Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * planner.c
4 : * The query optimizer external interface.
5 : *
6 : * Portions Copyright (c) 1996-2026, 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/acl.h"
62 : #include "utils/backend_status.h"
63 : #include "utils/lsyscache.h"
64 : #include "utils/rel.h"
65 : #include "utils/selfuncs.h"
66 :
67 : /* GUC parameters */
68 : double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
69 : int debug_parallel_query = DEBUG_PARALLEL_OFF;
70 : bool parallel_leader_participation = true;
71 : bool enable_distinct_reordering = true;
72 :
73 : /* Hook for plugins to get control in planner() */
74 : planner_hook_type planner_hook = NULL;
75 :
76 : /* Hook for plugins to get control after PlannerGlobal is initialized */
77 : planner_setup_hook_type planner_setup_hook = NULL;
78 :
79 : /* Hook for plugins to get control before PlannerGlobal is discarded */
80 : planner_shutdown_hook_type planner_shutdown_hook = NULL;
81 :
82 : /* Hook for plugins to get control when grouping_planner() plans upper rels */
83 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
84 :
85 :
86 : /* Expression kind codes for preprocess_expression */
87 : #define EXPRKIND_QUAL 0
88 : #define EXPRKIND_TARGET 1
89 : #define EXPRKIND_RTFUNC 2
90 : #define EXPRKIND_RTFUNC_LATERAL 3
91 : #define EXPRKIND_VALUES 4
92 : #define EXPRKIND_VALUES_LATERAL 5
93 : #define EXPRKIND_LIMIT 6
94 : #define EXPRKIND_APPINFO 7
95 : #define EXPRKIND_PHV 8
96 : #define EXPRKIND_TABLESAMPLE 9
97 : #define EXPRKIND_ARBITER_ELEM 10
98 : #define EXPRKIND_TABLEFUNC 11
99 : #define EXPRKIND_TABLEFUNC_LATERAL 12
100 : #define EXPRKIND_GROUPEXPR 13
101 :
102 : /*
103 : * Data specific to grouping sets
104 : */
105 : typedef struct
106 : {
107 : List *rollups;
108 : List *hash_sets_idx;
109 : double dNumHashGroups;
110 : bool any_hashable;
111 : Bitmapset *unsortable_refs;
112 : Bitmapset *unhashable_refs;
113 : List *unsortable_sets;
114 : int *tleref_to_colnum_map;
115 : } grouping_sets_data;
116 :
117 : /*
118 : * Temporary structure for use during WindowClause reordering in order to be
119 : * able to sort WindowClauses on partitioning/ordering prefix.
120 : */
121 : typedef struct
122 : {
123 : WindowClause *wc;
124 : List *uniqueOrder; /* A List of unique ordering/partitioning
125 : * clauses per Window */
126 : } WindowClauseSortData;
127 :
128 : /* Passthrough data for standard_qp_callback */
129 : typedef struct
130 : {
131 : List *activeWindows; /* active windows, if any */
132 : grouping_sets_data *gset_data; /* grouping sets data, if any */
133 : SetOperationStmt *setop; /* parent set operation or NULL if not a
134 : * subquery belonging to a set operation */
135 : } standard_qp_extra;
136 :
137 : /* Local functions */
138 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
139 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
140 : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
141 : SetOperationStmt *setops);
142 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
143 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
144 : int *tleref_to_colnum_map);
145 : static void preprocess_rowmarks(PlannerInfo *root);
146 : static double preprocess_limit(PlannerInfo *root,
147 : double tuple_fraction,
148 : int64 *offset_est, int64 *count_est);
149 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
150 : static List *extract_rollup_sets(List *groupingSets);
151 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
152 : static void standard_qp_callback(PlannerInfo *root, void *extra);
153 : static double get_number_of_groups(PlannerInfo *root,
154 : double path_rows,
155 : grouping_sets_data *gd,
156 : List *target_list);
157 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
158 : RelOptInfo *input_rel,
159 : PathTarget *target,
160 : bool target_parallel_safe,
161 : grouping_sets_data *gd);
162 : static bool is_degenerate_grouping(PlannerInfo *root);
163 : static void create_degenerate_grouping_paths(PlannerInfo *root,
164 : RelOptInfo *input_rel,
165 : RelOptInfo *grouped_rel);
166 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
167 : PathTarget *target, bool target_parallel_safe,
168 : Node *havingQual);
169 : static void create_ordinary_grouping_paths(PlannerInfo *root,
170 : RelOptInfo *input_rel,
171 : RelOptInfo *grouped_rel,
172 : const AggClauseCosts *agg_costs,
173 : grouping_sets_data *gd,
174 : GroupPathExtraData *extra,
175 : RelOptInfo **partially_grouped_rel_p);
176 : static void consider_groupingsets_paths(PlannerInfo *root,
177 : RelOptInfo *grouped_rel,
178 : Path *path,
179 : bool is_sorted,
180 : bool can_hash,
181 : grouping_sets_data *gd,
182 : const AggClauseCosts *agg_costs,
183 : double dNumGroups);
184 : static RelOptInfo *create_window_paths(PlannerInfo *root,
185 : RelOptInfo *input_rel,
186 : PathTarget *input_target,
187 : PathTarget *output_target,
188 : bool output_target_parallel_safe,
189 : WindowFuncLists *wflists,
190 : List *activeWindows);
191 : static void create_one_window_path(PlannerInfo *root,
192 : RelOptInfo *window_rel,
193 : Path *path,
194 : PathTarget *input_target,
195 : PathTarget *output_target,
196 : WindowFuncLists *wflists,
197 : List *activeWindows);
198 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
199 : RelOptInfo *input_rel,
200 : PathTarget *target);
201 : static void create_partial_distinct_paths(PlannerInfo *root,
202 : RelOptInfo *input_rel,
203 : RelOptInfo *final_distinct_rel,
204 : PathTarget *target);
205 : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
206 : RelOptInfo *input_rel,
207 : RelOptInfo *distinct_rel);
208 : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
209 : List *needed_pathkeys,
210 : List *path_pathkeys);
211 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
212 : RelOptInfo *input_rel,
213 : PathTarget *target,
214 : bool target_parallel_safe,
215 : double limit_tuples);
216 : static PathTarget *make_group_input_target(PlannerInfo *root,
217 : PathTarget *final_target);
218 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
219 : PathTarget *grouping_target,
220 : Node *havingQual);
221 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
222 : static void optimize_window_clauses(PlannerInfo *root,
223 : WindowFuncLists *wflists);
224 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
225 : static void name_active_windows(List *activeWindows);
226 : static PathTarget *make_window_input_target(PlannerInfo *root,
227 : PathTarget *final_target,
228 : List *activeWindows);
229 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
230 : List *tlist);
231 : static PathTarget *make_sort_input_target(PlannerInfo *root,
232 : PathTarget *final_target,
233 : bool *have_postponed_srfs);
234 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
235 : List *targets, List *targets_contain_srfs);
236 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
237 : RelOptInfo *grouped_rel,
238 : RelOptInfo *partially_grouped_rel,
239 : const AggClauseCosts *agg_costs,
240 : grouping_sets_data *gd,
241 : GroupPathExtraData *extra);
242 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
243 : RelOptInfo *grouped_rel,
244 : RelOptInfo *input_rel,
245 : grouping_sets_data *gd,
246 : GroupPathExtraData *extra,
247 : bool force_rel_creation);
248 : static Path *make_ordered_path(PlannerInfo *root,
249 : RelOptInfo *rel,
250 : Path *path,
251 : Path *cheapest_path,
252 : List *pathkeys,
253 : double limit_tuples);
254 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
255 : static bool can_partial_agg(PlannerInfo *root);
256 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
257 : RelOptInfo *rel,
258 : List *scanjoin_targets,
259 : List *scanjoin_targets_contain_srfs,
260 : bool scanjoin_target_parallel_safe,
261 : bool tlist_same_exprs);
262 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
263 : RelOptInfo *input_rel,
264 : RelOptInfo *grouped_rel,
265 : RelOptInfo *partially_grouped_rel,
266 : const AggClauseCosts *agg_costs,
267 : grouping_sets_data *gd,
268 : PartitionwiseAggregateType patype,
269 : GroupPathExtraData *extra);
270 : static bool group_by_has_partkey(RelOptInfo *input_rel,
271 : List *targetList,
272 : List *groupClause);
273 : static int common_prefix_cmp(const void *a, const void *b);
274 : static List *generate_setop_child_grouplist(SetOperationStmt *op,
275 : List *targetlist);
276 : static void create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
277 : List *sortPathkeys, List *groupClause,
278 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
279 : static void create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
280 : List *sortPathkeys, List *groupClause,
281 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
282 :
283 :
284 : /*****************************************************************************
285 : *
286 : * Query optimizer entry point
287 : *
288 : * Inputs:
289 : * parse: an analyzed-and-rewritten query tree for an optimizable statement
290 : * query_string: source text for the query tree (used for error reports)
291 : * cursorOptions: bitmask of CURSOR_OPT_XXX flags, see parsenodes.h
292 : * boundParams: passed-in parameter values, or NULL if none
293 : * es: ExplainState if being called from EXPLAIN, else NULL
294 : *
295 : * The result is a PlannedStmt tree.
296 : *
297 : * PARAM_EXTERN Param nodes within the parse tree can be replaced by Consts
298 : * using values from boundParams, if those values are marked PARAM_FLAG_CONST.
299 : * Parameter values not so marked are still relied on for estimation purposes.
300 : *
301 : * The ExplainState pointer is not currently used by the core planner, but it
302 : * is passed through to some planner hooks so that they can report information
303 : * back to EXPLAIN extension hooks.
304 : *
305 : * To support loadable plugins that monitor or modify planner behavior,
306 : * we provide a hook variable that lets a plugin get control before and
307 : * after the standard planning process. The plugin would normally call
308 : * standard_planner().
309 : *
310 : * Note to plugin authors: standard_planner() scribbles on its Query input,
311 : * so you'd better copy that data structure if you want to plan more than once.
312 : *
313 : *****************************************************************************/
314 : PlannedStmt *
315 470866 : planner(Query *parse, const char *query_string, int cursorOptions,
316 : ParamListInfo boundParams, ExplainState *es)
317 : {
318 : PlannedStmt *result;
319 :
320 470866 : if (planner_hook)
321 97266 : result = (*planner_hook) (parse, query_string, cursorOptions,
322 : boundParams, es);
323 : else
324 373600 : result = standard_planner(parse, query_string, cursorOptions,
325 : boundParams, es);
326 :
327 466018 : pgstat_report_plan_id(result->planId, false);
328 :
329 466018 : return result;
330 : }
331 :
332 : PlannedStmt *
333 470866 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
334 : ParamListInfo boundParams, ExplainState *es)
335 : {
336 : PlannedStmt *result;
337 : PlannerGlobal *glob;
338 : double tuple_fraction;
339 : PlannerInfo *root;
340 : RelOptInfo *final_rel;
341 : Path *best_path;
342 : Plan *top_plan;
343 : ListCell *lp,
344 : *lr;
345 :
346 : /*
347 : * Set up global state for this planner invocation. This data is needed
348 : * across all levels of sub-Query that might exist in the given command,
349 : * so we keep it in a separate struct that's linked to by each per-Query
350 : * PlannerInfo.
351 : */
352 470866 : glob = makeNode(PlannerGlobal);
353 :
354 470866 : glob->boundParams = boundParams;
355 470866 : glob->subplans = NIL;
356 470866 : glob->subpaths = NIL;
357 470866 : glob->subroots = NIL;
358 470866 : glob->rewindPlanIDs = NULL;
359 470866 : glob->finalrtable = NIL;
360 470866 : glob->allRelids = NULL;
361 470866 : glob->prunableRelids = NULL;
362 470866 : glob->finalrteperminfos = NIL;
363 470866 : glob->finalrowmarks = NIL;
364 470866 : glob->resultRelations = NIL;
365 470866 : glob->appendRelations = NIL;
366 470866 : glob->partPruneInfos = NIL;
367 470866 : glob->relationOids = NIL;
368 470866 : glob->invalItems = NIL;
369 470866 : glob->paramExecTypes = NIL;
370 470866 : glob->lastPHId = 0;
371 470866 : glob->lastRowMarkId = 0;
372 470866 : glob->lastPlanNodeId = 0;
373 470866 : glob->transientPlan = false;
374 470866 : glob->dependsOnRole = false;
375 470866 : glob->partition_directory = NULL;
376 470866 : glob->rel_notnullatts_hash = NULL;
377 :
378 : /*
379 : * Assess whether it's feasible to use parallel mode for this query. We
380 : * can't do this in a standalone backend, or if the command will try to
381 : * modify any data, or if this is a cursor operation, or if GUCs are set
382 : * to values that don't permit parallelism, or if parallel-unsafe
383 : * functions are present in the query tree.
384 : *
385 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
386 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
387 : * the command is writing into a completely new table which workers won't
388 : * be able to see. If the workers could see the table, the fact that
389 : * group locking would cause them to ignore the leader's heavyweight GIN
390 : * page locks would make this unsafe. We'll have to fix that somehow if
391 : * we want to allow parallel inserts in general; updates and deletes have
392 : * additional problems especially around combo CIDs.)
393 : *
394 : * For now, we don't try to use parallel mode if we're running inside a
395 : * parallel worker. We might eventually be able to relax this
396 : * restriction, but for now it seems best not to have parallel workers
397 : * trying to create their own parallel workers.
398 : */
399 470866 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
400 442274 : IsUnderPostmaster &&
401 442274 : parse->commandType == CMD_SELECT &&
402 353320 : !parse->hasModifyingCTE &&
403 353176 : max_parallel_workers_per_gather > 0 &&
404 352542 : !IsParallelWorker())
405 : {
406 : /* all the cheap tests pass, so scan the query tree */
407 352494 : glob->maxParallelHazard = max_parallel_hazard(parse);
408 352494 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
409 : }
410 : else
411 : {
412 : /* skip the query tree scan, just assume it's unsafe */
413 118372 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
414 118372 : glob->parallelModeOK = false;
415 : }
416 :
417 : /*
418 : * glob->parallelModeNeeded is normally set to false here and changed to
419 : * true during plan creation if a Gather or Gather Merge plan is actually
420 : * created (cf. create_gather_plan, create_gather_merge_plan).
421 : *
422 : * However, if debug_parallel_query = on or debug_parallel_query =
423 : * regress, then we impose parallel mode whenever it's safe to do so, even
424 : * if the final plan doesn't use parallelism. It's not safe to do so if
425 : * the query contains anything parallel-unsafe; parallelModeOK will be
426 : * false in that case. Note that parallelModeOK can't change after this
427 : * point. Otherwise, everything in the query is either parallel-safe or
428 : * parallel-restricted, and in either case it should be OK to impose
429 : * parallel-mode restrictions. If that ends up breaking something, then
430 : * either some function the user included in the query is incorrectly
431 : * labeled as parallel-safe or parallel-restricted when in reality it's
432 : * parallel-unsafe, or else the query planner itself has a bug.
433 : */
434 767112 : glob->parallelModeNeeded = glob->parallelModeOK &&
435 296246 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
436 :
437 : /* Determine what fraction of the plan is likely to be scanned */
438 470866 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
439 : {
440 : /*
441 : * We have no real idea how many tuples the user will ultimately FETCH
442 : * from a cursor, but it is often the case that he doesn't want 'em
443 : * all, or would prefer a fast-start plan anyway so that he can
444 : * process some of the tuples sooner. Use a GUC parameter to decide
445 : * what fraction to optimize for.
446 : */
447 4712 : tuple_fraction = cursor_tuple_fraction;
448 :
449 : /*
450 : * We document cursor_tuple_fraction as simply being a fraction, which
451 : * means the edge cases 0 and 1 have to be treated specially here. We
452 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
453 : */
454 4712 : if (tuple_fraction >= 1.0)
455 0 : tuple_fraction = 0.0;
456 4712 : else if (tuple_fraction <= 0.0)
457 0 : tuple_fraction = 1e-10;
458 : }
459 : else
460 : {
461 : /* Default assumption is we need all the tuples */
462 466154 : tuple_fraction = 0.0;
463 : }
464 :
465 : /*
466 : * Compute the initial path generation strategy mask.
467 : *
468 : * Some strategies, such as PGS_FOREIGNJOIN, have no corresponding enable_*
469 : * GUC, and so the corresponding bits are always set in the default
470 : * strategy mask.
471 : *
472 : * It may seem surprising that enable_indexscan sets both PGS_INDEXSCAN
473 : * and PGS_INDEXONLYSCAN. However, the historical behavior of this GUC
474 : * corresponds to this exactly: enable_indexscan=off disables both
475 : * index-scan and index-only scan paths, whereas enable_indexonlyscan=off
476 : * converts the index-only scan paths that we would have considered into
477 : * index scan paths.
478 : */
479 470866 : glob->default_pgs_mask = PGS_APPEND | PGS_MERGE_APPEND | PGS_FOREIGNJOIN |
480 : PGS_GATHER | PGS_CONSIDER_NONPARTIAL;
481 470866 : if (enable_tidscan)
482 470866 : glob->default_pgs_mask |= PGS_TIDSCAN;
483 470866 : if (enable_seqscan)
484 445688 : glob->default_pgs_mask |= PGS_SEQSCAN;
485 470866 : if (enable_indexscan)
486 468188 : glob->default_pgs_mask |= PGS_INDEXSCAN | PGS_INDEXONLYSCAN;
487 470866 : if (enable_indexonlyscan)
488 469208 : glob->default_pgs_mask |= PGS_CONSIDER_INDEXONLY;
489 470866 : if (enable_bitmapscan)
490 459376 : glob->default_pgs_mask |= PGS_BITMAPSCAN;
491 470866 : if (enable_mergejoin)
492 : {
493 468648 : glob->default_pgs_mask |= PGS_MERGEJOIN_PLAIN;
494 468648 : if (enable_material)
495 468566 : glob->default_pgs_mask |= PGS_MERGEJOIN_MATERIALIZE;
496 : }
497 470866 : if (enable_nestloop)
498 : {
499 470470 : glob->default_pgs_mask |= PGS_NESTLOOP_PLAIN;
500 470470 : if (enable_material)
501 470242 : glob->default_pgs_mask |= PGS_NESTLOOP_MATERIALIZE;
502 470470 : if (enable_memoize)
503 470338 : glob->default_pgs_mask |= PGS_NESTLOOP_MEMOIZE;
504 : }
505 470866 : if (enable_hashjoin)
506 468256 : glob->default_pgs_mask |= PGS_HASHJOIN;
507 470866 : if (enable_gathermerge)
508 470866 : glob->default_pgs_mask |= PGS_GATHER_MERGE;
509 470866 : if (enable_partitionwise_join)
510 2484 : glob->default_pgs_mask |= PGS_CONSIDER_PARTITIONWISE;
511 :
512 : /* Allow plugins to take control after we've initialized "glob" */
513 470866 : if (planner_setup_hook)
514 0 : (*planner_setup_hook) (glob, parse, query_string, cursorOptions,
515 : &tuple_fraction, es);
516 :
517 : /* primary planning entry point (may recurse for subqueries) */
518 470866 : root = subquery_planner(glob, parse, NULL, NULL, false, tuple_fraction,
519 : NULL);
520 :
521 : /* Select best Path and turn it into a Plan */
522 466414 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
523 466414 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
524 :
525 466414 : top_plan = create_plan(root, best_path);
526 :
527 : /*
528 : * If creating a plan for a scrollable cursor, make sure it can run
529 : * backwards on demand. Add a Material node at the top at need.
530 : */
531 466018 : if (cursorOptions & CURSOR_OPT_SCROLL)
532 : {
533 266 : if (!ExecSupportsBackwardScan(top_plan))
534 32 : top_plan = materialize_finished_plan(top_plan);
535 : }
536 :
537 : /*
538 : * Optionally add a Gather node for testing purposes, provided this is
539 : * actually a safe thing to do.
540 : *
541 : * We can add Gather even when top_plan has parallel-safe initPlans, but
542 : * then we have to move the initPlans to the Gather node because of
543 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
544 : * regression tests when debug_parallel_query = regress, because initPlans
545 : * that would normally appear on the top_plan move to the Gather, causing
546 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
547 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
548 : */
549 466018 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
550 194 : top_plan->parallel_safe &&
551 128 : (top_plan->initPlan == NIL ||
552 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
553 : {
554 128 : Gather *gather = makeNode(Gather);
555 : Cost initplan_cost;
556 : bool unsafe_initplans;
557 :
558 128 : gather->plan.targetlist = top_plan->targetlist;
559 128 : gather->plan.qual = NIL;
560 128 : gather->plan.lefttree = top_plan;
561 128 : gather->plan.righttree = NULL;
562 128 : gather->num_workers = 1;
563 128 : gather->single_copy = true;
564 128 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
565 :
566 : /* Transfer any initPlans to the new top node */
567 128 : gather->plan.initPlan = top_plan->initPlan;
568 128 : top_plan->initPlan = NIL;
569 :
570 : /*
571 : * Since this Gather has no parallel-aware descendants to signal to,
572 : * we don't need a rescan Param.
573 : */
574 128 : gather->rescan_param = -1;
575 :
576 : /*
577 : * Ideally we'd use cost_gather here, but setting up dummy path data
578 : * to satisfy it doesn't seem much cleaner than knowing what it does.
579 : */
580 128 : gather->plan.startup_cost = top_plan->startup_cost +
581 : parallel_setup_cost;
582 128 : gather->plan.total_cost = top_plan->total_cost +
583 128 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
584 128 : gather->plan.plan_rows = top_plan->plan_rows;
585 128 : gather->plan.plan_width = top_plan->plan_width;
586 128 : gather->plan.parallel_aware = false;
587 128 : gather->plan.parallel_safe = false;
588 :
589 : /*
590 : * Delete the initplans' cost from top_plan. We needn't add it to the
591 : * Gather node, since the above coding already included it there.
592 : */
593 128 : SS_compute_initplan_cost(gather->plan.initPlan,
594 : &initplan_cost, &unsafe_initplans);
595 128 : top_plan->startup_cost -= initplan_cost;
596 128 : top_plan->total_cost -= initplan_cost;
597 :
598 : /* use parallel mode for parallel plans. */
599 128 : root->glob->parallelModeNeeded = true;
600 :
601 128 : top_plan = &gather->plan;
602 : }
603 :
604 : /*
605 : * If any Params were generated, run through the plan tree and compute
606 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
607 : * set_plan_references' tree traversal, but for now it has to be separate
608 : * because we need to visit subplans before not after main plan.
609 : */
610 466018 : if (glob->paramExecTypes != NIL)
611 : {
612 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
613 211958 : forboth(lp, glob->subplans, lr, glob->subroots)
614 : {
615 44282 : Plan *subplan = (Plan *) lfirst(lp);
616 44282 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
617 :
618 44282 : SS_finalize_plan(subroot, subplan);
619 : }
620 167676 : SS_finalize_plan(root, top_plan);
621 : }
622 :
623 : /* final cleanup of the plan */
624 : Assert(glob->finalrtable == NIL);
625 : Assert(glob->finalrteperminfos == NIL);
626 : Assert(glob->finalrowmarks == NIL);
627 : Assert(glob->resultRelations == NIL);
628 : Assert(glob->appendRelations == NIL);
629 466018 : top_plan = set_plan_references(root, top_plan);
630 : /* ... and the subplans (both regular subplans and initplans) */
631 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
632 510300 : forboth(lp, glob->subplans, lr, glob->subroots)
633 : {
634 44282 : Plan *subplan = (Plan *) lfirst(lp);
635 44282 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
636 :
637 44282 : lfirst(lp) = set_plan_references(subroot, subplan);
638 : }
639 :
640 : /* build the PlannedStmt result */
641 466018 : result = makeNode(PlannedStmt);
642 :
643 466018 : result->commandType = parse->commandType;
644 466018 : result->queryId = parse->queryId;
645 466018 : result->planOrigin = PLAN_STMT_STANDARD;
646 466018 : result->hasReturning = (parse->returningList != NIL);
647 466018 : result->hasModifyingCTE = parse->hasModifyingCTE;
648 466018 : result->canSetTag = parse->canSetTag;
649 466018 : result->transientPlan = glob->transientPlan;
650 466018 : result->dependsOnRole = glob->dependsOnRole;
651 466018 : result->parallelModeNeeded = glob->parallelModeNeeded;
652 466018 : result->planTree = top_plan;
653 466018 : result->partPruneInfos = glob->partPruneInfos;
654 466018 : result->rtable = glob->finalrtable;
655 932036 : result->unprunableRelids = bms_difference(glob->allRelids,
656 466018 : glob->prunableRelids);
657 466018 : result->permInfos = glob->finalrteperminfos;
658 466018 : result->resultRelations = glob->resultRelations;
659 466018 : result->appendRelations = glob->appendRelations;
660 466018 : result->subplans = glob->subplans;
661 466018 : result->rewindPlanIDs = glob->rewindPlanIDs;
662 466018 : result->rowMarks = glob->finalrowmarks;
663 466018 : result->relationOids = glob->relationOids;
664 466018 : result->invalItems = glob->invalItems;
665 466018 : result->paramExecTypes = glob->paramExecTypes;
666 : /* utilityStmt should be null, but we might as well copy it */
667 466018 : result->utilityStmt = parse->utilityStmt;
668 466018 : result->stmt_location = parse->stmt_location;
669 466018 : result->stmt_len = parse->stmt_len;
670 :
671 466018 : result->jitFlags = PGJIT_NONE;
672 466018 : if (jit_enabled && jit_above_cost >= 0 &&
673 465236 : top_plan->total_cost > jit_above_cost)
674 : {
675 928 : result->jitFlags |= PGJIT_PERFORM;
676 :
677 : /*
678 : * Decide how much effort should be put into generating better code.
679 : */
680 928 : if (jit_optimize_above_cost >= 0 &&
681 928 : top_plan->total_cost > jit_optimize_above_cost)
682 368 : result->jitFlags |= PGJIT_OPT3;
683 928 : if (jit_inline_above_cost >= 0 &&
684 928 : top_plan->total_cost > jit_inline_above_cost)
685 368 : result->jitFlags |= PGJIT_INLINE;
686 :
687 : /*
688 : * Decide which operations should be JITed.
689 : */
690 928 : if (jit_expressions)
691 928 : result->jitFlags |= PGJIT_EXPR;
692 928 : if (jit_tuple_deforming)
693 928 : result->jitFlags |= PGJIT_DEFORM;
694 : }
695 :
696 : /* Allow plugins to take control before we discard "glob" */
697 466018 : if (planner_shutdown_hook)
698 0 : (*planner_shutdown_hook) (glob, parse, query_string, result);
699 :
700 466018 : if (glob->partition_directory != NULL)
701 12182 : DestroyPartitionDirectory(glob->partition_directory);
702 :
703 466018 : return result;
704 : }
705 :
706 :
707 : /*--------------------
708 : * subquery_planner
709 : * Invokes the planner on a subquery. We recurse to here for each
710 : * sub-SELECT found in the query tree.
711 : *
712 : * glob is the global state for the current planner run.
713 : * parse is the querytree produced by the parser & rewriter.
714 : * plan_name is the name to assign to this subplan (NULL at the top level).
715 : * parent_root is the immediate parent Query's info (NULL at the top level).
716 : * hasRecursion is true if this is a recursive WITH query.
717 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
718 : * tuple_fraction is interpreted as explained for grouping_planner, below.
719 : * setops is used for set operation subqueries to provide the subquery with
720 : * the context in which it's being used so that Paths correctly sorted for the
721 : * set operation can be generated. NULL when not planning a set operation
722 : * child, or when a child of a set op that isn't interested in sorted input.
723 : *
724 : * Basically, this routine does the stuff that should only be done once
725 : * per Query object. It then calls grouping_planner. At one time,
726 : * grouping_planner could be invoked recursively on the same Query object;
727 : * that's not currently true, but we keep the separation between the two
728 : * routines anyway, in case we need it again someday.
729 : *
730 : * subquery_planner will be called recursively to handle sub-Query nodes
731 : * found within the query's expressions and rangetable.
732 : *
733 : * Returns the PlannerInfo struct ("root") that contains all data generated
734 : * while planning the subquery. In particular, the Path(s) attached to
735 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
736 : * cheapest way(s) to implement the query. The top level will select the
737 : * best Path and pass it through createplan.c to produce a finished Plan.
738 : *--------------------
739 : */
740 : PlannerInfo *
741 555154 : subquery_planner(PlannerGlobal *glob, Query *parse, char *plan_name,
742 : PlannerInfo *parent_root, bool hasRecursion,
743 : double tuple_fraction, SetOperationStmt *setops)
744 : {
745 : PlannerInfo *root;
746 : List *newWithCheckOptions;
747 : List *newHaving;
748 : bool hasOuterJoins;
749 : bool hasResultRTEs;
750 : RelOptInfo *final_rel;
751 : ListCell *l;
752 :
753 : /* Create a PlannerInfo data structure for this subquery */
754 555154 : root = makeNode(PlannerInfo);
755 555154 : root->parse = parse;
756 555154 : root->glob = glob;
757 555154 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
758 555154 : root->plan_name = plan_name;
759 555154 : root->parent_root = parent_root;
760 555154 : root->plan_params = NIL;
761 555154 : root->outer_params = NULL;
762 555154 : root->planner_cxt = CurrentMemoryContext;
763 555154 : root->init_plans = NIL;
764 555154 : root->cte_plan_ids = NIL;
765 555154 : root->multiexpr_params = NIL;
766 555154 : root->join_domains = NIL;
767 555154 : root->eq_classes = NIL;
768 555154 : root->ec_merging_done = false;
769 555154 : root->last_rinfo_serial = 0;
770 555154 : root->all_result_relids =
771 555154 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
772 555154 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
773 555154 : root->append_rel_list = NIL;
774 555154 : root->row_identity_vars = NIL;
775 555154 : root->rowMarks = NIL;
776 555154 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
777 555154 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
778 555154 : root->processed_groupClause = NIL;
779 555154 : root->processed_distinctClause = NIL;
780 555154 : root->processed_tlist = NIL;
781 555154 : root->update_colnos = NIL;
782 555154 : root->grouping_map = NULL;
783 555154 : root->minmax_aggs = NIL;
784 555154 : root->qual_security_level = 0;
785 555154 : root->hasPseudoConstantQuals = false;
786 555154 : root->hasAlternativeSubPlans = false;
787 555154 : root->placeholdersFrozen = false;
788 555154 : root->hasRecursion = hasRecursion;
789 555154 : root->assumeReplanning = false;
790 555154 : if (hasRecursion)
791 938 : root->wt_param_id = assign_special_exec_param(root);
792 : else
793 554216 : root->wt_param_id = -1;
794 555154 : root->non_recursive_path = NULL;
795 :
796 : /*
797 : * Create the top-level join domain. This won't have valid contents until
798 : * deconstruct_jointree fills it in, but the node needs to exist before
799 : * that so we can build EquivalenceClasses referencing it.
800 : */
801 555154 : root->join_domains = list_make1(makeNode(JoinDomain));
802 :
803 : /*
804 : * If there is a WITH list, process each WITH query and either convert it
805 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
806 : */
807 555154 : if (parse->cteList)
808 3038 : SS_process_ctes(root);
809 :
810 : /*
811 : * If it's a MERGE command, transform the joinlist as appropriate.
812 : */
813 555148 : transform_MERGE_to_join(parse);
814 :
815 : /*
816 : * Scan the rangetable for relation RTEs and retrieve the necessary
817 : * catalog information for each relation. Using this information, clear
818 : * the inh flag for any relation that has no children, collect not-null
819 : * attribute numbers for any relation that has column not-null
820 : * constraints, and expand virtual generated columns for any relation that
821 : * contains them. Note that this step does not descend into sublinks and
822 : * subqueries; if we pull up any sublinks or subqueries below, their
823 : * relation RTEs are processed just before pulling them up.
824 : */
825 555148 : parse = root->parse = preprocess_relation_rtes(root);
826 :
827 : /*
828 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
829 : * that we don't need so many special cases to deal with that situation.
830 : */
831 555148 : replace_empty_jointree(parse);
832 :
833 : /*
834 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
835 : * to transform them into joins. Note that this step does not descend
836 : * into subqueries; if we pull up any subqueries below, their SubLinks are
837 : * processed just before pulling them up.
838 : */
839 555148 : if (parse->hasSubLinks)
840 37228 : pull_up_sublinks(root);
841 :
842 : /*
843 : * Scan the rangetable for function RTEs, do const-simplification on them,
844 : * and then inline them if possible (producing subqueries that might get
845 : * pulled up next). Recursion issues here are handled in the same way as
846 : * for SubLinks.
847 : */
848 555148 : preprocess_function_rtes(root);
849 :
850 : /*
851 : * Check to see if any subqueries in the jointree can be merged into this
852 : * query.
853 : */
854 555142 : pull_up_subqueries(root);
855 :
856 : /*
857 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
858 : * do this now because it requires applying pull_up_subqueries to the leaf
859 : * queries of the UNION ALL, which weren't touched above because they
860 : * weren't referenced by the jointree (they will be after we do this).
861 : */
862 555136 : if (parse->setOperations)
863 7010 : flatten_simple_union_all(root);
864 :
865 : /*
866 : * Survey the rangetable to see what kinds of entries are present. We can
867 : * skip some later processing if relevant SQL features are not used; for
868 : * example if there are no JOIN RTEs we can avoid the expense of doing
869 : * flatten_join_alias_vars(). This must be done after we have finished
870 : * adding rangetable entries, of course. (Note: actually, processing of
871 : * inherited or partitioned rels can cause RTEs for their child tables to
872 : * get added later; but those must all be RTE_RELATION entries, so they
873 : * don't invalidate the conclusions drawn here.)
874 : */
875 555136 : root->hasJoinRTEs = false;
876 555136 : root->hasLateralRTEs = false;
877 555136 : root->group_rtindex = 0;
878 555136 : hasOuterJoins = false;
879 555136 : hasResultRTEs = false;
880 1516392 : foreach(l, parse->rtable)
881 : {
882 961256 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
883 :
884 961256 : switch (rte->rtekind)
885 : {
886 97312 : case RTE_JOIN:
887 97312 : root->hasJoinRTEs = true;
888 97312 : if (IS_OUTER_JOIN(rte->jointype))
889 50564 : hasOuterJoins = true;
890 97312 : break;
891 211166 : case RTE_RESULT:
892 211166 : hasResultRTEs = true;
893 211166 : break;
894 5074 : case RTE_GROUP:
895 : Assert(parse->hasGroupRTE);
896 5074 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
897 5074 : break;
898 647704 : default:
899 : /* No work here for other RTE types */
900 647704 : break;
901 : }
902 :
903 961256 : if (rte->lateral)
904 11448 : root->hasLateralRTEs = true;
905 :
906 : /*
907 : * We can also determine the maximum security level required for any
908 : * securityQuals now. Addition of inheritance-child RTEs won't affect
909 : * this, because child tables don't have their own securityQuals; see
910 : * expand_single_inheritance_child().
911 : */
912 961256 : if (rte->securityQuals)
913 2784 : root->qual_security_level = Max(root->qual_security_level,
914 : list_length(rte->securityQuals));
915 : }
916 :
917 : /*
918 : * If we have now verified that the query target relation is
919 : * non-inheriting, mark it as a leaf target.
920 : */
921 555136 : if (parse->resultRelation)
922 : {
923 95996 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
924 :
925 95996 : if (!rte->inh)
926 93044 : root->leaf_result_relids =
927 93044 : bms_make_singleton(parse->resultRelation);
928 : }
929 :
930 : /*
931 : * This would be a convenient time to check access permissions for all
932 : * relations mentioned in the query, since it would be better to fail now,
933 : * before doing any detailed planning. However, for historical reasons,
934 : * we leave this to be done at executor startup.
935 : *
936 : * Note, however, that we do need to check access permissions for any view
937 : * relations mentioned in the query, in order to prevent information being
938 : * leaked by selectivity estimation functions, which only check view owner
939 : * permissions on underlying tables (see all_rows_selectable() and its
940 : * callers). This is a little ugly, because it means that access
941 : * permissions for views will be checked twice, which is another reason
942 : * why it would be better to do all the ACL checks here.
943 : */
944 1515268 : foreach(l, parse->rtable)
945 : {
946 960520 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
947 :
948 960520 : if (rte->perminfoindex != 0 &&
949 512270 : rte->relkind == RELKIND_VIEW)
950 : {
951 : RTEPermissionInfo *perminfo;
952 : bool result;
953 :
954 21816 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
955 21816 : result = ExecCheckOneRelPerms(perminfo);
956 21816 : if (!result)
957 388 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
958 388 : get_rel_name(perminfo->relid));
959 : }
960 : }
961 :
962 : /*
963 : * Preprocess RowMark information. We need to do this after subquery
964 : * pullup, so that all base relations are present.
965 : */
966 554748 : preprocess_rowmarks(root);
967 :
968 : /*
969 : * Set hasHavingQual to remember if HAVING clause is present. Needed
970 : * because preprocess_expression will reduce a constant-true condition to
971 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
972 : */
973 554748 : root->hasHavingQual = (parse->havingQual != NULL);
974 :
975 : /*
976 : * Do expression preprocessing on targetlist and quals, as well as other
977 : * random expressions in the querytree. Note that we do not need to
978 : * handle sort/group expressions explicitly, because they are actually
979 : * part of the targetlist.
980 : */
981 550768 : parse->targetList = (List *)
982 554748 : preprocess_expression(root, (Node *) parse->targetList,
983 : EXPRKIND_TARGET);
984 :
985 550768 : newWithCheckOptions = NIL;
986 553706 : foreach(l, parse->withCheckOptions)
987 : {
988 2938 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
989 :
990 2938 : wco->qual = preprocess_expression(root, wco->qual,
991 : EXPRKIND_QUAL);
992 2938 : if (wco->qual != NULL)
993 2538 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
994 : }
995 550768 : parse->withCheckOptions = newWithCheckOptions;
996 :
997 550768 : parse->returningList = (List *)
998 550768 : preprocess_expression(root, (Node *) parse->returningList,
999 : EXPRKIND_TARGET);
1000 :
1001 550768 : preprocess_qual_conditions(root, (Node *) parse->jointree);
1002 :
1003 550768 : parse->havingQual = preprocess_expression(root, parse->havingQual,
1004 : EXPRKIND_QUAL);
1005 :
1006 553702 : foreach(l, parse->windowClause)
1007 : {
1008 2934 : WindowClause *wc = lfirst_node(WindowClause, l);
1009 :
1010 : /* partitionClause/orderClause are sort/group expressions */
1011 2934 : wc->startOffset = preprocess_expression(root, wc->startOffset,
1012 : EXPRKIND_LIMIT);
1013 2934 : wc->endOffset = preprocess_expression(root, wc->endOffset,
1014 : EXPRKIND_LIMIT);
1015 : }
1016 :
1017 550768 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
1018 : EXPRKIND_LIMIT);
1019 550768 : parse->limitCount = preprocess_expression(root, parse->limitCount,
1020 : EXPRKIND_LIMIT);
1021 :
1022 550768 : if (parse->onConflict)
1023 : {
1024 3836 : parse->onConflict->arbiterElems = (List *)
1025 1918 : preprocess_expression(root,
1026 1918 : (Node *) parse->onConflict->arbiterElems,
1027 : EXPRKIND_ARBITER_ELEM);
1028 3836 : parse->onConflict->arbiterWhere =
1029 1918 : preprocess_expression(root,
1030 1918 : parse->onConflict->arbiterWhere,
1031 : EXPRKIND_QUAL);
1032 3836 : parse->onConflict->onConflictSet = (List *)
1033 1918 : preprocess_expression(root,
1034 1918 : (Node *) parse->onConflict->onConflictSet,
1035 : EXPRKIND_TARGET);
1036 1918 : parse->onConflict->onConflictWhere =
1037 1918 : preprocess_expression(root,
1038 1918 : parse->onConflict->onConflictWhere,
1039 : EXPRKIND_QUAL);
1040 : /* exclRelTlist contains only Vars, so no preprocessing needed */
1041 : }
1042 :
1043 553650 : foreach(l, parse->mergeActionList)
1044 : {
1045 2882 : MergeAction *action = (MergeAction *) lfirst(l);
1046 :
1047 2882 : action->targetList = (List *)
1048 2882 : preprocess_expression(root,
1049 2882 : (Node *) action->targetList,
1050 : EXPRKIND_TARGET);
1051 2882 : action->qual =
1052 2882 : preprocess_expression(root,
1053 : (Node *) action->qual,
1054 : EXPRKIND_QUAL);
1055 : }
1056 :
1057 550768 : parse->mergeJoinCondition =
1058 550768 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1059 :
1060 550768 : root->append_rel_list = (List *)
1061 550768 : preprocess_expression(root, (Node *) root->append_rel_list,
1062 : EXPRKIND_APPINFO);
1063 :
1064 : /* Also need to preprocess expressions within RTEs */
1065 1506324 : foreach(l, parse->rtable)
1066 : {
1067 955556 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1068 : int kind;
1069 : ListCell *lcsq;
1070 :
1071 955556 : if (rte->rtekind == RTE_RELATION)
1072 : {
1073 494648 : if (rte->tablesample)
1074 228 : rte->tablesample = (TableSampleClause *)
1075 228 : preprocess_expression(root,
1076 228 : (Node *) rte->tablesample,
1077 : EXPRKIND_TABLESAMPLE);
1078 : }
1079 460908 : else if (rte->rtekind == RTE_SUBQUERY)
1080 : {
1081 : /*
1082 : * We don't want to do all preprocessing yet on the subquery's
1083 : * expressions, since that will happen when we plan it. But if it
1084 : * contains any join aliases of our level, those have to get
1085 : * expanded now, because planning of the subquery won't do it.
1086 : * That's only possible if the subquery is LATERAL.
1087 : */
1088 83422 : if (rte->lateral && root->hasJoinRTEs)
1089 1842 : rte->subquery = (Query *)
1090 1842 : flatten_join_alias_vars(root, root->parse,
1091 1842 : (Node *) rte->subquery);
1092 : }
1093 377486 : else if (rte->rtekind == RTE_FUNCTION)
1094 : {
1095 : /* Preprocess the function expression(s) fully */
1096 53124 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1097 53124 : rte->functions = (List *)
1098 53124 : preprocess_expression(root, (Node *) rte->functions, kind);
1099 : }
1100 324362 : else if (rte->rtekind == RTE_TABLEFUNC)
1101 : {
1102 : /* Preprocess the function expression(s) fully */
1103 626 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1104 626 : rte->tablefunc = (TableFunc *)
1105 626 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1106 : }
1107 323736 : else if (rte->rtekind == RTE_VALUES)
1108 : {
1109 : /* Preprocess the values lists fully */
1110 8556 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1111 8556 : rte->values_lists = (List *)
1112 8556 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1113 : }
1114 315180 : else if (rte->rtekind == RTE_GROUP)
1115 : {
1116 : /* Preprocess the groupexprs list fully */
1117 5074 : rte->groupexprs = (List *)
1118 5074 : preprocess_expression(root, (Node *) rte->groupexprs,
1119 : EXPRKIND_GROUPEXPR);
1120 : }
1121 :
1122 : /*
1123 : * Process each element of the securityQuals list as if it were a
1124 : * separate qual expression (as indeed it is). We need to do it this
1125 : * way to get proper canonicalization of AND/OR structure. Note that
1126 : * this converts each element into an implicit-AND sublist.
1127 : */
1128 958734 : foreach(lcsq, rte->securityQuals)
1129 : {
1130 3178 : lfirst(lcsq) = preprocess_expression(root,
1131 3178 : (Node *) lfirst(lcsq),
1132 : EXPRKIND_QUAL);
1133 : }
1134 : }
1135 :
1136 : /*
1137 : * Now that we are done preprocessing expressions, and in particular done
1138 : * flattening join alias variables, get rid of the joinaliasvars lists.
1139 : * They no longer match what expressions in the rest of the tree look
1140 : * like, because we have not preprocessed expressions in those lists (and
1141 : * do not want to; for example, expanding a SubLink there would result in
1142 : * a useless unreferenced subplan). Leaving them in place simply creates
1143 : * a hazard for later scans of the tree. We could try to prevent that by
1144 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1145 : * but that doesn't sound very reliable.
1146 : */
1147 550768 : if (root->hasJoinRTEs)
1148 : {
1149 336124 : foreach(l, parse->rtable)
1150 : {
1151 276974 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1152 :
1153 276974 : rte->joinaliasvars = NIL;
1154 : }
1155 : }
1156 :
1157 : /*
1158 : * Replace any Vars in the subquery's targetlist and havingQual that
1159 : * reference GROUP outputs with the underlying grouping expressions.
1160 : *
1161 : * Note that we need to perform this replacement after we've preprocessed
1162 : * the grouping expressions. This is to ensure that there is only one
1163 : * instance of SubPlan for each SubLink contained within the grouping
1164 : * expressions.
1165 : */
1166 550768 : if (parse->hasGroupRTE)
1167 : {
1168 5074 : parse->targetList = (List *)
1169 5074 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1170 5074 : parse->havingQual =
1171 5074 : flatten_group_exprs(root, root->parse, parse->havingQual);
1172 : }
1173 :
1174 : /* Constant-folding might have removed all set-returning functions */
1175 550768 : if (parse->hasTargetSRFs)
1176 12084 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1177 :
1178 : /*
1179 : * If we have grouping sets, expand the groupingSets tree of this query to
1180 : * a flat list of grouping sets. We need to do this before optimizing
1181 : * HAVING, since we can't easily tell if there's an empty grouping set
1182 : * until we have this representation.
1183 : */
1184 550768 : if (parse->groupingSets)
1185 : {
1186 1020 : parse->groupingSets =
1187 1020 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1188 : }
1189 :
1190 : /*
1191 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1192 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1193 : * volatile functions (since a HAVING clause is supposed to be executed
1194 : * only once per group). We also can't do this if there are any grouping
1195 : * sets and the clause references any columns that are nullable by the
1196 : * grouping sets; the nulled values of those columns are not available
1197 : * before the grouping step. (The test on groupClause might seem wrong,
1198 : * but it's okay: it's just an optimization to avoid running pull_varnos
1199 : * when there cannot be any Vars in the HAVING clause.)
1200 : *
1201 : * Also, it may be that the clause is so expensive to execute that we're
1202 : * better off doing it only once per group, despite the loss of
1203 : * selectivity. This is hard to estimate short of doing the entire
1204 : * planning process twice, so we use a heuristic: clauses containing
1205 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1206 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1207 : * instead of after.
1208 : *
1209 : * If the query has no empty grouping set then we can simply move such a
1210 : * clause into WHERE; any group that fails the clause will not be in the
1211 : * output because none of its tuples will reach the grouping or
1212 : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1213 : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1214 : * clause must be degenerate (variable-free), so we can copy it into WHERE
1215 : * so that query_planner() can use it in a gating Result node. (This could
1216 : * be done better, but it seems not worth optimizing.)
1217 : *
1218 : * Note that a HAVING clause may contain expressions that are not fully
1219 : * preprocessed. This can happen if these expressions are part of
1220 : * grouping items. In such cases, they are replaced with GROUP Vars in
1221 : * the parser and then replaced back after we're done with expression
1222 : * preprocessing on havingQual. This is not an issue if the clause
1223 : * remains in HAVING, because these expressions will be matched to lower
1224 : * target items in setrefs.c. However, if the clause is moved or copied
1225 : * into WHERE, we need to ensure that these expressions are fully
1226 : * preprocessed.
1227 : *
1228 : * Note that both havingQual and parse->jointree->quals are in
1229 : * implicitly-ANDed-list form at this point, even though they are declared
1230 : * as Node *.
1231 : */
1232 550768 : newHaving = NIL;
1233 552156 : foreach(l, (List *) parse->havingQual)
1234 : {
1235 1388 : Node *havingclause = (Node *) lfirst(l);
1236 :
1237 1904 : if (contain_agg_clause(havingclause) ||
1238 1032 : contain_volatile_functions(havingclause) ||
1239 516 : contain_subplans(havingclause) ||
1240 636 : (parse->groupClause && parse->groupingSets &&
1241 120 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1242 : {
1243 : /* keep it in HAVING */
1244 944 : newHaving = lappend(newHaving, havingclause);
1245 : }
1246 444 : else if (parse->groupClause &&
1247 408 : (parse->groupingSets == NIL ||
1248 48 : (List *) linitial(parse->groupingSets) != NIL))
1249 396 : {
1250 : /* There is GROUP BY, but no empty grouping set */
1251 : Node *whereclause;
1252 :
1253 : /* Preprocess the HAVING clause fully */
1254 396 : whereclause = preprocess_expression(root, havingclause,
1255 : EXPRKIND_QUAL);
1256 : /* ... and move it to WHERE */
1257 396 : parse->jointree->quals = (Node *)
1258 396 : list_concat((List *) parse->jointree->quals,
1259 : (List *) whereclause);
1260 : }
1261 : else
1262 : {
1263 : /* There is an empty grouping set (perhaps implicitly) */
1264 : Node *whereclause;
1265 :
1266 : /* Preprocess the HAVING clause fully */
1267 48 : whereclause = preprocess_expression(root, copyObject(havingclause),
1268 : EXPRKIND_QUAL);
1269 : /* ... and put a copy in WHERE */
1270 96 : parse->jointree->quals = (Node *)
1271 48 : list_concat((List *) parse->jointree->quals,
1272 : (List *) whereclause);
1273 : /* ... and also keep it in HAVING */
1274 48 : newHaving = lappend(newHaving, havingclause);
1275 : }
1276 : }
1277 550768 : parse->havingQual = (Node *) newHaving;
1278 :
1279 : /*
1280 : * If we have any outer joins, try to reduce them to plain inner joins.
1281 : * This step is most easily done after we've done expression
1282 : * preprocessing.
1283 : */
1284 550768 : if (hasOuterJoins)
1285 34998 : reduce_outer_joins(root);
1286 :
1287 : /*
1288 : * If we have any RTE_RESULT relations, see if they can be deleted from
1289 : * the jointree. We also rely on this processing to flatten single-child
1290 : * FromExprs underneath outer joins. This step is most effectively done
1291 : * after we've done expression preprocessing and outer join reduction.
1292 : */
1293 550768 : if (hasResultRTEs || hasOuterJoins)
1294 240622 : remove_useless_result_rtes(root);
1295 :
1296 : /*
1297 : * Do the main planning.
1298 : */
1299 550768 : grouping_planner(root, tuple_fraction, setops);
1300 :
1301 : /*
1302 : * Capture the set of outer-level param IDs we have access to, for use in
1303 : * extParam/allParam calculations later.
1304 : */
1305 550696 : SS_identify_outer_params(root);
1306 :
1307 : /*
1308 : * If any initPlans were created in this query level, adjust the surviving
1309 : * Paths' costs and parallel-safety flags to account for them. The
1310 : * initPlans won't actually get attached to the plan tree till
1311 : * create_plan() runs, but we must include their effects now.
1312 : */
1313 550696 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1314 550696 : SS_charge_for_initplans(root, final_rel);
1315 :
1316 : /*
1317 : * Make sure we've identified the cheapest Path for the final rel. (By
1318 : * doing this here not in grouping_planner, we include initPlan costs in
1319 : * the decision, though it's unlikely that will change anything.)
1320 : */
1321 550696 : set_cheapest(final_rel);
1322 :
1323 550696 : return root;
1324 : }
1325 :
1326 : /*
1327 : * preprocess_expression
1328 : * Do subquery_planner's preprocessing work for an expression,
1329 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1330 : * conditions), a HAVING clause, or a few other things.
1331 : */
1332 : static Node *
1333 4624306 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1334 : {
1335 : /*
1336 : * Fall out quickly if expression is empty. This occurs often enough to
1337 : * be worth checking. Note that null->null is the correct conversion for
1338 : * implicit-AND result format, too.
1339 : */
1340 4624306 : if (expr == NULL)
1341 3650534 : return NULL;
1342 :
1343 : /*
1344 : * If the query has any join RTEs, replace join alias variables with
1345 : * base-relation variables. We must do this first, since any expressions
1346 : * we may extract from the joinaliasvars lists have not been preprocessed.
1347 : * For example, if we did this after sublink processing, sublinks expanded
1348 : * out from join aliases would not get processed. But we can skip this in
1349 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1350 : * they can't contain any Vars of the current query level.
1351 : */
1352 973772 : if (root->hasJoinRTEs &&
1353 427112 : !(kind == EXPRKIND_RTFUNC ||
1354 213378 : kind == EXPRKIND_VALUES ||
1355 : kind == EXPRKIND_TABLESAMPLE ||
1356 : kind == EXPRKIND_TABLEFUNC))
1357 213360 : expr = flatten_join_alias_vars(root, root->parse, expr);
1358 :
1359 : /*
1360 : * Simplify constant expressions. For function RTEs, this was already
1361 : * done by preprocess_function_rtes. (But note we must do it again for
1362 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1363 : * un-simplified subexpressions inserted by flattening of subqueries or
1364 : * join alias variables.)
1365 : *
1366 : * Note: an essential effect of this is to convert named-argument function
1367 : * calls to positional notation and insert the current actual values of
1368 : * any default arguments for functions. To ensure that happens, we *must*
1369 : * process all expressions here. Previous PG versions sometimes skipped
1370 : * const-simplification if it didn't seem worth the trouble, but we can't
1371 : * do that anymore.
1372 : *
1373 : * Note: this also flattens nested AND and OR expressions into N-argument
1374 : * form. All processing of a qual expression after this point must be
1375 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1376 : * with AND directly under AND, nor OR directly under OR.
1377 : */
1378 973772 : if (kind != EXPRKIND_RTFUNC)
1379 929318 : expr = eval_const_expressions(root, expr);
1380 :
1381 : /*
1382 : * If it's a qual or havingQual, canonicalize it.
1383 : */
1384 969792 : if (kind == EXPRKIND_QUAL)
1385 : {
1386 349694 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1387 :
1388 : #ifdef OPTIMIZER_DEBUG
1389 : printf("After canonicalize_qual()\n");
1390 : pprint(expr);
1391 : #endif
1392 : }
1393 :
1394 : /*
1395 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1396 : * hashfuncid of any that might execute more quickly by using hash lookups
1397 : * instead of a linear search.
1398 : */
1399 969792 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1400 : {
1401 887718 : convert_saop_to_hashed_saop(expr);
1402 : }
1403 :
1404 : /* Expand SubLinks to SubPlans */
1405 969792 : if (root->parse->hasSubLinks)
1406 108792 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1407 :
1408 : /*
1409 : * XXX do not insert anything here unless you have grokked the comments in
1410 : * SS_replace_correlation_vars ...
1411 : */
1412 :
1413 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1414 969792 : if (root->query_level > 1)
1415 191794 : expr = SS_replace_correlation_vars(root, expr);
1416 :
1417 : /*
1418 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1419 : * don't want to do this before eval_const_expressions, since the latter
1420 : * would be unable to simplify a top-level AND correctly. Also,
1421 : * SS_process_sublinks expects explicit-AND format.)
1422 : */
1423 969792 : if (kind == EXPRKIND_QUAL)
1424 349694 : expr = (Node *) make_ands_implicit((Expr *) expr);
1425 :
1426 969792 : return expr;
1427 : }
1428 :
1429 : /*
1430 : * preprocess_qual_conditions
1431 : * Recursively scan the query's jointree and do subquery_planner's
1432 : * preprocessing work on each qual condition found therein.
1433 : */
1434 : static void
1435 1372392 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1436 : {
1437 1372392 : if (jtnode == NULL)
1438 0 : return;
1439 1372392 : if (IsA(jtnode, RangeTblRef))
1440 : {
1441 : /* nothing to do here */
1442 : }
1443 671388 : else if (IsA(jtnode, FromExpr))
1444 : {
1445 566228 : FromExpr *f = (FromExpr *) jtnode;
1446 : ListCell *l;
1447 :
1448 1177532 : foreach(l, f->fromlist)
1449 611304 : preprocess_qual_conditions(root, lfirst(l));
1450 :
1451 566228 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1452 : }
1453 105160 : else if (IsA(jtnode, JoinExpr))
1454 : {
1455 105160 : JoinExpr *j = (JoinExpr *) jtnode;
1456 :
1457 105160 : preprocess_qual_conditions(root, j->larg);
1458 105160 : preprocess_qual_conditions(root, j->rarg);
1459 :
1460 105160 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1461 : }
1462 : else
1463 0 : elog(ERROR, "unrecognized node type: %d",
1464 : (int) nodeTag(jtnode));
1465 : }
1466 :
1467 : /*
1468 : * preprocess_phv_expression
1469 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1470 : *
1471 : * If a LATERAL subquery references an output of another subquery, and that
1472 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1473 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1474 : * and later pull it back up during find_lateral_references, which runs after
1475 : * subquery_planner has preprocessed all the expressions that were in the
1476 : * current query level to start with. So we need to preprocess it then.
1477 : */
1478 : Expr *
1479 90 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1480 : {
1481 90 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1482 : }
1483 :
1484 : /*--------------------
1485 : * grouping_planner
1486 : * Perform planning steps related to grouping, aggregation, etc.
1487 : *
1488 : * This function adds all required top-level processing to the scan/join
1489 : * Path(s) produced by query_planner.
1490 : *
1491 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1492 : * tuple_fraction is interpreted as follows:
1493 : * 0: expect all tuples to be retrieved (normal case)
1494 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1495 : * from the plan to be retrieved
1496 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1497 : * expected to be retrieved (ie, a LIMIT specification).
1498 : * setops is used for set operation subqueries to provide the subquery with
1499 : * the context in which it's being used so that Paths correctly sorted for the
1500 : * set operation can be generated. NULL when not planning a set operation
1501 : * child, or when a child of a set op that isn't interested in sorted input.
1502 : *
1503 : * Returns nothing; the useful output is in the Paths we attach to the
1504 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1505 : * root->processed_tlist contains the final processed targetlist.
1506 : *
1507 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1508 : * to leave this to the caller.
1509 : *--------------------
1510 : */
1511 : static void
1512 550768 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1513 : SetOperationStmt *setops)
1514 : {
1515 550768 : Query *parse = root->parse;
1516 550768 : int64 offset_est = 0;
1517 550768 : int64 count_est = 0;
1518 550768 : double limit_tuples = -1.0;
1519 550768 : bool have_postponed_srfs = false;
1520 : PathTarget *final_target;
1521 : List *final_targets;
1522 : List *final_targets_contain_srfs;
1523 : bool final_target_parallel_safe;
1524 : RelOptInfo *current_rel;
1525 : RelOptInfo *final_rel;
1526 : FinalPathExtraData extra;
1527 : ListCell *lc;
1528 :
1529 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1530 550768 : if (parse->limitCount || parse->limitOffset)
1531 : {
1532 5246 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1533 : &offset_est, &count_est);
1534 :
1535 : /*
1536 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1537 : * estimate the effects of using a bounded sort.
1538 : */
1539 5246 : if (count_est > 0 && offset_est >= 0)
1540 4694 : limit_tuples = (double) count_est + (double) offset_est;
1541 : }
1542 :
1543 : /* Make tuple_fraction accessible to lower-level routines */
1544 550768 : root->tuple_fraction = tuple_fraction;
1545 :
1546 550768 : if (parse->setOperations)
1547 : {
1548 : /*
1549 : * Construct Paths for set operations. The results will not need any
1550 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1551 : * special work for recursive unions is the responsibility of
1552 : * plan_set_operations.
1553 : */
1554 6218 : current_rel = plan_set_operations(root);
1555 :
1556 : /*
1557 : * We should not need to call preprocess_targetlist, since we must be
1558 : * in a SELECT query node. Instead, use the processed_tlist returned
1559 : * by plan_set_operations (since this tells whether it returned any
1560 : * resjunk columns!), and transfer any sort key information from the
1561 : * original tlist.
1562 : */
1563 : Assert(parse->commandType == CMD_SELECT);
1564 :
1565 : /* for safety, copy processed_tlist instead of modifying in-place */
1566 6212 : root->processed_tlist =
1567 6212 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1568 : parse->targetList);
1569 :
1570 : /* Also extract the PathTarget form of the setop result tlist */
1571 6212 : final_target = current_rel->cheapest_total_path->pathtarget;
1572 :
1573 : /* And check whether it's parallel safe */
1574 : final_target_parallel_safe =
1575 6212 : is_parallel_safe(root, (Node *) final_target->exprs);
1576 :
1577 : /* The setop result tlist couldn't contain any SRFs */
1578 : Assert(!parse->hasTargetSRFs);
1579 6212 : final_targets = final_targets_contain_srfs = NIL;
1580 :
1581 : /*
1582 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1583 : * checked already, but let's make sure).
1584 : */
1585 6212 : if (parse->rowMarks)
1586 0 : ereport(ERROR,
1587 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1588 : /*------
1589 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1590 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1591 : LCS_asString(linitial_node(RowMarkClause,
1592 : parse->rowMarks)->strength))));
1593 :
1594 : /*
1595 : * Calculate pathkeys that represent result ordering requirements
1596 : */
1597 : Assert(parse->distinctClause == NIL);
1598 6212 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1599 : parse->sortClause,
1600 : root->processed_tlist);
1601 : }
1602 : else
1603 : {
1604 : /* No set operations, do regular planning */
1605 : PathTarget *sort_input_target;
1606 : List *sort_input_targets;
1607 : List *sort_input_targets_contain_srfs;
1608 : bool sort_input_target_parallel_safe;
1609 : PathTarget *grouping_target;
1610 : List *grouping_targets;
1611 : List *grouping_targets_contain_srfs;
1612 : bool grouping_target_parallel_safe;
1613 : PathTarget *scanjoin_target;
1614 : List *scanjoin_targets;
1615 : List *scanjoin_targets_contain_srfs;
1616 : bool scanjoin_target_parallel_safe;
1617 : bool scanjoin_target_same_exprs;
1618 : bool have_grouping;
1619 544550 : WindowFuncLists *wflists = NULL;
1620 544550 : List *activeWindows = NIL;
1621 544550 : grouping_sets_data *gset_data = NULL;
1622 : standard_qp_extra qp_extra;
1623 :
1624 : /* A recursive query should always have setOperations */
1625 : Assert(!root->hasRecursion);
1626 :
1627 : /* Preprocess grouping sets and GROUP BY clause, if any */
1628 544550 : if (parse->groupingSets)
1629 : {
1630 1020 : gset_data = preprocess_grouping_sets(root);
1631 : }
1632 543530 : else if (parse->groupClause)
1633 : {
1634 : /* Preprocess regular GROUP BY clause, if any */
1635 4132 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1636 : }
1637 :
1638 : /*
1639 : * Preprocess targetlist. Note that much of the remaining planning
1640 : * work will be done with the PathTarget representation of tlists, but
1641 : * we must also maintain the full representation of the final tlist so
1642 : * that we can transfer its decoration (resnames etc) to the topmost
1643 : * tlist of the finished Plan. This is kept in processed_tlist.
1644 : */
1645 544544 : preprocess_targetlist(root);
1646 :
1647 : /*
1648 : * Mark all the aggregates with resolved aggtranstypes, and detect
1649 : * aggregates that are duplicates or can share transition state. We
1650 : * must do this before slicing and dicing the tlist into various
1651 : * pathtargets, else some copies of the Aggref nodes might escape
1652 : * being marked.
1653 : */
1654 544544 : if (parse->hasAggs)
1655 : {
1656 45198 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1657 45198 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1658 : }
1659 :
1660 : /*
1661 : * Locate any window functions in the tlist. (We don't need to look
1662 : * anywhere else, since expressions used in ORDER BY will be in there
1663 : * too.) Note that they could all have been eliminated by constant
1664 : * folding, in which case we don't need to do any more work.
1665 : */
1666 544544 : if (parse->hasWindowFuncs)
1667 : {
1668 2682 : wflists = find_window_functions((Node *) root->processed_tlist,
1669 2682 : list_length(parse->windowClause));
1670 2682 : if (wflists->numWindowFuncs > 0)
1671 : {
1672 : /*
1673 : * See if any modifications can be made to each WindowClause
1674 : * to allow the executor to execute the WindowFuncs more
1675 : * quickly.
1676 : */
1677 2676 : optimize_window_clauses(root, wflists);
1678 :
1679 : /* Extract the list of windows actually in use. */
1680 2676 : activeWindows = select_active_windows(root, wflists);
1681 :
1682 : /* Make sure they all have names, for EXPLAIN's use. */
1683 2676 : name_active_windows(activeWindows);
1684 : }
1685 : else
1686 6 : parse->hasWindowFuncs = false;
1687 : }
1688 :
1689 : /*
1690 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1691 : * adding logic between here and the query_planner() call. Anything
1692 : * that is needed in MIN/MAX-optimizable cases will have to be
1693 : * duplicated in planagg.c.
1694 : */
1695 544544 : if (parse->hasAggs)
1696 45198 : preprocess_minmax_aggregates(root);
1697 :
1698 : /*
1699 : * Figure out whether there's a hard limit on the number of rows that
1700 : * query_planner's result subplan needs to return. Even if we know a
1701 : * hard limit overall, it doesn't apply if the query has any
1702 : * grouping/aggregation operations, or SRFs in the tlist.
1703 : */
1704 544544 : if (parse->groupClause ||
1705 539476 : parse->groupingSets ||
1706 539398 : parse->distinctClause ||
1707 536408 : parse->hasAggs ||
1708 495648 : parse->hasWindowFuncs ||
1709 493116 : parse->hasTargetSRFs ||
1710 481550 : root->hasHavingQual)
1711 63018 : root->limit_tuples = -1.0;
1712 : else
1713 481526 : root->limit_tuples = limit_tuples;
1714 :
1715 : /* Set up data needed by standard_qp_callback */
1716 544544 : qp_extra.activeWindows = activeWindows;
1717 544544 : qp_extra.gset_data = gset_data;
1718 :
1719 : /*
1720 : * If we're a subquery for a set operation, store the SetOperationStmt
1721 : * in qp_extra.
1722 : */
1723 544544 : qp_extra.setop = setops;
1724 :
1725 : /*
1726 : * Generate the best unsorted and presorted paths for the scan/join
1727 : * portion of this Query, ie the processing represented by the
1728 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1729 : * We also generate (in standard_qp_callback) pathkey representations
1730 : * of the query's sort clause, distinct clause, etc.
1731 : */
1732 544544 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1733 :
1734 : /*
1735 : * Convert the query's result tlist into PathTarget format.
1736 : *
1737 : * Note: this cannot be done before query_planner() has performed
1738 : * appendrel expansion, because that might add resjunk entries to
1739 : * root->processed_tlist. Waiting till afterwards is also helpful
1740 : * because the target width estimates can use per-Var width numbers
1741 : * that were obtained within query_planner().
1742 : */
1743 544490 : final_target = create_pathtarget(root, root->processed_tlist);
1744 : final_target_parallel_safe =
1745 544490 : is_parallel_safe(root, (Node *) final_target->exprs);
1746 :
1747 : /*
1748 : * If ORDER BY was given, consider whether we should use a post-sort
1749 : * projection, and compute the adjusted target for preceding steps if
1750 : * so.
1751 : */
1752 544490 : if (parse->sortClause)
1753 : {
1754 72102 : sort_input_target = make_sort_input_target(root,
1755 : final_target,
1756 : &have_postponed_srfs);
1757 : sort_input_target_parallel_safe =
1758 72102 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1759 : }
1760 : else
1761 : {
1762 472388 : sort_input_target = final_target;
1763 472388 : sort_input_target_parallel_safe = final_target_parallel_safe;
1764 : }
1765 :
1766 : /*
1767 : * If we have window functions to deal with, the output from any
1768 : * grouping step needs to be what the window functions want;
1769 : * otherwise, it should be sort_input_target.
1770 : */
1771 544490 : if (activeWindows)
1772 : {
1773 2676 : grouping_target = make_window_input_target(root,
1774 : final_target,
1775 : activeWindows);
1776 : grouping_target_parallel_safe =
1777 2676 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1778 : }
1779 : else
1780 : {
1781 541814 : grouping_target = sort_input_target;
1782 541814 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1783 : }
1784 :
1785 : /*
1786 : * If we have grouping or aggregation to do, the topmost scan/join
1787 : * plan node must emit what the grouping step wants; otherwise, it
1788 : * should emit grouping_target.
1789 : */
1790 539422 : have_grouping = (parse->groupClause || parse->groupingSets ||
1791 1083912 : parse->hasAggs || root->hasHavingQual);
1792 544490 : if (have_grouping)
1793 : {
1794 45968 : scanjoin_target = make_group_input_target(root, final_target);
1795 : scanjoin_target_parallel_safe =
1796 45968 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1797 : }
1798 : else
1799 : {
1800 498522 : scanjoin_target = grouping_target;
1801 498522 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1802 : }
1803 :
1804 : /*
1805 : * If there are any SRFs in the targetlist, we must separate each of
1806 : * these PathTargets into SRF-computing and SRF-free targets. Replace
1807 : * each of the named targets with a SRF-free version, and remember the
1808 : * list of additional projection steps we need to add afterwards.
1809 : */
1810 544490 : if (parse->hasTargetSRFs)
1811 : {
1812 : /* final_target doesn't recompute any SRFs in sort_input_target */
1813 12084 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1814 : &final_targets,
1815 : &final_targets_contain_srfs);
1816 12084 : final_target = linitial_node(PathTarget, final_targets);
1817 : Assert(!linitial_int(final_targets_contain_srfs));
1818 : /* likewise for sort_input_target vs. grouping_target */
1819 12084 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1820 : &sort_input_targets,
1821 : &sort_input_targets_contain_srfs);
1822 12084 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
1823 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1824 : /* likewise for grouping_target vs. scanjoin_target */
1825 12084 : split_pathtarget_at_srfs_grouping(root,
1826 : grouping_target, scanjoin_target,
1827 : &grouping_targets,
1828 : &grouping_targets_contain_srfs);
1829 12084 : grouping_target = linitial_node(PathTarget, grouping_targets);
1830 : Assert(!linitial_int(grouping_targets_contain_srfs));
1831 : /* scanjoin_target will not have any SRFs precomputed for it */
1832 12084 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1833 : &scanjoin_targets,
1834 : &scanjoin_targets_contain_srfs);
1835 12084 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1836 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1837 : }
1838 : else
1839 : {
1840 : /* initialize lists; for most of these, dummy values are OK */
1841 532406 : final_targets = final_targets_contain_srfs = NIL;
1842 532406 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1843 532406 : grouping_targets = grouping_targets_contain_srfs = NIL;
1844 532406 : scanjoin_targets = list_make1(scanjoin_target);
1845 532406 : scanjoin_targets_contain_srfs = NIL;
1846 : }
1847 :
1848 : /* Apply scan/join target. */
1849 544490 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1850 544490 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1851 544490 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1852 : scanjoin_targets_contain_srfs,
1853 : scanjoin_target_parallel_safe,
1854 : scanjoin_target_same_exprs);
1855 :
1856 : /*
1857 : * Save the various upper-rel PathTargets we just computed into
1858 : * root->upper_targets[]. The core code doesn't use this, but it
1859 : * provides a convenient place for extensions to get at the info. For
1860 : * consistency, we save all the intermediate targets, even though some
1861 : * of the corresponding upperrels might not be needed for this query.
1862 : */
1863 544490 : root->upper_targets[UPPERREL_FINAL] = final_target;
1864 544490 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1865 544490 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1866 544490 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1867 544490 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1868 544490 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1869 :
1870 : /*
1871 : * If we have grouping and/or aggregation, consider ways to implement
1872 : * that. We build a new upperrel representing the output of this
1873 : * phase.
1874 : */
1875 544490 : if (have_grouping)
1876 : {
1877 45968 : current_rel = create_grouping_paths(root,
1878 : current_rel,
1879 : grouping_target,
1880 : grouping_target_parallel_safe,
1881 : gset_data);
1882 : /* Fix things up if grouping_target contains SRFs */
1883 45962 : if (parse->hasTargetSRFs)
1884 476 : adjust_paths_for_srfs(root, current_rel,
1885 : grouping_targets,
1886 : grouping_targets_contain_srfs);
1887 : }
1888 :
1889 : /*
1890 : * If we have window functions, consider ways to implement those. We
1891 : * build a new upperrel representing the output of this phase.
1892 : */
1893 544484 : if (activeWindows)
1894 : {
1895 2676 : current_rel = create_window_paths(root,
1896 : current_rel,
1897 : grouping_target,
1898 : sort_input_target,
1899 : sort_input_target_parallel_safe,
1900 : wflists,
1901 : activeWindows);
1902 : /* Fix things up if sort_input_target contains SRFs */
1903 2676 : if (parse->hasTargetSRFs)
1904 12 : adjust_paths_for_srfs(root, current_rel,
1905 : sort_input_targets,
1906 : sort_input_targets_contain_srfs);
1907 : }
1908 :
1909 : /*
1910 : * If there is a DISTINCT clause, consider ways to implement that. We
1911 : * build a new upperrel representing the output of this phase.
1912 : */
1913 544484 : if (parse->distinctClause)
1914 : {
1915 3024 : current_rel = create_distinct_paths(root,
1916 : current_rel,
1917 : sort_input_target);
1918 : }
1919 : } /* end of if (setOperations) */
1920 :
1921 : /*
1922 : * If ORDER BY was given, consider ways to implement that, and generate a
1923 : * new upperrel containing only paths that emit the correct ordering and
1924 : * project the correct final_target. We can apply the original
1925 : * limit_tuples limit in sort costing here, but only if there are no
1926 : * postponed SRFs.
1927 : */
1928 550696 : if (parse->sortClause)
1929 : {
1930 76072 : current_rel = create_ordered_paths(root,
1931 : current_rel,
1932 : final_target,
1933 : final_target_parallel_safe,
1934 : have_postponed_srfs ? -1.0 :
1935 : limit_tuples);
1936 : /* Fix things up if final_target contains SRFs */
1937 76072 : if (parse->hasTargetSRFs)
1938 220 : adjust_paths_for_srfs(root, current_rel,
1939 : final_targets,
1940 : final_targets_contain_srfs);
1941 : }
1942 :
1943 : /*
1944 : * Now we are prepared to build the final-output upperrel.
1945 : */
1946 550696 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1947 :
1948 : /*
1949 : * If the input rel is marked consider_parallel and there's nothing that's
1950 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1951 : * consider_parallel as well. Note that if the query has rowMarks or is
1952 : * not a SELECT, consider_parallel will be false for every relation in the
1953 : * query.
1954 : */
1955 731132 : if (current_rel->consider_parallel &&
1956 360848 : is_parallel_safe(root, parse->limitOffset) &&
1957 180412 : is_parallel_safe(root, parse->limitCount))
1958 180406 : final_rel->consider_parallel = true;
1959 :
1960 : /*
1961 : * If the current_rel belongs to a single FDW, so does the final_rel.
1962 : */
1963 550696 : final_rel->serverid = current_rel->serverid;
1964 550696 : final_rel->userid = current_rel->userid;
1965 550696 : final_rel->useridiscurrent = current_rel->useridiscurrent;
1966 550696 : final_rel->fdwroutine = current_rel->fdwroutine;
1967 :
1968 : /*
1969 : * Generate paths for the final_rel. Insert all surviving paths, with
1970 : * LockRows, Limit, and/or ModifyTable steps added if needed.
1971 : */
1972 1122432 : foreach(lc, current_rel->pathlist)
1973 : {
1974 571736 : Path *path = (Path *) lfirst(lc);
1975 :
1976 : /*
1977 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1978 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
1979 : * here. If there are only non-locking rowmarks, they should be
1980 : * handled by the ModifyTable node instead. However, root->rowMarks
1981 : * is what goes into the LockRows node.)
1982 : */
1983 571736 : if (parse->rowMarks)
1984 : {
1985 14032 : path = (Path *) create_lockrows_path(root, final_rel, path,
1986 : root->rowMarks,
1987 : assign_special_exec_param(root));
1988 : }
1989 :
1990 : /*
1991 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1992 : */
1993 571736 : if (limit_needed(parse))
1994 : {
1995 6270 : path = (Path *) create_limit_path(root, final_rel, path,
1996 : parse->limitOffset,
1997 : parse->limitCount,
1998 : parse->limitOption,
1999 : offset_est, count_est);
2000 : }
2001 :
2002 : /*
2003 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
2004 : */
2005 571736 : if (parse->commandType != CMD_SELECT)
2006 : {
2007 : Index rootRelation;
2008 95474 : List *resultRelations = NIL;
2009 95474 : List *updateColnosLists = NIL;
2010 95474 : List *withCheckOptionLists = NIL;
2011 95474 : List *returningLists = NIL;
2012 95474 : List *mergeActionLists = NIL;
2013 95474 : List *mergeJoinConditions = NIL;
2014 : List *rowMarks;
2015 :
2016 95474 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
2017 : {
2018 : /* Inherited UPDATE/DELETE/MERGE */
2019 2916 : RelOptInfo *top_result_rel = find_base_rel(root,
2020 : parse->resultRelation);
2021 2916 : int resultRelation = -1;
2022 :
2023 : /* Pass the root result rel forward to the executor. */
2024 2916 : rootRelation = parse->resultRelation;
2025 :
2026 : /* Add only leaf children to ModifyTable. */
2027 8474 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
2028 8474 : resultRelation)) >= 0)
2029 : {
2030 5558 : RelOptInfo *this_result_rel = find_base_rel(root,
2031 : resultRelation);
2032 :
2033 : /*
2034 : * Also exclude any leaf rels that have turned dummy since
2035 : * being added to the list, for example, by being excluded
2036 : * by constraint exclusion.
2037 : */
2038 5558 : if (IS_DUMMY_REL(this_result_rel))
2039 180 : continue;
2040 :
2041 : /* Build per-target-rel lists needed by ModifyTable */
2042 5378 : resultRelations = lappend_int(resultRelations,
2043 : resultRelation);
2044 5378 : if (parse->commandType == CMD_UPDATE)
2045 : {
2046 3696 : List *update_colnos = root->update_colnos;
2047 :
2048 3696 : if (this_result_rel != top_result_rel)
2049 : update_colnos =
2050 3696 : adjust_inherited_attnums_multilevel(root,
2051 : update_colnos,
2052 : this_result_rel->relid,
2053 : top_result_rel->relid);
2054 3696 : updateColnosLists = lappend(updateColnosLists,
2055 : update_colnos);
2056 : }
2057 5378 : if (parse->withCheckOptions)
2058 : {
2059 504 : List *withCheckOptions = parse->withCheckOptions;
2060 :
2061 504 : if (this_result_rel != top_result_rel)
2062 : withCheckOptions = (List *)
2063 504 : adjust_appendrel_attrs_multilevel(root,
2064 : (Node *) withCheckOptions,
2065 : this_result_rel,
2066 : top_result_rel);
2067 504 : withCheckOptionLists = lappend(withCheckOptionLists,
2068 : withCheckOptions);
2069 : }
2070 5378 : if (parse->returningList)
2071 : {
2072 846 : List *returningList = parse->returningList;
2073 :
2074 846 : if (this_result_rel != top_result_rel)
2075 : returningList = (List *)
2076 846 : adjust_appendrel_attrs_multilevel(root,
2077 : (Node *) returningList,
2078 : this_result_rel,
2079 : top_result_rel);
2080 846 : returningLists = lappend(returningLists,
2081 : returningList);
2082 : }
2083 5378 : if (parse->mergeActionList)
2084 : {
2085 : ListCell *l;
2086 542 : List *mergeActionList = NIL;
2087 :
2088 : /*
2089 : * Copy MergeActions and translate stuff that
2090 : * references attribute numbers.
2091 : */
2092 1692 : foreach(l, parse->mergeActionList)
2093 : {
2094 1150 : MergeAction *action = lfirst(l),
2095 1150 : *leaf_action = copyObject(action);
2096 :
2097 1150 : leaf_action->qual =
2098 1150 : adjust_appendrel_attrs_multilevel(root,
2099 : (Node *) action->qual,
2100 : this_result_rel,
2101 : top_result_rel);
2102 1150 : leaf_action->targetList = (List *)
2103 1150 : adjust_appendrel_attrs_multilevel(root,
2104 1150 : (Node *) action->targetList,
2105 : this_result_rel,
2106 : top_result_rel);
2107 1150 : if (leaf_action->commandType == CMD_UPDATE)
2108 642 : leaf_action->updateColnos =
2109 642 : adjust_inherited_attnums_multilevel(root,
2110 : action->updateColnos,
2111 : this_result_rel->relid,
2112 : top_result_rel->relid);
2113 1150 : mergeActionList = lappend(mergeActionList,
2114 : leaf_action);
2115 : }
2116 :
2117 542 : mergeActionLists = lappend(mergeActionLists,
2118 : mergeActionList);
2119 : }
2120 5378 : if (parse->commandType == CMD_MERGE)
2121 : {
2122 542 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2123 :
2124 542 : if (this_result_rel != top_result_rel)
2125 : mergeJoinCondition =
2126 542 : adjust_appendrel_attrs_multilevel(root,
2127 : mergeJoinCondition,
2128 : this_result_rel,
2129 : top_result_rel);
2130 542 : mergeJoinConditions = lappend(mergeJoinConditions,
2131 : mergeJoinCondition);
2132 : }
2133 : }
2134 :
2135 2916 : if (resultRelations == NIL)
2136 : {
2137 : /*
2138 : * We managed to exclude every child rel, so generate a
2139 : * dummy one-relation plan using info for the top target
2140 : * rel (even though that may not be a leaf target).
2141 : * Although it's clear that no data will be updated or
2142 : * deleted, we still need to have a ModifyTable node so
2143 : * that any statement triggers will be executed. (This
2144 : * could be cleaner if we fixed nodeModifyTable.c to allow
2145 : * zero target relations, but that probably wouldn't be a
2146 : * net win.)
2147 : */
2148 36 : resultRelations = list_make1_int(parse->resultRelation);
2149 36 : if (parse->commandType == CMD_UPDATE)
2150 32 : updateColnosLists = list_make1(root->update_colnos);
2151 36 : if (parse->withCheckOptions)
2152 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2153 36 : if (parse->returningList)
2154 18 : returningLists = list_make1(parse->returningList);
2155 36 : if (parse->mergeActionList)
2156 2 : mergeActionLists = list_make1(parse->mergeActionList);
2157 36 : if (parse->commandType == CMD_MERGE)
2158 2 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2159 : }
2160 : }
2161 : else
2162 : {
2163 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2164 92558 : rootRelation = 0; /* there's no separate root rel */
2165 92558 : resultRelations = list_make1_int(parse->resultRelation);
2166 92558 : if (parse->commandType == CMD_UPDATE)
2167 12694 : updateColnosLists = list_make1(root->update_colnos);
2168 92558 : if (parse->withCheckOptions)
2169 1040 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2170 92558 : if (parse->returningList)
2171 2570 : returningLists = list_make1(parse->returningList);
2172 92558 : if (parse->mergeActionList)
2173 1602 : mergeActionLists = list_make1(parse->mergeActionList);
2174 92558 : if (parse->commandType == CMD_MERGE)
2175 1602 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2176 : }
2177 :
2178 : /*
2179 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2180 : * will have dealt with fetching non-locked marked rows, else we
2181 : * need to have ModifyTable do that.
2182 : */
2183 95474 : if (parse->rowMarks)
2184 0 : rowMarks = NIL;
2185 : else
2186 95474 : rowMarks = root->rowMarks;
2187 :
2188 : path = (Path *)
2189 95474 : create_modifytable_path(root, final_rel,
2190 : path,
2191 : parse->commandType,
2192 95474 : parse->canSetTag,
2193 95474 : parse->resultRelation,
2194 : rootRelation,
2195 : resultRelations,
2196 : updateColnosLists,
2197 : withCheckOptionLists,
2198 : returningLists,
2199 : rowMarks,
2200 : parse->onConflict,
2201 : mergeActionLists,
2202 : mergeJoinConditions,
2203 : assign_special_exec_param(root));
2204 : }
2205 :
2206 : /* And shove it into final_rel */
2207 571736 : add_path(final_rel, path);
2208 : }
2209 :
2210 : /*
2211 : * Generate partial paths for final_rel, too, if outer query levels might
2212 : * be able to make use of them.
2213 : */
2214 550696 : if (final_rel->consider_parallel && root->query_level > 1 &&
2215 31002 : !limit_needed(parse))
2216 : {
2217 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2218 30936 : foreach(lc, current_rel->partial_pathlist)
2219 : {
2220 120 : Path *partial_path = (Path *) lfirst(lc);
2221 :
2222 120 : add_partial_path(final_rel, partial_path);
2223 : }
2224 : }
2225 :
2226 550696 : extra.limit_needed = limit_needed(parse);
2227 550696 : extra.limit_tuples = limit_tuples;
2228 550696 : extra.count_est = count_est;
2229 550696 : extra.offset_est = offset_est;
2230 :
2231 : /*
2232 : * If there is an FDW that's responsible for all baserels of the query,
2233 : * let it consider adding ForeignPaths.
2234 : */
2235 550696 : if (final_rel->fdwroutine &&
2236 1264 : final_rel->fdwroutine->GetForeignUpperPaths)
2237 1192 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2238 : current_rel, final_rel,
2239 : &extra);
2240 :
2241 : /* Let extensions possibly add some more paths */
2242 550696 : if (create_upper_paths_hook)
2243 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2244 : current_rel, final_rel, &extra);
2245 :
2246 : /* Note: currently, we leave it to callers to do set_cheapest() */
2247 550696 : }
2248 :
2249 : /*
2250 : * Do preprocessing for groupingSets clause and related data.
2251 : *
2252 : * We expect that parse->groupingSets has already been expanded into a flat
2253 : * list of grouping sets (that is, just integer Lists of ressortgroupref
2254 : * numbers) by expand_grouping_sets(). This function handles the preliminary
2255 : * steps of organizing the grouping sets into lists of rollups, and preparing
2256 : * annotations which will later be filled in with size estimates.
2257 : */
2258 : static grouping_sets_data *
2259 1020 : preprocess_grouping_sets(PlannerInfo *root)
2260 : {
2261 1020 : Query *parse = root->parse;
2262 : List *sets;
2263 1020 : int maxref = 0;
2264 : ListCell *lc_set;
2265 1020 : grouping_sets_data *gd = palloc0_object(grouping_sets_data);
2266 :
2267 : /*
2268 : * We don't currently make any attempt to optimize the groupClause when
2269 : * there are grouping sets, so just duplicate it in processed_groupClause.
2270 : */
2271 1020 : root->processed_groupClause = parse->groupClause;
2272 :
2273 : /* Detect unhashable and unsortable grouping expressions */
2274 1020 : gd->any_hashable = false;
2275 1020 : gd->unhashable_refs = NULL;
2276 1020 : gd->unsortable_refs = NULL;
2277 1020 : gd->unsortable_sets = NIL;
2278 :
2279 1020 : if (parse->groupClause)
2280 : {
2281 : ListCell *lc;
2282 :
2283 2980 : foreach(lc, parse->groupClause)
2284 : {
2285 2038 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2286 2038 : Index ref = gc->tleSortGroupRef;
2287 :
2288 2038 : if (ref > maxref)
2289 1990 : maxref = ref;
2290 :
2291 2038 : if (!gc->hashable)
2292 30 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2293 :
2294 2038 : if (!OidIsValid(gc->sortop))
2295 42 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2296 : }
2297 : }
2298 :
2299 : /* Allocate workspace array for remapping */
2300 1020 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2301 :
2302 : /*
2303 : * If we have any unsortable sets, we must extract them before trying to
2304 : * prepare rollups. Unsortable sets don't go through
2305 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2306 : * here.
2307 : */
2308 1020 : if (!bms_is_empty(gd->unsortable_refs))
2309 : {
2310 42 : List *sortable_sets = NIL;
2311 : ListCell *lc;
2312 :
2313 126 : foreach(lc, parse->groupingSets)
2314 : {
2315 90 : List *gset = (List *) lfirst(lc);
2316 :
2317 90 : if (bms_overlap_list(gd->unsortable_refs, gset))
2318 : {
2319 48 : GroupingSetData *gs = makeNode(GroupingSetData);
2320 :
2321 48 : gs->set = gset;
2322 48 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2323 :
2324 : /*
2325 : * We must enforce here that an unsortable set is hashable;
2326 : * later code assumes this. Parse analysis only checks that
2327 : * every individual column is either hashable or sortable.
2328 : *
2329 : * Note that passing this test doesn't guarantee we can
2330 : * generate a plan; there might be other showstoppers.
2331 : */
2332 48 : if (bms_overlap_list(gd->unhashable_refs, gset))
2333 6 : ereport(ERROR,
2334 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2335 : errmsg("could not implement GROUP BY"),
2336 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2337 : }
2338 : else
2339 42 : sortable_sets = lappend(sortable_sets, gset);
2340 : }
2341 :
2342 36 : if (sortable_sets)
2343 30 : sets = extract_rollup_sets(sortable_sets);
2344 : else
2345 6 : sets = NIL;
2346 : }
2347 : else
2348 978 : sets = extract_rollup_sets(parse->groupingSets);
2349 :
2350 2654 : foreach(lc_set, sets)
2351 : {
2352 1640 : List *current_sets = (List *) lfirst(lc_set);
2353 1640 : RollupData *rollup = makeNode(RollupData);
2354 : GroupingSetData *gs;
2355 :
2356 : /*
2357 : * Reorder the current list of grouping sets into correct prefix
2358 : * order. If only one aggregation pass is needed, try to make the
2359 : * list match the ORDER BY clause; if more than one pass is needed, we
2360 : * don't bother with that.
2361 : *
2362 : * Note that this reorders the sets from smallest-member-first to
2363 : * largest-member-first, and applies the GroupingSetData annotations,
2364 : * though the data will be filled in later.
2365 : */
2366 1640 : current_sets = reorder_grouping_sets(current_sets,
2367 1640 : (list_length(sets) == 1
2368 : ? parse->sortClause
2369 : : NIL));
2370 :
2371 : /*
2372 : * Get the initial (and therefore largest) grouping set.
2373 : */
2374 1640 : gs = linitial_node(GroupingSetData, current_sets);
2375 :
2376 : /*
2377 : * Order the groupClause appropriately. If the first grouping set is
2378 : * empty, then the groupClause must also be empty; otherwise we have
2379 : * to force the groupClause to match that grouping set's order.
2380 : *
2381 : * (The first grouping set can be empty even though parse->groupClause
2382 : * is not empty only if all non-empty grouping sets are unsortable.
2383 : * The groupClauses for hashed grouping sets are built later on.)
2384 : */
2385 1640 : if (gs->set)
2386 1562 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2387 : else
2388 78 : rollup->groupClause = NIL;
2389 :
2390 : /*
2391 : * Is it hashable? We pretend empty sets are hashable even though we
2392 : * actually force them not to be hashed later. But don't bother if
2393 : * there's nothing but empty sets (since in that case we can't hash
2394 : * anything).
2395 : */
2396 1640 : if (gs->set &&
2397 1562 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2398 : {
2399 1538 : rollup->hashable = true;
2400 1538 : gd->any_hashable = true;
2401 : }
2402 :
2403 : /*
2404 : * Now that we've pinned down an order for the groupClause for this
2405 : * list of grouping sets, we need to remap the entries in the grouping
2406 : * sets from sortgrouprefs to plain indices (0-based) into the
2407 : * groupClause for this collection of grouping sets. We keep the
2408 : * original form for later use, though.
2409 : */
2410 1640 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2411 : current_sets,
2412 : gd->tleref_to_colnum_map);
2413 1640 : rollup->gsets_data = current_sets;
2414 :
2415 1640 : gd->rollups = lappend(gd->rollups, rollup);
2416 : }
2417 :
2418 1014 : if (gd->unsortable_sets)
2419 : {
2420 : /*
2421 : * We have not yet pinned down a groupclause for this, but we will
2422 : * need index-based lists for estimation purposes. Construct
2423 : * hash_sets_idx based on the entire original groupclause for now.
2424 : */
2425 36 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2426 : gd->unsortable_sets,
2427 : gd->tleref_to_colnum_map);
2428 36 : gd->any_hashable = true;
2429 : }
2430 :
2431 1014 : return gd;
2432 : }
2433 :
2434 : /*
2435 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2436 : * (without annotation) mapped to indexes into the given groupclause.
2437 : */
2438 : static List *
2439 4628 : remap_to_groupclause_idx(List *groupClause,
2440 : List *gsets,
2441 : int *tleref_to_colnum_map)
2442 : {
2443 4628 : int ref = 0;
2444 4628 : List *result = NIL;
2445 : ListCell *lc;
2446 :
2447 11132 : foreach(lc, groupClause)
2448 : {
2449 6504 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2450 :
2451 6504 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2452 : }
2453 :
2454 10640 : foreach(lc, gsets)
2455 : {
2456 6012 : List *set = NIL;
2457 : ListCell *lc2;
2458 6012 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2459 :
2460 13378 : foreach(lc2, gs->set)
2461 : {
2462 7366 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2463 : }
2464 :
2465 6012 : result = lappend(result, set);
2466 : }
2467 :
2468 4628 : return result;
2469 : }
2470 :
2471 :
2472 : /*
2473 : * preprocess_rowmarks - set up PlanRowMarks if needed
2474 : */
2475 : static void
2476 554748 : preprocess_rowmarks(PlannerInfo *root)
2477 : {
2478 554748 : Query *parse = root->parse;
2479 : Bitmapset *rels;
2480 : List *prowmarks;
2481 : ListCell *l;
2482 : int i;
2483 :
2484 554748 : if (parse->rowMarks)
2485 : {
2486 : /*
2487 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2488 : * grouping, since grouping renders a reference to individual tuple
2489 : * CTIDs invalid. This is also checked at parse time, but that's
2490 : * insufficient because of rule substitution, query pullup, etc.
2491 : */
2492 13544 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2493 : parse->rowMarks)->strength);
2494 : }
2495 : else
2496 : {
2497 : /*
2498 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2499 : * UPDATE/SHARE.
2500 : */
2501 541204 : if (parse->commandType != CMD_UPDATE &&
2502 526466 : parse->commandType != CMD_DELETE &&
2503 522082 : parse->commandType != CMD_MERGE)
2504 520228 : return;
2505 : }
2506 :
2507 : /*
2508 : * We need to have rowmarks for all base relations except the target. We
2509 : * make a bitmapset of all base rels and then remove the items we don't
2510 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2511 : */
2512 34520 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2513 34520 : if (parse->resultRelation)
2514 20976 : rels = bms_del_member(rels, parse->resultRelation);
2515 :
2516 : /*
2517 : * Convert RowMarkClauses to PlanRowMark representation.
2518 : */
2519 34520 : prowmarks = NIL;
2520 48294 : foreach(l, parse->rowMarks)
2521 : {
2522 13774 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2523 13774 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2524 : PlanRowMark *newrc;
2525 :
2526 : /*
2527 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2528 : * applied to an update/delete target rel. If that ever becomes
2529 : * possible, we should drop the target from the PlanRowMark list.
2530 : */
2531 : Assert(rc->rti != parse->resultRelation);
2532 :
2533 : /*
2534 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2535 : * can't support true locking. Subqueries that got flattened into the
2536 : * main query should be ignored completely. Any that didn't will get
2537 : * ROW_MARK_COPY items in the next loop.
2538 : */
2539 13774 : if (rte->rtekind != RTE_RELATION)
2540 60 : continue;
2541 :
2542 13714 : rels = bms_del_member(rels, rc->rti);
2543 :
2544 13714 : newrc = makeNode(PlanRowMark);
2545 13714 : newrc->rti = newrc->prti = rc->rti;
2546 13714 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2547 13714 : newrc->markType = select_rowmark_type(rte, rc->strength);
2548 13714 : newrc->allMarkTypes = (1 << newrc->markType);
2549 13714 : newrc->strength = rc->strength;
2550 13714 : newrc->waitPolicy = rc->waitPolicy;
2551 13714 : newrc->isParent = false;
2552 :
2553 13714 : prowmarks = lappend(prowmarks, newrc);
2554 : }
2555 :
2556 : /*
2557 : * Now, add rowmarks for any non-target, non-locked base relations.
2558 : */
2559 34520 : i = 0;
2560 79900 : foreach(l, parse->rtable)
2561 : {
2562 45380 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2563 : PlanRowMark *newrc;
2564 :
2565 45380 : i++;
2566 45380 : if (!bms_is_member(i, rels))
2567 41610 : continue;
2568 :
2569 3770 : newrc = makeNode(PlanRowMark);
2570 3770 : newrc->rti = newrc->prti = i;
2571 3770 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2572 3770 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2573 3770 : newrc->allMarkTypes = (1 << newrc->markType);
2574 3770 : newrc->strength = LCS_NONE;
2575 3770 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2576 3770 : newrc->isParent = false;
2577 :
2578 3770 : prowmarks = lappend(prowmarks, newrc);
2579 : }
2580 :
2581 34520 : root->rowMarks = prowmarks;
2582 : }
2583 :
2584 : /*
2585 : * Select RowMarkType to use for a given table
2586 : */
2587 : RowMarkType
2588 19900 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2589 : {
2590 19900 : if (rte->rtekind != RTE_RELATION)
2591 : {
2592 : /* If it's not a table at all, use ROW_MARK_COPY */
2593 1438 : return ROW_MARK_COPY;
2594 : }
2595 18462 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2596 : {
2597 : /* Let the FDW select the rowmark type, if it wants to */
2598 228 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2599 :
2600 228 : if (fdwroutine->GetForeignRowMarkType != NULL)
2601 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2602 : /* Otherwise, use ROW_MARK_COPY by default */
2603 228 : return ROW_MARK_COPY;
2604 : }
2605 : else
2606 : {
2607 : /* Regular table, apply the appropriate lock type */
2608 18234 : switch (strength)
2609 : {
2610 2546 : case LCS_NONE:
2611 :
2612 : /*
2613 : * We don't need a tuple lock, only the ability to re-fetch
2614 : * the row.
2615 : */
2616 2546 : return ROW_MARK_REFERENCE;
2617 : break;
2618 13756 : case LCS_FORKEYSHARE:
2619 13756 : return ROW_MARK_KEYSHARE;
2620 : break;
2621 306 : case LCS_FORSHARE:
2622 306 : return ROW_MARK_SHARE;
2623 : break;
2624 78 : case LCS_FORNOKEYUPDATE:
2625 78 : return ROW_MARK_NOKEYEXCLUSIVE;
2626 : break;
2627 1548 : case LCS_FORUPDATE:
2628 1548 : return ROW_MARK_EXCLUSIVE;
2629 : break;
2630 : }
2631 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2632 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2633 : }
2634 : }
2635 :
2636 : /*
2637 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2638 : *
2639 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2640 : * results back in *count_est and *offset_est. These variables are set to
2641 : * 0 if the corresponding clause is not present, and -1 if it's present
2642 : * but we couldn't estimate the value for it. (The "0" convention is OK
2643 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2644 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2645 : * usual practice of never estimating less than one row.) These values will
2646 : * be passed to create_limit_path, which see if you change this code.
2647 : *
2648 : * The return value is the suitably adjusted tuple_fraction to use for
2649 : * planning the query. This adjustment is not overridable, since it reflects
2650 : * plan actions that grouping_planner() will certainly take, not assumptions
2651 : * about context.
2652 : */
2653 : static double
2654 5246 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2655 : int64 *offset_est, int64 *count_est)
2656 : {
2657 5246 : Query *parse = root->parse;
2658 : Node *est;
2659 : double limit_fraction;
2660 :
2661 : /* Should not be called unless LIMIT or OFFSET */
2662 : Assert(parse->limitCount || parse->limitOffset);
2663 :
2664 : /*
2665 : * Try to obtain the clause values. We use estimate_expression_value
2666 : * primarily because it can sometimes do something useful with Params.
2667 : */
2668 5246 : if (parse->limitCount)
2669 : {
2670 4718 : est = estimate_expression_value(root, parse->limitCount);
2671 4718 : if (est && IsA(est, Const))
2672 : {
2673 4712 : if (((Const *) est)->constisnull)
2674 : {
2675 : /* NULL indicates LIMIT ALL, ie, no limit */
2676 0 : *count_est = 0; /* treat as not present */
2677 : }
2678 : else
2679 : {
2680 4712 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2681 4712 : if (*count_est <= 0)
2682 150 : *count_est = 1; /* force to at least 1 */
2683 : }
2684 : }
2685 : else
2686 6 : *count_est = -1; /* can't estimate */
2687 : }
2688 : else
2689 528 : *count_est = 0; /* not present */
2690 :
2691 5246 : if (parse->limitOffset)
2692 : {
2693 900 : est = estimate_expression_value(root, parse->limitOffset);
2694 900 : if (est && IsA(est, Const))
2695 : {
2696 876 : if (((Const *) est)->constisnull)
2697 : {
2698 : /* Treat NULL as no offset; the executor will too */
2699 0 : *offset_est = 0; /* treat as not present */
2700 : }
2701 : else
2702 : {
2703 876 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2704 876 : if (*offset_est < 0)
2705 0 : *offset_est = 0; /* treat as not present */
2706 : }
2707 : }
2708 : else
2709 24 : *offset_est = -1; /* can't estimate */
2710 : }
2711 : else
2712 4346 : *offset_est = 0; /* not present */
2713 :
2714 5246 : if (*count_est != 0)
2715 : {
2716 : /*
2717 : * A LIMIT clause limits the absolute number of tuples returned.
2718 : * However, if it's not a constant LIMIT then we have to guess; for
2719 : * lack of a better idea, assume 10% of the plan's result is wanted.
2720 : */
2721 4718 : if (*count_est < 0 || *offset_est < 0)
2722 : {
2723 : /* LIMIT or OFFSET is an expression ... punt ... */
2724 24 : limit_fraction = 0.10;
2725 : }
2726 : else
2727 : {
2728 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2729 4694 : limit_fraction = (double) *count_est + (double) *offset_est;
2730 : }
2731 :
2732 : /*
2733 : * If we have absolute limits from both caller and LIMIT, use the
2734 : * smaller value; likewise if they are both fractional. If one is
2735 : * fractional and the other absolute, we can't easily determine which
2736 : * is smaller, but we use the heuristic that the absolute will usually
2737 : * be smaller.
2738 : */
2739 4718 : if (tuple_fraction >= 1.0)
2740 : {
2741 6 : if (limit_fraction >= 1.0)
2742 : {
2743 : /* both absolute */
2744 6 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2745 : }
2746 : else
2747 : {
2748 : /* caller absolute, limit fractional; use caller's value */
2749 : }
2750 : }
2751 4712 : else if (tuple_fraction > 0.0)
2752 : {
2753 148 : if (limit_fraction >= 1.0)
2754 : {
2755 : /* caller fractional, limit absolute; use limit */
2756 148 : tuple_fraction = limit_fraction;
2757 : }
2758 : else
2759 : {
2760 : /* both fractional */
2761 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2762 : }
2763 : }
2764 : else
2765 : {
2766 : /* no info from caller, just use limit */
2767 4564 : tuple_fraction = limit_fraction;
2768 : }
2769 : }
2770 528 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2771 : {
2772 : /*
2773 : * We have an OFFSET but no LIMIT. This acts entirely differently
2774 : * from the LIMIT case: here, we need to increase rather than decrease
2775 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2776 : * tuples to be fetched instead of fewer. This only matters if we got
2777 : * a tuple_fraction > 0, however.
2778 : *
2779 : * As above, use 10% if OFFSET is present but unestimatable.
2780 : */
2781 16 : if (*offset_est < 0)
2782 0 : limit_fraction = 0.10;
2783 : else
2784 16 : limit_fraction = (double) *offset_est;
2785 :
2786 : /*
2787 : * If we have absolute counts from both caller and OFFSET, add them
2788 : * together; likewise if they are both fractional. If one is
2789 : * fractional and the other absolute, we want to take the larger, and
2790 : * we heuristically assume that's the fractional one.
2791 : */
2792 16 : if (tuple_fraction >= 1.0)
2793 : {
2794 0 : if (limit_fraction >= 1.0)
2795 : {
2796 : /* both absolute, so add them together */
2797 0 : tuple_fraction += limit_fraction;
2798 : }
2799 : else
2800 : {
2801 : /* caller absolute, limit fractional; use limit */
2802 0 : tuple_fraction = limit_fraction;
2803 : }
2804 : }
2805 : else
2806 : {
2807 16 : if (limit_fraction >= 1.0)
2808 : {
2809 : /* caller fractional, limit absolute; use caller's value */
2810 : }
2811 : else
2812 : {
2813 : /* both fractional, so add them together */
2814 0 : tuple_fraction += limit_fraction;
2815 0 : if (tuple_fraction >= 1.0)
2816 0 : tuple_fraction = 0.0; /* assume fetch all */
2817 : }
2818 : }
2819 : }
2820 :
2821 5246 : return tuple_fraction;
2822 : }
2823 :
2824 : /*
2825 : * limit_needed - do we actually need a Limit plan node?
2826 : *
2827 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2828 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2829 : * locution for an optimization fence. (Because other places in the planner
2830 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2831 : * an optimization fence --- we're just suppressing unnecessary run-time
2832 : * overhead.)
2833 : *
2834 : * This might look like it could be merged into preprocess_limit, but there's
2835 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2836 : * in preprocess_limit it's good enough to consider estimated values.
2837 : */
2838 : bool
2839 1173724 : limit_needed(Query *parse)
2840 : {
2841 : Node *node;
2842 :
2843 1173724 : node = parse->limitCount;
2844 1173724 : if (node)
2845 : {
2846 11304 : if (IsA(node, Const))
2847 : {
2848 : /* NULL indicates LIMIT ALL, ie, no limit */
2849 11068 : if (!((Const *) node)->constisnull)
2850 11068 : return true; /* LIMIT with a constant value */
2851 : }
2852 : else
2853 236 : return true; /* non-constant LIMIT */
2854 : }
2855 :
2856 1162420 : node = parse->limitOffset;
2857 1162420 : if (node)
2858 : {
2859 1534 : if (IsA(node, Const))
2860 : {
2861 : /* Treat NULL as no offset; the executor would too */
2862 1226 : if (!((Const *) node)->constisnull)
2863 : {
2864 1226 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2865 :
2866 1226 : if (offset != 0)
2867 146 : return true; /* OFFSET with a nonzero value */
2868 : }
2869 : }
2870 : else
2871 308 : return true; /* non-constant OFFSET */
2872 : }
2873 :
2874 1161966 : return false; /* don't need a Limit plan node */
2875 : }
2876 :
2877 : /*
2878 : * preprocess_groupclause - do preparatory work on GROUP BY clause
2879 : *
2880 : * The idea here is to adjust the ordering of the GROUP BY elements
2881 : * (which in itself is semantically insignificant) to match ORDER BY,
2882 : * thereby allowing a single sort operation to both implement the ORDER BY
2883 : * requirement and set up for a Unique step that implements GROUP BY.
2884 : * We also consider partial match between GROUP BY and ORDER BY elements,
2885 : * which could allow to implement ORDER BY using the incremental sort.
2886 : *
2887 : * We also consider other orderings of the GROUP BY elements, which could
2888 : * match the sort ordering of other possible plans (eg an indexscan) and
2889 : * thereby reduce cost. This is implemented during the generation of grouping
2890 : * paths. See get_useful_group_keys_orderings() for details.
2891 : *
2892 : * Note: we need no comparable processing of the distinctClause because
2893 : * the parser already enforced that that matches ORDER BY.
2894 : *
2895 : * Note: we return a fresh List, but its elements are the same
2896 : * SortGroupClauses appearing in parse->groupClause. This is important
2897 : * because later processing may modify the processed_groupClause list.
2898 : *
2899 : * For grouping sets, the order of items is instead forced to agree with that
2900 : * of the grouping set (and items not in the grouping set are skipped). The
2901 : * work of sorting the order of grouping set elements to match the ORDER BY if
2902 : * possible is done elsewhere.
2903 : */
2904 : static List *
2905 8646 : preprocess_groupclause(PlannerInfo *root, List *force)
2906 : {
2907 8646 : Query *parse = root->parse;
2908 8646 : List *new_groupclause = NIL;
2909 : ListCell *sl;
2910 : ListCell *gl;
2911 :
2912 : /* For grouping sets, we need to force the ordering */
2913 8646 : if (force)
2914 : {
2915 10940 : foreach(sl, force)
2916 : {
2917 6426 : Index ref = lfirst_int(sl);
2918 6426 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2919 :
2920 6426 : new_groupclause = lappend(new_groupclause, cl);
2921 : }
2922 :
2923 4514 : return new_groupclause;
2924 : }
2925 :
2926 : /* If no ORDER BY, nothing useful to do here */
2927 4132 : if (parse->sortClause == NIL)
2928 2316 : return list_copy(parse->groupClause);
2929 :
2930 : /*
2931 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2932 : * items, but only as far as we can make a matching prefix.
2933 : *
2934 : * This code assumes that the sortClause contains no duplicate items.
2935 : */
2936 3538 : foreach(sl, parse->sortClause)
2937 : {
2938 2368 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2939 :
2940 3464 : foreach(gl, parse->groupClause)
2941 : {
2942 2818 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2943 :
2944 2818 : if (equal(gc, sc))
2945 : {
2946 1722 : new_groupclause = lappend(new_groupclause, gc);
2947 1722 : break;
2948 : }
2949 : }
2950 2368 : if (gl == NULL)
2951 646 : break; /* no match, so stop scanning */
2952 : }
2953 :
2954 :
2955 : /* If no match at all, no point in reordering GROUP BY */
2956 1816 : if (new_groupclause == NIL)
2957 298 : return list_copy(parse->groupClause);
2958 :
2959 : /*
2960 : * Add any remaining GROUP BY items to the new list. We don't require a
2961 : * complete match, because even partial match allows ORDER BY to be
2962 : * implemented using incremental sort. Also, give up if there are any
2963 : * non-sortable GROUP BY items, since then there's no hope anyway.
2964 : */
2965 3406 : foreach(gl, parse->groupClause)
2966 : {
2967 1888 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2968 :
2969 1888 : if (list_member_ptr(new_groupclause, gc))
2970 1722 : continue; /* it matched an ORDER BY item */
2971 166 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2972 0 : return list_copy(parse->groupClause);
2973 166 : new_groupclause = lappend(new_groupclause, gc);
2974 : }
2975 :
2976 : /* Success --- install the rearranged GROUP BY list */
2977 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2978 1518 : return new_groupclause;
2979 : }
2980 :
2981 : /*
2982 : * Extract lists of grouping sets that can be implemented using a single
2983 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2984 : *
2985 : * Input must be sorted with smallest sets first. Result has each sublist
2986 : * sorted with smallest sets first.
2987 : *
2988 : * We want to produce the absolute minimum possible number of lists here to
2989 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2990 : * of finding the minimal partition of a partially-ordered set into chains
2991 : * (which is what we need, taking the list of grouping sets as a poset ordered
2992 : * by set inclusion) can be mapped to the problem of finding the maximum
2993 : * cardinality matching on a bipartite graph, which is solvable in polynomial
2994 : * time with a worst case of no worse than O(n^2.5) and usually much
2995 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
2996 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
2997 : * half a second on my modest system even with optimization off and assertions
2998 : * on.)
2999 : */
3000 : static List *
3001 1008 : extract_rollup_sets(List *groupingSets)
3002 : {
3003 1008 : int num_sets_raw = list_length(groupingSets);
3004 1008 : int num_empty = 0;
3005 1008 : int num_sets = 0; /* distinct sets */
3006 1008 : int num_chains = 0;
3007 1008 : List *result = NIL;
3008 : List **results;
3009 : List **orig_sets;
3010 : Bitmapset **set_masks;
3011 : int *chains;
3012 : short **adjacency;
3013 : short *adjacency_buf;
3014 : BipartiteMatchState *state;
3015 : int i;
3016 : int j;
3017 : int j_size;
3018 1008 : ListCell *lc1 = list_head(groupingSets);
3019 : ListCell *lc;
3020 :
3021 : /*
3022 : * Start by stripping out empty sets. The algorithm doesn't require this,
3023 : * but the planner currently needs all empty sets to be returned in the
3024 : * first list, so we strip them here and add them back after.
3025 : */
3026 1712 : while (lc1 && lfirst(lc1) == NIL)
3027 : {
3028 704 : ++num_empty;
3029 704 : lc1 = lnext(groupingSets, lc1);
3030 : }
3031 :
3032 : /* bail out now if it turns out that all we had were empty sets. */
3033 1008 : if (!lc1)
3034 78 : return list_make1(groupingSets);
3035 :
3036 : /*----------
3037 : * We don't strictly need to remove duplicate sets here, but if we don't,
3038 : * they tend to become scattered through the result, which is a bit
3039 : * confusing (and irritating if we ever decide to optimize them out).
3040 : * So we remove them here and add them back after.
3041 : *
3042 : * For each non-duplicate set, we fill in the following:
3043 : *
3044 : * orig_sets[i] = list of the original set lists
3045 : * set_masks[i] = bitmapset for testing inclusion
3046 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3047 : *
3048 : * chains[i] will be the result group this set is assigned to.
3049 : *
3050 : * We index all of these from 1 rather than 0 because it is convenient
3051 : * to leave 0 free for the NIL node in the graph algorithm.
3052 : *----------
3053 : */
3054 930 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3055 930 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3056 930 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3057 930 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3058 :
3059 930 : j_size = 0;
3060 930 : j = 0;
3061 930 : i = 1;
3062 :
3063 3244 : for_each_cell(lc, groupingSets, lc1)
3064 : {
3065 2314 : List *candidate = (List *) lfirst(lc);
3066 2314 : Bitmapset *candidate_set = NULL;
3067 : ListCell *lc2;
3068 2314 : int dup_of = 0;
3069 :
3070 5546 : foreach(lc2, candidate)
3071 : {
3072 3232 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3073 : }
3074 :
3075 : /* we can only be a dup if we're the same length as a previous set */
3076 2314 : if (j_size == list_length(candidate))
3077 : {
3078 : int k;
3079 :
3080 2072 : for (k = j; k < i; ++k)
3081 : {
3082 1344 : if (bms_equal(set_masks[k], candidate_set))
3083 : {
3084 158 : dup_of = k;
3085 158 : break;
3086 : }
3087 : }
3088 : }
3089 1428 : else if (j_size < list_length(candidate))
3090 : {
3091 1428 : j_size = list_length(candidate);
3092 1428 : j = i;
3093 : }
3094 :
3095 2314 : if (dup_of > 0)
3096 : {
3097 158 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3098 158 : bms_free(candidate_set);
3099 : }
3100 : else
3101 : {
3102 : int k;
3103 2156 : int n_adj = 0;
3104 :
3105 2156 : orig_sets[i] = list_make1(candidate);
3106 2156 : set_masks[i] = candidate_set;
3107 :
3108 : /* fill in adjacency list; no need to compare equal-size sets */
3109 :
3110 3452 : for (k = j - 1; k > 0; --k)
3111 : {
3112 1296 : if (bms_is_subset(set_masks[k], candidate_set))
3113 1134 : adjacency_buf[++n_adj] = k;
3114 : }
3115 :
3116 2156 : if (n_adj > 0)
3117 : {
3118 622 : adjacency_buf[0] = n_adj;
3119 622 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3120 622 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3121 : }
3122 : else
3123 1534 : adjacency[i] = NULL;
3124 :
3125 2156 : ++i;
3126 : }
3127 : }
3128 :
3129 930 : num_sets = i - 1;
3130 :
3131 : /*
3132 : * Apply the graph matching algorithm to do the work.
3133 : */
3134 930 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3135 :
3136 : /*
3137 : * Now, the state->pair* fields have the info we need to assign sets to
3138 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3139 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3140 : * it in one pass)
3141 : */
3142 930 : chains = palloc0((num_sets + 1) * sizeof(int));
3143 :
3144 3086 : for (i = 1; i <= num_sets; ++i)
3145 : {
3146 2156 : int u = state->pair_vu[i];
3147 2156 : int v = state->pair_uv[i];
3148 :
3149 2156 : if (u > 0 && u < i)
3150 0 : chains[i] = chains[u];
3151 2156 : else if (v > 0 && v < i)
3152 594 : chains[i] = chains[v];
3153 : else
3154 1562 : chains[i] = ++num_chains;
3155 : }
3156 :
3157 : /* build result lists. */
3158 930 : results = palloc0((num_chains + 1) * sizeof(List *));
3159 :
3160 3086 : for (i = 1; i <= num_sets; ++i)
3161 : {
3162 2156 : int c = chains[i];
3163 :
3164 : Assert(c > 0);
3165 :
3166 2156 : results[c] = list_concat(results[c], orig_sets[i]);
3167 : }
3168 :
3169 : /* push any empty sets back on the first list. */
3170 1496 : while (num_empty-- > 0)
3171 566 : results[1] = lcons(NIL, results[1]);
3172 :
3173 : /* make result list */
3174 2492 : for (i = 1; i <= num_chains; ++i)
3175 1562 : result = lappend(result, results[i]);
3176 :
3177 : /*
3178 : * Free all the things.
3179 : *
3180 : * (This is over-fussy for small sets but for large sets we could have
3181 : * tied up a nontrivial amount of memory.)
3182 : */
3183 930 : BipartiteMatchFree(state);
3184 930 : pfree(results);
3185 930 : pfree(chains);
3186 3086 : for (i = 1; i <= num_sets; ++i)
3187 2156 : if (adjacency[i])
3188 622 : pfree(adjacency[i]);
3189 930 : pfree(adjacency);
3190 930 : pfree(adjacency_buf);
3191 930 : pfree(orig_sets);
3192 3086 : for (i = 1; i <= num_sets; ++i)
3193 2156 : bms_free(set_masks[i]);
3194 930 : pfree(set_masks);
3195 :
3196 930 : return result;
3197 : }
3198 :
3199 : /*
3200 : * Reorder the elements of a list of grouping sets such that they have correct
3201 : * prefix relationships. Also inserts the GroupingSetData annotations.
3202 : *
3203 : * The input must be ordered with smallest sets first; the result is returned
3204 : * with largest sets first. Note that the result shares no list substructure
3205 : * with the input, so it's safe for the caller to modify it later.
3206 : *
3207 : * If we're passed in a sortclause, we follow its order of columns to the
3208 : * extent possible, to minimize the chance that we add unnecessary sorts.
3209 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3210 : * gets implemented in one pass.)
3211 : */
3212 : static List *
3213 1640 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3214 : {
3215 : ListCell *lc;
3216 1640 : List *previous = NIL;
3217 1640 : List *result = NIL;
3218 :
3219 4658 : foreach(lc, groupingSets)
3220 : {
3221 3018 : List *candidate = (List *) lfirst(lc);
3222 3018 : List *new_elems = list_difference_int(candidate, previous);
3223 3018 : GroupingSetData *gs = makeNode(GroupingSetData);
3224 :
3225 3194 : while (list_length(sortclause) > list_length(previous) &&
3226 : new_elems != NIL)
3227 : {
3228 296 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3229 296 : int ref = sc->tleSortGroupRef;
3230 :
3231 296 : if (list_member_int(new_elems, ref))
3232 : {
3233 176 : previous = lappend_int(previous, ref);
3234 176 : new_elems = list_delete_int(new_elems, ref);
3235 : }
3236 : else
3237 : {
3238 : /* diverged from the sortclause; give up on it */
3239 120 : sortclause = NIL;
3240 120 : break;
3241 : }
3242 : }
3243 :
3244 3018 : previous = list_concat(previous, new_elems);
3245 :
3246 3018 : gs->set = list_copy(previous);
3247 3018 : result = lcons(gs, result);
3248 : }
3249 :
3250 1640 : list_free(previous);
3251 :
3252 1640 : return result;
3253 : }
3254 :
3255 : /*
3256 : * has_volatile_pathkey
3257 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3258 : * containing a volatile function. Otherwise returns false.
3259 : */
3260 : static bool
3261 2836 : has_volatile_pathkey(List *keys)
3262 : {
3263 : ListCell *lc;
3264 :
3265 5816 : foreach(lc, keys)
3266 : {
3267 2998 : PathKey *pathkey = lfirst_node(PathKey, lc);
3268 :
3269 2998 : if (pathkey->pk_eclass->ec_has_volatile)
3270 18 : return true;
3271 : }
3272 :
3273 2818 : return false;
3274 : }
3275 :
3276 : /*
3277 : * adjust_group_pathkeys_for_groupagg
3278 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3279 : * pre-ordered input for ordered aggregates.
3280 : *
3281 : * We define "best" as the pathkeys that suit the largest number of
3282 : * aggregate functions. We find these by looking at the first ORDER BY /
3283 : * DISTINCT aggregate and take the pathkeys for that before searching for
3284 : * other aggregates that require the same or a more strict variation of the
3285 : * same pathkeys. We then repeat that process for any remaining aggregates
3286 : * with different pathkeys and if we find another set of pathkeys that suits a
3287 : * larger number of aggregates then we select those pathkeys instead.
3288 : *
3289 : * When the best pathkeys are found we also mark each Aggref that can use
3290 : * those pathkeys as aggpresorted = true.
3291 : *
3292 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3293 : * volatile functions, we never make use of these pathkeys. We want to ensure
3294 : * that sorts using volatile functions are done independently in each Aggref
3295 : * rather than once at the query level. If we were to allow this then Aggrefs
3296 : * with compatible sort orders would all transition their rows in the same
3297 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3298 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3299 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3300 : * left to perform their sorts individually. To avoid this inconsistent
3301 : * behavior which could make Aggref results depend on what other Aggrefs the
3302 : * query contains, we always force Aggrefs with volatile functions to perform
3303 : * their own sorts.
3304 : */
3305 : static void
3306 2440 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3307 : {
3308 2440 : List *grouppathkeys = root->group_pathkeys;
3309 : List *bestpathkeys;
3310 : Bitmapset *bestaggs;
3311 : Bitmapset *unprocessed_aggs;
3312 : ListCell *lc;
3313 : int i;
3314 :
3315 : /* Shouldn't be here if there are grouping sets */
3316 : Assert(root->parse->groupingSets == NIL);
3317 : /* Shouldn't be here unless there are some ordered aggregates */
3318 : Assert(root->numOrderedAggs > 0);
3319 :
3320 : /* Do nothing if disabled */
3321 2440 : if (!enable_presorted_aggregate)
3322 6 : return;
3323 :
3324 : /*
3325 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3326 : * the indexes of all AggInfos to be processed below.
3327 : */
3328 2434 : unprocessed_aggs = NULL;
3329 5552 : foreach(lc, root->agginfos)
3330 : {
3331 3118 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3332 3118 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3333 :
3334 3118 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3335 264 : continue;
3336 :
3337 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3338 2854 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3339 300 : continue;
3340 :
3341 : /* Additional safety checks are needed if there's a FILTER clause */
3342 2554 : if (aggref->aggfilter != NULL)
3343 : {
3344 : ListCell *lc2;
3345 54 : bool allow_presort = true;
3346 :
3347 : /*
3348 : * When the Aggref has a FILTER clause, it's possible that the
3349 : * filter removes rows that cannot be sorted because the
3350 : * expression to sort by results in an error during its
3351 : * evaluation. This is a problem for presorting as that happens
3352 : * before the FILTER, whereas without presorting, the Aggregate
3353 : * node will apply the FILTER *before* sorting. So that we never
3354 : * try to sort anything that might error, here we aim to skip over
3355 : * any Aggrefs with arguments with expressions which, when
3356 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3357 : * may be more cases that should be allowed, but more thought
3358 : * needs to be given. Err on the side of caution.
3359 : */
3360 102 : foreach(lc2, aggref->args)
3361 : {
3362 72 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3363 72 : Expr *expr = tle->expr;
3364 :
3365 84 : while (IsA(expr, RelabelType))
3366 12 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3367 :
3368 : /* Common case, Vars and Consts are ok */
3369 72 : if (IsA(expr, Var) || IsA(expr, Const))
3370 48 : continue;
3371 :
3372 : /* Unsupported. Don't try to presort for this Aggref */
3373 24 : allow_presort = false;
3374 24 : break;
3375 : }
3376 :
3377 : /* Skip unsupported Aggrefs */
3378 54 : if (!allow_presort)
3379 24 : continue;
3380 : }
3381 :
3382 2530 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3383 : foreach_current_index(lc));
3384 : }
3385 :
3386 : /*
3387 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3388 : * for the given set of aggregates.
3389 : *
3390 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3391 : * this during the first loop using the pathkeys for the very first
3392 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3393 : * a more strict set of compatible pathkeys. Once the outer loop is
3394 : * complete, we mark off all the aggregates with compatible pathkeys then
3395 : * remove those from the unprocessed_aggs and repeat the process to try to
3396 : * find another set of pathkeys that are suitable for a larger number of
3397 : * aggregates. The outer loop will stop when there are not enough
3398 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3399 : * to suit a larger number of aggregates.
3400 : */
3401 2434 : bestpathkeys = NIL;
3402 2434 : bestaggs = NULL;
3403 4802 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3404 : {
3405 2368 : Bitmapset *aggindexes = NULL;
3406 2368 : List *currpathkeys = NIL;
3407 :
3408 2368 : i = -1;
3409 5204 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3410 : {
3411 2836 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3412 2836 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3413 : List *sortlist;
3414 : List *pathkeys;
3415 :
3416 2836 : if (aggref->aggdistinct != NIL)
3417 724 : sortlist = aggref->aggdistinct;
3418 : else
3419 2112 : sortlist = aggref->aggorder;
3420 :
3421 2836 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3422 : aggref->args);
3423 :
3424 : /*
3425 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3426 : * or DISTINCT clause.
3427 : */
3428 2836 : if (has_volatile_pathkey(pathkeys))
3429 : {
3430 18 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3431 18 : continue;
3432 : }
3433 :
3434 : /*
3435 : * When not set yet, take the pathkeys from the first unprocessed
3436 : * aggregate.
3437 : */
3438 2818 : if (currpathkeys == NIL)
3439 : {
3440 2362 : currpathkeys = pathkeys;
3441 :
3442 : /* include the GROUP BY pathkeys, if they exist */
3443 2362 : if (grouppathkeys != NIL)
3444 276 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3445 : currpathkeys);
3446 :
3447 : /* record that we found pathkeys for this aggregate */
3448 2362 : aggindexes = bms_add_member(aggindexes, i);
3449 : }
3450 : else
3451 : {
3452 : /* now look for a stronger set of matching pathkeys */
3453 :
3454 : /* include the GROUP BY pathkeys, if they exist */
3455 456 : if (grouppathkeys != NIL)
3456 288 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3457 : pathkeys);
3458 :
3459 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3460 456 : switch (compare_pathkeys(currpathkeys, pathkeys))
3461 : {
3462 12 : case PATHKEYS_BETTER2:
3463 : /* 'pathkeys' are stronger, use these ones instead */
3464 12 : currpathkeys = pathkeys;
3465 : /* FALLTHROUGH */
3466 :
3467 66 : case PATHKEYS_BETTER1:
3468 : /* 'pathkeys' are less strict */
3469 : /* FALLTHROUGH */
3470 :
3471 : case PATHKEYS_EQUAL:
3472 : /* mark this aggregate as covered by 'currpathkeys' */
3473 66 : aggindexes = bms_add_member(aggindexes, i);
3474 66 : break;
3475 :
3476 390 : case PATHKEYS_DIFFERENT:
3477 390 : break;
3478 : }
3479 : }
3480 : }
3481 :
3482 : /* remove the aggregates that we've just processed */
3483 2368 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3484 :
3485 : /*
3486 : * If this pass included more aggregates than the previous best then
3487 : * use these ones as the best set.
3488 : */
3489 2368 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3490 : {
3491 2260 : bestaggs = aggindexes;
3492 2260 : bestpathkeys = currpathkeys;
3493 : }
3494 : }
3495 :
3496 : /*
3497 : * If we found any ordered aggregates, update root->group_pathkeys to add
3498 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3499 : * the original GROUP BY pathkeys already.
3500 : */
3501 2434 : if (bestpathkeys != NIL)
3502 2200 : root->group_pathkeys = bestpathkeys;
3503 :
3504 : /*
3505 : * Now that we've found the best set of aggregates we can set the
3506 : * presorted flag to indicate to the executor that it needn't bother
3507 : * performing a sort for these Aggrefs. We're able to do this now as
3508 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3509 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3510 : * of ordered aggregates.
3511 : */
3512 2434 : i = -1;
3513 4730 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3514 : {
3515 2296 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3516 :
3517 4610 : foreach(lc, agginfo->aggrefs)
3518 : {
3519 2314 : Aggref *aggref = lfirst_node(Aggref, lc);
3520 :
3521 2314 : aggref->aggpresorted = true;
3522 : }
3523 : }
3524 : }
3525 :
3526 : /*
3527 : * Compute query_pathkeys and other pathkeys during plan generation
3528 : */
3529 : static void
3530 544526 : standard_qp_callback(PlannerInfo *root, void *extra)
3531 : {
3532 544526 : Query *parse = root->parse;
3533 544526 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3534 544526 : List *tlist = root->processed_tlist;
3535 544526 : List *activeWindows = qp_extra->activeWindows;
3536 :
3537 : /*
3538 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3539 : * aggregate requirements.
3540 : */
3541 544526 : if (qp_extra->gset_data)
3542 : {
3543 : /*
3544 : * With grouping sets, just use the first RollupData's groupClause. We
3545 : * don't make any effort to optimize grouping clauses when there are
3546 : * grouping sets, nor can we combine aggregate ordering keys with
3547 : * grouping.
3548 : */
3549 1014 : List *rollups = qp_extra->gset_data->rollups;
3550 1014 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3551 :
3552 1014 : if (grouping_is_sortable(groupClause))
3553 : {
3554 : bool sortable;
3555 :
3556 : /*
3557 : * The groupClause is logically below the grouping step. So if
3558 : * there is an RTE entry for the grouping step, we need to remove
3559 : * its RT index from the sort expressions before we make PathKeys
3560 : * for them.
3561 : */
3562 1014 : root->group_pathkeys =
3563 1014 : make_pathkeys_for_sortclauses_extended(root,
3564 : &groupClause,
3565 : tlist,
3566 : false,
3567 1014 : parse->hasGroupRTE,
3568 : &sortable,
3569 : false);
3570 : Assert(sortable);
3571 1014 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3572 : }
3573 : else
3574 : {
3575 0 : root->group_pathkeys = NIL;
3576 0 : root->num_groupby_pathkeys = 0;
3577 : }
3578 : }
3579 543512 : else if (parse->groupClause || root->numOrderedAggs > 0)
3580 6328 : {
3581 : /*
3582 : * With a plain GROUP BY list, we can remove any grouping items that
3583 : * are proven redundant by EquivalenceClass processing. For example,
3584 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3585 : * especially common cases, but they're nearly free to detect. Note
3586 : * that we remove redundant items from processed_groupClause but not
3587 : * the original parse->groupClause.
3588 : */
3589 : bool sortable;
3590 :
3591 : /*
3592 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3593 : * EquivalenceClass'es if it's not set yet.
3594 : */
3595 6328 : root->group_pathkeys =
3596 6328 : make_pathkeys_for_sortclauses_extended(root,
3597 : &root->processed_groupClause,
3598 : tlist,
3599 : true,
3600 : false,
3601 : &sortable,
3602 : true);
3603 6328 : if (!sortable)
3604 : {
3605 : /* Can't sort; no point in considering aggregate ordering either */
3606 0 : root->group_pathkeys = NIL;
3607 0 : root->num_groupby_pathkeys = 0;
3608 : }
3609 : else
3610 : {
3611 6328 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3612 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3613 6328 : if (root->numOrderedAggs > 0)
3614 2440 : adjust_group_pathkeys_for_groupagg(root);
3615 : }
3616 : }
3617 : else
3618 : {
3619 537184 : root->group_pathkeys = NIL;
3620 537184 : root->num_groupby_pathkeys = 0;
3621 : }
3622 :
3623 : /* We consider only the first (bottom) window in pathkeys logic */
3624 544526 : if (activeWindows != NIL)
3625 : {
3626 2676 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3627 :
3628 2676 : root->window_pathkeys = make_pathkeys_for_window(root,
3629 : wc,
3630 : tlist);
3631 : }
3632 : else
3633 541850 : root->window_pathkeys = NIL;
3634 :
3635 : /*
3636 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3637 : * redundant by EquivalenceClass processing. The non-redundant list is
3638 : * kept in root->processed_distinctClause, leaving the original
3639 : * parse->distinctClause alone.
3640 : */
3641 544526 : if (parse->distinctClause)
3642 : {
3643 : bool sortable;
3644 :
3645 : /* Make a copy since pathkey processing can modify the list */
3646 3024 : root->processed_distinctClause = list_copy(parse->distinctClause);
3647 3024 : root->distinct_pathkeys =
3648 3024 : make_pathkeys_for_sortclauses_extended(root,
3649 : &root->processed_distinctClause,
3650 : tlist,
3651 : true,
3652 : false,
3653 : &sortable,
3654 : false);
3655 3024 : if (!sortable)
3656 6 : root->distinct_pathkeys = NIL;
3657 : }
3658 : else
3659 541502 : root->distinct_pathkeys = NIL;
3660 :
3661 544526 : root->sort_pathkeys =
3662 544526 : make_pathkeys_for_sortclauses(root,
3663 : parse->sortClause,
3664 : tlist);
3665 :
3666 : /* setting setop_pathkeys might be useful to the union planner */
3667 544526 : if (qp_extra->setop != NULL)
3668 : {
3669 : List *groupClauses;
3670 : bool sortable;
3671 :
3672 12786 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3673 :
3674 12786 : root->setop_pathkeys =
3675 12786 : make_pathkeys_for_sortclauses_extended(root,
3676 : &groupClauses,
3677 : tlist,
3678 : false,
3679 : false,
3680 : &sortable,
3681 : false);
3682 12786 : if (!sortable)
3683 208 : root->setop_pathkeys = NIL;
3684 : }
3685 : else
3686 531740 : root->setop_pathkeys = NIL;
3687 :
3688 : /*
3689 : * Figure out whether we want a sorted result from query_planner.
3690 : *
3691 : * If we have a sortable GROUP BY clause, then we want a result sorted
3692 : * properly for grouping. Otherwise, if we have window functions to
3693 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3694 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3695 : * we try to produce output that's sufficiently well sorted for the
3696 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3697 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3698 : * for a set operation which can benefit from presorted results and have a
3699 : * sortable targetlist, we want to sort by the target list.
3700 : *
3701 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3702 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3703 : * that might just leave us failing to exploit an available sort order at
3704 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3705 : * much easier, since we know that the parser ensured that one is a
3706 : * superset of the other.
3707 : */
3708 544526 : if (root->group_pathkeys)
3709 6904 : root->query_pathkeys = root->group_pathkeys;
3710 537622 : else if (root->window_pathkeys)
3711 2110 : root->query_pathkeys = root->window_pathkeys;
3712 1071024 : else if (list_length(root->distinct_pathkeys) >
3713 535512 : list_length(root->sort_pathkeys))
3714 2528 : root->query_pathkeys = root->distinct_pathkeys;
3715 532984 : else if (root->sort_pathkeys)
3716 69318 : root->query_pathkeys = root->sort_pathkeys;
3717 463666 : else if (root->setop_pathkeys != NIL)
3718 11378 : root->query_pathkeys = root->setop_pathkeys;
3719 : else
3720 452288 : root->query_pathkeys = NIL;
3721 544526 : }
3722 :
3723 : /*
3724 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3725 : *
3726 : * path_rows: number of output rows from scan/join step
3727 : * gd: grouping sets data including list of grouping sets and their clauses
3728 : * target_list: target list containing group clause references
3729 : *
3730 : * If doing grouping sets, we also annotate the gsets data with the estimates
3731 : * for each set and each individual rollup list, with a view to later
3732 : * determining whether some combination of them could be hashed instead.
3733 : */
3734 : static double
3735 54368 : get_number_of_groups(PlannerInfo *root,
3736 : double path_rows,
3737 : grouping_sets_data *gd,
3738 : List *target_list)
3739 : {
3740 54368 : Query *parse = root->parse;
3741 : double dNumGroups;
3742 :
3743 54368 : if (parse->groupClause)
3744 : {
3745 : List *groupExprs;
3746 :
3747 10406 : if (parse->groupingSets)
3748 : {
3749 : /* Add up the estimates for each grouping set */
3750 : ListCell *lc;
3751 :
3752 : Assert(gd); /* keep Coverity happy */
3753 :
3754 936 : dNumGroups = 0;
3755 :
3756 2498 : foreach(lc, gd->rollups)
3757 : {
3758 1562 : RollupData *rollup = lfirst_node(RollupData, lc);
3759 : ListCell *lc2;
3760 : ListCell *lc3;
3761 :
3762 1562 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3763 : target_list);
3764 :
3765 1562 : rollup->numGroups = 0.0;
3766 :
3767 4442 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3768 : {
3769 2880 : List *gset = (List *) lfirst(lc2);
3770 2880 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3771 2880 : double numGroups = estimate_num_groups(root,
3772 : groupExprs,
3773 : path_rows,
3774 : &gset,
3775 : NULL);
3776 :
3777 2880 : gs->numGroups = numGroups;
3778 2880 : rollup->numGroups += numGroups;
3779 : }
3780 :
3781 1562 : dNumGroups += rollup->numGroups;
3782 : }
3783 :
3784 936 : if (gd->hash_sets_idx)
3785 : {
3786 : ListCell *lc2;
3787 :
3788 36 : gd->dNumHashGroups = 0;
3789 :
3790 36 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3791 : target_list);
3792 :
3793 78 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3794 : {
3795 42 : List *gset = (List *) lfirst(lc);
3796 42 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3797 42 : double numGroups = estimate_num_groups(root,
3798 : groupExprs,
3799 : path_rows,
3800 : &gset,
3801 : NULL);
3802 :
3803 42 : gs->numGroups = numGroups;
3804 42 : gd->dNumHashGroups += numGroups;
3805 : }
3806 :
3807 36 : dNumGroups += gd->dNumHashGroups;
3808 : }
3809 : }
3810 : else
3811 : {
3812 : /* Plain GROUP BY -- estimate based on optimized groupClause */
3813 9470 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3814 : target_list);
3815 :
3816 9470 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3817 : NULL, NULL);
3818 : }
3819 : }
3820 43962 : else if (parse->groupingSets)
3821 : {
3822 : /* Empty grouping sets ... one result row for each one */
3823 60 : dNumGroups = list_length(parse->groupingSets);
3824 : }
3825 43902 : else if (parse->hasAggs || root->hasHavingQual)
3826 : {
3827 : /* Plain aggregation, one result row */
3828 43902 : dNumGroups = 1;
3829 : }
3830 : else
3831 : {
3832 : /* Not grouping */
3833 0 : dNumGroups = 1;
3834 : }
3835 :
3836 54368 : return dNumGroups;
3837 : }
3838 :
3839 : /*
3840 : * create_grouping_paths
3841 : *
3842 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3843 : * Along the way, we also build an upperrel for Paths which are partially
3844 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3845 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3846 : * the only partially grouped paths we build are also partial paths; that
3847 : * is, they need a Gather and then a FinalizeAggregate.
3848 : *
3849 : * input_rel: contains the source-data Paths
3850 : * target: the pathtarget for the result Paths to compute
3851 : * gd: grouping sets data including list of grouping sets and their clauses
3852 : *
3853 : * Note: all Paths in input_rel are expected to return the target computed
3854 : * by make_group_input_target.
3855 : */
3856 : static RelOptInfo *
3857 45968 : create_grouping_paths(PlannerInfo *root,
3858 : RelOptInfo *input_rel,
3859 : PathTarget *target,
3860 : bool target_parallel_safe,
3861 : grouping_sets_data *gd)
3862 : {
3863 45968 : Query *parse = root->parse;
3864 : RelOptInfo *grouped_rel;
3865 : RelOptInfo *partially_grouped_rel;
3866 : AggClauseCosts agg_costs;
3867 :
3868 275808 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3869 45968 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3870 :
3871 : /*
3872 : * Create grouping relation to hold fully aggregated grouping and/or
3873 : * aggregation paths.
3874 : */
3875 45968 : grouped_rel = make_grouping_rel(root, input_rel, target,
3876 : target_parallel_safe, parse->havingQual);
3877 :
3878 : /*
3879 : * Create either paths for a degenerate grouping or paths for ordinary
3880 : * grouping, as appropriate.
3881 : */
3882 45968 : if (is_degenerate_grouping(root))
3883 42 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3884 : else
3885 : {
3886 45926 : int flags = 0;
3887 : GroupPathExtraData extra;
3888 :
3889 : /*
3890 : * Determine whether it's possible to perform sort-based
3891 : * implementations of grouping. (Note that if processed_groupClause
3892 : * is empty, grouping_is_sortable() is trivially true, and all the
3893 : * pathkeys_contained_in() tests will succeed too, so that we'll
3894 : * consider every surviving input path.)
3895 : *
3896 : * If we have grouping sets, we might be able to sort some but not all
3897 : * of them; in this case, we need can_sort to be true as long as we
3898 : * must consider any sorted-input plan.
3899 : */
3900 45926 : if ((gd && gd->rollups != NIL)
3901 44936 : || grouping_is_sortable(root->processed_groupClause))
3902 45920 : flags |= GROUPING_CAN_USE_SORT;
3903 :
3904 : /*
3905 : * Determine whether we should consider hash-based implementations of
3906 : * grouping.
3907 : *
3908 : * Hashed aggregation only applies if we're grouping. If we have
3909 : * grouping sets, some groups might be hashable but others not; in
3910 : * this case we set can_hash true as long as there is nothing globally
3911 : * preventing us from hashing (and we should therefore consider plans
3912 : * with hashes).
3913 : *
3914 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3915 : * BY aggregates. (Doing so would imply storing *all* the input
3916 : * values in the hash table, and/or running many sorts in parallel,
3917 : * either of which seems like a certain loser.) We similarly don't
3918 : * support ordered-set aggregates in hashed aggregation, but that case
3919 : * is also included in the numOrderedAggs count.
3920 : *
3921 : * Note: grouping_is_hashable() is much more expensive to check than
3922 : * the other gating conditions, so we want to do it last.
3923 : */
3924 45926 : if ((parse->groupClause != NIL &&
3925 9856 : root->numOrderedAggs == 0 &&
3926 4788 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3927 4784 : flags |= GROUPING_CAN_USE_HASH;
3928 :
3929 : /*
3930 : * Determine whether partial aggregation is possible.
3931 : */
3932 45926 : if (can_partial_agg(root))
3933 41034 : flags |= GROUPING_CAN_PARTIAL_AGG;
3934 :
3935 45926 : extra.flags = flags;
3936 45926 : extra.target_parallel_safe = target_parallel_safe;
3937 45926 : extra.havingQual = parse->havingQual;
3938 45926 : extra.targetList = parse->targetList;
3939 45926 : extra.partial_costs_set = false;
3940 :
3941 : /*
3942 : * Determine whether partitionwise aggregation is in theory possible.
3943 : * It can be disabled by the user, and for now, we don't try to
3944 : * support grouping sets. create_ordinary_grouping_paths() will check
3945 : * additional conditions, such as whether input_rel is partitioned.
3946 : */
3947 45926 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3948 700 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3949 : else
3950 45226 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3951 :
3952 45926 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3953 : &agg_costs, gd, &extra,
3954 : &partially_grouped_rel);
3955 : }
3956 :
3957 45962 : set_cheapest(grouped_rel);
3958 45962 : return grouped_rel;
3959 : }
3960 :
3961 : /*
3962 : * make_grouping_rel
3963 : *
3964 : * Create a new grouping rel and set basic properties.
3965 : *
3966 : * input_rel represents the underlying scan/join relation.
3967 : * target is the output expected from the grouping relation.
3968 : */
3969 : static RelOptInfo *
3970 48134 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3971 : PathTarget *target, bool target_parallel_safe,
3972 : Node *havingQual)
3973 : {
3974 : RelOptInfo *grouped_rel;
3975 :
3976 48134 : if (IS_OTHER_REL(input_rel))
3977 : {
3978 2166 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3979 : input_rel->relids);
3980 2166 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3981 : }
3982 : else
3983 : {
3984 : /*
3985 : * By tradition, the relids set for the main grouping relation is
3986 : * NULL. (This could be changed, but might require adjustments
3987 : * elsewhere.)
3988 : */
3989 45968 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3990 : }
3991 :
3992 : /* Set target. */
3993 48134 : grouped_rel->reltarget = target;
3994 :
3995 : /*
3996 : * If the input relation is not parallel-safe, then the grouped relation
3997 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3998 : * target list and HAVING quals are parallel-safe.
3999 : */
4000 77644 : if (input_rel->consider_parallel && target_parallel_safe &&
4001 29510 : is_parallel_safe(root, havingQual))
4002 29486 : grouped_rel->consider_parallel = true;
4003 :
4004 : /* Assume that the same path generation strategies are allowed */
4005 48134 : grouped_rel->pgs_mask = input_rel->pgs_mask;
4006 :
4007 : /*
4008 : * If the input rel belongs to a single FDW, so does the grouped rel.
4009 : */
4010 48134 : grouped_rel->serverid = input_rel->serverid;
4011 48134 : grouped_rel->userid = input_rel->userid;
4012 48134 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
4013 48134 : grouped_rel->fdwroutine = input_rel->fdwroutine;
4014 :
4015 48134 : return grouped_rel;
4016 : }
4017 :
4018 : /*
4019 : * is_degenerate_grouping
4020 : *
4021 : * A degenerate grouping is one in which the query has a HAVING qual and/or
4022 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
4023 : * grouping sets are all empty).
4024 : */
4025 : static bool
4026 45968 : is_degenerate_grouping(PlannerInfo *root)
4027 : {
4028 45968 : Query *parse = root->parse;
4029 :
4030 44718 : return (root->hasHavingQual || parse->groupingSets) &&
4031 90686 : !parse->hasAggs && parse->groupClause == NIL;
4032 : }
4033 :
4034 : /*
4035 : * create_degenerate_grouping_paths
4036 : *
4037 : * When the grouping is degenerate (see is_degenerate_grouping), we are
4038 : * supposed to emit either zero or one row for each grouping set depending on
4039 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
4040 : * either HAVING or the targetlist, so we actually do not need the FROM table
4041 : * at all! We can just throw away the plan-so-far and generate a Result node.
4042 : * This is a sufficiently unusual corner case that it's not worth contorting
4043 : * the structure of this module to avoid having to generate the earlier paths
4044 : * in the first place.
4045 : */
4046 : static void
4047 42 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4048 : RelOptInfo *grouped_rel)
4049 : {
4050 42 : Query *parse = root->parse;
4051 : int nrows;
4052 : Path *path;
4053 :
4054 42 : nrows = list_length(parse->groupingSets);
4055 42 : if (nrows > 1)
4056 : {
4057 : /*
4058 : * Doesn't seem worthwhile writing code to cons up a generate_series
4059 : * or a values scan to emit multiple rows. Instead just make N clones
4060 : * and append them. (With a volatile HAVING clause, this means you
4061 : * might get between 0 and N output rows. Offhand I think that's
4062 : * desired.)
4063 : */
4064 12 : List *paths = NIL;
4065 :
4066 36 : while (--nrows >= 0)
4067 : {
4068 : path = (Path *)
4069 24 : create_group_result_path(root, grouped_rel,
4070 24 : grouped_rel->reltarget,
4071 24 : (List *) parse->havingQual);
4072 24 : paths = lappend(paths, path);
4073 : }
4074 : path = (Path *)
4075 12 : create_append_path(root,
4076 : grouped_rel,
4077 : paths,
4078 : NIL,
4079 : NIL,
4080 : NULL,
4081 : 0,
4082 : false,
4083 : -1);
4084 : }
4085 : else
4086 : {
4087 : /* No grouping sets, or just one, so one output row */
4088 : path = (Path *)
4089 30 : create_group_result_path(root, grouped_rel,
4090 30 : grouped_rel->reltarget,
4091 30 : (List *) parse->havingQual);
4092 : }
4093 :
4094 42 : add_path(grouped_rel, path);
4095 42 : }
4096 :
4097 : /*
4098 : * create_ordinary_grouping_paths
4099 : *
4100 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4101 : *
4102 : * We need to consider sorted and hashed aggregation in the same function,
4103 : * because otherwise (1) it would be harder to throw an appropriate error
4104 : * message if neither way works, and (2) we should not allow hashtable size
4105 : * considerations to dissuade us from using hashing if sorting is not possible.
4106 : *
4107 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4108 : * function creates, or to NULL if it doesn't create one.
4109 : */
4110 : static void
4111 48092 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4112 : RelOptInfo *grouped_rel,
4113 : const AggClauseCosts *agg_costs,
4114 : grouping_sets_data *gd,
4115 : GroupPathExtraData *extra,
4116 : RelOptInfo **partially_grouped_rel_p)
4117 : {
4118 48092 : RelOptInfo *partially_grouped_rel = NULL;
4119 48092 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4120 :
4121 : /*
4122 : * If this is the topmost grouping relation or if the parent relation is
4123 : * doing some form of partitionwise aggregation, then we may be able to do
4124 : * it at this level also. However, if the input relation is not
4125 : * partitioned, partitionwise aggregate is impossible.
4126 : */
4127 48092 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4128 2866 : IS_PARTITIONED_REL(input_rel))
4129 : {
4130 : /*
4131 : * If this is the topmost relation or if the parent relation is doing
4132 : * full partitionwise aggregation, then we can do full partitionwise
4133 : * aggregation provided that the GROUP BY clause contains all of the
4134 : * partitioning columns at this level and the collation used by GROUP
4135 : * BY matches the partitioning collation. Otherwise, we can do at
4136 : * most partial partitionwise aggregation. But if partial aggregation
4137 : * is not supported in general then we can't use it for partitionwise
4138 : * aggregation either.
4139 : *
4140 : * Check parse->groupClause not processed_groupClause, because it's
4141 : * okay if some of the partitioning columns were proved redundant.
4142 : */
4143 1640 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4144 772 : group_by_has_partkey(input_rel, extra->targetList,
4145 772 : root->parse->groupClause))
4146 488 : patype = PARTITIONWISE_AGGREGATE_FULL;
4147 380 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4148 338 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4149 : else
4150 42 : patype = PARTITIONWISE_AGGREGATE_NONE;
4151 : }
4152 :
4153 : /*
4154 : * Before generating paths for grouped_rel, we first generate any possible
4155 : * partially grouped paths; that way, later code can easily consider both
4156 : * parallel and non-parallel approaches to grouping.
4157 : */
4158 48092 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4159 : {
4160 : bool force_rel_creation;
4161 :
4162 : /*
4163 : * If we're doing partitionwise aggregation at this level, force
4164 : * creation of a partially_grouped_rel so we can add partitionwise
4165 : * paths to it.
4166 : */
4167 43128 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4168 :
4169 : partially_grouped_rel =
4170 43128 : create_partial_grouping_paths(root,
4171 : grouped_rel,
4172 : input_rel,
4173 : gd,
4174 : extra,
4175 : force_rel_creation);
4176 : }
4177 :
4178 : /* Set out parameter. */
4179 48092 : *partially_grouped_rel_p = partially_grouped_rel;
4180 :
4181 : /* Apply partitionwise aggregation technique, if possible. */
4182 48092 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4183 826 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4184 : partially_grouped_rel, agg_costs,
4185 : gd, patype, extra);
4186 :
4187 : /* If we are doing partial aggregation only, return. */
4188 48092 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4189 : {
4190 : Assert(partially_grouped_rel);
4191 :
4192 858 : if (partially_grouped_rel->pathlist)
4193 858 : set_cheapest(partially_grouped_rel);
4194 :
4195 858 : return;
4196 : }
4197 :
4198 : /* Gather any partially grouped partial paths. */
4199 47234 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4200 2772 : gather_grouping_paths(root, partially_grouped_rel);
4201 :
4202 : /* Now choose the best path(s) for partially_grouped_rel. */
4203 47234 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4204 3000 : set_cheapest(partially_grouped_rel);
4205 :
4206 : /* Build final grouping paths */
4207 47234 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4208 : partially_grouped_rel, agg_costs, gd,
4209 : extra);
4210 :
4211 : /* Give a helpful error if we failed to find any implementation */
4212 47234 : if (grouped_rel->pathlist == NIL)
4213 6 : ereport(ERROR,
4214 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4215 : errmsg("could not implement GROUP BY"),
4216 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4217 :
4218 : /*
4219 : * If there is an FDW that's responsible for all baserels of the query,
4220 : * let it consider adding ForeignPaths.
4221 : */
4222 47228 : if (grouped_rel->fdwroutine &&
4223 338 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4224 336 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4225 : input_rel, grouped_rel,
4226 : extra);
4227 :
4228 : /* Let extensions possibly add some more paths */
4229 47228 : if (create_upper_paths_hook)
4230 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4231 : input_rel, grouped_rel,
4232 : extra);
4233 : }
4234 :
4235 : /*
4236 : * For a given input path, consider the possible ways of doing grouping sets on
4237 : * it, by combinations of hashing and sorting. This can be called multiple
4238 : * times, so it's important that it not scribble on input. No result is
4239 : * returned, but any generated paths are added to grouped_rel.
4240 : */
4241 : static void
4242 1962 : consider_groupingsets_paths(PlannerInfo *root,
4243 : RelOptInfo *grouped_rel,
4244 : Path *path,
4245 : bool is_sorted,
4246 : bool can_hash,
4247 : grouping_sets_data *gd,
4248 : const AggClauseCosts *agg_costs,
4249 : double dNumGroups)
4250 : {
4251 1962 : Query *parse = root->parse;
4252 1962 : Size hash_mem_limit = get_hash_memory_limit();
4253 :
4254 : /*
4255 : * If we're not being offered sorted input, then only consider plans that
4256 : * can be done entirely by hashing.
4257 : *
4258 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4259 : * the input is actually sorted despite not being advertised as such, we
4260 : * prefer to make use of that in order to use less memory.
4261 : *
4262 : * If none of the grouping sets are sortable, then ignore the hash_mem
4263 : * limit and generate a path anyway, since otherwise we'll just fail.
4264 : */
4265 1962 : if (!is_sorted)
4266 : {
4267 900 : List *new_rollups = NIL;
4268 900 : RollupData *unhashed_rollup = NULL;
4269 : List *sets_data;
4270 900 : List *empty_sets_data = NIL;
4271 900 : List *empty_sets = NIL;
4272 : ListCell *lc;
4273 900 : ListCell *l_start = list_head(gd->rollups);
4274 900 : AggStrategy strat = AGG_HASHED;
4275 : double hashsize;
4276 900 : double exclude_groups = 0.0;
4277 :
4278 : Assert(can_hash);
4279 :
4280 : /*
4281 : * If the input is coincidentally sorted usefully (which can happen
4282 : * even if is_sorted is false, since that only means that our caller
4283 : * has set up the sorting for us), then save some hashtable space by
4284 : * making use of that. But we need to watch out for degenerate cases:
4285 : *
4286 : * 1) If there are any empty grouping sets, then group_pathkeys might
4287 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4288 : * there will be a rollup containing only empty groups, and the
4289 : * pathkeys_contained_in test is vacuously true; this is ok.
4290 : *
4291 : * XXX: the above relies on the fact that group_pathkeys is generated
4292 : * from the first rollup. If we add the ability to consider multiple
4293 : * sort orders for grouping input, this assumption might fail.
4294 : *
4295 : * 2) If there are no empty sets and only unsortable sets, then the
4296 : * rollups list will be empty (and thus l_start == NULL), and
4297 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4298 : * pathkeys_contained_in test doesn't cause us to crash.
4299 : */
4300 1794 : if (l_start != NULL &&
4301 894 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4302 : {
4303 36 : unhashed_rollup = lfirst_node(RollupData, l_start);
4304 36 : exclude_groups = unhashed_rollup->numGroups;
4305 36 : l_start = lnext(gd->rollups, l_start);
4306 : }
4307 :
4308 900 : hashsize = estimate_hashagg_tablesize(root,
4309 : path,
4310 : agg_costs,
4311 : dNumGroups - exclude_groups);
4312 :
4313 : /*
4314 : * gd->rollups is empty if we have only unsortable columns to work
4315 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4316 : * sorted-input case to generate usable mixed paths.
4317 : */
4318 900 : if (hashsize > hash_mem_limit && gd->rollups)
4319 18 : return; /* nope, won't fit */
4320 :
4321 : /*
4322 : * We need to burst the existing rollups list into individual grouping
4323 : * sets and recompute a groupClause for each set.
4324 : */
4325 882 : sets_data = list_copy(gd->unsortable_sets);
4326 :
4327 2228 : for_each_cell(lc, gd->rollups, l_start)
4328 : {
4329 1370 : RollupData *rollup = lfirst_node(RollupData, lc);
4330 :
4331 : /*
4332 : * If we find an unhashable rollup that's not been skipped by the
4333 : * "actually sorted" check above, we can't cope; we'd need sorted
4334 : * input (with a different sort order) but we can't get that here.
4335 : * So bail out; we'll get a valid path from the is_sorted case
4336 : * instead.
4337 : *
4338 : * The mere presence of empty grouping sets doesn't make a rollup
4339 : * unhashable (see preprocess_grouping_sets), we handle those
4340 : * specially below.
4341 : */
4342 1370 : if (!rollup->hashable)
4343 24 : return;
4344 :
4345 1346 : sets_data = list_concat(sets_data, rollup->gsets_data);
4346 : }
4347 3480 : foreach(lc, sets_data)
4348 : {
4349 2622 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4350 2622 : List *gset = gs->set;
4351 : RollupData *rollup;
4352 :
4353 2622 : if (gset == NIL)
4354 : {
4355 : /* Empty grouping sets can't be hashed. */
4356 530 : empty_sets_data = lappend(empty_sets_data, gs);
4357 530 : empty_sets = lappend(empty_sets, NIL);
4358 : }
4359 : else
4360 : {
4361 2092 : rollup = makeNode(RollupData);
4362 :
4363 2092 : rollup->groupClause = preprocess_groupclause(root, gset);
4364 2092 : rollup->gsets_data = list_make1(gs);
4365 2092 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4366 : rollup->gsets_data,
4367 : gd->tleref_to_colnum_map);
4368 2092 : rollup->numGroups = gs->numGroups;
4369 2092 : rollup->hashable = true;
4370 2092 : rollup->is_hashed = true;
4371 2092 : new_rollups = lappend(new_rollups, rollup);
4372 : }
4373 : }
4374 :
4375 : /*
4376 : * If we didn't find anything nonempty to hash, then bail. We'll
4377 : * generate a path from the is_sorted case.
4378 : */
4379 858 : if (new_rollups == NIL)
4380 0 : return;
4381 :
4382 : /*
4383 : * If there were empty grouping sets they should have been in the
4384 : * first rollup.
4385 : */
4386 : Assert(!unhashed_rollup || !empty_sets);
4387 :
4388 858 : if (unhashed_rollup)
4389 : {
4390 36 : new_rollups = lappend(new_rollups, unhashed_rollup);
4391 36 : strat = AGG_MIXED;
4392 : }
4393 822 : else if (empty_sets)
4394 : {
4395 482 : RollupData *rollup = makeNode(RollupData);
4396 :
4397 482 : rollup->groupClause = NIL;
4398 482 : rollup->gsets_data = empty_sets_data;
4399 482 : rollup->gsets = empty_sets;
4400 482 : rollup->numGroups = list_length(empty_sets);
4401 482 : rollup->hashable = false;
4402 482 : rollup->is_hashed = false;
4403 482 : new_rollups = lappend(new_rollups, rollup);
4404 482 : strat = AGG_MIXED;
4405 : }
4406 :
4407 858 : add_path(grouped_rel, (Path *)
4408 858 : create_groupingsets_path(root,
4409 : grouped_rel,
4410 : path,
4411 858 : (List *) parse->havingQual,
4412 : strat,
4413 : new_rollups,
4414 : agg_costs));
4415 858 : return;
4416 : }
4417 :
4418 : /*
4419 : * If we have sorted input but nothing we can do with it, bail.
4420 : */
4421 1062 : if (gd->rollups == NIL)
4422 0 : return;
4423 :
4424 : /*
4425 : * Given sorted input, we try and make two paths: one sorted and one mixed
4426 : * sort/hash. (We need to try both because hashagg might be disabled, or
4427 : * some columns might not be sortable.)
4428 : *
4429 : * can_hash is passed in as false if some obstacle elsewhere (such as
4430 : * ordered aggs) means that we shouldn't consider hashing at all.
4431 : */
4432 1062 : if (can_hash && gd->any_hashable)
4433 : {
4434 966 : List *rollups = NIL;
4435 966 : List *hash_sets = list_copy(gd->unsortable_sets);
4436 966 : double availspace = hash_mem_limit;
4437 : ListCell *lc;
4438 :
4439 : /*
4440 : * Account first for space needed for groups we can't sort at all.
4441 : */
4442 966 : availspace -= estimate_hashagg_tablesize(root,
4443 : path,
4444 : agg_costs,
4445 : gd->dNumHashGroups);
4446 :
4447 966 : if (availspace > 0 && list_length(gd->rollups) > 1)
4448 : {
4449 : double scale;
4450 492 : int num_rollups = list_length(gd->rollups);
4451 : int k_capacity;
4452 492 : int *k_weights = palloc(num_rollups * sizeof(int));
4453 492 : Bitmapset *hash_items = NULL;
4454 : int i;
4455 :
4456 : /*
4457 : * We treat this as a knapsack problem: the knapsack capacity
4458 : * represents hash_mem, the item weights are the estimated memory
4459 : * usage of the hashtables needed to implement a single rollup,
4460 : * and we really ought to use the cost saving as the item value;
4461 : * however, currently the costs assigned to sort nodes don't
4462 : * reflect the comparison costs well, and so we treat all items as
4463 : * of equal value (each rollup we hash instead saves us one sort).
4464 : *
4465 : * To use the discrete knapsack, we need to scale the values to a
4466 : * reasonably small bounded range. We choose to allow a 5% error
4467 : * margin; we have no more than 4096 rollups in the worst possible
4468 : * case, which with a 5% error margin will require a bit over 42MB
4469 : * of workspace. (Anyone wanting to plan queries that complex had
4470 : * better have the memory for it. In more reasonable cases, with
4471 : * no more than a couple of dozen rollups, the memory usage will
4472 : * be negligible.)
4473 : *
4474 : * k_capacity is naturally bounded, but we clamp the values for
4475 : * scale and weight (below) to avoid overflows or underflows (or
4476 : * uselessly trying to use a scale factor less than 1 byte).
4477 : */
4478 492 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4479 492 : k_capacity = (int) floor(availspace / scale);
4480 :
4481 : /*
4482 : * We leave the first rollup out of consideration since it's the
4483 : * one that matches the input sort order. We assign indexes "i"
4484 : * to only those entries considered for hashing; the second loop,
4485 : * below, must use the same condition.
4486 : */
4487 492 : i = 0;
4488 1248 : for_each_from(lc, gd->rollups, 1)
4489 : {
4490 756 : RollupData *rollup = lfirst_node(RollupData, lc);
4491 :
4492 756 : if (rollup->hashable)
4493 : {
4494 756 : double sz = estimate_hashagg_tablesize(root,
4495 : path,
4496 : agg_costs,
4497 : rollup->numGroups);
4498 :
4499 : /*
4500 : * If sz is enormous, but hash_mem (and hence scale) is
4501 : * small, avoid integer overflow here.
4502 : */
4503 756 : k_weights[i] = (int) Min(floor(sz / scale),
4504 : k_capacity + 1.0);
4505 756 : ++i;
4506 : }
4507 : }
4508 :
4509 : /*
4510 : * Apply knapsack algorithm; compute the set of items which
4511 : * maximizes the value stored (in this case the number of sorts
4512 : * saved) while keeping the total size (approximately) within
4513 : * capacity.
4514 : */
4515 492 : if (i > 0)
4516 492 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4517 :
4518 492 : if (!bms_is_empty(hash_items))
4519 : {
4520 492 : rollups = list_make1(linitial(gd->rollups));
4521 :
4522 492 : i = 0;
4523 1248 : for_each_from(lc, gd->rollups, 1)
4524 : {
4525 756 : RollupData *rollup = lfirst_node(RollupData, lc);
4526 :
4527 756 : if (rollup->hashable)
4528 : {
4529 756 : if (bms_is_member(i, hash_items))
4530 720 : hash_sets = list_concat(hash_sets,
4531 720 : rollup->gsets_data);
4532 : else
4533 36 : rollups = lappend(rollups, rollup);
4534 756 : ++i;
4535 : }
4536 : else
4537 0 : rollups = lappend(rollups, rollup);
4538 : }
4539 : }
4540 : }
4541 :
4542 966 : if (!rollups && hash_sets)
4543 24 : rollups = list_copy(gd->rollups);
4544 :
4545 1826 : foreach(lc, hash_sets)
4546 : {
4547 860 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4548 860 : RollupData *rollup = makeNode(RollupData);
4549 :
4550 : Assert(gs->set != NIL);
4551 :
4552 860 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4553 860 : rollup->gsets_data = list_make1(gs);
4554 860 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4555 : rollup->gsets_data,
4556 : gd->tleref_to_colnum_map);
4557 860 : rollup->numGroups = gs->numGroups;
4558 860 : rollup->hashable = true;
4559 860 : rollup->is_hashed = true;
4560 860 : rollups = lcons(rollup, rollups);
4561 : }
4562 :
4563 966 : if (rollups)
4564 : {
4565 516 : add_path(grouped_rel, (Path *)
4566 516 : create_groupingsets_path(root,
4567 : grouped_rel,
4568 : path,
4569 516 : (List *) parse->havingQual,
4570 : AGG_MIXED,
4571 : rollups,
4572 : agg_costs));
4573 : }
4574 : }
4575 :
4576 : /*
4577 : * Now try the simple sorted case.
4578 : */
4579 1062 : if (!gd->unsortable_sets)
4580 1032 : add_path(grouped_rel, (Path *)
4581 1032 : create_groupingsets_path(root,
4582 : grouped_rel,
4583 : path,
4584 1032 : (List *) parse->havingQual,
4585 : AGG_SORTED,
4586 : gd->rollups,
4587 : agg_costs));
4588 : }
4589 :
4590 : /*
4591 : * create_window_paths
4592 : *
4593 : * Build a new upperrel containing Paths for window-function evaluation.
4594 : *
4595 : * input_rel: contains the source-data Paths
4596 : * input_target: result of make_window_input_target
4597 : * output_target: what the topmost WindowAggPath should return
4598 : * wflists: result of find_window_functions
4599 : * activeWindows: result of select_active_windows
4600 : *
4601 : * Note: all Paths in input_rel are expected to return input_target.
4602 : */
4603 : static RelOptInfo *
4604 2676 : create_window_paths(PlannerInfo *root,
4605 : RelOptInfo *input_rel,
4606 : PathTarget *input_target,
4607 : PathTarget *output_target,
4608 : bool output_target_parallel_safe,
4609 : WindowFuncLists *wflists,
4610 : List *activeWindows)
4611 : {
4612 : RelOptInfo *window_rel;
4613 : ListCell *lc;
4614 :
4615 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4616 2676 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4617 :
4618 : /*
4619 : * If the input relation is not parallel-safe, then the window relation
4620 : * can't be parallel-safe, either. Otherwise, we need to examine the
4621 : * target list and active windows for non-parallel-safe constructs.
4622 : */
4623 2676 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4624 0 : is_parallel_safe(root, (Node *) activeWindows))
4625 0 : window_rel->consider_parallel = true;
4626 :
4627 : /*
4628 : * If the input rel belongs to a single FDW, so does the window rel.
4629 : */
4630 2676 : window_rel->serverid = input_rel->serverid;
4631 2676 : window_rel->userid = input_rel->userid;
4632 2676 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4633 2676 : window_rel->fdwroutine = input_rel->fdwroutine;
4634 :
4635 : /*
4636 : * Consider computing window functions starting from the existing
4637 : * cheapest-total path (which will likely require a sort) as well as any
4638 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4639 : */
4640 5692 : foreach(lc, input_rel->pathlist)
4641 : {
4642 3016 : Path *path = (Path *) lfirst(lc);
4643 : int presorted_keys;
4644 :
4645 3356 : if (path == input_rel->cheapest_total_path ||
4646 340 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4647 152 : &presorted_keys) ||
4648 152 : presorted_keys > 0)
4649 2890 : create_one_window_path(root,
4650 : window_rel,
4651 : path,
4652 : input_target,
4653 : output_target,
4654 : wflists,
4655 : activeWindows);
4656 : }
4657 :
4658 : /*
4659 : * If there is an FDW that's responsible for all baserels of the query,
4660 : * let it consider adding ForeignPaths.
4661 : */
4662 2676 : if (window_rel->fdwroutine &&
4663 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4664 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4665 : input_rel, window_rel,
4666 : NULL);
4667 :
4668 : /* Let extensions possibly add some more paths */
4669 2676 : if (create_upper_paths_hook)
4670 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4671 : input_rel, window_rel, NULL);
4672 :
4673 : /* Now choose the best path(s) */
4674 2676 : set_cheapest(window_rel);
4675 :
4676 2676 : return window_rel;
4677 : }
4678 :
4679 : /*
4680 : * Stack window-function implementation steps atop the given Path, and
4681 : * add the result to window_rel.
4682 : *
4683 : * window_rel: upperrel to contain result
4684 : * path: input Path to use (must return input_target)
4685 : * input_target: result of make_window_input_target
4686 : * output_target: what the topmost WindowAggPath should return
4687 : * wflists: result of find_window_functions
4688 : * activeWindows: result of select_active_windows
4689 : */
4690 : static void
4691 2890 : create_one_window_path(PlannerInfo *root,
4692 : RelOptInfo *window_rel,
4693 : Path *path,
4694 : PathTarget *input_target,
4695 : PathTarget *output_target,
4696 : WindowFuncLists *wflists,
4697 : List *activeWindows)
4698 : {
4699 : PathTarget *window_target;
4700 : ListCell *l;
4701 2890 : List *topqual = NIL;
4702 :
4703 : /*
4704 : * Since each window clause could require a different sort order, we stack
4705 : * up a WindowAgg node for each clause, with sort steps between them as
4706 : * needed. (We assume that select_active_windows chose a good order for
4707 : * executing the clauses in.)
4708 : *
4709 : * input_target should contain all Vars and Aggs needed for the result.
4710 : * (In some cases we wouldn't need to propagate all of these all the way
4711 : * to the top, since they might only be needed as inputs to WindowFuncs.
4712 : * It's probably not worth trying to optimize that though.) It must also
4713 : * contain all window partitioning and sorting expressions, to ensure
4714 : * they're computed only once at the bottom of the stack (that's critical
4715 : * for volatile functions). As we climb up the stack, we'll add outputs
4716 : * for the WindowFuncs computed at each level.
4717 : */
4718 2890 : window_target = input_target;
4719 :
4720 5966 : foreach(l, activeWindows)
4721 : {
4722 3076 : WindowClause *wc = lfirst_node(WindowClause, l);
4723 : List *window_pathkeys;
4724 3076 : List *runcondition = NIL;
4725 : int presorted_keys;
4726 : bool is_sorted;
4727 : bool topwindow;
4728 : ListCell *lc2;
4729 :
4730 3076 : window_pathkeys = make_pathkeys_for_window(root,
4731 : wc,
4732 : root->processed_tlist);
4733 :
4734 3076 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4735 : path->pathkeys,
4736 : &presorted_keys);
4737 :
4738 : /* Sort if necessary */
4739 3076 : if (!is_sorted)
4740 : {
4741 : /*
4742 : * No presorted keys or incremental sort disabled, just perform a
4743 : * complete sort.
4744 : */
4745 2184 : if (presorted_keys == 0 || !enable_incremental_sort)
4746 2122 : path = (Path *) create_sort_path(root, window_rel,
4747 : path,
4748 : window_pathkeys,
4749 : -1.0);
4750 : else
4751 : {
4752 : /*
4753 : * Since we have presorted keys and incremental sort is
4754 : * enabled, just use incremental sort.
4755 : */
4756 62 : path = (Path *) create_incremental_sort_path(root,
4757 : window_rel,
4758 : path,
4759 : window_pathkeys,
4760 : presorted_keys,
4761 : -1.0);
4762 : }
4763 : }
4764 :
4765 3076 : if (lnext(activeWindows, l))
4766 : {
4767 : /*
4768 : * Add the current WindowFuncs to the output target for this
4769 : * intermediate WindowAggPath. We must copy window_target to
4770 : * avoid changing the previous path's target.
4771 : *
4772 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4773 : * we do need to account for the increase in tlist width.
4774 : */
4775 186 : int64 tuple_width = window_target->width;
4776 :
4777 186 : window_target = copy_pathtarget(window_target);
4778 444 : foreach(lc2, wflists->windowFuncs[wc->winref])
4779 : {
4780 258 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4781 :
4782 258 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4783 258 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4784 : }
4785 186 : window_target->width = clamp_width_est(tuple_width);
4786 : }
4787 : else
4788 : {
4789 : /* Install the goal target in the topmost WindowAgg */
4790 2890 : window_target = output_target;
4791 : }
4792 :
4793 : /* mark the final item in the list as the top-level window */
4794 3076 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4795 :
4796 : /*
4797 : * Collect the WindowFuncRunConditions from each WindowFunc and
4798 : * convert them into OpExprs
4799 : */
4800 7034 : foreach(lc2, wflists->windowFuncs[wc->winref])
4801 : {
4802 : ListCell *lc3;
4803 3958 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4804 :
4805 4138 : foreach(lc3, wfunc->runCondition)
4806 : {
4807 180 : WindowFuncRunCondition *wfuncrc =
4808 : lfirst_node(WindowFuncRunCondition, lc3);
4809 : Expr *opexpr;
4810 : Expr *leftop;
4811 : Expr *rightop;
4812 :
4813 180 : if (wfuncrc->wfunc_left)
4814 : {
4815 162 : leftop = (Expr *) copyObject(wfunc);
4816 162 : rightop = copyObject(wfuncrc->arg);
4817 : }
4818 : else
4819 : {
4820 18 : leftop = copyObject(wfuncrc->arg);
4821 18 : rightop = (Expr *) copyObject(wfunc);
4822 : }
4823 :
4824 180 : opexpr = make_opclause(wfuncrc->opno,
4825 : BOOLOID,
4826 : false,
4827 : leftop,
4828 : rightop,
4829 : InvalidOid,
4830 : wfuncrc->inputcollid);
4831 :
4832 180 : runcondition = lappend(runcondition, opexpr);
4833 :
4834 180 : if (!topwindow)
4835 24 : topqual = lappend(topqual, opexpr);
4836 : }
4837 : }
4838 :
4839 : path = (Path *)
4840 3076 : create_windowagg_path(root, window_rel, path, window_target,
4841 3076 : wflists->windowFuncs[wc->winref],
4842 : runcondition, wc,
4843 : topwindow ? topqual : NIL, topwindow);
4844 : }
4845 :
4846 2890 : add_path(window_rel, path);
4847 2890 : }
4848 :
4849 : /*
4850 : * create_distinct_paths
4851 : *
4852 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4853 : *
4854 : * input_rel: contains the source-data Paths
4855 : * target: the pathtarget for the result Paths to compute
4856 : *
4857 : * Note: input paths should already compute the desired pathtarget, since
4858 : * Sort/Unique won't project anything.
4859 : */
4860 : static RelOptInfo *
4861 3024 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4862 : PathTarget *target)
4863 : {
4864 : RelOptInfo *distinct_rel;
4865 :
4866 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4867 3024 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4868 :
4869 : /*
4870 : * We don't compute anything at this level, so distinct_rel will be
4871 : * parallel-safe if the input rel is parallel-safe. In particular, if
4872 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4873 : * output those expressions, and will not be parallel-safe unless those
4874 : * expressions are parallel-safe.
4875 : */
4876 3024 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4877 :
4878 : /*
4879 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4880 : */
4881 3024 : distinct_rel->serverid = input_rel->serverid;
4882 3024 : distinct_rel->userid = input_rel->userid;
4883 3024 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4884 3024 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4885 :
4886 : /* build distinct paths based on input_rel's pathlist */
4887 3024 : create_final_distinct_paths(root, input_rel, distinct_rel);
4888 :
4889 : /* now build distinct paths based on input_rel's partial_pathlist */
4890 3024 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4891 :
4892 : /* Give a helpful error if we failed to create any paths */
4893 3024 : if (distinct_rel->pathlist == NIL)
4894 0 : ereport(ERROR,
4895 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4896 : errmsg("could not implement DISTINCT"),
4897 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4898 :
4899 : /*
4900 : * If there is an FDW that's responsible for all baserels of the query,
4901 : * let it consider adding ForeignPaths.
4902 : */
4903 3024 : if (distinct_rel->fdwroutine &&
4904 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4905 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4906 : UPPERREL_DISTINCT,
4907 : input_rel,
4908 : distinct_rel,
4909 : NULL);
4910 :
4911 : /* Let extensions possibly add some more paths */
4912 3024 : if (create_upper_paths_hook)
4913 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4914 : distinct_rel, NULL);
4915 :
4916 : /* Now choose the best path(s) */
4917 3024 : set_cheapest(distinct_rel);
4918 :
4919 3024 : return distinct_rel;
4920 : }
4921 :
4922 : /*
4923 : * create_partial_distinct_paths
4924 : *
4925 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4926 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4927 : * paths on top and add a final unique/aggregate path to remove any duplicate
4928 : * produced from combining rows from parallel workers.
4929 : */
4930 : static void
4931 3024 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4932 : RelOptInfo *final_distinct_rel,
4933 : PathTarget *target)
4934 : {
4935 : RelOptInfo *partial_distinct_rel;
4936 : Query *parse;
4937 : List *distinctExprs;
4938 : double numDistinctRows;
4939 : Path *cheapest_partial_path;
4940 : ListCell *lc;
4941 :
4942 : /* nothing to do when there are no partial paths in the input rel */
4943 3024 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4944 2916 : return;
4945 :
4946 108 : parse = root->parse;
4947 :
4948 : /* can't do parallel DISTINCT ON */
4949 108 : if (parse->hasDistinctOn)
4950 0 : return;
4951 :
4952 108 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4953 : NULL);
4954 108 : partial_distinct_rel->reltarget = target;
4955 108 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4956 :
4957 : /*
4958 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4959 : */
4960 108 : partial_distinct_rel->serverid = input_rel->serverid;
4961 108 : partial_distinct_rel->userid = input_rel->userid;
4962 108 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4963 108 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4964 :
4965 108 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4966 :
4967 108 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4968 : parse->targetList);
4969 :
4970 : /* estimate how many distinct rows we'll get from each worker */
4971 108 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4972 : cheapest_partial_path->rows,
4973 : NULL, NULL);
4974 :
4975 : /*
4976 : * Try sorting the cheapest path and incrementally sorting any paths with
4977 : * presorted keys and put a unique paths atop of those. We'll also
4978 : * attempt to reorder the required pathkeys to match the input path's
4979 : * pathkeys as much as possible, in hopes of avoiding a possible need to
4980 : * re-sort.
4981 : */
4982 108 : if (grouping_is_sortable(root->processed_distinctClause))
4983 : {
4984 234 : foreach(lc, input_rel->partial_pathlist)
4985 : {
4986 126 : Path *input_path = (Path *) lfirst(lc);
4987 : Path *sorted_path;
4988 126 : List *useful_pathkeys_list = NIL;
4989 :
4990 : useful_pathkeys_list =
4991 126 : get_useful_pathkeys_for_distinct(root,
4992 : root->distinct_pathkeys,
4993 : input_path->pathkeys);
4994 : Assert(list_length(useful_pathkeys_list) > 0);
4995 :
4996 390 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
4997 : {
4998 138 : sorted_path = make_ordered_path(root,
4999 : partial_distinct_rel,
5000 : input_path,
5001 : cheapest_partial_path,
5002 : useful_pathkeys,
5003 : -1.0);
5004 :
5005 138 : if (sorted_path == NULL)
5006 12 : continue;
5007 :
5008 : /*
5009 : * An empty distinct_pathkeys means all tuples have the same
5010 : * value for the DISTINCT clause. See
5011 : * create_final_distinct_paths()
5012 : */
5013 126 : if (root->distinct_pathkeys == NIL)
5014 : {
5015 : Node *limitCount;
5016 :
5017 6 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5018 : sizeof(int64),
5019 : Int64GetDatum(1), false,
5020 : true);
5021 :
5022 : /*
5023 : * Apply a LimitPath onto the partial path to restrict the
5024 : * tuples from each worker to 1.
5025 : * create_final_distinct_paths will need to apply an
5026 : * additional LimitPath to restrict this to a single row
5027 : * after the Gather node. If the query already has a
5028 : * LIMIT clause, then we could end up with three Limit
5029 : * nodes in the final plan. Consolidating the top two of
5030 : * these could be done, but does not seem worth troubling
5031 : * over.
5032 : */
5033 6 : add_partial_path(partial_distinct_rel, (Path *)
5034 6 : create_limit_path(root, partial_distinct_rel,
5035 : sorted_path,
5036 : NULL,
5037 : limitCount,
5038 : LIMIT_OPTION_COUNT,
5039 : 0, 1));
5040 : }
5041 : else
5042 : {
5043 120 : add_partial_path(partial_distinct_rel, (Path *)
5044 120 : create_unique_path(root, partial_distinct_rel,
5045 : sorted_path,
5046 120 : list_length(root->distinct_pathkeys),
5047 : numDistinctRows));
5048 : }
5049 : }
5050 : }
5051 : }
5052 :
5053 : /*
5054 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5055 : * we're not on the hook to ensure we do our best to create at least one
5056 : * path here, we treat enable_hashagg as a hard off-switch rather than the
5057 : * slightly softer variant in create_final_distinct_paths.
5058 : */
5059 108 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5060 : {
5061 78 : add_partial_path(partial_distinct_rel, (Path *)
5062 78 : create_agg_path(root,
5063 : partial_distinct_rel,
5064 : cheapest_partial_path,
5065 : cheapest_partial_path->pathtarget,
5066 : AGG_HASHED,
5067 : AGGSPLIT_SIMPLE,
5068 : root->processed_distinctClause,
5069 : NIL,
5070 : NULL,
5071 : numDistinctRows));
5072 : }
5073 :
5074 : /*
5075 : * If there is an FDW that's responsible for all baserels of the query,
5076 : * let it consider adding ForeignPaths.
5077 : */
5078 108 : if (partial_distinct_rel->fdwroutine &&
5079 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5080 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5081 : UPPERREL_PARTIAL_DISTINCT,
5082 : input_rel,
5083 : partial_distinct_rel,
5084 : NULL);
5085 :
5086 : /* Let extensions possibly add some more partial paths */
5087 108 : if (create_upper_paths_hook)
5088 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5089 : input_rel, partial_distinct_rel, NULL);
5090 :
5091 108 : if (partial_distinct_rel->partial_pathlist != NIL)
5092 : {
5093 108 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5094 108 : set_cheapest(partial_distinct_rel);
5095 :
5096 : /*
5097 : * Finally, create paths to distinctify the final result. This step
5098 : * is needed to remove any duplicates due to combining rows from
5099 : * parallel workers.
5100 : */
5101 108 : create_final_distinct_paths(root, partial_distinct_rel,
5102 : final_distinct_rel);
5103 : }
5104 : }
5105 :
5106 : /*
5107 : * create_final_distinct_paths
5108 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5109 : *
5110 : * input_rel: contains the source-data paths
5111 : * distinct_rel: destination relation for storing created paths
5112 : */
5113 : static RelOptInfo *
5114 3132 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5115 : RelOptInfo *distinct_rel)
5116 : {
5117 3132 : Query *parse = root->parse;
5118 3132 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5119 : double numDistinctRows;
5120 : bool allow_hash;
5121 :
5122 : /* Estimate number of distinct rows there will be */
5123 3132 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5124 3058 : root->hasHavingQual)
5125 : {
5126 : /*
5127 : * If there was grouping or aggregation, use the number of input rows
5128 : * as the estimated number of DISTINCT rows (ie, assume the input is
5129 : * already mostly unique).
5130 : */
5131 74 : numDistinctRows = cheapest_input_path->rows;
5132 : }
5133 : else
5134 : {
5135 : /*
5136 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5137 : */
5138 : List *distinctExprs;
5139 :
5140 3058 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5141 : parse->targetList);
5142 3058 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5143 : cheapest_input_path->rows,
5144 : NULL, NULL);
5145 : }
5146 :
5147 : /*
5148 : * Consider sort-based implementations of DISTINCT, if possible.
5149 : */
5150 3132 : if (grouping_is_sortable(root->processed_distinctClause))
5151 : {
5152 : /*
5153 : * Firstly, if we have any adequately-presorted paths, just stick a
5154 : * Unique node on those. We also, consider doing an explicit sort of
5155 : * the cheapest input path and Unique'ing that. If any paths have
5156 : * presorted keys then we'll create an incremental sort atop of those
5157 : * before adding a unique node on the top. We'll also attempt to
5158 : * reorder the required pathkeys to match the input path's pathkeys as
5159 : * much as possible, in hopes of avoiding a possible need to re-sort.
5160 : *
5161 : * When we have DISTINCT ON, we must sort by the more rigorous of
5162 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5163 : * Also, if we do have to do an explicit sort, we might as well use
5164 : * the more rigorous ordering to avoid a second sort later. (Note
5165 : * that the parser will have ensured that one clause is a prefix of
5166 : * the other.)
5167 : */
5168 : List *needed_pathkeys;
5169 : ListCell *lc;
5170 3126 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5171 :
5172 3376 : if (parse->hasDistinctOn &&
5173 250 : list_length(root->distinct_pathkeys) <
5174 250 : list_length(root->sort_pathkeys))
5175 54 : needed_pathkeys = root->sort_pathkeys;
5176 : else
5177 3072 : needed_pathkeys = root->distinct_pathkeys;
5178 :
5179 8430 : foreach(lc, input_rel->pathlist)
5180 : {
5181 5304 : Path *input_path = (Path *) lfirst(lc);
5182 : Path *sorted_path;
5183 5304 : List *useful_pathkeys_list = NIL;
5184 :
5185 : useful_pathkeys_list =
5186 5304 : get_useful_pathkeys_for_distinct(root,
5187 : needed_pathkeys,
5188 : input_path->pathkeys);
5189 : Assert(list_length(useful_pathkeys_list) > 0);
5190 :
5191 16732 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5192 : {
5193 6124 : sorted_path = make_ordered_path(root,
5194 : distinct_rel,
5195 : input_path,
5196 : cheapest_input_path,
5197 : useful_pathkeys,
5198 : limittuples);
5199 :
5200 6124 : if (sorted_path == NULL)
5201 878 : continue;
5202 :
5203 : /*
5204 : * distinct_pathkeys may have become empty if all of the
5205 : * pathkeys were determined to be redundant. If all of the
5206 : * pathkeys are redundant then each DISTINCT target must only
5207 : * allow a single value, therefore all resulting tuples must
5208 : * be identical (or at least indistinguishable by an equality
5209 : * check). We can uniquify these tuples simply by just taking
5210 : * the first tuple. All we do here is add a path to do "LIMIT
5211 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5212 : * still have a non-NIL sort_pathkeys list, so we must still
5213 : * only do this with paths which are correctly sorted by
5214 : * sort_pathkeys.
5215 : */
5216 5246 : if (root->distinct_pathkeys == NIL)
5217 : {
5218 : Node *limitCount;
5219 :
5220 138 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5221 : sizeof(int64),
5222 : Int64GetDatum(1), false,
5223 : true);
5224 :
5225 : /*
5226 : * If the query already has a LIMIT clause, then we could
5227 : * end up with a duplicate LimitPath in the final plan.
5228 : * That does not seem worth troubling over too much.
5229 : */
5230 138 : add_path(distinct_rel, (Path *)
5231 138 : create_limit_path(root, distinct_rel, sorted_path,
5232 : NULL, limitCount,
5233 : LIMIT_OPTION_COUNT, 0, 1));
5234 : }
5235 : else
5236 : {
5237 5108 : add_path(distinct_rel, (Path *)
5238 5108 : create_unique_path(root, distinct_rel,
5239 : sorted_path,
5240 5108 : list_length(root->distinct_pathkeys),
5241 : numDistinctRows));
5242 : }
5243 : }
5244 : }
5245 : }
5246 :
5247 : /*
5248 : * Consider hash-based implementations of DISTINCT, if possible.
5249 : *
5250 : * If we were not able to make any other types of path, we *must* hash or
5251 : * die trying. If we do have other choices, there are two things that
5252 : * should prevent selection of hashing: if the query uses DISTINCT ON
5253 : * (because it won't really have the expected behavior if we hash), or if
5254 : * enable_hashagg is off.
5255 : *
5256 : * Note: grouping_is_hashable() is much more expensive to check than the
5257 : * other gating conditions, so we want to do it last.
5258 : */
5259 3132 : if (distinct_rel->pathlist == NIL)
5260 6 : allow_hash = true; /* we have no alternatives */
5261 3126 : else if (parse->hasDistinctOn || !enable_hashagg)
5262 400 : allow_hash = false; /* policy-based decision not to hash */
5263 : else
5264 2726 : allow_hash = true; /* default */
5265 :
5266 3132 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5267 : {
5268 : /* Generate hashed aggregate path --- no sort needed */
5269 2732 : add_path(distinct_rel, (Path *)
5270 2732 : create_agg_path(root,
5271 : distinct_rel,
5272 : cheapest_input_path,
5273 : cheapest_input_path->pathtarget,
5274 : AGG_HASHED,
5275 : AGGSPLIT_SIMPLE,
5276 : root->processed_distinctClause,
5277 : NIL,
5278 : NULL,
5279 : numDistinctRows));
5280 : }
5281 :
5282 3132 : return distinct_rel;
5283 : }
5284 :
5285 : /*
5286 : * get_useful_pathkeys_for_distinct
5287 : * Get useful orderings of pathkeys for distinctClause by reordering
5288 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5289 : *
5290 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5291 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5292 : */
5293 : static List *
5294 5430 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5295 : List *path_pathkeys)
5296 : {
5297 5430 : List *useful_pathkeys_list = NIL;
5298 5430 : List *useful_pathkeys = NIL;
5299 :
5300 : /* always include the given 'needed_pathkeys' */
5301 5430 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5302 : needed_pathkeys);
5303 :
5304 5430 : if (!enable_distinct_reordering)
5305 0 : return useful_pathkeys_list;
5306 :
5307 : /*
5308 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5309 : * that match 'needed_pathkeys', but only up to the longest matching
5310 : * prefix.
5311 : *
5312 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5313 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5314 : * the desired behavior.
5315 : */
5316 13442 : foreach_node(PathKey, pathkey, path_pathkeys)
5317 : {
5318 : /*
5319 : * The PathKey nodes are canonical, so they can be checked for
5320 : * equality by simple pointer comparison.
5321 : */
5322 2610 : if (!list_member_ptr(needed_pathkeys, pathkey))
5323 10 : break;
5324 2600 : if (root->parse->hasDistinctOn &&
5325 202 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5326 18 : break;
5327 :
5328 2582 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5329 : }
5330 :
5331 : /* If no match at all, no point in reordering needed_pathkeys */
5332 5430 : if (useful_pathkeys == NIL)
5333 3112 : return useful_pathkeys_list;
5334 :
5335 : /*
5336 : * If not full match, the resulting pathkey list is not useful without
5337 : * incremental sort.
5338 : */
5339 2318 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5340 1562 : !enable_incremental_sort)
5341 60 : return useful_pathkeys_list;
5342 :
5343 : /* Append the remaining PathKey nodes in needed_pathkeys */
5344 2258 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5345 : needed_pathkeys);
5346 :
5347 : /*
5348 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5349 : * just drop it.
5350 : */
5351 2258 : if (compare_pathkeys(needed_pathkeys,
5352 : useful_pathkeys) == PATHKEYS_EQUAL)
5353 1426 : return useful_pathkeys_list;
5354 :
5355 832 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5356 : useful_pathkeys);
5357 :
5358 832 : return useful_pathkeys_list;
5359 : }
5360 :
5361 : /*
5362 : * create_ordered_paths
5363 : *
5364 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5365 : *
5366 : * All paths in the result must satisfy the ORDER BY ordering.
5367 : * The only new paths we need consider are an explicit full sort
5368 : * and incremental sort on the cheapest-total existing path.
5369 : *
5370 : * input_rel: contains the source-data Paths
5371 : * target: the output tlist the result Paths must emit
5372 : * limit_tuples: estimated bound on the number of output tuples,
5373 : * or -1 if no LIMIT or couldn't estimate
5374 : *
5375 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5376 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5377 : */
5378 : static RelOptInfo *
5379 76072 : create_ordered_paths(PlannerInfo *root,
5380 : RelOptInfo *input_rel,
5381 : PathTarget *target,
5382 : bool target_parallel_safe,
5383 : double limit_tuples)
5384 : {
5385 76072 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5386 : RelOptInfo *ordered_rel;
5387 : ListCell *lc;
5388 :
5389 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5390 76072 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5391 :
5392 : /*
5393 : * If the input relation is not parallel-safe, then the ordered relation
5394 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5395 : * target list is parallel-safe.
5396 : */
5397 76072 : if (input_rel->consider_parallel && target_parallel_safe)
5398 53048 : ordered_rel->consider_parallel = true;
5399 :
5400 : /* Assume that the same path generation strategies are allowed. */
5401 76072 : ordered_rel->pgs_mask = input_rel->pgs_mask;
5402 :
5403 : /*
5404 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5405 : */
5406 76072 : ordered_rel->serverid = input_rel->serverid;
5407 76072 : ordered_rel->userid = input_rel->userid;
5408 76072 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5409 76072 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5410 :
5411 192126 : foreach(lc, input_rel->pathlist)
5412 : {
5413 116054 : Path *input_path = (Path *) lfirst(lc);
5414 : Path *sorted_path;
5415 : bool is_sorted;
5416 : int presorted_keys;
5417 :
5418 116054 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5419 : input_path->pathkeys, &presorted_keys);
5420 :
5421 116054 : if (is_sorted)
5422 43546 : sorted_path = input_path;
5423 : else
5424 : {
5425 : /*
5426 : * Try at least sorting the cheapest path and also try
5427 : * incrementally sorting any path which is partially sorted
5428 : * already (no need to deal with paths which have presorted keys
5429 : * when incremental sort is disabled unless it's the cheapest
5430 : * input path).
5431 : */
5432 72508 : if (input_path != cheapest_input_path &&
5433 6336 : (presorted_keys == 0 || !enable_incremental_sort))
5434 2232 : continue;
5435 :
5436 : /*
5437 : * We've no need to consider both a sort and incremental sort.
5438 : * We'll just do a sort if there are no presorted keys and an
5439 : * incremental sort when there are presorted keys.
5440 : */
5441 70276 : if (presorted_keys == 0 || !enable_incremental_sort)
5442 65544 : sorted_path = (Path *) create_sort_path(root,
5443 : ordered_rel,
5444 : input_path,
5445 : root->sort_pathkeys,
5446 : limit_tuples);
5447 : else
5448 4732 : sorted_path = (Path *) create_incremental_sort_path(root,
5449 : ordered_rel,
5450 : input_path,
5451 : root->sort_pathkeys,
5452 : presorted_keys,
5453 : limit_tuples);
5454 : }
5455 :
5456 : /*
5457 : * If the pathtarget of the result path has different expressions from
5458 : * the target to be applied, a projection step is needed.
5459 : */
5460 113822 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5461 394 : sorted_path = apply_projection_to_path(root, ordered_rel,
5462 : sorted_path, target);
5463 :
5464 113822 : add_path(ordered_rel, sorted_path);
5465 : }
5466 :
5467 : /*
5468 : * generate_gather_paths() will have already generated a simple Gather
5469 : * path for the best parallel path, if any, and the loop above will have
5470 : * considered sorting it. Similarly, generate_gather_paths() will also
5471 : * have generated order-preserving Gather Merge plans which can be used
5472 : * without sorting if they happen to match the sort_pathkeys, and the loop
5473 : * above will have handled those as well. However, there's one more
5474 : * possibility: it may make sense to sort the cheapest partial path or
5475 : * incrementally sort any partial path that is partially sorted according
5476 : * to the required output order and then use Gather Merge.
5477 : */
5478 76072 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5479 52844 : input_rel->partial_pathlist != NIL)
5480 : {
5481 : Path *cheapest_partial_path;
5482 :
5483 2866 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5484 :
5485 6526 : foreach(lc, input_rel->partial_pathlist)
5486 : {
5487 3660 : Path *input_path = (Path *) lfirst(lc);
5488 : Path *sorted_path;
5489 : bool is_sorted;
5490 : int presorted_keys;
5491 : double total_groups;
5492 :
5493 3660 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5494 : input_path->pathkeys,
5495 : &presorted_keys);
5496 :
5497 3660 : if (is_sorted)
5498 674 : continue;
5499 :
5500 : /*
5501 : * Try at least sorting the cheapest path and also try
5502 : * incrementally sorting any path which is partially sorted
5503 : * already (no need to deal with paths which have presorted keys
5504 : * when incremental sort is disabled unless it's the cheapest
5505 : * partial path).
5506 : */
5507 2986 : if (input_path != cheapest_partial_path &&
5508 150 : (presorted_keys == 0 || !enable_incremental_sort))
5509 0 : continue;
5510 :
5511 : /*
5512 : * We've no need to consider both a sort and incremental sort.
5513 : * We'll just do a sort if there are no presorted keys and an
5514 : * incremental sort when there are presorted keys.
5515 : */
5516 2986 : if (presorted_keys == 0 || !enable_incremental_sort)
5517 2818 : sorted_path = (Path *) create_sort_path(root,
5518 : ordered_rel,
5519 : input_path,
5520 : root->sort_pathkeys,
5521 : limit_tuples);
5522 : else
5523 168 : sorted_path = (Path *) create_incremental_sort_path(root,
5524 : ordered_rel,
5525 : input_path,
5526 : root->sort_pathkeys,
5527 : presorted_keys,
5528 : limit_tuples);
5529 2986 : total_groups = compute_gather_rows(sorted_path);
5530 : sorted_path = (Path *)
5531 2986 : create_gather_merge_path(root, ordered_rel,
5532 : sorted_path,
5533 : sorted_path->pathtarget,
5534 : root->sort_pathkeys, NULL,
5535 : &total_groups);
5536 :
5537 : /*
5538 : * If the pathtarget of the result path has different expressions
5539 : * from the target to be applied, a projection step is needed.
5540 : */
5541 2986 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5542 6 : sorted_path = apply_projection_to_path(root, ordered_rel,
5543 : sorted_path, target);
5544 :
5545 2986 : add_path(ordered_rel, sorted_path);
5546 : }
5547 : }
5548 :
5549 : /*
5550 : * If there is an FDW that's responsible for all baserels of the query,
5551 : * let it consider adding ForeignPaths.
5552 : */
5553 76072 : if (ordered_rel->fdwroutine &&
5554 386 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5555 370 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5556 : input_rel, ordered_rel,
5557 : NULL);
5558 :
5559 : /* Let extensions possibly add some more paths */
5560 76072 : if (create_upper_paths_hook)
5561 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5562 : input_rel, ordered_rel, NULL);
5563 :
5564 : /*
5565 : * No need to bother with set_cheapest here; grouping_planner does not
5566 : * need us to do it.
5567 : */
5568 : Assert(ordered_rel->pathlist != NIL);
5569 :
5570 76072 : return ordered_rel;
5571 : }
5572 :
5573 :
5574 : /*
5575 : * make_group_input_target
5576 : * Generate appropriate PathTarget for initial input to grouping nodes.
5577 : *
5578 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5579 : * the query's final targetlist; for example, it certainly can't emit any
5580 : * aggregate function calls. This routine generates the correct target
5581 : * for the scan/join subplan.
5582 : *
5583 : * The query target list passed from the parser already contains entries
5584 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5585 : * for variables used only in HAVING clauses; so we need to add those
5586 : * variables to the subplan target list. Also, we flatten all expressions
5587 : * except GROUP BY items into their component variables; other expressions
5588 : * will be computed by the upper plan nodes rather than by the subplan.
5589 : * For example, given a query like
5590 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5591 : * we want to pass this targetlist to the subplan:
5592 : * a+b,c,d
5593 : * where the a+b target will be used by the Sort/Group steps, and the
5594 : * other targets will be used for computing the final results.
5595 : *
5596 : * 'final_target' is the query's final target list (in PathTarget form)
5597 : *
5598 : * The result is the PathTarget to be computed by the Paths returned from
5599 : * query_planner().
5600 : */
5601 : static PathTarget *
5602 45968 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5603 : {
5604 45968 : Query *parse = root->parse;
5605 : PathTarget *input_target;
5606 : List *non_group_cols;
5607 : List *non_group_vars;
5608 : int i;
5609 : ListCell *lc;
5610 :
5611 : /*
5612 : * We must build a target containing all grouping columns, plus any other
5613 : * Vars mentioned in the query's targetlist and HAVING qual.
5614 : */
5615 45968 : input_target = create_empty_pathtarget();
5616 45968 : non_group_cols = NIL;
5617 :
5618 45968 : i = 0;
5619 114032 : foreach(lc, final_target->exprs)
5620 : {
5621 68064 : Expr *expr = (Expr *) lfirst(lc);
5622 68064 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5623 :
5624 77744 : if (sgref && root->processed_groupClause &&
5625 9680 : get_sortgroupref_clause_noerr(sgref,
5626 : root->processed_groupClause) != NULL)
5627 : {
5628 : /*
5629 : * It's a grouping column, so add it to the input target as-is.
5630 : *
5631 : * Note that the target is logically below the grouping step. So
5632 : * with grouping sets we need to remove the RT index of the
5633 : * grouping step if there is any from the target expression.
5634 : */
5635 7810 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5636 : {
5637 : Assert(root->group_rtindex > 0);
5638 : expr = (Expr *)
5639 2026 : remove_nulling_relids((Node *) expr,
5640 2026 : bms_make_singleton(root->group_rtindex),
5641 : NULL);
5642 : }
5643 7810 : add_column_to_pathtarget(input_target, expr, sgref);
5644 : }
5645 : else
5646 : {
5647 : /*
5648 : * Non-grouping column, so just remember the expression for later
5649 : * call to pull_var_clause.
5650 : */
5651 60254 : non_group_cols = lappend(non_group_cols, expr);
5652 : }
5653 :
5654 68064 : i++;
5655 : }
5656 :
5657 : /*
5658 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5659 : */
5660 45968 : if (parse->havingQual)
5661 962 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5662 :
5663 : /*
5664 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5665 : * add them to the input target if not already present. (A Var used
5666 : * directly as a GROUP BY item will be present already.) Note this
5667 : * includes Vars used in resjunk items, so we are covering the needs of
5668 : * ORDER BY and window specifications. Vars used within Aggrefs and
5669 : * WindowFuncs will be pulled out here, too.
5670 : *
5671 : * Note that the target is logically below the grouping step. So with
5672 : * grouping sets we need to remove the RT index of the grouping step if
5673 : * there is any from the non-group Vars.
5674 : */
5675 45968 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5676 : PVC_RECURSE_AGGREGATES |
5677 : PVC_RECURSE_WINDOWFUNCS |
5678 : PVC_INCLUDE_PLACEHOLDERS);
5679 45968 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5680 : {
5681 : Assert(root->group_rtindex > 0);
5682 : non_group_vars = (List *)
5683 936 : remove_nulling_relids((Node *) non_group_vars,
5684 936 : bms_make_singleton(root->group_rtindex),
5685 : NULL);
5686 : }
5687 45968 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5688 :
5689 : /* clean up cruft */
5690 45968 : list_free(non_group_vars);
5691 45968 : list_free(non_group_cols);
5692 :
5693 : /* XXX this causes some redundant cost calculation ... */
5694 45968 : return set_pathtarget_cost_width(root, input_target);
5695 : }
5696 :
5697 : /*
5698 : * make_partial_grouping_target
5699 : * Generate appropriate PathTarget for output of partial aggregate
5700 : * (or partial grouping, if there are no aggregates) nodes.
5701 : *
5702 : * A partial aggregation node needs to emit all the same aggregates that
5703 : * a regular aggregation node would, plus any aggregates used in HAVING;
5704 : * except that the Aggref nodes should be marked as partial aggregates.
5705 : *
5706 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5707 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5708 : * these would be Vars that are grouped by or used in grouping expressions.)
5709 : *
5710 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5711 : * havingQual represents the HAVING clause.
5712 : */
5713 : static PathTarget *
5714 3858 : make_partial_grouping_target(PlannerInfo *root,
5715 : PathTarget *grouping_target,
5716 : Node *havingQual)
5717 : {
5718 : PathTarget *partial_target;
5719 : List *non_group_cols;
5720 : List *non_group_exprs;
5721 : int i;
5722 : ListCell *lc;
5723 :
5724 3858 : partial_target = create_empty_pathtarget();
5725 3858 : non_group_cols = NIL;
5726 :
5727 3858 : i = 0;
5728 12758 : foreach(lc, grouping_target->exprs)
5729 : {
5730 8900 : Expr *expr = (Expr *) lfirst(lc);
5731 8900 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5732 :
5733 13676 : if (sgref && root->processed_groupClause &&
5734 4776 : get_sortgroupref_clause_noerr(sgref,
5735 : root->processed_groupClause) != NULL)
5736 : {
5737 : /*
5738 : * It's a grouping column, so add it to the partial_target as-is.
5739 : * (This allows the upper agg step to repeat the grouping calcs.)
5740 : */
5741 2828 : add_column_to_pathtarget(partial_target, expr, sgref);
5742 : }
5743 : else
5744 : {
5745 : /*
5746 : * Non-grouping column, so just remember the expression for later
5747 : * call to pull_var_clause.
5748 : */
5749 6072 : non_group_cols = lappend(non_group_cols, expr);
5750 : }
5751 :
5752 8900 : i++;
5753 : }
5754 :
5755 : /*
5756 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5757 : */
5758 3858 : if (havingQual)
5759 878 : non_group_cols = lappend(non_group_cols, havingQual);
5760 :
5761 : /*
5762 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5763 : * non-group cols (plus HAVING), and add them to the partial_target if not
5764 : * already present. (An expression used directly as a GROUP BY item will
5765 : * be present already.) Note this includes Vars used in resjunk items, so
5766 : * we are covering the needs of ORDER BY and window specifications.
5767 : */
5768 3858 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5769 : PVC_INCLUDE_AGGREGATES |
5770 : PVC_RECURSE_WINDOWFUNCS |
5771 : PVC_INCLUDE_PLACEHOLDERS);
5772 :
5773 3858 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5774 :
5775 : /*
5776 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5777 : * are at the top level of the target list, so we can just scan the list
5778 : * rather than recursing through the expression trees.
5779 : */
5780 13354 : foreach(lc, partial_target->exprs)
5781 : {
5782 9496 : Aggref *aggref = (Aggref *) lfirst(lc);
5783 :
5784 9496 : if (IsA(aggref, Aggref))
5785 : {
5786 : Aggref *newaggref;
5787 :
5788 : /*
5789 : * We shouldn't need to copy the substructure of the Aggref node,
5790 : * but flat-copy the node itself to avoid damaging other trees.
5791 : */
5792 6638 : newaggref = makeNode(Aggref);
5793 6638 : memcpy(newaggref, aggref, sizeof(Aggref));
5794 :
5795 : /* For now, assume serialization is required */
5796 6638 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5797 :
5798 6638 : lfirst(lc) = newaggref;
5799 : }
5800 : }
5801 :
5802 : /* clean up cruft */
5803 3858 : list_free(non_group_exprs);
5804 3858 : list_free(non_group_cols);
5805 :
5806 : /* XXX this causes some redundant cost calculation ... */
5807 3858 : return set_pathtarget_cost_width(root, partial_target);
5808 : }
5809 :
5810 : /*
5811 : * mark_partial_aggref
5812 : * Adjust an Aggref to make it represent a partial-aggregation step.
5813 : *
5814 : * The Aggref node is modified in-place; caller must do any copying required.
5815 : */
5816 : void
5817 18588 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5818 : {
5819 : /* aggtranstype should be computed by this point */
5820 : Assert(OidIsValid(agg->aggtranstype));
5821 : /* ... but aggsplit should still be as the parser left it */
5822 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5823 :
5824 : /* Mark the Aggref with the intended partial-aggregation mode */
5825 18588 : agg->aggsplit = aggsplit;
5826 :
5827 : /*
5828 : * Adjust result type if needed. Normally, a partial aggregate returns
5829 : * the aggregate's transition type; but if that's INTERNAL and we're
5830 : * serializing, it returns BYTEA instead.
5831 : */
5832 18588 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5833 : {
5834 16280 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5835 314 : agg->aggtype = BYTEAOID;
5836 : else
5837 15966 : agg->aggtype = agg->aggtranstype;
5838 : }
5839 18588 : }
5840 :
5841 : /*
5842 : * postprocess_setop_tlist
5843 : * Fix up targetlist returned by plan_set_operations().
5844 : *
5845 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5846 : * NOTE: this would not be good enough if we supported resjunk sort keys
5847 : * for results of set operations --- then, we'd need to project a whole
5848 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5849 : * find any resjunk columns in orig_tlist.
5850 : */
5851 : static List *
5852 6212 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5853 : {
5854 : ListCell *l;
5855 6212 : ListCell *orig_tlist_item = list_head(orig_tlist);
5856 :
5857 23846 : foreach(l, new_tlist)
5858 : {
5859 17634 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5860 : TargetEntry *orig_tle;
5861 :
5862 : /* ignore resjunk columns in setop result */
5863 17634 : if (new_tle->resjunk)
5864 0 : continue;
5865 :
5866 : Assert(orig_tlist_item != NULL);
5867 17634 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5868 17634 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5869 17634 : if (orig_tle->resjunk) /* should not happen */
5870 0 : elog(ERROR, "resjunk output columns are not implemented");
5871 : Assert(new_tle->resno == orig_tle->resno);
5872 17634 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5873 : }
5874 6212 : if (orig_tlist_item != NULL)
5875 0 : elog(ERROR, "resjunk output columns are not implemented");
5876 6212 : return new_tlist;
5877 : }
5878 :
5879 : /*
5880 : * optimize_window_clauses
5881 : * Call each WindowFunc's prosupport function to see if we're able to
5882 : * make any adjustments to any of the WindowClause's so that the executor
5883 : * can execute the window functions in a more optimal way.
5884 : *
5885 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5886 : * may allow more things to be done here in the future.
5887 : */
5888 : static void
5889 2676 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5890 : {
5891 2676 : List *windowClause = root->parse->windowClause;
5892 : ListCell *lc;
5893 :
5894 5598 : foreach(lc, windowClause)
5895 : {
5896 2922 : WindowClause *wc = lfirst_node(WindowClause, lc);
5897 : ListCell *lc2;
5898 2922 : int optimizedFrameOptions = 0;
5899 :
5900 : Assert(wc->winref <= wflists->maxWinRef);
5901 :
5902 : /* skip any WindowClauses that have no WindowFuncs */
5903 2922 : if (wflists->windowFuncs[wc->winref] == NIL)
5904 24 : continue;
5905 :
5906 3552 : foreach(lc2, wflists->windowFuncs[wc->winref])
5907 : {
5908 : SupportRequestOptimizeWindowClause req;
5909 : SupportRequestOptimizeWindowClause *res;
5910 2940 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5911 : Oid prosupport;
5912 :
5913 2940 : prosupport = get_func_support(wfunc->winfnoid);
5914 :
5915 : /* Check if there's a support function for 'wfunc' */
5916 2940 : if (!OidIsValid(prosupport))
5917 2286 : break; /* can't optimize this WindowClause */
5918 :
5919 880 : req.type = T_SupportRequestOptimizeWindowClause;
5920 880 : req.window_clause = wc;
5921 880 : req.window_func = wfunc;
5922 880 : req.frameOptions = wc->frameOptions;
5923 :
5924 : /* call the support function */
5925 : res = (SupportRequestOptimizeWindowClause *)
5926 880 : DatumGetPointer(OidFunctionCall1(prosupport,
5927 : PointerGetDatum(&req)));
5928 :
5929 : /*
5930 : * Skip to next WindowClause if the support function does not
5931 : * support this request type.
5932 : */
5933 880 : if (res == NULL)
5934 226 : break;
5935 :
5936 : /*
5937 : * Save these frameOptions for the first WindowFunc for this
5938 : * WindowClause.
5939 : */
5940 654 : if (foreach_current_index(lc2) == 0)
5941 630 : optimizedFrameOptions = res->frameOptions;
5942 :
5943 : /*
5944 : * On subsequent WindowFuncs, if the frameOptions are not the same
5945 : * then we're unable to optimize the frameOptions for this
5946 : * WindowClause.
5947 : */
5948 24 : else if (optimizedFrameOptions != res->frameOptions)
5949 0 : break; /* skip to the next WindowClause, if any */
5950 : }
5951 :
5952 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
5953 2898 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5954 : {
5955 : ListCell *lc3;
5956 :
5957 : /* apply the new frame options */
5958 612 : wc->frameOptions = optimizedFrameOptions;
5959 :
5960 : /*
5961 : * We now check to see if changing the frameOptions has caused
5962 : * this WindowClause to be a duplicate of some other WindowClause.
5963 : * This can only happen if we have multiple WindowClauses, so
5964 : * don't bother if there's only 1.
5965 : */
5966 612 : if (list_length(windowClause) == 1)
5967 522 : continue;
5968 :
5969 : /*
5970 : * Do the duplicate check and reuse the existing WindowClause if
5971 : * we find a duplicate.
5972 : */
5973 228 : foreach(lc3, windowClause)
5974 : {
5975 174 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5976 :
5977 : /* skip over the WindowClause we're currently editing */
5978 174 : if (existing_wc == wc)
5979 54 : continue;
5980 :
5981 : /*
5982 : * Perform the same duplicate check that is done in
5983 : * transformWindowFuncCall.
5984 : */
5985 240 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
5986 120 : equal(wc->orderClause, existing_wc->orderClause) &&
5987 120 : wc->frameOptions == existing_wc->frameOptions &&
5988 72 : equal(wc->startOffset, existing_wc->startOffset) &&
5989 36 : equal(wc->endOffset, existing_wc->endOffset))
5990 : {
5991 : ListCell *lc4;
5992 :
5993 : /*
5994 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
5995 : * This required adjusting each WindowFunc's winref and
5996 : * moving the WindowFuncs in 'wc' to the list of
5997 : * WindowFuncs in 'existing_wc'.
5998 : */
5999 78 : foreach(lc4, wflists->windowFuncs[wc->winref])
6000 : {
6001 42 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
6002 :
6003 42 : wfunc->winref = existing_wc->winref;
6004 : }
6005 :
6006 : /* move list items */
6007 72 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
6008 36 : wflists->windowFuncs[wc->winref]);
6009 36 : wflists->windowFuncs[wc->winref] = NIL;
6010 :
6011 : /*
6012 : * transformWindowFuncCall() should have made sure there
6013 : * are no other duplicates, so we needn't bother looking
6014 : * any further.
6015 : */
6016 36 : break;
6017 : }
6018 : }
6019 : }
6020 : }
6021 2676 : }
6022 :
6023 : /*
6024 : * select_active_windows
6025 : * Create a list of the "active" window clauses (ie, those referenced
6026 : * by non-deleted WindowFuncs) in the order they are to be executed.
6027 : */
6028 : static List *
6029 2676 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
6030 : {
6031 2676 : List *windowClause = root->parse->windowClause;
6032 2676 : List *result = NIL;
6033 : ListCell *lc;
6034 2676 : int nActive = 0;
6035 2676 : WindowClauseSortData *actives = palloc_array(WindowClauseSortData,
6036 : list_length(windowClause));
6037 :
6038 : /* First, construct an array of the active windows */
6039 5598 : foreach(lc, windowClause)
6040 : {
6041 2922 : WindowClause *wc = lfirst_node(WindowClause, lc);
6042 :
6043 : /* It's only active if wflists shows some related WindowFuncs */
6044 : Assert(wc->winref <= wflists->maxWinRef);
6045 2922 : if (wflists->windowFuncs[wc->winref] == NIL)
6046 60 : continue;
6047 :
6048 2862 : actives[nActive].wc = wc; /* original clause */
6049 :
6050 : /*
6051 : * For sorting, we want the list of partition keys followed by the
6052 : * list of sort keys. But pathkeys construction will remove duplicates
6053 : * between the two, so we can as well (even though we can't detect all
6054 : * of the duplicates, since some may come from ECs - that might mean
6055 : * we miss optimization chances here). We must, however, ensure that
6056 : * the order of entries is preserved with respect to the ones we do
6057 : * keep.
6058 : *
6059 : * partitionClause and orderClause had their own duplicates removed in
6060 : * parse analysis, so we're only concerned here with removing
6061 : * orderClause entries that also appear in partitionClause.
6062 : */
6063 5724 : actives[nActive].uniqueOrder =
6064 2862 : list_concat_unique(list_copy(wc->partitionClause),
6065 2862 : wc->orderClause);
6066 2862 : nActive++;
6067 : }
6068 :
6069 : /*
6070 : * Sort active windows by their partitioning/ordering clauses, ignoring
6071 : * any framing clauses, so that the windows that need the same sorting are
6072 : * adjacent in the list. When we come to generate paths, this will avoid
6073 : * inserting additional Sort nodes.
6074 : *
6075 : * This is how we implement a specific requirement from the SQL standard,
6076 : * which says that when two or more windows are order-equivalent (i.e.
6077 : * have matching partition and order clauses, even if their names or
6078 : * framing clauses differ), then all peer rows must be presented in the
6079 : * same order in all of them. If we allowed multiple sort nodes for such
6080 : * cases, we'd risk having the peer rows end up in different orders in
6081 : * equivalent windows due to sort instability. (See General Rule 4 of
6082 : * <window clause> in SQL2008 - SQL2016.)
6083 : *
6084 : * Additionally, if the entire list of clauses of one window is a prefix
6085 : * of another, put first the window with stronger sorting requirements.
6086 : * This way we will first sort for stronger window, and won't have to sort
6087 : * again for the weaker one.
6088 : */
6089 2676 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6090 :
6091 : /* build ordered list of the original WindowClause nodes */
6092 5538 : for (int i = 0; i < nActive; i++)
6093 2862 : result = lappend(result, actives[i].wc);
6094 :
6095 2676 : pfree(actives);
6096 :
6097 2676 : return result;
6098 : }
6099 :
6100 : /*
6101 : * name_active_windows
6102 : * Ensure all active windows have unique names.
6103 : *
6104 : * The parser will have checked that user-assigned window names are unique
6105 : * within the Query. Here we assign made-up names to any unnamed
6106 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6107 : * at parse time, because it'd mess up decompilation of views.)
6108 : *
6109 : * activeWindows: result of select_active_windows
6110 : */
6111 : static void
6112 2676 : name_active_windows(List *activeWindows)
6113 : {
6114 2676 : int next_n = 1;
6115 : char newname[16];
6116 : ListCell *lc;
6117 :
6118 5538 : foreach(lc, activeWindows)
6119 : {
6120 2862 : WindowClause *wc = lfirst_node(WindowClause, lc);
6121 :
6122 : /* Nothing to do if it has a name already. */
6123 2862 : if (wc->name)
6124 576 : continue;
6125 :
6126 : /* Select a name not currently present in the list. */
6127 : for (;;)
6128 6 : {
6129 : ListCell *lc2;
6130 :
6131 2292 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6132 4932 : foreach(lc2, activeWindows)
6133 : {
6134 2646 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6135 :
6136 2646 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6137 6 : break; /* matched */
6138 : }
6139 2292 : if (lc2 == NULL)
6140 2286 : break; /* reached the end with no match */
6141 : }
6142 2286 : wc->name = pstrdup(newname);
6143 : }
6144 2676 : }
6145 :
6146 : /*
6147 : * common_prefix_cmp
6148 : * QSort comparison function for WindowClauseSortData
6149 : *
6150 : * Sort the windows by the required sorting clauses. First, compare the sort
6151 : * clauses themselves. Second, if one window's clauses are a prefix of another
6152 : * one's clauses, put the window with more sort clauses first.
6153 : *
6154 : * We purposefully sort by the highest tleSortGroupRef first. Since
6155 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6156 : * and because here we sort the lowest tleSortGroupRefs last, if a
6157 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6158 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6159 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6160 : * reducing the total number of sorts required for the query.
6161 : */
6162 : static int
6163 204 : common_prefix_cmp(const void *a, const void *b)
6164 : {
6165 204 : const WindowClauseSortData *wcsa = a;
6166 204 : const WindowClauseSortData *wcsb = b;
6167 : ListCell *item_a;
6168 : ListCell *item_b;
6169 :
6170 366 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6171 : {
6172 264 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6173 264 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6174 :
6175 264 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6176 102 : return -1;
6177 252 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6178 66 : return 1;
6179 186 : else if (sca->sortop > scb->sortop)
6180 0 : return -1;
6181 186 : else if (sca->sortop < scb->sortop)
6182 24 : return 1;
6183 162 : else if (sca->nulls_first && !scb->nulls_first)
6184 0 : return -1;
6185 162 : else if (!sca->nulls_first && scb->nulls_first)
6186 0 : return 1;
6187 : /* no need to compare eqop, since it is fully determined by sortop */
6188 : }
6189 :
6190 102 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6191 6 : return -1;
6192 96 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6193 30 : return 1;
6194 :
6195 66 : return 0;
6196 : }
6197 :
6198 : /*
6199 : * make_window_input_target
6200 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6201 : *
6202 : * When the query has window functions, this function computes the desired
6203 : * target to be computed by the node just below the first WindowAgg.
6204 : * This tlist must contain all values needed to evaluate the window functions,
6205 : * compute the final target list, and perform any required final sort step.
6206 : * If multiple WindowAggs are needed, each intermediate one adds its window
6207 : * function results onto this base tlist; only the topmost WindowAgg computes
6208 : * the actual desired target list.
6209 : *
6210 : * This function is much like make_group_input_target, though not quite enough
6211 : * like it to share code. As in that function, we flatten most expressions
6212 : * into their component variables. But we do not want to flatten window
6213 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6214 : * evaluations of them, which would be bad (possibly even resulting in
6215 : * inconsistent answers, if they contain volatile functions).
6216 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6217 : * make_group_input_target, because we may no longer have access to the
6218 : * individual Vars in them.
6219 : *
6220 : * Another key difference from make_group_input_target is that we don't
6221 : * flatten Aggref expressions, since those are to be computed below the
6222 : * window functions and just referenced like Vars above that.
6223 : *
6224 : * 'final_target' is the query's final target list (in PathTarget form)
6225 : * 'activeWindows' is the list of active windows previously identified by
6226 : * select_active_windows.
6227 : *
6228 : * The result is the PathTarget to be computed by the plan node immediately
6229 : * below the first WindowAgg node.
6230 : */
6231 : static PathTarget *
6232 2676 : make_window_input_target(PlannerInfo *root,
6233 : PathTarget *final_target,
6234 : List *activeWindows)
6235 : {
6236 : PathTarget *input_target;
6237 : Bitmapset *sgrefs;
6238 : List *flattenable_cols;
6239 : List *flattenable_vars;
6240 : int i;
6241 : ListCell *lc;
6242 :
6243 : Assert(root->parse->hasWindowFuncs);
6244 :
6245 : /*
6246 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6247 : * into a bitmapset for convenient reference below.
6248 : */
6249 2676 : sgrefs = NULL;
6250 5538 : foreach(lc, activeWindows)
6251 : {
6252 2862 : WindowClause *wc = lfirst_node(WindowClause, lc);
6253 : ListCell *lc2;
6254 :
6255 3618 : foreach(lc2, wc->partitionClause)
6256 : {
6257 756 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6258 :
6259 756 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6260 : }
6261 5134 : foreach(lc2, wc->orderClause)
6262 : {
6263 2272 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6264 :
6265 2272 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6266 : }
6267 : }
6268 :
6269 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6270 2868 : foreach(lc, root->processed_groupClause)
6271 : {
6272 192 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6273 :
6274 192 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6275 : }
6276 :
6277 : /*
6278 : * Construct a target containing all the non-flattenable targetlist items,
6279 : * and save aside the others for a moment.
6280 : */
6281 2676 : input_target = create_empty_pathtarget();
6282 2676 : flattenable_cols = NIL;
6283 :
6284 2676 : i = 0;
6285 11228 : foreach(lc, final_target->exprs)
6286 : {
6287 8552 : Expr *expr = (Expr *) lfirst(lc);
6288 8552 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6289 :
6290 : /*
6291 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6292 : * that such items can't contain window functions, so it's okay to
6293 : * compute them below the WindowAgg nodes.)
6294 : */
6295 8552 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6296 : {
6297 : /*
6298 : * Don't want to deconstruct this value, so add it to the input
6299 : * target as-is.
6300 : */
6301 2864 : add_column_to_pathtarget(input_target, expr, sgref);
6302 : }
6303 : else
6304 : {
6305 : /*
6306 : * Column is to be flattened, so just remember the expression for
6307 : * later call to pull_var_clause.
6308 : */
6309 5688 : flattenable_cols = lappend(flattenable_cols, expr);
6310 : }
6311 :
6312 8552 : i++;
6313 : }
6314 :
6315 : /*
6316 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6317 : * add them to the input target if not already present. (Some might be
6318 : * there already because they're used directly as window/group clauses.)
6319 : *
6320 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6321 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6322 : * at higher levels. On the other hand, we should recurse into
6323 : * WindowFuncs to make sure their input expressions are available.
6324 : */
6325 2676 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6326 : PVC_INCLUDE_AGGREGATES |
6327 : PVC_RECURSE_WINDOWFUNCS |
6328 : PVC_INCLUDE_PLACEHOLDERS);
6329 2676 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6330 :
6331 : /* clean up cruft */
6332 2676 : list_free(flattenable_vars);
6333 2676 : list_free(flattenable_cols);
6334 :
6335 : /* XXX this causes some redundant cost calculation ... */
6336 2676 : return set_pathtarget_cost_width(root, input_target);
6337 : }
6338 :
6339 : /*
6340 : * make_pathkeys_for_window
6341 : * Create a pathkeys list describing the required input ordering
6342 : * for the given WindowClause.
6343 : *
6344 : * Modifies wc's partitionClause to remove any clauses which are deemed
6345 : * redundant by the pathkey logic.
6346 : *
6347 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6348 : * In the future we might try to implement windowing using hashing, in which
6349 : * case the ordering could be relaxed, but for now we always sort.
6350 : */
6351 : static List *
6352 5752 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6353 : List *tlist)
6354 : {
6355 5752 : List *window_pathkeys = NIL;
6356 :
6357 : /* Throw error if can't sort */
6358 5752 : if (!grouping_is_sortable(wc->partitionClause))
6359 0 : ereport(ERROR,
6360 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6361 : errmsg("could not implement window PARTITION BY"),
6362 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6363 5752 : if (!grouping_is_sortable(wc->orderClause))
6364 0 : ereport(ERROR,
6365 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6366 : errmsg("could not implement window ORDER BY"),
6367 : errdetail("Window ordering columns must be of sortable datatypes.")));
6368 :
6369 : /*
6370 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6371 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6372 : */
6373 5752 : if (wc->partitionClause != NIL)
6374 : {
6375 : bool sortable;
6376 :
6377 1320 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6378 : &wc->partitionClause,
6379 : tlist,
6380 : true,
6381 : false,
6382 : &sortable,
6383 : false);
6384 :
6385 : Assert(sortable);
6386 : }
6387 :
6388 : /*
6389 : * In principle, we could also consider removing redundant ORDER BY items
6390 : * too as doing so does not alter the result of peer row checks done by
6391 : * the executor. However, we must *not* remove the ordering column for
6392 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6393 : * if it's known to be equal to some partitioning column.
6394 : */
6395 5752 : if (wc->orderClause != NIL)
6396 : {
6397 : List *orderby_pathkeys;
6398 :
6399 4450 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6400 : wc->orderClause,
6401 : tlist);
6402 :
6403 : /* Okay, make the combined pathkeys */
6404 4450 : if (window_pathkeys != NIL)
6405 946 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6406 : else
6407 3504 : window_pathkeys = orderby_pathkeys;
6408 : }
6409 :
6410 5752 : return window_pathkeys;
6411 : }
6412 :
6413 : /*
6414 : * make_sort_input_target
6415 : * Generate appropriate PathTarget for initial input to Sort step.
6416 : *
6417 : * If the query has ORDER BY, this function chooses the target to be computed
6418 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6419 : * project) steps. This might or might not be identical to the query's final
6420 : * output target.
6421 : *
6422 : * The main argument for keeping the sort-input tlist the same as the final
6423 : * is that we avoid a separate projection node (which will be needed if
6424 : * they're different, because Sort can't project). However, there are also
6425 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6426 : * a consistent order of evaluation for any volatile functions in the tlist,
6427 : * and if there's also a LIMIT, we can stop the query without ever computing
6428 : * tlist functions for later rows, which is beneficial for both volatile and
6429 : * expensive functions.
6430 : *
6431 : * Our current policy is to postpone volatile expressions till after the sort
6432 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6433 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6434 : * postpone set-returning expressions, because running them beforehand would
6435 : * bloat the sort dataset, and because it might cause unexpected output order
6436 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6437 : * in the tlist should be evaluated at the same plan step, so that they can
6438 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6439 : * mustn't postpone any SRFs. (Note that in principle that policy should
6440 : * probably get applied to the group/window input targetlists too, but we
6441 : * have not done that historically.) Lastly, expensive expressions are
6442 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6443 : * partial evaluation of the query is possible (if neither is true, we expect
6444 : * to have to evaluate the expressions for every row anyway), or if there are
6445 : * any volatile or set-returning expressions (since once we've put in a
6446 : * projection at all, it won't cost any more to postpone more stuff).
6447 : *
6448 : * Another issue that could potentially be considered here is that
6449 : * evaluating tlist expressions could result in data that's either wider
6450 : * or narrower than the input Vars, thus changing the volume of data that
6451 : * has to go through the Sort. However, we usually have only a very bad
6452 : * idea of the output width of any expression more complex than a Var,
6453 : * so for now it seems too risky to try to optimize on that basis.
6454 : *
6455 : * Note that if we do produce a modified sort-input target, and then the
6456 : * query ends up not using an explicit Sort, no particular harm is done:
6457 : * we'll initially use the modified target for the preceding path nodes,
6458 : * but then change them to the final target with apply_projection_to_path.
6459 : * Moreover, in such a case the guarantees about evaluation order of
6460 : * volatile functions still hold, since the rows are sorted already.
6461 : *
6462 : * This function has some things in common with make_group_input_target and
6463 : * make_window_input_target, though the detailed rules for what to do are
6464 : * different. We never flatten/postpone any grouping or ordering columns;
6465 : * those are needed before the sort. If we do flatten a particular
6466 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6467 : * computed earlier.
6468 : *
6469 : * 'final_target' is the query's final target list (in PathTarget form)
6470 : * 'have_postponed_srfs' is an output argument, see below
6471 : *
6472 : * The result is the PathTarget to be computed by the plan node immediately
6473 : * below the Sort step (and the Distinct step, if any). This will be
6474 : * exactly final_target if we decide a projection step wouldn't be helpful.
6475 : *
6476 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6477 : * any set-returning functions to after the Sort.
6478 : */
6479 : static PathTarget *
6480 72102 : make_sort_input_target(PlannerInfo *root,
6481 : PathTarget *final_target,
6482 : bool *have_postponed_srfs)
6483 : {
6484 72102 : Query *parse = root->parse;
6485 : PathTarget *input_target;
6486 : int ncols;
6487 : bool *col_is_srf;
6488 : bool *postpone_col;
6489 : bool have_srf;
6490 : bool have_volatile;
6491 : bool have_expensive;
6492 : bool have_srf_sortcols;
6493 : bool postpone_srfs;
6494 : List *postponable_cols;
6495 : List *postponable_vars;
6496 : int i;
6497 : ListCell *lc;
6498 :
6499 : /* Shouldn't get here unless query has ORDER BY */
6500 : Assert(parse->sortClause);
6501 :
6502 72102 : *have_postponed_srfs = false; /* default result */
6503 :
6504 : /* Inspect tlist and collect per-column information */
6505 72102 : ncols = list_length(final_target->exprs);
6506 72102 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6507 72102 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6508 72102 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6509 :
6510 72102 : i = 0;
6511 435140 : foreach(lc, final_target->exprs)
6512 : {
6513 363038 : Expr *expr = (Expr *) lfirst(lc);
6514 :
6515 : /*
6516 : * If the column has a sortgroupref, assume it has to be evaluated
6517 : * before sorting. Generally such columns would be ORDER BY, GROUP
6518 : * BY, etc targets. One exception is columns that were removed from
6519 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6520 : * only be Vars anyway. There don't seem to be any cases where it
6521 : * would be worth the trouble to double-check.
6522 : */
6523 363038 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6524 : {
6525 : /*
6526 : * Check for SRF or volatile functions. Check the SRF case first
6527 : * because we must know whether we have any postponed SRFs.
6528 : */
6529 261636 : if (parse->hasTargetSRFs &&
6530 216 : expression_returns_set((Node *) expr))
6531 : {
6532 : /* We'll decide below whether these are postponable */
6533 96 : col_is_srf[i] = true;
6534 96 : have_srf = true;
6535 : }
6536 261324 : else if (contain_volatile_functions((Node *) expr))
6537 : {
6538 : /* Unconditionally postpone */
6539 248 : postpone_col[i] = true;
6540 248 : have_volatile = true;
6541 : }
6542 : else
6543 : {
6544 : /*
6545 : * Else check the cost. XXX it's annoying to have to do this
6546 : * when set_pathtarget_cost_width() just did it. Refactor to
6547 : * allow sharing the work?
6548 : */
6549 : QualCost cost;
6550 :
6551 261076 : cost_qual_eval_node(&cost, (Node *) expr, root);
6552 :
6553 : /*
6554 : * We arbitrarily define "expensive" as "more than 10X
6555 : * cpu_operator_cost". Note this will take in any PL function
6556 : * with default cost.
6557 : */
6558 261076 : if (cost.per_tuple > 10 * cpu_operator_cost)
6559 : {
6560 16882 : postpone_col[i] = true;
6561 16882 : have_expensive = true;
6562 : }
6563 : }
6564 : }
6565 : else
6566 : {
6567 : /* For sortgroupref cols, just check if any contain SRFs */
6568 101618 : if (!have_srf_sortcols &&
6569 101928 : parse->hasTargetSRFs &&
6570 334 : expression_returns_set((Node *) expr))
6571 148 : have_srf_sortcols = true;
6572 : }
6573 :
6574 363038 : i++;
6575 : }
6576 :
6577 : /*
6578 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6579 : */
6580 72102 : postpone_srfs = (have_srf && !have_srf_sortcols);
6581 :
6582 : /*
6583 : * If we don't need a post-sort projection, just return final_target.
6584 : */
6585 72102 : if (!(postpone_srfs || have_volatile ||
6586 71798 : (have_expensive &&
6587 9922 : (parse->limitCount || root->tuple_fraction > 0))))
6588 71762 : return final_target;
6589 :
6590 : /*
6591 : * Report whether the post-sort projection will contain set-returning
6592 : * functions. This is important because it affects whether the Sort can
6593 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6594 : * to return.
6595 : */
6596 340 : *have_postponed_srfs = postpone_srfs;
6597 :
6598 : /*
6599 : * Construct the sort-input target, taking all non-postponable columns and
6600 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6601 : * the postponable ones.
6602 : */
6603 340 : input_target = create_empty_pathtarget();
6604 340 : postponable_cols = NIL;
6605 :
6606 340 : i = 0;
6607 2290 : foreach(lc, final_target->exprs)
6608 : {
6609 1950 : Expr *expr = (Expr *) lfirst(lc);
6610 :
6611 1950 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6612 398 : postponable_cols = lappend(postponable_cols, expr);
6613 : else
6614 1552 : add_column_to_pathtarget(input_target, expr,
6615 1552 : get_pathtarget_sortgroupref(final_target, i));
6616 :
6617 1950 : i++;
6618 : }
6619 :
6620 : /*
6621 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6622 : * postponable columns, and add them to the sort-input target if not
6623 : * already present. (Some might be there already.) We mustn't
6624 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6625 : * would be unable to recompute them.
6626 : */
6627 340 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6628 : PVC_INCLUDE_AGGREGATES |
6629 : PVC_INCLUDE_WINDOWFUNCS |
6630 : PVC_INCLUDE_PLACEHOLDERS);
6631 340 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6632 :
6633 : /* clean up cruft */
6634 340 : list_free(postponable_vars);
6635 340 : list_free(postponable_cols);
6636 :
6637 : /* XXX this represents even more redundant cost calculation ... */
6638 340 : return set_pathtarget_cost_width(root, input_target);
6639 : }
6640 :
6641 : /*
6642 : * get_cheapest_fractional_path
6643 : * Find the cheapest path for retrieving a specified fraction of all
6644 : * the tuples expected to be returned by the given relation.
6645 : *
6646 : * Do not consider parameterized paths. If the caller needs a path for upper
6647 : * rel, it can't have parameterized paths. If the caller needs an append
6648 : * subpath, it could become limited by the treatment of similar
6649 : * parameterization of all the subpaths.
6650 : *
6651 : * We interpret tuple_fraction the same way as grouping_planner.
6652 : *
6653 : * We assume set_cheapest() has been run on the given rel.
6654 : */
6655 : Path *
6656 507642 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6657 : {
6658 507642 : Path *best_path = rel->cheapest_total_path;
6659 : ListCell *l;
6660 :
6661 : /* If all tuples will be retrieved, just return the cheapest-total path */
6662 507642 : if (tuple_fraction <= 0.0)
6663 498078 : return best_path;
6664 :
6665 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6666 9564 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6667 3966 : tuple_fraction /= best_path->rows;
6668 :
6669 24720 : foreach(l, rel->pathlist)
6670 : {
6671 15156 : Path *path = (Path *) lfirst(l);
6672 :
6673 15156 : if (path->param_info)
6674 200 : continue;
6675 :
6676 20348 : if (path == rel->cheapest_total_path ||
6677 5392 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6678 14432 : continue;
6679 :
6680 524 : best_path = path;
6681 : }
6682 :
6683 9564 : return best_path;
6684 : }
6685 :
6686 : /*
6687 : * adjust_paths_for_srfs
6688 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6689 : *
6690 : * The executor can only handle set-returning functions that appear at the
6691 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6692 : * that are not at top level, we need to split up the evaluation into multiple
6693 : * plan levels in which each level satisfies this constraint. This function
6694 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6695 : * output tlist to insert appropriate projection steps.
6696 : *
6697 : * The given targets and targets_contain_srfs lists are from
6698 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6699 : * target in targets.
6700 : */
6701 : static void
6702 12792 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6703 : List *targets, List *targets_contain_srfs)
6704 : {
6705 : ListCell *lc;
6706 :
6707 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6708 : Assert(!linitial_int(targets_contain_srfs));
6709 :
6710 : /* If no SRFs appear at this plan level, nothing to do */
6711 12792 : if (list_length(targets) == 1)
6712 708 : return;
6713 :
6714 : /*
6715 : * Stack SRF-evaluation nodes atop each path for the rel.
6716 : *
6717 : * In principle we should re-run set_cheapest() here to identify the
6718 : * cheapest path, but it seems unlikely that adding the same tlist eval
6719 : * costs to all the paths would change that, so we don't bother. Instead,
6720 : * just assume that the cheapest-startup and cheapest-total paths remain
6721 : * so. (There should be no parameterized paths anymore, so we needn't
6722 : * worry about updating cheapest_parameterized_paths.)
6723 : */
6724 24206 : foreach(lc, rel->pathlist)
6725 : {
6726 12122 : Path *subpath = (Path *) lfirst(lc);
6727 12122 : Path *newpath = subpath;
6728 : ListCell *lc1,
6729 : *lc2;
6730 :
6731 : Assert(subpath->param_info == NULL);
6732 37564 : forboth(lc1, targets, lc2, targets_contain_srfs)
6733 : {
6734 25442 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6735 25442 : bool contains_srfs = (bool) lfirst_int(lc2);
6736 :
6737 : /* If this level doesn't contain SRFs, do regular projection */
6738 25442 : if (contains_srfs)
6739 12182 : newpath = (Path *) create_set_projection_path(root,
6740 : rel,
6741 : newpath,
6742 : thistarget);
6743 : else
6744 13260 : newpath = (Path *) apply_projection_to_path(root,
6745 : rel,
6746 : newpath,
6747 : thistarget);
6748 : }
6749 12122 : lfirst(lc) = newpath;
6750 12122 : if (subpath == rel->cheapest_startup_path)
6751 404 : rel->cheapest_startup_path = newpath;
6752 12122 : if (subpath == rel->cheapest_total_path)
6753 404 : rel->cheapest_total_path = newpath;
6754 : }
6755 :
6756 : /* Likewise for partial paths, if any */
6757 12102 : foreach(lc, rel->partial_pathlist)
6758 : {
6759 18 : Path *subpath = (Path *) lfirst(lc);
6760 18 : Path *newpath = subpath;
6761 : ListCell *lc1,
6762 : *lc2;
6763 :
6764 : Assert(subpath->param_info == NULL);
6765 72 : forboth(lc1, targets, lc2, targets_contain_srfs)
6766 : {
6767 54 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6768 54 : bool contains_srfs = (bool) lfirst_int(lc2);
6769 :
6770 : /* If this level doesn't contain SRFs, do regular projection */
6771 54 : if (contains_srfs)
6772 18 : newpath = (Path *) create_set_projection_path(root,
6773 : rel,
6774 : newpath,
6775 : thistarget);
6776 : else
6777 : {
6778 : /* avoid apply_projection_to_path, in case of multiple refs */
6779 36 : newpath = (Path *) create_projection_path(root,
6780 : rel,
6781 : newpath,
6782 : thistarget);
6783 : }
6784 : }
6785 18 : lfirst(lc) = newpath;
6786 : }
6787 : }
6788 :
6789 : /*
6790 : * expression_planner
6791 : * Perform planner's transformations on a standalone expression.
6792 : *
6793 : * Various utility commands need to evaluate expressions that are not part
6794 : * of a plannable query. They can do so using the executor's regular
6795 : * expression-execution machinery, but first the expression has to be fed
6796 : * through here to transform it from parser output to something executable.
6797 : *
6798 : * Currently, we disallow sublinks in standalone expressions, so there's no
6799 : * real "planning" involved here. (That might not always be true though.)
6800 : * What we must do is run eval_const_expressions to ensure that any function
6801 : * calls are converted to positional notation and function default arguments
6802 : * get inserted. The fact that constant subexpressions get simplified is a
6803 : * side-effect that is useful when the expression will get evaluated more than
6804 : * once. Also, we must fix operator function IDs.
6805 : *
6806 : * This does not return any information about dependencies of the expression.
6807 : * Hence callers should use the results only for the duration of the current
6808 : * query. Callers that would like to cache the results for longer should use
6809 : * expression_planner_with_deps, probably via the plancache.
6810 : *
6811 : * Note: this must not make any damaging changes to the passed-in expression
6812 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6813 : * we first do an expression_tree_mutator-based walk, what is returned will
6814 : * be a new node tree.) The result is constructed in the current memory
6815 : * context; beware that this can leak a lot of additional stuff there, too.
6816 : */
6817 : Expr *
6818 246478 : expression_planner(Expr *expr)
6819 : {
6820 : Node *result;
6821 :
6822 : /*
6823 : * Convert named-argument function calls, insert default arguments and
6824 : * simplify constant subexprs
6825 : */
6826 246478 : result = eval_const_expressions(NULL, (Node *) expr);
6827 :
6828 : /* Fill in opfuncid values if missing */
6829 246460 : fix_opfuncids(result);
6830 :
6831 246460 : return (Expr *) result;
6832 : }
6833 :
6834 : /*
6835 : * expression_planner_with_deps
6836 : * Perform planner's transformations on a standalone expression,
6837 : * returning expression dependency information along with the result.
6838 : *
6839 : * This is identical to expression_planner() except that it also returns
6840 : * information about possible dependencies of the expression, ie identities of
6841 : * objects whose definitions affect the result. As in a PlannedStmt, these
6842 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6843 : */
6844 : Expr *
6845 380 : expression_planner_with_deps(Expr *expr,
6846 : List **relationOids,
6847 : List **invalItems)
6848 : {
6849 : Node *result;
6850 : PlannerGlobal glob;
6851 : PlannerInfo root;
6852 :
6853 : /* Make up dummy planner state so we can use setrefs machinery */
6854 10260 : MemSet(&glob, 0, sizeof(glob));
6855 380 : glob.type = T_PlannerGlobal;
6856 380 : glob.relationOids = NIL;
6857 380 : glob.invalItems = NIL;
6858 :
6859 35340 : MemSet(&root, 0, sizeof(root));
6860 380 : root.type = T_PlannerInfo;
6861 380 : root.glob = &glob;
6862 :
6863 : /*
6864 : * Convert named-argument function calls, insert default arguments and
6865 : * simplify constant subexprs. Collect identities of inlined functions
6866 : * and elided domains, too.
6867 : */
6868 380 : result = eval_const_expressions(&root, (Node *) expr);
6869 :
6870 : /* Fill in opfuncid values if missing */
6871 380 : fix_opfuncids(result);
6872 :
6873 : /*
6874 : * Now walk the finished expression to find anything else we ought to
6875 : * record as an expression dependency.
6876 : */
6877 380 : (void) extract_query_dependencies_walker(result, &root);
6878 :
6879 380 : *relationOids = glob.relationOids;
6880 380 : *invalItems = glob.invalItems;
6881 :
6882 380 : return (Expr *) result;
6883 : }
6884 :
6885 :
6886 : /*
6887 : * plan_cluster_use_sort
6888 : * Use the planner to decide how CLUSTER should implement sorting
6889 : *
6890 : * tableOid is the OID of a table to be clustered on its index indexOid
6891 : * (which is already known to be a btree index). Decide whether it's
6892 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6893 : * Return true to use sorting, false to use an indexscan.
6894 : *
6895 : * Note: caller had better already hold some type of lock on the table.
6896 : */
6897 : bool
6898 188 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6899 : {
6900 : PlannerInfo *root;
6901 : Query *query;
6902 : PlannerGlobal *glob;
6903 : RangeTblEntry *rte;
6904 : RelOptInfo *rel;
6905 : IndexOptInfo *indexInfo;
6906 : QualCost indexExprCost;
6907 : Cost comparisonCost;
6908 : Path *seqScanPath;
6909 : Path seqScanAndSortPath;
6910 : IndexPath *indexScanPath;
6911 : ListCell *lc;
6912 :
6913 : /* We can short-circuit the cost comparison if indexscans are disabled */
6914 188 : if (!enable_indexscan)
6915 30 : return true; /* use sort */
6916 :
6917 : /* Set up mostly-dummy planner state */
6918 158 : query = makeNode(Query);
6919 158 : query->commandType = CMD_SELECT;
6920 :
6921 158 : glob = makeNode(PlannerGlobal);
6922 :
6923 158 : root = makeNode(PlannerInfo);
6924 158 : root->parse = query;
6925 158 : root->glob = glob;
6926 158 : root->query_level = 1;
6927 158 : root->planner_cxt = CurrentMemoryContext;
6928 158 : root->wt_param_id = -1;
6929 158 : root->join_domains = list_make1(makeNode(JoinDomain));
6930 :
6931 : /* Build a minimal RTE for the rel */
6932 158 : rte = makeNode(RangeTblEntry);
6933 158 : rte->rtekind = RTE_RELATION;
6934 158 : rte->relid = tableOid;
6935 158 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6936 158 : rte->rellockmode = AccessShareLock;
6937 158 : rte->lateral = false;
6938 158 : rte->inh = false;
6939 158 : rte->inFromCl = true;
6940 158 : query->rtable = list_make1(rte);
6941 158 : addRTEPermissionInfo(&query->rteperminfos, rte);
6942 :
6943 : /* Set up RTE/RelOptInfo arrays */
6944 158 : setup_simple_rel_arrays(root);
6945 :
6946 : /* Build RelOptInfo */
6947 158 : rel = build_simple_rel(root, 1, NULL);
6948 :
6949 : /* Locate IndexOptInfo for the target index */
6950 158 : indexInfo = NULL;
6951 196 : foreach(lc, rel->indexlist)
6952 : {
6953 196 : indexInfo = lfirst_node(IndexOptInfo, lc);
6954 196 : if (indexInfo->indexoid == indexOid)
6955 158 : break;
6956 : }
6957 :
6958 : /*
6959 : * It's possible that get_relation_info did not generate an IndexOptInfo
6960 : * for the desired index; this could happen if it's not yet reached its
6961 : * indcheckxmin usability horizon, or if it's a system index and we're
6962 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6963 : * trust the index contents but use seqscan-and-sort.
6964 : */
6965 158 : if (lc == NULL) /* not in the list? */
6966 0 : return true; /* use sort */
6967 :
6968 : /*
6969 : * Rather than doing all the pushups that would be needed to use
6970 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6971 : */
6972 158 : rel->rows = rel->tuples;
6973 158 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6974 :
6975 158 : root->total_table_pages = rel->pages;
6976 :
6977 : /*
6978 : * Determine eval cost of the index expressions, if any. We need to
6979 : * charge twice that amount for each tuple comparison that happens during
6980 : * the sort, since tuplesort.c will have to re-evaluate the index
6981 : * expressions each time. (XXX that's pretty inefficient...)
6982 : */
6983 158 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6984 158 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6985 :
6986 : /* Estimate the cost of seq scan + sort */
6987 158 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6988 158 : cost_sort(&seqScanAndSortPath, root, NIL,
6989 : seqScanPath->disabled_nodes,
6990 158 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6991 : comparisonCost, maintenance_work_mem, -1.0);
6992 :
6993 : /* Estimate the cost of index scan */
6994 158 : indexScanPath = create_index_path(root, indexInfo,
6995 : NIL, NIL, NIL, NIL,
6996 : ForwardScanDirection, false,
6997 : NULL, 1.0, false);
6998 :
6999 158 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
7000 : }
7001 :
7002 : /*
7003 : * plan_create_index_workers
7004 : * Use the planner to decide how many parallel worker processes
7005 : * CREATE INDEX should request for use
7006 : *
7007 : * tableOid is the table on which the index is to be built. indexOid is the
7008 : * OID of an index to be created or reindexed (which must be an index with
7009 : * support for parallel builds - currently btree, GIN, or BRIN).
7010 : *
7011 : * Return value is the number of parallel worker processes to request. It
7012 : * may be unsafe to proceed if this is 0. Note that this does not include the
7013 : * leader participating as a worker (value is always a number of parallel
7014 : * worker processes).
7015 : *
7016 : * Note: caller had better already hold some type of lock on the table and
7017 : * index.
7018 : */
7019 : int
7020 36562 : plan_create_index_workers(Oid tableOid, Oid indexOid)
7021 : {
7022 : PlannerInfo *root;
7023 : Query *query;
7024 : PlannerGlobal *glob;
7025 : RangeTblEntry *rte;
7026 : Relation heap;
7027 : Relation index;
7028 : RelOptInfo *rel;
7029 : int parallel_workers;
7030 : BlockNumber heap_blocks;
7031 : double reltuples;
7032 : double allvisfrac;
7033 :
7034 : /*
7035 : * We don't allow performing parallel operation in standalone backend or
7036 : * when parallelism is disabled.
7037 : */
7038 36562 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
7039 514 : return 0;
7040 :
7041 : /* Set up largely-dummy planner state */
7042 36048 : query = makeNode(Query);
7043 36048 : query->commandType = CMD_SELECT;
7044 :
7045 36048 : glob = makeNode(PlannerGlobal);
7046 :
7047 36048 : root = makeNode(PlannerInfo);
7048 36048 : root->parse = query;
7049 36048 : root->glob = glob;
7050 36048 : root->query_level = 1;
7051 36048 : root->planner_cxt = CurrentMemoryContext;
7052 36048 : root->wt_param_id = -1;
7053 36048 : root->join_domains = list_make1(makeNode(JoinDomain));
7054 :
7055 : /*
7056 : * Build a minimal RTE.
7057 : *
7058 : * Mark the RTE with inh = true. This is a kludge to prevent
7059 : * get_relation_info() from fetching index info, which is necessary
7060 : * because it does not expect that any IndexOptInfo is currently
7061 : * undergoing REINDEX.
7062 : */
7063 36048 : rte = makeNode(RangeTblEntry);
7064 36048 : rte->rtekind = RTE_RELATION;
7065 36048 : rte->relid = tableOid;
7066 36048 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7067 36048 : rte->rellockmode = AccessShareLock;
7068 36048 : rte->lateral = false;
7069 36048 : rte->inh = true;
7070 36048 : rte->inFromCl = true;
7071 36048 : query->rtable = list_make1(rte);
7072 36048 : addRTEPermissionInfo(&query->rteperminfos, rte);
7073 :
7074 : /* Set up RTE/RelOptInfo arrays */
7075 36048 : setup_simple_rel_arrays(root);
7076 :
7077 : /* Build RelOptInfo */
7078 36048 : rel = build_simple_rel(root, 1, NULL);
7079 :
7080 : /* Rels are assumed already locked by the caller */
7081 36048 : heap = table_open(tableOid, NoLock);
7082 36048 : index = index_open(indexOid, NoLock);
7083 :
7084 : /*
7085 : * Determine if it's safe to proceed.
7086 : *
7087 : * Currently, parallel workers can't access the leader's temporary tables.
7088 : * Furthermore, any index predicate or index expressions must be parallel
7089 : * safe.
7090 : */
7091 36048 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7092 33968 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7093 33828 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7094 : {
7095 2220 : parallel_workers = 0;
7096 2220 : goto done;
7097 : }
7098 :
7099 : /*
7100 : * If parallel_workers storage parameter is set for the table, accept that
7101 : * as the number of parallel worker processes to launch (though still cap
7102 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7103 : * consider any other factor when parallel_workers is set. (e.g., memory
7104 : * use by workers.)
7105 : */
7106 33828 : if (rel->rel_parallel_workers != -1)
7107 : {
7108 100 : parallel_workers = Min(rel->rel_parallel_workers,
7109 : max_parallel_maintenance_workers);
7110 100 : goto done;
7111 : }
7112 :
7113 : /*
7114 : * Estimate heap relation size ourselves, since rel->pages cannot be
7115 : * trusted (heap RTE was marked as inheritance parent)
7116 : */
7117 33728 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7118 :
7119 : /*
7120 : * Determine number of workers to scan the heap relation using generic
7121 : * model
7122 : */
7123 33728 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7124 : max_parallel_maintenance_workers);
7125 :
7126 : /*
7127 : * Cap workers based on available maintenance_work_mem as needed.
7128 : *
7129 : * Note that each tuplesort participant receives an even share of the
7130 : * total maintenance_work_mem budget. Aim to leave participants
7131 : * (including the leader as a participant) with no less than 32MB of
7132 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7133 : * immediately past the threshold of being capable of launching a single
7134 : * parallel worker to sort.
7135 : */
7136 33890 : while (parallel_workers > 0 &&
7137 326 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7138 162 : parallel_workers--;
7139 :
7140 33728 : done:
7141 36048 : index_close(index, NoLock);
7142 36048 : table_close(heap, NoLock);
7143 :
7144 36048 : return parallel_workers;
7145 : }
7146 :
7147 : /*
7148 : * add_paths_to_grouping_rel
7149 : *
7150 : * Add non-partial paths to grouping relation.
7151 : */
7152 : static void
7153 47234 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7154 : RelOptInfo *grouped_rel,
7155 : RelOptInfo *partially_grouped_rel,
7156 : const AggClauseCosts *agg_costs,
7157 : grouping_sets_data *gd,
7158 : GroupPathExtraData *extra)
7159 : {
7160 47234 : Query *parse = root->parse;
7161 47234 : Path *cheapest_path = input_rel->cheapest_total_path;
7162 47234 : Path *cheapest_partially_grouped_path = NULL;
7163 : ListCell *lc;
7164 47234 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7165 47234 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7166 47234 : List *havingQual = (List *) extra->havingQual;
7167 47234 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7168 47234 : double dNumGroups = 0;
7169 47234 : double dNumFinalGroups = 0;
7170 :
7171 : /*
7172 : * Estimate number of groups for non-split aggregation.
7173 : */
7174 47234 : dNumGroups = get_number_of_groups(root,
7175 : cheapest_path->rows,
7176 : gd,
7177 : extra->targetList);
7178 :
7179 47234 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7180 : {
7181 3000 : cheapest_partially_grouped_path =
7182 : partially_grouped_rel->cheapest_total_path;
7183 :
7184 : /*
7185 : * Estimate number of groups for final phase of partial aggregation.
7186 : */
7187 : dNumFinalGroups =
7188 3000 : get_number_of_groups(root,
7189 : cheapest_partially_grouped_path->rows,
7190 : gd,
7191 : extra->targetList);
7192 : }
7193 :
7194 47234 : if (can_sort)
7195 : {
7196 : /*
7197 : * Use any available suitably-sorted path as input, and also consider
7198 : * sorting the cheapest-total path and incremental sort on any paths
7199 : * with presorted keys.
7200 : */
7201 97886 : foreach(lc, input_rel->pathlist)
7202 : {
7203 : ListCell *lc2;
7204 50658 : Path *path = (Path *) lfirst(lc);
7205 50658 : Path *path_save = path;
7206 50658 : List *pathkey_orderings = NIL;
7207 :
7208 : /* generate alternative group orderings that might be useful */
7209 50658 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7210 :
7211 : Assert(list_length(pathkey_orderings) > 0);
7212 :
7213 101460 : foreach(lc2, pathkey_orderings)
7214 : {
7215 50802 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7216 :
7217 : /* restore the path (we replace it in the loop) */
7218 50802 : path = path_save;
7219 :
7220 50802 : path = make_ordered_path(root,
7221 : grouped_rel,
7222 : path,
7223 : cheapest_path,
7224 : info->pathkeys,
7225 : -1.0);
7226 50802 : if (path == NULL)
7227 386 : continue;
7228 :
7229 : /* Now decide what to stick atop it */
7230 50416 : if (parse->groupingSets)
7231 : {
7232 1062 : consider_groupingsets_paths(root, grouped_rel,
7233 : path, true, can_hash,
7234 : gd, agg_costs, dNumGroups);
7235 : }
7236 49354 : else if (parse->hasAggs)
7237 : {
7238 : /*
7239 : * We have aggregation, possibly with plain GROUP BY. Make
7240 : * an AggPath.
7241 : */
7242 48570 : add_path(grouped_rel, (Path *)
7243 48570 : create_agg_path(root,
7244 : grouped_rel,
7245 : path,
7246 48570 : grouped_rel->reltarget,
7247 48570 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7248 : AGGSPLIT_SIMPLE,
7249 : info->clauses,
7250 : havingQual,
7251 : agg_costs,
7252 : dNumGroups));
7253 : }
7254 784 : else if (parse->groupClause)
7255 : {
7256 : /*
7257 : * We have GROUP BY without aggregation or grouping sets.
7258 : * Make a GroupPath.
7259 : */
7260 784 : add_path(grouped_rel, (Path *)
7261 784 : create_group_path(root,
7262 : grouped_rel,
7263 : path,
7264 : info->clauses,
7265 : havingQual,
7266 : dNumGroups));
7267 : }
7268 : else
7269 : {
7270 : /* Other cases should have been handled above */
7271 : Assert(false);
7272 : }
7273 : }
7274 : }
7275 :
7276 : /*
7277 : * Instead of operating directly on the input relation, we can
7278 : * consider finalizing a partially aggregated path.
7279 : */
7280 47228 : if (partially_grouped_rel != NULL)
7281 : {
7282 7682 : foreach(lc, partially_grouped_rel->pathlist)
7283 : {
7284 : ListCell *lc2;
7285 4682 : Path *path = (Path *) lfirst(lc);
7286 4682 : Path *path_save = path;
7287 4682 : List *pathkey_orderings = NIL;
7288 :
7289 : /* generate alternative group orderings that might be useful */
7290 4682 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7291 :
7292 : Assert(list_length(pathkey_orderings) > 0);
7293 :
7294 : /* process all potentially interesting grouping reorderings */
7295 9364 : foreach(lc2, pathkey_orderings)
7296 : {
7297 4682 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7298 :
7299 : /* restore the path (we replace it in the loop) */
7300 4682 : path = path_save;
7301 :
7302 4682 : path = make_ordered_path(root,
7303 : grouped_rel,
7304 : path,
7305 : cheapest_partially_grouped_path,
7306 : info->pathkeys,
7307 : -1.0);
7308 :
7309 4682 : if (path == NULL)
7310 204 : continue;
7311 :
7312 4478 : if (parse->hasAggs)
7313 4230 : add_path(grouped_rel, (Path *)
7314 4230 : create_agg_path(root,
7315 : grouped_rel,
7316 : path,
7317 4230 : grouped_rel->reltarget,
7318 4230 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7319 : AGGSPLIT_FINAL_DESERIAL,
7320 : info->clauses,
7321 : havingQual,
7322 : agg_final_costs,
7323 : dNumFinalGroups));
7324 : else
7325 248 : add_path(grouped_rel, (Path *)
7326 248 : create_group_path(root,
7327 : grouped_rel,
7328 : path,
7329 : info->clauses,
7330 : havingQual,
7331 : dNumFinalGroups));
7332 :
7333 : }
7334 : }
7335 : }
7336 : }
7337 :
7338 47234 : if (can_hash)
7339 : {
7340 6020 : if (parse->groupingSets)
7341 : {
7342 : /*
7343 : * Try for a hash-only groupingsets path over unsorted input.
7344 : */
7345 900 : consider_groupingsets_paths(root, grouped_rel,
7346 : cheapest_path, false, true,
7347 : gd, agg_costs, dNumGroups);
7348 : }
7349 : else
7350 : {
7351 : /*
7352 : * Generate a HashAgg Path. We just need an Agg over the
7353 : * cheapest-total input path, since input order won't matter.
7354 : */
7355 5120 : add_path(grouped_rel, (Path *)
7356 5120 : create_agg_path(root, grouped_rel,
7357 : cheapest_path,
7358 5120 : grouped_rel->reltarget,
7359 : AGG_HASHED,
7360 : AGGSPLIT_SIMPLE,
7361 : root->processed_groupClause,
7362 : havingQual,
7363 : agg_costs,
7364 : dNumGroups));
7365 : }
7366 :
7367 : /*
7368 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7369 : * grouped path, assuming there is one
7370 : */
7371 6020 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7372 : {
7373 1448 : add_path(grouped_rel, (Path *)
7374 1448 : create_agg_path(root,
7375 : grouped_rel,
7376 : cheapest_partially_grouped_path,
7377 1448 : grouped_rel->reltarget,
7378 : AGG_HASHED,
7379 : AGGSPLIT_FINAL_DESERIAL,
7380 : root->processed_groupClause,
7381 : havingQual,
7382 : agg_final_costs,
7383 : dNumFinalGroups));
7384 : }
7385 : }
7386 :
7387 : /*
7388 : * When partitionwise aggregate is used, we might have fully aggregated
7389 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7390 : * consider a path for grouped_rel consisting of a Parallel Append of
7391 : * non-partial paths from each child.
7392 : */
7393 47234 : if (grouped_rel->partial_pathlist != NIL)
7394 318 : gather_grouping_paths(root, grouped_rel);
7395 47234 : }
7396 :
7397 : /*
7398 : * create_partial_grouping_paths
7399 : *
7400 : * Create a new upper relation representing the result of partial aggregation
7401 : * and populate it with appropriate paths. Note that we don't finalize the
7402 : * lists of paths here, so the caller can add additional partial or non-partial
7403 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7404 : * the returned upper relation.
7405 : *
7406 : * All paths for this new upper relation -- both partial and non-partial --
7407 : * have been partially aggregated but require a subsequent FinalizeAggregate
7408 : * step.
7409 : *
7410 : * NB: This function is allowed to return NULL if it determines that there is
7411 : * no real need to create a new RelOptInfo.
7412 : */
7413 : static RelOptInfo *
7414 43128 : create_partial_grouping_paths(PlannerInfo *root,
7415 : RelOptInfo *grouped_rel,
7416 : RelOptInfo *input_rel,
7417 : grouping_sets_data *gd,
7418 : GroupPathExtraData *extra,
7419 : bool force_rel_creation)
7420 : {
7421 43128 : Query *parse = root->parse;
7422 : RelOptInfo *partially_grouped_rel;
7423 43128 : RelOptInfo *eager_agg_rel = NULL;
7424 43128 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7425 43128 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7426 43128 : Path *cheapest_partial_path = NULL;
7427 43128 : Path *cheapest_total_path = NULL;
7428 43128 : double dNumPartialGroups = 0;
7429 43128 : double dNumPartialPartialGroups = 0;
7430 : ListCell *lc;
7431 43128 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7432 43128 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7433 :
7434 : /*
7435 : * Check whether any partially aggregated paths have been generated
7436 : * through eager aggregation.
7437 : */
7438 43128 : if (input_rel->grouped_rel &&
7439 958 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7440 958 : input_rel->grouped_rel->pathlist != NIL)
7441 898 : eager_agg_rel = input_rel->grouped_rel;
7442 :
7443 : /*
7444 : * Consider whether we should generate partially aggregated non-partial
7445 : * paths. We can only do this if we have a non-partial path, and only if
7446 : * the parent of the input rel is performing partial partitionwise
7447 : * aggregation. (Note that extra->patype is the type of partitionwise
7448 : * aggregation being used at the parent level, not this level.)
7449 : */
7450 43128 : if (input_rel->pathlist != NIL &&
7451 43128 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7452 858 : cheapest_total_path = input_rel->cheapest_total_path;
7453 :
7454 : /*
7455 : * If parallelism is possible for grouped_rel, then we should consider
7456 : * generating partially-grouped partial paths. However, if the input rel
7457 : * has no partial paths, then we can't.
7458 : */
7459 43128 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7460 3276 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7461 :
7462 : /*
7463 : * If we can't partially aggregate partial paths, and we can't partially
7464 : * aggregate non-partial paths, and no partially aggregated paths were
7465 : * generated by eager aggregation, then don't bother creating the new
7466 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7467 : */
7468 43128 : if (cheapest_total_path == NULL &&
7469 39498 : cheapest_partial_path == NULL &&
7470 39368 : eager_agg_rel == NULL &&
7471 39368 : !force_rel_creation)
7472 39270 : return NULL;
7473 :
7474 : /*
7475 : * Build a new upper relation to represent the result of partially
7476 : * aggregating the rows from the input relation.
7477 : */
7478 3858 : partially_grouped_rel = fetch_upper_rel(root,
7479 : UPPERREL_PARTIAL_GROUP_AGG,
7480 : grouped_rel->relids);
7481 3858 : partially_grouped_rel->consider_parallel =
7482 3858 : grouped_rel->consider_parallel;
7483 3858 : partially_grouped_rel->pgs_mask = grouped_rel->pgs_mask;
7484 3858 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7485 3858 : partially_grouped_rel->serverid = grouped_rel->serverid;
7486 3858 : partially_grouped_rel->userid = grouped_rel->userid;
7487 3858 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7488 3858 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7489 :
7490 : /*
7491 : * Build target list for partial aggregate paths. These paths cannot just
7492 : * emit the same tlist as regular aggregate paths, because (1) we must
7493 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7494 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7495 : */
7496 3858 : partially_grouped_rel->reltarget =
7497 3858 : make_partial_grouping_target(root, grouped_rel->reltarget,
7498 : extra->havingQual);
7499 :
7500 3858 : if (!extra->partial_costs_set)
7501 : {
7502 : /*
7503 : * Collect statistics about aggregates for estimating costs of
7504 : * performing aggregation in parallel.
7505 : */
7506 13896 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7507 13896 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7508 2316 : if (parse->hasAggs)
7509 : {
7510 : /* partial phase */
7511 2182 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7512 : agg_partial_costs);
7513 :
7514 : /* final phase */
7515 2182 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7516 : agg_final_costs);
7517 : }
7518 :
7519 2316 : extra->partial_costs_set = true;
7520 : }
7521 :
7522 : /* Estimate number of partial groups. */
7523 3858 : if (cheapest_total_path != NULL)
7524 : dNumPartialGroups =
7525 858 : get_number_of_groups(root,
7526 : cheapest_total_path->rows,
7527 : gd,
7528 : extra->targetList);
7529 3858 : if (cheapest_partial_path != NULL)
7530 : dNumPartialPartialGroups =
7531 3276 : get_number_of_groups(root,
7532 : cheapest_partial_path->rows,
7533 : gd,
7534 : extra->targetList);
7535 :
7536 3858 : if (can_sort && cheapest_total_path != NULL)
7537 : {
7538 : /* This should have been checked previously */
7539 : Assert(parse->hasAggs || parse->groupClause);
7540 :
7541 : /*
7542 : * Use any available suitably-sorted path as input, and also consider
7543 : * sorting the cheapest partial path.
7544 : */
7545 1716 : foreach(lc, input_rel->pathlist)
7546 : {
7547 : ListCell *lc2;
7548 858 : Path *path = (Path *) lfirst(lc);
7549 858 : Path *path_save = path;
7550 858 : List *pathkey_orderings = NIL;
7551 :
7552 : /* generate alternative group orderings that might be useful */
7553 858 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7554 :
7555 : Assert(list_length(pathkey_orderings) > 0);
7556 :
7557 : /* process all potentially interesting grouping reorderings */
7558 1716 : foreach(lc2, pathkey_orderings)
7559 : {
7560 858 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7561 :
7562 : /* restore the path (we replace it in the loop) */
7563 858 : path = path_save;
7564 :
7565 858 : path = make_ordered_path(root,
7566 : partially_grouped_rel,
7567 : path,
7568 : cheapest_total_path,
7569 : info->pathkeys,
7570 : -1.0);
7571 :
7572 858 : if (path == NULL)
7573 0 : continue;
7574 :
7575 858 : if (parse->hasAggs)
7576 786 : add_path(partially_grouped_rel, (Path *)
7577 786 : create_agg_path(root,
7578 : partially_grouped_rel,
7579 : path,
7580 786 : partially_grouped_rel->reltarget,
7581 786 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7582 : AGGSPLIT_INITIAL_SERIAL,
7583 : info->clauses,
7584 : NIL,
7585 : agg_partial_costs,
7586 : dNumPartialGroups));
7587 : else
7588 72 : add_path(partially_grouped_rel, (Path *)
7589 72 : create_group_path(root,
7590 : partially_grouped_rel,
7591 : path,
7592 : info->clauses,
7593 : NIL,
7594 : dNumPartialGroups));
7595 : }
7596 : }
7597 : }
7598 :
7599 3858 : if (can_sort && cheapest_partial_path != NULL)
7600 : {
7601 : /* Similar to above logic, but for partial paths. */
7602 7074 : foreach(lc, input_rel->partial_pathlist)
7603 : {
7604 : ListCell *lc2;
7605 3798 : Path *path = (Path *) lfirst(lc);
7606 3798 : Path *path_save = path;
7607 3798 : List *pathkey_orderings = NIL;
7608 :
7609 : /* generate alternative group orderings that might be useful */
7610 3798 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7611 :
7612 : Assert(list_length(pathkey_orderings) > 0);
7613 :
7614 : /* process all potentially interesting grouping reorderings */
7615 7596 : foreach(lc2, pathkey_orderings)
7616 : {
7617 3798 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7618 :
7619 :
7620 : /* restore the path (we replace it in the loop) */
7621 3798 : path = path_save;
7622 :
7623 3798 : path = make_ordered_path(root,
7624 : partially_grouped_rel,
7625 : path,
7626 : cheapest_partial_path,
7627 : info->pathkeys,
7628 : -1.0);
7629 :
7630 3798 : if (path == NULL)
7631 6 : continue;
7632 :
7633 3792 : if (parse->hasAggs)
7634 3670 : add_partial_path(partially_grouped_rel, (Path *)
7635 3670 : create_agg_path(root,
7636 : partially_grouped_rel,
7637 : path,
7638 3670 : partially_grouped_rel->reltarget,
7639 3670 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7640 : AGGSPLIT_INITIAL_SERIAL,
7641 : info->clauses,
7642 : NIL,
7643 : agg_partial_costs,
7644 : dNumPartialPartialGroups));
7645 : else
7646 122 : add_partial_path(partially_grouped_rel, (Path *)
7647 122 : create_group_path(root,
7648 : partially_grouped_rel,
7649 : path,
7650 : info->clauses,
7651 : NIL,
7652 : dNumPartialPartialGroups));
7653 : }
7654 : }
7655 : }
7656 :
7657 : /*
7658 : * Add a partially-grouped HashAgg Path where possible
7659 : */
7660 3858 : if (can_hash && cheapest_total_path != NULL)
7661 : {
7662 : /* Checked above */
7663 : Assert(parse->hasAggs || parse->groupClause);
7664 :
7665 858 : add_path(partially_grouped_rel, (Path *)
7666 858 : create_agg_path(root,
7667 : partially_grouped_rel,
7668 : cheapest_total_path,
7669 858 : partially_grouped_rel->reltarget,
7670 : AGG_HASHED,
7671 : AGGSPLIT_INITIAL_SERIAL,
7672 : root->processed_groupClause,
7673 : NIL,
7674 : agg_partial_costs,
7675 : dNumPartialGroups));
7676 : }
7677 :
7678 : /*
7679 : * Now add a partially-grouped HashAgg partial Path where possible
7680 : */
7681 3858 : if (can_hash && cheapest_partial_path != NULL)
7682 : {
7683 1724 : add_partial_path(partially_grouped_rel, (Path *)
7684 1724 : create_agg_path(root,
7685 : partially_grouped_rel,
7686 : cheapest_partial_path,
7687 1724 : partially_grouped_rel->reltarget,
7688 : AGG_HASHED,
7689 : AGGSPLIT_INITIAL_SERIAL,
7690 : root->processed_groupClause,
7691 : NIL,
7692 : agg_partial_costs,
7693 : dNumPartialPartialGroups));
7694 : }
7695 :
7696 : /*
7697 : * Add any partially aggregated paths generated by eager aggregation to
7698 : * the new upper relation after applying projection steps as needed.
7699 : */
7700 3858 : if (eager_agg_rel)
7701 : {
7702 : /* Add the paths */
7703 2348 : foreach(lc, eager_agg_rel->pathlist)
7704 : {
7705 1450 : Path *path = (Path *) lfirst(lc);
7706 :
7707 : /* Shouldn't have any parameterized paths anymore */
7708 : Assert(path->param_info == NULL);
7709 :
7710 1450 : path = (Path *) create_projection_path(root,
7711 : partially_grouped_rel,
7712 : path,
7713 1450 : partially_grouped_rel->reltarget);
7714 :
7715 1450 : add_path(partially_grouped_rel, path);
7716 : }
7717 :
7718 : /*
7719 : * Likewise add the partial paths, but only if parallelism is possible
7720 : * for partially_grouped_rel.
7721 : */
7722 898 : if (partially_grouped_rel->consider_parallel)
7723 : {
7724 2028 : foreach(lc, eager_agg_rel->partial_pathlist)
7725 : {
7726 1212 : Path *path = (Path *) lfirst(lc);
7727 :
7728 : /* Shouldn't have any parameterized paths anymore */
7729 : Assert(path->param_info == NULL);
7730 :
7731 1212 : path = (Path *) create_projection_path(root,
7732 : partially_grouped_rel,
7733 : path,
7734 1212 : partially_grouped_rel->reltarget);
7735 :
7736 1212 : add_partial_path(partially_grouped_rel, path);
7737 : }
7738 : }
7739 : }
7740 :
7741 : /*
7742 : * If there is an FDW that's responsible for all baserels of the query,
7743 : * let it consider adding partially grouped ForeignPaths.
7744 : */
7745 3858 : if (partially_grouped_rel->fdwroutine &&
7746 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7747 : {
7748 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7749 :
7750 6 : fdwroutine->GetForeignUpperPaths(root,
7751 : UPPERREL_PARTIAL_GROUP_AGG,
7752 : input_rel, partially_grouped_rel,
7753 : extra);
7754 : }
7755 :
7756 3858 : return partially_grouped_rel;
7757 : }
7758 :
7759 : /*
7760 : * make_ordered_path
7761 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7762 : * return NULL if it doesn't make sense to generate an ordered path in
7763 : * this case.
7764 : */
7765 : static Path *
7766 66402 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7767 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7768 : {
7769 : bool is_sorted;
7770 : int presorted_keys;
7771 :
7772 66402 : is_sorted = pathkeys_count_contained_in(pathkeys,
7773 : path->pathkeys,
7774 : &presorted_keys);
7775 :
7776 66402 : if (!is_sorted)
7777 : {
7778 : /*
7779 : * Try at least sorting the cheapest path and also try incrementally
7780 : * sorting any path which is partially sorted already (no need to deal
7781 : * with paths which have presorted keys when incremental sort is
7782 : * disabled unless it's the cheapest input path).
7783 : */
7784 17146 : if (path != cheapest_path &&
7785 3314 : (presorted_keys == 0 || !enable_incremental_sort))
7786 1486 : return NULL;
7787 :
7788 : /*
7789 : * We've no need to consider both a sort and incremental sort. We'll
7790 : * just do a sort if there are no presorted keys and an incremental
7791 : * sort when there are presorted keys.
7792 : */
7793 15660 : if (presorted_keys == 0 || !enable_incremental_sort)
7794 13658 : path = (Path *) create_sort_path(root,
7795 : rel,
7796 : path,
7797 : pathkeys,
7798 : limit_tuples);
7799 : else
7800 2002 : path = (Path *) create_incremental_sort_path(root,
7801 : rel,
7802 : path,
7803 : pathkeys,
7804 : presorted_keys,
7805 : limit_tuples);
7806 : }
7807 :
7808 64916 : return path;
7809 : }
7810 :
7811 : /*
7812 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7813 : * grouping relation.
7814 : *
7815 : * generate_useful_gather_paths does most of the work, but we also consider a
7816 : * special case: we could try sorting the data by the group_pathkeys and then
7817 : * applying Gather Merge.
7818 : *
7819 : * NB: This function shouldn't be used for anything other than a grouped or
7820 : * partially grouped relation not only because of the fact that it explicitly
7821 : * references group_pathkeys but we pass "true" as the third argument to
7822 : * generate_useful_gather_paths().
7823 : */
7824 : static void
7825 3090 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7826 : {
7827 : ListCell *lc;
7828 : Path *cheapest_partial_path;
7829 : List *groupby_pathkeys;
7830 :
7831 : /*
7832 : * This occurs after any partial aggregation has taken place, so trim off
7833 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7834 : */
7835 3090 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7836 18 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7837 : root->num_groupby_pathkeys);
7838 : else
7839 3072 : groupby_pathkeys = root->group_pathkeys;
7840 :
7841 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7842 3090 : generate_useful_gather_paths(root, rel, true);
7843 :
7844 3090 : cheapest_partial_path = linitial(rel->partial_pathlist);
7845 :
7846 : /* XXX Shouldn't this also consider the group-key-reordering? */
7847 7308 : foreach(lc, rel->partial_pathlist)
7848 : {
7849 4218 : Path *path = (Path *) lfirst(lc);
7850 : bool is_sorted;
7851 : int presorted_keys;
7852 : double total_groups;
7853 :
7854 4218 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7855 : path->pathkeys,
7856 : &presorted_keys);
7857 :
7858 4218 : if (is_sorted)
7859 2760 : continue;
7860 :
7861 : /*
7862 : * Try at least sorting the cheapest path and also try incrementally
7863 : * sorting any path which is partially sorted already (no need to deal
7864 : * with paths which have presorted keys when incremental sort is
7865 : * disabled unless it's the cheapest input path).
7866 : */
7867 1458 : if (path != cheapest_partial_path &&
7868 0 : (presorted_keys == 0 || !enable_incremental_sort))
7869 0 : continue;
7870 :
7871 : /*
7872 : * We've no need to consider both a sort and incremental sort. We'll
7873 : * just do a sort if there are no presorted keys and an incremental
7874 : * sort when there are presorted keys.
7875 : */
7876 1458 : if (presorted_keys == 0 || !enable_incremental_sort)
7877 1458 : path = (Path *) create_sort_path(root, rel, path,
7878 : groupby_pathkeys,
7879 : -1.0);
7880 : else
7881 0 : path = (Path *) create_incremental_sort_path(root,
7882 : rel,
7883 : path,
7884 : groupby_pathkeys,
7885 : presorted_keys,
7886 : -1.0);
7887 1458 : total_groups = compute_gather_rows(path);
7888 : path = (Path *)
7889 1458 : create_gather_merge_path(root,
7890 : rel,
7891 : path,
7892 1458 : rel->reltarget,
7893 : groupby_pathkeys,
7894 : NULL,
7895 : &total_groups);
7896 :
7897 1458 : add_path(rel, path);
7898 : }
7899 3090 : }
7900 :
7901 : /*
7902 : * can_partial_agg
7903 : *
7904 : * Determines whether or not partial grouping and/or aggregation is possible.
7905 : * Returns true when possible, false otherwise.
7906 : */
7907 : static bool
7908 45926 : can_partial_agg(PlannerInfo *root)
7909 : {
7910 45926 : Query *parse = root->parse;
7911 :
7912 45926 : if (!parse->hasAggs && parse->groupClause == NIL)
7913 : {
7914 : /*
7915 : * We don't know how to do parallel aggregation unless we have either
7916 : * some aggregates or a grouping clause.
7917 : */
7918 0 : return false;
7919 : }
7920 45926 : else if (parse->groupingSets)
7921 : {
7922 : /* We don't know how to do grouping sets in parallel. */
7923 996 : return false;
7924 : }
7925 44930 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7926 : {
7927 : /* Insufficient support for partial mode. */
7928 3896 : return false;
7929 : }
7930 :
7931 : /* Everything looks good. */
7932 41034 : return true;
7933 : }
7934 :
7935 : /*
7936 : * apply_scanjoin_target_to_paths
7937 : *
7938 : * Adjust the final scan/join relation, and recursively all of its children,
7939 : * to generate the final scan/join target. It would be more correct to model
7940 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7941 : * for each child relation, but doing it this way is noticeably cheaper.
7942 : * Maybe that problem can be solved at some point, but for now we do this.
7943 : *
7944 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7945 : * the same expressions as the existing reltarget, so we need only insert the
7946 : * appropriate sortgroupref information. By avoiding the creation of
7947 : * projection paths we save effort both immediately and at plan creation time.
7948 : */
7949 : static void
7950 569180 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7951 : RelOptInfo *rel,
7952 : List *scanjoin_targets,
7953 : List *scanjoin_targets_contain_srfs,
7954 : bool scanjoin_target_parallel_safe,
7955 : bool tlist_same_exprs)
7956 : {
7957 569180 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7958 : PathTarget *scanjoin_target;
7959 : ListCell *lc;
7960 :
7961 : /* This recurses, so be paranoid. */
7962 569180 : check_stack_depth();
7963 :
7964 : /*
7965 : * If the rel only has Append and MergeAppend paths, we want to drop its
7966 : * existing paths and generate new ones. This function would still be
7967 : * correct if we kept the existing paths: we'd modify them to generate the
7968 : * correct target above the partitioning Append, and then they'd compete
7969 : * on cost with paths generating the target below the Append. However, in
7970 : * our current cost model the latter way is always the same or cheaper
7971 : * cost, so modifying the existing paths would just be useless work.
7972 : * Moreover, when the cost is the same, varying roundoff errors might
7973 : * sometimes allow an existing path to be picked, resulting in undesirable
7974 : * cross-platform plan variations. So we drop old paths and thereby force
7975 : * the work to be done below the Append.
7976 : *
7977 : * However, there are several cases when this optimization is not safe. If
7978 : * the rel isn't partitioned, then none of the paths will be Append or
7979 : * MergeAppend paths, so we should definitely not do this. If it is
7980 : * partitioned but is a joinrel, it may have Append and MergeAppend paths,
7981 : * but it can also have join paths that we can't afford to discard.
7982 : *
7983 : * Some care is needed, because we have to allow
7984 : * generate_useful_gather_paths to see the old partial paths in the next
7985 : * stanza. Hence, zap the main pathlist here, then allow
7986 : * generate_useful_gather_paths to add path(s) to the main list, and
7987 : * finally zap the partial pathlist.
7988 : */
7989 569180 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
7990 11710 : rel->pathlist = NIL;
7991 :
7992 : /*
7993 : * If the scan/join target is not parallel-safe, partial paths cannot
7994 : * generate it.
7995 : */
7996 569180 : if (!scanjoin_target_parallel_safe)
7997 : {
7998 : /*
7999 : * Since we can't generate the final scan/join target in parallel
8000 : * workers, this is our last opportunity to use any partial paths that
8001 : * exist; so build Gather path(s) that use them and emit whatever the
8002 : * current reltarget is. We don't do this in the case where the
8003 : * target is parallel-safe, since we will be able to generate superior
8004 : * paths by doing it after the final scan/join target has been
8005 : * applied.
8006 : */
8007 81200 : generate_useful_gather_paths(root, rel, false);
8008 :
8009 : /* Can't use parallel query above this level. */
8010 81200 : rel->partial_pathlist = NIL;
8011 81200 : rel->consider_parallel = false;
8012 : }
8013 :
8014 : /* Finish dropping old paths for a partitioned rel, per comment above */
8015 569180 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8016 11710 : rel->partial_pathlist = NIL;
8017 :
8018 : /* Extract SRF-free scan/join target. */
8019 569180 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
8020 :
8021 : /*
8022 : * Apply the SRF-free scan/join target to each existing path.
8023 : *
8024 : * If the tlist exprs are the same, we can just inject the sortgroupref
8025 : * information into the existing pathtargets. Otherwise, replace each
8026 : * path with a projection path that generates the SRF-free scan/join
8027 : * target. This can't change the ordering of paths within rel->pathlist,
8028 : * so we just modify the list in place.
8029 : */
8030 1182270 : foreach(lc, rel->pathlist)
8031 : {
8032 613090 : Path *subpath = (Path *) lfirst(lc);
8033 :
8034 : /* Shouldn't have any parameterized paths anymore */
8035 : Assert(subpath->param_info == NULL);
8036 :
8037 613090 : if (tlist_same_exprs)
8038 216440 : subpath->pathtarget->sortgrouprefs =
8039 216440 : scanjoin_target->sortgrouprefs;
8040 : else
8041 : {
8042 : Path *newpath;
8043 :
8044 396650 : newpath = (Path *) create_projection_path(root, rel, subpath,
8045 : scanjoin_target);
8046 396650 : lfirst(lc) = newpath;
8047 : }
8048 : }
8049 :
8050 : /* Likewise adjust the targets for any partial paths. */
8051 594290 : foreach(lc, rel->partial_pathlist)
8052 : {
8053 25110 : Path *subpath = (Path *) lfirst(lc);
8054 :
8055 : /* Shouldn't have any parameterized paths anymore */
8056 : Assert(subpath->param_info == NULL);
8057 :
8058 25110 : if (tlist_same_exprs)
8059 20106 : subpath->pathtarget->sortgrouprefs =
8060 20106 : scanjoin_target->sortgrouprefs;
8061 : else
8062 : {
8063 : Path *newpath;
8064 :
8065 5004 : newpath = (Path *) create_projection_path(root, rel, subpath,
8066 : scanjoin_target);
8067 5004 : lfirst(lc) = newpath;
8068 : }
8069 : }
8070 :
8071 : /*
8072 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8073 : * atop each existing path. (Note that this function doesn't look at the
8074 : * cheapest-path fields, which is a good thing because they're bogus right
8075 : * now.)
8076 : */
8077 569180 : if (root->parse->hasTargetSRFs)
8078 12084 : adjust_paths_for_srfs(root, rel,
8079 : scanjoin_targets,
8080 : scanjoin_targets_contain_srfs);
8081 :
8082 : /*
8083 : * Update the rel's target to be the final (with SRFs) scan/join target.
8084 : * This now matches the actual output of all the paths, and we might get
8085 : * confused in createplan.c if they don't agree. We must do this now so
8086 : * that any append paths made in the next part will use the correct
8087 : * pathtarget (cf. create_append_path).
8088 : *
8089 : * Note that this is also necessary if GetForeignUpperPaths() gets called
8090 : * on the final scan/join relation or on any of its children, since the
8091 : * FDW might look at the rel's target to create ForeignPaths.
8092 : */
8093 569180 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8094 :
8095 : /*
8096 : * If the relation is partitioned, recursively apply the scan/join target
8097 : * to all partitions, and generate brand-new Append paths in which the
8098 : * scan/join target is computed below the Append rather than above it.
8099 : * Since Append is not projection-capable, that might save a separate
8100 : * Result node, and it also is important for partitionwise aggregate.
8101 : */
8102 569180 : if (rel_is_partitioned)
8103 : {
8104 13276 : List *live_children = NIL;
8105 : int i;
8106 :
8107 : /* Adjust each partition. */
8108 13276 : i = -1;
8109 38006 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8110 : {
8111 24730 : RelOptInfo *child_rel = rel->part_rels[i];
8112 : AppendRelInfo **appinfos;
8113 : int nappinfos;
8114 24730 : List *child_scanjoin_targets = NIL;
8115 :
8116 : Assert(child_rel != NULL);
8117 :
8118 : /* Dummy children can be ignored. */
8119 24730 : if (IS_DUMMY_REL(child_rel))
8120 42 : continue;
8121 :
8122 : /* Translate scan/join targets for this child. */
8123 24688 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8124 : &nappinfos);
8125 49376 : foreach(lc, scanjoin_targets)
8126 : {
8127 24688 : PathTarget *target = lfirst_node(PathTarget, lc);
8128 :
8129 24688 : target = copy_pathtarget(target);
8130 24688 : target->exprs = (List *)
8131 24688 : adjust_appendrel_attrs(root,
8132 24688 : (Node *) target->exprs,
8133 : nappinfos, appinfos);
8134 24688 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8135 : target);
8136 : }
8137 24688 : pfree(appinfos);
8138 :
8139 : /* Recursion does the real work. */
8140 24688 : apply_scanjoin_target_to_paths(root, child_rel,
8141 : child_scanjoin_targets,
8142 : scanjoin_targets_contain_srfs,
8143 : scanjoin_target_parallel_safe,
8144 : tlist_same_exprs);
8145 :
8146 : /* Save non-dummy children for Append paths. */
8147 24688 : if (!IS_DUMMY_REL(child_rel))
8148 24688 : live_children = lappend(live_children, child_rel);
8149 : }
8150 :
8151 : /* Build new paths for this relation by appending child paths. */
8152 13276 : add_paths_to_append_rel(root, rel, live_children);
8153 : }
8154 :
8155 : /*
8156 : * Consider generating Gather or Gather Merge paths. We must only do this
8157 : * if the relation is parallel safe, and we don't do it for child rels to
8158 : * avoid creating multiple Gather nodes within the same plan. We must do
8159 : * this after all paths have been generated and before set_cheapest, since
8160 : * one of the generated paths may turn out to be the cheapest one.
8161 : */
8162 569180 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
8163 181086 : generate_useful_gather_paths(root, rel, false);
8164 :
8165 : /*
8166 : * Reassess which paths are the cheapest, now that we've potentially added
8167 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8168 : * this relation.
8169 : */
8170 569180 : set_cheapest(rel);
8171 569180 : }
8172 :
8173 : /*
8174 : * create_partitionwise_grouping_paths
8175 : *
8176 : * If the partition keys of input relation are part of the GROUP BY clause, all
8177 : * the rows belonging to a given group come from a single partition. This
8178 : * allows aggregation/grouping over a partitioned relation to be broken down
8179 : * into aggregation/grouping on each partition. This should be no worse, and
8180 : * often better, than the normal approach.
8181 : *
8182 : * However, if the GROUP BY clause does not contain all the partition keys,
8183 : * rows from a given group may be spread across multiple partitions. In that
8184 : * case, we perform partial aggregation for each group, append the results,
8185 : * and then finalize aggregation. This is less certain to win than the
8186 : * previous case. It may win if the PartialAggregate stage greatly reduces
8187 : * the number of groups, because fewer rows will pass through the Append node.
8188 : * It may lose if we have lots of small groups.
8189 : */
8190 : static void
8191 826 : create_partitionwise_grouping_paths(PlannerInfo *root,
8192 : RelOptInfo *input_rel,
8193 : RelOptInfo *grouped_rel,
8194 : RelOptInfo *partially_grouped_rel,
8195 : const AggClauseCosts *agg_costs,
8196 : grouping_sets_data *gd,
8197 : PartitionwiseAggregateType patype,
8198 : GroupPathExtraData *extra)
8199 : {
8200 826 : List *grouped_live_children = NIL;
8201 826 : List *partially_grouped_live_children = NIL;
8202 826 : PathTarget *target = grouped_rel->reltarget;
8203 826 : bool partial_grouping_valid = true;
8204 : int i;
8205 :
8206 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8207 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8208 : partially_grouped_rel != NULL);
8209 :
8210 : /* Add paths for partitionwise aggregation/grouping. */
8211 826 : i = -1;
8212 2992 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8213 : {
8214 2166 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8215 : PathTarget *child_target;
8216 : AppendRelInfo **appinfos;
8217 : int nappinfos;
8218 : GroupPathExtraData child_extra;
8219 : RelOptInfo *child_grouped_rel;
8220 : RelOptInfo *child_partially_grouped_rel;
8221 :
8222 : Assert(child_input_rel != NULL);
8223 :
8224 : /* Dummy children can be ignored. */
8225 2166 : if (IS_DUMMY_REL(child_input_rel))
8226 0 : continue;
8227 :
8228 2166 : child_target = copy_pathtarget(target);
8229 :
8230 : /*
8231 : * Copy the given "extra" structure as is and then override the
8232 : * members specific to this child.
8233 : */
8234 2166 : memcpy(&child_extra, extra, sizeof(child_extra));
8235 :
8236 2166 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8237 : &nappinfos);
8238 :
8239 2166 : child_target->exprs = (List *)
8240 2166 : adjust_appendrel_attrs(root,
8241 2166 : (Node *) target->exprs,
8242 : nappinfos, appinfos);
8243 :
8244 : /* Translate havingQual and targetList. */
8245 2166 : child_extra.havingQual = (Node *)
8246 : adjust_appendrel_attrs(root,
8247 : extra->havingQual,
8248 : nappinfos, appinfos);
8249 2166 : child_extra.targetList = (List *)
8250 2166 : adjust_appendrel_attrs(root,
8251 2166 : (Node *) extra->targetList,
8252 : nappinfos, appinfos);
8253 :
8254 : /*
8255 : * extra->patype was the value computed for our parent rel; patype is
8256 : * the value for this relation. For the child, our value is its
8257 : * parent rel's value.
8258 : */
8259 2166 : child_extra.patype = patype;
8260 :
8261 : /*
8262 : * Create grouping relation to hold fully aggregated grouping and/or
8263 : * aggregation paths for the child.
8264 : */
8265 2166 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8266 : child_target,
8267 2166 : extra->target_parallel_safe,
8268 : child_extra.havingQual);
8269 :
8270 : /* Create grouping paths for this child relation. */
8271 2166 : create_ordinary_grouping_paths(root, child_input_rel,
8272 : child_grouped_rel,
8273 : agg_costs, gd, &child_extra,
8274 : &child_partially_grouped_rel);
8275 :
8276 2166 : if (child_partially_grouped_rel)
8277 : {
8278 : partially_grouped_live_children =
8279 1542 : lappend(partially_grouped_live_children,
8280 : child_partially_grouped_rel);
8281 : }
8282 : else
8283 624 : partial_grouping_valid = false;
8284 :
8285 2166 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8286 : {
8287 1308 : set_cheapest(child_grouped_rel);
8288 1308 : grouped_live_children = lappend(grouped_live_children,
8289 : child_grouped_rel);
8290 : }
8291 :
8292 2166 : pfree(appinfos);
8293 : }
8294 :
8295 : /*
8296 : * Try to create append paths for partially grouped children. For full
8297 : * partitionwise aggregation, we might have paths in the partial_pathlist
8298 : * if parallel aggregation is possible. For partial partitionwise
8299 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8300 : *
8301 : * NB: We must have a partially grouped path for every child in order to
8302 : * generate a partially grouped path for this relation.
8303 : */
8304 826 : if (partially_grouped_rel && partial_grouping_valid)
8305 : {
8306 : Assert(partially_grouped_live_children != NIL);
8307 :
8308 602 : add_paths_to_append_rel(root, partially_grouped_rel,
8309 : partially_grouped_live_children);
8310 : }
8311 :
8312 : /* If possible, create append paths for fully grouped children. */
8313 826 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8314 : {
8315 : Assert(grouped_live_children != NIL);
8316 :
8317 488 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8318 : }
8319 826 : }
8320 :
8321 : /*
8322 : * group_by_has_partkey
8323 : *
8324 : * Returns true if all the partition keys of the given relation are part of
8325 : * the GROUP BY clauses, including having matching collation, false otherwise.
8326 : */
8327 : static bool
8328 772 : group_by_has_partkey(RelOptInfo *input_rel,
8329 : List *targetList,
8330 : List *groupClause)
8331 : {
8332 772 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8333 772 : int cnt = 0;
8334 : int partnatts;
8335 :
8336 : /* Input relation should be partitioned. */
8337 : Assert(input_rel->part_scheme);
8338 :
8339 : /* Rule out early, if there are no partition keys present. */
8340 772 : if (!input_rel->partexprs)
8341 0 : return false;
8342 :
8343 772 : partnatts = input_rel->part_scheme->partnatts;
8344 :
8345 1296 : for (cnt = 0; cnt < partnatts; cnt++)
8346 : {
8347 808 : List *partexprs = input_rel->partexprs[cnt];
8348 : ListCell *lc;
8349 808 : bool found = false;
8350 :
8351 1206 : foreach(lc, partexprs)
8352 : {
8353 : ListCell *lg;
8354 934 : Expr *partexpr = lfirst(lc);
8355 934 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8356 :
8357 1452 : foreach(lg, groupexprs)
8358 : {
8359 1054 : Expr *groupexpr = lfirst(lg);
8360 1054 : Oid groupcoll = exprCollation((Node *) groupexpr);
8361 :
8362 : /*
8363 : * Note: we can assume there is at most one RelabelType node;
8364 : * eval_const_expressions() will have simplified if more than
8365 : * one.
8366 : */
8367 1054 : if (IsA(groupexpr, RelabelType))
8368 24 : groupexpr = ((RelabelType *) groupexpr)->arg;
8369 :
8370 1054 : if (equal(groupexpr, partexpr))
8371 : {
8372 : /*
8373 : * Reject a match if the grouping collation does not match
8374 : * the partitioning collation.
8375 : */
8376 536 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8377 : partcoll != groupcoll)
8378 12 : return false;
8379 :
8380 524 : found = true;
8381 524 : break;
8382 : }
8383 : }
8384 :
8385 922 : if (found)
8386 524 : break;
8387 : }
8388 :
8389 : /*
8390 : * If none of the partition key expressions match with any of the
8391 : * GROUP BY expression, return false.
8392 : */
8393 796 : if (!found)
8394 272 : return false;
8395 : }
8396 :
8397 488 : return true;
8398 : }
8399 :
8400 : /*
8401 : * generate_setop_child_grouplist
8402 : * Build a SortGroupClause list defining the sort/grouping properties
8403 : * of the child of a set operation.
8404 : *
8405 : * This is similar to generate_setop_grouplist() but differs as the setop
8406 : * child query's targetlist entries may already have a tleSortGroupRef
8407 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8408 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8409 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8410 : * any of the columns in the targetlist don't match to the setop's colTypes
8411 : * then we return an empty list. This may leave some TLEs with unreferenced
8412 : * ressortgroupref markings, but that's harmless.
8413 : */
8414 : static List *
8415 12786 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8416 : {
8417 12786 : List *grouplist = copyObject(op->groupClauses);
8418 : ListCell *lg;
8419 : ListCell *lt;
8420 : ListCell *ct;
8421 :
8422 12786 : lg = list_head(grouplist);
8423 12786 : ct = list_head(op->colTypes);
8424 49314 : foreach(lt, targetlist)
8425 : {
8426 36950 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8427 : SortGroupClause *sgc;
8428 : Oid coltype;
8429 :
8430 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8431 36950 : if (tle->resjunk)
8432 0 : continue;
8433 :
8434 : /*
8435 : * We expect every non-resjunk target to have a SortGroupClause and
8436 : * colTypes.
8437 : */
8438 : Assert(lg != NULL);
8439 : Assert(ct != NULL);
8440 36950 : sgc = (SortGroupClause *) lfirst(lg);
8441 36950 : coltype = lfirst_oid(ct);
8442 :
8443 : /* reject if target type isn't the same as the setop target type */
8444 36950 : if (coltype != exprType((Node *) tle->expr))
8445 422 : return NIL;
8446 :
8447 36528 : lg = lnext(grouplist, lg);
8448 36528 : ct = lnext(op->colTypes, ct);
8449 :
8450 : /* assign a tleSortGroupRef, or reuse the existing one */
8451 36528 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8452 : }
8453 :
8454 : Assert(lg == NULL);
8455 : Assert(ct == NULL);
8456 :
8457 12364 : return grouplist;
8458 : }
8459 :
8460 : /*
8461 : * create_unique_paths
8462 : * Build a new RelOptInfo containing Paths that represent elimination of
8463 : * distinct rows from the input data. Distinct-ness is defined according to
8464 : * the needs of the semijoin represented by sjinfo. If it is not possible
8465 : * to identify how to make the data unique, NULL is returned.
8466 : *
8467 : * If used at all, this is likely to be called repeatedly on the same rel,
8468 : * so we cache the result.
8469 : */
8470 : RelOptInfo *
8471 8926 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8472 : {
8473 : RelOptInfo *unique_rel;
8474 8926 : List *sortPathkeys = NIL;
8475 8926 : List *groupClause = NIL;
8476 : MemoryContext oldcontext;
8477 :
8478 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8479 : Assert(sjinfo->jointype == JOIN_SEMI);
8480 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8481 :
8482 : /* If result already cached, return it */
8483 8926 : if (rel->unique_rel)
8484 1836 : return rel->unique_rel;
8485 :
8486 : /* If it's not possible to unique-ify, return NULL */
8487 7090 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8488 132 : return NULL;
8489 :
8490 : /*
8491 : * Punt if this is a child relation and we failed to build a unique-ified
8492 : * relation for its parent. This can happen if all the RHS columns were
8493 : * found to be equated to constants when unique-ifying the parent table,
8494 : * leaving no columns to unique-ify.
8495 : */
8496 6958 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8497 12 : return NULL;
8498 :
8499 : /*
8500 : * When called during GEQO join planning, we are in a short-lived memory
8501 : * context. We must make sure that the unique rel and any subsidiary data
8502 : * structures created for a baserel survive the GEQO cycle, else the
8503 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8504 : * are creating those for a joinrel during GEQO, we don't want them to
8505 : * clutter the main planning context. Upshot is that the best solution is
8506 : * to explicitly allocate memory in the same context the given RelOptInfo
8507 : * is in.
8508 : */
8509 6946 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8510 :
8511 6946 : unique_rel = makeNode(RelOptInfo);
8512 6946 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8513 :
8514 : /*
8515 : * clear path info
8516 : */
8517 6946 : unique_rel->pathlist = NIL;
8518 6946 : unique_rel->ppilist = NIL;
8519 6946 : unique_rel->partial_pathlist = NIL;
8520 6946 : unique_rel->cheapest_startup_path = NULL;
8521 6946 : unique_rel->cheapest_total_path = NULL;
8522 6946 : unique_rel->cheapest_parameterized_paths = NIL;
8523 :
8524 : /*
8525 : * Build the target list for the unique rel. We also build the pathkeys
8526 : * that represent the ordering requirements for the sort-based
8527 : * implementation, and the list of SortGroupClause nodes that represent
8528 : * the columns to be grouped on for the hash-based implementation.
8529 : *
8530 : * For a child rel, we can construct these fields from those of its
8531 : * parent.
8532 : */
8533 6946 : if (IS_OTHER_REL(rel))
8534 432 : {
8535 : PathTarget *child_unique_target;
8536 : PathTarget *parent_unique_target;
8537 :
8538 432 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8539 :
8540 432 : child_unique_target = copy_pathtarget(parent_unique_target);
8541 :
8542 : /* Translate the target expressions */
8543 432 : child_unique_target->exprs = (List *)
8544 432 : adjust_appendrel_attrs_multilevel(root,
8545 432 : (Node *) parent_unique_target->exprs,
8546 : rel,
8547 432 : rel->top_parent);
8548 :
8549 432 : unique_rel->reltarget = child_unique_target;
8550 :
8551 432 : sortPathkeys = rel->top_parent->unique_pathkeys;
8552 432 : groupClause = rel->top_parent->unique_groupclause;
8553 : }
8554 : else
8555 : {
8556 : List *newtlist;
8557 : int nextresno;
8558 6514 : List *sortList = NIL;
8559 : ListCell *lc1;
8560 : ListCell *lc2;
8561 :
8562 : /*
8563 : * The values we are supposed to unique-ify may be expressions in the
8564 : * variables of the input rel's targetlist. We have to add any such
8565 : * expressions to the unique rel's targetlist.
8566 : *
8567 : * To complicate matters, some of the values to be unique-ified may be
8568 : * known redundant by the EquivalenceClass machinery (e.g., because
8569 : * they have been equated to constants). There is no need to compare
8570 : * such values during unique-ification, and indeed we had better not
8571 : * try because the Vars involved may not have propagated as high as
8572 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8573 : * detect such cases, which is a tad inefficient but it doesn't seem
8574 : * worth building specialized infrastructure for this.
8575 : */
8576 6514 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8577 6514 : nextresno = list_length(newtlist) + 1;
8578 :
8579 13262 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8580 : {
8581 6748 : Expr *uniqexpr = lfirst(lc1);
8582 6748 : Oid in_oper = lfirst_oid(lc2);
8583 : Oid sortop;
8584 : TargetEntry *tle;
8585 6748 : bool made_tle = false;
8586 :
8587 6748 : tle = tlist_member(uniqexpr, newtlist);
8588 6748 : if (!tle)
8589 : {
8590 3268 : tle = makeTargetEntry(uniqexpr,
8591 : nextresno,
8592 : NULL,
8593 : false);
8594 3268 : newtlist = lappend(newtlist, tle);
8595 3268 : nextresno++;
8596 3268 : made_tle = true;
8597 : }
8598 :
8599 : /*
8600 : * Try to build an ORDER BY list to sort the input compatibly. We
8601 : * do this for each sortable clause even when the clauses are not
8602 : * all sortable, so that we can detect clauses that are redundant
8603 : * according to the pathkey machinery.
8604 : */
8605 6748 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8606 6748 : if (OidIsValid(sortop))
8607 : {
8608 : Oid eqop;
8609 : SortGroupClause *sortcl;
8610 :
8611 : /*
8612 : * The Unique node will need equality operators. Normally
8613 : * these are the same as the IN clause operators, but if those
8614 : * are cross-type operators then the equality operators are
8615 : * the ones for the IN clause operators' RHS datatype.
8616 : */
8617 6748 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8618 6748 : if (!OidIsValid(eqop)) /* shouldn't happen */
8619 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8620 : sortop);
8621 :
8622 6748 : sortcl = makeNode(SortGroupClause);
8623 6748 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8624 6748 : sortcl->eqop = eqop;
8625 6748 : sortcl->sortop = sortop;
8626 6748 : sortcl->reverse_sort = false;
8627 6748 : sortcl->nulls_first = false;
8628 6748 : sortcl->hashable = false; /* no need to make this accurate */
8629 6748 : sortList = lappend(sortList, sortcl);
8630 :
8631 : /*
8632 : * At each step, convert the SortGroupClause list to pathkey
8633 : * form. If the just-added SortGroupClause is redundant, the
8634 : * result will be shorter than the SortGroupClause list.
8635 : */
8636 6748 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8637 : newtlist);
8638 6748 : if (list_length(sortPathkeys) != list_length(sortList))
8639 : {
8640 : /* Drop the redundant SortGroupClause */
8641 2052 : sortList = list_delete_last(sortList);
8642 : Assert(list_length(sortPathkeys) == list_length(sortList));
8643 : /* Undo tlist addition, if we made one */
8644 2052 : if (made_tle)
8645 : {
8646 12 : newtlist = list_delete_last(newtlist);
8647 12 : nextresno--;
8648 : }
8649 : /* We need not consider this clause for hashing, either */
8650 2052 : continue;
8651 : }
8652 : }
8653 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8654 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8655 : in_oper);
8656 :
8657 4696 : if (sjinfo->semi_can_hash)
8658 : {
8659 : /* Create a GROUP BY list for the Agg node to use */
8660 : Oid eq_oper;
8661 : SortGroupClause *groupcl;
8662 :
8663 : /*
8664 : * Get the hashable equality operators for the Agg node to
8665 : * use. Normally these are the same as the IN clause
8666 : * operators, but if those are cross-type operators then the
8667 : * equality operators are the ones for the IN clause
8668 : * operators' RHS datatype.
8669 : */
8670 4696 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8671 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8672 : in_oper);
8673 :
8674 4696 : groupcl = makeNode(SortGroupClause);
8675 4696 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8676 4696 : groupcl->eqop = eq_oper;
8677 4696 : groupcl->sortop = sortop;
8678 4696 : groupcl->reverse_sort = false;
8679 4696 : groupcl->nulls_first = false;
8680 4696 : groupcl->hashable = true;
8681 4696 : groupClause = lappend(groupClause, groupcl);
8682 : }
8683 : }
8684 :
8685 : /*
8686 : * Done building the sortPathkeys and groupClause. But the
8687 : * sortPathkeys are bogus if not all the clauses were sortable.
8688 : */
8689 6514 : if (!sjinfo->semi_can_btree)
8690 0 : sortPathkeys = NIL;
8691 :
8692 : /*
8693 : * It can happen that all the RHS columns are equated to constants.
8694 : * We'd have to do something special to unique-ify in that case, and
8695 : * it's such an unlikely-in-the-real-world case that it's not worth
8696 : * the effort. So just punt if we found no columns to unique-ify.
8697 : */
8698 6514 : if (sortPathkeys == NIL && groupClause == NIL)
8699 : {
8700 1950 : MemoryContextSwitchTo(oldcontext);
8701 1950 : return NULL;
8702 : }
8703 :
8704 : /* Convert the required targetlist back to PathTarget form */
8705 4564 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8706 : }
8707 :
8708 : /* build unique paths based on input rel's pathlist */
8709 4996 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8710 : sjinfo, unique_rel);
8711 :
8712 : /* build unique paths based on input rel's partial_pathlist */
8713 4996 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8714 : sjinfo, unique_rel);
8715 :
8716 : /* Now choose the best path(s) */
8717 4996 : set_cheapest(unique_rel);
8718 :
8719 : /*
8720 : * There shouldn't be any partial paths for the unique relation;
8721 : * otherwise, we won't be able to properly guarantee uniqueness.
8722 : */
8723 : Assert(unique_rel->partial_pathlist == NIL);
8724 :
8725 : /* Cache the result */
8726 4996 : rel->unique_rel = unique_rel;
8727 4996 : rel->unique_pathkeys = sortPathkeys;
8728 4996 : rel->unique_groupclause = groupClause;
8729 :
8730 4996 : MemoryContextSwitchTo(oldcontext);
8731 :
8732 4996 : return unique_rel;
8733 : }
8734 :
8735 : /*
8736 : * create_final_unique_paths
8737 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8738 : */
8739 : static void
8740 8744 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8741 : List *sortPathkeys, List *groupClause,
8742 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8743 : {
8744 8744 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8745 :
8746 : /* Estimate number of output rows */
8747 8744 : unique_rel->rows = estimate_num_groups(root,
8748 : sjinfo->semi_rhs_exprs,
8749 : cheapest_input_path->rows,
8750 : NULL,
8751 : NULL);
8752 :
8753 : /* Consider sort-based implementations, if possible. */
8754 8744 : if (sjinfo->semi_can_btree)
8755 : {
8756 : ListCell *lc;
8757 :
8758 : /*
8759 : * Use any available suitably-sorted path as input, and also consider
8760 : * sorting the cheapest-total path and incremental sort on any paths
8761 : * with presorted keys.
8762 : *
8763 : * To save planning time, we ignore parameterized input paths unless
8764 : * they are the cheapest-total path.
8765 : */
8766 19056 : foreach(lc, input_rel->pathlist)
8767 : {
8768 10312 : Path *input_path = (Path *) lfirst(lc);
8769 : Path *path;
8770 : bool is_sorted;
8771 : int presorted_keys;
8772 :
8773 : /*
8774 : * Ignore parameterized paths that are not the cheapest-total
8775 : * path.
8776 : */
8777 10312 : if (input_path->param_info &&
8778 : input_path != cheapest_input_path)
8779 922 : continue;
8780 :
8781 9440 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8782 : input_path->pathkeys,
8783 : &presorted_keys);
8784 :
8785 : /*
8786 : * Ignore paths that are not suitably or partially sorted, unless
8787 : * they are the cheapest total path (no need to deal with paths
8788 : * which have presorted keys when incremental sort is disabled).
8789 : */
8790 9440 : if (!is_sorted && input_path != cheapest_input_path &&
8791 98 : (presorted_keys == 0 || !enable_incremental_sort))
8792 50 : continue;
8793 :
8794 : /*
8795 : * Make a separate ProjectionPath in case we need a Result node.
8796 : */
8797 9390 : path = (Path *) create_projection_path(root,
8798 : unique_rel,
8799 : input_path,
8800 9390 : unique_rel->reltarget);
8801 :
8802 9390 : if (!is_sorted)
8803 : {
8804 : /*
8805 : * We've no need to consider both a sort and incremental sort.
8806 : * We'll just do a sort if there are no presorted keys and an
8807 : * incremental sort when there are presorted keys.
8808 : */
8809 4982 : if (presorted_keys == 0 || !enable_incremental_sort)
8810 4934 : path = (Path *) create_sort_path(root,
8811 : unique_rel,
8812 : path,
8813 : sortPathkeys,
8814 : -1.0);
8815 : else
8816 48 : path = (Path *) create_incremental_sort_path(root,
8817 : unique_rel,
8818 : path,
8819 : sortPathkeys,
8820 : presorted_keys,
8821 : -1.0);
8822 : }
8823 :
8824 9390 : path = (Path *) create_unique_path(root, unique_rel, path,
8825 : list_length(sortPathkeys),
8826 : unique_rel->rows);
8827 :
8828 9390 : add_path(unique_rel, path);
8829 : }
8830 : }
8831 :
8832 : /* Consider hash-based implementation, if possible. */
8833 8744 : if (sjinfo->semi_can_hash)
8834 : {
8835 : Path *path;
8836 :
8837 : /*
8838 : * Make a separate ProjectionPath in case we need a Result node.
8839 : */
8840 8744 : path = (Path *) create_projection_path(root,
8841 : unique_rel,
8842 : cheapest_input_path,
8843 8744 : unique_rel->reltarget);
8844 :
8845 8744 : path = (Path *) create_agg_path(root,
8846 : unique_rel,
8847 : path,
8848 : cheapest_input_path->pathtarget,
8849 : AGG_HASHED,
8850 : AGGSPLIT_SIMPLE,
8851 : groupClause,
8852 : NIL,
8853 : NULL,
8854 : unique_rel->rows);
8855 :
8856 8744 : add_path(unique_rel, path);
8857 : }
8858 8744 : }
8859 :
8860 : /*
8861 : * create_partial_unique_paths
8862 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
8863 : */
8864 : static void
8865 4996 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8866 : List *sortPathkeys, List *groupClause,
8867 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8868 : {
8869 : RelOptInfo *partial_unique_rel;
8870 : Path *cheapest_partial_path;
8871 :
8872 : /* nothing to do when there are no partial paths in the input rel */
8873 4996 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
8874 1248 : return;
8875 :
8876 : /*
8877 : * nothing to do if there's anything in the targetlist that's
8878 : * parallel-restricted.
8879 : */
8880 3748 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
8881 0 : return;
8882 :
8883 3748 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
8884 :
8885 3748 : partial_unique_rel = makeNode(RelOptInfo);
8886 3748 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
8887 :
8888 : /*
8889 : * clear path info
8890 : */
8891 3748 : partial_unique_rel->pathlist = NIL;
8892 3748 : partial_unique_rel->ppilist = NIL;
8893 3748 : partial_unique_rel->partial_pathlist = NIL;
8894 3748 : partial_unique_rel->cheapest_startup_path = NULL;
8895 3748 : partial_unique_rel->cheapest_total_path = NULL;
8896 3748 : partial_unique_rel->cheapest_parameterized_paths = NIL;
8897 :
8898 : /* Estimate number of output rows */
8899 3748 : partial_unique_rel->rows = estimate_num_groups(root,
8900 : sjinfo->semi_rhs_exprs,
8901 : cheapest_partial_path->rows,
8902 : NULL,
8903 : NULL);
8904 3748 : partial_unique_rel->reltarget = unique_rel->reltarget;
8905 :
8906 : /* Consider sort-based implementations, if possible. */
8907 3748 : if (sjinfo->semi_can_btree)
8908 : {
8909 : ListCell *lc;
8910 :
8911 : /*
8912 : * Use any available suitably-sorted path as input, and also consider
8913 : * sorting the cheapest partial path and incremental sort on any paths
8914 : * with presorted keys.
8915 : */
8916 7808 : foreach(lc, input_rel->partial_pathlist)
8917 : {
8918 4060 : Path *input_path = (Path *) lfirst(lc);
8919 : Path *path;
8920 : bool is_sorted;
8921 : int presorted_keys;
8922 :
8923 4060 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8924 : input_path->pathkeys,
8925 : &presorted_keys);
8926 :
8927 : /*
8928 : * Ignore paths that are not suitably or partially sorted, unless
8929 : * they are the cheapest partial path (no need to deal with paths
8930 : * which have presorted keys when incremental sort is disabled).
8931 : */
8932 4060 : if (!is_sorted && input_path != cheapest_partial_path &&
8933 0 : (presorted_keys == 0 || !enable_incremental_sort))
8934 0 : continue;
8935 :
8936 : /*
8937 : * Make a separate ProjectionPath in case we need a Result node.
8938 : */
8939 4060 : path = (Path *) create_projection_path(root,
8940 : partial_unique_rel,
8941 : input_path,
8942 4060 : partial_unique_rel->reltarget);
8943 :
8944 4060 : if (!is_sorted)
8945 : {
8946 : /*
8947 : * We've no need to consider both a sort and incremental sort.
8948 : * We'll just do a sort if there are no presorted keys and an
8949 : * incremental sort when there are presorted keys.
8950 : */
8951 3700 : if (presorted_keys == 0 || !enable_incremental_sort)
8952 3700 : path = (Path *) create_sort_path(root,
8953 : partial_unique_rel,
8954 : path,
8955 : sortPathkeys,
8956 : -1.0);
8957 : else
8958 0 : path = (Path *) create_incremental_sort_path(root,
8959 : partial_unique_rel,
8960 : path,
8961 : sortPathkeys,
8962 : presorted_keys,
8963 : -1.0);
8964 : }
8965 :
8966 4060 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
8967 : list_length(sortPathkeys),
8968 : partial_unique_rel->rows);
8969 :
8970 4060 : add_partial_path(partial_unique_rel, path);
8971 : }
8972 : }
8973 :
8974 : /* Consider hash-based implementation, if possible. */
8975 3748 : if (sjinfo->semi_can_hash)
8976 : {
8977 : Path *path;
8978 :
8979 : /*
8980 : * Make a separate ProjectionPath in case we need a Result node.
8981 : */
8982 3748 : path = (Path *) create_projection_path(root,
8983 : partial_unique_rel,
8984 : cheapest_partial_path,
8985 3748 : partial_unique_rel->reltarget);
8986 :
8987 3748 : path = (Path *) create_agg_path(root,
8988 : partial_unique_rel,
8989 : path,
8990 : cheapest_partial_path->pathtarget,
8991 : AGG_HASHED,
8992 : AGGSPLIT_SIMPLE,
8993 : groupClause,
8994 : NIL,
8995 : NULL,
8996 : partial_unique_rel->rows);
8997 :
8998 3748 : add_partial_path(partial_unique_rel, path);
8999 : }
9000 :
9001 3748 : if (partial_unique_rel->partial_pathlist != NIL)
9002 : {
9003 3748 : generate_useful_gather_paths(root, partial_unique_rel, true);
9004 3748 : set_cheapest(partial_unique_rel);
9005 :
9006 : /*
9007 : * Finally, create paths to unique-ify the final result. This step is
9008 : * needed to remove any duplicates due to combining rows from parallel
9009 : * workers.
9010 : */
9011 3748 : create_final_unique_paths(root, partial_unique_rel,
9012 : sortPathkeys, groupClause,
9013 : sjinfo, unique_rel);
9014 : }
9015 : }
9016 :
9017 : /*
9018 : * Choose a unique name for some subroot.
9019 : *
9020 : * Modifies glob->subplanNames to track names already used.
9021 : */
9022 : char *
9023 85398 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
9024 : {
9025 : unsigned n;
9026 :
9027 : /*
9028 : * If a numeric suffix is not required, then search the list of
9029 : * previously-assigned names for a match. If none is found, then we can
9030 : * use the provided name without modification.
9031 : */
9032 85398 : if (!always_number)
9033 : {
9034 26700 : bool found = false;
9035 :
9036 63988 : foreach_ptr(char, subplan_name, glob->subplanNames)
9037 : {
9038 16300 : if (strcmp(subplan_name, name) == 0)
9039 : {
9040 5712 : found = true;
9041 5712 : break;
9042 : }
9043 : }
9044 :
9045 26700 : if (!found)
9046 : {
9047 : /* pstrdup here is just to avoid cast-away-const */
9048 20988 : char *chosen_name = pstrdup(name);
9049 :
9050 20988 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
9051 20988 : return chosen_name;
9052 : }
9053 : }
9054 :
9055 : /*
9056 : * If a numeric suffix is required or if the un-suffixed name is already
9057 : * in use, then loop until we find a positive integer that produces a
9058 : * novel name.
9059 : */
9060 64410 : for (n = 1; true; ++n)
9061 55434 : {
9062 119844 : char *proposed_name = psprintf("%s_%u", name, n);
9063 119844 : bool found = false;
9064 :
9065 457176 : foreach_ptr(char, subplan_name, glob->subplanNames)
9066 : {
9067 272922 : if (strcmp(subplan_name, proposed_name) == 0)
9068 : {
9069 55434 : found = true;
9070 55434 : break;
9071 : }
9072 : }
9073 :
9074 119844 : if (!found)
9075 : {
9076 64410 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9077 64410 : return proposed_name;
9078 : }
9079 :
9080 55434 : pfree(proposed_name);
9081 : }
9082 : }
|
