
2852 int num_chains = 0;
2853 List *result = NIL;
2854 List **results;
2855 List **orig_sets;
2856 Bitmapset **set_masks;
2857 int *chains;
2858 short **adjacency;
2859 short *adjacency_buf;
2860 BipartiteMatchState *state;
2861 int i;
2862 int j;
2863 int j_size;
2864 ListCell *lc1 = list_head(groupingSets);
2865 ListCell *lc;
2866
2867 /*
2868 * Start by stripping out empty sets. The algorithm doesn't require this,
2869 * but the planner currently needs all empty sets to be returned in the
2870 * first list, so we strip them here and add them back after.
2871 */
2872 while (lc1 && lfirst(lc1) == NIL)
2873 {
2874 ++num_empty;
2875 lc1 = lnext(groupingSets, lc1);
2876 }
2877
2878 /* bail out now if it turns out that all we had were empty sets. */
2879 if (!lc1)
2880 return list_make1(groupingSets);
2881
2882 /*----------
2883 * We don't strictly need to remove duplicate sets here, but if we don't,
2884 * they tend to become scattered through the result, which is a bit
2885 * confusing (and irritating if we ever decide to optimize them out).
2886 * So we remove them here and add them back after.
2887 *
2888 * For each non-duplicate set, we fill in the following:
2889 *
2890 * orig_sets[i] = list of the original set lists
2891 * set_masks[i] = bitmapset for testing inclusion
2892 * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2893 *
2894 * chains[i] will be the result group this set is assigned to.
2895 *
2896 * We index all of these from 1 rather than 0 because it is convenient
2897 * to leave 0 free for the NIL node in the graph algorithm.
2898 *----------
2899 */
2900 orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
2901 set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
2902 adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
2903 adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
2904
2905 j_size = 0;
2906 j = 0;
2907 i = 1;
2908
2909 for_each_cell(lc, groupingSets, lc1)
2910 {
2911 List *candidate = (List *) lfirst(lc);
2912 Bitmapset *candidate_set = NULL;
2913 ListCell *lc2;
2914 int dup_of = 0;
2915
2916 foreach(lc2, candidate)
2917 {
2918 candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
2919 }
2920
2921 /* we can only be a dup if we're the same length as a previous set */
2922 if (j_size == list_length(candidate))
2923 {
2924 int k;
2925
2926 for (k = j; k < i; ++k)
2927 {
2928 if (bms_equal(set_masks[k], candidate_set))
2929 {
2930 dup_of = k;
2931 break;
2932 }
2933 }
2934 }
2935 else if (j_size < list_length(candidate))
2936 {
2937 j_size = list_length(candidate);
2938 j = i;
2939 }
2940
2941 if (dup_of > 0)
2942 {
2943 orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
2944 bms_free(candidate_set);
2945 }
2946 else
2947 {
2948 int k;
2949 int n_adj = 0;
2950
2951 orig_sets[i] = list_make1(candidate);
2952 set_masks[i] = candidate_set;
2953
2954 /* fill in adjacency list; no need to compare equal-size sets */
2955
2956 for (k = j - 1; k > 0; --k)
2957 {
2958 if (bms_is_subset(set_masks[k], candidate_set))
2959 adjacency_buf[++n_adj] = k;
2960 }
2961
2962 if (n_adj > 0)
2963 {
2964 adjacency_buf[0] = n_adj;
2965 adjacency[i] = palloc((n_adj + 1) * sizeof(short));
2966 memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
2967 }
2968 else
2969 adjacency[i] = NULL;
2970
2971 ++i;
2972 }
2973 }
2974
2975 num_sets = i - 1;
2976
2977 /*
2978 * Apply the graph matching algorithm to do the work.
2979 */
2980 state = BipartiteMatch(num_sets, num_sets, adjacency);
2981
2982 /*
2983 * Now, the state->pair* fields have the info we need to assign sets to
2984 * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
2985 * pair_vu[v] = u (both will be true, but we check both so that we can do
2986 * it in one pass)
2987 */
2988 chains = palloc0((num_sets + 1) * sizeof(int));
2989
2990 for (i = 1; i <= num_sets; ++i)
2991 {
2992 int u = state->pair_vu[i];
2993 int v = state->pair_uv[i];
2994
2995 if (u > 0 && u < i)
2996 chains[i] = chains[u];
2997 else if (v > 0 && v < i)
2998 chains[i] = chains[v];
2999 else
3000 chains[i] = ++num_chains;
3001 }
3002
3003 /* build result lists. */
3004 results = palloc0((num_chains + 1) * sizeof(List *));
3005
3006 for (i = 1; i <= num_sets; ++i)
3007 {
3008 int c = chains[i];
3009
3010 Assert(c > 0);
3011
3012 results[c] = list_concat(results[c], orig_sets[i]);
3013 }
3014
3015 /* push any empty sets back on the first list. */
3016 while (num_empty-- > 0)
3017 results[1] = lcons(NIL, results[1]);
3018
3019 /* make result list */
3020 for (i = 1; i <= num_chains; ++i)
3021 result = lappend(result, results[i]);
3022
3023 /*
3024 * Free all the things.
3025 *
3026 * (This is over-fussy for small sets but for large sets we could have
3027 * tied up a nontrivial amount of memory.)
3028 */
3029 BipartiteMatchFree(state);
3030 pfree(results);
3031 pfree(chains);
3032 for (i = 1; i <= num_sets; ++i)
3033 if (adjacency[i])
3034 pfree(adjacency[i]);
3035 pfree(adjacency);
3036 pfree(adjacency_buf);
3037 pfree(orig_sets);
3038 for (i = 1; i <= num_sets; ++i)
3039 bms_free(set_masks[i]);
3040 pfree(set_masks);
3041
3042 return result;
3043}
3044
3045/*
3046 * Reorder the elements of a list of grouping sets such that they have correct
3047 * prefix relationships. Also inserts the GroupingSetData annotations.
3048 *
3049 * The input must be ordered with smallest sets first; the result is returned
3050 * with largest sets first. Note that the result shares no list substructure
3051 * with the input, so it's safe for the caller to modify it later.
3052 *
3053 * If we're passed in a sortclause, we follow its order of columns to the
3054 * extent possible, to minimize the chance that we add unnecessary sorts.
3055 * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3056 * gets implemented in one pass.)
3057 */
3058static List *
3059reorder_grouping_sets(List *groupingSets, List *sortclause)
3060{
3061 ListCell *lc;
3062 List *previous = NIL;
3063 List *result = NIL;
3064
3065 foreach(lc, groupingSets)
3066 {
3067 List *candidate = (List *) lfirst(lc);
3068 List *new_elems = list_difference_int(candidate, previous);
3069 GroupingSetData *gs = makeNode(GroupingSetData);
3070
3071 while (list_length(sortclause) > list_length(previous) &&
3072 new_elems != NIL)
3073 {
3074 SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3075 int ref = sc->tleSortGroupRef;
3076
3077 if (list_member_int(new_elems, ref))
3078 {
3079 previous = lappend_int(previous, ref);
3080 new_elems = list_delete_int(new_elems, ref);
3081 }
3082 else
3083 {
3084 /* diverged from the sortclause; give up on it */
3085 sortclause = NIL;
3086 break;
3087 }
3088 }
3089
3090 previous = list_concat(previous, new_elems);
3091
3092 gs->set = list_copy(previous);
3093 result = lcons(gs, result);
3094 }
3095
3096 list_free(previous);
3097
3098 return result;
3099}
3100
3101/*
3102 * has_volatile_pathkey
3103 * Returns true if any PathKey in 'keys' has an EquivalenceClass
3104 * containing a volatile function. Otherwise returns false.
3105 */
3106static bool
3107has_volatile_pathkey(List *keys)
3108{
3109 ListCell *lc;
3110
3111 foreach(lc, keys)
3112 {
3113 PathKey *pathkey = lfirst_node(PathKey, lc);
3114
3115 if (pathkey->pk_eclass->ec_has_volatile)
3116 return true;
3117 }
3118
3119 return false;
3120}
3121
3122/*
3123 * adjust_group_pathkeys_for_groupagg
3124 * Add pathkeys to root->group_pathkeys to reflect the best set of
3125 * pre-ordered input for ordered aggregates.
3126 *
3127 * We define "best" as the pathkeys that suit the largest number of
3128 * aggregate functions. We find these by looking at the first ORDER BY /
3129 * DISTINCT aggregate and take the pathkeys for that before searching for
3130 * other aggregates that require the same or a more strict variation of the
3131 * same pathkeys. We then repeat that process for any remaining aggregates
3132 * with different pathkeys and if we find another set of pathkeys that suits a
3133 * larger number of aggregates then we select those pathkeys instead.
3134 *
3135 * When the best pathkeys are found we also mark each Aggref that can use
3136 * those pathkeys as aggpresorted = true.
3137 *
3138 * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3139 * volatile functions, we never make use of these pathkeys. We want to ensure
3140 * that sorts using volatile functions are done independently in each Aggref
3141 * rather than once at the query level. If we were to allow this then Aggrefs
3142 * with compatible sort orders would all transition their rows in the same
3143 * order if those pathkeys were deemed to be the best pathkeys to sort on.
3144 * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3145 * better pathkeys to sort on, then the volatile function Aggrefs would be
3146 * left to perform their sorts individually. To avoid this inconsistent
3147 * behavior which could make Aggref results depend on what other Aggrefs the
3148 * query contains, we always force Aggrefs with volatile functions to perform
3149 * their own sorts.
3150 */
3151static void
3152adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3153{
3154 List *grouppathkeys = root->group_pathkeys;
3155 List *bestpathkeys;
3156 Bitmapset *bestaggs;
3157 Bitmapset *unprocessed_aggs;
3158 ListCell *lc;
3159 int i;
3160
3161 /* Shouldn't be here if there are grouping sets */
3162 Assert(root->parse->groupingSets == NIL);
3163 /* Shouldn't be here unless there are some ordered aggregates */
3164 Assert(root->numOrderedAggs > 0);
3165
3166 /* Do nothing if disabled */
3167 if (!enable_presorted_aggregate)
3168 return;
3169
3170 /*
3171 * Make a first pass over all AggInfos to collect a Bitmapset containing
3172 * the indexes of all AggInfos to be processed below.
3173 */
3174 unprocessed_aggs = NULL;
3175 foreach(lc, root->agginfos)
3176 {
3177 AggInfo *agginfo = lfirst_node(AggInfo, lc);
3178 Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3179
3180 if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3181 continue;
3182
3183 /* only add aggregates with a DISTINCT or ORDER BY */
3184 if (aggref->aggdistinct != NIL || aggref->aggorder != NIL)
3185 unprocessed_aggs = bms_add_member(unprocessed_aggs,
3186 foreach_current_index(lc));
3187 }
3188
3189 /*
3190 * Now process all the unprocessed_aggs to find the best set of pathkeys
3191 * for the given set of aggregates.
3192 *
3193 * On the first outer loop here 'bestaggs' will be empty. We'll populate
3194 * this during the first loop using the pathkeys for the very first
3195 * AggInfo then taking any stronger pathkeys from any other AggInfos with
3196 * a more strict set of compatible pathkeys. Once the outer loop is
3197 * complete, we mark off all the aggregates with compatible pathkeys then
3198 * remove those from the unprocessed_aggs and repeat the process to try to
3199 * find another set of pathkeys that are suitable for a larger number of
3200 * aggregates. The outer loop will stop when there are not enough
3201 * unprocessed aggregates for it to be possible to find a set of pathkeys
3202 * to suit a larger number of aggregates.
3203 */
3204 bestpathkeys = NIL;
3205 bestaggs = NULL;
3206 while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3207 {
3208 Bitmapset *aggindexes = NULL;
3209 List *currpathkeys = NIL;
3210
3211 i = -1;
3212 while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3213 {
3214 AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3215 Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3216 List *sortlist;
3217 List *pathkeys;
3218
3219 if (aggref->aggdistinct != NIL)
3220 sortlist = aggref->aggdistinct;
3221 else
3222 sortlist = aggref->aggorder;
3223
3224 pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3225 aggref->args);
3226
3227 /*
3228 * Ignore Aggrefs which have volatile functions in their ORDER BY
3229 * or DISTINCT clause.
3230 */
3231 if (has_volatile_pathkey(pathkeys))
3232 {
3233 unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3234 continue;
3235 }
3236
3237 /*
3238 * When not set yet, take the pathkeys from the first unprocessed
3239 * aggregate.
3240 */
3241 if (currpathkeys == NIL)
3242 {
3243 currpathkeys = pathkeys;
3244
3245 /* include the GROUP BY pathkeys, if they exist */
3246 if (grouppathkeys != NIL)
3247 currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3248 currpathkeys);
3249
3250 /* record that we found pathkeys for this aggregate */
3251 aggindexes = bms_add_member(aggindexes, i);
3252 }
3253 else
3254 {
3255 /* now look for a stronger set of matching pathkeys */
3256
3257 /* include the GROUP BY pathkeys, if they exist */
3258 if (grouppathkeys != NIL)
3259 pathkeys = append_pathkeys(list_copy(grouppathkeys),
3260 pathkeys);
3261
3262 /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3263 switch (compare_pathkeys(currpathkeys, pathkeys))
3264 {
3265 case PATHKEYS_BETTER2:
3266 /* 'pathkeys' are stronger, use these ones instead */
3267 currpathkeys = pathkeys;
3268 /* FALLTHROUGH */
3269
3270 case PATHKEYS_BETTER1:
3271 /* 'pathkeys' are less strict */
3272 /* FALLTHROUGH */
3273
3274 case PATHKEYS_EQUAL:
3275 /* mark this aggregate as covered by 'currpathkeys' */
3276 aggindexes = bms_add_member(aggindexes, i);
3277 break;
3278
3279 case PATHKEYS_DIFFERENT:
3280 break;
3281 }
3282 }
3283 }
3284
3285 /* remove the aggregates that we've just processed */
3286 unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3287
3288 /*
3289 * If this pass included more aggregates than the previous best then
3290 * use these ones as the best set.
3291 */
3292 if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3293 {
3294 bestaggs = aggindexes;
3295 bestpathkeys = currpathkeys;
3296 }
3297 }
3298
3299 /*
3300 * If we found any ordered aggregates, update root->group_pathkeys to add
3301 * the best set of aggregate pathkeys. Note that bestpathkeys includes
3302 * the original GROUP BY pathkeys already.
3303 */
3304 if (bestpathkeys != NIL)
3305 root->group_pathkeys = bestpathkeys;
3306
3307 /*
3308 * Now that we've found the best set of aggregates we can set the
3309 * presorted flag to indicate to the executor that it needn't bother
3310 * performing a sort for these Aggrefs. We're able to do this now as
3311 * there's no chance of a Hash Aggregate plan as create_grouping_paths
3312 * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3313 * of ordered aggregates.
3314 */
3315 i = -1;
3316 while ((i = bms_next_member(bestaggs, i)) >= 0)
3317 {
3318 AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3319
3320 foreach(lc, agginfo->aggrefs)
3321 {
3322 Aggref *aggref = lfirst_node(Aggref, lc);
3323
3324 aggref->aggpresorted = true;
3325 }
3326 }
3327}
3328
3329/*
3330 * Compute query_pathkeys and other pathkeys during plan generation
3331 */
3332static void
3333standard_qp_callback(PlannerInfo *root, void *extra)
3334{
3335 Query *parse = root->parse;
3336 standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3337 List *tlist = root->processed_tlist;
3338 List *activeWindows = qp_extra->activeWindows;
3339
3340 /*
3341 * Calculate pathkeys that represent grouping/ordering and/or ordered
3342 * aggregate requirements.
3343 */
3344 if (qp_extra->gset_data)
3345 {
3346 /*
3347 * With grouping sets, just use the first RollupData's groupClause. We
3348 * don't make any effort to optimize grouping clauses when there are
3349 * grouping sets, nor can we combine aggregate ordering keys with
3350 * grouping.
3351 */
3352 List *rollups = qp_extra->gset_data->rollups;
3353 List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3354
3355 if (grouping_is_sortable(groupClause))
3356 {
3357 bool sortable;
3358
3359 /*
3360 * The groupClause is logically below the grouping step. So if
3361 * there is an RTE entry for the grouping step, we need to remove
3362 * its RT index from the sort expressions before we make PathKeys
3363 * for them.
3364 */
3365 root->group_pathkeys =
3366 make_pathkeys_for_sortclauses_extended(root,
3367 &groupClause,
3368 tlist,
3369 false,
3370 parse->hasGroupRTE,
3371 &sortable,
3372 false);
3373 Assert(sortable);
3374 root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3375 }
3376 else
3377 {
3378 root->group_pathkeys = NIL;
3379 root->num_groupby_pathkeys = 0;
3380 }
3381 }
3382 else if (parse->groupClause || root->numOrderedAggs > 0)
3383 {
3384 /*
3385 * With a plain GROUP BY list, we can remove any grouping items that
3386 * are proven redundant by EquivalenceClass processing. For example,
3387 * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3388 * especially common cases, but they're nearly free to detect. Note
3389 * that we remove redundant items from processed_groupClause but not
3390 * the original parse->groupClause.
3391 */
3392 bool sortable;
3393
3394 /*
3395 * Convert group clauses into pathkeys. Set the ec_sortref field of
3396 * EquivalenceClass'es if it's not set yet.
3397 */
3398 root->group_pathkeys =
3399 make_pathkeys_for_sortclauses_extended(root,
3400 &root->processed_groupClause,
3401 tlist,
3402 true,
3403 false,
3404 &sortable,
3405 true);
3406 if (!sortable)
3407 {
3408 /* Can't sort; no point in considering aggregate ordering either */
3409 root->group_pathkeys = NIL;
3410 root->num_groupby_pathkeys = 0;
3411 }
3412 else
3413 {
3414 root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3415 /* If we have ordered aggs, consider adding onto group_pathkeys */
3416 if (root->numOrderedAggs > 0)
3417 adjust_group_pathkeys_for_groupagg(root);
3418 }
3419 }
3420 else
3421 {
3422 root->group_pathkeys = NIL;
3423 root->num_groupby_pathkeys = 0;
3424 }
3425
3426 /* We consider only the first (bottom) window in pathkeys logic */
3427 if (activeWindows != NIL)
3428 {
3429 WindowClause *wc = linitial_node(WindowClause, activeWindows);
3430
3431 root->window_pathkeys = make_pathkeys_for_window(root,
3432 wc,
3433 tlist);
3434 }
3435 else
3436 root->window_pathkeys = NIL;
3437
3438 /*
3439 * As with GROUP BY, we can discard any DISTINCT items that are proven
3440 * redundant by EquivalenceClass processing. The non-redundant list is
3441 * kept in root->processed_distinctClause, leaving the original
3442 * parse->distinctClause alone.
3443 */
3444 if (parse->distinctClause)
3445 {
3446 bool sortable;
3447
3448 /* Make a copy since pathkey processing can modify the list */
3449 root->processed_distinctClause = list_copy(parse->distinctClause);
3450 root->distinct_pathkeys =
3451 make_pathkeys_for_sortclauses_extended(root,
3452 &root->processed_distinctClause,
3453 tlist,
3454 true,
3455 false,
3456 &sortable,
3457 false);
3458 if (!sortable)
3459 root->distinct_pathkeys = NIL;
3460 }
3461 else
3462 root->distinct_pathkeys = NIL;
3463
3464 root->sort_pathkeys =
3465 make_pathkeys_for_sortclauses(root,
3466 parse->sortClause,
3467 tlist);
3468
3469 /* setting setop_pathkeys might be useful to the union planner */
3470 if (qp_extra->setop != NULL)
3471 {
3472 List *groupClauses;
3473 bool sortable;
3474
3475 groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3476
3477 root->setop_pathkeys =
3478 make_pathkeys_for_sortclauses_extended(root,
3479 &groupClauses,
3480 tlist,
3481 false,
3482 false,
3483 &sortable,
3484 false);
3485 if (!sortable)
3486 root->setop_pathkeys = NIL;
3487 }
3488 else
3489 root->setop_pathkeys = NIL;
3490
3491 /*
3492 * Figure out whether we want a sorted result from query_planner.
3493 *
3494 * If we have a sortable GROUP BY clause, then we want a result sorted
3495 * properly for grouping. Otherwise, if we have window functions to
3496 * evaluate, we try to sort for the first window. Otherwise, if there's a
3497 * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3498 * we try to produce output that's sufficiently well sorted for the
3499 * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3500 * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3501 * for a set operation which can benefit from presorted results and have a
3502 * sortable targetlist, we want to sort by the target list.
3503 *
3504 * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3505 * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3506 * that might just leave us failing to exploit an available sort order at
3507 * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3508 * much easier, since we know that the parser ensured that one is a
3509 * superset of the other.
3510 */
3511 if (root->group_pathkeys)
3512 root->query_pathkeys = root->group_pathkeys;
3513 else if (root->window_pathkeys)
3514 root->query_pathkeys = root->window_pathkeys;
3515 else if (list_length(root->distinct_pathkeys) >
3516 list_length(root->sort_pathkeys))
3517 root->query_pathkeys = root->distinct_pathkeys;
3518 else if (root->sort_pathkeys)
3519 root->query_pathkeys = root->sort_pathkeys;
3520 else if (root->setop_pathkeys != NIL)
3521 root->query_pathkeys = root->setop_pathkeys;
3522 else
3523 root->query_pathkeys = NIL;
3524}
3525
3526/*
3527 * Estimate number of groups produced by grouping clauses (1 if not grouping)
3528 *
3529 * path_rows: number of output rows from scan/join step
3530 * gd: grouping sets data including list of grouping sets and their clauses
3531 * target_list: target list containing group clause references
3532 *
3533 * If doing grouping sets, we also annotate the gsets data with the estimates
3534 * for each set and each individual rollup list, with a view to later
3535 * determining whether some combination of them could be hashed instead.
3536 */
3537static double
3538get_number_of_groups(PlannerInfo *root,
3539 double path_rows,
3540 grouping_sets_data *gd,
3541 List *target_list)
3542{
3543 Query *parse = root->parse;
3544 double dNumGroups;
3545
3546 if (parse->groupClause)
3547 {
3548 List *groupExprs;
3549
3550 if (parse->groupingSets)
3551 {
3552 /* Add up the estimates for each grouping set */
3553 ListCell *lc;
3554
3555 Assert(gd); /* keep Coverity happy */
3556
3557 dNumGroups = 0;
3558
3559 foreach(lc, gd->rollups)
3560 {
3561 RollupData *rollup = lfirst_node(RollupData, lc);
3562 ListCell *lc2;
3563 ListCell *lc3;
3564
3565 groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3566 target_list);
3567
3568 rollup->numGroups = 0.0;
3569
3570 forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3571 {
3572 List *gset = (List *) lfirst(lc2);
3573 GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3574 double numGroups = estimate_num_groups(root,
3575 groupExprs,
3576 path_rows,
3577 &gset,
3578 NULL);
3579
3580 gs->numGroups = numGroups;
3581 rollup->numGroups += numGroups;
3582 }
3583
3584 dNumGroups += rollup->numGroups;
3585 }
3586
3587 if (gd->hash_sets_idx)
3588 {
3589 ListCell *lc2;
3590
3591 gd->dNumHashGroups = 0;
3592
3593 groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3594 target_list);
3595
3596 forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3597 {
3598 List *gset = (List *) lfirst(lc);
3599 GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3600 double numGroups = estimate_num_groups(root,
3601 groupExprs,
3602 path_rows,
3603 &gset,
3604 NULL);
3605
3606 gs->numGroups = numGroups;
3607 gd->dNumHashGroups += numGroups;
3608 }
3609
3610 dNumGroups += gd->dNumHashGroups;
3611 }
3612 }
3613 else
3614 {
3615 /* Plain GROUP BY -- estimate based on optimized groupClause */
3616 groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3617 target_list);
3618
3619 dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3620 NULL, NULL);
3621 }
3622 }
3623 else if (parse->groupingSets)
3624 {
3625 /* Empty grouping sets ... one result row for each one */
3626 dNumGroups = list_length(parse->groupingSets);
3627 }
3628 else if (parse->hasAggs || root->hasHavingQual)
3629 {
3630 /* Plain aggregation, one result row */
3631 dNumGroups = 1;
3632 }
3633 else
3634 {
3635 /* Not grouping */
3636 dNumGroups = 1;
3637 }
3638
3639 return dNumGroups;
3640}
3641
3642/*
3643 * create_grouping_paths
3644 *
3645 * Build a new upperrel containing Paths for grouping and/or aggregation.
3646 * Along the way, we also build an upperrel for Paths which are partially
3647 * grouped and/or aggregated. A partially grouped and/or aggregated path
3648 * needs a FinalizeAggregate node to complete the aggregation. Currently,
3649 * the only partially grouped paths we build are also partial paths; that
3650 * is, they need a Gather and then a FinalizeAggregate.
3651 *
3652 * input_rel: contains the source-data Paths
3653 * target: the pathtarget for the result Paths to compute
3654 * gd: grouping sets data including list of grouping sets and their clauses
3655 *
3656 * Note: all Paths in input_rel are expected to return the target computed
3657 * by make_group_input_target.
3658 */
3659static RelOptInfo *
3660create_grouping_paths(PlannerInfo *root,
3661 RelOptInfo *input_rel,
3662 PathTarget *target,
3663 bool target_parallel_safe,
3664 grouping_sets_data *gd)
3665{
3666 Query *parse = root->parse;
3667 RelOptInfo *grouped_rel;
3668 RelOptInfo *partially_grouped_rel;
3669 AggClauseCosts agg_costs;
3670
3671 MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3672 get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3673
3674 /*
3675 * Create grouping relation to hold fully aggregated grouping and/or
3676 * aggregation paths.
3677 */
3678 grouped_rel = make_grouping_rel(root, input_rel, target,
3679 target_parallel_safe, parse->havingQual);
3680
3681 /*
3682 * Create either paths for a degenerate grouping or paths for ordinary
3683 * grouping, as appropriate.
3684 */
3685 if (is_degenerate_grouping(root))
3686 create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3687 else
3688 {
3689 int flags = 0;
3690 GroupPathExtraData extra;
3691
3692 /*
3693 * Determine whether it's possible to perform sort-based
3694 * implementations of grouping. (Note that if processed_groupClause
3695 * is empty, grouping_is_sortable() is trivially true, and all the
3696 * pathkeys_contained_in() tests will succeed too, so that we'll
3697 * consider every surviving input path.)
3698 *
3699 * If we have grouping sets, we might be able to sort some but not all
3700 * of them; in this case, we need can_sort to be true as long as we
3701 * must consider any sorted-input plan.
3702 */
3703 if ((gd && gd->rollups != NIL)
3704 || grouping_is_sortable(root->processed_groupClause))
3705 flags |= GROUPING_CAN_USE_SORT;
3706
3707 /*
3708 * Determine whether we should consider hash-based implementations of
3709 * grouping.
3710 *
3711 * Hashed aggregation only applies if we're grouping. If we have
3712 * grouping sets, some groups might be hashable but others not; in
3713 * this case we set can_hash true as long as there is nothing globally
3714 * preventing us from hashing (and we should therefore consider plans
3715 * with hashes).
3716 *
3717 * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3718 * BY aggregates. (Doing so would imply storing *all* the input
3719 * values in the hash table, and/or running many sorts in parallel,
3720 * either of which seems like a certain loser.) We similarly don't
3721 * support ordered-set aggregates in hashed aggregation, but that case
3722 * is also included in the numOrderedAggs count.
3723 *
3724 * Note: grouping_is_hashable() is much more expensive to check than
3725 * the other gating conditions, so we want to do it last.
3726 */
3727 if ((parse->groupClause != NIL &&
3728 root->numOrderedAggs == 0 &&
3729 (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3730 flags |= GROUPING_CAN_USE_HASH;
3731
3732 /*
3733 * Determine whether partial aggregation is possible.
3734 */
3735 if (can_partial_agg(root))
3736 flags |= GROUPING_CAN_PARTIAL_AGG;
3737
3738 extra.flags = flags;
3739 extra.target_parallel_safe = target_parallel_safe;
3740 extra.havingQual = parse->havingQual;
3741 extra.targetList = parse->targetList;
3742 extra.partial_costs_set = false;
3743
3744 /*
3745 * Determine whether partitionwise aggregation is in theory possible.
3746 * It can be disabled by the user, and for now, we don't try to
3747 * support grouping sets. create_ordinary_grouping_paths() will check
3748 * additional conditions, such as whether input_rel is partitioned.
3749 */
3750 if (enable_partitionwise_aggregate && !parse->groupingSets)
3751 extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3752 else
3753 extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3754
3755 create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3756 &agg_costs, gd, &extra,
3757 &partially_grouped_rel);
3758 }
3759
3760 set_cheapest(grouped_rel);
3761 return grouped_rel;
3762}
3763
3764/*
3765 * make_grouping_rel
3766 *
3767 * Create a new grouping rel and set basic properties.
3768 *
3769 * input_rel represents the underlying scan/join relation.
3770 * target is the output expected from the grouping relation.
3771 */
3772static RelOptInfo *
3773make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3774 PathTarget *target, bool target_parallel_safe,
3775 Node *havingQual)
3776{
3777 RelOptInfo *grouped_rel;
3778
3779 if (IS_OTHER_REL(input_rel))
3780 {
3781 grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3782 input_rel->relids);
3783 grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3784 }
3785 else
3786 {
3787 /*
3788 * By tradition, the relids set for the main grouping relation is
3789 * NULL. (This could be changed, but might require adjustments
3790 * elsewhere.)
3791 */
3792 grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3793 }
3794
3795 /* Set target. */
3796 grouped_rel->reltarget = target;
3797
3798 /*
3799 * If the input relation is not parallel-safe, then the grouped relation
3800 * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3801 * target list and HAVING quals are parallel-safe.
3802 */
3803 if (input_rel->consider_parallel && target_parallel_safe &&
3804 is_parallel_safe(root, (Node *) havingQual))
3805 grouped_rel->consider_parallel = true;
3806
3807 /*
3808 * If the input rel belongs to a single FDW, so does the grouped rel.
3809 */
3810 grouped_rel->serverid = input_rel->serverid;
3811 grouped_rel->userid = input_rel->userid;
3812 grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3813 grouped_rel->fdwroutine = input_rel->fdwroutine;
3814
3815 return grouped_rel;
3816}
3817
3818/*
3819 * is_degenerate_grouping
3820 *
3821 * A degenerate grouping is one in which the query has a HAVING qual and/or
3822 * grouping sets, but no aggregates and no GROUP BY (which implies that the
3823 * grouping sets are all empty).
3824 */
3825static bool
3826is_degenerate_grouping(PlannerInfo *root)
3827{
3828 Query *parse = root->parse;
3829
3830 return (root->hasHavingQual || parse->groupingSets) &&
3831 !parse->hasAggs && parse->groupClause == NIL;
3832}
3833
3834/*
3835 * create_degenerate_grouping_paths
3836 *
3837 * When the grouping is degenerate (see is_degenerate_grouping), we are
3838 * supposed to emit either zero or one row for each grouping set depending on
3839 * whether HAVING succeeds. Furthermore, there cannot be any variables in
3840 * either HAVING or the targetlist, so we actually do not need the FROM table
3841 * at all! We can just throw away the plan-so-far and generate a Result node.
3842 * This is a sufficiently unusual corner case that it's not worth contorting
3843 * the structure of this module to avoid having to generate the earlier paths
3844 * in the first place.
3845 */
3846static void
3847create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3848 RelOptInfo *grouped_rel)
3849{
3850 Query *parse = root->parse;
3851 int nrows;
3852 Path *path;
3853
3854 nrows = list_length(parse->groupingSets);
3855 if (nrows > 1)
3856 {
3857 /*
3858 * Doesn't seem worthwhile writing code to cons up a generate_series
3859 * or a values scan to emit multiple rows. Instead just make N clones
3860 * and append them. (With a volatile HAVING clause, this means you
3861 * might get between 0 and N output rows. Offhand I think that's
3862 * desired.)
3863 */
3864 List *paths = NIL;
3865
3866 while (--nrows >= 0)
3867 {
3868 path = (Path *)
3869 create_group_result_path(root, grouped_rel,
3870 grouped_rel->reltarget,
3871 (List *) parse->havingQual);
3872 paths = lappend(paths, path);
3873 }
3874 path = (Path *)
3875 create_append_path(root,
3876 grouped_rel,
3877 paths,
3878 NIL,
3879 NIL,
3880 NULL,
3881 0,
3882 false,
3883 -1);
3884 }
3885 else
3886 {
3887 /* No grouping sets, or just one, so one output row */
3888 path = (Path *)
3889 create_group_result_path(root, grouped_rel,
3890 grouped_rel->reltarget,
3891 (List *) parse->havingQual);
3892 }
3893
3894 add_path(grouped_rel, path);
3895}
3896
3897/*
3898 * create_ordinary_grouping_paths
3899 *
3900 * Create grouping paths for the ordinary (that is, non-degenerate) case.
3901 *
3902 * We need to consider sorted and hashed aggregation in the same function,
3903 * because otherwise (1) it would be harder to throw an appropriate error
3904 * message if neither way works, and (2) we should not allow hashtable size
3905 * considerations to dissuade us from using hashing if sorting is not possible.
3906 *
3907 * *partially_grouped_rel_p will be set to the partially grouped rel which this
3908 * function creates, or to NULL if it doesn't create one.
3909 */
3910static void
3911create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3912 RelOptInfo *grouped_rel,
3913 const AggClauseCosts *agg_costs,
3914 grouping_sets_data *gd,
3915 GroupPathExtraData *extra,
3916 RelOptInfo **partially_grouped_rel_p)
3917{
3918 Path *cheapest_path = input_rel->cheapest_total_path;
3919 RelOptInfo *partially_grouped_rel = NULL;
3920 double dNumGroups;
3921 PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
3922
3923 /*
3924 * If this is the topmost grouping relation or if the parent relation is
3925 * doing some form of partitionwise aggregation, then we may be able to do
3926 * it at this level also. However, if the input relation is not
3927 * partitioned, partitionwise aggregate is impossible.
3928 */
3929 if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
3930 IS_PARTITIONED_REL(input_rel))
3931 {
3932 /*
3933 * If this is the topmost relation or if the parent relation is doing
3934 * full partitionwise aggregation, then we can do full partitionwise
3935 * aggregation provided that the GROUP BY clause contains all of the
3936 * partitioning columns at this level and the collation used by GROUP
3937 * BY matches the partitioning collation. Otherwise, we can do at
3938 * most partial partitionwise aggregation. But if partial aggregation
3939 * is not supported in general then we can't use it for partitionwise
3940 * aggregation either.
3941 *
3942 * Check parse->groupClause not processed_groupClause, because it's
3943 * okay if some of the partitioning columns were proved redundant.
3944 */
3945 if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
3946 group_by_has_partkey(input_rel, extra->targetList,
3947 root->parse->groupClause))
3948 patype = PARTITIONWISE_AGGREGATE_FULL;
3949 else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
3950 patype = PARTITIONWISE_AGGREGATE_PARTIAL;
3951 else
3952 patype = PARTITIONWISE_AGGREGATE_NONE;
3953 }
3954
3955 /*
3956 * Before generating paths for grouped_rel, we first generate any possible
3957 * partially grouped paths; that way, later code can easily consider both
3958 * parallel and non-parallel approaches to grouping.
3959 */
3960 if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
3961 {
3962 bool force_rel_creation;
3963
3964 /*
3965 * If we're doing partitionwise aggregation at this level, force
3966 * creation of a partially_grouped_rel so we can add partitionwise
3967 * paths to it.
3968 */
3969 force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
3970
3971 partially_grouped_rel =
3972 create_partial_grouping_paths(root,
3973 grouped_rel,
3974 input_rel,
3975 gd,
3976 extra,
3977 force_rel_creation);
3978 }
3979
3980 /* Set out parameter. */
3981 *partially_grouped_rel_p = partially_grouped_rel;
3982
3983 /* Apply partitionwise aggregation technique, if possible. */
3984 if (patype != PARTITIONWISE_AGGREGATE_NONE)
3985 create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
3986 partially_grouped_rel, agg_costs,
3987 gd, patype, extra);
3988
3989 /* If we are doing partial aggregation only, return. */
3990 if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
3991 {
3992 Assert(partially_grouped_rel);
3993
3994 if (partially_grouped_rel->pathlist)
3995 set_cheapest(partially_grouped_rel);
3996
3997 return;
3998 }
3999
4000 /* Gather any partially grouped partial paths. */
4001 if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4002 {
4003 gather_grouping_paths(root, partially_grouped_rel);
4004 set_cheapest(partially_grouped_rel);
4005 }
4006
4007 /*
4008 * Estimate number of groups.
4009 */
4010 dNumGroups = get_number_of_groups(root,
4011 cheapest_path->rows,
4012 gd,
4013 extra->targetList);
4014
4015 /* Build final grouping paths */
4016 add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4017 partially_grouped_rel, agg_costs, gd,
4018 dNumGroups, extra);
4019
4020 /* Give a helpful error if we failed to find any implementation */
4021 if (grouped_rel->pathlist == NIL)
4022 ereport(ERROR,
4023 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4024 errmsg("could not implement GROUP BY"),
4025 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4026
4027 /*
4028 * If there is an FDW that's responsible for all baserels of the query,
4029 * let it consider adding ForeignPaths.
4030 */
4031 if (grouped_rel->fdwroutine &&
4032 grouped_rel->fdwroutine->GetForeignUpperPaths)
4033 grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4034 input_rel, grouped_rel,
4035 extra);
4036
4037 /* Let extensions possibly add some more paths */
4038 if (create_upper_paths_hook)
4039 (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4040 input_rel, grouped_rel,
4041 extra);
4042}
4043
4044/*
4045 * For a given input path, consider the possible ways of doing grouping sets on
4046 * it, by combinations of hashing and sorting. This can be called multiple
4047 * times, so it's important that it not scribble on input. No result is
4048 * returned, but any generated paths are added to grouped_rel.
4049 */
4050static void
4051consider_groupingsets_paths(PlannerInfo *root,
4052 RelOptInfo *grouped_rel,
4053 Path *path,
4054 bool is_sorted,
4055 bool can_hash,
4056 grouping_sets_data *gd,
4057 const AggClauseCosts *agg_costs,
4058 double dNumGroups)
4059{
4060 Query *parse = root->parse;
4061 Size hash_mem_limit = get_hash_memory_limit();
4062
4063 /*
4064 * If we're not being offered sorted input, then only consider plans that
4065 * can be done entirely by hashing.
4066 *
4067 * We can hash everything if it looks like it'll fit in hash_mem. But if
4068 * the input is actually sorted despite not being advertised as such, we
4069 * prefer to make use of that in order to use less memory.
4070 *
4071 * If none of the grouping sets are sortable, then ignore the hash_mem
4072 * limit and generate a path anyway, since otherwise we'll just fail.
4073 */
4074 if (!is_sorted)
4075 {
4076 List *new_rollups = NIL;
4077 RollupData *unhashed_rollup = NULL;
4078 List *sets_data;
4079 List *empty_sets_data = NIL;
4080 List *empty_sets = NIL;
4081 ListCell *lc;
4082 ListCell *l_start = list_head(gd->rollups);
4083 AggStrategy strat = AGG_HASHED;
4084 double hashsize;
4085 double exclude_groups = 0.0;
4086
4087 Assert(can_hash);
4088
4089 /*
4090 * If the input is coincidentally sorted usefully (which can happen
4091 * even if is_sorted is false, since that only means that our caller
4092 * has set up the sorting for us), then save some hashtable space by
4093 * making use of that. But we need to watch out for degenerate cases:
4094 *
4095 * 1) If there are any empty grouping sets, then group_pathkeys might
4096 * be NIL if all non-empty grouping sets are unsortable. In this case,
4097 * there will be a rollup containing only empty groups, and the
4098 * pathkeys_contained_in test is vacuously true; this is ok.
4099 *
4100 * XXX: the above relies on the fact that group_pathkeys is generated
4101 * from the first rollup. If we add the ability to consider multiple
4102 * sort orders for grouping input, this assumption might fail.
4103 *
4104 * 2) If there are no empty sets and only unsortable sets, then the
4105 * rollups list will be empty (and thus l_start == NULL), and
4106 * group_pathkeys will be NIL; we must ensure that the vacuously-true
4107 * pathkeys_contained_in test doesn't cause us to crash.
4108 */
4109 if (l_start != NULL &&
4110 pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4111 {
4112 unhashed_rollup = lfirst_node(RollupData, l_start);
4113 exclude_groups = unhashed_rollup->numGroups;
4114 l_start = lnext(gd->rollups, l_start);
4115 }
4116
4117 hashsize = estimate_hashagg_tablesize(root,
4118 path,
4119 agg_costs,
4120 dNumGroups - exclude_groups);
4121
4122 /*
4123 * gd->rollups is empty if we have only unsortable columns to work
4124 * with. Override hash_mem in that case; otherwise, we'll rely on the
4125 * sorted-input case to generate usable mixed paths.
4126 */
4127 if (hashsize > hash_mem_limit && gd->rollups)
4128 return; /* nope, won't fit */
4129
4130 /*
4131 * We need to burst the existing rollups list into individual grouping
4132 * sets and recompute a groupClause for each set.
4133 */
4134 sets_data = list_copy(gd->unsortable_sets);
4135
4136 for_each_cell(lc, gd->rollups, l_start)
4137 {
4138 RollupData *rollup = lfirst_node(RollupData, lc);
4139
4140 /*
4141 * If we find an unhashable rollup that's not been skipped by the
4142 * "actually sorted" check above, we can't cope; we'd need sorted
4143 * input (with a different sort order) but we can't get that here.
4144 * So bail out; we'll get a valid path from the is_sorted case
4145 * instead.
4146 *
4147 * The mere presence of empty grouping sets doesn't make a rollup
4148 * unhashable (see preprocess_grouping_sets), we handle those
4149 * specially below.
4150 */
4151 if (!rollup->hashable)
4152 return;
4153
4154 sets_data = list_concat(sets_data, rollup->gsets_data);
4155 }
4156 foreach(lc, sets_data)
4157 {
4158 GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4159 List *gset = gs->set;
4160 RollupData *rollup;
4161
4162 if (gset == NIL)
4163 {
4164 /* Empty grouping sets can't be hashed. */
4165 empty_sets_data = lappend(empty_sets_data, gs);
4166 empty_sets = lappend(empty_sets, NIL);
4167 }
4168 else
4169 {
4170 rollup = makeNode(RollupData);
4171
4172 rollup->groupClause = preprocess_groupclause(root, gset);
4173 rollup->gsets_data = list_make1(gs);
4174 rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4175 rollup->gsets_data,
4176 gd->tleref_to_colnum_map);
4177 rollup->numGroups = gs->numGroups;
4178 rollup->hashable = true;
4179 rollup->is_hashed = true;
4180 new_rollups = lappend(new_rollups, rollup);
4181 }
4182 }
4183
4184 /*
4185 * If we didn't find anything nonempty to hash, then bail. We'll
4186 * generate a path from the is_sorted case.
4187 */
4188 if (new_rollups == NIL)
4189 return;
4190
4191 /*
4192 * If there were empty grouping sets they should have been in the
4193 * first rollup.
4194 */
4195 Assert(!unhashed_rollup || !empty_sets);
4196
4197 if (unhashed_rollup)
4198 {
4199 new_rollups = lappend(new_rollups, unhashed_rollup);
4200 strat = AGG_MIXED;
4201 }
4202 else if (empty_sets)
4203 {
4204 RollupData *rollup = makeNode(RollupData);
4205
4206 rollup->groupClause = NIL;
4207 rollup->gsets_data = empty_sets_data;
4208 rollup->gsets = empty_sets;
4209 rollup->numGroups = list_length(empty_sets);
4210 rollup->hashable = false;
4211 rollup->is_hashed = false;
4212 new_rollups = lappend(new_rollups, rollup);
4213 strat = AGG_MIXED;
4214 }
4215
4216 add_path(grouped_rel, (Path *)
4217 create_groupingsets_path(root,
4218 grouped_rel,
4219 path,
4220 (List *) parse->havingQual,
4221 strat,
4222 new_rollups,
4223 agg_costs));
4224 return;
4225 }
4226
4227 /*
4228 * If we have sorted input but nothing we can do with it, bail.
4229 */
4230 if (gd->rollups == NIL)
4231 return;
4232
4233 /*
4234 * Given sorted input, we try and make two paths: one sorted and one mixed
4235 * sort/hash. (We need to try both because hashagg might be disabled, or
4236 * some columns might not be sortable.)
4237 *
4238 * can_hash is passed in as false if some obstacle elsewhere (such as
4239 * ordered aggs) means that we shouldn't consider hashing at all.
4240 */
4241 if (can_hash && gd->any_hashable)
4242 {
4243 List *rollups = NIL;
4244 List *hash_sets = list_copy(gd->unsortable_sets);
4245 double availspace = hash_mem_limit;
4246 ListCell *lc;
4247
4248 /*
4249 * Account first for space needed for groups we can't sort at all.
4250 */
4251 availspace -= estimate_hashagg_tablesize(root,
4252 path,
4253 agg_costs,
4254 gd->dNumHashGroups);
4255
4256 if (availspace > 0 && list_length(gd->rollups) > 1)
4257 {
4258 double scale;
4259 int num_rollups = list_length(gd->rollups);
4260 int k_capacity;
4261 int *k_weights = palloc(num_rollups * sizeof(int));
4262 Bitmapset *hash_items = NULL;
4263 int i;
4264
4265 /*
4266 * We treat this as a knapsack problem: the knapsack capacity
4267 * represents hash_mem, the item weights are the estimated memory
4268 * usage of the hashtables needed to implement a single rollup,
4269 * and we really ought to use the cost saving as the item value;
4270 * however, currently the costs assigned to sort nodes don't
4271 * reflect the comparison costs well, and so we treat all items as
4272 * of equal value (each rollup we hash instead saves us one sort).
4273 *
4274 * To use the discrete knapsack, we need to scale the values to a
4275 * reasonably small bounded range. We choose to allow a 5% error
4276 * margin; we have no more than 4096 rollups in the worst possible
4277 * case, which with a 5% error margin will require a bit over 42MB
4278 * of workspace. (Anyone wanting to plan queries that complex had
4279 * better have the memory for it. In more reasonable cases, with
4280 * no more than a couple of dozen rollups, the memory usage will
4281 * be negligible.)
4282 *
4283 * k_capacity is naturally bounded, but we clamp the values for
4284 * scale and weight (below) to avoid overflows or underflows (or
4285 * uselessly trying to use a scale factor less than 1 byte).
4286 */
4287 scale = Max(availspace / (20.0 * num_rollups), 1.0);
4288 k_capacity = (int) floor(availspace / scale);
4289
4290 /*
4291 * We leave the first rollup out of consideration since it's the
4292 * one that matches the input sort order. We assign indexes "i"
4293 * to only those entries considered for hashing; the second loop,
4294 * below, must use the same condition.
4295 */
4296 i = 0;
4297 for_each_from(lc, gd->rollups, 1)
4298 {
4299 RollupData *rollup = lfirst_node(RollupData, lc);
4300
4301 if (rollup->hashable)
4302 {
4303 double sz = estimate_hashagg_tablesize(root,
4304 path,
4305 agg_costs,
4306 rollup->numGroups);
4307
4308 /*
4309 * If sz is enormous, but hash_mem (and hence scale) is
4310 * small, avoid integer overflow here.
4311 */
4312 k_weights[i] = (int) Min(floor(sz / scale),
4313 k_capacity + 1.0);
4314 ++i;
4315 }
4316 }
4317
4318 /*
4319 * Apply knapsack algorithm; compute the set of items which
4320 * maximizes the value stored (in this case the number of sorts
4321 * saved) while keeping the total size (approximately) within
4322 * capacity.
4323 */
4324 if (i > 0)
4325 hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4326
4327 if (!bms_is_empty(hash_items))
4328 {
4329 rollups = list_make1(linitial(gd->rollups));
4330
4331 i = 0;
4332 for_each_from(lc, gd->rollups, 1)
4333 {
4334 RollupData *rollup = lfirst_node(RollupData, lc);
4335
4336 if (rollup->hashable)
4337 {
4338 if (bms_is_member(i, hash_items))
4339 hash_sets = list_concat(hash_sets,
4340 rollup->gsets_data);
4341 else
4342 rollups = lappend(rollups, rollup);
4343 ++i;
4344 }
4345 else
4346 rollups = lappend(rollups, rollup);
4347 }
4348 }
4349 }
4350
4351 if (!rollups && hash_sets)
4352 rollups = list_copy(gd->rollups);
4353
4354 foreach(lc, hash_sets)
4355 {
4356 GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4357 RollupData *rollup = makeNode(RollupData);
4358
4359 Assert(gs->set != NIL);
4360
4361 rollup->groupClause = preprocess_groupclause(root, gs->set);
4362 rollup->gsets_data = list_make1(gs);
4363 rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4364 rollup->gsets_data,
4365 gd->tleref_to_colnum_map);
4366 rollup->numGroups = gs->numGroups;
4367 rollup->hashable = true;
4368 rollup->is_hashed = true;
4369 rollups = lcons(rollup, rollups);
4370 }
4371
4372 if (rollups)
4373 {
4374 add_path(grouped_rel, (Path *)
4375 create_groupingsets_path(root,
4376 grouped_rel,
4377 path,
4378 (List *) parse->havingQual,
4379 AGG_MIXED,
4380 rollups,
4381 agg_costs));
4382 }
4383 }
4384
4385 /*
4386 * Now try the simple sorted case.
4387 */
4388 if (!gd->unsortable_sets)
4389 add_path(grouped_rel, (Path *)
4390 create_groupingsets_path(root,
4391 grouped_rel,
4392 path,
4393 (List *) parse->havingQual,
4394 AGG_SORTED,
4395 gd->rollups,
4396 agg_costs));
4397}
4398
4399/*
4400 * create_window_paths
4401 *
4402 * Build a new upperrel containing Paths for window-function evaluation.
4403 *
4404 * input_rel: contains the source-data Paths
4405 * input_target: result of make_window_input_target
4406 * output_target: what the topmost WindowAggPath should return
4407 * wflists: result of find_window_functions
4408 * activeWindows: result of select_active_windows
4409 *
4410 * Note: all Paths in input_rel are expected to return input_target.
4411 */
4412static RelOptInfo *
4413create_window_paths(PlannerInfo *root,
4414 RelOptInfo *input_rel,
4415 PathTarget *input_target,
4416 PathTarget *output_target,
4417 bool output_target_parallel_safe,
4418 WindowFuncLists *wflists,
4419 List *activeWindows)
4420{
4421 RelOptInfo *window_rel;
4422 ListCell *lc;
4423
4424 /* For now, do all work in the (WINDOW, NULL) upperrel */
4425 window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4426
4427 /*
4428 * If the input relation is not parallel-safe, then the window relation
4429 * can't be parallel-safe, either. Otherwise, we need to examine the
4430 * target list and active windows for non-parallel-safe constructs.
4431 */
4432 if (input_rel->consider_parallel && output_target_parallel_safe &&
4433 is_parallel_safe(root, (Node *) activeWindows))
4434 window_rel->consider_parallel = true;
4435
4436 /*
4437 * If the input rel belongs to a single FDW, so does the window rel.
4438 */
4439 window_rel->serverid = input_rel->serverid;
4440 window_rel->userid = input_rel->userid;
4441 window_rel->useridiscurrent = input_rel->useridiscurrent;
4442 window_rel->fdwroutine = input_rel->fdwroutine;
4443
4444 /*
4445 * Consider computing window functions starting from the existing
4446 * cheapest-total path (which will likely require a sort) as well as any
4447 * existing paths that satisfy or partially satisfy root->window_pathkeys.
4448 */
4449 foreach(lc, input_rel->pathlist)
4450 {
4451 Path *path = (Path *) lfirst(lc);
4452 int presorted_keys;
4453
4454 if (path == input_rel->cheapest_total_path ||
4455 pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4456 &presorted_keys) ||
4457 presorted_keys > 0)
4458 create_one_window_path(root,
4459 window_rel,
4460 path,
4461 input_target,
4462 output_target,
4463 wflists,
4464 activeWindows);
4465 }
4466
4467 /*
4468 * If there is an FDW that's responsible for all baserels of the query,
4469 * let it consider adding ForeignPaths.
4470 */
4471 if (window_rel->fdwroutine &&
4472 window_rel->fdwroutine->GetForeignUpperPaths)
4473 window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4474 input_rel, window_rel,
4475 NULL);
4476
4477 /* Let extensions possibly add some more paths */
4478 if (create_upper_paths_hook)
4479 (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4480 input_rel, window_rel, NULL);
4481
4482 /* Now choose the best path(s) */
4483 set_cheapest(window_rel);
4484
4485 return window_rel;
4486}
4487
4488/*
4489 * Stack window-function implementation steps atop the given Path, and
4490 * add the result to window_rel.
4491 *
4492 * window_rel: upperrel to contain result
4493 * path: input Path to use (must return input_target)
4494 * input_target: result of make_window_input_target
4495 * output_target: what the topmost WindowAggPath should return
4496 * wflists: result of find_window_functions
4497 * activeWindows: result of select_active_windows
4498 */
4499static void
4500create_one_window_path(PlannerInfo *root,
4501 RelOptInfo *window_rel,
4502 Path *path,
4503 PathTarget *input_target,
4504 PathTarget *output_target,
4505 WindowFuncLists *wflists,
4506 List *activeWindows)
4507{
4508 PathTarget *window_target;
4509 ListCell *l;
4510 List *topqual = NIL;
4511
4512 /*
4513 * Since each window clause could require a different sort order, we stack
4514 * up a WindowAgg node for each clause, with sort steps between them as
4515 * needed. (We assume that select_active_windows chose a good order for
4516 * executing the clauses in.)
4517 *
4518 * input_target should contain all Vars and Aggs needed for the result.
4519 * (In some cases we wouldn't need to propagate all of these all the way
4520 * to the top, since they might only be needed as inputs to WindowFuncs.
4521 * It's probably not worth trying to optimize that though.) It must also
4522 * contain all window partitioning and sorting expressions, to ensure
4523 * they're computed only once at the bottom of the stack (that's critical
4524 * for volatile functions). As we climb up the stack, we'll add outputs
4525 * for the WindowFuncs computed at each level.
4526 */
4527 window_target = input_target;
4528
4529 foreach(l, activeWindows)
4530 {
4531 WindowClause *wc = lfirst_node(WindowClause, l);
4532 List *window_pathkeys;
4533 List *runcondition = NIL;
4534 int presorted_keys;
4535 bool is_sorted;
4536 bool topwindow;
4537 ListCell *lc2;
4538
4539 window_pathkeys = make_pathkeys_for_window(root,
4540 wc,
4541 root->processed_tlist);
4542
4543 is_sorted = pathkeys_count_contained_in(window_pathkeys,
4544 path->pathkeys,
4545 &presorted_keys);
4546
4547 /* Sort if necessary */
4548 if (!is_sorted)
4549 {
4550 /*
4551 * No presorted keys or incremental sort disabled, just perform a
4552 * complete sort.
4553 */
4554 if (presorted_keys == 0 || !enable_incremental_sort)
4555 path = (Path *) create_sort_path(root, window_rel,
4556 path,
4557 window_pathkeys,
4558 -1.0);
4559 else
4560 {
4561 /*
4562 * Since we have presorted keys and incremental sort is
4563 * enabled, just use incremental sort.
4564 */
4565 path = (Path *) create_incremental_sort_path(root,
4566 window_rel,
4567 path,
4568 window_pathkeys,
4569 presorted_keys,
4570 -1.0);
4571 }
4572 }
4573
4574 if (lnext(activeWindows, l))
4575 {
4576 /*
4577 * Add the current WindowFuncs to the output target for this
4578 * intermediate WindowAggPath. We must copy window_target to
4579 * avoid changing the previous path's target.
4580 *
4581 * Note: a WindowFunc adds nothing to the target's eval costs; but
4582 * we do need to account for the increase in tlist width.
4583 */
4584 int64 tuple_width = window_target->width;
4585
4586 window_target = copy_pathtarget(window_target);
4587 foreach(lc2, wflists->windowFuncs[wc->winref])
4588 {
4589 WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4590
4591 add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4592 tuple_width += get_typavgwidth(wfunc->wintype, -1);
4593 }
4594 window_target->width = clamp_width_est(tuple_width);
4595 }
4596 else
4597 {
4598 /* Install the goal target in the topmost WindowAgg */
4599 window_target = output_target;
4600 }
4601
4602 /* mark the final item in the list as the top-level window */
4603 topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4604
4605 /*
4606 * Collect the WindowFuncRunConditions from each WindowFunc and
4607 * convert them into OpExprs
4608 */
4609 foreach(lc2, wflists->windowFuncs[wc->winref])
4610 {
4611 ListCell *lc3;
4612 WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4613
4614 foreach(lc3, wfunc->runCondition)
4615 {
4616 WindowFuncRunCondition *wfuncrc =
4617 lfirst_node(WindowFuncRunCondition, lc3);
4618 Expr *opexpr;
4619 Expr *leftop;
4620 Expr *rightop;
4621
4622 if (wfuncrc->wfunc_left)
4623 {
4624 leftop = (Expr *) copyObject(wfunc);
4625 rightop = copyObject(wfuncrc->arg);
4626 }
4627 else
4628 {
4629 leftop = copyObject(wfuncrc->arg);
4630 rightop = (Expr *) copyObject(wfunc);
4631 }
4632
4633 opexpr = make_opclause(wfuncrc->opno,
4634 BOOLOID,
4635 false,
4636 leftop,
4637 rightop,
4638 InvalidOid,
4639 wfuncrc->inputcollid);
4640
4641 runcondition = lappend(runcondition, opexpr);
4642
4643 if (!topwindow)
4644 topqual = lappend(topqual, opexpr);
4645 }
4646 }
4647
4648 path = (Path *)
4649 create_windowagg_path(root, window_rel, path, window_target,
4650 wflists->windowFuncs[wc->winref],
4651 runcondition, wc,
4652 topwindow ? topqual : NIL, topwindow);
4653 }
4654
4655 add_path(window_rel, path);
4656}
4657
4658/*
4659 * create_distinct_paths
4660 *
4661 * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4662 *
4663 * input_rel: contains the source-data Paths
4664 * target: the pathtarget for the result Paths to compute
4665 *
4666 * Note: input paths should already compute the desired pathtarget, since
4667 * Sort/Unique won't project anything.
4668 */
4669static RelOptInfo *
4670create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4671 PathTarget *target)
4672{
4673 RelOptInfo *distinct_rel;
4674
4675 /* For now, do all work in the (DISTINCT, NULL) upperrel */
4676 distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4677
4678 /*
4679 * We don't compute anything at this level, so distinct_rel will be
4680 * parallel-safe if the input rel is parallel-safe. In particular, if
4681 * there is a DISTINCT ON (...) clause, any path for the input_rel will
4682 * output those expressions, and will not be parallel-safe unless those
4683 * expressions are parallel-safe.
4684 */
4685 distinct_rel->consider_parallel = input_rel->consider_parallel;
4686
4687 /*
4688 * If the input rel belongs to a single FDW, so does the distinct_rel.
4689 */
4690 distinct_rel->serverid = input_rel->serverid;
4691 distinct_rel->userid = input_rel->userid;
4692 distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4693 distinct_rel->fdwroutine = input_rel->fdwroutine;
4694
4695 /* build distinct paths based on input_rel's pathlist */
4696 create_final_distinct_paths(root, input_rel, distinct_rel);
4697
4698 /* now build distinct paths based on input_rel's partial_pathlist */
4699 create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4700
4701 /* Give a helpful error if we failed to create any paths */
4702 if (distinct_rel->pathlist == NIL)
4703 ereport(ERROR,
4704 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4705 errmsg("could not implement DISTINCT"),
4706 errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4707
4708 /*
4709 * If there is an FDW that's responsible for all baserels of the query,
4710 * let it consider adding ForeignPaths.
4711 */
4712 if (distinct_rel->fdwroutine &&
4713 distinct_rel->fdwroutine->GetForeignUpperPaths)
4714 distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4715 UPPERREL_DISTINCT,
4716 input_rel,
4717 distinct_rel,
4718 NULL);
4719
4720 /* Let extensions possibly add some more paths */
4721 if (create_upper_paths_hook)
4722 (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4723 distinct_rel, NULL);
4724
4725 /* Now choose the best path(s) */
4726 set_cheapest(distinct_rel);
4727
4728 return distinct_rel;
4729}
4730
4731/*
4732 * create_partial_distinct_paths
4733 *
4734 * Process 'input_rel' partial paths and add unique/aggregate paths to the
4735 * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4736 * paths on top and add a final unique/aggregate path to remove any duplicate
4737 * produced from combining rows from parallel workers.
4738 */
4739static void
4740create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4741 RelOptInfo *final_distinct_rel,
4742 PathTarget *target)
4743{
4744 RelOptInfo *partial_distinct_rel;
4745 Query *parse;
4746 List *distinctExprs;
4747 double numDistinctRows;
4748 Path *cheapest_partial_path;
4749 ListCell *lc;
4750
4751 /* nothing to do when there are no partial paths in the input rel */
4752 if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4753 return;
4754
4755 parse = root->parse;
4756
4757 /* can't do parallel DISTINCT ON */
4758 if (parse->hasDistinctOn)
4759 return;
4760
4761 partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4762 NULL);
4763 partial_distinct_rel->reltarget = target;
4764 partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4765
4766 /*
4767 * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4768 */
4769 partial_distinct_rel->serverid = input_rel->serverid;
4770 partial_distinct_rel->userid = input_rel->userid;
4771 partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4772 partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4773
4774 cheapest_partial_path = linitial(input_rel->partial_pathlist);
4775
4776 distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4777 parse->targetList);
4778
4779 /* estimate how many distinct rows we'll get from each worker */
4780 numDistinctRows = estimate_num_groups(root, distinctExprs,
4781 cheapest_partial_path->rows,
4782 NULL, NULL);
4783
4784 /*
4785 * Try sorting the cheapest path and incrementally sorting any paths with
4786 * presorted keys and put a unique paths atop of those. We'll also
4787 * attempt to reorder the required pathkeys to match the input path's
4788 * pathkeys as much as possible, in hopes of avoiding a possible need to
4789 * re-sort.
4790 */
4791 if (grouping_is_sortable(root->processed_distinctClause))
4792 {
4793 foreach(lc, input_rel->partial_pathlist)
4794 {
4795 Path *input_path = (Path *) lfirst(lc);
4796 Path *sorted_path;
4797 List *useful_pathkeys_list = NIL;
4798
4799 useful_pathkeys_list =
4800 get_useful_pathkeys_for_distinct(root,
4801 root->distinct_pathkeys,
4802 input_path->pathkeys);
4803 Assert(list_length(useful_pathkeys_list) > 0);
4804
4805 foreach_node(List, useful_pathkeys, useful_pathkeys_list)
4806 {
4807 sorted_path = make_ordered_path(root,
4808 partial_distinct_rel,
4809 input_path,
4810 cheapest_partial_path,
4811 useful_pathkeys,
4812 -1.0);
4813
4814 if (sorted_path == NULL)
4815 continue;
4816
4817 /*
4818 * An empty distinct_pathkeys means all tuples have the same
4819 * value for the DISTINCT clause. See
4820 * create_final_distinct_paths()
4821 */
4822 if (root->distinct_pathkeys == NIL)
4823 {
4824 Node *limitCount;
4825
4826 limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
4827 sizeof(int64),
4828 Int64GetDatum(1), false,
4829 FLOAT8PASSBYVAL);
4830
4831 /*
4832 * Apply a LimitPath onto the partial path to restrict the
4833 * tuples from each worker to 1.
4834 * create_final_distinct_paths will need to apply an
4835 * additional LimitPath to restrict this to a single row
4836 * after the Gather node. If the query already has a
4837 * LIMIT clause, then we could end up with three Limit
4838 * nodes in the final plan. Consolidating the top two of
4839 * these could be done, but does not seem worth troubling
4840 * over.
4841 */
4842 add_partial_path(partial_distinct_rel, (Path *)
4843 create_limit_path(root, partial_distinct_rel,
4844 sorted_path,
4845 NULL,
4846 limitCount,
4847 LIMIT_OPTION_COUNT,
4848 0, 1));
4849 }
4850 else
4851 {
4852 add_partial_path(partial_distinct_rel, (Path *)
4853 create_upper_unique_path(root, partial_distinct_rel,
4854 sorted_path,
4855 list_length(root->distinct_pathkeys),
4856 numDistinctRows));
4857 }
4858 }
4859 }
4860 }
4861
4862 /*
4863 * Now try hash aggregate paths, if enabled and hashing is possible. Since
4864 * we're not on the hook to ensure we do our best to create at least one
4865 * path here, we treat enable_hashagg as a hard off-switch rather than the
4866 * slightly softer variant in create_final_distinct_paths.
4867 */
4868 if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
4869 {
4870 add_partial_path(partial_distinct_rel, (Path *)
4871 create_agg_path(root,
4872 partial_distinct_rel,
4873 cheapest_partial_path,
4874 cheapest_partial_path->pathtarget,
4875 AGG_HASHED,
4876 AGGSPLIT_SIMPLE,
4877 root->processed_distinctClause,
4878 NIL,
4879 NULL,
4880 numDistinctRows));
4881 }
4882
4883 /*
4884 * If there is an FDW that's responsible for all baserels of the query,
4885 * let it consider adding ForeignPaths.
4886 */
4887 if (partial_distinct_rel->fdwroutine &&
4888 partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
4889 partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4890 UPPERREL_PARTIAL_DISTINCT,
4891 input_rel,
4892 partial_distinct_rel,
4893 NULL);
4894
4895 /* Let extensions possibly add some more partial paths */
4896 if (create_upper_paths_hook)
4897 (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
4898 input_rel, partial_distinct_rel, NULL);
4899
4900 if (partial_distinct_rel->partial_pathlist != NIL)
4901 {
4902 generate_useful_gather_paths(root, partial_distinct_rel, true);
4903 set_cheapest(partial_distinct_rel);
4904
4905 /*
4906 * Finally, create paths to distinctify the final result. This step
4907 * is needed to remove any duplicates due to combining rows from
4908 * parallel workers.
4909 */
4910 create_final_distinct_paths(root, partial_distinct_rel,
4911 final_distinct_rel);
4912 }
4913}
4914
4915/*
4916 * create_final_distinct_paths
4917 * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
4918 *
4919 * input_rel: contains the source-data paths
4920 * distinct_rel: destination relation for storing created paths
4921 */
4922static RelOptInfo *
4923create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4924 RelOptInfo *distinct_rel)
4925{
4926 Query *parse = root->parse;
4927 Path *cheapest_input_path = input_rel->cheapest_total_path;
4928 double numDistinctRows;
4929 bool allow_hash;
4930
4931 /* Estimate number of distinct rows there will be */
4932 if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4933 root->hasHavingQual)
4934 {
4935 /*
4936 * If there was grouping or aggregation, use the number of input rows
4937 * as the estimated number of DISTINCT rows (ie, assume the input is
4938 * already mostly unique).
4939 */
4940 numDistinctRows = cheapest_input_path->rows;
4941 }
4942 else
4943 {
4944 /*
4945 * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4946 */
4947 List *distinctExprs;
4948
4949 distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4950 parse->targetList);
4951 numDistinctRows = estimate_num_groups(root, distinctExprs,
4952 cheapest_input_path->rows,
4953 NULL, NULL);
4954 }
4955
4956 /*
4957 * Consider sort-based implementations of DISTINCT, if possible.
4958 */
4959 if (grouping_is_sortable(root->processed_distinctClause))
4960 {
4961 /*
4962 * Firstly, if we have any adequately-presorted paths, just stick a
4963 * Unique node on those. We also, consider doing an explicit sort of
4964 * the cheapest input path and Unique'ing that. If any paths have
4965 * presorted keys then we'll create an incremental sort atop of those
4966 * before adding a unique node on the top. We'll also attempt to
4967 * reorder the required pathkeys to match the input path's pathkeys as
4968 * much as possible, in hopes of avoiding a possible need to re-sort.
4969 *
4970 * When we have DISTINCT ON, we must sort by the more rigorous of
4971 * DISTINCT and ORDER BY, else it won't have the desired behavior.
4972 * Also, if we do have to do an explicit sort, we might as well use
4973 * the more rigorous ordering to avoid a second sort later. (Note
4974 * that the parser will have ensured that one clause is a prefix of
4975 * the other.)
4976 */
4977 List *needed_pathkeys;
4978 ListCell *lc;
4979 double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
4980
4981 if (parse->hasDistinctOn &&
4982 list_length(root->distinct_pathkeys) <
4983 list_length(root->sort_pathkeys))
4984 needed_pathkeys = root->sort_pathkeys;
4985 else
4986 needed_pathkeys = root->distinct_pathkeys;
4987
4988 foreach(lc, input_rel->pathlist)
4989 {
4990 Path *input_path = (Path *) lfirst(lc);
4991 Path *sorted_path;
4992 List *useful_pathkeys_list = NIL;
4993
4994 useful_pathkeys_list =
4995 get_useful_pathkeys_for_distinct(root,
4996 needed_pathkeys,
4997 input_path->pathkeys);
4998 Assert(list_length(useful_pathkeys_list) > 0);
4999
5000 foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5001 {
5002 sorted_path = make_ordered_path(root,
5003 distinct_rel,
5004 input_path,
5005 cheapest_input_path,
5006 useful_pathkeys,
5007 limittuples);
5008
5009 if (sorted_path == NULL)
5010 continue;
5011
5012 /*
5013 * distinct_pathkeys may have become empty if all of the
5014 * pathkeys were determined to be redundant. If all of the
5015 * pathkeys are redundant then each DISTINCT target must only
5016 * allow a single value, therefore all resulting tuples must
5017 * be identical (or at least indistinguishable by an equality
5018 * check). We can uniquify these tuples simply by just taking
5019 * the first tuple. All we do here is add a path to do "LIMIT
5020 * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5021 * still have a non-NIL sort_pathkeys list, so we must still
5022 * only do this with paths which are correctly sorted by
5023 * sort_pathkeys.
5024 */
5025 if (root->distinct_pathkeys == NIL)
5026 {
5027 Node *limitCount;
5028
5029 limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5030 sizeof(int64),
5031 Int64GetDatum(1), false,
5032 FLOAT8PASSBYVAL);
5033
5034 /*
5035 * If the query already has a LIMIT clause, then we could
5036 * end up with a duplicate LimitPath in the final plan.
5037 * That does not seem worth troubling over too much.
5038 */
5039 add_path(distinct_rel, (Path *)
5040 create_limit_path(root, distinct_rel, sorted_path,
5041 NULL, limitCount,
5042 LIMIT_OPTION_COUNT, 0, 1));
5043 }
5044 else
5045 {
5046 add_path(distinct_rel, (Path *)
5047 create_upper_unique_path(root, distinct_rel,
5048 sorted_path,
5049 list_length(root->distinct_pathkeys),
5050 numDistinctRows));
5051 }
5052 }
5053 }
5054 }
5055
5056 /*
5057 * Consider hash-based implementations of DISTINCT, if possible.
5058 *
5059 * If we were not able to make any other types of path, we *must* hash or
5060 * die trying. If we do have other choices, there are two things that
5061 * should prevent selection of hashing: if the query uses DISTINCT ON
5062 * (because it won't really have the expected behavior if we hash), or if
5063 * enable_hashagg is off.
5064 *
5065 * Note: grouping_is_hashable() is much more expensive to check than the
5066 * other gating conditions, so we want to do it last.
5067 */
5068 if (distinct_rel->pathlist == NIL)
5069 allow_hash = true; /* we have no alternatives */
5070 else if (parse->hasDistinctOn || !enable_hashagg)
5071 allow_hash = false; /* policy-based decision not to hash */
5072 else
5073 allow_hash = true; /* default */
5074
5075 if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5076 {
5077 /* Generate hashed aggregate path --- no sort needed */
5078 add_path(distinct_rel, (Path *)
5079 create_agg_path(root,
5080 distinct_rel,
5081 cheapest_input_path,
5082 cheapest_input_path->pathtarget,
5083 AGG_HASHED,
5084 AGGSPLIT_SIMPLE,
5085 root->processed_distinctClause,
5086 NIL,
5087 NULL,
5088 numDistinctRows));
5089 }
5090
5091 return distinct_rel;
5092}
5093
5094/*
5095 * get_useful_pathkeys_for_distinct
5096 * Get useful orderings of pathkeys for distinctClause by reordering
5097 * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5098 *
5099 * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5100 * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5101 */
5102static List *
5103get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5104 List *path_pathkeys)
5105{
5106 List *useful_pathkeys_list = NIL;
5107 List *useful_pathkeys = NIL;
5108
5109 /* always include the given 'needed_pathkeys' */
5110 useful_pathkeys_list = lappend(useful_pathkeys_list,
5111 needed_pathkeys);
5112
5113 if (!enable_distinct_reordering)
5114 return useful_pathkeys_list;
5115
5116 /*
5117 * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5118 * that match 'needed_pathkeys', but only up to the longest matching
5119 * prefix.
5120 *
5121 * When we have DISTINCT ON, we must ensure that the resulting pathkey
5122 * list matches initial distinctClause pathkeys; otherwise, it won't have
5123 * the desired behavior.
5124 */
5125 foreach_node(PathKey, pathkey, path_pathkeys)
5126 {
5127 /*
5128 * The PathKey nodes are canonical, so they can be checked for
5129 * equality by simple pointer comparison.
5130 */
5131 if (!list_member_ptr(needed_pathkeys, pathkey))
5132 break;
5133 if (root->parse->hasDistinctOn &&
5134 !list_member_ptr(root->distinct_pathkeys, pathkey))
5135 break;
5136
5137 useful_pathkeys = lappend(useful_pathkeys, pathkey);
5138 }
5139
5140 /* If no match at all, no point in reordering needed_pathkeys */
5141 if (useful_pathkeys == NIL)
5142 return useful_pathkeys_list;
5143
5144 /*
5145 * If not full match, the resulting pathkey list is not useful without
5146 * incremental sort.
5147 */
5148 if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5149 !enable_incremental_sort)
5150 return useful_pathkeys_list;
5151
5152 /* Append the remaining PathKey nodes in needed_pathkeys */
5153 useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5154 needed_pathkeys);
5155
5156 /*
5157 * If the resulting pathkey list is the same as the 'needed_pathkeys',
5158 * just drop it.
5159 */
5160 if (compare_pathkeys(needed_pathkeys,
5161 useful_pathkeys) == PATHKEYS_EQUAL)
5162 return useful_pathkeys_list;
5163
5164 useful_pathkeys_list = lappend(useful_pathkeys_list,
5165 useful_pathkeys);
5166
5167 return useful_pathkeys_list;
5168}
5169
5170/*
5171 * create_ordered_paths
5172 *
5173 * Build a new upperrel containing Paths for ORDER BY evaluation.
5174 *
5175 * All paths in the result must satisfy the ORDER BY ordering.
5176 * The only new paths we need consider are an explicit full sort
5177 * and incremental sort on the cheapest-total existing path.
5178 *
5179 * input_rel: contains the source-data Paths
5180 * target: the output tlist the result Paths must emit
5181 * limit_tuples: estimated bound on the number of output tuples,
5182 * or -1 if no LIMIT or couldn't estimate
5183 *
5184 * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5185 * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5186 */
5187static RelOptInfo *
5188create_ordered_paths(PlannerInfo *root,
5189 RelOptInfo *input_rel,
5190 PathTarget *target,
5191 bool target_parallel_safe,
5192 double limit_tuples)
5193{
5194 Path *cheapest_input_path = input_rel->cheapest_total_path;
5195 RelOptInfo *ordered_rel;
5196 ListCell *lc;
5197
5198 /* For now, do all work in the (ORDERED, NULL) upperrel */
5199 ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5200
5201 /*
5202 * If the input relation is not parallel-safe, then the ordered relation
5203 * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5204 * target list is parallel-safe.
5205 */
5206 if (input_rel->consider_parallel && target_parallel_safe)
5207 ordered_rel->consider_parallel = true;
5208
5209 /*
5210 * If the input rel belongs to a single FDW, so does the ordered_rel.
5211 */
5212 ordered_rel->serverid = input_rel->serverid;
5213 ordered_rel->userid = input_rel->userid;
5214 ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5215 ordered_rel->fdwroutine = input_rel->fdwroutine;
5216
5217 foreach(lc, input_rel->pathlist)
5218 {
5219 Path *input_path = (Path *) lfirst(lc);
5220 Path *sorted_path;
5221 bool is_sorted;
5222 int presorted_keys;
5223
5224 is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5225 input_path->pathkeys, &presorted_keys);
5226
5227 if (is_sorted)
5228 sorted_path = input_path;
5229 else
5230 {
5231 /*
5232 * Try at least sorting the cheapest path and also try
5233 * incrementally sorting any path which is partially sorted
5234 * already (no need to deal with paths which have presorted keys
5235 * when incremental sort is disabled unless it's the cheapest
5236 * input path).
5237 */
5238 if (input_path != cheapest_input_path &&
5239 (presorted_keys == 0 || !enable_incremental_sort))
5240 continue;
5241
5242 /*
5243 * We've no need to consider both a sort and incremental sort.
5244 * We'll just do a sort if there are no presorted keys and an
5245 * incremental sort when there are presorted keys.
5246 */
5247 if (presorted_keys == 0 || !enable_incremental_sort)
5248 sorted_path = (Path *) create_sort_path(root,
5249 ordered_rel,
5250 input_path,
5251 root->sort_pathkeys,
5252 limit_tuples);
5253 else
5254 sorted_path = (Path *) create_incremental_sort_path(root,
5255 ordered_rel,
5256 input_path,
5257 root->sort_pathkeys,
5258 presorted_keys,
5259 limit_tuples);
5260 }
5261
5262 /*
5263 * If the pathtarget of the result path has different expressions from
5264 * the target to be applied, a projection step is needed.
5265 */
5266 if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5267 sorted_path = apply_projection_to_path(root, ordered_rel,
5268 sorted_path, target);
5269
5270 add_path(ordered_rel, sorted_path);
5271 }
5272
5273 /*
5274 * generate_gather_paths() will have already generated a simple Gather
5275 * path for the best parallel path, if any, and the loop above will have
5276 * considered sorting it. Similarly, generate_gather_paths() will also
5277 * have generated order-preserving Gather Merge plans which can be used
5278 * without sorting if they happen to match the sort_pathkeys, and the loop
5279 * above will have handled those as well. However, there's one more
5280 * possibility: it may make sense to sort the cheapest partial path or
5281 * incrementally sort any partial path that is partially sorted according
5282 * to the required output order and then use Gather Merge.
5283 */
5284 if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5285 input_rel->partial_pathlist != NIL)
5286 {
5287 Path *cheapest_partial_path;
5288
5289 cheapest_partial_path = linitial(input_rel->partial_pathlist);
5290
5291 foreach(lc, input_rel->partial_pathlist)
5292 {
5293 Path *input_path = (Path *) lfirst(lc);
5294 Path *sorted_path;
5295 bool is_sorted;
5296 int presorted_keys;
5297 double total_groups;
5298
5299 is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5300 input_path->pathkeys,
5301 &presorted_keys);
5302
5303 if (is_sorted)
5304 continue;
5305
5306 /*
5307 * Try at least sorting the cheapest path and also try
5308 * incrementally sorting any path which is partially sorted
5309 * already (no need to deal with paths which have presorted keys
5310 * when incremental sort is disabled unless it's the cheapest
5311 * partial path).
5312 */
5313 if (input_path != cheapest_partial_path &&
5314 (presorted_keys == 0 || !enable_incremental_sort))
5315 continue;
5316
5317 /*
5318 * We've no need to consider both a sort and incremental sort.
5319 * We'll just do a sort if there are no presorted keys and an
5320 * incremental sort when there are presorted keys.
5321 */
5322 if (presorted_keys == 0 || !enable_incremental_sort)
5323 sorted_path = (Path *) create_sort_path(root,
5324 ordered_rel,
5325 input_path,
5326 root->sort_pathkeys,
5327 limit_tuples);
5328 else